2019-05-19 13:07:45 +01:00
|
|
|
|
2018-07-31 13:39:35 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
2022-08-23 17:20:33 +02:00
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
|
|
|
|
|
2025-08-29 17:15:27 +01:00
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-07-28 16:51:39 +00:00
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
|
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
zswap: shrink zswap pool based on memory pressure
Currently, we only shrink the zswap pool when the user-defined limit is
hit. This means that if we set the limit too high, cold data that are
unlikely to be used again will reside in the pool, wasting precious
memory. It is hard to predict how much zswap space will be needed ahead
of time, as this depends on the workload (specifically, on factors such as
memory access patterns and compressibility of the memory pages).
This patch implements a memcg- and NUMA-aware shrinker for zswap, that is
initiated when there is memory pressure. The shrinker does not have any
parameter that must be tuned by the user, and can be opted in or out on a
per-memcg basis.
Furthermore, to make it more robust for many workloads and prevent
overshrinking (i.e evicting warm pages that might be refaulted into
memory), we build in the following heuristics:
* Estimate the number of warm pages residing in zswap, and attempt to
protect this region of the zswap LRU.
* Scale the number of freeable objects by an estimate of the memory
saving factor. The better zswap compresses the data, the fewer pages
we will evict to swap (as we will otherwise incur IO for relatively
small memory saving).
* During reclaim, if the shrinker encounters a page that is also being
brought into memory, the shrinker will cautiously terminate its
shrinking action, as this is a sign that it is touching the warmer
region of the zswap LRU.
As a proof of concept, we ran the following synthetic benchmark: build the
linux kernel in a memory-limited cgroup, and allocate some cold data in
tmpfs to see if the shrinker could write them out and improved the overall
performance. Depending on the amount of cold data generated, we observe
from 14% to 35% reduction in kernel CPU time used in the kernel builds.
[nphamcs@gmail.com: check shrinker enablement early, use less costly stat flushing]
Link: https://lkml.kernel.org/r/20231206194456.3234203-1-nphamcs@gmail.com
Link: https://lkml.kernel.org/r/20231130194023.4102148-7-nphamcs@gmail.com
Signed-off-by: Nhat Pham <nphamcs@gmail.com>
Acked-by: Johannes Weiner <hannes@cmpxchg.org>
Tested-by: Bagas Sanjaya <bagasdotme@gmail.com>
Cc: Chris Li <chrisl@kernel.org>
Cc: Dan Streetman <ddstreet@ieee.org>
Cc: Domenico Cerasuolo <cerasuolodomenico@gmail.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Roman Gushchin <roman.gushchin@linux.dev>
Cc: Seth Jennings <sjenning@redhat.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Vitaly Wool <vitaly.wool@konsulko.com>
Cc: Yosry Ahmed <yosryahmed@google.com>
Cc: Chengming Zhou <chengming.zhou@linux.dev>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-11-30 11:40:23 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
2025-08-29 17:15:27 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-01-18 09:52:09 +09:00
|
|
|
|
|
|
|
|
|
2023-01-18 09:52:10 +09:00
|
|
|
|
2023-01-18 09:52:09 +09:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-08-29 17:15:27 +01:00
|
|
|
|
|
|
|
|
|
2023-10-02 15:43:03 +02:00
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
|
|
|
|
|
2023-10-02 15:43:03 +02:00
|
|
|
|
2025-09-08 18:00:07 -07:00
|
|
|
|
2023-05-23 09:06:34 +02:00
|
|
|
|
2025-02-03 10:28:50 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-03-24 22:35:12 +01:00
|
|
|
|
2025-02-03 10:28:50 +01:00
|
|
|
|
2022-11-14 18:13:45 +01:00
|
|
|
|
2023-10-02 15:43:03 +02:00
|
|
|
|
2025-04-06 10:00:04 -07:00
|
|
|
|
2022-11-14 18:13:45 +01:00
|
|
|
|
|
|
|
|
|
2023-10-02 15:43:03 +02:00
|
|
|
|
2022-11-14 18:13:45 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-10-02 15:43:03 +02:00
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-10-02 15:43:03 +02:00
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-10-02 15:43:03 +02:00
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
2024-07-01 12:12:59 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-05-25 13:25:59 +02:00
|
|
|
|
|
|
|
|
|
2023-10-02 15:43:03 +02:00
|
|
|
|
|
|
|
|
|
2022-05-25 13:25:59 +02:00
|
|
|
|
2023-10-02 15:43:03 +02:00
|
|
|
|
2022-05-25 13:25:59 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
slab: support for compiler-assisted type-based slab cache partitioning
Rework the general infrastructure around RANDOM_KMALLOC_CACHES into more
flexible KMALLOC_PARTITION_CACHES, with the former being a partitioning
mode of the latter.
Introduce a new mode, KMALLOC_PARTITION_TYPED, which leverages a feature
available in Clang 22 and later, called "allocation tokens" via
__builtin_infer_alloc_token() [1]. Unlike KMALLOC_PARTITION_RANDOM
(formerly RANDOM_KMALLOC_CACHES), this mode deterministically assigns a
slab cache to an allocation of type T, regardless of allocation site.
The builtin __builtin_infer_alloc_token(<malloc-args>, ...) instructs
the compiler to infer an allocation type from arguments commonly passed
to memory-allocating functions and returns a type-derived token ID. The
implementation passes kmalloc-args to the builtin: the compiler performs
best-effort type inference, and then recognizes common patterns such as
`kmalloc(sizeof(T), ...)`, `kmalloc(sizeof(T) * n, ...)`, but also
`(T *)kmalloc(...)`. Where the compiler fails to infer a type the
fallback token (default: 0) is chosen.
Note: kmalloc_obj(..) APIs fix the pattern how size and result type are
expressed, and therefore ensures there's not much drift in which
patterns the compiler needs to recognize. Specifically, kmalloc_obj()
and friends expand to `(TYPE *)KMALLOC(__obj_size, GFP)`, which the
compiler recognizes via the cast to TYPE*.
Clang's default token ID calculation is described as [1]:
typehashpointersplit: This mode assigns a token ID based on the hash
of the allocated type's name, where the top half ID-space is reserved
for types that contain pointers and the bottom half for types that do
not contain pointers.
Separating pointer-containing objects from pointerless objects and data
allocations can help mitigate certain classes of memory corruption
exploits [2]: attackers who gains a buffer overflow on a primitive
buffer cannot use it to directly corrupt pointers or other critical
metadata in an object residing in a different, isolated heap region.
It is important to note that heap isolation strategies offer a
best-effort approach, and do not provide a 100% security guarantee,
albeit achievable at relatively low performance cost. Note that this
also does not prevent cross-cache attacks: while waiting for future
features like SLAB_VIRTUAL [3] to provide physical page isolation, this
feature should be deployed alongside SHUFFLE_PAGE_ALLOCATOR and
init_on_free=1 to mitigate cross-cache attacks and page-reuse attacks as
much as possible today.
With all that, my kernel (x86 defconfig) shows me a histogram of slab
cache object distribution per /proc/slabinfo (after boot):
<slab cache> <objs> <hist>
kmalloc-part-15 1465 ++++++++++++++
kmalloc-part-14 2988 +++++++++++++++++++++++++++++
kmalloc-part-13 1656 ++++++++++++++++
kmalloc-part-12 1045 ++++++++++
kmalloc-part-11 1697 ++++++++++++++++
kmalloc-part-10 1489 ++++++++++++++
kmalloc-part-09 965 +++++++++
kmalloc-part-08 710 +++++++
kmalloc-part-07 100 +
kmalloc-part-06 217 ++
kmalloc-part-05 105 +
kmalloc-part-04 4047 ++++++++++++++++++++++++++++++++++++++++
kmalloc-part-03 183 +
kmalloc-part-02 283 ++
kmalloc-part-01 316 +++
kmalloc 1422 ++++++++++++++
The above /proc/slabinfo snapshot shows me there are 6673 allocated
objects (slabs 00 - 07) that the compiler claims contain no pointers or
it was unable to infer the type of, and 12015 objects that contain
pointers (slabs 08 - 15). On a whole, this looks relatively sane.
Additionally, when I compile my kernel with -Rpass=alloc-token, which
provides diagnostics where (after dead-code elimination) type inference
failed, I see 186 allocation sites where the compiler failed to identify
a type (down from 966 when I sent the RFC [4]). Some initial review
confirms these are mostly variable sized buffers, but also include
structs with trailing flexible length arrays.
Link: https://clang.llvm.org/docs/AllocToken.html [1]
Link: https://blog.dfsec.com/ios/2025/05/30/blasting-past-ios-18/ [2]
Link: https://lwn.net/Articles/944647/ [3]
Link: https://lore.kernel.org/all/20250825154505.1558444-1-elver@google.com/ [4]
Link: https://discourse.llvm.org/t/rfc-a-framework-for-allocator-partitioning-hints/87434
Acked-by: GONG Ruiqi <gongruiqi1@huawei.com>
Co-developed-by: Harry Yoo (Oracle) <harry@kernel.org>
Signed-off-by: Harry Yoo (Oracle) <harry@kernel.org>
Signed-off-by: Marco Elver <elver@google.com>
Reviewed-by: Harry Yoo (Oracle) <harry@kernel.org>
Link: https://patch.msgid.link/20260511200136.3201646-1-elver@google.com
Signed-off-by: Vlastimil Babka (SUSE) <vbabka@kernel.org>
2026-05-11 22:00:48 +02:00
|
|
|
|
2023-10-02 15:43:03 +02:00
|
|
|
|
slab: support for compiler-assisted type-based slab cache partitioning
Rework the general infrastructure around RANDOM_KMALLOC_CACHES into more
flexible KMALLOC_PARTITION_CACHES, with the former being a partitioning
mode of the latter.
Introduce a new mode, KMALLOC_PARTITION_TYPED, which leverages a feature
available in Clang 22 and later, called "allocation tokens" via
__builtin_infer_alloc_token() [1]. Unlike KMALLOC_PARTITION_RANDOM
(formerly RANDOM_KMALLOC_CACHES), this mode deterministically assigns a
slab cache to an allocation of type T, regardless of allocation site.
The builtin __builtin_infer_alloc_token(<malloc-args>, ...) instructs
the compiler to infer an allocation type from arguments commonly passed
to memory-allocating functions and returns a type-derived token ID. The
implementation passes kmalloc-args to the builtin: the compiler performs
best-effort type inference, and then recognizes common patterns such as
`kmalloc(sizeof(T), ...)`, `kmalloc(sizeof(T) * n, ...)`, but also
`(T *)kmalloc(...)`. Where the compiler fails to infer a type the
fallback token (default: 0) is chosen.
Note: kmalloc_obj(..) APIs fix the pattern how size and result type are
expressed, and therefore ensures there's not much drift in which
patterns the compiler needs to recognize. Specifically, kmalloc_obj()
and friends expand to `(TYPE *)KMALLOC(__obj_size, GFP)`, which the
compiler recognizes via the cast to TYPE*.
Clang's default token ID calculation is described as [1]:
typehashpointersplit: This mode assigns a token ID based on the hash
of the allocated type's name, where the top half ID-space is reserved
for types that contain pointers and the bottom half for types that do
not contain pointers.
Separating pointer-containing objects from pointerless objects and data
allocations can help mitigate certain classes of memory corruption
exploits [2]: attackers who gains a buffer overflow on a primitive
buffer cannot use it to directly corrupt pointers or other critical
metadata in an object residing in a different, isolated heap region.
It is important to note that heap isolation strategies offer a
best-effort approach, and do not provide a 100% security guarantee,
albeit achievable at relatively low performance cost. Note that this
also does not prevent cross-cache attacks: while waiting for future
features like SLAB_VIRTUAL [3] to provide physical page isolation, this
feature should be deployed alongside SHUFFLE_PAGE_ALLOCATOR and
init_on_free=1 to mitigate cross-cache attacks and page-reuse attacks as
much as possible today.
With all that, my kernel (x86 defconfig) shows me a histogram of slab
cache object distribution per /proc/slabinfo (after boot):
<slab cache> <objs> <hist>
kmalloc-part-15 1465 ++++++++++++++
kmalloc-part-14 2988 +++++++++++++++++++++++++++++
kmalloc-part-13 1656 ++++++++++++++++
kmalloc-part-12 1045 ++++++++++
kmalloc-part-11 1697 ++++++++++++++++
kmalloc-part-10 1489 ++++++++++++++
kmalloc-part-09 965 +++++++++
kmalloc-part-08 710 +++++++
kmalloc-part-07 100 +
kmalloc-part-06 217 ++
kmalloc-part-05 105 +
kmalloc-part-04 4047 ++++++++++++++++++++++++++++++++++++++++
kmalloc-part-03 183 +
kmalloc-part-02 283 ++
kmalloc-part-01 316 +++
kmalloc 1422 ++++++++++++++
The above /proc/slabinfo snapshot shows me there are 6673 allocated
objects (slabs 00 - 07) that the compiler claims contain no pointers or
it was unable to infer the type of, and 12015 objects that contain
pointers (slabs 08 - 15). On a whole, this looks relatively sane.
Additionally, when I compile my kernel with -Rpass=alloc-token, which
provides diagnostics where (after dead-code elimination) type inference
failed, I see 186 allocation sites where the compiler failed to identify
a type (down from 966 when I sent the RFC [4]). Some initial review
confirms these are mostly variable sized buffers, but also include
structs with trailing flexible length arrays.
Link: https://clang.llvm.org/docs/AllocToken.html [1]
Link: https://blog.dfsec.com/ios/2025/05/30/blasting-past-ios-18/ [2]
Link: https://lwn.net/Articles/944647/ [3]
Link: https://lore.kernel.org/all/20250825154505.1558444-1-elver@google.com/ [4]
Link: https://discourse.llvm.org/t/rfc-a-framework-for-allocator-partitioning-hints/87434
Acked-by: GONG Ruiqi <gongruiqi1@huawei.com>
Co-developed-by: Harry Yoo (Oracle) <harry@kernel.org>
Signed-off-by: Harry Yoo (Oracle) <harry@kernel.org>
Signed-off-by: Marco Elver <elver@google.com>
Reviewed-by: Harry Yoo (Oracle) <harry@kernel.org>
Link: https://patch.msgid.link/20260511200136.3201646-1-elver@google.com
Signed-off-by: Vlastimil Babka (SUSE) <vbabka@kernel.org>
2026-05-11 22:00:48 +02:00
|
|
|
|
|
|
|
|
|
Randomized slab caches for kmalloc()
When exploiting memory vulnerabilities, "heap spraying" is a common
technique targeting those related to dynamic memory allocation (i.e. the
"heap"), and it plays an important role in a successful exploitation.
Basically, it is to overwrite the memory area of vulnerable object by
triggering allocation in other subsystems or modules and therefore
getting a reference to the targeted memory location. It's usable on
various types of vulnerablity including use after free (UAF), heap out-
of-bound write and etc.
There are (at least) two reasons why the heap can be sprayed: 1) generic
slab caches are shared among different subsystems and modules, and
2) dedicated slab caches could be merged with the generic ones.
Currently these two factors cannot be prevented at a low cost: the first
one is a widely used memory allocation mechanism, and shutting down slab
merging completely via `slub_nomerge` would be overkill.
To efficiently prevent heap spraying, we propose the following approach:
to create multiple copies of generic slab caches that will never be
merged, and random one of them will be used at allocation. The random
selection is based on the address of code that calls `kmalloc()`, which
means it is static at runtime (rather than dynamically determined at
each time of allocation, which could be bypassed by repeatedly spraying
in brute force). In other words, the randomness of cache selection will
be with respect to the code address rather than time, i.e. allocations
in different code paths would most likely pick different caches,
although kmalloc() at each place would use the same cache copy whenever
it is executed. In this way, the vulnerable object and memory allocated
in other subsystems and modules will (most probably) be on different
slab caches, which prevents the object from being sprayed.
Meanwhile, the static random selection is further enhanced with a
per-boot random seed, which prevents the attacker from finding a usable
kmalloc that happens to pick the same cache with the vulnerable
subsystem/module by analyzing the open source code. In other words, with
the per-boot seed, the random selection is static during each time the
system starts and runs, but not across different system startups.
The overhead of performance has been tested on a 40-core x86 server by
comparing the results of `perf bench all` between the kernels with and
without this patch based on the latest linux-next kernel, which shows
minor difference. A subset of benchmarks are listed below:
sched/ sched/ syscall/ mem/ mem/
messaging pipe basic memcpy memset
(sec) (sec) (sec) (GB/sec) (GB/sec)
control1 0.019 5.459 0.733 15.258789 51.398026
control2 0.019 5.439 0.730 16.009221 48.828125
control3 0.019 5.282 0.735 16.009221 48.828125
control_avg 0.019 5.393 0.733 15.759077 49.684759
experiment1 0.019 5.374 0.741 15.500992 46.502976
experiment2 0.019 5.440 0.746 16.276042 51.398026
experiment3 0.019 5.242 0.752 15.258789 51.398026
experiment_avg 0.019 5.352 0.746 15.678608 49.766343
The overhead of memory usage was measured by executing `free` after boot
on a QEMU VM with 1GB total memory, and as expected, it's positively
correlated with # of cache copies:
control 4 copies 8 copies 16 copies
total 969.8M 968.2M 968.2M 968.2M
used 20.0M 21.9M 24.1M 26.7M
free 936.9M 933.6M 931.4M 928.6M
available 932.2M 928.8M 926.6M 923.9M
Co-developed-by: Xiu Jianfeng <xiujianfeng@huawei.com>
Signed-off-by: Xiu Jianfeng <xiujianfeng@huawei.com>
Signed-off-by: GONG, Ruiqi <gongruiqi@huaweicloud.com>
Reviewed-by: Kees Cook <keescook@chromium.org>
Reviewed-by: Hyeonggon Yoo <42.hyeyoo@gmail.com>
Acked-by: Dennis Zhou <dennis@kernel.org> # percpu
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
2023-07-14 14:44:22 +08:00
|
|
|
|
slab: support for compiler-assisted type-based slab cache partitioning
Rework the general infrastructure around RANDOM_KMALLOC_CACHES into more
flexible KMALLOC_PARTITION_CACHES, with the former being a partitioning
mode of the latter.
Introduce a new mode, KMALLOC_PARTITION_TYPED, which leverages a feature
available in Clang 22 and later, called "allocation tokens" via
__builtin_infer_alloc_token() [1]. Unlike KMALLOC_PARTITION_RANDOM
(formerly RANDOM_KMALLOC_CACHES), this mode deterministically assigns a
slab cache to an allocation of type T, regardless of allocation site.
The builtin __builtin_infer_alloc_token(<malloc-args>, ...) instructs
the compiler to infer an allocation type from arguments commonly passed
to memory-allocating functions and returns a type-derived token ID. The
implementation passes kmalloc-args to the builtin: the compiler performs
best-effort type inference, and then recognizes common patterns such as
`kmalloc(sizeof(T), ...)`, `kmalloc(sizeof(T) * n, ...)`, but also
`(T *)kmalloc(...)`. Where the compiler fails to infer a type the
fallback token (default: 0) is chosen.
Note: kmalloc_obj(..) APIs fix the pattern how size and result type are
expressed, and therefore ensures there's not much drift in which
patterns the compiler needs to recognize. Specifically, kmalloc_obj()
and friends expand to `(TYPE *)KMALLOC(__obj_size, GFP)`, which the
compiler recognizes via the cast to TYPE*.
Clang's default token ID calculation is described as [1]:
typehashpointersplit: This mode assigns a token ID based on the hash
of the allocated type's name, where the top half ID-space is reserved
for types that contain pointers and the bottom half for types that do
not contain pointers.
Separating pointer-containing objects from pointerless objects and data
allocations can help mitigate certain classes of memory corruption
exploits [2]: attackers who gains a buffer overflow on a primitive
buffer cannot use it to directly corrupt pointers or other critical
metadata in an object residing in a different, isolated heap region.
It is important to note that heap isolation strategies offer a
best-effort approach, and do not provide a 100% security guarantee,
albeit achievable at relatively low performance cost. Note that this
also does not prevent cross-cache attacks: while waiting for future
features like SLAB_VIRTUAL [3] to provide physical page isolation, this
feature should be deployed alongside SHUFFLE_PAGE_ALLOCATOR and
init_on_free=1 to mitigate cross-cache attacks and page-reuse attacks as
much as possible today.
With all that, my kernel (x86 defconfig) shows me a histogram of slab
cache object distribution per /proc/slabinfo (after boot):
<slab cache> <objs> <hist>
kmalloc-part-15 1465 ++++++++++++++
kmalloc-part-14 2988 +++++++++++++++++++++++++++++
kmalloc-part-13 1656 ++++++++++++++++
kmalloc-part-12 1045 ++++++++++
kmalloc-part-11 1697 ++++++++++++++++
kmalloc-part-10 1489 ++++++++++++++
kmalloc-part-09 965 +++++++++
kmalloc-part-08 710 +++++++
kmalloc-part-07 100 +
kmalloc-part-06 217 ++
kmalloc-part-05 105 +
kmalloc-part-04 4047 ++++++++++++++++++++++++++++++++++++++++
kmalloc-part-03 183 +
kmalloc-part-02 283 ++
kmalloc-part-01 316 +++
kmalloc 1422 ++++++++++++++
The above /proc/slabinfo snapshot shows me there are 6673 allocated
objects (slabs 00 - 07) that the compiler claims contain no pointers or
it was unable to infer the type of, and 12015 objects that contain
pointers (slabs 08 - 15). On a whole, this looks relatively sane.
Additionally, when I compile my kernel with -Rpass=alloc-token, which
provides diagnostics where (after dead-code elimination) type inference
failed, I see 186 allocation sites where the compiler failed to identify
a type (down from 966 when I sent the RFC [4]). Some initial review
confirms these are mostly variable sized buffers, but also include
structs with trailing flexible length arrays.
Link: https://clang.llvm.org/docs/AllocToken.html [1]
Link: https://blog.dfsec.com/ios/2025/05/30/blasting-past-ios-18/ [2]
Link: https://lwn.net/Articles/944647/ [3]
Link: https://lore.kernel.org/all/20250825154505.1558444-1-elver@google.com/ [4]
Link: https://discourse.llvm.org/t/rfc-a-framework-for-allocator-partitioning-hints/87434
Acked-by: GONG Ruiqi <gongruiqi1@huawei.com>
Co-developed-by: Harry Yoo (Oracle) <harry@kernel.org>
Signed-off-by: Harry Yoo (Oracle) <harry@kernel.org>
Signed-off-by: Marco Elver <elver@google.com>
Reviewed-by: Harry Yoo (Oracle) <harry@kernel.org>
Link: https://patch.msgid.link/20260511200136.3201646-1-elver@google.com
Signed-off-by: Vlastimil Babka (SUSE) <vbabka@kernel.org>
2026-05-11 22:00:48 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Randomized slab caches for kmalloc()
When exploiting memory vulnerabilities, "heap spraying" is a common
technique targeting those related to dynamic memory allocation (i.e. the
"heap"), and it plays an important role in a successful exploitation.
Basically, it is to overwrite the memory area of vulnerable object by
triggering allocation in other subsystems or modules and therefore
getting a reference to the targeted memory location. It's usable on
various types of vulnerablity including use after free (UAF), heap out-
of-bound write and etc.
There are (at least) two reasons why the heap can be sprayed: 1) generic
slab caches are shared among different subsystems and modules, and
2) dedicated slab caches could be merged with the generic ones.
Currently these two factors cannot be prevented at a low cost: the first
one is a widely used memory allocation mechanism, and shutting down slab
merging completely via `slub_nomerge` would be overkill.
To efficiently prevent heap spraying, we propose the following approach:
to create multiple copies of generic slab caches that will never be
merged, and random one of them will be used at allocation. The random
selection is based on the address of code that calls `kmalloc()`, which
means it is static at runtime (rather than dynamically determined at
each time of allocation, which could be bypassed by repeatedly spraying
in brute force). In other words, the randomness of cache selection will
be with respect to the code address rather than time, i.e. allocations
in different code paths would most likely pick different caches,
although kmalloc() at each place would use the same cache copy whenever
it is executed. In this way, the vulnerable object and memory allocated
in other subsystems and modules will (most probably) be on different
slab caches, which prevents the object from being sprayed.
Meanwhile, the static random selection is further enhanced with a
per-boot random seed, which prevents the attacker from finding a usable
kmalloc that happens to pick the same cache with the vulnerable
subsystem/module by analyzing the open source code. In other words, with
the per-boot seed, the random selection is static during each time the
system starts and runs, but not across different system startups.
The overhead of performance has been tested on a 40-core x86 server by
comparing the results of `perf bench all` between the kernels with and
without this patch based on the latest linux-next kernel, which shows
minor difference. A subset of benchmarks are listed below:
sched/ sched/ syscall/ mem/ mem/
messaging pipe basic memcpy memset
(sec) (sec) (sec) (GB/sec) (GB/sec)
control1 0.019 5.459 0.733 15.258789 51.398026
control2 0.019 5.439 0.730 16.009221 48.828125
control3 0.019 5.282 0.735 16.009221 48.828125
control_avg 0.019 5.393 0.733 15.759077 49.684759
experiment1 0.019 5.374 0.741 15.500992 46.502976
experiment2 0.019 5.440 0.746 16.276042 51.398026
experiment3 0.019 5.242 0.752 15.258789 51.398026
experiment_avg 0.019 5.352 0.746 15.678608 49.766343
The overhead of memory usage was measured by executing `free` after boot
on a QEMU VM with 1GB total memory, and as expected, it's positively
correlated with # of cache copies:
control 4 copies 8 copies 16 copies
total 969.8M 968.2M 968.2M 968.2M
used 20.0M 21.9M 24.1M 26.7M
free 936.9M 933.6M 931.4M 928.6M
available 932.2M 928.8M 926.6M 923.9M
Co-developed-by: Xiu Jianfeng <xiujianfeng@huawei.com>
Signed-off-by: Xiu Jianfeng <xiujianfeng@huawei.com>
Signed-off-by: GONG, Ruiqi <gongruiqi@huaweicloud.com>
Reviewed-by: Kees Cook <keescook@chromium.org>
Reviewed-by: Hyeonggon Yoo <42.hyeyoo@gmail.com>
Acked-by: Dennis Zhou <dennis@kernel.org> # percpu
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
2023-07-14 14:44:22 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
slab: support for compiler-assisted type-based slab cache partitioning
Rework the general infrastructure around RANDOM_KMALLOC_CACHES into more
flexible KMALLOC_PARTITION_CACHES, with the former being a partitioning
mode of the latter.
Introduce a new mode, KMALLOC_PARTITION_TYPED, which leverages a feature
available in Clang 22 and later, called "allocation tokens" via
__builtin_infer_alloc_token() [1]. Unlike KMALLOC_PARTITION_RANDOM
(formerly RANDOM_KMALLOC_CACHES), this mode deterministically assigns a
slab cache to an allocation of type T, regardless of allocation site.
The builtin __builtin_infer_alloc_token(<malloc-args>, ...) instructs
the compiler to infer an allocation type from arguments commonly passed
to memory-allocating functions and returns a type-derived token ID. The
implementation passes kmalloc-args to the builtin: the compiler performs
best-effort type inference, and then recognizes common patterns such as
`kmalloc(sizeof(T), ...)`, `kmalloc(sizeof(T) * n, ...)`, but also
`(T *)kmalloc(...)`. Where the compiler fails to infer a type the
fallback token (default: 0) is chosen.
Note: kmalloc_obj(..) APIs fix the pattern how size and result type are
expressed, and therefore ensures there's not much drift in which
patterns the compiler needs to recognize. Specifically, kmalloc_obj()
and friends expand to `(TYPE *)KMALLOC(__obj_size, GFP)`, which the
compiler recognizes via the cast to TYPE*.
Clang's default token ID calculation is described as [1]:
typehashpointersplit: This mode assigns a token ID based on the hash
of the allocated type's name, where the top half ID-space is reserved
for types that contain pointers and the bottom half for types that do
not contain pointers.
Separating pointer-containing objects from pointerless objects and data
allocations can help mitigate certain classes of memory corruption
exploits [2]: attackers who gains a buffer overflow on a primitive
buffer cannot use it to directly corrupt pointers or other critical
metadata in an object residing in a different, isolated heap region.
It is important to note that heap isolation strategies offer a
best-effort approach, and do not provide a 100% security guarantee,
albeit achievable at relatively low performance cost. Note that this
also does not prevent cross-cache attacks: while waiting for future
features like SLAB_VIRTUAL [3] to provide physical page isolation, this
feature should be deployed alongside SHUFFLE_PAGE_ALLOCATOR and
init_on_free=1 to mitigate cross-cache attacks and page-reuse attacks as
much as possible today.
With all that, my kernel (x86 defconfig) shows me a histogram of slab
cache object distribution per /proc/slabinfo (after boot):
<slab cache> <objs> <hist>
kmalloc-part-15 1465 ++++++++++++++
kmalloc-part-14 2988 +++++++++++++++++++++++++++++
kmalloc-part-13 1656 ++++++++++++++++
kmalloc-part-12 1045 ++++++++++
kmalloc-part-11 1697 ++++++++++++++++
kmalloc-part-10 1489 ++++++++++++++
kmalloc-part-09 965 +++++++++
kmalloc-part-08 710 +++++++
kmalloc-part-07 100 +
kmalloc-part-06 217 ++
kmalloc-part-05 105 +
kmalloc-part-04 4047 ++++++++++++++++++++++++++++++++++++++++
kmalloc-part-03 183 +
kmalloc-part-02 283 ++
kmalloc-part-01 316 +++
kmalloc 1422 ++++++++++++++
The above /proc/slabinfo snapshot shows me there are 6673 allocated
objects (slabs 00 - 07) that the compiler claims contain no pointers or
it was unable to infer the type of, and 12015 objects that contain
pointers (slabs 08 - 15). On a whole, this looks relatively sane.
Additionally, when I compile my kernel with -Rpass=alloc-token, which
provides diagnostics where (after dead-code elimination) type inference
failed, I see 186 allocation sites where the compiler failed to identify
a type (down from 966 when I sent the RFC [4]). Some initial review
confirms these are mostly variable sized buffers, but also include
structs with trailing flexible length arrays.
Link: https://clang.llvm.org/docs/AllocToken.html [1]
Link: https://blog.dfsec.com/ios/2025/05/30/blasting-past-ios-18/ [2]
Link: https://lwn.net/Articles/944647/ [3]
Link: https://lore.kernel.org/all/20250825154505.1558444-1-elver@google.com/ [4]
Link: https://discourse.llvm.org/t/rfc-a-framework-for-allocator-partitioning-hints/87434
Acked-by: GONG Ruiqi <gongruiqi1@huawei.com>
Co-developed-by: Harry Yoo (Oracle) <harry@kernel.org>
Signed-off-by: Harry Yoo (Oracle) <harry@kernel.org>
Signed-off-by: Marco Elver <elver@google.com>
Reviewed-by: Harry Yoo (Oracle) <harry@kernel.org>
Link: https://patch.msgid.link/20260511200136.3201646-1-elver@google.com
Signed-off-by: Vlastimil Babka (SUSE) <vbabka@kernel.org>
2026-05-11 22:00:48 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Randomized slab caches for kmalloc()
When exploiting memory vulnerabilities, "heap spraying" is a common
technique targeting those related to dynamic memory allocation (i.e. the
"heap"), and it plays an important role in a successful exploitation.
Basically, it is to overwrite the memory area of vulnerable object by
triggering allocation in other subsystems or modules and therefore
getting a reference to the targeted memory location. It's usable on
various types of vulnerablity including use after free (UAF), heap out-
of-bound write and etc.
There are (at least) two reasons why the heap can be sprayed: 1) generic
slab caches are shared among different subsystems and modules, and
2) dedicated slab caches could be merged with the generic ones.
Currently these two factors cannot be prevented at a low cost: the first
one is a widely used memory allocation mechanism, and shutting down slab
merging completely via `slub_nomerge` would be overkill.
To efficiently prevent heap spraying, we propose the following approach:
to create multiple copies of generic slab caches that will never be
merged, and random one of them will be used at allocation. The random
selection is based on the address of code that calls `kmalloc()`, which
means it is static at runtime (rather than dynamically determined at
each time of allocation, which could be bypassed by repeatedly spraying
in brute force). In other words, the randomness of cache selection will
be with respect to the code address rather than time, i.e. allocations
in different code paths would most likely pick different caches,
although kmalloc() at each place would use the same cache copy whenever
it is executed. In this way, the vulnerable object and memory allocated
in other subsystems and modules will (most probably) be on different
slab caches, which prevents the object from being sprayed.
Meanwhile, the static random selection is further enhanced with a
per-boot random seed, which prevents the attacker from finding a usable
kmalloc that happens to pick the same cache with the vulnerable
subsystem/module by analyzing the open source code. In other words, with
the per-boot seed, the random selection is static during each time the
system starts and runs, but not across different system startups.
The overhead of performance has been tested on a 40-core x86 server by
comparing the results of `perf bench all` between the kernels with and
without this patch based on the latest linux-next kernel, which shows
minor difference. A subset of benchmarks are listed below:
sched/ sched/ syscall/ mem/ mem/
messaging pipe basic memcpy memset
(sec) (sec) (sec) (GB/sec) (GB/sec)
control1 0.019 5.459 0.733 15.258789 51.398026
control2 0.019 5.439 0.730 16.009221 48.828125
control3 0.019 5.282 0.735 16.009221 48.828125
control_avg 0.019 5.393 0.733 15.759077 49.684759
experiment1 0.019 5.374 0.741 15.500992 46.502976
experiment2 0.019 5.440 0.746 16.276042 51.398026
experiment3 0.019 5.242 0.752 15.258789 51.398026
experiment_avg 0.019 5.352 0.746 15.678608 49.766343
The overhead of memory usage was measured by executing `free` after boot
on a QEMU VM with 1GB total memory, and as expected, it's positively
correlated with # of cache copies:
control 4 copies 8 copies 16 copies
total 969.8M 968.2M 968.2M 968.2M
used 20.0M 21.9M 24.1M 26.7M
free 936.9M 933.6M 931.4M 928.6M
available 932.2M 928.8M 926.6M 923.9M
Co-developed-by: Xiu Jianfeng <xiujianfeng@huawei.com>
Signed-off-by: Xiu Jianfeng <xiujianfeng@huawei.com>
Signed-off-by: GONG, Ruiqi <gongruiqi@huaweicloud.com>
Reviewed-by: Kees Cook <keescook@chromium.org>
Reviewed-by: Hyeonggon Yoo <42.hyeyoo@gmail.com>
Acked-by: Dennis Zhou <dennis@kernel.org> # percpu
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
2023-07-14 14:44:22 +08:00
|
|
|
|
2023-10-02 15:43:03 +02:00
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-12-28 17:47:04 +03:00
|
|
|
|
2023-03-15 14:31:33 +03:00
|
|
|
|
|
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-04-22 11:18:40 -03:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-05-25 13:25:59 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-06-23 00:07:49 -07:00
|
|
|
|
|
|
|
|
|
2013-01-16 18:54:17 -08:00
|
|
|
|
2005-06-23 00:07:49 -07:00
|
|
|
|
2005-06-23 00:07:42 -07:00
|
|
|
|
|
|
|
|
|
2005-06-23 00:07:49 -07:00
|
|
|
|
[PATCH] sparsemem memory model
Sparsemem abstracts the use of discontiguous mem_maps[]. This kind of
mem_map[] is needed by discontiguous memory machines (like in the old
CONFIG_DISCONTIGMEM case) as well as memory hotplug systems. Sparsemem
replaces DISCONTIGMEM when enabled, and it is hoped that it can eventually
become a complete replacement.
A significant advantage over DISCONTIGMEM is that it's completely separated
from CONFIG_NUMA. When producing this patch, it became apparent in that NUMA
and DISCONTIG are often confused.
Another advantage is that sparse doesn't require each NUMA node's ranges to be
contiguous. It can handle overlapping ranges between nodes with no problems,
where DISCONTIGMEM currently throws away that memory.
Sparsemem uses an array to provide different pfn_to_page() translations for
each SECTION_SIZE area of physical memory. This is what allows the mem_map[]
to be chopped up.
In order to do quick pfn_to_page() operations, the section number of the page
is encoded in page->flags. Part of the sparsemem infrastructure enables
sharing of these bits more dynamically (at compile-time) between the
page_zone() and sparsemem operations. However, on 32-bit architectures, the
number of bits is quite limited, and may require growing the size of the
page->flags type in certain conditions. Several things might force this to
occur: a decrease in the SECTION_SIZE (if you want to hotplug smaller areas of
memory), an increase in the physical address space, or an increase in the
number of used page->flags.
One thing to note is that, once sparsemem is present, the NUMA node
information no longer needs to be stored in the page->flags. It might provide
speed increases on certain platforms and will be stored there if there is
room. But, if out of room, an alternate (theoretically slower) mechanism is
used.
This patch introduces CONFIG_FLATMEM. It is used in almost all cases where
there used to be an #ifndef DISCONTIG, because SPARSEMEM and DISCONTIGMEM
often have to compile out the same areas of code.
Signed-off-by: Andy Whitcroft <apw@shadowen.org>
Signed-off-by: Dave Hansen <haveblue@us.ibm.com>
Signed-off-by: Martin Bligh <mbligh@aracnet.com>
Signed-off-by: Adrian Bunk <bunk@stusta.de>
Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com>
Signed-off-by: Bob Picco <bob.picco@hp.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-23 00:07:54 -07:00
|
|
|
|
2005-06-23 00:07:49 -07:00
|
|
|
|
2019-05-13 17:23:05 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-06-23 00:07:42 -07:00
|
|
|
|
2005-06-23 00:07:49 -07:00
|
|
|
|
2005-06-23 00:07:42 -07:00
|
|
|
|
2026-08-05 00:11:41 +09:00
|
|
|
|
2021-06-28 19:42:52 -07:00
|
|
|
|
2005-06-23 00:07:42 -07:00
|
|
|
|
2019-05-13 17:23:05 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-06-23 00:07:42 -07:00
|
|
|
|
2019-05-13 17:23:05 -07:00
|
|
|
|
|
|
|
|
|
2019-11-30 17:58:26 -08:00
|
|
|
|
[PATCH] sparsemem memory model
Sparsemem abstracts the use of discontiguous mem_maps[]. This kind of
mem_map[] is needed by discontiguous memory machines (like in the old
CONFIG_DISCONTIGMEM case) as well as memory hotplug systems. Sparsemem
replaces DISCONTIGMEM when enabled, and it is hoped that it can eventually
become a complete replacement.
A significant advantage over DISCONTIGMEM is that it's completely separated
from CONFIG_NUMA. When producing this patch, it became apparent in that NUMA
and DISCONTIG are often confused.
Another advantage is that sparse doesn't require each NUMA node's ranges to be
contiguous. It can handle overlapping ranges between nodes with no problems,
where DISCONTIGMEM currently throws away that memory.
Sparsemem uses an array to provide different pfn_to_page() translations for
each SECTION_SIZE area of physical memory. This is what allows the mem_map[]
to be chopped up.
In order to do quick pfn_to_page() operations, the section number of the page
is encoded in page->flags. Part of the sparsemem infrastructure enables
sharing of these bits more dynamically (at compile-time) between the
page_zone() and sparsemem operations. However, on 32-bit architectures, the
number of bits is quite limited, and may require growing the size of the
page->flags type in certain conditions. Several things might force this to
occur: a decrease in the SECTION_SIZE (if you want to hotplug smaller areas of
memory), an increase in the physical address space, or an increase in the
number of used page->flags.
One thing to note is that, once sparsemem is present, the NUMA node
information no longer needs to be stored in the page->flags. It might provide
speed increases on certain platforms and will be stored there if there is
room. But, if out of room, an alternate (theoretically slower) mechanism is
used.
This patch introduces CONFIG_FLATMEM. It is used in almost all cases where
there used to be an #ifndef DISCONTIG, because SPARSEMEM and DISCONTIGMEM
often have to compile out the same areas of code.
Signed-off-by: Andy Whitcroft <apw@shadowen.org>
Signed-off-by: Dave Hansen <haveblue@us.ibm.com>
Signed-off-by: Martin Bligh <mbligh@aracnet.com>
Signed-off-by: Adrian Bunk <bunk@stusta.de>
Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com>
Signed-off-by: Bob Picco <bob.picco@hp.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-23 00:07:54 -07:00
|
|
|
|
|
|
|
|
|
2005-06-23 00:07:42 -07:00
|
|
|
|
[PATCH] sparsemem memory model
Sparsemem abstracts the use of discontiguous mem_maps[]. This kind of
mem_map[] is needed by discontiguous memory machines (like in the old
CONFIG_DISCONTIGMEM case) as well as memory hotplug systems. Sparsemem
replaces DISCONTIGMEM when enabled, and it is hoped that it can eventually
become a complete replacement.
A significant advantage over DISCONTIGMEM is that it's completely separated
from CONFIG_NUMA. When producing this patch, it became apparent in that NUMA
and DISCONTIG are often confused.
Another advantage is that sparse doesn't require each NUMA node's ranges to be
contiguous. It can handle overlapping ranges between nodes with no problems,
where DISCONTIGMEM currently throws away that memory.
Sparsemem uses an array to provide different pfn_to_page() translations for
each SECTION_SIZE area of physical memory. This is what allows the mem_map[]
to be chopped up.
In order to do quick pfn_to_page() operations, the section number of the page
is encoded in page->flags. Part of the sparsemem infrastructure enables
sharing of these bits more dynamically (at compile-time) between the
page_zone() and sparsemem operations. However, on 32-bit architectures, the
number of bits is quite limited, and may require growing the size of the
page->flags type in certain conditions. Several things might force this to
occur: a decrease in the SECTION_SIZE (if you want to hotplug smaller areas of
memory), an increase in the physical address space, or an increase in the
number of used page->flags.
One thing to note is that, once sparsemem is present, the NUMA node
information no longer needs to be stored in the page->flags. It might provide
speed increases on certain platforms and will be stored there if there is
room. But, if out of room, an alternate (theoretically slower) mechanism is
used.
This patch introduces CONFIG_FLATMEM. It is used in almost all cases where
there used to be an #ifndef DISCONTIG, because SPARSEMEM and DISCONTIGMEM
often have to compile out the same areas of code.
Signed-off-by: Andy Whitcroft <apw@shadowen.org>
Signed-off-by: Dave Hansen <haveblue@us.ibm.com>
Signed-off-by: Martin Bligh <mbligh@aracnet.com>
Signed-off-by: Adrian Bunk <bunk@stusta.de>
Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com>
Signed-off-by: Bob Picco <bob.picco@hp.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-23 00:07:54 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-05-13 17:23:05 -07:00
|
|
|
|
[PATCH] sparsemem memory model
Sparsemem abstracts the use of discontiguous mem_maps[]. This kind of
mem_map[] is needed by discontiguous memory machines (like in the old
CONFIG_DISCONTIGMEM case) as well as memory hotplug systems. Sparsemem
replaces DISCONTIGMEM when enabled, and it is hoped that it can eventually
become a complete replacement.
A significant advantage over DISCONTIGMEM is that it's completely separated
from CONFIG_NUMA. When producing this patch, it became apparent in that NUMA
and DISCONTIG are often confused.
Another advantage is that sparse doesn't require each NUMA node's ranges to be
contiguous. It can handle overlapping ranges between nodes with no problems,
where DISCONTIGMEM currently throws away that memory.
Sparsemem uses an array to provide different pfn_to_page() translations for
each SECTION_SIZE area of physical memory. This is what allows the mem_map[]
to be chopped up.
In order to do quick pfn_to_page() operations, the section number of the page
is encoded in page->flags. Part of the sparsemem infrastructure enables
sharing of these bits more dynamically (at compile-time) between the
page_zone() and sparsemem operations. However, on 32-bit architectures, the
number of bits is quite limited, and may require growing the size of the
page->flags type in certain conditions. Several things might force this to
occur: a decrease in the SECTION_SIZE (if you want to hotplug smaller areas of
memory), an increase in the physical address space, or an increase in the
number of used page->flags.
One thing to note is that, once sparsemem is present, the NUMA node
information no longer needs to be stored in the page->flags. It might provide
speed increases on certain platforms and will be stored there if there is
room. But, if out of room, an alternate (theoretically slower) mechanism is
used.
This patch introduces CONFIG_FLATMEM. It is used in almost all cases where
there used to be an #ifndef DISCONTIG, because SPARSEMEM and DISCONTIGMEM
often have to compile out the same areas of code.
Signed-off-by: Andy Whitcroft <apw@shadowen.org>
Signed-off-by: Dave Hansen <haveblue@us.ibm.com>
Signed-off-by: Martin Bligh <mbligh@aracnet.com>
Signed-off-by: Adrian Bunk <bunk@stusta.de>
Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com>
Signed-off-by: Bob Picco <bob.picco@hp.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-23 00:07:54 -07:00
|
|
|
|
2019-05-13 17:23:05 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
[PATCH] sparsemem memory model
Sparsemem abstracts the use of discontiguous mem_maps[]. This kind of
mem_map[] is needed by discontiguous memory machines (like in the old
CONFIG_DISCONTIGMEM case) as well as memory hotplug systems. Sparsemem
replaces DISCONTIGMEM when enabled, and it is hoped that it can eventually
become a complete replacement.
A significant advantage over DISCONTIGMEM is that it's completely separated
from CONFIG_NUMA. When producing this patch, it became apparent in that NUMA
and DISCONTIG are often confused.
Another advantage is that sparse doesn't require each NUMA node's ranges to be
contiguous. It can handle overlapping ranges between nodes with no problems,
where DISCONTIGMEM currently throws away that memory.
Sparsemem uses an array to provide different pfn_to_page() translations for
each SECTION_SIZE area of physical memory. This is what allows the mem_map[]
to be chopped up.
In order to do quick pfn_to_page() operations, the section number of the page
is encoded in page->flags. Part of the sparsemem infrastructure enables
sharing of these bits more dynamically (at compile-time) between the
page_zone() and sparsemem operations. However, on 32-bit architectures, the
number of bits is quite limited, and may require growing the size of the
page->flags type in certain conditions. Several things might force this to
occur: a decrease in the SECTION_SIZE (if you want to hotplug smaller areas of
memory), an increase in the physical address space, or an increase in the
number of used page->flags.
One thing to note is that, once sparsemem is present, the NUMA node
information no longer needs to be stored in the page->flags. It might provide
speed increases on certain platforms and will be stored there if there is
room. But, if out of room, an alternate (theoretically slower) mechanism is
used.
This patch introduces CONFIG_FLATMEM. It is used in almost all cases where
there used to be an #ifndef DISCONTIG, because SPARSEMEM and DISCONTIGMEM
often have to compile out the same areas of code.
Signed-off-by: Andy Whitcroft <apw@shadowen.org>
Signed-off-by: Dave Hansen <haveblue@us.ibm.com>
Signed-off-by: Martin Bligh <mbligh@aracnet.com>
Signed-off-by: Adrian Bunk <bunk@stusta.de>
Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com>
Signed-off-by: Bob Picco <bob.picco@hp.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-23 00:07:54 -07:00
|
|
|
|
2019-05-13 17:23:05 -07:00
|
|
|
|
[PATCH] sparsemem memory model
Sparsemem abstracts the use of discontiguous mem_maps[]. This kind of
mem_map[] is needed by discontiguous memory machines (like in the old
CONFIG_DISCONTIGMEM case) as well as memory hotplug systems. Sparsemem
replaces DISCONTIGMEM when enabled, and it is hoped that it can eventually
become a complete replacement.
A significant advantage over DISCONTIGMEM is that it's completely separated
from CONFIG_NUMA. When producing this patch, it became apparent in that NUMA
and DISCONTIG are often confused.
Another advantage is that sparse doesn't require each NUMA node's ranges to be
contiguous. It can handle overlapping ranges between nodes with no problems,
where DISCONTIGMEM currently throws away that memory.
Sparsemem uses an array to provide different pfn_to_page() translations for
each SECTION_SIZE area of physical memory. This is what allows the mem_map[]
to be chopped up.
In order to do quick pfn_to_page() operations, the section number of the page
is encoded in page->flags. Part of the sparsemem infrastructure enables
sharing of these bits more dynamically (at compile-time) between the
page_zone() and sparsemem operations. However, on 32-bit architectures, the
number of bits is quite limited, and may require growing the size of the
page->flags type in certain conditions. Several things might force this to
occur: a decrease in the SECTION_SIZE (if you want to hotplug smaller areas of
memory), an increase in the physical address space, or an increase in the
number of used page->flags.
One thing to note is that, once sparsemem is present, the NUMA node
information no longer needs to be stored in the page->flags. It might provide
speed increases on certain platforms and will be stored there if there is
room. But, if out of room, an alternate (theoretically slower) mechanism is
used.
This patch introduces CONFIG_FLATMEM. It is used in almost all cases where
there used to be an #ifndef DISCONTIG, because SPARSEMEM and DISCONTIGMEM
often have to compile out the same areas of code.
Signed-off-by: Andy Whitcroft <apw@shadowen.org>
Signed-off-by: Dave Hansen <haveblue@us.ibm.com>
Signed-off-by: Martin Bligh <mbligh@aracnet.com>
Signed-off-by: Adrian Bunk <bunk@stusta.de>
Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com>
Signed-off-by: Bob Picco <bob.picco@hp.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-23 00:07:54 -07:00
|
|
|
|
2005-06-23 00:07:42 -07:00
|
|
|
|
|
|
|
|
|
[PATCH] sparsemem memory model
Sparsemem abstracts the use of discontiguous mem_maps[]. This kind of
mem_map[] is needed by discontiguous memory machines (like in the old
CONFIG_DISCONTIGMEM case) as well as memory hotplug systems. Sparsemem
replaces DISCONTIGMEM when enabled, and it is hoped that it can eventually
become a complete replacement.
A significant advantage over DISCONTIGMEM is that it's completely separated
from CONFIG_NUMA. When producing this patch, it became apparent in that NUMA
and DISCONTIG are often confused.
Another advantage is that sparse doesn't require each NUMA node's ranges to be
contiguous. It can handle overlapping ranges between nodes with no problems,
where DISCONTIGMEM currently throws away that memory.
Sparsemem uses an array to provide different pfn_to_page() translations for
each SECTION_SIZE area of physical memory. This is what allows the mem_map[]
to be chopped up.
In order to do quick pfn_to_page() operations, the section number of the page
is encoded in page->flags. Part of the sparsemem infrastructure enables
sharing of these bits more dynamically (at compile-time) between the
page_zone() and sparsemem operations. However, on 32-bit architectures, the
number of bits is quite limited, and may require growing the size of the
page->flags type in certain conditions. Several things might force this to
occur: a decrease in the SECTION_SIZE (if you want to hotplug smaller areas of
memory), an increase in the physical address space, or an increase in the
number of used page->flags.
One thing to note is that, once sparsemem is present, the NUMA node
information no longer needs to be stored in the page->flags. It might provide
speed increases on certain platforms and will be stored there if there is
room. But, if out of room, an alternate (theoretically slower) mechanism is
used.
This patch introduces CONFIG_FLATMEM. It is used in almost all cases where
there used to be an #ifndef DISCONTIG, because SPARSEMEM and DISCONTIGMEM
often have to compile out the same areas of code.
Signed-off-by: Andy Whitcroft <apw@shadowen.org>
Signed-off-by: Dave Hansen <haveblue@us.ibm.com>
Signed-off-by: Martin Bligh <mbligh@aracnet.com>
Signed-off-by: Adrian Bunk <bunk@stusta.de>
Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com>
Signed-off-by: Bob Picco <bob.picco@hp.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-23 00:07:54 -07:00
|
|
|
|
|
|
|
|
|
2009-10-26 16:50:12 -07:00
|
|
|
|
[PATCH] sparsemem memory model
Sparsemem abstracts the use of discontiguous mem_maps[]. This kind of
mem_map[] is needed by discontiguous memory machines (like in the old
CONFIG_DISCONTIGMEM case) as well as memory hotplug systems. Sparsemem
replaces DISCONTIGMEM when enabled, and it is hoped that it can eventually
become a complete replacement.
A significant advantage over DISCONTIGMEM is that it's completely separated
from CONFIG_NUMA. When producing this patch, it became apparent in that NUMA
and DISCONTIG are often confused.
Another advantage is that sparse doesn't require each NUMA node's ranges to be
contiguous. It can handle overlapping ranges between nodes with no problems,
where DISCONTIGMEM currently throws away that memory.
Sparsemem uses an array to provide different pfn_to_page() translations for
each SECTION_SIZE area of physical memory. This is what allows the mem_map[]
to be chopped up.
In order to do quick pfn_to_page() operations, the section number of the page
is encoded in page->flags. Part of the sparsemem infrastructure enables
sharing of these bits more dynamically (at compile-time) between the
page_zone() and sparsemem operations. However, on 32-bit architectures, the
number of bits is quite limited, and may require growing the size of the
page->flags type in certain conditions. Several things might force this to
occur: a decrease in the SECTION_SIZE (if you want to hotplug smaller areas of
memory), an increase in the physical address space, or an increase in the
number of used page->flags.
One thing to note is that, once sparsemem is present, the NUMA node
information no longer needs to be stored in the page->flags. It might provide
speed increases on certain platforms and will be stored there if there is
room. But, if out of room, an alternate (theoretically slower) mechanism is
used.
This patch introduces CONFIG_FLATMEM. It is used in almost all cases where
there used to be an #ifndef DISCONTIG, because SPARSEMEM and DISCONTIGMEM
often have to compile out the same areas of code.
Signed-off-by: Andy Whitcroft <apw@shadowen.org>
Signed-off-by: Dave Hansen <haveblue@us.ibm.com>
Signed-off-by: Martin Bligh <mbligh@aracnet.com>
Signed-off-by: Adrian Bunk <bunk@stusta.de>
Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com>
Signed-off-by: Bob Picco <bob.picco@hp.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-23 00:07:54 -07:00
|
|
|
|
2005-06-23 00:07:49 -07:00
|
|
|
|
|
|
|
|
|
2026-08-05 00:11:41 +09:00
|
|
|
|
2021-06-28 19:42:52 -07:00
|
|
|
|
[PATCH] sparsemem memory model
Sparsemem abstracts the use of discontiguous mem_maps[]. This kind of
mem_map[] is needed by discontiguous memory machines (like in the old
CONFIG_DISCONTIGMEM case) as well as memory hotplug systems. Sparsemem
replaces DISCONTIGMEM when enabled, and it is hoped that it can eventually
become a complete replacement.
A significant advantage over DISCONTIGMEM is that it's completely separated
from CONFIG_NUMA. When producing this patch, it became apparent in that NUMA
and DISCONTIG are often confused.
Another advantage is that sparse doesn't require each NUMA node's ranges to be
contiguous. It can handle overlapping ranges between nodes with no problems,
where DISCONTIGMEM currently throws away that memory.
Sparsemem uses an array to provide different pfn_to_page() translations for
each SECTION_SIZE area of physical memory. This is what allows the mem_map[]
to be chopped up.
In order to do quick pfn_to_page() operations, the section number of the page
is encoded in page->flags. Part of the sparsemem infrastructure enables
sharing of these bits more dynamically (at compile-time) between the
page_zone() and sparsemem operations. However, on 32-bit architectures, the
number of bits is quite limited, and may require growing the size of the
page->flags type in certain conditions. Several things might force this to
occur: a decrease in the SECTION_SIZE (if you want to hotplug smaller areas of
memory), an increase in the physical address space, or an increase in the
number of used page->flags.
One thing to note is that, once sparsemem is present, the NUMA node
information no longer needs to be stored in the page->flags. It might provide
speed increases on certain platforms and will be stored there if there is
room. But, if out of room, an alternate (theoretically slower) mechanism is
used.
This patch introduces CONFIG_FLATMEM. It is used in almost all cases where
there used to be an #ifndef DISCONTIG, because SPARSEMEM and DISCONTIGMEM
often have to compile out the same areas of code.
Signed-off-by: Andy Whitcroft <apw@shadowen.org>
Signed-off-by: Dave Hansen <haveblue@us.ibm.com>
Signed-off-by: Martin Bligh <mbligh@aracnet.com>
Signed-off-by: Adrian Bunk <bunk@stusta.de>
Signed-off-by: Yasunori Goto <y-goto@jp.fujitsu.com>
Signed-off-by: Bob Picco <bob.picco@hp.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-06-23 00:07:54 -07:00
|
|
|
|
2005-09-03 15:54:28 -07:00
|
|
|
|
|
|
|
|
|
2020-08-06 23:24:02 -07:00
|
|
|
|
2005-09-03 15:54:28 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2008-10-15 22:01:38 -07:00
|
|
|
|
2005-09-03 15:54:28 -07:00
|
|
|
|
2005-09-03 15:54:26 -07:00
|
|
|
|
2006-10-03 22:34:14 +02:00
|
|
|
|
2005-09-03 15:54:26 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-09-03 15:54:28 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
[PATCH] mm: split page table lock
Christoph Lameter demonstrated very poor scalability on the SGI 512-way, with
a many-threaded application which concurrently initializes different parts of
a large anonymous area.
This patch corrects that, by using a separate spinlock per page table page, to
guard the page table entries in that page, instead of using the mm's single
page_table_lock. (But even then, page_table_lock is still used to guard page
table allocation, and anon_vma allocation.)
In this implementation, the spinlock is tucked inside the struct page of the
page table page: with a BUILD_BUG_ON in case it overflows - which it would in
the case of 32-bit PA-RISC with spinlock debugging enabled.
Splitting the lock is not quite for free: another cacheline access. Ideally,
I suppose we would use split ptlock only for multi-threaded processes on
multi-cpu machines; but deciding that dynamically would have its own costs.
So for now enable it by config, at some number of cpus - since the Kconfig
language doesn't support inequalities, let preprocessor compare that with
NR_CPUS. But I don't think it's worth being user-configurable: for good
testing of both split and unsplit configs, split now at 4 cpus, and perhaps
change that to 8 later.
There is a benefit even for singly threaded processes: kswapd can be attacking
one part of the mm while another part is busy faulting.
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-10-29 18:16:40 -07:00
|
|
|
|
2007-10-16 01:24:14 -07:00
|
|
|
|
2008-10-15 22:01:38 -07:00
|
|
|
|
2007-10-16 01:24:14 -07:00
|
|
|
|
|
|
|
|
|
mm: stop making SPARSEMEM_VMEMMAP user-selectable
Patch series "mm: remove nth_page()", v2.
As discussed recently with Linus, nth_page() is just nasty and we would
like to remove it.
To recap, the reason we currently need nth_page() within a folio is
because on some kernel configs (SPARSEMEM without SPARSEMEM_VMEMMAP), the
memmap is allocated per memory section.
While buddy allocations cannot cross memory section boundaries, hugetlb
and dax folios can.
So crossing a memory section means that "page++" could do the wrong thing.
Instead, nth_page() on these problematic configs always goes from
page->pfn, to the go from (++pfn)->page, which is rather nasty.
Likely, many people have no idea when nth_page() is required and when it
might be dropped.
We refer to such problematic PFN ranges and "non-contiguous pages". If we
only deal with "contiguous pages", there is not need for nth_page().
Besides that "obvious" folio case, we might end up using nth_page() within
CMA allocations (again, could span memory sections), and in one corner
case (kfence) when processing memblock allocations (again, could span
memory sections).
So let's handle all that, add sanity checks, and remove nth_page().
Patch #1 -> #5 : stop making SPARSEMEM_VMEMMAP user-selectable + cleanups
Patch #6 -> #13 : disallow folios to have non-contiguous pages
Patch #14 -> #20 : remove nth_page() usage within folios
Patch #22 : disallow CMA allocations of non-contiguous pages
Patch #23 -> #33 : sanity+check + remove nth_page() usage within SG entry
Patch #34 : sanity-check + remove nth_page() usage in
unpin_user_page_range_dirty_lock()
Patch #35 : remove nth_page() in kfence
Patch #36 : adjust stale comment regarding nth_page
Patch #37 : mm: remove nth_page()
A lot of this is inspired from the discussion at [1] between Linus, Jason
and me, so cudos to them.
This patch (of 37):
In an ideal world, we wouldn't have to deal with SPARSEMEM without
SPARSEMEM_VMEMMAP, but in particular for 32bit SPARSEMEM_VMEMMAP is
considered too costly and consequently not supported.
However, if an architecture does support SPARSEMEM with SPARSEMEM_VMEMMAP,
let's forbid the user to disable VMEMMAP: just like we already do for
arm64, s390 and x86.
So if SPARSEMEM_VMEMMAP is supported, don't allow to use SPARSEMEM without
SPARSEMEM_VMEMMAP.
This implies that the option to not use SPARSEMEM_VMEMMAP will now be gone
for loongarch, powerpc, riscv and sparc. All architectures only enable
SPARSEMEM_VMEMMAP with 64bit support, so there should not really be a big
downside to using the VMEMMAP (quite the contrary).
This is a preparation for not supporting
(1) folio sizes that exceed a single memory section
(2) CMA allocations of non-contiguous page ranges
in SPARSEMEM without SPARSEMEM_VMEMMAP configs, whereby we want to limit
possible impact as much as possible (e.g., gigantic hugetlb page
allocations suddenly fails).
Link: https://lkml.kernel.org/r/20250901150359.867252-1-david@redhat.com
Link: https://lkml.kernel.org/r/20250901150359.867252-2-david@redhat.com
Link: https://lore.kernel.org/all/CAHk-=wiCYfNp4AJLBORU-c7ZyRBUp66W2-Et6cdQ4REx-GyQ_A@mail.gmail.com/T/#u [1]
Signed-off-by: David Hildenbrand <david@redhat.com>
Acked-by: Zi Yan <ziy@nvidia.com>
Acked-by: Mike Rapoport (Microsoft) <rppt@kernel.org>
Acked-by: SeongJae Park <sj@kernel.org>
Reviewed-by: Wei Yang <richard.weiyang@gmail.com>
Reviewed-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Reviewed-by: Liam R. Howlett <Liam.Howlett@oracle.com>
Cc: Huacai Chen <chenhuacai@kernel.org>
Cc: WANG Xuerui <kernel@xen0n.name>
Cc: Madhavan Srinivasan <maddy@linux.ibm.com>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Nicholas Piggin <npiggin@gmail.com>
Cc: Christophe Leroy <christophe.leroy@csgroup.eu>
Cc: Paul Walmsley <paul.walmsley@sifive.com>
Cc: Palmer Dabbelt <palmer@dabbelt.com>
Cc: Albert Ou <aou@eecs.berkeley.edu>
Cc: Alexandre Ghiti <alex@ghiti.fr>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Andreas Larsson <andreas@gaisler.com>
Cc: Alexander Gordeev <agordeev@linux.ibm.com>
Cc: Alexander Potapenko <glider@google.com>
Cc: Alexandru Elisei <alexandru.elisei@arm.com>
Cc: Alex Dubov <oakad@yahoo.com>
Cc: Alex Willamson <alex.williamson@redhat.com>
Cc: Bart van Assche <bvanassche@acm.org>
Cc: Borislav Betkov <bp@alien8.de>
Cc: Brendan Jackman <jackmanb@google.com>
Cc: Brett Creeley <brett.creeley@amd.com>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Christian Borntraeger <borntraeger@linux.ibm.com>
Cc: Christoph Lameter (Ampere) <cl@gentwo.org>
Cc: Damien Le Maol <dlemoal@kernel.org>
Cc: Dave Airlie <airlied@gmail.com>
Cc: Dennis Zhou <dennis@kernel.org>
Cc: Dmitriy Vyukov <dvyukov@google.com>
Cc: Doug Gilbert <dgilbert@interlog.com>
Cc: Heiko Carstens <hca@linux.ibm.com>
Cc: Herbert Xu <herbert@gondor.apana.org.au>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Inki Dae <m.szyprowski@samsung.com>
Cc: James Bottomley <james.bottomley@HansenPartnership.com>
Cc: Jani Nikula <jani.nikula@linux.intel.com>
Cc: Jason A. Donenfeld <jason@zx2c4.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: Jason Gunthorpe <jgg@ziepe.ca>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Jesper Nilsson <jesper.nilsson@axis.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Jonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: Kevin Tian <kevin.tian@intel.com>
Cc: Lars Persson <lars.persson@axis.com>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Marco Elver <elver@google.com>
Cc: "Martin K. Petersen" <martin.petersen@oracle.com>
Cc: Maxim Levitky <maximlevitsky@gmail.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Niklas Cassel <cassel@kernel.org>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Pavel Begunkov <asml.silence@gmail.com>
Cc: Peter Xu <peterx@redhat.com>
Cc: Robin Murohy <robin.murphy@arm.com>
Cc: Rodrigo Vivi <rodrigo.vivi@intel.com>
Cc: Shameerali Kolothum Thodi <shameerali.kolothum.thodi@huawei.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Sven Schnelle <svens@linux.ibm.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Thomas Bogendoerfer <tsbogend@alpha.franken.de>
Cc: Thomas Gleinxer <tglx@linutronix.de>
Cc: Tvrtko Ursulin <tursulin@ursulin.net>
Cc: Ulf Hansson <ulf.hansson@linaro.org>
Cc: Vasily Gorbik <gor@linux.ibm.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Cc: Yishai Hadas <yishaih@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-09-01 17:03:22 +02:00
|
|
|
|
2007-12-17 16:19:53 -08:00
|
|
|
|
|
|
|
|
|
2019-11-30 17:58:23 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/sparse: allow for alternate vmemmap section init at boot
Add functions that are called just before the per-section memmap is
initialized and just before the memmap page structures are initialized.
They are called sparse_vmemmap_init_nid_early and
sparse_vmemmap_init_nid_late, respectively.
This allows for mm subsystems to add calls to initialize memmap and page
structures in a specific way, if using SPARSEMEM_VMEMMAP. Specifically,
hugetlb can pre-HVO bootmem allocated pages that way, so that no time and
resources are wasted on allocating vmemmap pages, only to free them later
(and possibly unnecessarily running the system out of memory in the
process).
Refactor some code and export a few convenience functions for external
use.
In sparse_init_nid, skip any sections that are already initialized, e.g.
they have been initialized by sparse_vmemmap_init_nid_early already.
The hugetlb code to use these functions will be added in a later commit.
Export section_map_size, as any alternate memmap init code will want to
use it.
The internal config option to enable this is SPARSEMEM_VMEMMAP_PREINIT,
which is selected if an architecture-specific option,
ARCH_WANT_HUGETLB_VMEMMAP_PREINIT, is set. In the future, if other
subsystems want to do preinit too, they can do it in a similar fashion.
The internal config option is there because a section flag is used, and
the number of flags available is architecture-dependent (see mmzone.h).
Architecures can decide if there is room for the flag when enabling
options that select SPARSEMEM_VMEMMAP_PREINIT.
Fortunately, as of right now, all sparse vmemmap using architectures do
have room.
Link: https://lkml.kernel.org/r/20250228182928.2645936-11-fvdl@google.com
Signed-off-by: Frank van der Linden <fvdl@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Alexander Gordeev <agordeev@linux.ibm.com>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Heiko Carstens <hca@linux.ibm.com>
Cc: Joao Martins <joao.m.martins@oracle.com>
Cc: Madhavan Srinivasan <maddy@linux.ibm.com>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Roman Gushchin (Cruise) <roman.gushchin@linux.dev>
Cc: Usama Arif <usamaarif642@gmail.com>
Cc: Vasily Gorbik <gor@linux.ibm.com>
Cc: Yu Zhao <yuzhao@google.com>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-02-28 18:29:11 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-04-12 10:30:25 +05:30
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-07-25 00:37:53 +05:30
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-04-12 10:30:25 +05:30
|
|
|
|
2007-10-16 01:24:14 -07:00
|
|
|
|
mm/sparse: allow for alternate vmemmap section init at boot
Add functions that are called just before the per-section memmap is
initialized and just before the memmap page structures are initialized.
They are called sparse_vmemmap_init_nid_early and
sparse_vmemmap_init_nid_late, respectively.
This allows for mm subsystems to add calls to initialize memmap and page
structures in a specific way, if using SPARSEMEM_VMEMMAP. Specifically,
hugetlb can pre-HVO bootmem allocated pages that way, so that no time and
resources are wasted on allocating vmemmap pages, only to free them later
(and possibly unnecessarily running the system out of memory in the
process).
Refactor some code and export a few convenience functions for external
use.
In sparse_init_nid, skip any sections that are already initialized, e.g.
they have been initialized by sparse_vmemmap_init_nid_early already.
The hugetlb code to use these functions will be added in a later commit.
Export section_map_size, as any alternate memmap init code will want to
use it.
The internal config option to enable this is SPARSEMEM_VMEMMAP_PREINIT,
which is selected if an architecture-specific option,
ARCH_WANT_HUGETLB_VMEMMAP_PREINIT, is set. In the future, if other
subsystems want to do preinit too, they can do it in a similar fashion.
The internal config option is there because a section flag is used, and
the number of flags available is architecture-dependent (see mmzone.h).
Architecures can decide if there is room for the flag when enabling
options that select SPARSEMEM_VMEMMAP_PREINIT.
Fortunately, as of right now, all sparse vmemmap using architectures do
have room.
Link: https://lkml.kernel.org/r/20250228182928.2645936-11-fvdl@google.com
Signed-off-by: Frank van der Linden <fvdl@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Alexander Gordeev <agordeev@linux.ibm.com>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Heiko Carstens <hca@linux.ibm.com>
Cc: Joao Martins <joao.m.martins@oracle.com>
Cc: Madhavan Srinivasan <maddy@linux.ibm.com>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Oscar Salvador <osalvador@suse.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Roman Gushchin (Cruise) <roman.gushchin@linux.dev>
Cc: Usama Arif <usamaarif642@gmail.com>
Cc: Vasily Gorbik <gor@linux.ibm.com>
Cc: Yu Zhao <yuzhao@google.com>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-02-28 18:29:11 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-01-29 18:16:01 +01:00
|
|
|
|
2014-12-20 15:41:11 -05:00
|
|
|
|
2014-01-29 18:16:01 +01:00
|
|
|
|
2024-04-02 14:55:15 +02:00
|
|
|
|
2019-07-11 20:57:21 -07:00
|
|
|
|
2014-12-20 15:41:11 -05:00
|
|
|
|
2014-10-09 15:29:14 -07:00
|
|
|
|
2025-05-09 00:46:20 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-06-04 16:48:38 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-05-13 17:22:59 -07:00
|
|
|
|
2014-12-20 15:41:11 -05:00
|
|
|
|
2011-07-14 11:46:03 +02:00
|
|
|
|
2020-02-16 12:01:04 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-07-31 16:43:50 -07:00
|
|
|
|
2014-12-20 15:41:11 -05:00
|
|
|
|
2012-07-31 16:43:50 -07:00
|
|
|
|
kernel/resource: disallow access to exclusive system RAM regions
virtio-mem dynamically exposes memory inside a device memory region as
system RAM to Linux, coordinating with the hypervisor which parts are
actually "plugged" and consequently usable/accessible.
On the one hand, the virtio-mem driver adds/removes whole memory blocks,
creating/removing busy IORESOURCE_SYSTEM_RAM resources, on the other
hand, it logically (un)plugs memory inside added memory blocks,
dynamically either exposing them to the buddy or hiding them from the
buddy and marking them PG_offline.
In contrast to physical devices, like a DIMM, the virtio-mem driver is
required to actually make use of any of the device-provided memory,
because it performs the handshake with the hypervisor. virtio-mem
memory cannot simply be access via /dev/mem without a driver.
There is no safe way to:
a) Access plugged memory blocks via /dev/mem, as they might contain
unplugged holes or might get silently unplugged by the virtio-mem
driver and consequently turned inaccessible.
b) Access unplugged memory blocks via /dev/mem because the virtio-mem
driver is required to make them actually accessible first.
The virtio-spec states that unplugged memory blocks MUST NOT be written,
and only selected unplugged memory blocks MAY be read. We want to make
sure, this is the case in sane environments -- where the virtio-mem driver
was loaded.
We want to make sure that in a sane environment, nobody "accidentially"
accesses unplugged memory inside the device managed region. For example,
a user might spot a memory region in /proc/iomem and try accessing it via
/dev/mem via gdb or dumping it via something else. By the time the mmap()
happens, the memory might already have been removed by the virtio-mem
driver silently: the mmap() would succeeed and user space might
accidentially access unplugged memory.
So once the driver was loaded and detected the device along the
device-managed region, we just want to disallow any access via /dev/mem to
it.
In an ideal world, we would mark the whole region as busy ("owned by a
driver") and exclude it; however, that would be wrong, as we don't really
have actual system RAM at these ranges added to Linux ("busy system RAM").
Instead, we want to mark such ranges as "not actual busy system RAM but
still soft-reserved and prepared by a driver for future use."
Let's teach iomem_is_exclusive() to reject access to any range with
"IORESOURCE_SYSTEM_RAM | IORESOURCE_EXCLUSIVE", even if not busy and even
if "iomem=relaxed" is set. Introduce EXCLUSIVE_SYSTEM_RAM to make it
easier for applicable drivers to depend on this setting in their Kconfig.
For now, there are no applicable ranges and we'll modify virtio-mem next
to properly set IORESOURCE_EXCLUSIVE on the parent resource container it
creates to contain all actual busy system RAM added via
add_memory_driver_managed().
Link: https://lkml.kernel.org/r/20210920142856.17758-3-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Reviewed-by: Dan Williams <dan.j.williams@intel.com>
Cc: Andy Shevchenko <andy.shevchenko@gmail.com>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: Hanjun Guo <guohanjun@huawei.com>
Cc: Jason Wang <jasowang@redhat.com>
Cc: "Michael S. Tsirkin" <mst@redhat.com>
Cc: "Rafael J. Wysocki" <rafael.j.wysocki@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-11-08 18:35:50 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-05-04 18:38:17 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-10-29 18:16:54 -07:00
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
|
|
|
|
|
2020-10-15 20:08:23 -07:00
|
|
|
|
2026-03-20 23:13:36 +01:00
|
|
|
|
2013-05-21 13:49:35 +10:00
|
|
|
|
2021-11-05 13:44:28 -07:00
|
|
|
|
2020-02-16 12:01:04 -08:00
|
|
|
|
2005-10-29 18:16:54 -07:00
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
|
|
|
|
|
2024-12-20 16:07:09 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-11-30 17:58:23 -08:00
|
|
|
|
2024-12-20 16:07:09 -05:00
|
|
|
|
|
|
|
|
|
2016-05-19 17:13:03 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-12-20 16:07:09 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2019-06-07 15:54:32 -03:00
|
|
|
|
2016-05-19 17:13:03 -07:00
|
|
|
|
2024-12-20 16:07:09 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-05-19 17:13:03 -07:00
|
|
|
|
2007-10-16 01:26:12 -07:00
|
|
|
|
|
|
|
|
|
2026-03-19 09:19:40 +01:00
|
|
|
|
2007-10-16 01:26:12 -07:00
|
|
|
|
mm,memory_hotplug: allocate memmap from the added memory range
Physical memory hotadd has to allocate a memmap (struct page array) for
the newly added memory section. Currently, alloc_pages_node() is used
for those allocations.
This has some disadvantages:
a) an existing memory is consumed for that purpose
(eg: ~2MB per 128MB memory section on x86_64)
This can even lead to extreme cases where system goes OOM because
the physically hotplugged memory depletes the available memory before
it is onlined.
b) if the whole node is movable then we have off-node struct pages
which has performance drawbacks.
c) It might be there are no PMD_ALIGNED chunks so memmap array gets
populated with base pages.
This can be improved when CONFIG_SPARSEMEM_VMEMMAP is enabled.
Vmemap page tables can map arbitrary memory. That means that we can
reserve a part of the physically hotadded memory to back vmemmap page
tables. This implementation uses the beginning of the hotplugged memory
for that purpose.
There are some non-obviously things to consider though.
Vmemmap pages are allocated/freed during the memory hotplug events
(add_memory_resource(), try_remove_memory()) when the memory is
added/removed. This means that the reserved physical range is not
online although it is used. The most obvious side effect is that
pfn_to_online_page() returns NULL for those pfns. The current design
expects that this should be OK as the hotplugged memory is considered a
garbage until it is onlined. For example hibernation wouldn't save the
content of those vmmemmaps into the image so it wouldn't be restored on
resume but this should be OK as there no real content to recover anyway
while metadata is reachable from other data structures (e.g. vmemmap
page tables).
The reserved space is therefore (de)initialized during the {on,off}line
events (mhp_{de}init_memmap_on_memory). That is done by extracting page
allocator independent initialization from the regular onlining path.
The primary reason to handle the reserved space outside of
{on,off}line_pages is to make each initialization specific to the
purpose rather than special case them in a single function.
As per above, the functions that are introduced are:
- mhp_init_memmap_on_memory:
Initializes vmemmap pages by calling move_pfn_range_to_zone(), calls
kasan_add_zero_shadow(), and onlines as many sections as vmemmap pages
fully span.
- mhp_deinit_memmap_on_memory:
Offlines as many sections as vmemmap pages fully span, removes the
range from zhe zone by remove_pfn_range_from_zone(), and calls
kasan_remove_zero_shadow() for the range.
The new function memory_block_online() calls mhp_init_memmap_on_memory()
before doing the actual online_pages(). Should online_pages() fail, we
clean up by calling mhp_deinit_memmap_on_memory(). Adjusting of
present_pages is done at the end once we know that online_pages()
succedeed.
On offline, memory_block_offline() needs to unaccount vmemmap pages from
present_pages() before calling offline_pages(). This is necessary because
offline_pages() tears down some structures based on the fact whether the
node or the zone become empty. If offline_pages() fails, we account back
vmemmap pages. If it succeeds, we call mhp_deinit_memmap_on_memory().
Hot-remove:
We need to be careful when removing memory, as adding and
removing memory needs to be done with the same granularity.
To check that this assumption is not violated, we check the
memory range we want to remove and if a) any memory block has
vmemmap pages and b) the range spans more than a single memory
block, we scream out loud and refuse to proceed.
If all is good and the range was using memmap on memory (aka vmemmap pages),
we construct an altmap structure so free_hugepage_table does the right
thing and calls vmem_altmap_free instead of free_pagetable.
Link: https://lkml.kernel.org/r/20210421102701.25051-5-osalvador@suse.de
Signed-off-by: Oscar Salvador <osalvador@suse.de>
Reviewed-by: David Hildenbrand <david@redhat.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Anshuman Khandual <anshuman.khandual@arm.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-05-04 18:39:42 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
|
|
|
|
|
2023-08-08 14:44:56 +05:30
|
|
|
|
|
|
|
|
|
|
|
|
|
|
[PATCH] mm: split page table lock
Christoph Lameter demonstrated very poor scalability on the SGI 512-way, with
a many-threaded application which concurrently initializes different parts of
a large anonymous area.
This patch corrects that, by using a separate spinlock per page table page, to
guard the page table entries in that page, instead of using the mm's single
page_table_lock. (But even then, page_table_lock is still used to guard page
table allocation, and anon_vma allocation.)
In this implementation, the spinlock is tucked inside the struct page of the
page table page: with a BUILD_BUG_ON in case it overflows - which it would in
the case of 32-bit PA-RISC with spinlock debugging enabled.
Splitting the lock is not quite for free: another cacheline access. Ideally,
I suppose we would use split ptlock only for multi-threaded processes on
multi-cpu machines; but deciding that dynamically would have its own costs.
So for now enable it by config, at some number of cpus - since the Kconfig
language doesn't support inequalities, let preprocessor compare that with
NR_CPUS. But I don't think it's worth being user-configurable: for good
testing of both split and unsplit configs, split now at 4 cpus, and perhaps
change that to 8 later.
There is a benefit even for singly threaded processes: kswapd can be attacking
one part of the mm while another part is busy faulting.
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-10-29 18:16:40 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2005-11-23 13:37:37 -08:00
|
|
|
|
2020-05-26 18:33:01 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2009-12-14 17:59:02 -08:00
|
|
|
|
[PATCH] mm: split page table lock
Christoph Lameter demonstrated very poor scalability on the SGI 512-way, with
a many-threaded application which concurrently initializes different parts of
a large anonymous area.
This patch corrects that, by using a separate spinlock per page table page, to
guard the page table entries in that page, instead of using the mm's single
page_table_lock. (But even then, page_table_lock is still used to guard page
table allocation, and anon_vma allocation.)
In this implementation, the spinlock is tucked inside the struct page of the
page table page: with a BUILD_BUG_ON in case it overflows - which it would in
the case of 32-bit PA-RISC with spinlock debugging enabled.
Splitting the lock is not quite for free: another cacheline access. Ideally,
I suppose we would use split ptlock only for multi-threaded processes on
multi-cpu machines; but deciding that dynamically would have its own costs.
So for now enable it by config, at some number of cpus - since the Kconfig
language doesn't support inequalities, let preprocessor compare that with
NR_CPUS. But I don't think it's worth being user-configurable: for good
testing of both split and unsplit configs, split now at 4 cpus, and perhaps
change that to 8 later.
There is a benefit even for singly threaded processes: kswapd can be attacking
one part of the mm while another part is busy faulting.
Signed-off-by: Hugh Dickins <hugh@veritas.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2005-10-29 18:16:40 -07:00
|
|
|
|
2024-07-26 17:07:26 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-09-24 08:42:05 -07:00
|
|
|
|
2024-07-26 17:07:26 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-03-03 07:52:23 +08:00
|
|
|
|
2006-01-08 01:00:49 -08:00
|
|
|
|
2013-11-14 14:31:07 -08:00
|
|
|
|
2014-12-20 15:41:11 -05:00
|
|
|
|
2013-11-14 14:31:07 -08:00
|
|
|
|
2024-07-26 17:07:26 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-10-09 15:29:32 -07:00
|
|
|
|
|
|
|
|
|
2026-01-20 00:01:31 +01:00
|
|
|
|
2014-12-20 15:41:11 -05:00
|
|
|
|
2014-10-09 15:29:32 -07:00
|
|
|
|
2012-12-11 16:02:38 -08:00
|
|
|
|
2026-01-20 00:01:30 +01:00
|
|
|
|
|
|
|
|
|
2026-01-20 00:01:29 +01:00
|
|
|
|
2024-02-11 21:48:08 +09:00
|
|
|
|
2026-01-20 00:01:31 +01:00
|
|
|
|
2012-12-11 16:02:38 -08:00
|
|
|
|
2026-01-20 00:01:29 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2012-12-11 16:02:38 -08:00
|
|
|
|
2010-05-24 14:32:21 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-02-11 21:48:08 +09:00
|
|
|
|
2010-05-24 14:32:21 -07:00
|
|
|
|
2011-01-25 15:07:25 -08:00
|
|
|
|
2010-05-24 14:32:21 -07:00
|
|
|
|
2019-11-30 17:58:23 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2010-05-24 14:32:21 -07:00
|
|
|
|
2022-08-25 18:41:29 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-04-06 20:04:56 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: introduce CONFIG_NUMA_MIGRATION and simplify CONFIG_MIGRATION
CONFIG_MEMORY_HOTREMOVE, CONFIG_COMPACTION and CONFIG_CMA all select
CONFIG_MIGRATION, because they require it to work (users).
Only CONFIG_NUMA_BALANCING and CONFIG_BALLOON_MIGRATION depend on
CONFIG_MIGRATION. CONFIG_BALLOON_MIGRATION is not an actual user, but an
implementation of migration support, so the dependency is correct
(CONFIG_BALLOON_MIGRATION does not make any sense without
CONFIG_MIGRATION).
However, kconfig-language.rst clearly states "In general use select only
for non-visible symbols". So far CONFIG_MIGRATION is user-visible ...
and the dependencies rather confusing.
The whole reason why CONFIG_MIGRATION is user-visible is because of
CONFIG_NUMA: some users might want CONFIG_NUMA but not page migration
support.
Let's clean all that up by introducing a dedicated CONFIG_NUMA_MIGRATION
config option for that purpose only. Make CONFIG_NUMA_BALANCING that so
far depended on CONFIG_NUMA && CONFIG_MIGRATION to depend on
CONFIG_MIGRATION instead. CONFIG_NUMA_MIGRATION will depend on
CONFIG_NUMA && CONFIG_MMU.
CONFIG_NUMA_MIGRATION is user-visible and will default to "y". We use
that default so new configs will automatically enable it, just like it was
the case with CONFIG_MIGRATION. The downside is that some configs that
used to have CONFIG_MIGRATION=n might get it re-enabled by
CONFIG_NUMA_MIGRATION=y, which shouldn't be a problem.
CONFIG_MIGRATION is now a non-visible config option. Any code that select
CONFIG_MIGRATION (as before) must depend directly or indirectly on
CONFIG_MMU.
CONFIG_NUMA_MIGRATION is responsible for any NUMA migration code, which is
mempolicy migration code, memory-tiering code, and move_pages() code in
migrate.c. CONFIG_NUMA_BALANCING uses its functionality.
Note that this implies that with CONFIG_NUMA_MIGRATION=n, move_pages()
will not be available even though CONFIG_MIGRATION=y, which is an expected
change.
In migrate.c, we can remove the CONFIG_NUMA check as both
CONFIG_NUMA_MIGRATION and CONFIG_NUMA_BALANCING depend on it.
With this change, CONFIG_MIGRATION is an internal config, all users of
migration selects CONFIG_MIGRATION, and only CONFIG_BALLOON_MIGRATION
depends on it.
Link: https://lkml.kernel.org/r/20260319-config_migration-v1-2-42270124966f@kernel.org
Signed-off-by: David Hildenbrand (Arm) <david@kernel.org>
Reviewed-by: Lorenzo Stoakes (Oracle) <ljs@kernel.org>
Acked-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Jonathan Cameron <jonathan.cameron@huawei.com>
Cc: Albert Ou <aou@eecs.berkeley.edu>
Cc: Alexander Gordeev <agordeev@linux.ibm.com>
Cc: Alexandre Ghiti <alex@ghiti.fr>
Cc: Alistair Popple <apopple@nvidia.com>
Cc: "Borislav Petkov (AMD)" <bp@alien8.de>
Cc: Byungchul Park <byungchul@sk.com>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Christian Borntraeger <borntraeger@linux.ibm.com>
Cc: Gregory Price <gourry@gourry.net>
Cc: Heiko Carstens <hca@linux.ibm.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Huacai Chen <chenhuacai@kernel.org>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Joshua Hahn <joshua.hahnjy@gmail.com>
Cc: Liam Howlett <liam.howlett@oracle.com>
Cc: Madhavan Srinivasan <maddy@linux.ibm.com>
Cc: Matthew Brost <matthew.brost@intel.com>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Nicholas Piggin <npiggin@gmail.com>
Cc: Palmer Dabbelt <palmer@dabbelt.com>
Cc: Rakie Kim <rakie.kim@sk.com>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Sven Schnelle <svens@linux.ibm.com>
Cc: Vasily Gorbik <gor@linux.ibm.com>
Cc: WANG Xuerui <kernel@xen0n.name>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2026-03-19 09:19:41 +01:00
|
|
|
|
|
|
|
|
|
2024-02-11 21:48:08 +09:00
|
|
|
|
mm: introduce CONFIG_NUMA_MIGRATION and simplify CONFIG_MIGRATION
CONFIG_MEMORY_HOTREMOVE, CONFIG_COMPACTION and CONFIG_CMA all select
CONFIG_MIGRATION, because they require it to work (users).
Only CONFIG_NUMA_BALANCING and CONFIG_BALLOON_MIGRATION depend on
CONFIG_MIGRATION. CONFIG_BALLOON_MIGRATION is not an actual user, but an
implementation of migration support, so the dependency is correct
(CONFIG_BALLOON_MIGRATION does not make any sense without
CONFIG_MIGRATION).
However, kconfig-language.rst clearly states "In general use select only
for non-visible symbols". So far CONFIG_MIGRATION is user-visible ...
and the dependencies rather confusing.
The whole reason why CONFIG_MIGRATION is user-visible is because of
CONFIG_NUMA: some users might want CONFIG_NUMA but not page migration
support.
Let's clean all that up by introducing a dedicated CONFIG_NUMA_MIGRATION
config option for that purpose only. Make CONFIG_NUMA_BALANCING that so
far depended on CONFIG_NUMA && CONFIG_MIGRATION to depend on
CONFIG_MIGRATION instead. CONFIG_NUMA_MIGRATION will depend on
CONFIG_NUMA && CONFIG_MMU.
CONFIG_NUMA_MIGRATION is user-visible and will default to "y". We use
that default so new configs will automatically enable it, just like it was
the case with CONFIG_MIGRATION. The downside is that some configs that
used to have CONFIG_MIGRATION=n might get it re-enabled by
CONFIG_NUMA_MIGRATION=y, which shouldn't be a problem.
CONFIG_MIGRATION is now a non-visible config option. Any code that select
CONFIG_MIGRATION (as before) must depend directly or indirectly on
CONFIG_MMU.
CONFIG_NUMA_MIGRATION is responsible for any NUMA migration code, which is
mempolicy migration code, memory-tiering code, and move_pages() code in
migrate.c. CONFIG_NUMA_BALANCING uses its functionality.
Note that this implies that with CONFIG_NUMA_MIGRATION=n, move_pages()
will not be available even though CONFIG_MIGRATION=y, which is an expected
change.
In migrate.c, we can remove the CONFIG_NUMA check as both
CONFIG_NUMA_MIGRATION and CONFIG_NUMA_BALANCING depend on it.
With this change, CONFIG_MIGRATION is an internal config, all users of
migration selects CONFIG_MIGRATION, and only CONFIG_BALLOON_MIGRATION
depends on it.
Link: https://lkml.kernel.org/r/20260319-config_migration-v1-2-42270124966f@kernel.org
Signed-off-by: David Hildenbrand (Arm) <david@kernel.org>
Reviewed-by: Lorenzo Stoakes (Oracle) <ljs@kernel.org>
Acked-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Jonathan Cameron <jonathan.cameron@huawei.com>
Cc: Albert Ou <aou@eecs.berkeley.edu>
Cc: Alexander Gordeev <agordeev@linux.ibm.com>
Cc: Alexandre Ghiti <alex@ghiti.fr>
Cc: Alistair Popple <apopple@nvidia.com>
Cc: "Borislav Petkov (AMD)" <bp@alien8.de>
Cc: Byungchul Park <byungchul@sk.com>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Christian Borntraeger <borntraeger@linux.ibm.com>
Cc: Gregory Price <gourry@gourry.net>
Cc: Heiko Carstens <hca@linux.ibm.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Huacai Chen <chenhuacai@kernel.org>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Joshua Hahn <joshua.hahnjy@gmail.com>
Cc: Liam Howlett <liam.howlett@oracle.com>
Cc: Madhavan Srinivasan <maddy@linux.ibm.com>
Cc: Matthew Brost <matthew.brost@intel.com>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Nicholas Piggin <npiggin@gmail.com>
Cc: Palmer Dabbelt <palmer@dabbelt.com>
Cc: Rakie Kim <rakie.kim@sk.com>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Sven Schnelle <svens@linux.ibm.com>
Cc: Vasily Gorbik <gor@linux.ibm.com>
Cc: WANG Xuerui <kernel@xen0n.name>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2026-03-19 09:19:41 +01:00
|
|
|
|
|
|
|
|
|
2006-03-22 00:09:12 -08:00
|
|
|
|
mm: introduce CONFIG_NUMA_MIGRATION and simplify CONFIG_MIGRATION
CONFIG_MEMORY_HOTREMOVE, CONFIG_COMPACTION and CONFIG_CMA all select
CONFIG_MIGRATION, because they require it to work (users).
Only CONFIG_NUMA_BALANCING and CONFIG_BALLOON_MIGRATION depend on
CONFIG_MIGRATION. CONFIG_BALLOON_MIGRATION is not an actual user, but an
implementation of migration support, so the dependency is correct
(CONFIG_BALLOON_MIGRATION does not make any sense without
CONFIG_MIGRATION).
However, kconfig-language.rst clearly states "In general use select only
for non-visible symbols". So far CONFIG_MIGRATION is user-visible ...
and the dependencies rather confusing.
The whole reason why CONFIG_MIGRATION is user-visible is because of
CONFIG_NUMA: some users might want CONFIG_NUMA but not page migration
support.
Let's clean all that up by introducing a dedicated CONFIG_NUMA_MIGRATION
config option for that purpose only. Make CONFIG_NUMA_BALANCING that so
far depended on CONFIG_NUMA && CONFIG_MIGRATION to depend on
CONFIG_MIGRATION instead. CONFIG_NUMA_MIGRATION will depend on
CONFIG_NUMA && CONFIG_MMU.
CONFIG_NUMA_MIGRATION is user-visible and will default to "y". We use
that default so new configs will automatically enable it, just like it was
the case with CONFIG_MIGRATION. The downside is that some configs that
used to have CONFIG_MIGRATION=n might get it re-enabled by
CONFIG_NUMA_MIGRATION=y, which shouldn't be a problem.
CONFIG_MIGRATION is now a non-visible config option. Any code that select
CONFIG_MIGRATION (as before) must depend directly or indirectly on
CONFIG_MMU.
CONFIG_NUMA_MIGRATION is responsible for any NUMA migration code, which is
mempolicy migration code, memory-tiering code, and move_pages() code in
migrate.c. CONFIG_NUMA_BALANCING uses its functionality.
Note that this implies that with CONFIG_NUMA_MIGRATION=n, move_pages()
will not be available even though CONFIG_MIGRATION=y, which is an expected
change.
In migrate.c, we can remove the CONFIG_NUMA check as both
CONFIG_NUMA_MIGRATION and CONFIG_NUMA_BALANCING depend on it.
With this change, CONFIG_MIGRATION is an internal config, all users of
migration selects CONFIG_MIGRATION, and only CONFIG_BALLOON_MIGRATION
depends on it.
Link: https://lkml.kernel.org/r/20260319-config_migration-v1-2-42270124966f@kernel.org
Signed-off-by: David Hildenbrand (Arm) <david@kernel.org>
Reviewed-by: Lorenzo Stoakes (Oracle) <ljs@kernel.org>
Acked-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Jonathan Cameron <jonathan.cameron@huawei.com>
Cc: Albert Ou <aou@eecs.berkeley.edu>
Cc: Alexander Gordeev <agordeev@linux.ibm.com>
Cc: Alexandre Ghiti <alex@ghiti.fr>
Cc: Alistair Popple <apopple@nvidia.com>
Cc: "Borislav Petkov (AMD)" <bp@alien8.de>
Cc: Byungchul Park <byungchul@sk.com>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Christian Borntraeger <borntraeger@linux.ibm.com>
Cc: Gregory Price <gourry@gourry.net>
Cc: Heiko Carstens <hca@linux.ibm.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Huacai Chen <chenhuacai@kernel.org>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Joshua Hahn <joshua.hahnjy@gmail.com>
Cc: Liam Howlett <liam.howlett@oracle.com>
Cc: Madhavan Srinivasan <maddy@linux.ibm.com>
Cc: Matthew Brost <matthew.brost@intel.com>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Nicholas Piggin <npiggin@gmail.com>
Cc: Palmer Dabbelt <palmer@dabbelt.com>
Cc: Rakie Kim <rakie.kim@sk.com>
Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de>
Cc: Steven Rostedt <rostedt@goodmis.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Sven Schnelle <svens@linux.ibm.com>
Cc: Vasily Gorbik <gor@linux.ibm.com>
Cc: WANG Xuerui <kernel@xen0n.name>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2026-03-19 09:19:41 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2006-06-12 17:11:31 -07:00
|
|
|
|
2022-02-16 15:31:38 +11:00
|
|
|
|
2022-02-16 15:31:38 +11:00
|
|
|
|
2022-02-16 15:31:38 +11:00
|
|
|
|
2014-06-04 16:05:35 -07:00
|
|
|
|
2014-12-20 15:41:11 -05:00
|
|
|
|
2014-06-04 16:05:35 -07:00
|
|
|
|
2026-07-06 04:42:08 -07:00
|
|
|
|
2017-09-08 16:10:53 -07:00
|
|
|
|
|
|
|
|
|
2021-05-04 18:33:19 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-12-28 17:47:04 +03:00
|
|
|
|
|
|
|
|
|
2022-03-22 14:43:20 -07:00
|
|
|
|
2019-05-13 17:19:00 -07:00
|
|
|
|
2019-11-30 17:58:23 -08:00
|
|
|
|
2019-05-13 17:19:00 -07:00
|
|
|
|
mm: restrict the pcp batch scale factor to avoid too long latency
In page allocator, PCP (Per-CPU Pageset) is refilled and drained in
batches to increase page allocation throughput, reduce page
allocation/freeing latency per page, and reduce zone lock contention. But
too large batch size will cause too long maximal allocation/freeing
latency, which may punish arbitrary users. So the default batch size is
chosen carefully (in zone_batchsize(), the value is 63 for zone > 1GB) to
avoid that.
In commit 3b12e7e97938 ("mm/page_alloc: scale the number of pages that are
batch freed"), the batch size will be scaled for large number of page
freeing to improve page freeing performance and reduce zone lock
contention. Similar optimization can be used for large number of pages
allocation too.
To find out a suitable max batch scale factor (that is, max effective
batch size), some tests and measurement on some machines were done as
follows.
A set of debug patches are implemented as follows,
- Set PCP high to be 2 * batch to reduce the effect of PCP high
- Disable free batch size scaling to get the raw performance.
- The code with zone lock held is extracted from rmqueue_bulk() and
free_pcppages_bulk() to 2 separate functions to make it easy to
measure the function run time with ftrace function_graph tracer.
- The batch size is hard coded to be 63 (default), 127, 255, 511,
1023, 2047, 4095.
Then will-it-scale/page_fault1 is used to generate the page
allocation/freeing workload. The page allocation/freeing throughput
(page/s) is measured via will-it-scale. The page allocation/freeing
average latency (alloc/free latency avg, in us) and allocation/freeing
latency at 99 percentile (alloc/free latency 99%, in us) are measured with
ftrace function_graph tracer.
The test results are as follows,
Sapphire Rapids Server
======================
Batch throughput free latency free latency alloc latency alloc latency
page/s avg / us 99% / us avg / us 99% / us
----- ---------- ------------ ------------ ------------- -------------
63 513633.4 2.33 3.57 2.67 6.83
127 517616.7 4.35 6.65 4.22 13.03
255 520822.8 8.29 13.32 7.52 25.24
511 524122.0 15.79 23.42 14.02 49.35
1023 525980.5 30.25 44.19 25.36 94.88
2047 526793.6 59.39 84.50 45.22 140.81
Ice Lake Server
===============
Batch throughput free latency free latency alloc latency alloc latency
page/s avg / us 99% / us avg / us 99% / us
----- ---------- ------------ ------------ ------------- -------------
63 620210.3 2.21 3.68 2.02 4.35
127 627003.0 4.09 6.86 3.51 8.28
255 630777.5 7.70 13.50 6.17 15.97
511 633651.5 14.85 22.62 11.66 31.08
1023 637071.1 28.55 42.02 20.81 54.36
2047 638089.7 56.54 84.06 39.28 91.68
Cascade Lake Server
===================
Batch throughput free latency free latency alloc latency alloc latency
page/s avg / us 99% / us avg / us 99% / us
----- ---------- ------------ ------------ ------------- -------------
63 404706.7 3.29 5.03 3.53 4.75
127 422475.2 6.12 9.09 6.36 8.76
255 411522.2 11.68 16.97 10.90 16.39
511 428124.1 22.54 31.28 19.86 32.25
1023 414718.4 43.39 62.52 40.00 66.33
2047 429848.7 86.64 120.34 71.14 106.08
Commet Lake Desktop
===================
Batch throughput free latency free latency alloc latency alloc latency
page/s avg / us 99% / us avg / us 99% / us
----- ---------- ------------ ------------ ------------- -------------
63 795183.13 2.18 3.55 2.03 3.05
127 803067.85 3.91 6.56 3.85 5.52
255 812771.10 7.35 10.80 7.14 10.20
511 817723.48 14.17 27.54 13.43 30.31
1023 818870.19 27.72 40.10 27.89 46.28
Coffee Lake Desktop
===================
Batch throughput free latency free latency alloc latency alloc latency
page/s avg / us 99% / us avg / us 99% / us
----- ---------- ------------ ------------ ------------- -------------
63 510542.8 3.13 4.40 2.48 3.43
127 514288.6 5.97 7.89 4.65 6.04
255 516889.7 11.86 15.58 8.96 12.55
511 519802.4 23.10 28.81 16.95 26.19
1023 520802.7 45.30 52.51 33.19 45.95
2047 519997.1 90.63 104.00 65.26 81.74
From the above data, to restrict the allocation/freeing latency to be less
than 100 us in most times, the max batch scale factor needs to be less
than or equal to 5.
Although it is reasonable to use 5 as max batch scale factor for the
systems tested, there are also slower systems. Where smaller value should
be used to constrain the page allocation/freeing latency.
So, in this patch, a new kconfig option (PCP_BATCH_SCALE_MAX) is added to
set the max batch scale factor. Whose default value is 5, and users can
reduce it when necessary.
Link: https://lkml.kernel.org/r/20231016053002.756205-5-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Acked-by: Andrew Morton <akpm@linux-foundation.org>
Acked-by: Mel Gorman <mgorman@techsingularity.net>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: David Hildenbrand <david@redhat.com>
Cc: Johannes Weiner <jweiner@redhat.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Christoph Lameter <cl@linux.com>
Cc: Arjan van de Ven <arjan@linux.intel.com>
Cc: Sudeep Holla <sudeep.holla@arm.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-10-16 13:29:57 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2008-09-11 01:31:45 -07:00
|
|
|
|
2018-04-03 16:24:20 +02:00
|
|
|
|
2008-09-11 01:31:45 -07:00
|
|
|
|
mmu-notifiers: core
With KVM/GFP/XPMEM there isn't just the primary CPU MMU pointing to pages.
There are secondary MMUs (with secondary sptes and secondary tlbs) too.
sptes in the kvm case are shadow pagetables, but when I say spte in
mmu-notifier context, I mean "secondary pte". In GRU case there's no
actual secondary pte and there's only a secondary tlb because the GRU
secondary MMU has no knowledge about sptes and every secondary tlb miss
event in the MMU always generates a page fault that has to be resolved by
the CPU (this is not the case of KVM where the a secondary tlb miss will
walk sptes in hardware and it will refill the secondary tlb transparently
to software if the corresponding spte is present). The same way
zap_page_range has to invalidate the pte before freeing the page, the spte
(and secondary tlb) must also be invalidated before any page is freed and
reused.
Currently we take a page_count pin on every page mapped by sptes, but that
means the pages can't be swapped whenever they're mapped by any spte
because they're part of the guest working set. Furthermore a spte unmap
event can immediately lead to a page to be freed when the pin is released
(so requiring the same complex and relatively slow tlb_gather smp safe
logic we have in zap_page_range and that can be avoided completely if the
spte unmap event doesn't require an unpin of the page previously mapped in
the secondary MMU).
The mmu notifiers allow kvm/GRU/XPMEM to attach to the tsk->mm and know
when the VM is swapping or freeing or doing anything on the primary MMU so
that the secondary MMU code can drop sptes before the pages are freed,
avoiding all page pinning and allowing 100% reliable swapping of guest
physical address space. Furthermore it avoids the code that teardown the
mappings of the secondary MMU, to implement a logic like tlb_gather in
zap_page_range that would require many IPI to flush other cpu tlbs, for
each fixed number of spte unmapped.
To make an example: if what happens on the primary MMU is a protection
downgrade (from writeable to wrprotect) the secondary MMU mappings will be
invalidated, and the next secondary-mmu-page-fault will call
get_user_pages and trigger a do_wp_page through get_user_pages if it
called get_user_pages with write=1, and it'll re-establishing an updated
spte or secondary-tlb-mapping on the copied page. Or it will setup a
readonly spte or readonly tlb mapping if it's a guest-read, if it calls
get_user_pages with write=0. This is just an example.
This allows to map any page pointed by any pte (and in turn visible in the
primary CPU MMU), into a secondary MMU (be it a pure tlb like GRU, or an
full MMU with both sptes and secondary-tlb like the shadow-pagetable layer
with kvm), or a remote DMA in software like XPMEM (hence needing of
schedule in XPMEM code to send the invalidate to the remote node, while no
need to schedule in kvm/gru as it's an immediate event like invalidating
primary-mmu pte).
At least for KVM without this patch it's impossible to swap guests
reliably. And having this feature and removing the page pin allows
several other optimizations that simplify life considerably.
Dependencies:
1) mm_take_all_locks() to register the mmu notifier when the whole VM
isn't doing anything with "mm". This allows mmu notifier users to keep
track if the VM is in the middle of the invalidate_range_begin/end
critical section with an atomic counter incraese in range_begin and
decreased in range_end. No secondary MMU page fault is allowed to map
any spte or secondary tlb reference, while the VM is in the middle of
range_begin/end as any page returned by get_user_pages in that critical
section could later immediately be freed without any further
->invalidate_page notification (invalidate_range_begin/end works on
ranges and ->invalidate_page isn't called immediately before freeing
the page). To stop all page freeing and pagetable overwrites the
mmap_sem must be taken in write mode and all other anon_vma/i_mmap
locks must be taken too.
2) It'd be a waste to add branches in the VM if nobody could possibly
run KVM/GRU/XPMEM on the kernel, so mmu notifiers will only enabled if
CONFIG_KVM=m/y. In the current kernel kvm won't yet take advantage of
mmu notifiers, but this already allows to compile a KVM external module
against a kernel with mmu notifiers enabled and from the next pull from
kvm.git we'll start using them. And GRU/XPMEM will also be able to
continue the development by enabling KVM=m in their config, until they
submit all GRU/XPMEM GPLv2 code to the mainline kernel. Then they can
also enable MMU_NOTIFIERS in the same way KVM does it (even if KVM=n).
This guarantees nobody selects MMU_NOTIFIER=y if KVM and GRU and XPMEM
are all =n.
The mmu_notifier_register call can fail because mm_take_all_locks may be
interrupted by a signal and return -EINTR. Because mmu_notifier_reigster
is used when a driver startup, a failure can be gracefully handled. Here
an example of the change applied to kvm to register the mmu notifiers.
Usually when a driver startups other allocations are required anyway and
-ENOMEM failure paths exists already.
struct kvm *kvm_arch_create_vm(void)
{
struct kvm *kvm = kzalloc(sizeof(struct kvm), GFP_KERNEL);
+ int err;
if (!kvm)
return ERR_PTR(-ENOMEM);
INIT_LIST_HEAD(&kvm->arch.active_mmu_pages);
+ kvm->arch.mmu_notifier.ops = &kvm_mmu_notifier_ops;
+ err = mmu_notifier_register(&kvm->arch.mmu_notifier, current->mm);
+ if (err) {
+ kfree(kvm);
+ return ERR_PTR(err);
+ }
+
return kvm;
}
mmu_notifier_unregister returns void and it's reliable.
The patch also adds a few needed but missing includes that would prevent
kernel to compile after these changes on non-x86 archs (x86 didn't need
them by luck).
[akpm@linux-foundation.org: coding-style fixes]
[akpm@linux-foundation.org: fix mm/filemap_xip.c build]
[akpm@linux-foundation.org: fix mm/mmu_notifier.c build]
Signed-off-by: Andrea Arcangeli <andrea@qumranet.com>
Signed-off-by: Nick Piggin <npiggin@suse.de>
Signed-off-by: Christoph Lameter <cl@linux-foundation.org>
Cc: Jack Steiner <steiner@sgi.com>
Cc: Robin Holt <holt@sgi.com>
Cc: Nick Piggin <npiggin@suse.de>
Cc: Peter Zijlstra <a.p.zijlstra@chello.nl>
Cc: Kanoj Sarcar <kanojsarcar@yahoo.com>
Cc: Roland Dreier <rdreier@cisco.com>
Cc: Steve Wise <swise@opengridcomputing.com>
Cc: Avi Kivity <avi@qumranet.com>
Cc: Hugh Dickins <hugh@veritas.com>
Cc: Rusty Russell <rusty@rustcorp.com.au>
Cc: Anthony Liguori <aliguori@us.ibm.com>
Cc: Chris Wright <chrisw@redhat.com>
Cc: Marcelo Tosatti <marcelo@kvack.org>
Cc: Eric Dumazet <dada1@cosmosbay.com>
Cc: "Paul E. McKenney" <paulmck@us.ibm.com>
Cc: Izik Eidus <izike@qumranet.com>
Cc: Anthony Liguori <aliguori@us.ibm.com>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2008-07-28 15:46:29 -07:00
|
|
|
|
|
|
|
|
|
mm/mmu_notifier: add an interval tree notifier
Of the 13 users of mmu_notifiers, 8 of them use only
invalidate_range_start/end() and immediately intersect the
mmu_notifier_range with some kind of internal list of VAs. 4 use an
interval tree (i915_gem, radeon_mn, umem_odp, hfi1). 4 use a linked list
of some kind (scif_dma, vhost, gntdev, hmm)
And the remaining 5 either don't use invalidate_range_start() or do some
special thing with it.
It turns out that building a correct scheme with an interval tree is
pretty complicated, particularly if the use case is synchronizing against
another thread doing get_user_pages(). Many of these implementations have
various subtle and difficult to fix races.
This approach puts the interval tree as common code at the top of the mmu
notifier call tree and implements a shareable locking scheme.
It includes:
- An interval tree tracking VA ranges, with per-range callbacks
- A read/write locking scheme for the interval tree that avoids
sleeping in the notifier path (for OOM killer)
- A sequence counter based collision-retry locking scheme to tell
device page fault that a VA range is being concurrently invalidated.
This is based on various ideas:
- hmm accumulates invalidated VA ranges and releases them when all
invalidates are done, via active_invalidate_ranges count.
This approach avoids having to intersect the interval tree twice (as
umem_odp does) at the potential cost of a longer device page fault.
- kvm/umem_odp use a sequence counter to drive the collision retry,
via invalidate_seq
- a deferred work todo list on unlock scheme like RTNL, via deferred_list.
This makes adding/removing interval tree members more deterministic
- seqlock, except this version makes the seqlock idea multi-holder on the
write side by protecting it with active_invalidate_ranges and a spinlock
To minimize MM overhead when only the interval tree is being used, the
entire SRCU and hlist overheads are dropped using some simple
branches. Similarly the interval tree overhead is dropped when in hlist
mode.
The overhead from the mandatory spinlock is broadly the same as most of
existing users which already had a lock (or two) of some sort on the
invalidation path.
Link: https://lore.kernel.org/r/20191112202231.3856-3-jgg@ziepe.ca
Acked-by: Christian König <christian.koenig@amd.com>
Tested-by: Philip Yang <Philip.Yang@amd.com>
Tested-by: Ralph Campbell <rcampbell@nvidia.com>
Reviewed-by: John Hubbard <jhubbard@nvidia.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
2019-11-12 16:22:19 -04:00
|
|
|
|
2009-05-06 16:03:05 -07:00
|
|
|
|
2009-09-21 17:01:57 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-12-28 00:34:05 -08:00
|
|
|
|
2009-09-21 17:01:57 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2009-12-14 17:59:34 -08:00
|
|
|
|
2009-09-21 17:01:57 -07:00
|
|
|
|
|
|
|
|
|
2022-06-27 09:00:26 +03:00
|
|
|
|
2009-10-07 16:32:22 -07:00
|
|
|
|
|
|
|
|
|
2009-09-21 17:01:57 -07:00
|
|
|
|
2009-06-03 16:04:31 -04:00
|
|
|
|
2019-11-30 17:58:23 -08:00
|
|
|
|
2009-12-15 19:27:45 +00:00
|
|
|
|
2019-11-30 17:58:23 -08:00
|
|
|
|
|
|
|
|
|
2009-06-03 16:04:31 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-06-19 10:30:38 +02:00
|
|
|
|
2009-06-03 16:04:31 -04:00
|
|
|
|
|
|
|
|
|
2009-07-31 12:54:11 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2009-06-03 16:04:31 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2009-09-26 09:35:07 -07:00
|
|
|
|
|
|
|
|
|
2009-06-03 16:04:31 -04:00
|
|
|
|
2009-09-16 11:50:15 +02:00
|
|
|
|
|
|
|
|
|
2009-09-26 09:35:07 -07:00
|
|
|
|
2009-09-16 11:50:15 +02:00
|
|
|
|
2025-11-02 18:44:33 +00:00
|
|
|
|
mm/Kconfig: make MEMORY_FAILURE select MIGRATION
For embedded devices, lacking support for NUMA, memory hotplug/hotremove,
CMA and huge pages is a quite common scenario. In this scenario, the
demand for contiguous physical memory allocation is very low. To reduce
the kernel image size, some devices disable the compaction. However,
their SoCs do support DDR ECC, meaning that memory-failure may be needed.
Migration is very useful for soft_offline_page() in memory-failure, which
may be triggered by correctable memory errors. Most anonymous and
file-mapped faulty pages can be migrated to other healthy pages.
Currently, MEMORY_FAILURE does not explicitly select MIGRATION. When
COMPACTION, MEMORY_HOTREMOVE, NUMA_MIGRATION and CMA are all disabled,
MEMORY_FAILURE can be enabled, but MIGRATION cannot be selected.
Make MEMORY_FAILURE select MIGRATION to handle this situation.
Link: https://lore.kernel.org/20260813134916.292733-1-xieyuanbin1@huawei.com
Signed-off-by: Xie Yuanbin <xieyuanbin1@huawei.com>
Suggested-by: Mike Rapoport <rppt@kernel.org>
Reviewed-by: Lorenzo Stoakes (ARM) <ljs@kernel.org>
Reviewed-by: Mike Rapoport (Microsoft) <rppt@kernel.org>
Acked-by: Zi Yan <ziy@nvidia.com>
Acked-by: David Hildenbrand (Arm) <david@kernel.org>
Acked-by: Miaohe Lin <linmiaohe@huawei.com>
Cc: Alistair Popple <apopple@nvidia.com>
Cc: "Borislav Petkov (AMD)" <bp@alien8.de>
Cc: Byungchul Park <byungchul@sk.com>
Cc: David Hildenbrand <david@kernel.org>
Cc: Gregory Price <gourry@gourry.net>
Cc: "Huang, Ying" <ying.huang@linux.alibaba.com>
Cc: Joshua Hahn <joshua.hahnjy@gmail.com>
Cc: Liam R. Howlett <liam@infradead.org>
Cc: liaohua <liaohua4@huawei.com>
Cc: "Luck, Tony" <tony.luck@intel.com>
Cc: Matthew Brost <matthew.brost@intel.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Naoya Horiguchi <nao.horiguchi@gmail.com>
Cc: Rakie Kim <rakie.kim@sk.com>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Vlastimil Babka <vbabka@kernel.org>
Cc: Yuanbin Xie <xieyuanbin1@huawei.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2026-08-13 21:49:16 +08:00
|
|
|
|
2009-09-16 11:50:15 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2009-09-16 11:50:17 +02:00
|
|
|
|
2009-12-16 12:20:00 +01:00
|
|
|
|
2009-12-21 19:56:42 +01:00
|
|
|
|
2009-12-16 12:19:59 +01:00
|
|
|
|
2009-09-16 11:50:17 +02:00
|
|
|
|
2009-05-06 16:03:05 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-08-12 11:22:30 +02:00
|
|
|
|
2010-09-03 18:22:48 +02:00
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-08-11 10:41:11 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/rmap: basic MM owner tracking for large folios (!hugetlb)
For small folios, we traditionally use the mapcount to decide whether it
was "certainly mapped exclusively" by a single MM (mapcount == 1) or
whether it "maybe mapped shared" by multiple MMs (mapcount > 1). For
PMD-sized folios that were PMD-mapped, we were able to use a similar
mechanism (single PMD mapping), but for PTE-mapped folios and in the
future folios that span multiple PMDs, this does not work.
So we need a different mechanism to handle large folios. Let's add a new
mechanism to detect whether a large folio is "certainly mapped
exclusively", or whether it is "maybe mapped shared".
We'll use this information next to optimize CoW reuse for PTE-mapped
anonymous THP, and to convert folio_likely_mapped_shared() to
folio_maybe_mapped_shared(), independent of per-page mapcounts.
For each large folio, we'll have two slots, whereby a slot stores:
(1) an MM id: unique id assigned to each MM
(2) a per-MM mapcount
If a slot is unoccupied, it can be taken by the next MM that maps folio
page.
In addition, we'll remember the current state -- "mapped exclusively" vs.
"maybe mapped shared" -- and use a bit spinlock to sync on updates and to
reduce the total number of atomic accesses on updates. In the future, it
might be possible to squeeze a proper spinlock into "struct folio". For
now, keep it simple, as we require the whole thing with THP only, that is
incompatible with RT.
As we have to squeeze this information into the "struct folio" of even
folios of order-1 (2 pages), and we generally want to reduce the required
metadata, we'll assign each MM a unique ID that can fit into an int. In
total, we can squeeze everything into 4x int (2x long) on 64bit.
32bit support is a bit challenging, because we only have 2x long == 2x int
in order-1 folios. But we can make it work for now, because we neither
expect many MMs nor very large folios on 32bit.
We will reliably detect folios as "mapped exclusively" vs. "mapped
shared" as long as only two MMs map pages of a folio at one point in time
-- for example with fork() and short-lived child processes, or with apps
that hand over state from one instance to another.
As soon as three MMs are involved at the same time, we might detect "maybe
mapped shared" although the folio is "mapped exclusively".
Example 1:
(1) App1 faults in a (shmem/file-backed) folio page -> Tracked as MM0
(2) App2 faults in a folio page -> Tracked as MM1
(4) App1 unmaps all folio pages
-> We will detect "mapped exclusively".
Example 2:
(1) App1 faults in a (shmem/file-backed) folio page -> Tracked as MM0
(2) App2 faults in a folio page -> Tracked as MM1
(3) App3 faults in a folio page -> No slot available, tracked as "unknown"
(4) App1 and App2 unmap all folio pages
-> We will detect "maybe mapped shared".
Make use of __always_inline to keep possible performance degradation when
(un)mapping large folios to a minimum.
Note: by squeezing the two flags into the "unsigned long" that stores the
MM ids, we can use non-atomic __bit_spin_unlock() and non-atomic
setting/clearing of the "maybe mapped shared" bit, effectively not adding
any new atomics on the hot path when updating the large mapcount + new
metadata, which further helps reduce the runtime overhead in
micro-benchmarks.
Link: https://lkml.kernel.org/r/20250303163014.1128035-13-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Cc: Andy Lutomirks^H^Hski <luto@kernel.org>
Cc: Borislav Betkov <bp@alien8.de>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Jann Horn <jannh@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Lance Yang <ioworker0@gmail.com>
Cc: Liam Howlett <liam.howlett@oracle.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Matthew Wilcow (Oracle) <willy@infradead.org>
Cc: Michal Koutn <mkoutny@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: tejun heo <tj@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Zefan Li <lizefan.x@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-03-03 17:30:05 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
2011-01-13 15:47:07 -08:00
|
|
|
|
2021-11-05 13:35:27 -07:00
|
|
|
|
2011-01-13 15:47:07 -08:00
|
|
|
|
2018-09-22 16:14:30 -04:00
|
|
|
|
mm/rmap: basic MM owner tracking for large folios (!hugetlb)
For small folios, we traditionally use the mapcount to decide whether it
was "certainly mapped exclusively" by a single MM (mapcount == 1) or
whether it "maybe mapped shared" by multiple MMs (mapcount > 1). For
PMD-sized folios that were PMD-mapped, we were able to use a similar
mechanism (single PMD mapping), but for PTE-mapped folios and in the
future folios that span multiple PMDs, this does not work.
So we need a different mechanism to handle large folios. Let's add a new
mechanism to detect whether a large folio is "certainly mapped
exclusively", or whether it is "maybe mapped shared".
We'll use this information next to optimize CoW reuse for PTE-mapped
anonymous THP, and to convert folio_likely_mapped_shared() to
folio_maybe_mapped_shared(), independent of per-page mapcounts.
For each large folio, we'll have two slots, whereby a slot stores:
(1) an MM id: unique id assigned to each MM
(2) a per-MM mapcount
If a slot is unoccupied, it can be taken by the next MM that maps folio
page.
In addition, we'll remember the current state -- "mapped exclusively" vs.
"maybe mapped shared" -- and use a bit spinlock to sync on updates and to
reduce the total number of atomic accesses on updates. In the future, it
might be possible to squeeze a proper spinlock into "struct folio". For
now, keep it simple, as we require the whole thing with THP only, that is
incompatible with RT.
As we have to squeeze this information into the "struct folio" of even
folios of order-1 (2 pages), and we generally want to reduce the required
metadata, we'll assign each MM a unique ID that can fit into an int. In
total, we can squeeze everything into 4x int (2x long) on 64bit.
32bit support is a bit challenging, because we only have 2x long == 2x int
in order-1 folios. But we can make it work for now, because we neither
expect many MMs nor very large folios on 32bit.
We will reliably detect folios as "mapped exclusively" vs. "mapped
shared" as long as only two MMs map pages of a folio at one point in time
-- for example with fork() and short-lived child processes, or with apps
that hand over state from one instance to another.
As soon as three MMs are involved at the same time, we might detect "maybe
mapped shared" although the folio is "mapped exclusively".
Example 1:
(1) App1 faults in a (shmem/file-backed) folio page -> Tracked as MM0
(2) App2 faults in a folio page -> Tracked as MM1
(4) App1 unmaps all folio pages
-> We will detect "mapped exclusively".
Example 2:
(1) App1 faults in a (shmem/file-backed) folio page -> Tracked as MM0
(2) App2 faults in a folio page -> Tracked as MM1
(3) App3 faults in a folio page -> No slot available, tracked as "unknown"
(4) App1 and App2 unmap all folio pages
-> We will detect "maybe mapped shared".
Make use of __always_inline to keep possible performance degradation when
(un)mapping large folios to a minimum.
Note: by squeezing the two flags into the "unsigned long" that stores the
MM ids, we can use non-atomic __bit_spin_unlock() and non-atomic
setting/clearing of the "maybe mapped shared" bit, effectively not adding
any new atomics on the hot path when updating the large mapcount + new
metadata, which further helps reduce the runtime overhead in
micro-benchmarks.
Link: https://lkml.kernel.org/r/20250303163014.1128035-13-david@redhat.com
Signed-off-by: David Hildenbrand <david@redhat.com>
Cc: Andy Lutomirks^H^Hski <luto@kernel.org>
Cc: Borislav Betkov <bp@alien8.de>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Jann Horn <jannh@google.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Lance Yang <ioworker0@gmail.com>
Cc: Liam Howlett <liam.howlett@oracle.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Matthew Wilcow (Oracle) <willy@infradead.org>
Cc: Michal Koutn <mkoutny@suse.com>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: tejun heo <tj@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Zefan Li <lizefan.x@bytedance.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-03-03 17:30:05 +01:00
|
|
|
|
2011-01-13 15:46:39 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
|
|
|
|
|
2011-01-13 15:47:07 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/thp: add CONFIG_TRANSPARENT_HUGEPAGE_NEVER option
Currently enabling THP support (CONFIG_TRANSPARENT_HUGEPAGE) requires
enabling either CONFIG_TRANSPARENT_HUGEPAGE_ALWAYS or
CONFIG_TRANSPARENT_HUGEPAGE_MADVISE, which both cause khugepaged starting
by default at kernel bootup. Add the third choice
CONFIG_TRANSPARENT_HUGEPAGE_NEVER, in line with the existing kernel
command line setting transparent_hugepage=never, to disable THP by default
(in particular, to prevent starting khugepaged by default) but still allow
enabling it at runtime via sysfs.
Rationale: khugepaged has its own non-negligible memory cost even if it is
not used by any applications, since it bumps up vm.min_free_kbytes to its
own required minimum in set_recommended_min_free_kbytes(). For example,
on a machine with 4GB RAM, with 3 mm zones and pageblock_order ==
MAX_ORDER, starting khugepaged causes vm.min_free_kbytes increase from 8MB
to 132MB.
So if we use THP on machines with e.g. >=8GB of memory for better
performance, but avoid using it on lower-memory machines to avoid its
memory overhead, then for the same reason we also want to avoid even
starting khugepaged on those <8GB machines. So with
CONFIG_TRANSPARENT_HUGEPAGE_NEVER we can use the same kernel image on both
>=8GB and <8GB machines, with THP support enabled but khugepaged not
started by default. The userspace can then decide to enable THP via sysfs
if needed, based on the total amount of memory.
This could also be achieved with the existing transparent_hugepage=never
setting in the kernel command line instead. But it seems cleaner to avoid
tweaking the command line for such a basic setting.
P.S. I see that CONFIG_TRANSPARENT_HUGEPAGE_NEVER was already proposed
in the past [1] but without an explanation of the purpose.
[1] https://lore.kernel.org/all/202211301651462590168@zte.com.cn/
Link: https://lkml.kernel.org/r/20231205170244.2746210-1-dmaluka@chromium.org
Link: https://lore.kernel.org/all/20231204163254.2636289-1-dmaluka@chromium.org/
Signed-off-by: Dmytro Maluka <dmaluka@chromium.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-12-05 18:02:44 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2011-01-13 15:47:07 -08:00
|
|
|
|
|
|
|
|
|
mm: shmem/tmpfs hugepage defaults config choice
Allow to override defaults for shemem and tmpfs at config time. This is
consistent with how transparent hugepages can be configured.
Same results can be achieved with the existing
'transparent_hugepage_shmem' and 'transparent_hugepage_tmpfs' settings in
the kernel command line, but it is more convenient to define basic
settings at config time instead of changing kernel command line later.
Defaults for shmem and tmpfs were not changed. They are remained the same
as before: 'never' for both cases. Options 'deny' and 'force' are omitted
intentionally since these are special values and supposed to be used for
emergencies or testing and are not expected to be permanent ones.
Primary motivation for adding config option is to enable policy
enforcement at build time. In large-scale production environments (Meta's
for example), the kernel configuration is often maintained centrally close
to the kernel code itself and owned by the kernel engineers, while boot
parameters are managed independently (e.g. by provisioning systems). In
such setups, the kernel build defines the supported and expected behavior
in a single place, but there is no reliable or uniform control over the
kernel command line options.
A build-time default allows kernel integrators to enforce a predictable
hugepage policy for shmem/tmpfs on a base layer, ensuring reproducible
behavior and avoiding configuration drift caused by possible boot-time
differences.
In short, primary benefit is mostly operational: it provides a way to
codify preferred policy in the kernel configuration, which is versioned,
reviewed, and tested as part of the kernel build process, rather than
depending on potentially variable boot parameters.
[d@ilvokhin.com: v2]
Link: https://lkml.kernel.org/r/aQECPpjd-fU_TC79@shell.ilvokhin.com
Link: https://lkml.kernel.org/r/aPpv8sAa2sYgNu3L@shell.ilvokhin.com
Signed-off-by: Dmitry Ilvokhin <d@ilvokhin.com>
Reviewed-by: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: Baolin Wang <baolin.wang@linux.alibaba.com>
Acked-by: Kiryl Shutsemau <kas@kernel.org>
Cc: David Hildenbrand <david@redhat.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Liam Howlett <liam.howlett@oracle.com>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-10-23 18:12:02 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm, THP, swap: delay splitting THP during swap out
Patch series "THP swap: Delay splitting THP during swapping out", v11.
This patchset is to optimize the performance of Transparent Huge Page
(THP) swap.
Recently, the performance of the storage devices improved so fast that
we cannot saturate the disk bandwidth with single logical CPU when do
page swap out even on a high-end server machine. Because the
performance of the storage device improved faster than that of single
logical CPU. And it seems that the trend will not change in the near
future. On the other hand, the THP becomes more and more popular
because of increased memory size. So it becomes necessary to optimize
THP swap performance.
The advantages of the THP swap support include:
- Batch the swap operations for the THP to reduce lock
acquiring/releasing, including allocating/freeing the swap space,
adding/deleting to/from the swap cache, and writing/reading the swap
space, etc. This will help improve the performance of the THP swap.
- The THP swap space read/write will be 2M sequential IO. It is
particularly helpful for the swap read, which are usually 4k random
IO. This will improve the performance of the THP swap too.
- It will help the memory fragmentation, especially when the THP is
heavily used by the applications. The 2M continuous pages will be
free up after THP swapping out.
- It will improve the THP utilization on the system with the swap
turned on. Because the speed for khugepaged to collapse the normal
pages into the THP is quite slow. After the THP is split during the
swapping out, it will take quite long time for the normal pages to
collapse back into the THP after being swapped in. The high THP
utilization helps the efficiency of the page based memory management
too.
There are some concerns regarding THP swap in, mainly because possible
enlarged read/write IO size (for swap in/out) may put more overhead on
the storage device. To deal with that, the THP swap in should be turned
on only when necessary. For example, it can be selected via
"always/never/madvise" logic, to be turned on globally, turned off
globally, or turned on only for VMA with MADV_HUGEPAGE, etc.
This patchset is the first step for the THP swap support. The plan is
to delay splitting THP step by step, finally avoid splitting THP during
the THP swapping out and swap out/in the THP as a whole.
As the first step, in this patchset, the splitting huge page is delayed
from almost the first step of swapping out to after allocating the swap
space for the THP and adding the THP into the swap cache. This will
reduce lock acquiring/releasing for the locks used for the swap cache
management.
With the patchset, the swap out throughput improves 15.5% (from about
3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case
with 8 processes. The test is done on a Xeon E5 v3 system. The swap
device used is a RAM simulated PMEM (persistent memory) device. To test
the sequential swapping out, the test case creates 8 processes, which
sequentially allocate and write to the anonymous pages until the RAM and
part of the swap device is used up.
This patch (of 5):
In this patch, splitting huge page is delayed from almost the first step
of swapping out to after allocating the swap space for the THP
(Transparent Huge Page) and adding the THP into the swap cache. This
will batch the corresponding operation, thus improve THP swap out
throughput.
This is the first step for the THP swap optimization. The plan is to
delay splitting the THP step by step and avoid splitting the THP
finally.
In this patch, one swap cluster is used to hold the contents of each THP
swapped out. So, the size of the swap cluster is changed to that of the
THP (Transparent Huge Page) on x86_64 architecture (512). For other
architectures which want such THP swap optimization,
ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for
the architecture. In effect, this will enlarge swap cluster size by 2
times on x86_64. Which may make it harder to find a free cluster when
the swap space becomes fragmented. So that, this may reduce the
continuous swap space allocation and sequential write in theory. The
performance test in 0day shows no regressions caused by this.
In the future of THP swap optimization, some information of the swapped
out THP (such as compound map count) will be recorded in the
swap_cluster_info data structure.
The mem cgroup swap accounting functions are enhanced to support charge
or uncharge a swap cluster backing a THP as a whole.
The swap cluster allocate/free functions are added to allocate/free a
swap cluster for a THP. A fair simple algorithm is used for swap
cluster allocation, that is, only the first swap device in priority list
will be tried to allocate the swap cluster. The function will fail if
the trying is not successful, and the caller will fallback to allocate a
single swap slot instead. This works good enough for normal cases. If
the difference of the number of the free swap clusters among multiple
swap devices is significant, it is possible that some THPs are split
earlier than necessary. For example, this could be caused by big size
difference among multiple swap devices.
The swap cache functions is enhanced to support add/delete THP to/from
the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be
enhanced in the future with multi-order radix tree. But because we will
split the THP soon during swapping out, that optimization doesn't make
much sense for this first step.
The THP splitting functions are enhanced to support to split THP in swap
cache during swapping out. The page lock will be held during allocating
the swap cluster, adding the THP into the swap cache and splitting the
THP. So in the code path other than swapping out, if the THP need to be
split, the PageSwapCache(THP) will be always false.
The swap cluster is only available for SSD, so the THP swap optimization
in this patchset has no effect for HDD.
[ying.huang@intel.com: fix two issues in THP optimize patch]
Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com
[hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size]
Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option]
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h]
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Shaohua Li <shli@kernel.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-06 15:37:18 -07:00
|
|
|
|
|
|
|
|
|
2026-03-31 08:07:30 +01:00
|
|
|
|
mm, THP, swap: delay splitting THP during swap out
Patch series "THP swap: Delay splitting THP during swapping out", v11.
This patchset is to optimize the performance of Transparent Huge Page
(THP) swap.
Recently, the performance of the storage devices improved so fast that
we cannot saturate the disk bandwidth with single logical CPU when do
page swap out even on a high-end server machine. Because the
performance of the storage device improved faster than that of single
logical CPU. And it seems that the trend will not change in the near
future. On the other hand, the THP becomes more and more popular
because of increased memory size. So it becomes necessary to optimize
THP swap performance.
The advantages of the THP swap support include:
- Batch the swap operations for the THP to reduce lock
acquiring/releasing, including allocating/freeing the swap space,
adding/deleting to/from the swap cache, and writing/reading the swap
space, etc. This will help improve the performance of the THP swap.
- The THP swap space read/write will be 2M sequential IO. It is
particularly helpful for the swap read, which are usually 4k random
IO. This will improve the performance of the THP swap too.
- It will help the memory fragmentation, especially when the THP is
heavily used by the applications. The 2M continuous pages will be
free up after THP swapping out.
- It will improve the THP utilization on the system with the swap
turned on. Because the speed for khugepaged to collapse the normal
pages into the THP is quite slow. After the THP is split during the
swapping out, it will take quite long time for the normal pages to
collapse back into the THP after being swapped in. The high THP
utilization helps the efficiency of the page based memory management
too.
There are some concerns regarding THP swap in, mainly because possible
enlarged read/write IO size (for swap in/out) may put more overhead on
the storage device. To deal with that, the THP swap in should be turned
on only when necessary. For example, it can be selected via
"always/never/madvise" logic, to be turned on globally, turned off
globally, or turned on only for VMA with MADV_HUGEPAGE, etc.
This patchset is the first step for the THP swap support. The plan is
to delay splitting THP step by step, finally avoid splitting THP during
the THP swapping out and swap out/in the THP as a whole.
As the first step, in this patchset, the splitting huge page is delayed
from almost the first step of swapping out to after allocating the swap
space for the THP and adding the THP into the swap cache. This will
reduce lock acquiring/releasing for the locks used for the swap cache
management.
With the patchset, the swap out throughput improves 15.5% (from about
3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case
with 8 processes. The test is done on a Xeon E5 v3 system. The swap
device used is a RAM simulated PMEM (persistent memory) device. To test
the sequential swapping out, the test case creates 8 processes, which
sequentially allocate and write to the anonymous pages until the RAM and
part of the swap device is used up.
This patch (of 5):
In this patch, splitting huge page is delayed from almost the first step
of swapping out to after allocating the swap space for the THP
(Transparent Huge Page) and adding the THP into the swap cache. This
will batch the corresponding operation, thus improve THP swap out
throughput.
This is the first step for the THP swap optimization. The plan is to
delay splitting the THP step by step and avoid splitting the THP
finally.
In this patch, one swap cluster is used to hold the contents of each THP
swapped out. So, the size of the swap cluster is changed to that of the
THP (Transparent Huge Page) on x86_64 architecture (512). For other
architectures which want such THP swap optimization,
ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for
the architecture. In effect, this will enlarge swap cluster size by 2
times on x86_64. Which may make it harder to find a free cluster when
the swap space becomes fragmented. So that, this may reduce the
continuous swap space allocation and sequential write in theory. The
performance test in 0day shows no regressions caused by this.
In the future of THP swap optimization, some information of the swapped
out THP (such as compound map count) will be recorded in the
swap_cluster_info data structure.
The mem cgroup swap accounting functions are enhanced to support charge
or uncharge a swap cluster backing a THP as a whole.
The swap cluster allocate/free functions are added to allocate/free a
swap cluster for a THP. A fair simple algorithm is used for swap
cluster allocation, that is, only the first swap device in priority list
will be tried to allocate the swap cluster. The function will fail if
the trying is not successful, and the caller will fallback to allocate a
single swap slot instead. This works good enough for normal cases. If
the difference of the number of the free swap clusters among multiple
swap devices is significant, it is possible that some THPs are split
earlier than necessary. For example, this could be caused by big size
difference among multiple swap devices.
The swap cache functions is enhanced to support add/delete THP to/from
the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be
enhanced in the future with multi-order radix tree. But because we will
split the THP soon during swapping out, that optimization doesn't make
much sense for this first step.
The THP splitting functions are enhanced to support to split THP in swap
cache during swapping out. The page lock will be held during allocating
the swap cluster, adding the THP into the swap cache and splitting the
THP. So in the code path other than swapping out, if the THP need to be
split, the PageSwapCache(THP) will be always false.
The swap cluster is only available for SSD, so the THP swap optimization
in this patchset has no effect for HDD.
[ying.huang@intel.com: fix two issues in THP optimize patch]
Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com
[hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size]
Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option]
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h]
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Shaohua Li <shli@kernel.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-06 15:37:18 -07:00
|
|
|
|
|
|
|
|
|
2018-08-17 15:49:41 -07:00
|
|
|
|
|
|
|
|
|
mm, THP, swap: delay splitting THP during swap out
Patch series "THP swap: Delay splitting THP during swapping out", v11.
This patchset is to optimize the performance of Transparent Huge Page
(THP) swap.
Recently, the performance of the storage devices improved so fast that
we cannot saturate the disk bandwidth with single logical CPU when do
page swap out even on a high-end server machine. Because the
performance of the storage device improved faster than that of single
logical CPU. And it seems that the trend will not change in the near
future. On the other hand, the THP becomes more and more popular
because of increased memory size. So it becomes necessary to optimize
THP swap performance.
The advantages of the THP swap support include:
- Batch the swap operations for the THP to reduce lock
acquiring/releasing, including allocating/freeing the swap space,
adding/deleting to/from the swap cache, and writing/reading the swap
space, etc. This will help improve the performance of the THP swap.
- The THP swap space read/write will be 2M sequential IO. It is
particularly helpful for the swap read, which are usually 4k random
IO. This will improve the performance of the THP swap too.
- It will help the memory fragmentation, especially when the THP is
heavily used by the applications. The 2M continuous pages will be
free up after THP swapping out.
- It will improve the THP utilization on the system with the swap
turned on. Because the speed for khugepaged to collapse the normal
pages into the THP is quite slow. After the THP is split during the
swapping out, it will take quite long time for the normal pages to
collapse back into the THP after being swapped in. The high THP
utilization helps the efficiency of the page based memory management
too.
There are some concerns regarding THP swap in, mainly because possible
enlarged read/write IO size (for swap in/out) may put more overhead on
the storage device. To deal with that, the THP swap in should be turned
on only when necessary. For example, it can be selected via
"always/never/madvise" logic, to be turned on globally, turned off
globally, or turned on only for VMA with MADV_HUGEPAGE, etc.
This patchset is the first step for the THP swap support. The plan is
to delay splitting THP step by step, finally avoid splitting THP during
the THP swapping out and swap out/in the THP as a whole.
As the first step, in this patchset, the splitting huge page is delayed
from almost the first step of swapping out to after allocating the swap
space for the THP and adding the THP into the swap cache. This will
reduce lock acquiring/releasing for the locks used for the swap cache
management.
With the patchset, the swap out throughput improves 15.5% (from about
3.73GB/s to about 4.31GB/s) in the vm-scalability swap-w-seq test case
with 8 processes. The test is done on a Xeon E5 v3 system. The swap
device used is a RAM simulated PMEM (persistent memory) device. To test
the sequential swapping out, the test case creates 8 processes, which
sequentially allocate and write to the anonymous pages until the RAM and
part of the swap device is used up.
This patch (of 5):
In this patch, splitting huge page is delayed from almost the first step
of swapping out to after allocating the swap space for the THP
(Transparent Huge Page) and adding the THP into the swap cache. This
will batch the corresponding operation, thus improve THP swap out
throughput.
This is the first step for the THP swap optimization. The plan is to
delay splitting the THP step by step and avoid splitting the THP
finally.
In this patch, one swap cluster is used to hold the contents of each THP
swapped out. So, the size of the swap cluster is changed to that of the
THP (Transparent Huge Page) on x86_64 architecture (512). For other
architectures which want such THP swap optimization,
ARCH_USES_THP_SWAP_CLUSTER needs to be selected in the Kconfig file for
the architecture. In effect, this will enlarge swap cluster size by 2
times on x86_64. Which may make it harder to find a free cluster when
the swap space becomes fragmented. So that, this may reduce the
continuous swap space allocation and sequential write in theory. The
performance test in 0day shows no regressions caused by this.
In the future of THP swap optimization, some information of the swapped
out THP (such as compound map count) will be recorded in the
swap_cluster_info data structure.
The mem cgroup swap accounting functions are enhanced to support charge
or uncharge a swap cluster backing a THP as a whole.
The swap cluster allocate/free functions are added to allocate/free a
swap cluster for a THP. A fair simple algorithm is used for swap
cluster allocation, that is, only the first swap device in priority list
will be tried to allocate the swap cluster. The function will fail if
the trying is not successful, and the caller will fallback to allocate a
single swap slot instead. This works good enough for normal cases. If
the difference of the number of the free swap clusters among multiple
swap devices is significant, it is possible that some THPs are split
earlier than necessary. For example, this could be caused by big size
difference among multiple swap devices.
The swap cache functions is enhanced to support add/delete THP to/from
the swap cache as a set of (HPAGE_PMD_NR) sub-pages. This may be
enhanced in the future with multi-order radix tree. But because we will
split the THP soon during swapping out, that optimization doesn't make
much sense for this first step.
The THP splitting functions are enhanced to support to split THP in swap
cache during swapping out. The page lock will be held during allocating
the swap cluster, adding the THP into the swap cache and splitting the
THP. So in the code path other than swapping out, if the THP need to be
split, the PageSwapCache(THP) will be always false.
The swap cluster is only available for SSD, so the THP swap optimization
in this patchset has no effect for HDD.
[ying.huang@intel.com: fix two issues in THP optimize patch]
Link: http://lkml.kernel.org/r/87k25ed8zo.fsf@yhuang-dev.intel.com
[hannes@cmpxchg.org: extensive cleanups and simplifications, reduce code size]
Link: http://lkml.kernel.org/r/20170515112522.32457-2-ying.huang@intel.com
Signed-off-by: "Huang, Ying" <ying.huang@intel.com>
Signed-off-by: Johannes Weiner <hannes@cmpxchg.org>
Suggested-by: Andrew Morton <akpm@linux-foundation.org> [for config option]
Acked-by: Kirill A. Shutemov <kirill.shutemov@linux.intel.com> [for changes in huge_memory.c and huge_mm.h]
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Ebru Akagunduz <ebru.akagunduz@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Cc: Hugh Dickins <hughd@google.com>
Cc: Shaohua Li <shli@kernel.org>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Rik van Riel <riel@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-06 15:37:18 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-03-03 17:30:08 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-05-19 14:08:53 -07:00
|
|
|
|
|
|
|
|
|
2025-03-03 17:30:08 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/Kconfig: CONFIG_PGTABLE_HAS_HUGE_LEAVES
Patch series "mm/gup: Unify hugetlb, part 2", v4.
The series removes the hugetlb slow gup path after a previous refactor
work [1], so that slow gup now uses the exact same path to process all
kinds of memory including hugetlb.
For the long term, we may want to remove most, if not all, call sites of
huge_pte_offset(). It'll be ideal if that API can be completely dropped
from arch hugetlb API. This series is one small step towards merging
hugetlb specific codes into generic mm paths. From that POV, this series
removes one reference to huge_pte_offset() out of many others.
One goal of such a route is that we can reconsider merging hugetlb
features like High Granularity Mapping (HGM). It was not accepted in the
past because it may add lots of hugetlb specific codes and make the mm
code even harder to maintain. With a merged codeset, features like HGM
can hopefully share some code with THP, legacy (PMD+) or modern
(continuous PTEs).
To make it work, the generic slow gup code will need to at least
understand hugepd, which is already done like so in fast-gup. Due to the
specialty of hugepd to be software-only solution (no hardware recognizes
the hugepd format, so it's purely artificial structures), there's chance
we can merge some or all hugepd formats with cont_pte in the future. That
question is yet unsettled from Power side to have an acknowledgement. As
of now for this series, I kept the hugepd handling because we may still
need to do so before getting a clearer picture of the future of hugepd.
The other reason is simply that we did it already for fast-gup and most
codes are still around to be reused. It'll make more sense to keep
slow/fast gup behave the same before a decision is made to remove hugepd.
There's one major difference for slow-gup on cont_pte / cont_pmd handling,
currently supported on three architectures (aarch64, riscv, ppc). Before
the series, slow gup will be able to recognize e.g. cont_pte entries with
the help of huge_pte_offset() when hstate is around. Now it's gone but
still working, by looking up pgtable entries one by one.
It's not ideal, but hopefully this change should not affect yet on major
workloads. There's some more information in the commit message of the
last patch. If this would be a concern, we can consider teaching slow gup
to recognize cont pte/pmd entries, and that should recover the lost
performance. But I doubt its necessity for now, so I kept it as simple as
it can be.
Patch layout
=============
Patch 1-8: Preparation works, or cleanups in relevant code paths
Patch 9-11: Teach slow gup with all kinds of huge entries (pXd, hugepd)
Patch 12: Drop hugetlb_follow_page_mask()
More information can be found in the commit messages of each patch.
[1] https://lore.kernel.org/all/20230628215310.73782-1-peterx@redhat.com
[2] https://lore.kernel.org/r/20240321215047.678172-1-peterx@redhat.com
Introduce a config option that will be selected as long as huge leaves are
involved in pgtable (thp or hugetlbfs). It would be useful to mark any
code with this new config that can process either hugetlb or thp pages in
any level that is higher than pte level.
Link: https://lkml.kernel.org/r/20240327152332.950956-1-peterx@redhat.com
Link: https://lkml.kernel.org/r/20240327152332.950956-2-peterx@redhat.com
Signed-off-by: Peter Xu <peterx@redhat.com>
Reviewed-by: Jason Gunthorpe <jgg@nvidia.com>
Tested-by: Ryan Roberts <ryan.roberts@arm.com>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Cc: Andrew Jones <andrew.jones@linux.dev>
Cc: Aneesh Kumar K.V (IBM) <aneesh.kumar@kernel.org>
Cc: Axel Rasmussen <axelrasmussen@google.com>
Cc: Christophe Leroy <christophe.leroy@csgroup.eu>
Cc: Christoph Hellwig <hch@infradead.org>
Cc: David Hildenbrand <david@redhat.com>
Cc: James Houghton <jthoughton@google.com>
Cc: John Hubbard <jhubbard@nvidia.com>
Cc: Kirill A. Shutemov <kirill@shutemov.name>
Cc: Lorenzo Stoakes <lstoakes@gmail.com>
Cc: Matthew Wilcox (Oracle) <willy@infradead.org>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: "Mike Rapoport (IBM)" <rppt@kernel.org>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Rik van Riel <riel@surriel.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Yang Shi <shy828301@gmail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-03-27 11:23:20 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: fix MAX_FOLIO_ORDER on powerpc configs with hugetlb
In the past, CONFIG_ARCH_HAS_GIGANTIC_PAGE indicated that we support
runtime allocation of gigantic hugetlb folios. In the meantime it evolved
into a generic way for the architecture to state that it supports gigantic
hugetlb folios.
In commit fae7d834c43c ("mm: add __dump_folio()") we started using
CONFIG_ARCH_HAS_GIGANTIC_PAGE to decide MAX_FOLIO_ORDER: whether we could
have folios larger than what the buddy can handle. In the context of that
commit, we started using MAX_FOLIO_ORDER to detect page corruptions when
dumping tail pages of folios. Before that commit, we assumed that we
cannot have folios larger than the highest buddy order, which was
obviously wrong.
In commit 7b4f21f5e038 ("mm/hugetlb: check for unreasonable folio sizes
when registering hstate"), we used MAX_FOLIO_ORDER to detect
inconsistencies, and in fact, we found some now.
Powerpc allows for configs that can allocate gigantic folio during boot
(not at runtime), that do not set CONFIG_ARCH_HAS_GIGANTIC_PAGE and can
exceed PUD_ORDER.
To fix it, let's make powerpc select CONFIG_ARCH_HAS_GIGANTIC_PAGE with
hugetlb on powerpc, and increase the maximum folio size with hugetlb to 16
GiB on 64bit (possible on arm64 and powerpc) and 1 GiB on 32 bit
(powerpc). Note that on some powerpc configurations, whether we actually
have gigantic pages depends on the setting of CONFIG_ARCH_FORCE_MAX_ORDER,
but there is nothing really problematic about setting it unconditionally:
we just try to keep the value small so we can better detect problems in
__dump_folio() and inconsistencies around the expected largest folio in
the system.
Ideally, we'd have a better way to obtain the maximum hugetlb folio size
and detect ourselves whether we really end up with gigantic folios. Let's
defer bigger changes and fix the warnings first.
While at it, handle gigantic DAX folios more clearly: DAX can only end up
creating gigantic folios with HAVE_ARCH_TRANSPARENT_HUGEPAGE_PUD.
Add a new Kconfig option HAVE_GIGANTIC_FOLIOS to make both cases clearer.
In particular, worry about ARCH_HAS_GIGANTIC_PAGE only with HUGETLB_PAGE.
Note: with enabling CONFIG_ARCH_HAS_GIGANTIC_PAGE on powerpc, we will now
also allow for runtime allocations of folios in some more powerpc configs.
I don't think this is a problem, but if it is we could handle it through
__HAVE_ARCH_GIGANTIC_PAGE_RUNTIME_SUPPORTED.
While __dump_page()/__dump_folio was also problematic (not handling
dumping of tail pages of such gigantic folios correctly), it doesn't seem
critical enough to mark it as a fix.
Link: https://lkml.kernel.org/r/20251114214920.2550676-1-david@kernel.org
Fixes: 7b4f21f5e038 ("mm/hugetlb: check for unreasonable folio sizes when registering hstate")
Reported-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Closes: https://lore.kernel.org/r/3e043453-3f27-48ad-b987-cc39f523060a@csgroup.eu/
Reported-by: Sourabh Jain <sourabhjain@linux.ibm.com>
Closes: https://lore.kernel.org/r/94377f5c-d4f0-4c0f-b0f6-5bf1cd7305b1@linux.ibm.com/
Signed-off-by: David Hildenbrand (Red Hat) <david@kernel.org>
Cc: Ritesh Harjani (IBM) <ritesh.list@gmail.com>
Cc: Madhavan Srinivasan <maddy@linux.ibm.com>
Cc: Donet Tom <donettom@linux.ibm.com>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Nicholas Piggin <npiggin@gmail.com>
Cc: Christophe Leroy <christophe.leroy@csgroup.eu>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: "Liam R. Howlett" <Liam.Howlett@oracle.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Michal Hocko <mhocko@suse.com>
Cc: Nathan Chancellor <nathan@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-11-14 22:49:20 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-10-22 16:26:33 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: introduce ARCH_SUPPORTS_HUGE_PFNMAP and special bits to pmd/pud
Patch series "mm: Support huge pfnmaps", v2.
Overview
========
This series implements huge pfnmaps support for mm in general. Huge
pfnmap allows e.g. VM_PFNMAP vmas to map in either PMD or PUD levels,
similar to what we do with dax / thp / hugetlb so far to benefit from TLB
hits. Now we extend that idea to PFN mappings, e.g. PCI MMIO bars where
it can grow as large as 8GB or even bigger.
Currently, only x86_64 (1G+2M) and arm64 (2M) are supported. The last
patch (from Alex Williamson) will be the first user of huge pfnmap, so as
to enable vfio-pci driver to fault in huge pfn mappings.
Implementation
==============
In reality, it's relatively simple to add such support comparing to many
other types of mappings, because of PFNMAP's specialties when there's no
vmemmap backing it, so that most of the kernel routines on huge mappings
should simply already fail for them, like GUPs or old-school follow_page()
(which is recently rewritten to be folio_walk* APIs by David).
One trick here is that we're still unmature on PUDs in generic paths here
and there, as DAX is so far the only user. This patchset will add the 2nd
user of it. Hugetlb can be a 3rd user if the hugetlb unification work can
go on smoothly, but to be discussed later.
The other trick is how to allow gup-fast working for such huge mappings
even if there's no direct sign of knowing whether it's a normal page or
MMIO mapping. This series chose to keep the pte_special solution, so that
it reuses similar idea on setting a special bit to pfnmap PMDs/PUDs so
that gup-fast will be able to identify them and fail properly.
Along the way, we'll also notice that the major pgtable pfn walker, aka,
follow_pte(), will need to retire soon due to the fact that it only works
with ptes. A new set of simple API is introduced (follow_pfnmap* API) to
be able to do whatever follow_pte() can already do, plus that it can also
process huge pfnmaps now. Half of this series is about that and
converting all existing pfnmap walkers to use the new API properly.
Hopefully the new API also looks better to avoid exposing e.g. pgtable
lock details into the callers, so that it can be used in an even more
straightforward way.
Here, three more options will be introduced and involved in huge pfnmap:
- ARCH_SUPPORTS_HUGE_PFNMAP
Arch developers will need to select this option when huge pfnmap is
supported in arch's Kconfig. After this patchset applied, both x86_64
and arm64 will start to enable it by default.
- ARCH_SUPPORTS_PMD_PFNMAP / ARCH_SUPPORTS_PUD_PFNMAP
These options are for driver developers to identify whether current
arch / config supports huge pfnmaps, making decision on whether it can
use the huge pfnmap APIs to inject them. One can refer to the last
vfio-pci patch from Alex on the use of them properly in a device
driver.
So after the whole set applied, and if one would enable some dynamic debug
lines in vfio-pci core files, we should observe things like:
vfio-pci 0000:00:06.0: vfio_pci_mmap_huge_fault(,order = 9) BAR 0 page offset 0x0: 0x100
vfio-pci 0000:00:06.0: vfio_pci_mmap_huge_fault(,order = 9) BAR 0 page offset 0x200: 0x100
vfio-pci 0000:00:06.0: vfio_pci_mmap_huge_fault(,order = 9) BAR 0 page offset 0x400: 0x100
In this specific case, it says that vfio-pci faults in PMDs properly for a
few BAR0 offsets.
Patch Layout
============
Patch 1: Introduce the new options mentioned above for huge PFNMAPs
Patch 2: A tiny cleanup
Patch 3-8: Preparation patches for huge pfnmap (include introduce
special bit for pmd/pud)
Patch 9-16: Introduce follow_pfnmap*() API, use it everywhere, and
then drop follow_pte() API
Patch 17: Add huge pfnmap support for x86_64
Patch 18: Add huge pfnmap support for arm64
Patch 19: Add vfio-pci support for all kinds of huge pfnmaps (Alex)
TODO
====
More architectures / More page sizes
------------------------------------
Currently only x86_64 (2M+1G) and arm64 (2M) are supported. There seems
to have plan to support arm64 1G later on top of this series [2].
Any arch will need to first support THP / THP_1G, then provide a special
bit in pmds/puds to support huge pfnmaps.
remap_pfn_range() support
-------------------------
Currently, remap_pfn_range() still only maps PTEs. With the new option,
remap_pfn_range() can logically start to inject either PMDs or PUDs when
the alignment requirements match on the VAs.
When the support is there, it should be able to silently benefit all
drivers that is using remap_pfn_range() in its mmap() handler on better
TLB hit rate and overall faster MMIO accesses similar to processor on
hugepages.
More driver support
-------------------
VFIO is so far the only consumer for the huge pfnmaps after this series
applied. Besides above remap_pfn_range() generic optimization, device
driver can also try to optimize its mmap() on a better VA alignment for
either PMD/PUD sizes. This may, iiuc, normally require userspace changes,
as the driver doesn't normally decide the VA to map a bar. But I don't
think I know all the drivers to know the full picture.
Credits all go to Alex on help testing the GPU/NIC use cases above.
[0] https://lore.kernel.org/r/73ad9540-3fb8-4154-9a4f-30a0a2b03d41@lucifer.local
[1] https://lore.kernel.org/r/20240807194812.819412-1-peterx@redhat.com
[2] https://lore.kernel.org/r/498e0731-81a4-4f75-95b4-a8ad0bcc7665@huawei.com
This patch (of 19):
This patch introduces the option to introduce special pte bit into
pmd/puds. Archs can start to define pmd_special / pud_special when
supported by selecting the new option. Per-arch support will be added
later.
Before that, create fallbacks for these helpers so that they are always
available.
Link: https://lkml.kernel.org/r/20240826204353.2228736-1-peterx@redhat.com
Link: https://lkml.kernel.org/r/20240826204353.2228736-2-peterx@redhat.com
Signed-off-by: Peter Xu <peterx@redhat.com>
Cc: Alexander Gordeev <agordeev@linux.ibm.com>
Cc: Alex Williamson <alex.williamson@redhat.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Christian Borntraeger <borntraeger@linux.ibm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Gavin Shan <gshan@redhat.com>
Cc: Gerald Schaefer <gerald.schaefer@linux.ibm.com>
Cc: Heiko Carstens <hca@linux.ibm.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Jason Gunthorpe <jgg@nvidia.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Niklas Schnelle <schnelle@linux.ibm.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Ryan Roberts <ryan.roberts@arm.com>
Cc: Sean Christopherson <seanjc@google.com>
Cc: Sven Schnelle <svens@linux.ibm.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Vasily Gorbik <gor@linux.ibm.com>
Cc: Will Deacon <will@kernel.org>
Cc: Zi Yan <ziy@nvidia.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-08-26 16:43:35 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-06-18 09:58:09 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2010-09-03 18:22:48 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-11-30 17:29:54 +00:00
|
|
|
|
2010-09-03 18:22:48 +02:00
|
|
|
|
|
|
|
|
|
2011-05-26 10:01:36 -06:00
|
|
|
|
mm: percpu: generalize percpu related config
Patch series "mm: percpu: Cleanup percpu first chunk function".
When supporting page mapping percpu first chunk allocator on arm64, we
found there are lots of duplicated codes in percpu embed/page first chunk
allocator. This patchset is aimed to cleanup them and should no function
change.
The currently supported status about 'embed' and 'page' in Archs shows
below,
embed: NEED_PER_CPU_PAGE_FIRST_CHUNK
page: NEED_PER_CPU_EMBED_FIRST_CHUNK
embed page
------------------------
arm64 Y Y
mips Y N
powerpc Y Y
riscv Y N
sparc Y Y
x86 Y Y
------------------------
There are two interfaces about percpu first chunk allocator,
extern int __init pcpu_embed_first_chunk(size_t reserved_size, size_t dyn_size,
size_t atom_size,
pcpu_fc_cpu_distance_fn_t cpu_distance_fn,
- pcpu_fc_alloc_fn_t alloc_fn,
- pcpu_fc_free_fn_t free_fn);
+ pcpu_fc_cpu_to_node_fn_t cpu_to_nd_fn);
extern int __init pcpu_page_first_chunk(size_t reserved_size,
- pcpu_fc_alloc_fn_t alloc_fn,
- pcpu_fc_free_fn_t free_fn,
- pcpu_fc_populate_pte_fn_t populate_pte_fn);
+ pcpu_fc_cpu_to_node_fn_t cpu_to_nd_fn);
The pcpu_fc_alloc_fn_t/pcpu_fc_free_fn_t is killed, we provide generic
pcpu_fc_alloc() and pcpu_fc_free() function, which are called in the
pcpu_embed/page_first_chunk().
1) For pcpu_embed_first_chunk(), pcpu_fc_cpu_to_node_fn_t is needed to be
provided when archs supported NUMA.
2) For pcpu_page_first_chunk(), the pcpu_fc_populate_pte_fn_t is killed too,
a generic pcpu_populate_pte() which marked '__weak' is provided, if you
need a different function to populate pte on the arch(like x86), please
provide its own implementation.
[1] https://github.com/kevin78/linux.git percpu-cleanup
This patch (of 4):
The HAVE_SETUP_PER_CPU_AREA/NEED_PER_CPU_EMBED_FIRST_CHUNK/
NEED_PER_CPU_PAGE_FIRST_CHUNK/USE_PERCPU_NUMA_NODE_ID configs, which have
duplicate definitions on platforms that subscribe it.
Move them into mm, drop these redundant definitions and instead just
select it on applicable platforms.
Link: https://lkml.kernel.org/r/20211216112359.103822-1-wangkefeng.wang@huawei.com
Link: https://lkml.kernel.org/r/20211216112359.103822-2-wangkefeng.wang@huawei.com
Signed-off-by: Kefeng Wang <wangkefeng.wang@huawei.com>
Acked-by: Catalin Marinas <catalin.marinas@arm.com> [arm64]
Cc: Will Deacon <will@kernel.org>
Cc: Thomas Bogendoerfer <tsbogend@alpha.franken.de>
Cc: Michael Ellerman <mpe@ellerman.id.au>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Paul Walmsley <paul.walmsley@sifive.com>
Cc: Palmer Dabbelt <palmer@dabbelt.com>
Cc: Albert Ou <aou@eecs.berkeley.edu>
Cc: "David S. Miller" <davem@davemloft.net>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Christoph Lameter <cl@linux.com>
Cc: Dennis Zhou <dennis@kernel.org>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Cc: "Rafael J. Wysocki" <rafael@kernel.org>
Cc: Tejun Heo <tj@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-01-19 18:07:41 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-07-02 11:15:15 +05:30
|
|
|
|
|
|
|
|
|
2018-10-30 15:07:44 -07:00
|
|
|
|
2013-07-02 11:15:15 +05:30
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-04-14 15:44:57 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-05-04 18:37:28 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-08-06 16:05:25 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-02-05 10:49:29 +05:30
|
|
|
|
|
|
|
|
|
2014-08-06 16:05:25 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-02-05 10:49:29 +05:30
|
|
|
|
2014-08-06 16:05:25 -07:00
|
|
|
|
mm: add CONFIG_PAGE_BLOCK_ORDER to select page block order
Problem: On large page size configurations (16KiB, 64KiB), the CMA
alignment requirement (CMA_MIN_ALIGNMENT_BYTES) increases considerably,
and this causes the CMA reservations to be larger than necessary. This
means that system will have less available MIGRATE_UNMOVABLE and
MIGRATE_RECLAIMABLE page blocks since MIGRATE_CMA can't fallback to them.
The CMA_MIN_ALIGNMENT_BYTES increases because it depends on MAX_PAGE_ORDER
which depends on ARCH_FORCE_MAX_ORDER. The value of ARCH_FORCE_MAX_ORDER
increases on 16k and 64k kernels.
For example, in ARM, the CMA alignment requirement when:
- CONFIG_ARCH_FORCE_MAX_ORDER default value is used
- CONFIG_TRANSPARENT_HUGEPAGE is set:
PAGE_SIZE | MAX_PAGE_ORDER | pageblock_order | CMA_MIN_ALIGNMENT_BYTES
-----------------------------------------------------------------------
4KiB | 10 | 9 | 4KiB * (2 ^ 9) = 2MiB
16Kib | 11 | 11 | 16KiB * (2 ^ 11) = 32MiB
64KiB | 13 | 13 | 64KiB * (2 ^ 13) = 512MiB
There are some extreme cases for the CMA alignment requirement when:
- CONFIG_ARCH_FORCE_MAX_ORDER maximum value is set
- CONFIG_TRANSPARENT_HUGEPAGE is NOT set:
- CONFIG_HUGETLB_PAGE is NOT set
PAGE_SIZE | MAX_PAGE_ORDER | pageblock_order | CMA_MIN_ALIGNMENT_BYTES
------------------------------------------------------------------------
4KiB | 15 | 15 | 4KiB * (2 ^ 15) = 128MiB
16Kib | 13 | 13 | 16KiB * (2 ^ 13) = 128MiB
64KiB | 13 | 13 | 64KiB * (2 ^ 13) = 512MiB
This affects the CMA reservations for the drivers. If a driver in a
4KiB kernel needs 4MiB of CMA memory, in a 16KiB kernel, the minimal
reservation has to be 32MiB due to the alignment requirements:
reserved-memory {
...
cma_test_reserve: cma_test_reserve {
compatible = "shared-dma-pool";
size = <0x0 0x400000>; /* 4 MiB */
...
};
};
reserved-memory {
...
cma_test_reserve: cma_test_reserve {
compatible = "shared-dma-pool";
size = <0x0 0x2000000>; /* 32 MiB */
...
};
};
Solution: Add a new config CONFIG_PAGE_BLOCK_ORDER that allows to set the
page block order in all the architectures. The maximum page block order
will be given by ARCH_FORCE_MAX_ORDER.
By default, CONFIG_PAGE_BLOCK_ORDER will have the same value that
ARCH_FORCE_MAX_ORDER. This will make sure that current kernel
configurations won't be affected by this change. It is a opt-in change.
This patch will allow to have the same CMA alignment requirements for
large page sizes (16KiB, 64KiB) as that in 4kb kernels by setting a lower
pageblock_order.
Tests:
- Verified that HugeTLB pages work when pageblock_order is 1, 7, 10 on
4k and 16k kernels.
- Verified that Transparent Huge Pages work when pageblock_order is 1,
7, 10 on 4k and 16k kernels.
- Verified that dma-buf heaps allocations work when pageblock_order is
1, 7, 10 on 4k and 16k kernels.
Benchmarks:
The benchmarks compare 16kb kernels with pageblock_order 10 and 7. The
reason for the pageblock_order 7 is because this value makes the min CMA
alignment requirement the same as that in 4kb kernels (2MB).
- Perform 100K dma-buf heaps (/dev/dma_heap/system) allocations of
SZ_8M, SZ_4M, SZ_2M, SZ_1M, SZ_64, SZ_8, SZ_4. Use simpleperf
(https://developer.android.com/ndk/guides/simpleperf) to measure the #
of instructions and page-faults on 16k kernels. The benchmark was
executed 10 times. The averages are below:
# instructions | #page-faults
order 10 | order 7 | order 10 | order 7
--------------------------------------------------------
13,891,765,770 | 11,425,777,314 | 220 | 217
14,456,293,487 | 12,660,819,302 | 224 | 219
13,924,261,018 | 13,243,970,736 | 217 | 221
13,910,886,504 | 13,845,519,630 | 217 | 221
14,388,071,190 | 13,498,583,098 | 223 | 224
13,656,442,167 | 12,915,831,681 | 216 | 218
13,300,268,343 | 12,930,484,776 | 222 | 218
13,625,470,223 | 14,234,092,777 | 219 | 218
13,508,964,965 | 13,432,689,094 | 225 | 219
13,368,950,667 | 13,683,587,37 | 219 | 225
-------------------------------------------------------------------
13,803,137,433 | 13,131,974,268 | 220 | 220 Averages
There were 4.85% #instructions when order was 7, in comparison with order
10.
13,803,137,433 - 13,131,974,268 = -671,163,166 (-4.86%)
The number of page faults in order 7 and 10 were the same.
These results didn't show any significant regression when the
pageblock_order is set to 7 on 16kb kernels.
- Run speedometer 3.1 (https://browserbench.org/Speedometer3.1/) 5 times
on the 16k kernels with pageblock_order 7 and 10.
order 10 | order 7 | order 7 - order 10 | (order 7 - order 10) %
-------------------------------------------------------------------
15.8 | 16.4 | 0.6 | 3.80%
16.4 | 16.2 | -0.2 | -1.22%
16.6 | 16.3 | -0.3 | -1.81%
16.8 | 16.3 | -0.5 | -2.98%
16.6 | 16.8 | 0.2 | 1.20%
-------------------------------------------------------------------
16.44 16.4 -0.04 -0.24% Averages
The results didn't show any significant regression when the
pageblock_order is set to 7 on 16kb kernels.
Link: https://lkml.kernel.org/r/20250521215807.1860663-1-jyescas@google.com
Signed-off-by: Juan Yescas <jyescas@google.com>
Acked-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Liam R. Howlett <Liam.Howlett@oracle.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Minchan Kim <minchan@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-05-21 14:57:45 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-06-04 17:14:27 -04:00
|
|
|
|
|
|
|
|
|
mm: add CONFIG_PAGE_BLOCK_ORDER to select page block order
Problem: On large page size configurations (16KiB, 64KiB), the CMA
alignment requirement (CMA_MIN_ALIGNMENT_BYTES) increases considerably,
and this causes the CMA reservations to be larger than necessary. This
means that system will have less available MIGRATE_UNMOVABLE and
MIGRATE_RECLAIMABLE page blocks since MIGRATE_CMA can't fallback to them.
The CMA_MIN_ALIGNMENT_BYTES increases because it depends on MAX_PAGE_ORDER
which depends on ARCH_FORCE_MAX_ORDER. The value of ARCH_FORCE_MAX_ORDER
increases on 16k and 64k kernels.
For example, in ARM, the CMA alignment requirement when:
- CONFIG_ARCH_FORCE_MAX_ORDER default value is used
- CONFIG_TRANSPARENT_HUGEPAGE is set:
PAGE_SIZE | MAX_PAGE_ORDER | pageblock_order | CMA_MIN_ALIGNMENT_BYTES
-----------------------------------------------------------------------
4KiB | 10 | 9 | 4KiB * (2 ^ 9) = 2MiB
16Kib | 11 | 11 | 16KiB * (2 ^ 11) = 32MiB
64KiB | 13 | 13 | 64KiB * (2 ^ 13) = 512MiB
There are some extreme cases for the CMA alignment requirement when:
- CONFIG_ARCH_FORCE_MAX_ORDER maximum value is set
- CONFIG_TRANSPARENT_HUGEPAGE is NOT set:
- CONFIG_HUGETLB_PAGE is NOT set
PAGE_SIZE | MAX_PAGE_ORDER | pageblock_order | CMA_MIN_ALIGNMENT_BYTES
------------------------------------------------------------------------
4KiB | 15 | 15 | 4KiB * (2 ^ 15) = 128MiB
16Kib | 13 | 13 | 16KiB * (2 ^ 13) = 128MiB
64KiB | 13 | 13 | 64KiB * (2 ^ 13) = 512MiB
This affects the CMA reservations for the drivers. If a driver in a
4KiB kernel needs 4MiB of CMA memory, in a 16KiB kernel, the minimal
reservation has to be 32MiB due to the alignment requirements:
reserved-memory {
...
cma_test_reserve: cma_test_reserve {
compatible = "shared-dma-pool";
size = <0x0 0x400000>; /* 4 MiB */
...
};
};
reserved-memory {
...
cma_test_reserve: cma_test_reserve {
compatible = "shared-dma-pool";
size = <0x0 0x2000000>; /* 32 MiB */
...
};
};
Solution: Add a new config CONFIG_PAGE_BLOCK_ORDER that allows to set the
page block order in all the architectures. The maximum page block order
will be given by ARCH_FORCE_MAX_ORDER.
By default, CONFIG_PAGE_BLOCK_ORDER will have the same value that
ARCH_FORCE_MAX_ORDER. This will make sure that current kernel
configurations won't be affected by this change. It is a opt-in change.
This patch will allow to have the same CMA alignment requirements for
large page sizes (16KiB, 64KiB) as that in 4kb kernels by setting a lower
pageblock_order.
Tests:
- Verified that HugeTLB pages work when pageblock_order is 1, 7, 10 on
4k and 16k kernels.
- Verified that Transparent Huge Pages work when pageblock_order is 1,
7, 10 on 4k and 16k kernels.
- Verified that dma-buf heaps allocations work when pageblock_order is
1, 7, 10 on 4k and 16k kernels.
Benchmarks:
The benchmarks compare 16kb kernels with pageblock_order 10 and 7. The
reason for the pageblock_order 7 is because this value makes the min CMA
alignment requirement the same as that in 4kb kernels (2MB).
- Perform 100K dma-buf heaps (/dev/dma_heap/system) allocations of
SZ_8M, SZ_4M, SZ_2M, SZ_1M, SZ_64, SZ_8, SZ_4. Use simpleperf
(https://developer.android.com/ndk/guides/simpleperf) to measure the #
of instructions and page-faults on 16k kernels. The benchmark was
executed 10 times. The averages are below:
# instructions | #page-faults
order 10 | order 7 | order 10 | order 7
--------------------------------------------------------
13,891,765,770 | 11,425,777,314 | 220 | 217
14,456,293,487 | 12,660,819,302 | 224 | 219
13,924,261,018 | 13,243,970,736 | 217 | 221
13,910,886,504 | 13,845,519,630 | 217 | 221
14,388,071,190 | 13,498,583,098 | 223 | 224
13,656,442,167 | 12,915,831,681 | 216 | 218
13,300,268,343 | 12,930,484,776 | 222 | 218
13,625,470,223 | 14,234,092,777 | 219 | 218
13,508,964,965 | 13,432,689,094 | 225 | 219
13,368,950,667 | 13,683,587,37 | 219 | 225
-------------------------------------------------------------------
13,803,137,433 | 13,131,974,268 | 220 | 220 Averages
There were 4.85% #instructions when order was 7, in comparison with order
10.
13,803,137,433 - 13,131,974,268 = -671,163,166 (-4.86%)
The number of page faults in order 7 and 10 were the same.
These results didn't show any significant regression when the
pageblock_order is set to 7 on 16kb kernels.
- Run speedometer 3.1 (https://browserbench.org/Speedometer3.1/) 5 times
on the 16k kernels with pageblock_order 7 and 10.
order 10 | order 7 | order 7 - order 10 | (order 7 - order 10) %
-------------------------------------------------------------------
15.8 | 16.4 | 0.6 | 3.80%
16.4 | 16.2 | -0.2 | -1.22%
16.6 | 16.3 | -0.3 | -1.81%
16.8 | 16.3 | -0.5 | -2.98%
16.6 | 16.8 | 0.2 | 1.20%
-------------------------------------------------------------------
16.44 16.4 -0.04 -0.24% Averages
The results didn't show any significant regression when the
pageblock_order is set to 7 on 16kb kernels.
Link: https://lkml.kernel.org/r/20250521215807.1860663-1-jyescas@google.com
Signed-off-by: Juan Yescas <jyescas@google.com>
Acked-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Liam R. Howlett <Liam.Howlett@oracle.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Minchan Kim <minchan@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-05-21 14:57:45 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-06-04 17:14:27 -04:00
|
|
|
|
|
|
|
|
|
mm: add CONFIG_PAGE_BLOCK_ORDER to select page block order
Problem: On large page size configurations (16KiB, 64KiB), the CMA
alignment requirement (CMA_MIN_ALIGNMENT_BYTES) increases considerably,
and this causes the CMA reservations to be larger than necessary. This
means that system will have less available MIGRATE_UNMOVABLE and
MIGRATE_RECLAIMABLE page blocks since MIGRATE_CMA can't fallback to them.
The CMA_MIN_ALIGNMENT_BYTES increases because it depends on MAX_PAGE_ORDER
which depends on ARCH_FORCE_MAX_ORDER. The value of ARCH_FORCE_MAX_ORDER
increases on 16k and 64k kernels.
For example, in ARM, the CMA alignment requirement when:
- CONFIG_ARCH_FORCE_MAX_ORDER default value is used
- CONFIG_TRANSPARENT_HUGEPAGE is set:
PAGE_SIZE | MAX_PAGE_ORDER | pageblock_order | CMA_MIN_ALIGNMENT_BYTES
-----------------------------------------------------------------------
4KiB | 10 | 9 | 4KiB * (2 ^ 9) = 2MiB
16Kib | 11 | 11 | 16KiB * (2 ^ 11) = 32MiB
64KiB | 13 | 13 | 64KiB * (2 ^ 13) = 512MiB
There are some extreme cases for the CMA alignment requirement when:
- CONFIG_ARCH_FORCE_MAX_ORDER maximum value is set
- CONFIG_TRANSPARENT_HUGEPAGE is NOT set:
- CONFIG_HUGETLB_PAGE is NOT set
PAGE_SIZE | MAX_PAGE_ORDER | pageblock_order | CMA_MIN_ALIGNMENT_BYTES
------------------------------------------------------------------------
4KiB | 15 | 15 | 4KiB * (2 ^ 15) = 128MiB
16Kib | 13 | 13 | 16KiB * (2 ^ 13) = 128MiB
64KiB | 13 | 13 | 64KiB * (2 ^ 13) = 512MiB
This affects the CMA reservations for the drivers. If a driver in a
4KiB kernel needs 4MiB of CMA memory, in a 16KiB kernel, the minimal
reservation has to be 32MiB due to the alignment requirements:
reserved-memory {
...
cma_test_reserve: cma_test_reserve {
compatible = "shared-dma-pool";
size = <0x0 0x400000>; /* 4 MiB */
...
};
};
reserved-memory {
...
cma_test_reserve: cma_test_reserve {
compatible = "shared-dma-pool";
size = <0x0 0x2000000>; /* 32 MiB */
...
};
};
Solution: Add a new config CONFIG_PAGE_BLOCK_ORDER that allows to set the
page block order in all the architectures. The maximum page block order
will be given by ARCH_FORCE_MAX_ORDER.
By default, CONFIG_PAGE_BLOCK_ORDER will have the same value that
ARCH_FORCE_MAX_ORDER. This will make sure that current kernel
configurations won't be affected by this change. It is a opt-in change.
This patch will allow to have the same CMA alignment requirements for
large page sizes (16KiB, 64KiB) as that in 4kb kernels by setting a lower
pageblock_order.
Tests:
- Verified that HugeTLB pages work when pageblock_order is 1, 7, 10 on
4k and 16k kernels.
- Verified that Transparent Huge Pages work when pageblock_order is 1,
7, 10 on 4k and 16k kernels.
- Verified that dma-buf heaps allocations work when pageblock_order is
1, 7, 10 on 4k and 16k kernels.
Benchmarks:
The benchmarks compare 16kb kernels with pageblock_order 10 and 7. The
reason for the pageblock_order 7 is because this value makes the min CMA
alignment requirement the same as that in 4kb kernels (2MB).
- Perform 100K dma-buf heaps (/dev/dma_heap/system) allocations of
SZ_8M, SZ_4M, SZ_2M, SZ_1M, SZ_64, SZ_8, SZ_4. Use simpleperf
(https://developer.android.com/ndk/guides/simpleperf) to measure the #
of instructions and page-faults on 16k kernels. The benchmark was
executed 10 times. The averages are below:
# instructions | #page-faults
order 10 | order 7 | order 10 | order 7
--------------------------------------------------------
13,891,765,770 | 11,425,777,314 | 220 | 217
14,456,293,487 | 12,660,819,302 | 224 | 219
13,924,261,018 | 13,243,970,736 | 217 | 221
13,910,886,504 | 13,845,519,630 | 217 | 221
14,388,071,190 | 13,498,583,098 | 223 | 224
13,656,442,167 | 12,915,831,681 | 216 | 218
13,300,268,343 | 12,930,484,776 | 222 | 218
13,625,470,223 | 14,234,092,777 | 219 | 218
13,508,964,965 | 13,432,689,094 | 225 | 219
13,368,950,667 | 13,683,587,37 | 219 | 225
-------------------------------------------------------------------
13,803,137,433 | 13,131,974,268 | 220 | 220 Averages
There were 4.85% #instructions when order was 7, in comparison with order
10.
13,803,137,433 - 13,131,974,268 = -671,163,166 (-4.86%)
The number of page faults in order 7 and 10 were the same.
These results didn't show any significant regression when the
pageblock_order is set to 7 on 16kb kernels.
- Run speedometer 3.1 (https://browserbench.org/Speedometer3.1/) 5 times
on the 16k kernels with pageblock_order 7 and 10.
order 10 | order 7 | order 7 - order 10 | (order 7 - order 10) %
-------------------------------------------------------------------
15.8 | 16.4 | 0.6 | 3.80%
16.4 | 16.2 | -0.2 | -1.22%
16.6 | 16.3 | -0.3 | -1.81%
16.8 | 16.3 | -0.5 | -2.98%
16.6 | 16.8 | 0.2 | 1.20%
-------------------------------------------------------------------
16.44 16.4 -0.04 -0.24% Averages
The results didn't show any significant regression when the
pageblock_order is set to 7 on 16kb kernels.
Link: https://lkml.kernel.org/r/20250521215807.1860663-1-jyescas@google.com
Signed-off-by: Juan Yescas <jyescas@google.com>
Acked-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Liam R. Howlett <Liam.Howlett@oracle.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Minchan Kim <minchan@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-05-21 14:57:45 -07:00
|
|
|
|
2025-06-04 17:14:27 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: add CONFIG_PAGE_BLOCK_ORDER to select page block order
Problem: On large page size configurations (16KiB, 64KiB), the CMA
alignment requirement (CMA_MIN_ALIGNMENT_BYTES) increases considerably,
and this causes the CMA reservations to be larger than necessary. This
means that system will have less available MIGRATE_UNMOVABLE and
MIGRATE_RECLAIMABLE page blocks since MIGRATE_CMA can't fallback to them.
The CMA_MIN_ALIGNMENT_BYTES increases because it depends on MAX_PAGE_ORDER
which depends on ARCH_FORCE_MAX_ORDER. The value of ARCH_FORCE_MAX_ORDER
increases on 16k and 64k kernels.
For example, in ARM, the CMA alignment requirement when:
- CONFIG_ARCH_FORCE_MAX_ORDER default value is used
- CONFIG_TRANSPARENT_HUGEPAGE is set:
PAGE_SIZE | MAX_PAGE_ORDER | pageblock_order | CMA_MIN_ALIGNMENT_BYTES
-----------------------------------------------------------------------
4KiB | 10 | 9 | 4KiB * (2 ^ 9) = 2MiB
16Kib | 11 | 11 | 16KiB * (2 ^ 11) = 32MiB
64KiB | 13 | 13 | 64KiB * (2 ^ 13) = 512MiB
There are some extreme cases for the CMA alignment requirement when:
- CONFIG_ARCH_FORCE_MAX_ORDER maximum value is set
- CONFIG_TRANSPARENT_HUGEPAGE is NOT set:
- CONFIG_HUGETLB_PAGE is NOT set
PAGE_SIZE | MAX_PAGE_ORDER | pageblock_order | CMA_MIN_ALIGNMENT_BYTES
------------------------------------------------------------------------
4KiB | 15 | 15 | 4KiB * (2 ^ 15) = 128MiB
16Kib | 13 | 13 | 16KiB * (2 ^ 13) = 128MiB
64KiB | 13 | 13 | 64KiB * (2 ^ 13) = 512MiB
This affects the CMA reservations for the drivers. If a driver in a
4KiB kernel needs 4MiB of CMA memory, in a 16KiB kernel, the minimal
reservation has to be 32MiB due to the alignment requirements:
reserved-memory {
...
cma_test_reserve: cma_test_reserve {
compatible = "shared-dma-pool";
size = <0x0 0x400000>; /* 4 MiB */
...
};
};
reserved-memory {
...
cma_test_reserve: cma_test_reserve {
compatible = "shared-dma-pool";
size = <0x0 0x2000000>; /* 32 MiB */
...
};
};
Solution: Add a new config CONFIG_PAGE_BLOCK_ORDER that allows to set the
page block order in all the architectures. The maximum page block order
will be given by ARCH_FORCE_MAX_ORDER.
By default, CONFIG_PAGE_BLOCK_ORDER will have the same value that
ARCH_FORCE_MAX_ORDER. This will make sure that current kernel
configurations won't be affected by this change. It is a opt-in change.
This patch will allow to have the same CMA alignment requirements for
large page sizes (16KiB, 64KiB) as that in 4kb kernels by setting a lower
pageblock_order.
Tests:
- Verified that HugeTLB pages work when pageblock_order is 1, 7, 10 on
4k and 16k kernels.
- Verified that Transparent Huge Pages work when pageblock_order is 1,
7, 10 on 4k and 16k kernels.
- Verified that dma-buf heaps allocations work when pageblock_order is
1, 7, 10 on 4k and 16k kernels.
Benchmarks:
The benchmarks compare 16kb kernels with pageblock_order 10 and 7. The
reason for the pageblock_order 7 is because this value makes the min CMA
alignment requirement the same as that in 4kb kernels (2MB).
- Perform 100K dma-buf heaps (/dev/dma_heap/system) allocations of
SZ_8M, SZ_4M, SZ_2M, SZ_1M, SZ_64, SZ_8, SZ_4. Use simpleperf
(https://developer.android.com/ndk/guides/simpleperf) to measure the #
of instructions and page-faults on 16k kernels. The benchmark was
executed 10 times. The averages are below:
# instructions | #page-faults
order 10 | order 7 | order 10 | order 7
--------------------------------------------------------
13,891,765,770 | 11,425,777,314 | 220 | 217
14,456,293,487 | 12,660,819,302 | 224 | 219
13,924,261,018 | 13,243,970,736 | 217 | 221
13,910,886,504 | 13,845,519,630 | 217 | 221
14,388,071,190 | 13,498,583,098 | 223 | 224
13,656,442,167 | 12,915,831,681 | 216 | 218
13,300,268,343 | 12,930,484,776 | 222 | 218
13,625,470,223 | 14,234,092,777 | 219 | 218
13,508,964,965 | 13,432,689,094 | 225 | 219
13,368,950,667 | 13,683,587,37 | 219 | 225
-------------------------------------------------------------------
13,803,137,433 | 13,131,974,268 | 220 | 220 Averages
There were 4.85% #instructions when order was 7, in comparison with order
10.
13,803,137,433 - 13,131,974,268 = -671,163,166 (-4.86%)
The number of page faults in order 7 and 10 were the same.
These results didn't show any significant regression when the
pageblock_order is set to 7 on 16kb kernels.
- Run speedometer 3.1 (https://browserbench.org/Speedometer3.1/) 5 times
on the 16k kernels with pageblock_order 7 and 10.
order 10 | order 7 | order 7 - order 10 | (order 7 - order 10) %
-------------------------------------------------------------------
15.8 | 16.4 | 0.6 | 3.80%
16.4 | 16.2 | -0.2 | -1.22%
16.6 | 16.3 | -0.3 | -1.81%
16.8 | 16.3 | -0.5 | -2.98%
16.6 | 16.8 | 0.2 | 1.20%
-------------------------------------------------------------------
16.44 16.4 -0.04 -0.24% Averages
The results didn't show any significant regression when the
pageblock_order is set to 7 on 16kb kernels.
Link: https://lkml.kernel.org/r/20250521215807.1860663-1-jyescas@google.com
Signed-off-by: Juan Yescas <jyescas@google.com>
Acked-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Liam R. Howlett <Liam.Howlett@oracle.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Minchan Kim <minchan@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-05-21 14:57:45 -07:00
|
|
|
|
|
|
|
|
|
2025-06-03 08:13:03 +02:00
|
|
|
|
mm: add CONFIG_PAGE_BLOCK_ORDER to select page block order
Problem: On large page size configurations (16KiB, 64KiB), the CMA
alignment requirement (CMA_MIN_ALIGNMENT_BYTES) increases considerably,
and this causes the CMA reservations to be larger than necessary. This
means that system will have less available MIGRATE_UNMOVABLE and
MIGRATE_RECLAIMABLE page blocks since MIGRATE_CMA can't fallback to them.
The CMA_MIN_ALIGNMENT_BYTES increases because it depends on MAX_PAGE_ORDER
which depends on ARCH_FORCE_MAX_ORDER. The value of ARCH_FORCE_MAX_ORDER
increases on 16k and 64k kernels.
For example, in ARM, the CMA alignment requirement when:
- CONFIG_ARCH_FORCE_MAX_ORDER default value is used
- CONFIG_TRANSPARENT_HUGEPAGE is set:
PAGE_SIZE | MAX_PAGE_ORDER | pageblock_order | CMA_MIN_ALIGNMENT_BYTES
-----------------------------------------------------------------------
4KiB | 10 | 9 | 4KiB * (2 ^ 9) = 2MiB
16Kib | 11 | 11 | 16KiB * (2 ^ 11) = 32MiB
64KiB | 13 | 13 | 64KiB * (2 ^ 13) = 512MiB
There are some extreme cases for the CMA alignment requirement when:
- CONFIG_ARCH_FORCE_MAX_ORDER maximum value is set
- CONFIG_TRANSPARENT_HUGEPAGE is NOT set:
- CONFIG_HUGETLB_PAGE is NOT set
PAGE_SIZE | MAX_PAGE_ORDER | pageblock_order | CMA_MIN_ALIGNMENT_BYTES
------------------------------------------------------------------------
4KiB | 15 | 15 | 4KiB * (2 ^ 15) = 128MiB
16Kib | 13 | 13 | 16KiB * (2 ^ 13) = 128MiB
64KiB | 13 | 13 | 64KiB * (2 ^ 13) = 512MiB
This affects the CMA reservations for the drivers. If a driver in a
4KiB kernel needs 4MiB of CMA memory, in a 16KiB kernel, the minimal
reservation has to be 32MiB due to the alignment requirements:
reserved-memory {
...
cma_test_reserve: cma_test_reserve {
compatible = "shared-dma-pool";
size = <0x0 0x400000>; /* 4 MiB */
...
};
};
reserved-memory {
...
cma_test_reserve: cma_test_reserve {
compatible = "shared-dma-pool";
size = <0x0 0x2000000>; /* 32 MiB */
...
};
};
Solution: Add a new config CONFIG_PAGE_BLOCK_ORDER that allows to set the
page block order in all the architectures. The maximum page block order
will be given by ARCH_FORCE_MAX_ORDER.
By default, CONFIG_PAGE_BLOCK_ORDER will have the same value that
ARCH_FORCE_MAX_ORDER. This will make sure that current kernel
configurations won't be affected by this change. It is a opt-in change.
This patch will allow to have the same CMA alignment requirements for
large page sizes (16KiB, 64KiB) as that in 4kb kernels by setting a lower
pageblock_order.
Tests:
- Verified that HugeTLB pages work when pageblock_order is 1, 7, 10 on
4k and 16k kernels.
- Verified that Transparent Huge Pages work when pageblock_order is 1,
7, 10 on 4k and 16k kernels.
- Verified that dma-buf heaps allocations work when pageblock_order is
1, 7, 10 on 4k and 16k kernels.
Benchmarks:
The benchmarks compare 16kb kernels with pageblock_order 10 and 7. The
reason for the pageblock_order 7 is because this value makes the min CMA
alignment requirement the same as that in 4kb kernels (2MB).
- Perform 100K dma-buf heaps (/dev/dma_heap/system) allocations of
SZ_8M, SZ_4M, SZ_2M, SZ_1M, SZ_64, SZ_8, SZ_4. Use simpleperf
(https://developer.android.com/ndk/guides/simpleperf) to measure the #
of instructions and page-faults on 16k kernels. The benchmark was
executed 10 times. The averages are below:
# instructions | #page-faults
order 10 | order 7 | order 10 | order 7
--------------------------------------------------------
13,891,765,770 | 11,425,777,314 | 220 | 217
14,456,293,487 | 12,660,819,302 | 224 | 219
13,924,261,018 | 13,243,970,736 | 217 | 221
13,910,886,504 | 13,845,519,630 | 217 | 221
14,388,071,190 | 13,498,583,098 | 223 | 224
13,656,442,167 | 12,915,831,681 | 216 | 218
13,300,268,343 | 12,930,484,776 | 222 | 218
13,625,470,223 | 14,234,092,777 | 219 | 218
13,508,964,965 | 13,432,689,094 | 225 | 219
13,368,950,667 | 13,683,587,37 | 219 | 225
-------------------------------------------------------------------
13,803,137,433 | 13,131,974,268 | 220 | 220 Averages
There were 4.85% #instructions when order was 7, in comparison with order
10.
13,803,137,433 - 13,131,974,268 = -671,163,166 (-4.86%)
The number of page faults in order 7 and 10 were the same.
These results didn't show any significant regression when the
pageblock_order is set to 7 on 16kb kernels.
- Run speedometer 3.1 (https://browserbench.org/Speedometer3.1/) 5 times
on the 16k kernels with pageblock_order 7 and 10.
order 10 | order 7 | order 7 - order 10 | (order 7 - order 10) %
-------------------------------------------------------------------
15.8 | 16.4 | 0.6 | 3.80%
16.4 | 16.2 | -0.2 | -1.22%
16.6 | 16.3 | -0.3 | -1.81%
16.8 | 16.3 | -0.5 | -2.98%
16.6 | 16.8 | 0.2 | 1.20%
-------------------------------------------------------------------
16.44 16.4 -0.04 -0.24% Averages
The results didn't show any significant regression when the
pageblock_order is set to 7 on 16kb kernels.
Link: https://lkml.kernel.org/r/20250521215807.1860663-1-jyescas@google.com
Signed-off-by: Juan Yescas <jyescas@google.com>
Acked-by: Zi Yan <ziy@nvidia.com>
Reviewed-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Liam R. Howlett <Liam.Howlett@oracle.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Suren Baghdasaryan <surenb@google.com>
Cc: Minchan Kim <minchan@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2025-05-21 14:57:45 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-08-06 16:08:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2013-07-10 16:04:55 -07:00
|
|
|
|
2014-08-06 16:08:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-04-18 11:07:49 +03:00
|
|
|
|
2013-07-10 16:04:55 -07:00
|
|
|
|
2014-04-07 15:39:48 -07:00
|
|
|
|
|
|
|
|
|
2014-04-30 23:26:02 +02:00
|
|
|
|
2020-11-06 19:41:36 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2014-04-30 23:26:02 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-11-06 19:41:36 +01:00
|
|
|
|
2014-04-30 23:26:02 +02:00
|
|
|
|
2020-11-06 19:41:36 +01:00
|
|
|
|
2015-06-30 14:57:02 -07:00
|
|
|
|
|
|
|
|
|
2016-02-05 15:36:21 -08:00
|
|
|
|
mm: make DEFERRED_STRUCT_PAGE_INIT explicitly depend on SPARSEMEM
The deferred memory initialization relies on section definitions, e.g
PAGES_PER_SECTION, that are only available when CONFIG_SPARSEMEM=y on
most architectures.
Initially DEFERRED_STRUCT_PAGE_INIT depended on explicit
ARCH_SUPPORTS_DEFERRED_STRUCT_PAGE_INIT configuration option, but since
the commit 2e3ca40f03bb13709df4 ("mm: relax deferred struct page
requirements") this requirement was relaxed and now it is possible to
enable DEFERRED_STRUCT_PAGE_INIT on architectures that support
DISCONTINGMEM and NO_BOOTMEM which causes build failures.
For instance, setting SMP=y and DEFERRED_STRUCT_PAGE_INIT=y on arc
causes the following build failure:
CC mm/page_alloc.o
mm/page_alloc.c: In function 'update_defer_init':
mm/page_alloc.c:321:14: error: 'PAGES_PER_SECTION'
undeclared (first use in this function); did you mean 'USEC_PER_SEC'?
(pfn & (PAGES_PER_SECTION - 1)) == 0) {
^~~~~~~~~~~~~~~~~
USEC_PER_SEC
mm/page_alloc.c:321:14: note: each undeclared identifier is reported only once for each function it appears in
In file included from include/linux/cache.h:5:0,
from include/linux/printk.h:9,
from include/linux/kernel.h:14,
from include/asm-generic/bug.h:18,
from arch/arc/include/asm/bug.h:32,
from include/linux/bug.h:5,
from include/linux/mmdebug.h:5,
from include/linux/mm.h:9,
from mm/page_alloc.c:18:
mm/page_alloc.c: In function 'deferred_grow_zone':
mm/page_alloc.c:1624:52: error: 'PAGES_PER_SECTION' undeclared (first use in this function); did you mean 'USEC_PER_SEC'?
unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
^
include/uapi/linux/kernel.h:11:47: note: in definition of macro '__ALIGN_KERNEL_MASK'
#define __ALIGN_KERNEL_MASK(x, mask) (((x) + (mask)) & ~(mask))
^~~~
include/linux/kernel.h:58:22: note: in expansion of macro '__ALIGN_KERNEL'
#define ALIGN(x, a) __ALIGN_KERNEL((x), (a))
^~~~~~~~~~~~~~
mm/page_alloc.c:1624:34: note: in expansion of macro 'ALIGN'
unsigned long nr_pages_needed = ALIGN(1 << order, PAGES_PER_SECTION);
^~~~~
In file included from include/asm-generic/bug.h:18:0,
from arch/arc/include/asm/bug.h:32,
from include/linux/bug.h:5,
from include/linux/mmdebug.h:5,
from include/linux/mm.h:9,
from mm/page_alloc.c:18:
mm/page_alloc.c: In function 'free_area_init_node':
mm/page_alloc.c:6379:50: error: 'PAGES_PER_SECTION' undeclared (first use in this function); did you mean 'USEC_PER_SEC'?
pgdat->static_init_pgcnt = min_t(unsigned long, PAGES_PER_SECTION,
^
include/linux/kernel.h:812:22: note: in definition of macro '__typecheck'
(!!(sizeof((typeof(x) *)1 == (typeof(y) *)1)))
^
include/linux/kernel.h:836:24: note: in expansion of macro '__safe_cmp'
__builtin_choose_expr(__safe_cmp(x, y), \
^~~~~~~~~~
include/linux/kernel.h:904:27: note: in expansion of macro '__careful_cmp'
#define min_t(type, x, y) __careful_cmp((type)(x), (type)(y), <)
^~~~~~~~~~~~~
mm/page_alloc.c:6379:29: note: in expansion of macro 'min_t'
pgdat->static_init_pgcnt = min_t(unsigned long, PAGES_PER_SECTION,
^~~~~
include/linux/kernel.h:836:2: error: first argument to '__builtin_choose_expr' not a constant
__builtin_choose_expr(__safe_cmp(x, y), \
^
include/linux/kernel.h:904:27: note: in expansion of macro '__careful_cmp'
#define min_t(type, x, y) __careful_cmp((type)(x), (type)(y), <)
^~~~~~~~~~~~~
mm/page_alloc.c:6379:29: note: in expansion of macro 'min_t'
pgdat->static_init_pgcnt = min_t(unsigned long, PAGES_PER_SECTION,
^~~~~
scripts/Makefile.build:317: recipe for target 'mm/page_alloc.o' failed
Let's make the DEFERRED_STRUCT_PAGE_INIT explicitly depend on SPARSEMEM
as the systems that support DISCONTIGMEM do not seem to have that huge
amounts of memory that would make DEFERRED_STRUCT_PAGE_INIT relevant.
Link: http://lkml.kernel.org/r/1530279308-24988-1-git-send-email-rppt@linux.vnet.ibm.com
Signed-off-by: Mike Rapoport <rppt@linux.vnet.ibm.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Reviewed-by: Pavel Tatashin <pasha.tatashin@oracle.com>
Tested-by: Randy Dunlap <rdunlap@infradead.org>
Cc: Pasha Tatashin <Pavel.Tatashin@microsoft.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-08-17 15:47:07 -07:00
|
|
|
|
2018-05-18 16:09:13 -07:00
|
|
|
|
2018-09-20 12:22:30 -07:00
|
|
|
|
2024-06-21 13:34:47 +02:00
|
|
|
|
mm: parallelize deferred_init_memmap()
Deferred struct page init is a significant bottleneck in kernel boot.
Optimizing it maximizes availability for large-memory systems and allows
spinning up short-lived VMs as needed without having to leave them
running. It also benefits bare metal machines hosting VMs that are
sensitive to downtime. In projects such as VMM Fast Restart[1], where
guest state is preserved across kexec reboot, it helps prevent application
and network timeouts in the guests.
Multithread to take full advantage of system memory bandwidth.
The maximum number of threads is capped at the number of CPUs on the node
because speedups always improve with additional threads on every system
tested, and at this phase of boot, the system is otherwise idle and
waiting on page init to finish.
Helper threads operate on section-aligned ranges to both avoid false
sharing when setting the pageblock's migrate type and to avoid accessing
uninitialized buddy pages, though max order alignment is enough for the
latter.
The minimum chunk size is also a section. There was benefit to using
multiple threads even on relatively small memory (1G) systems, and this is
the smallest size that the alignment allows.
The time (milliseconds) is the slowest node to initialize since boot
blocks until all nodes finish. intel_pstate is loaded in active mode
without hwp and with turbo enabled, and intel_idle is active as well.
Intel(R) Xeon(R) Platinum 8167M CPU @ 2.00GHz (Skylake, bare metal)
2 nodes * 26 cores * 2 threads = 104 CPUs
384G/node = 768G memory
kernel boot deferred init
------------------------ ------------------------
node% (thr) speedup time_ms (stdev) speedup time_ms (stdev)
( 0) -- 4089.7 ( 8.1) -- 1785.7 ( 7.6)
2% ( 1) 1.7% 4019.3 ( 1.5) 3.8% 1717.7 ( 11.8)
12% ( 6) 34.9% 2662.7 ( 2.9) 79.9% 359.3 ( 0.6)
25% ( 13) 39.9% 2459.0 ( 3.6) 91.2% 157.0 ( 0.0)
37% ( 19) 39.2% 2485.0 ( 29.7) 90.4% 172.0 ( 28.6)
50% ( 26) 39.3% 2482.7 ( 25.7) 90.3% 173.7 ( 30.0)
75% ( 39) 39.0% 2495.7 ( 5.5) 89.4% 190.0 ( 1.0)
100% ( 52) 40.2% 2443.7 ( 3.8) 92.3% 138.0 ( 1.0)
Intel(R) Xeon(R) CPU E5-2699C v4 @ 2.20GHz (Broadwell, kvm guest)
1 node * 16 cores * 2 threads = 32 CPUs
192G/node = 192G memory
kernel boot deferred init
------------------------ ------------------------
node% (thr) speedup time_ms (stdev) speedup time_ms (stdev)
( 0) -- 1988.7 ( 9.6) -- 1096.0 ( 11.5)
3% ( 1) 1.1% 1967.0 ( 17.6) 0.3% 1092.7 ( 11.0)
12% ( 4) 41.1% 1170.3 ( 14.2) 73.8% 287.0 ( 3.6)
25% ( 8) 47.1% 1052.7 ( 21.9) 83.9% 177.0 ( 13.5)
38% ( 12) 48.9% 1016.3 ( 12.1) 86.8% 144.7 ( 1.5)
50% ( 16) 48.9% 1015.7 ( 8.1) 87.8% 134.0 ( 4.4)
75% ( 24) 49.1% 1012.3 ( 3.1) 88.1% 130.3 ( 2.3)
100% ( 32) 49.5% 1004.0 ( 5.3) 88.5% 125.7 ( 2.1)
Intel(R) Xeon(R) CPU E5-2699 v3 @ 2.30GHz (Haswell, bare metal)
2 nodes * 18 cores * 2 threads = 72 CPUs
128G/node = 256G memory
kernel boot deferred init
------------------------ ------------------------
node% (thr) speedup time_ms (stdev) speedup time_ms (stdev)
( 0) -- 1680.0 ( 4.6) -- 627.0 ( 4.0)
3% ( 1) 0.3% 1675.7 ( 4.5) -0.2% 628.0 ( 3.6)
11% ( 4) 25.6% 1250.7 ( 2.1) 67.9% 201.0 ( 0.0)
25% ( 9) 30.7% 1164.0 ( 17.3) 81.8% 114.3 ( 17.7)
36% ( 13) 31.4% 1152.7 ( 10.8) 84.0% 100.3 ( 17.9)
50% ( 18) 31.5% 1150.7 ( 9.3) 83.9% 101.0 ( 14.1)
75% ( 27) 31.7% 1148.0 ( 5.6) 84.5% 97.3 ( 6.4)
100% ( 36) 32.0% 1142.3 ( 4.0) 85.6% 90.0 ( 1.0)
AMD EPYC 7551 32-Core Processor (Zen, kvm guest)
1 node * 8 cores * 2 threads = 16 CPUs
64G/node = 64G memory
kernel boot deferred init
------------------------ ------------------------
node% (thr) speedup time_ms (stdev) speedup time_ms (stdev)
( 0) -- 1029.3 ( 25.1) -- 240.7 ( 1.5)
6% ( 1) -0.6% 1036.0 ( 7.8) -2.2% 246.0 ( 0.0)
12% ( 2) 11.8% 907.7 ( 8.6) 44.7% 133.0 ( 1.0)
25% ( 4) 13.9% 886.0 ( 10.6) 62.6% 90.0 ( 6.0)
38% ( 6) 17.8% 845.7 ( 14.2) 69.1% 74.3 ( 3.8)
50% ( 8) 16.8% 856.0 ( 22.1) 72.9% 65.3 ( 5.7)
75% ( 12) 15.4% 871.0 ( 29.2) 79.8% 48.7 ( 7.4)
100% ( 16) 21.0% 813.7 ( 21.0) 80.5% 47.0 ( 5.2)
Server-oriented distros that enable deferred page init sometimes run in
small VMs, and they still benefit even though the fraction of boot time
saved is smaller:
AMD EPYC 7551 32-Core Processor (Zen, kvm guest)
1 node * 2 cores * 2 threads = 4 CPUs
16G/node = 16G memory
kernel boot deferred init
------------------------ ------------------------
node% (thr) speedup time_ms (stdev) speedup time_ms (stdev)
( 0) -- 716.0 ( 14.0) -- 49.7 ( 0.6)
25% ( 1) 1.8% 703.0 ( 5.3) -4.0% 51.7 ( 0.6)
50% ( 2) 1.6% 704.7 ( 1.2) 43.0% 28.3 ( 0.6)
75% ( 3) 2.7% 696.7 ( 13.1) 49.7% 25.0 ( 0.0)
100% ( 4) 4.1% 687.0 ( 10.4) 55.7% 22.0 ( 0.0)
Intel(R) Xeon(R) CPU E5-2699 v3 @ 2.30GHz (Haswell, kvm guest)
1 node * 2 cores * 2 threads = 4 CPUs
14G/node = 14G memory
kernel boot deferred init
------------------------ ------------------------
node% (thr) speedup time_ms (stdev) speedup time_ms (stdev)
( 0) -- 787.7 ( 6.4) -- 122.3 ( 0.6)
25% ( 1) 0.2% 786.3 ( 10.8) -2.5% 125.3 ( 2.1)
50% ( 2) 5.9% 741.0 ( 13.9) 37.6% 76.3 ( 19.7)
75% ( 3) 8.3% 722.0 ( 19.0) 49.9% 61.3 ( 3.2)
100% ( 4) 9.3% 714.7 ( 9.5) 56.4% 53.3 ( 1.5)
On Josh's 96-CPU and 192G memory system:
Without this patch series:
[ 0.487132] node 0 initialised, 23398907 pages in 292ms
[ 0.499132] node 1 initialised, 24189223 pages in 304ms
...
[ 0.629376] Run /sbin/init as init process
With this patch series:
[ 0.231435] node 1 initialised, 24189223 pages in 32ms
[ 0.236718] node 0 initialised, 23398907 pages in 36ms
[1] https://static.sched.com/hosted_files/kvmforum2019/66/VMM-fast-restart_kvmforum2019.pdf
Signed-off-by: Daniel Jordan <daniel.m.jordan@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Tested-by: Josh Triplett <josh@joshtriplett.org>
Reviewed-by: Alexander Duyck <alexander.h.duyck@linux.intel.com>
Cc: Alex Williamson <alex.williamson@redhat.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Herbert Xu <herbert@gondor.apana.org.au>
Cc: Jason Gunthorpe <jgg@ziepe.ca>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Kirill Tkhai <ktkhai@virtuozzo.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Pavel Machek <pavel@ucw.cz>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Robert Elliott <elliott@hpe.com>
Cc: Shile Zhang <shile.zhang@linux.alibaba.com>
Cc: Steffen Klassert <steffen.klassert@secunet.com>
Cc: Steven Sistare <steven.sistare@oracle.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Zi Yan <ziy@nvidia.com>
Link: http://lkml.kernel.org/r/20200527173608.2885243-7-daniel.m.jordan@oracle.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-06-03 15:59:51 -07:00
|
|
|
|
2015-06-30 14:57:02 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: parallelize deferred_init_memmap()
Deferred struct page init is a significant bottleneck in kernel boot.
Optimizing it maximizes availability for large-memory systems and allows
spinning up short-lived VMs as needed without having to leave them
running. It also benefits bare metal machines hosting VMs that are
sensitive to downtime. In projects such as VMM Fast Restart[1], where
guest state is preserved across kexec reboot, it helps prevent application
and network timeouts in the guests.
Multithread to take full advantage of system memory bandwidth.
The maximum number of threads is capped at the number of CPUs on the node
because speedups always improve with additional threads on every system
tested, and at this phase of boot, the system is otherwise idle and
waiting on page init to finish.
Helper threads operate on section-aligned ranges to both avoid false
sharing when setting the pageblock's migrate type and to avoid accessing
uninitialized buddy pages, though max order alignment is enough for the
latter.
The minimum chunk size is also a section. There was benefit to using
multiple threads even on relatively small memory (1G) systems, and this is
the smallest size that the alignment allows.
The time (milliseconds) is the slowest node to initialize since boot
blocks until all nodes finish. intel_pstate is loaded in active mode
without hwp and with turbo enabled, and intel_idle is active as well.
Intel(R) Xeon(R) Platinum 8167M CPU @ 2.00GHz (Skylake, bare metal)
2 nodes * 26 cores * 2 threads = 104 CPUs
384G/node = 768G memory
kernel boot deferred init
------------------------ ------------------------
node% (thr) speedup time_ms (stdev) speedup time_ms (stdev)
( 0) -- 4089.7 ( 8.1) -- 1785.7 ( 7.6)
2% ( 1) 1.7% 4019.3 ( 1.5) 3.8% 1717.7 ( 11.8)
12% ( 6) 34.9% 2662.7 ( 2.9) 79.9% 359.3 ( 0.6)
25% ( 13) 39.9% 2459.0 ( 3.6) 91.2% 157.0 ( 0.0)
37% ( 19) 39.2% 2485.0 ( 29.7) 90.4% 172.0 ( 28.6)
50% ( 26) 39.3% 2482.7 ( 25.7) 90.3% 173.7 ( 30.0)
75% ( 39) 39.0% 2495.7 ( 5.5) 89.4% 190.0 ( 1.0)
100% ( 52) 40.2% 2443.7 ( 3.8) 92.3% 138.0 ( 1.0)
Intel(R) Xeon(R) CPU E5-2699C v4 @ 2.20GHz (Broadwell, kvm guest)
1 node * 16 cores * 2 threads = 32 CPUs
192G/node = 192G memory
kernel boot deferred init
------------------------ ------------------------
node% (thr) speedup time_ms (stdev) speedup time_ms (stdev)
( 0) -- 1988.7 ( 9.6) -- 1096.0 ( 11.5)
3% ( 1) 1.1% 1967.0 ( 17.6) 0.3% 1092.7 ( 11.0)
12% ( 4) 41.1% 1170.3 ( 14.2) 73.8% 287.0 ( 3.6)
25% ( 8) 47.1% 1052.7 ( 21.9) 83.9% 177.0 ( 13.5)
38% ( 12) 48.9% 1016.3 ( 12.1) 86.8% 144.7 ( 1.5)
50% ( 16) 48.9% 1015.7 ( 8.1) 87.8% 134.0 ( 4.4)
75% ( 24) 49.1% 1012.3 ( 3.1) 88.1% 130.3 ( 2.3)
100% ( 32) 49.5% 1004.0 ( 5.3) 88.5% 125.7 ( 2.1)
Intel(R) Xeon(R) CPU E5-2699 v3 @ 2.30GHz (Haswell, bare metal)
2 nodes * 18 cores * 2 threads = 72 CPUs
128G/node = 256G memory
kernel boot deferred init
------------------------ ------------------------
node% (thr) speedup time_ms (stdev) speedup time_ms (stdev)
( 0) -- 1680.0 ( 4.6) -- 627.0 ( 4.0)
3% ( 1) 0.3% 1675.7 ( 4.5) -0.2% 628.0 ( 3.6)
11% ( 4) 25.6% 1250.7 ( 2.1) 67.9% 201.0 ( 0.0)
25% ( 9) 30.7% 1164.0 ( 17.3) 81.8% 114.3 ( 17.7)
36% ( 13) 31.4% 1152.7 ( 10.8) 84.0% 100.3 ( 17.9)
50% ( 18) 31.5% 1150.7 ( 9.3) 83.9% 101.0 ( 14.1)
75% ( 27) 31.7% 1148.0 ( 5.6) 84.5% 97.3 ( 6.4)
100% ( 36) 32.0% 1142.3 ( 4.0) 85.6% 90.0 ( 1.0)
AMD EPYC 7551 32-Core Processor (Zen, kvm guest)
1 node * 8 cores * 2 threads = 16 CPUs
64G/node = 64G memory
kernel boot deferred init
------------------------ ------------------------
node% (thr) speedup time_ms (stdev) speedup time_ms (stdev)
( 0) -- 1029.3 ( 25.1) -- 240.7 ( 1.5)
6% ( 1) -0.6% 1036.0 ( 7.8) -2.2% 246.0 ( 0.0)
12% ( 2) 11.8% 907.7 ( 8.6) 44.7% 133.0 ( 1.0)
25% ( 4) 13.9% 886.0 ( 10.6) 62.6% 90.0 ( 6.0)
38% ( 6) 17.8% 845.7 ( 14.2) 69.1% 74.3 ( 3.8)
50% ( 8) 16.8% 856.0 ( 22.1) 72.9% 65.3 ( 5.7)
75% ( 12) 15.4% 871.0 ( 29.2) 79.8% 48.7 ( 7.4)
100% ( 16) 21.0% 813.7 ( 21.0) 80.5% 47.0 ( 5.2)
Server-oriented distros that enable deferred page init sometimes run in
small VMs, and they still benefit even though the fraction of boot time
saved is smaller:
AMD EPYC 7551 32-Core Processor (Zen, kvm guest)
1 node * 2 cores * 2 threads = 4 CPUs
16G/node = 16G memory
kernel boot deferred init
------------------------ ------------------------
node% (thr) speedup time_ms (stdev) speedup time_ms (stdev)
( 0) -- 716.0 ( 14.0) -- 49.7 ( 0.6)
25% ( 1) 1.8% 703.0 ( 5.3) -4.0% 51.7 ( 0.6)
50% ( 2) 1.6% 704.7 ( 1.2) 43.0% 28.3 ( 0.6)
75% ( 3) 2.7% 696.7 ( 13.1) 49.7% 25.0 ( 0.0)
100% ( 4) 4.1% 687.0 ( 10.4) 55.7% 22.0 ( 0.0)
Intel(R) Xeon(R) CPU E5-2699 v3 @ 2.30GHz (Haswell, kvm guest)
1 node * 2 cores * 2 threads = 4 CPUs
14G/node = 14G memory
kernel boot deferred init
------------------------ ------------------------
node% (thr) speedup time_ms (stdev) speedup time_ms (stdev)
( 0) -- 787.7 ( 6.4) -- 122.3 ( 0.6)
25% ( 1) 0.2% 786.3 ( 10.8) -2.5% 125.3 ( 2.1)
50% ( 2) 5.9% 741.0 ( 13.9) 37.6% 76.3 ( 19.7)
75% ( 3) 8.3% 722.0 ( 19.0) 49.9% 61.3 ( 3.2)
100% ( 4) 9.3% 714.7 ( 9.5) 56.4% 53.3 ( 1.5)
On Josh's 96-CPU and 192G memory system:
Without this patch series:
[ 0.487132] node 0 initialised, 23398907 pages in 292ms
[ 0.499132] node 1 initialised, 24189223 pages in 304ms
...
[ 0.629376] Run /sbin/init as init process
With this patch series:
[ 0.231435] node 1 initialised, 24189223 pages in 32ms
[ 0.236718] node 0 initialised, 23398907 pages in 36ms
[1] https://static.sched.com/hosted_files/kvmforum2019/66/VMM-fast-restart_kvmforum2019.pdf
Signed-off-by: Daniel Jordan <daniel.m.jordan@oracle.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Tested-by: Josh Triplett <josh@joshtriplett.org>
Reviewed-by: Alexander Duyck <alexander.h.duyck@linux.intel.com>
Cc: Alex Williamson <alex.williamson@redhat.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: Herbert Xu <herbert@gondor.apana.org.au>
Cc: Jason Gunthorpe <jgg@ziepe.ca>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Kirill Tkhai <ktkhai@virtuozzo.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Pavel Machek <pavel@ucw.cz>
Cc: Pavel Tatashin <pasha.tatashin@soleen.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Randy Dunlap <rdunlap@infradead.org>
Cc: Robert Elliott <elliott@hpe.com>
Cc: Shile Zhang <shile.zhang@linux.alibaba.com>
Cc: Steffen Klassert <steffen.klassert@secunet.com>
Cc: Steven Sistare <steven.sistare@oracle.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Zi Yan <ziy@nvidia.com>
Link: http://lkml.kernel.org/r/20200527173608.2885243-7-daniel.m.jordan@oracle.com
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-06-03 15:59:51 -07:00
|
|
|
|
|
|
|
|
|
2016-02-05 15:36:21 -08:00
|
|
|
|
|
|
|
|
|
2015-08-09 15:29:06 -04:00
|
|
|
|
mm/idle_page_tracking: make PG_idle reusable
PG_idle and PG_young allow the two PTE Accessed bit users, Idle Page
Tracking and the reclaim logic concurrently work while not interfering
with each other. That is, when they need to clear the Accessed bit, they
set PG_young to represent the previous state of the bit, respectively.
And when they need to read the bit, if the bit is cleared, they further
read the PG_young to know whether the other has cleared the bit meanwhile
or not.
For yet another user of the PTE Accessed bit, we could add another page
flag, or extend the mechanism to use the flags. For the DAMON usecase,
however, we don't need to do that just yet. IDLE_PAGE_TRACKING and DAMON
are mutually exclusive, so there's only ever going to be one user of the
current set of flags.
In this commit, we split out the CONFIG options to allow for the use of
PG_young and PG_idle outside of idle page tracking.
In the next commit, DAMON's reference implementation of the virtual memory
address space monitoring primitives will use it.
[sjpark@amazon.de: set PAGE_EXTENSION for non-64BIT]
Link: https://lkml.kernel.org/r/20210806095153.6444-1-sj38.park@gmail.com
[akpm@linux-foundation.org: tweak Kconfig text]
[sjpark@amazon.de: hide PAGE_IDLE_FLAG from users]
Link: https://lkml.kernel.org/r/20210813081238.34705-1-sj38.park@gmail.com
Link: https://lkml.kernel.org/r/20210716081449.22187-5-sj38.park@gmail.com
Signed-off-by: SeongJae Park <sjpark@amazon.de>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Reviewed-by: Fernand Sieber <sieberf@amazon.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Amit Shah <amit@kernel.org>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Brendan Higgins <brendanhiggins@google.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: David Woodhouse <dwmw@amazon.com>
Cc: Fan Du <fan.du@intel.com>
Cc: Greg Kroah-Hartman <greg@kroah.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Joe Perches <joe@perches.com>
Cc: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Leonard Foerster <foersleo@amazon.de>
Cc: Marco Elver <elver@google.com>
Cc: Markus Boehme <markubo@amazon.de>
Cc: Maximilian Heyne <mheyne@amazon.de>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Rik van Riel <riel@surriel.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Steven Rostedt (VMware) <rostedt@goodmis.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-09-07 19:56:40 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: introduce idle page tracking
Knowing the portion of memory that is not used by a certain application or
memory cgroup (idle memory) can be useful for partitioning the system
efficiently, e.g. by setting memory cgroup limits appropriately.
Currently, the only means to estimate the amount of idle memory provided
by the kernel is /proc/PID/{clear_refs,smaps}: the user can clear the
access bit for all pages mapped to a particular process by writing 1 to
clear_refs, wait for some time, and then count smaps:Referenced. However,
this method has two serious shortcomings:
- it does not count unmapped file pages
- it affects the reclaimer logic
To overcome these drawbacks, this patch introduces two new page flags,
Idle and Young, and a new sysfs file, /sys/kernel/mm/page_idle/bitmap.
A page's Idle flag can only be set from userspace by setting bit in
/sys/kernel/mm/page_idle/bitmap at the offset corresponding to the page,
and it is cleared whenever the page is accessed either through page tables
(it is cleared in page_referenced() in this case) or using the read(2)
system call (mark_page_accessed()). Thus by setting the Idle flag for
pages of a particular workload, which can be found e.g. by reading
/proc/PID/pagemap, waiting for some time to let the workload access its
working set, and then reading the bitmap file, one can estimate the amount
of pages that are not used by the workload.
The Young page flag is used to avoid interference with the memory
reclaimer. A page's Young flag is set whenever the Access bit of a page
table entry pointing to the page is cleared by writing to the bitmap file.
If page_referenced() is called on a Young page, it will add 1 to its
return value, therefore concealing the fact that the Access bit was
cleared.
Note, since there is no room for extra page flags on 32 bit, this feature
uses extended page flags when compiled on 32 bit.
[akpm@linux-foundation.org: fix build]
[akpm@linux-foundation.org: kpageidle requires an MMU]
[akpm@linux-foundation.org: decouple from page-flags rework]
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Reviewed-by: Andres Lagar-Cavilla <andreslc@google.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Greg Thelen <gthelen@google.com>
Cc: Michel Lespinasse <walken@google.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-09-09 15:35:45 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm/idle_page_tracking: make PG_idle reusable
PG_idle and PG_young allow the two PTE Accessed bit users, Idle Page
Tracking and the reclaim logic concurrently work while not interfering
with each other. That is, when they need to clear the Accessed bit, they
set PG_young to represent the previous state of the bit, respectively.
And when they need to read the bit, if the bit is cleared, they further
read the PG_young to know whether the other has cleared the bit meanwhile
or not.
For yet another user of the PTE Accessed bit, we could add another page
flag, or extend the mechanism to use the flags. For the DAMON usecase,
however, we don't need to do that just yet. IDLE_PAGE_TRACKING and DAMON
are mutually exclusive, so there's only ever going to be one user of the
current set of flags.
In this commit, we split out the CONFIG options to allow for the use of
PG_young and PG_idle outside of idle page tracking.
In the next commit, DAMON's reference implementation of the virtual memory
address space monitoring primitives will use it.
[sjpark@amazon.de: set PAGE_EXTENSION for non-64BIT]
Link: https://lkml.kernel.org/r/20210806095153.6444-1-sj38.park@gmail.com
[akpm@linux-foundation.org: tweak Kconfig text]
[sjpark@amazon.de: hide PAGE_IDLE_FLAG from users]
Link: https://lkml.kernel.org/r/20210813081238.34705-1-sj38.park@gmail.com
Link: https://lkml.kernel.org/r/20210716081449.22187-5-sj38.park@gmail.com
Signed-off-by: SeongJae Park <sjpark@amazon.de>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Reviewed-by: Fernand Sieber <sieberf@amazon.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Amit Shah <amit@kernel.org>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Brendan Higgins <brendanhiggins@google.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: David Woodhouse <dwmw@amazon.com>
Cc: Fan Du <fan.du@intel.com>
Cc: Greg Kroah-Hartman <greg@kroah.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Joe Perches <joe@perches.com>
Cc: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Leonard Foerster <foersleo@amazon.de>
Cc: Marco Elver <elver@google.com>
Cc: Markus Boehme <markubo@amazon.de>
Cc: Maximilian Heyne <mheyne@amazon.de>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Rik van Riel <riel@surriel.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Steven Rostedt (VMware) <rostedt@goodmis.org>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Vlastimil Babka <vbabka@suse.cz>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-09-07 19:56:40 -07:00
|
|
|
|
mm: introduce idle page tracking
Knowing the portion of memory that is not used by a certain application or
memory cgroup (idle memory) can be useful for partitioning the system
efficiently, e.g. by setting memory cgroup limits appropriately.
Currently, the only means to estimate the amount of idle memory provided
by the kernel is /proc/PID/{clear_refs,smaps}: the user can clear the
access bit for all pages mapped to a particular process by writing 1 to
clear_refs, wait for some time, and then count smaps:Referenced. However,
this method has two serious shortcomings:
- it does not count unmapped file pages
- it affects the reclaimer logic
To overcome these drawbacks, this patch introduces two new page flags,
Idle and Young, and a new sysfs file, /sys/kernel/mm/page_idle/bitmap.
A page's Idle flag can only be set from userspace by setting bit in
/sys/kernel/mm/page_idle/bitmap at the offset corresponding to the page,
and it is cleared whenever the page is accessed either through page tables
(it is cleared in page_referenced() in this case) or using the read(2)
system call (mark_page_accessed()). Thus by setting the Idle flag for
pages of a particular workload, which can be found e.g. by reading
/proc/PID/pagemap, waiting for some time to let the workload access its
working set, and then reading the bitmap file, one can estimate the amount
of pages that are not used by the workload.
The Young page flag is used to avoid interference with the memory
reclaimer. A page's Young flag is set whenever the Access bit of a page
table entry pointing to the page is cleared by writing to the bitmap file.
If page_referenced() is called on a Young page, it will add 1 to its
return value, therefore concealing the fact that the Access bit was
cleared.
Note, since there is no room for extra page flags on 32 bit, this feature
uses extended page flags when compiled on 32 bit.
[akpm@linux-foundation.org: fix build]
[akpm@linux-foundation.org: kpageidle requires an MMU]
[akpm@linux-foundation.org: decouple from page-flags rework]
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Reviewed-by: Andres Lagar-Cavilla <andreslc@google.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Greg Thelen <gthelen@google.com>
Cc: Michel Lespinasse <walken@google.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-09-09 15:35:45 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-04-18 11:07:49 +03:00
|
|
|
|
|
|
|
|
|
mm: introduce idle page tracking
Knowing the portion of memory that is not used by a certain application or
memory cgroup (idle memory) can be useful for partitioning the system
efficiently, e.g. by setting memory cgroup limits appropriately.
Currently, the only means to estimate the amount of idle memory provided
by the kernel is /proc/PID/{clear_refs,smaps}: the user can clear the
access bit for all pages mapped to a particular process by writing 1 to
clear_refs, wait for some time, and then count smaps:Referenced. However,
this method has two serious shortcomings:
- it does not count unmapped file pages
- it affects the reclaimer logic
To overcome these drawbacks, this patch introduces two new page flags,
Idle and Young, and a new sysfs file, /sys/kernel/mm/page_idle/bitmap.
A page's Idle flag can only be set from userspace by setting bit in
/sys/kernel/mm/page_idle/bitmap at the offset corresponding to the page,
and it is cleared whenever the page is accessed either through page tables
(it is cleared in page_referenced() in this case) or using the read(2)
system call (mark_page_accessed()). Thus by setting the Idle flag for
pages of a particular workload, which can be found e.g. by reading
/proc/PID/pagemap, waiting for some time to let the workload access its
working set, and then reading the bitmap file, one can estimate the amount
of pages that are not used by the workload.
The Young page flag is used to avoid interference with the memory
reclaimer. A page's Young flag is set whenever the Access bit of a page
table entry pointing to the page is cleared by writing to the bitmap file.
If page_referenced() is called on a Young page, it will add 1 to its
return value, therefore concealing the fact that the Access bit was
cleared.
Note, since there is no room for extra page flags on 32 bit, this feature
uses extended page flags when compiled on 32 bit.
[akpm@linux-foundation.org: fix build]
[akpm@linux-foundation.org: kpageidle requires an MMU]
[akpm@linux-foundation.org: decouple from page-flags rework]
Signed-off-by: Vladimir Davydov <vdavydov@parallels.com>
Reviewed-by: Andres Lagar-Cavilla <andreslc@google.com>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Raghavendra K T <raghavendra.kt@linux.vnet.ibm.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Michal Hocko <mhocko@suse.cz>
Cc: Greg Thelen <gthelen@google.com>
Cc: Michel Lespinasse <walken@google.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Pavel Emelyanov <xemul@parallels.com>
Cc: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-09-09 15:35:45 -07:00
|
|
|
|
2024-02-15 09:46:32 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-05-04 18:38:09 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-02-16 12:05:28 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-06-30 18:52:20 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-08-09 15:29:06 -04:00
|
|
|
|
2017-09-08 16:11:43 -07:00
|
|
|
|
2015-08-09 15:29:06 -04:00
|
|
|
|
2018-09-22 16:14:30 -04:00
|
|
|
|
2015-08-09 15:29:06 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-11-06 15:13:50 -08:00
|
|
|
|
|
|
|
|
|
2015-08-09 15:29:06 -04:00
|
|
|
|
2025-11-06 15:13:50 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2015-09-11 16:42:39 -07:00
|
|
|
|
2019-08-06 19:05:52 +03:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-09-08 16:11:27 -07:00
|
|
|
|
2019-08-06 19:05:52 +03:00
|
|
|
|
2019-08-06 19:05:51 +03:00
|
|
|
|
2026-07-27 17:31:14 -07:00
|
|
|
|
2017-09-08 16:11:27 -07:00
|
|
|
|
2022-05-20 13:41:24 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-09-08 16:11:43 -07:00
|
|
|
|
|
|
|
|
|
2019-06-26 14:27:22 +02:00
|
|
|
|
2022-05-20 13:41:24 -07:00
|
|
|
|
2017-09-08 16:11:43 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2020-10-17 16:15:10 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2016-02-12 13:02:08 -08:00
|
|
|
|
|
|
|
|
|
2016-02-12 13:02:32 -08:00
|
|
|
|
|
|
|
|
|
2017-06-19 19:28:31 -04:00
|
|
|
|
2024-08-21 20:34:43 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-11-03 18:10:34 -07:00
|
|
|
|
|
|
|
|
|
2022-05-25 13:25:59 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-06-19 19:28:31 -04:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-11-17 15:31:22 -08:00
|
|
|
|
mm/gup_benchmark: rename to mm/gup_test
Patch series "selftests/vm: gup_test, hmm-tests, assorted improvements", v3.
Summary: This series provides two main things, and a number of smaller
supporting goodies. The two main points are:
1) Add a new sub-test to gup_test, which in turn is a renamed version
of gup_benchmark. This sub-test allows nicer testing of dump_pages(),
at least on user-space pages.
For quite a while, I was doing a quick hack to gup_test.c whenever I
wanted to try out changes to dump_page(). Then Matthew Wilcox asked me
what I meant when I said "I used my dump_page() unit test", and I
realized that it might be nice to check in a polished up version of
that.
Details about how it works and how to use it are in the commit
description for patch #6 ("selftests/vm: gup_test: introduce the
dump_pages() sub-test").
2) Fixes a limitation of hmm-tests: these tests are incredibly useful,
but only if people actually build and run them. And it turns out that
libhugetlbfs is a little too effective at throwing a wrench in the
works, there. So I've added a little configuration check that removes
just two of the 21 hmm-tests, if libhugetlbfs is not available.
Further details in the commit description of patch #8
("selftests/vm: hmm-tests: remove the libhugetlbfs dependency").
Other smaller things that this series does:
a) Remove code duplication by creating gup_test.h.
b) Clear up the sub-test organization, and their invocation within
run_vmtests.sh.
c) Other minor assorted improvements.
[1] v2 is here:
https://lore.kernel.org/linux-doc/20200929212747.251804-1-jhubbard@nvidia.com/
[2] https://lore.kernel.org/r/CAHk-=wgh-TMPHLY3jueHX7Y2fWh3D+nMBqVS__AZm6-oorquWA@mail.gmail.com
This patch (of 9):
Rename nearly every "gup_benchmark" reference and file name to "gup_test".
The one exception is for the actual gup benchmark test itself.
The current code already does a *little* bit more than benchmarking, and
definitely covers more than get_user_pages_fast(). More importantly,
however, subsequent patches are about to add some functionality that is
non-benchmark related.
Closely related changes:
* Kconfig: in addition to renaming the options from GUP_BENCHMARK to
GUP_TEST, update the help text to reflect that it's no longer a
benchmark-only test.
Link: https://lkml.kernel.org/r/20201026064021.3545418-1-jhubbard@nvidia.com
Link: https://lkml.kernel.org/r/20201026064021.3545418-2-jhubbard@nvidia.com
Signed-off-by: John Hubbard <jhubbard@nvidia.com>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Jérôme Glisse <jglisse@redhat.com>
Cc: Ralph Campbell <rcampbell@nvidia.com>
Cc: Shuah Khan <shuah@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-12-14 19:05:05 -08:00
|
|
|
|
|
|
|
|
|
2020-12-14 19:05:38 -08:00
|
|
|
|
2017-11-17 15:31:22 -08:00
|
|
|
|
mm/gup_benchmark: rename to mm/gup_test
Patch series "selftests/vm: gup_test, hmm-tests, assorted improvements", v3.
Summary: This series provides two main things, and a number of smaller
supporting goodies. The two main points are:
1) Add a new sub-test to gup_test, which in turn is a renamed version
of gup_benchmark. This sub-test allows nicer testing of dump_pages(),
at least on user-space pages.
For quite a while, I was doing a quick hack to gup_test.c whenever I
wanted to try out changes to dump_page(). Then Matthew Wilcox asked me
what I meant when I said "I used my dump_page() unit test", and I
realized that it might be nice to check in a polished up version of
that.
Details about how it works and how to use it are in the commit
description for patch #6 ("selftests/vm: gup_test: introduce the
dump_pages() sub-test").
2) Fixes a limitation of hmm-tests: these tests are incredibly useful,
but only if people actually build and run them. And it turns out that
libhugetlbfs is a little too effective at throwing a wrench in the
works, there. So I've added a little configuration check that removes
just two of the 21 hmm-tests, if libhugetlbfs is not available.
Further details in the commit description of patch #8
("selftests/vm: hmm-tests: remove the libhugetlbfs dependency").
Other smaller things that this series does:
a) Remove code duplication by creating gup_test.h.
b) Clear up the sub-test organization, and their invocation within
run_vmtests.sh.
c) Other minor assorted improvements.
[1] v2 is here:
https://lore.kernel.org/linux-doc/20200929212747.251804-1-jhubbard@nvidia.com/
[2] https://lore.kernel.org/r/CAHk-=wgh-TMPHLY3jueHX7Y2fWh3D+nMBqVS__AZm6-oorquWA@mail.gmail.com
This patch (of 9):
Rename nearly every "gup_benchmark" reference and file name to "gup_test".
The one exception is for the actual gup benchmark test itself.
The current code already does a *little* bit more than benchmarking, and
definitely covers more than get_user_pages_fast(). More importantly,
however, subsequent patches are about to add some functionality that is
non-benchmark related.
Closely related changes:
* Kconfig: in addition to renaming the options from GUP_BENCHMARK to
GUP_TEST, update the help text to reflect that it's no longer a
benchmark-only test.
Link: https://lkml.kernel.org/r/20201026064021.3545418-1-jhubbard@nvidia.com
Link: https://lkml.kernel.org/r/20201026064021.3545418-2-jhubbard@nvidia.com
Signed-off-by: John Hubbard <jhubbard@nvidia.com>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Jérôme Glisse <jglisse@redhat.com>
Cc: Ralph Campbell <rcampbell@nvidia.com>
Cc: Shuah Khan <shuah@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-12-14 19:05:05 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2017-11-17 15:31:22 -08:00
|
|
|
|
mm/gup_benchmark: rename to mm/gup_test
Patch series "selftests/vm: gup_test, hmm-tests, assorted improvements", v3.
Summary: This series provides two main things, and a number of smaller
supporting goodies. The two main points are:
1) Add a new sub-test to gup_test, which in turn is a renamed version
of gup_benchmark. This sub-test allows nicer testing of dump_pages(),
at least on user-space pages.
For quite a while, I was doing a quick hack to gup_test.c whenever I
wanted to try out changes to dump_page(). Then Matthew Wilcox asked me
what I meant when I said "I used my dump_page() unit test", and I
realized that it might be nice to check in a polished up version of
that.
Details about how it works and how to use it are in the commit
description for patch #6 ("selftests/vm: gup_test: introduce the
dump_pages() sub-test").
2) Fixes a limitation of hmm-tests: these tests are incredibly useful,
but only if people actually build and run them. And it turns out that
libhugetlbfs is a little too effective at throwing a wrench in the
works, there. So I've added a little configuration check that removes
just two of the 21 hmm-tests, if libhugetlbfs is not available.
Further details in the commit description of patch #8
("selftests/vm: hmm-tests: remove the libhugetlbfs dependency").
Other smaller things that this series does:
a) Remove code duplication by creating gup_test.h.
b) Clear up the sub-test organization, and their invocation within
run_vmtests.sh.
c) Other minor assorted improvements.
[1] v2 is here:
https://lore.kernel.org/linux-doc/20200929212747.251804-1-jhubbard@nvidia.com/
[2] https://lore.kernel.org/r/CAHk-=wgh-TMPHLY3jueHX7Y2fWh3D+nMBqVS__AZm6-oorquWA@mail.gmail.com
This patch (of 9):
Rename nearly every "gup_benchmark" reference and file name to "gup_test".
The one exception is for the actual gup benchmark test itself.
The current code already does a *little* bit more than benchmarking, and
definitely covers more than get_user_pages_fast(). More importantly,
however, subsequent patches are about to add some functionality that is
non-benchmark related.
Closely related changes:
* Kconfig: in addition to renaming the options from GUP_BENCHMARK to
GUP_TEST, update the help text to reflect that it's no longer a
benchmark-only test.
Link: https://lkml.kernel.org/r/20201026064021.3545418-1-jhubbard@nvidia.com
Link: https://lkml.kernel.org/r/20201026064021.3545418-2-jhubbard@nvidia.com
Signed-off-by: John Hubbard <jhubbard@nvidia.com>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Jérôme Glisse <jglisse@redhat.com>
Cc: Ralph Campbell <rcampbell@nvidia.com>
Cc: Shuah Khan <shuah@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-12-14 19:05:05 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
selftests/vm: gup_test: introduce the dump_pages() sub-test
For quite a while, I was doing a quick hack to gup_test.c (previously,
gup_benchmark.c) whenever I wanted to try out my changes to dump_page().
This makes that hack unnecessary, and instead allows anyone to easily get
the same coverage from a user space program. That saves a lot of time
because you don't have to change the kernel, in order to test different
pages and options.
The new sub-test takes advantage of the existing gup_test infrastructure,
which already provides a simple user space program, some allocated user
space pages, an ioctl call, pinning of those pages (via either
get_user_pages or pin_user_pages) and a corresponding kernel-side test
invocation. There's not much more required, mainly just a couple of
inputs from the user.
In fact, the new test re-uses the existing command line options in order
to get various helpful combinations (THP or normal, _fast or slow gup, gup
vs. pup, and more).
New command line options are: which pages to dump, and what type of
"get/pin" to use.
In order to figure out which pages to dump, the logic is:
* If the user doesn't specify anything, the page 0 (the first page in
the address range that the program sets up for testing) is dumped.
* Or, the user can type up to 8 page indices anywhere on the command
line. If you type more than 8, then it uses the first 8 and ignores the
remaining items.
For example:
./gup_test -ct -F 1 0 19 0x1000
Meaning:
-c: dump pages sub-test
-t: use THP pages
-F 1: use pin_user_pages() instead of get_user_pages()
0 19 0x1000: dump pages 0, 19, and 4096
Link: https://lkml.kernel.org/r/20201026064021.3545418-7-jhubbard@nvidia.com
Signed-off-by: John Hubbard <jhubbard@nvidia.com>
Cc: Jérôme Glisse <jglisse@redhat.com>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Ralph Campbell <rcampbell@nvidia.com>
Cc: Shuah Khan <shuah@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2020-12-14 19:05:21 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-01-03 18:07:53 +00:00
|
|
|
|
2018-06-07 17:06:08 -07:00
|
|
|
|
2020-12-14 19:05:38 -08:00
|
|
|
|
|
|
|
|
|
2018-06-07 17:06:08 -07:00
|
|
|
|
2022-10-21 14:51:44 +02:00
|
|
|
|
2019-07-11 20:56:49 -07:00
|
|
|
|
|
|
|
|
|
2023-01-26 13:51:14 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2018-06-07 17:06:08 -07:00
|
|
|
|
|
|
|
|
|
2018-07-31 13:39:35 +02:00
|
|
|
|
2019-03-19 13:12:30 +01:00
|
|
|
|
2026-04-07 11:09:34 +05:30
|
|
|
|
2019-03-19 13:12:30 +01:00
|
|
|
|
2020-11-03 10:27:18 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2021-11-19 16:43:55 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-06-30 11:08:53 +02:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: introduce memfd_secret system call to create "secret" memory areas
Introduce "memfd_secret" system call with the ability to create memory
areas visible only in the context of the owning process and not mapped not
only to other processes but in the kernel page tables as well.
The secretmem feature is off by default and the user must explicitly
enable it at the boot time.
Once secretmem is enabled, the user will be able to create a file
descriptor using the memfd_secret() system call. The memory areas created
by mmap() calls from this file descriptor will be unmapped from the kernel
direct map and they will be only mapped in the page table of the processes
that have access to the file descriptor.
Secretmem is designed to provide the following protections:
* Enhanced protection (in conjunction with all the other in-kernel
attack prevention systems) against ROP attacks. Seceretmem makes
"simple" ROP insufficient to perform exfiltration, which increases the
required complexity of the attack. Along with other protections like
the kernel stack size limit and address space layout randomization which
make finding gadgets is really hard, absence of any in-kernel primitive
for accessing secret memory means the one gadget ROP attack can't work.
Since the only way to access secret memory is to reconstruct the missing
mapping entry, the attacker has to recover the physical page and insert
a PTE pointing to it in the kernel and then retrieve the contents. That
takes at least three gadgets which is a level of difficulty beyond most
standard attacks.
* Prevent cross-process secret userspace memory exposures. Once the
secret memory is allocated, the user can't accidentally pass it into the
kernel to be transmitted somewhere. The secreremem pages cannot be
accessed via the direct map and they are disallowed in GUP.
* Harden against exploited kernel flaws. In order to access secretmem,
a kernel-side attack would need to either walk the page tables and
create new ones, or spawn a new privileged uiserspace process to perform
secrets exfiltration using ptrace.
The file descriptor based memory has several advantages over the
"traditional" mm interfaces, such as mlock(), mprotect(), madvise(). File
descriptor approach allows explicit and controlled sharing of the memory
areas, it allows to seal the operations. Besides, file descriptor based
memory paves the way for VMMs to remove the secret memory range from the
userspace hipervisor process, for instance QEMU. Andy Lutomirski says:
"Getting fd-backed memory into a guest will take some possibly major
work in the kernel, but getting vma-backed memory into a guest without
mapping it in the host user address space seems much, much worse."
memfd_secret() is made a dedicated system call rather than an extension to
memfd_create() because it's purpose is to allow the user to create more
secure memory mappings rather than to simply allow file based access to
the memory. Nowadays a new system call cost is negligible while it is way
simpler for userspace to deal with a clear-cut system calls than with a
multiplexer or an overloaded syscall. Moreover, the initial
implementation of memfd_secret() is completely distinct from
memfd_create() so there is no much sense in overloading memfd_create() to
begin with. If there will be a need for code sharing between these
implementation it can be easily achieved without a need to adjust user
visible APIs.
The secret memory remains accessible in the process context using uaccess
primitives, but it is not exposed to the kernel otherwise; secret memory
areas are removed from the direct map and functions in the
follow_page()/get_user_page() family will refuse to return a page that
belongs to the secret memory area.
Once there will be a use case that will require exposing secretmem to the
kernel it will be an opt-in request in the system call flags so that user
would have to decide what data can be exposed to the kernel.
Removing of the pages from the direct map may cause its fragmentation on
architectures that use large pages to map the physical memory which
affects the system performance. However, the original Kconfig text for
CONFIG_DIRECT_GBPAGES said that gigabyte pages in the direct map "... can
improve the kernel's performance a tiny bit ..." (commit 00d1c5e05736
("x86: add gbpages switches")) and the recent report [1] showed that "...
although 1G mappings are a good default choice, there is no compelling
evidence that it must be the only choice". Hence, it is sufficient to
have secretmem disabled by default with the ability of a system
administrator to enable it at boot time.
Pages in the secretmem regions are unevictable and unmovable to avoid
accidental exposure of the sensitive data via swap or during page
migration.
Since the secretmem mappings are locked in memory they cannot exceed
RLIMIT_MEMLOCK. Since these mappings are already locked independently
from mlock(), an attempt to mlock()/munlock() secretmem range would fail
and mlockall()/munlockall() will ignore secretmem mappings.
However, unlike mlock()ed memory, secretmem currently behaves more like
long-term GUP: secretmem mappings are unmovable mappings directly consumed
by user space. With default limits, there is no excessive use of
secretmem and it poses no real problem in combination with
ZONE_MOVABLE/CMA, but in the future this should be addressed to allow
balanced use of large amounts of secretmem along with ZONE_MOVABLE/CMA.
A page that was a part of the secret memory area is cleared when it is
freed to ensure the data is not exposed to the next user of that page.
The following example demonstrates creation of a secret mapping (error
handling is omitted):
fd = memfd_secret(0);
ftruncate(fd, MAP_SIZE);
ptr = mmap(NULL, MAP_SIZE, PROT_READ | PROT_WRITE,
MAP_SHARED, fd, 0);
[1] https://lore.kernel.org/linux-mm/213b4567-46ce-f116-9cdf-bbd0c884eb3c@linux.intel.com/
[akpm@linux-foundation.org: suppress Kconfig whine]
Link: https://lkml.kernel.org/r/20210518072034.31572-5-rppt@kernel.org
Signed-off-by: Mike Rapoport <rppt@linux.ibm.com>
Acked-by: Hagen Paul Pfeifer <hagen@jauu.net>
Acked-by: James Bottomley <James.Bottomley@HansenPartnership.com>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Christopher Lameter <cl@linux.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Elena Reshetova <elena.reshetova@intel.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: James Bottomley <jejb@linux.ibm.com>
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Palmer Dabbelt <palmer@dabbelt.com>
Cc: Palmer Dabbelt <palmerdabbelt@google.com>
Cc: Paul Walmsley <paul.walmsley@sifive.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Rick Edgecombe <rick.p.edgecombe@intel.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Tycho Andersen <tycho@tycho.ws>
Cc: Will Deacon <will@kernel.org>
Cc: David Hildenbrand <david@redhat.com>
Cc: kernel test robot <lkp@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-07-07 18:08:03 -07:00
|
|
|
|
mm: Kconfig: make config SECRETMEM visible with EXPERT
Commit 6a108a14fa35 ("kconfig: rename CONFIG_EMBEDDED to CONFIG_EXPERT")
introduces CONFIG_EXPERT to carry the previous intent of CONFIG_EMBEDDED
and just gives that intent a much better name. That has been clearly a
good and long overdue renaming, and it is clearly an improvement to the
kernel build configuration that has shown to help managing the kernel
build configuration in the last decade.
However, rather than bravely and radically just deleting CONFIG_EMBEDDED,
this commit gives CONFIG_EMBEDDED a new intended semantics, but keeps it
open for future contributors to implement that intended semantics:
A new CONFIG_EMBEDDED option is added that automatically selects
CONFIG_EXPERT when enabled and can be used in the future to isolate
options that should only be considered for embedded systems (RISC
architectures, SLOB, etc).
Since then, this CONFIG_EMBEDDED implicitly had two purposes:
- It can make even more options visible beyond what CONFIG_EXPERT makes
visible. In other words, it may introduce another level of enabling the
visibility of configuration options: always visible, visible with
CONFIG_EXPERT and visible with CONFIG_EMBEDDED.
- Set certain default values of some configurations differently,
following the assumption that configuring a kernel build for an
embedded system generally starts with a different set of default values
compared to kernel builds for all other kind of systems.
Considering the second purpose, note that already probably arguing that a
kernel build for an embedded system would choose some values differently
is already tricky: the set of embedded systems with Linux kernels is
already quite diverse. Many embedded system have powerful CPUs and it
would not be clear that all embedded systems just optimize towards one
specific aspect, e.g., a smaller kernel image size. So, it is unclear if
starting with "one set of default configuration" that is induced by
CONFIG_EMBEDDED is a good offer for developers configuring their kernels.
Also, the differences of needed user-space features in an embedded system
compared to a non-embedded system are probably difficult or even
impossible to name in some generic way.
So it is not surprising that in the last decade hardly anyone has
contributed changes to make something default differently in case of
CONFIG_EMBEDDED=y.
Currently, in v6.0-rc4, SECRETMEM is the only config switched off if
CONFIG_EMBEDDED=y.
As long as that is actually the only option that currently is selected or
deselected, it is better to just make SECRETMEM configurable at build time
by experts using menuconfig instead.
Make SECRETMEM configurable when EXPERT is set and otherwise default to
yes. Further, SECRETMEM needs ARCH_HAS_SET_DIRECT_MAP.
This allows us to remove CONFIG_EMBEDDED in the close future.
Link: https://lkml.kernel.org/r/20221116131922.25533-1-lukas.bulwahn@gmail.com
Signed-off-by: Lukas Bulwahn <lukas.bulwahn@gmail.com>
Acked-by: Mike Rapoport <rppt@linux.ibm.com>
Acked-by: Arnd Bergmann <arnd@arndb.de>
Reviewed-by: Masahiro Yamada <masahiroy@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-11-16 14:19:22 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: introduce memfd_secret system call to create "secret" memory areas
Introduce "memfd_secret" system call with the ability to create memory
areas visible only in the context of the owning process and not mapped not
only to other processes but in the kernel page tables as well.
The secretmem feature is off by default and the user must explicitly
enable it at the boot time.
Once secretmem is enabled, the user will be able to create a file
descriptor using the memfd_secret() system call. The memory areas created
by mmap() calls from this file descriptor will be unmapped from the kernel
direct map and they will be only mapped in the page table of the processes
that have access to the file descriptor.
Secretmem is designed to provide the following protections:
* Enhanced protection (in conjunction with all the other in-kernel
attack prevention systems) against ROP attacks. Seceretmem makes
"simple" ROP insufficient to perform exfiltration, which increases the
required complexity of the attack. Along with other protections like
the kernel stack size limit and address space layout randomization which
make finding gadgets is really hard, absence of any in-kernel primitive
for accessing secret memory means the one gadget ROP attack can't work.
Since the only way to access secret memory is to reconstruct the missing
mapping entry, the attacker has to recover the physical page and insert
a PTE pointing to it in the kernel and then retrieve the contents. That
takes at least three gadgets which is a level of difficulty beyond most
standard attacks.
* Prevent cross-process secret userspace memory exposures. Once the
secret memory is allocated, the user can't accidentally pass it into the
kernel to be transmitted somewhere. The secreremem pages cannot be
accessed via the direct map and they are disallowed in GUP.
* Harden against exploited kernel flaws. In order to access secretmem,
a kernel-side attack would need to either walk the page tables and
create new ones, or spawn a new privileged uiserspace process to perform
secrets exfiltration using ptrace.
The file descriptor based memory has several advantages over the
"traditional" mm interfaces, such as mlock(), mprotect(), madvise(). File
descriptor approach allows explicit and controlled sharing of the memory
areas, it allows to seal the operations. Besides, file descriptor based
memory paves the way for VMMs to remove the secret memory range from the
userspace hipervisor process, for instance QEMU. Andy Lutomirski says:
"Getting fd-backed memory into a guest will take some possibly major
work in the kernel, but getting vma-backed memory into a guest without
mapping it in the host user address space seems much, much worse."
memfd_secret() is made a dedicated system call rather than an extension to
memfd_create() because it's purpose is to allow the user to create more
secure memory mappings rather than to simply allow file based access to
the memory. Nowadays a new system call cost is negligible while it is way
simpler for userspace to deal with a clear-cut system calls than with a
multiplexer or an overloaded syscall. Moreover, the initial
implementation of memfd_secret() is completely distinct from
memfd_create() so there is no much sense in overloading memfd_create() to
begin with. If there will be a need for code sharing between these
implementation it can be easily achieved without a need to adjust user
visible APIs.
The secret memory remains accessible in the process context using uaccess
primitives, but it is not exposed to the kernel otherwise; secret memory
areas are removed from the direct map and functions in the
follow_page()/get_user_page() family will refuse to return a page that
belongs to the secret memory area.
Once there will be a use case that will require exposing secretmem to the
kernel it will be an opt-in request in the system call flags so that user
would have to decide what data can be exposed to the kernel.
Removing of the pages from the direct map may cause its fragmentation on
architectures that use large pages to map the physical memory which
affects the system performance. However, the original Kconfig text for
CONFIG_DIRECT_GBPAGES said that gigabyte pages in the direct map "... can
improve the kernel's performance a tiny bit ..." (commit 00d1c5e05736
("x86: add gbpages switches")) and the recent report [1] showed that "...
although 1G mappings are a good default choice, there is no compelling
evidence that it must be the only choice". Hence, it is sufficient to
have secretmem disabled by default with the ability of a system
administrator to enable it at boot time.
Pages in the secretmem regions are unevictable and unmovable to avoid
accidental exposure of the sensitive data via swap or during page
migration.
Since the secretmem mappings are locked in memory they cannot exceed
RLIMIT_MEMLOCK. Since these mappings are already locked independently
from mlock(), an attempt to mlock()/munlock() secretmem range would fail
and mlockall()/munlockall() will ignore secretmem mappings.
However, unlike mlock()ed memory, secretmem currently behaves more like
long-term GUP: secretmem mappings are unmovable mappings directly consumed
by user space. With default limits, there is no excessive use of
secretmem and it poses no real problem in combination with
ZONE_MOVABLE/CMA, but in the future this should be addressed to allow
balanced use of large amounts of secretmem along with ZONE_MOVABLE/CMA.
A page that was a part of the secret memory area is cleared when it is
freed to ensure the data is not exposed to the next user of that page.
The following example demonstrates creation of a secret mapping (error
handling is omitted):
fd = memfd_secret(0);
ftruncate(fd, MAP_SIZE);
ptr = mmap(NULL, MAP_SIZE, PROT_READ | PROT_WRITE,
MAP_SHARED, fd, 0);
[1] https://lore.kernel.org/linux-mm/213b4567-46ce-f116-9cdf-bbd0c884eb3c@linux.intel.com/
[akpm@linux-foundation.org: suppress Kconfig whine]
Link: https://lkml.kernel.org/r/20210518072034.31572-5-rppt@kernel.org
Signed-off-by: Mike Rapoport <rppt@linux.ibm.com>
Acked-by: Hagen Paul Pfeifer <hagen@jauu.net>
Acked-by: James Bottomley <James.Bottomley@HansenPartnership.com>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Arnd Bergmann <arnd@arndb.de>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Christopher Lameter <cl@linux.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Elena Reshetova <elena.reshetova@intel.com>
Cc: "H. Peter Anvin" <hpa@zytor.com>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: James Bottomley <jejb@linux.ibm.com>
Cc: "Kirill A. Shutemov" <kirill@shutemov.name>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mark Rutland <mark.rutland@arm.com>
Cc: Michael Kerrisk <mtk.manpages@gmail.com>
Cc: Palmer Dabbelt <palmer@dabbelt.com>
Cc: Palmer Dabbelt <palmerdabbelt@google.com>
Cc: Paul Walmsley <paul.walmsley@sifive.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Rick Edgecombe <rick.p.edgecombe@intel.com>
Cc: Roman Gushchin <guro@fb.com>
Cc: Shakeel Butt <shakeelb@google.com>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Tycho Andersen <tycho@tycho.ws>
Cc: Will Deacon <will@kernel.org>
Cc: David Hildenbrand <david@redhat.com>
Cc: kernel test robot <lkp@intel.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-07-07 18:08:03 -07:00
|
|
|
|
mm: add a field to store names for private anonymous memory
In many userspace applications, and especially in VM based applications
like Android uses heavily, there are multiple different allocators in
use. At a minimum there is libc malloc and the stack, and in many cases
there are libc malloc, the stack, direct syscalls to mmap anonymous
memory, and multiple VM heaps (one for small objects, one for big
objects, etc.). Each of these layers usually has its own tools to
inspect its usage; malloc by compiling a debug version, the VM through
heap inspection tools, and for direct syscalls there is usually no way
to track them.
On Android we heavily use a set of tools that use an extended version of
the logic covered in Documentation/vm/pagemap.txt to walk all pages
mapped in userspace and slice their usage by process, shared (COW) vs.
unique mappings, backing, etc. This can account for real physical
memory usage even in cases like fork without exec (which Android uses
heavily to share as many private COW pages as possible between
processes), Kernel SamePage Merging, and clean zero pages. It produces
a measurement of the pages that only exist in that process (USS, for
unique), and a measurement of the physical memory usage of that process
with the cost of shared pages being evenly split between processes that
share them (PSS).
If all anonymous memory is indistinguishable then figuring out the real
physical memory usage (PSS) of each heap requires either a pagemap
walking tool that can understand the heap debugging of every layer, or
for every layer's heap debugging tools to implement the pagemap walking
logic, in which case it is hard to get a consistent view of memory
across the whole system.
Tracking the information in userspace leads to all sorts of problems.
It either needs to be stored inside the process, which means every
process has to have an API to export its current heap information upon
request, or it has to be stored externally in a filesystem that somebody
needs to clean up on crashes. It needs to be readable while the process
is still running, so it has to have some sort of synchronization with
every layer of userspace. Efficiently tracking the ranges requires
reimplementing something like the kernel vma trees, and linking to it
from every layer of userspace. It requires more memory, more syscalls,
more runtime cost, and more complexity to separately track regions that
the kernel is already tracking.
This patch adds a field to /proc/pid/maps and /proc/pid/smaps to show a
userspace-provided name for anonymous vmas. The names of named
anonymous vmas are shown in /proc/pid/maps and /proc/pid/smaps as
[anon:<name>].
Userspace can set the name for a region of memory by calling
prctl(PR_SET_VMA, PR_SET_VMA_ANON_NAME, start, len, (unsigned long)name)
Setting the name to NULL clears it. The name length limit is 80 bytes
including NUL-terminator and is checked to contain only printable ascii
characters (including space), except '[',']','\','$' and '`'.
Ascii strings are being used to have a descriptive identifiers for vmas,
which can be understood by the users reading /proc/pid/maps or
/proc/pid/smaps. Names can be standardized for a given system and they
can include some variable parts such as the name of the allocator or a
library, tid of the thread using it, etc.
The name is stored in a pointer in the shared union in vm_area_struct
that points to a null terminated string. Anonymous vmas with the same
name (equivalent strings) and are otherwise mergeable will be merged.
The name pointers are not shared between vmas even if they contain the
same name. The name pointer is stored in a union with fields that are
only used on file-backed mappings, so it does not increase memory usage.
CONFIG_ANON_VMA_NAME kernel configuration is introduced to enable this
feature. It keeps the feature disabled by default to prevent any
additional memory overhead and to avoid confusing procfs parsers on
systems which are not ready to support named anonymous vmas.
The patch is based on the original patch developed by Colin Cross, more
specifically on its latest version [1] posted upstream by Sumit Semwal.
It used a userspace pointer to store vma names. In that design, name
pointers could be shared between vmas. However during the last
upstreaming attempt, Kees Cook raised concerns [2] about this approach
and suggested to copy the name into kernel memory space, perform
validity checks [3] and store as a string referenced from
vm_area_struct.
One big concern is about fork() performance which would need to strdup
anonymous vma names. Dave Hansen suggested experimenting with
worst-case scenario of forking a process with 64k vmas having longest
possible names [4]. I ran this experiment on an ARM64 Android device
and recorded a worst-case regression of almost 40% when forking such a
process.
This regression is addressed in the followup patch which replaces the
pointer to a name with a refcounted structure that allows sharing the
name pointer between vmas of the same name. Instead of duplicating the
string during fork() or when splitting a vma it increments the refcount.
[1] https://lore.kernel.org/linux-mm/20200901161459.11772-4-sumit.semwal@linaro.org/
[2] https://lore.kernel.org/linux-mm/202009031031.D32EF57ED@keescook/
[3] https://lore.kernel.org/linux-mm/202009031022.3834F692@keescook/
[4] https://lore.kernel.org/linux-mm/5d0358ab-8c47-2f5f-8e43-23b89d6a8e95@intel.com/
Changes for prctl(2) manual page (in the options section):
PR_SET_VMA
Sets an attribute specified in arg2 for virtual memory areas
starting from the address specified in arg3 and spanning the
size specified in arg4. arg5 specifies the value of the attribute
to be set. Note that assigning an attribute to a virtual memory
area might prevent it from being merged with adjacent virtual
memory areas due to the difference in that attribute's value.
Currently, arg2 must be one of:
PR_SET_VMA_ANON_NAME
Set a name for anonymous virtual memory areas. arg5 should
be a pointer to a null-terminated string containing the
name. The name length including null byte cannot exceed
80 bytes. If arg5 is NULL, the name of the appropriate
anonymous virtual memory areas will be reset. The name
can contain only printable ascii characters (including
space), except '[',']','\','$' and '`'.
This feature is available only if the kernel is built with
the CONFIG_ANON_VMA_NAME option enabled.
[surenb@google.com: docs: proc.rst: /proc/PID/maps: fix malformed table]
Link: https://lkml.kernel.org/r/20211123185928.2513763-1-surenb@google.com
[surenb: rebased over v5.15-rc6, replaced userpointer with a kernel copy,
added input sanitization and CONFIG_ANON_VMA_NAME config. The bulk of the
work here was done by Colin Cross, therefore, with his permission, keeping
him as the author]
Link: https://lkml.kernel.org/r/20211019215511.3771969-2-surenb@google.com
Signed-off-by: Colin Cross <ccross@google.com>
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
Reviewed-by: Kees Cook <keescook@chromium.org>
Cc: Stephen Rothwell <sfr@canb.auug.org.au>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Cyrill Gorcunov <gorcunov@openvz.org>
Cc: Dave Hansen <dave.hansen@intel.com>
Cc: David Rientjes <rientjes@google.com>
Cc: "Eric W. Biederman" <ebiederm@xmission.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Ingo Molnar <mingo@kernel.org>
Cc: Jan Glauber <jan.glauber@gmail.com>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: John Stultz <john.stultz@linaro.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Oleg Nesterov <oleg@redhat.com>
Cc: Pekka Enberg <penberg@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Rob Landley <rob@landley.net>
Cc: "Serge E. Hallyn" <serge.hallyn@ubuntu.com>
Cc: Shaohua Li <shli@fusionio.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2022-01-14 14:05:59 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-05-12 20:22:56 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-07-08 12:14:10 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-11-23 17:42:04 -05:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-05-12 20:22:52 -07:00
|
|
|
|
2022-05-12 20:22:56 -07:00
|
|
|
|
|
|
|
|
|
2026-07-12 17:33:26 +08:00
|
|
|
|
2022-05-12 20:22:52 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 02:00:03 -06:00
|
|
|
|
mm: multi-gen LRU: groundwork
Evictable pages are divided into multiple generations for each lruvec.
The youngest generation number is stored in lrugen->max_seq for both
anon and file types as they are aged on an equal footing. The oldest
generation numbers are stored in lrugen->min_seq[] separately for anon
and file types as clean file pages can be evicted regardless of swap
constraints. These three variables are monotonically increasing.
Generation numbers are truncated into order_base_2(MAX_NR_GENS+1) bits
in order to fit into the gen counter in folio->flags. Each truncated
generation number is an index to lrugen->lists[]. The sliding window
technique is used to track at least MIN_NR_GENS and at most
MAX_NR_GENS generations. The gen counter stores a value within [1,
MAX_NR_GENS] while a page is on one of lrugen->lists[]. Otherwise it
stores 0.
There are two conceptually independent procedures: "the aging", which
produces young generations, and "the eviction", which consumes old
generations. They form a closed-loop system, i.e., "the page reclaim".
Both procedures can be invoked from userspace for the purposes of working
set estimation and proactive reclaim. These techniques are commonly used
to optimize job scheduling (bin packing) in data centers [1][2].
To avoid confusion, the terms "hot" and "cold" will be applied to the
multi-gen LRU, as a new convention; the terms "active" and "inactive" will
be applied to the active/inactive LRU, as usual.
The protection of hot pages and the selection of cold pages are based
on page access channels and patterns. There are two access channels:
one through page tables and the other through file descriptors. The
protection of the former channel is by design stronger because:
1. The uncertainty in determining the access patterns of the former
channel is higher due to the approximation of the accessed bit.
2. The cost of evicting the former channel is higher due to the TLB
flushes required and the likelihood of encountering the dirty bit.
3. The penalty of underprotecting the former channel is higher because
applications usually do not prepare themselves for major page
faults like they do for blocked I/O. E.g., GUI applications
commonly use dedicated I/O threads to avoid blocking rendering
threads.
There are also two access patterns: one with temporal locality and the
other without. For the reasons listed above, the former channel is
assumed to follow the former pattern unless VM_SEQ_READ or VM_RAND_READ is
present; the latter channel is assumed to follow the latter pattern unless
outlying refaults have been observed [3][4].
The next patch will address the "outlying refaults". Three macros, i.e.,
LRU_REFS_WIDTH, LRU_REFS_PGOFF and LRU_REFS_MASK, used later are added in
this patch to make the entire patchset less diffy.
A page is added to the youngest generation on faulting. The aging needs
to check the accessed bit at least twice before handing this page over to
the eviction. The first check takes care of the accessed bit set on the
initial fault; the second check makes sure this page has not been used
since then. This protocol, AKA second chance, requires a minimum of two
generations, hence MIN_NR_GENS.
[1] https://dl.acm.org/doi/10.1145/3297858.3304053
[2] https://dl.acm.org/doi/10.1145/3503222.3507731
[3] https://lwn.net/Articles/495543/
[4] https://lwn.net/Articles/815342/
Link: https://lkml.kernel.org/r/20220918080010.2920238-6-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 02:00:02 -06:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2022-09-18 02:00:10 -06:00
|
|
|
|
|
|
|
|
|
mm: multi-gen LRU: groundwork
Evictable pages are divided into multiple generations for each lruvec.
The youngest generation number is stored in lrugen->max_seq for both
anon and file types as they are aged on an equal footing. The oldest
generation numbers are stored in lrugen->min_seq[] separately for anon
and file types as clean file pages can be evicted regardless of swap
constraints. These three variables are monotonically increasing.
Generation numbers are truncated into order_base_2(MAX_NR_GENS+1) bits
in order to fit into the gen counter in folio->flags. Each truncated
generation number is an index to lrugen->lists[]. The sliding window
technique is used to track at least MIN_NR_GENS and at most
MAX_NR_GENS generations. The gen counter stores a value within [1,
MAX_NR_GENS] while a page is on one of lrugen->lists[]. Otherwise it
stores 0.
There are two conceptually independent procedures: "the aging", which
produces young generations, and "the eviction", which consumes old
generations. They form a closed-loop system, i.e., "the page reclaim".
Both procedures can be invoked from userspace for the purposes of working
set estimation and proactive reclaim. These techniques are commonly used
to optimize job scheduling (bin packing) in data centers [1][2].
To avoid confusion, the terms "hot" and "cold" will be applied to the
multi-gen LRU, as a new convention; the terms "active" and "inactive" will
be applied to the active/inactive LRU, as usual.
The protection of hot pages and the selection of cold pages are based
on page access channels and patterns. There are two access channels:
one through page tables and the other through file descriptors. The
protection of the former channel is by design stronger because:
1. The uncertainty in determining the access patterns of the former
channel is higher due to the approximation of the accessed bit.
2. The cost of evicting the former channel is higher due to the TLB
flushes required and the likelihood of encountering the dirty bit.
3. The penalty of underprotecting the former channel is higher because
applications usually do not prepare themselves for major page
faults like they do for blocked I/O. E.g., GUI applications
commonly use dedicated I/O threads to avoid blocking rendering
threads.
There are also two access patterns: one with temporal locality and the
other without. For the reasons listed above, the former channel is
assumed to follow the former pattern unless VM_SEQ_READ or VM_RAND_READ is
present; the latter channel is assumed to follow the latter pattern unless
outlying refaults have been observed [3][4].
The next patch will address the "outlying refaults". Three macros, i.e.,
LRU_REFS_WIDTH, LRU_REFS_PGOFF and LRU_REFS_MASK, used later are added in
this patch to make the entire patchset less diffy.
A page is added to the youngest generation on faulting. The aging needs
to check the accessed bit at least twice before handing this page over to
the eviction. The first check takes care of the accessed bit set on the
initial fault; the second check makes sure this page has not been used
since then. This protocol, AKA second chance, requires a minimum of two
generations, hence MIN_NR_GENS.
[1] https://dl.acm.org/doi/10.1145/3297858.3304053
[2] https://dl.acm.org/doi/10.1145/3503222.3507731
[3] https://lwn.net/Articles/495543/
[4] https://lwn.net/Articles/815342/
Link: https://lkml.kernel.org/r/20220918080010.2920238-6-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 02:00:02 -06:00
|
|
|
|
mm: multi-gen LRU: kill switch
Add /sys/kernel/mm/lru_gen/enabled as a kill switch. Components that
can be disabled include:
0x0001: the multi-gen LRU core
0x0002: walking page table, when arch_has_hw_pte_young() returns
true
0x0004: clearing the accessed bit in non-leaf PMD entries, when
CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG=y
[yYnN]: apply to all the components above
E.g.,
echo y >/sys/kernel/mm/lru_gen/enabled
cat /sys/kernel/mm/lru_gen/enabled
0x0007
echo 5 >/sys/kernel/mm/lru_gen/enabled
cat /sys/kernel/mm/lru_gen/enabled
0x0005
NB: the page table walks happen on the scale of seconds under heavy memory
pressure, in which case the mmap_lock contention is a lesser concern,
compared with the LRU lock contention and the I/O congestion. So far the
only well-known case of the mmap_lock contention happens on Android, due
to Scudo [1] which allocates several thousand VMAs for merely a few
hundred MBs. The SPF and the Maple Tree also have provided their own
assessments [2][3]. However, if walking page tables does worsen the
mmap_lock contention, the kill switch can be used to disable it. In this
case the multi-gen LRU will suffer a minor performance degradation, as
shown previously.
Clearing the accessed bit in non-leaf PMD entries can also be disabled,
since this behavior was not tested on x86 varieties other than Intel and
AMD.
[1] https://source.android.com/devices/tech/debug/scudo
[2] https://lore.kernel.org/r/20220128131006.67712-1-michel@lespinasse.org/
[3] https://lore.kernel.org/r/20220426150616.3937571-1-Liam.Howlett@oracle.com/
Link: https://lkml.kernel.org/r/20220918080010.2920238-11-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 02:00:07 -06:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 02:00:03 -06:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-12-27 14:12:02 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: multi-gen LRU: minimal implementation
To avoid confusion, the terms "promotion" and "demotion" will be applied
to the multi-gen LRU, as a new convention; the terms "activation" and
"deactivation" will be applied to the active/inactive LRU, as usual.
The aging produces young generations. Given an lruvec, it increments
max_seq when max_seq-min_seq+1 approaches MIN_NR_GENS. The aging promotes
hot pages to the youngest generation when it finds them accessed through
page tables; the demotion of cold pages happens consequently when it
increments max_seq. Promotion in the aging path does not involve any LRU
list operations, only the updates of the gen counter and
lrugen->nr_pages[]; demotion, unless as the result of the increment of
max_seq, requires LRU list operations, e.g., lru_deactivate_fn(). The
aging has the complexity O(nr_hot_pages), since it is only interested in
hot pages.
The eviction consumes old generations. Given an lruvec, it increments
min_seq when lrugen->lists[] indexed by min_seq%MAX_NR_GENS becomes empty.
A feedback loop modeled after the PID controller monitors refaults over
anon and file types and decides which type to evict when both types are
available from the same generation.
The protection of pages accessed multiple times through file descriptors
takes place in the eviction path. Each generation is divided into
multiple tiers. A page accessed N times through file descriptors is in
tier order_base_2(N). Tiers do not have dedicated lrugen->lists[], only
bits in folio->flags. The aforementioned feedback loop also monitors
refaults over all tiers and decides when to protect pages in which tiers
(N>1), using the first tier (N=0,1) as a baseline. The first tier
contains single-use unmapped clean pages, which are most likely the best
choices. In contrast to promotion in the aging path, the protection of a
page in the eviction path is achieved by moving this page to the next
generation, i.e., min_seq+1, if the feedback loop decides so. This
approach has the following advantages:
1. It removes the cost of activation in the buffered access path by
inferring whether pages accessed multiple times through file
descriptors are statistically hot and thus worth protecting in the
eviction path.
2. It takes pages accessed through page tables into account and avoids
overprotecting pages accessed multiple times through file
descriptors. (Pages accessed through page tables are in the first
tier, since N=0.)
3. More tiers provide better protection for pages accessed more than
twice through file descriptors, when under heavy buffered I/O
workloads.
Server benchmark results:
Single workload:
fio (buffered I/O): +[30, 32]%
IOPS BW
5.19-rc1: 2673k 10.2GiB/s
patch1-6: 3491k 13.3GiB/s
Single workload:
memcached (anon): -[4, 6]%
Ops/sec KB/sec
5.19-rc1: 1161501.04 45177.25
patch1-6: 1106168.46 43025.04
Configurations:
CPU: two Xeon 6154
Mem: total 256G
Node 1 was only used as a ram disk to reduce the variance in the
results.
patch drivers/block/brd.c <<EOF
99,100c99,100
< gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM;
< page = alloc_page(gfp_flags);
---
> gfp_flags = GFP_NOIO | __GFP_ZERO | __GFP_HIGHMEM | __GFP_THISNODE;
> page = alloc_pages_node(1, gfp_flags, 0);
EOF
cat >>/etc/systemd/system.conf <<EOF
CPUAffinity=numa
NUMAPolicy=bind
NUMAMask=0
EOF
cat >>/etc/memcached.conf <<EOF
-m 184320
-s /var/run/memcached/memcached.sock
-a 0766
-t 36
-B binary
EOF
cat fio.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkfs.ext4 /dev/ram0
mount -t ext4 /dev/ram0 /mnt
mkdir /sys/fs/cgroup/user.slice/test
echo 38654705664 >/sys/fs/cgroup/user.slice/test/memory.max
echo $$ >/sys/fs/cgroup/user.slice/test/cgroup.procs
fio -name=mglru --numjobs=72 --directory=/mnt --size=1408m \
--buffered=1 --ioengine=io_uring --iodepth=128 \
--iodepth_batch_submit=32 --iodepth_batch_complete=32 \
--rw=randread --random_distribution=random --norandommap \
--time_based --ramp_time=10m --runtime=5m --group_reporting
cat memcached.sh
modprobe brd rd_nr=1 rd_size=113246208
swapoff -a
mkswap /dev/ram0
swapon /dev/ram0
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=P:P -c 1 -t 36 \
--ratio 1:0 --pipeline 8 -d 2000
memtier_benchmark -S /var/run/memcached/memcached.sock \
-P memcache_binary -n allkeys --key-minimum=1 \
--key-maximum=65000000 --key-pattern=R:R -c 1 -t 36 \
--ratio 0:1 --pipeline 8 --randomize --distinct-client-seed
Client benchmark results:
kswapd profiles:
5.19-rc1
40.33% page_vma_mapped_walk (overhead)
21.80% lzo1x_1_do_compress (real work)
7.53% do_raw_spin_lock
3.95% _raw_spin_unlock_irq
2.52% vma_interval_tree_iter_next
2.37% folio_referenced_one
2.28% vma_interval_tree_subtree_search
1.97% anon_vma_interval_tree_iter_first
1.60% ptep_clear_flush
1.06% __zram_bvec_write
patch1-6
39.03% lzo1x_1_do_compress (real work)
18.47% page_vma_mapped_walk (overhead)
6.74% _raw_spin_unlock_irq
3.97% do_raw_spin_lock
2.49% ptep_clear_flush
2.48% anon_vma_interval_tree_iter_first
1.92% folio_referenced_one
1.88% __zram_bvec_write
1.48% memmove
1.31% vma_interval_tree_iter_next
Configurations:
CPU: single Snapdragon 7c
Mem: total 4G
ChromeOS MemoryPressure [1]
[1] https://chromium.googlesource.com/chromiumos/platform/tast-tests/
Link: https://lkml.kernel.org/r/20220918080010.2920238-7-yuzhao@google.com
Signed-off-by: Yu Zhao <yuzhao@google.com>
Acked-by: Brian Geffon <bgeffon@google.com>
Acked-by: Jan Alexander Steffens (heftig) <heftig@archlinux.org>
Acked-by: Oleksandr Natalenko <oleksandr@natalenko.name>
Acked-by: Steven Barrett <steven@liquorix.net>
Acked-by: Suleiman Souhlal <suleiman@google.com>
Tested-by: Daniel Byrne <djbyrne@mtu.edu>
Tested-by: Donald Carr <d@chaos-reins.com>
Tested-by: Holger Hoffstätte <holger@applied-asynchrony.com>
Tested-by: Konstantin Kharlamov <Hi-Angel@yandex.ru>
Tested-by: Shuang Zhai <szhai2@cs.rochester.edu>
Tested-by: Sofia Trinh <sofia.trinh@edi.works>
Tested-by: Vaibhav Jain <vaibhav@linux.ibm.com>
Cc: Andi Kleen <ak@linux.intel.com>
Cc: Aneesh Kumar K.V <aneesh.kumar@linux.ibm.com>
Cc: Barry Song <baohua@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: Hillf Danton <hdanton@sina.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Johannes Weiner <hannes@cmpxchg.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Miaohe Lin <linmiaohe@huawei.com>
Cc: Michael Larabel <Michael@MichaelLarabel.com>
Cc: Michal Hocko <mhocko@kernel.org>
Cc: Mike Rapoport <rppt@kernel.org>
Cc: Mike Rapoport <rppt@linux.ibm.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Tejun Heo <tj@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Will Deacon <will@kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2022-09-18 02:00:03 -06:00
|
|
|
|
|
|
|
|
|
mm: introduce CONFIG_PER_VMA_LOCK
Patch series "Per-VMA locks", v4.
LWN article describing the feature: https://lwn.net/Articles/906852/
Per-vma locks idea that was discussed during SPF [1] discussion at LSF/MM
last year [2], which concluded with suggestion that “a reader/writer
semaphore could be put into the VMA itself; that would have the effect of
using the VMA as a sort of range lock. There would still be contention at
the VMA level, but it would be an improvement.” This patchset implements
this suggested approach.
When handling page faults we lookup the VMA that contains the faulting
page under RCU protection and try to acquire its lock. If that fails we
fall back to using mmap_lock, similar to how SPF handled this situation.
One notable way the implementation deviates from the proposal is the way
VMAs are read-locked. During some of mm updates, multiple VMAs need to be
locked until the end of the update (e.g. vma_merge, split_vma, etc).
Tracking all the locked VMAs, avoiding recursive locks, figuring out when
it's safe to unlock previously locked VMAs would make the code more
complex. So, instead of the usual lock/unlock pattern, the proposed
solution marks a VMA as locked and provides an efficient way to:
1. Identify locked VMAs.
2. Unlock all locked VMAs in bulk.
We also postpone unlocking the locked VMAs until the end of the update,
when we do mmap_write_unlock. Potentially this keeps a VMA locked for
longer than is absolutely necessary but it results in a big reduction of
code complexity.
Read-locking a VMA is done using two sequence numbers - one in the
vm_area_struct and one in the mm_struct. VMA is considered read-locked
when these sequence numbers are equal. To read-lock a VMA we set the
sequence number in vm_area_struct to be equal to the sequence number in
mm_struct. To unlock all VMAs we increment mm_struct's seq number. This
allows for an efficient way to track locked VMAs and to drop the locks on
all VMAs at the end of the update.
The patchset implements per-VMA locking only for anonymous pages which are
not in swap and avoids userfaultfs as their implementation is more
complex. Additional support for file-back page faults, swapped and user
pages can be added incrementally.
Performance benchmarks show similar although slightly smaller benefits as
with SPF patchset (~75% of SPF benefits). Still, with lower complexity
this approach might be more desirable.
Since RFC was posted in September 2022, two separate Google teams outside
of Android evaluated the patchset and confirmed positive results. Here
are the known usecases when per-VMA locks show benefits:
Android:
Apps with high number of threads (~100) launch times improve by up to 20%.
Each thread mmaps several areas upon startup (Stack and Thread-local
storage (TLS), thread signal stack, indirect ref table), which requires
taking mmap_lock in write mode. Page faults take mmap_lock in read mode.
During app launch, both thread creation and page faults establishing the
active workinget are happening in parallel and that causes lock contention
between mm writers and readers even if updates and page faults are
happening in different VMAs. Per-vma locks prevent this contention by
providing more granular lock.
Google Fibers:
We have several dynamically sized thread pools that spawn new threads
under increased load and reduce their number when idling. For example,
Google's in-process scheduling/threading framework, UMCG/Fibers, is backed
by such a thread pool. When idling, only a small number of idle worker
threads are available; when a spike of incoming requests arrive, each
request is handled in its own "fiber", which is a work item posted onto a
UMCG worker thread; quite often these spikes lead to a number of new
threads spawning. Each new thread needs to allocate and register an RSEQ
section on its TLS, then register itself with the kernel as a UMCG worker
thread, and only after that it can be considered by the in-process
UMCG/Fiber scheduler as available to do useful work. In short, during an
incoming workload spike new threads have to be spawned, and they perform
several syscalls (RSEQ registration, UMCG worker registration, memory
allocations) before they can actually start doing useful work. Removing
any bottlenecks on this thread startup path will greatly improve our
services' latencies when faced with request/workload spikes.
At high scale, mmap_lock contention during thread creation and stack page
faults leads to user-visible multi-second serving latencies in a similar
pattern to Android app startup. Per-VMA locking patchset has been run
successfully in limited experiments with user-facing production workloads.
In these experiments, we observed that the peak thread creation rate was
high enough that thread creation is no longer a bottleneck.
TCP zerocopy receive:
From the point of view of TCP zerocopy receive, the per-vma lock patch is
massively beneficial.
In today's implementation, a process with N threads where N - 1 are
performing zerocopy receive and 1 thread is performing madvise() with the
write lock taken (e.g. needs to change vm_flags) will result in all N -1
receive threads blocking until the madvise is done. Conversely, on a busy
process receiving a lot of data, an madvise operation that does need to
take the mmap lock in write mode will need to wait for all of the receives
to be done - a lose:lose proposition. Per-VMA locking _removes_ by
definition this source of contention entirely.
There are other benefits for receive as well, chiefly a reduction in
cacheline bouncing across receiving threads for locking/unlocking the
single mmap lock. On an RPC style synthetic workload with 4KB RPCs:
1a) The find+lock+unlock VMA path in the base case, without the
per-vma lock patchset, is about 0.7% of cycles as measured by perf.
1b) mmap_read_lock + mmap_read_unlock in the base case is about 0.5%
cycles overall - most of this is within the TCP read hotpath (a small
fraction is 'other' usage in the system).
2a) The find+lock+unlock VMA path, with the per-vma patchset and a
trivial patch written to take advantage of it in TCP, is about 0.4% of
cycles (down from 0.7% above)
2b) mmap_read_lock + mmap_read_unlock in the per-vma patchset is <
0.1% cycles and is out of the TCP read hotpath entirely (down from
0.5% before, the remaining usage is the 'other' usage in the system).
So, in addition to entirely removing an onerous source of contention,
it also reduces the CPU cycles of TCP receive zerocopy by about 0.5%+
(compared to overall cycles in perf) for the 'small' RPC scenario.
In https://lkml.kernel.org/r/87fsaqouyd.fsf_-_@stealth, Punit
demonstrated throughput improvements of as much as 188% from this
patchset.
This patch (of 25):
This configuration variable will be used to build the support for VMA
locking during page fault handling.
This is enabled on supported architectures with SMP and MMU set.
The architecture support is needed since the page fault handler is called
from the architecture's page faulting code which needs modifications to
handle faults under VMA lock.
Link: https://lkml.kernel.org/r/20230227173632.3292573-1-surenb@google.com
Link: https://lkml.kernel.org/r/20230227173632.3292573-10-surenb@google.com
Signed-off-by: Suren Baghdasaryan <surenb@google.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2023-02-27 09:36:08 -08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: introduce new 'lock_mm_and_find_vma()' page fault helper
.. and make x86 use it.
This basically extracts the existing x86 "find and expand faulting vma"
code, but extends it to also take the mmap lock for writing in case we
actually do need to expand the vma.
We've historically short-circuited that case, and have some rather ugly
special logic to serialize the stack segment expansion (since we only
hold the mmap lock for reading) that doesn't match the normal VM
locking.
That slight violation of locking worked well, right up until it didn't:
the maple tree code really does want proper locking even for simple
extension of an existing vma.
So extract the code for "look up the vma of the fault" from x86, fix it
up to do the necessary write locking, and make it available as a helper
function for other architectures that can use the common helper.
Note: I say "common helper", but it really only handles the normal
stack-grows-down case. Which is all architectures except for PA-RISC
and IA64. So some rare architectures can't use the helper, but if they
care they'll just need to open-code this logic.
It's also worth pointing out that this code really would like to have an
optimistic "mmap_upgrade_trylock()" to make it quicker to go from a
read-lock (for the common case) to taking the write lock (for having to
extend the vma) in the normal single-threaded situation where there is
no other locking activity.
But that _is_ all the very uncommon special case, so while it would be
nice to have such an operation, it probably doesn't matter in reality.
I did put in the skeleton code for such a possible future expansion,
even if it only acts as pseudo-documentation for what we're doing.
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2023-06-15 15:17:36 -07:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2023-10-27 08:05:20 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-05-05 19:06:18 +03:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-08-07 09:41:01 +03:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-08-07 09:41:03 +03:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-05-30 21:45:43 +08:00
|
|
|
|
2024-08-07 09:41:03 +03:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-08-06 09:58:38 +03:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2024-10-01 23:58:40 +01:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-04-07 11:09:34 +05:30
|
|
|
|
2024-10-01 23:58:40 +01:00
|
|
|
|
2026-01-27 20:13:00 +08:00
|
|
|
|
mm: pgtable: reclaim empty PTE page in madvise(MADV_DONTNEED)
Now in order to pursue high performance, applications mostly use some
high-performance user-mode memory allocators, such as jemalloc or
tcmalloc. These memory allocators use madvise(MADV_DONTNEED or MADV_FREE)
to release physical memory, but neither MADV_DONTNEED nor MADV_FREE will
release page table memory, which may cause huge page table memory usage.
The following are a memory usage snapshot of one process which actually
happened on our server:
VIRT: 55t
RES: 590g
VmPTE: 110g
In this case, most of the page table entries are empty. For such a PTE
page where all entries are empty, we can actually free it back to the
system for others to use.
As a first step, this commit aims to synchronously free the empty PTE
pages in madvise(MADV_DONTNEED) case. We will detect and free empty PTE
pages in zap_pte_range(), and will add zap_details.reclaim_pt to exclude
cases other than madvise(MADV_DONTNEED).
Once an empty PTE is detected, we first try to hold the pmd lock within
the pte lock. If successful, we clear the pmd entry directly (fast path).
Otherwise, we wait until the pte lock is released, then re-hold the pmd
and pte locks and loop PTRS_PER_PTE times to check pte_none() to re-detect
whether the PTE page is empty and free it (slow path).
For other cases such as madvise(MADV_FREE), consider scanning and freeing
empty PTE pages asynchronously in the future.
The following code snippet can show the effect of optimization:
mmap 50G
while (1) {
for (; i < 1024 * 25; i++) {
touch 2M memory
madvise MADV_DONTNEED 2M
}
}
As we can see, the memory usage of VmPTE is reduced:
before after
VIRT 50.0 GB 50.0 GB
RES 3.1 MB 3.1 MB
VmPTE 102640 KB 240 KB
[zhengqi.arch@bytedance.com: fix uninitialized symbol 'ptl']
Link: https://lkml.kernel.org/r/20241206112348.51570-1-zhengqi.arch@bytedance.com
Link: https://lore.kernel.org/linux-mm/224e6a4e-43b5-4080-bdd8-b0a6fb2f0853@stanley.mountain/
Link: https://lkml.kernel.org/r/92aba2b319a734913f18ba41e7d86a265f0b84e2.1733305182.git.zhengqi.arch@bytedance.com
Signed-off-by: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jann Horn <jannh@google.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Peter Xu <peterx@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Will Deacon <will@kernel.org>
Cc: Zach O'Keefe <zokeefe@google.com>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-12-04 19:09:49 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2026-01-27 20:13:01 +08:00
|
|
|
|
|
|
|
|
|
mm: pgtable: reclaim empty PTE page in madvise(MADV_DONTNEED)
Now in order to pursue high performance, applications mostly use some
high-performance user-mode memory allocators, such as jemalloc or
tcmalloc. These memory allocators use madvise(MADV_DONTNEED or MADV_FREE)
to release physical memory, but neither MADV_DONTNEED nor MADV_FREE will
release page table memory, which may cause huge page table memory usage.
The following are a memory usage snapshot of one process which actually
happened on our server:
VIRT: 55t
RES: 590g
VmPTE: 110g
In this case, most of the page table entries are empty. For such a PTE
page where all entries are empty, we can actually free it back to the
system for others to use.
As a first step, this commit aims to synchronously free the empty PTE
pages in madvise(MADV_DONTNEED) case. We will detect and free empty PTE
pages in zap_pte_range(), and will add zap_details.reclaim_pt to exclude
cases other than madvise(MADV_DONTNEED).
Once an empty PTE is detected, we first try to hold the pmd lock within
the pte lock. If successful, we clear the pmd entry directly (fast path).
Otherwise, we wait until the pte lock is released, then re-hold the pmd
and pte locks and loop PTRS_PER_PTE times to check pte_none() to re-detect
whether the PTE page is empty and free it (slow path).
For other cases such as madvise(MADV_FREE), consider scanning and freeing
empty PTE pages asynchronously in the future.
The following code snippet can show the effect of optimization:
mmap 50G
while (1) {
for (; i < 1024 * 25; i++) {
touch 2M memory
madvise MADV_DONTNEED 2M
}
}
As we can see, the memory usage of VmPTE is reduced:
before after
VIRT 50.0 GB 50.0 GB
RES 3.1 MB 3.1 MB
VmPTE 102640 KB 240 KB
[zhengqi.arch@bytedance.com: fix uninitialized symbol 'ptl']
Link: https://lkml.kernel.org/r/20241206112348.51570-1-zhengqi.arch@bytedance.com
Link: https://lore.kernel.org/linux-mm/224e6a4e-43b5-4080-bdd8-b0a6fb2f0853@stanley.mountain/
Link: https://lkml.kernel.org/r/92aba2b319a734913f18ba41e7d86a265f0b84e2.1733305182.git.zhengqi.arch@bytedance.com
Signed-off-by: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jann Horn <jannh@google.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Peter Xu <peterx@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Will Deacon <will@kernel.org>
Cc: Zach O'Keefe <zokeefe@google.com>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-12-04 19:09:49 +08:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-08-11 13:26:31 +02:00
|
|
|
|
|
|
|
|
|
mm: pgtable: reclaim empty PTE page in madvise(MADV_DONTNEED)
Now in order to pursue high performance, applications mostly use some
high-performance user-mode memory allocators, such as jemalloc or
tcmalloc. These memory allocators use madvise(MADV_DONTNEED or MADV_FREE)
to release physical memory, but neither MADV_DONTNEED nor MADV_FREE will
release page table memory, which may cause huge page table memory usage.
The following are a memory usage snapshot of one process which actually
happened on our server:
VIRT: 55t
RES: 590g
VmPTE: 110g
In this case, most of the page table entries are empty. For such a PTE
page where all entries are empty, we can actually free it back to the
system for others to use.
As a first step, this commit aims to synchronously free the empty PTE
pages in madvise(MADV_DONTNEED) case. We will detect and free empty PTE
pages in zap_pte_range(), and will add zap_details.reclaim_pt to exclude
cases other than madvise(MADV_DONTNEED).
Once an empty PTE is detected, we first try to hold the pmd lock within
the pte lock. If successful, we clear the pmd entry directly (fast path).
Otherwise, we wait until the pte lock is released, then re-hold the pmd
and pte locks and loop PTRS_PER_PTE times to check pte_none() to re-detect
whether the PTE page is empty and free it (slow path).
For other cases such as madvise(MADV_FREE), consider scanning and freeing
empty PTE pages asynchronously in the future.
The following code snippet can show the effect of optimization:
mmap 50G
while (1) {
for (; i < 1024 * 25; i++) {
touch 2M memory
madvise MADV_DONTNEED 2M
}
}
As we can see, the memory usage of VmPTE is reduced:
before after
VIRT 50.0 GB 50.0 GB
RES 3.1 MB 3.1 MB
VmPTE 102640 KB 240 KB
[zhengqi.arch@bytedance.com: fix uninitialized symbol 'ptl']
Link: https://lkml.kernel.org/r/20241206112348.51570-1-zhengqi.arch@bytedance.com
Link: https://lore.kernel.org/linux-mm/224e6a4e-43b5-4080-bdd8-b0a6fb2f0853@stanley.mountain/
Link: https://lkml.kernel.org/r/92aba2b319a734913f18ba41e7d86a265f0b84e2.1733305182.git.zhengqi.arch@bytedance.com
Signed-off-by: Qi Zheng <zhengqi.arch@bytedance.com>
Cc: Andy Lutomirski <luto@kernel.org>
Cc: Catalin Marinas <catalin.marinas@arm.com>
Cc: Dave Hansen <dave.hansen@linux.intel.com>
Cc: David Hildenbrand <david@redhat.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Hugh Dickins <hughd@google.com>
Cc: Jann Horn <jannh@google.com>
Cc: Lorenzo Stoakes <lorenzo.stoakes@oracle.com>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Muchun Song <muchun.song@linux.dev>
Cc: Peter Xu <peterx@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Will Deacon <will@kernel.org>
Cc: Zach O'Keefe <zokeefe@google.com>
Cc: Dan Carpenter <dan.carpenter@linaro.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
2024-12-04 19:09:49 +08:00
|
|
|
|
2025-12-15 15:03:15 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
2025-12-15 15:03:23 +00:00
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
mm: introduce Data Access MONitor (DAMON)
Patch series "Introduce Data Access MONitor (DAMON)", v34.
Introduction
============
DAMON is a data access monitoring framework for the Linux kernel. The
core mechanisms of DAMON called 'region based sampling' and 'adaptive
regions adjustment' (refer to 'mechanisms.rst' in the 11th patch of this
patchset for the detail) make it
- accurate (The monitored information is useful for DRAM level memory
management. It might not appropriate for Cache-level accuracy,
though.),
- light-weight (The monitoring overhead is low enough to be applied
online while making no impact on the performance of the target
workloads.), and
- scalable (the upper-bound of the instrumentation overhead is
controllable regardless of the size of target workloads.).
Using this framework, therefore, several memory management mechanisms such
as reclamation and THP can be optimized to aware real data access
patterns. Experimental access pattern aware memory management
optimization works that incurring high instrumentation overhead will be
able to have another try.
Though DAMON is for kernel subsystems, it can be easily exposed to the
user space by writing a DAMON-wrapper kernel subsystem. Then, user space
users who have some special workloads will be able to write personalized
tools or applications for deeper understanding and specialized
optimizations of their systems.
DAMON is also merged in two public Amazon Linux kernel trees that based on
v5.4.y[1] and v5.10.y[2].
[1] https://github.com/amazonlinux/linux/tree/amazon-5.4.y/master/mm/damon
[2] https://github.com/amazonlinux/linux/tree/amazon-5.10.y/master/mm/damon
The userspace tool[1] is available, released under GPLv2, and actively
being maintained. I am also planning to implement another basic user
interface in perf[2]. Also, the basic test suite for DAMON is available
under GPLv2[3].
[1] https://github.com/awslabs/damo
[2] https://lore.kernel.org/linux-mm/20210107120729.22328-1-sjpark@amazon.com/
[3] https://github.com/awslabs/damon-tests
Long-term Plan
--------------
DAMON is a part of a project called Data Access-aware Operating System
(DAOS). As the name implies, I want to improve the performance and
efficiency of systems using fine-grained data access patterns. The
optimizations are for both kernel and user spaces. I will therefore
modify or create kernel subsystems, export some of those to user space and
implement user space library / tools. Below shows the layers and
components for the project.
---------------------------------------------------------------------------
Primitives: PTE Accessed bit, PG_idle, rmap, (Intel CMT), ...
Framework: DAMON
Features: DAMOS, virtual addr, physical addr, ...
Applications: DAMON-debugfs, (DARC), ...
^^^^^^^^^^^^^^^^^^^^^^^ KERNEL SPACE ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
Raw Interface: debugfs, (sysfs), (damonfs), tracepoints, (sys_damon), ...
vvvvvvvvvvvvvvvvvvvvvvv USER SPACE vvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvv
Library: (libdamon), ...
Tools: DAMO, (perf), ...
---------------------------------------------------------------------------
The components in parentheses or marked as '...' are not implemented yet
but in the future plan. IOW, those are the TODO tasks of DAOS project.
For more detail, please refer to the plans:
https://lore.kernel.org/linux-mm/20201202082731.24828-1-sjpark@amazon.com/
Evaluations
===========
We evaluated DAMON's overhead, monitoring quality and usefulness using 24
realistic workloads on my QEMU/KVM based virtual machine running a kernel
that v24 DAMON patchset is applied.
DAMON is lightweight. It increases system memory usage by 0.39% and slows
target workloads down by 1.16%.
DAMON is accurate and useful for memory management optimizations. An
experimental DAMON-based operation scheme for THP, namely 'ethp', removes
76.15% of THP memory overheads while preserving 51.25% of THP speedup.
Another experimental DAMON-based 'proactive reclamation' implementation,
'prcl', reduces 93.38% of residential sets and 23.63% of system memory
footprint while incurring only 1.22% runtime overhead in the best case
(parsec3/freqmine).
NOTE that the experimental THP optimization and proactive reclamation are
not for production but only for proof of concepts.
Please refer to the official document[1] or "Documentation/admin-guide/mm:
Add a document for DAMON" patch in this patchset for detailed evaluation
setup and results.
[1] https://damonitor.github.io/doc/html/latest-damon/admin-guide/mm/damon/eval.html
Real-world User Story
=====================
In summary, DAMON has used on production systems and proved its usefulness.
DAMON as a profiler
-------------------
We analyzed characteristics of a large scale production systems of our
customers using DAMON. The systems utilize 70GB DRAM and 36 CPUs. From
this, we were able to find interesting things below.
There were obviously different access pattern under idle workload and
active workload. Under the idle workload, it accessed large memory
regions with low frequency, while the active workload accessed small
memory regions with high freuqnecy.
DAMON found a 7GB memory region that showing obviously high access
frequency under the active workload. We believe this is the
performance-effective working set and need to be protected.
There was a 4KB memory region that showing highest access frequency under
not only active but also idle workloads. We think this must be a hottest
code section like thing that should never be paged out.
For this analysis, DAMON used only 0.3-1% of single CPU time. Because we
used recording-based analysis, it consumed about 3-12 MB of disk space per
20 minutes. This is only small amount of disk space, but we can further
reduce the disk usage by using non-recording-based DAMON features. I'd
like to argue that only DAMON can do such detailed analysis (finding 4KB
highest region in 70GB memory) with the light overhead.
DAMON as a system optimization tool
-----------------------------------
We also found below potential performance problems on the systems and made
DAMON-based solutions.
The system doesn't want to make the workload suffer from the page
reclamation and thus it utilizes enough DRAM but no swap device. However,
we found the system is actively reclaiming file-backed pages, because the
system has intensive file IO. The file IO turned out to be not
performance critical for the workload, but the customer wanted to ensure
performance critical file-backed pages like code section to not mistakenly
be evicted.
Using direct IO should or `mlock()` would be a straightforward solution,
but modifying the user space code is not easy for the customer.
Alternatively, we could use DAMON-based operation scheme[1]. By using it,
we can ask DAMON to track access frequency of each region and make
'process_madvise(MADV_WILLNEED)[2]' call for regions having specific size
and access frequency for a time interval.
We also found the system is having high number of TLB misses. We tried
'always' THP enabled policy and it greatly reduced TLB misses, but the
page reclamation also been more frequent due to the THP internal
fragmentation caused memory bloat. We could try another DAMON-based
operation scheme that applies 'MADV_HUGEPAGE' to memory regions having
>=2MB size and high access frequency, while applying 'MADV_NOHUGEPAGE' to
regions having <2MB size and low access frequency.
We do not own the systems so we only reported the analysis results and
possible optimization solutions to the customers. The customers satisfied
about the analysis results and promised to try the optimization guides.
[1] https://lore.kernel.org/linux-mm/20201006123931.5847-1-sjpark@amazon.com/
[2] https://lore.kernel.org/linux-api/20200622192900.22757-4-minchan@kernel.org/
Comparison with Idle Page Tracking
==================================
Idle Page Tracking allows users to set and read idleness of pages using a
bitmap file which represents each page with each bit of the file. One
recommended usage of it is working set size detection. Users can do that
by
1. find PFN of each page for workloads in interest,
2. set all the pages as idle by doing writes to the bitmap file,
3. wait until the workload accesses its working set, and
4. read the idleness of the pages again and count pages became not idle.
NOTE: While Idle Page Tracking is for user space users, DAMON is primarily
designed for kernel subsystems though it can easily exposed to the user
space. Hence, this section only assumes such user space use of DAMON.
For what use cases Idle Page Tracking would be better?
------------------------------------------------------
1. Flexible usecases other than hotness monitoring.
Because Idle Page Tracking allows users to control the primitive (Page
idleness) by themselves, Idle Page Tracking users can do anything they
want. Meanwhile, DAMON is primarily designed to monitor the hotness of
each memory region. For this, DAMON asks users to provide sampling
interval and aggregation interval. For the reason, there could be some
use case that using Idle Page Tracking is simpler.
2. Physical memory monitoring.
Idle Page Tracking receives PFN range as input, so natively supports
physical memory monitoring.
DAMON is designed to be extensible for multiple address spaces and use
cases by implementing and using primitives for the given use case.
Therefore, by theory, DAMON has no limitation in the type of target
address space as long as primitives for the given address space exists.
However, the default primitives introduced by this patchset supports only
virtual address spaces.
Therefore, for physical memory monitoring, you should implement your own
primitives and use it, or simply use Idle Page Tracking.
Nonetheless, RFC patchsets[1] for the physical memory address space
primitives is already available. It also supports user memory same to
Idle Page Tracking.
[1] https://lore.kernel.org/linux-mm/20200831104730.28970-1-sjpark@amazon.com/
For what use cases DAMON is better?
-----------------------------------
1. Hotness Monitoring.
Idle Page Tracking let users know only if a page frame is accessed or not.
For hotness check, the user should write more code and use more memory.
DAMON do that by itself.
2. Low Monitoring Overhead
DAMON receives user's monitoring request with one step and then provide
the results. So, roughly speaking, DAMON require only O(1) user/kernel
context switches.
In case of Idle Page Tracking, however, because the interface receives
contiguous page frames, the number of user/kernel context switches
increases as the monitoring target becomes complex and huge. As a result,
the context switch overhead could be not negligible.
Moreover, DAMON is born to handle with the monitoring overhead. Because
the core mechanism is pure logical, Idle Page Tracking users might be able
to implement the mechanism on their own, but it would be time consuming
and the user/kernel context switching will still more frequent than that
of DAMON. Also, the kernel subsystems cannot use the logic in this case.
3. Page granularity working set size detection.
Until v22 of this patchset, this was categorized as the thing Idle Page
Tracking could do better, because DAMON basically maintains additional
metadata for each of the monitoring target regions. So, in the page
granularity working set size detection use case, DAMON would incur (number
of monitoring target pages * size of metadata) memory overhead. Size of
the single metadata item is about 54 bytes, so assuming 4KB pages, about
1.3% of monitoring target pages will be additionally used.
All essential metadata for Idle Page Tracking are embedded in 'struct
page' and page table entries. Therefore, in this use case, only one
counter variable for working set size accounting is required if Idle Page
Tracking is used.
There are more details to consider, but roughly speaking, this is true in
most cases.
However, the situation changed from v23. Now DAMON supports arbitrary
types of monitoring targets, which don't use the metadata. Using that,
DAMON can do the working set size detection with no additional space
overhead but less user-kernel context switch. A first draft for the
implementation of monitoring primitives for this usage is available in a
DAMON development tree[1]. An RFC patchset for it based on this patchset
will also be available soon.
Since v24, the arbitrary type support is dropped from this patchset
because this patchset doesn't introduce real use of the type. You can
still get it from the DAMON development tree[2], though.
[1] https://github.com/sjp38/linux/tree/damon/pgidle_hack
[2] https://github.com/sjp38/linux/tree/damon/master
4. More future usecases
While Idle Page Tracking has tight coupling with base primitives (PG_Idle
and page table Accessed bits), DAMON is designed to be extensible for many
use cases and address spaces. If you need some special address type or
want to use special h/w access check primitives, you can write your own
primitives for that and configure DAMON to use those. Therefore, if your
use case could be changed a lot in future, using DAMON could be better.
Can I use both Idle Page Tracking and DAMON?
--------------------------------------------
Yes, though using them concurrently for overlapping memory regions could
result in interference to each other. Nevertheless, such use case would
be rare or makes no sense at all. Even in the case, the noise would bot
be really significant. So, you can choose whatever you want depending on
the characteristics of your use cases.
More Information
================
We prepared a showcase web site[1] that you can get more information.
There are
- the official documentations[2],
- the heatmap format dynamic access pattern of various realistic workloads for
heap area[3], mmap()-ed area[4], and stack[5] area,
- the dynamic working set size distribution[6] and chronological working set
size changes[7], and
- the latest performance test results[8].
[1] https://damonitor.github.io/_index
[2] https://damonitor.github.io/doc/html/latest-damon
[3] https://damonitor.github.io/test/result/visual/latest/rec.heatmap.0.png.html
[4] https://damonitor.github.io/test/result/visual/latest/rec.heatmap.1.png.html
[5] https://damonitor.github.io/test/result/visual/latest/rec.heatmap.2.png.html
[6] https://damonitor.github.io/test/result/visual/latest/rec.wss_sz.png.html
[7] https://damonitor.github.io/test/result/visual/latest/rec.wss_time.png.html
[8] https://damonitor.github.io/test/result/perf/latest/html/index.html
Baseline and Complete Git Trees
===============================
The patches are based on the latest -mm tree, specifically
v5.14-rc1-mmots-2021-07-15-18-47 of https://github.com/hnaz/linux-mm. You can
also clone the complete git tree:
$ git clone git://github.com/sjp38/linux -b damon/patches/v34
The web is also available:
https://github.com/sjp38/linux/releases/tag/damon/patches/v34
Development Trees
-----------------
There are a couple of trees for entire DAMON patchset series and features
for future release.
- For latest release: https://github.com/sjp38/linux/tree/damon/master
- For next release: https://github.com/sjp38/linux/tree/damon/next
Long-term Support Trees
-----------------------
For people who want to test DAMON but using LTS kernels, there are another
couple of trees based on two latest LTS kernels respectively and
containing the 'damon/master' backports.
- For v5.4.y: https://github.com/sjp38/linux/tree/damon/for-v5.4.y
- For v5.10.y: https://github.com/sjp38/linux/tree/damon/for-v5.10.y
Amazon Linux Kernel Trees
-------------------------
DAMON is also merged in two public Amazon Linux kernel trees that based on
v5.4.y[1] and v5.10.y[2].
[1] https://github.com/amazonlinux/linux/tree/amazon-5.4.y/master/mm/damon
[2] https://github.com/amazonlinux/linux/tree/amazon-5.10.y/master/mm/damon
Git Tree for Diff of Patches
============================
For easy review of diff between different versions of each patch, I
prepared a git tree containing all versions of the DAMON patchset series:
https://github.com/sjp38/damon-patches
You can clone it and use 'diff' for easy review of changes between
different versions of the patchset. For example:
$ git clone https://github.com/sjp38/damon-patches && cd damon-patches
$ diff -u damon/v33 damon/v34
Sequence Of Patches
===================
First three patches implement the core logics of DAMON. The 1st patch
introduces basic sampling based hotness monitoring for arbitrary types of
targets. Following two patches implement the core mechanisms for control
of overhead and accuracy, namely regions based sampling (patch 2) and
adaptive regions adjustment (patch 3).
Now the essential parts of DAMON is complete, but it cannot work unless
someone provides monitoring primitives for a specific use case. The
following two patches make it just work for virtual address spaces
monitoring. The 4th patch makes 'PG_idle' can be used by DAMON and the
5th patch implements the virtual memory address space specific monitoring
primitives using page table Accessed bits and the 'PG_idle' page flag.
Now DAMON just works for virtual address space monitoring via the kernel
space api. To let the user space users can use DAMON, following four
patches add interfaces for them. The 6th patch adds a tracepoint for
monitoring results. The 7th patch implements a DAMON application kernel
module, namely damon-dbgfs, that simply wraps DAMON and exposes DAMON
interface to the user space via the debugfs interface. The 8th patch
further exports pid of monitoring thread (kdamond) to user space for
easier cpu usage accounting, and the 9th patch makes the debugfs interface
to support multiple contexts.
Three patches for maintainability follows. The 10th patch adds
documentations for both the user space and the kernel space. The 11th
patch provides unit tests (based on the kunit) while the 12th patch adds
user space tests (based on the kselftest).
Finally, the last patch (13th) updates the MAINTAINERS file.
This patch (of 13):
DAMON is a data access monitoring framework for the Linux kernel. The
core mechanisms of DAMON make it
- accurate (the monitoring output is useful enough for DRAM level
performance-centric memory management; It might be inappropriate for
CPU cache levels, though),
- light-weight (the monitoring overhead is normally low enough to be
applied online), and
- scalable (the upper-bound of the overhead is in constant range
regardless of the size of target workloads).
Using this framework, hence, we can easily write efficient kernel space
data access monitoring applications. For example, the kernel's memory
management mechanisms can make advanced decisions using this.
Experimental data access aware optimization works that incurring high
access monitoring overhead could again be implemented on top of this.
Due to its simple and flexible interface, providing user space interface
would be also easy. Then, user space users who have some special
workloads can write personalized applications for better understanding and
optimizations of their workloads and systems.
===
Nevertheless, this commit is defining and implementing only basic access
check part without the overhead-accuracy handling core logic. The basic
access check is as below.
The output of DAMON says what memory regions are how frequently accessed
for a given duration. The resolution of the access frequency is
controlled by setting ``sampling interval`` and ``aggregation interval``.
In detail, DAMON checks access to each page per ``sampling interval`` and
aggregates the results. In other words, counts the number of the accesses
to each region. After each ``aggregation interval`` passes, DAMON calls
callback functions that previously registered by users so that users can
read the aggregated results and then clears the results. This can be
described in below simple pseudo-code::
init()
while monitoring_on:
for page in monitoring_target:
if accessed(page):
nr_accesses[page] += 1
if time() % aggregation_interval == 0:
for callback in user_registered_callbacks:
callback(monitoring_target, nr_accesses)
for page in monitoring_target:
nr_accesses[page] = 0
if time() % update_interval == 0:
update()
sleep(sampling interval)
The target regions constructed at the beginning of the monitoring and
updated after each ``regions_update_interval``, because the target regions
could be dynamically changed (e.g., mmap() or memory hotplug). The
monitoring overhead of this mechanism will arbitrarily increase as the
size of the target workload grows.
The basic monitoring primitives for actual access check and dynamic target
regions construction aren't in the core part of DAMON. Instead, it allows
users to implement their own primitives that are optimized for their use
case and configure DAMON to use those. In other words, users cannot use
current version of DAMON without some additional works.
Following commits will implement the core mechanisms for the
overhead-accuracy control and default primitives implementations.
Link: https://lkml.kernel.org/r/20210716081449.22187-1-sj38.park@gmail.com
Link: https://lkml.kernel.org/r/20210716081449.22187-2-sj38.park@gmail.com
Signed-off-by: SeongJae Park <sjpark@amazon.de>
Reviewed-by: Leonard Foerster <foersleo@amazon.de>
Reviewed-by: Fernand Sieber <sieberf@amazon.com>
Acked-by: Shakeel Butt <shakeelb@google.com>
Cc: Jonathan Cameron <Jonathan.Cameron@huawei.com>
Cc: Alexander Shishkin <alexander.shishkin@linux.intel.com>
Cc: Amit Shah <amit@kernel.org>
Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org>
Cc: Jonathan Corbet <corbet@lwn.net>
Cc: David Hildenbrand <david@redhat.com>
Cc: David Woodhouse <dwmw@amazon.com>
Cc: Marco Elver <elver@google.com>
Cc: Fan Du <fan.du@intel.com>
Cc: Greg Kroah-Hartman <greg@kroah.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: Joe Perches <joe@perches.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Maximilian Heyne <mheyne@amazon.de>
Cc: Minchan Kim <minchan@kernel.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Namhyung Kim <namhyung@kernel.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Rik van Riel <riel@surriel.com>
Cc: David Rientjes <rientjes@google.com>
Cc: Steven Rostedt (VMware) <rostedt@goodmis.org>
Cc: Shuah Khan <shuah@kernel.org>
Cc: Vlastimil Babka <vbabka@suse.cz>
Cc: Vladimir Davydov <vdavydov.dev@gmail.com>
Cc: Brendan Higgins <brendanhiggins@google.com>
Cc: Markus Boehme <markubo@amazon.de>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2021-09-07 19:56:28 -07:00
|
|
|
|
|
|
|
|
|
2018-07-31 13:39:35 +02:00
|
|
|
|
