linux/include/linux/sched.h

1900 lines
52 KiB

License cleanup: add SPDX GPL-2.0 license identifier to files with no license Many source files in the tree are missing licensing information, which makes it harder for compliance tools to determine the correct license. By default all files without license information are under the default license of the kernel, which is GPL version 2. Update the files which contain no license information with the 'GPL-2.0' SPDX license identifier. The SPDX identifier is a legally binding shorthand, which can be used instead of the full boiler plate text. This patch is based on work done by Thomas Gleixner and Kate Stewart and Philippe Ombredanne. How this work was done: Patches were generated and checked against linux-4.14-rc6 for a subset of the use cases: - file had no licensing information it it. - file was a */uapi/* one with no licensing information in it, - file was a */uapi/* one with existing licensing information, Further patches will be generated in subsequent months to fix up cases where non-standard license headers were used, and references to license had to be inferred by heuristics based on keywords. The analysis to determine which SPDX License Identifier to be applied to a file was done in a spreadsheet of side by side results from of the output of two independent scanners (ScanCode & Windriver) producing SPDX tag:value files created by Philippe Ombredanne. Philippe prepared the base worksheet, and did an initial spot review of a few 1000 files. The 4.13 kernel was the starting point of the analysis with 60,537 files assessed. Kate Stewart did a file by file comparison of the scanner results in the spreadsheet to determine which SPDX license identifier(s) to be applied to the file. She confirmed any determination that was not immediately clear with lawyers working with the Linux Foundation. Criteria used to select files for SPDX license identifier tagging was: - Files considered eligible had to be source code files. - Make and config files were included as candidates if they contained >5 lines of source - File already had some variant of a license header in it (even if <5 lines). All documentation files were explicitly excluded. The following heuristics were used to determine which SPDX license identifiers to apply. - when both scanners couldn't find any license traces, file was considered to have no license information in it, and the top level COPYING file license applied. For non */uapi/* files that summary was: SPDX license identifier # files ---------------------------------------------------|------- GPL-2.0 11139 and resulted in the first patch in this series. If that file was a */uapi/* path one, it was "GPL-2.0 WITH Linux-syscall-note" otherwise it was "GPL-2.0". Results of that was: SPDX license identifier # files ---------------------------------------------------|------- GPL-2.0 WITH Linux-syscall-note 930 and resulted in the second patch in this series. - if a file had some form of licensing information in it, and was one of the */uapi/* ones, it was denoted with the Linux-syscall-note if any GPL family license was found in the file or had no licensing in it (per prior point). Results summary: SPDX license identifier # files ---------------------------------------------------|------ GPL-2.0 WITH Linux-syscall-note 270 GPL-2.0+ WITH Linux-syscall-note 169 ((GPL-2.0 WITH Linux-syscall-note) OR BSD-2-Clause) 21 ((GPL-2.0 WITH Linux-syscall-note) OR BSD-3-Clause) 17 LGPL-2.1+ WITH Linux-syscall-note 15 GPL-1.0+ WITH Linux-syscall-note 14 ((GPL-2.0+ WITH Linux-syscall-note) OR BSD-3-Clause) 5 LGPL-2.0+ WITH Linux-syscall-note 4 LGPL-2.1 WITH Linux-syscall-note 3 ((GPL-2.0 WITH Linux-syscall-note) OR MIT) 3 ((GPL-2.0 WITH Linux-syscall-note) AND MIT) 1 and that resulted in the third patch in this series. - when the two scanners agreed on the detected license(s), that became the concluded license(s). - when there was disagreement between the two scanners (one detected a license but the other didn't, or they both detected different licenses) a manual inspection of the file occurred. - In most cases a manual inspection of the information in the file resulted in a clear resolution of the license that should apply (and which scanner probably needed to revisit its heuristics). - When it was not immediately clear, the license identifier was confirmed with lawyers working with the Linux Foundation. - If there was any question as to the appropriate license identifier, the file was flagged for further research and to be revisited later in time. In total, over 70 hours of logged manual review was done on the spreadsheet to determine the SPDX license identifiers to apply to the source files by Kate, Philippe, Thomas and, in some cases, confirmation by lawyers working with the Linux Foundation. Kate also obtained a third independent scan of the 4.13 code base from FOSSology, and compared selected files where the other two scanners disagreed against that SPDX file, to see if there was new insights. The Windriver scanner is based on an older version of FOSSology in part, so they are related. Thomas did random spot checks in about 500 files from the spreadsheets for the uapi headers and agreed with SPDX license identifier in the files he inspected. For the non-uapi files Thomas did random spot checks in about 15000 files. In initial set of patches against 4.14-rc6, 3 files were found to have copy/paste license identifier errors, and have been fixed to reflect the correct identifier. Additionally Philippe spent 10 hours this week doing a detailed manual inspection and review of the 12,461 patched files from the initial patch version early this week with: - a full scancode scan run, collecting the matched texts, detected license ids and scores - reviewing anything where there was a license detected (about 500+ files) to ensure that the applied SPDX license was correct - reviewing anything where there was no detection but the patch license was not GPL-2.0 WITH Linux-syscall-note to ensure that the applied SPDX license was correct This produced a worksheet with 20 files needing minor correction. This worksheet was then exported into 3 different .csv files for the different types of files to be modified. These .csv files were then reviewed by Greg. Thomas wrote a script to parse the csv files and add the proper SPDX tag to the file, in the format that the file expected. This script was further refined by Greg based on the output to detect more types of files automatically and to distinguish between header and source .c files (which need different comment types.) Finally Greg ran the script using the .csv files to generate the patches. Reviewed-by: Kate Stewart <kstewart@linuxfoundation.org> Reviewed-by: Philippe Ombredanne <pombredanne@nexb.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2017-11-01 15:07:57 +01:00
rseq: Introduce restartable sequences system call Expose a new system call allowing each thread to register one userspace memory area to be used as an ABI between kernel and user-space for two purposes: user-space restartable sequences and quick access to read the current CPU number value from user-space. * Restartable sequences (per-cpu atomics) Restartables sequences allow user-space to perform update operations on per-cpu data without requiring heavy-weight atomic operations. The restartable critical sections (percpu atomics) work has been started by Paul Turner and Andrew Hunter. It lets the kernel handle restart of critical sections. [1] [2] The re-implementation proposed here brings a few simplifications to the ABI which facilitates porting to other architectures and speeds up the user-space fast path. Here are benchmarks of various rseq use-cases. Test hardware: arm32: ARMv7 Processor rev 4 (v7l) "Cubietruck", 2-core x86-64: Intel E5-2630 v3@2.40GHz, 16-core, hyperthreading The following benchmarks were all performed on a single thread. * Per-CPU statistic counter increment getcpu+atomic (ns/op) rseq (ns/op) speedup arm32: 344.0 31.4 11.0 x86-64: 15.3 2.0 7.7 * LTTng-UST: write event 32-bit header, 32-bit payload into tracer per-cpu buffer getcpu+atomic (ns/op) rseq (ns/op) speedup arm32: 2502.0 2250.0 1.1 x86-64: 117.4 98.0 1.2 * liburcu percpu: lock-unlock pair, dereference, read/compare word getcpu+atomic (ns/op) rseq (ns/op) speedup arm32: 751.0 128.5 5.8 x86-64: 53.4 28.6 1.9 * jemalloc memory allocator adapted to use rseq Using rseq with per-cpu memory pools in jemalloc at Facebook (based on rseq 2016 implementation): The production workload response-time has 1-2% gain avg. latency, and the P99 overall latency drops by 2-3%. * Reading the current CPU number Speeding up reading the current CPU number on which the caller thread is running is done by keeping the current CPU number up do date within the cpu_id field of the memory area registered by the thread. This is done by making scheduler preemption set the TIF_NOTIFY_RESUME flag on the current thread. Upon return to user-space, a notify-resume handler updates the current CPU value within the registered user-space memory area. User-space can then read the current CPU number directly from memory. Keeping the current cpu id in a memory area shared between kernel and user-space is an improvement over current mechanisms available to read the current CPU number, which has the following benefits over alternative approaches: - 35x speedup on ARM vs system call through glibc - 20x speedup on x86 compared to calling glibc, which calls vdso executing a "lsl" instruction, - 14x speedup on x86 compared to inlined "lsl" instruction, - Unlike vdso approaches, this cpu_id value can be read from an inline assembly, which makes it a useful building block for restartable sequences. - The approach of reading the cpu id through memory mapping shared between kernel and user-space is portable (e.g. ARM), which is not the case for the lsl-based x86 vdso. On x86, yet another possible approach would be to use the gs segment selector to point to user-space per-cpu data. This approach performs similarly to the cpu id cache, but it has two disadvantages: it is not portable, and it is incompatible with existing applications already using the gs segment selector for other purposes. Benchmarking various approaches for reading the current CPU number: ARMv7 Processor rev 4 (v7l) Machine model: Cubietruck - Baseline (empty loop): 8.4 ns - Read CPU from rseq cpu_id: 16.7 ns - Read CPU from rseq cpu_id (lazy register): 19.8 ns - glibc 2.19-0ubuntu6.6 getcpu: 301.8 ns - getcpu system call: 234.9 ns x86-64 Intel(R) Xeon(R) CPU E5-2630 v3 @ 2.40GHz: - Baseline (empty loop): 0.8 ns - Read CPU from rseq cpu_id: 0.8 ns - Read CPU from rseq cpu_id (lazy register): 0.8 ns - Read using gs segment selector: 0.8 ns - "lsl" inline assembly: 13.0 ns - glibc 2.19-0ubuntu6 getcpu: 16.6 ns - getcpu system call: 53.9 ns - Speed (benchmark taken on v8 of patchset) Running 10 runs of hackbench -l 100000 seems to indicate, contrary to expectations, that enabling CONFIG_RSEQ slightly accelerates the scheduler: Configuration: 2 sockets * 8-core Intel(R) Xeon(R) CPU E5-2630 v3 @ 2.40GHz (directly on hardware, hyperthreading disabled in BIOS, energy saving disabled in BIOS, turboboost disabled in BIOS, cpuidle.off=1 kernel parameter), with a Linux v4.6 defconfig+localyesconfig, restartable sequences series applied. * CONFIG_RSEQ=n avg.: 41.37 s std.dev.: 0.36 s * CONFIG_RSEQ=y avg.: 40.46 s std.dev.: 0.33 s - Size On x86-64, between CONFIG_RSEQ=n/y, the text size increase of vmlinux is 567 bytes, and the data size increase of vmlinux is 5696 bytes. [1] https://lwn.net/Articles/650333/ [2] http://www.linuxplumbersconf.org/2013/ocw/system/presentations/1695/original/LPC%20-%20PerCpu%20Atomics.pdf Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Joel Fernandes <joelaf@google.com> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Dave Watson <davejwatson@fb.com> Cc: Will Deacon <will.deacon@arm.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: "H . Peter Anvin" <hpa@zytor.com> Cc: Chris Lameter <cl@linux.com> Cc: Russell King <linux@arm.linux.org.uk> Cc: Andrew Hunter <ahh@google.com> Cc: Michael Kerrisk <mtk.manpages@gmail.com> Cc: "Paul E . McKenney" <paulmck@linux.vnet.ibm.com> Cc: Paul Turner <pjt@google.com> Cc: Boqun Feng <boqun.feng@gmail.com> Cc: Josh Triplett <josh@joshtriplett.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ben Maurer <bmaurer@fb.com> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: linux-api@vger.kernel.org Cc: Andy Lutomirski <luto@amacapital.net> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/20151027235635.16059.11630.stgit@pjt-glaptop.roam.corp.google.com Link: http://lkml.kernel.org/r/20150624222609.6116.86035.stgit@kitami.mtv.corp.google.com Link: https://lkml.kernel.org/r/20180602124408.8430-3-mathieu.desnoyers@efficios.com
2018-06-02 08:43:54 -04:00
sched: Prevent recursion in io_schedule() io_schedule() calls blk_flush_plug() which, depending on the contents of current->plug, can initiate arbitrary blk-io requests. Note that this contrasts with blk_schedule_flush_plug() which requires all non-trivial work to be handed off to a separate thread. This makes it possible for io_schedule() to recurse, and initiating block requests could possibly call mempool_alloc() which, in times of memory pressure, uses io_schedule(). Apart from any stack usage issues, io_schedule() will not behave correctly when called recursively as delayacct_blkio_start() does not allow for repeated calls. So: - use ->in_iowait to detect recursion. Set it earlier, and restore it to the old value. - move the call to "raw_rq" after the call to blk_flush_plug(). As this is some sort of per-cpu thing, we want some chance that we are on the right CPU - When io_schedule() is called recurively, use blk_schedule_flush_plug() which cannot further recurse. - as this makes io_schedule() a lot more complex and as io_schedule() must match io_schedule_timeout(), but all the changes in io_schedule_timeout() and make io_schedule a simple wrapper for that. Signed-off-by: NeilBrown <neilb@suse.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> [ Moved the now rudimentary io_schedule() into sched.h. ] Cc: Jens Axboe <axboe@kernel.dk> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Tony Battersby <tonyb@cybernetics.com> Link: http://lkml.kernel.org/r/20150213162600.059fffb2@notabene.brown Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-02-13 15:49:17 +11:00
sched: Prevent recursion in io_schedule() io_schedule() calls blk_flush_plug() which, depending on the contents of current->plug, can initiate arbitrary blk-io requests. Note that this contrasts with blk_schedule_flush_plug() which requires all non-trivial work to be handed off to a separate thread. This makes it possible for io_schedule() to recurse, and initiating block requests could possibly call mempool_alloc() which, in times of memory pressure, uses io_schedule(). Apart from any stack usage issues, io_schedule() will not behave correctly when called recursively as delayacct_blkio_start() does not allow for repeated calls. So: - use ->in_iowait to detect recursion. Set it earlier, and restore it to the old value. - move the call to "raw_rq" after the call to blk_flush_plug(). As this is some sort of per-cpu thing, we want some chance that we are on the right CPU - When io_schedule() is called recurively, use blk_schedule_flush_plug() which cannot further recurse. - as this makes io_schedule() a lot more complex and as io_schedule() must match io_schedule_timeout(), but all the changes in io_schedule_timeout() and make io_schedule a simple wrapper for that. Signed-off-by: NeilBrown <neilb@suse.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> [ Moved the now rudimentary io_schedule() into sched.h. ] Cc: Jens Axboe <axboe@kernel.dk> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Tony Battersby <tonyb@cybernetics.com> Link: http://lkml.kernel.org/r/20150213162600.059fffb2@notabene.brown Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-02-13 15:49:17 +11:00
timers: fix itimer/many thread hang Overview This patch reworks the handling of POSIX CPU timers, including the ITIMER_PROF, ITIMER_VIRT timers and rlimit handling. It was put together with the help of Roland McGrath, the owner and original writer of this code. The problem we ran into, and the reason for this rework, has to do with using a profiling timer in a process with a large number of threads. It appears that the performance of the old implementation of run_posix_cpu_timers() was at least O(n*3) (where "n" is the number of threads in a process) or worse. Everything is fine with an increasing number of threads until the time taken for that routine to run becomes the same as or greater than the tick time, at which point things degrade rather quickly. This patch fixes bug 9906, "Weird hang with NPTL and SIGPROF." Code Changes This rework corrects the implementation of run_posix_cpu_timers() to make it run in constant time for a particular machine. (Performance may vary between one machine and another depending upon whether the kernel is built as single- or multiprocessor and, in the latter case, depending upon the number of running processors.) To do this, at each tick we now update fields in signal_struct as well as task_struct. The run_posix_cpu_timers() function uses those fields to make its decisions. We define a new structure, "task_cputime," to contain user, system and scheduler times and use these in appropriate places: struct task_cputime { cputime_t utime; cputime_t stime; unsigned long long sum_exec_runtime; }; This is included in the structure "thread_group_cputime," which is a new substructure of signal_struct and which varies for uniprocessor versus multiprocessor kernels. For uniprocessor kernels, it uses "task_cputime" as a simple substructure, while for multiprocessor kernels it is a pointer: struct thread_group_cputime { struct task_cputime totals; }; struct thread_group_cputime { struct task_cputime *totals; }; We also add a new task_cputime substructure directly to signal_struct, to cache the earliest expiration of process-wide timers, and task_cputime also replaces the it_*_expires fields of task_struct (used for earliest expiration of thread timers). The "thread_group_cputime" structure contains process-wide timers that are updated via account_user_time() and friends. In the non-SMP case the structure is a simple aggregator; unfortunately in the SMP case that simplicity was not achievable due to cache-line contention between CPUs (in one measured case performance was actually _worse_ on a 16-cpu system than the same test on a 4-cpu system, due to this contention). For SMP, the thread_group_cputime counters are maintained as a per-cpu structure allocated using alloc_percpu(). The timer functions update only the timer field in the structure corresponding to the running CPU, obtained using per_cpu_ptr(). We define a set of inline functions in sched.h that we use to maintain the thread_group_cputime structure and hide the differences between UP and SMP implementations from the rest of the kernel. The thread_group_cputime_init() function initializes the thread_group_cputime structure for the given task. The thread_group_cputime_alloc() is a no-op for UP; for SMP it calls the out-of-line function thread_group_cputime_alloc_smp() to allocate and fill in the per-cpu structures and fields. The thread_group_cputime_free() function, also a no-op for UP, in SMP frees the per-cpu structures. The thread_group_cputime_clone_thread() function (also a UP no-op) for SMP calls thread_group_cputime_alloc() if the per-cpu structures haven't yet been allocated. The thread_group_cputime() function fills the task_cputime structure it is passed with the contents of the thread_group_cputime fields; in UP it's that simple but in SMP it must also safely check that tsk->signal is non-NULL (if it is it just uses the appropriate fields of task_struct) and, if so, sums the per-cpu values for each online CPU. Finally, the three functions account_group_user_time(), account_group_system_time() and account_group_exec_runtime() are used by timer functions to update the respective fields of the thread_group_cputime structure. Non-SMP operation is trivial and will not be mentioned further. The per-cpu structure is always allocated when a task creates its first new thread, via a call to thread_group_cputime_clone_thread() from copy_signal(). It is freed at process exit via a call to thread_group_cputime_free() from cleanup_signal(). All functions that formerly summed utime/stime/sum_sched_runtime values from from all threads in the thread group now use thread_group_cputime() to snapshot the values in the thread_group_cputime structure or the values in the task structure itself if the per-cpu structure hasn't been allocated. Finally, the code in kernel/posix-cpu-timers.c has changed quite a bit. The run_posix_cpu_timers() function has been split into a fast path and a slow path; the former safely checks whether there are any expired thread timers and, if not, just returns, while the slow path does the heavy lifting. With the dedicated thread group fields, timers are no longer "rebalanced" and the process_timer_rebalance() function and related code has gone away. All summing loops are gone and all code that used them now uses the thread_group_cputime() inline. When process-wide timers are set, the new task_cputime structure in signal_struct is used to cache the earliest expiration; this is checked in the fast path. Performance The fix appears not to add significant overhead to existing operations. It generally performs the same as the current code except in two cases, one in which it performs slightly worse (Case 5 below) and one in which it performs very significantly better (Case 2 below). Overall it's a wash except in those two cases. I've since done somewhat more involved testing on a dual-core Opteron system. Case 1: With no itimer running, for a test with 100,000 threads, the fixed kernel took 1428.5 seconds, 513 seconds more than the unfixed system, all of which was spent in the system. There were twice as many voluntary context switches with the fix as without it. Case 2: With an itimer running at .01 second ticks and 4000 threads (the most an unmodified kernel can handle), the fixed kernel ran the test in eight percent of the time (5.8 seconds as opposed to 70 seconds) and had better tick accuracy (.012 seconds per tick as opposed to .023 seconds per tick). Case 3: A 4000-thread test with an initial timer tick of .01 second and an interval of 10,000 seconds (i.e. a timer that ticks only once) had very nearly the same performance in both cases: 6.3 seconds elapsed for the fixed kernel versus 5.5 seconds for the unfixed kernel. With fewer threads (eight in these tests), the Case 1 test ran in essentially the same time on both the modified and unmodified kernels (5.2 seconds versus 5.8 seconds). The Case 2 test ran in about the same time as well, 5.9 seconds versus 5.4 seconds but again with much better tick accuracy, .013 seconds per tick versus .025 seconds per tick for the unmodified kernel. Since the fix affected the rlimit code, I also tested soft and hard CPU limits. Case 4: With a hard CPU limit of 20 seconds and eight threads (and an itimer running), the modified kernel was very slightly favored in that while it killed the process in 19.997 seconds of CPU time (5.002 seconds of wall time), only .003 seconds of that was system time, the rest was user time. The unmodified kernel killed the process in 20.001 seconds of CPU (5.014 seconds of wall time) of which .016 seconds was system time. Really, though, the results were too close to call. The results were essentially the same with no itimer running. Case 5: With a soft limit of 20 seconds and a hard limit of 2000 seconds (where the hard limit would never be reached) and an itimer running, the modified kernel exhibited worse tick accuracy than the unmodified kernel: .050 seconds/tick versus .028 seconds/tick. Otherwise, performance was almost indistinguishable. With no itimer running this test exhibited virtually identical behavior and times in both cases. In times past I did some limited performance testing. those results are below. On a four-cpu Opteron system without this fix, a sixteen-thread test executed in 3569.991 seconds, of which user was 3568.435s and system was 1.556s. On the same system with the fix, user and elapsed time were about the same, but system time dropped to 0.007 seconds. Performance with eight, four and one thread were comparable. Interestingly, the timer ticks with the fix seemed more accurate: The sixteen-thread test with the fix received 149543 ticks for 0.024 seconds per tick, while the same test without the fix received 58720 for 0.061 seconds per tick. Both cases were configured for an interval of 0.01 seconds. Again, the other tests were comparable. Each thread in this test computed the primes up to 25,000,000. I also did a test with a large number of threads, 100,000 threads, which is impossible without the fix. In this case each thread computed the primes only up to 10,000 (to make the runtime manageable). System time dominated, at 1546.968 seconds out of a total 2176.906 seconds (giving a user time of 629.938s). It received 147651 ticks for 0.015 seconds per tick, still quite accurate. There is obviously no comparable test without the fix. Signed-off-by: Frank Mayhar <fmayhar@google.com> Cc: Roland McGrath <roland@redhat.com> Cc: Alexey Dobriyan <adobriyan@gmail.com> Cc: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-09-12 09:54:39 -07:00
timers: fix itimer/many thread hang Overview This patch reworks the handling of POSIX CPU timers, including the ITIMER_PROF, ITIMER_VIRT timers and rlimit handling. It was put together with the help of Roland McGrath, the owner and original writer of this code. The problem we ran into, and the reason for this rework, has to do with using a profiling timer in a process with a large number of threads. It appears that the performance of the old implementation of run_posix_cpu_timers() was at least O(n*3) (where "n" is the number of threads in a process) or worse. Everything is fine with an increasing number of threads until the time taken for that routine to run becomes the same as or greater than the tick time, at which point things degrade rather quickly. This patch fixes bug 9906, "Weird hang with NPTL and SIGPROF." Code Changes This rework corrects the implementation of run_posix_cpu_timers() to make it run in constant time for a particular machine. (Performance may vary between one machine and another depending upon whether the kernel is built as single- or multiprocessor and, in the latter case, depending upon the number of running processors.) To do this, at each tick we now update fields in signal_struct as well as task_struct. The run_posix_cpu_timers() function uses those fields to make its decisions. We define a new structure, "task_cputime," to contain user, system and scheduler times and use these in appropriate places: struct task_cputime { cputime_t utime; cputime_t stime; unsigned long long sum_exec_runtime; }; This is included in the structure "thread_group_cputime," which is a new substructure of signal_struct and which varies for uniprocessor versus multiprocessor kernels. For uniprocessor kernels, it uses "task_cputime" as a simple substructure, while for multiprocessor kernels it is a pointer: struct thread_group_cputime { struct task_cputime totals; }; struct thread_group_cputime { struct task_cputime *totals; }; We also add a new task_cputime substructure directly to signal_struct, to cache the earliest expiration of process-wide timers, and task_cputime also replaces the it_*_expires fields of task_struct (used for earliest expiration of thread timers). The "thread_group_cputime" structure contains process-wide timers that are updated via account_user_time() and friends. In the non-SMP case the structure is a simple aggregator; unfortunately in the SMP case that simplicity was not achievable due to cache-line contention between CPUs (in one measured case performance was actually _worse_ on a 16-cpu system than the same test on a 4-cpu system, due to this contention). For SMP, the thread_group_cputime counters are maintained as a per-cpu structure allocated using alloc_percpu(). The timer functions update only the timer field in the structure corresponding to the running CPU, obtained using per_cpu_ptr(). We define a set of inline functions in sched.h that we use to maintain the thread_group_cputime structure and hide the differences between UP and SMP implementations from the rest of the kernel. The thread_group_cputime_init() function initializes the thread_group_cputime structure for the given task. The thread_group_cputime_alloc() is a no-op for UP; for SMP it calls the out-of-line function thread_group_cputime_alloc_smp() to allocate and fill in the per-cpu structures and fields. The thread_group_cputime_free() function, also a no-op for UP, in SMP frees the per-cpu structures. The thread_group_cputime_clone_thread() function (also a UP no-op) for SMP calls thread_group_cputime_alloc() if the per-cpu structures haven't yet been allocated. The thread_group_cputime() function fills the task_cputime structure it is passed with the contents of the thread_group_cputime fields; in UP it's that simple but in SMP it must also safely check that tsk->signal is non-NULL (if it is it just uses the appropriate fields of task_struct) and, if so, sums the per-cpu values for each online CPU. Finally, the three functions account_group_user_time(), account_group_system_time() and account_group_exec_runtime() are used by timer functions to update the respective fields of the thread_group_cputime structure. Non-SMP operation is trivial and will not be mentioned further. The per-cpu structure is always allocated when a task creates its first new thread, via a call to thread_group_cputime_clone_thread() from copy_signal(). It is freed at process exit via a call to thread_group_cputime_free() from cleanup_signal(). All functions that formerly summed utime/stime/sum_sched_runtime values from from all threads in the thread group now use thread_group_cputime() to snapshot the values in the thread_group_cputime structure or the values in the task structure itself if the per-cpu structure hasn't been allocated. Finally, the code in kernel/posix-cpu-timers.c has changed quite a bit. The run_posix_cpu_timers() function has been split into a fast path and a slow path; the former safely checks whether there are any expired thread timers and, if not, just returns, while the slow path does the heavy lifting. With the dedicated thread group fields, timers are no longer "rebalanced" and the process_timer_rebalance() function and related code has gone away. All summing loops are gone and all code that used them now uses the thread_group_cputime() inline. When process-wide timers are set, the new task_cputime structure in signal_struct is used to cache the earliest expiration; this is checked in the fast path. Performance The fix appears not to add significant overhead to existing operations. It generally performs the same as the current code except in two cases, one in which it performs slightly worse (Case 5 below) and one in which it performs very significantly better (Case 2 below). Overall it's a wash except in those two cases. I've since done somewhat more involved testing on a dual-core Opteron system. Case 1: With no itimer running, for a test with 100,000 threads, the fixed kernel took 1428.5 seconds, 513 seconds more than the unfixed system, all of which was spent in the system. There were twice as many voluntary context switches with the fix as without it. Case 2: With an itimer running at .01 second ticks and 4000 threads (the most an unmodified kernel can handle), the fixed kernel ran the test in eight percent of the time (5.8 seconds as opposed to 70 seconds) and had better tick accuracy (.012 seconds per tick as opposed to .023 seconds per tick). Case 3: A 4000-thread test with an initial timer tick of .01 second and an interval of 10,000 seconds (i.e. a timer that ticks only once) had very nearly the same performance in both cases: 6.3 seconds elapsed for the fixed kernel versus 5.5 seconds for the unfixed kernel. With fewer threads (eight in these tests), the Case 1 test ran in essentially the same time on both the modified and unmodified kernels (5.2 seconds versus 5.8 seconds). The Case 2 test ran in about the same time as well, 5.9 seconds versus 5.4 seconds but again with much better tick accuracy, .013 seconds per tick versus .025 seconds per tick for the unmodified kernel. Since the fix affected the rlimit code, I also tested soft and hard CPU limits. Case 4: With a hard CPU limit of 20 seconds and eight threads (and an itimer running), the modified kernel was very slightly favored in that while it killed the process in 19.997 seconds of CPU time (5.002 seconds of wall time), only .003 seconds of that was system time, the rest was user time. The unmodified kernel killed the process in 20.001 seconds of CPU (5.014 seconds of wall time) of which .016 seconds was system time. Really, though, the results were too close to call. The results were essentially the same with no itimer running. Case 5: With a soft limit of 20 seconds and a hard limit of 2000 seconds (where the hard limit would never be reached) and an itimer running, the modified kernel exhibited worse tick accuracy than the unmodified kernel: .050 seconds/tick versus .028 seconds/tick. Otherwise, performance was almost indistinguishable. With no itimer running this test exhibited virtually identical behavior and times in both cases. In times past I did some limited performance testing. those results are below. On a four-cpu Opteron system without this fix, a sixteen-thread test executed in 3569.991 seconds, of which user was 3568.435s and system was 1.556s. On the same system with the fix, user and elapsed time were about the same, but system time dropped to 0.007 seconds. Performance with eight, four and one thread were comparable. Interestingly, the timer ticks with the fix seemed more accurate: The sixteen-thread test with the fix received 149543 ticks for 0.024 seconds per tick, while the same test without the fix received 58720 for 0.061 seconds per tick. Both cases were configured for an interval of 0.01 seconds. Again, the other tests were comparable. Each thread in this test computed the primes up to 25,000,000. I also did a test with a large number of threads, 100,000 threads, which is impossible without the fix. In this case each thread computed the primes only up to 10,000 (to make the runtime manageable). System time dominated, at 1546.968 seconds out of a total 2176.906 seconds (giving a user time of 629.938s). It received 147651 ticks for 0.015 seconds per tick, still quite accurate. There is obviously no comparable test without the fix. Signed-off-by: Frank Mayhar <fmayhar@google.com> Cc: Roland McGrath <roland@redhat.com> Cc: Alexey Dobriyan <adobriyan@gmail.com> Cc: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-09-12 09:54:39 -07:00
timers: fix itimer/many thread hang Overview This patch reworks the handling of POSIX CPU timers, including the ITIMER_PROF, ITIMER_VIRT timers and rlimit handling. It was put together with the help of Roland McGrath, the owner and original writer of this code. The problem we ran into, and the reason for this rework, has to do with using a profiling timer in a process with a large number of threads. It appears that the performance of the old implementation of run_posix_cpu_timers() was at least O(n*3) (where "n" is the number of threads in a process) or worse. Everything is fine with an increasing number of threads until the time taken for that routine to run becomes the same as or greater than the tick time, at which point things degrade rather quickly. This patch fixes bug 9906, "Weird hang with NPTL and SIGPROF." Code Changes This rework corrects the implementation of run_posix_cpu_timers() to make it run in constant time for a particular machine. (Performance may vary between one machine and another depending upon whether the kernel is built as single- or multiprocessor and, in the latter case, depending upon the number of running processors.) To do this, at each tick we now update fields in signal_struct as well as task_struct. The run_posix_cpu_timers() function uses those fields to make its decisions. We define a new structure, "task_cputime," to contain user, system and scheduler times and use these in appropriate places: struct task_cputime { cputime_t utime; cputime_t stime; unsigned long long sum_exec_runtime; }; This is included in the structure "thread_group_cputime," which is a new substructure of signal_struct and which varies for uniprocessor versus multiprocessor kernels. For uniprocessor kernels, it uses "task_cputime" as a simple substructure, while for multiprocessor kernels it is a pointer: struct thread_group_cputime { struct task_cputime totals; }; struct thread_group_cputime { struct task_cputime *totals; }; We also add a new task_cputime substructure directly to signal_struct, to cache the earliest expiration of process-wide timers, and task_cputime also replaces the it_*_expires fields of task_struct (used for earliest expiration of thread timers). The "thread_group_cputime" structure contains process-wide timers that are updated via account_user_time() and friends. In the non-SMP case the structure is a simple aggregator; unfortunately in the SMP case that simplicity was not achievable due to cache-line contention between CPUs (in one measured case performance was actually _worse_ on a 16-cpu system than the same test on a 4-cpu system, due to this contention). For SMP, the thread_group_cputime counters are maintained as a per-cpu structure allocated using alloc_percpu(). The timer functions update only the timer field in the structure corresponding to the running CPU, obtained using per_cpu_ptr(). We define a set of inline functions in sched.h that we use to maintain the thread_group_cputime structure and hide the differences between UP and SMP implementations from the rest of the kernel. The thread_group_cputime_init() function initializes the thread_group_cputime structure for the given task. The thread_group_cputime_alloc() is a no-op for UP; for SMP it calls the out-of-line function thread_group_cputime_alloc_smp() to allocate and fill in the per-cpu structures and fields. The thread_group_cputime_free() function, also a no-op for UP, in SMP frees the per-cpu structures. The thread_group_cputime_clone_thread() function (also a UP no-op) for SMP calls thread_group_cputime_alloc() if the per-cpu structures haven't yet been allocated. The thread_group_cputime() function fills the task_cputime structure it is passed with the contents of the thread_group_cputime fields; in UP it's that simple but in SMP it must also safely check that tsk->signal is non-NULL (if it is it just uses the appropriate fields of task_struct) and, if so, sums the per-cpu values for each online CPU. Finally, the three functions account_group_user_time(), account_group_system_time() and account_group_exec_runtime() are used by timer functions to update the respective fields of the thread_group_cputime structure. Non-SMP operation is trivial and will not be mentioned further. The per-cpu structure is always allocated when a task creates its first new thread, via a call to thread_group_cputime_clone_thread() from copy_signal(). It is freed at process exit via a call to thread_group_cputime_free() from cleanup_signal(). All functions that formerly summed utime/stime/sum_sched_runtime values from from all threads in the thread group now use thread_group_cputime() to snapshot the values in the thread_group_cputime structure or the values in the task structure itself if the per-cpu structure hasn't been allocated. Finally, the code in kernel/posix-cpu-timers.c has changed quite a bit. The run_posix_cpu_timers() function has been split into a fast path and a slow path; the former safely checks whether there are any expired thread timers and, if not, just returns, while the slow path does the heavy lifting. With the dedicated thread group fields, timers are no longer "rebalanced" and the process_timer_rebalance() function and related code has gone away. All summing loops are gone and all code that used them now uses the thread_group_cputime() inline. When process-wide timers are set, the new task_cputime structure in signal_struct is used to cache the earliest expiration; this is checked in the fast path. Performance The fix appears not to add significant overhead to existing operations. It generally performs the same as the current code except in two cases, one in which it performs slightly worse (Case 5 below) and one in which it performs very significantly better (Case 2 below). Overall it's a wash except in those two cases. I've since done somewhat more involved testing on a dual-core Opteron system. Case 1: With no itimer running, for a test with 100,000 threads, the fixed kernel took 1428.5 seconds, 513 seconds more than the unfixed system, all of which was spent in the system. There were twice as many voluntary context switches with the fix as without it. Case 2: With an itimer running at .01 second ticks and 4000 threads (the most an unmodified kernel can handle), the fixed kernel ran the test in eight percent of the time (5.8 seconds as opposed to 70 seconds) and had better tick accuracy (.012 seconds per tick as opposed to .023 seconds per tick). Case 3: A 4000-thread test with an initial timer tick of .01 second and an interval of 10,000 seconds (i.e. a timer that ticks only once) had very nearly the same performance in both cases: 6.3 seconds elapsed for the fixed kernel versus 5.5 seconds for the unfixed kernel. With fewer threads (eight in these tests), the Case 1 test ran in essentially the same time on both the modified and unmodified kernels (5.2 seconds versus 5.8 seconds). The Case 2 test ran in about the same time as well, 5.9 seconds versus 5.4 seconds but again with much better tick accuracy, .013 seconds per tick versus .025 seconds per tick for the unmodified kernel. Since the fix affected the rlimit code, I also tested soft and hard CPU limits. Case 4: With a hard CPU limit of 20 seconds and eight threads (and an itimer running), the modified kernel was very slightly favored in that while it killed the process in 19.997 seconds of CPU time (5.002 seconds of wall time), only .003 seconds of that was system time, the rest was user time. The unmodified kernel killed the process in 20.001 seconds of CPU (5.014 seconds of wall time) of which .016 seconds was system time. Really, though, the results were too close to call. The results were essentially the same with no itimer running. Case 5: With a soft limit of 20 seconds and a hard limit of 2000 seconds (where the hard limit would never be reached) and an itimer running, the modified kernel exhibited worse tick accuracy than the unmodified kernel: .050 seconds/tick versus .028 seconds/tick. Otherwise, performance was almost indistinguishable. With no itimer running this test exhibited virtually identical behavior and times in both cases. In times past I did some limited performance testing. those results are below. On a four-cpu Opteron system without this fix, a sixteen-thread test executed in 3569.991 seconds, of which user was 3568.435s and system was 1.556s. On the same system with the fix, user and elapsed time were about the same, but system time dropped to 0.007 seconds. Performance with eight, four and one thread were comparable. Interestingly, the timer ticks with the fix seemed more accurate: The sixteen-thread test with the fix received 149543 ticks for 0.024 seconds per tick, while the same test without the fix received 58720 for 0.061 seconds per tick. Both cases were configured for an interval of 0.01 seconds. Again, the other tests were comparable. Each thread in this test computed the primes up to 25,000,000. I also did a test with a large number of threads, 100,000 threads, which is impossible without the fix. In this case each thread computed the primes only up to 10,000 (to make the runtime manageable). System time dominated, at 1546.968 seconds out of a total 2176.906 seconds (giving a user time of 629.938s). It received 147651 ticks for 0.015 seconds per tick, still quite accurate. There is obviously no comparable test without the fix. Signed-off-by: Frank Mayhar <fmayhar@google.com> Cc: Roland McGrath <roland@redhat.com> Cc: Alexey Dobriyan <adobriyan@gmail.com> Cc: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-09-12 09:54:39 -07:00
timers: fix itimer/many thread hang Overview This patch reworks the handling of POSIX CPU timers, including the ITIMER_PROF, ITIMER_VIRT timers and rlimit handling. It was put together with the help of Roland McGrath, the owner and original writer of this code. The problem we ran into, and the reason for this rework, has to do with using a profiling timer in a process with a large number of threads. It appears that the performance of the old implementation of run_posix_cpu_timers() was at least O(n*3) (where "n" is the number of threads in a process) or worse. Everything is fine with an increasing number of threads until the time taken for that routine to run becomes the same as or greater than the tick time, at which point things degrade rather quickly. This patch fixes bug 9906, "Weird hang with NPTL and SIGPROF." Code Changes This rework corrects the implementation of run_posix_cpu_timers() to make it run in constant time for a particular machine. (Performance may vary between one machine and another depending upon whether the kernel is built as single- or multiprocessor and, in the latter case, depending upon the number of running processors.) To do this, at each tick we now update fields in signal_struct as well as task_struct. The run_posix_cpu_timers() function uses those fields to make its decisions. We define a new structure, "task_cputime," to contain user, system and scheduler times and use these in appropriate places: struct task_cputime { cputime_t utime; cputime_t stime; unsigned long long sum_exec_runtime; }; This is included in the structure "thread_group_cputime," which is a new substructure of signal_struct and which varies for uniprocessor versus multiprocessor kernels. For uniprocessor kernels, it uses "task_cputime" as a simple substructure, while for multiprocessor kernels it is a pointer: struct thread_group_cputime { struct task_cputime totals; }; struct thread_group_cputime { struct task_cputime *totals; }; We also add a new task_cputime substructure directly to signal_struct, to cache the earliest expiration of process-wide timers, and task_cputime also replaces the it_*_expires fields of task_struct (used for earliest expiration of thread timers). The "thread_group_cputime" structure contains process-wide timers that are updated via account_user_time() and friends. In the non-SMP case the structure is a simple aggregator; unfortunately in the SMP case that simplicity was not achievable due to cache-line contention between CPUs (in one measured case performance was actually _worse_ on a 16-cpu system than the same test on a 4-cpu system, due to this contention). For SMP, the thread_group_cputime counters are maintained as a per-cpu structure allocated using alloc_percpu(). The timer functions update only the timer field in the structure corresponding to the running CPU, obtained using per_cpu_ptr(). We define a set of inline functions in sched.h that we use to maintain the thread_group_cputime structure and hide the differences between UP and SMP implementations from the rest of the kernel. The thread_group_cputime_init() function initializes the thread_group_cputime structure for the given task. The thread_group_cputime_alloc() is a no-op for UP; for SMP it calls the out-of-line function thread_group_cputime_alloc_smp() to allocate and fill in the per-cpu structures and fields. The thread_group_cputime_free() function, also a no-op for UP, in SMP frees the per-cpu structures. The thread_group_cputime_clone_thread() function (also a UP no-op) for SMP calls thread_group_cputime_alloc() if the per-cpu structures haven't yet been allocated. The thread_group_cputime() function fills the task_cputime structure it is passed with the contents of the thread_group_cputime fields; in UP it's that simple but in SMP it must also safely check that tsk->signal is non-NULL (if it is it just uses the appropriate fields of task_struct) and, if so, sums the per-cpu values for each online CPU. Finally, the three functions account_group_user_time(), account_group_system_time() and account_group_exec_runtime() are used by timer functions to update the respective fields of the thread_group_cputime structure. Non-SMP operation is trivial and will not be mentioned further. The per-cpu structure is always allocated when a task creates its first new thread, via a call to thread_group_cputime_clone_thread() from copy_signal(). It is freed at process exit via a call to thread_group_cputime_free() from cleanup_signal(). All functions that formerly summed utime/stime/sum_sched_runtime values from from all threads in the thread group now use thread_group_cputime() to snapshot the values in the thread_group_cputime structure or the values in the task structure itself if the per-cpu structure hasn't been allocated. Finally, the code in kernel/posix-cpu-timers.c has changed quite a bit. The run_posix_cpu_timers() function has been split into a fast path and a slow path; the former safely checks whether there are any expired thread timers and, if not, just returns, while the slow path does the heavy lifting. With the dedicated thread group fields, timers are no longer "rebalanced" and the process_timer_rebalance() function and related code has gone away. All summing loops are gone and all code that used them now uses the thread_group_cputime() inline. When process-wide timers are set, the new task_cputime structure in signal_struct is used to cache the earliest expiration; this is checked in the fast path. Performance The fix appears not to add significant overhead to existing operations. It generally performs the same as the current code except in two cases, one in which it performs slightly worse (Case 5 below) and one in which it performs very significantly better (Case 2 below). Overall it's a wash except in those two cases. I've since done somewhat more involved testing on a dual-core Opteron system. Case 1: With no itimer running, for a test with 100,000 threads, the fixed kernel took 1428.5 seconds, 513 seconds more than the unfixed system, all of which was spent in the system. There were twice as many voluntary context switches with the fix as without it. Case 2: With an itimer running at .01 second ticks and 4000 threads (the most an unmodified kernel can handle), the fixed kernel ran the test in eight percent of the time (5.8 seconds as opposed to 70 seconds) and had better tick accuracy (.012 seconds per tick as opposed to .023 seconds per tick). Case 3: A 4000-thread test with an initial timer tick of .01 second and an interval of 10,000 seconds (i.e. a timer that ticks only once) had very nearly the same performance in both cases: 6.3 seconds elapsed for the fixed kernel versus 5.5 seconds for the unfixed kernel. With fewer threads (eight in these tests), the Case 1 test ran in essentially the same time on both the modified and unmodified kernels (5.2 seconds versus 5.8 seconds). The Case 2 test ran in about the same time as well, 5.9 seconds versus 5.4 seconds but again with much better tick accuracy, .013 seconds per tick versus .025 seconds per tick for the unmodified kernel. Since the fix affected the rlimit code, I also tested soft and hard CPU limits. Case 4: With a hard CPU limit of 20 seconds and eight threads (and an itimer running), the modified kernel was very slightly favored in that while it killed the process in 19.997 seconds of CPU time (5.002 seconds of wall time), only .003 seconds of that was system time, the rest was user time. The unmodified kernel killed the process in 20.001 seconds of CPU (5.014 seconds of wall time) of which .016 seconds was system time. Really, though, the results were too close to call. The results were essentially the same with no itimer running. Case 5: With a soft limit of 20 seconds and a hard limit of 2000 seconds (where the hard limit would never be reached) and an itimer running, the modified kernel exhibited worse tick accuracy than the unmodified kernel: .050 seconds/tick versus .028 seconds/tick. Otherwise, performance was almost indistinguishable. With no itimer running this test exhibited virtually identical behavior and times in both cases. In times past I did some limited performance testing. those results are below. On a four-cpu Opteron system without this fix, a sixteen-thread test executed in 3569.991 seconds, of which user was 3568.435s and system was 1.556s. On the same system with the fix, user and elapsed time were about the same, but system time dropped to 0.007 seconds. Performance with eight, four and one thread were comparable. Interestingly, the timer ticks with the fix seemed more accurate: The sixteen-thread test with the fix received 149543 ticks for 0.024 seconds per tick, while the same test without the fix received 58720 for 0.061 seconds per tick. Both cases were configured for an interval of 0.01 seconds. Again, the other tests were comparable. Each thread in this test computed the primes up to 25,000,000. I also did a test with a large number of threads, 100,000 threads, which is impossible without the fix. In this case each thread computed the primes only up to 10,000 (to make the runtime manageable). System time dominated, at 1546.968 seconds out of a total 2176.906 seconds (giving a user time of 629.938s). It received 147651 ticks for 0.015 seconds per tick, still quite accurate. There is obviously no comparable test without the fix. Signed-off-by: Frank Mayhar <fmayhar@google.com> Cc: Roland McGrath <roland@redhat.com> Cc: Alexey Dobriyan <adobriyan@gmail.com> Cc: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-09-12 09:54:39 -07:00
timers: fix itimer/many thread hang Overview This patch reworks the handling of POSIX CPU timers, including the ITIMER_PROF, ITIMER_VIRT timers and rlimit handling. It was put together with the help of Roland McGrath, the owner and original writer of this code. The problem we ran into, and the reason for this rework, has to do with using a profiling timer in a process with a large number of threads. It appears that the performance of the old implementation of run_posix_cpu_timers() was at least O(n*3) (where "n" is the number of threads in a process) or worse. Everything is fine with an increasing number of threads until the time taken for that routine to run becomes the same as or greater than the tick time, at which point things degrade rather quickly. This patch fixes bug 9906, "Weird hang with NPTL and SIGPROF." Code Changes This rework corrects the implementation of run_posix_cpu_timers() to make it run in constant time for a particular machine. (Performance may vary between one machine and another depending upon whether the kernel is built as single- or multiprocessor and, in the latter case, depending upon the number of running processors.) To do this, at each tick we now update fields in signal_struct as well as task_struct. The run_posix_cpu_timers() function uses those fields to make its decisions. We define a new structure, "task_cputime," to contain user, system and scheduler times and use these in appropriate places: struct task_cputime { cputime_t utime; cputime_t stime; unsigned long long sum_exec_runtime; }; This is included in the structure "thread_group_cputime," which is a new substructure of signal_struct and which varies for uniprocessor versus multiprocessor kernels. For uniprocessor kernels, it uses "task_cputime" as a simple substructure, while for multiprocessor kernels it is a pointer: struct thread_group_cputime { struct task_cputime totals; }; struct thread_group_cputime { struct task_cputime *totals; }; We also add a new task_cputime substructure directly to signal_struct, to cache the earliest expiration of process-wide timers, and task_cputime also replaces the it_*_expires fields of task_struct (used for earliest expiration of thread timers). The "thread_group_cputime" structure contains process-wide timers that are updated via account_user_time() and friends. In the non-SMP case the structure is a simple aggregator; unfortunately in the SMP case that simplicity was not achievable due to cache-line contention between CPUs (in one measured case performance was actually _worse_ on a 16-cpu system than the same test on a 4-cpu system, due to this contention). For SMP, the thread_group_cputime counters are maintained as a per-cpu structure allocated using alloc_percpu(). The timer functions update only the timer field in the structure corresponding to the running CPU, obtained using per_cpu_ptr(). We define a set of inline functions in sched.h that we use to maintain the thread_group_cputime structure and hide the differences between UP and SMP implementations from the rest of the kernel. The thread_group_cputime_init() function initializes the thread_group_cputime structure for the given task. The thread_group_cputime_alloc() is a no-op for UP; for SMP it calls the out-of-line function thread_group_cputime_alloc_smp() to allocate and fill in the per-cpu structures and fields. The thread_group_cputime_free() function, also a no-op for UP, in SMP frees the per-cpu structures. The thread_group_cputime_clone_thread() function (also a UP no-op) for SMP calls thread_group_cputime_alloc() if the per-cpu structures haven't yet been allocated. The thread_group_cputime() function fills the task_cputime structure it is passed with the contents of the thread_group_cputime fields; in UP it's that simple but in SMP it must also safely check that tsk->signal is non-NULL (if it is it just uses the appropriate fields of task_struct) and, if so, sums the per-cpu values for each online CPU. Finally, the three functions account_group_user_time(), account_group_system_time() and account_group_exec_runtime() are used by timer functions to update the respective fields of the thread_group_cputime structure. Non-SMP operation is trivial and will not be mentioned further. The per-cpu structure is always allocated when a task creates its first new thread, via a call to thread_group_cputime_clone_thread() from copy_signal(). It is freed at process exit via a call to thread_group_cputime_free() from cleanup_signal(). All functions that formerly summed utime/stime/sum_sched_runtime values from from all threads in the thread group now use thread_group_cputime() to snapshot the values in the thread_group_cputime structure or the values in the task structure itself if the per-cpu structure hasn't been allocated. Finally, the code in kernel/posix-cpu-timers.c has changed quite a bit. The run_posix_cpu_timers() function has been split into a fast path and a slow path; the former safely checks whether there are any expired thread timers and, if not, just returns, while the slow path does the heavy lifting. With the dedicated thread group fields, timers are no longer "rebalanced" and the process_timer_rebalance() function and related code has gone away. All summing loops are gone and all code that used them now uses the thread_group_cputime() inline. When process-wide timers are set, the new task_cputime structure in signal_struct is used to cache the earliest expiration; this is checked in the fast path. Performance The fix appears not to add significant overhead to existing operations. It generally performs the same as the current code except in two cases, one in which it performs slightly worse (Case 5 below) and one in which it performs very significantly better (Case 2 below). Overall it's a wash except in those two cases. I've since done somewhat more involved testing on a dual-core Opteron system. Case 1: With no itimer running, for a test with 100,000 threads, the fixed kernel took 1428.5 seconds, 513 seconds more than the unfixed system, all of which was spent in the system. There were twice as many voluntary context switches with the fix as without it. Case 2: With an itimer running at .01 second ticks and 4000 threads (the most an unmodified kernel can handle), the fixed kernel ran the test in eight percent of the time (5.8 seconds as opposed to 70 seconds) and had better tick accuracy (.012 seconds per tick as opposed to .023 seconds per tick). Case 3: A 4000-thread test with an initial timer tick of .01 second and an interval of 10,000 seconds (i.e. a timer that ticks only once) had very nearly the same performance in both cases: 6.3 seconds elapsed for the fixed kernel versus 5.5 seconds for the unfixed kernel. With fewer threads (eight in these tests), the Case 1 test ran in essentially the same time on both the modified and unmodified kernels (5.2 seconds versus 5.8 seconds). The Case 2 test ran in about the same time as well, 5.9 seconds versus 5.4 seconds but again with much better tick accuracy, .013 seconds per tick versus .025 seconds per tick for the unmodified kernel. Since the fix affected the rlimit code, I also tested soft and hard CPU limits. Case 4: With a hard CPU limit of 20 seconds and eight threads (and an itimer running), the modified kernel was very slightly favored in that while it killed the process in 19.997 seconds of CPU time (5.002 seconds of wall time), only .003 seconds of that was system time, the rest was user time. The unmodified kernel killed the process in 20.001 seconds of CPU (5.014 seconds of wall time) of which .016 seconds was system time. Really, though, the results were too close to call. The results were essentially the same with no itimer running. Case 5: With a soft limit of 20 seconds and a hard limit of 2000 seconds (where the hard limit would never be reached) and an itimer running, the modified kernel exhibited worse tick accuracy than the unmodified kernel: .050 seconds/tick versus .028 seconds/tick. Otherwise, performance was almost indistinguishable. With no itimer running this test exhibited virtually identical behavior and times in both cases. In times past I did some limited performance testing. those results are below. On a four-cpu Opteron system without this fix, a sixteen-thread test executed in 3569.991 seconds, of which user was 3568.435s and system was 1.556s. On the same system with the fix, user and elapsed time were about the same, but system time dropped to 0.007 seconds. Performance with eight, four and one thread were comparable. Interestingly, the timer ticks with the fix seemed more accurate: The sixteen-thread test with the fix received 149543 ticks for 0.024 seconds per tick, while the same test without the fix received 58720 for 0.061 seconds per tick. Both cases were configured for an interval of 0.01 seconds. Again, the other tests were comparable. Each thread in this test computed the primes up to 25,000,000. I also did a test with a large number of threads, 100,000 threads, which is impossible without the fix. In this case each thread computed the primes only up to 10,000 (to make the runtime manageable). System time dominated, at 1546.968 seconds out of a total 2176.906 seconds (giving a user time of 629.938s). It received 147651 ticks for 0.015 seconds per tick, still quite accurate. There is obviously no comparable test without the fix. Signed-off-by: Frank Mayhar <fmayhar@google.com> Cc: Roland McGrath <roland@redhat.com> Cc: Alexey Dobriyan <adobriyan@gmail.com> Cc: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2008-09-12 09:54:39 -07:00
sched/cputime: Accumulate vtime on top of nsec clocksource Currently the cputime source used by vtime is jiffies. When we cross a context boundary and jiffies have changed since the last snapshot, the pending cputime is accounted to the switching out context. This system works ok if the ticks are not aligned across CPUs. If they instead are aligned (ie: all fire at the same time) and the CPUs run in userspace, the jiffies change is only observed on tick exit and therefore the user cputime is accounted as system cputime. This is because the CPU that maintains timekeeping fires its tick at the same time as the others. It updates jiffies in the middle of the tick and the other CPUs see that update on IRQ exit: CPU 0 (timekeeper) CPU 1 ------------------- ------------- jiffies = N ... run in userspace for a jiffy tick entry tick entry (sees jiffies = N) set jiffies = N + 1 tick exit tick exit (sees jiffies = N + 1) account 1 jiffy as stime Fix this with using a nanosec clock source instead of jiffies. The cputime is then accumulated and flushed everytime the pending delta reaches a jiffy in order to mitigate the accounting overhead. [ fweisbec: changelog, rebase on struct vtime, field renames, add delta on cputime readers, keep idle vtime as-is (low overhead accounting), harmonize clock sources. ] Suggested-by: Thomas Gleixner <tglx@linutronix.de> Reported-by: Luiz Capitulino <lcapitulino@redhat.com> Tested-by: Luiz Capitulino <lcapitulino@redhat.com> Signed-off-by: Wanpeng Li <wanpeng.li@hotmail.com> Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Rik van Riel <riel@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Wanpeng Li <kernellwp@gmail.com> Link: http://lkml.kernel.org/r/1498756511-11714-6-git-send-email-fweisbec@gmail.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-06-29 19:15:11 +02:00
sched/fair: Add util_est on top of PELT The util_avg signal computed by PELT is too variable for some use-cases. For example, a big task waking up after a long sleep period will have its utilization almost completely decayed. This introduces some latency before schedutil will be able to pick the best frequency to run a task. The same issue can affect task placement. Indeed, since the task utilization is already decayed at wakeup, when the task is enqueued in a CPU, this can result in a CPU running a big task as being temporarily represented as being almost empty. This leads to a race condition where other tasks can be potentially allocated on a CPU which just started to run a big task which slept for a relatively long period. Moreover, the PELT utilization of a task can be updated every [ms], thus making it a continuously changing value for certain longer running tasks. This means that the instantaneous PELT utilization of a RUNNING task is not really meaningful to properly support scheduler decisions. For all these reasons, a more stable signal can do a better job of representing the expected/estimated utilization of a task/cfs_rq. Such a signal can be easily created on top of PELT by still using it as an estimator which produces values to be aggregated on meaningful events. This patch adds a simple implementation of util_est, a new signal built on top of PELT's util_avg where: util_est(task) = max(task::util_avg, f(task::util_avg@dequeue)) This allows to remember how big a task has been reported by PELT in its previous activations via f(task::util_avg@dequeue), which is the new _task_util_est(struct task_struct*) function added by this patch. If a task should change its behavior and it runs longer in a new activation, after a certain time its util_est will just track the original PELT signal (i.e. task::util_avg). The estimated utilization of cfs_rq is defined only for root ones. That's because the only sensible consumer of this signal are the scheduler and schedutil when looking for the overall CPU utilization due to FAIR tasks. For this reason, the estimated utilization of a root cfs_rq is simply defined as: util_est(cfs_rq) = max(cfs_rq::util_avg, cfs_rq::util_est::enqueued) where: cfs_rq::util_est::enqueued = sum(_task_util_est(task)) for each RUNNABLE task on that root cfs_rq It's worth noting that the estimated utilization is tracked only for objects of interests, specifically: - Tasks: to better support tasks placement decisions - root cfs_rqs: to better support both tasks placement decisions as well as frequencies selection Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Joel Fernandes <joelaf@google.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Rafael J . Wysocki <rafael.j.wysocki@intel.com> Cc: Steve Muckle <smuckle@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Todd Kjos <tkjos@android.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Link: http://lkml.kernel.org/r/20180309095245.11071-2-patrick.bellasi@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-03-09 09:52:42 +00:00
sched/core: Force proper alignment of 'struct util_est' For some as yet not understood reason, Tony gets unaligned access traps on IA64 because of: struct util_est ue = READ_ONCE(p->se.avg.util_est); and: WRITE_ONCE(p->se.avg.util_est, ue); introduced by commit: d519329f72a6 ("sched/fair: Update util_est only on util_avg updates") Normally those two fields should end up on an 8-byte aligned location, but UP and RANDSTRUCT can mess that up so enforce the alignment explicitly. Also make the alignment on sched_avg unconditional, as it is really about data locality, not false-sharing. With or without this patch the layout for sched_avg on a ia64-defconfig build looks like: $ pahole -EC sched_avg ia64-defconfig/kernel/sched/core.o die__process_function: tag not supported (INVALID)! struct sched_avg { /* typedef u64 */ long long unsigned int last_update_time; /* 0 8 */ /* typedef u64 */ long long unsigned int load_sum; /* 8 8 */ /* typedef u64 */ long long unsigned int runnable_load_sum; /* 16 8 */ /* typedef u32 */ unsigned int util_sum; /* 24 4 */ /* typedef u32 */ unsigned int period_contrib; /* 28 4 */ long unsigned int load_avg; /* 32 8 */ long unsigned int runnable_load_avg; /* 40 8 */ long unsigned int util_avg; /* 48 8 */ struct util_est { unsigned int enqueued; /* 56 4 */ unsigned int ewma; /* 60 4 */ } util_est; /* 56 8 */ /* --- cacheline 1 boundary (64 bytes) --- */ /* size: 64, cachelines: 1, members: 9 */ }; Reported-and-Tested-by: Tony Luck <tony.luck@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Frederic Weisbecker <frederic@kernel.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Norbert Manthey <nmanthey@amazon.de> Cc: Patrick Bellasi <patrick.bellasi@arm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tony <tony.luck@intel.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Fixes: d519329f72a6 ("sched/fair: Update util_est only on util_avg updates") Link: http://lkml.kernel.org/r/20180405080521.GG4129@hirez.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-04-05 10:05:21 +02:00
sched/fair: Add util_est on top of PELT The util_avg signal computed by PELT is too variable for some use-cases. For example, a big task waking up after a long sleep period will have its utilization almost completely decayed. This introduces some latency before schedutil will be able to pick the best frequency to run a task. The same issue can affect task placement. Indeed, since the task utilization is already decayed at wakeup, when the task is enqueued in a CPU, this can result in a CPU running a big task as being temporarily represented as being almost empty. This leads to a race condition where other tasks can be potentially allocated on a CPU which just started to run a big task which slept for a relatively long period. Moreover, the PELT utilization of a task can be updated every [ms], thus making it a continuously changing value for certain longer running tasks. This means that the instantaneous PELT utilization of a RUNNING task is not really meaningful to properly support scheduler decisions. For all these reasons, a more stable signal can do a better job of representing the expected/estimated utilization of a task/cfs_rq. Such a signal can be easily created on top of PELT by still using it as an estimator which produces values to be aggregated on meaningful events. This patch adds a simple implementation of util_est, a new signal built on top of PELT's util_avg where: util_est(task) = max(task::util_avg, f(task::util_avg@dequeue)) This allows to remember how big a task has been reported by PELT in its previous activations via f(task::util_avg@dequeue), which is the new _task_util_est(struct task_struct*) function added by this patch. If a task should change its behavior and it runs longer in a new activation, after a certain time its util_est will just track the original PELT signal (i.e. task::util_avg). The estimated utilization of cfs_rq is defined only for root ones. That's because the only sensible consumer of this signal are the scheduler and schedutil when looking for the overall CPU utilization due to FAIR tasks. For this reason, the estimated utilization of a root cfs_rq is simply defined as: util_est(cfs_rq) = max(cfs_rq::util_avg, cfs_rq::util_est::enqueued) where: cfs_rq::util_est::enqueued = sum(_task_util_est(task)) for each RUNNABLE task on that root cfs_rq It's worth noting that the estimated utilization is tracked only for objects of interests, specifically: - Tasks: to better support tasks placement decisions - root cfs_rqs: to better support both tasks placement decisions as well as frequencies selection Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Joel Fernandes <joelaf@google.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Rafael J . Wysocki <rafael.j.wysocki@intel.com> Cc: Steve Muckle <smuckle@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Todd Kjos <tkjos@android.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Link: http://lkml.kernel.org/r/20180309095245.11071-2-patrick.bellasi@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-03-09 09:52:42 +00:00
sched/fair: Rewrite runnable load and utilization average tracking The idea of runnable load average (let runnable time contribute to weight) was proposed by Paul Turner and Ben Segall, and it is still followed by this rewrite. This rewrite aims to solve the following issues: 1. cfs_rq's load average (namely runnable_load_avg and blocked_load_avg) is updated at the granularity of an entity at a time, which results in the cfs_rq's load average is stale or partially updated: at any time, only one entity is up to date, all other entities are effectively lagging behind. This is undesirable. To illustrate, if we have n runnable entities in the cfs_rq, as time elapses, they certainly become outdated: t0: cfs_rq { e1_old, e2_old, ..., en_old } and when we update: t1: update e1, then we have cfs_rq { e1_new, e2_old, ..., en_old } t2: update e2, then we have cfs_rq { e1_old, e2_new, ..., en_old } ... We solve this by combining all runnable entities' load averages together in cfs_rq's avg, and update the cfs_rq's avg as a whole. This is based on the fact that if we regard the update as a function, then: w * update(e) = update(w * e) and update(e1) + update(e2) = update(e1 + e2), then w1 * update(e1) + w2 * update(e2) = update(w1 * e1 + w2 * e2) therefore, by this rewrite, we have an entirely updated cfs_rq at the time we update it: t1: update cfs_rq { e1_new, e2_new, ..., en_new } t2: update cfs_rq { e1_new, e2_new, ..., en_new } ... 2. cfs_rq's load average is different between top rq->cfs_rq and other task_group's per CPU cfs_rqs in whether or not blocked_load_average contributes to the load. The basic idea behind runnable load average (the same for utilization) is that the blocked state is taken into account as opposed to only accounting for the currently runnable state. Therefore, the average should include both the runnable/running and blocked load averages. This rewrite does that. In addition, we also combine runnable/running and blocked averages of all entities into the cfs_rq's average, and update it together at once. This is based on the fact that: update(runnable) + update(blocked) = update(runnable + blocked) This significantly reduces the code as we don't need to separately maintain/update runnable/running load and blocked load. 3. How task_group entities' share is calculated is complex and imprecise. We reduce the complexity in this rewrite to allow a very simple rule: the task_group's load_avg is aggregated from its per CPU cfs_rqs's load_avgs. Then group entity's weight is simply proportional to its own cfs_rq's load_avg / task_group's load_avg. To illustrate, if a task_group has { cfs_rq1, cfs_rq2, ..., cfs_rqn }, then, task_group_avg = cfs_rq1_avg + cfs_rq2_avg + ... + cfs_rqn_avg, then cfs_rqx's entity's share = cfs_rqx_avg / task_group_avg * task_group's share To sum up, this rewrite in principle is equivalent to the current one, but fixes the issues described above. Turns out, it significantly reduces the code complexity and hence increases clarity and efficiency. In addition, the new averages are more smooth/continuous (no spurious spikes and valleys) and updated more consistently and quickly to reflect the load dynamics. As a result, we have less load tracking overhead, better performance, and especially better power efficiency due to more balanced load. Signed-off-by: Yuyang Du <yuyang.du@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: arjan@linux.intel.com Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: fengguang.wu@intel.com Cc: len.brown@intel.com Cc: morten.rasmussen@arm.com Cc: pjt@google.com Cc: rafael.j.wysocki@intel.com Cc: umgwanakikbuti@gmail.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1436918682-4971-3-git-send-email-yuyang.du@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-07-15 08:04:37 +08:00
sched/fair: Rewrite runnable load and utilization average tracking The idea of runnable load average (let runnable time contribute to weight) was proposed by Paul Turner and Ben Segall, and it is still followed by this rewrite. This rewrite aims to solve the following issues: 1. cfs_rq's load average (namely runnable_load_avg and blocked_load_avg) is updated at the granularity of an entity at a time, which results in the cfs_rq's load average is stale or partially updated: at any time, only one entity is up to date, all other entities are effectively lagging behind. This is undesirable. To illustrate, if we have n runnable entities in the cfs_rq, as time elapses, they certainly become outdated: t0: cfs_rq { e1_old, e2_old, ..., en_old } and when we update: t1: update e1, then we have cfs_rq { e1_new, e2_old, ..., en_old } t2: update e2, then we have cfs_rq { e1_old, e2_new, ..., en_old } ... We solve this by combining all runnable entities' load averages together in cfs_rq's avg, and update the cfs_rq's avg as a whole. This is based on the fact that if we regard the update as a function, then: w * update(e) = update(w * e) and update(e1) + update(e2) = update(e1 + e2), then w1 * update(e1) + w2 * update(e2) = update(w1 * e1 + w2 * e2) therefore, by this rewrite, we have an entirely updated cfs_rq at the time we update it: t1: update cfs_rq { e1_new, e2_new, ..., en_new } t2: update cfs_rq { e1_new, e2_new, ..., en_new } ... 2. cfs_rq's load average is different between top rq->cfs_rq and other task_group's per CPU cfs_rqs in whether or not blocked_load_average contributes to the load. The basic idea behind runnable load average (the same for utilization) is that the blocked state is taken into account as opposed to only accounting for the currently runnable state. Therefore, the average should include both the runnable/running and blocked load averages. This rewrite does that. In addition, we also combine runnable/running and blocked averages of all entities into the cfs_rq's average, and update it together at once. This is based on the fact that: update(runnable) + update(blocked) = update(runnable + blocked) This significantly reduces the code as we don't need to separately maintain/update runnable/running load and blocked load. 3. How task_group entities' share is calculated is complex and imprecise. We reduce the complexity in this rewrite to allow a very simple rule: the task_group's load_avg is aggregated from its per CPU cfs_rqs's load_avgs. Then group entity's weight is simply proportional to its own cfs_rq's load_avg / task_group's load_avg. To illustrate, if a task_group has { cfs_rq1, cfs_rq2, ..., cfs_rqn }, then, task_group_avg = cfs_rq1_avg + cfs_rq2_avg + ... + cfs_rqn_avg, then cfs_rqx's entity's share = cfs_rqx_avg / task_group_avg * task_group's share To sum up, this rewrite in principle is equivalent to the current one, but fixes the issues described above. Turns out, it significantly reduces the code complexity and hence increases clarity and efficiency. In addition, the new averages are more smooth/continuous (no spurious spikes and valleys) and updated more consistently and quickly to reflect the load dynamics. As a result, we have less load tracking overhead, better performance, and especially better power efficiency due to more balanced load. Signed-off-by: Yuyang Du <yuyang.du@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: arjan@linux.intel.com Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: fengguang.wu@intel.com Cc: len.brown@intel.com Cc: morten.rasmussen@arm.com Cc: pjt@google.com Cc: rafael.j.wysocki@intel.com Cc: umgwanakikbuti@gmail.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1436918682-4971-3-git-send-email-yuyang.du@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-07-15 08:04:37 +08:00
sched/fair: Rewrite runnable load and utilization average tracking The idea of runnable load average (let runnable time contribute to weight) was proposed by Paul Turner and Ben Segall, and it is still followed by this rewrite. This rewrite aims to solve the following issues: 1. cfs_rq's load average (namely runnable_load_avg and blocked_load_avg) is updated at the granularity of an entity at a time, which results in the cfs_rq's load average is stale or partially updated: at any time, only one entity is up to date, all other entities are effectively lagging behind. This is undesirable. To illustrate, if we have n runnable entities in the cfs_rq, as time elapses, they certainly become outdated: t0: cfs_rq { e1_old, e2_old, ..., en_old } and when we update: t1: update e1, then we have cfs_rq { e1_new, e2_old, ..., en_old } t2: update e2, then we have cfs_rq { e1_old, e2_new, ..., en_old } ... We solve this by combining all runnable entities' load averages together in cfs_rq's avg, and update the cfs_rq's avg as a whole. This is based on the fact that if we regard the update as a function, then: w * update(e) = update(w * e) and update(e1) + update(e2) = update(e1 + e2), then w1 * update(e1) + w2 * update(e2) = update(w1 * e1 + w2 * e2) therefore, by this rewrite, we have an entirely updated cfs_rq at the time we update it: t1: update cfs_rq { e1_new, e2_new, ..., en_new } t2: update cfs_rq { e1_new, e2_new, ..., en_new } ... 2. cfs_rq's load average is different between top rq->cfs_rq and other task_group's per CPU cfs_rqs in whether or not blocked_load_average contributes to the load. The basic idea behind runnable load average (the same for utilization) is that the blocked state is taken into account as opposed to only accounting for the currently runnable state. Therefore, the average should include both the runnable/running and blocked load averages. This rewrite does that. In addition, we also combine runnable/running and blocked averages of all entities into the cfs_rq's average, and update it together at once. This is based on the fact that: update(runnable) + update(blocked) = update(runnable + blocked) This significantly reduces the code as we don't need to separately maintain/update runnable/running load and blocked load. 3. How task_group entities' share is calculated is complex and imprecise. We reduce the complexity in this rewrite to allow a very simple rule: the task_group's load_avg is aggregated from its per CPU cfs_rqs's load_avgs. Then group entity's weight is simply proportional to its own cfs_rq's load_avg / task_group's load_avg. To illustrate, if a task_group has { cfs_rq1, cfs_rq2, ..., cfs_rqn }, then, task_group_avg = cfs_rq1_avg + cfs_rq2_avg + ... + cfs_rqn_avg, then cfs_rqx's entity's share = cfs_rqx_avg / task_group_avg * task_group's share To sum up, this rewrite in principle is equivalent to the current one, but fixes the issues described above. Turns out, it significantly reduces the code complexity and hence increases clarity and efficiency. In addition, the new averages are more smooth/continuous (no spurious spikes and valleys) and updated more consistently and quickly to reflect the load dynamics. As a result, we have less load tracking overhead, better performance, and especially better power efficiency due to more balanced load. Signed-off-by: Yuyang Du <yuyang.du@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: arjan@linux.intel.com Cc: bsegall@google.com Cc: dietmar.eggemann@arm.com Cc: fengguang.wu@intel.com Cc: len.brown@intel.com Cc: morten.rasmussen@arm.com Cc: pjt@google.com Cc: rafael.j.wysocki@intel.com Cc: umgwanakikbuti@gmail.com Cc: vincent.guittot@linaro.org Link: http://lkml.kernel.org/r/1436918682-4971-3-git-send-email-yuyang.du@intel.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-07-15 08:04:37 +08:00
sched/fair: Add util_est on top of PELT The util_avg signal computed by PELT is too variable for some use-cases. For example, a big task waking up after a long sleep period will have its utilization almost completely decayed. This introduces some latency before schedutil will be able to pick the best frequency to run a task. The same issue can affect task placement. Indeed, since the task utilization is already decayed at wakeup, when the task is enqueued in a CPU, this can result in a CPU running a big task as being temporarily represented as being almost empty. This leads to a race condition where other tasks can be potentially allocated on a CPU which just started to run a big task which slept for a relatively long period. Moreover, the PELT utilization of a task can be updated every [ms], thus making it a continuously changing value for certain longer running tasks. This means that the instantaneous PELT utilization of a RUNNING task is not really meaningful to properly support scheduler decisions. For all these reasons, a more stable signal can do a better job of representing the expected/estimated utilization of a task/cfs_rq. Such a signal can be easily created on top of PELT by still using it as an estimator which produces values to be aggregated on meaningful events. This patch adds a simple implementation of util_est, a new signal built on top of PELT's util_avg where: util_est(task) = max(task::util_avg, f(task::util_avg@dequeue)) This allows to remember how big a task has been reported by PELT in its previous activations via f(task::util_avg@dequeue), which is the new _task_util_est(struct task_struct*) function added by this patch. If a task should change its behavior and it runs longer in a new activation, after a certain time its util_est will just track the original PELT signal (i.e. task::util_avg). The estimated utilization of cfs_rq is defined only for root ones. That's because the only sensible consumer of this signal are the scheduler and schedutil when looking for the overall CPU utilization due to FAIR tasks. For this reason, the estimated utilization of a root cfs_rq is simply defined as: util_est(cfs_rq) = max(cfs_rq::util_avg, cfs_rq::util_est::enqueued) where: cfs_rq::util_est::enqueued = sum(_task_util_est(task)) for each RUNNABLE task on that root cfs_rq It's worth noting that the estimated utilization is tracked only for objects of interests, specifically: - Tasks: to better support tasks placement decisions - root cfs_rqs: to better support both tasks placement decisions as well as frequencies selection Signed-off-by: Patrick Bellasi <patrick.bellasi@arm.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Reviewed-by: Dietmar Eggemann <dietmar.eggemann@arm.com> Cc: Joel Fernandes <joelaf@google.com> Cc: Juri Lelli <juri.lelli@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Morten Rasmussen <morten.rasmussen@arm.com> Cc: Paul Turner <pjt@google.com> Cc: Rafael J . Wysocki <rafael.j.wysocki@intel.com> Cc: Steve Muckle <smuckle@google.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Todd Kjos <tkjos@android.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Cc: Viresh Kumar <viresh.kumar@linaro.org> Link: http://lkml.kernel.org/r/20180309095245.11071-2-patrick.bellasi@arm.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-03-09 09:52:42 +00:00
sched/core: Force proper alignment of 'struct util_est' For some as yet not understood reason, Tony gets unaligned access traps on IA64 because of: struct util_est ue = READ_ONCE(p->se.avg.util_est); and: WRITE_ONCE(p->se.avg.util_est, ue); introduced by commit: d519329f72a6 ("sched/fair: Update util_est only on util_avg updates") Normally those two fields should end up on an 8-byte aligned location, but UP and RANDSTRUCT can mess that up so enforce the alignment explicitly. Also make the alignment on sched_avg unconditional, as it is really about data locality, not false-sharing. With or without this patch the layout for sched_avg on a ia64-defconfig build looks like: $ pahole -EC sched_avg ia64-defconfig/kernel/sched/core.o die__process_function: tag not supported (INVALID)! struct sched_avg { /* typedef u64 */ long long unsigned int last_update_time; /* 0 8 */ /* typedef u64 */ long long unsigned int load_sum; /* 8 8 */ /* typedef u64 */ long long unsigned int runnable_load_sum; /* 16 8 */ /* typedef u32 */ unsigned int util_sum; /* 24 4 */ /* typedef u32 */ unsigned int period_contrib; /* 28 4 */ long unsigned int load_avg; /* 32 8 */ long unsigned int runnable_load_avg; /* 40 8 */ long unsigned int util_avg; /* 48 8 */ struct util_est { unsigned int enqueued; /* 56 4 */ unsigned int ewma; /* 60 4 */ } util_est; /* 56 8 */ /* --- cacheline 1 boundary (64 bytes) --- */ /* size: 64, cachelines: 1, members: 9 */ }; Reported-and-Tested-by: Tony Luck <tony.luck@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Frederic Weisbecker <frederic@kernel.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Norbert Manthey <nmanthey@amazon.de> Cc: Patrick Bellasi <patrick.bellasi@arm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tony <tony.luck@intel.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Fixes: d519329f72a6 ("sched/fair: Update util_est only on util_avg updates") Link: http://lkml.kernel.org/r/20180405080521.GG4129@hirez.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-04-05 10:05:21 +02:00
sched/core: Move sched_entity::avg into separate cache line The sched_entity::avg collides with read-mostly sched_entity data. The perf c2c tool showed many read HITM accesses across many CPUs for sched_entity's cfs_rq and my_q, while having at the same time tons of stores for avg. After placing sched_entity::avg into separate cache line, the perf bench sched pipe showed around 20 seconds speedup. NOTE I cut out all perf events except for cycles and instructions from following output. Before: $ perf stat -r 5 perf bench sched pipe -l 10000000 # Running 'sched/pipe' benchmark: # Executed 10000000 pipe operations between two processes Total time: 270.348 [sec] 27.034805 usecs/op 36989 ops/sec ... 245,537,074,035 cycles # 1.433 GHz 187,264,548,519 instructions # 0.77 insns per cycle 272.653840535 seconds time elapsed ( +- 1.31% ) After: $ perf stat -r 5 perf bench sched pipe -l 10000000 # Running 'sched/pipe' benchmark: # Executed 10000000 pipe operations between two processes Total time: 251.076 [sec] 25.107678 usecs/op 39828 ops/sec ... 244,573,513,928 cycles # 1.572 GHz 187,409,641,157 instructions # 0.76 insns per cycle 251.679315188 seconds time elapsed ( +- 0.31% ) Signed-off-by: Jiri Olsa <jolsa@kernel.org> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Arnaldo Carvalho de Melo <acme@kernel.org> Cc: Don Zickus <dzickus@redhat.com> Cc: Joe Mario <jmario@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/1449606239-28602-1-git-send-email-jolsa@kernel.org Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-12-08 21:23:59 +01:00
sched/core: Force proper alignment of 'struct util_est' For some as yet not understood reason, Tony gets unaligned access traps on IA64 because of: struct util_est ue = READ_ONCE(p->se.avg.util_est); and: WRITE_ONCE(p->se.avg.util_est, ue); introduced by commit: d519329f72a6 ("sched/fair: Update util_est only on util_avg updates") Normally those two fields should end up on an 8-byte aligned location, but UP and RANDSTRUCT can mess that up so enforce the alignment explicitly. Also make the alignment on sched_avg unconditional, as it is really about data locality, not false-sharing. With or without this patch the layout for sched_avg on a ia64-defconfig build looks like: $ pahole -EC sched_avg ia64-defconfig/kernel/sched/core.o die__process_function: tag not supported (INVALID)! struct sched_avg { /* typedef u64 */ long long unsigned int last_update_time; /* 0 8 */ /* typedef u64 */ long long unsigned int load_sum; /* 8 8 */ /* typedef u64 */ long long unsigned int runnable_load_sum; /* 16 8 */ /* typedef u32 */ unsigned int util_sum; /* 24 4 */ /* typedef u32 */ unsigned int period_contrib; /* 28 4 */ long unsigned int load_avg; /* 32 8 */ long unsigned int runnable_load_avg; /* 40 8 */ long unsigned int util_avg; /* 48 8 */ struct util_est { unsigned int enqueued; /* 56 4 */ unsigned int ewma; /* 60 4 */ } util_est; /* 56 8 */ /* --- cacheline 1 boundary (64 bytes) --- */ /* size: 64, cachelines: 1, members: 9 */ }; Reported-and-Tested-by: Tony Luck <tony.luck@intel.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Frederic Weisbecker <frederic@kernel.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Norbert Manthey <nmanthey@amazon.de> Cc: Patrick Bellasi <patrick.bellasi@arm.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tony <tony.luck@intel.com> Cc: Vincent Guittot <vincent.guittot@linaro.org> Fixes: d519329f72a6 ("sched/fair: Update util_est only on util_avg updates") Link: http://lkml.kernel.org/r/20180405080521.GG4129@hirez.programming.kicks-ass.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-04-05 10:05:21 +02:00
sched/deadline: Add SCHED_DEADLINE structures & implementation Introduces the data structures, constants and symbols needed for SCHED_DEADLINE implementation. Core data structure of SCHED_DEADLINE are defined, along with their initializers. Hooks for checking if a task belong to the new policy are also added where they are needed. Adds a scheduling class, in sched/dl.c and a new policy called SCHED_DEADLINE. It is an implementation of the Earliest Deadline First (EDF) scheduling algorithm, augmented with a mechanism (called Constant Bandwidth Server, CBS) that makes it possible to isolate the behaviour of tasks between each other. The typical -deadline task will be made up of a computation phase (instance) which is activated on a periodic or sporadic fashion. The expected (maximum) duration of such computation is called the task's runtime; the time interval by which each instance need to be completed is called the task's relative deadline. The task's absolute deadline is dynamically calculated as the time instant a task (better, an instance) activates plus the relative deadline. The EDF algorithms selects the task with the smallest absolute deadline as the one to be executed first, while the CBS ensures each task to run for at most its runtime every (relative) deadline length time interval, avoiding any interference between different tasks (bandwidth isolation). Thanks to this feature, also tasks that do not strictly comply with the computational model sketched above can effectively use the new policy. To summarize, this patch: - introduces the data structures, constants and symbols needed; - implements the core logic of the scheduling algorithm in the new scheduling class file; - provides all the glue code between the new scheduling class and the core scheduler and refines the interactions between sched/dl and the other existing scheduling classes. Signed-off-by: Dario Faggioli <raistlin@linux.it> Signed-off-by: Michael Trimarchi <michael@amarulasolutions.com> Signed-off-by: Fabio Checconi <fchecconi@gmail.com> Signed-off-by: Juri Lelli <juri.lelli@gmail.com> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1383831828-15501-4-git-send-email-juri.lelli@gmail.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-11-28 11:14:43 +01:00
sched/deadline: Add SCHED_DEADLINE structures & implementation Introduces the data structures, constants and symbols needed for SCHED_DEADLINE implementation. Core data structure of SCHED_DEADLINE are defined, along with their initializers. Hooks for checking if a task belong to the new policy are also added where they are needed. Adds a scheduling class, in sched/dl.c and a new policy called SCHED_DEADLINE. It is an implementation of the Earliest Deadline First (EDF) scheduling algorithm, augmented with a mechanism (called Constant Bandwidth Server, CBS) that makes it possible to isolate the behaviour of tasks between each other. The typical -deadline task will be made up of a computation phase (instance) which is activated on a periodic or sporadic fashion. The expected (maximum) duration of such computation is called the task's runtime; the time interval by which each instance need to be completed is called the task's relative deadline. The task's absolute deadline is dynamically calculated as the time instant a task (better, an instance) activates plus the relative deadline. The EDF algorithms selects the task with the smallest absolute deadline as the one to be executed first, while the CBS ensures each task to run for at most its runtime every (relative) deadline length time interval, avoiding any interference between different tasks (bandwidth isolation). Thanks to this feature, also tasks that do not strictly comply with the computational model sketched above can effectively use the new policy. To summarize, this patch: - introduces the data structures, constants and symbols needed; - implements the core logic of the scheduling algorithm in the new scheduling class file; - provides all the glue code between the new scheduling class and the core scheduler and refines the interactions between sched/dl and the other existing scheduling classes. Signed-off-by: Dario Faggioli <raistlin@linux.it> Signed-off-by: Michael Trimarchi <michael@amarulasolutions.com> Signed-off-by: Fabio Checconi <fchecconi@gmail.com> Signed-off-by: Juri Lelli <juri.lelli@gmail.com> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1383831828-15501-4-git-send-email-juri.lelli@gmail.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-11-28 11:14:43 +01:00
sched/deadline: Add SCHED_DEADLINE structures & implementation Introduces the data structures, constants and symbols needed for SCHED_DEADLINE implementation. Core data structure of SCHED_DEADLINE are defined, along with their initializers. Hooks for checking if a task belong to the new policy are also added where they are needed. Adds a scheduling class, in sched/dl.c and a new policy called SCHED_DEADLINE. It is an implementation of the Earliest Deadline First (EDF) scheduling algorithm, augmented with a mechanism (called Constant Bandwidth Server, CBS) that makes it possible to isolate the behaviour of tasks between each other. The typical -deadline task will be made up of a computation phase (instance) which is activated on a periodic or sporadic fashion. The expected (maximum) duration of such computation is called the task's runtime; the time interval by which each instance need to be completed is called the task's relative deadline. The task's absolute deadline is dynamically calculated as the time instant a task (better, an instance) activates plus the relative deadline. The EDF algorithms selects the task with the smallest absolute deadline as the one to be executed first, while the CBS ensures each task to run for at most its runtime every (relative) deadline length time interval, avoiding any interference between different tasks (bandwidth isolation). Thanks to this feature, also tasks that do not strictly comply with the computational model sketched above can effectively use the new policy. To summarize, this patch: - introduces the data structures, constants and symbols needed; - implements the core logic of the scheduling algorithm in the new scheduling class file; - provides all the glue code between the new scheduling class and the core scheduler and refines the interactions between sched/dl and the other existing scheduling classes. Signed-off-by: Dario Faggioli <raistlin@linux.it> Signed-off-by: Michael Trimarchi <michael@amarulasolutions.com> Signed-off-by: Fabio Checconi <fchecconi@gmail.com> Signed-off-by: Juri Lelli <juri.lelli@gmail.com> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1383831828-15501-4-git-send-email-juri.lelli@gmail.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-11-28 11:14:43 +01:00
sched/deadline: Use the revised wakeup rule for suspending constrained dl tasks We have been facing some problems with self-suspending constrained deadline tasks. The main reason is that the original CBS was not designed for such sort of tasks. One problem reported by Xunlei Pang takes place when a task suspends, and then is awakened before the deadline, but so close to the deadline that its remaining runtime can cause the task to have an absolute density higher than allowed. In such situation, the original CBS assumes that the task is facing an early activation, and so it replenishes the task and set another deadline, one deadline in the future. This rule works fine for implicit deadline tasks. Moreover, it allows the system to adapt the period of a task in which the external event source suffered from a clock drift. However, this opens the window for bandwidth leakage for constrained deadline tasks. For instance, a task with the following parameters: runtime = 5 ms deadline = 7 ms [density] = 5 / 7 = 0.71 period = 1000 ms If the task runs for 1 ms, and then suspends for another 1ms, it will be awakened with the following parameters: remaining runtime = 4 laxity = 5 presenting a absolute density of 4 / 5 = 0.80. In this case, the original CBS would assume the task had an early wakeup. Then, CBS will reset the runtime, and the absolute deadline will be postponed by one relative deadline, allowing the task to run. The problem is that, if the task runs this pattern forever, it will keep receiving bandwidth, being able to run 1ms every 2ms. Following this behavior, the task would be able to run 500 ms in 1 sec. Thus running more than the 5 ms / 1 sec the admission control allowed it to run. Trying to address the self-suspending case, Luca Abeni, Giuseppe Lipari, and Juri Lelli [1] revisited the CBS in order to deal with self-suspending tasks. In the new approach, rather than replenishing/postponing the absolute deadline, the revised wakeup rule adjusts the remaining runtime, reducing it to fit into the allowed density. A revised version of the idea is: At a given time t, the maximum absolute density of a task cannot be higher than its relative density, that is: runtime / (deadline - t) <= dl_runtime / dl_deadline Knowing the laxity of a task (deadline - t), it is possible to move it to the other side of the equality, thus enabling to define max remaining runtime a task can use within the absolute deadline, without over-running the allowed density: runtime = (dl_runtime / dl_deadline) * (deadline - t) For instance, in our previous example, the task could still run: runtime = ( 5 / 7 ) * 5 runtime = 3.57 ms Without causing damage for other deadline tasks. It is note worthy that the laxity cannot be negative because that would cause a negative runtime. Thus, this patch depends on the patch: df8eac8cafce ("sched/deadline: Throttle a constrained deadline task activated after the deadline") Which throttles a constrained deadline task activated after the deadline. Finally, it is also possible to use the revised wakeup rule for all other tasks, but that would require some more discussions about pros and cons. Reported-by: Xunlei Pang <xpang@redhat.com> Signed-off-by: Daniel Bristot de Oliveira <bristot@redhat.com> [peterz: replaced dl_is_constrained with dl_is_implicit] Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Juri Lelli <juri.lelli@arm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Luca Abeni <luca.abeni@santannapisa.it> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Romulo Silva de Oliveira <romulo.deoliveira@ufsc.br> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Tommaso Cucinotta <tommaso.cucinotta@sssup.it> Link: http://lkml.kernel.org/r/5c800ab3a74a168a84ee5f3f84d12a02e11383be.1495803804.git.bristot@redhat.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2017-05-29 16:24:03 +02:00
sched/deadline: Add SCHED_DEADLINE structures & implementation Introduces the data structures, constants and symbols needed for SCHED_DEADLINE implementation. Core data structure of SCHED_DEADLINE are defined, along with their initializers. Hooks for checking if a task belong to the new policy are also added where they are needed. Adds a scheduling class, in sched/dl.c and a new policy called SCHED_DEADLINE. It is an implementation of the Earliest Deadline First (EDF) scheduling algorithm, augmented with a mechanism (called Constant Bandwidth Server, CBS) that makes it possible to isolate the behaviour of tasks between each other. The typical -deadline task will be made up of a computation phase (instance) which is activated on a periodic or sporadic fashion. The expected (maximum) duration of such computation is called the task's runtime; the time interval by which each instance need to be completed is called the task's relative deadline. The task's absolute deadline is dynamically calculated as the time instant a task (better, an instance) activates plus the relative deadline. The EDF algorithms selects the task with the smallest absolute deadline as the one to be executed first, while the CBS ensures each task to run for at most its runtime every (relative) deadline length time interval, avoiding any interference between different tasks (bandwidth isolation). Thanks to this feature, also tasks that do not strictly comply with the computational model sketched above can effectively use the new policy. To summarize, this patch: - introduces the data structures, constants and symbols needed; - implements the core logic of the scheduling algorithm in the new scheduling class file; - provides all the glue code between the new scheduling class and the core scheduler and refines the interactions between sched/dl and the other existing scheduling classes. Signed-off-by: Dario Faggioli <raistlin@linux.it> Signed-off-by: Michael Trimarchi <michael@amarulasolutions.com> Signed-off-by: Fabio Checconi <fchecconi@gmail.com> Signed-off-by: Juri Lelli <juri.lelli@gmail.com> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1383831828-15501-4-git-send-email-juri.lelli@gmail.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-11-28 11:14:43 +01:00
sched/deadline: Add SCHED_DEADLINE structures & implementation Introduces the data structures, constants and symbols needed for SCHED_DEADLINE implementation. Core data structure of SCHED_DEADLINE are defined, along with their initializers. Hooks for checking if a task belong to the new policy are also added where they are needed. Adds a scheduling class, in sched/dl.c and a new policy called SCHED_DEADLINE. It is an implementation of the Earliest Deadline First (EDF) scheduling algorithm, augmented with a mechanism (called Constant Bandwidth Server, CBS) that makes it possible to isolate the behaviour of tasks between each other. The typical -deadline task will be made up of a computation phase (instance) which is activated on a periodic or sporadic fashion. The expected (maximum) duration of such computation is called the task's runtime; the time interval by which each instance need to be completed is called the task's relative deadline. The task's absolute deadline is dynamically calculated as the time instant a task (better, an instance) activates plus the relative deadline. The EDF algorithms selects the task with the smallest absolute deadline as the one to be executed first, while the CBS ensures each task to run for at most its runtime every (relative) deadline length time interval, avoiding any interference between different tasks (bandwidth isolation). Thanks to this feature, also tasks that do not strictly comply with the computational model sketched above can effectively use the new policy. To summarize, this patch: - introduces the data structures, constants and symbols needed; - implements the core logic of the scheduling algorithm in the new scheduling class file; - provides all the glue code between the new scheduling class and the core scheduler and refines the interactions between sched/dl and the other existing scheduling classes. Signed-off-by: Dario Faggioli <raistlin@linux.it> Signed-off-by: Michael Trimarchi <michael@amarulasolutions.com> Signed-off-by: Fabio Checconi <fchecconi@gmail.com> Signed-off-by: Juri Lelli <juri.lelli@gmail.com> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1383831828-15501-4-git-send-email-juri.lelli@gmail.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-11-28 11:14:43 +01:00
sched/deadline: Add SCHED_DEADLINE inheritance logic Some method to deal with rt-mutexes and make sched_dl interact with the current PI-coded is needed, raising all but trivial issues, that needs (according to us) to be solved with some restructuring of the pi-code (i.e., going toward a proxy execution-ish implementation). This is under development, in the meanwhile, as a temporary solution, what this commits does is: - ensure a pi-lock owner with waiters is never throttled down. Instead, when it runs out of runtime, it immediately gets replenished and it's deadline is postponed; - the scheduling parameters (relative deadline and default runtime) used for that replenishments --during the whole period it holds the pi-lock-- are the ones of the waiting task with earliest deadline. Acting this way, we provide some kind of boosting to the lock-owner, still by using the existing (actually, slightly modified by the previous commit) pi-architecture. We would stress the fact that this is only a surely needed, all but clean solution to the problem. In the end it's only a way to re-start discussion within the community. So, as always, comments, ideas, rants, etc.. are welcome! :-) Signed-off-by: Dario Faggioli <raistlin@linux.it> Signed-off-by: Juri Lelli <juri.lelli@gmail.com> [ Added !RT_MUTEXES build fix. ] Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1383831828-15501-11-git-send-email-juri.lelli@gmail.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-11-07 14:43:44 +01:00
sched/deadline: Add SCHED_DEADLINE structures & implementation Introduces the data structures, constants and symbols needed for SCHED_DEADLINE implementation. Core data structure of SCHED_DEADLINE are defined, along with their initializers. Hooks for checking if a task belong to the new policy are also added where they are needed. Adds a scheduling class, in sched/dl.c and a new policy called SCHED_DEADLINE. It is an implementation of the Earliest Deadline First (EDF) scheduling algorithm, augmented with a mechanism (called Constant Bandwidth Server, CBS) that makes it possible to isolate the behaviour of tasks between each other. The typical -deadline task will be made up of a computation phase (instance) which is activated on a periodic or sporadic fashion. The expected (maximum) duration of such computation is called the task's runtime; the time interval by which each instance need to be completed is called the task's relative deadline. The task's absolute deadline is dynamically calculated as the time instant a task (better, an instance) activates plus the relative deadline. The EDF algorithms selects the task with the smallest absolute deadline as the one to be executed first, while the CBS ensures each task to run for at most its runtime every (relative) deadline length time interval, avoiding any interference between different tasks (bandwidth isolation). Thanks to this feature, also tasks that do not strictly comply with the computational model sketched above can effectively use the new policy. To summarize, this patch: - introduces the data structures, constants and symbols needed; - implements the core logic of the scheduling algorithm in the new scheduling class file; - provides all the glue code between the new scheduling class and the core scheduler and refines the interactions between sched/dl and the other existing scheduling classes. Signed-off-by: Dario Faggioli <raistlin@linux.it> Signed-off-by: Michael Trimarchi <michael@amarulasolutions.com> Signed-off-by: Fabio Checconi <fchecconi@gmail.com> Signed-off-by: Juri Lelli <juri.lelli@gmail.com> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1383831828-15501-4-git-send-email-juri.lelli@gmail.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-11-28 11:14:43 +01:00
sched/deadline: Add SCHED_DEADLINE structures & implementation Introduces the data structures, constants and symbols needed for SCHED_DEADLINE implementation. Core data structure of SCHED_DEADLINE are defined, along with their initializers. Hooks for checking if a task belong to the new policy are also added where they are needed. Adds a scheduling class, in sched/dl.c and a new policy called SCHED_DEADLINE. It is an implementation of the Earliest Deadline First (EDF) scheduling algorithm, augmented with a mechanism (called Constant Bandwidth Server, CBS) that makes it possible to isolate the behaviour of tasks between each other. The typical -deadline task will be made up of a computation phase (instance) which is activated on a periodic or sporadic fashion. The expected (maximum) duration of such computation is called the task's runtime; the time interval by which each instance need to be completed is called the task's relative deadline. The task's absolute deadline is dynamically calculated as the time instant a task (better, an instance) activates plus the relative deadline. The EDF algorithms selects the task with the smallest absolute deadline as the one to be executed first, while the CBS ensures each task to run for at most its runtime every (relative) deadline length time interval, avoiding any interference between different tasks (bandwidth isolation). Thanks to this feature, also tasks that do not strictly comply with the computational model sketched above can effectively use the new policy. To summarize, this patch: - introduces the data structures, constants and symbols needed; - implements the core logic of the scheduling algorithm in the new scheduling class file; - provides all the glue code between the new scheduling class and the core scheduler and refines the interactions between sched/dl and the other existing scheduling classes. Signed-off-by: Dario Faggioli <raistlin@linux.it> Signed-off-by: Michael Trimarchi <michael@amarulasolutions.com> Signed-off-by: Fabio Checconi <fchecconi@gmail.com> Signed-off-by: Juri Lelli <juri.lelli@gmail.com> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1383831828-15501-4-git-send-email-juri.lelli@gmail.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-11-28 11:14:43 +01:00
sched/deadline: Add SCHED_DEADLINE structures & implementation Introduces the data structures, constants and symbols needed for SCHED_DEADLINE implementation. Core data structure of SCHED_DEADLINE are defined, along with their initializers. Hooks for checking if a task belong to the new policy are also added where they are needed. Adds a scheduling class, in sched/dl.c and a new policy called SCHED_DEADLINE. It is an implementation of the Earliest Deadline First (EDF) scheduling algorithm, augmented with a mechanism (called Constant Bandwidth Server, CBS) that makes it possible to isolate the behaviour of tasks between each other. The typical -deadline task will be made up of a computation phase (instance) which is activated on a periodic or sporadic fashion. The expected (maximum) duration of such computation is called the task's runtime; the time interval by which each instance need to be completed is called the task's relative deadline. The task's absolute deadline is dynamically calculated as the time instant a task (better, an instance) activates plus the relative deadline. The EDF algorithms selects the task with the smallest absolute deadline as the one to be executed first, while the CBS ensures each task to run for at most its runtime every (relative) deadline length time interval, avoiding any interference between different tasks (bandwidth isolation). Thanks to this feature, also tasks that do not strictly comply with the computational model sketched above can effectively use the new policy. To summarize, this patch: - introduces the data structures, constants and symbols needed; - implements the core logic of the scheduling algorithm in the new scheduling class file; - provides all the glue code between the new scheduling class and the core scheduler and refines the interactions between sched/dl and the other existing scheduling classes. Signed-off-by: Dario Faggioli <raistlin@linux.it> Signed-off-by: Michael Trimarchi <michael@amarulasolutions.com> Signed-off-by: Fabio Checconi <fchecconi@gmail.com> Signed-off-by: Juri Lelli <juri.lelli@gmail.com> Signed-off-by: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/1383831828-15501-4-git-send-email-juri.lelli@gmail.com Signed-off-by: Ingo Molnar <mingo@kernel.org>
2013-11-28 11:14:43 +01:00
[PATCH] sched: implement smpnice Problem: The introduction of separate run queues per CPU has brought with it "nice" enforcement problems that are best described by a simple example. For the sake of argument suppose that on a single CPU machine with a nice==19 hard spinner and a nice==0 hard spinner running that the nice==0 task gets 95% of the CPU and the nice==19 task gets 5% of the CPU. Now suppose that there is a system with 2 CPUs and 2 nice==19 hard spinners and 2 nice==0 hard spinners running. The user of this system would be entitled to expect that the nice==0 tasks each get 95% of a CPU and the nice==19 tasks only get 5% each. However, whether this expectation is met is pretty much down to luck as there are four equally likely distributions of the tasks to the CPUs that the load balancing code will consider to be balanced with loads of 2.0 for each CPU. Two of these distributions involve one nice==0 and one nice==19 task per CPU and in these circumstances the users expectations will be met. The other two distributions both involve both nice==0 tasks being on one CPU and both nice==19 being on the other CPU and each task will get 50% of a CPU and the user's expectations will not be met. Solution: The solution to this problem that is implemented in the attached patch is to use weighted loads when determining if the system is balanced and, when an imbalance is detected, to move an amount of weighted load between run queues (as opposed to a number of tasks) to restore the balance. Once again, the easiest way to explain why both of these measures are necessary is to use a simple example. Suppose that (in a slight variation of the above example) that we have a two CPU system with 4 nice==0 and 4 nice=19 hard spinning tasks running and that the 4 nice==0 tasks are on one CPU and the 4 nice==19 tasks are on the other CPU. The weighted loads for the two CPUs would be 4.0 and 0.2 respectively and the load balancing code would move 2 tasks resulting in one CPU with a load of 2.0 and the other with load of 2.2. If this was considered to be a big enough imbalance to justify moving a task and that task was moved using the current move_tasks() then it would move the highest priority task that it found and this would result in one CPU with a load of 3.0 and the other with a load of 1.2 which would result in the movement of a task in the opposite direction and so on -- infinite loop. If, on the other hand, an amount of load to be moved is calculated from the imbalance (in this case 0.1) and move_tasks() skips tasks until it find ones whose contributions to the weighted load are less than this amount it would move two of the nice==19 tasks resulting in a system with 2 nice==0 and 2 nice=19 on each CPU with loads of 2.1 for each CPU. One of the advantages of this mechanism is that on a system where all tasks have nice==0 the load balancing calculations would be mathematically identical to the current load balancing code. Notes: struct task_struct: has a new field load_weight which (in a trade off of space for speed) stores the contribution that this task makes to a CPU's weighted load when it is runnable. struct runqueue: has a new field raw_weighted_load which is the sum of the load_weight values for the currently runnable tasks on this run queue. This field always needs to be updated when nr_running is updated so two new inline functions inc_nr_running() and dec_nr_running() have been created to make sure that this happens. This also offers a convenient way to optimize away this part of the smpnice mechanism when CONFIG_SMP is not defined. int try_to_wake_up(): in this function the value SCHED_LOAD_BALANCE is used to represent the load contribution of a single task in various calculations in the code that decides which CPU to put the waking task on. While this would be a valid on a system where the nice values for the runnable tasks were distributed evenly around zero it will lead to anomalous load balancing if the distribution is skewed in either direction. To overcome this problem SCHED_LOAD_SCALE has been replaced by the load_weight for the relevant task or by the average load_weight per task for the queue in question (as appropriate). int move_tasks(): The modifications to this function were complicated by the fact that active_load_balance() uses it to move exactly one task without checking whether an imbalance actually exists. This precluded the simple overloading of max_nr_move with max_load_move and necessitated the addition of the latter as an extra argument to the function. The internal implementation is then modified to move up to max_nr_move tasks and max_load_move of weighted load. This slightly complicates the code where move_tasks() is called and if ever active_load_balance() is changed to not use move_tasks() the implementation of move_tasks() should be simplified accordingly. struct sched_group *find_busiest_group(): Similar to try_to_wake_up(), there are places in this function where SCHED_LOAD_SCALE is used to represent the load contribution of a single task and the same issues are created. A similar solution is adopted except that it is now the average per task contribution to a group's load (as opposed to a run queue) that is required. As this value is not directly available from the group it is calculated on the fly as the queues in the groups are visited when determining the busiest group. A key change to this function is that it is no longer to scale down *imbalance on exit as move_tasks() uses the load in its scaled form. void set_user_nice(): has been modified to update the task's load_weight field when it's nice value and also to ensure that its run queue's raw_weighted_load field is updated if it was runnable. From: "Siddha, Suresh B" <suresh.b.siddha@intel.com> With smpnice, sched groups with highest priority tasks can mask the imbalance between the other sched groups with in the same domain. This patch fixes some of the listed down scenarios by not considering the sched groups which are lightly loaded. a) on a simple 4-way MP system, if we have one high priority and 4 normal priority tasks, with smpnice we would like to see the high priority task scheduled on one cpu, two other cpus getting one normal task each and the fourth cpu getting the remaining two normal tasks. but with current smpnice extra normal priority task keeps jumping from one cpu to another cpu having the normal priority task. This is because of the busiest_has_loaded_cpus, nr_loaded_cpus logic.. We are not including the cpu with high priority task in max_load calculations but including that in total and avg_load calcuations.. leading to max_load < avg_load and load balance between cpus running normal priority tasks(2 Vs 1) will always show imbalanace as one normal priority and the extra normal priority task will keep moving from one cpu to another cpu having normal priority task.. b) 4-way system with HT (8 logical processors). Package-P0 T0 has a highest priority task, T1 is idle. Package-P1 Both T0 and T1 have 1 normal priority task each.. P2 and P3 are idle. With this patch, one of the normal priority tasks on P1 will be moved to P2 or P3.. c) With the current weighted smp nice calculations, it doesn't always make sense to look at the highest weighted runqueue in the busy group.. Consider a load balance scenario on a DP with HT system, with Package-0 containing one high priority and one low priority, Package-1 containing one low priority(with other thread being idle).. Package-1 thinks that it need to take the low priority thread from Package-0. And find_busiest_queue() returns the cpu thread with highest priority task.. And ultimately(with help of active load balance) we move high priority task to Package-1. And same continues with Package-0 now, moving high priority task from package-1 to package-0.. Even without the presence of active load balance, load balance will fail to balance the above scenario.. Fix find_busiest_queue to use "imbalance" when it is lightly loaded. [kernel@kolivas.org: sched: store weighted load on up] [kernel@kolivas.org: sched: add discrete weighted cpu load function] [suresh.b.siddha@intel.com: sched: remove dead code] Signed-off-by: Peter Williams <pwil3058@bigpond.com.au> Cc: "Siddha, Suresh B" <suresh.b.siddha@intel.com> Cc: "Chen, Kenneth W" <kenneth.w.chen@intel.com> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Signed-off-by: Con Kolivas <kernel@kolivas.org> Cc: John Hawkes <hawkes@sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 02:54:34 -07:00
sched/fair: Use a recently used CPU as an idle candidate and the basis for SIS The select_idle_sibling() (SIS) rewrite in commit: 10e2f1acd010 ("sched/core: Rewrite and improve select_idle_siblings()") ... replaced a domain iteration with a search that broadly speaking does a wrapped walk of the scheduler domain sharing a last-level-cache. While this had a number of improvements, one consequence is that two tasks that share a waker/wakee relationship push each other around a socket. Even though two tasks may be active, all cores are evenly used. This is great from a search perspective and spreads a load across individual cores, but it has adverse consequences for cpufreq. As each CPU has relatively low utilisation, cpufreq may decide the utilisation is too low to used a higher P-state and overall computation throughput suffers. While individual cpufreq and cpuidle drivers may compensate by artifically boosting P-state (at c0) or avoiding lower C-states (during idle), it does not help if hardware-based cpufreq (e.g. HWP) is used. This patch tracks a recently used CPU based on what CPU a task was running on when it last was a waker a CPU it was recently using when a task is a wakee. During SIS, the recently used CPU is used as a target if it's still allowed by the task and is idle. The benefit may be non-obvious so consider an example of two tasks communicating back and forth. Task A may be an application doing IO where task B is a kworker or kthread like journald. Task A may issue IO, wake B and B wakes up A on completion. With the existing scheme this may look like the following (potentially different IDs if SMT is in use but similar principal applies). A (cpu 0) wake B (wakes on cpu 1) B (cpu 1) wake A (wakes on cpu 2) A (cpu 2) wake B (wakes on cpu 3) etc. A careful reader may wonder why CPU 0 was not idle when B wakes A the first time and it's simply due to the fact that A can be rescheduled to another CPU and the pattern is that prev == target when B tries to wakeup A and the information about CPU 0 has been lost. With this patch, the pattern is more likely to be: A (cpu 0) wake B (wakes on cpu 1) B (cpu 1) wake A (wakes on cpu 0) A (cpu 0) wake B (wakes on cpu 1) etc i.e. two communicating casts are more likely to use just two cores instead of all available cores sharing a LLC. The most dramatic speedup was noticed on dbench using the XFS filesystem on UMA as clients interact heavily with workqueues in that configuration. Note that a similar speedup is not observed on ext4 as the wakeup pattern is different: 4.15.0-rc9 4.15.0-rc9 waprev-v1 biasancestor-v1 Hmean 1 287.54 ( 0.00%) 817.01 ( 184.14%) Hmean 2 1268.12 ( 0.00%) 1781.24 ( 40.46%) Hmean 4 1739.68 ( 0.00%) 1594.47 ( -8.35%) Hmean 8 2464.12 ( 0.00%) 2479.56 ( 0.63%) Hmean 64 1455.57 ( 0.00%) 1434.68 ( -1.44%) The results can be less dramatic on NUMA where automatic balancing interferes with the test. It's also known that network benchmarks running on localhost also benefit quite a bit from this patch (roughly 10% on netperf RR for UDP and TCP depending on the machine). Hackbench also seens small improvements (6-11% depending on machine and thread count). The facebook schbench was also tested but in most cases showed little or no different to wakeup latencies. Signed-off-by: Mel Gorman <mgorman@techsingularity.net> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Matt Fleming <matt@codeblueprint.co.uk> Cc: Mike Galbraith <efault@gmx.de> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Link: http://lkml.kernel.org/r/20180130104555.4125-5-mgorman@techsingularity.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2018-01-30 10:45:55 +00:00
[PATCH] sched: implement smpnice Problem: The introduction of separate run queues per CPU has brought with it "nice" enforcement problems that are best described by a simple example. For the sake of argument suppose that on a single CPU machine with a nice==19 hard spinner and a nice==0 hard spinner running that the nice==0 task gets 95% of the CPU and the nice==19 task gets 5% of the CPU. Now suppose that there is a system with 2 CPUs and 2 nice==19 hard spinners and 2 nice==0 hard spinners running. The user of this system would be entitled to expect that the nice==0 tasks each get 95% of a CPU and the nice==19 tasks only get 5% each. However, whether this expectation is met is pretty much down to luck as there are four equally likely distributions of the tasks to the CPUs that the load balancing code will consider to be balanced with loads of 2.0 for each CPU. Two of these distributions involve one nice==0 and one nice==19 task per CPU and in these circumstances the users expectations will be met. The other two distributions both involve both nice==0 tasks being on one CPU and both nice==19 being on the other CPU and each task will get 50% of a CPU and the user's expectations will not be met. Solution: The solution to this problem that is implemented in the attached patch is to use weighted loads when determining if the system is balanced and, when an imbalance is detected, to move an amount of weighted load between run queues (as opposed to a number of tasks) to restore the balance. Once again, the easiest way to explain why both of these measures are necessary is to use a simple example. Suppose that (in a slight variation of the above example) that we have a two CPU system with 4 nice==0 and 4 nice=19 hard spinning tasks running and that the 4 nice==0 tasks are on one CPU and the 4 nice==19 tasks are on the other CPU. The weighted loads for the two CPUs would be 4.0 and 0.2 respectively and the load balancing code would move 2 tasks resulting in one CPU with a load of 2.0 and the other with load of 2.2. If this was considered to be a big enough imbalance to justify moving a task and that task was moved using the current move_tasks() then it would move the highest priority task that it found and this would result in one CPU with a load of 3.0 and the other with a load of 1.2 which would result in the movement of a task in the opposite direction and so on -- infinite loop. If, on the other hand, an amount of load to be moved is calculated from the imbalance (in this case 0.1) and move_tasks() skips tasks until it find ones whose contributions to the weighted load are less than this amount it would move two of the nice==19 tasks resulting in a system with 2 nice==0 and 2 nice=19 on each CPU with loads of 2.1 for each CPU. One of the advantages of this mechanism is that on a system where all tasks have nice==0 the load balancing calculations would be mathematically identical to the current load balancing code. Notes: struct task_struct: has a new field load_weight which (in a trade off of space for speed) stores the contribution that this task makes to a CPU's weighted load when it is runnable. struct runqueue: has a new field raw_weighted_load which is the sum of the load_weight values for the currently runnable tasks on this run queue. This field always needs to be updated when nr_running is updated so two new inline functions inc_nr_running() and dec_nr_running() have been created to make sure that this happens. This also offers a convenient way to optimize away this part of the smpnice mechanism when CONFIG_SMP is not defined. int try_to_wake_up(): in this function the value SCHED_LOAD_BALANCE is used to represent the load contribution of a single task in various calculations in the code that decides which CPU to put the waking task on. While this would be a valid on a system where the nice values for the runnable tasks were distributed evenly around zero it will lead to anomalous load balancing if the distribution is skewed in either direction. To overcome this problem SCHED_LOAD_SCALE has been replaced by the load_weight for the relevant task or by the average load_weight per task for the queue in question (as appropriate). int move_tasks(): The modifications to this function were complicated by the fact that active_load_balance() uses it to move exactly one task without checking whether an imbalance actually exists. This precluded the simple overloading of max_nr_move with max_load_move and necessitated the addition of the latter as an extra argument to the function. The internal implementation is then modified to move up to max_nr_move tasks and max_load_move of weighted load. This slightly complicates the code where move_tasks() is called and if ever active_load_balance() is changed to not use move_tasks() the implementation of move_tasks() should be simplified accordingly. struct sched_group *find_busiest_group(): Similar to try_to_wake_up(), there are places in this function where SCHED_LOAD_SCALE is used to represent the load contribution of a single task and the same issues are created. A similar solution is adopted except that it is now the average per task contribution to a group's load (as opposed to a run queue) that is required. As this value is not directly available from the group it is calculated on the fly as the queues in the groups are visited when determining the busiest group. A key change to this function is that it is no longer to scale down *imbalance on exit as move_tasks() uses the load in its scaled form. void set_user_nice(): has been modified to update the task's load_weight field when it's nice value and also to ensure that its run queue's raw_weighted_load field is updated if it was runnable. From: "Siddha, Suresh B" <suresh.b.siddha@intel.com> With smpnice, sched groups with highest priority tasks can mask the imbalance between the other sched groups with in the same domain. This patch fixes some of the listed down scenarios by not considering the sched groups which are lightly loaded. a) on a simple 4-way MP system, if we have one high priority and 4 normal priority tasks, with smpnice we would like to see the high priority task scheduled on one cpu, two other cpus getting one normal task each and the fourth cpu getting the remaining two normal tasks. but with current smpnice extra normal priority task keeps jumping from one cpu to another cpu having the normal priority task. This is because of the busiest_has_loaded_cpus, nr_loaded_cpus logic.. We are not including the cpu with high priority task in max_load calculations but including that in total and avg_load calcuations.. leading to max_load < avg_load and load balance between cpus running normal priority tasks(2 Vs 1) will always show imbalanace as one normal priority and the extra normal priority task will keep moving from one cpu to another cpu having normal priority task.. b) 4-way system with HT (8 logical processors). Package-P0 T0 has a highest priority task, T1 is idle. Package-P1 Both T0 and T1 have 1 normal priority task each.. P2 and P3 are idle. With this patch, one of the normal priority tasks on P1 will be moved to P2 or P3.. c) With the current weighted smp nice calculations, it doesn't always make sense to look at the highest weighted runqueue in the busy group.. Consider a load balance scenario on a DP with HT system, with Package-0 containing one high priority and one low priority, Package-1 containing one low priority(with other thread being idle).. Package-1 thinks that it need to take the low priority thread from Package-0. And find_busiest_queue() returns the cpu thread with highest priority task.. And ultimately(with help of active load balance) we move high priority task to Package-1. And same continues with Package-0 now, moving high priority task from package-1 to package-0.. Even without the presence of active load balance, load balance will fail to balance the above scenario.. Fix find_busiest_queue to use "imbalance" when it is lightly loaded. [kernel@kolivas.org: sched: store weighted load on up] [kernel@kolivas.org: sched: add discrete weighted cpu load function] [suresh.b.siddha@intel.com: sched: remove dead code] Signed-off-by: Peter Williams <pwil3058@bigpond.com.au> Cc: "Siddha, Suresh B" <suresh.b.siddha@intel.com> Cc: "Chen, Kenneth W" <kenneth.w.chen@intel.com> Acked-by: Ingo Molnar <mingo@elte.hu> Cc: Nick Piggin <nickpiggin@yahoo.com.au> Signed-off-by: Con Kolivas <kernel@kolivas.org> Cc: John Hawkes <hawkes@sgi.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-06-27 02:54:34 -07:00
rcu: Add a TINY_PREEMPT_RCU Implement a small-memory-footprint uniprocessor-only implementation of preemptible RCU. This implementation uses but a single blocked-tasks list rather than the combinatorial number used per leaf rcu_node by TREE_PREEMPT_RCU, which reduces memory consumption and greatly simplifies processing. This version also takes advantage of uniprocessor execution to accelerate grace periods in the case where there are no readers. The general design is otherwise broadly similar to that of TREE_PREEMPT_RCU. This implementation is a step towards having RCU implementation driven off of the SMP and PREEMPT kernel configuration variables, which can happen once this implementation has accumulated sufficient experience. Removed ACCESS_ONCE() from __rcu_read_unlock() and added barrier() as suggested by Steve Rostedt in order to avoid the compiler-reordering issue noted by Mathieu Desnoyers (http://lkml.org/lkml/2010/8/16/183). As can be seen below, CONFIG_TINY_PREEMPT_RCU represents almost 5Kbyte savings compared to CONFIG_TREE_PREEMPT_RCU. Of course, for non-real-time workloads, CONFIG_TINY_RCU is even better. CONFIG_TREE_PREEMPT_RCU text data bss dec filename 13 0 0 13 kernel/rcupdate.o 6170 825 28 7023 kernel/rcutree.o ---- 7026 Total CONFIG_TINY_PREEMPT_RCU text data bss dec filename 13 0 0 13 kernel/rcupdate.o 2081 81 8 2170 kernel/rcutiny.o ---- 2183 Total CONFIG_TINY_RCU (non-preemptible) text data bss dec filename 13 0 0 13 kernel/rcupdate.o 719 25 0 744 kernel/rcutiny.o --- 757 Total Requested-by: Loïc Minier <loic.minier@canonical.com> Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
2010-06-29 16:49:16 -07:00
signal: consolidate {TS,TLF}_RESTORE_SIGMASK code In general, there's no need for the "restore sigmask" flag to live in ti->flags. alpha, ia64, microblaze, powerpc, sh, sparc (64-bit only), tile, and x86 use essentially identical alternative implementations, placing the flag in ti->status. Replace those optimized implementations with an equally good common implementation that stores it in a bitfield in struct task_struct and drop the custom implementations. Additional architectures can opt in by removing their TIF_RESTORE_SIGMASK defines. Link: http://lkml.kernel.org/r/8a14321d64a28e40adfddc90e18a96c086a6d6f9.1468522723.git.luto@kernel.org Signed-off-by: Andy Lutomirski <luto@kernel.org> Tested-by: Michael Ellerman <mpe@ellerman.id.au> [powerpc] Cc: Richard Henderson <rth@twiddle.net> Cc: Ivan Kokshaysky <ink@jurassic.park.msu.ru> Cc: Matt Turner <mattst88@gmail.com> Cc: Tony Luck <tony.luck@intel.com> Cc: Fenghua Yu <fenghua.yu@intel.com> Cc: Michal Simek <monstr@monstr.eu> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Cc: Yoshinori Sato <ysato@users.sourceforge.jp> Cc: Rich Felker <dalias@libc.org> Cc: "David S. Miller" <davem@davemloft.net> Cc: Chris Metcalf <cmetcalf@mellanox.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Borislav Petkov <bp@suse.de> Cc: Brian Gerst <brgerst@gmail.com> Cc: Dmitry Safonov <dsafonov@virtuozzo.com> Cc: Oleg Nesterov <oleg@redhat.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-08-02 14:05:36 -07:00
cgroup, kthread: close race window where new kthreads can be migrated to non-root cgroups Creation of a kthread goes through a couple interlocked stages between the kthread itself and its creator. Once the new kthread starts running, it initializes itself and wakes up the creator. The creator then can further configure the kthread and then let it start doing its job by waking it up. In this configuration-by-creator stage, the creator is the only one that can wake it up but the kthread is visible to userland. When altering the kthread's attributes from userland is allowed, this is fine; however, for cases where CPU affinity is critical, kthread_bind() is used to first disable affinity changes from userland and then set the affinity. This also prevents the kthread from being migrated into non-root cgroups as that can affect the CPU affinity and many other things. Unfortunately, the cgroup side of protection is racy. While the PF_NO_SETAFFINITY flag prevents further migrations, userland can win the race before the creator sets the flag with kthread_bind() and put the kthread in a non-root cgroup, which can lead to all sorts of problems including incorrect CPU affinity and starvation. This bug got triggered by userland which periodically tries to migrate all processes in the root cpuset cgroup to a non-root one. Per-cpu workqueue workers got caught while being created and ended up with incorrected CPU affinity breaking concurrency management and sometimes stalling workqueue execution. This patch adds task->no_cgroup_migration which disallows the task to be migrated by userland. kthreadd starts with the flag set making every child kthread start in the root cgroup with migration disallowed. The flag is cleared after the kthread finishes initialization by which time PF_NO_SETAFFINITY is set if the kthread should stay in the root cgroup. It'd be better to wait for the initialization instead of failing but I couldn't think of a way of implementing that without adding either a new PF flag, or sleeping and retrying from waiting side. Even if userland depends on changing cgroup membership of a kthread, it either has to be synchronized with kthread_create() or periodically repeat, so it's unlikely that this would break anything. v2: Switch to a simpler implementation using a new task_struct bit field suggested by Oleg. Signed-off-by: Tejun Heo <tj@kernel.org> Suggested-by: Oleg Nesterov <oleg@redhat.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Thomas Gleixner <tglx@linutronix.de> Reported-and-debugged-by: Chris Mason <clm@fb.com> Cc: stable@vger.kernel.org # v4.3+ (we can't close the race on < v4.3) Signed-off-by: Tejun Heo <tj@kernel.org>
2017-03-16 16:54:24 -04:00
all arches, signal: move restart_block to struct task_struct If an attacker can cause a controlled kernel stack overflow, overwriting the restart block is a very juicy exploit target. This is because the restart_block is held in the same memory allocation as the kernel stack. Moving the restart block to struct task_struct prevents this exploit by making the restart_block harder to locate. Note that there are other fields in thread_info that are also easy targets, at least on some architectures. It's also a decent simplification, since the restart code is more or less identical on all architectures. [james.hogan@imgtec.com: metag: align thread_info::supervisor_stack] Signed-off-by: Andy Lutomirski <luto@amacapital.net> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Al Viro <viro@zeniv.linux.org.uk> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Kees Cook <keescook@chromium.org> Cc: David Miller <davem@davemloft.net> Acked-by: Richard Weinberger <richard@nod.at> Cc: Richard Henderson <rth@twiddle.net> Cc: Ivan Kokshaysky <ink@jurassic.park.msu.ru> Cc: Matt Turner <mattst88@gmail.com> Cc: Vineet Gupta <vgupta@synopsys.com> Cc: Russell King <rmk@arm.linux.org.uk> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Will Deacon <will.deacon@arm.com> Cc: Haavard Skinnemoen <hskinnemoen@gmail.com> Cc: Hans-Christian Egtvedt <egtvedt@samfundet.no> Cc: Steven Miao <realmz6@gmail.com> Cc: Mark Salter <msalter@redhat.com> Cc: Aurelien Jacquiot <a-jacquiot@ti.com> Cc: Mikael Starvik <starvik@axis.com> Cc: Jesper Nilsson <jesper.nilsson@axis.com> Cc: David Howells <dhowells@redhat.com> Cc: Richard Kuo <rkuo@codeaurora.org> Cc: "Luck, Tony" <tony.luck@intel.com> Cc: Geert Uytterhoeven <geert@linux-m68k.org> Cc: Michal Simek <monstr@monstr.eu> Cc: Ralf Baechle <ralf@linux-mips.org> Cc: Jonas Bonn <jonas@southpole.se> Cc: "James E.J. Bottomley" <jejb@parisc-linux.org> Cc: Helge Deller <deller@gmx.de> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: Paul Mackerras <paulus@samba.org> Acked-by: Michael Ellerman <mpe@ellerman.id.au> (powerpc) Tested-by: Michael Ellerman <mpe@ellerman.id.au> (powerpc) Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: Heiko Carstens <heiko.carstens@de.ibm.com> Cc: Chen Liqin <liqin.linux@gmail.com> Cc: Lennox Wu <lennox.wu@gmail.com> Cc: Chris Metcalf <cmetcalf@ezchip.com> Cc: Guan Xuetao <gxt@mprc.pku.edu.cn> Cc: Chris Zankel <chris@zankel.net> Cc: Max Filippov <jcmvbkbc@gmail.com> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Guenter Roeck <linux@roeck-us.net> Signed-off-by: James Hogan <james.hogan@imgtec.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-12 15:01:14 -08:00
Kbuild: rename CC_STACKPROTECTOR[_STRONG] config variables The changes to automatically test for working stack protector compiler support in the Kconfig files removed the special STACKPROTECTOR_AUTO option that picked the strongest stack protector that the compiler supported. That was all a nice cleanup - it makes no sense to have the AUTO case now that the Kconfig phase can just determine the compiler support directly. HOWEVER. It also meant that doing "make oldconfig" would now _disable_ the strong stackprotector if you had AUTO enabled, because in a legacy config file, the sane stack protector configuration would look like CONFIG_HAVE_CC_STACKPROTECTOR=y # CONFIG_CC_STACKPROTECTOR_NONE is not set # CONFIG_CC_STACKPROTECTOR_REGULAR is not set # CONFIG_CC_STACKPROTECTOR_STRONG is not set CONFIG_CC_STACKPROTECTOR_AUTO=y and when you ran this through "make oldconfig" with the Kbuild changes, it would ask you about the regular CONFIG_CC_STACKPROTECTOR (that had been renamed from CONFIG_CC_STACKPROTECTOR_REGULAR to just CONFIG_CC_STACKPROTECTOR), but it would think that the STRONG version used to be disabled (because it was really enabled by AUTO), and would disable it in the new config, resulting in: CONFIG_HAVE_CC_STACKPROTECTOR=y CONFIG_CC_HAS_STACKPROTECTOR_NONE=y CONFIG_CC_STACKPROTECTOR=y # CONFIG_CC_STACKPROTECTOR_STRONG is not set CONFIG_CC_HAS_SANE_STACKPROTECTOR=y That's dangerously subtle - people could suddenly find themselves with the weaker stack protector setup without even realizing. The solution here is to just rename not just the old RECULAR stack protector option, but also the strong one. This does that by just removing the CC_ prefix entirely for the user choices, because it really is not about the compiler support (the compiler support now instead automatially impacts _visibility_ of the options to users). This results in "make oldconfig" actually asking the user for their choice, so that we don't have any silent subtle security model changes. The end result would generally look like this: CONFIG_HAVE_CC_STACKPROTECTOR=y CONFIG_CC_HAS_STACKPROTECTOR_NONE=y CONFIG_STACKPROTECTOR=y CONFIG_STACKPROTECTOR_STRONG=y CONFIG_CC_HAS_SANE_STACKPROTECTOR=y where the "CC_" versions really are about internal compiler infrastructure, not the user selections. Acked-by: Masahiro Yamada <yamada.masahiro@socionext.com> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2018-06-14 12:21:18 +09:00
nohz: New tick dependency mask The tick dependency is evaluated on every IRQ and context switch. This consists is a batch of checks which determine whether it is safe to stop the tick or not. These checks are often split in many details: posix cpu timers, scheduler, sched clock, perf events.... each of which are made of smaller details: posix cpu timer involves checking process wide timers then thread wide timers. Perf involves checking freq events then more per cpu details. Checking these informations asynchronously every time we update the full dynticks state bring avoidable overhead and a messy layout. Let's introduce instead tick dependency masks: one for system wide dependency (unstable sched clock, freq based perf events), one for CPU wide dependency (sched, throttling perf events), and task/signal level dependencies (posix cpu timers). The subsystems are responsible for setting and clearing their dependency through a set of APIs that will take care of concurrent dependency mask modifications and kick targets to restart the relevant CPU tick whenever needed. This new dependency engine stays beside the old one until all subsystems having a tick dependency are converted to it. Suggested-by: Thomas Gleixner <tglx@linutronix.de> Suggested-by: Peter Zijlstra <peterz@infradead.org> Reviewed-by: Chris Metcalf <cmetcalf@ezchip.com> Cc: Christoph Lameter <cl@linux.com> Cc: Chris Metcalf <cmetcalf@ezchip.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Luiz Capitulino <lcapitulino@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Viresh Kumar <viresh.kumar@linaro.org> Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
2015-06-07 15:54:30 +02:00
nohz: New tick dependency mask The tick dependency is evaluated on every IRQ and context switch. This consists is a batch of checks which determine whether it is safe to stop the tick or not. These checks are often split in many details: posix cpu timers, scheduler, sched clock, perf events.... each of which are made of smaller details: posix cpu timer involves checking process wide timers then thread wide timers. Perf involves checking freq events then more per cpu details. Checking these informations asynchronously every time we update the full dynticks state bring avoidable overhead and a messy layout. Let's introduce instead tick dependency masks: one for system wide dependency (unstable sched clock, freq based perf events), one for CPU wide dependency (sched, throttling perf events), and task/signal level dependencies (posix cpu timers). The subsystems are responsible for setting and clearing their dependency through a set of APIs that will take care of concurrent dependency mask modifications and kick targets to restart the relevant CPU tick whenever needed. This new dependency engine stays beside the old one until all subsystems having a tick dependency are converted to it. Suggested-by: Thomas Gleixner <tglx@linutronix.de> Suggested-by: Peter Zijlstra <peterz@infradead.org> Reviewed-by: Chris Metcalf <cmetcalf@ezchip.com> Cc: Christoph Lameter <cl@linux.com> Cc: Chris Metcalf <cmetcalf@ezchip.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Luiz Capitulino <lcapitulino@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Rik van Riel <riel@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Viresh Kumar <viresh.kumar@linaro.org> Signed-off-by: Frederic Weisbecker <fweisbec@gmail.com>
2015-06-07 15:54:30 +02:00
softlockup: automatically detect hung TASK_UNINTERRUPTIBLE tasks this patch extends the soft-lockup detector to automatically detect hung TASK_UNINTERRUPTIBLE tasks. Such hung tasks are printed the following way: ------------------> INFO: task prctl:3042 blocked for more than 120 seconds. "echo 0 > /proc/sys/kernel/hung_task_timeout_secs" disables this message prctl D fd5e3793 0 3042 2997 f6050f38 00000046 00000001 fd5e3793 00000009 c06d8264 c06dae80 00000286 f6050f40 f6050f00 f7d34d90 f7d34fc8 c1e1be80 00000001 f6050000 00000000 f7e92d00 00000286 f6050f18 c0489d1a f6050f40 00006605 00000000 c0133a5b Call Trace: [<c04883a5>] schedule_timeout+0x6d/0x8b [<c04883d8>] schedule_timeout_uninterruptible+0x15/0x17 [<c0133a76>] msleep+0x10/0x16 [<c0138974>] sys_prctl+0x30/0x1e2 [<c0104c52>] sysenter_past_esp+0x5f/0xa5 ======================= 2 locks held by prctl/3042: #0: (&sb->s_type->i_mutex_key#5){--..}, at: [<c0197d11>] do_fsync+0x38/0x7a #1: (jbd_handle){--..}, at: [<c01ca3d2>] journal_start+0xc7/0xe9 <------------------ the current default timeout is 120 seconds. Such messages are printed up to 10 times per bootup. If the system has crashed already then the messages are not printed. if lockdep is enabled then all held locks are printed as well. this feature is a natural extension to the softlockup-detector (kernel locked up without scheduling) and to the NMI watchdog (kernel locked up with IRQs disabled). [ Gautham R Shenoy <ego@in.ibm.com>: CPU hotplug fixes. ] [ Andrew Morton <akpm@linux-foundation.org>: build warning fix. ] Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com>
2008-01-25 21:08:02 +01:00
task_work_add: generic process-context callbacks Provide a simple mechanism that allows running code in the (nonatomic) context of the arbitrary task. The caller does task_work_add(task, task_work) and this task executes task_work->func() either from do_notify_resume() or from do_exit(). The callback can rely on PF_EXITING to detect the latter case. "struct task_work" can be embedded in another struct, still it has "void *data" to handle the most common/simple case. This allows us to kill the ->replacement_session_keyring hack, and potentially this can have more users. Performance-wise, this adds 2 "unlikely(!hlist_empty())" checks into tracehook_notify_resume() and do_exit(). But at the same time we can remove the "replacement_session_keyring != NULL" checks from arch/*/signal.c and exit_creds(). Note: task_work_add/task_work_run abuses ->pi_lock. This is only because this lock is already used by lookup_pi_state() to synchronize with do_exit() setting PF_EXITING. Fortunately the scope of this lock in task_work.c is really tiny, and the code is unlikely anyway. Signed-off-by: Oleg Nesterov <oleg@redhat.com> Acked-by: David Howells <dhowells@redhat.com> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Richard Kuo <rkuo@codeaurora.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Alexander Gordeev <agordeev@redhat.com> Cc: Chris Zankel <chris@zankel.net> Cc: David Smith <dsmith@redhat.com> Cc: "Frank Ch. Eigler" <fche@redhat.com> Cc: Geert Uytterhoeven <geert@linux-m68k.org> Cc: Larry Woodman <lwoodman@redhat.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Tejun Heo <tj@kernel.org> Cc: Ingo Molnar <mingo@elte.hu> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
2012-05-11 10:59:07 +10:00
sched: Implement lockless wake-queues This is useful for locking primitives that can effect multiple wakeups per operation and want to avoid lock internal lock contention by delaying the wakeups until we've released the lock internal locks. Alternatively it can be used to avoid issuing multiple wakeups, and thus save a few cycles, in packet processing. Queue all target tasks and wakeup once you've processed all packets. That way you avoid waking the target task multiple times if there were multiple packets for the same task. Properties of a wake_q are: - Lockless, as queue head must reside on the stack. - Being a queue, maintains wakeup order passed by the callers. This can be important for otherwise, in scenarios where highly contended locks could affect any reliance on lock fairness. - A queued task cannot be added again until it is woken up. This patch adds the needed infrastructure into the scheduler code and uses the new wake_list to delay the futex wakeups until after we've released the hash bucket locks. Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> [tweaks, adjustments, comments, etc.] Signed-off-by: Davidlohr Bueso <dbueso@suse.de> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Thomas Gleixner <tglx@linutronix.de> Cc: Borislav Petkov <bp@alien8.de> Cc: Chris Mason <clm@fb.com> Cc: Davidlohr Bueso <dave@stgolabs.net> Cc: George Spelvin <linux@horizon.com> Cc: H. Peter Anvin <hpa@zytor.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Manfred Spraul <manfred@colorfullife.com> Cc: Sebastian Andrzej Siewior <bigeasy@linutronix.de> Cc: Steven Rostedt <rostedt@goodmis.org> Link: http://lkml.kernel.org/r/1430494072-30283-2-git-send-email-dave@stgolabs.net Signed-off-by: Ingo Molnar <mingo@kernel.org>
2015-05-01 08:27:50 -07:00
sched/rtmutex/deadline: Fix a PI crash for deadline tasks A crash happened while I was playing with deadline PI rtmutex. BUG: unable to handle kernel NULL pointer dereference at 0000000000000018 IP: [<ffffffff810eeb8f>] rt_mutex_get_top_task+0x1f/0x30 PGD 232a75067 PUD 230947067 PMD 0 Oops: 0000 [#1] SMP CPU: 1 PID: 10994 Comm: a.out Not tainted Call Trace: [<ffffffff810b658c>] enqueue_task+0x2c/0x80 [<ffffffff810ba763>] activate_task+0x23/0x30 [<ffffffff810d0ab5>] pull_dl_task+0x1d5/0x260 [<ffffffff810d0be6>] pre_schedule_dl+0x16/0x20 [<ffffffff8164e783>] __schedule+0xd3/0x900 [<ffffffff8164efd9>] schedule+0x29/0x70 [<ffffffff8165035b>] __rt_mutex_slowlock+0x4b/0xc0 [<ffffffff81650501>] rt_mutex_slowlock+0xd1/0x190 [<ffffffff810eeb33>] rt_mutex_timed_lock+0x53/0x60 [<ffffffff810ecbfc>] futex_lock_pi.isra.18+0x28c/0x390 [<ffffffff810ed8b0>] do_futex+0x190/0x5b0 [<ffffffff810edd50>] SyS_futex+0x80/0x180 This is because rt_mutex_enqueue_pi() and rt_mutex_dequeue_pi() are only protected by pi_lock when operating pi waiters, while rt_mutex_get_top_task(), will access them with rq lock held but not holding pi_lock. In order to tackle it, we introduce new "pi_top_task" pointer cached in task_struct, and add new rt_mutex_update_top_task() to update its value, it can be called by rt_mutex_setprio() which held both owner's pi_lock and rq lock. Thus "pi_top_task" can be safely accessed by enqueue_task_dl() under rq lock. Originally-From: Peter Zijlstra <peterz@infradead.org> Signed-off-by: Xunlei Pang <xlpang@redhat.com> Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org> Acked-by: Steven Rostedt <rostedt@goodmis.org> Reviewed-by: Thomas Gleixner <tglx@linutronix.de> Cc: juri.lelli@arm.com Cc: bigeasy@linutronix.de Cc: mathieu.desnoyers@efficios.com Cc: jdesfossez@efficios.com Cc: bristot@redhat.com Link: http://lkml.kernel.org/r/20170323150216.157682758@infradead.org Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
2017-03-23 15:56:08 +01:00
[PATCH] lockdep: core Do 'make oldconfig' and accept all the defaults for new config options - reboot into the kernel and if everything goes well it should boot up fine and you should have /proc/lockdep and /proc/lockdep_stats files. Typically if the lock validator finds some problem it will print out voluminous debug output that begins with "BUG: ..." and which syslog output can be used by kernel developers to figure out the precise locking scenario. What does the lock validator do? It "observes" and maps all locking rules as they occur dynamically (as triggered by the kernel's natural use of spinlocks, rwlocks, mutexes and rwsems). Whenever the lock validator subsystem detects a new locking scenario, it validates this new rule against the existing set of rules. If this new rule is consistent with the existing set of rules then the new rule is added transparently and the kernel continues as normal. If the new rule could create a deadlock scenario then this condition is printed out. When determining validity of locking, all possible "deadlock scenarios" are considered: assuming arbitrary number of CPUs, arbitrary irq context and task context constellations, running arbitrary combinations of all the existing locking scenarios. In a typical system this means millions of separate scenarios. This is why we call it a "locking correctness" validator - for all rules that are observed the lock validator proves it with mathematical certainty that a deadlock could not occur (assuming that the lock validator implementation itself is correct and its internal data structures are not corrupted by some other kernel subsystem). [see more details and conditionals of this statement in include/linux/lockdep.h and Documentation/lockdep-design.txt] Furthermore, this "all possible scenarios" property of the validator also enables the finding of complex, highly unlikely multi-CPU multi-context races via single single-context rules, increasing the likelyhood of finding bugs drastically. In practical terms: the lock validator already found a bug in the upstream kernel that could only occur on systems with 3 or more CPUs, and which needed 3 very unlikely code sequences to occur at once on the 3 CPUs. That bug was found and reported on a single-CPU system (!). So in essence a race will be found "piecemail-wise", triggering all the necessary components for the race, without having to reproduce the race scenario itself! In its short existence the lock validator found and reported many bugs before they actually caused a real deadlock. To further increase the efficiency of the validator, the mapping is not per "lock instance", but per "lock-class". For example, all struct inode objects in the kernel have inode->inotify_mutex. If there are 10,000 inodes cached, then there are 10,000 lock objects. But ->inotify_mutex is a single "lock type", and all locking activities that occur against ->inotify_mutex are "unified" into this single lock-class. The advantage of the lock-class approach is that all historical ->inotify_mutex uses are mapped into a single (and as narrow as possible) set of locking rules - regardless of how many different tasks or inode structures it took to build this set of rules. The set of rules persist during the lifetime of the kernel. To see the rough magnitude of checking that the lock validator does, here's a portion of /proc/lockdep_stats, fresh after bootup: lock-classes: 694 [max: 2048] direct dependencies: 1598 [max: 8192] indirect dependencies: 17896 all direct dependencies: 16206 dependency chains: 1910 [max: 8192] in-hardirq chains: 17 in-softirq chains: 105 in-process chains: 1065 stack-trace entries: 38761 [max: 131072] combined max dependencies: 2033928 hardirq-safe locks: 24 hardirq-unsafe locks: 176 softirq-safe locks: 53 softirq-unsafe locks: 137 irq-safe locks: 59 irq-unsafe locks: 176 The lock validator has observed 1598 actual single-thread locking patterns, and has validated all possible 2033928 distinct locking scenarios. More details about the design of the lock validator can be found in Documentation/lockdep-design.txt, which can also found at: http://redhat.com/~mingo/lockdep-patches/lockdep-design.txt [bunk@stusta.de: cleanups] Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Signed-off-by: Adrian Bunk <bunk@stusta.de> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-07-03 00:24:50 -07:00
[PATCH] lockdep: core Do 'make oldconfig' and accept all the defaults for new config options - reboot into the kernel and if everything goes well it should boot up fine and you should have /proc/lockdep and /proc/lockdep_stats files. Typically if the lock validator finds some problem it will print out voluminous debug output that begins with "BUG: ..." and which syslog output can be used by kernel developers to figure out the precise locking scenario. What does the lock validator do? It "observes" and maps all locking rules as they occur dynamically (as triggered by the kernel's natural use of spinlocks, rwlocks, mutexes and rwsems). Whenever the lock validator subsystem detects a new locking scenario, it validates this new rule against the existing set of rules. If this new rule is consistent with the existing set of rules then the new rule is added transparently and the kernel continues as normal. If the new rule could create a deadlock scenario then this condition is printed out. When determining validity of locking, all possible "deadlock scenarios" are considered: assuming arbitrary number of CPUs, arbitrary irq context and task context constellations, running arbitrary combinations of all the existing locking scenarios. In a typical system this means millions of separate scenarios. This is why we call it a "locking correctness" validator - for all rules that are observed the lock validator proves it with mathematical certainty that a deadlock could not occur (assuming that the lock validator implementation itself is correct and its internal data structures are not corrupted by some other kernel subsystem). [see more details and conditionals of this statement in include/linux/lockdep.h and Documentation/lockdep-design.txt] Furthermore, this "all possible scenarios" property of the validator also enables the finding of complex, highly unlikely multi-CPU multi-context races via single single-context rules, increasing the likelyhood of finding bugs drastically. In practical terms: the lock validator already found a bug in the upstream kernel that could only occur on systems with 3 or more CPUs, and which needed 3 very unlikely code sequences to occur at once on the 3 CPUs. That bug was found and reported on a single-CPU system (!). So in essence a race will be found "piecemail-wise", triggering all the necessary components for the race, without having to reproduce the race scenario itself! In its short existence the lock validator found and reported many bugs before they actually caused a real deadlock. To further increase the efficiency of the validator, the mapping is not per "lock instance", but per "lock-class". For example, all struct inode objects in the kernel have inode->inotify_mutex. If there are 10,000 inodes cached, then there are 10,000 lock objects. But ->inotify_mutex is a single "lock type", and all locking activities that occur against ->inotify_mutex are "unified" into this single lock-class. The advantage of the lock-class approach is that all historical ->inotify_mutex uses are mapped into a single (and as narrow as possible) set of locking rules - regardless of how many different tasks or inode structures it took to build this set of rules. The set of rules persist during the lifetime of the kernel. To see the rough magnitude of checking that the lock validator does, here's a portion of /proc/lockdep_stats, fresh after bootup: lock-classes: 694 [max: 2048] direct dependencies: 1598 [max: 8192] indirect dependencies: 17896 all direct dependencies: 16206 dependency chains: 1910 [max: 8192] in-hardirq chains: 17 in-softirq chains: 105 in-process chains: 1065 stack-trace entries: 38761 [max: 131072] combined max dependencies: 2033928 hardirq-safe locks: 24 hardirq-unsafe locks: 176 softirq-safe locks: 53 softirq-unsafe locks: 137 irq-safe locks: 59 irq-unsafe locks: 176 The lock validator has observed 1598 actual single-thread locking patterns, and has validated all possible 2033928 distinct locking scenarios. More details about the design of the lock validator can be found in Documentation/lockdep-design.txt, which can also found at: http://redhat.com/~mingo/lockdep-patches/lockdep-design.txt [bunk@stusta.de: cleanups] Signed-off-by: Ingo Molnar <mingo@elte.hu> Signed-off-by: Arjan van de Ven <arjan@linux.intel.com> Signed-off-by: Adrian Bunk <bunk@stusta.de> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-07-03 00:24:50 -07:00
UBSAN: run-time undefined behavior sanity checker UBSAN uses compile-time instrumentation to catch undefined behavior (UB). Compiler inserts code that perform certain kinds of checks before operations that could cause UB. If check fails (i.e. UB detected) __ubsan_handle_* function called to print error message. So the most of the work is done by compiler. This patch just implements ubsan handlers printing errors. GCC has this capability since 4.9.x [1] (see -fsanitize=undefined option and its suboptions). However GCC 5.x has more checkers implemented [2]. Article [3] has a bit more details about UBSAN in the GCC. [1] - https://gcc.gnu.org/onlinedocs/gcc-4.9.0/gcc/Debugging-Options.html [2] - https://gcc.gnu.org/onlinedocs/gcc/Debugging-Options.html [3] - http://developerblog.redhat.com/2014/10/16/gcc-undefined-behavior-sanitizer-ubsan/ Issues which UBSAN has found thus far are: Found bugs: * out-of-bounds access - 97840cb67ff5 ("netfilter: nfnetlink: fix insufficient validation in nfnetlink_bind") undefined shifts: * d48458d4a768 ("jbd2: use a better hash function for the revoke table") * 10632008b9e1 ("clockevents: Prevent shift out of bounds") * 'x << -1' shift in ext4 - http://lkml.kernel.org/r/<5444EF21.8020501@samsung.com> * undefined rol32(0) - http://lkml.kernel.org/r/<1449198241-20654-1-git-send-email-sasha.levin@oracle.com> * undefined dirty_ratelimit calculation - http://lkml.kernel.org/r/<566594E2.3050306@odin.com> * undefined roundown_pow_of_two(0) - http://lkml.kernel.org/r/<1449156616-11474-1-git-send-email-sasha.levin@oracle.com> * [WONTFIX] undefined shift in __bpf_prog_run - http://lkml.kernel.org/r/<CACT4Y+ZxoR3UjLgcNdUm4fECLMx2VdtfrENMtRRCdgHB2n0bJA@mail.gmail.com> WONTFIX here because it should be fixed in bpf program, not in kernel. signed overflows: * 32a8df4e0b33f ("sched: Fix odd values in effective_load() calculations") * mul overflow in ntp - http://lkml.kernel.org/r/<1449175608-1146-1-git-send-email-sasha.levin@oracle.com> * incorrect conversion into rtc_time in rtc_time64_to_tm() - http://lkml.kernel.org/r/<1449187944-11730-1-git-send-email-sasha.levin@oracle.com> * unvalidated timespec in io_getevents() - http://lkml.kernel.org/r/<CACT4Y+bBxVYLQ6LtOKrKtnLthqLHcw-BMp3aqP3mjdAvr9FULQ@mail.gmail.com> * [NOTABUG] signed overflow in ktime_add_safe() - http://lkml.kernel.org/r/<CACT4Y+aJ4muRnWxsUe1CMnA6P8nooO33kwG-c8YZg=0Xc8rJqw@mail.gmail.com> [akpm@linux-foundation.org: fix unused local warning] [akpm@linux-foundation.org: fix __int128 build woes] Signed-off-by: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Sasha Levin <sasha.levin@oracle.com> Cc: Randy Dunlap <rdunlap@infradead.org> Cc: Rasmus Villemoes <linux@rasmusvillemoes.dk> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Michal Marek <mmarek@suse.cz> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Ingo Molnar <mingo@redhat.com> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Yury Gribov <y.gribov@samsung.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Konstantin Khlebnikov <koct9i@gmail.com> Cc: Kostya Serebryany <kcc@google.com> Cc: Johannes Berg <johannes@sipsolutions.net> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-01-20 15:00:55 -08:00
UBSAN: run-time undefined behavior sanity checker UBSAN uses compile-time instrumentation to catch undefined behavior (UB). Compiler inserts code that perform certain kinds of checks before operations that could cause UB. If check fails (i.e. UB detected) __ubsan_handle_* function called to print error message. So the most of the work is done by compiler. This patch just implements ubsan handlers printing errors. GCC has this capability since 4.9.x [1] (see -fsanitize=undefined option and its suboptions). However GCC 5.x has more checkers implemented [2]. Article [3] has a bit more details about UBSAN in the GCC. [1] - https://gcc.gnu.org/onlinedocs/gcc-4.9.0/gcc/Debugging-Options.html [2] - https://gcc.gnu.org/onlinedocs/gcc/Debugging-Options.html [3] - http://developerblog.redhat.com/2014/10/16/gcc-undefined-behavior-sanitizer-ubsan/ Issues which UBSAN has found thus far are: Found bugs: * out-of-bounds access - 97840cb67ff5 ("netfilter: nfnetlink: fix insufficient validation in nfnetlink_bind") undefined shifts: * d48458d4a768 ("jbd2: use a better hash function for the revoke table") * 10632008b9e1 ("clockevents: Prevent shift out of bounds") * 'x << -1' shift in ext4 - http://lkml.kernel.org/r/<5444EF21.8020501@samsung.com> * undefined rol32(0) - http://lkml.kernel.org/r/<1449198241-20654-1-git-send-email-sasha.levin@oracle.com> * undefined dirty_ratelimit calculation - http://lkml.kernel.org/r/<566594E2.3050306@odin.com> * undefined roundown_pow_of_two(0) - http://lkml.kernel.org/r/<1449156616-11474-1-git-send-email-sasha.levin@oracle.com> * [WONTFIX] undefined shift in __bpf_prog_run - http://lkml.kernel.org/r/<CACT4Y+ZxoR3UjLgcNdUm4fECLMx2VdtfrENMtRRCdgHB2n0bJA@mail.gmail.com> WONTFIX here because it should be fixed in bpf program, not in kernel. signed overflows: * 32a8df4e0b33f ("sched: Fix odd values in effective_load() calculations") * mul overflow in ntp - http://lkml.kernel.org/r/<1449175608-1146-1-git-send-email-sasha.levin@oracle.com> * incorrect conversion into rtc_time in rtc_time64_to_tm() - http://lkml.kernel.org/r/<1449187944-11730-1-git-send-email-sasha.levin@oracle.com> * unvalidated timespec in io_getevents() - http://lkml.kernel.org/r/<CACT4Y+bBxVYLQ6LtOKrKtnLthqLHcw-BMp3aqP3mjdAvr9FULQ@mail.gmail.com> * [NOTABUG] signed overflow in ktime_add_safe() - http://lkml.kernel.org/r/<CACT4Y+aJ4muRnWxsUe1CMnA6P8nooO33kwG-c8YZg=0Xc8rJqw@mail.gmail.com> [akpm@linux-foundation.org: fix unused local warning] [akpm@linux-foundation.org: fix __int128 build woes] Signed-off-by: Andrey Ryabinin <aryabinin@virtuozzo.com> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Sasha Levin <sasha.levin@oracle.com> Cc: Randy Dunlap <rdunlap@infradead.org> Cc: Rasmus Villemoes <linux@rasmusvillemoes.dk> Cc: Jonathan Corbet <corbet@lwn.net> Cc: Michal Marek <mmarek@suse.cz> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: Ingo Molnar <mingo@redhat.com> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Yury Gribov <y.gribov@samsung.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Konstantin Khlebnikov <koct9i@gmail.com> Cc: Kostya Serebryany <kcc@google.com> Cc: Johannes Berg <johannes@sipsolutions.net> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-01-20 15:00:55 -08:00
When stacked block devices are in-use (e.g. md or dm), the recursive calls to generic_make_request can use up a lot of space, and we would rather they didn't. As generic_make_request is a void function, and as it is generally not expected that it will have any effect immediately, it is safe to delay any call to generic_make_request until there is sufficient stack space available. As ->bi_next is reserved for the driver to use, it can have no valid value when generic_make_request is called, and as __make_request implicitly assumes it will be NULL (ELEVATOR_BACK_MERGE fork of switch) we can be certain that all callers set it to NULL. We can therefore safely use bi_next to link pending requests together, providing we clear it before making the real call. So, we choose to allow each thread to only be active in one generic_make_request at a time. If a subsequent (recursive) call is made, the bio is linked into a per-thread list, and is handled when the active call completes. As the list of pending bios is per-thread, there are no locking issues to worry about. I say above that it is "safe to delay any call...". There are, however, some behaviours of a make_request_fn which would make it unsafe. These include any behaviour that assumes anything will have changed after a recursive call to generic_make_request. These could include: - waiting for that call to finish and call it's bi_end_io function. md use to sometimes do this (marking the superblock dirty before completing a write) but doesn't any more - inspecting the bio for fields that generic_make_request might change, such as bi_sector or bi_bdev. It is hard to see a good reason for this, and I don't think anyone actually does it. - inspecing the queue to see if, e.g. it is 'full' yet. Again, I think this is very unlikely to be useful, or to be done. Signed-off-by: Neil Brown <neilb@suse.de> Cc: Jens Axboe <axboe@kernel.dk> Cc: <dm-devel@redhat.com> Alasdair G Kergon <agk@redhat.com> said: I can see nothing wrong with this in principle. For device-mapper at the moment though it's essential that, while the bio mappings may now get delayed, they still get processed in exactly the same order as they were passed to generic_make_request(). My main concern is whether the timing changes implicit in this patch will make the rare data-corrupting races in the existing snapshot code more likely. (I'm working on a fix for these races, but the unfinished patch is already several hundred lines long.) It would be helpful if some people on this mailing list would test this patch in various scenarios and report back. Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
2007-05-01 09:53:42 +02:00
Task Control Groups: basic task cgroup framework Generic Process Control Groups -------------------------- There have recently been various proposals floating around for resource management/accounting and other task grouping subsystems in the kernel, including ResGroups, User BeanCounters, NSProxy cgroups, and others. These all need the basic abstraction of being able to group together multiple processes in an aggregate, in order to track/limit the resources permitted to those processes, or control other behaviour of the processes, and all implement this grouping in different ways. This patchset provides a framework for tracking and grouping processes into arbitrary "cgroups" and assigning arbitrary state to those groupings, in order to control the behaviour of the cgroup as an aggregate. The intention is that the various resource management and virtualization/cgroup efforts can also become task cgroup clients, with the result that: - the userspace APIs are (somewhat) normalised - it's easier to test e.g. the ResGroups CPU controller in conjunction with the BeanCounters memory controller, or use either of them as the resource-control portion of a virtual server system. - the additional kernel footprint of any of the competing resource management systems is substantially reduced, since it doesn't need to provide process grouping/containment, hence improving their chances of getting into the kernel This patch: Add the main task cgroups framework - the cgroup filesystem, and the basic structures for tracking membership and associating subsystem state objects to tasks. Signed-off-by: Paul Menage <menage@google.com> Cc: Serge E. Hallyn <serue@us.ibm.com> Cc: "Eric W. Biederman" <ebiederm@xmission.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Balbir Singh <balbir@in.ibm.com> Cc: Paul Jackson <pj@sgi.com> Cc: Kirill Korotaev <dev@openvz.org> Cc: Herbert Poetzl <herbert@13thfloor.at> Cc: Srivatsa Vaddagiri <vatsa@in.ibm.com> Cc: Cedric Le Goater <clg@fr.ibm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-18 23:39:30 -07:00
Task Control Groups: basic task cgroup framework Generic Process Control Groups -------------------------- There have recently been various proposals floating around for resource management/accounting and other task grouping subsystems in the kernel, including ResGroups, User BeanCounters, NSProxy cgroups, and others. These all need the basic abstraction of being able to group together multiple processes in an aggregate, in order to track/limit the resources permitted to those processes, or control other behaviour of the processes, and all implement this grouping in different ways. This patchset provides a framework for tracking and grouping processes into arbitrary "cgroups" and assigning arbitrary state to those groupings, in order to control the behaviour of the cgroup as an aggregate. The intention is that the various resource management and virtualization/cgroup efforts can also become task cgroup clients, with the result that: - the userspace APIs are (somewhat) normalised - it's easier to test e.g. the ResGroups CPU controller in conjunction with the BeanCounters memory controller, or use either of them as the resource-control portion of a virtual server system. - the additional kernel footprint of any of the competing resource management systems is substantially reduced, since it doesn't need to provide process grouping/containment, hence improving their chances of getting into the kernel This patch: Add the main task cgroups framework - the cgroup filesystem, and the basic structures for tracking membership and associating subsystem state objects to tasks. Signed-off-by: Paul Menage <menage@google.com> Cc: Serge E. Hallyn <serue@us.ibm.com> Cc: "Eric W. Biederman" <ebiederm@xmission.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Balbir Singh <balbir@in.ibm.com> Cc: Paul Jackson <pj@sgi.com> Cc: Kirill Korotaev <dev@openvz.org> Cc: Herbert Poetzl <herbert@13thfloor.at> Cc: Srivatsa Vaddagiri <vatsa@in.ibm.com> Cc: Cedric Le Goater <clg@fr.ibm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-18 23:39:30 -07:00
x86/intel_rdt/cqm: Add tasks file support The root directory, ctrl_mon and monitor groups are populated with a read/write file named "tasks". When read, it shows all the task IDs assigned to the resource group. Tasks can be added to groups by writing the PID to the file. A task can be present in one "ctrl_mon" group "and" one "monitor" group. IOW a PID_x can be seen in a ctrl_mon group and a monitor group at the same time. When a task is added to a ctrl_mon group, it is automatically removed from the previous ctrl_mon group where it belonged. Similarly if a task is moved to a monitor group it is removed from the previous monitor group . Also since the monitor groups can only have subset of tasks of parent ctrl_mon group, a task can be moved to a monitor group only if its already present in the parent ctrl_mon group. Task membership is indicated by a new field in the task_struct "u32 rmid" which holds the RMID for the task. RMID=0 is reserved for the default root group where the tasks belong to at mount. [tony: zero the rmid if rdtgroup was deleted when task was being moved] Signed-off-by: Tony Luck <tony.luck@linux.intel.com> Signed-off-by: Vikas Shivappa <vikas.shivappa@linux.intel.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Cc: ravi.v.shankar@intel.com Cc: tony.luck@intel.com Cc: fenghua.yu@intel.com Cc: peterz@infradead.org Cc: eranian@google.com Cc: vikas.shivappa@intel.com Cc: ak@linux.intel.com Cc: davidcc@google.com Cc: reinette.chatre@intel.com Link: http://lkml.kernel.org/r/1501017287-28083-16-git-send-email-vikas.shivappa@linux.intel.com
2017-07-25 14:14:34 -07:00
perf: Do the big rename: Performance Counters -> Performance Events Bye-bye Performance Counters, welcome Performance Events! In the past few months the perfcounters subsystem has grown out its initial role of counting hardware events, and has become (and is becoming) a much broader generic event enumeration, reporting, logging, monitoring, analysis facility. Naming its core object 'perf_counter' and naming the subsystem 'perfcounters' has become more and more of a misnomer. With pending code like hw-breakpoints support the 'counter' name is less and less appropriate. All in one, we've decided to rename the subsystem to 'performance events' and to propagate this rename through all fields, variables and API names. (in an ABI compatible fashion) The word 'event' is also a bit shorter than 'counter' - which makes it slightly more convenient to write/handle as well. Thanks goes to Stephane Eranian who first observed this misnomer and suggested a rename. User-space tooling and ABI compatibility is not affected - this patch should be function-invariant. (Also, defconfigs were not touched to keep the size down.) This patch has been generated via the following script: FILES=$(find * -type f | grep -vE 'oprofile|[^K]config') sed -i \ -e 's/PERF_EVENT_/PERF_RECORD_/g' \ -e 's/PERF_COUNTER/PERF_EVENT/g' \ -e 's/perf_counter/perf_event/g' \ -e 's/nb_counters/nb_events/g' \ -e 's/swcounter/swevent/g' \ -e 's/tpcounter_event/tp_event/g' \ $FILES for N in $(find . -name perf_counter.[ch]); do M=$(echo $N | sed 's/perf_counter/perf_event/g') mv $N $M done FILES=$(find . -name perf_event.*) sed -i \ -e 's/COUNTER_MASK/REG_MASK/g' \ -e 's/COUNTER/EVENT/g' \ -e 's/\<event\>/event_id/g' \ -e 's/counter/event/g' \ -e 's/Counter/Event/g' \ $FILES ... to keep it as correct as possible. This script can also be used by anyone who has pending perfcounters patches - it converts a Linux kernel tree over to the new naming. We tried to time this change to the point in time where the amount of pending patches is the smallest: the end of the merge window. Namespace clashes were fixed up in a preparatory patch - and some stylistic fallout will be fixed up in a subsequent patch. ( NOTE: 'counters' are still the proper terminology when we deal with hardware registers - and these sed scripts are a bit over-eager in renaming them. I've undone some of that, but in case there's something left where 'counter' would be better than 'event' we can undo that on an individual basis instead of touching an otherwise nicely automated patch. ) Suggested-by: Stephane Eranian <eranian@google.com> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Acked-by: Paul Mackerras <paulus@samba.org> Reviewed-by: Arjan van de Ven <arjan@linux.intel.com> Cc: Mike Galbraith <efault@gmx.de> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> Cc: Frederic Weisbecker <fweisbec@gmail.com> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Benjamin Herrenschmidt <benh@kernel.crashing.org> Cc: David Howells <dhowells@redhat.com> Cc: Kyle McMartin <kyle@mcmartin.ca> Cc: Martin Schwidefsky <schwidefsky@de.ibm.com> Cc: "David S. Miller" <davem@davemloft.net> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: <linux-arch@vger.kernel.org> LKML-Reference: <new-submission> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-09-21 12:02:48 +02:00
perf_counter: Dynamically allocate tasks' perf_counter_context struct This replaces the struct perf_counter_context in the task_struct with a pointer to a dynamically allocated perf_counter_context struct. The main reason for doing is this is to allow us to transfer a perf_counter_context from one task to another when we do lazy PMU switching in a later patch. This has a few side-benefits: the task_struct becomes a little smaller, we save some memory because only tasks that have perf_counters attached get a perf_counter_context allocated for them, and we can remove the inclusion of <linux/perf_counter.h> in sched.h, meaning that we don't end up recompiling nearly everything whenever perf_counter.h changes. The perf_counter_context structures are reference-counted and freed when the last reference is dropped. A context can have references from its task and the counters on its task. Counters can outlive the task so it is possible that a context will be freed well after its task has exited. Contexts are allocated on fork if the parent had a context, or otherwise the first time that a per-task counter is created on a task. In the latter case, we set the context pointer in the task struct locklessly using an atomic compare-and-exchange operation in case we raced with some other task in creating a context for the subject task. This also removes the task pointer from the perf_counter struct. The task pointer was not used anywhere and would make it harder to move a context from one task to another. Anything that needed to know which task a counter was attached to was already using counter->ctx->task. The __perf_counter_init_context function moves up in perf_counter.c so that it can be called from find_get_context, and now initializes the refcount, but is otherwise unchanged. We were potentially calling list_del_counter twice: once from __perf_counter_exit_task when the task exits and once from __perf_counter_remove_from_context when the counter's fd gets closed. This adds a check in list_del_counter so it doesn't do anything if the counter has already been removed from the lists. Since perf_counter_task_sched_in doesn't do anything if the task doesn't have a context, and leaves cpuctx->task_ctx = NULL, this adds code to __perf_install_in_context to set cpuctx->task_ctx if necessary, i.e. in the case where the current task adds the first counter to itself and thus creates a context for itself. This also adds similar code to __perf_counter_enable to handle a similar situation which can arise when the counters have been disabled using prctl; that also leaves cpuctx->task_ctx = NULL. [ Impact: refactor counter context management to prepare for new feature ] Signed-off-by: Paul Mackerras <paulus@samba.org> Acked-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Cc: Corey Ashford <cjashfor@linux.vnet.ibm.com> Cc: Marcelo Tosatti <mtosatti@redhat.com> Cc: Arnaldo Carvalho de Melo <acme@redhat.com> LKML-Reference: <18966.10075.781053.231153@cargo.ozlabs.ibm.com> Signed-off-by: Ingo Molnar <mingo@elte.hu>
2009-05-22 14:17:31 +10:00
mm, mempolicy: stop adjusting current->il_next in mpol_rebind_nodemask() The task->il_next variable stores the next allocation node id for task's MPOL_INTERLEAVE policy. mpol_rebind_nodemask() updates interleave and bind mempolicies due to changing cpuset mems. Currently it also tries to make sure that current->il_next is valid within the updated nodemask. This is bogus, because 1) we are updating potentially any task's mempolicy, not just current, and 2) we might be updating a per-vma mempolicy, not task one. The interleave_nodes() function that uses il_next can cope fine with the value not being within the currently allowed nodes, so this hasn't manifested as an actual issue. We can remove the need for updating il_next completely by changing it to il_prev and store the node id of the previous interleave allocation instead of the next id. Then interleave_nodes() can calculate the next id using the current nodemask and also store it as il_prev, except when querying the next node via do_get_mempolicy(). Link: http://lkml.kernel.org/r/20170517081140.30654-3-vbabka@suse.cz Signed-off-by: Vlastimil Babka <vbabka@suse.cz> Reviewed-by: Christoph Lameter <cl@linux.com> Cc: "Kirill A. Shutemov" <kirill.shutemov@linux.intel.com> Cc: Andrea Arcangeli <aarcange@redhat.com> Cc: Anshuman Khandual <khandual@linux.vnet.ibm.com> Cc: David Rientjes <rientjes@google.com> Cc: Dimitri Sivanich <sivanich@sgi.com> Cc: Hugh Dickins <hughd@google.com> Cc: Li Zefan <lizefan@huawei.com> Cc: Mel Gorman <mgorman@techsingularity.net> Cc: Michal Hocko <mhocko@suse.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-07-06 15:39:59 -07:00
mm: numa: slow PTE scan rate if migration failures occur Dave Chinner reported the following on https://lkml.org/lkml/2015/3/1/226 Across the board the 4.0-rc1 numbers are much slower, and the degradation is far worse when using the large memory footprint configs. Perf points straight at the cause - this is from 4.0-rc1 on the "-o bhash=101073" config: - 56.07% 56.07% [kernel] [k] default_send_IPI_mask_sequence_phys - default_send_IPI_mask_sequence_phys - 99.99% physflat_send_IPI_mask - 99.37% native_send_call_func_ipi smp_call_function_many - native_flush_tlb_others - 99.85% flush_tlb_page ptep_clear_flush try_to_unmap_one rmap_walk try_to_unmap migrate_pages migrate_misplaced_page - handle_mm_fault - 99.73% __do_page_fault trace_do_page_fault do_async_page_fault + async_page_fault 0.63% native_send_call_func_single_ipi generic_exec_single smp_call_function_single This is showing excessive migration activity even though excessive migrations are meant to get throttled. Normally, the scan rate is tuned on a per-task basis depending on the locality of faults. However, if migrations fail for any reason then the PTE scanner may scan faster if the faults continue to be remote. This means there is higher system CPU overhead and fault trapping at exactly the time we know that migrations cannot happen. This patch tracks when migration failures occur and slows the PTE scanner. Signed-off-by: Mel Gorman <mgorman@suse.de> Reported-by: Dave Chinner <david@fromorbit.com> Tested-by: Dave Chinner <david@fromorbit.com> Cc: Ingo Molnar <mingo@kernel.org> Cc: Aneesh Kumar <aneesh.kumar@linux.vnet.ibm.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-03-25 15:55:42 -07:00
rseq: Introduce restartable sequences system call Expose a new system call allowing each thread to register one userspace memory area to be used as an ABI between kernel and user-space for two purposes: user-space restartable sequences and quick access to read the current CPU number value from user-space. * Restartable sequences (per-cpu atomics) Restartables sequences allow user-space to perform update operations on per-cpu data without requiring heavy-weight atomic operations. The restartable critical sections (percpu atomics) work has been started by Paul Turner and Andrew Hunter. It lets the kernel handle restart of critical sections. [1] [2] The re-implementation proposed here brings a few simplifications to the ABI which facilitates porting to other architectures and speeds up the user-space fast path. Here are benchmarks of various rseq use-cases. Test hardware: arm32: ARMv7 Processor rev 4 (v7l) "Cubietruck", 2-core x86-64: Intel E5-2630 v3@2.40GHz, 16-core, hyperthreading The following benchmarks were all performed on a single thread. * Per-CPU statistic counter increment getcpu+atomic (ns/op) rseq (ns/op) speedup arm32: 344.0 31.4 11.0 x86-64: 15.3 2.0 7.7 * LTTng-UST: write event 32-bit header, 32-bit payload into tracer per-cpu buffer getcpu+atomic (ns/op) rseq (ns/op) speedup arm32: 2502.0 2250.0 1.1 x86-64: 117.4 98.0 1.2 * liburcu percpu: lock-unlock pair, dereference, read/compare word getcpu+atomic (ns/op) rseq (ns/op) speedup arm32: 751.0 128.5 5.8 x86-64: 53.4 28.6 1.9 * jemalloc memory allocator adapted to use rseq Using rseq with per-cpu memory pools in jemalloc at Facebook (based on rseq 2016 implementation): The production workload response-time has 1-2% gain avg. latency, and the P99 overall latency drops by 2-3%. * Reading the current CPU number Speeding up reading the current CPU number on which the caller thread is running is done by keeping the current CPU number up do date within the cpu_id field of the memory area registered by the thread. This is done by making scheduler preemption set the TIF_NOTIFY_RESUME flag on the current thread. Upon return to user-space, a notify-resume handler updates the current CPU value within the registered user-space memory area. User-space can then read the current CPU number directly from memory. Keeping the current cpu id in a memory area shared between kernel and user-space is an improvement over current mechanisms available to read the current CPU number, which has the following benefits over alternative approaches: - 35x speedup on ARM vs system call through glibc - 20x speedup on x86 compared to calling glibc, which calls vdso executing a "lsl" instruction, - 14x speedup on x86 compared to inlined "lsl" instruction, - Unlike vdso approaches, this cpu_id value can be read from an inline assembly, which makes it a useful building block for restartable sequences. - The approach of reading the cpu id through memory mapping shared between kernel and user-space is portable (e.g. ARM), which is not the case for the lsl-based x86 vdso. On x86, yet another possible approach would be to use the gs segment selector to point to user-space per-cpu data. This approach performs similarly to the cpu id cache, but it has two disadvantages: it is not portable, and it is incompatible with existing applications already using the gs segment selector for other purposes. Benchmarking various approaches for reading the current CPU number: ARMv7 Processor rev 4 (v7l) Machine model: Cubietruck - Baseline (empty loop): 8.4 ns - Read CPU from rseq cpu_id: 16.7 ns - Read CPU from rseq cpu_id (lazy register): 19.8 ns - glibc 2.19-0ubuntu6.6 getcpu: 301.8 ns - getcpu system call: 234.9 ns x86-64 Intel(R) Xeon(R) CPU E5-2630 v3 @ 2.40GHz: - Baseline (empty loop): 0.8 ns - Read CPU from rseq cpu_id: 0.8 ns - Read CPU from rseq cpu_id (lazy register): 0.8 ns - Read using gs segment selector: 0.8 ns - "lsl" inline assembly: 13.0 ns - glibc 2.19-0ubuntu6 getcpu: 16.6 ns - getcpu system call: 53.9 ns - Speed (benchmark taken on v8 of patchset) Running 10 runs of hackbench -l 100000 seems to indicate, contrary to expectations, that enabling CONFIG_RSEQ slightly accelerates the scheduler: Configuration: 2 sockets * 8-core Intel(R) Xeon(R) CPU E5-2630 v3 @ 2.40GHz (directly on hardware, hyperthreading disabled in BIOS, energy saving disabled in BIOS, turboboost disabled in BIOS, cpuidle.off=1 kernel parameter), with a Linux v4.6 defconfig+localyesconfig, restartable sequences series applied. * CONFIG_RSEQ=n avg.: 41.37 s std.dev.: 0.36 s * CONFIG_RSEQ=y avg.: 40.46 s std.dev.: 0.33 s - Size On x86-64, between CONFIG_RSEQ=n/y, the text size increase of vmlinux is 567 bytes, and the data size increase of vmlinux is 5696 bytes. [1] https://lwn.net/Articles/650333/ [2] http://www.linuxplumbersconf.org/2013/ocw/system/presentations/1695/original/LPC%20-%20PerCpu%20Atomics.pdf Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Joel Fernandes <joelaf@google.com> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Dave Watson <davejwatson@fb.com> Cc: Will Deacon <will.deacon@arm.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: "H . Peter Anvin" <hpa@zytor.com> Cc: Chris Lameter <cl@linux.com> Cc: Russell King <linux@arm.linux.org.uk> Cc: Andrew Hunter <ahh@google.com> Cc: Michael Kerrisk <mtk.manpages@gmail.com> Cc: "Paul E . McKenney" <paulmck@linux.vnet.ibm.com> Cc: Paul Turner <pjt@google.com> Cc: Boqun Feng <boqun.feng@gmail.com> Cc: Josh Triplett <josh@joshtriplett.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ben Maurer <bmaurer@fb.com> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: linux-api@vger.kernel.org Cc: Andy Lutomirski <luto@amacapital.net> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/20151027235635.16059.11630.stgit@pjt-glaptop.roam.corp.google.com Link: http://lkml.kernel.org/r/20150624222609.6116.86035.stgit@kitami.mtv.corp.google.com Link: https://lkml.kernel.org/r/20180602124408.8430-3-mathieu.desnoyers@efficios.com
2018-06-02 08:43:54 -04:00
mm: send one IPI per CPU to TLB flush all entries after unmapping pages An IPI is sent to flush remote TLBs when a page is unmapped that was potentially accesssed by other CPUs. There are many circumstances where this happens but the obvious one is kswapd reclaiming pages belonging to a running process as kswapd and the task are likely running on separate CPUs. On small machines, this is not a significant problem but as machine gets larger with more cores and more memory, the cost of these IPIs can be high. This patch uses a simple structure that tracks CPUs that potentially have TLB entries for pages being unmapped. When the unmapping is complete, the full TLB is flushed on the assumption that a refill cost is lower than flushing individual entries. Architectures wishing to do this must give the following guarantee. If a clean page is unmapped and not immediately flushed, the architecture must guarantee that a write to that linear address from a CPU with a cached TLB entry will trap a page fault. This is essentially what the kernel already depends on but the window is much larger with this patch applied and is worth highlighting. The architecture should consider whether the cost of the full TLB flush is higher than sending an IPI to flush each individual entry. An additional architecture helper called flush_tlb_local is required. It's a trivial wrapper with some accounting in the x86 case. The impact of this patch depends on the workload as measuring any benefit requires both mapped pages co-located on the LRU and memory pressure. The case with the biggest impact is multiple processes reading mapped pages taken from the vm-scalability test suite. The test case uses NR_CPU readers of mapped files that consume 10*RAM. Linear mapped reader on a 4-node machine with 64G RAM and 48 CPUs 4.2.0-rc1 4.2.0-rc1 vanilla flushfull-v7 Ops lru-file-mmap-read-elapsed 159.62 ( 0.00%) 120.68 ( 24.40%) Ops lru-file-mmap-read-time_range 30.59 ( 0.00%) 2.80 ( 90.85%) Ops lru-file-mmap-read-time_stddv 6.70 ( 0.00%) 0.64 ( 90.38%) 4.2.0-rc1 4.2.0-rc1 vanilla flushfull-v7 User 581.00 611.43 System 5804.93 4111.76 Elapsed 161.03 122.12 This is showing that the readers completed 24.40% faster with 29% less system CPU time. From vmstats, it is known that the vanilla kernel was interrupted roughly 900K times per second during the steady phase of the test and the patched kernel was interrupts 180K times per second. The impact is lower on a single socket machine. 4.2.0-rc1 4.2.0-rc1 vanilla flushfull-v7 Ops lru-file-mmap-read-elapsed 25.33 ( 0.00%) 20.38 ( 19.54%) Ops lru-file-mmap-read-time_range 0.91 ( 0.00%) 1.44 (-58.24%) Ops lru-file-mmap-read-time_stddv 0.28 ( 0.00%) 0.47 (-65.34%) 4.2.0-rc1 4.2.0-rc1 vanilla flushfull-v7 User 58.09 57.64 System 111.82 76.56 Elapsed 27.29 22.55 It's still a noticeable improvement with vmstat showing interrupts went from roughly 500K per second to 45K per second. The patch will have no impact on workloads with no memory pressure or have relatively few mapped pages. It will have an unpredictable impact on the workload running on the CPU being flushed as it'll depend on how many TLB entries need to be refilled and how long that takes. Worst case, the TLB will be completely cleared of active entries when the target PFNs were not resident at all. [sasha.levin@oracle.com: trace tlb flush after disabling preemption in try_to_unmap_flush] Signed-off-by: Mel Gorman <mgorman@suse.de> Reviewed-by: Rik van Riel <riel@redhat.com> Cc: Dave Hansen <dave.hansen@intel.com> Acked-by: Ingo Molnar <mingo@kernel.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Signed-off-by: Sasha Levin <sasha.levin@oracle.com> Cc: Michal Hocko <mhocko@suse.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-09-04 15:47:32 -07:00
net: use a per task frag allocator We currently use a per socket order-0 page cache for tcp_sendmsg() operations. This page is used to build fragments for skbs. Its done to increase probability of coalescing small write() into single segments in skbs still in write queue (not yet sent) But it wastes a lot of memory for applications handling many mostly idle sockets, since each socket holds one page in sk->sk_sndmsg_page Its also quite inefficient to build TSO 64KB packets, because we need about 16 pages per skb on arches where PAGE_SIZE = 4096, so we hit page allocator more than wanted. This patch adds a per task frag allocator and uses bigger pages, if available. An automatic fallback is done in case of memory pressure. (up to 32768 bytes per frag, thats order-3 pages on x86) This increases TCP stream performance by 20% on loopback device, but also benefits on other network devices, since 8x less frags are mapped on transmit and unmapped on tx completion. Alexander Duyck mentioned a probable performance win on systems with IOMMU enabled. Its possible some SG enabled hardware cant cope with bigger fragments, but their ndo_start_xmit() should already handle this, splitting a fragment in sub fragments, since some arches have PAGE_SIZE=65536 Successfully tested on various ethernet devices. (ixgbe, igb, bnx2x, tg3, mellanox mlx4) Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Ben Hutchings <bhutchings@solarflare.com> Cc: Vijay Subramanian <subramanian.vijay@gmail.com> Cc: Alexander Duyck <alexander.h.duyck@intel.com> Tested-by: Vijay Subramanian <subramanian.vijay@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2012-09-23 23:04:42 +00:00
net: use a per task frag allocator We currently use a per socket order-0 page cache for tcp_sendmsg() operations. This page is used to build fragments for skbs. Its done to increase probability of coalescing small write() into single segments in skbs still in write queue (not yet sent) But it wastes a lot of memory for applications handling many mostly idle sockets, since each socket holds one page in sk->sk_sndmsg_page Its also quite inefficient to build TSO 64KB packets, because we need about 16 pages per skb on arches where PAGE_SIZE = 4096, so we hit page allocator more than wanted. This patch adds a per task frag allocator and uses bigger pages, if available. An automatic fallback is done in case of memory pressure. (up to 32768 bytes per frag, thats order-3 pages on x86) This increases TCP stream performance by 20% on loopback device, but also benefits on other network devices, since 8x less frags are mapped on transmit and unmapped on tx completion. Alexander Duyck mentioned a probable performance win on systems with IOMMU enabled. Its possible some SG enabled hardware cant cope with bigger fragments, but their ndo_start_xmit() should already handle this, splitting a fragment in sub fragments, since some arches have PAGE_SIZE=65536 Successfully tested on various ethernet devices. (ixgbe, igb, bnx2x, tg3, mellanox mlx4) Signed-off-by: Eric Dumazet <edumazet@google.com> Cc: Ben Hutchings <bhutchings@solarflare.com> Cc: Vijay Subramanian <subramanian.vijay@gmail.com> Cc: Alexander Duyck <alexander.h.duyck@intel.com> Tested-by: Vijay Subramanian <subramanian.vijay@gmail.com> Signed-off-by: David S. Miller <davem@davemloft.net>
2012-09-23 23:04:42 +00:00
writeback: per task dirty rate limit Add two fields to task_struct. 1) account dirtied pages in the individual tasks, for accuracy 2) per-task balance_dirty_pages() call intervals, for flexibility The balance_dirty_pages() call interval (ie. nr_dirtied_pause) will scale near-sqrt to the safety gap between dirty pages and threshold. The main problem of per-task nr_dirtied is, if 1k+ tasks start dirtying pages at exactly the same time, each task will be assigned a large initial nr_dirtied_pause, so that the dirty threshold will be exceeded long before each task reached its nr_dirtied_pause and hence call balance_dirty_pages(). The solution is to watch for the number of pages dirtied on each CPU in between the calls into balance_dirty_pages(). If it exceeds ratelimit_pages (3% dirty threshold), force call balance_dirty_pages() for a chance to set bdi->dirty_exceeded. In normal situations, this safeguarding condition is not expected to trigger at all. On the sqrt in dirty_poll_interval(): It will serve as an initial guess when dirty pages are still in the freerun area. When dirty pages are floating inside the dirty control scope [freerun, limit], a followup patch will use some refined dirty poll interval to get the desired pause time. thresh-dirty (MB) sqrt 1 16 2 22 4 32 8 45 16 64 32 90 64 128 128 181 256 256 512 362 1024 512 The above table means, given 1MB (or 1GB) gap and the dd tasks polling balance_dirty_pages() on every 16 (or 512) pages, the dirty limit won't be exceeded as long as there are less than 16 (or 512) concurrent dd's. So sqrt naturally leads to less overheads and more safe concurrent tasks for large memory servers, which have large (thresh-freerun) gaps. peter: keep the per-CPU ratelimit for safeguarding the 1k+ tasks case CC: Peter Zijlstra <a.p.zijlstra@chello.nl> Reviewed-by: Andrea Righi <andrea@betterlinux.com> Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
2011-06-11 18:10:12 -06:00
writeback: per task dirty rate limit Add two fields to task_struct. 1) account dirtied pages in the individual tasks, for accuracy 2) per-task balance_dirty_pages() call intervals, for flexibility The balance_dirty_pages() call interval (ie. nr_dirtied_pause) will scale near-sqrt to the safety gap between dirty pages and threshold. The main problem of per-task nr_dirtied is, if 1k+ tasks start dirtying pages at exactly the same time, each task will be assigned a large initial nr_dirtied_pause, so that the dirty threshold will be exceeded long before each task reached its nr_dirtied_pause and hence call balance_dirty_pages(). The solution is to watch for the number of pages dirtied on each CPU in between the calls into balance_dirty_pages(). If it exceeds ratelimit_pages (3% dirty threshold), force call balance_dirty_pages() for a chance to set bdi->dirty_exceeded. In normal situations, this safeguarding condition is not expected to trigger at all. On the sqrt in dirty_poll_interval(): It will serve as an initial guess when dirty pages are still in the freerun area. When dirty pages are floating inside the dirty control scope [freerun, limit], a followup patch will use some refined dirty poll interval to get the desired pause time. thresh-dirty (MB) sqrt 1 16 2 22 4 32 8 45 16 64 32 90 64 128 128 181 256 256 512 362 1024 512 The above table means, given 1MB (or 1GB) gap and the dd tasks polling balance_dirty_pages() on every 16 (or 512) pages, the dirty limit won't be exceeded as long as there are less than 16 (or 512) concurrent dd's. So sqrt naturally leads to less overheads and more safe concurrent tasks for large memory servers, which have large (thresh-freerun) gaps. peter: keep the per-CPU ratelimit for safeguarding the 1k+ tasks case CC: Peter Zijlstra <a.p.zijlstra@chello.nl> Reviewed-by: Andrea Righi <andrea@betterlinux.com> Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
2011-06-11 18:10:12 -06:00
writeback: per task dirty rate limit Add two fields to task_struct. 1) account dirtied pages in the individual tasks, for accuracy 2) per-task balance_dirty_pages() call intervals, for flexibility The balance_dirty_pages() call interval (ie. nr_dirtied_pause) will scale near-sqrt to the safety gap between dirty pages and threshold. The main problem of per-task nr_dirtied is, if 1k+ tasks start dirtying pages at exactly the same time, each task will be assigned a large initial nr_dirtied_pause, so that the dirty threshold will be exceeded long before each task reached its nr_dirtied_pause and hence call balance_dirty_pages(). The solution is to watch for the number of pages dirtied on each CPU in between the calls into balance_dirty_pages(). If it exceeds ratelimit_pages (3% dirty threshold), force call balance_dirty_pages() for a chance to set bdi->dirty_exceeded. In normal situations, this safeguarding condition is not expected to trigger at all. On the sqrt in dirty_poll_interval(): It will serve as an initial guess when dirty pages are still in the freerun area. When dirty pages are floating inside the dirty control scope [freerun, limit], a followup patch will use some refined dirty poll interval to get the desired pause time. thresh-dirty (MB) sqrt 1 16 2 22 4 32 8 45 16 64 32 90 64 128 128 181 256 256 512 362 1024 512 The above table means, given 1MB (or 1GB) gap and the dd tasks polling balance_dirty_pages() on every 16 (or 512) pages, the dirty limit won't be exceeded as long as there are less than 16 (or 512) concurrent dd's. So sqrt naturally leads to less overheads and more safe concurrent tasks for large memory servers, which have large (thresh-freerun) gaps. peter: keep the per-CPU ratelimit for safeguarding the 1k+ tasks case CC: Peter Zijlstra <a.p.zijlstra@chello.nl> Reviewed-by: Andrea Righi <andrea@betterlinux.com> Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
2011-06-11 18:10:12 -06:00
kasan: add kernel address sanitizer infrastructure Kernel Address sanitizer (KASan) is a dynamic memory error detector. It provides fast and comprehensive solution for finding use-after-free and out-of-bounds bugs. KASAN uses compile-time instrumentation for checking every memory access, therefore GCC > v4.9.2 required. v4.9.2 almost works, but has issues with putting symbol aliases into the wrong section, which breaks kasan instrumentation of globals. This patch only adds infrastructure for kernel address sanitizer. It's not available for use yet. The idea and some code was borrowed from [1]. Basic idea: The main idea of KASAN is to use shadow memory to record whether each byte of memory is safe to access or not, and use compiler's instrumentation to check the shadow memory on each memory access. Address sanitizer uses 1/8 of the memory addressable in kernel for shadow memory and uses direct mapping with a scale and offset to translate a memory address to its corresponding shadow address. Here is function to translate address to corresponding shadow address: unsigned long kasan_mem_to_shadow(unsigned long addr) { return (addr >> KASAN_SHADOW_SCALE_SHIFT) + KASAN_SHADOW_OFFSET; } where KASAN_SHADOW_SCALE_SHIFT = 3. So for every 8 bytes there is one corresponding byte of shadow memory. The following encoding used for each shadow byte: 0 means that all 8 bytes of the corresponding memory region are valid for access; k (1 <= k <= 7) means that the first k bytes are valid for access, and other (8 - k) bytes are not; Any negative value indicates that the entire 8-bytes are inaccessible. Different negative values used to distinguish between different kinds of inaccessible memory (redzones, freed memory) (see mm/kasan/kasan.h). To be able to detect accesses to bad memory we need a special compiler. Such compiler inserts a specific function calls (__asan_load*(addr), __asan_store*(addr)) before each memory access of size 1, 2, 4, 8 or 16. These functions check whether memory region is valid to access or not by checking corresponding shadow memory. If access is not valid an error printed. Historical background of the address sanitizer from Dmitry Vyukov: "We've developed the set of tools, AddressSanitizer (Asan), ThreadSanitizer and MemorySanitizer, for user space. We actively use them for testing inside of Google (continuous testing, fuzzing, running prod services). To date the tools have found more than 10'000 scary bugs in Chromium, Google internal codebase and various open-source projects (Firefox, OpenSSL, gcc, clang, ffmpeg, MySQL and lots of others): [2] [3] [4]. The tools are part of both gcc and clang compilers. We have not yet done massive testing under the Kernel AddressSanitizer (it's kind of chicken and egg problem, you need it to be upstream to start applying it extensively). To date it has found about 50 bugs. Bugs that we've found in upstream kernel are listed in [5]. We've also found ~20 bugs in out internal version of the kernel. Also people from Samsung and Oracle have found some. [...] As others noted, the main feature of AddressSanitizer is its performance due to inline compiler instrumentation and simple linear shadow memory. User-space Asan has ~2x slowdown on computational programs and ~2x memory consumption increase. Taking into account that kernel usually consumes only small fraction of CPU and memory when running real user-space programs, I would expect that kernel Asan will have ~10-30% slowdown and similar memory consumption increase (when we finish all tuning). I agree that Asan can well replace kmemcheck. We have plans to start working on Kernel MemorySanitizer that finds uses of unitialized memory. Asan+Msan will provide feature-parity with kmemcheck. As others noted, Asan will unlikely replace debug slab and pagealloc that can be enabled at runtime. Asan uses compiler instrumentation, so even if it is disabled, it still incurs visible overheads. Asan technology is easily portable to other architectures. Compiler instrumentation is fully portable. Runtime has some arch-dependent parts like shadow mapping and atomic operation interception. They are relatively easy to port." Comparison with other debugging features: ======================================== KMEMCHECK: - KASan can do almost everything that kmemcheck can. KASan uses compile-time instrumentation, which makes it significantly faster than kmemcheck. The only advantage of kmemcheck over KASan is detection of uninitialized memory reads. Some brief performance testing showed that kasan could be x500-x600 times faster than kmemcheck: $ netperf -l 30 MIGRATED TCP STREAM TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to localhost (127.0.0.1) port 0 AF_INET Recv Send Send Socket Socket Message Elapsed Size Size Size Time Throughput bytes bytes bytes secs. 10^6bits/sec no debug: 87380 16384 16384 30.00 41624.72 kasan inline: 87380 16384 16384 30.00 12870.54 kasan outline: 87380 16384 16384 30.00 10586.39 kmemcheck: 87380 16384 16384 30.03 20.23 - Also kmemcheck couldn't work on several CPUs. It always sets number of CPUs to 1. KASan doesn't have such limitation. DEBUG_PAGEALLOC: - KASan is slower than DEBUG_PAGEALLOC, but KASan works on sub-page granularity level, so it able to find more bugs. SLUB_DEBUG (poisoning, redzones): - SLUB_DEBUG has lower overhead than KASan. - SLUB_DEBUG in most cases are not able to detect bad reads, KASan able to detect both reads and writes. - In some cases (e.g. redzone overwritten) SLUB_DEBUG detect bugs only on allocation/freeing of object. KASan catch bugs right before it will happen, so we always know exact place of first bad read/write. [1] https://code.google.com/p/address-sanitizer/wiki/AddressSanitizerForKernel [2] https://code.google.com/p/address-sanitizer/wiki/FoundBugs [3] https://code.google.com/p/thread-sanitizer/wiki/FoundBugs [4] https://code.google.com/p/memory-sanitizer/wiki/FoundBugs [5] https://code.google.com/p/address-sanitizer/wiki/AddressSanitizerForKernel#Trophies Based on work by Andrey Konovalov. Signed-off-by: Andrey Ryabinin <a.ryabinin@samsung.com> Acked-by: Michal Marek <mmarek@suse.cz> Signed-off-by: Andrey Konovalov <adech.fo@gmail.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Konstantin Serebryany <kcc@google.com> Cc: Dmitry Chernenkov <dmitryc@google.com> Cc: Yuri Gribov <tetra2005@gmail.com> Cc: Konstantin Khlebnikov <koct9i@gmail.com> Cc: Sasha Levin <sasha.levin@oracle.com> Cc: Christoph Lameter <cl@linux.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Ingo Molnar <mingo@elte.hu> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-13 14:39:17 -08:00
kasan: add kernel address sanitizer infrastructure Kernel Address sanitizer (KASan) is a dynamic memory error detector. It provides fast and comprehensive solution for finding use-after-free and out-of-bounds bugs. KASAN uses compile-time instrumentation for checking every memory access, therefore GCC > v4.9.2 required. v4.9.2 almost works, but has issues with putting symbol aliases into the wrong section, which breaks kasan instrumentation of globals. This patch only adds infrastructure for kernel address sanitizer. It's not available for use yet. The idea and some code was borrowed from [1]. Basic idea: The main idea of KASAN is to use shadow memory to record whether each byte of memory is safe to access or not, and use compiler's instrumentation to check the shadow memory on each memory access. Address sanitizer uses 1/8 of the memory addressable in kernel for shadow memory and uses direct mapping with a scale and offset to translate a memory address to its corresponding shadow address. Here is function to translate address to corresponding shadow address: unsigned long kasan_mem_to_shadow(unsigned long addr) { return (addr >> KASAN_SHADOW_SCALE_SHIFT) + KASAN_SHADOW_OFFSET; } where KASAN_SHADOW_SCALE_SHIFT = 3. So for every 8 bytes there is one corresponding byte of shadow memory. The following encoding used for each shadow byte: 0 means that all 8 bytes of the corresponding memory region are valid for access; k (1 <= k <= 7) means that the first k bytes are valid for access, and other (8 - k) bytes are not; Any negative value indicates that the entire 8-bytes are inaccessible. Different negative values used to distinguish between different kinds of inaccessible memory (redzones, freed memory) (see mm/kasan/kasan.h). To be able to detect accesses to bad memory we need a special compiler. Such compiler inserts a specific function calls (__asan_load*(addr), __asan_store*(addr)) before each memory access of size 1, 2, 4, 8 or 16. These functions check whether memory region is valid to access or not by checking corresponding shadow memory. If access is not valid an error printed. Historical background of the address sanitizer from Dmitry Vyukov: "We've developed the set of tools, AddressSanitizer (Asan), ThreadSanitizer and MemorySanitizer, for user space. We actively use them for testing inside of Google (continuous testing, fuzzing, running prod services). To date the tools have found more than 10'000 scary bugs in Chromium, Google internal codebase and various open-source projects (Firefox, OpenSSL, gcc, clang, ffmpeg, MySQL and lots of others): [2] [3] [4]. The tools are part of both gcc and clang compilers. We have not yet done massive testing under the Kernel AddressSanitizer (it's kind of chicken and egg problem, you need it to be upstream to start applying it extensively). To date it has found about 50 bugs. Bugs that we've found in upstream kernel are listed in [5]. We've also found ~20 bugs in out internal version of the kernel. Also people from Samsung and Oracle have found some. [...] As others noted, the main feature of AddressSanitizer is its performance due to inline compiler instrumentation and simple linear shadow memory. User-space Asan has ~2x slowdown on computational programs and ~2x memory consumption increase. Taking into account that kernel usually consumes only small fraction of CPU and memory when running real user-space programs, I would expect that kernel Asan will have ~10-30% slowdown and similar memory consumption increase (when we finish all tuning). I agree that Asan can well replace kmemcheck. We have plans to start working on Kernel MemorySanitizer that finds uses of unitialized memory. Asan+Msan will provide feature-parity with kmemcheck. As others noted, Asan will unlikely replace debug slab and pagealloc that can be enabled at runtime. Asan uses compiler instrumentation, so even if it is disabled, it still incurs visible overheads. Asan technology is easily portable to other architectures. Compiler instrumentation is fully portable. Runtime has some arch-dependent parts like shadow mapping and atomic operation interception. They are relatively easy to port." Comparison with other debugging features: ======================================== KMEMCHECK: - KASan can do almost everything that kmemcheck can. KASan uses compile-time instrumentation, which makes it significantly faster than kmemcheck. The only advantage of kmemcheck over KASan is detection of uninitialized memory reads. Some brief performance testing showed that kasan could be x500-x600 times faster than kmemcheck: $ netperf -l 30 MIGRATED TCP STREAM TEST from 0.0.0.0 (0.0.0.0) port 0 AF_INET to localhost (127.0.0.1) port 0 AF_INET Recv Send Send Socket Socket Message Elapsed Size Size Size Time Throughput bytes bytes bytes secs. 10^6bits/sec no debug: 87380 16384 16384 30.00 41624.72 kasan inline: 87380 16384 16384 30.00 12870.54 kasan outline: 87380 16384 16384 30.00 10586.39 kmemcheck: 87380 16384 16384 30.03 20.23 - Also kmemcheck couldn't work on several CPUs. It always sets number of CPUs to 1. KASan doesn't have such limitation. DEBUG_PAGEALLOC: - KASan is slower than DEBUG_PAGEALLOC, but KASan works on sub-page granularity level, so it able to find more bugs. SLUB_DEBUG (poisoning, redzones): - SLUB_DEBUG has lower overhead than KASan. - SLUB_DEBUG in most cases are not able to detect bad reads, KASan able to detect both reads and writes. - In some cases (e.g. redzone overwritten) SLUB_DEBUG detect bugs only on allocation/freeing of object. KASan catch bugs right before it will happen, so we always know exact place of first bad read/write. [1] https://code.google.com/p/address-sanitizer/wiki/AddressSanitizerForKernel [2] https://code.google.com/p/address-sanitizer/wiki/FoundBugs [3] https://code.google.com/p/thread-sanitizer/wiki/FoundBugs [4] https://code.google.com/p/memory-sanitizer/wiki/FoundBugs [5] https://code.google.com/p/address-sanitizer/wiki/AddressSanitizerForKernel#Trophies Based on work by Andrey Konovalov. Signed-off-by: Andrey Ryabinin <a.ryabinin@samsung.com> Acked-by: Michal Marek <mmarek@suse.cz> Signed-off-by: Andrey Konovalov <adech.fo@gmail.com> Cc: Dmitry Vyukov <dvyukov@google.com> Cc: Konstantin Serebryany <kcc@google.com> Cc: Dmitry Chernenkov <dmitryc@google.com> Cc: Yuri Gribov <tetra2005@gmail.com> Cc: Konstantin Khlebnikov <koct9i@gmail.com> Cc: Sasha Levin <sasha.levin@oracle.com> Cc: Christoph Lameter <cl@linux.com> Cc: Joonsoo Kim <iamjoonsoo.kim@lge.com> Cc: Dave Hansen <dave.hansen@intel.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Ingo Molnar <mingo@elte.hu> Cc: Thomas Gleixner <tglx@linutronix.de> Cc: "H. Peter Anvin" <hpa@zytor.com> Cc: Christoph Lameter <cl@linux.com> Cc: Pekka Enberg <penberg@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Stephen Rothwell <sfr@canb.auug.org.au> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-02-13 14:39:17 -08:00
memcg: punt high overage reclaim to return-to-userland path Currently, try_charge() tries to reclaim memory synchronously when the high limit is breached; however, if the allocation doesn't have __GFP_WAIT, synchronous reclaim is skipped. If a process performs only speculative allocations, it can blow way past the high limit. This is actually easily reproducible by simply doing "find /". slab/slub allocator tries speculative allocations first, so as long as there's memory which can be consumed without blocking, it can keep allocating memory regardless of the high limit. This patch makes try_charge() always punt the over-high reclaim to the return-to-userland path. If try_charge() detects that high limit is breached, it adds the overage to current->memcg_nr_pages_over_high and schedules execution of mem_cgroup_handle_over_high() which performs synchronous reclaim from the return-to-userland path. As long as kernel doesn't have a run-away allocation spree, this should provide enough protection while making kmemcg behave more consistently. It also has the following benefits. - All over-high reclaims can use GFP_KERNEL regardless of the specific gfp mask in use, e.g. GFP_NOFS, when the limit was breached. - It copes with prio inversion. Previously, a low-prio task with small memory.high might perform over-high reclaim with a bunch of locks held. If a higher prio task needed any of these locks, it would have to wait until the low prio task finished reclaim and released the locks. By handing over-high reclaim to the task exit path this issue can be avoided. Signed-off-by: Tejun Heo <tj@kernel.org> Acked-by: Michal Hocko <mhocko@kernel.org> Reviewed-by: Vladimir Davydov <vdavydov@parallels.com> Acked-by: Johannes Weiner <hannes@cmpxchg.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2015-11-05 18:46:11 -08:00
memcg: coalesce uncharge during unmap/truncate In massive parallel enviroment, res_counter can be a performance bottleneck. One strong techinque to reduce lock contention is reducing calls by coalescing some amount of calls into one. Considering charge/uncharge chatacteristic, - charge is done one by one via demand-paging. - uncharge is done by - in chunk at munmap, truncate, exit, execve... - one by one via vmscan/paging. It seems we have a chance to coalesce uncharges for improving scalability at unmap/truncation. This patch is a for coalescing uncharge. For avoiding scattering memcg's structure to functions under /mm, this patch adds memcg batch uncharge information to the task. A reason for per-task batching is for making use of caller's context information. We do batched uncharge (deleyed uncharge) when truncation/unmap occurs but do direct uncharge when uncharge is called by memory reclaim (vmscan.c). The degree of coalescing depends on callers - at invalidate/trucate... pagevec size - at unmap ....ZAP_BLOCK_SIZE (memory itself will be freed in this degree.) Then, we'll not coalescing too much. On x86-64 8cpu server, I tested overheads of memcg at page fault by running a program which does map/fault/unmap in a loop. Running a task per a cpu by taskset and see sum of the number of page faults in 60secs. [without memcg config] 40156968 page-faults # 0.085 M/sec ( +- 0.046% ) 27.67 cache-miss/faults [root cgroup] 36659599 page-faults # 0.077 M/sec ( +- 0.247% ) 31.58 miss/faults [in a child cgroup] 18444157 page-faults # 0.039 M/sec ( +- 0.133% ) 69.96 miss/faults [child with this patch] 27133719 page-faults # 0.057 M/sec ( +- 0.155% ) 47.16 miss/faults We can see some amounts of improvement. (root cgroup doesn't affected by this patch) Another patch for "charge" will follow this and above will be improved more. Changelog(since 2009/10/02): - renamed filed of memcg_batch (as pages to bytes, memsw to memsw_bytes) - some clean up and commentary/description updates. - added initialize code to copy_process(). (possible bug fix) Changelog(old): - fixed !CONFIG_MEM_CGROUP case. - rebased onto the latest mmotm + softlimit fix patches. - unified patch for callers - added commetns. - make ->do_batch as bool. - removed css_get() at el. We don't need it. Signed-off-by: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com> Cc: Balbir Singh <balbir@in.ibm.com> Cc: Daisuke Nishimura <nishimura@mxp.nes.nec.co.jp> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2009-12-15 16:47:03 -08:00
uprobes/core: Handle breakpoint and singlestep exceptions Uprobes uses exception notifiers to get to know if a thread hit a breakpoint or a singlestep exception. When a thread hits a uprobe or is singlestepping post a uprobe hit, the uprobe exception notifier sets its TIF_UPROBE bit, which will then be checked on its return to userspace path (do_notify_resume() ->uprobe_notify_resume()), where the consumers handlers are run (in task context) based on the defined filters. Uprobe hits are thread specific and hence we need to maintain information about if a task hit a uprobe, what uprobe was hit, the slot where the original instruction was copied for xol so that it can be singlestepped with appropriate fixups. In some cases, special care is needed for instructions that are executed out of line (xol). These are architecture specific artefacts, such as handling RIP relative instructions on x86_64. Since the instruction at which the uprobe was inserted is executed out of line, architecture specific fixups are added so that the thread continues normal execution in the presence of a uprobe. Postpone the signals until we execute the probed insn. post_xol() path does a recalc_sigpending() before return to user-mode, this ensures the signal can't be lost. Uprobes relies on DIE_DEBUG notification to notify if a singlestep is complete. Adds x86 specific uprobe exception notifiers and appropriate hooks needed to determine a uprobe hit and subsequent post processing. Add requisite x86 fixups for xol for uprobes. Specific cases needing fixups include relative jumps (x86_64), calls, etc. Where possible, we check and skip singlestepping the breakpointed instructions. For now we skip single byte as well as few multibyte nop instructions. However this can be extended to other instructions too. Credits to Oleg Nesterov for suggestions/patches related to signal, breakpoint, singlestep handling code. Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Ananth N Mavinakayanahalli <ananth@in.ibm.com> Cc: Jim Keniston <jkenisto@linux.vnet.ibm.com> Cc: Linux-mm <linux-mm@kvack.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Christoph Hellwig <hch@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Arnaldo Carvalho de Melo <acme@infradead.org> Cc: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com> Cc: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/20120313180011.29771.89027.sendpatchset@srdronam.in.ibm.com [ Performed various cleanliness edits ] Signed-off-by: Ingo Molnar <mingo@elte.hu>
2012-03-13 23:30:11 +05:30
uprobes/core: Handle breakpoint and singlestep exceptions Uprobes uses exception notifiers to get to know if a thread hit a breakpoint or a singlestep exception. When a thread hits a uprobe or is singlestepping post a uprobe hit, the uprobe exception notifier sets its TIF_UPROBE bit, which will then be checked on its return to userspace path (do_notify_resume() ->uprobe_notify_resume()), where the consumers handlers are run (in task context) based on the defined filters. Uprobe hits are thread specific and hence we need to maintain information about if a task hit a uprobe, what uprobe was hit, the slot where the original instruction was copied for xol so that it can be singlestepped with appropriate fixups. In some cases, special care is needed for instructions that are executed out of line (xol). These are architecture specific artefacts, such as handling RIP relative instructions on x86_64. Since the instruction at which the uprobe was inserted is executed out of line, architecture specific fixups are added so that the thread continues normal execution in the presence of a uprobe. Postpone the signals until we execute the probed insn. post_xol() path does a recalc_sigpending() before return to user-mode, this ensures the signal can't be lost. Uprobes relies on DIE_DEBUG notification to notify if a singlestep is complete. Adds x86 specific uprobe exception notifiers and appropriate hooks needed to determine a uprobe hit and subsequent post processing. Add requisite x86 fixups for xol for uprobes. Specific cases needing fixups include relative jumps (x86_64), calls, etc. Where possible, we check and skip singlestepping the breakpointed instructions. For now we skip single byte as well as few multibyte nop instructions. However this can be extended to other instructions too. Credits to Oleg Nesterov for suggestions/patches related to signal, breakpoint, singlestep handling code. Signed-off-by: Srikar Dronamraju <srikar@linux.vnet.ibm.com> Cc: Linus Torvalds <torvalds@linux-foundation.org> Cc: Ananth N Mavinakayanahalli <ananth@in.ibm.com> Cc: Jim Keniston <jkenisto@linux.vnet.ibm.com> Cc: Linux-mm <linux-mm@kvack.org> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: Christoph Hellwig <hch@infradead.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Arnaldo Carvalho de Melo <acme@infradead.org> Cc: Masami Hiramatsu <masami.hiramatsu.pt@hitachi.com> Cc: Peter Zijlstra <peterz@infradead.org> Link: http://lkml.kernel.org/r/20120313180011.29771.89027.sendpatchset@srdronam.in.ibm.com [ Performed various cleanliness edits ] Signed-off-by: Ingo Molnar <mingo@elte.hu>
2012-03-13 23:30:11 +05:30
mm, oom_reaper: implement OOM victims queuing wake_oom_reaper has allowed only 1 oom victim to be queued. The main reason for that was the simplicity as other solutions would require some way of queuing. The current approach is racy and that was deemed sufficient as the oom_reaper is considered a best effort approach to help with oom handling when the OOM victim cannot terminate in a reasonable time. The race could lead to missing an oom victim which can get stuck out_of_memory wake_oom_reaper cmpxchg // OK oom_reaper oom_reap_task __oom_reap_task oom_victim terminates atomic_inc_not_zero // fail out_of_memory wake_oom_reaper cmpxchg // fails task_to_reap = NULL This race requires 2 OOM invocations in a short time period which is not very likely but certainly not impossible. E.g. the original victim might have not released a lot of memory for some reason. The situation would improve considerably if wake_oom_reaper used a more robust queuing. This is what this patch implements. This means adding oom_reaper_list list_head into task_struct (eat a hole before embeded thread_struct for that purpose) and a oom_reaper_lock spinlock for queuing synchronization. wake_oom_reaper will then add the task on the queue and oom_reaper will dequeue it. Signed-off-by: Michal Hocko <mhocko@suse.com> Cc: Vladimir Davydov <vdavydov@virtuozzo.com> Cc: Andrea Argangeli <andrea@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Rik van Riel <riel@redhat.com> Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-03-25 14:20:33 -07:00
mm, oom_reaper: implement OOM victims queuing wake_oom_reaper has allowed only 1 oom victim to be queued. The main reason for that was the simplicity as other solutions would require some way of queuing. The current approach is racy and that was deemed sufficient as the oom_reaper is considered a best effort approach to help with oom handling when the OOM victim cannot terminate in a reasonable time. The race could lead to missing an oom victim which can get stuck out_of_memory wake_oom_reaper cmpxchg // OK oom_reaper oom_reap_task __oom_reap_task oom_victim terminates atomic_inc_not_zero // fail out_of_memory wake_oom_reaper cmpxchg // fails task_to_reap = NULL This race requires 2 OOM invocations in a short time period which is not very likely but certainly not impossible. E.g. the original victim might have not released a lot of memory for some reason. The situation would improve considerably if wake_oom_reaper used a more robust queuing. This is what this patch implements. This means adding oom_reaper_list list_head into task_struct (eat a hole before embeded thread_struct for that purpose) and a oom_reaper_lock spinlock for queuing synchronization. wake_oom_reaper will then add the task on the queue and oom_reaper will dequeue it. Signed-off-by: Michal Hocko <mhocko@suse.com> Cc: Vladimir Davydov <vdavydov@virtuozzo.com> Cc: Andrea Argangeli <andrea@kernel.org> Cc: David Rientjes <rientjes@google.com> Cc: Hugh Dickins <hughd@google.com> Cc: Johannes Weiner <hannes@cmpxchg.org> Cc: Mel Gorman <mgorman@suse.de> Cc: Oleg Nesterov <oleg@redhat.com> Cc: Rik van Riel <riel@redhat.com> Cc: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2016-03-25 14:20:33 -07:00
livepatch: change to a per-task consistency model Change livepatch to use a basic per-task consistency model. This is the foundation which will eventually enable us to patch those ~10% of security patches which change function or data semantics. This is the biggest remaining piece needed to make livepatch more generally useful. This code stems from the design proposal made by Vojtech [1] in November 2014. It's a hybrid of kGraft and kpatch: it uses kGraft's per-task consistency and syscall barrier switching combined with kpatch's stack trace switching. There are also a number of fallback options which make it quite flexible. Patches are applied on a per-task basis, when the task is deemed safe to switch over. When a patch is enabled, livepatch enters into a transition state where tasks are converging to the patched state. Usually this transition state can complete in a few seconds. The same sequence occurs when a patch is disabled, except the tasks converge from the patched state to the unpatched state. An interrupt handler inherits the patched state of the task it interrupts. The same is true for forked tasks: the child inherits the patched state of the parent. Livepatch uses several complementary approaches to determine when it's safe to patch tasks: 1. The first and most effective approach is stack checking of sleeping tasks. If no affected functions are on the stack of a given task, the task is patched. In most cases this will patch most or all of the tasks on the first try. Otherwise it'll keep trying periodically. This option is only available if the architecture has reliable stacks (HAVE_RELIABLE_STACKTRACE). 2. The second approach, if needed, is kernel exit switching. A task is switched when it returns to user space from a system call, a user space IRQ, or a signal. It's useful in the following cases: a) Patching I/O-bound user tasks which are sleeping on an affected function. In this case you have to send SIGSTOP and SIGCONT to force it to exit the kernel and be patched. b) Patching CPU-bound user tasks. If the task is highly CPU-bound then it will get patched the next time it gets interrupted by an IRQ. c) In the future it could be useful for applying patches for architectures which don't yet have HAVE_RELIABLE_STACKTRACE. In this case you would have to signal most of the tasks on the system. However this isn't supported yet because there's currently no way to patch kthreads without HAVE_RELIABLE_STACKTRACE. 3. For idle "swapper" tasks, since they don't ever exit the kernel, they instead have a klp_update_patch_state() call in the idle loop which allows them to be patched before the CPU enters the idle state. (Note there's not yet such an approach for kthreads.) All the above approaches may be skipped by setting the 'immediate' flag in the 'klp_patch' struct, which will disable per-task consistency and patch all tasks immediately. This can be useful if the patch doesn't change any function or data semantics. Note that, even with this flag set, it's possible that some tasks may still be running with an old version of the function, until that function returns. There's also an 'immediate' flag in the 'klp_func' struct which allows you to specify that certain functions in the patch can be applied without per-task consistency. This might be useful if you want to patch a common function like schedule(), and the function change doesn't need consistency but the rest of the patch does. For architectures which don't have HAVE_RELIABLE_STACKTRACE, the user must set patch->immediate which causes all tasks to be patched immediately. This option should be used with care, only when the patch doesn't change any function or data semantics. In the future, architectures which don't have HAVE_RELIABLE_STACKTRACE may be allowed to use per-task consistency if we can come up with another way to patch kthreads. The /sys/kernel/livepatch/<patch>/transition file shows whether a patch is in transition. Only a single patch (the topmost patch on the stack) can be in transition at a given time. A patch can remain in transition indefinitely, if any of the tasks are stuck in the initial patch state. A transition can be reversed and effectively canceled by writing the opposite value to the /sys/kernel/livepatch/<patch>/enabled file while the transition is in progress. Then all the tasks will attempt to converge back to the original patch state. [1] https://lkml.kernel.org/r/20141107140458.GA21774@suse.cz Signed-off-by: Josh Poimboeuf <jpoimboe@redhat.com> Acked-by: Miroslav Benes <mbenes@suse.cz> Acked-by: Ingo Molnar <mingo@kernel.org> # for the scheduler changes Signed-off-by: Jiri Kosina <jkosina@suse.cz>
2017-02-13 19:42:40 -06:00
Merge branch 'next' of git://git.kernel.org/pub/scm/linux/kernel/git/jmorris/linux-security Pull security subsystem updates from James Morris: "Highlights: IMA: - provide ">" and "<" operators for fowner/uid/euid rules KEYS: - add a system blacklist keyring - add KEYCTL_RESTRICT_KEYRING, exposes keyring link restriction functionality to userland via keyctl() LSM: - harden LSM API with __ro_after_init - add prlmit security hook, implement for SELinux - revive security_task_alloc hook TPM: - implement contextual TPM command 'spaces'" * 'next' of git://git.kernel.org/pub/scm/linux/kernel/git/jmorris/linux-security: (98 commits) tpm: Fix reference count to main device tpm_tis: convert to using locality callbacks tpm: fix handling of the TPM 2.0 event logs tpm_crb: remove a cruft constant keys: select CONFIG_CRYPTO when selecting DH / KDF apparmor: Make path_max parameter readonly apparmor: fix parameters so that the permission test is bypassed at boot apparmor: fix invalid reference to index variable of iterator line 836 apparmor: use SHASH_DESC_ON_STACK security/apparmor/lsm.c: set debug messages apparmor: fix boolreturn.cocci warnings Smack: Use GFP_KERNEL for smk_netlbl_mls(). smack: fix double free in smack_parse_opts_str() KEYS: add SP800-56A KDF support for DH KEYS: Keyring asymmetric key restrict method with chaining KEYS: Restrict asymmetric key linkage using a specific keychain KEYS: Add a lookup_restriction function for the asymmetric key type KEYS: Add KEYCTL_RESTRICT_KEYRING KEYS: Consistent ordering for __key_link_begin and restrict check KEYS: Add an optional lookup_restriction hook to key_type ...
2017-05-03 08:50:52 -07:00
LSM: Revive security_task_alloc() hook and per "struct task_struct" security blob. We switched from "struct task_struct"->security to "struct cred"->security in Linux 2.6.29. But not all LSM modules were happy with that change. TOMOYO LSM module is an example which want to use per "struct task_struct" security blob, for TOMOYO's security context is defined based on "struct task_struct" rather than "struct cred". AppArmor LSM module is another example which want to use it, for AppArmor is currently abusing the cred a little bit to store the change_hat and setexeccon info. Although security_task_free() hook was revived in Linux 3.4 because Yama LSM module wanted to release per "struct task_struct" security blob, security_task_alloc() hook and "struct task_struct"->security field were not revived. Nowadays, we are getting proposals of lightweight LSM modules which want to use per "struct task_struct" security blob. We are already allowing multiple concurrent LSM modules (up to one fully armored module which uses "struct cred"->security field or exclusive hooks like security_xfrm_state_pol_flow_match(), plus unlimited number of lightweight modules which do not use "struct cred"->security nor exclusive hooks) as long as they are built into the kernel. But this patch does not implement variable length "struct task_struct"->security field which will become needed when multiple LSM modules want to use "struct task_struct"-> security field. Although it won't be difficult to implement variable length "struct task_struct"->security field, let's think about it after we merged this patch. Signed-off-by: Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp> Acked-by: John Johansen <john.johansen@canonical.com> Acked-by: Serge Hallyn <serge@hallyn.com> Acked-by: Casey Schaufler <casey@schaufler-ca.com> Tested-by: Djalal Harouni <tixxdz@gmail.com> Acked-by: José Bollo <jobol@nonadev.net> Cc: Paul Moore <paul@paul-moore.com> Cc: Stephen Smalley <sds@tycho.nsa.gov> Cc: Eric Paris <eparis@parisplace.org> Cc: Kees Cook <keescook@chromium.org> Cc: James Morris <james.l.morris@oracle.com> Cc: José Bollo <jobol@nonadev.net> Signed-off-by: James Morris <james.l.morris@oracle.com>
2017-03-24 20:46:33 +09:00
[PATCH] pid: implement access helpers for a tacks various process groups In the last round of cleaning up the pid hash table a more general struct pid was introduced, that can be referenced counted. With the more general struct pid most if not all places where we store a pid_t we can now store a struct pid * and remove the need for a hash table lookup, and avoid any possible problems with pid roll over. Looking forward to the pid namespaces struct pid * gives us an absolute form a pid so we can compare and use them without caring which pid namespace we are in. This patchset introduces the infrastructure needed to use struct pid instead of pid_t, and then it goes on to convert two different kernel users that currently store a pid_t value. There are a lot more places to go but this is enough to get the basic idea. Before we can merge a pid namespace patch all of the kernel pid_t users need to be examined. Those that deal with user space processes need to be converted to using a struct pid *. Those that deal with kernel processes need to converted to using the kthread api. A rare few that only use their current processes pid values get to be left alone. This patch: task_session returns the struct pid of a tasks session. task_pgrp returns the struct pid of a tasks process group. task_tgid returns the struct pid of a tasks thread group. task_pid returns the struct pid of a tasks process id. These can be used to avoid unnecessary hash table lookups, and to implement safe pid comparisions in the face of a pid namespace. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-10-02 02:17:09 -07:00
[PATCH] pid: implement access helpers for a tacks various process groups In the last round of cleaning up the pid hash table a more general struct pid was introduced, that can be referenced counted. With the more general struct pid most if not all places where we store a pid_t we can now store a struct pid * and remove the need for a hash table lookup, and avoid any possible problems with pid roll over. Looking forward to the pid namespaces struct pid * gives us an absolute form a pid so we can compare and use them without caring which pid namespace we are in. This patchset introduces the infrastructure needed to use struct pid instead of pid_t, and then it goes on to convert two different kernel users that currently store a pid_t value. There are a lot more places to go but this is enough to get the basic idea. Before we can merge a pid namespace patch all of the kernel pid_t users need to be examined. Those that deal with user space processes need to be converted to using a struct pid *. Those that deal with kernel processes need to converted to using the kthread api. A rare few that only use their current processes pid values get to be left alone. This patch: task_session returns the struct pid of a tasks session. task_pgrp returns the struct pid of a tasks process group. task_tgid returns the struct pid of a tasks thread group. task_pid returns the struct pid of a tasks process id. These can be used to avoid unnecessary hash table lookups, and to implement safe pid comparisions in the face of a pid namespace. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-10-02 02:17:09 -07:00
[PATCH] pid: implement access helpers for a tacks various process groups In the last round of cleaning up the pid hash table a more general struct pid was introduced, that can be referenced counted. With the more general struct pid most if not all places where we store a pid_t we can now store a struct pid * and remove the need for a hash table lookup, and avoid any possible problems with pid roll over. Looking forward to the pid namespaces struct pid * gives us an absolute form a pid so we can compare and use them without caring which pid namespace we are in. This patchset introduces the infrastructure needed to use struct pid instead of pid_t, and then it goes on to convert two different kernel users that currently store a pid_t value. There are a lot more places to go but this is enough to get the basic idea. Before we can merge a pid namespace patch all of the kernel pid_t users need to be examined. Those that deal with user space processes need to be converted to using a struct pid *. Those that deal with kernel processes need to converted to using the kthread api. A rare few that only use their current processes pid values get to be left alone. This patch: task_session returns the struct pid of a tasks session. task_pgrp returns the struct pid of a tasks process group. task_tgid returns the struct pid of a tasks thread group. task_pid returns the struct pid of a tasks process id. These can be used to avoid unnecessary hash table lookups, and to implement safe pid comparisions in the face of a pid namespace. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-10-02 02:17:09 -07:00
[PATCH] pid: implement access helpers for a tacks various process groups In the last round of cleaning up the pid hash table a more general struct pid was introduced, that can be referenced counted. With the more general struct pid most if not all places where we store a pid_t we can now store a struct pid * and remove the need for a hash table lookup, and avoid any possible problems with pid roll over. Looking forward to the pid namespaces struct pid * gives us an absolute form a pid so we can compare and use them without caring which pid namespace we are in. This patchset introduces the infrastructure needed to use struct pid instead of pid_t, and then it goes on to convert two different kernel users that currently store a pid_t value. There are a lot more places to go but this is enough to get the basic idea. Before we can merge a pid namespace patch all of the kernel pid_t users need to be examined. Those that deal with user space processes need to be converted to using a struct pid *. Those that deal with kernel processes need to converted to using the kthread api. A rare few that only use their current processes pid values get to be left alone. This patch: task_session returns the struct pid of a tasks session. task_pgrp returns the struct pid of a tasks process group. task_tgid returns the struct pid of a tasks thread group. task_pid returns the struct pid of a tasks process id. These can be used to avoid unnecessary hash table lookups, and to implement safe pid comparisions in the face of a pid namespace. Signed-off-by: Eric W. Biederman <ebiederm@xmission.com> Signed-off-by: Andrew Morton <akpm@osdl.org> Signed-off-by: Linus Torvalds <torvalds@osdl.org>
2006-10-02 02:17:09 -07:00
pid namespaces: define is_global_init() and is_container_init() is_init() is an ambiguous name for the pid==1 check. Split it into is_global_init() and is_container_init(). A cgroup init has it's tsk->pid == 1. A global init also has it's tsk->pid == 1 and it's active pid namespace is the init_pid_ns. But rather than check the active pid namespace, compare the task structure with 'init_pid_ns.child_reaper', which is initialized during boot to the /sbin/init process and never changes. Changelog: 2.6.22-rc4-mm2-pidns1: - Use 'init_pid_ns.child_reaper' to determine if a given task is the global init (/sbin/init) process. This would improve performance and remove dependence on the task_pid(). 2.6.21-mm2-pidns2: - [Sukadev Bhattiprolu] Changed is_container_init() calls in {powerpc, ppc,avr32}/traps.c for the _exception() call to is_global_init(). This way, we kill only the cgroup if the cgroup's init has a bug rather than force a kernel panic. [akpm@linux-foundation.org: fix comment] [sukadev@us.ibm.com: Use is_global_init() in arch/m32r/mm/fault.c] [bunk@stusta.de: kernel/pid.c: remove unused exports] [sukadev@us.ibm.com: Fix capability.c to work with threaded init] Signed-off-by: Serge E. Hallyn <serue@us.ibm.com> Signed-off-by: Sukadev Bhattiprolu <sukadev@us.ibm.com> Acked-by: Pavel Emelianov <xemul@openvz.org> Cc: Eric W. Biederman <ebiederm@xmission.com> Cc: Cedric Le Goater <clg@fr.ibm.com> Cc: Dave Hansen <haveblue@us.ibm.com> Cc: Herbert Poetzel <herbert@13thfloor.at> Cc: Kirill Korotaev <dev@sw.ru> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2007-10-18 23:39:52 -07:00
mm: introduce memalloc_nofs_{save,restore} API GFP_NOFS context is used for the following 5 reasons currently: - to prevent from deadlocks when the lock held by the allocation context would be needed during the memory reclaim - to prevent from stack overflows during the reclaim because the allocation is performed from a deep context already - to prevent lockups when the allocation context depends on other reclaimers to make a forward progress indirectly - just in case because this would be safe from the fs POV - silence lockdep false positives Unfortunately overuse of this allocation context brings some problems to the MM. Memory reclaim is much weaker (especially during heavy FS metadata workloads), OOM killer cannot be invoked because the MM layer doesn't have enough information about how much memory is freeable by the FS layer. In many cases it is far from clear why the weaker context is even used and so it might be used unnecessarily. We would like to get rid of those as much as possible. One way to do that is to use the flag in scopes rather than isolated cases. Such a scope is declared when really necessary, tracked per task and all the allocation requests from within the context will simply inherit the GFP_NOFS semantic. Not only this is easier to understand and maintain because there are much less problematic contexts than specific allocation requests, this also helps code paths where FS layer interacts with other layers (e.g. crypto, security modules, MM etc...) and there is no easy way to convey the allocation context between the layers. Introduce memalloc_nofs_{save,restore} API to control the scope of GFP_NOFS allocation context. This is basically copying memalloc_noio_{save,restore} API we have for other restricted allocation context GFP_NOIO. The PF_MEMALLOC_NOFS flag already exists and it is just an alias for PF_FSTRANS which has been xfs specific until recently. There are no more PF_FSTRANS users anymore so let's just drop it. PF_MEMALLOC_NOFS is now checked in the MM layer and drops __GFP_FS implicitly same as PF_MEMALLOC_NOIO drops __GFP_IO. memalloc_noio_flags is renamed to current_gfp_context because it now cares about both PF_MEMALLOC_NOFS and PF_MEMALLOC_NOIO contexts. Xfs code paths preserve their semantic. kmem_flags_convert() doesn't need to evaluate the flag anymore. This patch shouldn't introduce any functional changes. Let's hope that filesystems will drop direct GFP_NOFS (resp. ~__GFP_FS) usage as much as possible and only use a properly documented memalloc_nofs_{save,restore} checkpoints where they are appropriate. [akpm@linux-foundation.org: fix comment typo, reflow comment] Link: http://lkml.kernel.org/r/20170306131408.9828-5-mhocko@kernel.org Signed-off-by: Michal Hocko <mhocko@suse.com> Acked-by: Vlastimil Babka <vbabka@suse.cz> Cc: Dave Chinner <david@fromorbit.com> Cc: Theodore Ts'o <tytso@mit.edu> Cc: Chris Mason <clm@fb.com> Cc: David Sterba <dsterba@suse.cz> Cc: Jan Kara <jack@suse.cz> Cc: Brian Foster <bfoster@redhat.com> Cc: Darrick J. Wong <darrick.wong@oracle.com> Cc: Nikolay Borisov <nborisov@suse.com> Cc: Peter Zijlstra <peterz@infradead.org> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-05-03 14:53:15 -07:00
mm: allow PF_MEMALLOC from softirq context This is needed to allow network softirq packet processing to make use of PF_MEMALLOC. Currently softirq context cannot use PF_MEMALLOC due to it not being associated with a task, and therefore not having task flags to fiddle with - thus the gfp to alloc flag mapping ignores the task flags when in interrupts (hard or soft) context. Allowing softirqs to make use of PF_MEMALLOC therefore requires some trickery. This patch borrows the task flags from whatever process happens to be preempted by the softirq. It then modifies the gfp to alloc flags mapping to not exclude task flags in softirq context, and modify the softirq code to save, clear and restore the PF_MEMALLOC flag. The save and clear, ensures the preempted task's PF_MEMALLOC flag doesn't leak into the softirq. The restore ensures a softirq's PF_MEMALLOC flag cannot leak back into the preempted process. This should be safe due to the following reasons Softirqs can run on multiple CPUs sure but the same task should not be executing the same softirq code. Neither should the softirq handler be preempted by any other softirq handler so the flags should not leak to an unrelated softirq. Softirqs re-enable hardware interrupts in __do_softirq() so can be preempted by hardware interrupts so PF_MEMALLOC is inherited by the hard IRQ. However, this is similar to a process in reclaim being preempted by a hardirq. While PF_MEMALLOC is set, gfp_to_alloc_flags() distinguishes between hard and soft irqs and avoids giving a hardirq the ALLOC_NO_WATERMARKS flag. If the softirq is deferred to ksoftirq then its flags may be used instead of a normal tasks but as the softirq cannot be preempted, the PF_MEMALLOC flag does not leak to other code by accident. [davem@davemloft.net: Document why PF_MEMALLOC is safe] Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by: Mel Gorman <mgorman@suse.de> Cc: David Miller <davem@davemloft.net> Cc: Neil Brown <neilb@suse.de> Cc: Mike Christie <michaelc@cs.wisc.edu> Cc: Eric B Munson <emunson@mgebm.net> Cc: Eric Dumazet <eric.dumazet@gmail.com> Cc: Sebastian Andrzej Siewior <sebastian@breakpoint.cc> Cc: Mel Gorman <mgorman@suse.de> Cc: Christoph Lameter <cl@linux.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-07-31 16:44:07 -07:00
mm: allow PF_MEMALLOC from softirq context This is needed to allow network softirq packet processing to make use of PF_MEMALLOC. Currently softirq context cannot use PF_MEMALLOC due to it not being associated with a task, and therefore not having task flags to fiddle with - thus the gfp to alloc flag mapping ignores the task flags when in interrupts (hard or soft) context. Allowing softirqs to make use of PF_MEMALLOC therefore requires some trickery. This patch borrows the task flags from whatever process happens to be preempted by the softirq. It then modifies the gfp to alloc flags mapping to not exclude task flags in softirq context, and modify the softirq code to save, clear and restore the PF_MEMALLOC flag. The save and clear, ensures the preempted task's PF_MEMALLOC flag doesn't leak into the softirq. The restore ensures a softirq's PF_MEMALLOC flag cannot leak back into the preempted process. This should be safe due to the following reasons Softirqs can run on multiple CPUs sure but the same task should not be executing the same softirq code. Neither should the softirq handler be preempted by any other softirq handler so the flags should not leak to an unrelated softirq. Softirqs re-enable hardware interrupts in __do_softirq() so can be preempted by hardware interrupts so PF_MEMALLOC is inherited by the hard IRQ. However, this is similar to a process in reclaim being preempted by a hardirq. While PF_MEMALLOC is set, gfp_to_alloc_flags() distinguishes between hard and soft irqs and avoids giving a hardirq the ALLOC_NO_WATERMARKS flag. If the softirq is deferred to ksoftirq then its flags may be used instead of a normal tasks but as the softirq cannot be preempted, the PF_MEMALLOC flag does not leak to other code by accident. [davem@davemloft.net: Document why PF_MEMALLOC is safe] Signed-off-by: Peter Zijlstra <a.p.zijlstra@chello.nl> Signed-off-by: Mel Gorman <mgorman@suse.de> Cc: David Miller <davem@davemloft.net> Cc: Neil Brown <neilb@suse.de> Cc: Mike Christie <michaelc@cs.wisc.edu> Cc: Eric B Munson <emunson@mgebm.net> Cc: Eric Dumazet <eric.dumazet@gmail.com> Cc: Sebastian Andrzej Siewior <sebastian@breakpoint.cc> Cc: Mel Gorman <mgorman@suse.de> Cc: Christoph Lameter <cl@linux.com> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2012-07-31 16:44:07 -07:00
exec: avoid gcc-8 warning for get_task_comm gcc-8 warns about using strncpy() with the source size as the limit: fs/exec.c:1223:32: error: argument to 'sizeof' in 'strncpy' call is the same expression as the source; did you mean to use the size of the destination? [-Werror=sizeof-pointer-memaccess] This is indeed slightly suspicious, as it protects us from source arguments without NUL-termination, but does not guarantee that the destination is terminated. This keeps the strncpy() to ensure we have properly padded target buffer, but ensures that we use the correct length, by passing the actual length of the destination buffer as well as adding a build-time check to ensure it is exactly TASK_COMM_LEN. There are only 23 callsites which I all reviewed to ensure this is currently the case. We could get away with doing only the check or passing the right length, but it doesn't hurt to do both. Link: http://lkml.kernel.org/r/20171205151724.1764896-1-arnd@arndb.de Signed-off-by: Arnd Bergmann <arnd@arndb.de> Suggested-by: Kees Cook <keescook@chromium.org> Acked-by: Kees Cook <keescook@chromium.org> Acked-by: Ingo Molnar <mingo@kernel.org> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: Peter Zijlstra <peterz@infradead.org> Cc: Serge Hallyn <serge@hallyn.com> Cc: James Morris <james.l.morris@oracle.com> Cc: Aleksa Sarai <asarai@suse.de> Cc: "Eric W. Biederman" <ebiederm@xmission.com> Cc: Frederic Weisbecker <frederic@kernel.org> Cc: Thomas Gleixner <tglx@linutronix.de> Signed-off-by: Andrew Morton <akpm@linux-foundation.org> Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
2017-12-14 15:32:41 -08:00
rseq: Introduce restartable sequences system call Expose a new system call allowing each thread to register one userspace memory area to be used as an ABI between kernel and user-space for two purposes: user-space restartable sequences and quick access to read the current CPU number value from user-space. * Restartable sequences (per-cpu atomics) Restartables sequences allow user-space to perform update operations on per-cpu data without requiring heavy-weight atomic operations. The restartable critical sections (percpu atomics) work has been started by Paul Turner and Andrew Hunter. It lets the kernel handle restart of critical sections. [1] [2] The re-implementation proposed here brings a few simplifications to the ABI which facilitates porting to other architectures and speeds up the user-space fast path. Here are benchmarks of various rseq use-cases. Test hardware: arm32: ARMv7 Processor rev 4 (v7l) "Cubietruck", 2-core x86-64: Intel E5-2630 v3@2.40GHz, 16-core, hyperthreading The following benchmarks were all performed on a single thread. * Per-CPU statistic counter increment getcpu+atomic (ns/op) rseq (ns/op) speedup arm32: 344.0 31.4 11.0 x86-64: 15.3 2.0 7.7 * LTTng-UST: write event 32-bit header, 32-bit payload into tracer per-cpu buffer getcpu+atomic (ns/op) rseq (ns/op) speedup arm32: 2502.0 2250.0 1.1 x86-64: 117.4 98.0 1.2 * liburcu percpu: lock-unlock pair, dereference, read/compare word getcpu+atomic (ns/op) rseq (ns/op) speedup arm32: 751.0 128.5 5.8 x86-64: 53.4 28.6 1.9 * jemalloc memory allocator adapted to use rseq Using rseq with per-cpu memory pools in jemalloc at Facebook (based on rseq 2016 implementation): The production workload response-time has 1-2% gain avg. latency, and the P99 overall latency drops by 2-3%. * Reading the current CPU number Speeding up reading the current CPU number on which the caller thread is running is done by keeping the current CPU number up do date within the cpu_id field of the memory area registered by the thread. This is done by making scheduler preemption set the TIF_NOTIFY_RESUME flag on the current thread. Upon return to user-space, a notify-resume handler updates the current CPU value within the registered user-space memory area. User-space can then read the current CPU number directly from memory. Keeping the current cpu id in a memory area shared between kernel and user-space is an improvement over current mechanisms available to read the current CPU number, which has the following benefits over alternative approaches: - 35x speedup on ARM vs system call through glibc - 20x speedup on x86 compared to calling glibc, which calls vdso executing a "lsl" instruction, - 14x speedup on x86 compared to inlined "lsl" instruction, - Unlike vdso approaches, this cpu_id value can be read from an inline assembly, which makes it a useful building block for restartable sequences. - The approach of reading the cpu id through memory mapping shared between kernel and user-space is portable (e.g. ARM), which is not the case for the lsl-based x86 vdso. On x86, yet another possible approach would be to use the gs segment selector to point to user-space per-cpu data. This approach performs similarly to the cpu id cache, but it has two disadvantages: it is not portable, and it is incompatible with existing applications already using the gs segment selector for other purposes. Benchmarking various approaches for reading the current CPU number: ARMv7 Processor rev 4 (v7l) Machine model: Cubietruck - Baseline (empty loop): 8.4 ns - Read CPU from rseq cpu_id: 16.7 ns - Read CPU from rseq cpu_id (lazy register): 19.8 ns - glibc 2.19-0ubuntu6.6 getcpu: 301.8 ns - getcpu system call: 234.9 ns x86-64 Intel(R) Xeon(R) CPU E5-2630 v3 @ 2.40GHz: - Baseline (empty loop): 0.8 ns - Read CPU from rseq cpu_id: 0.8 ns - Read CPU from rseq cpu_id (lazy register): 0.8 ns - Read using gs segment selector: 0.8 ns - "lsl" inline assembly: 13.0 ns - glibc 2.19-0ubuntu6 getcpu: 16.6 ns - getcpu system call: 53.9 ns - Speed (benchmark taken on v8 of patchset) Running 10 runs of hackbench -l 100000 seems to indicate, contrary to expectations, that enabling CONFIG_RSEQ slightly accelerates the scheduler: Configuration: 2 sockets * 8-core Intel(R) Xeon(R) CPU E5-2630 v3 @ 2.40GHz (directly on hardware, hyperthreading disabled in BIOS, energy saving disabled in BIOS, turboboost disabled in BIOS, cpuidle.off=1 kernel parameter), with a Linux v4.6 defconfig+localyesconfig, restartable sequences series applied. * CONFIG_RSEQ=n avg.: 41.37 s std.dev.: 0.36 s * CONFIG_RSEQ=y avg.: 40.46 s std.dev.: 0.33 s - Size On x86-64, between CONFIG_RSEQ=n/y, the text size increase of vmlinux is 567 bytes, and the data size increase of vmlinux is 5696 bytes. [1] https://lwn.net/Articles/650333/ [2] http://www.linuxplumbersconf.org/2013/ocw/system/presentations/1695/original/LPC%20-%20PerCpu%20Atomics.pdf Signed-off-by: Mathieu Desnoyers <mathieu.desnoyers@efficios.com> Signed-off-by: Thomas Gleixner <tglx@linutronix.de> Acked-by: Peter Zijlstra (Intel) <peterz@infradead.org> Cc: Joel Fernandes <joelaf@google.com> Cc: Catalin Marinas <catalin.marinas@arm.com> Cc: Dave Watson <davejwatson@fb.com> Cc: Will Deacon <will.deacon@arm.com> Cc: Andi Kleen <andi@firstfloor.org> Cc: "H . Peter Anvin" <hpa@zytor.com> Cc: Chris Lameter <cl@linux.com> Cc: Russell King <linux@arm.linux.org.uk> Cc: Andrew Hunter <ahh@google.com> Cc: Michael Kerrisk <mtk.manpages@gmail.com> Cc: "Paul E . McKenney" <paulmck@linux.vnet.ibm.com> Cc: Paul Turner <pjt@google.com> Cc: Boqun Feng <boqun.feng@gmail.com> Cc: Josh Triplett <josh@joshtriplett.org> Cc: Steven Rostedt <rostedt@goodmis.org> Cc: Ben Maurer <bmaurer@fb.com> Cc: Alexander Viro <viro@zeniv.linux.org.uk> Cc: linux-api@vger.kernel.org Cc: Andy Lutomirski <luto@amacapital.net> Cc: Andrew Morton <akpm@linux-foundation.org> Cc: Linus Torvalds <torvalds@linux-foundation.org> Link: http://lkml.kernel.org/r/20151027235635.16059.11630.stgit@pjt-glaptop.roam.corp.google.com Link: http://lkml.kernel.org/r/20150624222609.6116.86035.stgit@kitami.mtv.corp.google.com Link: https://lkml.kernel.org/r/20180602124408.8430-3-mathieu.desnoyers@efficios.com
2018-06-02 08:43:54 -04:00