commit 00312419f0 upstream.
On systems using the hash MMU, there is a software SLB preload cache that
mirrors the entries loaded into the hardware SLB buffer. This preload
cache is subject to periodic eviction — typically after every 256 context
switches — to remove old entry.
To optimize performance, the kernel skips switch_mmu_context() in
switch_mm_irqs_off() when the prev and next mm_struct are the same.
However, on hash MMU systems, this can lead to inconsistencies between
the hardware SLB and the software preload cache.
If an SLB entry for a process is evicted from the software cache on one
CPU, and the same process later runs on another CPU without executing
switch_mmu_context(), the hardware SLB may retain stale entries. If the
kernel then attempts to reload that entry, it can trigger an SLB
multi-hit error.
The following timeline shows how stale SLB entries are created and can
cause a multi-hit error when a process moves between CPUs without a
MMU context switch.
CPU 0 CPU 1
----- -----
Process P
exec swapper/1
load_elf_binary
begin_new_exc
activate_mm
switch_mm_irqs_off
switch_mmu_context
switch_slb
/*
* This invalidates all
* the entries in the HW
* and setup the new HW
* SLB entries as per the
* preload cache.
*/
context_switch
sched_migrate_task migrates process P to cpu-1
Process swapper/0 context switch (to process P)
(uses mm_struct of Process P) switch_mm_irqs_off()
switch_slb
load_slb++
/*
* load_slb becomes 0 here
* and we evict an entry from
* the preload cache with
* preload_age(). We still
* keep HW SLB and preload
* cache in sync, that is
* because all HW SLB entries
* anyways gets evicted in
* switch_slb during SLBIA.
* We then only add those
* entries back in HW SLB,
* which are currently
* present in preload_cache
* (after eviction).
*/
load_elf_binary continues...
setup_new_exec()
slb_setup_new_exec()
sched_switch event
sched_migrate_task migrates
process P to cpu-0
context_switch from swapper/0 to Process P
switch_mm_irqs_off()
/*
* Since both prev and next mm struct are same we don't call
* switch_mmu_context(). This will cause the HW SLB and SW preload
* cache to go out of sync in preload_new_slb_context. Because there
* was an SLB entry which was evicted from both HW and preload cache
* on cpu-1. Now later in preload_new_slb_context(), when we will try
* to add the same preload entry again, we will add this to the SW
* preload cache and then will add it to the HW SLB. Since on cpu-0
* this entry was never invalidated, hence adding this entry to the HW
* SLB will cause a SLB multi-hit error.
*/
load_elf_binary continues...
START_THREAD
start_thread
preload_new_slb_context
/*
* This tries to add a new EA to preload cache which was earlier
* evicted from both cpu-1 HW SLB and preload cache. This caused the
* HW SLB of cpu-0 to go out of sync with the SW preload cache. The
* reason for this was, that when we context switched back on CPU-0,
* we should have ideally called switch_mmu_context() which will
* bring the HW SLB entries on CPU-0 in sync with SW preload cache
* entries by setting up the mmu context properly. But we didn't do
* that since the prev mm_struct running on cpu-0 was same as the
* next mm_struct (which is true for swapper / kernel threads). So
* now when we try to add this new entry into the HW SLB of cpu-0,
* we hit a SLB multi-hit error.
*/
WARNING: CPU: 0 PID: 1810970 at arch/powerpc/mm/book3s64/slb.c:62
assert_slb_presence+0x2c/0x50(48 results) 02:47:29 [20157/42149]
Modules linked in:
CPU: 0 UID: 0 PID: 1810970 Comm: dd Not tainted 6.16.0-rc3-dirty #12
VOLUNTARY
Hardware name: IBM pSeries (emulated by qemu) POWER8 (architected)
0x4d0200 0xf000004 of:SLOF,HEAD hv:linux,kvm pSeries
NIP: c00000000015426c LR: c0000000001543b4 CTR: 0000000000000000
REGS: c0000000497c77e0 TRAP: 0700 Not tainted (6.16.0-rc3-dirty)
MSR: 8000000002823033 <SF,VEC,VSX,FP,ME,IR,DR,RI,LE> CR: 28888482 XER: 00000000
CFAR: c0000000001543b0 IRQMASK: 3
<...>
NIP [c00000000015426c] assert_slb_presence+0x2c/0x50
LR [c0000000001543b4] slb_insert_entry+0x124/0x390
Call Trace:
0x7fffceb5ffff (unreliable)
preload_new_slb_context+0x100/0x1a0
start_thread+0x26c/0x420
load_elf_binary+0x1b04/0x1c40
bprm_execve+0x358/0x680
do_execveat_common+0x1f8/0x240
sys_execve+0x58/0x70
system_call_exception+0x114/0x300
system_call_common+0x160/0x2c4
>From the above analysis, during early exec the hardware SLB is cleared,
and entries from the software preload cache are reloaded into hardware
by switch_slb. However, preload_new_slb_context and slb_setup_new_exec
also attempt to load some of the same entries, which can trigger a
multi-hit. In most cases, these additional preloads simply hit existing
entries and add nothing new. Removing these functions avoids redundant
preloads and eliminates the multi-hit issue. This patch removes these
two functions.
We tested process switching performance using the context_switch
benchmark on POWER9/hash, and observed no regression.
Without this patch: 129041 ops/sec
With this patch: 129341 ops/sec
We also measured SLB faults during boot, and the counts are essentially
the same with and without this patch.
SLB faults without this patch: 19727
SLB faults with this patch: 19786
Fixes: 5434ae7462 ("powerpc/64s/hash: Add a SLB preload cache")
cc: stable@vger.kernel.org
Suggested-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Donet Tom <donettom@linux.ibm.com>
Signed-off-by: Ritesh Harjani (IBM) <ritesh.list@gmail.com>
Signed-off-by: Madhavan Srinivasan <maddy@linux.ibm.com>
Link: https://patch.msgid.link/0ac694ae683494fe8cadbd911a1a5018d5d3c541.1761834163.git.ritesh.list@gmail.com
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
895123c309