|
Jiri Slaby |
4505d7 |
From: Vlastimil Babka <vbabka@suse.cz>
|
|
Jiri Slaby |
4505d7 |
Date: Mon, 15 May 2017 15:13:16 +0200
|
|
Jiri Slaby |
4505d7 |
Subject: [PATCH] sched/numa: Use down_read_trylock() for the mmap_sem
|
|
Jiri Slaby |
4505d7 |
References: bnc#1060662
|
|
Thomas Zimmermann |
1d81d2 |
Patch-mainline: v4.12.1
|
|
Jiri Slaby |
4505d7 |
Git-commit: 8655d5497735b288f8a9b458bd22e7d1bf95bb61
|
|
Jiri Slaby |
4505d7 |
|
|
Jiri Slaby |
4505d7 |
commit 8655d5497735b288f8a9b458bd22e7d1bf95bb61 upstream.
|
|
Jiri Slaby |
4505d7 |
|
|
Jiri Slaby |
4505d7 |
A customer has reported a soft-lockup when running an intensive
|
|
Jiri Slaby |
4505d7 |
memory stress test, where the trace on multiple CPU's looks like this:
|
|
Jiri Slaby |
4505d7 |
|
|
Jiri Slaby |
4505d7 |
RIP: 0010:[<ffffffff810c53fe>]
|
|
Jiri Slaby |
4505d7 |
[<ffffffff810c53fe>] native_queued_spin_lock_slowpath+0x10e/0x190
|
|
Jiri Slaby |
4505d7 |
...
|
|
Jiri Slaby |
4505d7 |
Call Trace:
|
|
Jiri Slaby |
4505d7 |
[<ffffffff81182d07>] queued_spin_lock_slowpath+0x7/0xa
|
|
Jiri Slaby |
4505d7 |
[<ffffffff811bc331>] change_protection_range+0x3b1/0x930
|
|
Jiri Slaby |
4505d7 |
[<ffffffff811d4be8>] change_prot_numa+0x18/0x30
|
|
Jiri Slaby |
4505d7 |
[<ffffffff810adefe>] task_numa_work+0x1fe/0x310
|
|
Jiri Slaby |
4505d7 |
[<ffffffff81098322>] task_work_run+0x72/0x90
|
|
Jiri Slaby |
4505d7 |
|
|
Jiri Slaby |
4505d7 |
Further investigation showed that the lock contention here is pmd_lock().
|
|
Jiri Slaby |
4505d7 |
|
|
Jiri Slaby |
4505d7 |
The task_numa_work() function makes sure that only one thread is let to perform
|
|
Jiri Slaby |
4505d7 |
the work in a single scan period (via cmpxchg), but if there's a thread with
|
|
Jiri Slaby |
4505d7 |
mmap_sem locked for writing for several periods, multiple threads in
|
|
Jiri Slaby |
4505d7 |
task_numa_work() can build up a convoy waiting for mmap_sem for read and then
|
|
Jiri Slaby |
4505d7 |
all get unblocked at once.
|
|
Jiri Slaby |
4505d7 |
|
|
Jiri Slaby |
4505d7 |
This patch changes the down_read() to the trylock version, which prevents the
|
|
Jiri Slaby |
4505d7 |
build up. For a workload experiencing mmap_sem contention, it's probably better
|
|
Jiri Slaby |
4505d7 |
to postpone the NUMA balancing work anyway. This seems to have fixed the soft
|
|
Jiri Slaby |
4505d7 |
lockups involving pmd_lock(), which is in line with the convoy theory.
|
|
Jiri Slaby |
4505d7 |
|
|
Jiri Slaby |
4505d7 |
Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
|
|
Jiri Slaby |
4505d7 |
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
|
|
Jiri Slaby |
4505d7 |
Acked-by: Rik van Riel <riel@redhat.com>
|
|
Jiri Slaby |
4505d7 |
Acked-by: Mel Gorman <mgorman@techsingularity.net>
|
|
Jiri Slaby |
4505d7 |
Cc: Linus Torvalds <torvalds@linux-foundation.org>
|
|
Jiri Slaby |
4505d7 |
Cc: Peter Zijlstra <peterz@infradead.org>
|
|
Jiri Slaby |
4505d7 |
Cc: Thomas Gleixner <tglx@linutronix.de>
|
|
Jiri Slaby |
4505d7 |
Link: http://lkml.kernel.org/r/20170515131316.21909-1-vbabka@suse.cz
|
|
Jiri Slaby |
4505d7 |
Signed-off-by: Ingo Molnar <mingo@kernel.org>
|
|
Jiri Slaby |
4505d7 |
Signed-off-by: Mel Gorman <mgorman@techsingularity.net>
|
|
Jiri Slaby |
4505d7 |
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
|
|
Jiri Slaby |
4505d7 |
Signed-off-by: Jiri Slaby <jslaby@suse.cz>
|
|
Jiri Slaby |
4505d7 |
---
|
|
Jiri Slaby |
4505d7 |
kernel/sched/fair.c | 3 ++-
|
|
Jiri Slaby |
4505d7 |
1 file changed, 2 insertions(+), 1 deletion(-)
|
|
Jiri Slaby |
4505d7 |
|
|
Jiri Slaby |
4505d7 |
diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
|
|
Jiri Slaby |
4505d7 |
index 19e2469f7aea..2c89ad8a3eba 100644
|
|
Jiri Slaby |
4505d7 |
--- a/kernel/sched/fair.c
|
|
Jiri Slaby |
4505d7 |
+++ b/kernel/sched/fair.c
|
|
Jiri Slaby |
4505d7 |
@@ -2469,7 +2469,8 @@ void task_numa_work(struct callback_head *work)
|
|
Jiri Slaby |
4505d7 |
return;
|
|
Jiri Slaby |
4505d7 |
|
|
Jiri Slaby |
4505d7 |
|
|
Jiri Slaby |
4505d7 |
- down_read(&mm->mmap_sem);
|
|
Jiri Slaby |
4505d7 |
+ if (!down_read_trylock(&mm->mmap_sem))
|
|
Jiri Slaby |
4505d7 |
+ return;
|
|
Jiri Slaby |
4505d7 |
vma = find_vma(mm, start);
|
|
Jiri Slaby |
4505d7 |
if (!vma) {
|
|
Jiri Slaby |
4505d7 |
reset_ptenuma_scan(p);
|
|
Jiri Slaby |
4505d7 |
--
|
|
Jiri Slaby |
4505d7 |
2.14.2
|
|
Jiri Slaby |
4505d7 |
|