[PATCH] mm: memory-failure: update statistics before emitting the trace event

[PATCH] mm: memory-failure: update statistics before emitting the trace event

[Mauro Carvalho Chehab]

From: Mauro Carvalho Chehab <m.chehab@huawei.com>

Today's rasdaemon-ci run showed a race condition at the page poison
logic affecting userspace (rasdaemon): action_result() emits a
memory_failure_event before incrementing num_poisoned_pages and
updating the per-node memory failure statistics.

That causes userspace to read /proc/meminfo to detect HardwareCorrupted
trace events to warn if the value exceeds POISON_STAT_THRESHOLD.
This works most of the cases, but there is a chance that the value
not being updated in time before libtraceevent notification is handled.

Root cause is commit e240ac52f7da ("mm/memory-failure.c: simplify num_poisoned_pages_inc/dec")
which changed the order of updates.

Rasdaemod-ci run up to Oct 27 worked properly, but on today's run, it
failed to get the HardwareCorrupted data from /proc/meminfo because
it was updated only after trigging the hardware error event:
	https://mchehab.github.io/rasdaemon-ci/failures/20261028T035110Z-40629018388/daily/aarch64/consumer-poison-page-stat.html

E.g. it failed to log this:
	rasdaemon: Poison page statistics exceeded threshold: 4 kB (threshold: 0 kB)

This is important for rasdaemon, as the consumer's action monitors a failed
threshold before poisoning pages. So, if the counter doesn't update properly,
it will end not poisoning pages at the expected threshold.

Fixes: e240ac52f7da ("mm/memory-failure.c: simplify num_poisoned_pages_inc/dec")
Closes: https://github.com/mchehab/rasdaemon-ci/actions/runs/40629018388
Signed-off-by: Mauro Carvalho Chehab <mchehab+huawei@kernel.org>
---
 mm/memory-failure.c | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/mm/memory-failure.c b/mm/memory-failure.c
index a8b03e2920ba..eba6fe1dfdea 100644
--- a/mm/memory-failure.c
+++ b/mm/memory-failure.c
@@ -1281,13 +1281,13 @@ static bool panic_on_unrecoverable_mf(enum mf_action_page_type type,
 static int action_result(unsigned long pfn, enum mf_action_page_type type,
 			 enum mf_result result)
 {
-	trace_memory_failure_event(pfn, type, result);
-
 	if (type != MF_MSG_ALREADY_POISONED && type != MF_MSG_PFN_MAP) {
 		num_poisoned_pages_inc(pfn);
 		update_per_node_mf_stats(pfn, result);
 	}
 
+	trace_memory_failure_event(pfn, type, result);
+
 	pr_err("%#lx: recovery action for %s: %s\n",
 		pfn, action_page_types[type], action_name[result]);
 
-- 
2.56.0

Re: [PATCH] mm: memory-failure: update statistics before emitting the trace event

[Miaohe Lin]

On 2026/10/28 16:09, Mauro Carvalho Chehab wrote:
> From: Mauro Carvalho Chehab <m.chehab@huawei.com>
> 
> Today's rasdaemon-ci run showed a race condition at the page poison
> logic affecting userspace (rasdaemon): action_result() emits a
> memory_failure_event before incrementing num_poisoned_pages and
> updating the per-node memory failure statistics.
> 
> That causes userspace to read /proc/meminfo to detect HardwareCorrupted
> trace events to warn if the value exceeds POISON_STAT_THRESHOLD.
> This works most of the cases, but there is a chance that the value
> not being updated in time before libtraceevent notification is handled.
> 
> Root cause is commit e240ac52f7da ("mm/memory-failure.c: simplify num_poisoned_pages_inc/dec")
> which changed the order of updates.
> 
> Rasdaemod-ci run up to Oct 27 worked properly, but on today's run, it
> failed to get the HardwareCorrupted data from /proc/meminfo because
> it was updated only after trigging the hardware error event:
> 	https://mchehab.github.io/rasdaemon-ci/failures/20261028T035110Z-40629018388/daily/aarch64/consumer-poison-page-stat.html
> 
> E.g. it failed to log this:
> 	rasdaemon: Poison page statistics exceeded threshold: 4 kB (threshold: 0 kB)
> 
> This is important for rasdaemon, as the consumer's action monitors a failed
> threshold before poisoning pages. So, if the counter doesn't update properly,
> it will end not poisoning pages at the expected threshold.
> 
> Fixes: e240ac52f7da ("mm/memory-failure.c: simplify num_poisoned_pages_inc/dec")
> Closes: https://github.com/mchehab/rasdaemon-ci/actions/runs/40629018388
> Signed-off-by: Mauro Carvalho Chehab <mchehab+huawei@kernel.org>

Acked-by: Miaohe Lin <linmiaohe@huawei.com>

Thanks.
.