* [PATCH v2] mm/page_owner: fix double decrement of stack count on unpoison
@ 2026-11-19 23:48 Nilay Vaish
2026-11-20 3:46 ` Miaohe Lin
0 siblings, 1 reply; 2+ messages in thread
From: Nilay Vaish @ 2026-11-19 23:48 UTC (permalink / raw)
To: Andrew Morton, Vlastimil Babka, Oscar Salvador
Cc: Jose A . Perez de Azpillaga, Johannes Weiner, Suren Baghdasaryan,
Michal Hocko, Brendan Jackman, Zi Yan, Marco Elver,
Andrey Konovalov, Miaohe Lin, Naoya Horiguchi, linux-mm,
linux-kernel, Nilay Vaish, stable
page_owner counts, for each allocation stack, how many pages allocated
from that stack are still in use. The count is incremented when a page
is allocated and decremented in __reset_page_owner() when it is freed.
A page that is hwpoisoned while in use and later unpoisoned passes
through free_pages_prepare() twice:
1. On poisoning, the last reference is dropped. free_pages_prepare()
calls reset_page_owner() but, because the page is PageHWPoison(),
keeps it off the free lists.
2. On unpoisoning, unpoison_memory() drops the poison reference and
free_pages_prepare() calls reset_page_owner() again, this time
really freeing the page.
__reset_page_owner() decrements the stack count both times, although
the page was allocated only once. The first call clears
PAGE_EXT_OWNER_ALLOCATED, but the second call never checks it.
If this drives the count to zero while the stack is still in use,
stackdepot treats the record as freed and warns on every later
allocation or free that uses the same stack:
dec_stack_record_count: refcount went to 0 for 92274844 handle
WARNING: lib/stackdepot.c:517 at depot_fetch_stack+0x5c/0x70
Call Trace:
__set_page_owner+0x110/0x290
get_page_from_freelist+0x1225/0x1440
__alloc_frozen_pages_noprof+0x1fc/0x1b00
__folio_alloc_noprof+0x12/0xa0
alloc_migration_target+0x81/0x1e0
migrate_pages_batch+0x1f2/0x14c0
migrate_pages+0x9c3/0xca0
soft_offline_page+0x189/0x570
A single soft offline followed by an unpoison is enough to trigger it,
because the migration-target stack that the offlined page is accounted
to usually has very few pages.
Fix it by decrementing the stack count only if PAGE_EXT_OWNER_ALLOCATED
was still set, i.e. only the first time the page is freed.
To reproduce, boot with page_owner=on and CONFIG_HWPOISON_INJECT, soft
offline anonymous pages with madvise(MADV_SOFT_OFFLINE), and unpoison
each old PFN through debugfs hwpoison/unpoison-pfn. Without this patch
the first cycle prints the message above and 2200 cycles produce 6597
warnings. With it, the same run is clean and all stack counts return
to their initial values.
Fixes: 217b2119b9e2 ("mm,page_owner: implement the tracking of the stacks count")
Cc: stable@vger.kernel.org
Reviewed-by: Jose A. Perez de Azpillaga <azpijr@gmail.com>
Reviewed-by: Oscar Salvador (SUSE) <osalvador@kernel.org>
Signed-off-by: Nilay Vaish <nilayvaish@google.com>
---
v2:
- Expand the comment in __reset_page_owner() to explain the allocated
check (Oscar)
- Collect Reviewed-by tags from Jose and Oscar
v1: https://lore.kernel.org/linux-mm/20261118212819.403059-1-nilayvaish@google.com/
mm/page_owner.c | 12 +++++++++++-
1 file changed, 11 insertions(+), 1 deletion(-)
diff --git a/mm/page_owner.c b/mm/page_owner.c
index 02bb304b53cf..09a3c2958079 100644
--- a/mm/page_owner.c
+++ b/mm/page_owner.c
@@ -320,11 +320,13 @@ void __reset_page_owner(struct page *page, unsigned short order)
depot_stack_handle_t alloc_handle;
struct page_owner *page_owner;
u64 free_ts_nsec = local_clock();
+ bool allocated;
page_ext = page_ext_get(page);
if (unlikely(!page_ext))
return;
+ allocated = test_bit(PAGE_EXT_OWNER_ALLOCATED, &page_ext->flags);
page_owner = get_page_owner(page_ext);
alloc_handle = page_owner->handle;
page_ext_put(page_ext);
@@ -339,8 +341,15 @@ void __reset_page_owner(struct page *page, unsigned short order)
__update_page_owner_free_handle(page, handle, order, current->pid,
current->tgid, free_ts_nsec);
- if (alloc_handle != early_handle)
+ if (allocated && alloc_handle != early_handle) {
/*
+ * Not every page that gets here is still accounted as
+ * allocated: a hwpoisoned page that is later unpoisoned is
+ * freed twice, and the first free already cleared
+ * PAGE_EXT_OWNER_ALLOCATED and dropped the count. Only
+ * decrement the allocation stack's count if the page was
+ * still marked as allocated.
+ *
* early_handle is being set as a handle for all those
* early allocated pages. See init_pages_in_zone().
* Since their refcount is not being incremented because
@@ -348,6 +357,7 @@ void __reset_page_owner(struct page *page, unsigned short order)
* their refcount either.
*/
dec_stack_record_count(alloc_handle, 1 << order);
+ }
}
noinline void __set_page_owner(struct page *page, unsigned short order,
--
2.56.1.750.gf70dc682cf-goog
^ permalink raw reply [flat|nested] 2+ messages in thread
* Re: [PATCH v2] mm/page_owner: fix double decrement of stack count on unpoison
2026-11-19 23:48 [PATCH v2] mm/page_owner: fix double decrement of stack count on unpoison Nilay Vaish
@ 2026-11-20 3:46 ` Miaohe Lin
0 siblings, 0 replies; 2+ messages in thread
From: Miaohe Lin @ 2026-11-20 3:46 UTC (permalink / raw)
To: Nilay Vaish
Cc: Jose A . Perez de Azpillaga, Johannes Weiner, Suren Baghdasaryan,
Michal Hocko, Brendan Jackman, Zi Yan, Marco Elver,
Andrey Konovalov, Naoya Horiguchi, linux-mm, linux-kernel,
stable, Andrew Morton, Vlastimil Babka, Oscar Salvador
On 2026/11/20 7:48, Nilay Vaish wrote:
> page_owner counts, for each allocation stack, how many pages allocated
> from that stack are still in use. The count is incremented when a page
> is allocated and decremented in __reset_page_owner() when it is freed.
>
> A page that is hwpoisoned while in use and later unpoisoned passes
> through free_pages_prepare() twice:
>
> 1. On poisoning, the last reference is dropped. free_pages_prepare()
> calls reset_page_owner() but, because the page is PageHWPoison(),
> keeps it off the free lists.
>
> 2. On unpoisoning, unpoison_memory() drops the poison reference and
> free_pages_prepare() calls reset_page_owner() again, this time
> really freeing the page.
>
> __reset_page_owner() decrements the stack count both times, although
> the page was allocated only once. The first call clears
> PAGE_EXT_OWNER_ALLOCATED, but the second call never checks it.
>
> If this drives the count to zero while the stack is still in use,
> stackdepot treats the record as freed and warns on every later
> allocation or free that uses the same stack:
>
> dec_stack_record_count: refcount went to 0 for 92274844 handle
> WARNING: lib/stackdepot.c:517 at depot_fetch_stack+0x5c/0x70
> Call Trace:
> __set_page_owner+0x110/0x290
> get_page_from_freelist+0x1225/0x1440
> __alloc_frozen_pages_noprof+0x1fc/0x1b00
> __folio_alloc_noprof+0x12/0xa0
> alloc_migration_target+0x81/0x1e0
> migrate_pages_batch+0x1f2/0x14c0
> migrate_pages+0x9c3/0xca0
> soft_offline_page+0x189/0x570
>
> A single soft offline followed by an unpoison is enough to trigger it,
> because the migration-target stack that the offlined page is accounted
> to usually has very few pages.
>
> Fix it by decrementing the stack count only if PAGE_EXT_OWNER_ALLOCATED
> was still set, i.e. only the first time the page is freed.
>
> To reproduce, boot with page_owner=on and CONFIG_HWPOISON_INJECT, soft
> offline anonymous pages with madvise(MADV_SOFT_OFFLINE), and unpoison
> each old PFN through debugfs hwpoison/unpoison-pfn. Without this patch
Oh wow, this issue should be very easy to reproduce, and it has been around for
quite some time. When I tested earlier, CONFIG_PAGE_OWNER was enabled, but I didn't
add page_owner=on to the boot parameters. :(
> the first cycle prints the message above and 2200 cycles produce 6597
> warnings. With it, the same run is clean and all stack counts return
> to their initial values.
>
> Fixes: 217b2119b9e2 ("mm,page_owner: implement the tracking of the stacks count")
> Cc: stable@vger.kernel.org
> Reviewed-by: Jose A. Perez de Azpillaga <azpijr@gmail.com>
> Reviewed-by: Oscar Salvador (SUSE) <osalvador@kernel.org>
> Signed-off-by: Nilay Vaish <nilayvaish@google.com>
Reviewed-by: Miaohe Lin <linmiaohe@huawei.com>
Thanks.
.
^ permalink raw reply [flat|nested] 2+ messages in thread
end of thread, other threads:[~2026-11-20 3:46 UTC | newest]
Thread overview: 2+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2026-11-19 23:48 [PATCH v2] mm/page_owner: fix double decrement of stack count on unpoison Nilay Vaish
2026-11-20 3:46 ` Miaohe Lin
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox
all inboxes | Powered by JetHome®