mm/swap_cgroup: fix kernel BUG in swap_cgroup_record
Subsystems (5)
PATCH 1/1 mm/swap_cgroup: fix kernel BUG in swap_cgroup_record
When using MADV_PAGEOUT, pages can remain in swapcache with their swap entries assigned. If MADV_PAGEOUT is called again on these pages, they reuse the same swap entries, causing memcg1_swapout() to call swap_cgroup_record() with an already-recorded entry.
The existing code assumes swap entries are always being recorded for the first time (oldid == 0), triggering VM_BUG_ON when it encounters an already-recorded entry:
------------[ cut here ]------------ kernel BUG at mm/swap_cgroup.c:78! Oops: invalid opcode: 0000 [#1] SMP KASAN PTI CPU: 0 UID: 0 PID: 6176 Comm: syz.0.30 Not tainted RIP: 0010:swap_cgroup_record+0x19c/0x1c0 mm/swap_cgroup.c:78 Call Trace: memcg1_swapout+0x2fa/0x830 mm/memcontrol-v1.c:623 __remove_mapping+0xac5/0xe30 mm/vmscan.c:773 shrink_folio_list+0x2786/0x4f40 mm/vmscan.c:1528 reclaim_folio_list+0xeb/0x4e0 mm/vmscan.c:2208 reclaim_pages+0x454/0x520 mm/vmscan.c:2245 madvise_cold_or_pageout_pte_range+0x19a0/0x1ce0 mm/madvise.c:563 ... do_madvise+0x1bc/0x270 mm/madvise.c:2030 __do_sys_madvise mm/madvise.c:2039
This bug occurs because pages in swapcache can be targeted by MADV_PAGEOUT multiple times without being swapped in between. Each time, the same swap entry is reused, but swap_cgroup_record() expects to only record new, unused entries.
Fix this by checking if the swap entry already has the correct cgroup ID recorded before attempting to record it. Use the existing lookup_swap_cgroup_id() to read the current cgroup ID, and return early from memcg1_swapout() if the entry is already correctly recorded. Only call swap_cgroup_record() when the entry needs to be set or updated.
This approach avoids unnecessary atomic operations, reference count manipulations, and statistics updates when the entry is already correct.
--- mm/memcontrol-v1.c | 11 +++++++++++ 1 file changed, 11 insertions(+)
mm/memcontrol-v1.c +11 -0
@@ -614,6 +614,7 @@ void memcg1_swapout(struct folio *folio, swp_entry_t entry) {struct mem_cgroup *memcg, *swap_memcg;unsigned int nr_entries; +unsigned short oldid;VM_BUG_ON_FOLIO (folio_test_lru (folio), folio);VM_BUG_ON_FOLIO (folio_ref_count (folio), folio);
@@ -630,6 +631,16 @@ void memcg1_swapout(struct folio *folio, swp_entry_t entry)if (!memcg)return ; +/* +* Check if this swap entry is already recorded. This can happen +* when MADV_PAGEOUT is called multiple times on pages that remain +* in swapcache, reusing the same swap entries. +*/ + oldid =lookup_swap_cgroup_id (entry); +if (oldid ==mem_cgroup_id (memcg)) +return ; +VM_WARN_ON_ONCE (oldid !=0 ); +/* * In case the memcg owning these pages has been offlined and doesn't * have an ID allocated to it anymore, charge the closest online
Signature · 1 line
2.43.0
Andrew Morton
maintainerQuoted 30 lines · On Sat, 10 Jan 2026 12:16:13 +0530 Deepanshu Kartikey <[email protected]>…
When using
MADV_PAGEOUT, pages can remain in swapcache with their swap entries assigned. IfMADV_PAGEOUTis called again on these pages, they reuse the same swap entries, causingmemcg1_swapout()to callswap_cgroup_record()with an already-recorded entry.The existing code assumes swap entries are always being recorded for the first time (oldid == 0), triggering
VM_BUG_ONwhen it encounters an already-recorded entry:------------[ cut here ]------------ kernel BUG at mm/swap_cgroup.c:78! Oops: invalid opcode: 0000 [#1] SMP KASAN PTI CPU: 0 UID: 0 PID: 6176 Comm: syz.0.30 Not tainted RIP: 0010:swap_cgroup_record+0x19c/0x1c0 mm/swap_cgroup.c:78 Call Trace: memcg1_swapout+0x2fa/0x830 mm/memcontrol-v1.c:623 __remove_mapping+0xac5/0xe30 mm/vmscan.c:773 shrink_folio_list+0x2786/0x4f40 mm/vmscan.c:1528 reclaim_folio_list+0xeb/0x4e0 mm/vmscan.c:2208 reclaim_pages+0x454/0x520 mm/vmscan.c:2245 madvise_cold_or_pageout_pte_range+0x19a0/0x1ce0 mm/madvise.c:563 ... do_madvise+0x1bc/0x270 mm/madvise.c:2030 __do_sys_madvise mm/madvise.c:2039This bug occurs because pages in swapcache can be targeted by
MADV_PAGEOUTmultiple times without being swapped in between. Each time, the same swap entry is reused, butswap_cgroup_record()expects to only record new, unused entries.
Quoted 8 lines
Fix this by checking if the swap entry already has the correct cgroup ID recorded before attempting to record it. Use the existing
lookup_swap_cgroup_id()to read the current cgroup ID, and return early frommemcg1_swapout()if the entry is already correctly recorded. Only callswap_cgroup_record()when the entry needs to be set or updated.This approach avoids unnecessary atomic operations, reference count manipulations, and statistics updates when the entry is already correct.
Thanks. This looks like a fairly old bug and it annoyingly predates a lot of memcg code movement.
What do people think? Should we backport this into -stable kernels? If so, can some intrepid soul please help figure out what it Fixes:?
Deepanshu, if we do decide to put a cc:stable on this then some -stable maintainers will complain that the patch alters things in a file which doesn't exist and they'll hope that you can help. Which means backporting the fix into kernels which predate 89ce924f0bd44.
Quoted 25 lines · Deepanshu Kartikey
When using
MADV_PAGEOUT, pages can remain in swapcache with their swap entries assigned. IfMADV_PAGEOUTis called again on these pages, they reuse the same swap entries, causingmemcg1_swapout()to callswap_cgroup_record()with an already-recorded entry.The existing code assumes swap entries are always being recorded for the first time (oldid == 0), triggering
VM_BUG_ONwhen it encounters an already-recorded entry:------------[ cut here ]------------ kernel BUG at mm/swap_cgroup.c:78! Oops: invalid opcode: 0000 [#1] SMP KASAN PTI CPU: 0 UID: 0 PID: 6176 Comm: syz.0.30 Not tainted RIP: 0010:swap_cgroup_record+0x19c/0x1c0 mm/swap_cgroup.c:78 Call Trace: memcg1_swapout+0x2fa/0x830 mm/memcontrol-v1.c:623 __remove_mapping+0xac5/0xe30 mm/vmscan.c:773 shrink_folio_list+0x2786/0x4f40 mm/vmscan.c:1528 reclaim_folio_list+0xeb/0x4e0 mm/vmscan.c:2208 reclaim_pages+0x454/0x520 mm/vmscan.c:2245 madvise_cold_or_pageout_pte_range+0x19a0/0x1ce0 mm/madvise.c:563 ... do_madvise+0x1bc/0x270 mm/madvise.c:2030 __do_sys_madvise mm/madvise.c:2039
Quoted 4 lines
This bug occurs because pages in swapcache can be targeted by
MADV_PAGEOUTmultiple times without being swapped in between. Each time, the same swap entry is reused, butswap_cgroup_record()expects to only record new, unused entries.
Shouldn't madvise path avoid paging out swap cache pages instead? IIRC this is what the normal reclaim path does.
Quoted 6 lines
Fix this by checking if the swap entry already has the correct cgroup ID recorded before attempting to record it. Use the existing
lookup_swap_cgroup_id()to read the current cgroup ID, and return early frommemcg1_swapout()if the entry is already correctly recorded. Only callswap_cgroup_record()when the entry needs to be set or updated.
Quoted 8 lines
This approach avoids unnecessary atomic operations, reference count manipulations, and statistics updates when the entry is already correct.
Reported-by [email protected]Tested-by [email protected]Signed-off-by Deepanshu Kartikey <[email protected]>
I would use
Signature · 2 lines
Michal Hocko SUSE Labs
Johannes Weiner
maintainerQuoted 2 lines · Deepanshu Kartikey
When using
MADV_PAGEOUT, pages can remain in swapcache with their swap entries assigned. IfMADV_PAGEOUTis called again on these pages,
This doesn't add up to me - maybe I'm missing something.
memcg1_swapout() is called at the very end of reclaim, from __remove_mapping(), which *removes the folio from swapcache*. At this point the folio is exclusive to *that* thread - there are no more present ptes that another madvise could even be acting on.
How could we reach here twice for the same swap entry?
It seems more likely that we're missing a swapin notification, fail to clear the swap entry from the cgroup records, and then trip up when the entry is recycled to a totally different page down the line. No?
Quoted 12 lines · Johannes Weiner
Quoted 2 lines · Deepanshu Kartikey
When using
MADV_PAGEOUT, pages can remain in swapcache with their swap entries assigned. IfMADV_PAGEOUTis called again on these pages,This doesn't add up to me - maybe I'm missing something.
memcg1_swapout()is called at the very end of reclaim, from__remove_mapping(), which *removes the folio from swapcache*. At this point the folio is exclusive to *that* thread - there are no more present ptes that another madvise could even be acting on.
Quoted 5 lines
How could we reach here twice for the same swap entry?
It seems more likely that we're missing a swapin notification, fail to clear the swap entry from the cgroup records, and then trip up when the entry is recycled to a totally different page down the line. No?
Thank you so much for Ccing me!
Deepanshu's patch is helpful but that's not the root cause. I did see the problem locally sometime ago, but I completely forgot about this one :(
I think the following fix should be good?
mm/vmscan.c +0 -0
@@ -758,8 +758,8 @@ static int __remove_mapping(struct address_space
*mapping, struct folio *folio,
if (reclaimed && !mapping_exiting(mapping))
shadow = workingset_eviction(folio, target_memcg);
- __swap_cache_del_folio(ci, folio, swap, shadow);
memcg1_swapout(folio, swap);
+ __swap_cache_del_folio(ci, folio, swap, shadow);
swap_cluster_unlock_irq(ci);
} else {
void (*free_folio)(struct folio *);---
It's caused by https://lore.kernel.org/linux-mm/[email protected]/
Before that patch, if the folio's swap entries are already freed and have swap count of zero, memcg1_swapout records a stalled value but the put_swap_folio below will clear the cgroup info as it frees the folio's swap slots. After that commit, put_swap_folio is merged into __swap_cache_del_folio so the stalled value will stay.