mm/swap: Add cond_resched() in swap_reclaim_full_clusters to prevent softlockup
Subsystems (2)
PATCH 1/1 mm/swap: Add cond_resched() in swap_reclaim_full_clusters to prevent softlockup
Add periodic cond_resched() calls during large full_clusters reclaim operations to prevent softlockup issues.
--- mm/swapfile.c | 1 + 1 file changed, 1 insertion(+)
mm/swapfile.c +1 -0
@@ -1054,6 +1054,7 @@ static void swap_reclaim_full_clusters(struct swap_info_struct *si, bool force)swap_cluster_unlock (ci);if (to_scan <=0 )break ; +cond_resched (); } }
Signature · 1 line
2.43.5
Baoquan He
reviewerHi Keray,
Quoted 9 lines · kerayhuang
Add periodic
cond_resched()calls during largefull_clustersreclaim operations to prevent softlockup issues.Signed-off-by kerayhuang <[email protected]>Reviewed-by Kairui Song <[email protected]>Reviewed-by Hao Peng <[email protected]>--- mm/swapfile.c | 1 + 1 file changed, 1 insertion(+)
Thanks for the patch. The change looks good to me, however there are still small concerns.
For patch log, it might be better to provide more details, e.g did you observe this issue in a product environment, or just a code exploring? If observed in a product environment, what does the backtrace look like when softlockup happened?
mm/swapfile.c static void swap_reclaim_full_clusters(struct swap_info_struct *si, bool force) · line 1057
Quoted 10 lines
mm/swapfile.c +1
-0@@ -1054,6 +1054,7 @@ static void swap_reclaim_full_clusters(struct swap_info_struct *si, bool force)swap_cluster_unlock (ci);if (to_scan <=0 )break ; +cond_resched ();
Besides, is it a little bit too aggressive to call cond_resched() for each cluster reclaiming compared with the old code? Do you consider to make it gentle, e.g calling cond_resched() every several clusters, 8, 16 or other number decided based on your testing performance statistics.
Thanks
Baoquan
Hi Baoquan,
Thanks for the review!
Quoted 13 lines
Hi Keray,
Quoted 9 lines · kerayhuang
Add periodic
cond_resched()calls during largefull_clustersreclaim operations to prevent softlockup issues.Signed-off-by kerayhuang <[email protected]>Reviewed-by Kairui Song <[email protected]>Reviewed-by Hao Peng <[email protected]>--- mm/swapfile.c | 1 + 1 file changed, 1 insertion(+)
Quoted 7 lines
Thanks for the patch. The change looks good to me, however there are still small concerns.
For patch log, it might be better to provide more details, e.g did you observe this issue in a product environment, or just a code exploring? If observed in a product environment, what does the backtrace look like when softlockup happened?
We hit a real softlockup in an internal stress test environment. The workload was LTP memory/swap stress on a large arm64 machine, with 320 CPUs, about 1TB memory and an 8.6GB swap device. The system was under heavy load and the swap device had a large number of full clusters. The softlockup was triggered during a stress test after about 3 days.
The backtrace looks like:
PID: 3817773 TASK: ffff0883bb28b780 CPU: 48 COMMAND: "kworker/48:7"
#0 [ffff800080183d10] __crash_kexec at ffffa4c1361e5de4 #1 [ffff800080183d90] panic at ffffa4c1360d5e9c
#2 [ffff800080183e20] watchdog_timer_fn at ffffa4c136231fa8
...
#16 [ffff8000c4ad3cb0] swap_cache_del_folio at ffffa4c1363e1614 #17 [ffff8000c4ad3ce0] __try_to_reclaim_swap at ffffa4c1363e4bfc #18 [ffff8000c4ad3d40] swap_reclaim_full_clusters at ffffa4c1363e5474 #19 [ffff8000c4ad3da0] swap_reclaim_work at ffffa4c1363e550c #20 [ffff8000c4ad3dc0] process_one_work at ffffa4c136102edc #21 [ffff8000c4ad3e10] worker_thread at ffffa4c136103398 #22 [ffff8000c4ad3e70] kthread at ffffa4c13610d95c
From the vmcore analysis, swap_reclaim_work() called swap_reclaim_full_clusters() with force=true, which sets to_scan to 1551 clusters. At the time of the softlockup, there were still 1427 full clusters remaining in the full_clusters list.
I will add these details to the commit log in v2.
Quoted 6 lines
Quoted 6 lines
mm/swapfile.c +0
-0@@ -1054,6 +1054,7 @@ static void swap_reclaim_full_clusters(struct swap_info_struct *si, bool force)swap_cluster_unlock (ci);
Quoted 8 lines
mm/swapfile.c static void swap_reclaim_full_clusters(struct swap_info_struct *si, bool force) · line 1057
Quoted 3 lines
mm/swapfile.c +1
-0if (to_scan <=0 )break ; +cond_resched ();Besides, is it a little bit too aggressive to call
cond_resched()for each cluster reclaiming compared with the old code? Do you consider to make it gentle, e.g callingcond_resched()every several clusters, 8, 16 or other number decided based on your testing performance statistics.
I think calling cond_resched() per cluster is reasonable here because:
1) Each cluster iteration already involves scanning up to 512 slots, and each slot reclaim may call __try_to_reclaim_swap() which does non-trivial work (lock/unlock, folio lookup, swap cache deletion, and potentially slab freeing). So the work per cluster is already substantial.
2) cond_resched() is a lightweight check - it only actually reschedules when need_resched() is set, so in the common case it's just a flag check with negligible overhead. Therefore calling it once per cluster gives a bounded latency without forcing an actual context switch every time. If we call it only every 8 or 16 clusters, the worst-case non-preemptible window can still become quite large on machines with many full clusters.
2) This is a workqueue context (swap_reclaim_work), not a hot fast path, so the slight overhead is acceptable.
Thanks,
Keray
Baoquan He
reviewerQuoted 45 lines · kerayhuang
Hi Baoquan,
Thanks for the review!Quoted 13 lines
Hi Keray,
Quoted 9 lines · kerayhuang
Add periodic
cond_resched()calls during largefull_clustersreclaim operations to prevent softlockup issues.Signed-off-by kerayhuang <[email protected]>Reviewed-by Kairui Song <[email protected]>Reviewed-by Hao Peng <[email protected]>--- mm/swapfile.c | 1 + 1 file changed, 1 insertion(+)Quoted 7 lines
Thanks for the patch. The change looks good to me, however there are still small concerns.
For patch log, it might be better to provide more details, e.g did you observe this issue in a product environment, or just a code exploring? If observed in a product environment, what does the backtrace look like when softlockup happened?
We hit a real softlockup in an internal stress test environment. The workload was LTP memory/swap stress on a large arm64 machine, with 320 CPUs, about 1TB memory and an 8.6GB swap device. The system was under heavy load and the swap device had a large number of full clusters. The softlockup was triggered during a stress test after about 3 days.
The backtrace looks like:
PID: 3817773 TASK:
ffff0883bb28b780CPU: 48 COMMAND: "kworker/48:7"
#0 [ffff800080183d10]__crash_kexecatffffa4c1361e5de4#1 [ffff800080183d90] panic atffffa4c1360d5e9c
#2 [ffff800080183e20]watchdog_timer_fnatffffa4c136231fa8
...
#16 [ffff8000c4ad3cb0]swap_cache_del_folioatffffa4c1363e1614#17 [ffff8000c4ad3ce0]__try_to_reclaim_swapatffffa4c1363e4bfc#18 [ffff8000c4ad3d40]swap_reclaim_full_clustersatffffa4c1363e5474#19 [ffff8000c4ad3da0]swap_reclaim_workatffffa4c1363e550c#20 [ffff8000c4ad3dc0]process_one_workatffffa4c136102edc#21 [ffff8000c4ad3e10]worker_threadatffffa4c136103398#22 [ffff8000c4ad3e70] kthread atffffa4c13610d95c
Quoted 6 lines
From the vmcore analysis,
swap_reclaim_work()calledswap_reclaim_full_clusters()with force=true, which setsto_scanto 1551 clusters. At the time of the softlockup, there were still 1427 full clusters remaining in thefull_clusterslist.I will add these details to the commit log in v2.
Sounds like a very great root cause digging. Adding these into patch log will be very helpful.
By the way, is it worth a Fixes tag?
Quoted 33 lines
Quoted 15 lines
mm/swapfile.c static void swap_reclaim_full_clusters(struct swap_info_struct *si, bool force) · line 1057
Quoted 10 lines
mm/swapfile.c +1
-0@@ -1054,6 +1054,7 @@ static void swap_reclaim_full_clusters(struct swap_info_struct *si, bool force)swap_cluster_unlock (ci);if (to_scan <=0 )break ; +cond_resched ();Besides, is it a little bit too aggressive to call
cond_resched()for each cluster reclaiming compared with the old code? Do you consider to make it gentle, e.g callingcond_resched()every several clusters, 8, 16 or other number decided based on your testing performance statistics.I think calling
cond_resched()per cluster is reasonable here because:1) Each cluster iteration already involves scanning up to 512 slots, and each slot reclaim may call
__try_to_reclaim_swap()which does non-trivial work (lock/unlock, folio lookup, swap cache deletion, and potentially slab freeing). So the work per cluster is already substantial.2)
cond_resched()is a lightweight check - it only actually reschedules whenneed_resched()is set, so in the common case it's just a flag check with negligible overhead. Therefore calling it once per cluster gives a bounded latency without forcing an actual context switch every time. If we call it only every 8 or 16 clusters, the worst-case non-preemptible window can still become quite large on machines with many full clusters.
Quoted 2 lines
2) This is a workqueue context (
swap_reclaim_work), not a hot fast path, so the slight overhead is acceptable.
OK, that sounds good. Just when system is under heavy stress, it could yield after each cluster reclaiming. Imagine a system with a bigger swap disk, it will alwasy need to check if swap is 50% full and if it's in workqueue. Anyway, maybe I am overthinking. The overrall looks very great to me, good catch, good root cause digging and good fix. Let's see if other people have any concern.
Thanks
Baoquan
PATCH v2 1/1 mm/swap: Add cond_resched() in swap_reclaim_full_clusters to prevent softlockup
We hit a real softlockup in an internal stress test environment. The workload was LTP memory/swap stress on a large arm64 machine, with 320 CPUs, about 1TB memory and an 8.6GB swap device. The system was under heavy load and the swap device had a large number of full clusters. The softlockup was triggered during a stress test after about 3 days.
So, add periodic cond_resched() calls during large full_clusters reclaim operations to prevent softlockup issues.
Detailed call trace as follow:
PID: 3817773 TASK: ffff0883bb28b780 CPU: 48 COMMAND: "kworker/48:7"
#0 [ffff800080183d10] __crash_kexec at ffffa4c1361e5de4 #1 [ffff800080183d90] panic at ffffa4c1360d5e9c
#2 [ffff800080183e20] watchdog_timer_fn at ffffa4c136231fa8
...
#16 [ffff8000c4ad3cb0] swap_cache_del_folio at ffffa4c1363e1614 #17 [ffff8000c4ad3ce0] __try_to_reclaim_swap at ffffa4c1363e4bfc #18 [ffff8000c4ad3d40] swap_reclaim_full_clusters at ffffa4c1363e5474 #19 [ffff8000c4ad3da0] swap_reclaim_work at ffffa4c1363e550c #20 [ffff8000c4ad3dc0] process_one_work at ffffa4c136102edc #21 [ffff8000c4ad3e10] worker_thread at ffffa4c136103398 #22 [ffff8000c4ad3e70] kthread at ffffa4c13610d95c
--- mm/swapfile.c | 1 + 1 file changed, 1 insertion(+)
mm/swapfile.c +1 -0
@@ -1054,6 +1054,7 @@ static void swap_reclaim_full_clusters(struct swap_info_struct *si, bool force)swap_cluster_unlock (ci);if (to_scan <=0 )break ; +cond_resched (); } }
Signature · 1 line
2.43.5
Quoted 39 lines · Zijiang Huang
We hit a real softlockup in an internal stress test environment. The workload was LTP memory/swap stress on a large arm64 machine, with 320 CPUs, about 1TB memory and an 8.6GB swap device. The system was under heavy load and the swap device had a large number of full clusters. The softlockup was triggered during a stress test after about 3 days.
So, add periodic
cond_resched()calls during largefull_clustersreclaim operations to prevent softlockup issues.Detailed call trace as follow:
PID: 3817773 TASK:
ffff0883bb28b780CPU: 48 COMMAND: "kworker/48:7"
#0 [ffff800080183d10]__crash_kexecatffffa4c1361e5de4#1 [ffff800080183d90] panic atffffa4c1360d5e9c
#2 [ffff800080183e20]watchdog_timer_fnatffffa4c136231fa8
...
#16 [ffff8000c4ad3cb0]swap_cache_del_folioatffffa4c1363e1614#17 [ffff8000c4ad3ce0]__try_to_reclaim_swapatffffa4c1363e4bfc#18 [ffff8000c4ad3d40]swap_reclaim_full_clustersatffffa4c1363e5474#19 [ffff8000c4ad3da0]swap_reclaim_workatffffa4c1363e550c#20 [ffff8000c4ad3dc0]process_one_workatffffa4c136102edc#21 [ffff8000c4ad3e10]worker_threadatffffa4c136103398#22 [ffff8000c4ad3e70] kthread atffffa4c13610d95cFixes 5168a68eb78f ("mm, swap: avoid over reclaim of full clusters")Signed-off-by Zijiang Huang <[email protected]>Reviewed-by Kairui Song <[email protected]>Reviewed-by Hao Peng <[email protected]>Reviewed-by albinwyang <[email protected]>--- mm/swapfile.c | 1 + 1 file changed, 1 insertion(+)mm/swapfile.c +0
-0
mm/swapfile.c static void swap_reclaim_full_clusters(struct swap_info_struct *si, bool force) · line 1059
Quoted 6 lines
mm/swapfile.c +1
-0@@ -1054,6 +1054,7 @@ static void swap_reclaim_full_clusters(struct swap_info_struct *si, bool force)swap_cluster_unlock (ci);if (to_scan <=0 )break ; +cond_resched (); } }
LGTM,
Quoted 10 lines · Zijiang Huang
We hit a real softlockup in an internal stress test environment. The workload was LTP memory/swap stress on a large arm64 machine, with 320 CPUs, about 1TB memory and an 8.6GB swap device. The system was under heavy load and the swap device had a large number of full clusters. The softlockup was triggered during a stress test after about 3 days.
So, add periodic
cond_resched()calls during largefull_clustersreclaim operations to prevent softlockup issues.
Thank you for reporting and fixing this issue.
Can you add that to the patch commit log? This background information is very important for the reader why this change is needed.
Assume you will update the commit log.
Chris
Quoted 14 lines
Quoted 10 lines · On Wed, May 6, 2026 at 4:56=E2=80=AFPM Zijiang Huang <[email protected]=…
We hit a real softlockup in an internal stress test environment. The workload was LTP memory/swap stress on a large arm64 machine, with 320 CPUs, about 1TB memory and an 8.6GB swap device. The system was under heavy load and the swap device had a large number of full clusters. The softlockup was triggered during a stress test after about 3 days.
So, add periodic
cond_resched()calls during largefull_clustersreclaim operations to prevent softlockup issues.Thank you for reporting and fixing this issue.
Quoted 8 lines
Can you add that to the patch commit log? This background information is very important for the reader why this change is needed.
Assume you will update the commit log.
Acked-by Chris Li <[email protected]>Chris
Hi Chris, I've updated the commit log with the background information in v2. Thanks!
Baoquan He
reviewerQuoted 17 lines · Zijiang Huang
Quoted 17 lines
Quoted 10 lines · On Wed, May 6, 2026 at 4:56=E2=80=AFPM Zijiang Huang <[email protected]=…
We hit a real softlockup in an internal stress test environment. The workload was LTP memory/swap stress on a large arm64 machine, with 320 CPUs, about 1TB memory and an 8.6GB swap device. The system was under heavy load and the swap device had a large number of full clusters. The softlockup was triggered during a stress test after about 3 days.
So, add periodic
cond_resched()calls during largefull_clustersreclaim operations to prevent softlockup issues.Thank you for reporting and fixing this issue.
Can you add that to the patch commit log? This background information is very important for the reader why this change is needed.
Quoted 7 lines
Quoted 5 lines
Assume you will update the commit log.
Acked-by Chris Li <[email protected]>Chris
Hi Chris, I've updated the commit log with the background information in v2. Thanks!
From my mail client, I see your v2 is in the same thread as v1. People may mistaken it as a discussion reply. It's better to post v2 separately. You can resend v2 with reviewers' ack tag collected.
Andrew Morton
maintainerQuoted 12 lines · On Wed, 6 May 2026 21:09:19 +0800 Zijiang Huang <[email protected]> wrote:
We hit a real softlockup in an internal stress test environment. The workload was LTP memory/swap stress on a large arm64 machine, with 320 CPUs, about 1TB memory and an 8.6GB swap device. The system was under heavy load and the swap device had a large number of full clusters. The softlockup was triggered during a stress test after about 3 days.
So, add periodic
cond_resched()calls during largefull_clustersreclaim operations to prevent softlockup issues.Detailed call trace as follow:
Quoted 12 lines
PID: 3817773 TASK:
ffff0883bb28b780CPU: 48 COMMAND: "kworker/48:7"
#0 [ffff800080183d10]__crash_kexecatffffa4c1361e5de4#1 [ffff800080183d90] panic atffffa4c1360d5e9c
#2 [ffff800080183e20]watchdog_timer_fnatffffa4c136231fa8
...
#16 [ffff8000c4ad3cb0]swap_cache_del_folioatffffa4c1363e1614#17 [ffff8000c4ad3ce0]__try_to_reclaim_swapatffffa4c1363e4bfc#18 [ffff8000c4ad3d40]swap_reclaim_full_clustersatffffa4c1363e5474#19 [ffff8000c4ad3da0]swap_reclaim_workatffffa4c1363e550c#20 [ffff8000c4ad3dc0]process_one_workatffffa4c136102edc#21 [ffff8000c4ad3e10]worker_threadatffffa4c136103398#22 [ffff8000c4ad3e70] kthread atffffa4c13610d95c
Thanks.
Quoted 1 line
Fixes 5168a68eb78f ("mm, swap: avoid over reclaim of full clusters")
I'll add a cc:stable to this to help ensure that earlier kernels don't hit this.
Quoted 4 lines
Signed-off-by Zijiang Huang <[email protected]>Reviewed-by Kairui Song <[email protected]>Reviewed-by Hao Peng <[email protected]>Reviewed-by albinwyang <[email protected]>