Skip to contentSkip to conversation

bpf/helpers: Use __GFP_HIGH instead of GFP_ATOMIC in __bpf_async_init()

PATCH Superseded Peilin Ye 2 participants 4 messages Sep 5, 2025 – Sep 8, 2025

Subsystems (2)

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore

Peilin Ye

author
#0 lore raw

PATCH 1/1 bpf/helpers: Use __GFP_HIGH instead of GFP_ATOMIC in __bpf_async_init()

Currently, calling bpf_map_kmalloc_node() from __bpf_async_init() can cause various locking issues; see the following stack trace (edited for style) as one example:

...

 [10.011566]  do_raw_spin_lock.cold
 [10.011570]  try_to_wake_up             (5) double-acquiring the same
 [10.011575]  kick_pool                      rq_lock, causing a hardlockup
 [10.011579]  __queue_work
 [10.011582]  queue_work_on
 [10.011585]  kernfs_notify
 [10.011589]  cgroup_file_notify
 [10.011593]  try_charge_memcg           (4) memcg accounting raises an
 [10.011597]  obj_cgroup_charge_pages        MEMCG_MAX event
 [10.011599]  obj_cgroup_charge_account
 [10.011600]  __memcg_slab_post_alloc_hook
 [10.011603]  __kmalloc_node_noprof
...
 [10.011611]  bpf_map_kmalloc_node
 [10.011612]  __bpf_async_init
 [10.011615]  bpf_timer_init             (3) BPF calls bpf_timer_init()
 [10.011617]  bpf_prog_xxxxxxxxxxxxxxxx_fcg_runnable
 [10.011619]  bpf__sched_ext_ops_runnable
 [10.011620]  enqueue_task_scx           (2) BPF runs with rq_lock held
 [10.011622]  enqueue_task
 [10.011626]  ttwu_do_activate
 [10.011629]  sched_ttwu_pending         (1) grabs rq_lock

...

The above was reproduced on bpf-next (b338cf849ec8) by modifying ./tools/sched_ext/scx_flatcg.bpf.c to call bpf_timer_init() during ops.runnable(), and hacking [1] the memcg accounting code a bit to make a bpf_timer_init() call much more likely to raise an MEMCG_MAX event.

We have also run into other similar variants (both internally and on bpf-next), including double-acquiring cgroup_file_kn_lock, the same worker_pool::lock, etc.

As suggested by Shakeel, fix this by using __GFP_HIGH instead of GFP_ATOMIC in __bpf_async_init(), so that if try_charge_memcg() raises an MEMCG_MAX event, we call __memcg_memory_event() with @allow_spinning=false and skip calling cgroup_file_notify(), in order to avoid the locking issues described above.

Depends on mm patch "memcg: skip cgroup_file_notify if spinning is not allowed". Tested with vmtest.sh (llvm-18, x86-64):

 $ ./test_progs -a '*timer*' -a '*wq*'
...
 Summary: 7/12 PASSED, 0 SKIPPED, 0 FAILED

[1] Making bpf_timer_init() much more likely to raise an MEMCG_MAX event (gist-only, for brevity):

kernel/bpf/helpers.c:__bpf_async_init():
- cb = bpf_map_kmalloc_node(map, size, GFP_ATOMIC, map->numa_node);
+ cb = bpf_map_kmalloc_node(map, size, GFP_ATOMIC | __GFP_HACK,
+ map->numa_node);

mm/memcontrol.c:try_charge_memcg():

          if (!do_memsw_account() ||
 -            page_counter_try_charge(&memcg->memsw, batch, &counter)) {
 -                if (page_counter_try_charge(&memcg->memory, batch, &counter))
 +            page_counter_try_charge_hack(&memcg->memsw, batch, &counter,
 +                                         gfp_mask & __GFP_HACK)) {
 +                if (page_counter_try_charge_hack(&memcg->memory, batch,
 +                                                 &counter,
 +                                                 gfp_mask & __GFP_HACK))
                          goto done_restock;

mm/page_counter.c:page_counter_try_charge():
-bool page_counter_try_charge(struct page_counter *counter,
- unsigned long nr_pages,
- struct page_counter **fail)
+bool page_counter_try_charge_hack(struct page_counter *counter,
+ unsigned long nr_pages,
+ struct page_counter **fail, bool hack)
{
...
- if (new > c->max) {

 +                if (hack || new > c->max) {     // goto failed;
                          atomic_long_sub(nr_pages, &c->usage);
Fixes b00628b1c7d5 ("bpf: Introduce bpf timers.")
---
 kernel/bpf/helpers.c | 10 ++++++++--
 1 file changed, 8 insertions(+), 2 deletions(-)
kernel/bpf/helpers.c +8 -2
index b9b0c5fe33f6..508b13c24778 100644
@@ -1274,8 +1274,14 @@ static int __bpf_async_init(struct bpf_async_kern *async, struct bpf_map *map, u
 		goto out;
 	}
 
-	/* allocate hrtimer via map_kmalloc to use memcg accounting */
-	cb = bpf_map_kmalloc_node(map, size, GFP_ATOMIC, map->numa_node);
+	/* Allocate via bpf_map_kmalloc_node() for memcg accounting. Use
+	 * __GFP_HIGH instead of GFP_ATOMIC to avoid calling
+	 * cgroup_file_notify() if an MEMCG_MAX event is raised by
+	 * try_charge_memcg(). This prevents various locking issues, including
+	 * double-acquiring locks that may already be held here (e.g.,
+	 * cgroup_file_kn_lock, rq_lock).
+	 */
+	cb = bpf_map_kmalloc_node(map, size, __GFP_HIGH, map->numa_node);
 	if (!cb) {
 		ret = -ENOMEM;
 		goto out;
Signature · 1 line
2.51.0.355.g5224444f11-goog
#2 lore raw
In reply to Peilin Ye
Quoted 96 lines · Peilin Ye

Currently, calling bpf_map_kmalloc_node() from __bpf_async_init() can cause various locking issues; see the following stack trace (edited for style) as one example:

...

 [10.011566]  do_raw_spin_lock.cold
 [10.011570]  try_to_wake_up             (5) double-acquiring the same
 [10.011575]  kick_pool                      rq_lock, causing a hardlockup
 [10.011579]  __queue_work
 [10.011582]  queue_work_on
 [10.011585]  kernfs_notify
 [10.011589]  cgroup_file_notify
 [10.011593]  try_charge_memcg           (4) memcg accounting raises an
 [10.011597]  obj_cgroup_charge_pages        MEMCG_MAX event
 [10.011599]  obj_cgroup_charge_account
 [10.011600]  __memcg_slab_post_alloc_hook
 [10.011603]  __kmalloc_node_noprof
...
 [10.011611]  bpf_map_kmalloc_node
 [10.011612]  __bpf_async_init
 [10.011615]  bpf_timer_init             (3) BPF calls bpf_timer_init()
 [10.011617]  bpf_prog_xxxxxxxxxxxxxxxx_fcg_runnable
 [10.011619]  bpf__sched_ext_ops_runnable
 [10.011620]  enqueue_task_scx           (2) BPF runs with rq_lock held
 [10.011622]  enqueue_task
 [10.011626]  ttwu_do_activate
 [10.011629]  sched_ttwu_pending         (1) grabs rq_lock

...

The above was reproduced on bpf-next (b338cf849ec8) by modifying ./tools/sched_ext/scx_flatcg.bpf.c to call bpf_timer_init() during ops.runnable(), and hacking [1] the memcg accounting code a bit to make a bpf_timer_init() call much more likely to raise an MEMCG_MAX event.

We have also run into other similar variants (both internally and on bpf-next), including double-acquiring cgroup_file_kn_lock, the same worker_pool::lock, etc.

As suggested by Shakeel, fix this by using __GFP_HIGH instead of GFP_ATOMIC in __bpf_async_init(), so that if try_charge_memcg() raises an MEMCG_MAX event, we call __memcg_memory_event() with @allow_spinning=false and skip calling cgroup_file_notify(), in order to avoid the locking issues described above.

Depends on mm patch "memcg: skip cgroup_file_notify if spinning is not allowed". Tested with vmtest.sh (llvm-18, x86-64):

 $ ./test_progs -a '*timer*' -a '*wq*'
...
 Summary: 7/12 PASSED, 0 SKIPPED, 0 FAILED

[1] Making bpf_timer_init() much more likely to raise an MEMCG_MAX event (gist-only, for brevity):

kernel/bpf/helpers.c:__bpf_async_init():
- cb = bpf_map_kmalloc_node(map, size, GFP_ATOMIC, map->numa_node);
+ cb = bpf_map_kmalloc_node(map, size, GFP_ATOMIC | __GFP_HACK,
+ map->numa_node);

mm/memcontrol.c:try_charge_memcg():

          if (!do_memsw_account() ||
 -            page_counter_try_charge(&memcg->memsw, batch, &counter)) {
 -                if (page_counter_try_charge(&memcg->memory, batch, &counter))
 +            page_counter_try_charge_hack(&memcg->memsw, batch, &counter,
 +                                         gfp_mask & __GFP_HACK)) {
 +                if (page_counter_try_charge_hack(&memcg->memory, batch,
 +                                                 &counter,
 +                                                 gfp_mask & __GFP_HACK))
                          goto done_restock;

mm/page_counter.c:page_counter_try_charge():
-bool page_counter_try_charge(struct page_counter *counter,
- unsigned long nr_pages,
- struct page_counter **fail)
+bool page_counter_try_charge_hack(struct page_counter *counter,
+ unsigned long nr_pages,
+ struct page_counter **fail, bool hack)
{
...
- if (new > c->max) {

 +                if (hack || new > c->max) {     // goto failed;
                          atomic_long_sub(nr_pages, &c->usage);
Fixes b00628b1c7d5 ("bpf: Introduce bpf timers.")
---
 kernel/bpf/helpers.c | 10 ++++++++--
 1 file changed, 8 insertions(+), 2 deletions(-)
kernel/bpf/helpers.c +0 -0
index b9b0c5fe33f6..508b13c24778 100644
@@ -1274,8 +1274,14 @@ static int __bpf_async_init(struct bpf_async_kern *async, struct bpf_map *map, u
 		goto out;

kernel/bpf/helpers.c static int __bpf_async_init(struct bpf_async_kern *async, struct bpf_map *map, u · line 1282

Quoted 10 lines
kernel/bpf/helpers.c +6 -2
 	}
 
-	/* allocate hrtimer via map_kmalloc to use memcg accounting */
-	cb = bpf_map_kmalloc_node(map, size, GFP_ATOMIC, map->numa_node);
+	/* Allocate via bpf_map_kmalloc_node() for memcg accounting. Use
+	 * __GFP_HIGH instead of GFP_ATOMIC to avoid calling
+	 * cgroup_file_notify() if an MEMCG_MAX event is raised by
+	 * try_charge_memcg(). This prevents various locking issues, including
+	 * double-acquiring locks that may already be held here (e.g.,
+	 * cgroup_file_kn_lock, rq_lock).

Too much unnecessary information in the comment. Just mention that we want nolock allocations and for that we need to remove __GFP_RECLAIM flags until nolock allocation interfaces are available.

author
#3 lore raw
Quoted 9 lines · Shakeel Butt
Quoted 8 lines · Peilin Ye
kernel/bpf/helpers.c +0 -0
index b9b0c5fe33f6..508b13c24778 100644
@@ -1274,8 +1274,14 @@ static int __bpf_async_init(struct bpf_async_kern *async, struct bpf_map *map, u
 		goto out;
 	}
 
Quoted 12 lines

kernel/bpf/helpers.c static int __bpf_async_init(struct bpf_async_kern *async, struct bpf_map *map, u · line 1282

Quoted 8 lines
kernel/bpf/helpers.c +6 -2
-	/* allocate hrtimer via map_kmalloc to use memcg accounting */
-	cb = bpf_map_kmalloc_node(map, size, GFP_ATOMIC, map->numa_node);
+	/* Allocate via bpf_map_kmalloc_node() for memcg accounting. Use
+	 * __GFP_HIGH instead of GFP_ATOMIC to avoid calling
+	 * cgroup_file_notify() if an MEMCG_MAX event is raised by
+	 * try_charge_memcg(). This prevents various locking issues, including
+	 * double-acquiring locks that may already be held here (e.g.,
+	 * cgroup_file_kn_lock, rq_lock).

Too much unnecessary information in the comment. Just mention that we want nolock allocations and for that we need to remove __GFP_RECLAIM flags until nolock allocation interfaces are available.

Got it - I'll reword in v2.

Thanks,
Peilin Ye