On Mon, Aug 18, 2025 at 10:33:51AM +0800, yangshiguang wrote:
At 2025-08-18 10:22:36, "Harry Yoo" harry.yoo@oracle.com wrote:
On Mon, Aug 18, 2025 at 10:07:40AM +0800, yangshiguang wrote:
At 2025-08-16 18:46:12, "Harry Yoo" harry.yoo@oracle.com wrote:
On Sat, Aug 16, 2025 at 06:05:15PM +0800, yangshiguang wrote:
At 2025-08-16 16:25:25, "Harry Yoo" harry.yoo@oracle.com wrote:
On Thu, Aug 14, 2025 at 07:16:42PM +0800, yangshiguang1011@163.com wrote: > From: yangshiguang yangshiguang@xiaomi.com > > From: yangshiguang yangshiguang@xiaomi.com > > set_track_prepare() can incur lock recursion. > The issue is that it is called from hrtimer_start_range_ns > holding the per_cpu(hrtimer_bases)[n].lock, but when enabled > CONFIG_DEBUG_OBJECTS_TIMERS, may wake up kswapd in set_track_prepare, > and try to hold the per_cpu(hrtimer_bases)[n].lock. > > So avoid waking up kswapd.The oops looks something like:
Hi yangshiguang,
In the next revision, could you please elaborate the commit message to reflect how this change avoids waking up kswapd?
of course. Thanks for the reminder.
> BUG: spinlock recursion on CPU#3, swapper/3/0 > lock: 0xffffff8a4bf29c80, .magic: dead4ead, .owner: swapper/3/0, .owner_cpu: 3 > Hardware name: Qualcomm Technologies, Inc. Popsicle based on SM8850 (DT) > Call trace: > spin_bug+0x0 > _raw_spin_lock_irqsave+0x80 > hrtimer_try_to_cancel+0x94 > task_contending+0x10c > enqueue_dl_entity+0x2a4 > dl_server_start+0x74 > enqueue_task_fair+0x568 > enqueue_task+0xac > do_activate_task+0x14c > ttwu_do_activate+0xcc > try_to_wake_up+0x6c8 > default_wake_function+0x20 > autoremove_wake_function+0x1c > __wake_up+0xac > wakeup_kswapd+0x19c > wake_all_kswapds+0x78 > __alloc_pages_slowpath+0x1ac > __alloc_pages_noprof+0x298 > stack_depot_save_flags+0x6b0 > stack_depot_save+0x14 > set_track_prepare+0x5c > ___slab_alloc+0xccc > __kmalloc_cache_noprof+0x470 > __set_page_owner+0x2bc > post_alloc_hook[jt]+0x1b8 > prep_new_page+0x28 > get_page_from_freelist+0x1edc > __alloc_pages_noprof+0x13c > alloc_slab_page+0x244 > allocate_slab+0x7c > ___slab_alloc+0x8e8 > kmem_cache_alloc_noprof+0x450 > debug_objects_fill_pool+0x22c > debug_object_activate+0x40 > enqueue_hrtimer[jt]+0xdc > hrtimer_start_range_ns+0x5f8 > ... > > Signed-off-by: yangshiguang yangshiguang@xiaomi.com > Fixes: 5cf909c553e9 ("mm/slub: use stackdepot to save stack trace in objects") > --- > v1 -> v2: > propagate gfp flags to set_track_prepare() > > [1] https://urldefense.com/v3/__https://lore.kernel.org/all/20250801065121.87679... > --- > mm/slub.c | 21 +++++++++++---------- > 1 file changed, 11 insertions(+), 10 deletions(-) > > diff --git a/mm/slub.c b/mm/slub.c > index 30003763d224..dba905bf1e03 100644 > --- a/mm/slub.c > +++ b/mm/slub.c > @@ -962,19 +962,20 @@ static struct track *get_track(struct kmem_cache *s, void *object, > } > > #ifdef CONFIG_STACKDEPOT > -static noinline depot_stack_handle_t set_track_prepare(void) > +static noinline depot_stack_handle_t set_track_prepare(gfp_t gfp_flags) > { > depot_stack_handle_t handle; > unsigned long entries[TRACK_ADDRS_COUNT]; > unsigned int nr_entries; > + gfp_flags &= GFP_NOWAIT;
Is there any reason to downgrade it to GFP_NOWAIT when the gfp flag allows direct reclamation?
Hi Harry,
The original allocation is GFP_NOWAIT. So I think it's better not to increase the allocation cost here.
I don't think the allocation cost is important here, because collecting a stack trace for each alloc/free is quite slow anyway. And we don't really care about performance in debug caches (it isn't designed to be performant).
I think it was GFP_NOWAIT because it was considered safe without regard to the GFP flags passed, rather than due to performance considerations.
Hi harry,
Is that so? gfp_flags &= (GFP_NOWAIT | __GFP_DIRECT_RECLAIM);
This still clears gfp flags passed by the caller to the allocator. Why not use gfp_flags directly without clearing some flags?
Hi Harry,
This introduces new problems.
call stackļ¼ dump_backtrace+0xfc/0x17c show_stack+0x18/0x28 dump_stack_lvl+0x40/0xc0 dump_stack+0x18/0x24 __might_resched+0x164/0x184 __might_sleep+0x38/0x84 prepare_alloc_pages+0xc0/0x17c __alloc_pages_noprof+0x130/0x3f8 stack_depot_save_flags+0x5a8/0x6bc stack_depot_save+0x14/0x24 set_track_prepare+0x64/0x90 ___slab_alloc+0xc14/0xc48 __kmalloc_cache_noprof+0x398/0x568 __kthread_create_on_node+0x8c/0x1f0 kthread_create_on_node+0x4c/0x74 create_worker+0xe0/0x298 workqueue_init+0x228/0x324 kernel_init_freeable+0x124/0x1c8 kernel_init+0x20/0x1ac ret_from_fork+0x10/0x20
Ok, because preemption is disabled in ___slab_alloc(), blocking allocations are not allowed even when gfp_flags allows it. So __GFP_DIRECT_RECLAIM should be cleared.
So,
/* Preemption is disabled in ___slab_alloc() */ gfp_flags &= ~(__GFP_DIRECT_RECLAIM);
should work?
Of course there are other problems.
So it is best to limit gtp flags.