Re:Re: [PATCH v4] mm: slub: avoid wake up kswapd in set_track_prepare

[yangshiguang <yangshiguang1011@163.com>]

At 2025-09-01 16:46:13, "Vlastimil Babka" <vbabka@suse.cz> wrote:
>On 9/1/25 10:29, yangshiguang wrote:
>> 
>> 
>> At 2025-09-01 16:15:04, "Vlastimil Babka" <vbabka@suse.cz> wrote:
>>>On 9/1/25 09:50, David Rientjes wrote:
>>>> On Sat, 30 Aug 2025, yangshiguang1011@163.com wrote:
>>>> 
>>>>> From: yangshiguang <yangshiguang@xiaomi.com>
>>>>> 
>>>>> From: yangshiguang <yangshiguang@xiaomi.com>
>>>>> 
>>>> 
>>>> Duplicate lines.
>>>> 
>>>>> 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.
>>>>> 
>>>>> Avoid deadlock caused by implicitly waking up kswapd by
>>>>> passing in allocation flags. And the slab caller context has
>>>>> preemption disabled, so __GFP_KSWAPD_RECLAIM must not appear in gfp_flags.
>>>>> 
>>>> 
>>>> This mentions __GFP_KSWAPD_RECLAIM, but the patch actually masks off 
>>>> __GFP_DIRECT_RECLAIM which would be a heavierweight operation.  Disabling 
>>>> direct reclaim does not necessarily imply that kswapd will be disabled as 
>>>> well.
>>>
>>>Yeah I think the changelog should say __GFP_DIRECT_RECLAIM.
>>>
>>>> Are you meaning to clear __GFP_RECLAIM in set_track_prepare()?
>>>
>>>No because if the context context (e.g. the hrtimers) can't support
>>>__GFP_KSWAPD_RECLAIM it won't have it in gfp_flags and we now pass them to
>> 
>>>set_track_prepare() so it already won't be there.
>> 
>> 
>> Sry. Should be __GFP_DIRECT_RECLAIM. I will resend the patch.
>
>I have adjusted it locally already. Also moved the masking of
>__GFP_DIRECT_RECLAIM to ___slab_alloc itself as that's where
>the preemption is disabled so it's more obvious.
>
>Does the result look good to you?

This looks good.
Currently only ___slab_alloc disables preemption context calls to set_track. 
In the future, not all callers will disable preemption. 


>
>commit 1b7052bc536650f8ca29b4f6f8682dc9f5692d16
>Author: yangshiguang <yangshiguang@xiaomi.com>
>Date:   Sat Aug 30 10:09:46 2025 +0800
>
>    mm: slub: avoid wake up kswapd in set_track_prepare
>    
>    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.
>    
>    Avoid deadlock caused by implicitly waking up kswapd by passing in
>    allocation flags, which do not contain __GFP_KSWAPD_RECLAIM in the
>    debug_objects_fill_pool() case.
>    Since ___slab_alloc() has preemption disabled and thus was using
>    GFP_NOWAIT, we instead mask out __GFP_DIRECT_RECLAIM from the flags.
>    
>    The oops looks something like:
>    
>    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")
>    Cc: stable@vger.kernel.org
>    Signed-off-by: Vlastimil Babka <vbabka@suse.cz>
>
>diff --git a/mm/slub.c b/mm/slub.c
>index 1787e4d51e48..d257141896c9 100644
>--- a/mm/slub.c
>+++ b/mm/slub.c
>@@ -962,19 +962,19 @@ 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;
> 
> 	nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 3);
>-	handle = stack_depot_save(entries, nr_entries, GFP_NOWAIT);
>+	handle = stack_depot_save(entries, nr_entries, gfp_flags);
> 
> 	return handle;
> }
> #else
>-static inline depot_stack_handle_t set_track_prepare(void)
>+static inline depot_stack_handle_t set_track_prepare(gfp_t gfp_flags)
> {
> 	return 0;
> }
>@@ -996,9 +996,9 @@ static void set_track_update(struct kmem_cache *s, void *object,
> }
> 
> static __always_inline void set_track(struct kmem_cache *s, void *object,
>-				      enum track_item alloc, unsigned long addr)
>+				      enum track_item alloc, unsigned long addr, gfp_t gfp_flags)
> {
>-	depot_stack_handle_t handle = set_track_prepare();
>+	depot_stack_handle_t handle = set_track_prepare(gfp_flags);
> 
> 	set_track_update(s, object, alloc, addr, handle);
> }
>@@ -1926,9 +1926,9 @@ static inline bool free_debug_processing(struct kmem_cache *s,
> static inline void slab_pad_check(struct kmem_cache *s, struct slab *slab) {}
> static inline int check_object(struct kmem_cache *s, struct slab *slab,
> 			void *object, u8 val) { return 1; }
>-static inline depot_stack_handle_t set_track_prepare(void) { return 0; }
>+static inline depot_stack_handle_t set_track_prepare(gfp_t gfp_flags) { return 0; }
> static inline void set_track(struct kmem_cache *s, void *object,
>-			     enum track_item alloc, unsigned long addr) {}
>+			     enum track_item alloc, unsigned long addr, gfp_t gfp_flags) {}
> static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
> 					struct slab *slab) {}
> static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
>@@ -3881,9 +3881,14 @@ static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
> 			 * For debug caches here we had to go through
> 			 * alloc_single_from_partial() so just store the
> 			 * tracking info and return the object.
>+			 *
>+			 * Due to disabled preemption we need to disallow
>+			 * blocking. The flags are further adjusted by
>+			 * gfp_nested_mask() in stack_depot itself.
> 			 */
> 			if (s->flags & SLAB_STORE_USER)
>-				set_track(s, freelist, TRACK_ALLOC, addr);
>+				set_track(s, freelist, TRACK_ALLOC, addr,
>+					  gfpflags & ~(__GFP_DIRECT_RECLAIM));
> 
> 			return freelist;
> 		}
>@@ -3915,7 +3920,8 @@ static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
> 			goto new_objects;
> 
> 		if (s->flags & SLAB_STORE_USER)
>-			set_track(s, freelist, TRACK_ALLOC, addr);
>+			set_track(s, freelist, TRACK_ALLOC, addr,
>+				  gfpflags & ~(__GFP_DIRECT_RECLAIM));
> 
> 		return freelist;
> 	}
>@@ -4426,8 +4432,12 @@ static noinline void free_to_partial_list(
> 	unsigned long flags;
> 	depot_stack_handle_t handle = 0;
> 
>+	/*
>+	 * We cannot use GFP_NOWAIT as there are callsites where waking up
>+	 * kswapd could deadlock
>+	 */
> 	if (s->flags & SLAB_STORE_USER)
>-		handle = set_track_prepare();
>+		handle = set_track_prepare(__GFP_NOWARN);
> 
> 	spin_lock_irqsave(&n->list_lock, flags);
>