The patch titled Subject: kmemleak: avoid RCU stalls when freeing metadata for per-CPU pointers has been added to the -mm mm-unstable branch. Its filename is kmemleak-avoid-rcu-stalls-when-freeing-metadata-for-per-cpu-pointers.patch This patch will shortly appear at https://git.kernel.org/pub/scm/linux/kernel/git/akpm/25-new.git/tree/patches/kmemleak-avoid-rcu-stalls-when-freeing-metadata-for-per-cpu-pointers.patch This patch will later appear in the mm-unstable branch at git://git.kernel.org/pub/scm/linux/kernel/git/akpm/mm Before you just go and hit "reply", please: a) Consider who else should be cc'ed b) Prefer to cc a suitable mailing list as well c) Ideally: find the original patch on the mailing list and do a reply-to-all to that, adding suitable additional cc's *** Remember to use Documentation/process/submit-checklist.rst when testing your code *** The -mm tree is included into linux-next via the mm-everything branch at git://git.kernel.org/pub/scm/linux/kernel/git/akpm/mm and is updated there every 2-3 working days ------------------------------------------------------ From: Catalin Marinas <catalin.marinas@xxxxxxx> Subject: kmemleak: avoid RCU stalls when freeing metadata for per-CPU pointers Date: Fri, 1 Dec 2023 19:08:29 +0000 On systems with large number of CPUs, the following soft lockup splat might sometimes happen: [ 2656.001617] watchdog: BUG: soft lockup - CPU#364 stuck for 21s! [ksoftirqd/364:2206] : [ 2656.141194] RIP: 0010:_raw_spin_unlock_irqrestore+0x3d/0x70 : 2656.241214] Call Trace: [ 2656.243971] <IRQ> [ 2656.246237] ? show_trace_log_lvl+0x1c4/0x2df [ 2656.251152] ? show_trace_log_lvl+0x1c4/0x2df [ 2656.256066] ? kmemleak_free_percpu+0x11f/0x1f0 [ 2656.261173] ? watchdog_timer_fn+0x379/0x470 [ 2656.265984] ? __pfx_watchdog_timer_fn+0x10/0x10 [ 2656.271179] ? __hrtimer_run_queues+0x5f3/0xd00 [ 2656.276283] ? __pfx___hrtimer_run_queues+0x10/0x10 [ 2656.281783] ? ktime_get_update_offsets_now+0x95/0x2c0 [ 2656.287573] ? ktime_get_update_offsets_now+0xdd/0x2c0 [ 2656.293380] ? hrtimer_interrupt+0x2e9/0x780 [ 2656.298221] ? __sysvec_apic_timer_interrupt+0x184/0x640 [ 2656.304211] ? sysvec_apic_timer_interrupt+0x8e/0xc0 [ 2656.309807] </IRQ> [ 2656.312169] <TASK> [ 2656.326110] kmemleak_free_percpu+0x11f/0x1f0 [ 2656.331015] free_percpu.part.0+0x1b/0xe70 [ 2656.335635] free_vfsmnt+0xb9/0x100 [ 2656.339567] rcu_do_batch+0x3c8/0xe30 [ 2656.363693] rcu_core+0x3de/0x5a0 [ 2656.367433] __do_softirq+0x2d0/0x9a8 [ 2656.381119] run_ksoftirqd+0x36/0x60 [ 2656.385145] smpboot_thread_fn+0x556/0x910 [ 2656.394971] kthread+0x2a4/0x350 [ 2656.402826] ret_from_fork+0x29/0x50 [ 2656.406861] </TASK> The issue is caused by kmemleak registering each per_cpu_ptr() corresponding to the __percpu pointer. This is unnecessary since such individual per-CPU pointers are not tracked anyway. Create a new object_percpu_tree_root rbtree that stores a single __percpu pointer together with an OBJECT_PERCPU flag for the kmemleak metadata. Scanning needs to be done for all per_cpu_ptr() pointers with a cond_resched() between each CPU iteration to avoid RCU stalls. Link: https://lkml.kernel.org/r/20231201190829.825856-1-catalin.marinas@xxxxxxx Signed-off-by: Catalin Marinas <catalin.marinas@xxxxxxx> Reported-by: Waiman Long <longman@xxxxxxxxxx> Closes: https://lore.kernel.org/r/20231127194153.289626-1-longman@xxxxxxxxxx Signed-off-by: Andrew Morton <akpm@xxxxxxxxxxxxxxxxxxxx> --- mm/kmemleak.c | 178 ++++++++++++++++++++++++++---------------------- 1 file changed, 97 insertions(+), 81 deletions(-) --- a/mm/kmemleak.c~kmemleak-avoid-rcu-stalls-when-freeing-metadata-for-per-cpu-pointers +++ a/mm/kmemleak.c @@ -14,17 +14,15 @@ * The following locks and mutexes are used by kmemleak: * * - kmemleak_lock (raw_spinlock_t): protects the object_list as well as - * del_state modifications and accesses to the object_tree_root (or - * object_phys_tree_root). The object_list is the main list holding the - * metadata (struct kmemleak_object) for the allocated memory blocks. - * The object_tree_root and object_phys_tree_root are red - * black trees used to look-up metadata based on a pointer to the - * corresponding memory block. The object_phys_tree_root is for objects - * allocated with physical address. The kmemleak_object structures are - * added to the object_list and object_tree_root (or object_phys_tree_root) - * in the create_object() function called from the kmemleak_alloc() (or - * kmemleak_alloc_phys()) callback and removed in delete_object() called from - * the kmemleak_free() callback + * del_state modifications and accesses to the object trees + * (object_tree_root, object_phys_tree_root, object_percpu_tree_root). The + * object_list is the main list holding the metadata (struct + * kmemleak_object) for the allocated memory blocks. The object trees are + * red black trees used to look-up metadata based on a pointer to the + * corresponding memory block. The kmemleak_object structures are added to + * the object_list and the object tree root in the create_object() function + * called from the kmemleak_alloc() (or kmemleak_alloc_phys()) callback and + * removed in delete_object() called from the kmemleak_free() callback * - kmemleak_object.lock (raw_spinlock_t): protects a kmemleak_object. * Accesses to the metadata (e.g. count) are protected by this lock. Note * that some members of this structure may be protected by other means @@ -178,6 +176,8 @@ struct kmemleak_object { #define OBJECT_FULL_SCAN (1 << 3) /* flag set for object allocated with physical address */ #define OBJECT_PHYS (1 << 4) +/* flag set for per-CPU pointers */ +#define OBJECT_PERCPU (1 << 5) /* set when __remove_object() called */ #define DELSTATE_REMOVED (1 << 0) @@ -206,6 +206,8 @@ static LIST_HEAD(mem_pool_free_list); static struct rb_root object_tree_root = RB_ROOT; /* search tree for object (with OBJECT_PHYS flag) boundaries */ static struct rb_root object_phys_tree_root = RB_ROOT; +/* search tree for object (with OBJECT_PERCPU flag) boundaries */ +static struct rb_root object_percpu_tree_root = RB_ROOT; /* protecting the access to object_list, object_tree_root (or object_phys_tree_root) */ static DEFINE_RAW_SPINLOCK(kmemleak_lock); @@ -298,7 +300,7 @@ static void hex_dump_object(struct seq_f const u8 *ptr = (const u8 *)object->pointer; size_t len; - if (WARN_ON_ONCE(object->flags & OBJECT_PHYS)) + if (WARN_ON_ONCE(object->flags & (OBJECT_PHYS | OBJECT_PERCPU))) return; /* limit the number of lines to HEX_MAX_LINES */ @@ -390,6 +392,15 @@ static void dump_object_info(struct kmem stack_depot_print(object->trace_handle); } +static struct rb_root *object_tree(unsigned long objflags) +{ + if (objflags & OBJECT_PHYS) + return &object_phys_tree_root; + if (objflags & OBJECT_PERCPU) + return &object_percpu_tree_root; + return &object_tree_root; +} + /* * Look-up a memory block metadata (kmemleak_object) in the object search * tree based on a pointer value. If alias is 0, only values pointing to the @@ -397,10 +408,9 @@ static void dump_object_info(struct kmem * when calling this function. */ static struct kmemleak_object *__lookup_object(unsigned long ptr, int alias, - bool is_phys) + unsigned int objflags) { - struct rb_node *rb = is_phys ? object_phys_tree_root.rb_node : - object_tree_root.rb_node; + struct rb_node *rb = object_tree(objflags)->rb_node; unsigned long untagged_ptr = (unsigned long)kasan_reset_tag((void *)ptr); while (rb) { @@ -429,7 +439,7 @@ static struct kmemleak_object *__lookup_ /* Look-up a kmemleak object which allocated with virtual address. */ static struct kmemleak_object *lookup_object(unsigned long ptr, int alias) { - return __lookup_object(ptr, alias, false); + return __lookup_object(ptr, alias, 0); } /* @@ -542,14 +552,14 @@ static void put_object(struct kmemleak_o * Look up an object in the object search tree and increase its use_count. */ static struct kmemleak_object *__find_and_get_object(unsigned long ptr, int alias, - bool is_phys) + unsigned int objflags) { unsigned long flags; struct kmemleak_object *object; rcu_read_lock(); raw_spin_lock_irqsave(&kmemleak_lock, flags); - object = __lookup_object(ptr, alias, is_phys); + object = __lookup_object(ptr, alias, objflags); raw_spin_unlock_irqrestore(&kmemleak_lock, flags); /* check whether the object is still available */ @@ -563,19 +573,16 @@ static struct kmemleak_object *__find_an /* Look up and get an object which allocated with virtual address. */ static struct kmemleak_object *find_and_get_object(unsigned long ptr, int alias) { - return __find_and_get_object(ptr, alias, false); + return __find_and_get_object(ptr, alias, 0); } /* - * Remove an object from the object_tree_root (or object_phys_tree_root) - * and object_list. Must be called with the kmemleak_lock held _if_ kmemleak - * is still enabled. + * Remove an object from its object tree and object_list. Must be called with + * the kmemleak_lock held _if_ kmemleak is still enabled. */ static void __remove_object(struct kmemleak_object *object) { - rb_erase(&object->rb_node, object->flags & OBJECT_PHYS ? - &object_phys_tree_root : - &object_tree_root); + rb_erase(&object->rb_node, object_tree(object->flags)); if (!(object->del_state & DELSTATE_NO_DELETE)) list_del_rcu(&object->object_list); object->del_state |= DELSTATE_REMOVED; @@ -583,11 +590,11 @@ static void __remove_object(struct kmeml static struct kmemleak_object *__find_and_remove_object(unsigned long ptr, int alias, - bool is_phys) + unsigned int objflags) { struct kmemleak_object *object; - object = __lookup_object(ptr, alias, is_phys); + object = __lookup_object(ptr, alias, objflags); if (object) __remove_object(object); @@ -595,19 +602,18 @@ static struct kmemleak_object *__find_an } /* - * Look up an object in the object search tree and remove it from both - * object_tree_root (or object_phys_tree_root) and object_list. The - * returned object's use_count should be at least 1, as initially set - * by create_object(). + * Look up an object in the object search tree and remove it from both object + * tree root and object_list. The returned object's use_count should be at + * least 1, as initially set by create_object(). */ static struct kmemleak_object *find_and_remove_object(unsigned long ptr, int alias, - bool is_phys) + unsigned int objflags) { unsigned long flags; struct kmemleak_object *object; raw_spin_lock_irqsave(&kmemleak_lock, flags); - object = __find_and_remove_object(ptr, alias, is_phys); + object = __find_and_remove_object(ptr, alias, objflags); raw_spin_unlock_irqrestore(&kmemleak_lock, flags); return object; @@ -678,7 +684,7 @@ static struct kmemleak_object *__alloc_o } static int __link_object(struct kmemleak_object *object, unsigned long ptr, - size_t size, int min_count, bool is_phys) + size_t size, int min_count, unsigned int objflags) { struct kmemleak_object *parent; @@ -686,7 +692,7 @@ static int __link_object(struct kmemleak unsigned long untagged_ptr; unsigned long untagged_objp; - object->flags = OBJECT_ALLOCATED | (is_phys ? OBJECT_PHYS : 0); + object->flags = OBJECT_ALLOCATED | objflags; object->pointer = ptr; object->size = kfence_ksize((void *)ptr) ?: size; object->min_count = min_count; @@ -697,12 +703,11 @@ static int __link_object(struct kmemleak * Only update min_addr and max_addr with object * storing virtual address. */ - if (!is_phys) { + if (!(objflags & (OBJECT_PHYS | OBJECT_PERCPU))) { min_addr = min(min_addr, untagged_ptr); max_addr = max(max_addr, untagged_ptr + size); } - link = is_phys ? &object_phys_tree_root.rb_node : - &object_tree_root.rb_node; + link = &object_tree(objflags)->rb_node; rb_parent = NULL; while (*link) { rb_parent = *link; @@ -724,8 +729,7 @@ static int __link_object(struct kmemleak } } rb_link_node(&object->rb_node, rb_parent, link); - rb_insert_color(&object->rb_node, is_phys ? &object_phys_tree_root : - &object_tree_root); + rb_insert_color(&object->rb_node, object_tree(objflags)); list_add_tail_rcu(&object->object_list, &object_list); return 0; @@ -733,11 +737,10 @@ static int __link_object(struct kmemleak /* * Create the metadata (struct kmemleak_object) corresponding to an allocated - * memory block and add it to the object_list and object_tree_root (or - * object_phys_tree_root). + * memory block and add it to the object_list and object tree. */ static void __create_object(unsigned long ptr, size_t size, - int min_count, gfp_t gfp, bool is_phys) + int min_count, gfp_t gfp, unsigned int objflags) { struct kmemleak_object *object; unsigned long flags; @@ -748,7 +751,7 @@ static void __create_object(unsigned lon return; raw_spin_lock_irqsave(&kmemleak_lock, flags); - ret = __link_object(object, ptr, size, min_count, is_phys); + ret = __link_object(object, ptr, size, min_count, objflags); raw_spin_unlock_irqrestore(&kmemleak_lock, flags); if (ret) mem_pool_free(object); @@ -758,14 +761,21 @@ static void __create_object(unsigned lon static void create_object(unsigned long ptr, size_t size, int min_count, gfp_t gfp) { - __create_object(ptr, size, min_count, gfp, false); + __create_object(ptr, size, min_count, gfp, 0); } /* Create kmemleak object which allocated with physical address. */ static void create_object_phys(unsigned long ptr, size_t size, int min_count, gfp_t gfp) { - __create_object(ptr, size, min_count, gfp, true); + __create_object(ptr, size, min_count, gfp, OBJECT_PHYS); +} + +/* Create kmemleak object corresponding to a per-CPU allocation. */ +static void create_object_percpu(unsigned long ptr, size_t size, + int min_count, gfp_t gfp) +{ + __create_object(ptr, size, min_count, gfp, OBJECT_PERCPU); } /* @@ -792,11 +802,11 @@ static void __delete_object(struct kmeml * Look up the metadata (struct kmemleak_object) corresponding to ptr and * delete it. */ -static void delete_object_full(unsigned long ptr) +static void delete_object_full(unsigned long ptr, unsigned int objflags) { struct kmemleak_object *object; - object = find_and_remove_object(ptr, 0, false); + object = find_and_remove_object(ptr, 0, objflags); if (!object) { #ifdef DEBUG kmemleak_warn("Freeing unknown object at 0x%08lx\n", @@ -812,7 +822,8 @@ static void delete_object_full(unsigned * delete it. If the memory block is partially freed, the function may create * additional metadata for the remaining parts of the block. */ -static void delete_object_part(unsigned long ptr, size_t size, bool is_phys) +static void delete_object_part(unsigned long ptr, size_t size, + unsigned int objflags) { struct kmemleak_object *object, *object_l, *object_r; unsigned long start, end, flags; @@ -826,7 +837,7 @@ static void delete_object_part(unsigned goto out; raw_spin_lock_irqsave(&kmemleak_lock, flags); - object = __find_and_remove_object(ptr, 1, is_phys); + object = __find_and_remove_object(ptr, 1, objflags); if (!object) { #ifdef DEBUG kmemleak_warn("Partially freeing unknown object at 0x%08lx (size %zu)\n", @@ -844,11 +855,11 @@ static void delete_object_part(unsigned end = object->pointer + object->size; if ((ptr > start) && !__link_object(object_l, start, ptr - start, - object->min_count, is_phys)) + object->min_count, objflags)) object_l = NULL; if ((ptr + size < end) && !__link_object(object_r, ptr + size, end - ptr - size, - object->min_count, is_phys)) + object->min_count, objflags)) object_r = NULL; unlock: @@ -879,11 +890,11 @@ static void paint_it(struct kmemleak_obj raw_spin_unlock_irqrestore(&object->lock, flags); } -static void paint_ptr(unsigned long ptr, int color, bool is_phys) +static void paint_ptr(unsigned long ptr, int color, unsigned int objflags) { struct kmemleak_object *object; - object = __find_and_get_object(ptr, 0, is_phys); + object = __find_and_get_object(ptr, 0, objflags); if (!object) { kmemleak_warn("Trying to color unknown object at 0x%08lx as %s\n", ptr, @@ -901,16 +912,16 @@ static void paint_ptr(unsigned long ptr, */ static void make_gray_object(unsigned long ptr) { - paint_ptr(ptr, KMEMLEAK_GREY, false); + paint_ptr(ptr, KMEMLEAK_GREY, 0); } /* * Mark the object as black-colored so that it is ignored from scans and * reporting. */ -static void make_black_object(unsigned long ptr, bool is_phys) +static void make_black_object(unsigned long ptr, unsigned int objflags) { - paint_ptr(ptr, KMEMLEAK_BLACK, is_phys); + paint_ptr(ptr, KMEMLEAK_BLACK, objflags); } /* @@ -1046,8 +1057,6 @@ EXPORT_SYMBOL_GPL(kmemleak_alloc); void __ref kmemleak_alloc_percpu(const void __percpu *ptr, size_t size, gfp_t gfp) { - unsigned int cpu; - pr_debug("%s(0x%px, %zu)\n", __func__, ptr, size); /* @@ -1055,9 +1064,7 @@ void __ref kmemleak_alloc_percpu(const v * (min_count is set to 0). */ if (kmemleak_enabled && ptr && !IS_ERR(ptr)) - for_each_possible_cpu(cpu) - create_object((unsigned long)per_cpu_ptr(ptr, cpu), - size, 0, gfp); + create_object_percpu((unsigned long)ptr, size, 0, gfp); } EXPORT_SYMBOL_GPL(kmemleak_alloc_percpu); @@ -1098,7 +1105,7 @@ void __ref kmemleak_free(const void *ptr pr_debug("%s(0x%px)\n", __func__, ptr); if (kmemleak_free_enabled && ptr && !IS_ERR(ptr)) - delete_object_full((unsigned long)ptr); + delete_object_full((unsigned long)ptr, 0); } EXPORT_SYMBOL_GPL(kmemleak_free); @@ -1116,7 +1123,7 @@ void __ref kmemleak_free_part(const void pr_debug("%s(0x%px)\n", __func__, ptr); if (kmemleak_enabled && ptr && !IS_ERR(ptr)) - delete_object_part((unsigned long)ptr, size, false); + delete_object_part((unsigned long)ptr, size, 0); } EXPORT_SYMBOL_GPL(kmemleak_free_part); @@ -1129,14 +1136,10 @@ EXPORT_SYMBOL_GPL(kmemleak_free_part); */ void __ref kmemleak_free_percpu(const void __percpu *ptr) { - unsigned int cpu; - pr_debug("%s(0x%px)\n", __func__, ptr); if (kmemleak_free_enabled && ptr && !IS_ERR(ptr)) - for_each_possible_cpu(cpu) - delete_object_full((unsigned long)per_cpu_ptr(ptr, - cpu)); + delete_object_full((unsigned long)ptr, OBJECT_PERCPU); } EXPORT_SYMBOL_GPL(kmemleak_free_percpu); @@ -1206,7 +1209,7 @@ void __ref kmemleak_ignore(const void *p pr_debug("%s(0x%px)\n", __func__, ptr); if (kmemleak_enabled && ptr && !IS_ERR(ptr)) - make_black_object((unsigned long)ptr, false); + make_black_object((unsigned long)ptr, 0); } EXPORT_SYMBOL(kmemleak_ignore); @@ -1280,7 +1283,7 @@ void __ref kmemleak_free_part_phys(phys_ pr_debug("%s(0x%px)\n", __func__, &phys); if (kmemleak_enabled) - delete_object_part((unsigned long)phys, size, true); + delete_object_part((unsigned long)phys, size, OBJECT_PHYS); } EXPORT_SYMBOL(kmemleak_free_part_phys); @@ -1294,7 +1297,7 @@ void __ref kmemleak_ignore_phys(phys_add pr_debug("%s(0x%px)\n", __func__, &phys); if (kmemleak_enabled) - make_black_object((unsigned long)phys, true); + make_black_object((unsigned long)phys, OBJECT_PHYS); } EXPORT_SYMBOL(kmemleak_ignore_phys); @@ -1305,7 +1308,7 @@ static bool update_checksum(struct kmeml { u32 old_csum = object->checksum; - if (WARN_ON_ONCE(object->flags & OBJECT_PHYS)) + if (WARN_ON_ONCE(object->flags & (OBJECT_PHYS | OBJECT_PERCPU))) return false; kasan_disable_current(); @@ -1461,7 +1464,6 @@ static void scan_object(struct kmemleak_ { struct kmemleak_scan_area *area; unsigned long flags; - void *obj_ptr; /* * Once the object->lock is acquired, the corresponding memory block @@ -1474,14 +1476,27 @@ static void scan_object(struct kmemleak_ /* already freed object */ goto out; - obj_ptr = object->flags & OBJECT_PHYS ? - __va((phys_addr_t)object->pointer) : - (void *)object->pointer; + if (object->flags & OBJECT_PERCPU) { + unsigned int cpu; + + for_each_possible_cpu(cpu) { + void *start = per_cpu_ptr((void __percpu *)object->pointer, cpu); + void *end = start + object->size; - if (hlist_empty(&object->area_list) || + scan_block(start, end, object); + + raw_spin_unlock_irqrestore(&object->lock, flags); + cond_resched(); + raw_spin_lock_irqsave(&object->lock, flags); + if (!(object->flags & OBJECT_ALLOCATED)) + break; + } + } else if (hlist_empty(&object->area_list) || object->flags & OBJECT_FULL_SCAN) { - void *start = obj_ptr; - void *end = obj_ptr + object->size; + void *start = object->flags & OBJECT_PHYS ? + __va((phys_addr_t)object->pointer) : + (void *)object->pointer; + void *end = start + object->size; void *next; do { @@ -1496,11 +1511,12 @@ static void scan_object(struct kmemleak_ cond_resched(); raw_spin_lock_irqsave(&object->lock, flags); } while (object->flags & OBJECT_ALLOCATED); - } else + } else { hlist_for_each_entry(area, &object->area_list, node) scan_block((void *)area->start, (void *)(area->start + area->size), object); + } out: raw_spin_unlock_irqrestore(&object->lock, flags); } _ Patches currently in -mm which might be from catalin.marinas@xxxxxxx are kmemleak-avoid-rcu-stalls-when-freeing-metadata-for-per-cpu-pointers.patch