1 // SPDX-License-Identifier: GPL-2.0
4 * Written by Mark Hemment, 1996/97.
5 * (markhe@nextd.demon.co.uk)
7 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
9 * Major cleanup, different bufctl logic, per-cpu arrays
10 * (c) 2000 Manfred Spraul
12 * Cleanup, make the head arrays unconditional, preparation for NUMA
13 * (c) 2002 Manfred Spraul
15 * An implementation of the Slab Allocator as described in outline in;
16 * UNIX Internals: The New Frontiers by Uresh Vahalia
17 * Pub: Prentice Hall ISBN 0-13-101908-2
18 * or with a little more detail in;
19 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
20 * Jeff Bonwick (Sun Microsystems).
21 * Presented at: USENIX Summer 1994 Technical Conference
23 * The memory is organized in caches, one cache for each object type.
24 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
25 * Each cache consists out of many slabs (they are small (usually one
26 * page long) and always contiguous), and each slab contains multiple
27 * initialized objects.
29 * This means, that your constructor is used only for newly allocated
30 * slabs and you must pass objects with the same initializations to
33 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
34 * normal). If you need a special memory type, then must create a new
35 * cache for that memory type.
37 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
38 * full slabs with 0 free objects
40 * empty slabs with no allocated objects
42 * If partial slabs exist, then new allocations come from these slabs,
43 * otherwise from empty slabs or new slabs are allocated.
45 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
46 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
48 * Each cache has a short per-cpu head array, most allocs
49 * and frees go into that array, and if that array overflows, then 1/2
50 * of the entries in the array are given back into the global cache.
51 * The head array is strictly LIFO and should improve the cache hit rates.
52 * On SMP, it additionally reduces the spinlock operations.
54 * The c_cpuarray may not be read with enabled local interrupts -
55 * it's changed with a smp_call_function().
57 * SMP synchronization:
58 * constructors and destructors are called without any locking.
59 * Several members in struct kmem_cache and struct slab never change, they
60 * are accessed without any locking.
61 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
62 * and local interrupts are disabled so slab code is preempt-safe.
63 * The non-constant members are protected with a per-cache irq spinlock.
65 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
66 * in 2000 - many ideas in the current implementation are derived from
69 * Further notes from the original documentation:
71 * 11 April '97. Started multi-threading - markhe
72 * The global cache-chain is protected by the mutex 'slab_mutex'.
73 * The sem is only needed when accessing/extending the cache-chain, which
74 * can never happen inside an interrupt (kmem_cache_create(),
75 * kmem_cache_shrink() and kmem_cache_reap()).
77 * At present, each engine can be growing a cache. This should be blocked.
79 * 15 March 2005. NUMA slab allocator.
80 * Shai Fultheim <shai@scalex86.org>.
81 * Shobhit Dayal <shobhit@calsoftinc.com>
82 * Alok N Kataria <alokk@calsoftinc.com>
83 * Christoph Lameter <christoph@lameter.com>
85 * Modified the slab allocator to be node aware on NUMA systems.
86 * Each node has its own list of partial, free and full slabs.
87 * All object allocations for a node occur from node specific slab lists.
90 #include <linux/slab.h>
92 #include <linux/poison.h>
93 #include <linux/swap.h>
94 #include <linux/cache.h>
95 #include <linux/interrupt.h>
96 #include <linux/init.h>
97 #include <linux/compiler.h>
98 #include <linux/cpuset.h>
99 #include <linux/proc_fs.h>
100 #include <linux/seq_file.h>
101 #include <linux/notifier.h>
102 #include <linux/kallsyms.h>
103 #include <linux/kfence.h>
104 #include <linux/cpu.h>
105 #include <linux/sysctl.h>
106 #include <linux/module.h>
107 #include <linux/rcupdate.h>
108 #include <linux/string.h>
109 #include <linux/uaccess.h>
110 #include <linux/nodemask.h>
111 #include <linux/kmemleak.h>
112 #include <linux/mempolicy.h>
113 #include <linux/mutex.h>
114 #include <linux/fault-inject.h>
115 #include <linux/rtmutex.h>
116 #include <linux/reciprocal_div.h>
117 #include <linux/debugobjects.h>
118 #include <linux/memory.h>
119 #include <linux/prefetch.h>
120 #include <linux/sched/task_stack.h>
122 #include <net/sock.h>
124 #include <asm/cacheflush.h>
125 #include <asm/tlbflush.h>
126 #include <asm/page.h>
128 #include <trace/events/kmem.h>
130 #include "internal.h"
135 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
136 * 0 for faster, smaller code (especially in the critical paths).
138 * STATS - 1 to collect stats for /proc/slabinfo.
139 * 0 for faster, smaller code (especially in the critical paths).
141 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
144 #ifdef CONFIG_DEBUG_SLAB
147 #define FORCED_DEBUG 1
151 #define FORCED_DEBUG 0
154 /* Shouldn't this be in a header file somewhere? */
155 #define BYTES_PER_WORD sizeof(void *)
156 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
158 #ifndef ARCH_KMALLOC_FLAGS
159 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
162 #define FREELIST_BYTE_INDEX (((PAGE_SIZE >> BITS_PER_BYTE) \
163 <= SLAB_OBJ_MIN_SIZE) ? 1 : 0)
165 #if FREELIST_BYTE_INDEX
166 typedef unsigned char freelist_idx_t;
168 typedef unsigned short freelist_idx_t;
171 #define SLAB_OBJ_MAX_NUM ((1 << sizeof(freelist_idx_t) * BITS_PER_BYTE) - 1)
177 * - LIFO ordering, to hand out cache-warm objects from _alloc
178 * - reduce the number of linked list operations
179 * - reduce spinlock operations
181 * The limit is stored in the per-cpu structure to reduce the data cache
188 unsigned int batchcount;
189 unsigned int touched;
191 * Must have this definition in here for the proper
192 * alignment of array_cache. Also simplifies accessing
199 struct array_cache ac;
203 * Need this for bootstrapping a per node allocator.
205 #define NUM_INIT_LISTS (2 * MAX_NUMNODES)
206 static struct kmem_cache_node __initdata init_kmem_cache_node[NUM_INIT_LISTS];
207 #define CACHE_CACHE 0
208 #define SIZE_NODE (MAX_NUMNODES)
210 static int drain_freelist(struct kmem_cache *cache,
211 struct kmem_cache_node *n, int tofree);
212 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
213 int node, struct list_head *list);
214 static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list);
215 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
216 static void cache_reap(struct work_struct *unused);
218 static inline void fixup_objfreelist_debug(struct kmem_cache *cachep,
220 static inline void fixup_slab_list(struct kmem_cache *cachep,
221 struct kmem_cache_node *n, struct page *page,
223 static int slab_early_init = 1;
225 #define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
227 static void kmem_cache_node_init(struct kmem_cache_node *parent)
229 INIT_LIST_HEAD(&parent->slabs_full);
230 INIT_LIST_HEAD(&parent->slabs_partial);
231 INIT_LIST_HEAD(&parent->slabs_free);
232 parent->total_slabs = 0;
233 parent->free_slabs = 0;
234 parent->shared = NULL;
235 parent->alien = NULL;
236 parent->colour_next = 0;
237 spin_lock_init(&parent->list_lock);
238 parent->free_objects = 0;
239 parent->free_touched = 0;
242 #define MAKE_LIST(cachep, listp, slab, nodeid) \
244 INIT_LIST_HEAD(listp); \
245 list_splice(&get_node(cachep, nodeid)->slab, listp); \
248 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
250 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
251 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
252 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
255 #define CFLGS_OBJFREELIST_SLAB ((slab_flags_t __force)0x40000000U)
256 #define CFLGS_OFF_SLAB ((slab_flags_t __force)0x80000000U)
257 #define OBJFREELIST_SLAB(x) ((x)->flags & CFLGS_OBJFREELIST_SLAB)
258 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
260 #define BATCHREFILL_LIMIT 16
262 * Optimization question: fewer reaps means less probability for unnessary
263 * cpucache drain/refill cycles.
265 * OTOH the cpuarrays can contain lots of objects,
266 * which could lock up otherwise freeable slabs.
268 #define REAPTIMEOUT_AC (2*HZ)
269 #define REAPTIMEOUT_NODE (4*HZ)
272 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
273 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
274 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
275 #define STATS_INC_GROWN(x) ((x)->grown++)
276 #define STATS_ADD_REAPED(x, y) ((x)->reaped += (y))
277 #define STATS_SET_HIGH(x) \
279 if ((x)->num_active > (x)->high_mark) \
280 (x)->high_mark = (x)->num_active; \
282 #define STATS_INC_ERR(x) ((x)->errors++)
283 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
284 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
285 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
286 #define STATS_SET_FREEABLE(x, i) \
288 if ((x)->max_freeable < i) \
289 (x)->max_freeable = i; \
291 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
292 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
293 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
294 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
296 #define STATS_INC_ACTIVE(x) do { } while (0)
297 #define STATS_DEC_ACTIVE(x) do { } while (0)
298 #define STATS_INC_ALLOCED(x) do { } while (0)
299 #define STATS_INC_GROWN(x) do { } while (0)
300 #define STATS_ADD_REAPED(x, y) do { (void)(y); } while (0)
301 #define STATS_SET_HIGH(x) do { } while (0)
302 #define STATS_INC_ERR(x) do { } while (0)
303 #define STATS_INC_NODEALLOCS(x) do { } while (0)
304 #define STATS_INC_NODEFREES(x) do { } while (0)
305 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
306 #define STATS_SET_FREEABLE(x, i) do { } while (0)
307 #define STATS_INC_ALLOCHIT(x) do { } while (0)
308 #define STATS_INC_ALLOCMISS(x) do { } while (0)
309 #define STATS_INC_FREEHIT(x) do { } while (0)
310 #define STATS_INC_FREEMISS(x) do { } while (0)
316 * memory layout of objects:
318 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
319 * the end of an object is aligned with the end of the real
320 * allocation. Catches writes behind the end of the allocation.
321 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
323 * cachep->obj_offset: The real object.
324 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
325 * cachep->size - 1* BYTES_PER_WORD: last caller address
326 * [BYTES_PER_WORD long]
328 static int obj_offset(struct kmem_cache *cachep)
330 return cachep->obj_offset;
333 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
335 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
336 return (unsigned long long *) (objp + obj_offset(cachep) -
337 sizeof(unsigned long long));
340 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
342 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
343 if (cachep->flags & SLAB_STORE_USER)
344 return (unsigned long long *)(objp + cachep->size -
345 sizeof(unsigned long long) -
347 return (unsigned long long *) (objp + cachep->size -
348 sizeof(unsigned long long));
351 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
353 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
354 return (void **)(objp + cachep->size - BYTES_PER_WORD);
359 #define obj_offset(x) 0
360 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
361 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
362 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
367 * Do not go above this order unless 0 objects fit into the slab or
368 * overridden on the command line.
370 #define SLAB_MAX_ORDER_HI 1
371 #define SLAB_MAX_ORDER_LO 0
372 static int slab_max_order = SLAB_MAX_ORDER_LO;
373 static bool slab_max_order_set __initdata;
375 static inline void *index_to_obj(struct kmem_cache *cache, struct page *page,
378 return page->s_mem + cache->size * idx;
381 #define BOOT_CPUCACHE_ENTRIES 1
382 /* internal cache of cache description objs */
383 static struct kmem_cache kmem_cache_boot = {
385 .limit = BOOT_CPUCACHE_ENTRIES,
387 .size = sizeof(struct kmem_cache),
388 .name = "kmem_cache",
391 static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
393 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
395 return this_cpu_ptr(cachep->cpu_cache);
399 * Calculate the number of objects and left-over bytes for a given buffer size.
401 static unsigned int cache_estimate(unsigned long gfporder, size_t buffer_size,
402 slab_flags_t flags, size_t *left_over)
405 size_t slab_size = PAGE_SIZE << gfporder;
408 * The slab management structure can be either off the slab or
409 * on it. For the latter case, the memory allocated for a
412 * - @buffer_size bytes for each object
413 * - One freelist_idx_t for each object
415 * We don't need to consider alignment of freelist because
416 * freelist will be at the end of slab page. The objects will be
417 * at the correct alignment.
419 * If the slab management structure is off the slab, then the
420 * alignment will already be calculated into the size. Because
421 * the slabs are all pages aligned, the objects will be at the
422 * correct alignment when allocated.
424 if (flags & (CFLGS_OBJFREELIST_SLAB | CFLGS_OFF_SLAB)) {
425 num = slab_size / buffer_size;
426 *left_over = slab_size % buffer_size;
428 num = slab_size / (buffer_size + sizeof(freelist_idx_t));
429 *left_over = slab_size %
430 (buffer_size + sizeof(freelist_idx_t));
437 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
439 static void __slab_error(const char *function, struct kmem_cache *cachep,
442 pr_err("slab error in %s(): cache `%s': %s\n",
443 function, cachep->name, msg);
445 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
450 * By default on NUMA we use alien caches to stage the freeing of
451 * objects allocated from other nodes. This causes massive memory
452 * inefficiencies when using fake NUMA setup to split memory into a
453 * large number of small nodes, so it can be disabled on the command
457 static int use_alien_caches __read_mostly = 1;
458 static int __init noaliencache_setup(char *s)
460 use_alien_caches = 0;
463 __setup("noaliencache", noaliencache_setup);
465 static int __init slab_max_order_setup(char *str)
467 get_option(&str, &slab_max_order);
468 slab_max_order = slab_max_order < 0 ? 0 :
469 min(slab_max_order, MAX_ORDER - 1);
470 slab_max_order_set = true;
474 __setup("slab_max_order=", slab_max_order_setup);
478 * Special reaping functions for NUMA systems called from cache_reap().
479 * These take care of doing round robin flushing of alien caches (containing
480 * objects freed on different nodes from which they were allocated) and the
481 * flushing of remote pcps by calling drain_node_pages.
483 static DEFINE_PER_CPU(unsigned long, slab_reap_node);
485 static void init_reap_node(int cpu)
487 per_cpu(slab_reap_node, cpu) = next_node_in(cpu_to_mem(cpu),
491 static void next_reap_node(void)
493 int node = __this_cpu_read(slab_reap_node);
495 node = next_node_in(node, node_online_map);
496 __this_cpu_write(slab_reap_node, node);
500 #define init_reap_node(cpu) do { } while (0)
501 #define next_reap_node(void) do { } while (0)
505 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
506 * via the workqueue/eventd.
507 * Add the CPU number into the expiration time to minimize the possibility of
508 * the CPUs getting into lockstep and contending for the global cache chain
511 static void start_cpu_timer(int cpu)
513 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
515 if (reap_work->work.func == NULL) {
517 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
518 schedule_delayed_work_on(cpu, reap_work,
519 __round_jiffies_relative(HZ, cpu));
523 static void init_arraycache(struct array_cache *ac, int limit, int batch)
528 ac->batchcount = batch;
533 static struct array_cache *alloc_arraycache(int node, int entries,
534 int batchcount, gfp_t gfp)
536 size_t memsize = sizeof(void *) * entries + sizeof(struct array_cache);
537 struct array_cache *ac = NULL;
539 ac = kmalloc_node(memsize, gfp, node);
541 * The array_cache structures contain pointers to free object.
542 * However, when such objects are allocated or transferred to another
543 * cache the pointers are not cleared and they could be counted as
544 * valid references during a kmemleak scan. Therefore, kmemleak must
545 * not scan such objects.
547 kmemleak_no_scan(ac);
548 init_arraycache(ac, entries, batchcount);
552 static noinline void cache_free_pfmemalloc(struct kmem_cache *cachep,
553 struct page *page, void *objp)
555 struct kmem_cache_node *n;
559 page_node = page_to_nid(page);
560 n = get_node(cachep, page_node);
562 spin_lock(&n->list_lock);
563 free_block(cachep, &objp, 1, page_node, &list);
564 spin_unlock(&n->list_lock);
566 slabs_destroy(cachep, &list);
570 * Transfer objects in one arraycache to another.
571 * Locking must be handled by the caller.
573 * Return the number of entries transferred.
575 static int transfer_objects(struct array_cache *to,
576 struct array_cache *from, unsigned int max)
578 /* Figure out how many entries to transfer */
579 int nr = min3(from->avail, max, to->limit - to->avail);
584 memcpy(to->entry + to->avail, from->entry + from->avail - nr,
592 /* &alien->lock must be held by alien callers. */
593 static __always_inline void __free_one(struct array_cache *ac, void *objp)
595 /* Avoid trivial double-free. */
596 if (IS_ENABLED(CONFIG_SLAB_FREELIST_HARDENED) &&
597 WARN_ON_ONCE(ac->avail > 0 && ac->entry[ac->avail - 1] == objp))
599 ac->entry[ac->avail++] = objp;
604 #define drain_alien_cache(cachep, alien) do { } while (0)
605 #define reap_alien(cachep, n) do { } while (0)
607 static inline struct alien_cache **alloc_alien_cache(int node,
608 int limit, gfp_t gfp)
613 static inline void free_alien_cache(struct alien_cache **ac_ptr)
617 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
622 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
628 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
629 gfp_t flags, int nodeid)
634 static inline gfp_t gfp_exact_node(gfp_t flags)
636 return flags & ~__GFP_NOFAIL;
639 #else /* CONFIG_NUMA */
641 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
642 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
644 static struct alien_cache *__alloc_alien_cache(int node, int entries,
645 int batch, gfp_t gfp)
647 size_t memsize = sizeof(void *) * entries + sizeof(struct alien_cache);
648 struct alien_cache *alc = NULL;
650 alc = kmalloc_node(memsize, gfp, node);
652 kmemleak_no_scan(alc);
653 init_arraycache(&alc->ac, entries, batch);
654 spin_lock_init(&alc->lock);
659 static struct alien_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
661 struct alien_cache **alc_ptr;
666 alc_ptr = kcalloc_node(nr_node_ids, sizeof(void *), gfp, node);
671 if (i == node || !node_online(i))
673 alc_ptr[i] = __alloc_alien_cache(node, limit, 0xbaadf00d, gfp);
675 for (i--; i >= 0; i--)
684 static void free_alien_cache(struct alien_cache **alc_ptr)
695 static void __drain_alien_cache(struct kmem_cache *cachep,
696 struct array_cache *ac, int node,
697 struct list_head *list)
699 struct kmem_cache_node *n = get_node(cachep, node);
702 spin_lock(&n->list_lock);
704 * Stuff objects into the remote nodes shared array first.
705 * That way we could avoid the overhead of putting the objects
706 * into the free lists and getting them back later.
709 transfer_objects(n->shared, ac, ac->limit);
711 free_block(cachep, ac->entry, ac->avail, node, list);
713 spin_unlock(&n->list_lock);
718 * Called from cache_reap() to regularly drain alien caches round robin.
720 static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
722 int node = __this_cpu_read(slab_reap_node);
725 struct alien_cache *alc = n->alien[node];
726 struct array_cache *ac;
730 if (ac->avail && spin_trylock_irq(&alc->lock)) {
733 __drain_alien_cache(cachep, ac, node, &list);
734 spin_unlock_irq(&alc->lock);
735 slabs_destroy(cachep, &list);
741 static void drain_alien_cache(struct kmem_cache *cachep,
742 struct alien_cache **alien)
745 struct alien_cache *alc;
746 struct array_cache *ac;
749 for_each_online_node(i) {
755 spin_lock_irqsave(&alc->lock, flags);
756 __drain_alien_cache(cachep, ac, i, &list);
757 spin_unlock_irqrestore(&alc->lock, flags);
758 slabs_destroy(cachep, &list);
763 static int __cache_free_alien(struct kmem_cache *cachep, void *objp,
764 int node, int page_node)
766 struct kmem_cache_node *n;
767 struct alien_cache *alien = NULL;
768 struct array_cache *ac;
771 n = get_node(cachep, node);
772 STATS_INC_NODEFREES(cachep);
773 if (n->alien && n->alien[page_node]) {
774 alien = n->alien[page_node];
776 spin_lock(&alien->lock);
777 if (unlikely(ac->avail == ac->limit)) {
778 STATS_INC_ACOVERFLOW(cachep);
779 __drain_alien_cache(cachep, ac, page_node, &list);
781 __free_one(ac, objp);
782 spin_unlock(&alien->lock);
783 slabs_destroy(cachep, &list);
785 n = get_node(cachep, page_node);
786 spin_lock(&n->list_lock);
787 free_block(cachep, &objp, 1, page_node, &list);
788 spin_unlock(&n->list_lock);
789 slabs_destroy(cachep, &list);
794 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
796 int page_node = page_to_nid(virt_to_page(objp));
797 int node = numa_mem_id();
799 * Make sure we are not freeing a object from another node to the array
802 if (likely(node == page_node))
805 return __cache_free_alien(cachep, objp, node, page_node);
809 * Construct gfp mask to allocate from a specific node but do not reclaim or
810 * warn about failures.
812 static inline gfp_t gfp_exact_node(gfp_t flags)
814 return (flags | __GFP_THISNODE | __GFP_NOWARN) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
818 static int init_cache_node(struct kmem_cache *cachep, int node, gfp_t gfp)
820 struct kmem_cache_node *n;
823 * Set up the kmem_cache_node for cpu before we can
824 * begin anything. Make sure some other cpu on this
825 * node has not already allocated this
827 n = get_node(cachep, node);
829 spin_lock_irq(&n->list_lock);
830 n->free_limit = (1 + nr_cpus_node(node)) * cachep->batchcount +
832 spin_unlock_irq(&n->list_lock);
837 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
841 kmem_cache_node_init(n);
842 n->next_reap = jiffies + REAPTIMEOUT_NODE +
843 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
846 (1 + nr_cpus_node(node)) * cachep->batchcount + cachep->num;
849 * The kmem_cache_nodes don't come and go as CPUs
850 * come and go. slab_mutex is sufficient
853 cachep->node[node] = n;
858 #if (defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)) || defined(CONFIG_SMP)
860 * Allocates and initializes node for a node on each slab cache, used for
861 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
862 * will be allocated off-node since memory is not yet online for the new node.
863 * When hotplugging memory or a cpu, existing node are not replaced if
866 * Must hold slab_mutex.
868 static int init_cache_node_node(int node)
871 struct kmem_cache *cachep;
873 list_for_each_entry(cachep, &slab_caches, list) {
874 ret = init_cache_node(cachep, node, GFP_KERNEL);
883 static int setup_kmem_cache_node(struct kmem_cache *cachep,
884 int node, gfp_t gfp, bool force_change)
887 struct kmem_cache_node *n;
888 struct array_cache *old_shared = NULL;
889 struct array_cache *new_shared = NULL;
890 struct alien_cache **new_alien = NULL;
893 if (use_alien_caches) {
894 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
899 if (cachep->shared) {
900 new_shared = alloc_arraycache(node,
901 cachep->shared * cachep->batchcount, 0xbaadf00d, gfp);
906 ret = init_cache_node(cachep, node, gfp);
910 n = get_node(cachep, node);
911 spin_lock_irq(&n->list_lock);
912 if (n->shared && force_change) {
913 free_block(cachep, n->shared->entry,
914 n->shared->avail, node, &list);
915 n->shared->avail = 0;
918 if (!n->shared || force_change) {
919 old_shared = n->shared;
920 n->shared = new_shared;
925 n->alien = new_alien;
929 spin_unlock_irq(&n->list_lock);
930 slabs_destroy(cachep, &list);
933 * To protect lockless access to n->shared during irq disabled context.
934 * If n->shared isn't NULL in irq disabled context, accessing to it is
935 * guaranteed to be valid until irq is re-enabled, because it will be
936 * freed after synchronize_rcu().
938 if (old_shared && force_change)
944 free_alien_cache(new_alien);
951 static void cpuup_canceled(long cpu)
953 struct kmem_cache *cachep;
954 struct kmem_cache_node *n = NULL;
955 int node = cpu_to_mem(cpu);
956 const struct cpumask *mask = cpumask_of_node(node);
958 list_for_each_entry(cachep, &slab_caches, list) {
959 struct array_cache *nc;
960 struct array_cache *shared;
961 struct alien_cache **alien;
964 n = get_node(cachep, node);
968 spin_lock_irq(&n->list_lock);
970 /* Free limit for this kmem_cache_node */
971 n->free_limit -= cachep->batchcount;
973 /* cpu is dead; no one can alloc from it. */
974 nc = per_cpu_ptr(cachep->cpu_cache, cpu);
975 free_block(cachep, nc->entry, nc->avail, node, &list);
978 if (!cpumask_empty(mask)) {
979 spin_unlock_irq(&n->list_lock);
985 free_block(cachep, shared->entry,
986 shared->avail, node, &list);
993 spin_unlock_irq(&n->list_lock);
997 drain_alien_cache(cachep, alien);
998 free_alien_cache(alien);
1002 slabs_destroy(cachep, &list);
1005 * In the previous loop, all the objects were freed to
1006 * the respective cache's slabs, now we can go ahead and
1007 * shrink each nodelist to its limit.
1009 list_for_each_entry(cachep, &slab_caches, list) {
1010 n = get_node(cachep, node);
1013 drain_freelist(cachep, n, INT_MAX);
1017 static int cpuup_prepare(long cpu)
1019 struct kmem_cache *cachep;
1020 int node = cpu_to_mem(cpu);
1024 * We need to do this right in the beginning since
1025 * alloc_arraycache's are going to use this list.
1026 * kmalloc_node allows us to add the slab to the right
1027 * kmem_cache_node and not this cpu's kmem_cache_node
1029 err = init_cache_node_node(node);
1034 * Now we can go ahead with allocating the shared arrays and
1037 list_for_each_entry(cachep, &slab_caches, list) {
1038 err = setup_kmem_cache_node(cachep, node, GFP_KERNEL, false);
1045 cpuup_canceled(cpu);
1049 int slab_prepare_cpu(unsigned int cpu)
1053 mutex_lock(&slab_mutex);
1054 err = cpuup_prepare(cpu);
1055 mutex_unlock(&slab_mutex);
1060 * This is called for a failed online attempt and for a successful
1063 * Even if all the cpus of a node are down, we don't free the
1064 * kmem_cache_node of any cache. This to avoid a race between cpu_down, and
1065 * a kmalloc allocation from another cpu for memory from the node of
1066 * the cpu going down. The kmem_cache_node structure is usually allocated from
1067 * kmem_cache_create() and gets destroyed at kmem_cache_destroy().
1069 int slab_dead_cpu(unsigned int cpu)
1071 mutex_lock(&slab_mutex);
1072 cpuup_canceled(cpu);
1073 mutex_unlock(&slab_mutex);
1078 static int slab_online_cpu(unsigned int cpu)
1080 start_cpu_timer(cpu);
1084 static int slab_offline_cpu(unsigned int cpu)
1087 * Shutdown cache reaper. Note that the slab_mutex is held so
1088 * that if cache_reap() is invoked it cannot do anything
1089 * expensive but will only modify reap_work and reschedule the
1092 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1093 /* Now the cache_reaper is guaranteed to be not running. */
1094 per_cpu(slab_reap_work, cpu).work.func = NULL;
1098 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1100 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1101 * Returns -EBUSY if all objects cannot be drained so that the node is not
1104 * Must hold slab_mutex.
1106 static int __meminit drain_cache_node_node(int node)
1108 struct kmem_cache *cachep;
1111 list_for_each_entry(cachep, &slab_caches, list) {
1112 struct kmem_cache_node *n;
1114 n = get_node(cachep, node);
1118 drain_freelist(cachep, n, INT_MAX);
1120 if (!list_empty(&n->slabs_full) ||
1121 !list_empty(&n->slabs_partial)) {
1129 static int __meminit slab_memory_callback(struct notifier_block *self,
1130 unsigned long action, void *arg)
1132 struct memory_notify *mnb = arg;
1136 nid = mnb->status_change_nid;
1141 case MEM_GOING_ONLINE:
1142 mutex_lock(&slab_mutex);
1143 ret = init_cache_node_node(nid);
1144 mutex_unlock(&slab_mutex);
1146 case MEM_GOING_OFFLINE:
1147 mutex_lock(&slab_mutex);
1148 ret = drain_cache_node_node(nid);
1149 mutex_unlock(&slab_mutex);
1153 case MEM_CANCEL_ONLINE:
1154 case MEM_CANCEL_OFFLINE:
1158 return notifier_from_errno(ret);
1160 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1163 * swap the static kmem_cache_node with kmalloced memory
1165 static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
1168 struct kmem_cache_node *ptr;
1170 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
1173 memcpy(ptr, list, sizeof(struct kmem_cache_node));
1175 * Do not assume that spinlocks can be initialized via memcpy:
1177 spin_lock_init(&ptr->list_lock);
1179 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1180 cachep->node[nodeid] = ptr;
1184 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1185 * size of kmem_cache_node.
1187 static void __init set_up_node(struct kmem_cache *cachep, int index)
1191 for_each_online_node(node) {
1192 cachep->node[node] = &init_kmem_cache_node[index + node];
1193 cachep->node[node]->next_reap = jiffies +
1195 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1200 * Initialisation. Called after the page allocator have been initialised and
1201 * before smp_init().
1203 void __init kmem_cache_init(void)
1207 kmem_cache = &kmem_cache_boot;
1209 if (!IS_ENABLED(CONFIG_NUMA) || num_possible_nodes() == 1)
1210 use_alien_caches = 0;
1212 for (i = 0; i < NUM_INIT_LISTS; i++)
1213 kmem_cache_node_init(&init_kmem_cache_node[i]);
1216 * Fragmentation resistance on low memory - only use bigger
1217 * page orders on machines with more than 32MB of memory if
1218 * not overridden on the command line.
1220 if (!slab_max_order_set && totalram_pages() > (32 << 20) >> PAGE_SHIFT)
1221 slab_max_order = SLAB_MAX_ORDER_HI;
1223 /* Bootstrap is tricky, because several objects are allocated
1224 * from caches that do not exist yet:
1225 * 1) initialize the kmem_cache cache: it contains the struct
1226 * kmem_cache structures of all caches, except kmem_cache itself:
1227 * kmem_cache is statically allocated.
1228 * Initially an __init data area is used for the head array and the
1229 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1230 * array at the end of the bootstrap.
1231 * 2) Create the first kmalloc cache.
1232 * The struct kmem_cache for the new cache is allocated normally.
1233 * An __init data area is used for the head array.
1234 * 3) Create the remaining kmalloc caches, with minimally sized
1236 * 4) Replace the __init data head arrays for kmem_cache and the first
1237 * kmalloc cache with kmalloc allocated arrays.
1238 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1239 * the other cache's with kmalloc allocated memory.
1240 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1243 /* 1) create the kmem_cache */
1246 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1248 create_boot_cache(kmem_cache, "kmem_cache",
1249 offsetof(struct kmem_cache, node) +
1250 nr_node_ids * sizeof(struct kmem_cache_node *),
1251 SLAB_HWCACHE_ALIGN, 0, 0);
1252 list_add(&kmem_cache->list, &slab_caches);
1253 slab_state = PARTIAL;
1256 * Initialize the caches that provide memory for the kmem_cache_node
1257 * structures first. Without this, further allocations will bug.
1259 kmalloc_caches[KMALLOC_NORMAL][INDEX_NODE] = create_kmalloc_cache(
1260 kmalloc_info[INDEX_NODE].name[KMALLOC_NORMAL],
1261 kmalloc_info[INDEX_NODE].size,
1262 ARCH_KMALLOC_FLAGS, 0,
1263 kmalloc_info[INDEX_NODE].size);
1264 slab_state = PARTIAL_NODE;
1265 setup_kmalloc_cache_index_table();
1267 slab_early_init = 0;
1269 /* 5) Replace the bootstrap kmem_cache_node */
1273 for_each_online_node(nid) {
1274 init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
1276 init_list(kmalloc_caches[KMALLOC_NORMAL][INDEX_NODE],
1277 &init_kmem_cache_node[SIZE_NODE + nid], nid);
1281 create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
1284 void __init kmem_cache_init_late(void)
1286 struct kmem_cache *cachep;
1288 /* 6) resize the head arrays to their final sizes */
1289 mutex_lock(&slab_mutex);
1290 list_for_each_entry(cachep, &slab_caches, list)
1291 if (enable_cpucache(cachep, GFP_NOWAIT))
1293 mutex_unlock(&slab_mutex);
1300 * Register a memory hotplug callback that initializes and frees
1303 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1307 * The reap timers are started later, with a module init call: That part
1308 * of the kernel is not yet operational.
1312 static int __init cpucache_init(void)
1317 * Register the timers that return unneeded pages to the page allocator
1319 ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "SLAB online",
1320 slab_online_cpu, slab_offline_cpu);
1325 __initcall(cpucache_init);
1327 static noinline void
1328 slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1331 struct kmem_cache_node *n;
1332 unsigned long flags;
1334 static DEFINE_RATELIMIT_STATE(slab_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1335 DEFAULT_RATELIMIT_BURST);
1337 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slab_oom_rs))
1340 pr_warn("SLAB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
1341 nodeid, gfpflags, &gfpflags);
1342 pr_warn(" cache: %s, object size: %d, order: %d\n",
1343 cachep->name, cachep->size, cachep->gfporder);
1345 for_each_kmem_cache_node(cachep, node, n) {
1346 unsigned long total_slabs, free_slabs, free_objs;
1348 spin_lock_irqsave(&n->list_lock, flags);
1349 total_slabs = n->total_slabs;
1350 free_slabs = n->free_slabs;
1351 free_objs = n->free_objects;
1352 spin_unlock_irqrestore(&n->list_lock, flags);
1354 pr_warn(" node %d: slabs: %ld/%ld, objs: %ld/%ld\n",
1355 node, total_slabs - free_slabs, total_slabs,
1356 (total_slabs * cachep->num) - free_objs,
1357 total_slabs * cachep->num);
1363 * Interface to system's page allocator. No need to hold the
1364 * kmem_cache_node ->list_lock.
1366 * If we requested dmaable memory, we will get it. Even if we
1367 * did not request dmaable memory, we might get it, but that
1368 * would be relatively rare and ignorable.
1370 static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1375 flags |= cachep->allocflags;
1377 page = __alloc_pages_node(nodeid, flags, cachep->gfporder);
1379 slab_out_of_memory(cachep, flags, nodeid);
1383 account_slab_page(page, cachep->gfporder, cachep, flags);
1384 __SetPageSlab(page);
1385 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1386 if (sk_memalloc_socks() && page_is_pfmemalloc(page))
1387 SetPageSlabPfmemalloc(page);
1393 * Interface to system's page release.
1395 static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
1397 int order = cachep->gfporder;
1399 BUG_ON(!PageSlab(page));
1400 __ClearPageSlabPfmemalloc(page);
1401 __ClearPageSlab(page);
1402 page_mapcount_reset(page);
1403 /* In union with page->mapping where page allocator expects NULL */
1404 page->slab_cache = NULL;
1406 if (current->reclaim_state)
1407 current->reclaim_state->reclaimed_slab += 1 << order;
1408 unaccount_slab_page(page, order, cachep);
1409 __free_pages(page, order);
1412 static void kmem_rcu_free(struct rcu_head *head)
1414 struct kmem_cache *cachep;
1417 page = container_of(head, struct page, rcu_head);
1418 cachep = page->slab_cache;
1420 kmem_freepages(cachep, page);
1424 static bool is_debug_pagealloc_cache(struct kmem_cache *cachep)
1426 if (debug_pagealloc_enabled_static() && OFF_SLAB(cachep) &&
1427 (cachep->size % PAGE_SIZE) == 0)
1433 #ifdef CONFIG_DEBUG_PAGEALLOC
1434 static void slab_kernel_map(struct kmem_cache *cachep, void *objp, int map)
1436 if (!is_debug_pagealloc_cache(cachep))
1439 __kernel_map_pages(virt_to_page(objp), cachep->size / PAGE_SIZE, map);
1443 static inline void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1448 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1450 int size = cachep->object_size;
1451 addr = &((char *)addr)[obj_offset(cachep)];
1453 memset(addr, val, size);
1454 *(unsigned char *)(addr + size - 1) = POISON_END;
1457 static void dump_line(char *data, int offset, int limit)
1460 unsigned char error = 0;
1463 pr_err("%03x: ", offset);
1464 for (i = 0; i < limit; i++) {
1465 if (data[offset + i] != POISON_FREE) {
1466 error = data[offset + i];
1470 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1471 &data[offset], limit, 1);
1473 if (bad_count == 1) {
1474 error ^= POISON_FREE;
1475 if (!(error & (error - 1))) {
1476 pr_err("Single bit error detected. Probably bad RAM.\n");
1478 pr_err("Run memtest86+ or a similar memory test tool.\n");
1480 pr_err("Run a memory test tool.\n");
1489 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1494 if (cachep->flags & SLAB_RED_ZONE) {
1495 pr_err("Redzone: 0x%llx/0x%llx\n",
1496 *dbg_redzone1(cachep, objp),
1497 *dbg_redzone2(cachep, objp));
1500 if (cachep->flags & SLAB_STORE_USER)
1501 pr_err("Last user: (%pSR)\n", *dbg_userword(cachep, objp));
1502 realobj = (char *)objp + obj_offset(cachep);
1503 size = cachep->object_size;
1504 for (i = 0; i < size && lines; i += 16, lines--) {
1507 if (i + limit > size)
1509 dump_line(realobj, i, limit);
1513 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1519 if (is_debug_pagealloc_cache(cachep))
1522 realobj = (char *)objp + obj_offset(cachep);
1523 size = cachep->object_size;
1525 for (i = 0; i < size; i++) {
1526 char exp = POISON_FREE;
1529 if (realobj[i] != exp) {
1534 pr_err("Slab corruption (%s): %s start=%px, len=%d\n",
1535 print_tainted(), cachep->name,
1537 print_objinfo(cachep, objp, 0);
1539 /* Hexdump the affected line */
1542 if (i + limit > size)
1544 dump_line(realobj, i, limit);
1547 /* Limit to 5 lines */
1553 /* Print some data about the neighboring objects, if they
1556 struct page *page = virt_to_head_page(objp);
1559 objnr = obj_to_index(cachep, page, objp);
1561 objp = index_to_obj(cachep, page, objnr - 1);
1562 realobj = (char *)objp + obj_offset(cachep);
1563 pr_err("Prev obj: start=%px, len=%d\n", realobj, size);
1564 print_objinfo(cachep, objp, 2);
1566 if (objnr + 1 < cachep->num) {
1567 objp = index_to_obj(cachep, page, objnr + 1);
1568 realobj = (char *)objp + obj_offset(cachep);
1569 pr_err("Next obj: start=%px, len=%d\n", realobj, size);
1570 print_objinfo(cachep, objp, 2);
1577 static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1582 if (OBJFREELIST_SLAB(cachep) && cachep->flags & SLAB_POISON) {
1583 poison_obj(cachep, page->freelist - obj_offset(cachep),
1587 for (i = 0; i < cachep->num; i++) {
1588 void *objp = index_to_obj(cachep, page, i);
1590 if (cachep->flags & SLAB_POISON) {
1591 check_poison_obj(cachep, objp);
1592 slab_kernel_map(cachep, objp, 1);
1594 if (cachep->flags & SLAB_RED_ZONE) {
1595 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1596 slab_error(cachep, "start of a freed object was overwritten");
1597 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1598 slab_error(cachep, "end of a freed object was overwritten");
1603 static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1610 * slab_destroy - destroy and release all objects in a slab
1611 * @cachep: cache pointer being destroyed
1612 * @page: page pointer being destroyed
1614 * Destroy all the objs in a slab page, and release the mem back to the system.
1615 * Before calling the slab page must have been unlinked from the cache. The
1616 * kmem_cache_node ->list_lock is not held/needed.
1618 static void slab_destroy(struct kmem_cache *cachep, struct page *page)
1622 freelist = page->freelist;
1623 slab_destroy_debugcheck(cachep, page);
1624 if (unlikely(cachep->flags & SLAB_TYPESAFE_BY_RCU))
1625 call_rcu(&page->rcu_head, kmem_rcu_free);
1627 kmem_freepages(cachep, page);
1630 * From now on, we don't use freelist
1631 * although actual page can be freed in rcu context
1633 if (OFF_SLAB(cachep))
1634 kmem_cache_free(cachep->freelist_cache, freelist);
1638 * Update the size of the caches before calling slabs_destroy as it may
1639 * recursively call kfree.
1641 static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list)
1643 struct page *page, *n;
1645 list_for_each_entry_safe(page, n, list, slab_list) {
1646 list_del(&page->slab_list);
1647 slab_destroy(cachep, page);
1652 * calculate_slab_order - calculate size (page order) of slabs
1653 * @cachep: pointer to the cache that is being created
1654 * @size: size of objects to be created in this cache.
1655 * @flags: slab allocation flags
1657 * Also calculates the number of objects per slab.
1659 * This could be made much more intelligent. For now, try to avoid using
1660 * high order pages for slabs. When the gfp() functions are more friendly
1661 * towards high-order requests, this should be changed.
1663 * Return: number of left-over bytes in a slab
1665 static size_t calculate_slab_order(struct kmem_cache *cachep,
1666 size_t size, slab_flags_t flags)
1668 size_t left_over = 0;
1671 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
1675 num = cache_estimate(gfporder, size, flags, &remainder);
1679 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1680 if (num > SLAB_OBJ_MAX_NUM)
1683 if (flags & CFLGS_OFF_SLAB) {
1684 struct kmem_cache *freelist_cache;
1685 size_t freelist_size;
1687 freelist_size = num * sizeof(freelist_idx_t);
1688 freelist_cache = kmalloc_slab(freelist_size, 0u);
1689 if (!freelist_cache)
1693 * Needed to avoid possible looping condition
1694 * in cache_grow_begin()
1696 if (OFF_SLAB(freelist_cache))
1699 /* check if off slab has enough benefit */
1700 if (freelist_cache->size > cachep->size / 2)
1704 /* Found something acceptable - save it away */
1706 cachep->gfporder = gfporder;
1707 left_over = remainder;
1710 * A VFS-reclaimable slab tends to have most allocations
1711 * as GFP_NOFS and we really don't want to have to be allocating
1712 * higher-order pages when we are unable to shrink dcache.
1714 if (flags & SLAB_RECLAIM_ACCOUNT)
1718 * Large number of objects is good, but very large slabs are
1719 * currently bad for the gfp()s.
1721 if (gfporder >= slab_max_order)
1725 * Acceptable internal fragmentation?
1727 if (left_over * 8 <= (PAGE_SIZE << gfporder))
1733 static struct array_cache __percpu *alloc_kmem_cache_cpus(
1734 struct kmem_cache *cachep, int entries, int batchcount)
1738 struct array_cache __percpu *cpu_cache;
1740 size = sizeof(void *) * entries + sizeof(struct array_cache);
1741 cpu_cache = __alloc_percpu(size, sizeof(void *));
1746 for_each_possible_cpu(cpu) {
1747 init_arraycache(per_cpu_ptr(cpu_cache, cpu),
1748 entries, batchcount);
1754 static int __ref setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
1756 if (slab_state >= FULL)
1757 return enable_cpucache(cachep, gfp);
1759 cachep->cpu_cache = alloc_kmem_cache_cpus(cachep, 1, 1);
1760 if (!cachep->cpu_cache)
1763 if (slab_state == DOWN) {
1764 /* Creation of first cache (kmem_cache). */
1765 set_up_node(kmem_cache, CACHE_CACHE);
1766 } else if (slab_state == PARTIAL) {
1767 /* For kmem_cache_node */
1768 set_up_node(cachep, SIZE_NODE);
1772 for_each_online_node(node) {
1773 cachep->node[node] = kmalloc_node(
1774 sizeof(struct kmem_cache_node), gfp, node);
1775 BUG_ON(!cachep->node[node]);
1776 kmem_cache_node_init(cachep->node[node]);
1780 cachep->node[numa_mem_id()]->next_reap =
1781 jiffies + REAPTIMEOUT_NODE +
1782 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1784 cpu_cache_get(cachep)->avail = 0;
1785 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1786 cpu_cache_get(cachep)->batchcount = 1;
1787 cpu_cache_get(cachep)->touched = 0;
1788 cachep->batchcount = 1;
1789 cachep->limit = BOOT_CPUCACHE_ENTRIES;
1793 slab_flags_t kmem_cache_flags(unsigned int object_size,
1794 slab_flags_t flags, const char *name)
1800 __kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
1801 slab_flags_t flags, void (*ctor)(void *))
1803 struct kmem_cache *cachep;
1805 cachep = find_mergeable(size, align, flags, name, ctor);
1810 * Adjust the object sizes so that we clear
1811 * the complete object on kzalloc.
1813 cachep->object_size = max_t(int, cachep->object_size, size);
1818 static bool set_objfreelist_slab_cache(struct kmem_cache *cachep,
1819 size_t size, slab_flags_t flags)
1826 * If slab auto-initialization on free is enabled, store the freelist
1827 * off-slab, so that its contents don't end up in one of the allocated
1830 if (unlikely(slab_want_init_on_free(cachep)))
1833 if (cachep->ctor || flags & SLAB_TYPESAFE_BY_RCU)
1836 left = calculate_slab_order(cachep, size,
1837 flags | CFLGS_OBJFREELIST_SLAB);
1841 if (cachep->num * sizeof(freelist_idx_t) > cachep->object_size)
1844 cachep->colour = left / cachep->colour_off;
1849 static bool set_off_slab_cache(struct kmem_cache *cachep,
1850 size_t size, slab_flags_t flags)
1857 * Always use on-slab management when SLAB_NOLEAKTRACE
1858 * to avoid recursive calls into kmemleak.
1860 if (flags & SLAB_NOLEAKTRACE)
1864 * Size is large, assume best to place the slab management obj
1865 * off-slab (should allow better packing of objs).
1867 left = calculate_slab_order(cachep, size, flags | CFLGS_OFF_SLAB);
1872 * If the slab has been placed off-slab, and we have enough space then
1873 * move it on-slab. This is at the expense of any extra colouring.
1875 if (left >= cachep->num * sizeof(freelist_idx_t))
1878 cachep->colour = left / cachep->colour_off;
1883 static bool set_on_slab_cache(struct kmem_cache *cachep,
1884 size_t size, slab_flags_t flags)
1890 left = calculate_slab_order(cachep, size, flags);
1894 cachep->colour = left / cachep->colour_off;
1900 * __kmem_cache_create - Create a cache.
1901 * @cachep: cache management descriptor
1902 * @flags: SLAB flags
1904 * Returns a ptr to the cache on success, NULL on failure.
1905 * Cannot be called within a int, but can be interrupted.
1906 * The @ctor is run when new pages are allocated by the cache.
1910 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1911 * to catch references to uninitialised memory.
1913 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1914 * for buffer overruns.
1916 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1917 * cacheline. This can be beneficial if you're counting cycles as closely
1920 * Return: a pointer to the created cache or %NULL in case of error
1922 int __kmem_cache_create(struct kmem_cache *cachep, slab_flags_t flags)
1924 size_t ralign = BYTES_PER_WORD;
1927 unsigned int size = cachep->size;
1932 * Enable redzoning and last user accounting, except for caches with
1933 * large objects, if the increased size would increase the object size
1934 * above the next power of two: caches with object sizes just above a
1935 * power of two have a significant amount of internal fragmentation.
1937 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
1938 2 * sizeof(unsigned long long)))
1939 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
1940 if (!(flags & SLAB_TYPESAFE_BY_RCU))
1941 flags |= SLAB_POISON;
1946 * Check that size is in terms of words. This is needed to avoid
1947 * unaligned accesses for some archs when redzoning is used, and makes
1948 * sure any on-slab bufctl's are also correctly aligned.
1950 size = ALIGN(size, BYTES_PER_WORD);
1952 if (flags & SLAB_RED_ZONE) {
1953 ralign = REDZONE_ALIGN;
1954 /* If redzoning, ensure that the second redzone is suitably
1955 * aligned, by adjusting the object size accordingly. */
1956 size = ALIGN(size, REDZONE_ALIGN);
1959 /* 3) caller mandated alignment */
1960 if (ralign < cachep->align) {
1961 ralign = cachep->align;
1963 /* disable debug if necessary */
1964 if (ralign > __alignof__(unsigned long long))
1965 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
1969 cachep->align = ralign;
1970 cachep->colour_off = cache_line_size();
1971 /* Offset must be a multiple of the alignment. */
1972 if (cachep->colour_off < cachep->align)
1973 cachep->colour_off = cachep->align;
1975 if (slab_is_available())
1983 * Both debugging options require word-alignment which is calculated
1986 if (flags & SLAB_RED_ZONE) {
1987 /* add space for red zone words */
1988 cachep->obj_offset += sizeof(unsigned long long);
1989 size += 2 * sizeof(unsigned long long);
1991 if (flags & SLAB_STORE_USER) {
1992 /* user store requires one word storage behind the end of
1993 * the real object. But if the second red zone needs to be
1994 * aligned to 64 bits, we must allow that much space.
1996 if (flags & SLAB_RED_ZONE)
1997 size += REDZONE_ALIGN;
1999 size += BYTES_PER_WORD;
2003 kasan_cache_create(cachep, &size, &flags);
2005 size = ALIGN(size, cachep->align);
2007 * We should restrict the number of objects in a slab to implement
2008 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2010 if (FREELIST_BYTE_INDEX && size < SLAB_OBJ_MIN_SIZE)
2011 size = ALIGN(SLAB_OBJ_MIN_SIZE, cachep->align);
2015 * To activate debug pagealloc, off-slab management is necessary
2016 * requirement. In early phase of initialization, small sized slab
2017 * doesn't get initialized so it would not be possible. So, we need
2018 * to check size >= 256. It guarantees that all necessary small
2019 * sized slab is initialized in current slab initialization sequence.
2021 if (debug_pagealloc_enabled_static() && (flags & SLAB_POISON) &&
2022 size >= 256 && cachep->object_size > cache_line_size()) {
2023 if (size < PAGE_SIZE || size % PAGE_SIZE == 0) {
2024 size_t tmp_size = ALIGN(size, PAGE_SIZE);
2026 if (set_off_slab_cache(cachep, tmp_size, flags)) {
2027 flags |= CFLGS_OFF_SLAB;
2028 cachep->obj_offset += tmp_size - size;
2036 if (set_objfreelist_slab_cache(cachep, size, flags)) {
2037 flags |= CFLGS_OBJFREELIST_SLAB;
2041 if (set_off_slab_cache(cachep, size, flags)) {
2042 flags |= CFLGS_OFF_SLAB;
2046 if (set_on_slab_cache(cachep, size, flags))
2052 cachep->freelist_size = cachep->num * sizeof(freelist_idx_t);
2053 cachep->flags = flags;
2054 cachep->allocflags = __GFP_COMP;
2055 if (flags & SLAB_CACHE_DMA)
2056 cachep->allocflags |= GFP_DMA;
2057 if (flags & SLAB_CACHE_DMA32)
2058 cachep->allocflags |= GFP_DMA32;
2059 if (flags & SLAB_RECLAIM_ACCOUNT)
2060 cachep->allocflags |= __GFP_RECLAIMABLE;
2061 cachep->size = size;
2062 cachep->reciprocal_buffer_size = reciprocal_value(size);
2066 * If we're going to use the generic kernel_map_pages()
2067 * poisoning, then it's going to smash the contents of
2068 * the redzone and userword anyhow, so switch them off.
2070 if (IS_ENABLED(CONFIG_PAGE_POISONING) &&
2071 (cachep->flags & SLAB_POISON) &&
2072 is_debug_pagealloc_cache(cachep))
2073 cachep->flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2076 if (OFF_SLAB(cachep)) {
2077 cachep->freelist_cache =
2078 kmalloc_slab(cachep->freelist_size, 0u);
2081 err = setup_cpu_cache(cachep, gfp);
2083 __kmem_cache_release(cachep);
2091 static void check_irq_off(void)
2093 BUG_ON(!irqs_disabled());
2096 static void check_irq_on(void)
2098 BUG_ON(irqs_disabled());
2101 static void check_mutex_acquired(void)
2103 BUG_ON(!mutex_is_locked(&slab_mutex));
2106 static void check_spinlock_acquired(struct kmem_cache *cachep)
2110 assert_spin_locked(&get_node(cachep, numa_mem_id())->list_lock);
2114 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2118 assert_spin_locked(&get_node(cachep, node)->list_lock);
2123 #define check_irq_off() do { } while(0)
2124 #define check_irq_on() do { } while(0)
2125 #define check_mutex_acquired() do { } while(0)
2126 #define check_spinlock_acquired(x) do { } while(0)
2127 #define check_spinlock_acquired_node(x, y) do { } while(0)
2130 static void drain_array_locked(struct kmem_cache *cachep, struct array_cache *ac,
2131 int node, bool free_all, struct list_head *list)
2135 if (!ac || !ac->avail)
2138 tofree = free_all ? ac->avail : (ac->limit + 4) / 5;
2139 if (tofree > ac->avail)
2140 tofree = (ac->avail + 1) / 2;
2142 free_block(cachep, ac->entry, tofree, node, list);
2143 ac->avail -= tofree;
2144 memmove(ac->entry, &(ac->entry[tofree]), sizeof(void *) * ac->avail);
2147 static void do_drain(void *arg)
2149 struct kmem_cache *cachep = arg;
2150 struct array_cache *ac;
2151 int node = numa_mem_id();
2152 struct kmem_cache_node *n;
2156 ac = cpu_cache_get(cachep);
2157 n = get_node(cachep, node);
2158 spin_lock(&n->list_lock);
2159 free_block(cachep, ac->entry, ac->avail, node, &list);
2160 spin_unlock(&n->list_lock);
2162 slabs_destroy(cachep, &list);
2165 static void drain_cpu_caches(struct kmem_cache *cachep)
2167 struct kmem_cache_node *n;
2171 on_each_cpu(do_drain, cachep, 1);
2173 for_each_kmem_cache_node(cachep, node, n)
2175 drain_alien_cache(cachep, n->alien);
2177 for_each_kmem_cache_node(cachep, node, n) {
2178 spin_lock_irq(&n->list_lock);
2179 drain_array_locked(cachep, n->shared, node, true, &list);
2180 spin_unlock_irq(&n->list_lock);
2182 slabs_destroy(cachep, &list);
2187 * Remove slabs from the list of free slabs.
2188 * Specify the number of slabs to drain in tofree.
2190 * Returns the actual number of slabs released.
2192 static int drain_freelist(struct kmem_cache *cache,
2193 struct kmem_cache_node *n, int tofree)
2195 struct list_head *p;
2200 while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
2202 spin_lock_irq(&n->list_lock);
2203 p = n->slabs_free.prev;
2204 if (p == &n->slabs_free) {
2205 spin_unlock_irq(&n->list_lock);
2209 page = list_entry(p, struct page, slab_list);
2210 list_del(&page->slab_list);
2214 * Safe to drop the lock. The slab is no longer linked
2217 n->free_objects -= cache->num;
2218 spin_unlock_irq(&n->list_lock);
2219 slab_destroy(cache, page);
2226 bool __kmem_cache_empty(struct kmem_cache *s)
2229 struct kmem_cache_node *n;
2231 for_each_kmem_cache_node(s, node, n)
2232 if (!list_empty(&n->slabs_full) ||
2233 !list_empty(&n->slabs_partial))
2238 int __kmem_cache_shrink(struct kmem_cache *cachep)
2242 struct kmem_cache_node *n;
2244 drain_cpu_caches(cachep);
2247 for_each_kmem_cache_node(cachep, node, n) {
2248 drain_freelist(cachep, n, INT_MAX);
2250 ret += !list_empty(&n->slabs_full) ||
2251 !list_empty(&n->slabs_partial);
2253 return (ret ? 1 : 0);
2256 int __kmem_cache_shutdown(struct kmem_cache *cachep)
2258 return __kmem_cache_shrink(cachep);
2261 void __kmem_cache_release(struct kmem_cache *cachep)
2264 struct kmem_cache_node *n;
2266 cache_random_seq_destroy(cachep);
2268 free_percpu(cachep->cpu_cache);
2270 /* NUMA: free the node structures */
2271 for_each_kmem_cache_node(cachep, i, n) {
2273 free_alien_cache(n->alien);
2275 cachep->node[i] = NULL;
2280 * Get the memory for a slab management obj.
2282 * For a slab cache when the slab descriptor is off-slab, the
2283 * slab descriptor can't come from the same cache which is being created,
2284 * Because if it is the case, that means we defer the creation of
2285 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2286 * And we eventually call down to __kmem_cache_create(), which
2287 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2288 * This is a "chicken-and-egg" problem.
2290 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2291 * which are all initialized during kmem_cache_init().
2293 static void *alloc_slabmgmt(struct kmem_cache *cachep,
2294 struct page *page, int colour_off,
2295 gfp_t local_flags, int nodeid)
2298 void *addr = page_address(page);
2300 page->s_mem = addr + colour_off;
2303 if (OBJFREELIST_SLAB(cachep))
2305 else if (OFF_SLAB(cachep)) {
2306 /* Slab management obj is off-slab. */
2307 freelist = kmem_cache_alloc_node(cachep->freelist_cache,
2308 local_flags, nodeid);
2310 /* We will use last bytes at the slab for freelist */
2311 freelist = addr + (PAGE_SIZE << cachep->gfporder) -
2312 cachep->freelist_size;
2318 static inline freelist_idx_t get_free_obj(struct page *page, unsigned int idx)
2320 return ((freelist_idx_t *)page->freelist)[idx];
2323 static inline void set_free_obj(struct page *page,
2324 unsigned int idx, freelist_idx_t val)
2326 ((freelist_idx_t *)(page->freelist))[idx] = val;
2329 static void cache_init_objs_debug(struct kmem_cache *cachep, struct page *page)
2334 for (i = 0; i < cachep->num; i++) {
2335 void *objp = index_to_obj(cachep, page, i);
2337 if (cachep->flags & SLAB_STORE_USER)
2338 *dbg_userword(cachep, objp) = NULL;
2340 if (cachep->flags & SLAB_RED_ZONE) {
2341 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2342 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2345 * Constructors are not allowed to allocate memory from the same
2346 * cache which they are a constructor for. Otherwise, deadlock.
2347 * They must also be threaded.
2349 if (cachep->ctor && !(cachep->flags & SLAB_POISON)) {
2350 kasan_unpoison_object_data(cachep,
2351 objp + obj_offset(cachep));
2352 cachep->ctor(objp + obj_offset(cachep));
2353 kasan_poison_object_data(
2354 cachep, objp + obj_offset(cachep));
2357 if (cachep->flags & SLAB_RED_ZONE) {
2358 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2359 slab_error(cachep, "constructor overwrote the end of an object");
2360 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2361 slab_error(cachep, "constructor overwrote the start of an object");
2363 /* need to poison the objs? */
2364 if (cachep->flags & SLAB_POISON) {
2365 poison_obj(cachep, objp, POISON_FREE);
2366 slab_kernel_map(cachep, objp, 0);
2372 #ifdef CONFIG_SLAB_FREELIST_RANDOM
2373 /* Hold information during a freelist initialization */
2374 union freelist_init_state {
2380 struct rnd_state rnd_state;
2384 * Initialize the state based on the randomization methode available.
2385 * return true if the pre-computed list is available, false otherwize.
2387 static bool freelist_state_initialize(union freelist_init_state *state,
2388 struct kmem_cache *cachep,
2394 /* Use best entropy available to define a random shift */
2395 rand = get_random_int();
2397 /* Use a random state if the pre-computed list is not available */
2398 if (!cachep->random_seq) {
2399 prandom_seed_state(&state->rnd_state, rand);
2402 state->list = cachep->random_seq;
2403 state->count = count;
2404 state->pos = rand % count;
2410 /* Get the next entry on the list and randomize it using a random shift */
2411 static freelist_idx_t next_random_slot(union freelist_init_state *state)
2413 if (state->pos >= state->count)
2415 return state->list[state->pos++];
2418 /* Swap two freelist entries */
2419 static void swap_free_obj(struct page *page, unsigned int a, unsigned int b)
2421 swap(((freelist_idx_t *)page->freelist)[a],
2422 ((freelist_idx_t *)page->freelist)[b]);
2426 * Shuffle the freelist initialization state based on pre-computed lists.
2427 * return true if the list was successfully shuffled, false otherwise.
2429 static bool shuffle_freelist(struct kmem_cache *cachep, struct page *page)
2431 unsigned int objfreelist = 0, i, rand, count = cachep->num;
2432 union freelist_init_state state;
2438 precomputed = freelist_state_initialize(&state, cachep, count);
2440 /* Take a random entry as the objfreelist */
2441 if (OBJFREELIST_SLAB(cachep)) {
2443 objfreelist = count - 1;
2445 objfreelist = next_random_slot(&state);
2446 page->freelist = index_to_obj(cachep, page, objfreelist) +
2452 * On early boot, generate the list dynamically.
2453 * Later use a pre-computed list for speed.
2456 for (i = 0; i < count; i++)
2457 set_free_obj(page, i, i);
2459 /* Fisher-Yates shuffle */
2460 for (i = count - 1; i > 0; i--) {
2461 rand = prandom_u32_state(&state.rnd_state);
2463 swap_free_obj(page, i, rand);
2466 for (i = 0; i < count; i++)
2467 set_free_obj(page, i, next_random_slot(&state));
2470 if (OBJFREELIST_SLAB(cachep))
2471 set_free_obj(page, cachep->num - 1, objfreelist);
2476 static inline bool shuffle_freelist(struct kmem_cache *cachep,
2481 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
2483 static void cache_init_objs(struct kmem_cache *cachep,
2490 cache_init_objs_debug(cachep, page);
2492 /* Try to randomize the freelist if enabled */
2493 shuffled = shuffle_freelist(cachep, page);
2495 if (!shuffled && OBJFREELIST_SLAB(cachep)) {
2496 page->freelist = index_to_obj(cachep, page, cachep->num - 1) +
2500 for (i = 0; i < cachep->num; i++) {
2501 objp = index_to_obj(cachep, page, i);
2502 objp = kasan_init_slab_obj(cachep, objp);
2504 /* constructor could break poison info */
2505 if (DEBUG == 0 && cachep->ctor) {
2506 kasan_unpoison_object_data(cachep, objp);
2508 kasan_poison_object_data(cachep, objp);
2512 set_free_obj(page, i, i);
2516 static void *slab_get_obj(struct kmem_cache *cachep, struct page *page)
2520 objp = index_to_obj(cachep, page, get_free_obj(page, page->active));
2526 static void slab_put_obj(struct kmem_cache *cachep,
2527 struct page *page, void *objp)
2529 unsigned int objnr = obj_to_index(cachep, page, objp);
2533 /* Verify double free bug */
2534 for (i = page->active; i < cachep->num; i++) {
2535 if (get_free_obj(page, i) == objnr) {
2536 pr_err("slab: double free detected in cache '%s', objp %px\n",
2537 cachep->name, objp);
2543 if (!page->freelist)
2544 page->freelist = objp + obj_offset(cachep);
2546 set_free_obj(page, page->active, objnr);
2550 * Map pages beginning at addr to the given cache and slab. This is required
2551 * for the slab allocator to be able to lookup the cache and slab of a
2552 * virtual address for kfree, ksize, and slab debugging.
2554 static void slab_map_pages(struct kmem_cache *cache, struct page *page,
2557 page->slab_cache = cache;
2558 page->freelist = freelist;
2562 * Grow (by 1) the number of slabs within a cache. This is called by
2563 * kmem_cache_alloc() when there are no active objs left in a cache.
2565 static struct page *cache_grow_begin(struct kmem_cache *cachep,
2566 gfp_t flags, int nodeid)
2572 struct kmem_cache_node *n;
2576 * Be lazy and only check for valid flags here, keeping it out of the
2577 * critical path in kmem_cache_alloc().
2579 if (unlikely(flags & GFP_SLAB_BUG_MASK))
2580 flags = kmalloc_fix_flags(flags);
2582 WARN_ON_ONCE(cachep->ctor && (flags & __GFP_ZERO));
2583 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2586 if (gfpflags_allow_blocking(local_flags))
2590 * Get mem for the objs. Attempt to allocate a physical page from
2593 page = kmem_getpages(cachep, local_flags, nodeid);
2597 page_node = page_to_nid(page);
2598 n = get_node(cachep, page_node);
2600 /* Get colour for the slab, and cal the next value. */
2602 if (n->colour_next >= cachep->colour)
2605 offset = n->colour_next;
2606 if (offset >= cachep->colour)
2609 offset *= cachep->colour_off;
2612 * Call kasan_poison_slab() before calling alloc_slabmgmt(), so
2613 * page_address() in the latter returns a non-tagged pointer,
2614 * as it should be for slab pages.
2616 kasan_poison_slab(page);
2618 /* Get slab management. */
2619 freelist = alloc_slabmgmt(cachep, page, offset,
2620 local_flags & ~GFP_CONSTRAINT_MASK, page_node);
2621 if (OFF_SLAB(cachep) && !freelist)
2624 slab_map_pages(cachep, page, freelist);
2626 cache_init_objs(cachep, page);
2628 if (gfpflags_allow_blocking(local_flags))
2629 local_irq_disable();
2634 kmem_freepages(cachep, page);
2636 if (gfpflags_allow_blocking(local_flags))
2637 local_irq_disable();
2641 static void cache_grow_end(struct kmem_cache *cachep, struct page *page)
2643 struct kmem_cache_node *n;
2651 INIT_LIST_HEAD(&page->slab_list);
2652 n = get_node(cachep, page_to_nid(page));
2654 spin_lock(&n->list_lock);
2656 if (!page->active) {
2657 list_add_tail(&page->slab_list, &n->slabs_free);
2660 fixup_slab_list(cachep, n, page, &list);
2662 STATS_INC_GROWN(cachep);
2663 n->free_objects += cachep->num - page->active;
2664 spin_unlock(&n->list_lock);
2666 fixup_objfreelist_debug(cachep, &list);
2672 * Perform extra freeing checks:
2673 * - detect bad pointers.
2674 * - POISON/RED_ZONE checking
2676 static void kfree_debugcheck(const void *objp)
2678 if (!virt_addr_valid(objp)) {
2679 pr_err("kfree_debugcheck: out of range ptr %lxh\n",
2680 (unsigned long)objp);
2685 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2687 unsigned long long redzone1, redzone2;
2689 redzone1 = *dbg_redzone1(cache, obj);
2690 redzone2 = *dbg_redzone2(cache, obj);
2695 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2698 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2699 slab_error(cache, "double free detected");
2701 slab_error(cache, "memory outside object was overwritten");
2703 pr_err("%px: redzone 1:0x%llx, redzone 2:0x%llx\n",
2704 obj, redzone1, redzone2);
2707 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2708 unsigned long caller)
2713 BUG_ON(virt_to_cache(objp) != cachep);
2715 objp -= obj_offset(cachep);
2716 kfree_debugcheck(objp);
2717 page = virt_to_head_page(objp);
2719 if (cachep->flags & SLAB_RED_ZONE) {
2720 verify_redzone_free(cachep, objp);
2721 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2722 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2724 if (cachep->flags & SLAB_STORE_USER)
2725 *dbg_userword(cachep, objp) = (void *)caller;
2727 objnr = obj_to_index(cachep, page, objp);
2729 BUG_ON(objnr >= cachep->num);
2730 BUG_ON(objp != index_to_obj(cachep, page, objnr));
2732 if (cachep->flags & SLAB_POISON) {
2733 poison_obj(cachep, objp, POISON_FREE);
2734 slab_kernel_map(cachep, objp, 0);
2740 #define kfree_debugcheck(x) do { } while(0)
2741 #define cache_free_debugcheck(x, objp, z) (objp)
2744 static inline void fixup_objfreelist_debug(struct kmem_cache *cachep,
2752 objp = next - obj_offset(cachep);
2753 next = *(void **)next;
2754 poison_obj(cachep, objp, POISON_FREE);
2759 static inline void fixup_slab_list(struct kmem_cache *cachep,
2760 struct kmem_cache_node *n, struct page *page,
2763 /* move slabp to correct slabp list: */
2764 list_del(&page->slab_list);
2765 if (page->active == cachep->num) {
2766 list_add(&page->slab_list, &n->slabs_full);
2767 if (OBJFREELIST_SLAB(cachep)) {
2769 /* Poisoning will be done without holding the lock */
2770 if (cachep->flags & SLAB_POISON) {
2771 void **objp = page->freelist;
2777 page->freelist = NULL;
2780 list_add(&page->slab_list, &n->slabs_partial);
2783 /* Try to find non-pfmemalloc slab if needed */
2784 static noinline struct page *get_valid_first_slab(struct kmem_cache_node *n,
2785 struct page *page, bool pfmemalloc)
2793 if (!PageSlabPfmemalloc(page))
2796 /* No need to keep pfmemalloc slab if we have enough free objects */
2797 if (n->free_objects > n->free_limit) {
2798 ClearPageSlabPfmemalloc(page);
2802 /* Move pfmemalloc slab to the end of list to speed up next search */
2803 list_del(&page->slab_list);
2804 if (!page->active) {
2805 list_add_tail(&page->slab_list, &n->slabs_free);
2808 list_add_tail(&page->slab_list, &n->slabs_partial);
2810 list_for_each_entry(page, &n->slabs_partial, slab_list) {
2811 if (!PageSlabPfmemalloc(page))
2815 n->free_touched = 1;
2816 list_for_each_entry(page, &n->slabs_free, slab_list) {
2817 if (!PageSlabPfmemalloc(page)) {
2826 static struct page *get_first_slab(struct kmem_cache_node *n, bool pfmemalloc)
2830 assert_spin_locked(&n->list_lock);
2831 page = list_first_entry_or_null(&n->slabs_partial, struct page,
2834 n->free_touched = 1;
2835 page = list_first_entry_or_null(&n->slabs_free, struct page,
2841 if (sk_memalloc_socks())
2842 page = get_valid_first_slab(n, page, pfmemalloc);
2847 static noinline void *cache_alloc_pfmemalloc(struct kmem_cache *cachep,
2848 struct kmem_cache_node *n, gfp_t flags)
2854 if (!gfp_pfmemalloc_allowed(flags))
2857 spin_lock(&n->list_lock);
2858 page = get_first_slab(n, true);
2860 spin_unlock(&n->list_lock);
2864 obj = slab_get_obj(cachep, page);
2867 fixup_slab_list(cachep, n, page, &list);
2869 spin_unlock(&n->list_lock);
2870 fixup_objfreelist_debug(cachep, &list);
2876 * Slab list should be fixed up by fixup_slab_list() for existing slab
2877 * or cache_grow_end() for new slab
2879 static __always_inline int alloc_block(struct kmem_cache *cachep,
2880 struct array_cache *ac, struct page *page, int batchcount)
2883 * There must be at least one object available for
2886 BUG_ON(page->active >= cachep->num);
2888 while (page->active < cachep->num && batchcount--) {
2889 STATS_INC_ALLOCED(cachep);
2890 STATS_INC_ACTIVE(cachep);
2891 STATS_SET_HIGH(cachep);
2893 ac->entry[ac->avail++] = slab_get_obj(cachep, page);
2899 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
2902 struct kmem_cache_node *n;
2903 struct array_cache *ac, *shared;
2909 node = numa_mem_id();
2911 ac = cpu_cache_get(cachep);
2912 batchcount = ac->batchcount;
2913 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2915 * If there was little recent activity on this cache, then
2916 * perform only a partial refill. Otherwise we could generate
2919 batchcount = BATCHREFILL_LIMIT;
2921 n = get_node(cachep, node);
2923 BUG_ON(ac->avail > 0 || !n);
2924 shared = READ_ONCE(n->shared);
2925 if (!n->free_objects && (!shared || !shared->avail))
2928 spin_lock(&n->list_lock);
2929 shared = READ_ONCE(n->shared);
2931 /* See if we can refill from the shared array */
2932 if (shared && transfer_objects(ac, shared, batchcount)) {
2933 shared->touched = 1;
2937 while (batchcount > 0) {
2938 /* Get slab alloc is to come from. */
2939 page = get_first_slab(n, false);
2943 check_spinlock_acquired(cachep);
2945 batchcount = alloc_block(cachep, ac, page, batchcount);
2946 fixup_slab_list(cachep, n, page, &list);
2950 n->free_objects -= ac->avail;
2952 spin_unlock(&n->list_lock);
2953 fixup_objfreelist_debug(cachep, &list);
2956 if (unlikely(!ac->avail)) {
2957 /* Check if we can use obj in pfmemalloc slab */
2958 if (sk_memalloc_socks()) {
2959 void *obj = cache_alloc_pfmemalloc(cachep, n, flags);
2965 page = cache_grow_begin(cachep, gfp_exact_node(flags), node);
2968 * cache_grow_begin() can reenable interrupts,
2969 * then ac could change.
2971 ac = cpu_cache_get(cachep);
2972 if (!ac->avail && page)
2973 alloc_block(cachep, ac, page, batchcount);
2974 cache_grow_end(cachep, page);
2981 return ac->entry[--ac->avail];
2984 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2987 might_sleep_if(gfpflags_allow_blocking(flags));
2991 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
2992 gfp_t flags, void *objp, unsigned long caller)
2994 WARN_ON_ONCE(cachep->ctor && (flags & __GFP_ZERO));
2997 if (cachep->flags & SLAB_POISON) {
2998 check_poison_obj(cachep, objp);
2999 slab_kernel_map(cachep, objp, 1);
3000 poison_obj(cachep, objp, POISON_INUSE);
3002 if (cachep->flags & SLAB_STORE_USER)
3003 *dbg_userword(cachep, objp) = (void *)caller;
3005 if (cachep->flags & SLAB_RED_ZONE) {
3006 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3007 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3008 slab_error(cachep, "double free, or memory outside object was overwritten");
3009 pr_err("%px: redzone 1:0x%llx, redzone 2:0x%llx\n",
3010 objp, *dbg_redzone1(cachep, objp),
3011 *dbg_redzone2(cachep, objp));
3013 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3014 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3017 objp += obj_offset(cachep);
3018 if (cachep->ctor && cachep->flags & SLAB_POISON)
3020 if (ARCH_SLAB_MINALIGN &&
3021 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
3022 pr_err("0x%px: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3023 objp, (int)ARCH_SLAB_MINALIGN);
3028 #define cache_alloc_debugcheck_after(a, b, objp, d) (objp)
3031 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3034 struct array_cache *ac;
3038 ac = cpu_cache_get(cachep);
3039 if (likely(ac->avail)) {
3041 objp = ac->entry[--ac->avail];
3043 STATS_INC_ALLOCHIT(cachep);
3047 STATS_INC_ALLOCMISS(cachep);
3048 objp = cache_alloc_refill(cachep, flags);
3050 * the 'ac' may be updated by cache_alloc_refill(),
3051 * and kmemleak_erase() requires its correct value.
3053 ac = cpu_cache_get(cachep);
3057 * To avoid a false negative, if an object that is in one of the
3058 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3059 * treat the array pointers as a reference to the object.
3062 kmemleak_erase(&ac->entry[ac->avail]);
3068 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
3070 * If we are in_interrupt, then process context, including cpusets and
3071 * mempolicy, may not apply and should not be used for allocation policy.
3073 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3075 int nid_alloc, nid_here;
3077 if (in_interrupt() || (flags & __GFP_THISNODE))
3079 nid_alloc = nid_here = numa_mem_id();
3080 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3081 nid_alloc = cpuset_slab_spread_node();
3082 else if (current->mempolicy)
3083 nid_alloc = mempolicy_slab_node();
3084 if (nid_alloc != nid_here)
3085 return ____cache_alloc_node(cachep, flags, nid_alloc);
3090 * Fallback function if there was no memory available and no objects on a
3091 * certain node and fall back is permitted. First we scan all the
3092 * available node for available objects. If that fails then we
3093 * perform an allocation without specifying a node. This allows the page
3094 * allocator to do its reclaim / fallback magic. We then insert the
3095 * slab into the proper nodelist and then allocate from it.
3097 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3099 struct zonelist *zonelist;
3102 enum zone_type highest_zoneidx = gfp_zone(flags);
3106 unsigned int cpuset_mems_cookie;
3108 if (flags & __GFP_THISNODE)
3112 cpuset_mems_cookie = read_mems_allowed_begin();
3113 zonelist = node_zonelist(mempolicy_slab_node(), flags);
3117 * Look through allowed nodes for objects available
3118 * from existing per node queues.
3120 for_each_zone_zonelist(zone, z, zonelist, highest_zoneidx) {
3121 nid = zone_to_nid(zone);
3123 if (cpuset_zone_allowed(zone, flags) &&
3124 get_node(cache, nid) &&
3125 get_node(cache, nid)->free_objects) {
3126 obj = ____cache_alloc_node(cache,
3127 gfp_exact_node(flags), nid);
3135 * This allocation will be performed within the constraints
3136 * of the current cpuset / memory policy requirements.
3137 * We may trigger various forms of reclaim on the allowed
3138 * set and go into memory reserves if necessary.
3140 page = cache_grow_begin(cache, flags, numa_mem_id());
3141 cache_grow_end(cache, page);
3143 nid = page_to_nid(page);
3144 obj = ____cache_alloc_node(cache,
3145 gfp_exact_node(flags), nid);
3148 * Another processor may allocate the objects in
3149 * the slab since we are not holding any locks.
3156 if (unlikely(!obj && read_mems_allowed_retry(cpuset_mems_cookie)))
3162 * A interface to enable slab creation on nodeid
3164 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3168 struct kmem_cache_node *n;
3172 VM_BUG_ON(nodeid < 0 || nodeid >= MAX_NUMNODES);
3173 n = get_node(cachep, nodeid);
3177 spin_lock(&n->list_lock);
3178 page = get_first_slab(n, false);
3182 check_spinlock_acquired_node(cachep, nodeid);
3184 STATS_INC_NODEALLOCS(cachep);
3185 STATS_INC_ACTIVE(cachep);
3186 STATS_SET_HIGH(cachep);
3188 BUG_ON(page->active == cachep->num);
3190 obj = slab_get_obj(cachep, page);
3193 fixup_slab_list(cachep, n, page, &list);
3195 spin_unlock(&n->list_lock);
3196 fixup_objfreelist_debug(cachep, &list);
3200 spin_unlock(&n->list_lock);
3201 page = cache_grow_begin(cachep, gfp_exact_node(flags), nodeid);
3203 /* This slab isn't counted yet so don't update free_objects */
3204 obj = slab_get_obj(cachep, page);
3206 cache_grow_end(cachep, page);
3208 return obj ? obj : fallback_alloc(cachep, flags);
3211 static __always_inline void *
3212 slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid, size_t orig_size,
3213 unsigned long caller)
3215 unsigned long save_flags;
3217 int slab_node = numa_mem_id();
3218 struct obj_cgroup *objcg = NULL;
3220 flags &= gfp_allowed_mask;
3221 cachep = slab_pre_alloc_hook(cachep, &objcg, 1, flags);
3222 if (unlikely(!cachep))
3225 ptr = kfence_alloc(cachep, orig_size, flags);
3229 cache_alloc_debugcheck_before(cachep, flags);
3230 local_irq_save(save_flags);
3232 if (nodeid == NUMA_NO_NODE)
3235 if (unlikely(!get_node(cachep, nodeid))) {
3236 /* Node not bootstrapped yet */
3237 ptr = fallback_alloc(cachep, flags);
3241 if (nodeid == slab_node) {
3243 * Use the locally cached objects if possible.
3244 * However ____cache_alloc does not allow fallback
3245 * to other nodes. It may fail while we still have
3246 * objects on other nodes available.
3248 ptr = ____cache_alloc(cachep, flags);
3252 /* ___cache_alloc_node can fall back to other nodes */
3253 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3255 local_irq_restore(save_flags);
3256 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3258 if (unlikely(slab_want_init_on_alloc(flags, cachep)) && ptr)
3259 memset(ptr, 0, cachep->object_size);
3262 slab_post_alloc_hook(cachep, objcg, flags, 1, &ptr);
3266 static __always_inline void *
3267 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3271 if (current->mempolicy || cpuset_do_slab_mem_spread()) {
3272 objp = alternate_node_alloc(cache, flags);
3276 objp = ____cache_alloc(cache, flags);
3279 * We may just have run out of memory on the local node.
3280 * ____cache_alloc_node() knows how to locate memory on other nodes
3283 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3290 static __always_inline void *
3291 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3293 return ____cache_alloc(cachep, flags);
3296 #endif /* CONFIG_NUMA */
3298 static __always_inline void *
3299 slab_alloc(struct kmem_cache *cachep, gfp_t flags, size_t orig_size, unsigned long caller)
3301 unsigned long save_flags;
3303 struct obj_cgroup *objcg = NULL;
3305 flags &= gfp_allowed_mask;
3306 cachep = slab_pre_alloc_hook(cachep, &objcg, 1, flags);
3307 if (unlikely(!cachep))
3310 objp = kfence_alloc(cachep, orig_size, flags);
3314 cache_alloc_debugcheck_before(cachep, flags);
3315 local_irq_save(save_flags);
3316 objp = __do_cache_alloc(cachep, flags);
3317 local_irq_restore(save_flags);
3318 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3321 if (unlikely(slab_want_init_on_alloc(flags, cachep)) && objp)
3322 memset(objp, 0, cachep->object_size);
3325 slab_post_alloc_hook(cachep, objcg, flags, 1, &objp);
3330 * Caller needs to acquire correct kmem_cache_node's list_lock
3331 * @list: List of detached free slabs should be freed by caller
3333 static void free_block(struct kmem_cache *cachep, void **objpp,
3334 int nr_objects, int node, struct list_head *list)
3337 struct kmem_cache_node *n = get_node(cachep, node);
3340 n->free_objects += nr_objects;
3342 for (i = 0; i < nr_objects; i++) {
3348 page = virt_to_head_page(objp);
3349 list_del(&page->slab_list);
3350 check_spinlock_acquired_node(cachep, node);
3351 slab_put_obj(cachep, page, objp);
3352 STATS_DEC_ACTIVE(cachep);
3354 /* fixup slab chains */
3355 if (page->active == 0) {
3356 list_add(&page->slab_list, &n->slabs_free);
3359 /* Unconditionally move a slab to the end of the
3360 * partial list on free - maximum time for the
3361 * other objects to be freed, too.
3363 list_add_tail(&page->slab_list, &n->slabs_partial);
3367 while (n->free_objects > n->free_limit && !list_empty(&n->slabs_free)) {
3368 n->free_objects -= cachep->num;
3370 page = list_last_entry(&n->slabs_free, struct page, slab_list);
3371 list_move(&page->slab_list, list);
3377 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3380 struct kmem_cache_node *n;
3381 int node = numa_mem_id();
3384 batchcount = ac->batchcount;
3387 n = get_node(cachep, node);
3388 spin_lock(&n->list_lock);
3390 struct array_cache *shared_array = n->shared;
3391 int max = shared_array->limit - shared_array->avail;
3393 if (batchcount > max)
3395 memcpy(&(shared_array->entry[shared_array->avail]),
3396 ac->entry, sizeof(void *) * batchcount);
3397 shared_array->avail += batchcount;
3402 free_block(cachep, ac->entry, batchcount, node, &list);
3409 list_for_each_entry(page, &n->slabs_free, slab_list) {
3410 BUG_ON(page->active);
3414 STATS_SET_FREEABLE(cachep, i);
3417 spin_unlock(&n->list_lock);
3418 ac->avail -= batchcount;
3419 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3420 slabs_destroy(cachep, &list);
3424 * Release an obj back to its cache. If the obj has a constructed state, it must
3425 * be in this state _before_ it is released. Called with disabled ints.
3427 static __always_inline void __cache_free(struct kmem_cache *cachep, void *objp,
3428 unsigned long caller)
3430 if (is_kfence_address(objp)) {
3431 kmemleak_free_recursive(objp, cachep->flags);
3432 __kfence_free(objp);
3436 if (unlikely(slab_want_init_on_free(cachep)))
3437 memset(objp, 0, cachep->object_size);
3439 /* Put the object into the quarantine, don't touch it for now. */
3440 if (kasan_slab_free(cachep, objp))
3443 /* Use KCSAN to help debug racy use-after-free. */
3444 if (!(cachep->flags & SLAB_TYPESAFE_BY_RCU))
3445 __kcsan_check_access(objp, cachep->object_size,
3446 KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT);
3448 ___cache_free(cachep, objp, caller);
3451 void ___cache_free(struct kmem_cache *cachep, void *objp,
3452 unsigned long caller)
3454 struct array_cache *ac = cpu_cache_get(cachep);
3457 kmemleak_free_recursive(objp, cachep->flags);
3458 objp = cache_free_debugcheck(cachep, objp, caller);
3459 memcg_slab_free_hook(cachep, &objp, 1);
3462 * Skip calling cache_free_alien() when the platform is not numa.
3463 * This will avoid cache misses that happen while accessing slabp (which
3464 * is per page memory reference) to get nodeid. Instead use a global
3465 * variable to skip the call, which is mostly likely to be present in
3468 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3471 if (ac->avail < ac->limit) {
3472 STATS_INC_FREEHIT(cachep);
3474 STATS_INC_FREEMISS(cachep);
3475 cache_flusharray(cachep, ac);
3478 if (sk_memalloc_socks()) {
3479 struct page *page = virt_to_head_page(objp);
3481 if (unlikely(PageSlabPfmemalloc(page))) {
3482 cache_free_pfmemalloc(cachep, page, objp);
3487 __free_one(ac, objp);
3491 * kmem_cache_alloc - Allocate an object
3492 * @cachep: The cache to allocate from.
3493 * @flags: See kmalloc().
3495 * Allocate an object from this cache. The flags are only relevant
3496 * if the cache has no available objects.
3498 * Return: pointer to the new object or %NULL in case of error
3500 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3502 void *ret = slab_alloc(cachep, flags, cachep->object_size, _RET_IP_);
3504 trace_kmem_cache_alloc(_RET_IP_, ret,
3505 cachep->object_size, cachep->size, flags);
3509 EXPORT_SYMBOL(kmem_cache_alloc);
3511 static __always_inline void
3512 cache_alloc_debugcheck_after_bulk(struct kmem_cache *s, gfp_t flags,
3513 size_t size, void **p, unsigned long caller)
3517 for (i = 0; i < size; i++)
3518 p[i] = cache_alloc_debugcheck_after(s, flags, p[i], caller);
3521 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3525 struct obj_cgroup *objcg = NULL;
3527 s = slab_pre_alloc_hook(s, &objcg, size, flags);
3531 cache_alloc_debugcheck_before(s, flags);
3533 local_irq_disable();
3534 for (i = 0; i < size; i++) {
3535 void *objp = kfence_alloc(s, s->object_size, flags) ?: __do_cache_alloc(s, flags);
3537 if (unlikely(!objp))
3543 cache_alloc_debugcheck_after_bulk(s, flags, size, p, _RET_IP_);
3545 /* Clear memory outside IRQ disabled section */
3546 if (unlikely(slab_want_init_on_alloc(flags, s)))
3547 for (i = 0; i < size; i++)
3548 memset(p[i], 0, s->object_size);
3550 slab_post_alloc_hook(s, objcg, flags, size, p);
3551 /* FIXME: Trace call missing. Christoph would like a bulk variant */
3555 cache_alloc_debugcheck_after_bulk(s, flags, i, p, _RET_IP_);
3556 slab_post_alloc_hook(s, objcg, flags, i, p);
3557 __kmem_cache_free_bulk(s, i, p);
3560 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3562 #ifdef CONFIG_TRACING
3564 kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
3568 ret = slab_alloc(cachep, flags, size, _RET_IP_);
3570 ret = kasan_kmalloc(cachep, ret, size, flags);
3571 trace_kmalloc(_RET_IP_, ret,
3572 size, cachep->size, flags);
3575 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3580 * kmem_cache_alloc_node - Allocate an object on the specified node
3581 * @cachep: The cache to allocate from.
3582 * @flags: See kmalloc().
3583 * @nodeid: node number of the target node.
3585 * Identical to kmem_cache_alloc but it will allocate memory on the given
3586 * node, which can improve the performance for cpu bound structures.
3588 * Fallback to other node is possible if __GFP_THISNODE is not set.
3590 * Return: pointer to the new object or %NULL in case of error
3592 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3594 void *ret = slab_alloc_node(cachep, flags, nodeid, cachep->object_size, _RET_IP_);
3596 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3597 cachep->object_size, cachep->size,
3602 EXPORT_SYMBOL(kmem_cache_alloc_node);
3604 #ifdef CONFIG_TRACING
3605 void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
3612 ret = slab_alloc_node(cachep, flags, nodeid, size, _RET_IP_);
3614 ret = kasan_kmalloc(cachep, ret, size, flags);
3615 trace_kmalloc_node(_RET_IP_, ret,
3620 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3623 static __always_inline void *
3624 __do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
3626 struct kmem_cache *cachep;
3629 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3631 cachep = kmalloc_slab(size, flags);
3632 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3634 ret = kmem_cache_alloc_node_trace(cachep, flags, node, size);
3635 ret = kasan_kmalloc(cachep, ret, size, flags);
3640 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3642 return __do_kmalloc_node(size, flags, node, _RET_IP_);
3644 EXPORT_SYMBOL(__kmalloc_node);
3646 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3647 int node, unsigned long caller)
3649 return __do_kmalloc_node(size, flags, node, caller);
3651 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3652 #endif /* CONFIG_NUMA */
3654 void kmem_obj_info(struct kmem_obj_info *kpp, void *object, struct page *page)
3656 struct kmem_cache *cachep;
3660 kpp->kp_ptr = object;
3661 kpp->kp_page = page;
3662 cachep = page->slab_cache;
3663 kpp->kp_slab_cache = cachep;
3664 objp = object - obj_offset(cachep);
3665 kpp->kp_data_offset = obj_offset(cachep);
3666 page = virt_to_head_page(objp);
3667 objnr = obj_to_index(cachep, page, objp);
3668 objp = index_to_obj(cachep, page, objnr);
3669 kpp->kp_objp = objp;
3670 if (DEBUG && cachep->flags & SLAB_STORE_USER)
3671 kpp->kp_ret = *dbg_userword(cachep, objp);
3675 * __do_kmalloc - allocate memory
3676 * @size: how many bytes of memory are required.
3677 * @flags: the type of memory to allocate (see kmalloc).
3678 * @caller: function caller for debug tracking of the caller
3680 * Return: pointer to the allocated memory or %NULL in case of error
3682 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3683 unsigned long caller)
3685 struct kmem_cache *cachep;
3688 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3690 cachep = kmalloc_slab(size, flags);
3691 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3693 ret = slab_alloc(cachep, flags, size, caller);
3695 ret = kasan_kmalloc(cachep, ret, size, flags);
3696 trace_kmalloc(caller, ret,
3697 size, cachep->size, flags);
3702 void *__kmalloc(size_t size, gfp_t flags)
3704 return __do_kmalloc(size, flags, _RET_IP_);
3706 EXPORT_SYMBOL(__kmalloc);
3708 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3710 return __do_kmalloc(size, flags, caller);
3712 EXPORT_SYMBOL(__kmalloc_track_caller);
3715 * kmem_cache_free - Deallocate an object
3716 * @cachep: The cache the allocation was from.
3717 * @objp: The previously allocated object.
3719 * Free an object which was previously allocated from this
3722 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3724 unsigned long flags;
3725 cachep = cache_from_obj(cachep, objp);
3729 local_irq_save(flags);
3730 debug_check_no_locks_freed(objp, cachep->object_size);
3731 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3732 debug_check_no_obj_freed(objp, cachep->object_size);
3733 __cache_free(cachep, objp, _RET_IP_);
3734 local_irq_restore(flags);
3736 trace_kmem_cache_free(_RET_IP_, objp, cachep->name);
3738 EXPORT_SYMBOL(kmem_cache_free);
3740 void kmem_cache_free_bulk(struct kmem_cache *orig_s, size_t size, void **p)
3742 struct kmem_cache *s;
3745 local_irq_disable();
3746 for (i = 0; i < size; i++) {
3749 if (!orig_s) /* called via kfree_bulk */
3750 s = virt_to_cache(objp);
3752 s = cache_from_obj(orig_s, objp);
3756 debug_check_no_locks_freed(objp, s->object_size);
3757 if (!(s->flags & SLAB_DEBUG_OBJECTS))
3758 debug_check_no_obj_freed(objp, s->object_size);
3760 __cache_free(s, objp, _RET_IP_);
3764 /* FIXME: add tracing */
3766 EXPORT_SYMBOL(kmem_cache_free_bulk);
3769 * kfree - free previously allocated memory
3770 * @objp: pointer returned by kmalloc.
3772 * If @objp is NULL, no operation is performed.
3774 * Don't free memory not originally allocated by kmalloc()
3775 * or you will run into trouble.
3777 void kfree(const void *objp)
3779 struct kmem_cache *c;
3780 unsigned long flags;
3782 trace_kfree(_RET_IP_, objp);
3784 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3786 local_irq_save(flags);
3787 kfree_debugcheck(objp);
3788 c = virt_to_cache(objp);
3790 local_irq_restore(flags);
3793 debug_check_no_locks_freed(objp, c->object_size);
3795 debug_check_no_obj_freed(objp, c->object_size);
3796 __cache_free(c, (void *)objp, _RET_IP_);
3797 local_irq_restore(flags);
3799 EXPORT_SYMBOL(kfree);
3802 * This initializes kmem_cache_node or resizes various caches for all nodes.
3804 static int setup_kmem_cache_nodes(struct kmem_cache *cachep, gfp_t gfp)
3808 struct kmem_cache_node *n;
3810 for_each_online_node(node) {
3811 ret = setup_kmem_cache_node(cachep, node, gfp, true);
3820 if (!cachep->list.next) {
3821 /* Cache is not active yet. Roll back what we did */
3824 n = get_node(cachep, node);
3827 free_alien_cache(n->alien);
3829 cachep->node[node] = NULL;
3837 /* Always called with the slab_mutex held */
3838 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3839 int batchcount, int shared, gfp_t gfp)
3841 struct array_cache __percpu *cpu_cache, *prev;
3844 cpu_cache = alloc_kmem_cache_cpus(cachep, limit, batchcount);
3848 prev = cachep->cpu_cache;
3849 cachep->cpu_cache = cpu_cache;
3851 * Without a previous cpu_cache there's no need to synchronize remote
3852 * cpus, so skip the IPIs.
3855 kick_all_cpus_sync();
3858 cachep->batchcount = batchcount;
3859 cachep->limit = limit;
3860 cachep->shared = shared;
3865 for_each_online_cpu(cpu) {
3868 struct kmem_cache_node *n;
3869 struct array_cache *ac = per_cpu_ptr(prev, cpu);
3871 node = cpu_to_mem(cpu);
3872 n = get_node(cachep, node);
3873 spin_lock_irq(&n->list_lock);
3874 free_block(cachep, ac->entry, ac->avail, node, &list);
3875 spin_unlock_irq(&n->list_lock);
3876 slabs_destroy(cachep, &list);
3881 return setup_kmem_cache_nodes(cachep, gfp);
3884 /* Called with slab_mutex held always */
3885 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
3892 err = cache_random_seq_create(cachep, cachep->num, gfp);
3896 if (limit && shared && batchcount)
3899 * The head array serves three purposes:
3900 * - create a LIFO ordering, i.e. return objects that are cache-warm
3901 * - reduce the number of spinlock operations.
3902 * - reduce the number of linked list operations on the slab and
3903 * bufctl chains: array operations are cheaper.
3904 * The numbers are guessed, we should auto-tune as described by
3907 if (cachep->size > 131072)
3909 else if (cachep->size > PAGE_SIZE)
3911 else if (cachep->size > 1024)
3913 else if (cachep->size > 256)
3919 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3920 * allocation behaviour: Most allocs on one cpu, most free operations
3921 * on another cpu. For these cases, an efficient object passing between
3922 * cpus is necessary. This is provided by a shared array. The array
3923 * replaces Bonwick's magazine layer.
3924 * On uniprocessor, it's functionally equivalent (but less efficient)
3925 * to a larger limit. Thus disabled by default.
3928 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
3933 * With debugging enabled, large batchcount lead to excessively long
3934 * periods with disabled local interrupts. Limit the batchcount
3939 batchcount = (limit + 1) / 2;
3941 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3944 pr_err("enable_cpucache failed for %s, error %d\n",
3945 cachep->name, -err);
3950 * Drain an array if it contains any elements taking the node lock only if
3951 * necessary. Note that the node listlock also protects the array_cache
3952 * if drain_array() is used on the shared array.
3954 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
3955 struct array_cache *ac, int node)
3959 /* ac from n->shared can be freed if we don't hold the slab_mutex. */
3960 check_mutex_acquired();
3962 if (!ac || !ac->avail)
3970 spin_lock_irq(&n->list_lock);
3971 drain_array_locked(cachep, ac, node, false, &list);
3972 spin_unlock_irq(&n->list_lock);
3974 slabs_destroy(cachep, &list);
3978 * cache_reap - Reclaim memory from caches.
3979 * @w: work descriptor
3981 * Called from workqueue/eventd every few seconds.
3983 * - clear the per-cpu caches for this CPU.
3984 * - return freeable pages to the main free memory pool.
3986 * If we cannot acquire the cache chain mutex then just give up - we'll try
3987 * again on the next iteration.
3989 static void cache_reap(struct work_struct *w)
3991 struct kmem_cache *searchp;
3992 struct kmem_cache_node *n;
3993 int node = numa_mem_id();
3994 struct delayed_work *work = to_delayed_work(w);
3996 if (!mutex_trylock(&slab_mutex))
3997 /* Give up. Setup the next iteration. */
4000 list_for_each_entry(searchp, &slab_caches, list) {
4004 * We only take the node lock if absolutely necessary and we
4005 * have established with reasonable certainty that
4006 * we can do some work if the lock was obtained.
4008 n = get_node(searchp, node);
4010 reap_alien(searchp, n);
4012 drain_array(searchp, n, cpu_cache_get(searchp), node);
4015 * These are racy checks but it does not matter
4016 * if we skip one check or scan twice.
4018 if (time_after(n->next_reap, jiffies))
4021 n->next_reap = jiffies + REAPTIMEOUT_NODE;
4023 drain_array(searchp, n, n->shared, node);
4025 if (n->free_touched)
4026 n->free_touched = 0;
4030 freed = drain_freelist(searchp, n, (n->free_limit +
4031 5 * searchp->num - 1) / (5 * searchp->num));
4032 STATS_ADD_REAPED(searchp, freed);
4038 mutex_unlock(&slab_mutex);
4041 /* Set up the next iteration */
4042 schedule_delayed_work_on(smp_processor_id(), work,
4043 round_jiffies_relative(REAPTIMEOUT_AC));
4046 void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
4048 unsigned long active_objs, num_objs, active_slabs;
4049 unsigned long total_slabs = 0, free_objs = 0, shared_avail = 0;
4050 unsigned long free_slabs = 0;
4052 struct kmem_cache_node *n;
4054 for_each_kmem_cache_node(cachep, node, n) {
4056 spin_lock_irq(&n->list_lock);
4058 total_slabs += n->total_slabs;
4059 free_slabs += n->free_slabs;
4060 free_objs += n->free_objects;
4063 shared_avail += n->shared->avail;
4065 spin_unlock_irq(&n->list_lock);
4067 num_objs = total_slabs * cachep->num;
4068 active_slabs = total_slabs - free_slabs;
4069 active_objs = num_objs - free_objs;
4071 sinfo->active_objs = active_objs;
4072 sinfo->num_objs = num_objs;
4073 sinfo->active_slabs = active_slabs;
4074 sinfo->num_slabs = total_slabs;
4075 sinfo->shared_avail = shared_avail;
4076 sinfo->limit = cachep->limit;
4077 sinfo->batchcount = cachep->batchcount;
4078 sinfo->shared = cachep->shared;
4079 sinfo->objects_per_slab = cachep->num;
4080 sinfo->cache_order = cachep->gfporder;
4083 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4087 unsigned long high = cachep->high_mark;
4088 unsigned long allocs = cachep->num_allocations;
4089 unsigned long grown = cachep->grown;
4090 unsigned long reaped = cachep->reaped;
4091 unsigned long errors = cachep->errors;
4092 unsigned long max_freeable = cachep->max_freeable;
4093 unsigned long node_allocs = cachep->node_allocs;
4094 unsigned long node_frees = cachep->node_frees;
4095 unsigned long overflows = cachep->node_overflow;
4097 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu %4lu %4lu",
4098 allocs, high, grown,
4099 reaped, errors, max_freeable, node_allocs,
4100 node_frees, overflows);
4104 unsigned long allochit = atomic_read(&cachep->allochit);
4105 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4106 unsigned long freehit = atomic_read(&cachep->freehit);
4107 unsigned long freemiss = atomic_read(&cachep->freemiss);
4109 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4110 allochit, allocmiss, freehit, freemiss);
4115 #define MAX_SLABINFO_WRITE 128
4117 * slabinfo_write - Tuning for the slab allocator
4119 * @buffer: user buffer
4120 * @count: data length
4123 * Return: %0 on success, negative error code otherwise.
4125 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4126 size_t count, loff_t *ppos)
4128 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4129 int limit, batchcount, shared, res;
4130 struct kmem_cache *cachep;
4132 if (count > MAX_SLABINFO_WRITE)
4134 if (copy_from_user(&kbuf, buffer, count))
4136 kbuf[MAX_SLABINFO_WRITE] = '\0';
4138 tmp = strchr(kbuf, ' ');
4143 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4146 /* Find the cache in the chain of caches. */
4147 mutex_lock(&slab_mutex);
4149 list_for_each_entry(cachep, &slab_caches, list) {
4150 if (!strcmp(cachep->name, kbuf)) {
4151 if (limit < 1 || batchcount < 1 ||
4152 batchcount > limit || shared < 0) {
4155 res = do_tune_cpucache(cachep, limit,
4162 mutex_unlock(&slab_mutex);
4168 #ifdef CONFIG_HARDENED_USERCOPY
4170 * Rejects incorrectly sized objects and objects that are to be copied
4171 * to/from userspace but do not fall entirely within the containing slab
4172 * cache's usercopy region.
4174 * Returns NULL if check passes, otherwise const char * to name of cache
4175 * to indicate an error.
4177 void __check_heap_object(const void *ptr, unsigned long n, struct page *page,
4180 struct kmem_cache *cachep;
4182 unsigned long offset;
4184 ptr = kasan_reset_tag(ptr);
4186 /* Find and validate object. */
4187 cachep = page->slab_cache;
4188 objnr = obj_to_index(cachep, page, (void *)ptr);
4189 BUG_ON(objnr >= cachep->num);
4191 /* Find offset within object. */
4192 if (is_kfence_address(ptr))
4193 offset = ptr - kfence_object_start(ptr);
4195 offset = ptr - index_to_obj(cachep, page, objnr) - obj_offset(cachep);
4197 /* Allow address range falling entirely within usercopy region. */
4198 if (offset >= cachep->useroffset &&
4199 offset - cachep->useroffset <= cachep->usersize &&
4200 n <= cachep->useroffset - offset + cachep->usersize)
4204 * If the copy is still within the allocated object, produce
4205 * a warning instead of rejecting the copy. This is intended
4206 * to be a temporary method to find any missing usercopy
4209 if (usercopy_fallback &&
4210 offset <= cachep->object_size &&
4211 n <= cachep->object_size - offset) {
4212 usercopy_warn("SLAB object", cachep->name, to_user, offset, n);
4216 usercopy_abort("SLAB object", cachep->name, to_user, offset, n);
4218 #endif /* CONFIG_HARDENED_USERCOPY */
4221 * __ksize -- Uninstrumented ksize.
4222 * @objp: pointer to the object
4224 * Unlike ksize(), __ksize() is uninstrumented, and does not provide the same
4225 * safety checks as ksize() with KASAN instrumentation enabled.
4227 * Return: size of the actual memory used by @objp in bytes
4229 size_t __ksize(const void *objp)
4231 struct kmem_cache *c;
4235 if (unlikely(objp == ZERO_SIZE_PTR))
4238 c = virt_to_cache(objp);
4239 size = c ? c->object_size : 0;
4243 EXPORT_SYMBOL(__ksize);