Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * linux/mm/slab.c
4 * Written by Mark Hemment, 1996/97.
5 * (markhe@nextd.demon.co.uk)
6 *
7 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
8 *
9 * Major cleanup, different bufctl logic, per-cpu arrays
10 * (c) 2000 Manfred Spraul
11 *
12 * Cleanup, make the head arrays unconditional, preparation for NUMA
13 * (c) 2002 Manfred Spraul
14 *
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
22 *
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.
28 *
29 * This means, that your constructor is used only for newly allocated
30 * slabs and you must pass objects with the same initializations to
31 * kmem_cache_free.
32 *
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.
36 *
37 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
38 * full slabs with 0 free objects
39 * partial slabs
40 * empty slabs with no allocated objects
41 *
42 * If partial slabs exist, then new allocations come from these slabs,
43 * otherwise from empty slabs or new slabs are allocated.
44 *
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.
47 *
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.
53 *
54 * The c_cpuarray may not be read with enabled local interrupts -
55 * it's changed with a smp_call_function().
56 *
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.
64 *
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
67 * his patch.
68 *
69 * Further notes from the original documentation:
70 *
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()).
76 *
77 * At present, each engine can be growing a cache. This should be blocked.
78 *
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>
84 *
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.
88 */
89
90#include <linux/__KEEPIDENTS__B.h>
91#include <linux/__KEEPIDENTS__C.h>
92#include <linux/__KEEPIDENTS__D.h>
93#include <linux/__KEEPIDENTS__E.h>
94#include <linux/__KEEPIDENTS__F.h>
95#include <linux/__KEEPIDENTS__G.h>
96#include <linux/__KEEPIDENTS__H.h>
97#include <linux/__KEEPIDENTS__I.h>
98#include <linux/__KEEPIDENTS__J.h>
99#include <linux/proc_fs.h>
100#include <linux/__KEEPIDENTS__BA.h>
101#include <linux/__KEEPIDENTS__BB.h>
102#include <linux/__KEEPIDENTS__BC.h>
103#include <linux/kfence.h>
104#include <linux/cpu.h>
105#include <linux/__KEEPIDENTS__BD.h>
106#include <linux/__KEEPIDENTS__BE.h>
107#include <linux/rcupdate.h>
108#include <linux/__KEEPIDENTS__BF.h>
109#include <linux/__KEEPIDENTS__BG.h>
110#include <linux/__KEEPIDENTS__BH.h>
111#include <linux/kmemleak.h>
112#include <linux/__KEEPIDENTS__BI.h>
113#include <linux/__KEEPIDENTS__BJ.h>
114#include <linux/__KEEPIDENTS__CA-__KEEPIDENTS__CB.h>
115#include <linux/__KEEPIDENTS__CC.h>
116#include <linux/reciprocal_div.h>
117#include <linux/debugobjects.h>
118#include <linux/__KEEPIDENTS__CD.h>
119#include <linux/__KEEPIDENTS__CE.h>
120#include <linux/__KEEPIDENTS__CF/task_stack.h>
121
122#include <net/__KEEPIDENTS__CG.h>
123
124#include <asm/cacheflush.h>
125#include <asm/tlbflush.h>
126#include <asm/page.h>
127
128#include <trace/events/kmem.h>
129
130#include "internal.h"
131
132#include "slab.h"
133
134/*
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).
137 *
138 * STATS - 1 to collect stats for /proc/slabinfo.
139 * 0 for faster, smaller code (especially in the critical paths).
140 *
141 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
142 */
143
144#ifdef CONFIG_DEBUG_SLAB
145#define DEBUG 1
146#define STATS 1
147#define FORCED_DEBUG 1
148#else
149#define DEBUG 0
150#define STATS 0
151#define FORCED_DEBUG 0
152#endif
153
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))
157
158#ifndef ARCH_KMALLOC_FLAGS
159#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
160#endif
161
162#define FREELIST_BYTE_INDEX (((PAGE_SIZE >> BITS_PER_BYTE) \
163 <= SLAB_OBJ_MIN_SIZE) ? 1 : 0)
164
165#if FREELIST_BYTE_INDEX
166typedef unsigned char freelist_idx_t;
167#else
168typedef unsigned short freelist_idx_t;
169#endif
170
171#define SLAB_OBJ_MAX_NUM ((1 << sizeof(freelist_idx_t) * BITS_PER_BYTE) - 1)
172
173/*
174 * struct array_cache
175 *
176 * Purpose:
177 * - LIFO ordering, to hand out cache-warm objects from _alloc
178 * - reduce the number of linked list operations
179 * - reduce spinlock operations
180 *
181 * The limit is stored in the per-cpu structure to reduce the data cache
182 * footprint.
183 *
184 */
185struct array_cache {
186 unsigned int avail;
187 unsigned int limit;
188 unsigned int batchcount;
189 unsigned int touched;
190 void *entry[]; /*
191 * Must have this definition in here for the proper
192 * alignment of array_cache. Also simplifies accessing
193 * the entries.
194 */
195};
196
197struct alien_cache {
198 spinlock_t lock;
199 struct array_cache ac;
200};
201
202/*
203 * Need this for bootstrapping a per node allocator.
204 */
205#define NUM_INIT_LISTS (2 * MAX_NUMNODES)
206static struct kmem_cache_node __initdata init_kmem_cache_node[NUM_INIT_LISTS];
207#define CACHE_CACHE 0
208#define SIZE_NODE (MAX_NUMNODES)
209
210static int drain_freelist(struct kmem_cache *cache,
211 struct kmem_cache_node *n, int tofree);
212static void free_block(struct kmem_cache *cachep, void **objpp, int len,
213 int node, struct list_head *list);
214static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list);
215static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
216static void cache_reap(struct work_struct *unused);
217
218static inline void fixup_objfreelist_debug(struct kmem_cache *cachep,
219 void **list);
220static inline void fixup_slab_list(struct kmem_cache *cachep,
221 struct kmem_cache_node *n, struct slab *slab,
222 void **list);
223static int slab_early_init = 1;
224
225#define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
226
227static void kmem_cache_node_init(struct kmem_cache_node *parent)
228{
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 raw_spin_lock_init(&parent->list_lock);
238 parent->free_objects = 0;
239 parent->free_touched = 0;
240}
241
242#define MAKE_LIST(cachep, listp, slab, nodeid) \
243 do { \
244 INIT_LIST_HEAD(listp); \
245 list_splice(&get_node(cachep, nodeid)->slab, listp); \
246 } while (0)
247
248#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
249 do { \
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); \
253 } while (0)
254
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)
259
260#define BATCHREFILL_LIMIT 16
261/*
262 * Optimization question: fewer reaps means less probability for unnecessary
263 * cpucache drain/refill cycles.
264 *
265 * OTOH the cpuarrays can contain lots of objects,
266 * which could lock up otherwise freeable slabs.
267 */
268#define REAPTIMEOUT_AC (2*HZ)
269#define REAPTIMEOUT_NODE (4*HZ)
270
271#if STATS
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) \
278 do { \
279 if ((x)->num_active > (x)->high_mark) \
280 (x)->high_mark = (x)->num_active; \
281 } while (0)
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) \
287 do { \
288 if ((x)->max_freeable < i) \
289 (x)->max_freeable = i; \
290 } while (0)
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)
295#else
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)
311#endif
312
313#if DEBUG
314
315/*
316 * memory layout of objects:
317 * 0 : objp
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:
322 * redzone word.
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]
327 */
328static int obj_offset(struct kmem_cache *cachep)
329{
330 return cachep->obj_offset;
331}
332
333static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
334{
335 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
336 return (unsigned long long *) (objp + obj_offset(cachep) -
337 sizeof(unsigned long long));
338}
339
340static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
341{
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) -
346 REDZONE_ALIGN);
347 return (unsigned long long *) (objp + cachep->size -
348 sizeof(unsigned long long));
349}
350
351static void **dbg_userword(struct kmem_cache *cachep, void *objp)
352{
353 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
354 return (void **)(objp + cachep->size - BYTES_PER_WORD);
355}
356
357#else
358
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;})
363
364#endif
365
366/*
367 * Do not go above this order unless 0 objects fit into the slab or
368 * overridden on the command line.
369 */
370#define SLAB_MAX_ORDER_HI 1
371#define SLAB_MAX_ORDER_LO 0
372static int slab_max_order = SLAB_MAX_ORDER_LO;
373static bool slab_max_order_set __initdata;
374
375static inline void *index_to_obj(struct kmem_cache *cache,
376 const struct slab *slab, unsigned int idx)
377{
378 return slab->s_mem + cache->size * idx;
379}
380
381#define BOOT_CPUCACHE_ENTRIES 1
382/* internal cache of cache description objs */
383static struct kmem_cache kmem_cache_boot = {
384 .batchcount = 1,
385 .limit = BOOT_CPUCACHE_ENTRIES,
386 .shared = 1,
387 .size = sizeof(struct kmem_cache),
388 .name = "kmem_cache",
389};
390
391static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
392
393static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
394{
395 return this_cpu_ptr(cachep->cpu_cache);
396}
397
398/*
399 * Calculate the number of objects and left-over bytes for a given buffer size.
400 */
401static unsigned int cache_estimate(unsigned long gfporder, size_t buffer_size,
402 slab_flags_t flags, size_t *left_over)
403{
404 unsigned int num;
405 size_t slab_size = PAGE_SIZE << gfporder;
406
407 /*
408 * The slab management structure can be either off the slab or
409 * on it. For the latter case, the memory allocated for a
410 * slab is used for:
411 *
412 * - @buffer_size bytes for each object
413 * - One freelist_idx_t for each object
414 *
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.
418 *
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.
423 */
424 if (flags & (CFLGS_OBJFREELIST_SLAB | CFLGS_OFF_SLAB)) {
425 num = slab_size / buffer_size;
426 *left_over = slab_size % buffer_size;
427 } else {
428 num = slab_size / (buffer_size + sizeof(freelist_idx_t));
429 *left_over = slab_size %
430 (buffer_size + sizeof(freelist_idx_t));
431 }
432
433 return num;
434}
435
436#if DEBUG
437#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
438
439static void __slab_error(const char *function, struct kmem_cache *cachep,
440 char *msg)
441{
442 pr_err("slab error in %s(): cache `%s': %s\n",
443 function, cachep->name, msg);
444 dump_stack();
445 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
446}
447#endif
448
449/*
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
454 * line
455 */
456
457static int use_alien_caches __read_mostly = 1;
458static int __init noaliencache_setup(char *s)
459{
460 use_alien_caches = 0;
461 return 1;
462}
463__setup("noaliencache", noaliencache_setup);
464
465static int __init slab_max_order_setup(char *str)
466{
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;
471
472 return 1;
473}
474__setup("slab_max_order=", slab_max_order_setup);
475
476#ifdef CONFIG_NUMA
477/*
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.
482 */
483static DEFINE_PER_CPU(unsigned long, slab_reap_node);
484
485static void init_reap_node(int cpu)
486{
487 per_cpu(slab_reap_node, cpu) = next_node_in(cpu_to_mem(cpu),
488 node_online_map);
489}
490
491static void next_reap_node(void)
492{
493 int node = __this_cpu_read(slab_reap_node);
494
495 node = next_node_in(node, node_online_map);
496 __this_cpu_write(slab_reap_node, node);
497}
498
499#else
500#define init_reap_node(cpu) do { } while (0)
501#define next_reap_node(void) do { } while (0)
502#endif
503
504/*
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
509 * lock.
510 */
511static void start_cpu_timer(int cpu)
512{
513 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
514
515 if (reap_work->work.func == NULL) {
516 init_reap_node(cpu);
517 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
518 schedule_delayed_work_on(cpu, reap_work,
519 __round_jiffies_relative(HZ, cpu));
520 }
521}
522
523static void init_arraycache(struct array_cache *ac, int limit, int batch)
524{
525 if (ac) {
526 ac->avail = 0;
527 ac->limit = limit;
528 ac->batchcount = batch;
529 ac->touched = 0;
530 }
531}
532
533static struct array_cache *alloc_arraycache(int node, int entries,
534 int batchcount, gfp_t gfp)
535{
536 size_t memsize = sizeof(void *) * entries + sizeof(struct array_cache);
537 struct array_cache *ac = NULL;
538
539 ac = kmalloc_node(memsize, gfp, node);
540 /*
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.
546 */
547 kmemleak_no_scan(ac);
548 init_arraycache(ac, entries, batchcount);
549 return ac;
550}
551
552static noinline void cache_free_pfmemalloc(struct kmem_cache *cachep,
553 struct slab *slab, void *objp)
554{
555 struct kmem_cache_node *n;
556 int slab_node;
557 LIST_HEAD(list);
558
559 slab_node = slab_nid(slab);
560 n = get_node(cachep, slab_node);
561
562 raw_spin_lock(&n->list_lock);
563 free_block(cachep, &objp, 1, slab_node, &list);
564 raw_spin_unlock(&n->list_lock);
565
566 slabs_destroy(cachep, &list);
567}
568
569/*
570 * Transfer objects in one arraycache to another.
571 * Locking must be handled by the caller.
572 *
573 * Return the number of entries transferred.
574 */
575static int transfer_objects(struct array_cache *to,
576 struct array_cache *from, unsigned int max)
577{
578 /* Figure out how many entries to transfer */
579 int nr = min3(from->avail, max, to->limit - to->avail);
580
581 if (!nr)
582 return 0;
583
584 memcpy(to->entry + to->avail, from->entry + from->avail - nr,
585 sizeof(void *) *nr);
586
587 from->avail -= nr;
588 to->avail += nr;
589 return nr;
590}
591
592/* &alien->lock must be held by alien callers. */
593static __always_inline void __free_one(struct array_cache *ac, void *objp)
594{
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))
598 return;
599 ac->entry[ac->avail++] = objp;
600}
601
602#ifndef CONFIG_NUMA
603
604#define drain_alien_cache(cachep, alien) do { } while (0)
605#define reap_alien(cachep, n) do { } while (0)
606
607static inline struct alien_cache **alloc_alien_cache(int node,
608 int limit, gfp_t gfp)
609{
610 return NULL;
611}
612
613static inline void free_alien_cache(struct alien_cache **ac_ptr)
614{
615}
616
617static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
618{
619 return 0;
620}
621
622static inline gfp_t gfp_exact_node(gfp_t flags)
623{
624 return flags & ~__GFP_NOFAIL;
625}
626
627#else /* CONFIG_NUMA */
628
629static struct alien_cache *__alloc_alien_cache(int node, int entries,
630 int batch, gfp_t gfp)
631{
632 size_t memsize = sizeof(void *) * entries + sizeof(struct alien_cache);
633 struct alien_cache *alc = NULL;
634
635 alc = kmalloc_node(memsize, gfp, node);
636 if (alc) {
637 kmemleak_no_scan(alc);
638 init_arraycache(&alc->ac, entries, batch);
639 spin_lock_init(&alc->lock);
640 }
641 return alc;
642}
643
644static struct alien_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
645{
646 struct alien_cache **alc_ptr;
647 int i;
648
649 if (limit > 1)
650 limit = 12;
651 alc_ptr = kcalloc_node(nr_node_ids, sizeof(void *), gfp, node);
652 if (!alc_ptr)
653 return NULL;
654
655 for_each_node(i) {
656 if (i == node || !node_online(i))
657 continue;
658 alc_ptr[i] = __alloc_alien_cache(node, limit, 0xbaadf00d, gfp);
659 if (!alc_ptr[i]) {
660 for (i--; i >= 0; i--)
661 kfree(alc_ptr[i]);
662 kfree(alc_ptr);
663 return NULL;
664 }
665 }
666 return alc_ptr;
667}
668
669static void free_alien_cache(struct alien_cache **alc_ptr)
670{
671 int i;
672
673 if (!alc_ptr)
674 return;
675 for_each_node(i)
676 kfree(alc_ptr[i]);
677 kfree(alc_ptr);
678}
679
680static void __drain_alien_cache(struct kmem_cache *cachep,
681 struct array_cache *ac, int node,
682 struct list_head *list)
683{
684 struct kmem_cache_node *n = get_node(cachep, node);
685
686 if (ac->avail) {
687 raw_spin_lock(&n->list_lock);
688 /*
689 * Stuff objects into the remote nodes shared array first.
690 * That way we could avoid the overhead of putting the objects
691 * into the free lists and getting them back later.
692 */
693 if (n->shared)
694 transfer_objects(n->shared, ac, ac->limit);
695
696 free_block(cachep, ac->entry, ac->avail, node, list);
697 ac->avail = 0;
698 raw_spin_unlock(&n->list_lock);
699 }
700}
701
702/*
703 * Called from cache_reap() to regularly drain alien caches round robin.
704 */
705static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
706{
707 int node = __this_cpu_read(slab_reap_node);
708
709 if (n->alien) {
710 struct alien_cache *alc = n->alien[node];
711 struct array_cache *ac;
712
713 if (alc) {
714 ac = &alc->ac;
715 if (ac->avail && spin_trylock_irq(&alc->lock)) {
716 LIST_HEAD(list);
717
718 __drain_alien_cache(cachep, ac, node, &list);
719 spin_unlock_irq(&alc->lock);
720 slabs_destroy(cachep, &list);
721 }
722 }
723 }
724}
725
726static void drain_alien_cache(struct kmem_cache *cachep,
727 struct alien_cache **alien)
728{
729 int i = 0;
730 struct alien_cache *alc;
731 struct array_cache *ac;
732 unsigned long flags;
733
734 for_each_online_node(i) {
735 alc = alien[i];
736 if (alc) {
737 LIST_HEAD(list);
738
739 ac = &alc->ac;
740 spin_lock_irqsave(&alc->lock, flags);
741 __drain_alien_cache(cachep, ac, i, &list);
742 spin_unlock_irqrestore(&alc->lock, flags);
743 slabs_destroy(cachep, &list);
744 }
745 }
746}
747
748static int __cache_free_alien(struct kmem_cache *cachep, void *objp,
749 int node, int slab_node)
750{
751 struct kmem_cache_node *n;
752 struct alien_cache *alien = NULL;
753 struct array_cache *ac;
754 LIST_HEAD(list);
755
756 n = get_node(cachep, node);
757 STATS_INC_NODEFREES(cachep);
758 if (n->alien && n->alien[slab_node]) {
759 alien = n->alien[slab_node];
760 ac = &alien->ac;
761 spin_lock(&alien->lock);
762 if (unlikely(ac->avail == ac->limit)) {
763 STATS_INC_ACOVERFLOW(cachep);
764 __drain_alien_cache(cachep, ac, slab_node, &list);
765 }
766 __free_one(ac, objp);
767 spin_unlock(&alien->lock);
768 slabs_destroy(cachep, &list);
769 } else {
770 n = get_node(cachep, slab_node);
771 raw_spin_lock(&n->list_lock);
772 free_block(cachep, &objp, 1, slab_node, &list);
773 raw_spin_unlock(&n->list_lock);
774 slabs_destroy(cachep, &list);
775 }
776 return 1;
777}
778
779static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
780{
781 int slab_node = slab_nid(virt_to_slab(objp));
782 int node = numa_mem_id();
783 /*
784 * Make sure we are not freeing an object from another node to the array
785 * cache on this cpu.
786 */
787 if (likely(node == slab_node))
788 return 0;
789
790 return __cache_free_alien(cachep, objp, node, slab_node);
791}
792
793/*
794 * Construct gfp mask to allocate from a specific node but do not reclaim or
795 * warn about failures.
796 */
797static inline gfp_t gfp_exact_node(gfp_t flags)
798{
799 return (flags | __GFP_THISNODE | __GFP_NOWARN) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
800}
801#endif
802
803static int init_cache_node(struct kmem_cache *cachep, int node, gfp_t gfp)
804{
805 struct kmem_cache_node *n;
806
807 /*
808 * Set up the kmem_cache_node for cpu before we can
809 * begin anything. Make sure some other cpu on this
810 * node has not already allocated this
811 */
812 n = get_node(cachep, node);
813 if (n) {
814 raw_spin_lock_irq(&n->list_lock);
815 n->free_limit = (1 + nr_cpus_node(node)) * cachep->batchcount +
816 cachep->num;
817 raw_spin_unlock_irq(&n->list_lock);
818
819 return 0;
820 }
821
822 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
823 if (!n)
824 return -ENOMEM;
825
826 kmem_cache_node_init(n);
827 n->next_reap = jiffies + REAPTIMEOUT_NODE +
828 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
829
830 n->free_limit =
831 (1 + nr_cpus_node(node)) * cachep->batchcount + cachep->num;
832
833 /*
834 * The kmem_cache_nodes don't come and go as CPUs
835 * come and go. slab_mutex provides sufficient
836 * protection here.
837 */
838 cachep->node[node] = n;
839
840 return 0;
841}
842
843#if (defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)) || defined(CONFIG_SMP)
844/*
845 * Allocates and initializes node for a node on each slab cache, used for
846 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
847 * will be allocated off-node since memory is not yet online for the new node.
848 * When hotplugging memory or a cpu, existing nodes are not replaced if
849 * already in use.
850 *
851 * Must hold slab_mutex.
852 */
853static int init_cache_node_node(int node)
854{
855 int ret;
856 struct kmem_cache *cachep;
857
858 list_for_each_entry(cachep, &slab_caches, list) {
859 ret = init_cache_node(cachep, node, GFP_KERNEL);
860 if (ret)
861 return ret;
862 }
863
864 return 0;
865}
866#endif
867
868static int setup_kmem_cache_node(struct kmem_cache *cachep,
869 int node, gfp_t gfp, bool force_change)
870{
871 int ret = -ENOMEM;
872 struct kmem_cache_node *n;
873 struct array_cache *old_shared = NULL;
874 struct array_cache *new_shared = NULL;
875 struct alien_cache **new_alien = NULL;
876 LIST_HEAD(list);
877
878 if (use_alien_caches) {
879 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
880 if (!new_alien)
881 goto fail;
882 }
883
884 if (cachep->shared) {
885 new_shared = alloc_arraycache(node,
886 cachep->shared * cachep->batchcount, 0xbaadf00d, gfp);
887 if (!new_shared)
888 goto fail;
889 }
890
891 ret = init_cache_node(cachep, node, gfp);
892 if (ret)
893 goto fail;
894
895 n = get_node(cachep, node);
896 raw_spin_lock_irq(&n->list_lock);
897 if (n->shared && force_change) {
898 free_block(cachep, n->shared->entry,
899 n->shared->avail, node, &list);
900 n->shared->avail = 0;
901 }
902
903 if (!n->shared || force_change) {
904 old_shared = n->shared;
905 n->shared = new_shared;
906 new_shared = NULL;
907 }
908
909 if (!n->alien) {
910 n->alien = new_alien;
911 new_alien = NULL;
912 }
913
914 raw_spin_unlock_irq(&n->list_lock);
915 slabs_destroy(cachep, &list);
916
917 /*
918 * To protect lockless access to n->shared during irq disabled context.
919 * If n->shared isn't NULL in irq disabled context, accessing to it is
920 * guaranteed to be valid until irq is re-enabled, because it will be
921 * freed after synchronize_rcu().
922 */
923 if (old_shared && force_change)
924 synchronize_rcu();
925
926fail:
927 kfree(old_shared);
928 kfree(new_shared);
929 free_alien_cache(new_alien);
930
931 return ret;
932}
933
934#ifdef CONFIG_SMP
935
936static void cpuup_canceled(long cpu)
937{
938 struct kmem_cache *cachep;
939 struct kmem_cache_node *n = NULL;
940 int node = cpu_to_mem(cpu);
941 const struct cpumask *mask = cpumask_of_node(node);
942
943 list_for_each_entry(cachep, &slab_caches, list) {
944 struct array_cache *nc;
945 struct array_cache *shared;
946 struct alien_cache **alien;
947 LIST_HEAD(list);
948
949 n = get_node(cachep, node);
950 if (!n)
951 continue;
952
953 raw_spin_lock_irq(&n->list_lock);
954
955 /* Free limit for this kmem_cache_node */
956 n->free_limit -= cachep->batchcount;
957
958 /* cpu is dead; no one can alloc from it. */
959 nc = per_cpu_ptr(cachep->cpu_cache, cpu);
960 free_block(cachep, nc->entry, nc->avail, node, &list);
961 nc->avail = 0;
962
963 if (!cpumask_empty(mask)) {
964 raw_spin_unlock_irq(&n->list_lock);
965 goto free_slab;
966 }
967
968 shared = n->shared;
969 if (shared) {
970 free_block(cachep, shared->entry,
971 shared->avail, node, &list);
972 n->shared = NULL;
973 }
974
975 alien = n->alien;
976 n->alien = NULL;
977
978 raw_spin_unlock_irq(&n->list_lock);
979
980 kfree(shared);
981 if (alien) {
982 drain_alien_cache(cachep, alien);
983 free_alien_cache(alien);
984 }
985
986free_slab:
987 slabs_destroy(cachep, &list);
988 }
989 /*
990 * In the previous loop, all the objects were freed to
991 * the respective cache's slabs, now we can go ahead and
992 * shrink each nodelist to its limit.
993 */
994 list_for_each_entry(cachep, &slab_caches, list) {
995 n = get_node(cachep, node);
996 if (!n)
997 continue;
998 drain_freelist(cachep, n, INT_MAX);
999 }
1000}
1001
1002static int cpuup_prepare(long cpu)
1003{
1004 struct kmem_cache *cachep;
1005 int node = cpu_to_mem(cpu);
1006 int err;
1007
1008 /*
1009 * We need to do this right in the beginning since
1010 * alloc_arraycache's are going to use this list.
1011 * kmalloc_node allows us to add the slab to the right
1012 * kmem_cache_node and not this cpu's kmem_cache_node
1013 */
1014 err = init_cache_node_node(node);
1015 if (err < 0)
1016 goto bad;
1017
1018 /*
1019 * Now we can go ahead with allocating the shared arrays and
1020 * array caches
1021 */
1022 list_for_each_entry(cachep, &slab_caches, list) {
1023 err = setup_kmem_cache_node(cachep, node, GFP_KERNEL, false);
1024 if (err)
1025 goto bad;
1026 }
1027
1028 return 0;
1029bad:
1030 cpuup_canceled(cpu);
1031 return -ENOMEM;
1032}
1033
1034int slab_prepare_cpu(unsigned int cpu)
1035{
1036 int err;
1037
1038 mutex_lock(&slab_mutex);
1039 err = cpuup_prepare(cpu);
1040 mutex_unlock(&slab_mutex);
1041 return err;
1042}
1043
1044/*
1045 * This is called for a failed online attempt and for a successful
1046 * offline.
1047 *
1048 * Even if all the cpus of a node are down, we don't free the
1049 * kmem_cache_node of any cache. This is to avoid a race between cpu_down, and
1050 * a kmalloc allocation from another cpu for memory from the node of
1051 * the cpu going down. The kmem_cache_node structure is usually allocated from
1052 * kmem_cache_create() and gets destroyed at kmem_cache_destroy().
1053 */
1054int slab_dead_cpu(unsigned int cpu)
1055{
1056 mutex_lock(&slab_mutex);
1057 cpuup_canceled(cpu);
1058 mutex_unlock(&slab_mutex);
1059 return 0;
1060}
1061#endif
1062
1063static int slab_online_cpu(unsigned int cpu)
1064{
1065 start_cpu_timer(cpu);
1066 return 0;
1067}
1068
1069static int slab_offline_cpu(unsigned int cpu)
1070{
1071 /*
1072 * Shutdown cache reaper. Note that the slab_mutex is held so
1073 * that if cache_reap() is invoked it cannot do anything
1074 * expensive but will only modify reap_work and reschedule the
1075 * timer.
1076 */
1077 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1078 /* Now the cache_reaper is guaranteed to be not running. */
1079 per_cpu(slab_reap_work, cpu).work.func = NULL;
1080 return 0;
1081}
1082
1083#if defined(CONFIG_NUMA)
1084/*
1085 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1086 * Returns -EBUSY if all objects cannot be drained so that the node is not
1087 * removed.
1088 *
1089 * Must hold slab_mutex.
1090 */
1091static int __meminit drain_cache_node_node(int node)
1092{
1093 struct kmem_cache *cachep;
1094 int ret = 0;
1095
1096 list_for_each_entry(cachep, &slab_caches, list) {
1097 struct kmem_cache_node *n;
1098
1099 n = get_node(cachep, node);
1100 if (!n)
1101 continue;
1102
1103 drain_freelist(cachep, n, INT_MAX);
1104
1105 if (!list_empty(&n->slabs_full) ||
1106 !list_empty(&n->slabs_partial)) {
1107 ret = -EBUSY;
1108 break;
1109 }
1110 }
1111 return ret;
1112}
1113
1114static int __meminit slab_memory_callback(struct notifier_block *self,
1115 unsigned long action, void *arg)
1116{
1117 struct memory_notify *mnb = arg;
1118 int ret = 0;
1119 int nid;
1120
1121 nid = mnb->status_change_nid;
1122 if (nid < 0)
1123 goto out;
1124
1125 switch (action) {
1126 case MEM_GOING_ONLINE:
1127 mutex_lock(&slab_mutex);
1128 ret = init_cache_node_node(nid);
1129 mutex_unlock(&slab_mutex);
1130 break;
1131 case MEM_GOING_OFFLINE:
1132 mutex_lock(&slab_mutex);
1133 ret = drain_cache_node_node(nid);
1134 mutex_unlock(&slab_mutex);
1135 break;
1136 case MEM_ONLINE:
1137 case MEM_OFFLINE:
1138 case MEM_CANCEL_ONLINE:
1139 case MEM_CANCEL_OFFLINE:
1140 break;
1141 }
1142out:
1143 return notifier_from_errno(ret);
1144}
1145#endif /* CONFIG_NUMA */
1146
1147/*
1148 * swap the static kmem_cache_node with kmalloced memory
1149 */
1150static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
1151 int nodeid)
1152{
1153 struct kmem_cache_node *ptr;
1154
1155 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
1156 BUG_ON(!ptr);
1157
1158 memcpy(ptr, list, sizeof(struct kmem_cache_node));
1159 /*
1160 * Do not assume that spinlocks can be initialized via memcpy:
1161 */
1162 raw_spin_lock_init(&ptr->list_lock);
1163
1164 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1165 cachep->node[nodeid] = ptr;
1166}
1167
1168/*
1169 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1170 * size of kmem_cache_node.
1171 */
1172static void __init set_up_node(struct kmem_cache *cachep, int index)
1173{
1174 int node;
1175
1176 for_each_online_node(node) {
1177 cachep->node[node] = &init_kmem_cache_node[index + node];
1178 cachep->node[node]->next_reap = jiffies +
1179 REAPTIMEOUT_NODE +
1180 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1181 }
1182}
1183
1184/*
1185 * Initialisation. Called after the page allocator have been initialised and
1186 * before smp_init().
1187 */
1188void __init kmem_cache_init(void)
1189{
1190 int i;
1191
1192 kmem_cache = &kmem_cache_boot;
1193
1194 if (!IS_ENABLED(CONFIG_NUMA) || num_possible_nodes() == 1)
1195 use_alien_caches = 0;
1196
1197 for (i = 0; i < NUM_INIT_LISTS; i++)
1198 kmem_cache_node_init(&init_kmem_cache_node[i]);
1199
1200 /*
1201 * Fragmentation resistance on low memory - only use bigger
1202 * page orders on machines with more than 32MB of memory if
1203 * not overridden on the command line.
1204 */
1205 if (!slab_max_order_set && totalram_pages() > (32 << 20) >> PAGE_SHIFT)
1206 slab_max_order = SLAB_MAX_ORDER_HI;
1207
1208 /* Bootstrap is tricky, because several objects are allocated
1209 * from caches that do not exist yet:
1210 * 1) initialize the kmem_cache cache: it contains the struct
1211 * kmem_cache structures of all caches, except kmem_cache itself:
1212 * kmem_cache is statically allocated.
1213 * Initially an __init data area is used for the head array and the
1214 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1215 * array at the end of the bootstrap.
1216 * 2) Create the first kmalloc cache.
1217 * The struct kmem_cache for the new cache is allocated normally.
1218 * An __init data area is used for the head array.
1219 * 3) Create the remaining kmalloc caches, with minimally sized
1220 * head arrays.
1221 * 4) Replace the __init data head arrays for kmem_cache and the first
1222 * kmalloc cache with kmalloc allocated arrays.
1223 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1224 * the other cache's with kmalloc allocated memory.
1225 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1226 */
1227
1228 /* 1) create the kmem_cache */
1229
1230 /*
1231 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1232 */
1233 create_boot_cache(kmem_cache, "kmem_cache",
1234 offsetof(struct kmem_cache, node) +
1235 nr_node_ids * sizeof(struct kmem_cache_node *),
1236 SLAB_HWCACHE_ALIGN, 0, 0);
1237 list_add(&kmem_cache->list, &slab_caches);
1238 slab_state = PARTIAL;
1239
1240 /*
1241 * Initialize the caches that provide memory for the kmem_cache_node
1242 * structures first. Without this, further allocations will bug.
1243 */
1244 kmalloc_caches[KMALLOC_NORMAL][INDEX_NODE] = create_kmalloc_cache(
1245 kmalloc_info[INDEX_NODE].name[KMALLOC_NORMAL],
1246 kmalloc_info[INDEX_NODE].size,
1247 ARCH_KMALLOC_FLAGS, 0,
1248 kmalloc_info[INDEX_NODE].size);
1249 slab_state = PARTIAL_NODE;
1250 setup_kmalloc_cache_index_table();
1251
1252 slab_early_init = 0;
1253
1254 /* 5) Replace the bootstrap kmem_cache_node */
1255 {
1256 int nid;
1257
1258 for_each_online_node(nid) {
1259 init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
1260
1261 init_list(kmalloc_caches[KMALLOC_NORMAL][INDEX_NODE],
1262 &init_kmem_cache_node[SIZE_NODE + nid], nid);
1263 }
1264 }
1265
1266 create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
1267}
1268
1269void __init kmem_cache_init_late(void)
1270{
1271 struct kmem_cache *cachep;
1272
1273 /* 6) resize the head arrays to their final sizes */
1274 mutex_lock(&slab_mutex);
1275 list_for_each_entry(cachep, &slab_caches, list)
1276 if (enable_cpucache(cachep, GFP_NOWAIT))
1277 BUG();
1278 mutex_unlock(&slab_mutex);
1279
1280 /* Done! */
1281 slab_state = FULL;
1282
1283#ifdef CONFIG_NUMA
1284 /*
1285 * Register a memory hotplug callback that initializes and frees
1286 * node.
1287 */
1288 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1289#endif
1290
1291 /*
1292 * The reap timers are started later, with a module init call: That part
1293 * of the kernel is not yet operational.
1294 */
1295}
1296
1297static int __init cpucache_init(void)
1298{
1299 int ret;
1300
1301 /*
1302 * Register the timers that return unneeded pages to the page allocator
1303 */
1304 ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "SLAB online",
1305 slab_online_cpu, slab_offline_cpu);
1306 WARN_ON(ret < 0);
1307
1308 return 0;
1309}
1310__initcall(cpucache_init);
1311
1312static noinline void
1313slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1314{
1315#if DEBUG
1316 struct kmem_cache_node *n;
1317 unsigned long flags;
1318 int node;
1319 static DEFINE_RATELIMIT_STATE(slab_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1320 DEFAULT_RATELIMIT_BURST);
1321
1322 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slab_oom_rs))
1323 return;
1324
1325 pr_warn("SLAB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
1326 nodeid, gfpflags, &gfpflags);
1327 pr_warn(" cache: %s, object size: %d, order: %d\n",
1328 cachep->name, cachep->size, cachep->gfporder);
1329
1330 for_each_kmem_cache_node(cachep, node, n) {
1331 unsigned long total_slabs, free_slabs, free_objs;
1332
1333 raw_spin_lock_irqsave(&n->list_lock, flags);
1334 total_slabs = n->total_slabs;
1335 free_slabs = n->free_slabs;
1336 free_objs = n->free_objects;
1337 raw_spin_unlock_irqrestore(&n->list_lock, flags);
1338
1339 pr_warn(" node %d: slabs: %ld/%ld, objs: %ld/%ld\n",
1340 node, total_slabs - free_slabs, total_slabs,
1341 (total_slabs * cachep->num) - free_objs,
1342 total_slabs * cachep->num);
1343 }
1344#endif
1345}
1346
1347/*
1348 * Interface to system's page allocator. No need to hold the
1349 * kmem_cache_node ->list_lock.
1350 *
1351 * If we requested dmaable memory, we will get it. Even if we
1352 * did not request dmaable memory, we might get it, but that
1353 * would be relatively rare and ignorable.
1354 */
1355static struct slab *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1356 int nodeid)
1357{
1358 struct folio *folio;
1359 struct slab *slab;
1360
1361 flags |= cachep->allocflags;
1362
1363 folio = (struct folio *) __alloc_pages_node(nodeid, flags, cachep->gfporder);
1364 if (!folio) {
1365 slab_out_of_memory(cachep, flags, nodeid);
1366 return NULL;
1367 }
1368
1369 slab = folio_slab(folio);
1370
1371 account_slab(slab, cachep->gfporder, cachep, flags);
1372 __folio_set_slab(folio);
1373 /* Make the flag visible before any changes to folio->mapping */
1374 smp_wmb();
1375 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1376 if (sk_memalloc_socks() && page_is_pfmemalloc(folio_page(folio, 0)))
1377 slab_set_pfmemalloc(slab);
1378
1379 return slab;
1380}
1381
1382/*
1383 * Interface to system's page release.
1384 */
1385static void kmem_freepages(struct kmem_cache *cachep, struct slab *slab)
1386{
1387 int order = cachep->gfporder;
1388 struct folio *folio = slab_folio(slab);
1389
1390 BUG_ON(!folio_test_slab(folio));
1391 __slab_clear_pfmemalloc(slab);
1392 page_mapcount_reset(folio_page(folio, 0));
1393 folio->mapping = NULL;
1394 /* Make the mapping reset visible before clearing the flag */
1395 smp_wmb();
1396 __folio_clear_slab(folio);
1397
1398 if (current->reclaim_state)
1399 current->reclaim_state->reclaimed_slab += 1 << order;
1400 unaccount_slab(slab, order, cachep);
1401 __free_pages(folio_page(folio, 0), order);
1402}
1403
1404static void kmem_rcu_free(struct rcu_head *head)
1405{
1406 struct kmem_cache *cachep;
1407 struct slab *slab;
1408
1409 slab = container_of(head, struct slab, rcu_head);
1410 cachep = slab->slab_cache;
1411
1412 kmem_freepages(cachep, slab);
1413}
1414
1415#if DEBUG
1416static bool is_debug_pagealloc_cache(struct kmem_cache *cachep)
1417{
1418 if (debug_pagealloc_enabled_static() && OFF_SLAB(cachep) &&
1419 (cachep->size % PAGE_SIZE) == 0)
1420 return true;
1421
1422 return false;
1423}
1424
1425#ifdef CONFIG_DEBUG_PAGEALLOC
1426static void slab_kernel_map(struct kmem_cache *cachep, void *objp, int map)
1427{
1428 if (!is_debug_pagealloc_cache(cachep))
1429 return;
1430
1431 __kernel_map_pages(virt_to_page(objp), cachep->size / PAGE_SIZE, map);
1432}
1433
1434#else
1435static inline void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1436 int map) {}
1437
1438#endif
1439
1440static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1441{
1442 int size = cachep->object_size;
1443 addr = &((char *)addr)[obj_offset(cachep)];
1444
1445 memset(addr, val, size);
1446 *(unsigned char *)(addr + size - 1) = POISON_END;
1447}
1448
1449static void dump_line(char *data, int offset, int limit)
1450{
1451 int i;
1452 unsigned char error = 0;
1453 int bad_count = 0;
1454
1455 pr_err("%03x: ", offset);
1456 for (i = 0; i < limit; i++) {
1457 if (data[offset + i] != POISON_FREE) {
1458 error = data[offset + i];
1459 bad_count++;
1460 }
1461 }
1462 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1463 &data[offset], limit, 1);
1464
1465 if (bad_count == 1) {
1466 error ^= POISON_FREE;
1467 if (!(error & (error - 1))) {
1468 pr_err("Single bit error detected. Probably bad RAM.\n");
1469#ifdef CONFIG_X86
1470 pr_err("Run memtest86+ or a similar memory test tool.\n");
1471#else
1472 pr_err("Run a memory test tool.\n");
1473#endif
1474 }
1475 }
1476}
1477#endif
1478
1479#if DEBUG
1480
1481static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1482{
1483 int i, size;
1484 char *realobj;
1485
1486 if (cachep->flags & SLAB_RED_ZONE) {
1487 pr_err("Redzone: 0x%llx/0x%llx\n",
1488 *dbg_redzone1(cachep, objp),
1489 *dbg_redzone2(cachep, objp));
1490 }
1491
1492 if (cachep->flags & SLAB_STORE_USER)
1493 pr_err("Last user: (%pSR)\n", *dbg_userword(cachep, objp));
1494 realobj = (char *)objp + obj_offset(cachep);
1495 size = cachep->object_size;
1496 for (i = 0; i < size && lines; i += 16, lines--) {
1497 int limit;
1498 limit = 16;
1499 if (i + limit > size)
1500 limit = size - i;
1501 dump_line(realobj, i, limit);
1502 }
1503}
1504
1505static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1506{
1507 char *realobj;
1508 int size, i;
1509 int lines = 0;
1510
1511 if (is_debug_pagealloc_cache(cachep))
1512 return;
1513
1514 realobj = (char *)objp + obj_offset(cachep);
1515 size = cachep->object_size;
1516
1517 for (i = 0; i < size; i++) {
1518 char exp = POISON_FREE;
1519 if (i == size - 1)
1520 exp = POISON_END;
1521 if (realobj[i] != exp) {
1522 int limit;
1523 /* Mismatch ! */
1524 /* Print header */
1525 if (lines == 0) {
1526 pr_err("Slab corruption (%s): %s start=%px, len=%d\n",
1527 print_tainted(), cachep->name,
1528 realobj, size);
1529 print_objinfo(cachep, objp, 0);
1530 }
1531 /* Hexdump the affected line */
1532 i = (i / 16) * 16;
1533 limit = 16;
1534 if (i + limit > size)
1535 limit = size - i;
1536 dump_line(realobj, i, limit);
1537 i += 16;
1538 lines++;
1539 /* Limit to 5 lines */
1540 if (lines > 5)
1541 break;
1542 }
1543 }
1544 if (lines != 0) {
1545 /* Print some data about the neighboring objects, if they
1546 * exist:
1547 */
1548 struct slab *slab = virt_to_slab(objp);
1549 unsigned int objnr;
1550
1551 objnr = obj_to_index(cachep, slab, objp);
1552 if (objnr) {
1553 objp = index_to_obj(cachep, slab, objnr - 1);
1554 realobj = (char *)objp + obj_offset(cachep);
1555 pr_err("Prev obj: start=%px, len=%d\n", realobj, size);
1556 print_objinfo(cachep, objp, 2);
1557 }
1558 if (objnr + 1 < cachep->num) {
1559 objp = index_to_obj(cachep, slab, objnr + 1);
1560 realobj = (char *)objp + obj_offset(cachep);
1561 pr_err("Next obj: start=%px, len=%d\n", realobj, size);
1562 print_objinfo(cachep, objp, 2);
1563 }
1564 }
1565}
1566#endif
1567
1568#if DEBUG
1569static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1570 struct slab *slab)
1571{
1572 int i;
1573
1574 if (OBJFREELIST_SLAB(cachep) && cachep->flags & SLAB_POISON) {
1575 poison_obj(cachep, slab->freelist - obj_offset(cachep),
1576 POISON_FREE);
1577 }
1578
1579 for (i = 0; i < cachep->num; i++) {
1580 void *objp = index_to_obj(cachep, slab, i);
1581
1582 if (cachep->flags & SLAB_POISON) {
1583 check_poison_obj(cachep, objp);
1584 slab_kernel_map(cachep, objp, 1);
1585 }
1586 if (cachep->flags & SLAB_RED_ZONE) {
1587 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1588 slab_error(cachep, "start of a freed object was overwritten");
1589 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1590 slab_error(cachep, "end of a freed object was overwritten");
1591 }
1592 }
1593}
1594#else
1595static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1596 struct slab *slab)
1597{
1598}
1599#endif
1600
1601/**
1602 * slab_destroy - destroy and release all objects in a slab
1603 * @cachep: cache pointer being destroyed
1604 * @slab: slab being destroyed
1605 *
1606 * Destroy all the objs in a slab, and release the mem back to the system.
1607 * Before calling the slab must have been unlinked from the cache. The
1608 * kmem_cache_node ->list_lock is not held/needed.
1609 */
1610static void slab_destroy(struct kmem_cache *cachep, struct slab *slab)
1611{
1612 void *freelist;
1613
1614 freelist = slab->freelist;
1615 slab_destroy_debugcheck(cachep, slab);
1616 if (unlikely(cachep->flags & SLAB_TYPESAFE_BY_RCU))
1617 call_rcu(&slab->rcu_head, kmem_rcu_free);
1618 else
1619 kmem_freepages(cachep, slab);
1620
1621 /*
1622 * From now on, we don't use freelist
1623 * although actual page can be freed in rcu context
1624 */
1625 if (OFF_SLAB(cachep))
1626 kfree(freelist);
1627}
1628
1629/*
1630 * Update the size of the caches before calling slabs_destroy as it may
1631 * recursively call kfree.
1632 */
1633static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list)
1634{
1635 struct slab *slab, *n;
1636
1637 list_for_each_entry_safe(slab, n, list, slab_list) {
1638 list_del(&slab->slab_list);
1639 slab_destroy(cachep, slab);
1640 }
1641}
1642
1643/**
1644 * calculate_slab_order - calculate size (page order) of slabs
1645 * @cachep: pointer to the cache that is being created
1646 * @size: size of objects to be created in this cache.
1647 * @flags: slab allocation flags
1648 *
1649 * Also calculates the number of objects per slab.
1650 *
1651 * This could be made much more intelligent. For now, try to avoid using
1652 * high order pages for slabs. When the gfp() functions are more friendly
1653 * towards high-order requests, this should be changed.
1654 *
1655 * Return: number of left-over bytes in a slab
1656 */
1657static size_t calculate_slab_order(struct kmem_cache *cachep,
1658 size_t size, slab_flags_t flags)
1659{
1660 size_t left_over = 0;
1661 int gfporder;
1662
1663 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
1664 unsigned int num;
1665 size_t remainder;
1666
1667 num = cache_estimate(gfporder, size, flags, &remainder);
1668 if (!num)
1669 continue;
1670
1671 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1672 if (num > SLAB_OBJ_MAX_NUM)
1673 break;
1674
1675 if (flags & CFLGS_OFF_SLAB) {
1676 struct kmem_cache *freelist_cache;
1677 size_t freelist_size;
1678 size_t freelist_cache_size;
1679
1680 freelist_size = num * sizeof(freelist_idx_t);
1681 if (freelist_size > KMALLOC_MAX_CACHE_SIZE) {
1682 freelist_cache_size = PAGE_SIZE << get_order(freelist_size);
1683 } else {
1684 freelist_cache = kmalloc_slab(freelist_size, 0u);
1685 if (!freelist_cache)
1686 continue;
1687 freelist_cache_size = freelist_cache->size;
1688
1689 /*
1690 * Needed to avoid possible looping condition
1691 * in cache_grow_begin()
1692 */
1693 if (OFF_SLAB(freelist_cache))
1694 continue;
1695 }
1696
1697 /* check if off slab has enough benefit */
1698 if (freelist_cache_size > cachep->size / 2)
1699 continue;
1700 }
1701
1702 /* Found something acceptable - save it away */
1703 cachep->num = num;
1704 cachep->gfporder = gfporder;
1705 left_over = remainder;
1706
1707 /*
1708 * A VFS-reclaimable slab tends to have most allocations
1709 * as GFP_NOFS and we really don't want to have to be allocating
1710 * higher-order pages when we are unable to shrink dcache.
1711 */
1712 if (flags & SLAB_RECLAIM_ACCOUNT)
1713 break;
1714
1715 /*
1716 * Large number of objects is good, but very large slabs are
1717 * currently bad for the gfp()s.
1718 */
1719 if (gfporder >= slab_max_order)
1720 break;
1721
1722 /*
1723 * Acceptable internal fragmentation?
1724 */
1725 if (left_over * 8 <= (PAGE_SIZE << gfporder))
1726 break;
1727 }
1728 return left_over;
1729}
1730
1731static struct array_cache __percpu *alloc_kmem_cache_cpus(
1732 struct kmem_cache *cachep, int entries, int batchcount)
1733{
1734 int cpu;
1735 size_t size;
1736 struct array_cache __percpu *cpu_cache;
1737
1738 size = sizeof(void *) * entries + sizeof(struct array_cache);
1739 cpu_cache = __alloc_percpu(size, sizeof(void *));
1740
1741 if (!cpu_cache)
1742 return NULL;
1743
1744 for_each_possible_cpu(cpu) {
1745 init_arraycache(per_cpu_ptr(cpu_cache, cpu),
1746 entries, batchcount);
1747 }
1748
1749 return cpu_cache;
1750}
1751
1752static int __ref setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
1753{
1754 if (slab_state >= FULL)
1755 return enable_cpucache(cachep, gfp);
1756
1757 cachep->cpu_cache = alloc_kmem_cache_cpus(cachep, 1, 1);
1758 if (!cachep->cpu_cache)
1759 return 1;
1760
1761 if (slab_state == DOWN) {
1762 /* Creation of first cache (kmem_cache). */
1763 set_up_node(kmem_cache, CACHE_CACHE);
1764 } else if (slab_state == PARTIAL) {
1765 /* For kmem_cache_node */
1766 set_up_node(cachep, SIZE_NODE);
1767 } else {
1768 int node;
1769
1770 for_each_online_node(node) {
1771 cachep->node[node] = kmalloc_node(
1772 sizeof(struct kmem_cache_node), gfp, node);
1773 BUG_ON(!cachep->node[node]);
1774 kmem_cache_node_init(cachep->node[node]);
1775 }
1776 }
1777
1778 cachep->node[numa_mem_id()]->next_reap =
1779 jiffies + REAPTIMEOUT_NODE +
1780 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1781
1782 cpu_cache_get(cachep)->avail = 0;
1783 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1784 cpu_cache_get(cachep)->batchcount = 1;
1785 cpu_cache_get(cachep)->touched = 0;
1786 cachep->batchcount = 1;
1787 cachep->limit = BOOT_CPUCACHE_ENTRIES;
1788 return 0;
1789}
1790
1791slab_flags_t kmem_cache_flags(unsigned int object_size,
1792 slab_flags_t flags, const char *name)
1793{
1794 return flags;
1795}
1796
1797struct kmem_cache *
1798__kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
1799 slab_flags_t flags, void (*ctor)(void *))
1800{
1801 struct kmem_cache *cachep;
1802
1803 cachep = find_mergeable(size, align, flags, name, ctor);
1804 if (cachep) {
1805 cachep->refcount++;
1806
1807 /*
1808 * Adjust the object sizes so that we clear
1809 * the complete object on kzalloc.
1810 */
1811 cachep->object_size = max_t(int, cachep->object_size, size);
1812 }
1813 return cachep;
1814}
1815
1816static bool set_objfreelist_slab_cache(struct kmem_cache *cachep,
1817 size_t size, slab_flags_t flags)
1818{
1819 size_t left;
1820
1821 cachep->num = 0;
1822
1823 /*
1824 * If slab auto-initialization on free is enabled, store the freelist
1825 * off-slab, so that its contents don't end up in one of the allocated
1826 * objects.
1827 */
1828 if (unlikely(slab_want_init_on_free(cachep)))
1829 return false;
1830
1831 if (cachep->ctor || flags & SLAB_TYPESAFE_BY_RCU)
1832 return false;
1833
1834 left = calculate_slab_order(cachep, size,
1835 flags | CFLGS_OBJFREELIST_SLAB);
1836 if (!cachep->num)
1837 return false;
1838
1839 if (cachep->num * sizeof(freelist_idx_t) > cachep->object_size)
1840 return false;
1841
1842 cachep->colour = left / cachep->colour_off;
1843
1844 return true;
1845}
1846
1847static bool set_off_slab_cache(struct kmem_cache *cachep,
1848 size_t size, slab_flags_t flags)
1849{
1850 size_t left;
1851
1852 cachep->num = 0;
1853
1854 /*
1855 * Always use on-slab management when SLAB_NOLEAKTRACE
1856 * to avoid recursive calls into kmemleak.
1857 */
1858 if (flags & SLAB_NOLEAKTRACE)
1859 return false;
1860
1861 /*
1862 * Size is large, assume best to place the slab management obj
1863 * off-slab (should allow better packing of objs).
1864 */
1865 left = calculate_slab_order(cachep, size, flags | CFLGS_OFF_SLAB);
1866 if (!cachep->num)
1867 return false;
1868
1869 /*
1870 * If the slab has been placed off-slab, and we have enough space then
1871 * move it on-slab. This is at the expense of any extra colouring.
1872 */
1873 if (left >= cachep->num * sizeof(freelist_idx_t))
1874 return false;
1875
1876 cachep->colour = left / cachep->colour_off;
1877
1878 return true;
1879}
1880
1881static bool set_on_slab_cache(struct kmem_cache *cachep,
1882 size_t size, slab_flags_t flags)
1883{
1884 size_t left;
1885
1886 cachep->num = 0;
1887
1888 left = calculate_slab_order(cachep, size, flags);
1889 if (!cachep->num)
1890 return false;
1891
1892 cachep->colour = left / cachep->colour_off;
1893
1894 return true;
1895}
1896
1897/**
1898 * __kmem_cache_create - Create a cache.
1899 * @cachep: cache management descriptor
1900 * @flags: SLAB flags
1901 *
1902 * Returns a ptr to the cache on success, NULL on failure.
1903 * Cannot be called within an int, but can be interrupted.
1904 * The @ctor is run when new pages are allocated by the cache.
1905 *
1906 * The flags are
1907 *
1908 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1909 * to catch references to uninitialised memory.
1910 *
1911 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1912 * for buffer overruns.
1913 *
1914 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1915 * cacheline. This can be beneficial if you're counting cycles as closely
1916 * as davem.
1917 *
1918 * Return: a pointer to the created cache or %NULL in case of error
1919 */
1920int __kmem_cache_create(struct kmem_cache *cachep, slab_flags_t flags)
1921{
1922 size_t ralign = BYTES_PER_WORD;
1923 gfp_t gfp;
1924 int err;
1925 unsigned int size = cachep->size;
1926
1927#if DEBUG
1928#if FORCED_DEBUG
1929 /*
1930 * Enable redzoning and last user accounting, except for caches with
1931 * large objects, if the increased size would increase the object size
1932 * above the next power of two: caches with object sizes just above a
1933 * power of two have a significant amount of internal fragmentation.
1934 */
1935 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
1936 2 * sizeof(unsigned long long)))
1937 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
1938 if (!(flags & SLAB_TYPESAFE_BY_RCU))
1939 flags |= SLAB_POISON;
1940#endif
1941#endif
1942
1943 /*
1944 * Check that size is in terms of words. This is needed to avoid
1945 * unaligned accesses for some archs when redzoning is used, and makes
1946 * sure any on-slab bufctl's are also correctly aligned.
1947 */
1948 size = ALIGN(size, BYTES_PER_WORD);
1949
1950 if (flags & SLAB_RED_ZONE) {
1951 ralign = REDZONE_ALIGN;
1952 /* If redzoning, ensure that the second redzone is suitably
1953 * aligned, by adjusting the object size accordingly. */
1954 size = ALIGN(size, REDZONE_ALIGN);
1955 }
1956
1957 /* 3) caller mandated alignment */
1958 if (ralign < cachep->align) {
1959 ralign = cachep->align;
1960 }
1961 /* disable debug if necessary */
1962 if (ralign > __alignof__(unsigned long long))
1963 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
1964 /*
1965 * 4) Store it.
1966 */
1967 cachep->align = ralign;
1968 cachep->colour_off = cache_line_size();
1969 /* Offset must be a multiple of the alignment. */
1970 if (cachep->colour_off < cachep->align)
1971 cachep->colour_off = cachep->align;
1972
1973 if (slab_is_available())
1974 gfp = GFP_KERNEL;
1975 else
1976 gfp = GFP_NOWAIT;
1977
1978#if DEBUG
1979
1980 /*
1981 * Both debugging options require word-alignment which is calculated
1982 * into align above.
1983 */
1984 if (flags & SLAB_RED_ZONE) {
1985 /* add space for red zone words */
1986 cachep->obj_offset += sizeof(unsigned long long);
1987 size += 2 * sizeof(unsigned long long);
1988 }
1989 if (flags & SLAB_STORE_USER) {
1990 /* user store requires one word storage behind the end of
1991 * the real object. But if the second red zone needs to be
1992 * aligned to 64 bits, we must allow that much space.
1993 */
1994 if (flags & SLAB_RED_ZONE)
1995 size += REDZONE_ALIGN;
1996 else
1997 size += BYTES_PER_WORD;
1998 }
1999#endif
2000
2001 kasan_cache_create(cachep, &size, &flags);
2002
2003 size = ALIGN(size, cachep->align);
2004 /*
2005 * We should restrict the number of objects in a slab to implement
2006 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2007 */
2008 if (FREELIST_BYTE_INDEX && size < SLAB_OBJ_MIN_SIZE)
2009 size = ALIGN(SLAB_OBJ_MIN_SIZE, cachep->align);
2010
2011#if DEBUG
2012 /*
2013 * To activate debug pagealloc, off-slab management is necessary
2014 * requirement. In early phase of initialization, small sized slab
2015 * doesn't get initialized so it would not be possible. So, we need
2016 * to check size >= 256. It guarantees that all necessary small
2017 * sized slab is initialized in current slab initialization sequence.
2018 */
2019 if (debug_pagealloc_enabled_static() && (flags & SLAB_POISON) &&
2020 size >= 256 && cachep->object_size > cache_line_size()) {
2021 if (size < PAGE_SIZE || size % PAGE_SIZE == 0) {
2022 size_t tmp_size = ALIGN(size, PAGE_SIZE);
2023
2024 if (set_off_slab_cache(cachep, tmp_size, flags)) {
2025 flags |= CFLGS_OFF_SLAB;
2026 cachep->obj_offset += tmp_size - size;
2027 size = tmp_size;
2028 goto done;
2029 }
2030 }
2031 }
2032#endif
2033
2034 if (set_objfreelist_slab_cache(cachep, size, flags)) {
2035 flags |= CFLGS_OBJFREELIST_SLAB;
2036 goto done;
2037 }
2038
2039 if (set_off_slab_cache(cachep, size, flags)) {
2040 flags |= CFLGS_OFF_SLAB;
2041 goto done;
2042 }
2043
2044 if (set_on_slab_cache(cachep, size, flags))
2045 goto done;
2046
2047 return -E2BIG;
2048
2049done:
2050 cachep->freelist_size = cachep->num * sizeof(freelist_idx_t);
2051 cachep->flags = flags;
2052 cachep->allocflags = __GFP_COMP;
2053 if (flags & SLAB_CACHE_DMA)
2054 cachep->allocflags |= GFP_DMA;
2055 if (flags & SLAB_CACHE_DMA32)
2056 cachep->allocflags |= GFP_DMA32;
2057 if (flags & SLAB_RECLAIM_ACCOUNT)
2058 cachep->allocflags |= __GFP_RECLAIMABLE;
2059 cachep->size = size;
2060 cachep->reciprocal_buffer_size = reciprocal_value(size);
2061
2062#if DEBUG
2063 /*
2064 * If we're going to use the generic kernel_map_pages()
2065 * poisoning, then it's going to smash the contents of
2066 * the redzone and userword anyhow, so switch them off.
2067 */
2068 if (IS_ENABLED(CONFIG_PAGE_POISONING) &&
2069 (cachep->flags & SLAB_POISON) &&
2070 is_debug_pagealloc_cache(cachep))
2071 cachep->flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2072#endif
2073
2074 err = setup_cpu_cache(cachep, gfp);
2075 if (err) {
2076 __kmem_cache_release(cachep);
2077 return err;
2078 }
2079
2080 return 0;
2081}
2082
2083#if DEBUG
2084static void check_irq_off(void)
2085{
2086 BUG_ON(!irqs_disabled());
2087}
2088
2089static void check_irq_on(void)
2090{
2091 BUG_ON(irqs_disabled());
2092}
2093
2094static void check_mutex_acquired(void)
2095{
2096 BUG_ON(!mutex_is_locked(&slab_mutex));
2097}
2098
2099static void check_spinlock_acquired(struct kmem_cache *cachep)
2100{
2101#ifdef CONFIG_SMP
2102 check_irq_off();
2103 assert_raw_spin_locked(&get_node(cachep, numa_mem_id())->list_lock);
2104#endif
2105}
2106
2107static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2108{
2109#ifdef CONFIG_SMP
2110 check_irq_off();
2111 assert_raw_spin_locked(&get_node(cachep, node)->list_lock);
2112#endif
2113}
2114
2115#else
2116#define check_irq_off() do { } while(0)
2117#define check_irq_on() do { } while(0)
2118#define check_mutex_acquired() do { } while(0)
2119#define check_spinlock_acquired(x) do { } while(0)
2120#define check_spinlock_acquired_node(x, y) do { } while(0)
2121#endif
2122
2123static void drain_array_locked(struct kmem_cache *cachep, struct array_cache *ac,
2124 int node, bool free_all, struct list_head *list)
2125{
2126 int tofree;
2127
2128 if (!ac || !ac->avail)
2129 return;
2130
2131 tofree = free_all ? ac->avail : (ac->limit + 4) / 5;
2132 if (tofree > ac->avail)
2133 tofree = (ac->avail + 1) / 2;
2134
2135 free_block(cachep, ac->entry, tofree, node, list);
2136 ac->avail -= tofree;
2137 memmove(ac->entry, &(ac->entry[tofree]), sizeof(void *) * ac->avail);
2138}
2139
2140static void do_drain(void *arg)
2141{
2142 struct kmem_cache *cachep = arg;
2143 struct array_cache *ac;
2144 int node = numa_mem_id();
2145 struct kmem_cache_node *n;
2146 LIST_HEAD(list);
2147
2148 check_irq_off();
2149 ac = cpu_cache_get(cachep);
2150 n = get_node(cachep, node);
2151 raw_spin_lock(&n->list_lock);
2152 free_block(cachep, ac->entry, ac->avail, node, &list);
2153 raw_spin_unlock(&n->list_lock);
2154 ac->avail = 0;
2155 slabs_destroy(cachep, &list);
2156}
2157
2158static void drain_cpu_caches(struct kmem_cache *cachep)
2159{
2160 struct kmem_cache_node *n;
2161 int node;
2162 LIST_HEAD(list);
2163
2164 on_each_cpu(do_drain, cachep, 1);
2165 check_irq_on();
2166 for_each_kmem_cache_node(cachep, node, n)
2167 if (n->alien)
2168 drain_alien_cache(cachep, n->alien);
2169
2170 for_each_kmem_cache_node(cachep, node, n) {
2171 raw_spin_lock_irq(&n->list_lock);
2172 drain_array_locked(cachep, n->shared, node, true, &list);
2173 raw_spin_unlock_irq(&n->list_lock);
2174
2175 slabs_destroy(cachep, &list);
2176 }
2177}
2178
2179/*
2180 * Remove slabs from the list of free slabs.
2181 * Specify the number of slabs to drain in tofree.
2182 *
2183 * Returns the actual number of slabs released.
2184 */
2185static int drain_freelist(struct kmem_cache *cache,
2186 struct kmem_cache_node *n, int tofree)
2187{
2188 struct list_head *p;
2189 int nr_freed;
2190 struct slab *slab;
2191
2192 nr_freed = 0;
2193 while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
2194
2195 raw_spin_lock_irq(&n->list_lock);
2196 p = n->slabs_free.prev;
2197 if (p == &n->slabs_free) {
2198 raw_spin_unlock_irq(&n->list_lock);
2199 goto out;
2200 }
2201
2202 slab = list_entry(p, struct slab, slab_list);
2203 list_del(&slab->slab_list);
2204 n->free_slabs--;
2205 n->total_slabs--;
2206 /*
2207 * Safe to drop the lock. The slab is no longer linked
2208 * to the cache.
2209 */
2210 n->free_objects -= cache->num;
2211 raw_spin_unlock_irq(&n->list_lock);
2212 slab_destroy(cache, slab);
2213 nr_freed++;
2214
2215 cond_resched();
2216 }
2217out:
2218 return nr_freed;
2219}
2220
2221bool __kmem_cache_empty(struct kmem_cache *s)
2222{
2223 int node;
2224 struct kmem_cache_node *n;
2225
2226 for_each_kmem_cache_node(s, node, n)
2227 if (!list_empty(&n->slabs_full) ||
2228 !list_empty(&n->slabs_partial))
2229 return false;
2230 return true;
2231}
2232
2233int __kmem_cache_shrink(struct kmem_cache *cachep)
2234{
2235 int ret = 0;
2236 int node;
2237 struct kmem_cache_node *n;
2238
2239 drain_cpu_caches(cachep);
2240
2241 check_irq_on();
2242 for_each_kmem_cache_node(cachep, node, n) {
2243 drain_freelist(cachep, n, INT_MAX);
2244
2245 ret += !list_empty(&n->slabs_full) ||
2246 !list_empty(&n->slabs_partial);
2247 }
2248 return (ret ? 1 : 0);
2249}
2250
2251int __kmem_cache_shutdown(struct kmem_cache *cachep)
2252{
2253 return __kmem_cache_shrink(cachep);
2254}
2255
2256void __kmem_cache_release(struct kmem_cache *cachep)
2257{
2258 int i;
2259 struct kmem_cache_node *n;
2260
2261 cache_random_seq_destroy(cachep);
2262
2263 free_percpu(cachep->cpu_cache);
2264
2265 /* NUMA: free the node structures */
2266 for_each_kmem_cache_node(cachep, i, n) {
2267 kfree(n->shared);
2268 free_alien_cache(n->alien);
2269 kfree(n);
2270 cachep->node[i] = NULL;
2271 }
2272}
2273
2274/*
2275 * Get the memory for a slab management obj.
2276 *
2277 * For a slab cache when the slab descriptor is off-slab, the
2278 * slab descriptor can't come from the same cache which is being created,
2279 * Because if it is the case, that means we defer the creation of
2280 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2281 * And we eventually call down to __kmem_cache_create(), which
2282 * in turn looks up in the kmalloc_{dma,}_caches for the desired-size one.
2283 * This is a "chicken-and-egg" problem.
2284 *
2285 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2286 * which are all initialized during kmem_cache_init().
2287 */
2288static void *alloc_slabmgmt(struct kmem_cache *cachep,
2289 struct slab *slab, int colour_off,
2290 gfp_t local_flags, int nodeid)
2291{
2292 void *freelist;
2293 void *addr = slab_address(slab);
2294
2295 slab->s_mem = addr + colour_off;
2296 slab->active = 0;
2297
2298 if (OBJFREELIST_SLAB(cachep))
2299 freelist = NULL;
2300 else if (OFF_SLAB(cachep)) {
2301 /* Slab management obj is off-slab. */
2302 freelist = kmalloc_node(cachep->freelist_size,
2303 local_flags, nodeid);
2304 } else {
2305 /* We will use last bytes at the slab for freelist */
2306 freelist = addr + (PAGE_SIZE << cachep->gfporder) -
2307 cachep->freelist_size;
2308 }
2309
2310 return freelist;
2311}
2312
2313static inline freelist_idx_t get_free_obj(struct slab *slab, unsigned int idx)
2314{
2315 return ((freelist_idx_t *) slab->freelist)[idx];
2316}
2317
2318static inline void set_free_obj(struct slab *slab,
2319 unsigned int idx, freelist_idx_t val)
2320{
2321 ((freelist_idx_t *)(slab->freelist))[idx] = val;
2322}
2323
2324static void cache_init_objs_debug(struct kmem_cache *cachep, struct slab *slab)
2325{
2326#if DEBUG
2327 int i;
2328
2329 for (i = 0; i < cachep->num; i++) {
2330 void *objp = index_to_obj(cachep, slab, i);
2331
2332 if (cachep->flags & SLAB_STORE_USER)
2333 *dbg_userword(cachep, objp) = NULL;
2334
2335 if (cachep->flags & SLAB_RED_ZONE) {
2336 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2337 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2338 }
2339 /*
2340 * Constructors are not allowed to allocate memory from the same
2341 * cache which they are a constructor for. Otherwise, deadlock.
2342 * They must also be threaded.
2343 */
2344 if (cachep->ctor && !(cachep->flags & SLAB_POISON)) {
2345 kasan_unpoison_object_data(cachep,
2346 objp + obj_offset(cachep));
2347 cachep->ctor(objp + obj_offset(cachep));
2348 kasan_poison_object_data(
2349 cachep, objp + obj_offset(cachep));
2350 }
2351
2352 if (cachep->flags & SLAB_RED_ZONE) {
2353 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2354 slab_error(cachep, "constructor overwrote the end of an object");
2355 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2356 slab_error(cachep, "constructor overwrote the start of an object");
2357 }
2358 /* need to poison the objs? */
2359 if (cachep->flags & SLAB_POISON) {
2360 poison_obj(cachep, objp, POISON_FREE);
2361 slab_kernel_map(cachep, objp, 0);
2362 }
2363 }
2364#endif
2365}
2366
2367#ifdef CONFIG_SLAB_FREELIST_RANDOM
2368/* Hold information during a freelist initialization */
2369union freelist_init_state {
2370 struct {
2371 unsigned int pos;
2372 unsigned int *list;
2373 unsigned int count;
2374 };
2375 struct rnd_state rnd_state;
2376};
2377
2378/*
2379 * Initialize the state based on the randomization method available.
2380 * return true if the pre-computed list is available, false otherwise.
2381 */
2382static bool freelist_state_initialize(union freelist_init_state *state,
2383 struct kmem_cache *cachep,
2384 unsigned int count)
2385{
2386 bool ret;
2387 unsigned int rand;
2388
2389 /* Use best entropy available to define a random shift */
2390 rand = get_random_u32();
2391
2392 /* Use a random state if the pre-computed list is not available */
2393 if (!cachep->random_seq) {
2394 prandom_seed_state(&state->rnd_state, rand);
2395 ret = false;
2396 } else {
2397 state->list = cachep->random_seq;
2398 state->count = count;
2399 state->pos = rand % count;
2400 ret = true;
2401 }
2402 return ret;
2403}
2404
2405/* Get the next entry on the list and randomize it using a random shift */
2406static freelist_idx_t next_random_slot(union freelist_init_state *state)
2407{
2408 if (state->pos >= state->count)
2409 state->pos = 0;
2410 return state->list[state->pos++];
2411}
2412
2413/* Swap two freelist entries */
2414static void swap_free_obj(struct slab *slab, unsigned int a, unsigned int b)
2415{
2416 swap(((freelist_idx_t *) slab->freelist)[a],
2417 ((freelist_idx_t *) slab->freelist)[b]);
2418}
2419
2420/*
2421 * Shuffle the freelist initialization state based on pre-computed lists.
2422 * return true if the list was successfully shuffled, false otherwise.
2423 */
2424static bool shuffle_freelist(struct kmem_cache *cachep, struct slab *slab)
2425{
2426 unsigned int objfreelist = 0, i, rand, count = cachep->num;
2427 union freelist_init_state state;
2428 bool precomputed;
2429
2430 if (count < 2)
2431 return false;
2432
2433 precomputed = freelist_state_initialize(&state, cachep, count);
2434
2435 /* Take a random entry as the objfreelist */
2436 if (OBJFREELIST_SLAB(cachep)) {
2437 if (!precomputed)
2438 objfreelist = count - 1;
2439 else
2440 objfreelist = next_random_slot(&state);
2441 slab->freelist = index_to_obj(cachep, slab, objfreelist) +
2442 obj_offset(cachep);
2443 count--;
2444 }
2445
2446 /*
2447 * On early boot, generate the list dynamically.
2448 * Later use a pre-computed list for speed.
2449 */
2450 if (!precomputed) {
2451 for (i = 0; i < count; i++)
2452 set_free_obj(slab, i, i);
2453
2454 /* Fisher-Yates shuffle */
2455 for (i = count - 1; i > 0; i--) {
2456 rand = prandom_u32_state(&state.rnd_state);
2457 rand %= (i + 1);
2458 swap_free_obj(slab, i, rand);
2459 }
2460 } else {
2461 for (i = 0; i < count; i++)
2462 set_free_obj(slab, i, next_random_slot(&state));
2463 }
2464
2465 if (OBJFREELIST_SLAB(cachep))
2466 set_free_obj(slab, cachep->num - 1, objfreelist);
2467
2468 return true;
2469}
2470#else
2471static inline bool shuffle_freelist(struct kmem_cache *cachep,
2472 struct slab *slab)
2473{
2474 return false;
2475}
2476#endif /* CONFIG_SLAB_FREELIST_RANDOM */
2477
2478static void cache_init_objs(struct kmem_cache *cachep,
2479 struct slab *slab)
2480{
2481 int i;
2482 void *objp;
2483 bool shuffled;
2484
2485 cache_init_objs_debug(cachep, slab);
2486
2487 /* Try to randomize the freelist if enabled */
2488 shuffled = shuffle_freelist(cachep, slab);
2489
2490 if (!shuffled && OBJFREELIST_SLAB(cachep)) {
2491 slab->freelist = index_to_obj(cachep, slab, cachep->num - 1) +
2492 obj_offset(cachep);
2493 }
2494
2495 for (i = 0; i < cachep->num; i++) {
2496 objp = index_to_obj(cachep, slab, i);
2497 objp = kasan_init_slab_obj(cachep, objp);
2498
2499 /* constructor could break poison info */
2500 if (DEBUG == 0 && cachep->ctor) {
2501 kasan_unpoison_object_data(cachep, objp);
2502 cachep->ctor(objp);
2503 kasan_poison_object_data(cachep, objp);
2504 }
2505
2506 if (!shuffled)
2507 set_free_obj(slab, i, i);
2508 }
2509}
2510
2511static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slab)
2512{
2513 void *objp;
2514
2515 objp = index_to_obj(cachep, slab, get_free_obj(slab, slab->active));
2516 slab->active++;
2517
2518 return objp;
2519}
2520
2521static void slab_put_obj(struct kmem_cache *cachep,
2522 struct slab *slab, void *objp)
2523{
2524 unsigned int objnr = obj_to_index(cachep, slab, objp);
2525#if DEBUG
2526 unsigned int i;
2527
2528 /* Verify double free bug */
2529 for (i = slab->active; i < cachep->num; i++) {
2530 if (get_free_obj(slab, i) == objnr) {
2531 pr_err("slab: double free detected in cache '%s', objp %px\n",
2532 cachep->name, objp);
2533 BUG();
2534 }
2535 }
2536#endif
2537 slab->active--;
2538 if (!slab->freelist)
2539 slab->freelist = objp + obj_offset(cachep);
2540
2541 set_free_obj(slab, slab->active, objnr);
2542}
2543
2544/*
2545 * Grow (by 1) the number of slabs within a cache. This is called by
2546 * kmem_cache_alloc() when there are no active objs left in a cache.
2547 */
2548static struct slab *cache_grow_begin(struct kmem_cache *cachep,
2549 gfp_t flags, int nodeid)
2550{
2551 void *freelist;
2552 size_t offset;
2553 gfp_t local_flags;
2554 int slab_node;
2555 struct kmem_cache_node *n;
2556 struct slab *slab;
2557
2558 /*
2559 * Be lazy and only check for valid flags here, keeping it out of the
2560 * critical path in kmem_cache_alloc().
2561 */
2562 if (unlikely(flags & GFP_SLAB_BUG_MASK))
2563 flags = kmalloc_fix_flags(flags);
2564
2565 WARN_ON_ONCE(cachep->ctor && (flags & __GFP_ZERO));
2566 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2567
2568 check_irq_off();
2569 if (gfpflags_allow_blocking(local_flags))
2570 local_irq_enable();
2571
2572 /*
2573 * Get mem for the objs. Attempt to allocate a physical page from
2574 * 'nodeid'.
2575 */
2576 slab = kmem_getpages(cachep, local_flags, nodeid);
2577 if (!slab)
2578 goto failed;
2579
2580 slab_node = slab_nid(slab);
2581 n = get_node(cachep, slab_node);
2582
2583 /* Get colour for the slab, and cal the next value. */
2584 n->colour_next++;
2585 if (n->colour_next >= cachep->colour)
2586 n->colour_next = 0;
2587
2588 offset = n->colour_next;
2589 if (offset >= cachep->colour)
2590 offset = 0;
2591
2592 offset *= cachep->colour_off;
2593
2594 /*
2595 * Call kasan_poison_slab() before calling alloc_slabmgmt(), so
2596 * page_address() in the latter returns a non-tagged pointer,
2597 * as it should be for slab pages.
2598 */
2599 kasan_poison_slab(slab);
2600
2601 /* Get slab management. */
2602 freelist = alloc_slabmgmt(cachep, slab, offset,
2603 local_flags & ~GFP_CONSTRAINT_MASK, slab_node);
2604 if (OFF_SLAB(cachep) && !freelist)
2605 goto opps1;
2606
2607 slab->slab_cache = cachep;
2608 slab->freelist = freelist;
2609
2610 cache_init_objs(cachep, slab);
2611
2612 if (gfpflags_allow_blocking(local_flags))
2613 local_irq_disable();
2614
2615 return slab;
2616
2617opps1:
2618 kmem_freepages(cachep, slab);
2619failed:
2620 if (gfpflags_allow_blocking(local_flags))
2621 local_irq_disable();
2622 return NULL;
2623}
2624
2625static void cache_grow_end(struct kmem_cache *cachep, struct slab *slab)
2626{
2627 struct kmem_cache_node *n;
2628 void *list = NULL;
2629
2630 check_irq_off();
2631
2632 if (!slab)
2633 return;
2634
2635 INIT_LIST_HEAD(&slab->slab_list);
2636 n = get_node(cachep, slab_nid(slab));
2637
2638 raw_spin_lock(&n->list_lock);
2639 n->total_slabs++;
2640 if (!slab->active) {
2641 list_add_tail(&slab->slab_list, &n->slabs_free);
2642 n->free_slabs++;
2643 } else
2644 fixup_slab_list(cachep, n, slab, &list);
2645
2646 STATS_INC_GROWN(cachep);
2647 n->free_objects += cachep->num - slab->active;
2648 raw_spin_unlock(&n->list_lock);
2649
2650 fixup_objfreelist_debug(cachep, &list);
2651}
2652
2653#if DEBUG
2654
2655/*
2656 * Perform extra freeing checks:
2657 * - detect bad pointers.
2658 * - POISON/RED_ZONE checking
2659 */
2660static void kfree_debugcheck(const void *objp)
2661{
2662 if (!virt_addr_valid(objp)) {
2663 pr_err("kfree_debugcheck: out of range ptr %lxh\n",
2664 (unsigned long)objp);
2665 BUG();
2666 }
2667}
2668
2669static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2670{
2671 unsigned long long redzone1, redzone2;
2672
2673 redzone1 = *dbg_redzone1(cache, obj);
2674 redzone2 = *dbg_redzone2(cache, obj);
2675
2676 /*
2677 * Redzone is ok.
2678 */
2679 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2680 return;
2681
2682 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2683 slab_error(cache, "double free detected");
2684 else
2685 slab_error(cache, "memory outside object was overwritten");
2686
2687 pr_err("%px: redzone 1:0x%llx, redzone 2:0x%llx\n",
2688 obj, redzone1, redzone2);
2689}
2690
2691static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2692 unsigned long caller)
2693{
2694 unsigned int objnr;
2695 struct slab *slab;
2696
2697 BUG_ON(virt_to_cache(objp) != cachep);
2698
2699 objp -= obj_offset(cachep);
2700 kfree_debugcheck(objp);
2701 slab = virt_to_slab(objp);
2702
2703 if (cachep->flags & SLAB_RED_ZONE) {
2704 verify_redzone_free(cachep, objp);
2705 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2706 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2707 }
2708 if (cachep->flags & SLAB_STORE_USER)
2709 *dbg_userword(cachep, objp) = (void *)caller;
2710
2711 objnr = obj_to_index(cachep, slab, objp);
2712
2713 BUG_ON(objnr >= cachep->num);
2714 BUG_ON(objp != index_to_obj(cachep, slab, objnr));
2715
2716 if (cachep->flags & SLAB_POISON) {
2717 poison_obj(cachep, objp, POISON_FREE);
2718 slab_kernel_map(cachep, objp, 0);
2719 }
2720 return objp;
2721}
2722
2723#else
2724#define kfree_debugcheck(x) do { } while(0)
2725#define cache_free_debugcheck(x, objp, z) (objp)
2726#endif
2727
2728static inline void fixup_objfreelist_debug(struct kmem_cache *cachep,
2729 void **list)
2730{
2731#if DEBUG
2732 void *next = *list;
2733 void *objp;
2734
2735 while (next) {
2736 objp = next - obj_offset(cachep);
2737 next = *(void **)next;
2738 poison_obj(cachep, objp, POISON_FREE);
2739 }
2740#endif
2741}
2742
2743static inline void fixup_slab_list(struct kmem_cache *cachep,
2744 struct kmem_cache_node *n, struct slab *slab,
2745 void **list)
2746{
2747 /* move slabp to correct slabp list: */
2748 list_del(&slab->slab_list);
2749 if (slab->active == cachep->num) {
2750 list_add(&slab->slab_list, &n->slabs_full);
2751 if (OBJFREELIST_SLAB(cachep)) {
2752#if DEBUG
2753 /* Poisoning will be done without holding the lock */
2754 if (cachep->flags & SLAB_POISON) {
2755 void **objp = slab->freelist;
2756
2757 *objp = *list;
2758 *list = objp;
2759 }
2760#endif
2761 slab->freelist = NULL;
2762 }
2763 } else
2764 list_add(&slab->slab_list, &n->slabs_partial);
2765}
2766
2767/* Try to find non-pfmemalloc slab if needed */
2768static noinline struct slab *get_valid_first_slab(struct kmem_cache_node *n,
2769 struct slab *slab, bool pfmemalloc)
2770{
2771 if (!slab)
2772 return NULL;
2773
2774 if (pfmemalloc)
2775 return slab;
2776
2777 if (!slab_test_pfmemalloc(slab))
2778 return slab;
2779
2780 /* No need to keep pfmemalloc slab if we have enough free objects */
2781 if (n->free_objects > n->free_limit) {
2782 slab_clear_pfmemalloc(slab);
2783 return slab;
2784 }
2785
2786 /* Move pfmemalloc slab to the end of list to speed up next search */
2787 list_del(&slab->slab_list);
2788 if (!slab->active) {
2789 list_add_tail(&slab->slab_list, &n->slabs_free);
2790 n->free_slabs++;
2791 } else
2792 list_add_tail(&slab->slab_list, &n->slabs_partial);
2793
2794 list_for_each_entry(slab, &n->slabs_partial, slab_list) {
2795 if (!slab_test_pfmemalloc(slab))
2796 return slab;
2797 }
2798
2799 n->free_touched = 1;
2800 list_for_each_entry(slab, &n->slabs_free, slab_list) {
2801 if (!slab_test_pfmemalloc(slab)) {
2802 n->free_slabs--;
2803 return slab;
2804 }
2805 }
2806
2807 return NULL;
2808}
2809
2810static struct slab *get_first_slab(struct kmem_cache_node *n, bool pfmemalloc)
2811{
2812 struct slab *slab;
2813
2814 assert_raw_spin_locked(&n->list_lock);
2815 slab = list_first_entry_or_null(&n->slabs_partial, struct slab,
2816 slab_list);
2817 if (!slab) {
2818 n->free_touched = 1;
2819 slab = list_first_entry_or_null(&n->slabs_free, struct slab,
2820 slab_list);
2821 if (slab)
2822 n->free_slabs--;
2823 }
2824
2825 if (sk_memalloc_socks())
2826 slab = get_valid_first_slab(n, slab, pfmemalloc);
2827
2828 return slab;
2829}
2830
2831static noinline void *cache_alloc_pfmemalloc(struct kmem_cache *cachep,
2832 struct kmem_cache_node *n, gfp_t flags)
2833{
2834 struct slab *slab;
2835 void *obj;
2836 void *list = NULL;
2837
2838 if (!gfp_pfmemalloc_allowed(flags))
2839 return NULL;
2840
2841 raw_spin_lock(&n->list_lock);
2842 slab = get_first_slab(n, true);
2843 if (!slab) {
2844 raw_spin_unlock(&n->list_lock);
2845 return NULL;
2846 }
2847
2848 obj = slab_get_obj(cachep, slab);
2849 n->free_objects--;
2850
2851 fixup_slab_list(cachep, n, slab, &list);
2852
2853 raw_spin_unlock(&n->list_lock);
2854 fixup_objfreelist_debug(cachep, &list);
2855
2856 return obj;
2857}
2858
2859/*
2860 * Slab list should be fixed up by fixup_slab_list() for existing slab
2861 * or cache_grow_end() for new slab
2862 */
2863static __always_inline int alloc_block(struct kmem_cache *cachep,
2864 struct array_cache *ac, struct slab *slab, int batchcount)
2865{
2866 /*
2867 * There must be at least one object available for
2868 * allocation.
2869 */
2870 BUG_ON(slab->active >= cachep->num);
2871
2872 while (slab->active < cachep->num && batchcount--) {
2873 STATS_INC_ALLOCED(cachep);
2874 STATS_INC_ACTIVE(cachep);
2875 STATS_SET_HIGH(cachep);
2876
2877 ac->entry[ac->avail++] = slab_get_obj(cachep, slab);
2878 }
2879
2880 return batchcount;
2881}
2882
2883static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
2884{
2885 int batchcount;
2886 struct kmem_cache_node *n;
2887 struct array_cache *ac, *shared;
2888 int node;
2889 void *list = NULL;
2890 struct slab *slab;
2891
2892 check_irq_off();
2893 node = numa_mem_id();
2894
2895 ac = cpu_cache_get(cachep);
2896 batchcount = ac->batchcount;
2897 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2898 /*
2899 * If there was little recent activity on this cache, then
2900 * perform only a partial refill. Otherwise we could generate
2901 * refill bouncing.
2902 */
2903 batchcount = BATCHREFILL_LIMIT;
2904 }
2905 n = get_node(cachep, node);
2906
2907 BUG_ON(ac->avail > 0 || !n);
2908 shared = READ_ONCE(n->shared);
2909 if (!n->free_objects && (!shared || !shared->avail))
2910 goto direct_grow;
2911
2912 raw_spin_lock(&n->list_lock);
2913 shared = READ_ONCE(n->shared);
2914
2915 /* See if we can refill from the shared array */
2916 if (shared && transfer_objects(ac, shared, batchcount)) {
2917 shared->touched = 1;
2918 goto alloc_done;
2919 }
2920
2921 while (batchcount > 0) {
2922 /* Get slab alloc is to come from. */
2923 slab = get_first_slab(n, false);
2924 if (!slab)
2925 goto must_grow;
2926
2927 check_spinlock_acquired(cachep);
2928
2929 batchcount = alloc_block(cachep, ac, slab, batchcount);
2930 fixup_slab_list(cachep, n, slab, &list);
2931 }
2932
2933must_grow:
2934 n->free_objects -= ac->avail;
2935alloc_done:
2936 raw_spin_unlock(&n->list_lock);
2937 fixup_objfreelist_debug(cachep, &list);
2938
2939direct_grow:
2940 if (unlikely(!ac->avail)) {
2941 /* Check if we can use obj in pfmemalloc slab */
2942 if (sk_memalloc_socks()) {
2943 void *obj = cache_alloc_pfmemalloc(cachep, n, flags);
2944
2945 if (obj)
2946 return obj;
2947 }
2948
2949 slab = cache_grow_begin(cachep, gfp_exact_node(flags), node);
2950
2951 /*
2952 * cache_grow_begin() can reenable interrupts,
2953 * then ac could change.
2954 */
2955 ac = cpu_cache_get(cachep);
2956 if (!ac->avail && slab)
2957 alloc_block(cachep, ac, slab, batchcount);
2958 cache_grow_end(cachep, slab);
2959
2960 if (!ac->avail)
2961 return NULL;
2962 }
2963 ac->touched = 1;
2964
2965 return ac->entry[--ac->avail];
2966}
2967
2968#if DEBUG
2969static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
2970 gfp_t flags, void *objp, unsigned long caller)
2971{
2972 WARN_ON_ONCE(cachep->ctor && (flags & __GFP_ZERO));
2973 if (!objp || is_kfence_address(objp))
2974 return objp;
2975 if (cachep->flags & SLAB_POISON) {
2976 check_poison_obj(cachep, objp);
2977 slab_kernel_map(cachep, objp, 1);
2978 poison_obj(cachep, objp, POISON_INUSE);
2979 }
2980 if (cachep->flags & SLAB_STORE_USER)
2981 *dbg_userword(cachep, objp) = (void *)caller;
2982
2983 if (cachep->flags & SLAB_RED_ZONE) {
2984 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
2985 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2986 slab_error(cachep, "double free, or memory outside object was overwritten");
2987 pr_err("%px: redzone 1:0x%llx, redzone 2:0x%llx\n",
2988 objp, *dbg_redzone1(cachep, objp),
2989 *dbg_redzone2(cachep, objp));
2990 }
2991 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2992 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2993 }
2994
2995 objp += obj_offset(cachep);
2996 if (cachep->ctor && cachep->flags & SLAB_POISON)
2997 cachep->ctor(objp);
2998 if ((unsigned long)objp & (arch_slab_minalign() - 1)) {
2999 pr_err("0x%px: not aligned to arch_slab_minalign()=%u\n", objp,
3000 arch_slab_minalign());
3001 }
3002 return objp;
3003}
3004#else
3005#define cache_alloc_debugcheck_after(a, b, objp, d) (objp)
3006#endif
3007
3008static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3009{
3010 void *objp;
3011 struct array_cache *ac;
3012
3013 check_irq_off();
3014
3015 ac = cpu_cache_get(cachep);
3016 if (likely(ac->avail)) {
3017 ac->touched = 1;
3018 objp = ac->entry[--ac->avail];
3019
3020 STATS_INC_ALLOCHIT(cachep);
3021 goto out;
3022 }
3023
3024 STATS_INC_ALLOCMISS(cachep);
3025 objp = cache_alloc_refill(cachep, flags);
3026 /*
3027 * the 'ac' may be updated by cache_alloc_refill(),
3028 * and kmemleak_erase() requires its correct value.
3029 */
3030 ac = cpu_cache_get(cachep);
3031
3032out:
3033 /*
3034 * To avoid a false negative, if an object that is in one of the
3035 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3036 * treat the array pointers as a reference to the object.
3037 */
3038 if (objp)
3039 kmemleak_erase(&ac->entry[ac->avail]);
3040 return objp;
3041}
3042
3043#ifdef CONFIG_NUMA
3044static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
3045
3046/*
3047 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
3048 *
3049 * If we are in_interrupt, then process context, including cpusets and
3050 * mempolicy, may not apply and should not be used for allocation policy.
3051 */
3052static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3053{
3054 int nid_alloc, nid_here;
3055
3056 if (in_interrupt() || (flags & __GFP_THISNODE))
3057 return NULL;
3058 nid_alloc = nid_here = numa_mem_id();
3059 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3060 nid_alloc = cpuset_slab_spread_node();
3061 else if (current->mempolicy)
3062 nid_alloc = mempolicy_slab_node();
3063 if (nid_alloc != nid_here)
3064 return ____cache_alloc_node(cachep, flags, nid_alloc);
3065 return NULL;
3066}
3067
3068/*
3069 * Fallback function if there was no memory available and no objects on a
3070 * certain node and fall back is permitted. First we scan all the
3071 * available node for available objects. If that fails then we
3072 * perform an allocation without specifying a node. This allows the page
3073 * allocator to do its reclaim / fallback magic. We then insert the
3074 * slab into the proper nodelist and then allocate from it.
3075 */
3076static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3077{
3078 struct zonelist *zonelist;
3079 struct zoneref *z;
3080 struct zone *zone;
3081 enum zone_type highest_zoneidx = gfp_zone(flags);
3082 void *obj = NULL;
3083 struct slab *slab;
3084 int nid;
3085 unsigned int cpuset_mems_cookie;
3086
3087 if (flags & __GFP_THISNODE)
3088 return NULL;
3089
3090retry_cpuset:
3091 cpuset_mems_cookie = read_mems_allowed_begin();
3092 zonelist = node_zonelist(mempolicy_slab_node(), flags);
3093
3094retry:
3095 /*
3096 * Look through allowed nodes for objects available
3097 * from existing per node queues.
3098 */
3099 for_each_zone_zonelist(zone, z, zonelist, highest_zoneidx) {
3100 nid = zone_to_nid(zone);
3101
3102 if (cpuset_zone_allowed(zone, flags) &&
3103 get_node(cache, nid) &&
3104 get_node(cache, nid)->free_objects) {
3105 obj = ____cache_alloc_node(cache,
3106 gfp_exact_node(flags), nid);
3107 if (obj)
3108 break;
3109 }
3110 }
3111
3112 if (!obj) {
3113 /*
3114 * This allocation will be performed within the constraints
3115 * of the current cpuset / memory policy requirements.
3116 * We may trigger various forms of reclaim on the allowed
3117 * set and go into memory reserves if necessary.
3118 */
3119 slab = cache_grow_begin(cache, flags, numa_mem_id());
3120 cache_grow_end(cache, slab);
3121 if (slab) {
3122 nid = slab_nid(slab);
3123 obj = ____cache_alloc_node(cache,
3124 gfp_exact_node(flags), nid);
3125
3126 /*
3127 * Another processor may allocate the objects in
3128 * the slab since we are not holding any locks.
3129 */
3130 if (!obj)
3131 goto retry;
3132 }
3133 }
3134
3135 if (unlikely(!obj && read_mems_allowed_retry(cpuset_mems_cookie)))
3136 goto retry_cpuset;
3137 return obj;
3138}
3139
3140/*
3141 * An interface to enable slab creation on nodeid
3142 */
3143static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3144 int nodeid)
3145{
3146 struct slab *slab;
3147 struct kmem_cache_node *n;
3148 void *obj = NULL;
3149 void *list = NULL;
3150
3151 VM_BUG_ON(nodeid < 0 || nodeid >= MAX_NUMNODES);
3152 n = get_node(cachep, nodeid);
3153 BUG_ON(!n);
3154
3155 check_irq_off();
3156 raw_spin_lock(&n->list_lock);
3157 slab = get_first_slab(n, false);
3158 if (!slab)
3159 goto must_grow;
3160
3161 check_spinlock_acquired_node(cachep, nodeid);
3162
3163 STATS_INC_NODEALLOCS(cachep);
3164 STATS_INC_ACTIVE(cachep);
3165 STATS_SET_HIGH(cachep);
3166
3167 BUG_ON(slab->active == cachep->num);
3168
3169 obj = slab_get_obj(cachep, slab);
3170 n->free_objects--;
3171
3172 fixup_slab_list(cachep, n, slab, &list);
3173
3174 raw_spin_unlock(&n->list_lock);
3175 fixup_objfreelist_debug(cachep, &list);
3176 return obj;
3177
3178must_grow:
3179 raw_spin_unlock(&n->list_lock);
3180 slab = cache_grow_begin(cachep, gfp_exact_node(flags), nodeid);
3181 if (slab) {
3182 /* This slab isn't counted yet so don't update free_objects */
3183 obj = slab_get_obj(cachep, slab);
3184 }
3185 cache_grow_end(cachep, slab);
3186
3187 return obj ? obj : fallback_alloc(cachep, flags);
3188}
3189
3190static __always_inline void *
3191__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3192{
3193 void *objp = NULL;
3194 int slab_node = numa_mem_id();
3195
3196 if (nodeid == NUMA_NO_NODE) {
3197 if (current->mempolicy || cpuset_do_slab_mem_spread()) {
3198 objp = alternate_node_alloc(cachep, flags);
3199 if (objp)
3200 goto out;
3201 }
3202 /*
3203 * Use the locally cached objects if possible.
3204 * However ____cache_alloc does not allow fallback
3205 * to other nodes. It may fail while we still have
3206 * objects on other nodes available.
3207 */
3208 objp = ____cache_alloc(cachep, flags);
3209 nodeid = slab_node;
3210 } else if (nodeid == slab_node) {
3211 objp = ____cache_alloc(cachep, flags);
3212 } else if (!get_node(cachep, nodeid)) {
3213 /* Node not bootstrapped yet */
3214 objp = fallback_alloc(cachep, flags);
3215 goto out;
3216 }
3217
3218 /*
3219 * We may just have run out of memory on the local node.
3220 * ____cache_alloc_node() knows how to locate memory on other nodes
3221 */
3222 if (!objp)
3223 objp = ____cache_alloc_node(cachep, flags, nodeid);
3224out:
3225 return objp;
3226}
3227#else
3228
3229static __always_inline void *
3230__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags, int nodeid __maybe_unused)
3231{
3232 return ____cache_alloc(cachep, flags);
3233}
3234
3235#endif /* CONFIG_NUMA */
3236
3237static __always_inline void *
3238slab_alloc_node(struct kmem_cache *cachep, struct list_lru *lru, gfp_t flags,
3239 int nodeid, size_t orig_size, unsigned long caller)
3240{
3241 unsigned long save_flags;
3242 void *objp;
3243 struct obj_cgroup *objcg = NULL;
3244 bool init = false;
3245
3246 flags &= gfp_allowed_mask;
3247 cachep = slab_pre_alloc_hook(cachep, lru, &objcg, 1, flags);
3248 if (unlikely(!cachep))
3249 return NULL;
3250
3251 objp = kfence_alloc(cachep, orig_size, flags);
3252 if (unlikely(objp))
3253 goto out;
3254
3255 local_irq_save(save_flags);
3256 objp = __do_cache_alloc(cachep, flags, nodeid);
3257 local_irq_restore(save_flags);
3258 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3259 prefetchw(objp);
3260 init = slab_want_init_on_alloc(flags, cachep);
3261
3262out:
3263 slab_post_alloc_hook(cachep, objcg, flags, 1, &objp, init,
3264 cachep->object_size);
3265 return objp;
3266}
3267
3268static __always_inline void *
3269slab_alloc(struct kmem_cache *cachep, struct list_lru *lru, gfp_t flags,
3270 size_t orig_size, unsigned long caller)
3271{
3272 return slab_alloc_node(cachep, lru, flags, NUMA_NO_NODE, orig_size,
3273 caller);
3274}
3275
3276/*
3277 * Caller needs to acquire correct kmem_cache_node's list_lock
3278 * @list: List of detached free slabs should be freed by caller
3279 */
3280static void free_block(struct kmem_cache *cachep, void **objpp,
3281 int nr_objects, int node, struct list_head *list)
3282{
3283 int i;
3284 struct kmem_cache_node *n = get_node(cachep, node);
3285 struct slab *slab;
3286
3287 n->free_objects += nr_objects;
3288
3289 for (i = 0; i < nr_objects; i++) {
3290 void *objp;
3291 struct slab *slab;
3292
3293 objp = objpp[i];
3294
3295 slab = virt_to_slab(objp);
3296 list_del(&slab->slab_list);
3297 check_spinlock_acquired_node(cachep, node);
3298 slab_put_obj(cachep, slab, objp);
3299 STATS_DEC_ACTIVE(cachep);
3300
3301 /* fixup slab chains */
3302 if (slab->active == 0) {
3303 list_add(&slab->slab_list, &n->slabs_free);
3304 n->free_slabs++;
3305 } else {
3306 /* Unconditionally move a slab to the end of the
3307 * partial list on free - maximum time for the
3308 * other objects to be freed, too.
3309 */
3310 list_add_tail(&slab->slab_list, &n->slabs_partial);
3311 }
3312 }
3313
3314 while (n->free_objects > n->free_limit && !list_empty(&n->slabs_free)) {
3315 n->free_objects -= cachep->num;
3316
3317 slab = list_last_entry(&n->slabs_free, struct slab, slab_list);
3318 list_move(&slab->slab_list, list);
3319 n->free_slabs--;
3320 n->total_slabs--;
3321 }
3322}
3323
3324static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3325{
3326 int batchcount;
3327 struct kmem_cache_node *n;
3328 int node = numa_mem_id();
3329 LIST_HEAD(list);
3330
3331 batchcount = ac->batchcount;
3332
3333 check_irq_off();
3334 n = get_node(cachep, node);
3335 raw_spin_lock(&n->list_lock);
3336 if (n->shared) {
3337 struct array_cache *shared_array = n->shared;
3338 int max = shared_array->limit - shared_array->avail;
3339 if (max) {
3340 if (batchcount > max)
3341 batchcount = max;
3342 memcpy(&(shared_array->entry[shared_array->avail]),
3343 ac->entry, sizeof(void *) * batchcount);
3344 shared_array->avail += batchcount;
3345 goto free_done;
3346 }
3347 }
3348
3349 free_block(cachep, ac->entry, batchcount, node, &list);
3350free_done:
3351#if STATS
3352 {
3353 int i = 0;
3354 struct slab *slab;
3355
3356 list_for_each_entry(slab, &n->slabs_free, slab_list) {
3357 BUG_ON(slab->active);
3358
3359 i++;
3360 }
3361 STATS_SET_FREEABLE(cachep, i);
3362 }
3363#endif
3364 raw_spin_unlock(&n->list_lock);
3365 ac->avail -= batchcount;
3366 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3367 slabs_destroy(cachep, &list);
3368}
3369
3370/*
3371 * Release an obj back to its cache. If the obj has a constructed state, it must
3372 * be in this state _before_ it is released. Called with disabled ints.
3373 */
3374static __always_inline void __cache_free(struct kmem_cache *cachep, void *objp,
3375 unsigned long caller)
3376{
3377 bool init;
3378
3379 memcg_slab_free_hook(cachep, virt_to_slab(objp), &objp, 1);
3380
3381 if (is_kfence_address(objp)) {
3382 kmemleak_free_recursive(objp, cachep->flags);
3383 __kfence_free(objp);
3384 return;
3385 }
3386
3387 /*
3388 * As memory initialization might be integrated into KASAN,
3389 * kasan_slab_free and initialization memset must be
3390 * kept together to avoid discrepancies in behavior.
3391 */
3392 init = slab_want_init_on_free(cachep);
3393 if (init && !kasan_has_integrated_init())
3394 memset(objp, 0, cachep->object_size);
3395 /* KASAN might put objp into memory quarantine, delaying its reuse. */
3396 if (kasan_slab_free(cachep, objp, init))
3397 return;
3398
3399 /* Use KCSAN to help debug racy use-after-free. */
3400 if (!(cachep->flags & SLAB_TYPESAFE_BY_RCU))
3401 __kcsan_check_access(objp, cachep->object_size,
3402 KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT);
3403
3404 ___cache_free(cachep, objp, caller);
3405}
3406
3407void ___cache_free(struct kmem_cache *cachep, void *objp,
3408 unsigned long caller)
3409{
3410 struct array_cache *ac = cpu_cache_get(cachep);
3411
3412 check_irq_off();
3413 kmemleak_free_recursive(objp, cachep->flags);
3414 objp = cache_free_debugcheck(cachep, objp, caller);
3415
3416 /*
3417 * Skip calling cache_free_alien() when the platform is not numa.
3418 * This will avoid cache misses that happen while accessing slabp (which
3419 * is per page memory reference) to get nodeid. Instead use a global
3420 * variable to skip the call, which is mostly likely to be present in
3421 * the cache.
3422 */
3423 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3424 return;
3425
3426 if (ac->avail < ac->limit) {
3427 STATS_INC_FREEHIT(cachep);
3428 } else {
3429 STATS_INC_FREEMISS(cachep);
3430 cache_flusharray(cachep, ac);
3431 }
3432
3433 if (sk_memalloc_socks()) {
3434 struct slab *slab = virt_to_slab(objp);
3435
3436 if (unlikely(slab_test_pfmemalloc(slab))) {
3437 cache_free_pfmemalloc(cachep, slab, objp);
3438 return;
3439 }
3440 }
3441
3442 __free_one(ac, objp);
3443}
3444
3445static __always_inline
3446void *__kmem_cache_alloc_lru(struct kmem_cache *cachep, struct list_lru *lru,
3447 gfp_t flags)
3448{
3449 void *ret = slab_alloc(cachep, lru, flags, cachep->object_size, _RET_IP_);
3450
3451 trace_kmem_cache_alloc(_RET_IP_, ret, cachep, flags, NUMA_NO_NODE);
3452
3453 return ret;
3454}
3455
3456void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3457{
3458 return __kmem_cache_alloc_lru(cachep, NULL, flags);
3459}
3460EXPORT_SYMBOL(kmem_cache_alloc);
3461
3462void *kmem_cache_alloc_lru(struct kmem_cache *cachep, struct list_lru *lru,
3463 gfp_t flags)
3464{
3465 return __kmem_cache_alloc_lru(cachep, lru, flags);
3466}
3467EXPORT_SYMBOL(kmem_cache_alloc_lru);
3468
3469static __always_inline void
3470cache_alloc_debugcheck_after_bulk(struct kmem_cache *s, gfp_t flags,
3471 size_t size, void **p, unsigned long caller)
3472{
3473 size_t i;
3474
3475 for (i = 0; i < size; i++)
3476 p[i] = cache_alloc_debugcheck_after(s, flags, p[i], caller);
3477}
3478
3479int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3480 void **p)
3481{
3482 size_t i;
3483 struct obj_cgroup *objcg = NULL;
3484
3485 s = slab_pre_alloc_hook(s, NULL, &objcg, size, flags);
3486 if (!s)
3487 return 0;
3488
3489 local_irq_disable();
3490 for (i = 0; i < size; i++) {
3491 void *objp = kfence_alloc(s, s->object_size, flags) ?:
3492 __do_cache_alloc(s, flags, NUMA_NO_NODE);
3493
3494 if (unlikely(!objp))
3495 goto error;
3496 p[i] = objp;
3497 }
3498 local_irq_enable();
3499
3500 cache_alloc_debugcheck_after_bulk(s, flags, size, p, _RET_IP_);
3501
3502 /*
3503 * memcg and kmem_cache debug support and memory initialization.
3504 * Done outside of the IRQ disabled section.
3505 */
3506 slab_post_alloc_hook(s, objcg, flags, size, p,
3507 slab_want_init_on_alloc(flags, s), s->object_size);
3508 /* FIXME: Trace call missing. Christoph would like a bulk variant */
3509 return size;
3510error:
3511 local_irq_enable();
3512 cache_alloc_debugcheck_after_bulk(s, flags, i, p, _RET_IP_);
3513 slab_post_alloc_hook(s, objcg, flags, i, p, false, s->object_size);
3514 kmem_cache_free_bulk(s, i, p);
3515 return 0;
3516}
3517EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3518
3519/**
3520 * kmem_cache_alloc_node - Allocate an object on the specified node
3521 * @cachep: The cache to allocate from.
3522 * @flags: See kmalloc().
3523 * @nodeid: node number of the target node.
3524 *
3525 * Identical to kmem_cache_alloc but it will allocate memory on the given
3526 * node, which can improve the performance for cpu bound structures.
3527 *
3528 * Fallback to other node is possible if __GFP_THISNODE is not set.
3529 *
3530 * Return: pointer to the new object or %NULL in case of error
3531 */
3532void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3533{
3534 void *ret = slab_alloc_node(cachep, NULL, flags, nodeid, cachep->object_size, _RET_IP_);
3535
3536 trace_kmem_cache_alloc(_RET_IP_, ret, cachep, flags, nodeid);
3537
3538 return ret;
3539}
3540EXPORT_SYMBOL(kmem_cache_alloc_node);
3541
3542void *__kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3543 int nodeid, size_t orig_size,
3544 unsigned long caller)
3545{
3546 return slab_alloc_node(cachep, NULL, flags, nodeid,
3547 orig_size, caller);
3548}
3549
3550#ifdef CONFIG_PRINTK
3551void __kmem_obj_info(struct kmem_obj_info *kpp, void *object, struct slab *slab)
3552{
3553 struct kmem_cache *cachep;
3554 unsigned int objnr;
3555 void *objp;
3556
3557 kpp->kp_ptr = object;
3558 kpp->kp_slab = slab;
3559 cachep = slab->slab_cache;
3560 kpp->kp_slab_cache = cachep;
3561 objp = object - obj_offset(cachep);
3562 kpp->kp_data_offset = obj_offset(cachep);
3563 slab = virt_to_slab(objp);
3564 objnr = obj_to_index(cachep, slab, objp);
3565 objp = index_to_obj(cachep, slab, objnr);
3566 kpp->kp_objp = objp;
3567 if (DEBUG && cachep->flags & SLAB_STORE_USER)
3568 kpp->kp_ret = *dbg_userword(cachep, objp);
3569}
3570#endif
3571
3572static __always_inline
3573void __do_kmem_cache_free(struct kmem_cache *cachep, void *objp,
3574 unsigned long caller)
3575{
3576 unsigned long flags;
3577
3578 local_irq_save(flags);
3579 debug_check_no_locks_freed(objp, cachep->object_size);
3580 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3581 debug_check_no_obj_freed(objp, cachep->object_size);
3582 __cache_free(cachep, objp, caller);
3583 local_irq_restore(flags);
3584}
3585
3586void __kmem_cache_free(struct kmem_cache *cachep, void *objp,
3587 unsigned long caller)
3588{
3589 __do_kmem_cache_free(cachep, objp, caller);
3590}
3591
3592/**
3593 * kmem_cache_free - Deallocate an object
3594 * @cachep: The cache the allocation was from.
3595 * @objp: The previously allocated object.
3596 *
3597 * Free an object which was previously allocated from this
3598 * cache.
3599 */
3600void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3601{
3602 cachep = cache_from_obj(cachep, objp);
3603 if (!cachep)
3604 return;
3605
3606 trace_kmem_cache_free(_RET_IP_, objp, cachep);
3607 __do_kmem_cache_free(cachep, objp, _RET_IP_);
3608}
3609EXPORT_SYMBOL(kmem_cache_free);
3610
3611void kmem_cache_free_bulk(struct kmem_cache *orig_s, size_t size, void **p)
3612{
3613
3614 local_irq_disable();
3615 for (int i = 0; i < size; i++) {
3616 void *objp = p[i];
3617 struct kmem_cache *s;
3618
3619 if (!orig_s) {
3620 struct folio *folio = virt_to_folio(objp);
3621
3622 /* called via kfree_bulk */
3623 if (!folio_test_slab(folio)) {
3624 local_irq_enable();
3625 free_large_kmalloc(folio, objp);
3626 local_irq_disable();
3627 continue;
3628 }
3629 s = folio_slab(folio)->slab_cache;
3630 } else {
3631 s = cache_from_obj(orig_s, objp);
3632 }
3633
3634 if (!s)
3635 continue;
3636
3637 debug_check_no_locks_freed(objp, s->object_size);
3638 if (!(s->flags & SLAB_DEBUG_OBJECTS))
3639 debug_check_no_obj_freed(objp, s->object_size);
3640
3641 __cache_free(s, objp, _RET_IP_);
3642 }
3643 local_irq_enable();
3644
3645 /* FIXME: add tracing */
3646}
3647EXPORT_SYMBOL(kmem_cache_free_bulk);
3648
3649/*
3650 * This initializes kmem_cache_node or resizes various caches for all nodes.
3651 */
3652static int setup_kmem_cache_nodes(struct kmem_cache *cachep, gfp_t gfp)
3653{
3654 int ret;
3655 int node;
3656 struct kmem_cache_node *n;
3657
3658 for_each_online_node(node) {
3659 ret = setup_kmem_cache_node(cachep, node, gfp, true);
3660 if (ret)
3661 goto fail;
3662
3663 }
3664
3665 return 0;
3666
3667fail:
3668 if (!cachep->list.next) {
3669 /* Cache is not active yet. Roll back what we did */
3670 node--;
3671 while (node >= 0) {
3672 n = get_node(cachep, node);
3673 if (n) {
3674 kfree(n->shared);
3675 free_alien_cache(n->alien);
3676 kfree(n);
3677 cachep->node[node] = NULL;
3678 }
3679 node--;
3680 }
3681 }
3682 return -ENOMEM;
3683}
3684
3685/* Always called with the slab_mutex held */
3686static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3687 int batchcount, int shared, gfp_t gfp)
3688{
3689 struct array_cache __percpu *cpu_cache, *prev;
3690 int cpu;
3691
3692 cpu_cache = alloc_kmem_cache_cpus(cachep, limit, batchcount);
3693 if (!cpu_cache)
3694 return -ENOMEM;
3695
3696 prev = cachep->cpu_cache;
3697 cachep->cpu_cache = cpu_cache;
3698 /*
3699 * Without a previous cpu_cache there's no need to synchronize remote
3700 * cpus, so skip the IPIs.
3701 */
3702 if (prev)
3703 kick_all_cpus_sync();
3704
3705 check_irq_on();
3706 cachep->batchcount = batchcount;
3707 cachep->limit = limit;
3708 cachep->shared = shared;
3709
3710 if (!prev)
3711 goto setup_node;
3712
3713 for_each_online_cpu(cpu) {
3714 LIST_HEAD(list);
3715 int node;
3716 struct kmem_cache_node *n;
3717 struct array_cache *ac = per_cpu_ptr(prev, cpu);
3718
3719 node = cpu_to_mem(cpu);
3720 n = get_node(cachep, node);
3721 raw_spin_lock_irq(&n->list_lock);
3722 free_block(cachep, ac->entry, ac->avail, node, &list);
3723 raw_spin_unlock_irq(&n->list_lock);
3724 slabs_destroy(cachep, &list);
3725 }
3726 free_percpu(prev);
3727
3728setup_node:
3729 return setup_kmem_cache_nodes(cachep, gfp);
3730}
3731
3732/* Called with slab_mutex held always */
3733static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
3734{
3735 int err;
3736 int limit = 0;
3737 int shared = 0;
3738 int batchcount = 0;
3739
3740 err = cache_random_seq_create(cachep, cachep->num, gfp);
3741 if (err)
3742 goto end;
3743
3744 /*
3745 * The head array serves three purposes:
3746 * - create a LIFO ordering, i.e. return objects that are cache-warm
3747 * - reduce the number of spinlock operations.
3748 * - reduce the number of linked list operations on the slab and
3749 * bufctl chains: array operations are cheaper.
3750 * The numbers are guessed, we should auto-tune as described by
3751 * Bonwick.
3752 */
3753 if (cachep->size > 131072)
3754 limit = 1;
3755 else if (cachep->size > PAGE_SIZE)
3756 limit = 8;
3757 else if (cachep->size > 1024)
3758 limit = 24;
3759 else if (cachep->size > 256)
3760 limit = 54;
3761 else
3762 limit = 120;
3763
3764 /*
3765 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3766 * allocation behaviour: Most allocs on one cpu, most free operations
3767 * on another cpu. For these cases, an efficient object passing between
3768 * cpus is necessary. This is provided by a shared array. The array
3769 * replaces Bonwick's magazine layer.
3770 * On uniprocessor, it's functionally equivalent (but less efficient)
3771 * to a larger limit. Thus disabled by default.
3772 */
3773 shared = 0;
3774 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
3775 shared = 8;
3776
3777#if DEBUG
3778 /*
3779 * With debugging enabled, large batchcount lead to excessively long
3780 * periods with disabled local interrupts. Limit the batchcount
3781 */
3782 if (limit > 32)
3783 limit = 32;
3784#endif
3785 batchcount = (limit + 1) / 2;
3786 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3787end:
3788 if (err)
3789 pr_err("enable_cpucache failed for %s, error %d\n",
3790 cachep->name, -err);
3791 return err;
3792}
3793
3794/*
3795 * Drain an array if it contains any elements taking the node lock only if
3796 * necessary. Note that the node listlock also protects the array_cache
3797 * if drain_array() is used on the shared array.
3798 */
3799static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
3800 struct array_cache *ac, int node)
3801{
3802 LIST_HEAD(list);
3803
3804 /* ac from n->shared can be freed if we don't hold the slab_mutex. */
3805 check_mutex_acquired();
3806
3807 if (!ac || !ac->avail)
3808 return;
3809
3810 if (ac->touched) {
3811 ac->touched = 0;
3812 return;
3813 }
3814
3815 raw_spin_lock_irq(&n->list_lock);
3816 drain_array_locked(cachep, ac, node, false, &list);
3817 raw_spin_unlock_irq(&n->list_lock);
3818
3819 slabs_destroy(cachep, &list);
3820}
3821
3822/**
3823 * cache_reap - Reclaim memory from caches.
3824 * @w: work descriptor
3825 *
3826 * Called from workqueue/eventd every few seconds.
3827 * Purpose:
3828 * - clear the per-cpu caches for this CPU.
3829 * - return freeable pages to the main free memory pool.
3830 *
3831 * If we cannot acquire the cache chain mutex then just give up - we'll try
3832 * again on the next iteration.
3833 */
3834static void cache_reap(struct work_struct *w)
3835{
3836 struct kmem_cache *searchp;
3837 struct kmem_cache_node *n;
3838 int node = numa_mem_id();
3839 struct delayed_work *work = to_delayed_work(w);
3840
3841 if (!mutex_trylock(&slab_mutex))
3842 /* Give up. Setup the next iteration. */
3843 goto out;
3844
3845 list_for_each_entry(searchp, &slab_caches, list) {
3846 check_irq_on();
3847
3848 /*
3849 * We only take the node lock if absolutely necessary and we
3850 * have established with reasonable certainty that
3851 * we can do some work if the lock was obtained.
3852 */
3853 n = get_node(searchp, node);
3854
3855 reap_alien(searchp, n);
3856
3857 drain_array(searchp, n, cpu_cache_get(searchp), node);
3858
3859 /*
3860 * These are racy checks but it does not matter
3861 * if we skip one check or scan twice.
3862 */
3863 if (time_after(n->next_reap, jiffies))
3864 goto next;
3865
3866 n->next_reap = jiffies + REAPTIMEOUT_NODE;
3867
3868 drain_array(searchp, n, n->shared, node);
3869
3870 if (n->free_touched)
3871 n->free_touched = 0;
3872 else {
3873 int freed;
3874
3875 freed = drain_freelist(searchp, n, (n->free_limit +
3876 5 * searchp->num - 1) / (5 * searchp->num));
3877 STATS_ADD_REAPED(searchp, freed);
3878 }
3879next:
3880 cond_resched();
3881 }
3882 check_irq_on();
3883 mutex_unlock(&slab_mutex);
3884 next_reap_node();
3885out:
3886 /* Set up the next iteration */
3887 schedule_delayed_work_on(smp_processor_id(), work,
3888 round_jiffies_relative(REAPTIMEOUT_AC));
3889}
3890
3891void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
3892{
3893 unsigned long active_objs, num_objs, active_slabs;
3894 unsigned long total_slabs = 0, free_objs = 0, shared_avail = 0;
3895 unsigned long free_slabs = 0;
3896 int node;
3897 struct kmem_cache_node *n;
3898
3899 for_each_kmem_cache_node(cachep, node, n) {
3900 check_irq_on();
3901 raw_spin_lock_irq(&n->list_lock);
3902
3903 total_slabs += n->total_slabs;
3904 free_slabs += n->free_slabs;
3905 free_objs += n->free_objects;
3906
3907 if (n->shared)
3908 shared_avail += n->shared->avail;
3909
3910 raw_spin_unlock_irq(&n->list_lock);
3911 }
3912 num_objs = total_slabs * cachep->num;
3913 active_slabs = total_slabs - free_slabs;
3914 active_objs = num_objs - free_objs;
3915
3916 sinfo->active_objs = active_objs;
3917 sinfo->num_objs = num_objs;
3918 sinfo->active_slabs = active_slabs;
3919 sinfo->num_slabs = total_slabs;
3920 sinfo->shared_avail = shared_avail;
3921 sinfo->limit = cachep->limit;
3922 sinfo->batchcount = cachep->batchcount;
3923 sinfo->shared = cachep->shared;
3924 sinfo->objects_per_slab = cachep->num;
3925 sinfo->cache_order = cachep->gfporder;
3926}
3927
3928void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
3929{
3930#if STATS
3931 { /* node stats */
3932 unsigned long high = cachep->high_mark;
3933 unsigned long allocs = cachep->num_allocations;
3934 unsigned long grown = cachep->grown;
3935 unsigned long reaped = cachep->reaped;
3936 unsigned long errors = cachep->errors;
3937 unsigned long max_freeable = cachep->max_freeable;
3938 unsigned long node_allocs = cachep->node_allocs;
3939 unsigned long node_frees = cachep->node_frees;
3940 unsigned long overflows = cachep->node_overflow;
3941
3942 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu %4lu %4lu",
3943 allocs, high, grown,
3944 reaped, errors, max_freeable, node_allocs,
3945 node_frees, overflows);
3946 }
3947 /* cpu stats */
3948 {
3949 unsigned long allochit = atomic_read(&cachep->allochit);
3950 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
3951 unsigned long freehit = atomic_read(&cachep->freehit);
3952 unsigned long freemiss = atomic_read(&cachep->freemiss);
3953
3954 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
3955 allochit, allocmiss, freehit, freemiss);
3956 }
3957#endif
3958}
3959
3960#define MAX_SLABINFO_WRITE 128
3961/**
3962 * slabinfo_write - Tuning for the slab allocator
3963 * @file: unused
3964 * @buffer: user buffer
3965 * @count: data length
3966 * @ppos: unused
3967 *
3968 * Return: %0 on success, negative error code otherwise.
3969 */
3970ssize_t slabinfo_write(struct file *file, const char __user *buffer,
3971 size_t count, loff_t *ppos)
3972{
3973 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
3974 int limit, batchcount, shared, res;
3975 struct kmem_cache *cachep;
3976
3977 if (count > MAX_SLABINFO_WRITE)
3978 return -EINVAL;
3979 if (copy_from_user(&kbuf, buffer, count))
3980 return -EFAULT;
3981 kbuf[MAX_SLABINFO_WRITE] = '\0';
3982
3983 tmp = strchr(kbuf, ' ');
3984 if (!tmp)
3985 return -EINVAL;
3986 *tmp = '\0';
3987 tmp++;
3988 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
3989 return -EINVAL;
3990
3991 /* Find the cache in the chain of caches. */
3992 mutex_lock(&slab_mutex);
3993 res = -EINVAL;
3994 list_for_each_entry(cachep, &slab_caches, list) {
3995 if (!strcmp(cachep->name, kbuf)) {
3996 if (limit < 1 || batchcount < 1 ||
3997 batchcount > limit || shared < 0) {
3998 res = 0;
3999 } else {
4000 res = do_tune_cpucache(cachep, limit,
4001 batchcount, shared,
4002 GFP_KERNEL);
4003 }
4004 break;
4005 }
4006 }
4007 mutex_unlock(&slab_mutex);
4008 if (res >= 0)
4009 res = count;
4010 return res;
4011}
4012
4013#ifdef CONFIG_HARDENED_USERCOPY
4014/*
4015 * Rejects incorrectly sized objects and objects that are to be copied
4016 * to/from userspace but do not fall entirely within the containing slab
4017 * cache's usercopy region.
4018 *
4019 * Returns NULL if check passes, otherwise const char * to name of cache
4020 * to indicate an error.
4021 */
4022void __check_heap_object(const void *ptr, unsigned long n,
4023 const struct slab *slab, bool to_user)
4024{
4025 struct kmem_cache *cachep;
4026 unsigned int objnr;
4027 unsigned long offset;
4028
4029 ptr = kasan_reset_tag(ptr);
4030
4031 /* Find and validate object. */
4032 cachep = slab->slab_cache;
4033 objnr = obj_to_index(cachep, slab, (void *)ptr);
4034 BUG_ON(objnr >= cachep->num);
4035
4036 /* Find offset within object. */
4037 if (is_kfence_address(ptr))
4038 offset = ptr - kfence_object_start(ptr);
4039 else
4040 offset = ptr - index_to_obj(cachep, slab, objnr) - obj_offset(cachep);
4041
4042 /* Allow address range falling entirely within usercopy region. */
4043 if (offset >= cachep->useroffset &&
4044 offset - cachep->useroffset <= cachep->usersize &&
4045 n <= cachep->useroffset - offset + cachep->usersize)
4046 return;
4047
4048 usercopy_abort("SLAB object", cachep->name, to_user, offset, n);
4049}
4050#endif /* CONFIG_HARDENED_USERCOPY */
1/*
2 * linux/mm/slab.c
3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
5 *
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
7 *
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
10 *
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
13 *
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
21 *
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
27 *
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same initializations to
30 * kmem_cache_free.
31 *
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
35 *
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
38 * partial slabs
39 * empty slabs with no allocated objects
40 *
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
43 *
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
46 *
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
52 *
53 * The c_cpuarray may not be read with enabled local interrupts -
54 * it's changed with a smp_call_function().
55 *
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
58 * Several members in struct kmem_cache and struct slab never change, they
59 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
63 *
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
66 * his patch.
67 *
68 * Further notes from the original documentation:
69 *
70 * 11 April '97. Started multi-threading - markhe
71 * The global cache-chain is protected by the mutex 'slab_mutex'.
72 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
75 *
76 * At present, each engine can be growing a cache. This should be blocked.
77 *
78 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
83 *
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
87 */
88
89#include <linux/__KEEPIDENTS__B.h>
90#include <linux/__KEEPIDENTS__C.h>
91#include <linux/__KEEPIDENTS__D.h>
92#include <linux/__KEEPIDENTS__E.h>
93#include <linux/__KEEPIDENTS__F.h>
94#include <linux/__KEEPIDENTS__G.h>
95#include <linux/__KEEPIDENTS__H.h>
96#include <linux/__KEEPIDENTS__I.h>
97#include <linux/__KEEPIDENTS__J.h>
98#include <linux/proc_fs.h>
99#include <linux/__KEEPIDENTS__BA.h>
100#include <linux/__KEEPIDENTS__BB.h>
101#include <linux/__KEEPIDENTS__BC.h>
102#include <linux/cpu.h>
103#include <linux/__KEEPIDENTS__BD.h>
104#include <linux/__KEEPIDENTS__BE.h>
105#include <linux/rcupdate.h>
106#include <linux/__KEEPIDENTS__BF.h>
107#include <linux/__KEEPIDENTS__BG.h>
108#include <linux/__KEEPIDENTS__BH.h>
109#include <linux/kmemleak.h>
110#include <linux/__KEEPIDENTS__BI.h>
111#include <linux/__KEEPIDENTS__BJ.h>
112#include <linux/__KEEPIDENTS__CA-__KEEPIDENTS__CB.h>
113#include <linux/__KEEPIDENTS__CC.h>
114#include <linux/reciprocal_div.h>
115#include <linux/debugobjects.h>
116#include <linux/kmemcheck.h>
117#include <linux/__KEEPIDENTS__CD.h>
118#include <linux/__KEEPIDENTS__CE.h>
119
120#include <net/__KEEPIDENTS__CF.h>
121
122#include <asm/cacheflush.h>
123#include <asm/tlbflush.h>
124#include <asm/page.h>
125
126#include <trace/events/kmem.h>
127
128#include "internal.h"
129
130#include "slab.h"
131
132/*
133 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
134 * 0 for faster, smaller code (especially in the critical paths).
135 *
136 * STATS - 1 to collect stats for /proc/slabinfo.
137 * 0 for faster, smaller code (especially in the critical paths).
138 *
139 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
140 */
141
142#ifdef CONFIG_DEBUG_SLAB
143#define DEBUG 1
144#define STATS 1
145#define FORCED_DEBUG 1
146#else
147#define DEBUG 0
148#define STATS 0
149#define FORCED_DEBUG 0
150#endif
151
152/* Shouldn't this be in a header file somewhere? */
153#define BYTES_PER_WORD sizeof(void *)
154#define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
155
156#ifndef ARCH_KMALLOC_FLAGS
157#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
158#endif
159
160#define FREELIST_BYTE_INDEX (((PAGE_SIZE >> BITS_PER_BYTE) \
161 <= SLAB_OBJ_MIN_SIZE) ? 1 : 0)
162
163#if FREELIST_BYTE_INDEX
164typedef unsigned char freelist_idx_t;
165#else
166typedef unsigned short freelist_idx_t;
167#endif
168
169#define SLAB_OBJ_MAX_NUM ((1 << sizeof(freelist_idx_t) * BITS_PER_BYTE) - 1)
170
171/*
172 * struct array_cache
173 *
174 * Purpose:
175 * - LIFO ordering, to hand out cache-warm objects from _alloc
176 * - reduce the number of linked list operations
177 * - reduce spinlock operations
178 *
179 * The limit is stored in the per-cpu structure to reduce the data cache
180 * footprint.
181 *
182 */
183struct array_cache {
184 unsigned int avail;
185 unsigned int limit;
186 unsigned int batchcount;
187 unsigned int touched;
188 void *entry[]; /*
189 * Must have this definition in here for the proper
190 * alignment of array_cache. Also simplifies accessing
191 * the entries.
192 */
193};
194
195struct alien_cache {
196 spinlock_t lock;
197 struct array_cache ac;
198};
199
200/*
201 * Need this for bootstrapping a per node allocator.
202 */
203#define NUM_INIT_LISTS (2 * MAX_NUMNODES)
204static struct kmem_cache_node __initdata init_kmem_cache_node[NUM_INIT_LISTS];
205#define CACHE_CACHE 0
206#define SIZE_NODE (MAX_NUMNODES)
207
208static int drain_freelist(struct kmem_cache *cache,
209 struct kmem_cache_node *n, int tofree);
210static void free_block(struct kmem_cache *cachep, void **objpp, int len,
211 int node, struct list_head *list);
212static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list);
213static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
214static void cache_reap(struct work_struct *unused);
215
216static int slab_early_init = 1;
217
218#define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
219
220static void kmem_cache_node_init(struct kmem_cache_node *parent)
221{
222 INIT_LIST_HEAD(&parent->slabs_full);
223 INIT_LIST_HEAD(&parent->slabs_partial);
224 INIT_LIST_HEAD(&parent->slabs_free);
225 parent->shared = NULL;
226 parent->alien = NULL;
227 parent->colour_next = 0;
228 spin_lock_init(&parent->list_lock);
229 parent->free_objects = 0;
230 parent->free_touched = 0;
231}
232
233#define MAKE_LIST(cachep, listp, slab, nodeid) \
234 do { \
235 INIT_LIST_HEAD(listp); \
236 list_splice(&get_node(cachep, nodeid)->slab, listp); \
237 } while (0)
238
239#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
240 do { \
241 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
242 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
243 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
244 } while (0)
245
246#define CFLGS_OBJFREELIST_SLAB (0x40000000UL)
247#define CFLGS_OFF_SLAB (0x80000000UL)
248#define OBJFREELIST_SLAB(x) ((x)->flags & CFLGS_OBJFREELIST_SLAB)
249#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
250
251#define BATCHREFILL_LIMIT 16
252/*
253 * Optimization question: fewer reaps means less probability for unnessary
254 * cpucache drain/refill cycles.
255 *
256 * OTOH the cpuarrays can contain lots of objects,
257 * which could lock up otherwise freeable slabs.
258 */
259#define REAPTIMEOUT_AC (2*HZ)
260#define REAPTIMEOUT_NODE (4*HZ)
261
262#if STATS
263#define STATS_INC_ACTIVE(x) ((x)->num_active++)
264#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
265#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
266#define STATS_INC_GROWN(x) ((x)->grown++)
267#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
268#define STATS_SET_HIGH(x) \
269 do { \
270 if ((x)->num_active > (x)->high_mark) \
271 (x)->high_mark = (x)->num_active; \
272 } while (0)
273#define STATS_INC_ERR(x) ((x)->errors++)
274#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
275#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
276#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
277#define STATS_SET_FREEABLE(x, i) \
278 do { \
279 if ((x)->max_freeable < i) \
280 (x)->max_freeable = i; \
281 } while (0)
282#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
283#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
284#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
285#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
286#else
287#define STATS_INC_ACTIVE(x) do { } while (0)
288#define STATS_DEC_ACTIVE(x) do { } while (0)
289#define STATS_INC_ALLOCED(x) do { } while (0)
290#define STATS_INC_GROWN(x) do { } while (0)
291#define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
292#define STATS_SET_HIGH(x) do { } while (0)
293#define STATS_INC_ERR(x) do { } while (0)
294#define STATS_INC_NODEALLOCS(x) do { } while (0)
295#define STATS_INC_NODEFREES(x) do { } while (0)
296#define STATS_INC_ACOVERFLOW(x) do { } while (0)
297#define STATS_SET_FREEABLE(x, i) do { } while (0)
298#define STATS_INC_ALLOCHIT(x) do { } while (0)
299#define STATS_INC_ALLOCMISS(x) do { } while (0)
300#define STATS_INC_FREEHIT(x) do { } while (0)
301#define STATS_INC_FREEMISS(x) do { } while (0)
302#endif
303
304#if DEBUG
305
306/*
307 * memory layout of objects:
308 * 0 : objp
309 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
310 * the end of an object is aligned with the end of the real
311 * allocation. Catches writes behind the end of the allocation.
312 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
313 * redzone word.
314 * cachep->obj_offset: The real object.
315 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
316 * cachep->size - 1* BYTES_PER_WORD: last caller address
317 * [BYTES_PER_WORD long]
318 */
319static int obj_offset(struct kmem_cache *cachep)
320{
321 return cachep->obj_offset;
322}
323
324static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
325{
326 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
327 return (unsigned long long*) (objp + obj_offset(cachep) -
328 sizeof(unsigned long long));
329}
330
331static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
332{
333 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
334 if (cachep->flags & SLAB_STORE_USER)
335 return (unsigned long long *)(objp + cachep->size -
336 sizeof(unsigned long long) -
337 REDZONE_ALIGN);
338 return (unsigned long long *) (objp + cachep->size -
339 sizeof(unsigned long long));
340}
341
342static void **dbg_userword(struct kmem_cache *cachep, void *objp)
343{
344 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
345 return (void **)(objp + cachep->size - BYTES_PER_WORD);
346}
347
348#else
349
350#define obj_offset(x) 0
351#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
352#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
353#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
354
355#endif
356
357#ifdef CONFIG_DEBUG_SLAB_LEAK
358
359static inline bool is_store_user_clean(struct kmem_cache *cachep)
360{
361 return atomic_read(&cachep->store_user_clean) == 1;
362}
363
364static inline void set_store_user_clean(struct kmem_cache *cachep)
365{
366 atomic_set(&cachep->store_user_clean, 1);
367}
368
369static inline void set_store_user_dirty(struct kmem_cache *cachep)
370{
371 if (is_store_user_clean(cachep))
372 atomic_set(&cachep->store_user_clean, 0);
373}
374
375#else
376static inline void set_store_user_dirty(struct kmem_cache *cachep) {}
377
378#endif
379
380/*
381 * Do not go above this order unless 0 objects fit into the slab or
382 * overridden on the command line.
383 */
384#define SLAB_MAX_ORDER_HI 1
385#define SLAB_MAX_ORDER_LO 0
386static int slab_max_order = SLAB_MAX_ORDER_LO;
387static bool slab_max_order_set __initdata;
388
389static inline struct kmem_cache *virt_to_cache(const void *obj)
390{
391 struct page *page = virt_to_head_page(obj);
392 return page->slab_cache;
393}
394
395static inline void *index_to_obj(struct kmem_cache *cache, struct page *page,
396 unsigned int idx)
397{
398 return page->s_mem + cache->size * idx;
399}
400
401/*
402 * We want to avoid an expensive divide : (offset / cache->size)
403 * Using the fact that size is a constant for a particular cache,
404 * we can replace (offset / cache->size) by
405 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
406 */
407static inline unsigned int obj_to_index(const struct kmem_cache *cache,
408 const struct page *page, void *obj)
409{
410 u32 offset = (obj - page->s_mem);
411 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
412}
413
414#define BOOT_CPUCACHE_ENTRIES 1
415/* internal cache of cache description objs */
416static struct kmem_cache kmem_cache_boot = {
417 .batchcount = 1,
418 .limit = BOOT_CPUCACHE_ENTRIES,
419 .shared = 1,
420 .size = sizeof(struct kmem_cache),
421 .name = "kmem_cache",
422};
423
424#define BAD_ALIEN_MAGIC 0x01020304ul
425
426static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
427
428static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
429{
430 return this_cpu_ptr(cachep->cpu_cache);
431}
432
433/*
434 * Calculate the number of objects and left-over bytes for a given buffer size.
435 */
436static unsigned int cache_estimate(unsigned long gfporder, size_t buffer_size,
437 unsigned long flags, size_t *left_over)
438{
439 unsigned int num;
440 size_t slab_size = PAGE_SIZE << gfporder;
441
442 /*
443 * The slab management structure can be either off the slab or
444 * on it. For the latter case, the memory allocated for a
445 * slab is used for:
446 *
447 * - @buffer_size bytes for each object
448 * - One freelist_idx_t for each object
449 *
450 * We don't need to consider alignment of freelist because
451 * freelist will be at the end of slab page. The objects will be
452 * at the correct alignment.
453 *
454 * If the slab management structure is off the slab, then the
455 * alignment will already be calculated into the size. Because
456 * the slabs are all pages aligned, the objects will be at the
457 * correct alignment when allocated.
458 */
459 if (flags & (CFLGS_OBJFREELIST_SLAB | CFLGS_OFF_SLAB)) {
460 num = slab_size / buffer_size;
461 *left_over = slab_size % buffer_size;
462 } else {
463 num = slab_size / (buffer_size + sizeof(freelist_idx_t));
464 *left_over = slab_size %
465 (buffer_size + sizeof(freelist_idx_t));
466 }
467
468 return num;
469}
470
471#if DEBUG
472#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
473
474static void __slab_error(const char *function, struct kmem_cache *cachep,
475 char *msg)
476{
477 pr_err("slab error in %s(): cache `%s': %s\n",
478 function, cachep->name, msg);
479 dump_stack();
480 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
481}
482#endif
483
484/*
485 * By default on NUMA we use alien caches to stage the freeing of
486 * objects allocated from other nodes. This causes massive memory
487 * inefficiencies when using fake NUMA setup to split memory into a
488 * large number of small nodes, so it can be disabled on the command
489 * line
490 */
491
492static int use_alien_caches __read_mostly = 1;
493static int __init noaliencache_setup(char *s)
494{
495 use_alien_caches = 0;
496 return 1;
497}
498__setup("noaliencache", noaliencache_setup);
499
500static int __init slab_max_order_setup(char *str)
501{
502 get_option(&str, &slab_max_order);
503 slab_max_order = slab_max_order < 0 ? 0 :
504 min(slab_max_order, MAX_ORDER - 1);
505 slab_max_order_set = true;
506
507 return 1;
508}
509__setup("slab_max_order=", slab_max_order_setup);
510
511#ifdef CONFIG_NUMA
512/*
513 * Special reaping functions for NUMA systems called from cache_reap().
514 * These take care of doing round robin flushing of alien caches (containing
515 * objects freed on different nodes from which they were allocated) and the
516 * flushing of remote pcps by calling drain_node_pages.
517 */
518static DEFINE_PER_CPU(unsigned long, slab_reap_node);
519
520static void init_reap_node(int cpu)
521{
522 int node;
523
524 node = next_node(cpu_to_mem(cpu), node_online_map);
525 if (node == MAX_NUMNODES)
526 node = first_node(node_online_map);
527
528 per_cpu(slab_reap_node, cpu) = node;
529}
530
531static void next_reap_node(void)
532{
533 int node = __this_cpu_read(slab_reap_node);
534
535 node = next_node(node, node_online_map);
536 if (unlikely(node >= MAX_NUMNODES))
537 node = first_node(node_online_map);
538 __this_cpu_write(slab_reap_node, node);
539}
540
541#else
542#define init_reap_node(cpu) do { } while (0)
543#define next_reap_node(void) do { } while (0)
544#endif
545
546/*
547 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
548 * via the workqueue/eventd.
549 * Add the CPU number into the expiration time to minimize the possibility of
550 * the CPUs getting into lockstep and contending for the global cache chain
551 * lock.
552 */
553static void start_cpu_timer(int cpu)
554{
555 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
556
557 /*
558 * When this gets called from do_initcalls via cpucache_init(),
559 * init_workqueues() has already run, so keventd will be setup
560 * at that time.
561 */
562 if (keventd_up() && reap_work->work.func == NULL) {
563 init_reap_node(cpu);
564 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
565 schedule_delayed_work_on(cpu, reap_work,
566 __round_jiffies_relative(HZ, cpu));
567 }
568}
569
570static void init_arraycache(struct array_cache *ac, int limit, int batch)
571{
572 /*
573 * The array_cache structures contain pointers to free object.
574 * However, when such objects are allocated or transferred to another
575 * cache the pointers are not cleared and they could be counted as
576 * valid references during a kmemleak scan. Therefore, kmemleak must
577 * not scan such objects.
578 */
579 kmemleak_no_scan(ac);
580 if (ac) {
581 ac->avail = 0;
582 ac->limit = limit;
583 ac->batchcount = batch;
584 ac->touched = 0;
585 }
586}
587
588static struct array_cache *alloc_arraycache(int node, int entries,
589 int batchcount, gfp_t gfp)
590{
591 size_t memsize = sizeof(void *) * entries + sizeof(struct array_cache);
592 struct array_cache *ac = NULL;
593
594 ac = kmalloc_node(memsize, gfp, node);
595 init_arraycache(ac, entries, batchcount);
596 return ac;
597}
598
599static noinline void cache_free_pfmemalloc(struct kmem_cache *cachep,
600 struct page *page, void *objp)
601{
602 struct kmem_cache_node *n;
603 int page_node;
604 LIST_HEAD(list);
605
606 page_node = page_to_nid(page);
607 n = get_node(cachep, page_node);
608
609 spin_lock(&n->list_lock);
610 free_block(cachep, &objp, 1, page_node, &list);
611 spin_unlock(&n->list_lock);
612
613 slabs_destroy(cachep, &list);
614}
615
616/*
617 * Transfer objects in one arraycache to another.
618 * Locking must be handled by the caller.
619 *
620 * Return the number of entries transferred.
621 */
622static int transfer_objects(struct array_cache *to,
623 struct array_cache *from, unsigned int max)
624{
625 /* Figure out how many entries to transfer */
626 int nr = min3(from->avail, max, to->limit - to->avail);
627
628 if (!nr)
629 return 0;
630
631 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
632 sizeof(void *) *nr);
633
634 from->avail -= nr;
635 to->avail += nr;
636 return nr;
637}
638
639#ifndef CONFIG_NUMA
640
641#define drain_alien_cache(cachep, alien) do { } while (0)
642#define reap_alien(cachep, n) do { } while (0)
643
644static inline struct alien_cache **alloc_alien_cache(int node,
645 int limit, gfp_t gfp)
646{
647 return (struct alien_cache **)BAD_ALIEN_MAGIC;
648}
649
650static inline void free_alien_cache(struct alien_cache **ac_ptr)
651{
652}
653
654static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
655{
656 return 0;
657}
658
659static inline void *alternate_node_alloc(struct kmem_cache *cachep,
660 gfp_t flags)
661{
662 return NULL;
663}
664
665static inline void *____cache_alloc_node(struct kmem_cache *cachep,
666 gfp_t flags, int nodeid)
667{
668 return NULL;
669}
670
671static inline gfp_t gfp_exact_node(gfp_t flags)
672{
673 return flags & ~__GFP_NOFAIL;
674}
675
676#else /* CONFIG_NUMA */
677
678static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
679static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
680
681static struct alien_cache *__alloc_alien_cache(int node, int entries,
682 int batch, gfp_t gfp)
683{
684 size_t memsize = sizeof(void *) * entries + sizeof(struct alien_cache);
685 struct alien_cache *alc = NULL;
686
687 alc = kmalloc_node(memsize, gfp, node);
688 init_arraycache(&alc->ac, entries, batch);
689 spin_lock_init(&alc->lock);
690 return alc;
691}
692
693static struct alien_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
694{
695 struct alien_cache **alc_ptr;
696 size_t memsize = sizeof(void *) * nr_node_ids;
697 int i;
698
699 if (limit > 1)
700 limit = 12;
701 alc_ptr = kzalloc_node(memsize, gfp, node);
702 if (!alc_ptr)
703 return NULL;
704
705 for_each_node(i) {
706 if (i == node || !node_online(i))
707 continue;
708 alc_ptr[i] = __alloc_alien_cache(node, limit, 0xbaadf00d, gfp);
709 if (!alc_ptr[i]) {
710 for (i--; i >= 0; i--)
711 kfree(alc_ptr[i]);
712 kfree(alc_ptr);
713 return NULL;
714 }
715 }
716 return alc_ptr;
717}
718
719static void free_alien_cache(struct alien_cache **alc_ptr)
720{
721 int i;
722
723 if (!alc_ptr)
724 return;
725 for_each_node(i)
726 kfree(alc_ptr[i]);
727 kfree(alc_ptr);
728}
729
730static void __drain_alien_cache(struct kmem_cache *cachep,
731 struct array_cache *ac, int node,
732 struct list_head *list)
733{
734 struct kmem_cache_node *n = get_node(cachep, node);
735
736 if (ac->avail) {
737 spin_lock(&n->list_lock);
738 /*
739 * Stuff objects into the remote nodes shared array first.
740 * That way we could avoid the overhead of putting the objects
741 * into the free lists and getting them back later.
742 */
743 if (n->shared)
744 transfer_objects(n->shared, ac, ac->limit);
745
746 free_block(cachep, ac->entry, ac->avail, node, list);
747 ac->avail = 0;
748 spin_unlock(&n->list_lock);
749 }
750}
751
752/*
753 * Called from cache_reap() to regularly drain alien caches round robin.
754 */
755static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
756{
757 int node = __this_cpu_read(slab_reap_node);
758
759 if (n->alien) {
760 struct alien_cache *alc = n->alien[node];
761 struct array_cache *ac;
762
763 if (alc) {
764 ac = &alc->ac;
765 if (ac->avail && spin_trylock_irq(&alc->lock)) {
766 LIST_HEAD(list);
767
768 __drain_alien_cache(cachep, ac, node, &list);
769 spin_unlock_irq(&alc->lock);
770 slabs_destroy(cachep, &list);
771 }
772 }
773 }
774}
775
776static void drain_alien_cache(struct kmem_cache *cachep,
777 struct alien_cache **alien)
778{
779 int i = 0;
780 struct alien_cache *alc;
781 struct array_cache *ac;
782 unsigned long flags;
783
784 for_each_online_node(i) {
785 alc = alien[i];
786 if (alc) {
787 LIST_HEAD(list);
788
789 ac = &alc->ac;
790 spin_lock_irqsave(&alc->lock, flags);
791 __drain_alien_cache(cachep, ac, i, &list);
792 spin_unlock_irqrestore(&alc->lock, flags);
793 slabs_destroy(cachep, &list);
794 }
795 }
796}
797
798static int __cache_free_alien(struct kmem_cache *cachep, void *objp,
799 int node, int page_node)
800{
801 struct kmem_cache_node *n;
802 struct alien_cache *alien = NULL;
803 struct array_cache *ac;
804 LIST_HEAD(list);
805
806 n = get_node(cachep, node);
807 STATS_INC_NODEFREES(cachep);
808 if (n->alien && n->alien[page_node]) {
809 alien = n->alien[page_node];
810 ac = &alien->ac;
811 spin_lock(&alien->lock);
812 if (unlikely(ac->avail == ac->limit)) {
813 STATS_INC_ACOVERFLOW(cachep);
814 __drain_alien_cache(cachep, ac, page_node, &list);
815 }
816 ac->entry[ac->avail++] = objp;
817 spin_unlock(&alien->lock);
818 slabs_destroy(cachep, &list);
819 } else {
820 n = get_node(cachep, page_node);
821 spin_lock(&n->list_lock);
822 free_block(cachep, &objp, 1, page_node, &list);
823 spin_unlock(&n->list_lock);
824 slabs_destroy(cachep, &list);
825 }
826 return 1;
827}
828
829static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
830{
831 int page_node = page_to_nid(virt_to_page(objp));
832 int node = numa_mem_id();
833 /*
834 * Make sure we are not freeing a object from another node to the array
835 * cache on this cpu.
836 */
837 if (likely(node == page_node))
838 return 0;
839
840 return __cache_free_alien(cachep, objp, node, page_node);
841}
842
843/*
844 * Construct gfp mask to allocate from a specific node but do not reclaim or
845 * warn about failures.
846 */
847static inline gfp_t gfp_exact_node(gfp_t flags)
848{
849 return (flags | __GFP_THISNODE | __GFP_NOWARN) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
850}
851#endif
852
853/*
854 * Allocates and initializes node for a node on each slab cache, used for
855 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
856 * will be allocated off-node since memory is not yet online for the new node.
857 * When hotplugging memory or a cpu, existing node are not replaced if
858 * already in use.
859 *
860 * Must hold slab_mutex.
861 */
862static int init_cache_node_node(int node)
863{
864 struct kmem_cache *cachep;
865 struct kmem_cache_node *n;
866 const size_t memsize = sizeof(struct kmem_cache_node);
867
868 list_for_each_entry(cachep, &slab_caches, list) {
869 /*
870 * Set up the kmem_cache_node for cpu before we can
871 * begin anything. Make sure some other cpu on this
872 * node has not already allocated this
873 */
874 n = get_node(cachep, node);
875 if (!n) {
876 n = kmalloc_node(memsize, GFP_KERNEL, node);
877 if (!n)
878 return -ENOMEM;
879 kmem_cache_node_init(n);
880 n->next_reap = jiffies + REAPTIMEOUT_NODE +
881 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
882
883 /*
884 * The kmem_cache_nodes don't come and go as CPUs
885 * come and go. slab_mutex is sufficient
886 * protection here.
887 */
888 cachep->node[node] = n;
889 }
890
891 spin_lock_irq(&n->list_lock);
892 n->free_limit =
893 (1 + nr_cpus_node(node)) *
894 cachep->batchcount + cachep->num;
895 spin_unlock_irq(&n->list_lock);
896 }
897 return 0;
898}
899
900static inline int slabs_tofree(struct kmem_cache *cachep,
901 struct kmem_cache_node *n)
902{
903 return (n->free_objects + cachep->num - 1) / cachep->num;
904}
905
906static void cpuup_canceled(long cpu)
907{
908 struct kmem_cache *cachep;
909 struct kmem_cache_node *n = NULL;
910 int node = cpu_to_mem(cpu);
911 const struct cpumask *mask = cpumask_of_node(node);
912
913 list_for_each_entry(cachep, &slab_caches, list) {
914 struct array_cache *nc;
915 struct array_cache *shared;
916 struct alien_cache **alien;
917 LIST_HEAD(list);
918
919 n = get_node(cachep, node);
920 if (!n)
921 continue;
922
923 spin_lock_irq(&n->list_lock);
924
925 /* Free limit for this kmem_cache_node */
926 n->free_limit -= cachep->batchcount;
927
928 /* cpu is dead; no one can alloc from it. */
929 nc = per_cpu_ptr(cachep->cpu_cache, cpu);
930 if (nc) {
931 free_block(cachep, nc->entry, nc->avail, node, &list);
932 nc->avail = 0;
933 }
934
935 if (!cpumask_empty(mask)) {
936 spin_unlock_irq(&n->list_lock);
937 goto free_slab;
938 }
939
940 shared = n->shared;
941 if (shared) {
942 free_block(cachep, shared->entry,
943 shared->avail, node, &list);
944 n->shared = NULL;
945 }
946
947 alien = n->alien;
948 n->alien = NULL;
949
950 spin_unlock_irq(&n->list_lock);
951
952 kfree(shared);
953 if (alien) {
954 drain_alien_cache(cachep, alien);
955 free_alien_cache(alien);
956 }
957
958free_slab:
959 slabs_destroy(cachep, &list);
960 }
961 /*
962 * In the previous loop, all the objects were freed to
963 * the respective cache's slabs, now we can go ahead and
964 * shrink each nodelist to its limit.
965 */
966 list_for_each_entry(cachep, &slab_caches, list) {
967 n = get_node(cachep, node);
968 if (!n)
969 continue;
970 drain_freelist(cachep, n, slabs_tofree(cachep, n));
971 }
972}
973
974static int cpuup_prepare(long cpu)
975{
976 struct kmem_cache *cachep;
977 struct kmem_cache_node *n = NULL;
978 int node = cpu_to_mem(cpu);
979 int err;
980
981 /*
982 * We need to do this right in the beginning since
983 * alloc_arraycache's are going to use this list.
984 * kmalloc_node allows us to add the slab to the right
985 * kmem_cache_node and not this cpu's kmem_cache_node
986 */
987 err = init_cache_node_node(node);
988 if (err < 0)
989 goto bad;
990
991 /*
992 * Now we can go ahead with allocating the shared arrays and
993 * array caches
994 */
995 list_for_each_entry(cachep, &slab_caches, list) {
996 struct array_cache *shared = NULL;
997 struct alien_cache **alien = NULL;
998
999 if (cachep->shared) {
1000 shared = alloc_arraycache(node,
1001 cachep->shared * cachep->batchcount,
1002 0xbaadf00d, GFP_KERNEL);
1003 if (!shared)
1004 goto bad;
1005 }
1006 if (use_alien_caches) {
1007 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1008 if (!alien) {
1009 kfree(shared);
1010 goto bad;
1011 }
1012 }
1013 n = get_node(cachep, node);
1014 BUG_ON(!n);
1015
1016 spin_lock_irq(&n->list_lock);
1017 if (!n->shared) {
1018 /*
1019 * We are serialised from CPU_DEAD or
1020 * CPU_UP_CANCELLED by the cpucontrol lock
1021 */
1022 n->shared = shared;
1023 shared = NULL;
1024 }
1025#ifdef CONFIG_NUMA
1026 if (!n->alien) {
1027 n->alien = alien;
1028 alien = NULL;
1029 }
1030#endif
1031 spin_unlock_irq(&n->list_lock);
1032 kfree(shared);
1033 free_alien_cache(alien);
1034 }
1035
1036 return 0;
1037bad:
1038 cpuup_canceled(cpu);
1039 return -ENOMEM;
1040}
1041
1042static int cpuup_callback(struct notifier_block *nfb,
1043 unsigned long action, void *hcpu)
1044{
1045 long cpu = (long)hcpu;
1046 int err = 0;
1047
1048 switch (action) {
1049 case CPU_UP_PREPARE:
1050 case CPU_UP_PREPARE_FROZEN:
1051 mutex_lock(&slab_mutex);
1052 err = cpuup_prepare(cpu);
1053 mutex_unlock(&slab_mutex);
1054 break;
1055 case CPU_ONLINE:
1056 case CPU_ONLINE_FROZEN:
1057 start_cpu_timer(cpu);
1058 break;
1059#ifdef CONFIG_HOTPLUG_CPU
1060 case CPU_DOWN_PREPARE:
1061 case CPU_DOWN_PREPARE_FROZEN:
1062 /*
1063 * Shutdown cache reaper. Note that the slab_mutex is
1064 * held so that if cache_reap() is invoked it cannot do
1065 * anything expensive but will only modify reap_work
1066 * and reschedule the timer.
1067 */
1068 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1069 /* Now the cache_reaper is guaranteed to be not running. */
1070 per_cpu(slab_reap_work, cpu).work.func = NULL;
1071 break;
1072 case CPU_DOWN_FAILED:
1073 case CPU_DOWN_FAILED_FROZEN:
1074 start_cpu_timer(cpu);
1075 break;
1076 case CPU_DEAD:
1077 case CPU_DEAD_FROZEN:
1078 /*
1079 * Even if all the cpus of a node are down, we don't free the
1080 * kmem_cache_node of any cache. This to avoid a race between
1081 * cpu_down, and a kmalloc allocation from another cpu for
1082 * memory from the node of the cpu going down. The node
1083 * structure is usually allocated from kmem_cache_create() and
1084 * gets destroyed at kmem_cache_destroy().
1085 */
1086 /* fall through */
1087#endif
1088 case CPU_UP_CANCELED:
1089 case CPU_UP_CANCELED_FROZEN:
1090 mutex_lock(&slab_mutex);
1091 cpuup_canceled(cpu);
1092 mutex_unlock(&slab_mutex);
1093 break;
1094 }
1095 return notifier_from_errno(err);
1096}
1097
1098static struct notifier_block cpucache_notifier = {
1099 &cpuup_callback, NULL, 0
1100};
1101
1102#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1103/*
1104 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1105 * Returns -EBUSY if all objects cannot be drained so that the node is not
1106 * removed.
1107 *
1108 * Must hold slab_mutex.
1109 */
1110static int __meminit drain_cache_node_node(int node)
1111{
1112 struct kmem_cache *cachep;
1113 int ret = 0;
1114
1115 list_for_each_entry(cachep, &slab_caches, list) {
1116 struct kmem_cache_node *n;
1117
1118 n = get_node(cachep, node);
1119 if (!n)
1120 continue;
1121
1122 drain_freelist(cachep, n, slabs_tofree(cachep, n));
1123
1124 if (!list_empty(&n->slabs_full) ||
1125 !list_empty(&n->slabs_partial)) {
1126 ret = -EBUSY;
1127 break;
1128 }
1129 }
1130 return ret;
1131}
1132
1133static int __meminit slab_memory_callback(struct notifier_block *self,
1134 unsigned long action, void *arg)
1135{
1136 struct memory_notify *mnb = arg;
1137 int ret = 0;
1138 int nid;
1139
1140 nid = mnb->status_change_nid;
1141 if (nid < 0)
1142 goto out;
1143
1144 switch (action) {
1145 case MEM_GOING_ONLINE:
1146 mutex_lock(&slab_mutex);
1147 ret = init_cache_node_node(nid);
1148 mutex_unlock(&slab_mutex);
1149 break;
1150 case MEM_GOING_OFFLINE:
1151 mutex_lock(&slab_mutex);
1152 ret = drain_cache_node_node(nid);
1153 mutex_unlock(&slab_mutex);
1154 break;
1155 case MEM_ONLINE:
1156 case MEM_OFFLINE:
1157 case MEM_CANCEL_ONLINE:
1158 case MEM_CANCEL_OFFLINE:
1159 break;
1160 }
1161out:
1162 return notifier_from_errno(ret);
1163}
1164#endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1165
1166/*
1167 * swap the static kmem_cache_node with kmalloced memory
1168 */
1169static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
1170 int nodeid)
1171{
1172 struct kmem_cache_node *ptr;
1173
1174 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
1175 BUG_ON(!ptr);
1176
1177 memcpy(ptr, list, sizeof(struct kmem_cache_node));
1178 /*
1179 * Do not assume that spinlocks can be initialized via memcpy:
1180 */
1181 spin_lock_init(&ptr->list_lock);
1182
1183 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1184 cachep->node[nodeid] = ptr;
1185}
1186
1187/*
1188 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1189 * size of kmem_cache_node.
1190 */
1191static void __init set_up_node(struct kmem_cache *cachep, int index)
1192{
1193 int node;
1194
1195 for_each_online_node(node) {
1196 cachep->node[node] = &init_kmem_cache_node[index + node];
1197 cachep->node[node]->next_reap = jiffies +
1198 REAPTIMEOUT_NODE +
1199 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1200 }
1201}
1202
1203/*
1204 * Initialisation. Called after the page allocator have been initialised and
1205 * before smp_init().
1206 */
1207void __init kmem_cache_init(void)
1208{
1209 int i;
1210
1211 BUILD_BUG_ON(sizeof(((struct page *)NULL)->lru) <
1212 sizeof(struct rcu_head));
1213 kmem_cache = &kmem_cache_boot;
1214
1215 if (num_possible_nodes() == 1)
1216 use_alien_caches = 0;
1217
1218 for (i = 0; i < NUM_INIT_LISTS; i++)
1219 kmem_cache_node_init(&init_kmem_cache_node[i]);
1220
1221 /*
1222 * Fragmentation resistance on low memory - only use bigger
1223 * page orders on machines with more than 32MB of memory if
1224 * not overridden on the command line.
1225 */
1226 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
1227 slab_max_order = SLAB_MAX_ORDER_HI;
1228
1229 /* Bootstrap is tricky, because several objects are allocated
1230 * from caches that do not exist yet:
1231 * 1) initialize the kmem_cache cache: it contains the struct
1232 * kmem_cache structures of all caches, except kmem_cache itself:
1233 * kmem_cache is statically allocated.
1234 * Initially an __init data area is used for the head array and the
1235 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1236 * array at the end of the bootstrap.
1237 * 2) Create the first kmalloc cache.
1238 * The struct kmem_cache for the new cache is allocated normally.
1239 * An __init data area is used for the head array.
1240 * 3) Create the remaining kmalloc caches, with minimally sized
1241 * head arrays.
1242 * 4) Replace the __init data head arrays for kmem_cache and the first
1243 * kmalloc cache with kmalloc allocated arrays.
1244 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1245 * the other cache's with kmalloc allocated memory.
1246 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1247 */
1248
1249 /* 1) create the kmem_cache */
1250
1251 /*
1252 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1253 */
1254 create_boot_cache(kmem_cache, "kmem_cache",
1255 offsetof(struct kmem_cache, node) +
1256 nr_node_ids * sizeof(struct kmem_cache_node *),
1257 SLAB_HWCACHE_ALIGN);
1258 list_add(&kmem_cache->list, &slab_caches);
1259 slab_state = PARTIAL;
1260
1261 /*
1262 * Initialize the caches that provide memory for the kmem_cache_node
1263 * structures first. Without this, further allocations will bug.
1264 */
1265 kmalloc_caches[INDEX_NODE] = create_kmalloc_cache("kmalloc-node",
1266 kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS);
1267 slab_state = PARTIAL_NODE;
1268 setup_kmalloc_cache_index_table();
1269
1270 slab_early_init = 0;
1271
1272 /* 5) Replace the bootstrap kmem_cache_node */
1273 {
1274 int nid;
1275
1276 for_each_online_node(nid) {
1277 init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
1278
1279 init_list(kmalloc_caches[INDEX_NODE],
1280 &init_kmem_cache_node[SIZE_NODE + nid], nid);
1281 }
1282 }
1283
1284 create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
1285}
1286
1287void __init kmem_cache_init_late(void)
1288{
1289 struct kmem_cache *cachep;
1290
1291 slab_state = UP;
1292
1293 /* 6) resize the head arrays to their final sizes */
1294 mutex_lock(&slab_mutex);
1295 list_for_each_entry(cachep, &slab_caches, list)
1296 if (enable_cpucache(cachep, GFP_NOWAIT))
1297 BUG();
1298 mutex_unlock(&slab_mutex);
1299
1300 /* Done! */
1301 slab_state = FULL;
1302
1303 /*
1304 * Register a cpu startup notifier callback that initializes
1305 * cpu_cache_get for all new cpus
1306 */
1307 register_cpu_notifier(&cpucache_notifier);
1308
1309#ifdef CONFIG_NUMA
1310 /*
1311 * Register a memory hotplug callback that initializes and frees
1312 * node.
1313 */
1314 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1315#endif
1316
1317 /*
1318 * The reap timers are started later, with a module init call: That part
1319 * of the kernel is not yet operational.
1320 */
1321}
1322
1323static int __init cpucache_init(void)
1324{
1325 int cpu;
1326
1327 /*
1328 * Register the timers that return unneeded pages to the page allocator
1329 */
1330 for_each_online_cpu(cpu)
1331 start_cpu_timer(cpu);
1332
1333 /* Done! */
1334 slab_state = FULL;
1335 return 0;
1336}
1337__initcall(cpucache_init);
1338
1339static noinline void
1340slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1341{
1342#if DEBUG
1343 struct kmem_cache_node *n;
1344 struct page *page;
1345 unsigned long flags;
1346 int node;
1347 static DEFINE_RATELIMIT_STATE(slab_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1348 DEFAULT_RATELIMIT_BURST);
1349
1350 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slab_oom_rs))
1351 return;
1352
1353 pr_warn("SLAB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
1354 nodeid, gfpflags, &gfpflags);
1355 pr_warn(" cache: %s, object size: %d, order: %d\n",
1356 cachep->name, cachep->size, cachep->gfporder);
1357
1358 for_each_kmem_cache_node(cachep, node, n) {
1359 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1360 unsigned long active_slabs = 0, num_slabs = 0;
1361
1362 spin_lock_irqsave(&n->list_lock, flags);
1363 list_for_each_entry(page, &n->slabs_full, lru) {
1364 active_objs += cachep->num;
1365 active_slabs++;
1366 }
1367 list_for_each_entry(page, &n->slabs_partial, lru) {
1368 active_objs += page->active;
1369 active_slabs++;
1370 }
1371 list_for_each_entry(page, &n->slabs_free, lru)
1372 num_slabs++;
1373
1374 free_objects += n->free_objects;
1375 spin_unlock_irqrestore(&n->list_lock, flags);
1376
1377 num_slabs += active_slabs;
1378 num_objs = num_slabs * cachep->num;
1379 pr_warn(" node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1380 node, active_slabs, num_slabs, active_objs, num_objs,
1381 free_objects);
1382 }
1383#endif
1384}
1385
1386/*
1387 * Interface to system's page allocator. No need to hold the
1388 * kmem_cache_node ->list_lock.
1389 *
1390 * If we requested dmaable memory, we will get it. Even if we
1391 * did not request dmaable memory, we might get it, but that
1392 * would be relatively rare and ignorable.
1393 */
1394static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1395 int nodeid)
1396{
1397 struct page *page;
1398 int nr_pages;
1399
1400 flags |= cachep->allocflags;
1401 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1402 flags |= __GFP_RECLAIMABLE;
1403
1404 page = __alloc_pages_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1405 if (!page) {
1406 slab_out_of_memory(cachep, flags, nodeid);
1407 return NULL;
1408 }
1409
1410 if (memcg_charge_slab(page, flags, cachep->gfporder, cachep)) {
1411 __free_pages(page, cachep->gfporder);
1412 return NULL;
1413 }
1414
1415 nr_pages = (1 << cachep->gfporder);
1416 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1417 add_zone_page_state(page_zone(page),
1418 NR_SLAB_RECLAIMABLE, nr_pages);
1419 else
1420 add_zone_page_state(page_zone(page),
1421 NR_SLAB_UNRECLAIMABLE, nr_pages);
1422
1423 __SetPageSlab(page);
1424 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1425 if (sk_memalloc_socks() && page_is_pfmemalloc(page))
1426 SetPageSlabPfmemalloc(page);
1427
1428 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1429 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1430
1431 if (cachep->ctor)
1432 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1433 else
1434 kmemcheck_mark_unallocated_pages(page, nr_pages);
1435 }
1436
1437 return page;
1438}
1439
1440/*
1441 * Interface to system's page release.
1442 */
1443static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
1444{
1445 int order = cachep->gfporder;
1446 unsigned long nr_freed = (1 << order);
1447
1448 kmemcheck_free_shadow(page, order);
1449
1450 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1451 sub_zone_page_state(page_zone(page),
1452 NR_SLAB_RECLAIMABLE, nr_freed);
1453 else
1454 sub_zone_page_state(page_zone(page),
1455 NR_SLAB_UNRECLAIMABLE, nr_freed);
1456
1457 BUG_ON(!PageSlab(page));
1458 __ClearPageSlabPfmemalloc(page);
1459 __ClearPageSlab(page);
1460 page_mapcount_reset(page);
1461 page->mapping = NULL;
1462
1463 if (current->reclaim_state)
1464 current->reclaim_state->reclaimed_slab += nr_freed;
1465 memcg_uncharge_slab(page, order, cachep);
1466 __free_pages(page, order);
1467}
1468
1469static void kmem_rcu_free(struct rcu_head *head)
1470{
1471 struct kmem_cache *cachep;
1472 struct page *page;
1473
1474 page = container_of(head, struct page, rcu_head);
1475 cachep = page->slab_cache;
1476
1477 kmem_freepages(cachep, page);
1478}
1479
1480#if DEBUG
1481static bool is_debug_pagealloc_cache(struct kmem_cache *cachep)
1482{
1483 if (debug_pagealloc_enabled() && OFF_SLAB(cachep) &&
1484 (cachep->size % PAGE_SIZE) == 0)
1485 return true;
1486
1487 return false;
1488}
1489
1490#ifdef CONFIG_DEBUG_PAGEALLOC
1491static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1492 unsigned long caller)
1493{
1494 int size = cachep->object_size;
1495
1496 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1497
1498 if (size < 5 * sizeof(unsigned long))
1499 return;
1500
1501 *addr++ = 0x12345678;
1502 *addr++ = caller;
1503 *addr++ = smp_processor_id();
1504 size -= 3 * sizeof(unsigned long);
1505 {
1506 unsigned long *sptr = &caller;
1507 unsigned long svalue;
1508
1509 while (!kstack_end(sptr)) {
1510 svalue = *sptr++;
1511 if (kernel_text_address(svalue)) {
1512 *addr++ = svalue;
1513 size -= sizeof(unsigned long);
1514 if (size <= sizeof(unsigned long))
1515 break;
1516 }
1517 }
1518
1519 }
1520 *addr++ = 0x87654321;
1521}
1522
1523static void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1524 int map, unsigned long caller)
1525{
1526 if (!is_debug_pagealloc_cache(cachep))
1527 return;
1528
1529 if (caller)
1530 store_stackinfo(cachep, objp, caller);
1531
1532 kernel_map_pages(virt_to_page(objp), cachep->size / PAGE_SIZE, map);
1533}
1534
1535#else
1536static inline void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1537 int map, unsigned long caller) {}
1538
1539#endif
1540
1541static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1542{
1543 int size = cachep->object_size;
1544 addr = &((char *)addr)[obj_offset(cachep)];
1545
1546 memset(addr, val, size);
1547 *(unsigned char *)(addr + size - 1) = POISON_END;
1548}
1549
1550static void dump_line(char *data, int offset, int limit)
1551{
1552 int i;
1553 unsigned char error = 0;
1554 int bad_count = 0;
1555
1556 pr_err("%03x: ", offset);
1557 for (i = 0; i < limit; i++) {
1558 if (data[offset + i] != POISON_FREE) {
1559 error = data[offset + i];
1560 bad_count++;
1561 }
1562 }
1563 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1564 &data[offset], limit, 1);
1565
1566 if (bad_count == 1) {
1567 error ^= POISON_FREE;
1568 if (!(error & (error - 1))) {
1569 pr_err("Single bit error detected. Probably bad RAM.\n");
1570#ifdef CONFIG_X86
1571 pr_err("Run memtest86+ or a similar memory test tool.\n");
1572#else
1573 pr_err("Run a memory test tool.\n");
1574#endif
1575 }
1576 }
1577}
1578#endif
1579
1580#if DEBUG
1581
1582static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1583{
1584 int i, size;
1585 char *realobj;
1586
1587 if (cachep->flags & SLAB_RED_ZONE) {
1588 pr_err("Redzone: 0x%llx/0x%llx\n",
1589 *dbg_redzone1(cachep, objp),
1590 *dbg_redzone2(cachep, objp));
1591 }
1592
1593 if (cachep->flags & SLAB_STORE_USER) {
1594 pr_err("Last user: [<%p>](%pSR)\n",
1595 *dbg_userword(cachep, objp),
1596 *dbg_userword(cachep, objp));
1597 }
1598 realobj = (char *)objp + obj_offset(cachep);
1599 size = cachep->object_size;
1600 for (i = 0; i < size && lines; i += 16, lines--) {
1601 int limit;
1602 limit = 16;
1603 if (i + limit > size)
1604 limit = size - i;
1605 dump_line(realobj, i, limit);
1606 }
1607}
1608
1609static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1610{
1611 char *realobj;
1612 int size, i;
1613 int lines = 0;
1614
1615 if (is_debug_pagealloc_cache(cachep))
1616 return;
1617
1618 realobj = (char *)objp + obj_offset(cachep);
1619 size = cachep->object_size;
1620
1621 for (i = 0; i < size; i++) {
1622 char exp = POISON_FREE;
1623 if (i == size - 1)
1624 exp = POISON_END;
1625 if (realobj[i] != exp) {
1626 int limit;
1627 /* Mismatch ! */
1628 /* Print header */
1629 if (lines == 0) {
1630 pr_err("Slab corruption (%s): %s start=%p, len=%d\n",
1631 print_tainted(), cachep->name,
1632 realobj, size);
1633 print_objinfo(cachep, objp, 0);
1634 }
1635 /* Hexdump the affected line */
1636 i = (i / 16) * 16;
1637 limit = 16;
1638 if (i + limit > size)
1639 limit = size - i;
1640 dump_line(realobj, i, limit);
1641 i += 16;
1642 lines++;
1643 /* Limit to 5 lines */
1644 if (lines > 5)
1645 break;
1646 }
1647 }
1648 if (lines != 0) {
1649 /* Print some data about the neighboring objects, if they
1650 * exist:
1651 */
1652 struct page *page = virt_to_head_page(objp);
1653 unsigned int objnr;
1654
1655 objnr = obj_to_index(cachep, page, objp);
1656 if (objnr) {
1657 objp = index_to_obj(cachep, page, objnr - 1);
1658 realobj = (char *)objp + obj_offset(cachep);
1659 pr_err("Prev obj: start=%p, len=%d\n", realobj, size);
1660 print_objinfo(cachep, objp, 2);
1661 }
1662 if (objnr + 1 < cachep->num) {
1663 objp = index_to_obj(cachep, page, objnr + 1);
1664 realobj = (char *)objp + obj_offset(cachep);
1665 pr_err("Next obj: start=%p, len=%d\n", realobj, size);
1666 print_objinfo(cachep, objp, 2);
1667 }
1668 }
1669}
1670#endif
1671
1672#if DEBUG
1673static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1674 struct page *page)
1675{
1676 int i;
1677
1678 if (OBJFREELIST_SLAB(cachep) && cachep->flags & SLAB_POISON) {
1679 poison_obj(cachep, page->freelist - obj_offset(cachep),
1680 POISON_FREE);
1681 }
1682
1683 for (i = 0; i < cachep->num; i++) {
1684 void *objp = index_to_obj(cachep, page, i);
1685
1686 if (cachep->flags & SLAB_POISON) {
1687 check_poison_obj(cachep, objp);
1688 slab_kernel_map(cachep, objp, 1, 0);
1689 }
1690 if (cachep->flags & SLAB_RED_ZONE) {
1691 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1692 slab_error(cachep, "start of a freed object was overwritten");
1693 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1694 slab_error(cachep, "end of a freed object was overwritten");
1695 }
1696 }
1697}
1698#else
1699static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1700 struct page *page)
1701{
1702}
1703#endif
1704
1705/**
1706 * slab_destroy - destroy and release all objects in a slab
1707 * @cachep: cache pointer being destroyed
1708 * @page: page pointer being destroyed
1709 *
1710 * Destroy all the objs in a slab page, and release the mem back to the system.
1711 * Before calling the slab page must have been unlinked from the cache. The
1712 * kmem_cache_node ->list_lock is not held/needed.
1713 */
1714static void slab_destroy(struct kmem_cache *cachep, struct page *page)
1715{
1716 void *freelist;
1717
1718 freelist = page->freelist;
1719 slab_destroy_debugcheck(cachep, page);
1720 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
1721 call_rcu(&page->rcu_head, kmem_rcu_free);
1722 else
1723 kmem_freepages(cachep, page);
1724
1725 /*
1726 * From now on, we don't use freelist
1727 * although actual page can be freed in rcu context
1728 */
1729 if (OFF_SLAB(cachep))
1730 kmem_cache_free(cachep->freelist_cache, freelist);
1731}
1732
1733static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list)
1734{
1735 struct page *page, *n;
1736
1737 list_for_each_entry_safe(page, n, list, lru) {
1738 list_del(&page->lru);
1739 slab_destroy(cachep, page);
1740 }
1741}
1742
1743/**
1744 * calculate_slab_order - calculate size (page order) of slabs
1745 * @cachep: pointer to the cache that is being created
1746 * @size: size of objects to be created in this cache.
1747 * @flags: slab allocation flags
1748 *
1749 * Also calculates the number of objects per slab.
1750 *
1751 * This could be made much more intelligent. For now, try to avoid using
1752 * high order pages for slabs. When the gfp() functions are more friendly
1753 * towards high-order requests, this should be changed.
1754 */
1755static size_t calculate_slab_order(struct kmem_cache *cachep,
1756 size_t size, unsigned long flags)
1757{
1758 size_t left_over = 0;
1759 int gfporder;
1760
1761 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
1762 unsigned int num;
1763 size_t remainder;
1764
1765 num = cache_estimate(gfporder, size, flags, &remainder);
1766 if (!num)
1767 continue;
1768
1769 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1770 if (num > SLAB_OBJ_MAX_NUM)
1771 break;
1772
1773 if (flags & CFLGS_OFF_SLAB) {
1774 struct kmem_cache *freelist_cache;
1775 size_t freelist_size;
1776
1777 freelist_size = num * sizeof(freelist_idx_t);
1778 freelist_cache = kmalloc_slab(freelist_size, 0u);
1779 if (!freelist_cache)
1780 continue;
1781
1782 /*
1783 * Needed to avoid possible looping condition
1784 * in cache_grow()
1785 */
1786 if (OFF_SLAB(freelist_cache))
1787 continue;
1788
1789 /* check if off slab has enough benefit */
1790 if (freelist_cache->size > cachep->size / 2)
1791 continue;
1792 }
1793
1794 /* Found something acceptable - save it away */
1795 cachep->num = num;
1796 cachep->gfporder = gfporder;
1797 left_over = remainder;
1798
1799 /*
1800 * A VFS-reclaimable slab tends to have most allocations
1801 * as GFP_NOFS and we really don't want to have to be allocating
1802 * higher-order pages when we are unable to shrink dcache.
1803 */
1804 if (flags & SLAB_RECLAIM_ACCOUNT)
1805 break;
1806
1807 /*
1808 * Large number of objects is good, but very large slabs are
1809 * currently bad for the gfp()s.
1810 */
1811 if (gfporder >= slab_max_order)
1812 break;
1813
1814 /*
1815 * Acceptable internal fragmentation?
1816 */
1817 if (left_over * 8 <= (PAGE_SIZE << gfporder))
1818 break;
1819 }
1820 return left_over;
1821}
1822
1823static struct array_cache __percpu *alloc_kmem_cache_cpus(
1824 struct kmem_cache *cachep, int entries, int batchcount)
1825{
1826 int cpu;
1827 size_t size;
1828 struct array_cache __percpu *cpu_cache;
1829
1830 size = sizeof(void *) * entries + sizeof(struct array_cache);
1831 cpu_cache = __alloc_percpu(size, sizeof(void *));
1832
1833 if (!cpu_cache)
1834 return NULL;
1835
1836 for_each_possible_cpu(cpu) {
1837 init_arraycache(per_cpu_ptr(cpu_cache, cpu),
1838 entries, batchcount);
1839 }
1840
1841 return cpu_cache;
1842}
1843
1844static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
1845{
1846 if (slab_state >= FULL)
1847 return enable_cpucache(cachep, gfp);
1848
1849 cachep->cpu_cache = alloc_kmem_cache_cpus(cachep, 1, 1);
1850 if (!cachep->cpu_cache)
1851 return 1;
1852
1853 if (slab_state == DOWN) {
1854 /* Creation of first cache (kmem_cache). */
1855 set_up_node(kmem_cache, CACHE_CACHE);
1856 } else if (slab_state == PARTIAL) {
1857 /* For kmem_cache_node */
1858 set_up_node(cachep, SIZE_NODE);
1859 } else {
1860 int node;
1861
1862 for_each_online_node(node) {
1863 cachep->node[node] = kmalloc_node(
1864 sizeof(struct kmem_cache_node), gfp, node);
1865 BUG_ON(!cachep->node[node]);
1866 kmem_cache_node_init(cachep->node[node]);
1867 }
1868 }
1869
1870 cachep->node[numa_mem_id()]->next_reap =
1871 jiffies + REAPTIMEOUT_NODE +
1872 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1873
1874 cpu_cache_get(cachep)->avail = 0;
1875 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1876 cpu_cache_get(cachep)->batchcount = 1;
1877 cpu_cache_get(cachep)->touched = 0;
1878 cachep->batchcount = 1;
1879 cachep->limit = BOOT_CPUCACHE_ENTRIES;
1880 return 0;
1881}
1882
1883unsigned long kmem_cache_flags(unsigned long object_size,
1884 unsigned long flags, const char *name,
1885 void (*ctor)(void *))
1886{
1887 return flags;
1888}
1889
1890struct kmem_cache *
1891__kmem_cache_alias(const char *name, size_t size, size_t align,
1892 unsigned long flags, void (*ctor)(void *))
1893{
1894 struct kmem_cache *cachep;
1895
1896 cachep = find_mergeable(size, align, flags, name, ctor);
1897 if (cachep) {
1898 cachep->refcount++;
1899
1900 /*
1901 * Adjust the object sizes so that we clear
1902 * the complete object on kzalloc.
1903 */
1904 cachep->object_size = max_t(int, cachep->object_size, size);
1905 }
1906 return cachep;
1907}
1908
1909static bool set_objfreelist_slab_cache(struct kmem_cache *cachep,
1910 size_t size, unsigned long flags)
1911{
1912 size_t left;
1913
1914 cachep->num = 0;
1915
1916 if (cachep->ctor || flags & SLAB_DESTROY_BY_RCU)
1917 return false;
1918
1919 left = calculate_slab_order(cachep, size,
1920 flags | CFLGS_OBJFREELIST_SLAB);
1921 if (!cachep->num)
1922 return false;
1923
1924 if (cachep->num * sizeof(freelist_idx_t) > cachep->object_size)
1925 return false;
1926
1927 cachep->colour = left / cachep->colour_off;
1928
1929 return true;
1930}
1931
1932static bool set_off_slab_cache(struct kmem_cache *cachep,
1933 size_t size, unsigned long flags)
1934{
1935 size_t left;
1936
1937 cachep->num = 0;
1938
1939 /*
1940 * Always use on-slab management when SLAB_NOLEAKTRACE
1941 * to avoid recursive calls into kmemleak.
1942 */
1943 if (flags & SLAB_NOLEAKTRACE)
1944 return false;
1945
1946 /*
1947 * Size is large, assume best to place the slab management obj
1948 * off-slab (should allow better packing of objs).
1949 */
1950 left = calculate_slab_order(cachep, size, flags | CFLGS_OFF_SLAB);
1951 if (!cachep->num)
1952 return false;
1953
1954 /*
1955 * If the slab has been placed off-slab, and we have enough space then
1956 * move it on-slab. This is at the expense of any extra colouring.
1957 */
1958 if (left >= cachep->num * sizeof(freelist_idx_t))
1959 return false;
1960
1961 cachep->colour = left / cachep->colour_off;
1962
1963 return true;
1964}
1965
1966static bool set_on_slab_cache(struct kmem_cache *cachep,
1967 size_t size, unsigned long flags)
1968{
1969 size_t left;
1970
1971 cachep->num = 0;
1972
1973 left = calculate_slab_order(cachep, size, flags);
1974 if (!cachep->num)
1975 return false;
1976
1977 cachep->colour = left / cachep->colour_off;
1978
1979 return true;
1980}
1981
1982/**
1983 * __kmem_cache_create - Create a cache.
1984 * @cachep: cache management descriptor
1985 * @flags: SLAB flags
1986 *
1987 * Returns a ptr to the cache on success, NULL on failure.
1988 * Cannot be called within a int, but can be interrupted.
1989 * The @ctor is run when new pages are allocated by the cache.
1990 *
1991 * The flags are
1992 *
1993 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1994 * to catch references to uninitialised memory.
1995 *
1996 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1997 * for buffer overruns.
1998 *
1999 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2000 * cacheline. This can be beneficial if you're counting cycles as closely
2001 * as davem.
2002 */
2003int
2004__kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
2005{
2006 size_t ralign = BYTES_PER_WORD;
2007 gfp_t gfp;
2008 int err;
2009 size_t size = cachep->size;
2010
2011#if DEBUG
2012#if FORCED_DEBUG
2013 /*
2014 * Enable redzoning and last user accounting, except for caches with
2015 * large objects, if the increased size would increase the object size
2016 * above the next power of two: caches with object sizes just above a
2017 * power of two have a significant amount of internal fragmentation.
2018 */
2019 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2020 2 * sizeof(unsigned long long)))
2021 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2022 if (!(flags & SLAB_DESTROY_BY_RCU))
2023 flags |= SLAB_POISON;
2024#endif
2025#endif
2026
2027 /*
2028 * Check that size is in terms of words. This is needed to avoid
2029 * unaligned accesses for some archs when redzoning is used, and makes
2030 * sure any on-slab bufctl's are also correctly aligned.
2031 */
2032 if (size & (BYTES_PER_WORD - 1)) {
2033 size += (BYTES_PER_WORD - 1);
2034 size &= ~(BYTES_PER_WORD - 1);
2035 }
2036
2037 if (flags & SLAB_RED_ZONE) {
2038 ralign = REDZONE_ALIGN;
2039 /* If redzoning, ensure that the second redzone is suitably
2040 * aligned, by adjusting the object size accordingly. */
2041 size += REDZONE_ALIGN - 1;
2042 size &= ~(REDZONE_ALIGN - 1);
2043 }
2044
2045 /* 3) caller mandated alignment */
2046 if (ralign < cachep->align) {
2047 ralign = cachep->align;
2048 }
2049 /* disable debug if necessary */
2050 if (ralign > __alignof__(unsigned long long))
2051 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2052 /*
2053 * 4) Store it.
2054 */
2055 cachep->align = ralign;
2056 cachep->colour_off = cache_line_size();
2057 /* Offset must be a multiple of the alignment. */
2058 if (cachep->colour_off < cachep->align)
2059 cachep->colour_off = cachep->align;
2060
2061 if (slab_is_available())
2062 gfp = GFP_KERNEL;
2063 else
2064 gfp = GFP_NOWAIT;
2065
2066#if DEBUG
2067
2068 /*
2069 * Both debugging options require word-alignment which is calculated
2070 * into align above.
2071 */
2072 if (flags & SLAB_RED_ZONE) {
2073 /* add space for red zone words */
2074 cachep->obj_offset += sizeof(unsigned long long);
2075 size += 2 * sizeof(unsigned long long);
2076 }
2077 if (flags & SLAB_STORE_USER) {
2078 /* user store requires one word storage behind the end of
2079 * the real object. But if the second red zone needs to be
2080 * aligned to 64 bits, we must allow that much space.
2081 */
2082 if (flags & SLAB_RED_ZONE)
2083 size += REDZONE_ALIGN;
2084 else
2085 size += BYTES_PER_WORD;
2086 }
2087#endif
2088
2089 kasan_cache_create(cachep, &size, &flags);
2090
2091 size = ALIGN(size, cachep->align);
2092 /*
2093 * We should restrict the number of objects in a slab to implement
2094 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2095 */
2096 if (FREELIST_BYTE_INDEX && size < SLAB_OBJ_MIN_SIZE)
2097 size = ALIGN(SLAB_OBJ_MIN_SIZE, cachep->align);
2098
2099#if DEBUG
2100 /*
2101 * To activate debug pagealloc, off-slab management is necessary
2102 * requirement. In early phase of initialization, small sized slab
2103 * doesn't get initialized so it would not be possible. So, we need
2104 * to check size >= 256. It guarantees that all necessary small
2105 * sized slab is initialized in current slab initialization sequence.
2106 */
2107 if (debug_pagealloc_enabled() && (flags & SLAB_POISON) &&
2108 size >= 256 && cachep->object_size > cache_line_size()) {
2109 if (size < PAGE_SIZE || size % PAGE_SIZE == 0) {
2110 size_t tmp_size = ALIGN(size, PAGE_SIZE);
2111
2112 if (set_off_slab_cache(cachep, tmp_size, flags)) {
2113 flags |= CFLGS_OFF_SLAB;
2114 cachep->obj_offset += tmp_size - size;
2115 size = tmp_size;
2116 goto done;
2117 }
2118 }
2119 }
2120#endif
2121
2122 if (set_objfreelist_slab_cache(cachep, size, flags)) {
2123 flags |= CFLGS_OBJFREELIST_SLAB;
2124 goto done;
2125 }
2126
2127 if (set_off_slab_cache(cachep, size, flags)) {
2128 flags |= CFLGS_OFF_SLAB;
2129 goto done;
2130 }
2131
2132 if (set_on_slab_cache(cachep, size, flags))
2133 goto done;
2134
2135 return -E2BIG;
2136
2137done:
2138 cachep->freelist_size = cachep->num * sizeof(freelist_idx_t);
2139 cachep->flags = flags;
2140 cachep->allocflags = __GFP_COMP;
2141 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2142 cachep->allocflags |= GFP_DMA;
2143 cachep->size = size;
2144 cachep->reciprocal_buffer_size = reciprocal_value(size);
2145
2146#if DEBUG
2147 /*
2148 * If we're going to use the generic kernel_map_pages()
2149 * poisoning, then it's going to smash the contents of
2150 * the redzone and userword anyhow, so switch them off.
2151 */
2152 if (IS_ENABLED(CONFIG_PAGE_POISONING) &&
2153 (cachep->flags & SLAB_POISON) &&
2154 is_debug_pagealloc_cache(cachep))
2155 cachep->flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2156#endif
2157
2158 if (OFF_SLAB(cachep)) {
2159 cachep->freelist_cache =
2160 kmalloc_slab(cachep->freelist_size, 0u);
2161 }
2162
2163 err = setup_cpu_cache(cachep, gfp);
2164 if (err) {
2165 __kmem_cache_release(cachep);
2166 return err;
2167 }
2168
2169 return 0;
2170}
2171
2172#if DEBUG
2173static void check_irq_off(void)
2174{
2175 BUG_ON(!irqs_disabled());
2176}
2177
2178static void check_irq_on(void)
2179{
2180 BUG_ON(irqs_disabled());
2181}
2182
2183static void check_spinlock_acquired(struct kmem_cache *cachep)
2184{
2185#ifdef CONFIG_SMP
2186 check_irq_off();
2187 assert_spin_locked(&get_node(cachep, numa_mem_id())->list_lock);
2188#endif
2189}
2190
2191static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2192{
2193#ifdef CONFIG_SMP
2194 check_irq_off();
2195 assert_spin_locked(&get_node(cachep, node)->list_lock);
2196#endif
2197}
2198
2199#else
2200#define check_irq_off() do { } while(0)
2201#define check_irq_on() do { } while(0)
2202#define check_spinlock_acquired(x) do { } while(0)
2203#define check_spinlock_acquired_node(x, y) do { } while(0)
2204#endif
2205
2206static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
2207 struct array_cache *ac,
2208 int force, int node);
2209
2210static void do_drain(void *arg)
2211{
2212 struct kmem_cache *cachep = arg;
2213 struct array_cache *ac;
2214 int node = numa_mem_id();
2215 struct kmem_cache_node *n;
2216 LIST_HEAD(list);
2217
2218 check_irq_off();
2219 ac = cpu_cache_get(cachep);
2220 n = get_node(cachep, node);
2221 spin_lock(&n->list_lock);
2222 free_block(cachep, ac->entry, ac->avail, node, &list);
2223 spin_unlock(&n->list_lock);
2224 slabs_destroy(cachep, &list);
2225 ac->avail = 0;
2226}
2227
2228static void drain_cpu_caches(struct kmem_cache *cachep)
2229{
2230 struct kmem_cache_node *n;
2231 int node;
2232
2233 on_each_cpu(do_drain, cachep, 1);
2234 check_irq_on();
2235 for_each_kmem_cache_node(cachep, node, n)
2236 if (n->alien)
2237 drain_alien_cache(cachep, n->alien);
2238
2239 for_each_kmem_cache_node(cachep, node, n)
2240 drain_array(cachep, n, n->shared, 1, node);
2241}
2242
2243/*
2244 * Remove slabs from the list of free slabs.
2245 * Specify the number of slabs to drain in tofree.
2246 *
2247 * Returns the actual number of slabs released.
2248 */
2249static int drain_freelist(struct kmem_cache *cache,
2250 struct kmem_cache_node *n, int tofree)
2251{
2252 struct list_head *p;
2253 int nr_freed;
2254 struct page *page;
2255
2256 nr_freed = 0;
2257 while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
2258
2259 spin_lock_irq(&n->list_lock);
2260 p = n->slabs_free.prev;
2261 if (p == &n->slabs_free) {
2262 spin_unlock_irq(&n->list_lock);
2263 goto out;
2264 }
2265
2266 page = list_entry(p, struct page, lru);
2267 list_del(&page->lru);
2268 /*
2269 * Safe to drop the lock. The slab is no longer linked
2270 * to the cache.
2271 */
2272 n->free_objects -= cache->num;
2273 spin_unlock_irq(&n->list_lock);
2274 slab_destroy(cache, page);
2275 nr_freed++;
2276 }
2277out:
2278 return nr_freed;
2279}
2280
2281int __kmem_cache_shrink(struct kmem_cache *cachep, bool deactivate)
2282{
2283 int ret = 0;
2284 int node;
2285 struct kmem_cache_node *n;
2286
2287 drain_cpu_caches(cachep);
2288
2289 check_irq_on();
2290 for_each_kmem_cache_node(cachep, node, n) {
2291 drain_freelist(cachep, n, slabs_tofree(cachep, n));
2292
2293 ret += !list_empty(&n->slabs_full) ||
2294 !list_empty(&n->slabs_partial);
2295 }
2296 return (ret ? 1 : 0);
2297}
2298
2299int __kmem_cache_shutdown(struct kmem_cache *cachep)
2300{
2301 return __kmem_cache_shrink(cachep, false);
2302}
2303
2304void __kmem_cache_release(struct kmem_cache *cachep)
2305{
2306 int i;
2307 struct kmem_cache_node *n;
2308
2309 free_percpu(cachep->cpu_cache);
2310
2311 /* NUMA: free the node structures */
2312 for_each_kmem_cache_node(cachep, i, n) {
2313 kfree(n->shared);
2314 free_alien_cache(n->alien);
2315 kfree(n);
2316 cachep->node[i] = NULL;
2317 }
2318}
2319
2320/*
2321 * Get the memory for a slab management obj.
2322 *
2323 * For a slab cache when the slab descriptor is off-slab, the
2324 * slab descriptor can't come from the same cache which is being created,
2325 * Because if it is the case, that means we defer the creation of
2326 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2327 * And we eventually call down to __kmem_cache_create(), which
2328 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2329 * This is a "chicken-and-egg" problem.
2330 *
2331 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2332 * which are all initialized during kmem_cache_init().
2333 */
2334static void *alloc_slabmgmt(struct kmem_cache *cachep,
2335 struct page *page, int colour_off,
2336 gfp_t local_flags, int nodeid)
2337{
2338 void *freelist;
2339 void *addr = page_address(page);
2340
2341 page->s_mem = addr + colour_off;
2342 page->active = 0;
2343
2344 if (OBJFREELIST_SLAB(cachep))
2345 freelist = NULL;
2346 else if (OFF_SLAB(cachep)) {
2347 /* Slab management obj is off-slab. */
2348 freelist = kmem_cache_alloc_node(cachep->freelist_cache,
2349 local_flags, nodeid);
2350 if (!freelist)
2351 return NULL;
2352 } else {
2353 /* We will use last bytes at the slab for freelist */
2354 freelist = addr + (PAGE_SIZE << cachep->gfporder) -
2355 cachep->freelist_size;
2356 }
2357
2358 return freelist;
2359}
2360
2361static inline freelist_idx_t get_free_obj(struct page *page, unsigned int idx)
2362{
2363 return ((freelist_idx_t *)page->freelist)[idx];
2364}
2365
2366static inline void set_free_obj(struct page *page,
2367 unsigned int idx, freelist_idx_t val)
2368{
2369 ((freelist_idx_t *)(page->freelist))[idx] = val;
2370}
2371
2372static void cache_init_objs_debug(struct kmem_cache *cachep, struct page *page)
2373{
2374#if DEBUG
2375 int i;
2376
2377 for (i = 0; i < cachep->num; i++) {
2378 void *objp = index_to_obj(cachep, page, i);
2379
2380 if (cachep->flags & SLAB_STORE_USER)
2381 *dbg_userword(cachep, objp) = NULL;
2382
2383 if (cachep->flags & SLAB_RED_ZONE) {
2384 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2385 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2386 }
2387 /*
2388 * Constructors are not allowed to allocate memory from the same
2389 * cache which they are a constructor for. Otherwise, deadlock.
2390 * They must also be threaded.
2391 */
2392 if (cachep->ctor && !(cachep->flags & SLAB_POISON)) {
2393 kasan_unpoison_object_data(cachep,
2394 objp + obj_offset(cachep));
2395 cachep->ctor(objp + obj_offset(cachep));
2396 kasan_poison_object_data(
2397 cachep, objp + obj_offset(cachep));
2398 }
2399
2400 if (cachep->flags & SLAB_RED_ZONE) {
2401 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2402 slab_error(cachep, "constructor overwrote the end of an object");
2403 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2404 slab_error(cachep, "constructor overwrote the start of an object");
2405 }
2406 /* need to poison the objs? */
2407 if (cachep->flags & SLAB_POISON) {
2408 poison_obj(cachep, objp, POISON_FREE);
2409 slab_kernel_map(cachep, objp, 0, 0);
2410 }
2411 }
2412#endif
2413}
2414
2415static void cache_init_objs(struct kmem_cache *cachep,
2416 struct page *page)
2417{
2418 int i;
2419 void *objp;
2420
2421 cache_init_objs_debug(cachep, page);
2422
2423 if (OBJFREELIST_SLAB(cachep)) {
2424 page->freelist = index_to_obj(cachep, page, cachep->num - 1) +
2425 obj_offset(cachep);
2426 }
2427
2428 for (i = 0; i < cachep->num; i++) {
2429 /* constructor could break poison info */
2430 if (DEBUG == 0 && cachep->ctor) {
2431 objp = index_to_obj(cachep, page, i);
2432 kasan_unpoison_object_data(cachep, objp);
2433 cachep->ctor(objp);
2434 kasan_poison_object_data(cachep, objp);
2435 }
2436
2437 set_free_obj(page, i, i);
2438 }
2439}
2440
2441static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2442{
2443 if (CONFIG_ZONE_DMA_FLAG) {
2444 if (flags & GFP_DMA)
2445 BUG_ON(!(cachep->allocflags & GFP_DMA));
2446 else
2447 BUG_ON(cachep->allocflags & GFP_DMA);
2448 }
2449}
2450
2451static void *slab_get_obj(struct kmem_cache *cachep, struct page *page)
2452{
2453 void *objp;
2454
2455 objp = index_to_obj(cachep, page, get_free_obj(page, page->active));
2456 page->active++;
2457
2458#if DEBUG
2459 if (cachep->flags & SLAB_STORE_USER)
2460 set_store_user_dirty(cachep);
2461#endif
2462
2463 return objp;
2464}
2465
2466static void slab_put_obj(struct kmem_cache *cachep,
2467 struct page *page, void *objp)
2468{
2469 unsigned int objnr = obj_to_index(cachep, page, objp);
2470#if DEBUG
2471 unsigned int i;
2472
2473 /* Verify double free bug */
2474 for (i = page->active; i < cachep->num; i++) {
2475 if (get_free_obj(page, i) == objnr) {
2476 pr_err("slab: double free detected in cache '%s', objp %p\n",
2477 cachep->name, objp);
2478 BUG();
2479 }
2480 }
2481#endif
2482 page->active--;
2483 if (!page->freelist)
2484 page->freelist = objp + obj_offset(cachep);
2485
2486 set_free_obj(page, page->active, objnr);
2487}
2488
2489/*
2490 * Map pages beginning at addr to the given cache and slab. This is required
2491 * for the slab allocator to be able to lookup the cache and slab of a
2492 * virtual address for kfree, ksize, and slab debugging.
2493 */
2494static void slab_map_pages(struct kmem_cache *cache, struct page *page,
2495 void *freelist)
2496{
2497 page->slab_cache = cache;
2498 page->freelist = freelist;
2499}
2500
2501/*
2502 * Grow (by 1) the number of slabs within a cache. This is called by
2503 * kmem_cache_alloc() when there are no active objs left in a cache.
2504 */
2505static int cache_grow(struct kmem_cache *cachep,
2506 gfp_t flags, int nodeid, struct page *page)
2507{
2508 void *freelist;
2509 size_t offset;
2510 gfp_t local_flags;
2511 struct kmem_cache_node *n;
2512
2513 /*
2514 * Be lazy and only check for valid flags here, keeping it out of the
2515 * critical path in kmem_cache_alloc().
2516 */
2517 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
2518 pr_emerg("gfp: %u\n", flags & GFP_SLAB_BUG_MASK);
2519 BUG();
2520 }
2521 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2522
2523 /* Take the node list lock to change the colour_next on this node */
2524 check_irq_off();
2525 n = get_node(cachep, nodeid);
2526 spin_lock(&n->list_lock);
2527
2528 /* Get colour for the slab, and cal the next value. */
2529 offset = n->colour_next;
2530 n->colour_next++;
2531 if (n->colour_next >= cachep->colour)
2532 n->colour_next = 0;
2533 spin_unlock(&n->list_lock);
2534
2535 offset *= cachep->colour_off;
2536
2537 if (gfpflags_allow_blocking(local_flags))
2538 local_irq_enable();
2539
2540 /*
2541 * The test for missing atomic flag is performed here, rather than
2542 * the more obvious place, simply to reduce the critical path length
2543 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2544 * will eventually be caught here (where it matters).
2545 */
2546 kmem_flagcheck(cachep, flags);
2547
2548 /*
2549 * Get mem for the objs. Attempt to allocate a physical page from
2550 * 'nodeid'.
2551 */
2552 if (!page)
2553 page = kmem_getpages(cachep, local_flags, nodeid);
2554 if (!page)
2555 goto failed;
2556
2557 /* Get slab management. */
2558 freelist = alloc_slabmgmt(cachep, page, offset,
2559 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2560 if (OFF_SLAB(cachep) && !freelist)
2561 goto opps1;
2562
2563 slab_map_pages(cachep, page, freelist);
2564
2565 kasan_poison_slab(page);
2566 cache_init_objs(cachep, page);
2567
2568 if (gfpflags_allow_blocking(local_flags))
2569 local_irq_disable();
2570 check_irq_off();
2571 spin_lock(&n->list_lock);
2572
2573 /* Make slab active. */
2574 list_add_tail(&page->lru, &(n->slabs_free));
2575 STATS_INC_GROWN(cachep);
2576 n->free_objects += cachep->num;
2577 spin_unlock(&n->list_lock);
2578 return 1;
2579opps1:
2580 kmem_freepages(cachep, page);
2581failed:
2582 if (gfpflags_allow_blocking(local_flags))
2583 local_irq_disable();
2584 return 0;
2585}
2586
2587#if DEBUG
2588
2589/*
2590 * Perform extra freeing checks:
2591 * - detect bad pointers.
2592 * - POISON/RED_ZONE checking
2593 */
2594static void kfree_debugcheck(const void *objp)
2595{
2596 if (!virt_addr_valid(objp)) {
2597 pr_err("kfree_debugcheck: out of range ptr %lxh\n",
2598 (unsigned long)objp);
2599 BUG();
2600 }
2601}
2602
2603static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2604{
2605 unsigned long long redzone1, redzone2;
2606
2607 redzone1 = *dbg_redzone1(cache, obj);
2608 redzone2 = *dbg_redzone2(cache, obj);
2609
2610 /*
2611 * Redzone is ok.
2612 */
2613 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2614 return;
2615
2616 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2617 slab_error(cache, "double free detected");
2618 else
2619 slab_error(cache, "memory outside object was overwritten");
2620
2621 pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2622 obj, redzone1, redzone2);
2623}
2624
2625static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2626 unsigned long caller)
2627{
2628 unsigned int objnr;
2629 struct page *page;
2630
2631 BUG_ON(virt_to_cache(objp) != cachep);
2632
2633 objp -= obj_offset(cachep);
2634 kfree_debugcheck(objp);
2635 page = virt_to_head_page(objp);
2636
2637 if (cachep->flags & SLAB_RED_ZONE) {
2638 verify_redzone_free(cachep, objp);
2639 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2640 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2641 }
2642 if (cachep->flags & SLAB_STORE_USER) {
2643 set_store_user_dirty(cachep);
2644 *dbg_userword(cachep, objp) = (void *)caller;
2645 }
2646
2647 objnr = obj_to_index(cachep, page, objp);
2648
2649 BUG_ON(objnr >= cachep->num);
2650 BUG_ON(objp != index_to_obj(cachep, page, objnr));
2651
2652 if (cachep->flags & SLAB_POISON) {
2653 poison_obj(cachep, objp, POISON_FREE);
2654 slab_kernel_map(cachep, objp, 0, caller);
2655 }
2656 return objp;
2657}
2658
2659#else
2660#define kfree_debugcheck(x) do { } while(0)
2661#define cache_free_debugcheck(x,objp,z) (objp)
2662#endif
2663
2664static inline void fixup_objfreelist_debug(struct kmem_cache *cachep,
2665 void **list)
2666{
2667#if DEBUG
2668 void *next = *list;
2669 void *objp;
2670
2671 while (next) {
2672 objp = next - obj_offset(cachep);
2673 next = *(void **)next;
2674 poison_obj(cachep, objp, POISON_FREE);
2675 }
2676#endif
2677}
2678
2679static inline void fixup_slab_list(struct kmem_cache *cachep,
2680 struct kmem_cache_node *n, struct page *page,
2681 void **list)
2682{
2683 /* move slabp to correct slabp list: */
2684 list_del(&page->lru);
2685 if (page->active == cachep->num) {
2686 list_add(&page->lru, &n->slabs_full);
2687 if (OBJFREELIST_SLAB(cachep)) {
2688#if DEBUG
2689 /* Poisoning will be done without holding the lock */
2690 if (cachep->flags & SLAB_POISON) {
2691 void **objp = page->freelist;
2692
2693 *objp = *list;
2694 *list = objp;
2695 }
2696#endif
2697 page->freelist = NULL;
2698 }
2699 } else
2700 list_add(&page->lru, &n->slabs_partial);
2701}
2702
2703/* Try to find non-pfmemalloc slab if needed */
2704static noinline struct page *get_valid_first_slab(struct kmem_cache_node *n,
2705 struct page *page, bool pfmemalloc)
2706{
2707 if (!page)
2708 return NULL;
2709
2710 if (pfmemalloc)
2711 return page;
2712
2713 if (!PageSlabPfmemalloc(page))
2714 return page;
2715
2716 /* No need to keep pfmemalloc slab if we have enough free objects */
2717 if (n->free_objects > n->free_limit) {
2718 ClearPageSlabPfmemalloc(page);
2719 return page;
2720 }
2721
2722 /* Move pfmemalloc slab to the end of list to speed up next search */
2723 list_del(&page->lru);
2724 if (!page->active)
2725 list_add_tail(&page->lru, &n->slabs_free);
2726 else
2727 list_add_tail(&page->lru, &n->slabs_partial);
2728
2729 list_for_each_entry(page, &n->slabs_partial, lru) {
2730 if (!PageSlabPfmemalloc(page))
2731 return page;
2732 }
2733
2734 list_for_each_entry(page, &n->slabs_free, lru) {
2735 if (!PageSlabPfmemalloc(page))
2736 return page;
2737 }
2738
2739 return NULL;
2740}
2741
2742static struct page *get_first_slab(struct kmem_cache_node *n, bool pfmemalloc)
2743{
2744 struct page *page;
2745
2746 page = list_first_entry_or_null(&n->slabs_partial,
2747 struct page, lru);
2748 if (!page) {
2749 n->free_touched = 1;
2750 page = list_first_entry_or_null(&n->slabs_free,
2751 struct page, lru);
2752 }
2753
2754 if (sk_memalloc_socks())
2755 return get_valid_first_slab(n, page, pfmemalloc);
2756
2757 return page;
2758}
2759
2760static noinline void *cache_alloc_pfmemalloc(struct kmem_cache *cachep,
2761 struct kmem_cache_node *n, gfp_t flags)
2762{
2763 struct page *page;
2764 void *obj;
2765 void *list = NULL;
2766
2767 if (!gfp_pfmemalloc_allowed(flags))
2768 return NULL;
2769
2770 spin_lock(&n->list_lock);
2771 page = get_first_slab(n, true);
2772 if (!page) {
2773 spin_unlock(&n->list_lock);
2774 return NULL;
2775 }
2776
2777 obj = slab_get_obj(cachep, page);
2778 n->free_objects--;
2779
2780 fixup_slab_list(cachep, n, page, &list);
2781
2782 spin_unlock(&n->list_lock);
2783 fixup_objfreelist_debug(cachep, &list);
2784
2785 return obj;
2786}
2787
2788static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
2789{
2790 int batchcount;
2791 struct kmem_cache_node *n;
2792 struct array_cache *ac;
2793 int node;
2794 void *list = NULL;
2795
2796 check_irq_off();
2797 node = numa_mem_id();
2798
2799retry:
2800 ac = cpu_cache_get(cachep);
2801 batchcount = ac->batchcount;
2802 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2803 /*
2804 * If there was little recent activity on this cache, then
2805 * perform only a partial refill. Otherwise we could generate
2806 * refill bouncing.
2807 */
2808 batchcount = BATCHREFILL_LIMIT;
2809 }
2810 n = get_node(cachep, node);
2811
2812 BUG_ON(ac->avail > 0 || !n);
2813 spin_lock(&n->list_lock);
2814
2815 /* See if we can refill from the shared array */
2816 if (n->shared && transfer_objects(ac, n->shared, batchcount)) {
2817 n->shared->touched = 1;
2818 goto alloc_done;
2819 }
2820
2821 while (batchcount > 0) {
2822 struct page *page;
2823 /* Get slab alloc is to come from. */
2824 page = get_first_slab(n, false);
2825 if (!page)
2826 goto must_grow;
2827
2828 check_spinlock_acquired(cachep);
2829
2830 /*
2831 * The slab was either on partial or free list so
2832 * there must be at least one object available for
2833 * allocation.
2834 */
2835 BUG_ON(page->active >= cachep->num);
2836
2837 while (page->active < cachep->num && batchcount--) {
2838 STATS_INC_ALLOCED(cachep);
2839 STATS_INC_ACTIVE(cachep);
2840 STATS_SET_HIGH(cachep);
2841
2842 ac->entry[ac->avail++] = slab_get_obj(cachep, page);
2843 }
2844
2845 fixup_slab_list(cachep, n, page, &list);
2846 }
2847
2848must_grow:
2849 n->free_objects -= ac->avail;
2850alloc_done:
2851 spin_unlock(&n->list_lock);
2852 fixup_objfreelist_debug(cachep, &list);
2853
2854 if (unlikely(!ac->avail)) {
2855 int x;
2856
2857 /* Check if we can use obj in pfmemalloc slab */
2858 if (sk_memalloc_socks()) {
2859 void *obj = cache_alloc_pfmemalloc(cachep, n, flags);
2860
2861 if (obj)
2862 return obj;
2863 }
2864
2865 x = cache_grow(cachep, gfp_exact_node(flags), node, NULL);
2866
2867 /* cache_grow can reenable interrupts, then ac could change. */
2868 ac = cpu_cache_get(cachep);
2869 node = numa_mem_id();
2870
2871 /* no objects in sight? abort */
2872 if (!x && ac->avail == 0)
2873 return NULL;
2874
2875 if (!ac->avail) /* objects refilled by interrupt? */
2876 goto retry;
2877 }
2878 ac->touched = 1;
2879
2880 return ac->entry[--ac->avail];
2881}
2882
2883static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2884 gfp_t flags)
2885{
2886 might_sleep_if(gfpflags_allow_blocking(flags));
2887#if DEBUG
2888 kmem_flagcheck(cachep, flags);
2889#endif
2890}
2891
2892#if DEBUG
2893static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
2894 gfp_t flags, void *objp, unsigned long caller)
2895{
2896 if (!objp)
2897 return objp;
2898 if (cachep->flags & SLAB_POISON) {
2899 check_poison_obj(cachep, objp);
2900 slab_kernel_map(cachep, objp, 1, 0);
2901 poison_obj(cachep, objp, POISON_INUSE);
2902 }
2903 if (cachep->flags & SLAB_STORE_USER)
2904 *dbg_userword(cachep, objp) = (void *)caller;
2905
2906 if (cachep->flags & SLAB_RED_ZONE) {
2907 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
2908 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2909 slab_error(cachep, "double free, or memory outside object was overwritten");
2910 pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2911 objp, *dbg_redzone1(cachep, objp),
2912 *dbg_redzone2(cachep, objp));
2913 }
2914 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2915 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2916 }
2917
2918 objp += obj_offset(cachep);
2919 if (cachep->ctor && cachep->flags & SLAB_POISON)
2920 cachep->ctor(objp);
2921 if (ARCH_SLAB_MINALIGN &&
2922 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
2923 pr_err("0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
2924 objp, (int)ARCH_SLAB_MINALIGN);
2925 }
2926 return objp;
2927}
2928#else
2929#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2930#endif
2931
2932static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
2933{
2934 void *objp;
2935 struct array_cache *ac;
2936
2937 check_irq_off();
2938
2939 ac = cpu_cache_get(cachep);
2940 if (likely(ac->avail)) {
2941 ac->touched = 1;
2942 objp = ac->entry[--ac->avail];
2943
2944 STATS_INC_ALLOCHIT(cachep);
2945 goto out;
2946 }
2947
2948 STATS_INC_ALLOCMISS(cachep);
2949 objp = cache_alloc_refill(cachep, flags);
2950 /*
2951 * the 'ac' may be updated by cache_alloc_refill(),
2952 * and kmemleak_erase() requires its correct value.
2953 */
2954 ac = cpu_cache_get(cachep);
2955
2956out:
2957 /*
2958 * To avoid a false negative, if an object that is in one of the
2959 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
2960 * treat the array pointers as a reference to the object.
2961 */
2962 if (objp)
2963 kmemleak_erase(&ac->entry[ac->avail]);
2964 return objp;
2965}
2966
2967#ifdef CONFIG_NUMA
2968/*
2969 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
2970 *
2971 * If we are in_interrupt, then process context, including cpusets and
2972 * mempolicy, may not apply and should not be used for allocation policy.
2973 */
2974static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
2975{
2976 int nid_alloc, nid_here;
2977
2978 if (in_interrupt() || (flags & __GFP_THISNODE))
2979 return NULL;
2980 nid_alloc = nid_here = numa_mem_id();
2981 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
2982 nid_alloc = cpuset_slab_spread_node();
2983 else if (current->mempolicy)
2984 nid_alloc = mempolicy_slab_node();
2985 if (nid_alloc != nid_here)
2986 return ____cache_alloc_node(cachep, flags, nid_alloc);
2987 return NULL;
2988}
2989
2990/*
2991 * Fallback function if there was no memory available and no objects on a
2992 * certain node and fall back is permitted. First we scan all the
2993 * available node for available objects. If that fails then we
2994 * perform an allocation without specifying a node. This allows the page
2995 * allocator to do its reclaim / fallback magic. We then insert the
2996 * slab into the proper nodelist and then allocate from it.
2997 */
2998static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
2999{
3000 struct zonelist *zonelist;
3001 gfp_t local_flags;
3002 struct zoneref *z;
3003 struct zone *zone;
3004 enum zone_type high_zoneidx = gfp_zone(flags);
3005 void *obj = NULL;
3006 int nid;
3007 unsigned int cpuset_mems_cookie;
3008
3009 if (flags & __GFP_THISNODE)
3010 return NULL;
3011
3012 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3013
3014retry_cpuset:
3015 cpuset_mems_cookie = read_mems_allowed_begin();
3016 zonelist = node_zonelist(mempolicy_slab_node(), flags);
3017
3018retry:
3019 /*
3020 * Look through allowed nodes for objects available
3021 * from existing per node queues.
3022 */
3023 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3024 nid = zone_to_nid(zone);
3025
3026 if (cpuset_zone_allowed(zone, flags) &&
3027 get_node(cache, nid) &&
3028 get_node(cache, nid)->free_objects) {
3029 obj = ____cache_alloc_node(cache,
3030 gfp_exact_node(flags), nid);
3031 if (obj)
3032 break;
3033 }
3034 }
3035
3036 if (!obj) {
3037 /*
3038 * This allocation will be performed within the constraints
3039 * of the current cpuset / memory policy requirements.
3040 * We may trigger various forms of reclaim on the allowed
3041 * set and go into memory reserves if necessary.
3042 */
3043 struct page *page;
3044
3045 if (gfpflags_allow_blocking(local_flags))
3046 local_irq_enable();
3047 kmem_flagcheck(cache, flags);
3048 page = kmem_getpages(cache, local_flags, numa_mem_id());
3049 if (gfpflags_allow_blocking(local_flags))
3050 local_irq_disable();
3051 if (page) {
3052 /*
3053 * Insert into the appropriate per node queues
3054 */
3055 nid = page_to_nid(page);
3056 if (cache_grow(cache, flags, nid, page)) {
3057 obj = ____cache_alloc_node(cache,
3058 gfp_exact_node(flags), nid);
3059 if (!obj)
3060 /*
3061 * Another processor may allocate the
3062 * objects in the slab since we are
3063 * not holding any locks.
3064 */
3065 goto retry;
3066 } else {
3067 /* cache_grow already freed obj */
3068 obj = NULL;
3069 }
3070 }
3071 }
3072
3073 if (unlikely(!obj && read_mems_allowed_retry(cpuset_mems_cookie)))
3074 goto retry_cpuset;
3075 return obj;
3076}
3077
3078/*
3079 * A interface to enable slab creation on nodeid
3080 */
3081static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3082 int nodeid)
3083{
3084 struct page *page;
3085 struct kmem_cache_node *n;
3086 void *obj;
3087 void *list = NULL;
3088 int x;
3089
3090 VM_BUG_ON(nodeid < 0 || nodeid >= MAX_NUMNODES);
3091 n = get_node(cachep, nodeid);
3092 BUG_ON(!n);
3093
3094retry:
3095 check_irq_off();
3096 spin_lock(&n->list_lock);
3097 page = get_first_slab(n, false);
3098 if (!page)
3099 goto must_grow;
3100
3101 check_spinlock_acquired_node(cachep, nodeid);
3102
3103 STATS_INC_NODEALLOCS(cachep);
3104 STATS_INC_ACTIVE(cachep);
3105 STATS_SET_HIGH(cachep);
3106
3107 BUG_ON(page->active == cachep->num);
3108
3109 obj = slab_get_obj(cachep, page);
3110 n->free_objects--;
3111
3112 fixup_slab_list(cachep, n, page, &list);
3113
3114 spin_unlock(&n->list_lock);
3115 fixup_objfreelist_debug(cachep, &list);
3116 goto done;
3117
3118must_grow:
3119 spin_unlock(&n->list_lock);
3120 x = cache_grow(cachep, gfp_exact_node(flags), nodeid, NULL);
3121 if (x)
3122 goto retry;
3123
3124 return fallback_alloc(cachep, flags);
3125
3126done:
3127 return obj;
3128}
3129
3130static __always_inline void *
3131slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3132 unsigned long caller)
3133{
3134 unsigned long save_flags;
3135 void *ptr;
3136 int slab_node = numa_mem_id();
3137
3138 flags &= gfp_allowed_mask;
3139 cachep = slab_pre_alloc_hook(cachep, flags);
3140 if (unlikely(!cachep))
3141 return NULL;
3142
3143 cache_alloc_debugcheck_before(cachep, flags);
3144 local_irq_save(save_flags);
3145
3146 if (nodeid == NUMA_NO_NODE)
3147 nodeid = slab_node;
3148
3149 if (unlikely(!get_node(cachep, nodeid))) {
3150 /* Node not bootstrapped yet */
3151 ptr = fallback_alloc(cachep, flags);
3152 goto out;
3153 }
3154
3155 if (nodeid == slab_node) {
3156 /*
3157 * Use the locally cached objects if possible.
3158 * However ____cache_alloc does not allow fallback
3159 * to other nodes. It may fail while we still have
3160 * objects on other nodes available.
3161 */
3162 ptr = ____cache_alloc(cachep, flags);
3163 if (ptr)
3164 goto out;
3165 }
3166 /* ___cache_alloc_node can fall back to other nodes */
3167 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3168 out:
3169 local_irq_restore(save_flags);
3170 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3171
3172 if (unlikely(flags & __GFP_ZERO) && ptr)
3173 memset(ptr, 0, cachep->object_size);
3174
3175 slab_post_alloc_hook(cachep, flags, 1, &ptr);
3176 return ptr;
3177}
3178
3179static __always_inline void *
3180__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3181{
3182 void *objp;
3183
3184 if (current->mempolicy || cpuset_do_slab_mem_spread()) {
3185 objp = alternate_node_alloc(cache, flags);
3186 if (objp)
3187 goto out;
3188 }
3189 objp = ____cache_alloc(cache, flags);
3190
3191 /*
3192 * We may just have run out of memory on the local node.
3193 * ____cache_alloc_node() knows how to locate memory on other nodes
3194 */
3195 if (!objp)
3196 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3197
3198 out:
3199 return objp;
3200}
3201#else
3202
3203static __always_inline void *
3204__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3205{
3206 return ____cache_alloc(cachep, flags);
3207}
3208
3209#endif /* CONFIG_NUMA */
3210
3211static __always_inline void *
3212slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
3213{
3214 unsigned long save_flags;
3215 void *objp;
3216
3217 flags &= gfp_allowed_mask;
3218 cachep = slab_pre_alloc_hook(cachep, flags);
3219 if (unlikely(!cachep))
3220 return NULL;
3221
3222 cache_alloc_debugcheck_before(cachep, flags);
3223 local_irq_save(save_flags);
3224 objp = __do_cache_alloc(cachep, flags);
3225 local_irq_restore(save_flags);
3226 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3227 prefetchw(objp);
3228
3229 if (unlikely(flags & __GFP_ZERO) && objp)
3230 memset(objp, 0, cachep->object_size);
3231
3232 slab_post_alloc_hook(cachep, flags, 1, &objp);
3233 return objp;
3234}
3235
3236/*
3237 * Caller needs to acquire correct kmem_cache_node's list_lock
3238 * @list: List of detached free slabs should be freed by caller
3239 */
3240static void free_block(struct kmem_cache *cachep, void **objpp,
3241 int nr_objects, int node, struct list_head *list)
3242{
3243 int i;
3244 struct kmem_cache_node *n = get_node(cachep, node);
3245
3246 for (i = 0; i < nr_objects; i++) {
3247 void *objp;
3248 struct page *page;
3249
3250 objp = objpp[i];
3251
3252 page = virt_to_head_page(objp);
3253 list_del(&page->lru);
3254 check_spinlock_acquired_node(cachep, node);
3255 slab_put_obj(cachep, page, objp);
3256 STATS_DEC_ACTIVE(cachep);
3257 n->free_objects++;
3258
3259 /* fixup slab chains */
3260 if (page->active == 0) {
3261 if (n->free_objects > n->free_limit) {
3262 n->free_objects -= cachep->num;
3263 list_add_tail(&page->lru, list);
3264 } else {
3265 list_add(&page->lru, &n->slabs_free);
3266 }
3267 } else {
3268 /* Unconditionally move a slab to the end of the
3269 * partial list on free - maximum time for the
3270 * other objects to be freed, too.
3271 */
3272 list_add_tail(&page->lru, &n->slabs_partial);
3273 }
3274 }
3275}
3276
3277static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3278{
3279 int batchcount;
3280 struct kmem_cache_node *n;
3281 int node = numa_mem_id();
3282 LIST_HEAD(list);
3283
3284 batchcount = ac->batchcount;
3285
3286 check_irq_off();
3287 n = get_node(cachep, node);
3288 spin_lock(&n->list_lock);
3289 if (n->shared) {
3290 struct array_cache *shared_array = n->shared;
3291 int max = shared_array->limit - shared_array->avail;
3292 if (max) {
3293 if (batchcount > max)
3294 batchcount = max;
3295 memcpy(&(shared_array->entry[shared_array->avail]),
3296 ac->entry, sizeof(void *) * batchcount);
3297 shared_array->avail += batchcount;
3298 goto free_done;
3299 }
3300 }
3301
3302 free_block(cachep, ac->entry, batchcount, node, &list);
3303free_done:
3304#if STATS
3305 {
3306 int i = 0;
3307 struct page *page;
3308
3309 list_for_each_entry(page, &n->slabs_free, lru) {
3310 BUG_ON(page->active);
3311
3312 i++;
3313 }
3314 STATS_SET_FREEABLE(cachep, i);
3315 }
3316#endif
3317 spin_unlock(&n->list_lock);
3318 slabs_destroy(cachep, &list);
3319 ac->avail -= batchcount;
3320 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3321}
3322
3323/*
3324 * Release an obj back to its cache. If the obj has a constructed state, it must
3325 * be in this state _before_ it is released. Called with disabled ints.
3326 */
3327static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3328 unsigned long caller)
3329{
3330 struct array_cache *ac = cpu_cache_get(cachep);
3331
3332 kasan_slab_free(cachep, objp);
3333
3334 check_irq_off();
3335 kmemleak_free_recursive(objp, cachep->flags);
3336 objp = cache_free_debugcheck(cachep, objp, caller);
3337
3338 kmemcheck_slab_free(cachep, objp, cachep->object_size);
3339
3340 /*
3341 * Skip calling cache_free_alien() when the platform is not numa.
3342 * This will avoid cache misses that happen while accessing slabp (which
3343 * is per page memory reference) to get nodeid. Instead use a global
3344 * variable to skip the call, which is mostly likely to be present in
3345 * the cache.
3346 */
3347 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3348 return;
3349
3350 if (ac->avail < ac->limit) {
3351 STATS_INC_FREEHIT(cachep);
3352 } else {
3353 STATS_INC_FREEMISS(cachep);
3354 cache_flusharray(cachep, ac);
3355 }
3356
3357 if (sk_memalloc_socks()) {
3358 struct page *page = virt_to_head_page(objp);
3359
3360 if (unlikely(PageSlabPfmemalloc(page))) {
3361 cache_free_pfmemalloc(cachep, page, objp);
3362 return;
3363 }
3364 }
3365
3366 ac->entry[ac->avail++] = objp;
3367}
3368
3369/**
3370 * kmem_cache_alloc - Allocate an object
3371 * @cachep: The cache to allocate from.
3372 * @flags: See kmalloc().
3373 *
3374 * Allocate an object from this cache. The flags are only relevant
3375 * if the cache has no available objects.
3376 */
3377void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3378{
3379 void *ret = slab_alloc(cachep, flags, _RET_IP_);
3380
3381 kasan_slab_alloc(cachep, ret, flags);
3382 trace_kmem_cache_alloc(_RET_IP_, ret,
3383 cachep->object_size, cachep->size, flags);
3384
3385 return ret;
3386}
3387EXPORT_SYMBOL(kmem_cache_alloc);
3388
3389static __always_inline void
3390cache_alloc_debugcheck_after_bulk(struct kmem_cache *s, gfp_t flags,
3391 size_t size, void **p, unsigned long caller)
3392{
3393 size_t i;
3394
3395 for (i = 0; i < size; i++)
3396 p[i] = cache_alloc_debugcheck_after(s, flags, p[i], caller);
3397}
3398
3399int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3400 void **p)
3401{
3402 size_t i;
3403
3404 s = slab_pre_alloc_hook(s, flags);
3405 if (!s)
3406 return 0;
3407
3408 cache_alloc_debugcheck_before(s, flags);
3409
3410 local_irq_disable();
3411 for (i = 0; i < size; i++) {
3412 void *objp = __do_cache_alloc(s, flags);
3413
3414 if (unlikely(!objp))
3415 goto error;
3416 p[i] = objp;
3417 }
3418 local_irq_enable();
3419
3420 cache_alloc_debugcheck_after_bulk(s, flags, size, p, _RET_IP_);
3421
3422 /* Clear memory outside IRQ disabled section */
3423 if (unlikely(flags & __GFP_ZERO))
3424 for (i = 0; i < size; i++)
3425 memset(p[i], 0, s->object_size);
3426
3427 slab_post_alloc_hook(s, flags, size, p);
3428 /* FIXME: Trace call missing. Christoph would like a bulk variant */
3429 return size;
3430error:
3431 local_irq_enable();
3432 cache_alloc_debugcheck_after_bulk(s, flags, i, p, _RET_IP_);
3433 slab_post_alloc_hook(s, flags, i, p);
3434 __kmem_cache_free_bulk(s, i, p);
3435 return 0;
3436}
3437EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3438
3439#ifdef CONFIG_TRACING
3440void *
3441kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
3442{
3443 void *ret;
3444
3445 ret = slab_alloc(cachep, flags, _RET_IP_);
3446
3447 kasan_kmalloc(cachep, ret, size, flags);
3448 trace_kmalloc(_RET_IP_, ret,
3449 size, cachep->size, flags);
3450 return ret;
3451}
3452EXPORT_SYMBOL(kmem_cache_alloc_trace);
3453#endif
3454
3455#ifdef CONFIG_NUMA
3456/**
3457 * kmem_cache_alloc_node - Allocate an object on the specified node
3458 * @cachep: The cache to allocate from.
3459 * @flags: See kmalloc().
3460 * @nodeid: node number of the target node.
3461 *
3462 * Identical to kmem_cache_alloc but it will allocate memory on the given
3463 * node, which can improve the performance for cpu bound structures.
3464 *
3465 * Fallback to other node is possible if __GFP_THISNODE is not set.
3466 */
3467void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3468{
3469 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3470
3471 kasan_slab_alloc(cachep, ret, flags);
3472 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3473 cachep->object_size, cachep->size,
3474 flags, nodeid);
3475
3476 return ret;
3477}
3478EXPORT_SYMBOL(kmem_cache_alloc_node);
3479
3480#ifdef CONFIG_TRACING
3481void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
3482 gfp_t flags,
3483 int nodeid,
3484 size_t size)
3485{
3486 void *ret;
3487
3488 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3489
3490 kasan_kmalloc(cachep, ret, size, flags);
3491 trace_kmalloc_node(_RET_IP_, ret,
3492 size, cachep->size,
3493 flags, nodeid);
3494 return ret;
3495}
3496EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3497#endif
3498
3499static __always_inline void *
3500__do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
3501{
3502 struct kmem_cache *cachep;
3503 void *ret;
3504
3505 cachep = kmalloc_slab(size, flags);
3506 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3507 return cachep;
3508 ret = kmem_cache_alloc_node_trace(cachep, flags, node, size);
3509 kasan_kmalloc(cachep, ret, size, flags);
3510
3511 return ret;
3512}
3513
3514void *__kmalloc_node(size_t size, gfp_t flags, int node)
3515{
3516 return __do_kmalloc_node(size, flags, node, _RET_IP_);
3517}
3518EXPORT_SYMBOL(__kmalloc_node);
3519
3520void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3521 int node, unsigned long caller)
3522{
3523 return __do_kmalloc_node(size, flags, node, caller);
3524}
3525EXPORT_SYMBOL(__kmalloc_node_track_caller);
3526#endif /* CONFIG_NUMA */
3527
3528/**
3529 * __do_kmalloc - allocate memory
3530 * @size: how many bytes of memory are required.
3531 * @flags: the type of memory to allocate (see kmalloc).
3532 * @caller: function caller for debug tracking of the caller
3533 */
3534static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3535 unsigned long caller)
3536{
3537 struct kmem_cache *cachep;
3538 void *ret;
3539
3540 cachep = kmalloc_slab(size, flags);
3541 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3542 return cachep;
3543 ret = slab_alloc(cachep, flags, caller);
3544
3545 kasan_kmalloc(cachep, ret, size, flags);
3546 trace_kmalloc(caller, ret,
3547 size, cachep->size, flags);
3548
3549 return ret;
3550}
3551
3552void *__kmalloc(size_t size, gfp_t flags)
3553{
3554 return __do_kmalloc(size, flags, _RET_IP_);
3555}
3556EXPORT_SYMBOL(__kmalloc);
3557
3558void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3559{
3560 return __do_kmalloc(size, flags, caller);
3561}
3562EXPORT_SYMBOL(__kmalloc_track_caller);
3563
3564/**
3565 * kmem_cache_free - Deallocate an object
3566 * @cachep: The cache the allocation was from.
3567 * @objp: The previously allocated object.
3568 *
3569 * Free an object which was previously allocated from this
3570 * cache.
3571 */
3572void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3573{
3574 unsigned long flags;
3575 cachep = cache_from_obj(cachep, objp);
3576 if (!cachep)
3577 return;
3578
3579 local_irq_save(flags);
3580 debug_check_no_locks_freed(objp, cachep->object_size);
3581 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3582 debug_check_no_obj_freed(objp, cachep->object_size);
3583 __cache_free(cachep, objp, _RET_IP_);
3584 local_irq_restore(flags);
3585
3586 trace_kmem_cache_free(_RET_IP_, objp);
3587}
3588EXPORT_SYMBOL(kmem_cache_free);
3589
3590void kmem_cache_free_bulk(struct kmem_cache *orig_s, size_t size, void **p)
3591{
3592 struct kmem_cache *s;
3593 size_t i;
3594
3595 local_irq_disable();
3596 for (i = 0; i < size; i++) {
3597 void *objp = p[i];
3598
3599 if (!orig_s) /* called via kfree_bulk */
3600 s = virt_to_cache(objp);
3601 else
3602 s = cache_from_obj(orig_s, objp);
3603
3604 debug_check_no_locks_freed(objp, s->object_size);
3605 if (!(s->flags & SLAB_DEBUG_OBJECTS))
3606 debug_check_no_obj_freed(objp, s->object_size);
3607
3608 __cache_free(s, objp, _RET_IP_);
3609 }
3610 local_irq_enable();
3611
3612 /* FIXME: add tracing */
3613}
3614EXPORT_SYMBOL(kmem_cache_free_bulk);
3615
3616/**
3617 * kfree - free previously allocated memory
3618 * @objp: pointer returned by kmalloc.
3619 *
3620 * If @objp is NULL, no operation is performed.
3621 *
3622 * Don't free memory not originally allocated by kmalloc()
3623 * or you will run into trouble.
3624 */
3625void kfree(const void *objp)
3626{
3627 struct kmem_cache *c;
3628 unsigned long flags;
3629
3630 trace_kfree(_RET_IP_, objp);
3631
3632 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3633 return;
3634 local_irq_save(flags);
3635 kfree_debugcheck(objp);
3636 c = virt_to_cache(objp);
3637 debug_check_no_locks_freed(objp, c->object_size);
3638
3639 debug_check_no_obj_freed(objp, c->object_size);
3640 __cache_free(c, (void *)objp, _RET_IP_);
3641 local_irq_restore(flags);
3642}
3643EXPORT_SYMBOL(kfree);
3644
3645/*
3646 * This initializes kmem_cache_node or resizes various caches for all nodes.
3647 */
3648static int alloc_kmem_cache_node(struct kmem_cache *cachep, gfp_t gfp)
3649{
3650 int node;
3651 struct kmem_cache_node *n;
3652 struct array_cache *new_shared;
3653 struct alien_cache **new_alien = NULL;
3654
3655 for_each_online_node(node) {
3656
3657 if (use_alien_caches) {
3658 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3659 if (!new_alien)
3660 goto fail;
3661 }
3662
3663 new_shared = NULL;
3664 if (cachep->shared) {
3665 new_shared = alloc_arraycache(node,
3666 cachep->shared*cachep->batchcount,
3667 0xbaadf00d, gfp);
3668 if (!new_shared) {
3669 free_alien_cache(new_alien);
3670 goto fail;
3671 }
3672 }
3673
3674 n = get_node(cachep, node);
3675 if (n) {
3676 struct array_cache *shared = n->shared;
3677 LIST_HEAD(list);
3678
3679 spin_lock_irq(&n->list_lock);
3680
3681 if (shared)
3682 free_block(cachep, shared->entry,
3683 shared->avail, node, &list);
3684
3685 n->shared = new_shared;
3686 if (!n->alien) {
3687 n->alien = new_alien;
3688 new_alien = NULL;
3689 }
3690 n->free_limit = (1 + nr_cpus_node(node)) *
3691 cachep->batchcount + cachep->num;
3692 spin_unlock_irq(&n->list_lock);
3693 slabs_destroy(cachep, &list);
3694 kfree(shared);
3695 free_alien_cache(new_alien);
3696 continue;
3697 }
3698 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
3699 if (!n) {
3700 free_alien_cache(new_alien);
3701 kfree(new_shared);
3702 goto fail;
3703 }
3704
3705 kmem_cache_node_init(n);
3706 n->next_reap = jiffies + REAPTIMEOUT_NODE +
3707 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
3708 n->shared = new_shared;
3709 n->alien = new_alien;
3710 n->free_limit = (1 + nr_cpus_node(node)) *
3711 cachep->batchcount + cachep->num;
3712 cachep->node[node] = n;
3713 }
3714 return 0;
3715
3716fail:
3717 if (!cachep->list.next) {
3718 /* Cache is not active yet. Roll back what we did */
3719 node--;
3720 while (node >= 0) {
3721 n = get_node(cachep, node);
3722 if (n) {
3723 kfree(n->shared);
3724 free_alien_cache(n->alien);
3725 kfree(n);
3726 cachep->node[node] = NULL;
3727 }
3728 node--;
3729 }
3730 }
3731 return -ENOMEM;
3732}
3733
3734/* Always called with the slab_mutex held */
3735static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
3736 int batchcount, int shared, gfp_t gfp)
3737{
3738 struct array_cache __percpu *cpu_cache, *prev;
3739 int cpu;
3740
3741 cpu_cache = alloc_kmem_cache_cpus(cachep, limit, batchcount);
3742 if (!cpu_cache)
3743 return -ENOMEM;
3744
3745 prev = cachep->cpu_cache;
3746 cachep->cpu_cache = cpu_cache;
3747 kick_all_cpus_sync();
3748
3749 check_irq_on();
3750 cachep->batchcount = batchcount;
3751 cachep->limit = limit;
3752 cachep->shared = shared;
3753
3754 if (!prev)
3755 goto alloc_node;
3756
3757 for_each_online_cpu(cpu) {
3758 LIST_HEAD(list);
3759 int node;
3760 struct kmem_cache_node *n;
3761 struct array_cache *ac = per_cpu_ptr(prev, cpu);
3762
3763 node = cpu_to_mem(cpu);
3764 n = get_node(cachep, node);
3765 spin_lock_irq(&n->list_lock);
3766 free_block(cachep, ac->entry, ac->avail, node, &list);
3767 spin_unlock_irq(&n->list_lock);
3768 slabs_destroy(cachep, &list);
3769 }
3770 free_percpu(prev);
3771
3772alloc_node:
3773 return alloc_kmem_cache_node(cachep, gfp);
3774}
3775
3776static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3777 int batchcount, int shared, gfp_t gfp)
3778{
3779 int ret;
3780 struct kmem_cache *c;
3781
3782 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3783
3784 if (slab_state < FULL)
3785 return ret;
3786
3787 if ((ret < 0) || !is_root_cache(cachep))
3788 return ret;
3789
3790 lockdep_assert_held(&slab_mutex);
3791 for_each_memcg_cache(c, cachep) {
3792 /* return value determined by the root cache only */
3793 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
3794 }
3795
3796 return ret;
3797}
3798
3799/* Called with slab_mutex held always */
3800static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
3801{
3802 int err;
3803 int limit = 0;
3804 int shared = 0;
3805 int batchcount = 0;
3806
3807 if (!is_root_cache(cachep)) {
3808 struct kmem_cache *root = memcg_root_cache(cachep);
3809 limit = root->limit;
3810 shared = root->shared;
3811 batchcount = root->batchcount;
3812 }
3813
3814 if (limit && shared && batchcount)
3815 goto skip_setup;
3816 /*
3817 * The head array serves three purposes:
3818 * - create a LIFO ordering, i.e. return objects that are cache-warm
3819 * - reduce the number of spinlock operations.
3820 * - reduce the number of linked list operations on the slab and
3821 * bufctl chains: array operations are cheaper.
3822 * The numbers are guessed, we should auto-tune as described by
3823 * Bonwick.
3824 */
3825 if (cachep->size > 131072)
3826 limit = 1;
3827 else if (cachep->size > PAGE_SIZE)
3828 limit = 8;
3829 else if (cachep->size > 1024)
3830 limit = 24;
3831 else if (cachep->size > 256)
3832 limit = 54;
3833 else
3834 limit = 120;
3835
3836 /*
3837 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3838 * allocation behaviour: Most allocs on one cpu, most free operations
3839 * on another cpu. For these cases, an efficient object passing between
3840 * cpus is necessary. This is provided by a shared array. The array
3841 * replaces Bonwick's magazine layer.
3842 * On uniprocessor, it's functionally equivalent (but less efficient)
3843 * to a larger limit. Thus disabled by default.
3844 */
3845 shared = 0;
3846 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
3847 shared = 8;
3848
3849#if DEBUG
3850 /*
3851 * With debugging enabled, large batchcount lead to excessively long
3852 * periods with disabled local interrupts. Limit the batchcount
3853 */
3854 if (limit > 32)
3855 limit = 32;
3856#endif
3857 batchcount = (limit + 1) / 2;
3858skip_setup:
3859 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3860 if (err)
3861 pr_err("enable_cpucache failed for %s, error %d\n",
3862 cachep->name, -err);
3863 return err;
3864}
3865
3866/*
3867 * Drain an array if it contains any elements taking the node lock only if
3868 * necessary. Note that the node listlock also protects the array_cache
3869 * if drain_array() is used on the shared array.
3870 */
3871static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
3872 struct array_cache *ac, int force, int node)
3873{
3874 LIST_HEAD(list);
3875 int tofree;
3876
3877 if (!ac || !ac->avail)
3878 return;
3879 if (ac->touched && !force) {
3880 ac->touched = 0;
3881 } else {
3882 spin_lock_irq(&n->list_lock);
3883 if (ac->avail) {
3884 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3885 if (tofree > ac->avail)
3886 tofree = (ac->avail + 1) / 2;
3887 free_block(cachep, ac->entry, tofree, node, &list);
3888 ac->avail -= tofree;
3889 memmove(ac->entry, &(ac->entry[tofree]),
3890 sizeof(void *) * ac->avail);
3891 }
3892 spin_unlock_irq(&n->list_lock);
3893 slabs_destroy(cachep, &list);
3894 }
3895}
3896
3897/**
3898 * cache_reap - Reclaim memory from caches.
3899 * @w: work descriptor
3900 *
3901 * Called from workqueue/eventd every few seconds.
3902 * Purpose:
3903 * - clear the per-cpu caches for this CPU.
3904 * - return freeable pages to the main free memory pool.
3905 *
3906 * If we cannot acquire the cache chain mutex then just give up - we'll try
3907 * again on the next iteration.
3908 */
3909static void cache_reap(struct work_struct *w)
3910{
3911 struct kmem_cache *searchp;
3912 struct kmem_cache_node *n;
3913 int node = numa_mem_id();
3914 struct delayed_work *work = to_delayed_work(w);
3915
3916 if (!mutex_trylock(&slab_mutex))
3917 /* Give up. Setup the next iteration. */
3918 goto out;
3919
3920 list_for_each_entry(searchp, &slab_caches, list) {
3921 check_irq_on();
3922
3923 /*
3924 * We only take the node lock if absolutely necessary and we
3925 * have established with reasonable certainty that
3926 * we can do some work if the lock was obtained.
3927 */
3928 n = get_node(searchp, node);
3929
3930 reap_alien(searchp, n);
3931
3932 drain_array(searchp, n, cpu_cache_get(searchp), 0, node);
3933
3934 /*
3935 * These are racy checks but it does not matter
3936 * if we skip one check or scan twice.
3937 */
3938 if (time_after(n->next_reap, jiffies))
3939 goto next;
3940
3941 n->next_reap = jiffies + REAPTIMEOUT_NODE;
3942
3943 drain_array(searchp, n, n->shared, 0, node);
3944
3945 if (n->free_touched)
3946 n->free_touched = 0;
3947 else {
3948 int freed;
3949
3950 freed = drain_freelist(searchp, n, (n->free_limit +
3951 5 * searchp->num - 1) / (5 * searchp->num));
3952 STATS_ADD_REAPED(searchp, freed);
3953 }
3954next:
3955 cond_resched();
3956 }
3957 check_irq_on();
3958 mutex_unlock(&slab_mutex);
3959 next_reap_node();
3960out:
3961 /* Set up the next iteration */
3962 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_AC));
3963}
3964
3965#ifdef CONFIG_SLABINFO
3966void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
3967{
3968 struct page *page;
3969 unsigned long active_objs;
3970 unsigned long num_objs;
3971 unsigned long active_slabs = 0;
3972 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
3973 const char *name;
3974 char *error = NULL;
3975 int node;
3976 struct kmem_cache_node *n;
3977
3978 active_objs = 0;
3979 num_slabs = 0;
3980 for_each_kmem_cache_node(cachep, node, n) {
3981
3982 check_irq_on();
3983 spin_lock_irq(&n->list_lock);
3984
3985 list_for_each_entry(page, &n->slabs_full, lru) {
3986 if (page->active != cachep->num && !error)
3987 error = "slabs_full accounting error";
3988 active_objs += cachep->num;
3989 active_slabs++;
3990 }
3991 list_for_each_entry(page, &n->slabs_partial, lru) {
3992 if (page->active == cachep->num && !error)
3993 error = "slabs_partial accounting error";
3994 if (!page->active && !error)
3995 error = "slabs_partial accounting error";
3996 active_objs += page->active;
3997 active_slabs++;
3998 }
3999 list_for_each_entry(page, &n->slabs_free, lru) {
4000 if (page->active && !error)
4001 error = "slabs_free accounting error";
4002 num_slabs++;
4003 }
4004 free_objects += n->free_objects;
4005 if (n->shared)
4006 shared_avail += n->shared->avail;
4007
4008 spin_unlock_irq(&n->list_lock);
4009 }
4010 num_slabs += active_slabs;
4011 num_objs = num_slabs * cachep->num;
4012 if (num_objs - active_objs != free_objects && !error)
4013 error = "free_objects accounting error";
4014
4015 name = cachep->name;
4016 if (error)
4017 pr_err("slab: cache %s error: %s\n", name, error);
4018
4019 sinfo->active_objs = active_objs;
4020 sinfo->num_objs = num_objs;
4021 sinfo->active_slabs = active_slabs;
4022 sinfo->num_slabs = num_slabs;
4023 sinfo->shared_avail = shared_avail;
4024 sinfo->limit = cachep->limit;
4025 sinfo->batchcount = cachep->batchcount;
4026 sinfo->shared = cachep->shared;
4027 sinfo->objects_per_slab = cachep->num;
4028 sinfo->cache_order = cachep->gfporder;
4029}
4030
4031void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4032{
4033#if STATS
4034 { /* node stats */
4035 unsigned long high = cachep->high_mark;
4036 unsigned long allocs = cachep->num_allocations;
4037 unsigned long grown = cachep->grown;
4038 unsigned long reaped = cachep->reaped;
4039 unsigned long errors = cachep->errors;
4040 unsigned long max_freeable = cachep->max_freeable;
4041 unsigned long node_allocs = cachep->node_allocs;
4042 unsigned long node_frees = cachep->node_frees;
4043 unsigned long overflows = cachep->node_overflow;
4044
4045 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu %4lu %4lu",
4046 allocs, high, grown,
4047 reaped, errors, max_freeable, node_allocs,
4048 node_frees, overflows);
4049 }
4050 /* cpu stats */
4051 {
4052 unsigned long allochit = atomic_read(&cachep->allochit);
4053 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4054 unsigned long freehit = atomic_read(&cachep->freehit);
4055 unsigned long freemiss = atomic_read(&cachep->freemiss);
4056
4057 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4058 allochit, allocmiss, freehit, freemiss);
4059 }
4060#endif
4061}
4062
4063#define MAX_SLABINFO_WRITE 128
4064/**
4065 * slabinfo_write - Tuning for the slab allocator
4066 * @file: unused
4067 * @buffer: user buffer
4068 * @count: data length
4069 * @ppos: unused
4070 */
4071ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4072 size_t count, loff_t *ppos)
4073{
4074 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4075 int limit, batchcount, shared, res;
4076 struct kmem_cache *cachep;
4077
4078 if (count > MAX_SLABINFO_WRITE)
4079 return -EINVAL;
4080 if (copy_from_user(&kbuf, buffer, count))
4081 return -EFAULT;
4082 kbuf[MAX_SLABINFO_WRITE] = '\0';
4083
4084 tmp = strchr(kbuf, ' ');
4085 if (!tmp)
4086 return -EINVAL;
4087 *tmp = '\0';
4088 tmp++;
4089 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4090 return -EINVAL;
4091
4092 /* Find the cache in the chain of caches. */
4093 mutex_lock(&slab_mutex);
4094 res = -EINVAL;
4095 list_for_each_entry(cachep, &slab_caches, list) {
4096 if (!strcmp(cachep->name, kbuf)) {
4097 if (limit < 1 || batchcount < 1 ||
4098 batchcount > limit || shared < 0) {
4099 res = 0;
4100 } else {
4101 res = do_tune_cpucache(cachep, limit,
4102 batchcount, shared,
4103 GFP_KERNEL);
4104 }
4105 break;
4106 }
4107 }
4108 mutex_unlock(&slab_mutex);
4109 if (res >= 0)
4110 res = count;
4111 return res;
4112}
4113
4114#ifdef CONFIG_DEBUG_SLAB_LEAK
4115
4116static inline int add_caller(unsigned long *n, unsigned long v)
4117{
4118 unsigned long *p;
4119 int l;
4120 if (!v)
4121 return 1;
4122 l = n[1];
4123 p = n + 2;
4124 while (l) {
4125 int i = l/2;
4126 unsigned long *q = p + 2 * i;
4127 if (*q == v) {
4128 q[1]++;
4129 return 1;
4130 }
4131 if (*q > v) {
4132 l = i;
4133 } else {
4134 p = q + 2;
4135 l -= i + 1;
4136 }
4137 }
4138 if (++n[1] == n[0])
4139 return 0;
4140 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4141 p[0] = v;
4142 p[1] = 1;
4143 return 1;
4144}
4145
4146static void handle_slab(unsigned long *n, struct kmem_cache *c,
4147 struct page *page)
4148{
4149 void *p;
4150 int i, j;
4151 unsigned long v;
4152
4153 if (n[0] == n[1])
4154 return;
4155 for (i = 0, p = page->s_mem; i < c->num; i++, p += c->size) {
4156 bool active = true;
4157
4158 for (j = page->active; j < c->num; j++) {
4159 if (get_free_obj(page, j) == i) {
4160 active = false;
4161 break;
4162 }
4163 }
4164
4165 if (!active)
4166 continue;
4167
4168 /*
4169 * probe_kernel_read() is used for DEBUG_PAGEALLOC. page table
4170 * mapping is established when actual object allocation and
4171 * we could mistakenly access the unmapped object in the cpu
4172 * cache.
4173 */
4174 if (probe_kernel_read(&v, dbg_userword(c, p), sizeof(v)))
4175 continue;
4176
4177 if (!add_caller(n, v))
4178 return;
4179 }
4180}
4181
4182static void show_symbol(struct seq_file *m, unsigned long address)
4183{
4184#ifdef CONFIG_KALLSYMS
4185 unsigned long offset, size;
4186 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4187
4188 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4189 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4190 if (modname[0])
4191 seq_printf(m, " [%s]", modname);
4192 return;
4193 }
4194#endif
4195 seq_printf(m, "%p", (void *)address);
4196}
4197
4198static int leaks_show(struct seq_file *m, void *p)
4199{
4200 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4201 struct page *page;
4202 struct kmem_cache_node *n;
4203 const char *name;
4204 unsigned long *x = m->private;
4205 int node;
4206 int i;
4207
4208 if (!(cachep->flags & SLAB_STORE_USER))
4209 return 0;
4210 if (!(cachep->flags & SLAB_RED_ZONE))
4211 return 0;
4212
4213 /*
4214 * Set store_user_clean and start to grab stored user information
4215 * for all objects on this cache. If some alloc/free requests comes
4216 * during the processing, information would be wrong so restart
4217 * whole processing.
4218 */
4219 do {
4220 set_store_user_clean(cachep);
4221 drain_cpu_caches(cachep);
4222
4223 x[1] = 0;
4224
4225 for_each_kmem_cache_node(cachep, node, n) {
4226
4227 check_irq_on();
4228 spin_lock_irq(&n->list_lock);
4229
4230 list_for_each_entry(page, &n->slabs_full, lru)
4231 handle_slab(x, cachep, page);
4232 list_for_each_entry(page, &n->slabs_partial, lru)
4233 handle_slab(x, cachep, page);
4234 spin_unlock_irq(&n->list_lock);
4235 }
4236 } while (!is_store_user_clean(cachep));
4237
4238 name = cachep->name;
4239 if (x[0] == x[1]) {
4240 /* Increase the buffer size */
4241 mutex_unlock(&slab_mutex);
4242 m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4243 if (!m->private) {
4244 /* Too bad, we are really out */
4245 m->private = x;
4246 mutex_lock(&slab_mutex);
4247 return -ENOMEM;
4248 }
4249 *(unsigned long *)m->private = x[0] * 2;
4250 kfree(x);
4251 mutex_lock(&slab_mutex);
4252 /* Now make sure this entry will be retried */
4253 m->count = m->size;
4254 return 0;
4255 }
4256 for (i = 0; i < x[1]; i++) {
4257 seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4258 show_symbol(m, x[2*i+2]);
4259 seq_putc(m, '\n');
4260 }
4261
4262 return 0;
4263}
4264
4265static const struct seq_operations slabstats_op = {
4266 .start = slab_start,
4267 .next = slab_next,
4268 .stop = slab_stop,
4269 .show = leaks_show,
4270};
4271
4272static int slabstats_open(struct inode *inode, struct file *file)
4273{
4274 unsigned long *n;
4275
4276 n = __seq_open_private(file, &slabstats_op, PAGE_SIZE);
4277 if (!n)
4278 return -ENOMEM;
4279
4280 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4281
4282 return 0;
4283}
4284
4285static const struct file_operations proc_slabstats_operations = {
4286 .open = slabstats_open,
4287 .read = seq_read,
4288 .llseek = seq_lseek,
4289 .release = seq_release_private,
4290};
4291#endif
4292
4293static int __init slab_proc_init(void)
4294{
4295#ifdef CONFIG_DEBUG_SLAB_LEAK
4296 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4297#endif
4298 return 0;
4299}
4300module_init(slab_proc_init);
4301#endif
4302
4303/**
4304 * ksize - get the actual amount of memory allocated for a given object
4305 * @objp: Pointer to the object
4306 *
4307 * kmalloc may internally round up allocations and return more memory
4308 * than requested. ksize() can be used to determine the actual amount of
4309 * memory allocated. The caller may use this additional memory, even though
4310 * a smaller amount of memory was initially specified with the kmalloc call.
4311 * The caller must guarantee that objp points to a valid object previously
4312 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4313 * must not be freed during the duration of the call.
4314 */
4315size_t ksize(const void *objp)
4316{
4317 size_t size;
4318
4319 BUG_ON(!objp);
4320 if (unlikely(objp == ZERO_SIZE_PTR))
4321 return 0;
4322
4323 size = virt_to_cache(objp)->object_size;
4324 /* We assume that ksize callers could use the whole allocated area,
4325 * so we need to unpoison this area.
4326 */
4327 kasan_krealloc(objp, size, GFP_NOWAIT);
4328
4329 return size;
4330}
4331EXPORT_SYMBOL(ksize);