Loading...
Note: File does not exist in v6.8.
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/cpu.h>
104#include <linux/__KEEPIDENTS__BD.h>
105#include <linux/__KEEPIDENTS__BE.h>
106#include <linux/rcupdate.h>
107#include <linux/__KEEPIDENTS__BF.h>
108#include <linux/__KEEPIDENTS__BG.h>
109#include <linux/__KEEPIDENTS__BH.h>
110#include <linux/kmemleak.h>
111#include <linux/__KEEPIDENTS__BI.h>
112#include <linux/__KEEPIDENTS__BJ.h>
113#include <linux/__KEEPIDENTS__CA-__KEEPIDENTS__CB.h>
114#include <linux/__KEEPIDENTS__CC.h>
115#include <linux/reciprocal_div.h>
116#include <linux/debugobjects.h>
117#include <linux/__KEEPIDENTS__CD.h>
118#include <linux/__KEEPIDENTS__CE.h>
119#include <linux/__KEEPIDENTS__CF/task_stack.h>
120
121#include <net/__KEEPIDENTS__CG.h>
122
123#include <asm/cacheflush.h>
124#include <asm/tlbflush.h>
125#include <asm/page.h>
126
127#include <trace/events/kmem.h>
128
129#include "internal.h"
130
131#include "slab.h"
132
133/*
134 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
135 * 0 for faster, smaller code (especially in the critical paths).
136 *
137 * STATS - 1 to collect stats for /proc/slabinfo.
138 * 0 for faster, smaller code (especially in the critical paths).
139 *
140 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
141 */
142
143#ifdef CONFIG_DEBUG_SLAB
144#define DEBUG 1
145#define STATS 1
146#define FORCED_DEBUG 1
147#else
148#define DEBUG 0
149#define STATS 0
150#define FORCED_DEBUG 0
151#endif
152
153/* Shouldn't this be in a header file somewhere? */
154#define BYTES_PER_WORD sizeof(void *)
155#define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
156
157#ifndef ARCH_KMALLOC_FLAGS
158#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
159#endif
160
161#define FREELIST_BYTE_INDEX (((PAGE_SIZE >> BITS_PER_BYTE) \
162 <= SLAB_OBJ_MIN_SIZE) ? 1 : 0)
163
164#if FREELIST_BYTE_INDEX
165typedef unsigned char freelist_idx_t;
166#else
167typedef unsigned short freelist_idx_t;
168#endif
169
170#define SLAB_OBJ_MAX_NUM ((1 << sizeof(freelist_idx_t) * BITS_PER_BYTE) - 1)
171
172/*
173 * struct array_cache
174 *
175 * Purpose:
176 * - LIFO ordering, to hand out cache-warm objects from _alloc
177 * - reduce the number of linked list operations
178 * - reduce spinlock operations
179 *
180 * The limit is stored in the per-cpu structure to reduce the data cache
181 * footprint.
182 *
183 */
184struct array_cache {
185 unsigned int avail;
186 unsigned int limit;
187 unsigned int batchcount;
188 unsigned int touched;
189 void *entry[]; /*
190 * Must have this definition in here for the proper
191 * alignment of array_cache. Also simplifies accessing
192 * the entries.
193 */
194};
195
196struct alien_cache {
197 spinlock_t lock;
198 struct array_cache ac;
199};
200
201/*
202 * Need this for bootstrapping a per node allocator.
203 */
204#define NUM_INIT_LISTS (2 * MAX_NUMNODES)
205static struct kmem_cache_node __initdata init_kmem_cache_node[NUM_INIT_LISTS];
206#define CACHE_CACHE 0
207#define SIZE_NODE (MAX_NUMNODES)
208
209static int drain_freelist(struct kmem_cache *cache,
210 struct kmem_cache_node *n, int tofree);
211static void free_block(struct kmem_cache *cachep, void **objpp, int len,
212 int node, struct list_head *list);
213static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list);
214static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
215static void cache_reap(struct work_struct *unused);
216
217static inline void fixup_objfreelist_debug(struct kmem_cache *cachep,
218 void **list);
219static inline void fixup_slab_list(struct kmem_cache *cachep,
220 struct kmem_cache_node *n, struct page *page,
221 void **list);
222static int slab_early_init = 1;
223
224#define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
225
226static void kmem_cache_node_init(struct kmem_cache_node *parent)
227{
228 INIT_LIST_HEAD(&parent->slabs_full);
229 INIT_LIST_HEAD(&parent->slabs_partial);
230 INIT_LIST_HEAD(&parent->slabs_free);
231 parent->total_slabs = 0;
232 parent->free_slabs = 0;
233 parent->shared = NULL;
234 parent->alien = NULL;
235 parent->colour_next = 0;
236 spin_lock_init(&parent->list_lock);
237 parent->free_objects = 0;
238 parent->free_touched = 0;
239}
240
241#define MAKE_LIST(cachep, listp, slab, nodeid) \
242 do { \
243 INIT_LIST_HEAD(listp); \
244 list_splice(&get_node(cachep, nodeid)->slab, listp); \
245 } while (0)
246
247#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
248 do { \
249 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
250 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
251 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
252 } while (0)
253
254#define CFLGS_OBJFREELIST_SLAB ((slab_flags_t __force)0x40000000U)
255#define CFLGS_OFF_SLAB ((slab_flags_t __force)0x80000000U)
256#define OBJFREELIST_SLAB(x) ((x)->flags & CFLGS_OBJFREELIST_SLAB)
257#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
258
259#define BATCHREFILL_LIMIT 16
260/*
261 * Optimization question: fewer reaps means less probability for unnessary
262 * cpucache drain/refill cycles.
263 *
264 * OTOH the cpuarrays can contain lots of objects,
265 * which could lock up otherwise freeable slabs.
266 */
267#define REAPTIMEOUT_AC (2*HZ)
268#define REAPTIMEOUT_NODE (4*HZ)
269
270#if STATS
271#define STATS_INC_ACTIVE(x) ((x)->num_active++)
272#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
273#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
274#define STATS_INC_GROWN(x) ((x)->grown++)
275#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
276#define STATS_SET_HIGH(x) \
277 do { \
278 if ((x)->num_active > (x)->high_mark) \
279 (x)->high_mark = (x)->num_active; \
280 } while (0)
281#define STATS_INC_ERR(x) ((x)->errors++)
282#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
283#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
284#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
285#define STATS_SET_FREEABLE(x, i) \
286 do { \
287 if ((x)->max_freeable < i) \
288 (x)->max_freeable = i; \
289 } while (0)
290#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
291#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
292#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
293#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
294#else
295#define STATS_INC_ACTIVE(x) do { } while (0)
296#define STATS_DEC_ACTIVE(x) do { } while (0)
297#define STATS_INC_ALLOCED(x) do { } while (0)
298#define STATS_INC_GROWN(x) do { } while (0)
299#define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
300#define STATS_SET_HIGH(x) do { } while (0)
301#define STATS_INC_ERR(x) do { } while (0)
302#define STATS_INC_NODEALLOCS(x) do { } while (0)
303#define STATS_INC_NODEFREES(x) do { } while (0)
304#define STATS_INC_ACOVERFLOW(x) do { } while (0)
305#define STATS_SET_FREEABLE(x, i) do { } while (0)
306#define STATS_INC_ALLOCHIT(x) do { } while (0)
307#define STATS_INC_ALLOCMISS(x) do { } while (0)
308#define STATS_INC_FREEHIT(x) do { } while (0)
309#define STATS_INC_FREEMISS(x) do { } while (0)
310#endif
311
312#if DEBUG
313
314/*
315 * memory layout of objects:
316 * 0 : objp
317 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
318 * the end of an object is aligned with the end of the real
319 * allocation. Catches writes behind the end of the allocation.
320 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
321 * redzone word.
322 * cachep->obj_offset: The real object.
323 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
324 * cachep->size - 1* BYTES_PER_WORD: last caller address
325 * [BYTES_PER_WORD long]
326 */
327static int obj_offset(struct kmem_cache *cachep)
328{
329 return cachep->obj_offset;
330}
331
332static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
333{
334 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
335 return (unsigned long long*) (objp + obj_offset(cachep) -
336 sizeof(unsigned long long));
337}
338
339static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
340{
341 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
342 if (cachep->flags & SLAB_STORE_USER)
343 return (unsigned long long *)(objp + cachep->size -
344 sizeof(unsigned long long) -
345 REDZONE_ALIGN);
346 return (unsigned long long *) (objp + cachep->size -
347 sizeof(unsigned long long));
348}
349
350static void **dbg_userword(struct kmem_cache *cachep, void *objp)
351{
352 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
353 return (void **)(objp + cachep->size - BYTES_PER_WORD);
354}
355
356#else
357
358#define obj_offset(x) 0
359#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
360#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
361#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
362
363#endif
364
365/*
366 * Do not go above this order unless 0 objects fit into the slab or
367 * overridden on the command line.
368 */
369#define SLAB_MAX_ORDER_HI 1
370#define SLAB_MAX_ORDER_LO 0
371static int slab_max_order = SLAB_MAX_ORDER_LO;
372static bool slab_max_order_set __initdata;
373
374static inline void *index_to_obj(struct kmem_cache *cache, struct page *page,
375 unsigned int idx)
376{
377 return page->s_mem + cache->size * idx;
378}
379
380#define BOOT_CPUCACHE_ENTRIES 1
381/* internal cache of cache description objs */
382static struct kmem_cache kmem_cache_boot = {
383 .batchcount = 1,
384 .limit = BOOT_CPUCACHE_ENTRIES,
385 .shared = 1,
386 .size = sizeof(struct kmem_cache),
387 .name = "kmem_cache",
388};
389
390static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
391
392static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
393{
394 return this_cpu_ptr(cachep->cpu_cache);
395}
396
397/*
398 * Calculate the number of objects and left-over bytes for a given buffer size.
399 */
400static unsigned int cache_estimate(unsigned long gfporder, size_t buffer_size,
401 slab_flags_t flags, size_t *left_over)
402{
403 unsigned int num;
404 size_t slab_size = PAGE_SIZE << gfporder;
405
406 /*
407 * The slab management structure can be either off the slab or
408 * on it. For the latter case, the memory allocated for a
409 * slab is used for:
410 *
411 * - @buffer_size bytes for each object
412 * - One freelist_idx_t for each object
413 *
414 * We don't need to consider alignment of freelist because
415 * freelist will be at the end of slab page. The objects will be
416 * at the correct alignment.
417 *
418 * If the slab management structure is off the slab, then the
419 * alignment will already be calculated into the size. Because
420 * the slabs are all pages aligned, the objects will be at the
421 * correct alignment when allocated.
422 */
423 if (flags & (CFLGS_OBJFREELIST_SLAB | CFLGS_OFF_SLAB)) {
424 num = slab_size / buffer_size;
425 *left_over = slab_size % buffer_size;
426 } else {
427 num = slab_size / (buffer_size + sizeof(freelist_idx_t));
428 *left_over = slab_size %
429 (buffer_size + sizeof(freelist_idx_t));
430 }
431
432 return num;
433}
434
435#if DEBUG
436#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
437
438static void __slab_error(const char *function, struct kmem_cache *cachep,
439 char *msg)
440{
441 pr_err("slab error in %s(): cache `%s': %s\n",
442 function, cachep->name, msg);
443 dump_stack();
444 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
445}
446#endif
447
448/*
449 * By default on NUMA we use alien caches to stage the freeing of
450 * objects allocated from other nodes. This causes massive memory
451 * inefficiencies when using fake NUMA setup to split memory into a
452 * large number of small nodes, so it can be disabled on the command
453 * line
454 */
455
456static int use_alien_caches __read_mostly = 1;
457static int __init noaliencache_setup(char *s)
458{
459 use_alien_caches = 0;
460 return 1;
461}
462__setup("noaliencache", noaliencache_setup);
463
464static int __init slab_max_order_setup(char *str)
465{
466 get_option(&str, &slab_max_order);
467 slab_max_order = slab_max_order < 0 ? 0 :
468 min(slab_max_order, MAX_ORDER - 1);
469 slab_max_order_set = true;
470
471 return 1;
472}
473__setup("slab_max_order=", slab_max_order_setup);
474
475#ifdef CONFIG_NUMA
476/*
477 * Special reaping functions for NUMA systems called from cache_reap().
478 * These take care of doing round robin flushing of alien caches (containing
479 * objects freed on different nodes from which they were allocated) and the
480 * flushing of remote pcps by calling drain_node_pages.
481 */
482static DEFINE_PER_CPU(unsigned long, slab_reap_node);
483
484static void init_reap_node(int cpu)
485{
486 per_cpu(slab_reap_node, cpu) = next_node_in(cpu_to_mem(cpu),
487 node_online_map);
488}
489
490static void next_reap_node(void)
491{
492 int node = __this_cpu_read(slab_reap_node);
493
494 node = next_node_in(node, node_online_map);
495 __this_cpu_write(slab_reap_node, node);
496}
497
498#else
499#define init_reap_node(cpu) do { } while (0)
500#define next_reap_node(void) do { } while (0)
501#endif
502
503/*
504 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
505 * via the workqueue/eventd.
506 * Add the CPU number into the expiration time to minimize the possibility of
507 * the CPUs getting into lockstep and contending for the global cache chain
508 * lock.
509 */
510static void start_cpu_timer(int cpu)
511{
512 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
513
514 if (reap_work->work.func == NULL) {
515 init_reap_node(cpu);
516 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
517 schedule_delayed_work_on(cpu, reap_work,
518 __round_jiffies_relative(HZ, cpu));
519 }
520}
521
522static void init_arraycache(struct array_cache *ac, int limit, int batch)
523{
524 if (ac) {
525 ac->avail = 0;
526 ac->limit = limit;
527 ac->batchcount = batch;
528 ac->touched = 0;
529 }
530}
531
532static struct array_cache *alloc_arraycache(int node, int entries,
533 int batchcount, gfp_t gfp)
534{
535 size_t memsize = sizeof(void *) * entries + sizeof(struct array_cache);
536 struct array_cache *ac = NULL;
537
538 ac = kmalloc_node(memsize, gfp, node);
539 /*
540 * The array_cache structures contain pointers to free object.
541 * However, when such objects are allocated or transferred to another
542 * cache the pointers are not cleared and they could be counted as
543 * valid references during a kmemleak scan. Therefore, kmemleak must
544 * not scan such objects.
545 */
546 kmemleak_no_scan(ac);
547 init_arraycache(ac, entries, batchcount);
548 return ac;
549}
550
551static noinline void cache_free_pfmemalloc(struct kmem_cache *cachep,
552 struct page *page, void *objp)
553{
554 struct kmem_cache_node *n;
555 int page_node;
556 LIST_HEAD(list);
557
558 page_node = page_to_nid(page);
559 n = get_node(cachep, page_node);
560
561 spin_lock(&n->list_lock);
562 free_block(cachep, &objp, 1, page_node, &list);
563 spin_unlock(&n->list_lock);
564
565 slabs_destroy(cachep, &list);
566}
567
568/*
569 * Transfer objects in one arraycache to another.
570 * Locking must be handled by the caller.
571 *
572 * Return the number of entries transferred.
573 */
574static int transfer_objects(struct array_cache *to,
575 struct array_cache *from, unsigned int max)
576{
577 /* Figure out how many entries to transfer */
578 int nr = min3(from->avail, max, to->limit - to->avail);
579
580 if (!nr)
581 return 0;
582
583 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
584 sizeof(void *) *nr);
585
586 from->avail -= nr;
587 to->avail += nr;
588 return nr;
589}
590
591/* &alien->lock must be held by alien callers. */
592static __always_inline void __free_one(struct array_cache *ac, void *objp)
593{
594 /* Avoid trivial double-free. */
595 if (IS_ENABLED(CONFIG_SLAB_FREELIST_HARDENED) &&
596 WARN_ON_ONCE(ac->avail > 0 && ac->entry[ac->avail - 1] == objp))
597 return;
598 ac->entry[ac->avail++] = objp;
599}
600
601#ifndef CONFIG_NUMA
602
603#define drain_alien_cache(cachep, alien) do { } while (0)
604#define reap_alien(cachep, n) do { } while (0)
605
606static inline struct alien_cache **alloc_alien_cache(int node,
607 int limit, gfp_t gfp)
608{
609 return NULL;
610}
611
612static inline void free_alien_cache(struct alien_cache **ac_ptr)
613{
614}
615
616static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
617{
618 return 0;
619}
620
621static inline void *alternate_node_alloc(struct kmem_cache *cachep,
622 gfp_t flags)
623{
624 return NULL;
625}
626
627static inline void *____cache_alloc_node(struct kmem_cache *cachep,
628 gfp_t flags, int nodeid)
629{
630 return NULL;
631}
632
633static inline gfp_t gfp_exact_node(gfp_t flags)
634{
635 return flags & ~__GFP_NOFAIL;
636}
637
638#else /* CONFIG_NUMA */
639
640static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
641static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
642
643static struct alien_cache *__alloc_alien_cache(int node, int entries,
644 int batch, gfp_t gfp)
645{
646 size_t memsize = sizeof(void *) * entries + sizeof(struct alien_cache);
647 struct alien_cache *alc = NULL;
648
649 alc = kmalloc_node(memsize, gfp, node);
650 if (alc) {
651 kmemleak_no_scan(alc);
652 init_arraycache(&alc->ac, entries, batch);
653 spin_lock_init(&alc->lock);
654 }
655 return alc;
656}
657
658static struct alien_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
659{
660 struct alien_cache **alc_ptr;
661 int i;
662
663 if (limit > 1)
664 limit = 12;
665 alc_ptr = kcalloc_node(nr_node_ids, sizeof(void *), gfp, node);
666 if (!alc_ptr)
667 return NULL;
668
669 for_each_node(i) {
670 if (i == node || !node_online(i))
671 continue;
672 alc_ptr[i] = __alloc_alien_cache(node, limit, 0xbaadf00d, gfp);
673 if (!alc_ptr[i]) {
674 for (i--; i >= 0; i--)
675 kfree(alc_ptr[i]);
676 kfree(alc_ptr);
677 return NULL;
678 }
679 }
680 return alc_ptr;
681}
682
683static void free_alien_cache(struct alien_cache **alc_ptr)
684{
685 int i;
686
687 if (!alc_ptr)
688 return;
689 for_each_node(i)
690 kfree(alc_ptr[i]);
691 kfree(alc_ptr);
692}
693
694static void __drain_alien_cache(struct kmem_cache *cachep,
695 struct array_cache *ac, int node,
696 struct list_head *list)
697{
698 struct kmem_cache_node *n = get_node(cachep, node);
699
700 if (ac->avail) {
701 spin_lock(&n->list_lock);
702 /*
703 * Stuff objects into the remote nodes shared array first.
704 * That way we could avoid the overhead of putting the objects
705 * into the free lists and getting them back later.
706 */
707 if (n->shared)
708 transfer_objects(n->shared, ac, ac->limit);
709
710 free_block(cachep, ac->entry, ac->avail, node, list);
711 ac->avail = 0;
712 spin_unlock(&n->list_lock);
713 }
714}
715
716/*
717 * Called from cache_reap() to regularly drain alien caches round robin.
718 */
719static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
720{
721 int node = __this_cpu_read(slab_reap_node);
722
723 if (n->alien) {
724 struct alien_cache *alc = n->alien[node];
725 struct array_cache *ac;
726
727 if (alc) {
728 ac = &alc->ac;
729 if (ac->avail && spin_trylock_irq(&alc->lock)) {
730 LIST_HEAD(list);
731
732 __drain_alien_cache(cachep, ac, node, &list);
733 spin_unlock_irq(&alc->lock);
734 slabs_destroy(cachep, &list);
735 }
736 }
737 }
738}
739
740static void drain_alien_cache(struct kmem_cache *cachep,
741 struct alien_cache **alien)
742{
743 int i = 0;
744 struct alien_cache *alc;
745 struct array_cache *ac;
746 unsigned long flags;
747
748 for_each_online_node(i) {
749 alc = alien[i];
750 if (alc) {
751 LIST_HEAD(list);
752
753 ac = &alc->ac;
754 spin_lock_irqsave(&alc->lock, flags);
755 __drain_alien_cache(cachep, ac, i, &list);
756 spin_unlock_irqrestore(&alc->lock, flags);
757 slabs_destroy(cachep, &list);
758 }
759 }
760}
761
762static int __cache_free_alien(struct kmem_cache *cachep, void *objp,
763 int node, int page_node)
764{
765 struct kmem_cache_node *n;
766 struct alien_cache *alien = NULL;
767 struct array_cache *ac;
768 LIST_HEAD(list);
769
770 n = get_node(cachep, node);
771 STATS_INC_NODEFREES(cachep);
772 if (n->alien && n->alien[page_node]) {
773 alien = n->alien[page_node];
774 ac = &alien->ac;
775 spin_lock(&alien->lock);
776 if (unlikely(ac->avail == ac->limit)) {
777 STATS_INC_ACOVERFLOW(cachep);
778 __drain_alien_cache(cachep, ac, page_node, &list);
779 }
780 __free_one(ac, objp);
781 spin_unlock(&alien->lock);
782 slabs_destroy(cachep, &list);
783 } else {
784 n = get_node(cachep, page_node);
785 spin_lock(&n->list_lock);
786 free_block(cachep, &objp, 1, page_node, &list);
787 spin_unlock(&n->list_lock);
788 slabs_destroy(cachep, &list);
789 }
790 return 1;
791}
792
793static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
794{
795 int page_node = page_to_nid(virt_to_page(objp));
796 int node = numa_mem_id();
797 /*
798 * Make sure we are not freeing a object from another node to the array
799 * cache on this cpu.
800 */
801 if (likely(node == page_node))
802 return 0;
803
804 return __cache_free_alien(cachep, objp, node, page_node);
805}
806
807/*
808 * Construct gfp mask to allocate from a specific node but do not reclaim or
809 * warn about failures.
810 */
811static inline gfp_t gfp_exact_node(gfp_t flags)
812{
813 return (flags | __GFP_THISNODE | __GFP_NOWARN) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
814}
815#endif
816
817static int init_cache_node(struct kmem_cache *cachep, int node, gfp_t gfp)
818{
819 struct kmem_cache_node *n;
820
821 /*
822 * Set up the kmem_cache_node for cpu before we can
823 * begin anything. Make sure some other cpu on this
824 * node has not already allocated this
825 */
826 n = get_node(cachep, node);
827 if (n) {
828 spin_lock_irq(&n->list_lock);
829 n->free_limit = (1 + nr_cpus_node(node)) * cachep->batchcount +
830 cachep->num;
831 spin_unlock_irq(&n->list_lock);
832
833 return 0;
834 }
835
836 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
837 if (!n)
838 return -ENOMEM;
839
840 kmem_cache_node_init(n);
841 n->next_reap = jiffies + REAPTIMEOUT_NODE +
842 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
843
844 n->free_limit =
845 (1 + nr_cpus_node(node)) * cachep->batchcount + cachep->num;
846
847 /*
848 * The kmem_cache_nodes don't come and go as CPUs
849 * come and go. slab_mutex is sufficient
850 * protection here.
851 */
852 cachep->node[node] = n;
853
854 return 0;
855}
856
857#if (defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)) || defined(CONFIG_SMP)
858/*
859 * Allocates and initializes node for a node on each slab cache, used for
860 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
861 * will be allocated off-node since memory is not yet online for the new node.
862 * When hotplugging memory or a cpu, existing node are not replaced if
863 * already in use.
864 *
865 * Must hold slab_mutex.
866 */
867static int init_cache_node_node(int node)
868{
869 int ret;
870 struct kmem_cache *cachep;
871
872 list_for_each_entry(cachep, &slab_caches, list) {
873 ret = init_cache_node(cachep, node, GFP_KERNEL);
874 if (ret)
875 return ret;
876 }
877
878 return 0;
879}
880#endif
881
882static int setup_kmem_cache_node(struct kmem_cache *cachep,
883 int node, gfp_t gfp, bool force_change)
884{
885 int ret = -ENOMEM;
886 struct kmem_cache_node *n;
887 struct array_cache *old_shared = NULL;
888 struct array_cache *new_shared = NULL;
889 struct alien_cache **new_alien = NULL;
890 LIST_HEAD(list);
891
892 if (use_alien_caches) {
893 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
894 if (!new_alien)
895 goto fail;
896 }
897
898 if (cachep->shared) {
899 new_shared = alloc_arraycache(node,
900 cachep->shared * cachep->batchcount, 0xbaadf00d, gfp);
901 if (!new_shared)
902 goto fail;
903 }
904
905 ret = init_cache_node(cachep, node, gfp);
906 if (ret)
907 goto fail;
908
909 n = get_node(cachep, node);
910 spin_lock_irq(&n->list_lock);
911 if (n->shared && force_change) {
912 free_block(cachep, n->shared->entry,
913 n->shared->avail, node, &list);
914 n->shared->avail = 0;
915 }
916
917 if (!n->shared || force_change) {
918 old_shared = n->shared;
919 n->shared = new_shared;
920 new_shared = NULL;
921 }
922
923 if (!n->alien) {
924 n->alien = new_alien;
925 new_alien = NULL;
926 }
927
928 spin_unlock_irq(&n->list_lock);
929 slabs_destroy(cachep, &list);
930
931 /*
932 * To protect lockless access to n->shared during irq disabled context.
933 * If n->shared isn't NULL in irq disabled context, accessing to it is
934 * guaranteed to be valid until irq is re-enabled, because it will be
935 * freed after synchronize_rcu().
936 */
937 if (old_shared && force_change)
938 synchronize_rcu();
939
940fail:
941 kfree(old_shared);
942 kfree(new_shared);
943 free_alien_cache(new_alien);
944
945 return ret;
946}
947
948#ifdef CONFIG_SMP
949
950static void cpuup_canceled(long cpu)
951{
952 struct kmem_cache *cachep;
953 struct kmem_cache_node *n = NULL;
954 int node = cpu_to_mem(cpu);
955 const struct cpumask *mask = cpumask_of_node(node);
956
957 list_for_each_entry(cachep, &slab_caches, list) {
958 struct array_cache *nc;
959 struct array_cache *shared;
960 struct alien_cache **alien;
961 LIST_HEAD(list);
962
963 n = get_node(cachep, node);
964 if (!n)
965 continue;
966
967 spin_lock_irq(&n->list_lock);
968
969 /* Free limit for this kmem_cache_node */
970 n->free_limit -= cachep->batchcount;
971
972 /* cpu is dead; no one can alloc from it. */
973 nc = per_cpu_ptr(cachep->cpu_cache, cpu);
974 free_block(cachep, nc->entry, nc->avail, node, &list);
975 nc->avail = 0;
976
977 if (!cpumask_empty(mask)) {
978 spin_unlock_irq(&n->list_lock);
979 goto free_slab;
980 }
981
982 shared = n->shared;
983 if (shared) {
984 free_block(cachep, shared->entry,
985 shared->avail, node, &list);
986 n->shared = NULL;
987 }
988
989 alien = n->alien;
990 n->alien = NULL;
991
992 spin_unlock_irq(&n->list_lock);
993
994 kfree(shared);
995 if (alien) {
996 drain_alien_cache(cachep, alien);
997 free_alien_cache(alien);
998 }
999
1000free_slab:
1001 slabs_destroy(cachep, &list);
1002 }
1003 /*
1004 * In the previous loop, all the objects were freed to
1005 * the respective cache's slabs, now we can go ahead and
1006 * shrink each nodelist to its limit.
1007 */
1008 list_for_each_entry(cachep, &slab_caches, list) {
1009 n = get_node(cachep, node);
1010 if (!n)
1011 continue;
1012 drain_freelist(cachep, n, INT_MAX);
1013 }
1014}
1015
1016static int cpuup_prepare(long cpu)
1017{
1018 struct kmem_cache *cachep;
1019 int node = cpu_to_mem(cpu);
1020 int err;
1021
1022 /*
1023 * We need to do this right in the beginning since
1024 * alloc_arraycache's are going to use this list.
1025 * kmalloc_node allows us to add the slab to the right
1026 * kmem_cache_node and not this cpu's kmem_cache_node
1027 */
1028 err = init_cache_node_node(node);
1029 if (err < 0)
1030 goto bad;
1031
1032 /*
1033 * Now we can go ahead with allocating the shared arrays and
1034 * array caches
1035 */
1036 list_for_each_entry(cachep, &slab_caches, list) {
1037 err = setup_kmem_cache_node(cachep, node, GFP_KERNEL, false);
1038 if (err)
1039 goto bad;
1040 }
1041
1042 return 0;
1043bad:
1044 cpuup_canceled(cpu);
1045 return -ENOMEM;
1046}
1047
1048int slab_prepare_cpu(unsigned int cpu)
1049{
1050 int err;
1051
1052 mutex_lock(&slab_mutex);
1053 err = cpuup_prepare(cpu);
1054 mutex_unlock(&slab_mutex);
1055 return err;
1056}
1057
1058/*
1059 * This is called for a failed online attempt and for a successful
1060 * offline.
1061 *
1062 * Even if all the cpus of a node are down, we don't free the
1063 * kmem_cache_node of any cache. This to avoid a race between cpu_down, and
1064 * a kmalloc allocation from another cpu for memory from the node of
1065 * the cpu going down. The list3 structure is usually allocated from
1066 * kmem_cache_create() and gets destroyed at kmem_cache_destroy().
1067 */
1068int slab_dead_cpu(unsigned int cpu)
1069{
1070 mutex_lock(&slab_mutex);
1071 cpuup_canceled(cpu);
1072 mutex_unlock(&slab_mutex);
1073 return 0;
1074}
1075#endif
1076
1077static int slab_online_cpu(unsigned int cpu)
1078{
1079 start_cpu_timer(cpu);
1080 return 0;
1081}
1082
1083static int slab_offline_cpu(unsigned int cpu)
1084{
1085 /*
1086 * Shutdown cache reaper. Note that the slab_mutex is held so
1087 * that if cache_reap() is invoked it cannot do anything
1088 * expensive but will only modify reap_work and reschedule the
1089 * timer.
1090 */
1091 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1092 /* Now the cache_reaper is guaranteed to be not running. */
1093 per_cpu(slab_reap_work, cpu).work.func = NULL;
1094 return 0;
1095}
1096
1097#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1098/*
1099 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1100 * Returns -EBUSY if all objects cannot be drained so that the node is not
1101 * removed.
1102 *
1103 * Must hold slab_mutex.
1104 */
1105static int __meminit drain_cache_node_node(int node)
1106{
1107 struct kmem_cache *cachep;
1108 int ret = 0;
1109
1110 list_for_each_entry(cachep, &slab_caches, list) {
1111 struct kmem_cache_node *n;
1112
1113 n = get_node(cachep, node);
1114 if (!n)
1115 continue;
1116
1117 drain_freelist(cachep, n, INT_MAX);
1118
1119 if (!list_empty(&n->slabs_full) ||
1120 !list_empty(&n->slabs_partial)) {
1121 ret = -EBUSY;
1122 break;
1123 }
1124 }
1125 return ret;
1126}
1127
1128static int __meminit slab_memory_callback(struct notifier_block *self,
1129 unsigned long action, void *arg)
1130{
1131 struct memory_notify *mnb = arg;
1132 int ret = 0;
1133 int nid;
1134
1135 nid = mnb->status_change_nid;
1136 if (nid < 0)
1137 goto out;
1138
1139 switch (action) {
1140 case MEM_GOING_ONLINE:
1141 mutex_lock(&slab_mutex);
1142 ret = init_cache_node_node(nid);
1143 mutex_unlock(&slab_mutex);
1144 break;
1145 case MEM_GOING_OFFLINE:
1146 mutex_lock(&slab_mutex);
1147 ret = drain_cache_node_node(nid);
1148 mutex_unlock(&slab_mutex);
1149 break;
1150 case MEM_ONLINE:
1151 case MEM_OFFLINE:
1152 case MEM_CANCEL_ONLINE:
1153 case MEM_CANCEL_OFFLINE:
1154 break;
1155 }
1156out:
1157 return notifier_from_errno(ret);
1158}
1159#endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1160
1161/*
1162 * swap the static kmem_cache_node with kmalloced memory
1163 */
1164static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
1165 int nodeid)
1166{
1167 struct kmem_cache_node *ptr;
1168
1169 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
1170 BUG_ON(!ptr);
1171
1172 memcpy(ptr, list, sizeof(struct kmem_cache_node));
1173 /*
1174 * Do not assume that spinlocks can be initialized via memcpy:
1175 */
1176 spin_lock_init(&ptr->list_lock);
1177
1178 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1179 cachep->node[nodeid] = ptr;
1180}
1181
1182/*
1183 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1184 * size of kmem_cache_node.
1185 */
1186static void __init set_up_node(struct kmem_cache *cachep, int index)
1187{
1188 int node;
1189
1190 for_each_online_node(node) {
1191 cachep->node[node] = &init_kmem_cache_node[index + node];
1192 cachep->node[node]->next_reap = jiffies +
1193 REAPTIMEOUT_NODE +
1194 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1195 }
1196}
1197
1198/*
1199 * Initialisation. Called after the page allocator have been initialised and
1200 * before smp_init().
1201 */
1202void __init kmem_cache_init(void)
1203{
1204 int i;
1205
1206 kmem_cache = &kmem_cache_boot;
1207
1208 if (!IS_ENABLED(CONFIG_NUMA) || num_possible_nodes() == 1)
1209 use_alien_caches = 0;
1210
1211 for (i = 0; i < NUM_INIT_LISTS; i++)
1212 kmem_cache_node_init(&init_kmem_cache_node[i]);
1213
1214 /*
1215 * Fragmentation resistance on low memory - only use bigger
1216 * page orders on machines with more than 32MB of memory if
1217 * not overridden on the command line.
1218 */
1219 if (!slab_max_order_set && totalram_pages() > (32 << 20) >> PAGE_SHIFT)
1220 slab_max_order = SLAB_MAX_ORDER_HI;
1221
1222 /* Bootstrap is tricky, because several objects are allocated
1223 * from caches that do not exist yet:
1224 * 1) initialize the kmem_cache cache: it contains the struct
1225 * kmem_cache structures of all caches, except kmem_cache itself:
1226 * kmem_cache is statically allocated.
1227 * Initially an __init data area is used for the head array and the
1228 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1229 * array at the end of the bootstrap.
1230 * 2) Create the first kmalloc cache.
1231 * The struct kmem_cache for the new cache is allocated normally.
1232 * An __init data area is used for the head array.
1233 * 3) Create the remaining kmalloc caches, with minimally sized
1234 * head arrays.
1235 * 4) Replace the __init data head arrays for kmem_cache and the first
1236 * kmalloc cache with kmalloc allocated arrays.
1237 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1238 * the other cache's with kmalloc allocated memory.
1239 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1240 */
1241
1242 /* 1) create the kmem_cache */
1243
1244 /*
1245 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1246 */
1247 create_boot_cache(kmem_cache, "kmem_cache",
1248 offsetof(struct kmem_cache, node) +
1249 nr_node_ids * sizeof(struct kmem_cache_node *),
1250 SLAB_HWCACHE_ALIGN, 0, 0);
1251 list_add(&kmem_cache->list, &slab_caches);
1252 slab_state = PARTIAL;
1253
1254 /*
1255 * Initialize the caches that provide memory for the kmem_cache_node
1256 * structures first. Without this, further allocations will bug.
1257 */
1258 kmalloc_caches[KMALLOC_NORMAL][INDEX_NODE] = create_kmalloc_cache(
1259 kmalloc_info[INDEX_NODE].name[KMALLOC_NORMAL],
1260 kmalloc_info[INDEX_NODE].size,
1261 ARCH_KMALLOC_FLAGS, 0,
1262 kmalloc_info[INDEX_NODE].size);
1263 slab_state = PARTIAL_NODE;
1264 setup_kmalloc_cache_index_table();
1265
1266 slab_early_init = 0;
1267
1268 /* 5) Replace the bootstrap kmem_cache_node */
1269 {
1270 int nid;
1271
1272 for_each_online_node(nid) {
1273 init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
1274
1275 init_list(kmalloc_caches[KMALLOC_NORMAL][INDEX_NODE],
1276 &init_kmem_cache_node[SIZE_NODE + nid], nid);
1277 }
1278 }
1279
1280 create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
1281}
1282
1283void __init kmem_cache_init_late(void)
1284{
1285 struct kmem_cache *cachep;
1286
1287 /* 6) resize the head arrays to their final sizes */
1288 mutex_lock(&slab_mutex);
1289 list_for_each_entry(cachep, &slab_caches, list)
1290 if (enable_cpucache(cachep, GFP_NOWAIT))
1291 BUG();
1292 mutex_unlock(&slab_mutex);
1293
1294 /* Done! */
1295 slab_state = FULL;
1296
1297#ifdef CONFIG_NUMA
1298 /*
1299 * Register a memory hotplug callback that initializes and frees
1300 * node.
1301 */
1302 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1303#endif
1304
1305 /*
1306 * The reap timers are started later, with a module init call: That part
1307 * of the kernel is not yet operational.
1308 */
1309}
1310
1311static int __init cpucache_init(void)
1312{
1313 int ret;
1314
1315 /*
1316 * Register the timers that return unneeded pages to the page allocator
1317 */
1318 ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "SLAB online",
1319 slab_online_cpu, slab_offline_cpu);
1320 WARN_ON(ret < 0);
1321
1322 return 0;
1323}
1324__initcall(cpucache_init);
1325
1326static noinline void
1327slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1328{
1329#if DEBUG
1330 struct kmem_cache_node *n;
1331 unsigned long flags;
1332 int node;
1333 static DEFINE_RATELIMIT_STATE(slab_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1334 DEFAULT_RATELIMIT_BURST);
1335
1336 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slab_oom_rs))
1337 return;
1338
1339 pr_warn("SLAB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
1340 nodeid, gfpflags, &gfpflags);
1341 pr_warn(" cache: %s, object size: %d, order: %d\n",
1342 cachep->name, cachep->size, cachep->gfporder);
1343
1344 for_each_kmem_cache_node(cachep, node, n) {
1345 unsigned long total_slabs, free_slabs, free_objs;
1346
1347 spin_lock_irqsave(&n->list_lock, flags);
1348 total_slabs = n->total_slabs;
1349 free_slabs = n->free_slabs;
1350 free_objs = n->free_objects;
1351 spin_unlock_irqrestore(&n->list_lock, flags);
1352
1353 pr_warn(" node %d: slabs: %ld/%ld, objs: %ld/%ld\n",
1354 node, total_slabs - free_slabs, total_slabs,
1355 (total_slabs * cachep->num) - free_objs,
1356 total_slabs * cachep->num);
1357 }
1358#endif
1359}
1360
1361/*
1362 * Interface to system's page allocator. No need to hold the
1363 * kmem_cache_node ->list_lock.
1364 *
1365 * If we requested dmaable memory, we will get it. Even if we
1366 * did not request dmaable memory, we might get it, but that
1367 * would be relatively rare and ignorable.
1368 */
1369static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1370 int nodeid)
1371{
1372 struct page *page;
1373
1374 flags |= cachep->allocflags;
1375
1376 page = __alloc_pages_node(nodeid, flags, cachep->gfporder);
1377 if (!page) {
1378 slab_out_of_memory(cachep, flags, nodeid);
1379 return NULL;
1380 }
1381
1382 account_slab_page(page, cachep->gfporder, cachep);
1383 __SetPageSlab(page);
1384 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1385 if (sk_memalloc_socks() && page_is_pfmemalloc(page))
1386 SetPageSlabPfmemalloc(page);
1387
1388 return page;
1389}
1390
1391/*
1392 * Interface to system's page release.
1393 */
1394static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
1395{
1396 int order = cachep->gfporder;
1397
1398 BUG_ON(!PageSlab(page));
1399 __ClearPageSlabPfmemalloc(page);
1400 __ClearPageSlab(page);
1401 page_mapcount_reset(page);
1402 page->mapping = NULL;
1403
1404 if (current->reclaim_state)
1405 current->reclaim_state->reclaimed_slab += 1 << order;
1406 unaccount_slab_page(page, order, cachep);
1407 __free_pages(page, order);
1408}
1409
1410static void kmem_rcu_free(struct rcu_head *head)
1411{
1412 struct kmem_cache *cachep;
1413 struct page *page;
1414
1415 page = container_of(head, struct page, rcu_head);
1416 cachep = page->slab_cache;
1417
1418 kmem_freepages(cachep, page);
1419}
1420
1421#if DEBUG
1422static bool is_debug_pagealloc_cache(struct kmem_cache *cachep)
1423{
1424 if (debug_pagealloc_enabled_static() && OFF_SLAB(cachep) &&
1425 (cachep->size % PAGE_SIZE) == 0)
1426 return true;
1427
1428 return false;
1429}
1430
1431#ifdef CONFIG_DEBUG_PAGEALLOC
1432static void slab_kernel_map(struct kmem_cache *cachep, void *objp, int map)
1433{
1434 if (!is_debug_pagealloc_cache(cachep))
1435 return;
1436
1437 kernel_map_pages(virt_to_page(objp), cachep->size / PAGE_SIZE, map);
1438}
1439
1440#else
1441static inline void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1442 int map) {}
1443
1444#endif
1445
1446static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1447{
1448 int size = cachep->object_size;
1449 addr = &((char *)addr)[obj_offset(cachep)];
1450
1451 memset(addr, val, size);
1452 *(unsigned char *)(addr + size - 1) = POISON_END;
1453}
1454
1455static void dump_line(char *data, int offset, int limit)
1456{
1457 int i;
1458 unsigned char error = 0;
1459 int bad_count = 0;
1460
1461 pr_err("%03x: ", offset);
1462 for (i = 0; i < limit; i++) {
1463 if (data[offset + i] != POISON_FREE) {
1464 error = data[offset + i];
1465 bad_count++;
1466 }
1467 }
1468 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1469 &data[offset], limit, 1);
1470
1471 if (bad_count == 1) {
1472 error ^= POISON_FREE;
1473 if (!(error & (error - 1))) {
1474 pr_err("Single bit error detected. Probably bad RAM.\n");
1475#ifdef CONFIG_X86
1476 pr_err("Run memtest86+ or a similar memory test tool.\n");
1477#else
1478 pr_err("Run a memory test tool.\n");
1479#endif
1480 }
1481 }
1482}
1483#endif
1484
1485#if DEBUG
1486
1487static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1488{
1489 int i, size;
1490 char *realobj;
1491
1492 if (cachep->flags & SLAB_RED_ZONE) {
1493 pr_err("Redzone: 0x%llx/0x%llx\n",
1494 *dbg_redzone1(cachep, objp),
1495 *dbg_redzone2(cachep, objp));
1496 }
1497
1498 if (cachep->flags & SLAB_STORE_USER)
1499 pr_err("Last user: (%pSR)\n", *dbg_userword(cachep, objp));
1500 realobj = (char *)objp + obj_offset(cachep);
1501 size = cachep->object_size;
1502 for (i = 0; i < size && lines; i += 16, lines--) {
1503 int limit;
1504 limit = 16;
1505 if (i + limit > size)
1506 limit = size - i;
1507 dump_line(realobj, i, limit);
1508 }
1509}
1510
1511static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1512{
1513 char *realobj;
1514 int size, i;
1515 int lines = 0;
1516
1517 if (is_debug_pagealloc_cache(cachep))
1518 return;
1519
1520 realobj = (char *)objp + obj_offset(cachep);
1521 size = cachep->object_size;
1522
1523 for (i = 0; i < size; i++) {
1524 char exp = POISON_FREE;
1525 if (i == size - 1)
1526 exp = POISON_END;
1527 if (realobj[i] != exp) {
1528 int limit;
1529 /* Mismatch ! */
1530 /* Print header */
1531 if (lines == 0) {
1532 pr_err("Slab corruption (%s): %s start=%px, len=%d\n",
1533 print_tainted(), cachep->name,
1534 realobj, size);
1535 print_objinfo(cachep, objp, 0);
1536 }
1537 /* Hexdump the affected line */
1538 i = (i / 16) * 16;
1539 limit = 16;
1540 if (i + limit > size)
1541 limit = size - i;
1542 dump_line(realobj, i, limit);
1543 i += 16;
1544 lines++;
1545 /* Limit to 5 lines */
1546 if (lines > 5)
1547 break;
1548 }
1549 }
1550 if (lines != 0) {
1551 /* Print some data about the neighboring objects, if they
1552 * exist:
1553 */
1554 struct page *page = virt_to_head_page(objp);
1555 unsigned int objnr;
1556
1557 objnr = obj_to_index(cachep, page, objp);
1558 if (objnr) {
1559 objp = index_to_obj(cachep, page, objnr - 1);
1560 realobj = (char *)objp + obj_offset(cachep);
1561 pr_err("Prev obj: start=%px, len=%d\n", realobj, size);
1562 print_objinfo(cachep, objp, 2);
1563 }
1564 if (objnr + 1 < cachep->num) {
1565 objp = index_to_obj(cachep, page, objnr + 1);
1566 realobj = (char *)objp + obj_offset(cachep);
1567 pr_err("Next obj: start=%px, len=%d\n", realobj, size);
1568 print_objinfo(cachep, objp, 2);
1569 }
1570 }
1571}
1572#endif
1573
1574#if DEBUG
1575static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1576 struct page *page)
1577{
1578 int i;
1579
1580 if (OBJFREELIST_SLAB(cachep) && cachep->flags & SLAB_POISON) {
1581 poison_obj(cachep, page->freelist - obj_offset(cachep),
1582 POISON_FREE);
1583 }
1584
1585 for (i = 0; i < cachep->num; i++) {
1586 void *objp = index_to_obj(cachep, page, i);
1587
1588 if (cachep->flags & SLAB_POISON) {
1589 check_poison_obj(cachep, objp);
1590 slab_kernel_map(cachep, objp, 1);
1591 }
1592 if (cachep->flags & SLAB_RED_ZONE) {
1593 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1594 slab_error(cachep, "start of a freed object was overwritten");
1595 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1596 slab_error(cachep, "end of a freed object was overwritten");
1597 }
1598 }
1599}
1600#else
1601static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1602 struct page *page)
1603{
1604}
1605#endif
1606
1607/**
1608 * slab_destroy - destroy and release all objects in a slab
1609 * @cachep: cache pointer being destroyed
1610 * @page: page pointer being destroyed
1611 *
1612 * Destroy all the objs in a slab page, and release the mem back to the system.
1613 * Before calling the slab page must have been unlinked from the cache. The
1614 * kmem_cache_node ->list_lock is not held/needed.
1615 */
1616static void slab_destroy(struct kmem_cache *cachep, struct page *page)
1617{
1618 void *freelist;
1619
1620 freelist = page->freelist;
1621 slab_destroy_debugcheck(cachep, page);
1622 if (unlikely(cachep->flags & SLAB_TYPESAFE_BY_RCU))
1623 call_rcu(&page->rcu_head, kmem_rcu_free);
1624 else
1625 kmem_freepages(cachep, page);
1626
1627 /*
1628 * From now on, we don't use freelist
1629 * although actual page can be freed in rcu context
1630 */
1631 if (OFF_SLAB(cachep))
1632 kmem_cache_free(cachep->freelist_cache, freelist);
1633}
1634
1635/*
1636 * Update the size of the caches before calling slabs_destroy as it may
1637 * recursively call kfree.
1638 */
1639static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list)
1640{
1641 struct page *page, *n;
1642
1643 list_for_each_entry_safe(page, n, list, slab_list) {
1644 list_del(&page->slab_list);
1645 slab_destroy(cachep, page);
1646 }
1647}
1648
1649/**
1650 * calculate_slab_order - calculate size (page order) of slabs
1651 * @cachep: pointer to the cache that is being created
1652 * @size: size of objects to be created in this cache.
1653 * @flags: slab allocation flags
1654 *
1655 * Also calculates the number of objects per slab.
1656 *
1657 * This could be made much more intelligent. For now, try to avoid using
1658 * high order pages for slabs. When the gfp() functions are more friendly
1659 * towards high-order requests, this should be changed.
1660 *
1661 * Return: number of left-over bytes in a slab
1662 */
1663static size_t calculate_slab_order(struct kmem_cache *cachep,
1664 size_t size, slab_flags_t flags)
1665{
1666 size_t left_over = 0;
1667 int gfporder;
1668
1669 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
1670 unsigned int num;
1671 size_t remainder;
1672
1673 num = cache_estimate(gfporder, size, flags, &remainder);
1674 if (!num)
1675 continue;
1676
1677 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1678 if (num > SLAB_OBJ_MAX_NUM)
1679 break;
1680
1681 if (flags & CFLGS_OFF_SLAB) {
1682 struct kmem_cache *freelist_cache;
1683 size_t freelist_size;
1684
1685 freelist_size = num * sizeof(freelist_idx_t);
1686 freelist_cache = kmalloc_slab(freelist_size, 0u);
1687 if (!freelist_cache)
1688 continue;
1689
1690 /*
1691 * Needed to avoid possible looping condition
1692 * in cache_grow_begin()
1693 */
1694 if (OFF_SLAB(freelist_cache))
1695 continue;
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 void (*ctor)(void *))
1794{
1795 return flags;
1796}
1797
1798struct kmem_cache *
1799__kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
1800 slab_flags_t flags, void (*ctor)(void *))
1801{
1802 struct kmem_cache *cachep;
1803
1804 cachep = find_mergeable(size, align, flags, name, ctor);
1805 if (cachep) {
1806 cachep->refcount++;
1807
1808 /*
1809 * Adjust the object sizes so that we clear
1810 * the complete object on kzalloc.
1811 */
1812 cachep->object_size = max_t(int, cachep->object_size, size);
1813 }
1814 return cachep;
1815}
1816
1817static bool set_objfreelist_slab_cache(struct kmem_cache *cachep,
1818 size_t size, slab_flags_t flags)
1819{
1820 size_t left;
1821
1822 cachep->num = 0;
1823
1824 /*
1825 * If slab auto-initialization on free is enabled, store the freelist
1826 * off-slab, so that its contents don't end up in one of the allocated
1827 * objects.
1828 */
1829 if (unlikely(slab_want_init_on_free(cachep)))
1830 return false;
1831
1832 if (cachep->ctor || flags & SLAB_TYPESAFE_BY_RCU)
1833 return false;
1834
1835 left = calculate_slab_order(cachep, size,
1836 flags | CFLGS_OBJFREELIST_SLAB);
1837 if (!cachep->num)
1838 return false;
1839
1840 if (cachep->num * sizeof(freelist_idx_t) > cachep->object_size)
1841 return false;
1842
1843 cachep->colour = left / cachep->colour_off;
1844
1845 return true;
1846}
1847
1848static bool set_off_slab_cache(struct kmem_cache *cachep,
1849 size_t size, slab_flags_t flags)
1850{
1851 size_t left;
1852
1853 cachep->num = 0;
1854
1855 /*
1856 * Always use on-slab management when SLAB_NOLEAKTRACE
1857 * to avoid recursive calls into kmemleak.
1858 */
1859 if (flags & SLAB_NOLEAKTRACE)
1860 return false;
1861
1862 /*
1863 * Size is large, assume best to place the slab management obj
1864 * off-slab (should allow better packing of objs).
1865 */
1866 left = calculate_slab_order(cachep, size, flags | CFLGS_OFF_SLAB);
1867 if (!cachep->num)
1868 return false;
1869
1870 /*
1871 * If the slab has been placed off-slab, and we have enough space then
1872 * move it on-slab. This is at the expense of any extra colouring.
1873 */
1874 if (left >= cachep->num * sizeof(freelist_idx_t))
1875 return false;
1876
1877 cachep->colour = left / cachep->colour_off;
1878
1879 return true;
1880}
1881
1882static bool set_on_slab_cache(struct kmem_cache *cachep,
1883 size_t size, slab_flags_t flags)
1884{
1885 size_t left;
1886
1887 cachep->num = 0;
1888
1889 left = calculate_slab_order(cachep, size, flags);
1890 if (!cachep->num)
1891 return false;
1892
1893 cachep->colour = left / cachep->colour_off;
1894
1895 return true;
1896}
1897
1898/**
1899 * __kmem_cache_create - Create a cache.
1900 * @cachep: cache management descriptor
1901 * @flags: SLAB flags
1902 *
1903 * Returns a ptr to the cache on success, NULL on failure.
1904 * Cannot be called within a int, but can be interrupted.
1905 * The @ctor is run when new pages are allocated by the cache.
1906 *
1907 * The flags are
1908 *
1909 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1910 * to catch references to uninitialised memory.
1911 *
1912 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1913 * for buffer overruns.
1914 *
1915 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1916 * cacheline. This can be beneficial if you're counting cycles as closely
1917 * as davem.
1918 *
1919 * Return: a pointer to the created cache or %NULL in case of error
1920 */
1921int __kmem_cache_create(struct kmem_cache *cachep, slab_flags_t flags)
1922{
1923 size_t ralign = BYTES_PER_WORD;
1924 gfp_t gfp;
1925 int err;
1926 unsigned int size = cachep->size;
1927
1928#if DEBUG
1929#if FORCED_DEBUG
1930 /*
1931 * Enable redzoning and last user accounting, except for caches with
1932 * large objects, if the increased size would increase the object size
1933 * above the next power of two: caches with object sizes just above a
1934 * power of two have a significant amount of internal fragmentation.
1935 */
1936 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
1937 2 * sizeof(unsigned long long)))
1938 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
1939 if (!(flags & SLAB_TYPESAFE_BY_RCU))
1940 flags |= SLAB_POISON;
1941#endif
1942#endif
1943
1944 /*
1945 * Check that size is in terms of words. This is needed to avoid
1946 * unaligned accesses for some archs when redzoning is used, and makes
1947 * sure any on-slab bufctl's are also correctly aligned.
1948 */
1949 size = ALIGN(size, BYTES_PER_WORD);
1950
1951 if (flags & SLAB_RED_ZONE) {
1952 ralign = REDZONE_ALIGN;
1953 /* If redzoning, ensure that the second redzone is suitably
1954 * aligned, by adjusting the object size accordingly. */
1955 size = ALIGN(size, REDZONE_ALIGN);
1956 }
1957
1958 /* 3) caller mandated alignment */
1959 if (ralign < cachep->align) {
1960 ralign = cachep->align;
1961 }
1962 /* disable debug if necessary */
1963 if (ralign > __alignof__(unsigned long long))
1964 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
1965 /*
1966 * 4) Store it.
1967 */
1968 cachep->align = ralign;
1969 cachep->colour_off = cache_line_size();
1970 /* Offset must be a multiple of the alignment. */
1971 if (cachep->colour_off < cachep->align)
1972 cachep->colour_off = cachep->align;
1973
1974 if (slab_is_available())
1975 gfp = GFP_KERNEL;
1976 else
1977 gfp = GFP_NOWAIT;
1978
1979#if DEBUG
1980
1981 /*
1982 * Both debugging options require word-alignment which is calculated
1983 * into align above.
1984 */
1985 if (flags & SLAB_RED_ZONE) {
1986 /* add space for red zone words */
1987 cachep->obj_offset += sizeof(unsigned long long);
1988 size += 2 * sizeof(unsigned long long);
1989 }
1990 if (flags & SLAB_STORE_USER) {
1991 /* user store requires one word storage behind the end of
1992 * the real object. But if the second red zone needs to be
1993 * aligned to 64 bits, we must allow that much space.
1994 */
1995 if (flags & SLAB_RED_ZONE)
1996 size += REDZONE_ALIGN;
1997 else
1998 size += BYTES_PER_WORD;
1999 }
2000#endif
2001
2002 kasan_cache_create(cachep, &size, &flags);
2003
2004 size = ALIGN(size, cachep->align);
2005 /*
2006 * We should restrict the number of objects in a slab to implement
2007 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2008 */
2009 if (FREELIST_BYTE_INDEX && size < SLAB_OBJ_MIN_SIZE)
2010 size = ALIGN(SLAB_OBJ_MIN_SIZE, cachep->align);
2011
2012#if DEBUG
2013 /*
2014 * To activate debug pagealloc, off-slab management is necessary
2015 * requirement. In early phase of initialization, small sized slab
2016 * doesn't get initialized so it would not be possible. So, we need
2017 * to check size >= 256. It guarantees that all necessary small
2018 * sized slab is initialized in current slab initialization sequence.
2019 */
2020 if (debug_pagealloc_enabled_static() && (flags & SLAB_POISON) &&
2021 size >= 256 && cachep->object_size > cache_line_size()) {
2022 if (size < PAGE_SIZE || size % PAGE_SIZE == 0) {
2023 size_t tmp_size = ALIGN(size, PAGE_SIZE);
2024
2025 if (set_off_slab_cache(cachep, tmp_size, flags)) {
2026 flags |= CFLGS_OFF_SLAB;
2027 cachep->obj_offset += tmp_size - size;
2028 size = tmp_size;
2029 goto done;
2030 }
2031 }
2032 }
2033#endif
2034
2035 if (set_objfreelist_slab_cache(cachep, size, flags)) {
2036 flags |= CFLGS_OBJFREELIST_SLAB;
2037 goto done;
2038 }
2039
2040 if (set_off_slab_cache(cachep, size, flags)) {
2041 flags |= CFLGS_OFF_SLAB;
2042 goto done;
2043 }
2044
2045 if (set_on_slab_cache(cachep, size, flags))
2046 goto done;
2047
2048 return -E2BIG;
2049
2050done:
2051 cachep->freelist_size = cachep->num * sizeof(freelist_idx_t);
2052 cachep->flags = flags;
2053 cachep->allocflags = __GFP_COMP;
2054 if (flags & SLAB_CACHE_DMA)
2055 cachep->allocflags |= GFP_DMA;
2056 if (flags & SLAB_CACHE_DMA32)
2057 cachep->allocflags |= GFP_DMA32;
2058 if (flags & SLAB_RECLAIM_ACCOUNT)
2059 cachep->allocflags |= __GFP_RECLAIMABLE;
2060 cachep->size = size;
2061 cachep->reciprocal_buffer_size = reciprocal_value(size);
2062
2063#if DEBUG
2064 /*
2065 * If we're going to use the generic kernel_map_pages()
2066 * poisoning, then it's going to smash the contents of
2067 * the redzone and userword anyhow, so switch them off.
2068 */
2069 if (IS_ENABLED(CONFIG_PAGE_POISONING) &&
2070 (cachep->flags & SLAB_POISON) &&
2071 is_debug_pagealloc_cache(cachep))
2072 cachep->flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2073#endif
2074
2075 if (OFF_SLAB(cachep)) {
2076 cachep->freelist_cache =
2077 kmalloc_slab(cachep->freelist_size, 0u);
2078 }
2079
2080 err = setup_cpu_cache(cachep, gfp);
2081 if (err) {
2082 __kmem_cache_release(cachep);
2083 return err;
2084 }
2085
2086 return 0;
2087}
2088
2089#if DEBUG
2090static void check_irq_off(void)
2091{
2092 BUG_ON(!irqs_disabled());
2093}
2094
2095static void check_irq_on(void)
2096{
2097 BUG_ON(irqs_disabled());
2098}
2099
2100static void check_mutex_acquired(void)
2101{
2102 BUG_ON(!mutex_is_locked(&slab_mutex));
2103}
2104
2105static void check_spinlock_acquired(struct kmem_cache *cachep)
2106{
2107#ifdef CONFIG_SMP
2108 check_irq_off();
2109 assert_spin_locked(&get_node(cachep, numa_mem_id())->list_lock);
2110#endif
2111}
2112
2113static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2114{
2115#ifdef CONFIG_SMP
2116 check_irq_off();
2117 assert_spin_locked(&get_node(cachep, node)->list_lock);
2118#endif
2119}
2120
2121#else
2122#define check_irq_off() do { } while(0)
2123#define check_irq_on() do { } while(0)
2124#define check_mutex_acquired() do { } while(0)
2125#define check_spinlock_acquired(x) do { } while(0)
2126#define check_spinlock_acquired_node(x, y) do { } while(0)
2127#endif
2128
2129static void drain_array_locked(struct kmem_cache *cachep, struct array_cache *ac,
2130 int node, bool free_all, struct list_head *list)
2131{
2132 int tofree;
2133
2134 if (!ac || !ac->avail)
2135 return;
2136
2137 tofree = free_all ? ac->avail : (ac->limit + 4) / 5;
2138 if (tofree > ac->avail)
2139 tofree = (ac->avail + 1) / 2;
2140
2141 free_block(cachep, ac->entry, tofree, node, list);
2142 ac->avail -= tofree;
2143 memmove(ac->entry, &(ac->entry[tofree]), sizeof(void *) * ac->avail);
2144}
2145
2146static void do_drain(void *arg)
2147{
2148 struct kmem_cache *cachep = arg;
2149 struct array_cache *ac;
2150 int node = numa_mem_id();
2151 struct kmem_cache_node *n;
2152 LIST_HEAD(list);
2153
2154 check_irq_off();
2155 ac = cpu_cache_get(cachep);
2156 n = get_node(cachep, node);
2157 spin_lock(&n->list_lock);
2158 free_block(cachep, ac->entry, ac->avail, node, &list);
2159 spin_unlock(&n->list_lock);
2160 ac->avail = 0;
2161 slabs_destroy(cachep, &list);
2162}
2163
2164static void drain_cpu_caches(struct kmem_cache *cachep)
2165{
2166 struct kmem_cache_node *n;
2167 int node;
2168 LIST_HEAD(list);
2169
2170 on_each_cpu(do_drain, cachep, 1);
2171 check_irq_on();
2172 for_each_kmem_cache_node(cachep, node, n)
2173 if (n->alien)
2174 drain_alien_cache(cachep, n->alien);
2175
2176 for_each_kmem_cache_node(cachep, node, n) {
2177 spin_lock_irq(&n->list_lock);
2178 drain_array_locked(cachep, n->shared, node, true, &list);
2179 spin_unlock_irq(&n->list_lock);
2180
2181 slabs_destroy(cachep, &list);
2182 }
2183}
2184
2185/*
2186 * Remove slabs from the list of free slabs.
2187 * Specify the number of slabs to drain in tofree.
2188 *
2189 * Returns the actual number of slabs released.
2190 */
2191static int drain_freelist(struct kmem_cache *cache,
2192 struct kmem_cache_node *n, int tofree)
2193{
2194 struct list_head *p;
2195 int nr_freed;
2196 struct page *page;
2197
2198 nr_freed = 0;
2199 while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
2200
2201 spin_lock_irq(&n->list_lock);
2202 p = n->slabs_free.prev;
2203 if (p == &n->slabs_free) {
2204 spin_unlock_irq(&n->list_lock);
2205 goto out;
2206 }
2207
2208 page = list_entry(p, struct page, slab_list);
2209 list_del(&page->slab_list);
2210 n->free_slabs--;
2211 n->total_slabs--;
2212 /*
2213 * Safe to drop the lock. The slab is no longer linked
2214 * to the cache.
2215 */
2216 n->free_objects -= cache->num;
2217 spin_unlock_irq(&n->list_lock);
2218 slab_destroy(cache, page);
2219 nr_freed++;
2220 }
2221out:
2222 return nr_freed;
2223}
2224
2225bool __kmem_cache_empty(struct kmem_cache *s)
2226{
2227 int node;
2228 struct kmem_cache_node *n;
2229
2230 for_each_kmem_cache_node(s, node, n)
2231 if (!list_empty(&n->slabs_full) ||
2232 !list_empty(&n->slabs_partial))
2233 return false;
2234 return true;
2235}
2236
2237int __kmem_cache_shrink(struct kmem_cache *cachep)
2238{
2239 int ret = 0;
2240 int node;
2241 struct kmem_cache_node *n;
2242
2243 drain_cpu_caches(cachep);
2244
2245 check_irq_on();
2246 for_each_kmem_cache_node(cachep, node, n) {
2247 drain_freelist(cachep, n, INT_MAX);
2248
2249 ret += !list_empty(&n->slabs_full) ||
2250 !list_empty(&n->slabs_partial);
2251 }
2252 return (ret ? 1 : 0);
2253}
2254
2255int __kmem_cache_shutdown(struct kmem_cache *cachep)
2256{
2257 return __kmem_cache_shrink(cachep);
2258}
2259
2260void __kmem_cache_release(struct kmem_cache *cachep)
2261{
2262 int i;
2263 struct kmem_cache_node *n;
2264
2265 cache_random_seq_destroy(cachep);
2266
2267 free_percpu(cachep->cpu_cache);
2268
2269 /* NUMA: free the node structures */
2270 for_each_kmem_cache_node(cachep, i, n) {
2271 kfree(n->shared);
2272 free_alien_cache(n->alien);
2273 kfree(n);
2274 cachep->node[i] = NULL;
2275 }
2276}
2277
2278/*
2279 * Get the memory for a slab management obj.
2280 *
2281 * For a slab cache when the slab descriptor is off-slab, the
2282 * slab descriptor can't come from the same cache which is being created,
2283 * Because if it is the case, that means we defer the creation of
2284 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2285 * And we eventually call down to __kmem_cache_create(), which
2286 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2287 * This is a "chicken-and-egg" problem.
2288 *
2289 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2290 * which are all initialized during kmem_cache_init().
2291 */
2292static void *alloc_slabmgmt(struct kmem_cache *cachep,
2293 struct page *page, int colour_off,
2294 gfp_t local_flags, int nodeid)
2295{
2296 void *freelist;
2297 void *addr = page_address(page);
2298
2299 page->s_mem = addr + colour_off;
2300 page->active = 0;
2301
2302 if (OBJFREELIST_SLAB(cachep))
2303 freelist = NULL;
2304 else if (OFF_SLAB(cachep)) {
2305 /* Slab management obj is off-slab. */
2306 freelist = kmem_cache_alloc_node(cachep->freelist_cache,
2307 local_flags, nodeid);
2308 if (!freelist)
2309 return NULL;
2310 } else {
2311 /* We will use last bytes at the slab for freelist */
2312 freelist = addr + (PAGE_SIZE << cachep->gfporder) -
2313 cachep->freelist_size;
2314 }
2315
2316 return freelist;
2317}
2318
2319static inline freelist_idx_t get_free_obj(struct page *page, unsigned int idx)
2320{
2321 return ((freelist_idx_t *)page->freelist)[idx];
2322}
2323
2324static inline void set_free_obj(struct page *page,
2325 unsigned int idx, freelist_idx_t val)
2326{
2327 ((freelist_idx_t *)(page->freelist))[idx] = val;
2328}
2329
2330static void cache_init_objs_debug(struct kmem_cache *cachep, struct page *page)
2331{
2332#if DEBUG
2333 int i;
2334
2335 for (i = 0; i < cachep->num; i++) {
2336 void *objp = index_to_obj(cachep, page, i);
2337
2338 if (cachep->flags & SLAB_STORE_USER)
2339 *dbg_userword(cachep, objp) = NULL;
2340
2341 if (cachep->flags & SLAB_RED_ZONE) {
2342 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2343 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2344 }
2345 /*
2346 * Constructors are not allowed to allocate memory from the same
2347 * cache which they are a constructor for. Otherwise, deadlock.
2348 * They must also be threaded.
2349 */
2350 if (cachep->ctor && !(cachep->flags & SLAB_POISON)) {
2351 kasan_unpoison_object_data(cachep,
2352 objp + obj_offset(cachep));
2353 cachep->ctor(objp + obj_offset(cachep));
2354 kasan_poison_object_data(
2355 cachep, objp + obj_offset(cachep));
2356 }
2357
2358 if (cachep->flags & SLAB_RED_ZONE) {
2359 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2360 slab_error(cachep, "constructor overwrote the end of an object");
2361 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2362 slab_error(cachep, "constructor overwrote the start of an object");
2363 }
2364 /* need to poison the objs? */
2365 if (cachep->flags & SLAB_POISON) {
2366 poison_obj(cachep, objp, POISON_FREE);
2367 slab_kernel_map(cachep, objp, 0);
2368 }
2369 }
2370#endif
2371}
2372
2373#ifdef CONFIG_SLAB_FREELIST_RANDOM
2374/* Hold information during a freelist initialization */
2375union freelist_init_state {
2376 struct {
2377 unsigned int pos;
2378 unsigned int *list;
2379 unsigned int count;
2380 };
2381 struct rnd_state rnd_state;
2382};
2383
2384/*
2385 * Initialize the state based on the randomization methode available.
2386 * return true if the pre-computed list is available, false otherwize.
2387 */
2388static bool freelist_state_initialize(union freelist_init_state *state,
2389 struct kmem_cache *cachep,
2390 unsigned int count)
2391{
2392 bool ret;
2393 unsigned int rand;
2394
2395 /* Use best entropy available to define a random shift */
2396 rand = get_random_int();
2397
2398 /* Use a random state if the pre-computed list is not available */
2399 if (!cachep->random_seq) {
2400 prandom_seed_state(&state->rnd_state, rand);
2401 ret = false;
2402 } else {
2403 state->list = cachep->random_seq;
2404 state->count = count;
2405 state->pos = rand % count;
2406 ret = true;
2407 }
2408 return ret;
2409}
2410
2411/* Get the next entry on the list and randomize it using a random shift */
2412static freelist_idx_t next_random_slot(union freelist_init_state *state)
2413{
2414 if (state->pos >= state->count)
2415 state->pos = 0;
2416 return state->list[state->pos++];
2417}
2418
2419/* Swap two freelist entries */
2420static void swap_free_obj(struct page *page, unsigned int a, unsigned int b)
2421{
2422 swap(((freelist_idx_t *)page->freelist)[a],
2423 ((freelist_idx_t *)page->freelist)[b]);
2424}
2425
2426/*
2427 * Shuffle the freelist initialization state based on pre-computed lists.
2428 * return true if the list was successfully shuffled, false otherwise.
2429 */
2430static bool shuffle_freelist(struct kmem_cache *cachep, struct page *page)
2431{
2432 unsigned int objfreelist = 0, i, rand, count = cachep->num;
2433 union freelist_init_state state;
2434 bool precomputed;
2435
2436 if (count < 2)
2437 return false;
2438
2439 precomputed = freelist_state_initialize(&state, cachep, count);
2440
2441 /* Take a random entry as the objfreelist */
2442 if (OBJFREELIST_SLAB(cachep)) {
2443 if (!precomputed)
2444 objfreelist = count - 1;
2445 else
2446 objfreelist = next_random_slot(&state);
2447 page->freelist = index_to_obj(cachep, page, objfreelist) +
2448 obj_offset(cachep);
2449 count--;
2450 }
2451
2452 /*
2453 * On early boot, generate the list dynamically.
2454 * Later use a pre-computed list for speed.
2455 */
2456 if (!precomputed) {
2457 for (i = 0; i < count; i++)
2458 set_free_obj(page, i, i);
2459
2460 /* Fisher-Yates shuffle */
2461 for (i = count - 1; i > 0; i--) {
2462 rand = prandom_u32_state(&state.rnd_state);
2463 rand %= (i + 1);
2464 swap_free_obj(page, i, rand);
2465 }
2466 } else {
2467 for (i = 0; i < count; i++)
2468 set_free_obj(page, i, next_random_slot(&state));
2469 }
2470
2471 if (OBJFREELIST_SLAB(cachep))
2472 set_free_obj(page, cachep->num - 1, objfreelist);
2473
2474 return true;
2475}
2476#else
2477static inline bool shuffle_freelist(struct kmem_cache *cachep,
2478 struct page *page)
2479{
2480 return false;
2481}
2482#endif /* CONFIG_SLAB_FREELIST_RANDOM */
2483
2484static void cache_init_objs(struct kmem_cache *cachep,
2485 struct page *page)
2486{
2487 int i;
2488 void *objp;
2489 bool shuffled;
2490
2491 cache_init_objs_debug(cachep, page);
2492
2493 /* Try to randomize the freelist if enabled */
2494 shuffled = shuffle_freelist(cachep, page);
2495
2496 if (!shuffled && OBJFREELIST_SLAB(cachep)) {
2497 page->freelist = index_to_obj(cachep, page, cachep->num - 1) +
2498 obj_offset(cachep);
2499 }
2500
2501 for (i = 0; i < cachep->num; i++) {
2502 objp = index_to_obj(cachep, page, i);
2503 objp = kasan_init_slab_obj(cachep, objp);
2504
2505 /* constructor could break poison info */
2506 if (DEBUG == 0 && cachep->ctor) {
2507 kasan_unpoison_object_data(cachep, objp);
2508 cachep->ctor(objp);
2509 kasan_poison_object_data(cachep, objp);
2510 }
2511
2512 if (!shuffled)
2513 set_free_obj(page, i, i);
2514 }
2515}
2516
2517static void *slab_get_obj(struct kmem_cache *cachep, struct page *page)
2518{
2519 void *objp;
2520
2521 objp = index_to_obj(cachep, page, get_free_obj(page, page->active));
2522 page->active++;
2523
2524 return objp;
2525}
2526
2527static void slab_put_obj(struct kmem_cache *cachep,
2528 struct page *page, void *objp)
2529{
2530 unsigned int objnr = obj_to_index(cachep, page, objp);
2531#if DEBUG
2532 unsigned int i;
2533
2534 /* Verify double free bug */
2535 for (i = page->active; i < cachep->num; i++) {
2536 if (get_free_obj(page, i) == objnr) {
2537 pr_err("slab: double free detected in cache '%s', objp %px\n",
2538 cachep->name, objp);
2539 BUG();
2540 }
2541 }
2542#endif
2543 page->active--;
2544 if (!page->freelist)
2545 page->freelist = objp + obj_offset(cachep);
2546
2547 set_free_obj(page, page->active, objnr);
2548}
2549
2550/*
2551 * Map pages beginning at addr to the given cache and slab. This is required
2552 * for the slab allocator to be able to lookup the cache and slab of a
2553 * virtual address for kfree, ksize, and slab debugging.
2554 */
2555static void slab_map_pages(struct kmem_cache *cache, struct page *page,
2556 void *freelist)
2557{
2558 page->slab_cache = cache;
2559 page->freelist = freelist;
2560}
2561
2562/*
2563 * Grow (by 1) the number of slabs within a cache. This is called by
2564 * kmem_cache_alloc() when there are no active objs left in a cache.
2565 */
2566static struct page *cache_grow_begin(struct kmem_cache *cachep,
2567 gfp_t flags, int nodeid)
2568{
2569 void *freelist;
2570 size_t offset;
2571 gfp_t local_flags;
2572 int page_node;
2573 struct kmem_cache_node *n;
2574 struct page *page;
2575
2576 /*
2577 * Be lazy and only check for valid flags here, keeping it out of the
2578 * critical path in kmem_cache_alloc().
2579 */
2580 if (unlikely(flags & GFP_SLAB_BUG_MASK))
2581 flags = kmalloc_fix_flags(flags);
2582
2583 WARN_ON_ONCE(cachep->ctor && (flags & __GFP_ZERO));
2584 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2585
2586 check_irq_off();
2587 if (gfpflags_allow_blocking(local_flags))
2588 local_irq_enable();
2589
2590 /*
2591 * Get mem for the objs. Attempt to allocate a physical page from
2592 * 'nodeid'.
2593 */
2594 page = kmem_getpages(cachep, local_flags, nodeid);
2595 if (!page)
2596 goto failed;
2597
2598 page_node = page_to_nid(page);
2599 n = get_node(cachep, page_node);
2600
2601 /* Get colour for the slab, and cal the next value. */
2602 n->colour_next++;
2603 if (n->colour_next >= cachep->colour)
2604 n->colour_next = 0;
2605
2606 offset = n->colour_next;
2607 if (offset >= cachep->colour)
2608 offset = 0;
2609
2610 offset *= cachep->colour_off;
2611
2612 /*
2613 * Call kasan_poison_slab() before calling alloc_slabmgmt(), so
2614 * page_address() in the latter returns a non-tagged pointer,
2615 * as it should be for slab pages.
2616 */
2617 kasan_poison_slab(page);
2618
2619 /* Get slab management. */
2620 freelist = alloc_slabmgmt(cachep, page, offset,
2621 local_flags & ~GFP_CONSTRAINT_MASK, page_node);
2622 if (OFF_SLAB(cachep) && !freelist)
2623 goto opps1;
2624
2625 slab_map_pages(cachep, page, freelist);
2626
2627 cache_init_objs(cachep, page);
2628
2629 if (gfpflags_allow_blocking(local_flags))
2630 local_irq_disable();
2631
2632 return page;
2633
2634opps1:
2635 kmem_freepages(cachep, page);
2636failed:
2637 if (gfpflags_allow_blocking(local_flags))
2638 local_irq_disable();
2639 return NULL;
2640}
2641
2642static void cache_grow_end(struct kmem_cache *cachep, struct page *page)
2643{
2644 struct kmem_cache_node *n;
2645 void *list = NULL;
2646
2647 check_irq_off();
2648
2649 if (!page)
2650 return;
2651
2652 INIT_LIST_HEAD(&page->slab_list);
2653 n = get_node(cachep, page_to_nid(page));
2654
2655 spin_lock(&n->list_lock);
2656 n->total_slabs++;
2657 if (!page->active) {
2658 list_add_tail(&page->slab_list, &n->slabs_free);
2659 n->free_slabs++;
2660 } else
2661 fixup_slab_list(cachep, n, page, &list);
2662
2663 STATS_INC_GROWN(cachep);
2664 n->free_objects += cachep->num - page->active;
2665 spin_unlock(&n->list_lock);
2666
2667 fixup_objfreelist_debug(cachep, &list);
2668}
2669
2670#if DEBUG
2671
2672/*
2673 * Perform extra freeing checks:
2674 * - detect bad pointers.
2675 * - POISON/RED_ZONE checking
2676 */
2677static void kfree_debugcheck(const void *objp)
2678{
2679 if (!virt_addr_valid(objp)) {
2680 pr_err("kfree_debugcheck: out of range ptr %lxh\n",
2681 (unsigned long)objp);
2682 BUG();
2683 }
2684}
2685
2686static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2687{
2688 unsigned long long redzone1, redzone2;
2689
2690 redzone1 = *dbg_redzone1(cache, obj);
2691 redzone2 = *dbg_redzone2(cache, obj);
2692
2693 /*
2694 * Redzone is ok.
2695 */
2696 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2697 return;
2698
2699 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2700 slab_error(cache, "double free detected");
2701 else
2702 slab_error(cache, "memory outside object was overwritten");
2703
2704 pr_err("%px: redzone 1:0x%llx, redzone 2:0x%llx\n",
2705 obj, redzone1, redzone2);
2706}
2707
2708static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2709 unsigned long caller)
2710{
2711 unsigned int objnr;
2712 struct page *page;
2713
2714 BUG_ON(virt_to_cache(objp) != cachep);
2715
2716 objp -= obj_offset(cachep);
2717 kfree_debugcheck(objp);
2718 page = virt_to_head_page(objp);
2719
2720 if (cachep->flags & SLAB_RED_ZONE) {
2721 verify_redzone_free(cachep, objp);
2722 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2723 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2724 }
2725 if (cachep->flags & SLAB_STORE_USER)
2726 *dbg_userword(cachep, objp) = (void *)caller;
2727
2728 objnr = obj_to_index(cachep, page, objp);
2729
2730 BUG_ON(objnr >= cachep->num);
2731 BUG_ON(objp != index_to_obj(cachep, page, objnr));
2732
2733 if (cachep->flags & SLAB_POISON) {
2734 poison_obj(cachep, objp, POISON_FREE);
2735 slab_kernel_map(cachep, objp, 0);
2736 }
2737 return objp;
2738}
2739
2740#else
2741#define kfree_debugcheck(x) do { } while(0)
2742#define cache_free_debugcheck(x,objp,z) (objp)
2743#endif
2744
2745static inline void fixup_objfreelist_debug(struct kmem_cache *cachep,
2746 void **list)
2747{
2748#if DEBUG
2749 void *next = *list;
2750 void *objp;
2751
2752 while (next) {
2753 objp = next - obj_offset(cachep);
2754 next = *(void **)next;
2755 poison_obj(cachep, objp, POISON_FREE);
2756 }
2757#endif
2758}
2759
2760static inline void fixup_slab_list(struct kmem_cache *cachep,
2761 struct kmem_cache_node *n, struct page *page,
2762 void **list)
2763{
2764 /* move slabp to correct slabp list: */
2765 list_del(&page->slab_list);
2766 if (page->active == cachep->num) {
2767 list_add(&page->slab_list, &n->slabs_full);
2768 if (OBJFREELIST_SLAB(cachep)) {
2769#if DEBUG
2770 /* Poisoning will be done without holding the lock */
2771 if (cachep->flags & SLAB_POISON) {
2772 void **objp = page->freelist;
2773
2774 *objp = *list;
2775 *list = objp;
2776 }
2777#endif
2778 page->freelist = NULL;
2779 }
2780 } else
2781 list_add(&page->slab_list, &n->slabs_partial);
2782}
2783
2784/* Try to find non-pfmemalloc slab if needed */
2785static noinline struct page *get_valid_first_slab(struct kmem_cache_node *n,
2786 struct page *page, bool pfmemalloc)
2787{
2788 if (!page)
2789 return NULL;
2790
2791 if (pfmemalloc)
2792 return page;
2793
2794 if (!PageSlabPfmemalloc(page))
2795 return page;
2796
2797 /* No need to keep pfmemalloc slab if we have enough free objects */
2798 if (n->free_objects > n->free_limit) {
2799 ClearPageSlabPfmemalloc(page);
2800 return page;
2801 }
2802
2803 /* Move pfmemalloc slab to the end of list to speed up next search */
2804 list_del(&page->slab_list);
2805 if (!page->active) {
2806 list_add_tail(&page->slab_list, &n->slabs_free);
2807 n->free_slabs++;
2808 } else
2809 list_add_tail(&page->slab_list, &n->slabs_partial);
2810
2811 list_for_each_entry(page, &n->slabs_partial, slab_list) {
2812 if (!PageSlabPfmemalloc(page))
2813 return page;
2814 }
2815
2816 n->free_touched = 1;
2817 list_for_each_entry(page, &n->slabs_free, slab_list) {
2818 if (!PageSlabPfmemalloc(page)) {
2819 n->free_slabs--;
2820 return page;
2821 }
2822 }
2823
2824 return NULL;
2825}
2826
2827static struct page *get_first_slab(struct kmem_cache_node *n, bool pfmemalloc)
2828{
2829 struct page *page;
2830
2831 assert_spin_locked(&n->list_lock);
2832 page = list_first_entry_or_null(&n->slabs_partial, struct page,
2833 slab_list);
2834 if (!page) {
2835 n->free_touched = 1;
2836 page = list_first_entry_or_null(&n->slabs_free, struct page,
2837 slab_list);
2838 if (page)
2839 n->free_slabs--;
2840 }
2841
2842 if (sk_memalloc_socks())
2843 page = get_valid_first_slab(n, page, pfmemalloc);
2844
2845 return page;
2846}
2847
2848static noinline void *cache_alloc_pfmemalloc(struct kmem_cache *cachep,
2849 struct kmem_cache_node *n, gfp_t flags)
2850{
2851 struct page *page;
2852 void *obj;
2853 void *list = NULL;
2854
2855 if (!gfp_pfmemalloc_allowed(flags))
2856 return NULL;
2857
2858 spin_lock(&n->list_lock);
2859 page = get_first_slab(n, true);
2860 if (!page) {
2861 spin_unlock(&n->list_lock);
2862 return NULL;
2863 }
2864
2865 obj = slab_get_obj(cachep, page);
2866 n->free_objects--;
2867
2868 fixup_slab_list(cachep, n, page, &list);
2869
2870 spin_unlock(&n->list_lock);
2871 fixup_objfreelist_debug(cachep, &list);
2872
2873 return obj;
2874}
2875
2876/*
2877 * Slab list should be fixed up by fixup_slab_list() for existing slab
2878 * or cache_grow_end() for new slab
2879 */
2880static __always_inline int alloc_block(struct kmem_cache *cachep,
2881 struct array_cache *ac, struct page *page, int batchcount)
2882{
2883 /*
2884 * There must be at least one object available for
2885 * allocation.
2886 */
2887 BUG_ON(page->active >= cachep->num);
2888
2889 while (page->active < cachep->num && batchcount--) {
2890 STATS_INC_ALLOCED(cachep);
2891 STATS_INC_ACTIVE(cachep);
2892 STATS_SET_HIGH(cachep);
2893
2894 ac->entry[ac->avail++] = slab_get_obj(cachep, page);
2895 }
2896
2897 return batchcount;
2898}
2899
2900static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
2901{
2902 int batchcount;
2903 struct kmem_cache_node *n;
2904 struct array_cache *ac, *shared;
2905 int node;
2906 void *list = NULL;
2907 struct page *page;
2908
2909 check_irq_off();
2910 node = numa_mem_id();
2911
2912 ac = cpu_cache_get(cachep);
2913 batchcount = ac->batchcount;
2914 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2915 /*
2916 * If there was little recent activity on this cache, then
2917 * perform only a partial refill. Otherwise we could generate
2918 * refill bouncing.
2919 */
2920 batchcount = BATCHREFILL_LIMIT;
2921 }
2922 n = get_node(cachep, node);
2923
2924 BUG_ON(ac->avail > 0 || !n);
2925 shared = READ_ONCE(n->shared);
2926 if (!n->free_objects && (!shared || !shared->avail))
2927 goto direct_grow;
2928
2929 spin_lock(&n->list_lock);
2930 shared = READ_ONCE(n->shared);
2931
2932 /* See if we can refill from the shared array */
2933 if (shared && transfer_objects(ac, shared, batchcount)) {
2934 shared->touched = 1;
2935 goto alloc_done;
2936 }
2937
2938 while (batchcount > 0) {
2939 /* Get slab alloc is to come from. */
2940 page = get_first_slab(n, false);
2941 if (!page)
2942 goto must_grow;
2943
2944 check_spinlock_acquired(cachep);
2945
2946 batchcount = alloc_block(cachep, ac, page, batchcount);
2947 fixup_slab_list(cachep, n, page, &list);
2948 }
2949
2950must_grow:
2951 n->free_objects -= ac->avail;
2952alloc_done:
2953 spin_unlock(&n->list_lock);
2954 fixup_objfreelist_debug(cachep, &list);
2955
2956direct_grow:
2957 if (unlikely(!ac->avail)) {
2958 /* Check if we can use obj in pfmemalloc slab */
2959 if (sk_memalloc_socks()) {
2960 void *obj = cache_alloc_pfmemalloc(cachep, n, flags);
2961
2962 if (obj)
2963 return obj;
2964 }
2965
2966 page = cache_grow_begin(cachep, gfp_exact_node(flags), node);
2967
2968 /*
2969 * cache_grow_begin() can reenable interrupts,
2970 * then ac could change.
2971 */
2972 ac = cpu_cache_get(cachep);
2973 if (!ac->avail && page)
2974 alloc_block(cachep, ac, page, batchcount);
2975 cache_grow_end(cachep, page);
2976
2977 if (!ac->avail)
2978 return NULL;
2979 }
2980 ac->touched = 1;
2981
2982 return ac->entry[--ac->avail];
2983}
2984
2985static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2986 gfp_t flags)
2987{
2988 might_sleep_if(gfpflags_allow_blocking(flags));
2989}
2990
2991#if DEBUG
2992static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
2993 gfp_t flags, void *objp, unsigned long caller)
2994{
2995 WARN_ON_ONCE(cachep->ctor && (flags & __GFP_ZERO));
2996 if (!objp)
2997 return objp;
2998 if (cachep->flags & SLAB_POISON) {
2999 check_poison_obj(cachep, objp);
3000 slab_kernel_map(cachep, objp, 1);
3001 poison_obj(cachep, objp, POISON_INUSE);
3002 }
3003 if (cachep->flags & SLAB_STORE_USER)
3004 *dbg_userword(cachep, objp) = (void *)caller;
3005
3006 if (cachep->flags & SLAB_RED_ZONE) {
3007 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3008 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3009 slab_error(cachep, "double free, or memory outside object was overwritten");
3010 pr_err("%px: redzone 1:0x%llx, redzone 2:0x%llx\n",
3011 objp, *dbg_redzone1(cachep, objp),
3012 *dbg_redzone2(cachep, objp));
3013 }
3014 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3015 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3016 }
3017
3018 objp += obj_offset(cachep);
3019 if (cachep->ctor && cachep->flags & SLAB_POISON)
3020 cachep->ctor(objp);
3021 if (ARCH_SLAB_MINALIGN &&
3022 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
3023 pr_err("0x%px: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3024 objp, (int)ARCH_SLAB_MINALIGN);
3025 }
3026 return objp;
3027}
3028#else
3029#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3030#endif
3031
3032static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3033{
3034 void *objp;
3035 struct array_cache *ac;
3036
3037 check_irq_off();
3038
3039 ac = cpu_cache_get(cachep);
3040 if (likely(ac->avail)) {
3041 ac->touched = 1;
3042 objp = ac->entry[--ac->avail];
3043
3044 STATS_INC_ALLOCHIT(cachep);
3045 goto out;
3046 }
3047
3048 STATS_INC_ALLOCMISS(cachep);
3049 objp = cache_alloc_refill(cachep, flags);
3050 /*
3051 * the 'ac' may be updated by cache_alloc_refill(),
3052 * and kmemleak_erase() requires its correct value.
3053 */
3054 ac = cpu_cache_get(cachep);
3055
3056out:
3057 /*
3058 * To avoid a false negative, if an object that is in one of the
3059 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3060 * treat the array pointers as a reference to the object.
3061 */
3062 if (objp)
3063 kmemleak_erase(&ac->entry[ac->avail]);
3064 return objp;
3065}
3066
3067#ifdef CONFIG_NUMA
3068/*
3069 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
3070 *
3071 * If we are in_interrupt, then process context, including cpusets and
3072 * mempolicy, may not apply and should not be used for allocation policy.
3073 */
3074static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3075{
3076 int nid_alloc, nid_here;
3077
3078 if (in_interrupt() || (flags & __GFP_THISNODE))
3079 return NULL;
3080 nid_alloc = nid_here = numa_mem_id();
3081 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3082 nid_alloc = cpuset_slab_spread_node();
3083 else if (current->mempolicy)
3084 nid_alloc = mempolicy_slab_node();
3085 if (nid_alloc != nid_here)
3086 return ____cache_alloc_node(cachep, flags, nid_alloc);
3087 return NULL;
3088}
3089
3090/*
3091 * Fallback function if there was no memory available and no objects on a
3092 * certain node and fall back is permitted. First we scan all the
3093 * available node for available objects. If that fails then we
3094 * perform an allocation without specifying a node. This allows the page
3095 * allocator to do its reclaim / fallback magic. We then insert the
3096 * slab into the proper nodelist and then allocate from it.
3097 */
3098static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3099{
3100 struct zonelist *zonelist;
3101 struct zoneref *z;
3102 struct zone *zone;
3103 enum zone_type highest_zoneidx = gfp_zone(flags);
3104 void *obj = NULL;
3105 struct page *page;
3106 int nid;
3107 unsigned int cpuset_mems_cookie;
3108
3109 if (flags & __GFP_THISNODE)
3110 return NULL;
3111
3112retry_cpuset:
3113 cpuset_mems_cookie = read_mems_allowed_begin();
3114 zonelist = node_zonelist(mempolicy_slab_node(), flags);
3115
3116retry:
3117 /*
3118 * Look through allowed nodes for objects available
3119 * from existing per node queues.
3120 */
3121 for_each_zone_zonelist(zone, z, zonelist, highest_zoneidx) {
3122 nid = zone_to_nid(zone);
3123
3124 if (cpuset_zone_allowed(zone, flags) &&
3125 get_node(cache, nid) &&
3126 get_node(cache, nid)->free_objects) {
3127 obj = ____cache_alloc_node(cache,
3128 gfp_exact_node(flags), nid);
3129 if (obj)
3130 break;
3131 }
3132 }
3133
3134 if (!obj) {
3135 /*
3136 * This allocation will be performed within the constraints
3137 * of the current cpuset / memory policy requirements.
3138 * We may trigger various forms of reclaim on the allowed
3139 * set and go into memory reserves if necessary.
3140 */
3141 page = cache_grow_begin(cache, flags, numa_mem_id());
3142 cache_grow_end(cache, page);
3143 if (page) {
3144 nid = page_to_nid(page);
3145 obj = ____cache_alloc_node(cache,
3146 gfp_exact_node(flags), nid);
3147
3148 /*
3149 * Another processor may allocate the objects in
3150 * the slab since we are not holding any locks.
3151 */
3152 if (!obj)
3153 goto retry;
3154 }
3155 }
3156
3157 if (unlikely(!obj && read_mems_allowed_retry(cpuset_mems_cookie)))
3158 goto retry_cpuset;
3159 return obj;
3160}
3161
3162/*
3163 * A interface to enable slab creation on nodeid
3164 */
3165static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3166 int nodeid)
3167{
3168 struct page *page;
3169 struct kmem_cache_node *n;
3170 void *obj = NULL;
3171 void *list = NULL;
3172
3173 VM_BUG_ON(nodeid < 0 || nodeid >= MAX_NUMNODES);
3174 n = get_node(cachep, nodeid);
3175 BUG_ON(!n);
3176
3177 check_irq_off();
3178 spin_lock(&n->list_lock);
3179 page = get_first_slab(n, false);
3180 if (!page)
3181 goto must_grow;
3182
3183 check_spinlock_acquired_node(cachep, nodeid);
3184
3185 STATS_INC_NODEALLOCS(cachep);
3186 STATS_INC_ACTIVE(cachep);
3187 STATS_SET_HIGH(cachep);
3188
3189 BUG_ON(page->active == cachep->num);
3190
3191 obj = slab_get_obj(cachep, page);
3192 n->free_objects--;
3193
3194 fixup_slab_list(cachep, n, page, &list);
3195
3196 spin_unlock(&n->list_lock);
3197 fixup_objfreelist_debug(cachep, &list);
3198 return obj;
3199
3200must_grow:
3201 spin_unlock(&n->list_lock);
3202 page = cache_grow_begin(cachep, gfp_exact_node(flags), nodeid);
3203 if (page) {
3204 /* This slab isn't counted yet so don't update free_objects */
3205 obj = slab_get_obj(cachep, page);
3206 }
3207 cache_grow_end(cachep, page);
3208
3209 return obj ? obj : fallback_alloc(cachep, flags);
3210}
3211
3212static __always_inline void *
3213slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3214 unsigned long caller)
3215{
3216 unsigned long save_flags;
3217 void *ptr;
3218 int slab_node = numa_mem_id();
3219 struct obj_cgroup *objcg = NULL;
3220
3221 flags &= gfp_allowed_mask;
3222 cachep = slab_pre_alloc_hook(cachep, &objcg, 1, flags);
3223 if (unlikely(!cachep))
3224 return NULL;
3225
3226 cache_alloc_debugcheck_before(cachep, flags);
3227 local_irq_save(save_flags);
3228
3229 if (nodeid == NUMA_NO_NODE)
3230 nodeid = slab_node;
3231
3232 if (unlikely(!get_node(cachep, nodeid))) {
3233 /* Node not bootstrapped yet */
3234 ptr = fallback_alloc(cachep, flags);
3235 goto out;
3236 }
3237
3238 if (nodeid == slab_node) {
3239 /*
3240 * Use the locally cached objects if possible.
3241 * However ____cache_alloc does not allow fallback
3242 * to other nodes. It may fail while we still have
3243 * objects on other nodes available.
3244 */
3245 ptr = ____cache_alloc(cachep, flags);
3246 if (ptr)
3247 goto out;
3248 }
3249 /* ___cache_alloc_node can fall back to other nodes */
3250 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3251 out:
3252 local_irq_restore(save_flags);
3253 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3254
3255 if (unlikely(slab_want_init_on_alloc(flags, cachep)) && ptr)
3256 memset(ptr, 0, cachep->object_size);
3257
3258 slab_post_alloc_hook(cachep, objcg, flags, 1, &ptr);
3259 return ptr;
3260}
3261
3262static __always_inline void *
3263__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3264{
3265 void *objp;
3266
3267 if (current->mempolicy || cpuset_do_slab_mem_spread()) {
3268 objp = alternate_node_alloc(cache, flags);
3269 if (objp)
3270 goto out;
3271 }
3272 objp = ____cache_alloc(cache, flags);
3273
3274 /*
3275 * We may just have run out of memory on the local node.
3276 * ____cache_alloc_node() knows how to locate memory on other nodes
3277 */
3278 if (!objp)
3279 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3280
3281 out:
3282 return objp;
3283}
3284#else
3285
3286static __always_inline void *
3287__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3288{
3289 return ____cache_alloc(cachep, flags);
3290}
3291
3292#endif /* CONFIG_NUMA */
3293
3294static __always_inline void *
3295slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
3296{
3297 unsigned long save_flags;
3298 void *objp;
3299 struct obj_cgroup *objcg = NULL;
3300
3301 flags &= gfp_allowed_mask;
3302 cachep = slab_pre_alloc_hook(cachep, &objcg, 1, flags);
3303 if (unlikely(!cachep))
3304 return NULL;
3305
3306 cache_alloc_debugcheck_before(cachep, flags);
3307 local_irq_save(save_flags);
3308 objp = __do_cache_alloc(cachep, flags);
3309 local_irq_restore(save_flags);
3310 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3311 prefetchw(objp);
3312
3313 if (unlikely(slab_want_init_on_alloc(flags, cachep)) && objp)
3314 memset(objp, 0, cachep->object_size);
3315
3316 slab_post_alloc_hook(cachep, objcg, flags, 1, &objp);
3317 return objp;
3318}
3319
3320/*
3321 * Caller needs to acquire correct kmem_cache_node's list_lock
3322 * @list: List of detached free slabs should be freed by caller
3323 */
3324static void free_block(struct kmem_cache *cachep, void **objpp,
3325 int nr_objects, int node, struct list_head *list)
3326{
3327 int i;
3328 struct kmem_cache_node *n = get_node(cachep, node);
3329 struct page *page;
3330
3331 n->free_objects += nr_objects;
3332
3333 for (i = 0; i < nr_objects; i++) {
3334 void *objp;
3335 struct page *page;
3336
3337 objp = objpp[i];
3338
3339 page = virt_to_head_page(objp);
3340 list_del(&page->slab_list);
3341 check_spinlock_acquired_node(cachep, node);
3342 slab_put_obj(cachep, page, objp);
3343 STATS_DEC_ACTIVE(cachep);
3344
3345 /* fixup slab chains */
3346 if (page->active == 0) {
3347 list_add(&page->slab_list, &n->slabs_free);
3348 n->free_slabs++;
3349 } else {
3350 /* Unconditionally move a slab to the end of the
3351 * partial list on free - maximum time for the
3352 * other objects to be freed, too.
3353 */
3354 list_add_tail(&page->slab_list, &n->slabs_partial);
3355 }
3356 }
3357
3358 while (n->free_objects > n->free_limit && !list_empty(&n->slabs_free)) {
3359 n->free_objects -= cachep->num;
3360
3361 page = list_last_entry(&n->slabs_free, struct page, slab_list);
3362 list_move(&page->slab_list, list);
3363 n->free_slabs--;
3364 n->total_slabs--;
3365 }
3366}
3367
3368static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3369{
3370 int batchcount;
3371 struct kmem_cache_node *n;
3372 int node = numa_mem_id();
3373 LIST_HEAD(list);
3374
3375 batchcount = ac->batchcount;
3376
3377 check_irq_off();
3378 n = get_node(cachep, node);
3379 spin_lock(&n->list_lock);
3380 if (n->shared) {
3381 struct array_cache *shared_array = n->shared;
3382 int max = shared_array->limit - shared_array->avail;
3383 if (max) {
3384 if (batchcount > max)
3385 batchcount = max;
3386 memcpy(&(shared_array->entry[shared_array->avail]),
3387 ac->entry, sizeof(void *) * batchcount);
3388 shared_array->avail += batchcount;
3389 goto free_done;
3390 }
3391 }
3392
3393 free_block(cachep, ac->entry, batchcount, node, &list);
3394free_done:
3395#if STATS
3396 {
3397 int i = 0;
3398 struct page *page;
3399
3400 list_for_each_entry(page, &n->slabs_free, slab_list) {
3401 BUG_ON(page->active);
3402
3403 i++;
3404 }
3405 STATS_SET_FREEABLE(cachep, i);
3406 }
3407#endif
3408 spin_unlock(&n->list_lock);
3409 ac->avail -= batchcount;
3410 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3411 slabs_destroy(cachep, &list);
3412}
3413
3414/*
3415 * Release an obj back to its cache. If the obj has a constructed state, it must
3416 * be in this state _before_ it is released. Called with disabled ints.
3417 */
3418static __always_inline void __cache_free(struct kmem_cache *cachep, void *objp,
3419 unsigned long caller)
3420{
3421 /* Put the object into the quarantine, don't touch it for now. */
3422 if (kasan_slab_free(cachep, objp, _RET_IP_))
3423 return;
3424
3425 /* Use KCSAN to help debug racy use-after-free. */
3426 if (!(cachep->flags & SLAB_TYPESAFE_BY_RCU))
3427 __kcsan_check_access(objp, cachep->object_size,
3428 KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT);
3429
3430 ___cache_free(cachep, objp, caller);
3431}
3432
3433void ___cache_free(struct kmem_cache *cachep, void *objp,
3434 unsigned long caller)
3435{
3436 struct array_cache *ac = cpu_cache_get(cachep);
3437
3438 check_irq_off();
3439 if (unlikely(slab_want_init_on_free(cachep)))
3440 memset(objp, 0, cachep->object_size);
3441 kmemleak_free_recursive(objp, cachep->flags);
3442 objp = cache_free_debugcheck(cachep, objp, caller);
3443 memcg_slab_free_hook(cachep, virt_to_head_page(objp), objp);
3444
3445 /*
3446 * Skip calling cache_free_alien() when the platform is not numa.
3447 * This will avoid cache misses that happen while accessing slabp (which
3448 * is per page memory reference) to get nodeid. Instead use a global
3449 * variable to skip the call, which is mostly likely to be present in
3450 * the cache.
3451 */
3452 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3453 return;
3454
3455 if (ac->avail < ac->limit) {
3456 STATS_INC_FREEHIT(cachep);
3457 } else {
3458 STATS_INC_FREEMISS(cachep);
3459 cache_flusharray(cachep, ac);
3460 }
3461
3462 if (sk_memalloc_socks()) {
3463 struct page *page = virt_to_head_page(objp);
3464
3465 if (unlikely(PageSlabPfmemalloc(page))) {
3466 cache_free_pfmemalloc(cachep, page, objp);
3467 return;
3468 }
3469 }
3470
3471 __free_one(ac, objp);
3472}
3473
3474/**
3475 * kmem_cache_alloc - Allocate an object
3476 * @cachep: The cache to allocate from.
3477 * @flags: See kmalloc().
3478 *
3479 * Allocate an object from this cache. The flags are only relevant
3480 * if the cache has no available objects.
3481 *
3482 * Return: pointer to the new object or %NULL in case of error
3483 */
3484void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3485{
3486 void *ret = slab_alloc(cachep, flags, _RET_IP_);
3487
3488 trace_kmem_cache_alloc(_RET_IP_, ret,
3489 cachep->object_size, cachep->size, flags);
3490
3491 return ret;
3492}
3493EXPORT_SYMBOL(kmem_cache_alloc);
3494
3495static __always_inline void
3496cache_alloc_debugcheck_after_bulk(struct kmem_cache *s, gfp_t flags,
3497 size_t size, void **p, unsigned long caller)
3498{
3499 size_t i;
3500
3501 for (i = 0; i < size; i++)
3502 p[i] = cache_alloc_debugcheck_after(s, flags, p[i], caller);
3503}
3504
3505int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3506 void **p)
3507{
3508 size_t i;
3509 struct obj_cgroup *objcg = NULL;
3510
3511 s = slab_pre_alloc_hook(s, &objcg, size, flags);
3512 if (!s)
3513 return 0;
3514
3515 cache_alloc_debugcheck_before(s, flags);
3516
3517 local_irq_disable();
3518 for (i = 0; i < size; i++) {
3519 void *objp = __do_cache_alloc(s, flags);
3520
3521 if (unlikely(!objp))
3522 goto error;
3523 p[i] = objp;
3524 }
3525 local_irq_enable();
3526
3527 cache_alloc_debugcheck_after_bulk(s, flags, size, p, _RET_IP_);
3528
3529 /* Clear memory outside IRQ disabled section */
3530 if (unlikely(slab_want_init_on_alloc(flags, s)))
3531 for (i = 0; i < size; i++)
3532 memset(p[i], 0, s->object_size);
3533
3534 slab_post_alloc_hook(s, objcg, flags, size, p);
3535 /* FIXME: Trace call missing. Christoph would like a bulk variant */
3536 return size;
3537error:
3538 local_irq_enable();
3539 cache_alloc_debugcheck_after_bulk(s, flags, i, p, _RET_IP_);
3540 slab_post_alloc_hook(s, objcg, flags, i, p);
3541 __kmem_cache_free_bulk(s, i, p);
3542 return 0;
3543}
3544EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3545
3546#ifdef CONFIG_TRACING
3547void *
3548kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
3549{
3550 void *ret;
3551
3552 ret = slab_alloc(cachep, flags, _RET_IP_);
3553
3554 ret = kasan_kmalloc(cachep, ret, size, flags);
3555 trace_kmalloc(_RET_IP_, ret,
3556 size, cachep->size, flags);
3557 return ret;
3558}
3559EXPORT_SYMBOL(kmem_cache_alloc_trace);
3560#endif
3561
3562#ifdef CONFIG_NUMA
3563/**
3564 * kmem_cache_alloc_node - Allocate an object on the specified node
3565 * @cachep: The cache to allocate from.
3566 * @flags: See kmalloc().
3567 * @nodeid: node number of the target node.
3568 *
3569 * Identical to kmem_cache_alloc but it will allocate memory on the given
3570 * node, which can improve the performance for cpu bound structures.
3571 *
3572 * Fallback to other node is possible if __GFP_THISNODE is not set.
3573 *
3574 * Return: pointer to the new object or %NULL in case of error
3575 */
3576void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3577{
3578 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3579
3580 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3581 cachep->object_size, cachep->size,
3582 flags, nodeid);
3583
3584 return ret;
3585}
3586EXPORT_SYMBOL(kmem_cache_alloc_node);
3587
3588#ifdef CONFIG_TRACING
3589void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
3590 gfp_t flags,
3591 int nodeid,
3592 size_t size)
3593{
3594 void *ret;
3595
3596 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3597
3598 ret = kasan_kmalloc(cachep, ret, size, flags);
3599 trace_kmalloc_node(_RET_IP_, ret,
3600 size, cachep->size,
3601 flags, nodeid);
3602 return ret;
3603}
3604EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3605#endif
3606
3607static __always_inline void *
3608__do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
3609{
3610 struct kmem_cache *cachep;
3611 void *ret;
3612
3613 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3614 return NULL;
3615 cachep = kmalloc_slab(size, flags);
3616 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3617 return cachep;
3618 ret = kmem_cache_alloc_node_trace(cachep, flags, node, size);
3619 ret = kasan_kmalloc(cachep, ret, size, flags);
3620
3621 return ret;
3622}
3623
3624void *__kmalloc_node(size_t size, gfp_t flags, int node)
3625{
3626 return __do_kmalloc_node(size, flags, node, _RET_IP_);
3627}
3628EXPORT_SYMBOL(__kmalloc_node);
3629
3630void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3631 int node, unsigned long caller)
3632{
3633 return __do_kmalloc_node(size, flags, node, caller);
3634}
3635EXPORT_SYMBOL(__kmalloc_node_track_caller);
3636#endif /* CONFIG_NUMA */
3637
3638/**
3639 * __do_kmalloc - allocate memory
3640 * @size: how many bytes of memory are required.
3641 * @flags: the type of memory to allocate (see kmalloc).
3642 * @caller: function caller for debug tracking of the caller
3643 *
3644 * Return: pointer to the allocated memory or %NULL in case of error
3645 */
3646static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3647 unsigned long caller)
3648{
3649 struct kmem_cache *cachep;
3650 void *ret;
3651
3652 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3653 return NULL;
3654 cachep = kmalloc_slab(size, flags);
3655 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3656 return cachep;
3657 ret = slab_alloc(cachep, flags, caller);
3658
3659 ret = kasan_kmalloc(cachep, ret, size, flags);
3660 trace_kmalloc(caller, ret,
3661 size, cachep->size, flags);
3662
3663 return ret;
3664}
3665
3666void *__kmalloc(size_t size, gfp_t flags)
3667{
3668 return __do_kmalloc(size, flags, _RET_IP_);
3669}
3670EXPORT_SYMBOL(__kmalloc);
3671
3672void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3673{
3674 return __do_kmalloc(size, flags, caller);
3675}
3676EXPORT_SYMBOL(__kmalloc_track_caller);
3677
3678/**
3679 * kmem_cache_free - Deallocate an object
3680 * @cachep: The cache the allocation was from.
3681 * @objp: The previously allocated object.
3682 *
3683 * Free an object which was previously allocated from this
3684 * cache.
3685 */
3686void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3687{
3688 unsigned long flags;
3689 cachep = cache_from_obj(cachep, objp);
3690 if (!cachep)
3691 return;
3692
3693 local_irq_save(flags);
3694 debug_check_no_locks_freed(objp, cachep->object_size);
3695 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3696 debug_check_no_obj_freed(objp, cachep->object_size);
3697 __cache_free(cachep, objp, _RET_IP_);
3698 local_irq_restore(flags);
3699
3700 trace_kmem_cache_free(_RET_IP_, objp);
3701}
3702EXPORT_SYMBOL(kmem_cache_free);
3703
3704void kmem_cache_free_bulk(struct kmem_cache *orig_s, size_t size, void **p)
3705{
3706 struct kmem_cache *s;
3707 size_t i;
3708
3709 local_irq_disable();
3710 for (i = 0; i < size; i++) {
3711 void *objp = p[i];
3712
3713 if (!orig_s) /* called via kfree_bulk */
3714 s = virt_to_cache(objp);
3715 else
3716 s = cache_from_obj(orig_s, objp);
3717 if (!s)
3718 continue;
3719
3720 debug_check_no_locks_freed(objp, s->object_size);
3721 if (!(s->flags & SLAB_DEBUG_OBJECTS))
3722 debug_check_no_obj_freed(objp, s->object_size);
3723
3724 __cache_free(s, objp, _RET_IP_);
3725 }
3726 local_irq_enable();
3727
3728 /* FIXME: add tracing */
3729}
3730EXPORT_SYMBOL(kmem_cache_free_bulk);
3731
3732/**
3733 * kfree - free previously allocated memory
3734 * @objp: pointer returned by kmalloc.
3735 *
3736 * If @objp is NULL, no operation is performed.
3737 *
3738 * Don't free memory not originally allocated by kmalloc()
3739 * or you will run into trouble.
3740 */
3741void kfree(const void *objp)
3742{
3743 struct kmem_cache *c;
3744 unsigned long flags;
3745
3746 trace_kfree(_RET_IP_, objp);
3747
3748 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3749 return;
3750 local_irq_save(flags);
3751 kfree_debugcheck(objp);
3752 c = virt_to_cache(objp);
3753 if (!c) {
3754 local_irq_restore(flags);
3755 return;
3756 }
3757 debug_check_no_locks_freed(objp, c->object_size);
3758
3759 debug_check_no_obj_freed(objp, c->object_size);
3760 __cache_free(c, (void *)objp, _RET_IP_);
3761 local_irq_restore(flags);
3762}
3763EXPORT_SYMBOL(kfree);
3764
3765/*
3766 * This initializes kmem_cache_node or resizes various caches for all nodes.
3767 */
3768static int setup_kmem_cache_nodes(struct kmem_cache *cachep, gfp_t gfp)
3769{
3770 int ret;
3771 int node;
3772 struct kmem_cache_node *n;
3773
3774 for_each_online_node(node) {
3775 ret = setup_kmem_cache_node(cachep, node, gfp, true);
3776 if (ret)
3777 goto fail;
3778
3779 }
3780
3781 return 0;
3782
3783fail:
3784 if (!cachep->list.next) {
3785 /* Cache is not active yet. Roll back what we did */
3786 node--;
3787 while (node >= 0) {
3788 n = get_node(cachep, node);
3789 if (n) {
3790 kfree(n->shared);
3791 free_alien_cache(n->alien);
3792 kfree(n);
3793 cachep->node[node] = NULL;
3794 }
3795 node--;
3796 }
3797 }
3798 return -ENOMEM;
3799}
3800
3801/* Always called with the slab_mutex held */
3802static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3803 int batchcount, int shared, gfp_t gfp)
3804{
3805 struct array_cache __percpu *cpu_cache, *prev;
3806 int cpu;
3807
3808 cpu_cache = alloc_kmem_cache_cpus(cachep, limit, batchcount);
3809 if (!cpu_cache)
3810 return -ENOMEM;
3811
3812 prev = cachep->cpu_cache;
3813 cachep->cpu_cache = cpu_cache;
3814 /*
3815 * Without a previous cpu_cache there's no need to synchronize remote
3816 * cpus, so skip the IPIs.
3817 */
3818 if (prev)
3819 kick_all_cpus_sync();
3820
3821 check_irq_on();
3822 cachep->batchcount = batchcount;
3823 cachep->limit = limit;
3824 cachep->shared = shared;
3825
3826 if (!prev)
3827 goto setup_node;
3828
3829 for_each_online_cpu(cpu) {
3830 LIST_HEAD(list);
3831 int node;
3832 struct kmem_cache_node *n;
3833 struct array_cache *ac = per_cpu_ptr(prev, cpu);
3834
3835 node = cpu_to_mem(cpu);
3836 n = get_node(cachep, node);
3837 spin_lock_irq(&n->list_lock);
3838 free_block(cachep, ac->entry, ac->avail, node, &list);
3839 spin_unlock_irq(&n->list_lock);
3840 slabs_destroy(cachep, &list);
3841 }
3842 free_percpu(prev);
3843
3844setup_node:
3845 return setup_kmem_cache_nodes(cachep, gfp);
3846}
3847
3848/* Called with slab_mutex held always */
3849static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
3850{
3851 int err;
3852 int limit = 0;
3853 int shared = 0;
3854 int batchcount = 0;
3855
3856 err = cache_random_seq_create(cachep, cachep->num, gfp);
3857 if (err)
3858 goto end;
3859
3860 if (limit && shared && batchcount)
3861 goto skip_setup;
3862 /*
3863 * The head array serves three purposes:
3864 * - create a LIFO ordering, i.e. return objects that are cache-warm
3865 * - reduce the number of spinlock operations.
3866 * - reduce the number of linked list operations on the slab and
3867 * bufctl chains: array operations are cheaper.
3868 * The numbers are guessed, we should auto-tune as described by
3869 * Bonwick.
3870 */
3871 if (cachep->size > 131072)
3872 limit = 1;
3873 else if (cachep->size > PAGE_SIZE)
3874 limit = 8;
3875 else if (cachep->size > 1024)
3876 limit = 24;
3877 else if (cachep->size > 256)
3878 limit = 54;
3879 else
3880 limit = 120;
3881
3882 /*
3883 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3884 * allocation behaviour: Most allocs on one cpu, most free operations
3885 * on another cpu. For these cases, an efficient object passing between
3886 * cpus is necessary. This is provided by a shared array. The array
3887 * replaces Bonwick's magazine layer.
3888 * On uniprocessor, it's functionally equivalent (but less efficient)
3889 * to a larger limit. Thus disabled by default.
3890 */
3891 shared = 0;
3892 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
3893 shared = 8;
3894
3895#if DEBUG
3896 /*
3897 * With debugging enabled, large batchcount lead to excessively long
3898 * periods with disabled local interrupts. Limit the batchcount
3899 */
3900 if (limit > 32)
3901 limit = 32;
3902#endif
3903 batchcount = (limit + 1) / 2;
3904skip_setup:
3905 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3906end:
3907 if (err)
3908 pr_err("enable_cpucache failed for %s, error %d\n",
3909 cachep->name, -err);
3910 return err;
3911}
3912
3913/*
3914 * Drain an array if it contains any elements taking the node lock only if
3915 * necessary. Note that the node listlock also protects the array_cache
3916 * if drain_array() is used on the shared array.
3917 */
3918static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
3919 struct array_cache *ac, int node)
3920{
3921 LIST_HEAD(list);
3922
3923 /* ac from n->shared can be freed if we don't hold the slab_mutex. */
3924 check_mutex_acquired();
3925
3926 if (!ac || !ac->avail)
3927 return;
3928
3929 if (ac->touched) {
3930 ac->touched = 0;
3931 return;
3932 }
3933
3934 spin_lock_irq(&n->list_lock);
3935 drain_array_locked(cachep, ac, node, false, &list);
3936 spin_unlock_irq(&n->list_lock);
3937
3938 slabs_destroy(cachep, &list);
3939}
3940
3941/**
3942 * cache_reap - Reclaim memory from caches.
3943 * @w: work descriptor
3944 *
3945 * Called from workqueue/eventd every few seconds.
3946 * Purpose:
3947 * - clear the per-cpu caches for this CPU.
3948 * - return freeable pages to the main free memory pool.
3949 *
3950 * If we cannot acquire the cache chain mutex then just give up - we'll try
3951 * again on the next iteration.
3952 */
3953static void cache_reap(struct work_struct *w)
3954{
3955 struct kmem_cache *searchp;
3956 struct kmem_cache_node *n;
3957 int node = numa_mem_id();
3958 struct delayed_work *work = to_delayed_work(w);
3959
3960 if (!mutex_trylock(&slab_mutex))
3961 /* Give up. Setup the next iteration. */
3962 goto out;
3963
3964 list_for_each_entry(searchp, &slab_caches, list) {
3965 check_irq_on();
3966
3967 /*
3968 * We only take the node lock if absolutely necessary and we
3969 * have established with reasonable certainty that
3970 * we can do some work if the lock was obtained.
3971 */
3972 n = get_node(searchp, node);
3973
3974 reap_alien(searchp, n);
3975
3976 drain_array(searchp, n, cpu_cache_get(searchp), node);
3977
3978 /*
3979 * These are racy checks but it does not matter
3980 * if we skip one check or scan twice.
3981 */
3982 if (time_after(n->next_reap, jiffies))
3983 goto next;
3984
3985 n->next_reap = jiffies + REAPTIMEOUT_NODE;
3986
3987 drain_array(searchp, n, n->shared, node);
3988
3989 if (n->free_touched)
3990 n->free_touched = 0;
3991 else {
3992 int freed;
3993
3994 freed = drain_freelist(searchp, n, (n->free_limit +
3995 5 * searchp->num - 1) / (5 * searchp->num));
3996 STATS_ADD_REAPED(searchp, freed);
3997 }
3998next:
3999 cond_resched();
4000 }
4001 check_irq_on();
4002 mutex_unlock(&slab_mutex);
4003 next_reap_node();
4004out:
4005 /* Set up the next iteration */
4006 schedule_delayed_work_on(smp_processor_id(), work,
4007 round_jiffies_relative(REAPTIMEOUT_AC));
4008}
4009
4010void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
4011{
4012 unsigned long active_objs, num_objs, active_slabs;
4013 unsigned long total_slabs = 0, free_objs = 0, shared_avail = 0;
4014 unsigned long free_slabs = 0;
4015 int node;
4016 struct kmem_cache_node *n;
4017
4018 for_each_kmem_cache_node(cachep, node, n) {
4019 check_irq_on();
4020 spin_lock_irq(&n->list_lock);
4021
4022 total_slabs += n->total_slabs;
4023 free_slabs += n->free_slabs;
4024 free_objs += n->free_objects;
4025
4026 if (n->shared)
4027 shared_avail += n->shared->avail;
4028
4029 spin_unlock_irq(&n->list_lock);
4030 }
4031 num_objs = total_slabs * cachep->num;
4032 active_slabs = total_slabs - free_slabs;
4033 active_objs = num_objs - free_objs;
4034
4035 sinfo->active_objs = active_objs;
4036 sinfo->num_objs = num_objs;
4037 sinfo->active_slabs = active_slabs;
4038 sinfo->num_slabs = total_slabs;
4039 sinfo->shared_avail = shared_avail;
4040 sinfo->limit = cachep->limit;
4041 sinfo->batchcount = cachep->batchcount;
4042 sinfo->shared = cachep->shared;
4043 sinfo->objects_per_slab = cachep->num;
4044 sinfo->cache_order = cachep->gfporder;
4045}
4046
4047void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4048{
4049#if STATS
4050 { /* node stats */
4051 unsigned long high = cachep->high_mark;
4052 unsigned long allocs = cachep->num_allocations;
4053 unsigned long grown = cachep->grown;
4054 unsigned long reaped = cachep->reaped;
4055 unsigned long errors = cachep->errors;
4056 unsigned long max_freeable = cachep->max_freeable;
4057 unsigned long node_allocs = cachep->node_allocs;
4058 unsigned long node_frees = cachep->node_frees;
4059 unsigned long overflows = cachep->node_overflow;
4060
4061 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu %4lu %4lu",
4062 allocs, high, grown,
4063 reaped, errors, max_freeable, node_allocs,
4064 node_frees, overflows);
4065 }
4066 /* cpu stats */
4067 {
4068 unsigned long allochit = atomic_read(&cachep->allochit);
4069 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4070 unsigned long freehit = atomic_read(&cachep->freehit);
4071 unsigned long freemiss = atomic_read(&cachep->freemiss);
4072
4073 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4074 allochit, allocmiss, freehit, freemiss);
4075 }
4076#endif
4077}
4078
4079#define MAX_SLABINFO_WRITE 128
4080/**
4081 * slabinfo_write - Tuning for the slab allocator
4082 * @file: unused
4083 * @buffer: user buffer
4084 * @count: data length
4085 * @ppos: unused
4086 *
4087 * Return: %0 on success, negative error code otherwise.
4088 */
4089ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4090 size_t count, loff_t *ppos)
4091{
4092 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4093 int limit, batchcount, shared, res;
4094 struct kmem_cache *cachep;
4095
4096 if (count > MAX_SLABINFO_WRITE)
4097 return -EINVAL;
4098 if (copy_from_user(&kbuf, buffer, count))
4099 return -EFAULT;
4100 kbuf[MAX_SLABINFO_WRITE] = '\0';
4101
4102 tmp = strchr(kbuf, ' ');
4103 if (!tmp)
4104 return -EINVAL;
4105 *tmp = '\0';
4106 tmp++;
4107 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4108 return -EINVAL;
4109
4110 /* Find the cache in the chain of caches. */
4111 mutex_lock(&slab_mutex);
4112 res = -EINVAL;
4113 list_for_each_entry(cachep, &slab_caches, list) {
4114 if (!strcmp(cachep->name, kbuf)) {
4115 if (limit < 1 || batchcount < 1 ||
4116 batchcount > limit || shared < 0) {
4117 res = 0;
4118 } else {
4119 res = do_tune_cpucache(cachep, limit,
4120 batchcount, shared,
4121 GFP_KERNEL);
4122 }
4123 break;
4124 }
4125 }
4126 mutex_unlock(&slab_mutex);
4127 if (res >= 0)
4128 res = count;
4129 return res;
4130}
4131
4132#ifdef CONFIG_HARDENED_USERCOPY
4133/*
4134 * Rejects incorrectly sized objects and objects that are to be copied
4135 * to/from userspace but do not fall entirely within the containing slab
4136 * cache's usercopy region.
4137 *
4138 * Returns NULL if check passes, otherwise const char * to name of cache
4139 * to indicate an error.
4140 */
4141void __check_heap_object(const void *ptr, unsigned long n, struct page *page,
4142 bool to_user)
4143{
4144 struct kmem_cache *cachep;
4145 unsigned int objnr;
4146 unsigned long offset;
4147
4148 ptr = kasan_reset_tag(ptr);
4149
4150 /* Find and validate object. */
4151 cachep = page->slab_cache;
4152 objnr = obj_to_index(cachep, page, (void *)ptr);
4153 BUG_ON(objnr >= cachep->num);
4154
4155 /* Find offset within object. */
4156 offset = ptr - index_to_obj(cachep, page, objnr) - obj_offset(cachep);
4157
4158 /* Allow address range falling entirely within usercopy region. */
4159 if (offset >= cachep->useroffset &&
4160 offset - cachep->useroffset <= cachep->usersize &&
4161 n <= cachep->useroffset - offset + cachep->usersize)
4162 return;
4163
4164 /*
4165 * If the copy is still within the allocated object, produce
4166 * a warning instead of rejecting the copy. This is intended
4167 * to be a temporary method to find any missing usercopy
4168 * whitelists.
4169 */
4170 if (usercopy_fallback &&
4171 offset <= cachep->object_size &&
4172 n <= cachep->object_size - offset) {
4173 usercopy_warn("SLAB object", cachep->name, to_user, offset, n);
4174 return;
4175 }
4176
4177 usercopy_abort("SLAB object", cachep->name, to_user, offset, n);
4178}
4179#endif /* CONFIG_HARDENED_USERCOPY */
4180
4181/**
4182 * __ksize -- Uninstrumented ksize.
4183 * @objp: pointer to the object
4184 *
4185 * Unlike ksize(), __ksize() is uninstrumented, and does not provide the same
4186 * safety checks as ksize() with KASAN instrumentation enabled.
4187 *
4188 * Return: size of the actual memory used by @objp in bytes
4189 */
4190size_t __ksize(const void *objp)
4191{
4192 struct kmem_cache *c;
4193 size_t size;
4194
4195 BUG_ON(!objp);
4196 if (unlikely(objp == ZERO_SIZE_PTR))
4197 return 0;
4198
4199 c = virt_to_cache(objp);
4200 size = c ? c->object_size : 0;
4201
4202 return size;
4203}
4204EXPORT_SYMBOL(__ksize);