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#ifndef CONFIG_NUMA
592
593#define drain_alien_cache(cachep, alien) do { } while (0)
594#define reap_alien(cachep, n) do { } while (0)
595
596static inline struct alien_cache **alloc_alien_cache(int node,
597 int limit, gfp_t gfp)
598{
599 return NULL;
600}
601
602static inline void free_alien_cache(struct alien_cache **ac_ptr)
603{
604}
605
606static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
607{
608 return 0;
609}
610
611static inline void *alternate_node_alloc(struct kmem_cache *cachep,
612 gfp_t flags)
613{
614 return NULL;
615}
616
617static inline void *____cache_alloc_node(struct kmem_cache *cachep,
618 gfp_t flags, int nodeid)
619{
620 return NULL;
621}
622
623static inline gfp_t gfp_exact_node(gfp_t flags)
624{
625 return flags & ~__GFP_NOFAIL;
626}
627
628#else /* CONFIG_NUMA */
629
630static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
631static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
632
633static struct alien_cache *__alloc_alien_cache(int node, int entries,
634 int batch, gfp_t gfp)
635{
636 size_t memsize = sizeof(void *) * entries + sizeof(struct alien_cache);
637 struct alien_cache *alc = NULL;
638
639 alc = kmalloc_node(memsize, gfp, node);
640 if (alc) {
641 kmemleak_no_scan(alc);
642 init_arraycache(&alc->ac, entries, batch);
643 spin_lock_init(&alc->lock);
644 }
645 return alc;
646}
647
648static struct alien_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
649{
650 struct alien_cache **alc_ptr;
651 int i;
652
653 if (limit > 1)
654 limit = 12;
655 alc_ptr = kcalloc_node(nr_node_ids, sizeof(void *), gfp, node);
656 if (!alc_ptr)
657 return NULL;
658
659 for_each_node(i) {
660 if (i == node || !node_online(i))
661 continue;
662 alc_ptr[i] = __alloc_alien_cache(node, limit, 0xbaadf00d, gfp);
663 if (!alc_ptr[i]) {
664 for (i--; i >= 0; i--)
665 kfree(alc_ptr[i]);
666 kfree(alc_ptr);
667 return NULL;
668 }
669 }
670 return alc_ptr;
671}
672
673static void free_alien_cache(struct alien_cache **alc_ptr)
674{
675 int i;
676
677 if (!alc_ptr)
678 return;
679 for_each_node(i)
680 kfree(alc_ptr[i]);
681 kfree(alc_ptr);
682}
683
684static void __drain_alien_cache(struct kmem_cache *cachep,
685 struct array_cache *ac, int node,
686 struct list_head *list)
687{
688 struct kmem_cache_node *n = get_node(cachep, node);
689
690 if (ac->avail) {
691 spin_lock(&n->list_lock);
692 /*
693 * Stuff objects into the remote nodes shared array first.
694 * That way we could avoid the overhead of putting the objects
695 * into the free lists and getting them back later.
696 */
697 if (n->shared)
698 transfer_objects(n->shared, ac, ac->limit);
699
700 free_block(cachep, ac->entry, ac->avail, node, list);
701 ac->avail = 0;
702 spin_unlock(&n->list_lock);
703 }
704}
705
706/*
707 * Called from cache_reap() to regularly drain alien caches round robin.
708 */
709static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
710{
711 int node = __this_cpu_read(slab_reap_node);
712
713 if (n->alien) {
714 struct alien_cache *alc = n->alien[node];
715 struct array_cache *ac;
716
717 if (alc) {
718 ac = &alc->ac;
719 if (ac->avail && spin_trylock_irq(&alc->lock)) {
720 LIST_HEAD(list);
721
722 __drain_alien_cache(cachep, ac, node, &list);
723 spin_unlock_irq(&alc->lock);
724 slabs_destroy(cachep, &list);
725 }
726 }
727 }
728}
729
730static void drain_alien_cache(struct kmem_cache *cachep,
731 struct alien_cache **alien)
732{
733 int i = 0;
734 struct alien_cache *alc;
735 struct array_cache *ac;
736 unsigned long flags;
737
738 for_each_online_node(i) {
739 alc = alien[i];
740 if (alc) {
741 LIST_HEAD(list);
742
743 ac = &alc->ac;
744 spin_lock_irqsave(&alc->lock, flags);
745 __drain_alien_cache(cachep, ac, i, &list);
746 spin_unlock_irqrestore(&alc->lock, flags);
747 slabs_destroy(cachep, &list);
748 }
749 }
750}
751
752static int __cache_free_alien(struct kmem_cache *cachep, void *objp,
753 int node, int page_node)
754{
755 struct kmem_cache_node *n;
756 struct alien_cache *alien = NULL;
757 struct array_cache *ac;
758 LIST_HEAD(list);
759
760 n = get_node(cachep, node);
761 STATS_INC_NODEFREES(cachep);
762 if (n->alien && n->alien[page_node]) {
763 alien = n->alien[page_node];
764 ac = &alien->ac;
765 spin_lock(&alien->lock);
766 if (unlikely(ac->avail == ac->limit)) {
767 STATS_INC_ACOVERFLOW(cachep);
768 __drain_alien_cache(cachep, ac, page_node, &list);
769 }
770 ac->entry[ac->avail++] = objp;
771 spin_unlock(&alien->lock);
772 slabs_destroy(cachep, &list);
773 } else {
774 n = get_node(cachep, page_node);
775 spin_lock(&n->list_lock);
776 free_block(cachep, &objp, 1, page_node, &list);
777 spin_unlock(&n->list_lock);
778 slabs_destroy(cachep, &list);
779 }
780 return 1;
781}
782
783static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
784{
785 int page_node = page_to_nid(virt_to_page(objp));
786 int node = numa_mem_id();
787 /*
788 * Make sure we are not freeing a object from another node to the array
789 * cache on this cpu.
790 */
791 if (likely(node == page_node))
792 return 0;
793
794 return __cache_free_alien(cachep, objp, node, page_node);
795}
796
797/*
798 * Construct gfp mask to allocate from a specific node but do not reclaim or
799 * warn about failures.
800 */
801static inline gfp_t gfp_exact_node(gfp_t flags)
802{
803 return (flags | __GFP_THISNODE | __GFP_NOWARN) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
804}
805#endif
806
807static int init_cache_node(struct kmem_cache *cachep, int node, gfp_t gfp)
808{
809 struct kmem_cache_node *n;
810
811 /*
812 * Set up the kmem_cache_node for cpu before we can
813 * begin anything. Make sure some other cpu on this
814 * node has not already allocated this
815 */
816 n = get_node(cachep, node);
817 if (n) {
818 spin_lock_irq(&n->list_lock);
819 n->free_limit = (1 + nr_cpus_node(node)) * cachep->batchcount +
820 cachep->num;
821 spin_unlock_irq(&n->list_lock);
822
823 return 0;
824 }
825
826 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
827 if (!n)
828 return -ENOMEM;
829
830 kmem_cache_node_init(n);
831 n->next_reap = jiffies + REAPTIMEOUT_NODE +
832 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
833
834 n->free_limit =
835 (1 + nr_cpus_node(node)) * cachep->batchcount + cachep->num;
836
837 /*
838 * The kmem_cache_nodes don't come and go as CPUs
839 * come and go. slab_mutex is sufficient
840 * protection here.
841 */
842 cachep->node[node] = n;
843
844 return 0;
845}
846
847#if (defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)) || defined(CONFIG_SMP)
848/*
849 * Allocates and initializes node for a node on each slab cache, used for
850 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
851 * will be allocated off-node since memory is not yet online for the new node.
852 * When hotplugging memory or a cpu, existing node are not replaced if
853 * already in use.
854 *
855 * Must hold slab_mutex.
856 */
857static int init_cache_node_node(int node)
858{
859 int ret;
860 struct kmem_cache *cachep;
861
862 list_for_each_entry(cachep, &slab_caches, list) {
863 ret = init_cache_node(cachep, node, GFP_KERNEL);
864 if (ret)
865 return ret;
866 }
867
868 return 0;
869}
870#endif
871
872static int setup_kmem_cache_node(struct kmem_cache *cachep,
873 int node, gfp_t gfp, bool force_change)
874{
875 int ret = -ENOMEM;
876 struct kmem_cache_node *n;
877 struct array_cache *old_shared = NULL;
878 struct array_cache *new_shared = NULL;
879 struct alien_cache **new_alien = NULL;
880 LIST_HEAD(list);
881
882 if (use_alien_caches) {
883 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
884 if (!new_alien)
885 goto fail;
886 }
887
888 if (cachep->shared) {
889 new_shared = alloc_arraycache(node,
890 cachep->shared * cachep->batchcount, 0xbaadf00d, gfp);
891 if (!new_shared)
892 goto fail;
893 }
894
895 ret = init_cache_node(cachep, node, gfp);
896 if (ret)
897 goto fail;
898
899 n = get_node(cachep, node);
900 spin_lock_irq(&n->list_lock);
901 if (n->shared && force_change) {
902 free_block(cachep, n->shared->entry,
903 n->shared->avail, node, &list);
904 n->shared->avail = 0;
905 }
906
907 if (!n->shared || force_change) {
908 old_shared = n->shared;
909 n->shared = new_shared;
910 new_shared = NULL;
911 }
912
913 if (!n->alien) {
914 n->alien = new_alien;
915 new_alien = NULL;
916 }
917
918 spin_unlock_irq(&n->list_lock);
919 slabs_destroy(cachep, &list);
920
921 /*
922 * To protect lockless access to n->shared during irq disabled context.
923 * If n->shared isn't NULL in irq disabled context, accessing to it is
924 * guaranteed to be valid until irq is re-enabled, because it will be
925 * freed after synchronize_rcu().
926 */
927 if (old_shared && force_change)
928 synchronize_rcu();
929
930fail:
931 kfree(old_shared);
932 kfree(new_shared);
933 free_alien_cache(new_alien);
934
935 return ret;
936}
937
938#ifdef CONFIG_SMP
939
940static void cpuup_canceled(long cpu)
941{
942 struct kmem_cache *cachep;
943 struct kmem_cache_node *n = NULL;
944 int node = cpu_to_mem(cpu);
945 const struct cpumask *mask = cpumask_of_node(node);
946
947 list_for_each_entry(cachep, &slab_caches, list) {
948 struct array_cache *nc;
949 struct array_cache *shared;
950 struct alien_cache **alien;
951 LIST_HEAD(list);
952
953 n = get_node(cachep, node);
954 if (!n)
955 continue;
956
957 spin_lock_irq(&n->list_lock);
958
959 /* Free limit for this kmem_cache_node */
960 n->free_limit -= cachep->batchcount;
961
962 /* cpu is dead; no one can alloc from it. */
963 nc = per_cpu_ptr(cachep->cpu_cache, cpu);
964 free_block(cachep, nc->entry, nc->avail, node, &list);
965 nc->avail = 0;
966
967 if (!cpumask_empty(mask)) {
968 spin_unlock_irq(&n->list_lock);
969 goto free_slab;
970 }
971
972 shared = n->shared;
973 if (shared) {
974 free_block(cachep, shared->entry,
975 shared->avail, node, &list);
976 n->shared = NULL;
977 }
978
979 alien = n->alien;
980 n->alien = NULL;
981
982 spin_unlock_irq(&n->list_lock);
983
984 kfree(shared);
985 if (alien) {
986 drain_alien_cache(cachep, alien);
987 free_alien_cache(alien);
988 }
989
990free_slab:
991 slabs_destroy(cachep, &list);
992 }
993 /*
994 * In the previous loop, all the objects were freed to
995 * the respective cache's slabs, now we can go ahead and
996 * shrink each nodelist to its limit.
997 */
998 list_for_each_entry(cachep, &slab_caches, list) {
999 n = get_node(cachep, node);
1000 if (!n)
1001 continue;
1002 drain_freelist(cachep, n, INT_MAX);
1003 }
1004}
1005
1006static int cpuup_prepare(long cpu)
1007{
1008 struct kmem_cache *cachep;
1009 int node = cpu_to_mem(cpu);
1010 int err;
1011
1012 /*
1013 * We need to do this right in the beginning since
1014 * alloc_arraycache's are going to use this list.
1015 * kmalloc_node allows us to add the slab to the right
1016 * kmem_cache_node and not this cpu's kmem_cache_node
1017 */
1018 err = init_cache_node_node(node);
1019 if (err < 0)
1020 goto bad;
1021
1022 /*
1023 * Now we can go ahead with allocating the shared arrays and
1024 * array caches
1025 */
1026 list_for_each_entry(cachep, &slab_caches, list) {
1027 err = setup_kmem_cache_node(cachep, node, GFP_KERNEL, false);
1028 if (err)
1029 goto bad;
1030 }
1031
1032 return 0;
1033bad:
1034 cpuup_canceled(cpu);
1035 return -ENOMEM;
1036}
1037
1038int slab_prepare_cpu(unsigned int cpu)
1039{
1040 int err;
1041
1042 mutex_lock(&slab_mutex);
1043 err = cpuup_prepare(cpu);
1044 mutex_unlock(&slab_mutex);
1045 return err;
1046}
1047
1048/*
1049 * This is called for a failed online attempt and for a successful
1050 * offline.
1051 *
1052 * Even if all the cpus of a node are down, we don't free the
1053 * kmem_list3 of any cache. This to avoid a race between cpu_down, and
1054 * a kmalloc allocation from another cpu for memory from the node of
1055 * the cpu going down. The list3 structure is usually allocated from
1056 * kmem_cache_create() and gets destroyed at kmem_cache_destroy().
1057 */
1058int slab_dead_cpu(unsigned int cpu)
1059{
1060 mutex_lock(&slab_mutex);
1061 cpuup_canceled(cpu);
1062 mutex_unlock(&slab_mutex);
1063 return 0;
1064}
1065#endif
1066
1067static int slab_online_cpu(unsigned int cpu)
1068{
1069 start_cpu_timer(cpu);
1070 return 0;
1071}
1072
1073static int slab_offline_cpu(unsigned int cpu)
1074{
1075 /*
1076 * Shutdown cache reaper. Note that the slab_mutex is held so
1077 * that if cache_reap() is invoked it cannot do anything
1078 * expensive but will only modify reap_work and reschedule the
1079 * timer.
1080 */
1081 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1082 /* Now the cache_reaper is guaranteed to be not running. */
1083 per_cpu(slab_reap_work, cpu).work.func = NULL;
1084 return 0;
1085}
1086
1087#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1088/*
1089 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1090 * Returns -EBUSY if all objects cannot be drained so that the node is not
1091 * removed.
1092 *
1093 * Must hold slab_mutex.
1094 */
1095static int __meminit drain_cache_node_node(int node)
1096{
1097 struct kmem_cache *cachep;
1098 int ret = 0;
1099
1100 list_for_each_entry(cachep, &slab_caches, list) {
1101 struct kmem_cache_node *n;
1102
1103 n = get_node(cachep, node);
1104 if (!n)
1105 continue;
1106
1107 drain_freelist(cachep, n, INT_MAX);
1108
1109 if (!list_empty(&n->slabs_full) ||
1110 !list_empty(&n->slabs_partial)) {
1111 ret = -EBUSY;
1112 break;
1113 }
1114 }
1115 return ret;
1116}
1117
1118static int __meminit slab_memory_callback(struct notifier_block *self,
1119 unsigned long action, void *arg)
1120{
1121 struct memory_notify *mnb = arg;
1122 int ret = 0;
1123 int nid;
1124
1125 nid = mnb->status_change_nid;
1126 if (nid < 0)
1127 goto out;
1128
1129 switch (action) {
1130 case MEM_GOING_ONLINE:
1131 mutex_lock(&slab_mutex);
1132 ret = init_cache_node_node(nid);
1133 mutex_unlock(&slab_mutex);
1134 break;
1135 case MEM_GOING_OFFLINE:
1136 mutex_lock(&slab_mutex);
1137 ret = drain_cache_node_node(nid);
1138 mutex_unlock(&slab_mutex);
1139 break;
1140 case MEM_ONLINE:
1141 case MEM_OFFLINE:
1142 case MEM_CANCEL_ONLINE:
1143 case MEM_CANCEL_OFFLINE:
1144 break;
1145 }
1146out:
1147 return notifier_from_errno(ret);
1148}
1149#endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1150
1151/*
1152 * swap the static kmem_cache_node with kmalloced memory
1153 */
1154static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
1155 int nodeid)
1156{
1157 struct kmem_cache_node *ptr;
1158
1159 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
1160 BUG_ON(!ptr);
1161
1162 memcpy(ptr, list, sizeof(struct kmem_cache_node));
1163 /*
1164 * Do not assume that spinlocks can be initialized via memcpy:
1165 */
1166 spin_lock_init(&ptr->list_lock);
1167
1168 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1169 cachep->node[nodeid] = ptr;
1170}
1171
1172/*
1173 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1174 * size of kmem_cache_node.
1175 */
1176static void __init set_up_node(struct kmem_cache *cachep, int index)
1177{
1178 int node;
1179
1180 for_each_online_node(node) {
1181 cachep->node[node] = &init_kmem_cache_node[index + node];
1182 cachep->node[node]->next_reap = jiffies +
1183 REAPTIMEOUT_NODE +
1184 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1185 }
1186}
1187
1188/*
1189 * Initialisation. Called after the page allocator have been initialised and
1190 * before smp_init().
1191 */
1192void __init kmem_cache_init(void)
1193{
1194 int i;
1195
1196 kmem_cache = &kmem_cache_boot;
1197
1198 if (!IS_ENABLED(CONFIG_NUMA) || num_possible_nodes() == 1)
1199 use_alien_caches = 0;
1200
1201 for (i = 0; i < NUM_INIT_LISTS; i++)
1202 kmem_cache_node_init(&init_kmem_cache_node[i]);
1203
1204 /*
1205 * Fragmentation resistance on low memory - only use bigger
1206 * page orders on machines with more than 32MB of memory if
1207 * not overridden on the command line.
1208 */
1209 if (!slab_max_order_set && totalram_pages() > (32 << 20) >> PAGE_SHIFT)
1210 slab_max_order = SLAB_MAX_ORDER_HI;
1211
1212 /* Bootstrap is tricky, because several objects are allocated
1213 * from caches that do not exist yet:
1214 * 1) initialize the kmem_cache cache: it contains the struct
1215 * kmem_cache structures of all caches, except kmem_cache itself:
1216 * kmem_cache is statically allocated.
1217 * Initially an __init data area is used for the head array and the
1218 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1219 * array at the end of the bootstrap.
1220 * 2) Create the first kmalloc cache.
1221 * The struct kmem_cache for the new cache is allocated normally.
1222 * An __init data area is used for the head array.
1223 * 3) Create the remaining kmalloc caches, with minimally sized
1224 * head arrays.
1225 * 4) Replace the __init data head arrays for kmem_cache and the first
1226 * kmalloc cache with kmalloc allocated arrays.
1227 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1228 * the other cache's with kmalloc allocated memory.
1229 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1230 */
1231
1232 /* 1) create the kmem_cache */
1233
1234 /*
1235 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1236 */
1237 create_boot_cache(kmem_cache, "kmem_cache",
1238 offsetof(struct kmem_cache, node) +
1239 nr_node_ids * sizeof(struct kmem_cache_node *),
1240 SLAB_HWCACHE_ALIGN, 0, 0);
1241 list_add(&kmem_cache->list, &slab_caches);
1242 memcg_link_cache(kmem_cache, NULL);
1243 slab_state = PARTIAL;
1244
1245 /*
1246 * Initialize the caches that provide memory for the kmem_cache_node
1247 * structures first. Without this, further allocations will bug.
1248 */
1249 kmalloc_caches[KMALLOC_NORMAL][INDEX_NODE] = create_kmalloc_cache(
1250 kmalloc_info[INDEX_NODE].name,
1251 kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS,
1252 0, kmalloc_size(INDEX_NODE));
1253 slab_state = PARTIAL_NODE;
1254 setup_kmalloc_cache_index_table();
1255
1256 slab_early_init = 0;
1257
1258 /* 5) Replace the bootstrap kmem_cache_node */
1259 {
1260 int nid;
1261
1262 for_each_online_node(nid) {
1263 init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
1264
1265 init_list(kmalloc_caches[KMALLOC_NORMAL][INDEX_NODE],
1266 &init_kmem_cache_node[SIZE_NODE + nid], nid);
1267 }
1268 }
1269
1270 create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
1271}
1272
1273void __init kmem_cache_init_late(void)
1274{
1275 struct kmem_cache *cachep;
1276
1277 /* 6) resize the head arrays to their final sizes */
1278 mutex_lock(&slab_mutex);
1279 list_for_each_entry(cachep, &slab_caches, list)
1280 if (enable_cpucache(cachep, GFP_NOWAIT))
1281 BUG();
1282 mutex_unlock(&slab_mutex);
1283
1284 /* Done! */
1285 slab_state = FULL;
1286
1287#ifdef CONFIG_NUMA
1288 /*
1289 * Register a memory hotplug callback that initializes and frees
1290 * node.
1291 */
1292 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1293#endif
1294
1295 /*
1296 * The reap timers are started later, with a module init call: That part
1297 * of the kernel is not yet operational.
1298 */
1299}
1300
1301static int __init cpucache_init(void)
1302{
1303 int ret;
1304
1305 /*
1306 * Register the timers that return unneeded pages to the page allocator
1307 */
1308 ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "SLAB online",
1309 slab_online_cpu, slab_offline_cpu);
1310 WARN_ON(ret < 0);
1311
1312 return 0;
1313}
1314__initcall(cpucache_init);
1315
1316static noinline void
1317slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1318{
1319#if DEBUG
1320 struct kmem_cache_node *n;
1321 unsigned long flags;
1322 int node;
1323 static DEFINE_RATELIMIT_STATE(slab_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1324 DEFAULT_RATELIMIT_BURST);
1325
1326 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slab_oom_rs))
1327 return;
1328
1329 pr_warn("SLAB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
1330 nodeid, gfpflags, &gfpflags);
1331 pr_warn(" cache: %s, object size: %d, order: %d\n",
1332 cachep->name, cachep->size, cachep->gfporder);
1333
1334 for_each_kmem_cache_node(cachep, node, n) {
1335 unsigned long total_slabs, free_slabs, free_objs;
1336
1337 spin_lock_irqsave(&n->list_lock, flags);
1338 total_slabs = n->total_slabs;
1339 free_slabs = n->free_slabs;
1340 free_objs = n->free_objects;
1341 spin_unlock_irqrestore(&n->list_lock, flags);
1342
1343 pr_warn(" node %d: slabs: %ld/%ld, objs: %ld/%ld\n",
1344 node, total_slabs - free_slabs, total_slabs,
1345 (total_slabs * cachep->num) - free_objs,
1346 total_slabs * cachep->num);
1347 }
1348#endif
1349}
1350
1351/*
1352 * Interface to system's page allocator. No need to hold the
1353 * kmem_cache_node ->list_lock.
1354 *
1355 * If we requested dmaable memory, we will get it. Even if we
1356 * did not request dmaable memory, we might get it, but that
1357 * would be relatively rare and ignorable.
1358 */
1359static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1360 int nodeid)
1361{
1362 struct page *page;
1363
1364 flags |= cachep->allocflags;
1365
1366 page = __alloc_pages_node(nodeid, flags, cachep->gfporder);
1367 if (!page) {
1368 slab_out_of_memory(cachep, flags, nodeid);
1369 return NULL;
1370 }
1371
1372 if (charge_slab_page(page, flags, cachep->gfporder, cachep)) {
1373 __free_pages(page, cachep->gfporder);
1374 return NULL;
1375 }
1376
1377 __SetPageSlab(page);
1378 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1379 if (sk_memalloc_socks() && page_is_pfmemalloc(page))
1380 SetPageSlabPfmemalloc(page);
1381
1382 return page;
1383}
1384
1385/*
1386 * Interface to system's page release.
1387 */
1388static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
1389{
1390 int order = cachep->gfporder;
1391
1392 BUG_ON(!PageSlab(page));
1393 __ClearPageSlabPfmemalloc(page);
1394 __ClearPageSlab(page);
1395 page_mapcount_reset(page);
1396 page->mapping = NULL;
1397
1398 if (current->reclaim_state)
1399 current->reclaim_state->reclaimed_slab += 1 << order;
1400 uncharge_slab_page(page, order, cachep);
1401 __free_pages(page, order);
1402}
1403
1404static void kmem_rcu_free(struct rcu_head *head)
1405{
1406 struct kmem_cache *cachep;
1407 struct page *page;
1408
1409 page = container_of(head, struct page, rcu_head);
1410 cachep = page->slab_cache;
1411
1412 kmem_freepages(cachep, page);
1413}
1414
1415#if DEBUG
1416static bool is_debug_pagealloc_cache(struct kmem_cache *cachep)
1417{
1418 if (debug_pagealloc_enabled() && OFF_SLAB(cachep) &&
1419 (cachep->size % PAGE_SIZE) == 0)
1420 return true;
1421
1422 return false;
1423}
1424
1425#ifdef CONFIG_DEBUG_PAGEALLOC
1426static void slab_kernel_map(struct kmem_cache *cachep, void *objp, int map)
1427{
1428 if (!is_debug_pagealloc_cache(cachep))
1429 return;
1430
1431 kernel_map_pages(virt_to_page(objp), cachep->size / PAGE_SIZE, map);
1432}
1433
1434#else
1435static inline void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1436 int map) {}
1437
1438#endif
1439
1440static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1441{
1442 int size = cachep->object_size;
1443 addr = &((char *)addr)[obj_offset(cachep)];
1444
1445 memset(addr, val, size);
1446 *(unsigned char *)(addr + size - 1) = POISON_END;
1447}
1448
1449static void dump_line(char *data, int offset, int limit)
1450{
1451 int i;
1452 unsigned char error = 0;
1453 int bad_count = 0;
1454
1455 pr_err("%03x: ", offset);
1456 for (i = 0; i < limit; i++) {
1457 if (data[offset + i] != POISON_FREE) {
1458 error = data[offset + i];
1459 bad_count++;
1460 }
1461 }
1462 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1463 &data[offset], limit, 1);
1464
1465 if (bad_count == 1) {
1466 error ^= POISON_FREE;
1467 if (!(error & (error - 1))) {
1468 pr_err("Single bit error detected. Probably bad RAM.\n");
1469#ifdef CONFIG_X86
1470 pr_err("Run memtest86+ or a similar memory test tool.\n");
1471#else
1472 pr_err("Run a memory test tool.\n");
1473#endif
1474 }
1475 }
1476}
1477#endif
1478
1479#if DEBUG
1480
1481static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1482{
1483 int i, size;
1484 char *realobj;
1485
1486 if (cachep->flags & SLAB_RED_ZONE) {
1487 pr_err("Redzone: 0x%llx/0x%llx\n",
1488 *dbg_redzone1(cachep, objp),
1489 *dbg_redzone2(cachep, objp));
1490 }
1491
1492 if (cachep->flags & SLAB_STORE_USER)
1493 pr_err("Last user: (%pSR)\n", *dbg_userword(cachep, objp));
1494 realobj = (char *)objp + obj_offset(cachep);
1495 size = cachep->object_size;
1496 for (i = 0; i < size && lines; i += 16, lines--) {
1497 int limit;
1498 limit = 16;
1499 if (i + limit > size)
1500 limit = size - i;
1501 dump_line(realobj, i, limit);
1502 }
1503}
1504
1505static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1506{
1507 char *realobj;
1508 int size, i;
1509 int lines = 0;
1510
1511 if (is_debug_pagealloc_cache(cachep))
1512 return;
1513
1514 realobj = (char *)objp + obj_offset(cachep);
1515 size = cachep->object_size;
1516
1517 for (i = 0; i < size; i++) {
1518 char exp = POISON_FREE;
1519 if (i == size - 1)
1520 exp = POISON_END;
1521 if (realobj[i] != exp) {
1522 int limit;
1523 /* Mismatch ! */
1524 /* Print header */
1525 if (lines == 0) {
1526 pr_err("Slab corruption (%s): %s start=%px, len=%d\n",
1527 print_tainted(), cachep->name,
1528 realobj, size);
1529 print_objinfo(cachep, objp, 0);
1530 }
1531 /* Hexdump the affected line */
1532 i = (i / 16) * 16;
1533 limit = 16;
1534 if (i + limit > size)
1535 limit = size - i;
1536 dump_line(realobj, i, limit);
1537 i += 16;
1538 lines++;
1539 /* Limit to 5 lines */
1540 if (lines > 5)
1541 break;
1542 }
1543 }
1544 if (lines != 0) {
1545 /* Print some data about the neighboring objects, if they
1546 * exist:
1547 */
1548 struct page *page = virt_to_head_page(objp);
1549 unsigned int objnr;
1550
1551 objnr = obj_to_index(cachep, page, objp);
1552 if (objnr) {
1553 objp = index_to_obj(cachep, page, objnr - 1);
1554 realobj = (char *)objp + obj_offset(cachep);
1555 pr_err("Prev obj: start=%px, len=%d\n", realobj, size);
1556 print_objinfo(cachep, objp, 2);
1557 }
1558 if (objnr + 1 < cachep->num) {
1559 objp = index_to_obj(cachep, page, objnr + 1);
1560 realobj = (char *)objp + obj_offset(cachep);
1561 pr_err("Next obj: start=%px, len=%d\n", realobj, size);
1562 print_objinfo(cachep, objp, 2);
1563 }
1564 }
1565}
1566#endif
1567
1568#if DEBUG
1569static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1570 struct page *page)
1571{
1572 int i;
1573
1574 if (OBJFREELIST_SLAB(cachep) && cachep->flags & SLAB_POISON) {
1575 poison_obj(cachep, page->freelist - obj_offset(cachep),
1576 POISON_FREE);
1577 }
1578
1579 for (i = 0; i < cachep->num; i++) {
1580 void *objp = index_to_obj(cachep, page, i);
1581
1582 if (cachep->flags & SLAB_POISON) {
1583 check_poison_obj(cachep, objp);
1584 slab_kernel_map(cachep, objp, 1);
1585 }
1586 if (cachep->flags & SLAB_RED_ZONE) {
1587 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1588 slab_error(cachep, "start of a freed object was overwritten");
1589 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1590 slab_error(cachep, "end of a freed object was overwritten");
1591 }
1592 }
1593}
1594#else
1595static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1596 struct page *page)
1597{
1598}
1599#endif
1600
1601/**
1602 * slab_destroy - destroy and release all objects in a slab
1603 * @cachep: cache pointer being destroyed
1604 * @page: page pointer being destroyed
1605 *
1606 * Destroy all the objs in a slab page, and release the mem back to the system.
1607 * Before calling the slab page must have been unlinked from the cache. The
1608 * kmem_cache_node ->list_lock is not held/needed.
1609 */
1610static void slab_destroy(struct kmem_cache *cachep, struct page *page)
1611{
1612 void *freelist;
1613
1614 freelist = page->freelist;
1615 slab_destroy_debugcheck(cachep, page);
1616 if (unlikely(cachep->flags & SLAB_TYPESAFE_BY_RCU))
1617 call_rcu(&page->rcu_head, kmem_rcu_free);
1618 else
1619 kmem_freepages(cachep, page);
1620
1621 /*
1622 * From now on, we don't use freelist
1623 * although actual page can be freed in rcu context
1624 */
1625 if (OFF_SLAB(cachep))
1626 kmem_cache_free(cachep->freelist_cache, freelist);
1627}
1628
1629static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list)
1630{
1631 struct page *page, *n;
1632
1633 list_for_each_entry_safe(page, n, list, slab_list) {
1634 list_del(&page->slab_list);
1635 slab_destroy(cachep, page);
1636 }
1637}
1638
1639/**
1640 * calculate_slab_order - calculate size (page order) of slabs
1641 * @cachep: pointer to the cache that is being created
1642 * @size: size of objects to be created in this cache.
1643 * @flags: slab allocation flags
1644 *
1645 * Also calculates the number of objects per slab.
1646 *
1647 * This could be made much more intelligent. For now, try to avoid using
1648 * high order pages for slabs. When the gfp() functions are more friendly
1649 * towards high-order requests, this should be changed.
1650 *
1651 * Return: number of left-over bytes in a slab
1652 */
1653static size_t calculate_slab_order(struct kmem_cache *cachep,
1654 size_t size, slab_flags_t flags)
1655{
1656 size_t left_over = 0;
1657 int gfporder;
1658
1659 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
1660 unsigned int num;
1661 size_t remainder;
1662
1663 num = cache_estimate(gfporder, size, flags, &remainder);
1664 if (!num)
1665 continue;
1666
1667 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1668 if (num > SLAB_OBJ_MAX_NUM)
1669 break;
1670
1671 if (flags & CFLGS_OFF_SLAB) {
1672 struct kmem_cache *freelist_cache;
1673 size_t freelist_size;
1674
1675 freelist_size = num * sizeof(freelist_idx_t);
1676 freelist_cache = kmalloc_slab(freelist_size, 0u);
1677 if (!freelist_cache)
1678 continue;
1679
1680 /*
1681 * Needed to avoid possible looping condition
1682 * in cache_grow_begin()
1683 */
1684 if (OFF_SLAB(freelist_cache))
1685 continue;
1686
1687 /* check if off slab has enough benefit */
1688 if (freelist_cache->size > cachep->size / 2)
1689 continue;
1690 }
1691
1692 /* Found something acceptable - save it away */
1693 cachep->num = num;
1694 cachep->gfporder = gfporder;
1695 left_over = remainder;
1696
1697 /*
1698 * A VFS-reclaimable slab tends to have most allocations
1699 * as GFP_NOFS and we really don't want to have to be allocating
1700 * higher-order pages when we are unable to shrink dcache.
1701 */
1702 if (flags & SLAB_RECLAIM_ACCOUNT)
1703 break;
1704
1705 /*
1706 * Large number of objects is good, but very large slabs are
1707 * currently bad for the gfp()s.
1708 */
1709 if (gfporder >= slab_max_order)
1710 break;
1711
1712 /*
1713 * Acceptable internal fragmentation?
1714 */
1715 if (left_over * 8 <= (PAGE_SIZE << gfporder))
1716 break;
1717 }
1718 return left_over;
1719}
1720
1721static struct array_cache __percpu *alloc_kmem_cache_cpus(
1722 struct kmem_cache *cachep, int entries, int batchcount)
1723{
1724 int cpu;
1725 size_t size;
1726 struct array_cache __percpu *cpu_cache;
1727
1728 size = sizeof(void *) * entries + sizeof(struct array_cache);
1729 cpu_cache = __alloc_percpu(size, sizeof(void *));
1730
1731 if (!cpu_cache)
1732 return NULL;
1733
1734 for_each_possible_cpu(cpu) {
1735 init_arraycache(per_cpu_ptr(cpu_cache, cpu),
1736 entries, batchcount);
1737 }
1738
1739 return cpu_cache;
1740}
1741
1742static int __ref setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
1743{
1744 if (slab_state >= FULL)
1745 return enable_cpucache(cachep, gfp);
1746
1747 cachep->cpu_cache = alloc_kmem_cache_cpus(cachep, 1, 1);
1748 if (!cachep->cpu_cache)
1749 return 1;
1750
1751 if (slab_state == DOWN) {
1752 /* Creation of first cache (kmem_cache). */
1753 set_up_node(kmem_cache, CACHE_CACHE);
1754 } else if (slab_state == PARTIAL) {
1755 /* For kmem_cache_node */
1756 set_up_node(cachep, SIZE_NODE);
1757 } else {
1758 int node;
1759
1760 for_each_online_node(node) {
1761 cachep->node[node] = kmalloc_node(
1762 sizeof(struct kmem_cache_node), gfp, node);
1763 BUG_ON(!cachep->node[node]);
1764 kmem_cache_node_init(cachep->node[node]);
1765 }
1766 }
1767
1768 cachep->node[numa_mem_id()]->next_reap =
1769 jiffies + REAPTIMEOUT_NODE +
1770 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1771
1772 cpu_cache_get(cachep)->avail = 0;
1773 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1774 cpu_cache_get(cachep)->batchcount = 1;
1775 cpu_cache_get(cachep)->touched = 0;
1776 cachep->batchcount = 1;
1777 cachep->limit = BOOT_CPUCACHE_ENTRIES;
1778 return 0;
1779}
1780
1781slab_flags_t kmem_cache_flags(unsigned int object_size,
1782 slab_flags_t flags, const char *name,
1783 void (*ctor)(void *))
1784{
1785 return flags;
1786}
1787
1788struct kmem_cache *
1789__kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
1790 slab_flags_t flags, void (*ctor)(void *))
1791{
1792 struct kmem_cache *cachep;
1793
1794 cachep = find_mergeable(size, align, flags, name, ctor);
1795 if (cachep) {
1796 cachep->refcount++;
1797
1798 /*
1799 * Adjust the object sizes so that we clear
1800 * the complete object on kzalloc.
1801 */
1802 cachep->object_size = max_t(int, cachep->object_size, size);
1803 }
1804 return cachep;
1805}
1806
1807static bool set_objfreelist_slab_cache(struct kmem_cache *cachep,
1808 size_t size, slab_flags_t flags)
1809{
1810 size_t left;
1811
1812 cachep->num = 0;
1813
1814 /*
1815 * If slab auto-initialization on free is enabled, store the freelist
1816 * off-slab, so that its contents don't end up in one of the allocated
1817 * objects.
1818 */
1819 if (unlikely(slab_want_init_on_free(cachep)))
1820 return false;
1821
1822 if (cachep->ctor || flags & SLAB_TYPESAFE_BY_RCU)
1823 return false;
1824
1825 left = calculate_slab_order(cachep, size,
1826 flags | CFLGS_OBJFREELIST_SLAB);
1827 if (!cachep->num)
1828 return false;
1829
1830 if (cachep->num * sizeof(freelist_idx_t) > cachep->object_size)
1831 return false;
1832
1833 cachep->colour = left / cachep->colour_off;
1834
1835 return true;
1836}
1837
1838static bool set_off_slab_cache(struct kmem_cache *cachep,
1839 size_t size, slab_flags_t flags)
1840{
1841 size_t left;
1842
1843 cachep->num = 0;
1844
1845 /*
1846 * Always use on-slab management when SLAB_NOLEAKTRACE
1847 * to avoid recursive calls into kmemleak.
1848 */
1849 if (flags & SLAB_NOLEAKTRACE)
1850 return false;
1851
1852 /*
1853 * Size is large, assume best to place the slab management obj
1854 * off-slab (should allow better packing of objs).
1855 */
1856 left = calculate_slab_order(cachep, size, flags | CFLGS_OFF_SLAB);
1857 if (!cachep->num)
1858 return false;
1859
1860 /*
1861 * If the slab has been placed off-slab, and we have enough space then
1862 * move it on-slab. This is at the expense of any extra colouring.
1863 */
1864 if (left >= cachep->num * sizeof(freelist_idx_t))
1865 return false;
1866
1867 cachep->colour = left / cachep->colour_off;
1868
1869 return true;
1870}
1871
1872static bool set_on_slab_cache(struct kmem_cache *cachep,
1873 size_t size, slab_flags_t flags)
1874{
1875 size_t left;
1876
1877 cachep->num = 0;
1878
1879 left = calculate_slab_order(cachep, size, flags);
1880 if (!cachep->num)
1881 return false;
1882
1883 cachep->colour = left / cachep->colour_off;
1884
1885 return true;
1886}
1887
1888/**
1889 * __kmem_cache_create - Create a cache.
1890 * @cachep: cache management descriptor
1891 * @flags: SLAB flags
1892 *
1893 * Returns a ptr to the cache on success, NULL on failure.
1894 * Cannot be called within a int, but can be interrupted.
1895 * The @ctor is run when new pages are allocated by the cache.
1896 *
1897 * The flags are
1898 *
1899 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1900 * to catch references to uninitialised memory.
1901 *
1902 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1903 * for buffer overruns.
1904 *
1905 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1906 * cacheline. This can be beneficial if you're counting cycles as closely
1907 * as davem.
1908 *
1909 * Return: a pointer to the created cache or %NULL in case of error
1910 */
1911int __kmem_cache_create(struct kmem_cache *cachep, slab_flags_t flags)
1912{
1913 size_t ralign = BYTES_PER_WORD;
1914 gfp_t gfp;
1915 int err;
1916 unsigned int size = cachep->size;
1917
1918#if DEBUG
1919#if FORCED_DEBUG
1920 /*
1921 * Enable redzoning and last user accounting, except for caches with
1922 * large objects, if the increased size would increase the object size
1923 * above the next power of two: caches with object sizes just above a
1924 * power of two have a significant amount of internal fragmentation.
1925 */
1926 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
1927 2 * sizeof(unsigned long long)))
1928 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
1929 if (!(flags & SLAB_TYPESAFE_BY_RCU))
1930 flags |= SLAB_POISON;
1931#endif
1932#endif
1933
1934 /*
1935 * Check that size is in terms of words. This is needed to avoid
1936 * unaligned accesses for some archs when redzoning is used, and makes
1937 * sure any on-slab bufctl's are also correctly aligned.
1938 */
1939 size = ALIGN(size, BYTES_PER_WORD);
1940
1941 if (flags & SLAB_RED_ZONE) {
1942 ralign = REDZONE_ALIGN;
1943 /* If redzoning, ensure that the second redzone is suitably
1944 * aligned, by adjusting the object size accordingly. */
1945 size = ALIGN(size, REDZONE_ALIGN);
1946 }
1947
1948 /* 3) caller mandated alignment */
1949 if (ralign < cachep->align) {
1950 ralign = cachep->align;
1951 }
1952 /* disable debug if necessary */
1953 if (ralign > __alignof__(unsigned long long))
1954 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
1955 /*
1956 * 4) Store it.
1957 */
1958 cachep->align = ralign;
1959 cachep->colour_off = cache_line_size();
1960 /* Offset must be a multiple of the alignment. */
1961 if (cachep->colour_off < cachep->align)
1962 cachep->colour_off = cachep->align;
1963
1964 if (slab_is_available())
1965 gfp = GFP_KERNEL;
1966 else
1967 gfp = GFP_NOWAIT;
1968
1969#if DEBUG
1970
1971 /*
1972 * Both debugging options require word-alignment which is calculated
1973 * into align above.
1974 */
1975 if (flags & SLAB_RED_ZONE) {
1976 /* add space for red zone words */
1977 cachep->obj_offset += sizeof(unsigned long long);
1978 size += 2 * sizeof(unsigned long long);
1979 }
1980 if (flags & SLAB_STORE_USER) {
1981 /* user store requires one word storage behind the end of
1982 * the real object. But if the second red zone needs to be
1983 * aligned to 64 bits, we must allow that much space.
1984 */
1985 if (flags & SLAB_RED_ZONE)
1986 size += REDZONE_ALIGN;
1987 else
1988 size += BYTES_PER_WORD;
1989 }
1990#endif
1991
1992 kasan_cache_create(cachep, &size, &flags);
1993
1994 size = ALIGN(size, cachep->align);
1995 /*
1996 * We should restrict the number of objects in a slab to implement
1997 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
1998 */
1999 if (FREELIST_BYTE_INDEX && size < SLAB_OBJ_MIN_SIZE)
2000 size = ALIGN(SLAB_OBJ_MIN_SIZE, cachep->align);
2001
2002#if DEBUG
2003 /*
2004 * To activate debug pagealloc, off-slab management is necessary
2005 * requirement. In early phase of initialization, small sized slab
2006 * doesn't get initialized so it would not be possible. So, we need
2007 * to check size >= 256. It guarantees that all necessary small
2008 * sized slab is initialized in current slab initialization sequence.
2009 */
2010 if (debug_pagealloc_enabled() && (flags & SLAB_POISON) &&
2011 size >= 256 && cachep->object_size > cache_line_size()) {
2012 if (size < PAGE_SIZE || size % PAGE_SIZE == 0) {
2013 size_t tmp_size = ALIGN(size, PAGE_SIZE);
2014
2015 if (set_off_slab_cache(cachep, tmp_size, flags)) {
2016 flags |= CFLGS_OFF_SLAB;
2017 cachep->obj_offset += tmp_size - size;
2018 size = tmp_size;
2019 goto done;
2020 }
2021 }
2022 }
2023#endif
2024
2025 if (set_objfreelist_slab_cache(cachep, size, flags)) {
2026 flags |= CFLGS_OBJFREELIST_SLAB;
2027 goto done;
2028 }
2029
2030 if (set_off_slab_cache(cachep, size, flags)) {
2031 flags |= CFLGS_OFF_SLAB;
2032 goto done;
2033 }
2034
2035 if (set_on_slab_cache(cachep, size, flags))
2036 goto done;
2037
2038 return -E2BIG;
2039
2040done:
2041 cachep->freelist_size = cachep->num * sizeof(freelist_idx_t);
2042 cachep->flags = flags;
2043 cachep->allocflags = __GFP_COMP;
2044 if (flags & SLAB_CACHE_DMA)
2045 cachep->allocflags |= GFP_DMA;
2046 if (flags & SLAB_CACHE_DMA32)
2047 cachep->allocflags |= GFP_DMA32;
2048 if (flags & SLAB_RECLAIM_ACCOUNT)
2049 cachep->allocflags |= __GFP_RECLAIMABLE;
2050 cachep->size = size;
2051 cachep->reciprocal_buffer_size = reciprocal_value(size);
2052
2053#if DEBUG
2054 /*
2055 * If we're going to use the generic kernel_map_pages()
2056 * poisoning, then it's going to smash the contents of
2057 * the redzone and userword anyhow, so switch them off.
2058 */
2059 if (IS_ENABLED(CONFIG_PAGE_POISONING) &&
2060 (cachep->flags & SLAB_POISON) &&
2061 is_debug_pagealloc_cache(cachep))
2062 cachep->flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2063#endif
2064
2065 if (OFF_SLAB(cachep)) {
2066 cachep->freelist_cache =
2067 kmalloc_slab(cachep->freelist_size, 0u);
2068 }
2069
2070 err = setup_cpu_cache(cachep, gfp);
2071 if (err) {
2072 __kmem_cache_release(cachep);
2073 return err;
2074 }
2075
2076 return 0;
2077}
2078
2079#if DEBUG
2080static void check_irq_off(void)
2081{
2082 BUG_ON(!irqs_disabled());
2083}
2084
2085static void check_irq_on(void)
2086{
2087 BUG_ON(irqs_disabled());
2088}
2089
2090static void check_mutex_acquired(void)
2091{
2092 BUG_ON(!mutex_is_locked(&slab_mutex));
2093}
2094
2095static void check_spinlock_acquired(struct kmem_cache *cachep)
2096{
2097#ifdef CONFIG_SMP
2098 check_irq_off();
2099 assert_spin_locked(&get_node(cachep, numa_mem_id())->list_lock);
2100#endif
2101}
2102
2103static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2104{
2105#ifdef CONFIG_SMP
2106 check_irq_off();
2107 assert_spin_locked(&get_node(cachep, node)->list_lock);
2108#endif
2109}
2110
2111#else
2112#define check_irq_off() do { } while(0)
2113#define check_irq_on() do { } while(0)
2114#define check_mutex_acquired() do { } while(0)
2115#define check_spinlock_acquired(x) do { } while(0)
2116#define check_spinlock_acquired_node(x, y) do { } while(0)
2117#endif
2118
2119static void drain_array_locked(struct kmem_cache *cachep, struct array_cache *ac,
2120 int node, bool free_all, struct list_head *list)
2121{
2122 int tofree;
2123
2124 if (!ac || !ac->avail)
2125 return;
2126
2127 tofree = free_all ? ac->avail : (ac->limit + 4) / 5;
2128 if (tofree > ac->avail)
2129 tofree = (ac->avail + 1) / 2;
2130
2131 free_block(cachep, ac->entry, tofree, node, list);
2132 ac->avail -= tofree;
2133 memmove(ac->entry, &(ac->entry[tofree]), sizeof(void *) * ac->avail);
2134}
2135
2136static void do_drain(void *arg)
2137{
2138 struct kmem_cache *cachep = arg;
2139 struct array_cache *ac;
2140 int node = numa_mem_id();
2141 struct kmem_cache_node *n;
2142 LIST_HEAD(list);
2143
2144 check_irq_off();
2145 ac = cpu_cache_get(cachep);
2146 n = get_node(cachep, node);
2147 spin_lock(&n->list_lock);
2148 free_block(cachep, ac->entry, ac->avail, node, &list);
2149 spin_unlock(&n->list_lock);
2150 slabs_destroy(cachep, &list);
2151 ac->avail = 0;
2152}
2153
2154static void drain_cpu_caches(struct kmem_cache *cachep)
2155{
2156 struct kmem_cache_node *n;
2157 int node;
2158 LIST_HEAD(list);
2159
2160 on_each_cpu(do_drain, cachep, 1);
2161 check_irq_on();
2162 for_each_kmem_cache_node(cachep, node, n)
2163 if (n->alien)
2164 drain_alien_cache(cachep, n->alien);
2165
2166 for_each_kmem_cache_node(cachep, node, n) {
2167 spin_lock_irq(&n->list_lock);
2168 drain_array_locked(cachep, n->shared, node, true, &list);
2169 spin_unlock_irq(&n->list_lock);
2170
2171 slabs_destroy(cachep, &list);
2172 }
2173}
2174
2175/*
2176 * Remove slabs from the list of free slabs.
2177 * Specify the number of slabs to drain in tofree.
2178 *
2179 * Returns the actual number of slabs released.
2180 */
2181static int drain_freelist(struct kmem_cache *cache,
2182 struct kmem_cache_node *n, int tofree)
2183{
2184 struct list_head *p;
2185 int nr_freed;
2186 struct page *page;
2187
2188 nr_freed = 0;
2189 while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
2190
2191 spin_lock_irq(&n->list_lock);
2192 p = n->slabs_free.prev;
2193 if (p == &n->slabs_free) {
2194 spin_unlock_irq(&n->list_lock);
2195 goto out;
2196 }
2197
2198 page = list_entry(p, struct page, slab_list);
2199 list_del(&page->slab_list);
2200 n->free_slabs--;
2201 n->total_slabs--;
2202 /*
2203 * Safe to drop the lock. The slab is no longer linked
2204 * to the cache.
2205 */
2206 n->free_objects -= cache->num;
2207 spin_unlock_irq(&n->list_lock);
2208 slab_destroy(cache, page);
2209 nr_freed++;
2210 }
2211out:
2212 return nr_freed;
2213}
2214
2215bool __kmem_cache_empty(struct kmem_cache *s)
2216{
2217 int node;
2218 struct kmem_cache_node *n;
2219
2220 for_each_kmem_cache_node(s, node, n)
2221 if (!list_empty(&n->slabs_full) ||
2222 !list_empty(&n->slabs_partial))
2223 return false;
2224 return true;
2225}
2226
2227int __kmem_cache_shrink(struct kmem_cache *cachep)
2228{
2229 int ret = 0;
2230 int node;
2231 struct kmem_cache_node *n;
2232
2233 drain_cpu_caches(cachep);
2234
2235 check_irq_on();
2236 for_each_kmem_cache_node(cachep, node, n) {
2237 drain_freelist(cachep, n, INT_MAX);
2238
2239 ret += !list_empty(&n->slabs_full) ||
2240 !list_empty(&n->slabs_partial);
2241 }
2242 return (ret ? 1 : 0);
2243}
2244
2245#ifdef CONFIG_MEMCG
2246void __kmemcg_cache_deactivate(struct kmem_cache *cachep)
2247{
2248 __kmem_cache_shrink(cachep);
2249}
2250
2251void __kmemcg_cache_deactivate_after_rcu(struct kmem_cache *s)
2252{
2253}
2254#endif
2255
2256int __kmem_cache_shutdown(struct kmem_cache *cachep)
2257{
2258 return __kmem_cache_shrink(cachep);
2259}
2260
2261void __kmem_cache_release(struct kmem_cache *cachep)
2262{
2263 int i;
2264 struct kmem_cache_node *n;
2265
2266 cache_random_seq_destroy(cachep);
2267
2268 free_percpu(cachep->cpu_cache);
2269
2270 /* NUMA: free the node structures */
2271 for_each_kmem_cache_node(cachep, i, n) {
2272 kfree(n->shared);
2273 free_alien_cache(n->alien);
2274 kfree(n);
2275 cachep->node[i] = NULL;
2276 }
2277}
2278
2279/*
2280 * Get the memory for a slab management obj.
2281 *
2282 * For a slab cache when the slab descriptor is off-slab, the
2283 * slab descriptor can't come from the same cache which is being created,
2284 * Because if it is the case, that means we defer the creation of
2285 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2286 * And we eventually call down to __kmem_cache_create(), which
2287 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2288 * This is a "chicken-and-egg" problem.
2289 *
2290 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2291 * which are all initialized during kmem_cache_init().
2292 */
2293static void *alloc_slabmgmt(struct kmem_cache *cachep,
2294 struct page *page, int colour_off,
2295 gfp_t local_flags, int nodeid)
2296{
2297 void *freelist;
2298 void *addr = page_address(page);
2299
2300 page->s_mem = addr + colour_off;
2301 page->active = 0;
2302
2303 if (OBJFREELIST_SLAB(cachep))
2304 freelist = NULL;
2305 else if (OFF_SLAB(cachep)) {
2306 /* Slab management obj is off-slab. */
2307 freelist = kmem_cache_alloc_node(cachep->freelist_cache,
2308 local_flags, nodeid);
2309 if (!freelist)
2310 return NULL;
2311 } else {
2312 /* We will use last bytes at the slab for freelist */
2313 freelist = addr + (PAGE_SIZE << cachep->gfporder) -
2314 cachep->freelist_size;
2315 }
2316
2317 return freelist;
2318}
2319
2320static inline freelist_idx_t get_free_obj(struct page *page, unsigned int idx)
2321{
2322 return ((freelist_idx_t *)page->freelist)[idx];
2323}
2324
2325static inline void set_free_obj(struct page *page,
2326 unsigned int idx, freelist_idx_t val)
2327{
2328 ((freelist_idx_t *)(page->freelist))[idx] = val;
2329}
2330
2331static void cache_init_objs_debug(struct kmem_cache *cachep, struct page *page)
2332{
2333#if DEBUG
2334 int i;
2335
2336 for (i = 0; i < cachep->num; i++) {
2337 void *objp = index_to_obj(cachep, page, i);
2338
2339 if (cachep->flags & SLAB_STORE_USER)
2340 *dbg_userword(cachep, objp) = NULL;
2341
2342 if (cachep->flags & SLAB_RED_ZONE) {
2343 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2344 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2345 }
2346 /*
2347 * Constructors are not allowed to allocate memory from the same
2348 * cache which they are a constructor for. Otherwise, deadlock.
2349 * They must also be threaded.
2350 */
2351 if (cachep->ctor && !(cachep->flags & SLAB_POISON)) {
2352 kasan_unpoison_object_data(cachep,
2353 objp + obj_offset(cachep));
2354 cachep->ctor(objp + obj_offset(cachep));
2355 kasan_poison_object_data(
2356 cachep, objp + obj_offset(cachep));
2357 }
2358
2359 if (cachep->flags & SLAB_RED_ZONE) {
2360 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2361 slab_error(cachep, "constructor overwrote the end of an object");
2362 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2363 slab_error(cachep, "constructor overwrote the start of an object");
2364 }
2365 /* need to poison the objs? */
2366 if (cachep->flags & SLAB_POISON) {
2367 poison_obj(cachep, objp, POISON_FREE);
2368 slab_kernel_map(cachep, objp, 0);
2369 }
2370 }
2371#endif
2372}
2373
2374#ifdef CONFIG_SLAB_FREELIST_RANDOM
2375/* Hold information during a freelist initialization */
2376union freelist_init_state {
2377 struct {
2378 unsigned int pos;
2379 unsigned int *list;
2380 unsigned int count;
2381 };
2382 struct rnd_state rnd_state;
2383};
2384
2385/*
2386 * Initialize the state based on the randomization methode available.
2387 * return true if the pre-computed list is available, false otherwize.
2388 */
2389static bool freelist_state_initialize(union freelist_init_state *state,
2390 struct kmem_cache *cachep,
2391 unsigned int count)
2392{
2393 bool ret;
2394 unsigned int rand;
2395
2396 /* Use best entropy available to define a random shift */
2397 rand = get_random_int();
2398
2399 /* Use a random state if the pre-computed list is not available */
2400 if (!cachep->random_seq) {
2401 prandom_seed_state(&state->rnd_state, rand);
2402 ret = false;
2403 } else {
2404 state->list = cachep->random_seq;
2405 state->count = count;
2406 state->pos = rand % count;
2407 ret = true;
2408 }
2409 return ret;
2410}
2411
2412/* Get the next entry on the list and randomize it using a random shift */
2413static freelist_idx_t next_random_slot(union freelist_init_state *state)
2414{
2415 if (state->pos >= state->count)
2416 state->pos = 0;
2417 return state->list[state->pos++];
2418}
2419
2420/* Swap two freelist entries */
2421static void swap_free_obj(struct page *page, unsigned int a, unsigned int b)
2422{
2423 swap(((freelist_idx_t *)page->freelist)[a],
2424 ((freelist_idx_t *)page->freelist)[b]);
2425}
2426
2427/*
2428 * Shuffle the freelist initialization state based on pre-computed lists.
2429 * return true if the list was successfully shuffled, false otherwise.
2430 */
2431static bool shuffle_freelist(struct kmem_cache *cachep, struct page *page)
2432{
2433 unsigned int objfreelist = 0, i, rand, count = cachep->num;
2434 union freelist_init_state state;
2435 bool precomputed;
2436
2437 if (count < 2)
2438 return false;
2439
2440 precomputed = freelist_state_initialize(&state, cachep, count);
2441
2442 /* Take a random entry as the objfreelist */
2443 if (OBJFREELIST_SLAB(cachep)) {
2444 if (!precomputed)
2445 objfreelist = count - 1;
2446 else
2447 objfreelist = next_random_slot(&state);
2448 page->freelist = index_to_obj(cachep, page, objfreelist) +
2449 obj_offset(cachep);
2450 count--;
2451 }
2452
2453 /*
2454 * On early boot, generate the list dynamically.
2455 * Later use a pre-computed list for speed.
2456 */
2457 if (!precomputed) {
2458 for (i = 0; i < count; i++)
2459 set_free_obj(page, i, i);
2460
2461 /* Fisher-Yates shuffle */
2462 for (i = count - 1; i > 0; i--) {
2463 rand = prandom_u32_state(&state.rnd_state);
2464 rand %= (i + 1);
2465 swap_free_obj(page, i, rand);
2466 }
2467 } else {
2468 for (i = 0; i < count; i++)
2469 set_free_obj(page, i, next_random_slot(&state));
2470 }
2471
2472 if (OBJFREELIST_SLAB(cachep))
2473 set_free_obj(page, cachep->num - 1, objfreelist);
2474
2475 return true;
2476}
2477#else
2478static inline bool shuffle_freelist(struct kmem_cache *cachep,
2479 struct page *page)
2480{
2481 return false;
2482}
2483#endif /* CONFIG_SLAB_FREELIST_RANDOM */
2484
2485static void cache_init_objs(struct kmem_cache *cachep,
2486 struct page *page)
2487{
2488 int i;
2489 void *objp;
2490 bool shuffled;
2491
2492 cache_init_objs_debug(cachep, page);
2493
2494 /* Try to randomize the freelist if enabled */
2495 shuffled = shuffle_freelist(cachep, page);
2496
2497 if (!shuffled && OBJFREELIST_SLAB(cachep)) {
2498 page->freelist = index_to_obj(cachep, page, cachep->num - 1) +
2499 obj_offset(cachep);
2500 }
2501
2502 for (i = 0; i < cachep->num; i++) {
2503 objp = index_to_obj(cachep, page, i);
2504 objp = kasan_init_slab_obj(cachep, objp);
2505
2506 /* constructor could break poison info */
2507 if (DEBUG == 0 && cachep->ctor) {
2508 kasan_unpoison_object_data(cachep, objp);
2509 cachep->ctor(objp);
2510 kasan_poison_object_data(cachep, objp);
2511 }
2512
2513 if (!shuffled)
2514 set_free_obj(page, i, i);
2515 }
2516}
2517
2518static void *slab_get_obj(struct kmem_cache *cachep, struct page *page)
2519{
2520 void *objp;
2521
2522 objp = index_to_obj(cachep, page, get_free_obj(page, page->active));
2523 page->active++;
2524
2525 return objp;
2526}
2527
2528static void slab_put_obj(struct kmem_cache *cachep,
2529 struct page *page, void *objp)
2530{
2531 unsigned int objnr = obj_to_index(cachep, page, objp);
2532#if DEBUG
2533 unsigned int i;
2534
2535 /* Verify double free bug */
2536 for (i = page->active; i < cachep->num; i++) {
2537 if (get_free_obj(page, i) == objnr) {
2538 pr_err("slab: double free detected in cache '%s', objp %px\n",
2539 cachep->name, objp);
2540 BUG();
2541 }
2542 }
2543#endif
2544 page->active--;
2545 if (!page->freelist)
2546 page->freelist = objp + obj_offset(cachep);
2547
2548 set_free_obj(page, page->active, objnr);
2549}
2550
2551/*
2552 * Map pages beginning at addr to the given cache and slab. This is required
2553 * for the slab allocator to be able to lookup the cache and slab of a
2554 * virtual address for kfree, ksize, and slab debugging.
2555 */
2556static void slab_map_pages(struct kmem_cache *cache, struct page *page,
2557 void *freelist)
2558{
2559 page->slab_cache = cache;
2560 page->freelist = freelist;
2561}
2562
2563/*
2564 * Grow (by 1) the number of slabs within a cache. This is called by
2565 * kmem_cache_alloc() when there are no active objs left in a cache.
2566 */
2567static struct page *cache_grow_begin(struct kmem_cache *cachep,
2568 gfp_t flags, int nodeid)
2569{
2570 void *freelist;
2571 size_t offset;
2572 gfp_t local_flags;
2573 int page_node;
2574 struct kmem_cache_node *n;
2575 struct page *page;
2576
2577 /*
2578 * Be lazy and only check for valid flags here, keeping it out of the
2579 * critical path in kmem_cache_alloc().
2580 */
2581 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
2582 gfp_t invalid_mask = flags & GFP_SLAB_BUG_MASK;
2583 flags &= ~GFP_SLAB_BUG_MASK;
2584 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n",
2585 invalid_mask, &invalid_mask, flags, &flags);
2586 dump_stack();
2587 }
2588 WARN_ON_ONCE(cachep->ctor && (flags & __GFP_ZERO));
2589 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2590
2591 check_irq_off();
2592 if (gfpflags_allow_blocking(local_flags))
2593 local_irq_enable();
2594
2595 /*
2596 * Get mem for the objs. Attempt to allocate a physical page from
2597 * 'nodeid'.
2598 */
2599 page = kmem_getpages(cachep, local_flags, nodeid);
2600 if (!page)
2601 goto failed;
2602
2603 page_node = page_to_nid(page);
2604 n = get_node(cachep, page_node);
2605
2606 /* Get colour for the slab, and cal the next value. */
2607 n->colour_next++;
2608 if (n->colour_next >= cachep->colour)
2609 n->colour_next = 0;
2610
2611 offset = n->colour_next;
2612 if (offset >= cachep->colour)
2613 offset = 0;
2614
2615 offset *= cachep->colour_off;
2616
2617 /*
2618 * Call kasan_poison_slab() before calling alloc_slabmgmt(), so
2619 * page_address() in the latter returns a non-tagged pointer,
2620 * as it should be for slab pages.
2621 */
2622 kasan_poison_slab(page);
2623
2624 /* Get slab management. */
2625 freelist = alloc_slabmgmt(cachep, page, offset,
2626 local_flags & ~GFP_CONSTRAINT_MASK, page_node);
2627 if (OFF_SLAB(cachep) && !freelist)
2628 goto opps1;
2629
2630 slab_map_pages(cachep, page, freelist);
2631
2632 cache_init_objs(cachep, page);
2633
2634 if (gfpflags_allow_blocking(local_flags))
2635 local_irq_disable();
2636
2637 return page;
2638
2639opps1:
2640 kmem_freepages(cachep, page);
2641failed:
2642 if (gfpflags_allow_blocking(local_flags))
2643 local_irq_disable();
2644 return NULL;
2645}
2646
2647static void cache_grow_end(struct kmem_cache *cachep, struct page *page)
2648{
2649 struct kmem_cache_node *n;
2650 void *list = NULL;
2651
2652 check_irq_off();
2653
2654 if (!page)
2655 return;
2656
2657 INIT_LIST_HEAD(&page->slab_list);
2658 n = get_node(cachep, page_to_nid(page));
2659
2660 spin_lock(&n->list_lock);
2661 n->total_slabs++;
2662 if (!page->active) {
2663 list_add_tail(&page->slab_list, &n->slabs_free);
2664 n->free_slabs++;
2665 } else
2666 fixup_slab_list(cachep, n, page, &list);
2667
2668 STATS_INC_GROWN(cachep);
2669 n->free_objects += cachep->num - page->active;
2670 spin_unlock(&n->list_lock);
2671
2672 fixup_objfreelist_debug(cachep, &list);
2673}
2674
2675#if DEBUG
2676
2677/*
2678 * Perform extra freeing checks:
2679 * - detect bad pointers.
2680 * - POISON/RED_ZONE checking
2681 */
2682static void kfree_debugcheck(const void *objp)
2683{
2684 if (!virt_addr_valid(objp)) {
2685 pr_err("kfree_debugcheck: out of range ptr %lxh\n",
2686 (unsigned long)objp);
2687 BUG();
2688 }
2689}
2690
2691static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2692{
2693 unsigned long long redzone1, redzone2;
2694
2695 redzone1 = *dbg_redzone1(cache, obj);
2696 redzone2 = *dbg_redzone2(cache, obj);
2697
2698 /*
2699 * Redzone is ok.
2700 */
2701 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2702 return;
2703
2704 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2705 slab_error(cache, "double free detected");
2706 else
2707 slab_error(cache, "memory outside object was overwritten");
2708
2709 pr_err("%px: redzone 1:0x%llx, redzone 2:0x%llx\n",
2710 obj, redzone1, redzone2);
2711}
2712
2713static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2714 unsigned long caller)
2715{
2716 unsigned int objnr;
2717 struct page *page;
2718
2719 BUG_ON(virt_to_cache(objp) != cachep);
2720
2721 objp -= obj_offset(cachep);
2722 kfree_debugcheck(objp);
2723 page = virt_to_head_page(objp);
2724
2725 if (cachep->flags & SLAB_RED_ZONE) {
2726 verify_redzone_free(cachep, objp);
2727 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2728 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2729 }
2730 if (cachep->flags & SLAB_STORE_USER)
2731 *dbg_userword(cachep, objp) = (void *)caller;
2732
2733 objnr = obj_to_index(cachep, page, objp);
2734
2735 BUG_ON(objnr >= cachep->num);
2736 BUG_ON(objp != index_to_obj(cachep, page, objnr));
2737
2738 if (cachep->flags & SLAB_POISON) {
2739 poison_obj(cachep, objp, POISON_FREE);
2740 slab_kernel_map(cachep, objp, 0);
2741 }
2742 return objp;
2743}
2744
2745#else
2746#define kfree_debugcheck(x) do { } while(0)
2747#define cache_free_debugcheck(x,objp,z) (objp)
2748#endif
2749
2750static inline void fixup_objfreelist_debug(struct kmem_cache *cachep,
2751 void **list)
2752{
2753#if DEBUG
2754 void *next = *list;
2755 void *objp;
2756
2757 while (next) {
2758 objp = next - obj_offset(cachep);
2759 next = *(void **)next;
2760 poison_obj(cachep, objp, POISON_FREE);
2761 }
2762#endif
2763}
2764
2765static inline void fixup_slab_list(struct kmem_cache *cachep,
2766 struct kmem_cache_node *n, struct page *page,
2767 void **list)
2768{
2769 /* move slabp to correct slabp list: */
2770 list_del(&page->slab_list);
2771 if (page->active == cachep->num) {
2772 list_add(&page->slab_list, &n->slabs_full);
2773 if (OBJFREELIST_SLAB(cachep)) {
2774#if DEBUG
2775 /* Poisoning will be done without holding the lock */
2776 if (cachep->flags & SLAB_POISON) {
2777 void **objp = page->freelist;
2778
2779 *objp = *list;
2780 *list = objp;
2781 }
2782#endif
2783 page->freelist = NULL;
2784 }
2785 } else
2786 list_add(&page->slab_list, &n->slabs_partial);
2787}
2788
2789/* Try to find non-pfmemalloc slab if needed */
2790static noinline struct page *get_valid_first_slab(struct kmem_cache_node *n,
2791 struct page *page, bool pfmemalloc)
2792{
2793 if (!page)
2794 return NULL;
2795
2796 if (pfmemalloc)
2797 return page;
2798
2799 if (!PageSlabPfmemalloc(page))
2800 return page;
2801
2802 /* No need to keep pfmemalloc slab if we have enough free objects */
2803 if (n->free_objects > n->free_limit) {
2804 ClearPageSlabPfmemalloc(page);
2805 return page;
2806 }
2807
2808 /* Move pfmemalloc slab to the end of list to speed up next search */
2809 list_del(&page->slab_list);
2810 if (!page->active) {
2811 list_add_tail(&page->slab_list, &n->slabs_free);
2812 n->free_slabs++;
2813 } else
2814 list_add_tail(&page->slab_list, &n->slabs_partial);
2815
2816 list_for_each_entry(page, &n->slabs_partial, slab_list) {
2817 if (!PageSlabPfmemalloc(page))
2818 return page;
2819 }
2820
2821 n->free_touched = 1;
2822 list_for_each_entry(page, &n->slabs_free, slab_list) {
2823 if (!PageSlabPfmemalloc(page)) {
2824 n->free_slabs--;
2825 return page;
2826 }
2827 }
2828
2829 return NULL;
2830}
2831
2832static struct page *get_first_slab(struct kmem_cache_node *n, bool pfmemalloc)
2833{
2834 struct page *page;
2835
2836 assert_spin_locked(&n->list_lock);
2837 page = list_first_entry_or_null(&n->slabs_partial, struct page,
2838 slab_list);
2839 if (!page) {
2840 n->free_touched = 1;
2841 page = list_first_entry_or_null(&n->slabs_free, struct page,
2842 slab_list);
2843 if (page)
2844 n->free_slabs--;
2845 }
2846
2847 if (sk_memalloc_socks())
2848 page = get_valid_first_slab(n, page, pfmemalloc);
2849
2850 return page;
2851}
2852
2853static noinline void *cache_alloc_pfmemalloc(struct kmem_cache *cachep,
2854 struct kmem_cache_node *n, gfp_t flags)
2855{
2856 struct page *page;
2857 void *obj;
2858 void *list = NULL;
2859
2860 if (!gfp_pfmemalloc_allowed(flags))
2861 return NULL;
2862
2863 spin_lock(&n->list_lock);
2864 page = get_first_slab(n, true);
2865 if (!page) {
2866 spin_unlock(&n->list_lock);
2867 return NULL;
2868 }
2869
2870 obj = slab_get_obj(cachep, page);
2871 n->free_objects--;
2872
2873 fixup_slab_list(cachep, n, page, &list);
2874
2875 spin_unlock(&n->list_lock);
2876 fixup_objfreelist_debug(cachep, &list);
2877
2878 return obj;
2879}
2880
2881/*
2882 * Slab list should be fixed up by fixup_slab_list() for existing slab
2883 * or cache_grow_end() for new slab
2884 */
2885static __always_inline int alloc_block(struct kmem_cache *cachep,
2886 struct array_cache *ac, struct page *page, int batchcount)
2887{
2888 /*
2889 * There must be at least one object available for
2890 * allocation.
2891 */
2892 BUG_ON(page->active >= cachep->num);
2893
2894 while (page->active < cachep->num && batchcount--) {
2895 STATS_INC_ALLOCED(cachep);
2896 STATS_INC_ACTIVE(cachep);
2897 STATS_SET_HIGH(cachep);
2898
2899 ac->entry[ac->avail++] = slab_get_obj(cachep, page);
2900 }
2901
2902 return batchcount;
2903}
2904
2905static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
2906{
2907 int batchcount;
2908 struct kmem_cache_node *n;
2909 struct array_cache *ac, *shared;
2910 int node;
2911 void *list = NULL;
2912 struct page *page;
2913
2914 check_irq_off();
2915 node = numa_mem_id();
2916
2917 ac = cpu_cache_get(cachep);
2918 batchcount = ac->batchcount;
2919 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2920 /*
2921 * If there was little recent activity on this cache, then
2922 * perform only a partial refill. Otherwise we could generate
2923 * refill bouncing.
2924 */
2925 batchcount = BATCHREFILL_LIMIT;
2926 }
2927 n = get_node(cachep, node);
2928
2929 BUG_ON(ac->avail > 0 || !n);
2930 shared = READ_ONCE(n->shared);
2931 if (!n->free_objects && (!shared || !shared->avail))
2932 goto direct_grow;
2933
2934 spin_lock(&n->list_lock);
2935 shared = READ_ONCE(n->shared);
2936
2937 /* See if we can refill from the shared array */
2938 if (shared && transfer_objects(ac, shared, batchcount)) {
2939 shared->touched = 1;
2940 goto alloc_done;
2941 }
2942
2943 while (batchcount > 0) {
2944 /* Get slab alloc is to come from. */
2945 page = get_first_slab(n, false);
2946 if (!page)
2947 goto must_grow;
2948
2949 check_spinlock_acquired(cachep);
2950
2951 batchcount = alloc_block(cachep, ac, page, batchcount);
2952 fixup_slab_list(cachep, n, page, &list);
2953 }
2954
2955must_grow:
2956 n->free_objects -= ac->avail;
2957alloc_done:
2958 spin_unlock(&n->list_lock);
2959 fixup_objfreelist_debug(cachep, &list);
2960
2961direct_grow:
2962 if (unlikely(!ac->avail)) {
2963 /* Check if we can use obj in pfmemalloc slab */
2964 if (sk_memalloc_socks()) {
2965 void *obj = cache_alloc_pfmemalloc(cachep, n, flags);
2966
2967 if (obj)
2968 return obj;
2969 }
2970
2971 page = cache_grow_begin(cachep, gfp_exact_node(flags), node);
2972
2973 /*
2974 * cache_grow_begin() can reenable interrupts,
2975 * then ac could change.
2976 */
2977 ac = cpu_cache_get(cachep);
2978 if (!ac->avail && page)
2979 alloc_block(cachep, ac, page, batchcount);
2980 cache_grow_end(cachep, page);
2981
2982 if (!ac->avail)
2983 return NULL;
2984 }
2985 ac->touched = 1;
2986
2987 return ac->entry[--ac->avail];
2988}
2989
2990static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2991 gfp_t flags)
2992{
2993 might_sleep_if(gfpflags_allow_blocking(flags));
2994}
2995
2996#if DEBUG
2997static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
2998 gfp_t flags, void *objp, unsigned long caller)
2999{
3000 WARN_ON_ONCE(cachep->ctor && (flags & __GFP_ZERO));
3001 if (!objp)
3002 return objp;
3003 if (cachep->flags & SLAB_POISON) {
3004 check_poison_obj(cachep, objp);
3005 slab_kernel_map(cachep, objp, 1);
3006 poison_obj(cachep, objp, POISON_INUSE);
3007 }
3008 if (cachep->flags & SLAB_STORE_USER)
3009 *dbg_userword(cachep, objp) = (void *)caller;
3010
3011 if (cachep->flags & SLAB_RED_ZONE) {
3012 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3013 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3014 slab_error(cachep, "double free, or memory outside object was overwritten");
3015 pr_err("%px: redzone 1:0x%llx, redzone 2:0x%llx\n",
3016 objp, *dbg_redzone1(cachep, objp),
3017 *dbg_redzone2(cachep, objp));
3018 }
3019 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3020 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3021 }
3022
3023 objp += obj_offset(cachep);
3024 if (cachep->ctor && cachep->flags & SLAB_POISON)
3025 cachep->ctor(objp);
3026 if (ARCH_SLAB_MINALIGN &&
3027 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
3028 pr_err("0x%px: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3029 objp, (int)ARCH_SLAB_MINALIGN);
3030 }
3031 return objp;
3032}
3033#else
3034#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3035#endif
3036
3037static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3038{
3039 void *objp;
3040 struct array_cache *ac;
3041
3042 check_irq_off();
3043
3044 ac = cpu_cache_get(cachep);
3045 if (likely(ac->avail)) {
3046 ac->touched = 1;
3047 objp = ac->entry[--ac->avail];
3048
3049 STATS_INC_ALLOCHIT(cachep);
3050 goto out;
3051 }
3052
3053 STATS_INC_ALLOCMISS(cachep);
3054 objp = cache_alloc_refill(cachep, flags);
3055 /*
3056 * the 'ac' may be updated by cache_alloc_refill(),
3057 * and kmemleak_erase() requires its correct value.
3058 */
3059 ac = cpu_cache_get(cachep);
3060
3061out:
3062 /*
3063 * To avoid a false negative, if an object that is in one of the
3064 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3065 * treat the array pointers as a reference to the object.
3066 */
3067 if (objp)
3068 kmemleak_erase(&ac->entry[ac->avail]);
3069 return objp;
3070}
3071
3072#ifdef CONFIG_NUMA
3073/*
3074 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
3075 *
3076 * If we are in_interrupt, then process context, including cpusets and
3077 * mempolicy, may not apply and should not be used for allocation policy.
3078 */
3079static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3080{
3081 int nid_alloc, nid_here;
3082
3083 if (in_interrupt() || (flags & __GFP_THISNODE))
3084 return NULL;
3085 nid_alloc = nid_here = numa_mem_id();
3086 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3087 nid_alloc = cpuset_slab_spread_node();
3088 else if (current->mempolicy)
3089 nid_alloc = mempolicy_slab_node();
3090 if (nid_alloc != nid_here)
3091 return ____cache_alloc_node(cachep, flags, nid_alloc);
3092 return NULL;
3093}
3094
3095/*
3096 * Fallback function if there was no memory available and no objects on a
3097 * certain node and fall back is permitted. First we scan all the
3098 * available node for available objects. If that fails then we
3099 * perform an allocation without specifying a node. This allows the page
3100 * allocator to do its reclaim / fallback magic. We then insert the
3101 * slab into the proper nodelist and then allocate from it.
3102 */
3103static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3104{
3105 struct zonelist *zonelist;
3106 struct zoneref *z;
3107 struct zone *zone;
3108 enum zone_type high_zoneidx = gfp_zone(flags);
3109 void *obj = NULL;
3110 struct page *page;
3111 int nid;
3112 unsigned int cpuset_mems_cookie;
3113
3114 if (flags & __GFP_THISNODE)
3115 return NULL;
3116
3117retry_cpuset:
3118 cpuset_mems_cookie = read_mems_allowed_begin();
3119 zonelist = node_zonelist(mempolicy_slab_node(), flags);
3120
3121retry:
3122 /*
3123 * Look through allowed nodes for objects available
3124 * from existing per node queues.
3125 */
3126 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3127 nid = zone_to_nid(zone);
3128
3129 if (cpuset_zone_allowed(zone, flags) &&
3130 get_node(cache, nid) &&
3131 get_node(cache, nid)->free_objects) {
3132 obj = ____cache_alloc_node(cache,
3133 gfp_exact_node(flags), nid);
3134 if (obj)
3135 break;
3136 }
3137 }
3138
3139 if (!obj) {
3140 /*
3141 * This allocation will be performed within the constraints
3142 * of the current cpuset / memory policy requirements.
3143 * We may trigger various forms of reclaim on the allowed
3144 * set and go into memory reserves if necessary.
3145 */
3146 page = cache_grow_begin(cache, flags, numa_mem_id());
3147 cache_grow_end(cache, page);
3148 if (page) {
3149 nid = page_to_nid(page);
3150 obj = ____cache_alloc_node(cache,
3151 gfp_exact_node(flags), nid);
3152
3153 /*
3154 * Another processor may allocate the objects in
3155 * the slab since we are not holding any locks.
3156 */
3157 if (!obj)
3158 goto retry;
3159 }
3160 }
3161
3162 if (unlikely(!obj && read_mems_allowed_retry(cpuset_mems_cookie)))
3163 goto retry_cpuset;
3164 return obj;
3165}
3166
3167/*
3168 * A interface to enable slab creation on nodeid
3169 */
3170static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3171 int nodeid)
3172{
3173 struct page *page;
3174 struct kmem_cache_node *n;
3175 void *obj = NULL;
3176 void *list = NULL;
3177
3178 VM_BUG_ON(nodeid < 0 || nodeid >= MAX_NUMNODES);
3179 n = get_node(cachep, nodeid);
3180 BUG_ON(!n);
3181
3182 check_irq_off();
3183 spin_lock(&n->list_lock);
3184 page = get_first_slab(n, false);
3185 if (!page)
3186 goto must_grow;
3187
3188 check_spinlock_acquired_node(cachep, nodeid);
3189
3190 STATS_INC_NODEALLOCS(cachep);
3191 STATS_INC_ACTIVE(cachep);
3192 STATS_SET_HIGH(cachep);
3193
3194 BUG_ON(page->active == cachep->num);
3195
3196 obj = slab_get_obj(cachep, page);
3197 n->free_objects--;
3198
3199 fixup_slab_list(cachep, n, page, &list);
3200
3201 spin_unlock(&n->list_lock);
3202 fixup_objfreelist_debug(cachep, &list);
3203 return obj;
3204
3205must_grow:
3206 spin_unlock(&n->list_lock);
3207 page = cache_grow_begin(cachep, gfp_exact_node(flags), nodeid);
3208 if (page) {
3209 /* This slab isn't counted yet so don't update free_objects */
3210 obj = slab_get_obj(cachep, page);
3211 }
3212 cache_grow_end(cachep, page);
3213
3214 return obj ? obj : fallback_alloc(cachep, flags);
3215}
3216
3217static __always_inline void *
3218slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3219 unsigned long caller)
3220{
3221 unsigned long save_flags;
3222 void *ptr;
3223 int slab_node = numa_mem_id();
3224
3225 flags &= gfp_allowed_mask;
3226 cachep = slab_pre_alloc_hook(cachep, flags);
3227 if (unlikely(!cachep))
3228 return NULL;
3229
3230 cache_alloc_debugcheck_before(cachep, flags);
3231 local_irq_save(save_flags);
3232
3233 if (nodeid == NUMA_NO_NODE)
3234 nodeid = slab_node;
3235
3236 if (unlikely(!get_node(cachep, nodeid))) {
3237 /* Node not bootstrapped yet */
3238 ptr = fallback_alloc(cachep, flags);
3239 goto out;
3240 }
3241
3242 if (nodeid == slab_node) {
3243 /*
3244 * Use the locally cached objects if possible.
3245 * However ____cache_alloc does not allow fallback
3246 * to other nodes. It may fail while we still have
3247 * objects on other nodes available.
3248 */
3249 ptr = ____cache_alloc(cachep, flags);
3250 if (ptr)
3251 goto out;
3252 }
3253 /* ___cache_alloc_node can fall back to other nodes */
3254 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3255 out:
3256 local_irq_restore(save_flags);
3257 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3258
3259 if (unlikely(slab_want_init_on_alloc(flags, cachep)) && ptr)
3260 memset(ptr, 0, cachep->object_size);
3261
3262 slab_post_alloc_hook(cachep, flags, 1, &ptr);
3263 return ptr;
3264}
3265
3266static __always_inline void *
3267__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3268{
3269 void *objp;
3270
3271 if (current->mempolicy || cpuset_do_slab_mem_spread()) {
3272 objp = alternate_node_alloc(cache, flags);
3273 if (objp)
3274 goto out;
3275 }
3276 objp = ____cache_alloc(cache, flags);
3277
3278 /*
3279 * We may just have run out of memory on the local node.
3280 * ____cache_alloc_node() knows how to locate memory on other nodes
3281 */
3282 if (!objp)
3283 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3284
3285 out:
3286 return objp;
3287}
3288#else
3289
3290static __always_inline void *
3291__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3292{
3293 return ____cache_alloc(cachep, flags);
3294}
3295
3296#endif /* CONFIG_NUMA */
3297
3298static __always_inline void *
3299slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
3300{
3301 unsigned long save_flags;
3302 void *objp;
3303
3304 flags &= gfp_allowed_mask;
3305 cachep = slab_pre_alloc_hook(cachep, flags);
3306 if (unlikely(!cachep))
3307 return NULL;
3308
3309 cache_alloc_debugcheck_before(cachep, flags);
3310 local_irq_save(save_flags);
3311 objp = __do_cache_alloc(cachep, flags);
3312 local_irq_restore(save_flags);
3313 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3314 prefetchw(objp);
3315
3316 if (unlikely(slab_want_init_on_alloc(flags, cachep)) && objp)
3317 memset(objp, 0, cachep->object_size);
3318
3319 slab_post_alloc_hook(cachep, flags, 1, &objp);
3320 return objp;
3321}
3322
3323/*
3324 * Caller needs to acquire correct kmem_cache_node's list_lock
3325 * @list: List of detached free slabs should be freed by caller
3326 */
3327static void free_block(struct kmem_cache *cachep, void **objpp,
3328 int nr_objects, int node, struct list_head *list)
3329{
3330 int i;
3331 struct kmem_cache_node *n = get_node(cachep, node);
3332 struct page *page;
3333
3334 n->free_objects += nr_objects;
3335
3336 for (i = 0; i < nr_objects; i++) {
3337 void *objp;
3338 struct page *page;
3339
3340 objp = objpp[i];
3341
3342 page = virt_to_head_page(objp);
3343 list_del(&page->slab_list);
3344 check_spinlock_acquired_node(cachep, node);
3345 slab_put_obj(cachep, page, objp);
3346 STATS_DEC_ACTIVE(cachep);
3347
3348 /* fixup slab chains */
3349 if (page->active == 0) {
3350 list_add(&page->slab_list, &n->slabs_free);
3351 n->free_slabs++;
3352 } else {
3353 /* Unconditionally move a slab to the end of the
3354 * partial list on free - maximum time for the
3355 * other objects to be freed, too.
3356 */
3357 list_add_tail(&page->slab_list, &n->slabs_partial);
3358 }
3359 }
3360
3361 while (n->free_objects > n->free_limit && !list_empty(&n->slabs_free)) {
3362 n->free_objects -= cachep->num;
3363
3364 page = list_last_entry(&n->slabs_free, struct page, slab_list);
3365 list_move(&page->slab_list, list);
3366 n->free_slabs--;
3367 n->total_slabs--;
3368 }
3369}
3370
3371static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3372{
3373 int batchcount;
3374 struct kmem_cache_node *n;
3375 int node = numa_mem_id();
3376 LIST_HEAD(list);
3377
3378 batchcount = ac->batchcount;
3379
3380 check_irq_off();
3381 n = get_node(cachep, node);
3382 spin_lock(&n->list_lock);
3383 if (n->shared) {
3384 struct array_cache *shared_array = n->shared;
3385 int max = shared_array->limit - shared_array->avail;
3386 if (max) {
3387 if (batchcount > max)
3388 batchcount = max;
3389 memcpy(&(shared_array->entry[shared_array->avail]),
3390 ac->entry, sizeof(void *) * batchcount);
3391 shared_array->avail += batchcount;
3392 goto free_done;
3393 }
3394 }
3395
3396 free_block(cachep, ac->entry, batchcount, node, &list);
3397free_done:
3398#if STATS
3399 {
3400 int i = 0;
3401 struct page *page;
3402
3403 list_for_each_entry(page, &n->slabs_free, slab_list) {
3404 BUG_ON(page->active);
3405
3406 i++;
3407 }
3408 STATS_SET_FREEABLE(cachep, i);
3409 }
3410#endif
3411 spin_unlock(&n->list_lock);
3412 slabs_destroy(cachep, &list);
3413 ac->avail -= batchcount;
3414 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3415}
3416
3417/*
3418 * Release an obj back to its cache. If the obj has a constructed state, it must
3419 * be in this state _before_ it is released. Called with disabled ints.
3420 */
3421static __always_inline void __cache_free(struct kmem_cache *cachep, void *objp,
3422 unsigned long caller)
3423{
3424 /* Put the object into the quarantine, don't touch it for now. */
3425 if (kasan_slab_free(cachep, objp, _RET_IP_))
3426 return;
3427
3428 ___cache_free(cachep, objp, caller);
3429}
3430
3431void ___cache_free(struct kmem_cache *cachep, void *objp,
3432 unsigned long caller)
3433{
3434 struct array_cache *ac = cpu_cache_get(cachep);
3435
3436 check_irq_off();
3437 if (unlikely(slab_want_init_on_free(cachep)))
3438 memset(objp, 0, cachep->object_size);
3439 kmemleak_free_recursive(objp, cachep->flags);
3440 objp = cache_free_debugcheck(cachep, objp, caller);
3441
3442 /*
3443 * Skip calling cache_free_alien() when the platform is not numa.
3444 * This will avoid cache misses that happen while accessing slabp (which
3445 * is per page memory reference) to get nodeid. Instead use a global
3446 * variable to skip the call, which is mostly likely to be present in
3447 * the cache.
3448 */
3449 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3450 return;
3451
3452 if (ac->avail < ac->limit) {
3453 STATS_INC_FREEHIT(cachep);
3454 } else {
3455 STATS_INC_FREEMISS(cachep);
3456 cache_flusharray(cachep, ac);
3457 }
3458
3459 if (sk_memalloc_socks()) {
3460 struct page *page = virt_to_head_page(objp);
3461
3462 if (unlikely(PageSlabPfmemalloc(page))) {
3463 cache_free_pfmemalloc(cachep, page, objp);
3464 return;
3465 }
3466 }
3467
3468 ac->entry[ac->avail++] = objp;
3469}
3470
3471/**
3472 * kmem_cache_alloc - Allocate an object
3473 * @cachep: The cache to allocate from.
3474 * @flags: See kmalloc().
3475 *
3476 * Allocate an object from this cache. The flags are only relevant
3477 * if the cache has no available objects.
3478 *
3479 * Return: pointer to the new object or %NULL in case of error
3480 */
3481void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3482{
3483 void *ret = slab_alloc(cachep, flags, _RET_IP_);
3484
3485 trace_kmem_cache_alloc(_RET_IP_, ret,
3486 cachep->object_size, cachep->size, flags);
3487
3488 return ret;
3489}
3490EXPORT_SYMBOL(kmem_cache_alloc);
3491
3492static __always_inline void
3493cache_alloc_debugcheck_after_bulk(struct kmem_cache *s, gfp_t flags,
3494 size_t size, void **p, unsigned long caller)
3495{
3496 size_t i;
3497
3498 for (i = 0; i < size; i++)
3499 p[i] = cache_alloc_debugcheck_after(s, flags, p[i], caller);
3500}
3501
3502int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3503 void **p)
3504{
3505 size_t i;
3506
3507 s = slab_pre_alloc_hook(s, flags);
3508 if (!s)
3509 return 0;
3510
3511 cache_alloc_debugcheck_before(s, flags);
3512
3513 local_irq_disable();
3514 for (i = 0; i < size; i++) {
3515 void *objp = __do_cache_alloc(s, flags);
3516
3517 if (unlikely(!objp))
3518 goto error;
3519 p[i] = objp;
3520 }
3521 local_irq_enable();
3522
3523 cache_alloc_debugcheck_after_bulk(s, flags, size, p, _RET_IP_);
3524
3525 /* Clear memory outside IRQ disabled section */
3526 if (unlikely(slab_want_init_on_alloc(flags, s)))
3527 for (i = 0; i < size; i++)
3528 memset(p[i], 0, s->object_size);
3529
3530 slab_post_alloc_hook(s, flags, size, p);
3531 /* FIXME: Trace call missing. Christoph would like a bulk variant */
3532 return size;
3533error:
3534 local_irq_enable();
3535 cache_alloc_debugcheck_after_bulk(s, flags, i, p, _RET_IP_);
3536 slab_post_alloc_hook(s, flags, i, p);
3537 __kmem_cache_free_bulk(s, i, p);
3538 return 0;
3539}
3540EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3541
3542#ifdef CONFIG_TRACING
3543void *
3544kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
3545{
3546 void *ret;
3547
3548 ret = slab_alloc(cachep, flags, _RET_IP_);
3549
3550 ret = kasan_kmalloc(cachep, ret, size, flags);
3551 trace_kmalloc(_RET_IP_, ret,
3552 size, cachep->size, flags);
3553 return ret;
3554}
3555EXPORT_SYMBOL(kmem_cache_alloc_trace);
3556#endif
3557
3558#ifdef CONFIG_NUMA
3559/**
3560 * kmem_cache_alloc_node - Allocate an object on the specified node
3561 * @cachep: The cache to allocate from.
3562 * @flags: See kmalloc().
3563 * @nodeid: node number of the target node.
3564 *
3565 * Identical to kmem_cache_alloc but it will allocate memory on the given
3566 * node, which can improve the performance for cpu bound structures.
3567 *
3568 * Fallback to other node is possible if __GFP_THISNODE is not set.
3569 *
3570 * Return: pointer to the new object or %NULL in case of error
3571 */
3572void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3573{
3574 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3575
3576 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3577 cachep->object_size, cachep->size,
3578 flags, nodeid);
3579
3580 return ret;
3581}
3582EXPORT_SYMBOL(kmem_cache_alloc_node);
3583
3584#ifdef CONFIG_TRACING
3585void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
3586 gfp_t flags,
3587 int nodeid,
3588 size_t size)
3589{
3590 void *ret;
3591
3592 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3593
3594 ret = kasan_kmalloc(cachep, ret, size, flags);
3595 trace_kmalloc_node(_RET_IP_, ret,
3596 size, cachep->size,
3597 flags, nodeid);
3598 return ret;
3599}
3600EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3601#endif
3602
3603static __always_inline void *
3604__do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
3605{
3606 struct kmem_cache *cachep;
3607 void *ret;
3608
3609 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3610 return NULL;
3611 cachep = kmalloc_slab(size, flags);
3612 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3613 return cachep;
3614 ret = kmem_cache_alloc_node_trace(cachep, flags, node, size);
3615 ret = kasan_kmalloc(cachep, ret, size, flags);
3616
3617 return ret;
3618}
3619
3620void *__kmalloc_node(size_t size, gfp_t flags, int node)
3621{
3622 return __do_kmalloc_node(size, flags, node, _RET_IP_);
3623}
3624EXPORT_SYMBOL(__kmalloc_node);
3625
3626void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3627 int node, unsigned long caller)
3628{
3629 return __do_kmalloc_node(size, flags, node, caller);
3630}
3631EXPORT_SYMBOL(__kmalloc_node_track_caller);
3632#endif /* CONFIG_NUMA */
3633
3634/**
3635 * __do_kmalloc - allocate memory
3636 * @size: how many bytes of memory are required.
3637 * @flags: the type of memory to allocate (see kmalloc).
3638 * @caller: function caller for debug tracking of the caller
3639 *
3640 * Return: pointer to the allocated memory or %NULL in case of error
3641 */
3642static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3643 unsigned long caller)
3644{
3645 struct kmem_cache *cachep;
3646 void *ret;
3647
3648 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3649 return NULL;
3650 cachep = kmalloc_slab(size, flags);
3651 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3652 return cachep;
3653 ret = slab_alloc(cachep, flags, caller);
3654
3655 ret = kasan_kmalloc(cachep, ret, size, flags);
3656 trace_kmalloc(caller, ret,
3657 size, cachep->size, flags);
3658
3659 return ret;
3660}
3661
3662void *__kmalloc(size_t size, gfp_t flags)
3663{
3664 return __do_kmalloc(size, flags, _RET_IP_);
3665}
3666EXPORT_SYMBOL(__kmalloc);
3667
3668void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3669{
3670 return __do_kmalloc(size, flags, caller);
3671}
3672EXPORT_SYMBOL(__kmalloc_track_caller);
3673
3674/**
3675 * kmem_cache_free - Deallocate an object
3676 * @cachep: The cache the allocation was from.
3677 * @objp: The previously allocated object.
3678 *
3679 * Free an object which was previously allocated from this
3680 * cache.
3681 */
3682void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3683{
3684 unsigned long flags;
3685 cachep = cache_from_obj(cachep, objp);
3686 if (!cachep)
3687 return;
3688
3689 local_irq_save(flags);
3690 debug_check_no_locks_freed(objp, cachep->object_size);
3691 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3692 debug_check_no_obj_freed(objp, cachep->object_size);
3693 __cache_free(cachep, objp, _RET_IP_);
3694 local_irq_restore(flags);
3695
3696 trace_kmem_cache_free(_RET_IP_, objp);
3697}
3698EXPORT_SYMBOL(kmem_cache_free);
3699
3700void kmem_cache_free_bulk(struct kmem_cache *orig_s, size_t size, void **p)
3701{
3702 struct kmem_cache *s;
3703 size_t i;
3704
3705 local_irq_disable();
3706 for (i = 0; i < size; i++) {
3707 void *objp = p[i];
3708
3709 if (!orig_s) /* called via kfree_bulk */
3710 s = virt_to_cache(objp);
3711 else
3712 s = cache_from_obj(orig_s, objp);
3713 if (!s)
3714 continue;
3715
3716 debug_check_no_locks_freed(objp, s->object_size);
3717 if (!(s->flags & SLAB_DEBUG_OBJECTS))
3718 debug_check_no_obj_freed(objp, s->object_size);
3719
3720 __cache_free(s, objp, _RET_IP_);
3721 }
3722 local_irq_enable();
3723
3724 /* FIXME: add tracing */
3725}
3726EXPORT_SYMBOL(kmem_cache_free_bulk);
3727
3728/**
3729 * kfree - free previously allocated memory
3730 * @objp: pointer returned by kmalloc.
3731 *
3732 * If @objp is NULL, no operation is performed.
3733 *
3734 * Don't free memory not originally allocated by kmalloc()
3735 * or you will run into trouble.
3736 */
3737void kfree(const void *objp)
3738{
3739 struct kmem_cache *c;
3740 unsigned long flags;
3741
3742 trace_kfree(_RET_IP_, objp);
3743
3744 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3745 return;
3746 local_irq_save(flags);
3747 kfree_debugcheck(objp);
3748 c = virt_to_cache(objp);
3749 if (!c) {
3750 local_irq_restore(flags);
3751 return;
3752 }
3753 debug_check_no_locks_freed(objp, c->object_size);
3754
3755 debug_check_no_obj_freed(objp, c->object_size);
3756 __cache_free(c, (void *)objp, _RET_IP_);
3757 local_irq_restore(flags);
3758}
3759EXPORT_SYMBOL(kfree);
3760
3761/*
3762 * This initializes kmem_cache_node or resizes various caches for all nodes.
3763 */
3764static int setup_kmem_cache_nodes(struct kmem_cache *cachep, gfp_t gfp)
3765{
3766 int ret;
3767 int node;
3768 struct kmem_cache_node *n;
3769
3770 for_each_online_node(node) {
3771 ret = setup_kmem_cache_node(cachep, node, gfp, true);
3772 if (ret)
3773 goto fail;
3774
3775 }
3776
3777 return 0;
3778
3779fail:
3780 if (!cachep->list.next) {
3781 /* Cache is not active yet. Roll back what we did */
3782 node--;
3783 while (node >= 0) {
3784 n = get_node(cachep, node);
3785 if (n) {
3786 kfree(n->shared);
3787 free_alien_cache(n->alien);
3788 kfree(n);
3789 cachep->node[node] = NULL;
3790 }
3791 node--;
3792 }
3793 }
3794 return -ENOMEM;
3795}
3796
3797/* Always called with the slab_mutex held */
3798static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
3799 int batchcount, int shared, gfp_t gfp)
3800{
3801 struct array_cache __percpu *cpu_cache, *prev;
3802 int cpu;
3803
3804 cpu_cache = alloc_kmem_cache_cpus(cachep, limit, batchcount);
3805 if (!cpu_cache)
3806 return -ENOMEM;
3807
3808 prev = cachep->cpu_cache;
3809 cachep->cpu_cache = cpu_cache;
3810 /*
3811 * Without a previous cpu_cache there's no need to synchronize remote
3812 * cpus, so skip the IPIs.
3813 */
3814 if (prev)
3815 kick_all_cpus_sync();
3816
3817 check_irq_on();
3818 cachep->batchcount = batchcount;
3819 cachep->limit = limit;
3820 cachep->shared = shared;
3821
3822 if (!prev)
3823 goto setup_node;
3824
3825 for_each_online_cpu(cpu) {
3826 LIST_HEAD(list);
3827 int node;
3828 struct kmem_cache_node *n;
3829 struct array_cache *ac = per_cpu_ptr(prev, cpu);
3830
3831 node = cpu_to_mem(cpu);
3832 n = get_node(cachep, node);
3833 spin_lock_irq(&n->list_lock);
3834 free_block(cachep, ac->entry, ac->avail, node, &list);
3835 spin_unlock_irq(&n->list_lock);
3836 slabs_destroy(cachep, &list);
3837 }
3838 free_percpu(prev);
3839
3840setup_node:
3841 return setup_kmem_cache_nodes(cachep, gfp);
3842}
3843
3844static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3845 int batchcount, int shared, gfp_t gfp)
3846{
3847 int ret;
3848 struct kmem_cache *c;
3849
3850 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3851
3852 if (slab_state < FULL)
3853 return ret;
3854
3855 if ((ret < 0) || !is_root_cache(cachep))
3856 return ret;
3857
3858 lockdep_assert_held(&slab_mutex);
3859 for_each_memcg_cache(c, cachep) {
3860 /* return value determined by the root cache only */
3861 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
3862 }
3863
3864 return ret;
3865}
3866
3867/* Called with slab_mutex held always */
3868static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
3869{
3870 int err;
3871 int limit = 0;
3872 int shared = 0;
3873 int batchcount = 0;
3874
3875 err = cache_random_seq_create(cachep, cachep->num, gfp);
3876 if (err)
3877 goto end;
3878
3879 if (!is_root_cache(cachep)) {
3880 struct kmem_cache *root = memcg_root_cache(cachep);
3881 limit = root->limit;
3882 shared = root->shared;
3883 batchcount = root->batchcount;
3884 }
3885
3886 if (limit && shared && batchcount)
3887 goto skip_setup;
3888 /*
3889 * The head array serves three purposes:
3890 * - create a LIFO ordering, i.e. return objects that are cache-warm
3891 * - reduce the number of spinlock operations.
3892 * - reduce the number of linked list operations on the slab and
3893 * bufctl chains: array operations are cheaper.
3894 * The numbers are guessed, we should auto-tune as described by
3895 * Bonwick.
3896 */
3897 if (cachep->size > 131072)
3898 limit = 1;
3899 else if (cachep->size > PAGE_SIZE)
3900 limit = 8;
3901 else if (cachep->size > 1024)
3902 limit = 24;
3903 else if (cachep->size > 256)
3904 limit = 54;
3905 else
3906 limit = 120;
3907
3908 /*
3909 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3910 * allocation behaviour: Most allocs on one cpu, most free operations
3911 * on another cpu. For these cases, an efficient object passing between
3912 * cpus is necessary. This is provided by a shared array. The array
3913 * replaces Bonwick's magazine layer.
3914 * On uniprocessor, it's functionally equivalent (but less efficient)
3915 * to a larger limit. Thus disabled by default.
3916 */
3917 shared = 0;
3918 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
3919 shared = 8;
3920
3921#if DEBUG
3922 /*
3923 * With debugging enabled, large batchcount lead to excessively long
3924 * periods with disabled local interrupts. Limit the batchcount
3925 */
3926 if (limit > 32)
3927 limit = 32;
3928#endif
3929 batchcount = (limit + 1) / 2;
3930skip_setup:
3931 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3932end:
3933 if (err)
3934 pr_err("enable_cpucache failed for %s, error %d\n",
3935 cachep->name, -err);
3936 return err;
3937}
3938
3939/*
3940 * Drain an array if it contains any elements taking the node lock only if
3941 * necessary. Note that the node listlock also protects the array_cache
3942 * if drain_array() is used on the shared array.
3943 */
3944static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
3945 struct array_cache *ac, int node)
3946{
3947 LIST_HEAD(list);
3948
3949 /* ac from n->shared can be freed if we don't hold the slab_mutex. */
3950 check_mutex_acquired();
3951
3952 if (!ac || !ac->avail)
3953 return;
3954
3955 if (ac->touched) {
3956 ac->touched = 0;
3957 return;
3958 }
3959
3960 spin_lock_irq(&n->list_lock);
3961 drain_array_locked(cachep, ac, node, false, &list);
3962 spin_unlock_irq(&n->list_lock);
3963
3964 slabs_destroy(cachep, &list);
3965}
3966
3967/**
3968 * cache_reap - Reclaim memory from caches.
3969 * @w: work descriptor
3970 *
3971 * Called from workqueue/eventd every few seconds.
3972 * Purpose:
3973 * - clear the per-cpu caches for this CPU.
3974 * - return freeable pages to the main free memory pool.
3975 *
3976 * If we cannot acquire the cache chain mutex then just give up - we'll try
3977 * again on the next iteration.
3978 */
3979static void cache_reap(struct work_struct *w)
3980{
3981 struct kmem_cache *searchp;
3982 struct kmem_cache_node *n;
3983 int node = numa_mem_id();
3984 struct delayed_work *work = to_delayed_work(w);
3985
3986 if (!mutex_trylock(&slab_mutex))
3987 /* Give up. Setup the next iteration. */
3988 goto out;
3989
3990 list_for_each_entry(searchp, &slab_caches, list) {
3991 check_irq_on();
3992
3993 /*
3994 * We only take the node lock if absolutely necessary and we
3995 * have established with reasonable certainty that
3996 * we can do some work if the lock was obtained.
3997 */
3998 n = get_node(searchp, node);
3999
4000 reap_alien(searchp, n);
4001
4002 drain_array(searchp, n, cpu_cache_get(searchp), node);
4003
4004 /*
4005 * These are racy checks but it does not matter
4006 * if we skip one check or scan twice.
4007 */
4008 if (time_after(n->next_reap, jiffies))
4009 goto next;
4010
4011 n->next_reap = jiffies + REAPTIMEOUT_NODE;
4012
4013 drain_array(searchp, n, n->shared, node);
4014
4015 if (n->free_touched)
4016 n->free_touched = 0;
4017 else {
4018 int freed;
4019
4020 freed = drain_freelist(searchp, n, (n->free_limit +
4021 5 * searchp->num - 1) / (5 * searchp->num));
4022 STATS_ADD_REAPED(searchp, freed);
4023 }
4024next:
4025 cond_resched();
4026 }
4027 check_irq_on();
4028 mutex_unlock(&slab_mutex);
4029 next_reap_node();
4030out:
4031 /* Set up the next iteration */
4032 schedule_delayed_work_on(smp_processor_id(), work,
4033 round_jiffies_relative(REAPTIMEOUT_AC));
4034}
4035
4036void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
4037{
4038 unsigned long active_objs, num_objs, active_slabs;
4039 unsigned long total_slabs = 0, free_objs = 0, shared_avail = 0;
4040 unsigned long free_slabs = 0;
4041 int node;
4042 struct kmem_cache_node *n;
4043
4044 for_each_kmem_cache_node(cachep, node, n) {
4045 check_irq_on();
4046 spin_lock_irq(&n->list_lock);
4047
4048 total_slabs += n->total_slabs;
4049 free_slabs += n->free_slabs;
4050 free_objs += n->free_objects;
4051
4052 if (n->shared)
4053 shared_avail += n->shared->avail;
4054
4055 spin_unlock_irq(&n->list_lock);
4056 }
4057 num_objs = total_slabs * cachep->num;
4058 active_slabs = total_slabs - free_slabs;
4059 active_objs = num_objs - free_objs;
4060
4061 sinfo->active_objs = active_objs;
4062 sinfo->num_objs = num_objs;
4063 sinfo->active_slabs = active_slabs;
4064 sinfo->num_slabs = total_slabs;
4065 sinfo->shared_avail = shared_avail;
4066 sinfo->limit = cachep->limit;
4067 sinfo->batchcount = cachep->batchcount;
4068 sinfo->shared = cachep->shared;
4069 sinfo->objects_per_slab = cachep->num;
4070 sinfo->cache_order = cachep->gfporder;
4071}
4072
4073void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4074{
4075#if STATS
4076 { /* node stats */
4077 unsigned long high = cachep->high_mark;
4078 unsigned long allocs = cachep->num_allocations;
4079 unsigned long grown = cachep->grown;
4080 unsigned long reaped = cachep->reaped;
4081 unsigned long errors = cachep->errors;
4082 unsigned long max_freeable = cachep->max_freeable;
4083 unsigned long node_allocs = cachep->node_allocs;
4084 unsigned long node_frees = cachep->node_frees;
4085 unsigned long overflows = cachep->node_overflow;
4086
4087 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu %4lu %4lu",
4088 allocs, high, grown,
4089 reaped, errors, max_freeable, node_allocs,
4090 node_frees, overflows);
4091 }
4092 /* cpu stats */
4093 {
4094 unsigned long allochit = atomic_read(&cachep->allochit);
4095 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4096 unsigned long freehit = atomic_read(&cachep->freehit);
4097 unsigned long freemiss = atomic_read(&cachep->freemiss);
4098
4099 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4100 allochit, allocmiss, freehit, freemiss);
4101 }
4102#endif
4103}
4104
4105#define MAX_SLABINFO_WRITE 128
4106/**
4107 * slabinfo_write - Tuning for the slab allocator
4108 * @file: unused
4109 * @buffer: user buffer
4110 * @count: data length
4111 * @ppos: unused
4112 *
4113 * Return: %0 on success, negative error code otherwise.
4114 */
4115ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4116 size_t count, loff_t *ppos)
4117{
4118 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4119 int limit, batchcount, shared, res;
4120 struct kmem_cache *cachep;
4121
4122 if (count > MAX_SLABINFO_WRITE)
4123 return -EINVAL;
4124 if (copy_from_user(&kbuf, buffer, count))
4125 return -EFAULT;
4126 kbuf[MAX_SLABINFO_WRITE] = '\0';
4127
4128 tmp = strchr(kbuf, ' ');
4129 if (!tmp)
4130 return -EINVAL;
4131 *tmp = '\0';
4132 tmp++;
4133 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4134 return -EINVAL;
4135
4136 /* Find the cache in the chain of caches. */
4137 mutex_lock(&slab_mutex);
4138 res = -EINVAL;
4139 list_for_each_entry(cachep, &slab_caches, list) {
4140 if (!strcmp(cachep->name, kbuf)) {
4141 if (limit < 1 || batchcount < 1 ||
4142 batchcount > limit || shared < 0) {
4143 res = 0;
4144 } else {
4145 res = do_tune_cpucache(cachep, limit,
4146 batchcount, shared,
4147 GFP_KERNEL);
4148 }
4149 break;
4150 }
4151 }
4152 mutex_unlock(&slab_mutex);
4153 if (res >= 0)
4154 res = count;
4155 return res;
4156}
4157
4158#ifdef CONFIG_HARDENED_USERCOPY
4159/*
4160 * Rejects incorrectly sized objects and objects that are to be copied
4161 * to/from userspace but do not fall entirely within the containing slab
4162 * cache's usercopy region.
4163 *
4164 * Returns NULL if check passes, otherwise const char * to name of cache
4165 * to indicate an error.
4166 */
4167void __check_heap_object(const void *ptr, unsigned long n, struct page *page,
4168 bool to_user)
4169{
4170 struct kmem_cache *cachep;
4171 unsigned int objnr;
4172 unsigned long offset;
4173
4174 ptr = kasan_reset_tag(ptr);
4175
4176 /* Find and validate object. */
4177 cachep = page->slab_cache;
4178 objnr = obj_to_index(cachep, page, (void *)ptr);
4179 BUG_ON(objnr >= cachep->num);
4180
4181 /* Find offset within object. */
4182 offset = ptr - index_to_obj(cachep, page, objnr) - obj_offset(cachep);
4183
4184 /* Allow address range falling entirely within usercopy region. */
4185 if (offset >= cachep->useroffset &&
4186 offset - cachep->useroffset <= cachep->usersize &&
4187 n <= cachep->useroffset - offset + cachep->usersize)
4188 return;
4189
4190 /*
4191 * If the copy is still within the allocated object, produce
4192 * a warning instead of rejecting the copy. This is intended
4193 * to be a temporary method to find any missing usercopy
4194 * whitelists.
4195 */
4196 if (usercopy_fallback &&
4197 offset <= cachep->object_size &&
4198 n <= cachep->object_size - offset) {
4199 usercopy_warn("SLAB object", cachep->name, to_user, offset, n);
4200 return;
4201 }
4202
4203 usercopy_abort("SLAB object", cachep->name, to_user, offset, n);
4204}
4205#endif /* CONFIG_HARDENED_USERCOPY */
4206
4207/**
4208 * __ksize -- Uninstrumented ksize.
4209 * @objp: pointer to the object
4210 *
4211 * Unlike ksize(), __ksize() is uninstrumented, and does not provide the same
4212 * safety checks as ksize() with KASAN instrumentation enabled.
4213 *
4214 * Return: size of the actual memory used by @objp in bytes
4215 */
4216size_t __ksize(const void *objp)
4217{
4218 struct kmem_cache *c;
4219 size_t size;
4220
4221 BUG_ON(!objp);
4222 if (unlikely(objp == ZERO_SIZE_PTR))
4223 return 0;
4224
4225 c = virt_to_cache(objp);
4226 size = c ? c->object_size : 0;
4227
4228 return size;
4229}
4230EXPORT_SYMBOL(__ksize);