Loading...
Note: File does not exist in v3.1.
1/*
2 * Slab allocator functions that are independent of the allocator strategy
3 *
4 * (C) 2012 Christoph Lameter <cl@linux.com>
5 */
6#include <linux/slab.h>
7
8#include <linux/mm.h>
9#include <linux/poison.h>
10#include <linux/interrupt.h>
11#include <linux/memory.h>
12#include <linux/compiler.h>
13#include <linux/module.h>
14#include <linux/cpu.h>
15#include <linux/uaccess.h>
16#include <linux/seq_file.h>
17#include <linux/proc_fs.h>
18#include <asm/cacheflush.h>
19#include <asm/tlbflush.h>
20#include <asm/page.h>
21#include <linux/memcontrol.h>
22#include <trace/events/kmem.h>
23
24#include "slab.h"
25
26enum slab_state slab_state;
27LIST_HEAD(slab_caches);
28DEFINE_MUTEX(slab_mutex);
29struct kmem_cache *kmem_cache;
30
31#ifdef CONFIG_DEBUG_VM
32static int kmem_cache_sanity_check(const char *name, size_t size)
33{
34 struct kmem_cache *s = NULL;
35
36 if (!name || in_interrupt() || size < sizeof(void *) ||
37 size > KMALLOC_MAX_SIZE) {
38 pr_err("kmem_cache_create(%s) integrity check failed\n", name);
39 return -EINVAL;
40 }
41
42 list_for_each_entry(s, &slab_caches, list) {
43 char tmp;
44 int res;
45
46 /*
47 * This happens when the module gets unloaded and doesn't
48 * destroy its slab cache and no-one else reuses the vmalloc
49 * area of the module. Print a warning.
50 */
51 res = probe_kernel_address(s->name, tmp);
52 if (res) {
53 pr_err("Slab cache with size %d has lost its name\n",
54 s->object_size);
55 continue;
56 }
57
58#if !defined(CONFIG_SLUB) || !defined(CONFIG_SLUB_DEBUG_ON)
59 if (!strcmp(s->name, name)) {
60 pr_err("%s (%s): Cache name already exists.\n",
61 __func__, name);
62 dump_stack();
63 s = NULL;
64 return -EINVAL;
65 }
66#endif
67 }
68
69 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
70 return 0;
71}
72#else
73static inline int kmem_cache_sanity_check(const char *name, size_t size)
74{
75 return 0;
76}
77#endif
78
79#ifdef CONFIG_MEMCG_KMEM
80int memcg_update_all_caches(int num_memcgs)
81{
82 struct kmem_cache *s;
83 int ret = 0;
84 mutex_lock(&slab_mutex);
85
86 list_for_each_entry(s, &slab_caches, list) {
87 if (!is_root_cache(s))
88 continue;
89
90 ret = memcg_update_cache_size(s, num_memcgs);
91 /*
92 * See comment in memcontrol.c, memcg_update_cache_size:
93 * Instead of freeing the memory, we'll just leave the caches
94 * up to this point in an updated state.
95 */
96 if (ret)
97 goto out;
98 }
99
100 memcg_update_array_size(num_memcgs);
101out:
102 mutex_unlock(&slab_mutex);
103 return ret;
104}
105#endif
106
107/*
108 * Figure out what the alignment of the objects will be given a set of
109 * flags, a user specified alignment and the size of the objects.
110 */
111unsigned long calculate_alignment(unsigned long flags,
112 unsigned long align, unsigned long size)
113{
114 /*
115 * If the user wants hardware cache aligned objects then follow that
116 * suggestion if the object is sufficiently large.
117 *
118 * The hardware cache alignment cannot override the specified
119 * alignment though. If that is greater then use it.
120 */
121 if (flags & SLAB_HWCACHE_ALIGN) {
122 unsigned long ralign = cache_line_size();
123 while (size <= ralign / 2)
124 ralign /= 2;
125 align = max(align, ralign);
126 }
127
128 if (align < ARCH_SLAB_MINALIGN)
129 align = ARCH_SLAB_MINALIGN;
130
131 return ALIGN(align, sizeof(void *));
132}
133
134static struct kmem_cache *
135do_kmem_cache_create(char *name, size_t object_size, size_t size, size_t align,
136 unsigned long flags, void (*ctor)(void *),
137 struct mem_cgroup *memcg, struct kmem_cache *root_cache)
138{
139 struct kmem_cache *s;
140 int err;
141
142 err = -ENOMEM;
143 s = kmem_cache_zalloc(kmem_cache, GFP_KERNEL);
144 if (!s)
145 goto out;
146
147 s->name = name;
148 s->object_size = object_size;
149 s->size = size;
150 s->align = align;
151 s->ctor = ctor;
152
153 err = memcg_alloc_cache_params(memcg, s, root_cache);
154 if (err)
155 goto out_free_cache;
156
157 err = __kmem_cache_create(s, flags);
158 if (err)
159 goto out_free_cache;
160
161 s->refcount = 1;
162 list_add(&s->list, &slab_caches);
163 memcg_register_cache(s);
164out:
165 if (err)
166 return ERR_PTR(err);
167 return s;
168
169out_free_cache:
170 memcg_free_cache_params(s);
171 kfree(s);
172 goto out;
173}
174
175/*
176 * kmem_cache_create - Create a cache.
177 * @name: A string which is used in /proc/slabinfo to identify this cache.
178 * @size: The size of objects to be created in this cache.
179 * @align: The required alignment for the objects.
180 * @flags: SLAB flags
181 * @ctor: A constructor for the objects.
182 *
183 * Returns a ptr to the cache on success, NULL on failure.
184 * Cannot be called within a interrupt, but can be interrupted.
185 * The @ctor is run when new pages are allocated by the cache.
186 *
187 * The flags are
188 *
189 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
190 * to catch references to uninitialised memory.
191 *
192 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
193 * for buffer overruns.
194 *
195 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
196 * cacheline. This can be beneficial if you're counting cycles as closely
197 * as davem.
198 */
199struct kmem_cache *
200kmem_cache_create(const char *name, size_t size, size_t align,
201 unsigned long flags, void (*ctor)(void *))
202{
203 struct kmem_cache *s;
204 char *cache_name;
205 int err;
206
207 get_online_cpus();
208 mutex_lock(&slab_mutex);
209
210 err = kmem_cache_sanity_check(name, size);
211 if (err)
212 goto out_unlock;
213
214 /*
215 * Some allocators will constraint the set of valid flags to a subset
216 * of all flags. We expect them to define CACHE_CREATE_MASK in this
217 * case, and we'll just provide them with a sanitized version of the
218 * passed flags.
219 */
220 flags &= CACHE_CREATE_MASK;
221
222 s = __kmem_cache_alias(name, size, align, flags, ctor);
223 if (s)
224 goto out_unlock;
225
226 cache_name = kstrdup(name, GFP_KERNEL);
227 if (!cache_name) {
228 err = -ENOMEM;
229 goto out_unlock;
230 }
231
232 s = do_kmem_cache_create(cache_name, size, size,
233 calculate_alignment(flags, align, size),
234 flags, ctor, NULL, NULL);
235 if (IS_ERR(s)) {
236 err = PTR_ERR(s);
237 kfree(cache_name);
238 }
239
240out_unlock:
241 mutex_unlock(&slab_mutex);
242 put_online_cpus();
243
244 if (err) {
245 if (flags & SLAB_PANIC)
246 panic("kmem_cache_create: Failed to create slab '%s'. Error %d\n",
247 name, err);
248 else {
249 printk(KERN_WARNING "kmem_cache_create(%s) failed with error %d",
250 name, err);
251 dump_stack();
252 }
253 return NULL;
254 }
255 return s;
256}
257EXPORT_SYMBOL(kmem_cache_create);
258
259#ifdef CONFIG_MEMCG_KMEM
260/*
261 * kmem_cache_create_memcg - Create a cache for a memory cgroup.
262 * @memcg: The memory cgroup the new cache is for.
263 * @root_cache: The parent of the new cache.
264 *
265 * This function attempts to create a kmem cache that will serve allocation
266 * requests going from @memcg to @root_cache. The new cache inherits properties
267 * from its parent.
268 */
269void kmem_cache_create_memcg(struct mem_cgroup *memcg, struct kmem_cache *root_cache)
270{
271 struct kmem_cache *s;
272 char *cache_name;
273
274 get_online_cpus();
275 mutex_lock(&slab_mutex);
276
277 /*
278 * Since per-memcg caches are created asynchronously on first
279 * allocation (see memcg_kmem_get_cache()), several threads can try to
280 * create the same cache, but only one of them may succeed.
281 */
282 if (cache_from_memcg_idx(root_cache, memcg_cache_id(memcg)))
283 goto out_unlock;
284
285 cache_name = memcg_create_cache_name(memcg, root_cache);
286 if (!cache_name)
287 goto out_unlock;
288
289 s = do_kmem_cache_create(cache_name, root_cache->object_size,
290 root_cache->size, root_cache->align,
291 root_cache->flags, root_cache->ctor,
292 memcg, root_cache);
293 if (IS_ERR(s)) {
294 kfree(cache_name);
295 goto out_unlock;
296 }
297
298 s->allocflags |= __GFP_KMEMCG;
299
300out_unlock:
301 mutex_unlock(&slab_mutex);
302 put_online_cpus();
303}
304
305static int kmem_cache_destroy_memcg_children(struct kmem_cache *s)
306{
307 int rc;
308
309 if (!s->memcg_params ||
310 !s->memcg_params->is_root_cache)
311 return 0;
312
313 mutex_unlock(&slab_mutex);
314 rc = __kmem_cache_destroy_memcg_children(s);
315 mutex_lock(&slab_mutex);
316
317 return rc;
318}
319#else
320static int kmem_cache_destroy_memcg_children(struct kmem_cache *s)
321{
322 return 0;
323}
324#endif /* CONFIG_MEMCG_KMEM */
325
326void slab_kmem_cache_release(struct kmem_cache *s)
327{
328 kfree(s->name);
329 kmem_cache_free(kmem_cache, s);
330}
331
332void kmem_cache_destroy(struct kmem_cache *s)
333{
334 get_online_cpus();
335 mutex_lock(&slab_mutex);
336
337 s->refcount--;
338 if (s->refcount)
339 goto out_unlock;
340
341 if (kmem_cache_destroy_memcg_children(s) != 0)
342 goto out_unlock;
343
344 list_del(&s->list);
345 memcg_unregister_cache(s);
346
347 if (__kmem_cache_shutdown(s) != 0) {
348 list_add(&s->list, &slab_caches);
349 memcg_register_cache(s);
350 printk(KERN_ERR "kmem_cache_destroy %s: "
351 "Slab cache still has objects\n", s->name);
352 dump_stack();
353 goto out_unlock;
354 }
355
356 mutex_unlock(&slab_mutex);
357 if (s->flags & SLAB_DESTROY_BY_RCU)
358 rcu_barrier();
359
360 memcg_free_cache_params(s);
361#ifdef SLAB_SUPPORTS_SYSFS
362 sysfs_slab_remove(s);
363#else
364 slab_kmem_cache_release(s);
365#endif
366 goto out_put_cpus;
367
368out_unlock:
369 mutex_unlock(&slab_mutex);
370out_put_cpus:
371 put_online_cpus();
372}
373EXPORT_SYMBOL(kmem_cache_destroy);
374
375int slab_is_available(void)
376{
377 return slab_state >= UP;
378}
379
380#ifndef CONFIG_SLOB
381/* Create a cache during boot when no slab services are available yet */
382void __init create_boot_cache(struct kmem_cache *s, const char *name, size_t size,
383 unsigned long flags)
384{
385 int err;
386
387 s->name = name;
388 s->size = s->object_size = size;
389 s->align = calculate_alignment(flags, ARCH_KMALLOC_MINALIGN, size);
390 err = __kmem_cache_create(s, flags);
391
392 if (err)
393 panic("Creation of kmalloc slab %s size=%zu failed. Reason %d\n",
394 name, size, err);
395
396 s->refcount = -1; /* Exempt from merging for now */
397}
398
399struct kmem_cache *__init create_kmalloc_cache(const char *name, size_t size,
400 unsigned long flags)
401{
402 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
403
404 if (!s)
405 panic("Out of memory when creating slab %s\n", name);
406
407 create_boot_cache(s, name, size, flags);
408 list_add(&s->list, &slab_caches);
409 s->refcount = 1;
410 return s;
411}
412
413struct kmem_cache *kmalloc_caches[KMALLOC_SHIFT_HIGH + 1];
414EXPORT_SYMBOL(kmalloc_caches);
415
416#ifdef CONFIG_ZONE_DMA
417struct kmem_cache *kmalloc_dma_caches[KMALLOC_SHIFT_HIGH + 1];
418EXPORT_SYMBOL(kmalloc_dma_caches);
419#endif
420
421/*
422 * Conversion table for small slabs sizes / 8 to the index in the
423 * kmalloc array. This is necessary for slabs < 192 since we have non power
424 * of two cache sizes there. The size of larger slabs can be determined using
425 * fls.
426 */
427static s8 size_index[24] = {
428 3, /* 8 */
429 4, /* 16 */
430 5, /* 24 */
431 5, /* 32 */
432 6, /* 40 */
433 6, /* 48 */
434 6, /* 56 */
435 6, /* 64 */
436 1, /* 72 */
437 1, /* 80 */
438 1, /* 88 */
439 1, /* 96 */
440 7, /* 104 */
441 7, /* 112 */
442 7, /* 120 */
443 7, /* 128 */
444 2, /* 136 */
445 2, /* 144 */
446 2, /* 152 */
447 2, /* 160 */
448 2, /* 168 */
449 2, /* 176 */
450 2, /* 184 */
451 2 /* 192 */
452};
453
454static inline int size_index_elem(size_t bytes)
455{
456 return (bytes - 1) / 8;
457}
458
459/*
460 * Find the kmem_cache structure that serves a given size of
461 * allocation
462 */
463struct kmem_cache *kmalloc_slab(size_t size, gfp_t flags)
464{
465 int index;
466
467 if (unlikely(size > KMALLOC_MAX_SIZE)) {
468 WARN_ON_ONCE(!(flags & __GFP_NOWARN));
469 return NULL;
470 }
471
472 if (size <= 192) {
473 if (!size)
474 return ZERO_SIZE_PTR;
475
476 index = size_index[size_index_elem(size)];
477 } else
478 index = fls(size - 1);
479
480#ifdef CONFIG_ZONE_DMA
481 if (unlikely((flags & GFP_DMA)))
482 return kmalloc_dma_caches[index];
483
484#endif
485 return kmalloc_caches[index];
486}
487
488/*
489 * Create the kmalloc array. Some of the regular kmalloc arrays
490 * may already have been created because they were needed to
491 * enable allocations for slab creation.
492 */
493void __init create_kmalloc_caches(unsigned long flags)
494{
495 int i;
496
497 /*
498 * Patch up the size_index table if we have strange large alignment
499 * requirements for the kmalloc array. This is only the case for
500 * MIPS it seems. The standard arches will not generate any code here.
501 *
502 * Largest permitted alignment is 256 bytes due to the way we
503 * handle the index determination for the smaller caches.
504 *
505 * Make sure that nothing crazy happens if someone starts tinkering
506 * around with ARCH_KMALLOC_MINALIGN
507 */
508 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
509 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
510
511 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
512 int elem = size_index_elem(i);
513
514 if (elem >= ARRAY_SIZE(size_index))
515 break;
516 size_index[elem] = KMALLOC_SHIFT_LOW;
517 }
518
519 if (KMALLOC_MIN_SIZE >= 64) {
520 /*
521 * The 96 byte size cache is not used if the alignment
522 * is 64 byte.
523 */
524 for (i = 64 + 8; i <= 96; i += 8)
525 size_index[size_index_elem(i)] = 7;
526
527 }
528
529 if (KMALLOC_MIN_SIZE >= 128) {
530 /*
531 * The 192 byte sized cache is not used if the alignment
532 * is 128 byte. Redirect kmalloc to use the 256 byte cache
533 * instead.
534 */
535 for (i = 128 + 8; i <= 192; i += 8)
536 size_index[size_index_elem(i)] = 8;
537 }
538 for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++) {
539 if (!kmalloc_caches[i]) {
540 kmalloc_caches[i] = create_kmalloc_cache(NULL,
541 1 << i, flags);
542 }
543
544 /*
545 * Caches that are not of the two-to-the-power-of size.
546 * These have to be created immediately after the
547 * earlier power of two caches
548 */
549 if (KMALLOC_MIN_SIZE <= 32 && !kmalloc_caches[1] && i == 6)
550 kmalloc_caches[1] = create_kmalloc_cache(NULL, 96, flags);
551
552 if (KMALLOC_MIN_SIZE <= 64 && !kmalloc_caches[2] && i == 7)
553 kmalloc_caches[2] = create_kmalloc_cache(NULL, 192, flags);
554 }
555
556 /* Kmalloc array is now usable */
557 slab_state = UP;
558
559 for (i = 0; i <= KMALLOC_SHIFT_HIGH; i++) {
560 struct kmem_cache *s = kmalloc_caches[i];
561 char *n;
562
563 if (s) {
564 n = kasprintf(GFP_NOWAIT, "kmalloc-%d", kmalloc_size(i));
565
566 BUG_ON(!n);
567 s->name = n;
568 }
569 }
570
571#ifdef CONFIG_ZONE_DMA
572 for (i = 0; i <= KMALLOC_SHIFT_HIGH; i++) {
573 struct kmem_cache *s = kmalloc_caches[i];
574
575 if (s) {
576 int size = kmalloc_size(i);
577 char *n = kasprintf(GFP_NOWAIT,
578 "dma-kmalloc-%d", size);
579
580 BUG_ON(!n);
581 kmalloc_dma_caches[i] = create_kmalloc_cache(n,
582 size, SLAB_CACHE_DMA | flags);
583 }
584 }
585#endif
586}
587#endif /* !CONFIG_SLOB */
588
589#ifdef CONFIG_TRACING
590void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
591{
592 void *ret = kmalloc_order(size, flags, order);
593 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
594 return ret;
595}
596EXPORT_SYMBOL(kmalloc_order_trace);
597#endif
598
599#ifdef CONFIG_SLABINFO
600
601#ifdef CONFIG_SLAB
602#define SLABINFO_RIGHTS (S_IWUSR | S_IRUSR)
603#else
604#define SLABINFO_RIGHTS S_IRUSR
605#endif
606
607void print_slabinfo_header(struct seq_file *m)
608{
609 /*
610 * Output format version, so at least we can change it
611 * without _too_ many complaints.
612 */
613#ifdef CONFIG_DEBUG_SLAB
614 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
615#else
616 seq_puts(m, "slabinfo - version: 2.1\n");
617#endif
618 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
619 "<objperslab> <pagesperslab>");
620 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
621 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
622#ifdef CONFIG_DEBUG_SLAB
623 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
624 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
625 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
626#endif
627 seq_putc(m, '\n');
628}
629
630static void *s_start(struct seq_file *m, loff_t *pos)
631{
632 loff_t n = *pos;
633
634 mutex_lock(&slab_mutex);
635 if (!n)
636 print_slabinfo_header(m);
637
638 return seq_list_start(&slab_caches, *pos);
639}
640
641void *slab_next(struct seq_file *m, void *p, loff_t *pos)
642{
643 return seq_list_next(p, &slab_caches, pos);
644}
645
646void slab_stop(struct seq_file *m, void *p)
647{
648 mutex_unlock(&slab_mutex);
649}
650
651static void
652memcg_accumulate_slabinfo(struct kmem_cache *s, struct slabinfo *info)
653{
654 struct kmem_cache *c;
655 struct slabinfo sinfo;
656 int i;
657
658 if (!is_root_cache(s))
659 return;
660
661 for_each_memcg_cache_index(i) {
662 c = cache_from_memcg_idx(s, i);
663 if (!c)
664 continue;
665
666 memset(&sinfo, 0, sizeof(sinfo));
667 get_slabinfo(c, &sinfo);
668
669 info->active_slabs += sinfo.active_slabs;
670 info->num_slabs += sinfo.num_slabs;
671 info->shared_avail += sinfo.shared_avail;
672 info->active_objs += sinfo.active_objs;
673 info->num_objs += sinfo.num_objs;
674 }
675}
676
677int cache_show(struct kmem_cache *s, struct seq_file *m)
678{
679 struct slabinfo sinfo;
680
681 memset(&sinfo, 0, sizeof(sinfo));
682 get_slabinfo(s, &sinfo);
683
684 memcg_accumulate_slabinfo(s, &sinfo);
685
686 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
687 cache_name(s), sinfo.active_objs, sinfo.num_objs, s->size,
688 sinfo.objects_per_slab, (1 << sinfo.cache_order));
689
690 seq_printf(m, " : tunables %4u %4u %4u",
691 sinfo.limit, sinfo.batchcount, sinfo.shared);
692 seq_printf(m, " : slabdata %6lu %6lu %6lu",
693 sinfo.active_slabs, sinfo.num_slabs, sinfo.shared_avail);
694 slabinfo_show_stats(m, s);
695 seq_putc(m, '\n');
696 return 0;
697}
698
699static int s_show(struct seq_file *m, void *p)
700{
701 struct kmem_cache *s = list_entry(p, struct kmem_cache, list);
702
703 if (!is_root_cache(s))
704 return 0;
705 return cache_show(s, m);
706}
707
708/*
709 * slabinfo_op - iterator that generates /proc/slabinfo
710 *
711 * Output layout:
712 * cache-name
713 * num-active-objs
714 * total-objs
715 * object size
716 * num-active-slabs
717 * total-slabs
718 * num-pages-per-slab
719 * + further values on SMP and with statistics enabled
720 */
721static const struct seq_operations slabinfo_op = {
722 .start = s_start,
723 .next = slab_next,
724 .stop = slab_stop,
725 .show = s_show,
726};
727
728static int slabinfo_open(struct inode *inode, struct file *file)
729{
730 return seq_open(file, &slabinfo_op);
731}
732
733static const struct file_operations proc_slabinfo_operations = {
734 .open = slabinfo_open,
735 .read = seq_read,
736 .write = slabinfo_write,
737 .llseek = seq_lseek,
738 .release = seq_release,
739};
740
741static int __init slab_proc_init(void)
742{
743 proc_create("slabinfo", SLABINFO_RIGHTS, NULL,
744 &proc_slabinfo_operations);
745 return 0;
746}
747module_init(slab_proc_init);
748#endif /* CONFIG_SLABINFO */