2 * Slab allocator functions that are independent of the allocator strategy
4 * (C) 2012 Christoph Lameter <cl@linux.com>
6 #include <linux/slab.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>
21 #include <linux/memcontrol.h>
22 #include <trace/events/kmem.h>
26 enum slab_state slab_state;
27 LIST_HEAD(slab_caches);
28 DEFINE_MUTEX(slab_mutex);
29 struct kmem_cache *kmem_cache;
31 #ifdef CONFIG_DEBUG_VM
32 static int kmem_cache_sanity_check(struct mem_cgroup *memcg, const char *name,
35 struct kmem_cache *s = NULL;
37 if (!name || in_interrupt() || size < sizeof(void *) ||
38 size > KMALLOC_MAX_SIZE) {
39 pr_err("kmem_cache_create(%s) integrity check failed\n", name);
43 list_for_each_entry(s, &slab_caches, list) {
48 * This happens when the module gets unloaded and doesn't
49 * destroy its slab cache and no-one else reuses the vmalloc
50 * area of the module. Print a warning.
52 res = probe_kernel_address(s->name, tmp);
54 pr_err("Slab cache with size %d has lost its name\n",
60 * For simplicity, we won't check this in the list of memcg
61 * caches. We have control over memcg naming, and if there
62 * aren't duplicates in the global list, there won't be any
63 * duplicates in the memcg lists as well.
65 if (!memcg && !strcmp(s->name, name)) {
66 pr_err("%s (%s): Cache name already exists.\n",
74 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
78 static inline int kmem_cache_sanity_check(struct mem_cgroup *memcg,
79 const char *name, size_t size)
85 #ifdef CONFIG_MEMCG_KMEM
86 int memcg_update_all_caches(int num_memcgs)
90 mutex_lock(&slab_mutex);
92 list_for_each_entry(s, &slab_caches, list) {
93 if (!is_root_cache(s))
96 ret = memcg_update_cache_size(s, num_memcgs);
98 * See comment in memcontrol.c, memcg_update_cache_size:
99 * Instead of freeing the memory, we'll just leave the caches
100 * up to this point in an updated state.
106 memcg_update_array_size(num_memcgs);
108 mutex_unlock(&slab_mutex);
114 * Figure out what the alignment of the objects will be given a set of
115 * flags, a user specified alignment and the size of the objects.
117 unsigned long calculate_alignment(unsigned long flags,
118 unsigned long align, unsigned long size)
121 * If the user wants hardware cache aligned objects then follow that
122 * suggestion if the object is sufficiently large.
124 * The hardware cache alignment cannot override the specified
125 * alignment though. If that is greater then use it.
127 if (flags & SLAB_HWCACHE_ALIGN) {
128 unsigned long ralign = cache_line_size();
129 while (size <= ralign / 2)
131 align = max(align, ralign);
134 if (align < ARCH_SLAB_MINALIGN)
135 align = ARCH_SLAB_MINALIGN;
137 return ALIGN(align, sizeof(void *));
142 * kmem_cache_create - Create a cache.
143 * @name: A string which is used in /proc/slabinfo to identify this cache.
144 * @size: The size of objects to be created in this cache.
145 * @align: The required alignment for the objects.
147 * @ctor: A constructor for the objects.
149 * Returns a ptr to the cache on success, NULL on failure.
150 * Cannot be called within a interrupt, but can be interrupted.
151 * The @ctor is run when new pages are allocated by the cache.
155 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
156 * to catch references to uninitialised memory.
158 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
159 * for buffer overruns.
161 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
162 * cacheline. This can be beneficial if you're counting cycles as closely
167 kmem_cache_create_memcg(struct mem_cgroup *memcg, const char *name, size_t size,
168 size_t align, unsigned long flags, void (*ctor)(void *),
169 struct kmem_cache *parent_cache)
171 struct kmem_cache *s = NULL;
175 mutex_lock(&slab_mutex);
177 if (!kmem_cache_sanity_check(memcg, name, size) == 0)
181 * Some allocators will constraint the set of valid flags to a subset
182 * of all flags. We expect them to define CACHE_CREATE_MASK in this
183 * case, and we'll just provide them with a sanitized version of the
186 flags &= CACHE_CREATE_MASK;
188 s = __kmem_cache_alias(memcg, name, size, align, flags, ctor);
192 s = kmem_cache_zalloc(kmem_cache, GFP_KERNEL);
194 s->object_size = s->size = size;
195 s->align = calculate_alignment(flags, align, size);
198 if (memcg_register_cache(memcg, s, parent_cache)) {
199 kmem_cache_free(kmem_cache, s);
204 s->name = kstrdup(name, GFP_KERNEL);
206 kmem_cache_free(kmem_cache, s);
211 err = __kmem_cache_create(s, flags);
214 list_add(&s->list, &slab_caches);
215 memcg_cache_list_add(memcg, s);
218 kmem_cache_free(kmem_cache, s);
224 mutex_unlock(&slab_mutex);
229 if (flags & SLAB_PANIC)
230 panic("kmem_cache_create: Failed to create slab '%s'. Error %d\n",
233 printk(KERN_WARNING "kmem_cache_create(%s) failed with error %d",
245 kmem_cache_create(const char *name, size_t size, size_t align,
246 unsigned long flags, void (*ctor)(void *))
248 return kmem_cache_create_memcg(NULL, name, size, align, flags, ctor, NULL);
250 EXPORT_SYMBOL(kmem_cache_create);
252 void kmem_cache_destroy(struct kmem_cache *s)
254 /* Destroy all the children caches if we aren't a memcg cache */
255 kmem_cache_destroy_memcg_children(s);
258 mutex_lock(&slab_mutex);
263 if (!__kmem_cache_shutdown(s)) {
264 mutex_unlock(&slab_mutex);
265 if (s->flags & SLAB_DESTROY_BY_RCU)
268 memcg_release_cache(s);
270 kmem_cache_free(kmem_cache, s);
272 list_add(&s->list, &slab_caches);
273 mutex_unlock(&slab_mutex);
274 printk(KERN_ERR "kmem_cache_destroy %s: Slab cache still has objects\n",
279 mutex_unlock(&slab_mutex);
283 EXPORT_SYMBOL(kmem_cache_destroy);
285 int slab_is_available(void)
287 return slab_state >= UP;
291 /* Create a cache during boot when no slab services are available yet */
292 void __init create_boot_cache(struct kmem_cache *s, const char *name, size_t size,
298 s->size = s->object_size = size;
299 s->align = calculate_alignment(flags, ARCH_KMALLOC_MINALIGN, size);
300 err = __kmem_cache_create(s, flags);
303 panic("Creation of kmalloc slab %s size=%zu failed. Reason %d\n",
306 s->refcount = -1; /* Exempt from merging for now */
309 struct kmem_cache *__init create_kmalloc_cache(const char *name, size_t size,
312 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
315 panic("Out of memory when creating slab %s\n", name);
317 create_boot_cache(s, name, size, flags);
318 list_add(&s->list, &slab_caches);
323 struct kmem_cache *kmalloc_caches[KMALLOC_SHIFT_HIGH + 1];
324 EXPORT_SYMBOL(kmalloc_caches);
326 #ifdef CONFIG_ZONE_DMA
327 struct kmem_cache *kmalloc_dma_caches[KMALLOC_SHIFT_HIGH + 1];
328 EXPORT_SYMBOL(kmalloc_dma_caches);
332 * Conversion table for small slabs sizes / 8 to the index in the
333 * kmalloc array. This is necessary for slabs < 192 since we have non power
334 * of two cache sizes there. The size of larger slabs can be determined using
337 static s8 size_index[24] = {
364 static inline int size_index_elem(size_t bytes)
366 return (bytes - 1) / 8;
370 * Find the kmem_cache structure that serves a given size of
373 struct kmem_cache *kmalloc_slab(size_t size, gfp_t flags)
377 if (unlikely(size > KMALLOC_MAX_SIZE)) {
378 WARN_ON_ONCE(!(flags & __GFP_NOWARN));
384 return ZERO_SIZE_PTR;
386 index = size_index[size_index_elem(size)];
388 index = fls(size - 1);
390 #ifdef CONFIG_ZONE_DMA
391 if (unlikely((flags & GFP_DMA)))
392 return kmalloc_dma_caches[index];
395 return kmalloc_caches[index];
399 * Create the kmalloc array. Some of the regular kmalloc arrays
400 * may already have been created because they were needed to
401 * enable allocations for slab creation.
403 void __init create_kmalloc_caches(unsigned long flags)
408 * Patch up the size_index table if we have strange large alignment
409 * requirements for the kmalloc array. This is only the case for
410 * MIPS it seems. The standard arches will not generate any code here.
412 * Largest permitted alignment is 256 bytes due to the way we
413 * handle the index determination for the smaller caches.
415 * Make sure that nothing crazy happens if someone starts tinkering
416 * around with ARCH_KMALLOC_MINALIGN
418 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
419 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
421 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
422 int elem = size_index_elem(i);
424 if (elem >= ARRAY_SIZE(size_index))
426 size_index[elem] = KMALLOC_SHIFT_LOW;
429 if (KMALLOC_MIN_SIZE >= 64) {
431 * The 96 byte size cache is not used if the alignment
434 for (i = 64 + 8; i <= 96; i += 8)
435 size_index[size_index_elem(i)] = 7;
439 if (KMALLOC_MIN_SIZE >= 128) {
441 * The 192 byte sized cache is not used if the alignment
442 * is 128 byte. Redirect kmalloc to use the 256 byte cache
445 for (i = 128 + 8; i <= 192; i += 8)
446 size_index[size_index_elem(i)] = 8;
448 for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++) {
449 if (!kmalloc_caches[i]) {
450 kmalloc_caches[i] = create_kmalloc_cache(NULL,
455 * Caches that are not of the two-to-the-power-of size.
456 * These have to be created immediately after the
457 * earlier power of two caches
459 if (KMALLOC_MIN_SIZE <= 32 && !kmalloc_caches[1] && i == 6)
460 kmalloc_caches[1] = create_kmalloc_cache(NULL, 96, flags);
462 if (KMALLOC_MIN_SIZE <= 64 && !kmalloc_caches[2] && i == 7)
463 kmalloc_caches[2] = create_kmalloc_cache(NULL, 192, flags);
466 /* Kmalloc array is now usable */
469 for (i = 0; i <= KMALLOC_SHIFT_HIGH; i++) {
470 struct kmem_cache *s = kmalloc_caches[i];
474 n = kasprintf(GFP_NOWAIT, "kmalloc-%d", kmalloc_size(i));
481 #ifdef CONFIG_ZONE_DMA
482 for (i = 0; i <= KMALLOC_SHIFT_HIGH; i++) {
483 struct kmem_cache *s = kmalloc_caches[i];
486 int size = kmalloc_size(i);
487 char *n = kasprintf(GFP_NOWAIT,
488 "dma-kmalloc-%d", size);
491 kmalloc_dma_caches[i] = create_kmalloc_cache(n,
492 size, SLAB_CACHE_DMA | flags);
497 #endif /* !CONFIG_SLOB */
499 #ifdef CONFIG_TRACING
500 void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
502 void *ret = kmalloc_order(size, flags, order);
503 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
506 EXPORT_SYMBOL(kmalloc_order_trace);
509 #ifdef CONFIG_SLABINFO
512 #define SLABINFO_RIGHTS (S_IWUSR | S_IRUSR)
514 #define SLABINFO_RIGHTS S_IRUSR
517 void print_slabinfo_header(struct seq_file *m)
520 * Output format version, so at least we can change it
521 * without _too_ many complaints.
523 #ifdef CONFIG_DEBUG_SLAB
524 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
526 seq_puts(m, "slabinfo - version: 2.1\n");
528 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
529 "<objperslab> <pagesperslab>");
530 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
531 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
532 #ifdef CONFIG_DEBUG_SLAB
533 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
534 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
535 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
540 static void *s_start(struct seq_file *m, loff_t *pos)
544 mutex_lock(&slab_mutex);
546 print_slabinfo_header(m);
548 return seq_list_start(&slab_caches, *pos);
551 void *slab_next(struct seq_file *m, void *p, loff_t *pos)
553 return seq_list_next(p, &slab_caches, pos);
556 void slab_stop(struct seq_file *m, void *p)
558 mutex_unlock(&slab_mutex);
562 memcg_accumulate_slabinfo(struct kmem_cache *s, struct slabinfo *info)
564 struct kmem_cache *c;
565 struct slabinfo sinfo;
568 if (!is_root_cache(s))
571 for_each_memcg_cache_index(i) {
572 c = cache_from_memcg(s, i);
576 memset(&sinfo, 0, sizeof(sinfo));
577 get_slabinfo(c, &sinfo);
579 info->active_slabs += sinfo.active_slabs;
580 info->num_slabs += sinfo.num_slabs;
581 info->shared_avail += sinfo.shared_avail;
582 info->active_objs += sinfo.active_objs;
583 info->num_objs += sinfo.num_objs;
587 int cache_show(struct kmem_cache *s, struct seq_file *m)
589 struct slabinfo sinfo;
591 memset(&sinfo, 0, sizeof(sinfo));
592 get_slabinfo(s, &sinfo);
594 memcg_accumulate_slabinfo(s, &sinfo);
596 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
597 cache_name(s), sinfo.active_objs, sinfo.num_objs, s->size,
598 sinfo.objects_per_slab, (1 << sinfo.cache_order));
600 seq_printf(m, " : tunables %4u %4u %4u",
601 sinfo.limit, sinfo.batchcount, sinfo.shared);
602 seq_printf(m, " : slabdata %6lu %6lu %6lu",
603 sinfo.active_slabs, sinfo.num_slabs, sinfo.shared_avail);
604 slabinfo_show_stats(m, s);
609 static int s_show(struct seq_file *m, void *p)
611 struct kmem_cache *s = list_entry(p, struct kmem_cache, list);
613 if (!is_root_cache(s))
615 return cache_show(s, m);
619 * slabinfo_op - iterator that generates /proc/slabinfo
629 * + further values on SMP and with statistics enabled
631 static const struct seq_operations slabinfo_op = {
638 static int slabinfo_open(struct inode *inode, struct file *file)
640 return seq_open(file, &slabinfo_op);
643 static const struct file_operations proc_slabinfo_operations = {
644 .open = slabinfo_open,
646 .write = slabinfo_write,
648 .release = seq_release,
651 static int __init slab_proc_init(void)
653 proc_create("slabinfo", SLABINFO_RIGHTS, NULL,
654 &proc_slabinfo_operations);
657 module_init(slab_proc_init);
658 #endif /* CONFIG_SLABINFO */