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",
59 #if !defined(CONFIG_SLUB) || !defined(CONFIG_SLUB_DEBUG_ON)
61 * For simplicity, we won't check this in the list of memcg
62 * caches. We have control over memcg naming, and if there
63 * aren't duplicates in the global list, there won't be any
64 * duplicates in the memcg lists as well.
66 if (!memcg && !strcmp(s->name, name)) {
67 pr_err("%s (%s): Cache name already exists.\n",
76 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
80 static inline int kmem_cache_sanity_check(struct mem_cgroup *memcg,
81 const char *name, size_t size)
87 #ifdef CONFIG_MEMCG_KMEM
88 int memcg_update_all_caches(int num_memcgs)
92 mutex_lock(&slab_mutex);
94 list_for_each_entry(s, &slab_caches, list) {
95 if (!is_root_cache(s))
98 ret = memcg_update_cache_size(s, num_memcgs);
100 * See comment in memcontrol.c, memcg_update_cache_size:
101 * Instead of freeing the memory, we'll just leave the caches
102 * up to this point in an updated state.
108 memcg_update_array_size(num_memcgs);
110 mutex_unlock(&slab_mutex);
116 * Figure out what the alignment of the objects will be given a set of
117 * flags, a user specified alignment and the size of the objects.
119 unsigned long calculate_alignment(unsigned long flags,
120 unsigned long align, unsigned long size)
123 * If the user wants hardware cache aligned objects then follow that
124 * suggestion if the object is sufficiently large.
126 * The hardware cache alignment cannot override the specified
127 * alignment though. If that is greater then use it.
129 if (flags & SLAB_HWCACHE_ALIGN) {
130 unsigned long ralign = cache_line_size();
131 while (size <= ralign / 2)
133 align = max(align, ralign);
136 if (align < ARCH_SLAB_MINALIGN)
137 align = ARCH_SLAB_MINALIGN;
139 return ALIGN(align, sizeof(void *));
144 * kmem_cache_create - Create a cache.
145 * @name: A string which is used in /proc/slabinfo to identify this cache.
146 * @size: The size of objects to be created in this cache.
147 * @align: The required alignment for the objects.
149 * @ctor: A constructor for the objects.
151 * Returns a ptr to the cache on success, NULL on failure.
152 * Cannot be called within a interrupt, but can be interrupted.
153 * The @ctor is run when new pages are allocated by the cache.
157 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
158 * to catch references to uninitialised memory.
160 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
161 * for buffer overruns.
163 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
164 * cacheline. This can be beneficial if you're counting cycles as closely
169 kmem_cache_create_memcg(struct mem_cgroup *memcg, const char *name, size_t size,
170 size_t align, unsigned long flags, void (*ctor)(void *),
171 struct kmem_cache *parent_cache)
173 struct kmem_cache *s = NULL;
177 mutex_lock(&slab_mutex);
179 err = kmem_cache_sanity_check(memcg, name, size);
185 * Since per-memcg caches are created asynchronously on first
186 * allocation (see memcg_kmem_get_cache()), several threads can
187 * try to create the same cache, but only one of them may
188 * succeed. Therefore if we get here and see the cache has
189 * already been created, we silently return NULL.
191 if (cache_from_memcg_idx(parent_cache, memcg_cache_id(memcg)))
196 * Some allocators will constraint the set of valid flags to a subset
197 * of all flags. We expect them to define CACHE_CREATE_MASK in this
198 * case, and we'll just provide them with a sanitized version of the
201 flags &= CACHE_CREATE_MASK;
203 s = __kmem_cache_alias(memcg, name, size, align, flags, ctor);
208 s = kmem_cache_zalloc(kmem_cache, GFP_KERNEL);
212 s->object_size = s->size = size;
213 s->align = calculate_alignment(flags, align, size);
216 s->name = kstrdup(name, GFP_KERNEL);
220 err = memcg_alloc_cache_params(memcg, s, parent_cache);
224 err = __kmem_cache_create(s, flags);
229 list_add(&s->list, &slab_caches);
230 memcg_register_cache(s);
233 mutex_unlock(&slab_mutex);
238 * There is no point in flooding logs with warnings or
239 * especially crashing the system if we fail to create a cache
240 * for a memcg. In this case we will be accounting the memcg
241 * allocation to the root cgroup until we succeed to create its
242 * own cache, but it isn't that critical.
247 if (flags & SLAB_PANIC)
248 panic("kmem_cache_create: Failed to create slab '%s'. Error %d\n",
251 printk(KERN_WARNING "kmem_cache_create(%s) failed with error %d",
260 memcg_free_cache_params(s);
262 kmem_cache_free(kmem_cache, s);
267 kmem_cache_create(const char *name, size_t size, size_t align,
268 unsigned long flags, void (*ctor)(void *))
270 return kmem_cache_create_memcg(NULL, name, size, align, flags, ctor, NULL);
272 EXPORT_SYMBOL(kmem_cache_create);
274 void kmem_cache_destroy(struct kmem_cache *s)
276 /* Destroy all the children caches if we aren't a memcg cache */
277 kmem_cache_destroy_memcg_children(s);
280 mutex_lock(&slab_mutex);
285 if (!__kmem_cache_shutdown(s)) {
286 memcg_unregister_cache(s);
287 mutex_unlock(&slab_mutex);
288 if (s->flags & SLAB_DESTROY_BY_RCU)
291 memcg_free_cache_params(s);
293 kmem_cache_free(kmem_cache, s);
295 list_add(&s->list, &slab_caches);
296 mutex_unlock(&slab_mutex);
297 printk(KERN_ERR "kmem_cache_destroy %s: Slab cache still has objects\n",
302 mutex_unlock(&slab_mutex);
306 EXPORT_SYMBOL(kmem_cache_destroy);
308 int slab_is_available(void)
310 return slab_state >= UP;
314 /* Create a cache during boot when no slab services are available yet */
315 void __init create_boot_cache(struct kmem_cache *s, const char *name, size_t size,
321 s->size = s->object_size = size;
322 s->align = calculate_alignment(flags, ARCH_KMALLOC_MINALIGN, size);
323 err = __kmem_cache_create(s, flags);
326 panic("Creation of kmalloc slab %s size=%zu failed. Reason %d\n",
329 s->refcount = -1; /* Exempt from merging for now */
332 struct kmem_cache *__init create_kmalloc_cache(const char *name, size_t size,
335 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
338 panic("Out of memory when creating slab %s\n", name);
340 create_boot_cache(s, name, size, flags);
341 list_add(&s->list, &slab_caches);
346 struct kmem_cache *kmalloc_caches[KMALLOC_SHIFT_HIGH + 1];
347 EXPORT_SYMBOL(kmalloc_caches);
349 #ifdef CONFIG_ZONE_DMA
350 struct kmem_cache *kmalloc_dma_caches[KMALLOC_SHIFT_HIGH + 1];
351 EXPORT_SYMBOL(kmalloc_dma_caches);
355 * Conversion table for small slabs sizes / 8 to the index in the
356 * kmalloc array. This is necessary for slabs < 192 since we have non power
357 * of two cache sizes there. The size of larger slabs can be determined using
360 static s8 size_index[24] = {
387 static inline int size_index_elem(size_t bytes)
389 return (bytes - 1) / 8;
393 * Find the kmem_cache structure that serves a given size of
396 struct kmem_cache *kmalloc_slab(size_t size, gfp_t flags)
400 if (unlikely(size > KMALLOC_MAX_SIZE)) {
401 WARN_ON_ONCE(!(flags & __GFP_NOWARN));
407 return ZERO_SIZE_PTR;
409 index = size_index[size_index_elem(size)];
411 index = fls(size - 1);
413 #ifdef CONFIG_ZONE_DMA
414 if (unlikely((flags & GFP_DMA)))
415 return kmalloc_dma_caches[index];
418 return kmalloc_caches[index];
422 * Create the kmalloc array. Some of the regular kmalloc arrays
423 * may already have been created because they were needed to
424 * enable allocations for slab creation.
426 void __init create_kmalloc_caches(unsigned long flags)
431 * Patch up the size_index table if we have strange large alignment
432 * requirements for the kmalloc array. This is only the case for
433 * MIPS it seems. The standard arches will not generate any code here.
435 * Largest permitted alignment is 256 bytes due to the way we
436 * handle the index determination for the smaller caches.
438 * Make sure that nothing crazy happens if someone starts tinkering
439 * around with ARCH_KMALLOC_MINALIGN
441 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
442 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
444 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
445 int elem = size_index_elem(i);
447 if (elem >= ARRAY_SIZE(size_index))
449 size_index[elem] = KMALLOC_SHIFT_LOW;
452 if (KMALLOC_MIN_SIZE >= 64) {
454 * The 96 byte size cache is not used if the alignment
457 for (i = 64 + 8; i <= 96; i += 8)
458 size_index[size_index_elem(i)] = 7;
462 if (KMALLOC_MIN_SIZE >= 128) {
464 * The 192 byte sized cache is not used if the alignment
465 * is 128 byte. Redirect kmalloc to use the 256 byte cache
468 for (i = 128 + 8; i <= 192; i += 8)
469 size_index[size_index_elem(i)] = 8;
471 for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++) {
472 if (!kmalloc_caches[i]) {
473 kmalloc_caches[i] = create_kmalloc_cache(NULL,
478 * Caches that are not of the two-to-the-power-of size.
479 * These have to be created immediately after the
480 * earlier power of two caches
482 if (KMALLOC_MIN_SIZE <= 32 && !kmalloc_caches[1] && i == 6)
483 kmalloc_caches[1] = create_kmalloc_cache(NULL, 96, flags);
485 if (KMALLOC_MIN_SIZE <= 64 && !kmalloc_caches[2] && i == 7)
486 kmalloc_caches[2] = create_kmalloc_cache(NULL, 192, flags);
489 /* Kmalloc array is now usable */
492 for (i = 0; i <= KMALLOC_SHIFT_HIGH; i++) {
493 struct kmem_cache *s = kmalloc_caches[i];
497 n = kasprintf(GFP_NOWAIT, "kmalloc-%d", kmalloc_size(i));
504 #ifdef CONFIG_ZONE_DMA
505 for (i = 0; i <= KMALLOC_SHIFT_HIGH; i++) {
506 struct kmem_cache *s = kmalloc_caches[i];
509 int size = kmalloc_size(i);
510 char *n = kasprintf(GFP_NOWAIT,
511 "dma-kmalloc-%d", size);
514 kmalloc_dma_caches[i] = create_kmalloc_cache(n,
515 size, SLAB_CACHE_DMA | flags);
520 #endif /* !CONFIG_SLOB */
522 #ifdef CONFIG_TRACING
523 void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
525 void *ret = kmalloc_order(size, flags, order);
526 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
529 EXPORT_SYMBOL(kmalloc_order_trace);
532 #ifdef CONFIG_SLABINFO
535 #define SLABINFO_RIGHTS (S_IWUSR | S_IRUSR)
537 #define SLABINFO_RIGHTS S_IRUSR
540 void print_slabinfo_header(struct seq_file *m)
543 * Output format version, so at least we can change it
544 * without _too_ many complaints.
546 #ifdef CONFIG_DEBUG_SLAB
547 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
549 seq_puts(m, "slabinfo - version: 2.1\n");
551 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
552 "<objperslab> <pagesperslab>");
553 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
554 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
555 #ifdef CONFIG_DEBUG_SLAB
556 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
557 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
558 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
563 static void *s_start(struct seq_file *m, loff_t *pos)
567 mutex_lock(&slab_mutex);
569 print_slabinfo_header(m);
571 return seq_list_start(&slab_caches, *pos);
574 void *slab_next(struct seq_file *m, void *p, loff_t *pos)
576 return seq_list_next(p, &slab_caches, pos);
579 void slab_stop(struct seq_file *m, void *p)
581 mutex_unlock(&slab_mutex);
585 memcg_accumulate_slabinfo(struct kmem_cache *s, struct slabinfo *info)
587 struct kmem_cache *c;
588 struct slabinfo sinfo;
591 if (!is_root_cache(s))
594 for_each_memcg_cache_index(i) {
595 c = cache_from_memcg_idx(s, i);
599 memset(&sinfo, 0, sizeof(sinfo));
600 get_slabinfo(c, &sinfo);
602 info->active_slabs += sinfo.active_slabs;
603 info->num_slabs += sinfo.num_slabs;
604 info->shared_avail += sinfo.shared_avail;
605 info->active_objs += sinfo.active_objs;
606 info->num_objs += sinfo.num_objs;
610 int cache_show(struct kmem_cache *s, struct seq_file *m)
612 struct slabinfo sinfo;
614 memset(&sinfo, 0, sizeof(sinfo));
615 get_slabinfo(s, &sinfo);
617 memcg_accumulate_slabinfo(s, &sinfo);
619 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
620 cache_name(s), sinfo.active_objs, sinfo.num_objs, s->size,
621 sinfo.objects_per_slab, (1 << sinfo.cache_order));
623 seq_printf(m, " : tunables %4u %4u %4u",
624 sinfo.limit, sinfo.batchcount, sinfo.shared);
625 seq_printf(m, " : slabdata %6lu %6lu %6lu",
626 sinfo.active_slabs, sinfo.num_slabs, sinfo.shared_avail);
627 slabinfo_show_stats(m, s);
632 static int s_show(struct seq_file *m, void *p)
634 struct kmem_cache *s = list_entry(p, struct kmem_cache, list);
636 if (!is_root_cache(s))
638 return cache_show(s, m);
642 * slabinfo_op - iterator that generates /proc/slabinfo
652 * + further values on SMP and with statistics enabled
654 static const struct seq_operations slabinfo_op = {
661 static int slabinfo_open(struct inode *inode, struct file *file)
663 return seq_open(file, &slabinfo_op);
666 static const struct file_operations proc_slabinfo_operations = {
667 .open = slabinfo_open,
669 .write = slabinfo_write,
671 .release = seq_release,
674 static int __init slab_proc_init(void)
676 proc_create("slabinfo", SLABINFO_RIGHTS, NULL,
677 &proc_slabinfo_operations);
680 module_init(slab_proc_init);
681 #endif /* CONFIG_SLABINFO */