xref: /openbmc/linux/mm/slab_common.c (revision ee8a99bd)
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 
23 #include "slab.h"
24 
25 enum slab_state slab_state;
26 LIST_HEAD(slab_caches);
27 DEFINE_MUTEX(slab_mutex);
28 struct kmem_cache *kmem_cache;
29 
30 #ifdef CONFIG_DEBUG_VM
31 static int kmem_cache_sanity_check(struct mem_cgroup *memcg, const char *name,
32 				   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 		/*
59 		 * For simplicity, we won't check this in the list of memcg
60 		 * caches. We have control over memcg naming, and if there
61 		 * aren't duplicates in the global list, there won't be any
62 		 * duplicates in the memcg lists as well.
63 		 */
64 		if (!memcg && !strcmp(s->name, name)) {
65 			pr_err("%s (%s): Cache name already exists.\n",
66 			       __func__, name);
67 			dump_stack();
68 			s = NULL;
69 			return -EINVAL;
70 		}
71 	}
72 
73 	WARN_ON(strchr(name, ' '));	/* It confuses parsers */
74 	return 0;
75 }
76 #else
77 static inline int kmem_cache_sanity_check(struct mem_cgroup *memcg,
78 					  const char *name, size_t size)
79 {
80 	return 0;
81 }
82 #endif
83 
84 #ifdef CONFIG_MEMCG_KMEM
85 int memcg_update_all_caches(int num_memcgs)
86 {
87 	struct kmem_cache *s;
88 	int ret = 0;
89 	mutex_lock(&slab_mutex);
90 
91 	list_for_each_entry(s, &slab_caches, list) {
92 		if (!is_root_cache(s))
93 			continue;
94 
95 		ret = memcg_update_cache_size(s, num_memcgs);
96 		/*
97 		 * See comment in memcontrol.c, memcg_update_cache_size:
98 		 * Instead of freeing the memory, we'll just leave the caches
99 		 * up to this point in an updated state.
100 		 */
101 		if (ret)
102 			goto out;
103 	}
104 
105 	memcg_update_array_size(num_memcgs);
106 out:
107 	mutex_unlock(&slab_mutex);
108 	return ret;
109 }
110 #endif
111 
112 /*
113  * Figure out what the alignment of the objects will be given a set of
114  * flags, a user specified alignment and the size of the objects.
115  */
116 unsigned long calculate_alignment(unsigned long flags,
117 		unsigned long align, unsigned long size)
118 {
119 	/*
120 	 * If the user wants hardware cache aligned objects then follow that
121 	 * suggestion if the object is sufficiently large.
122 	 *
123 	 * The hardware cache alignment cannot override the specified
124 	 * alignment though. If that is greater then use it.
125 	 */
126 	if (flags & SLAB_HWCACHE_ALIGN) {
127 		unsigned long ralign = cache_line_size();
128 		while (size <= ralign / 2)
129 			ralign /= 2;
130 		align = max(align, ralign);
131 	}
132 
133 	if (align < ARCH_SLAB_MINALIGN)
134 		align = ARCH_SLAB_MINALIGN;
135 
136 	return ALIGN(align, sizeof(void *));
137 }
138 
139 
140 /*
141  * kmem_cache_create - Create a cache.
142  * @name: A string which is used in /proc/slabinfo to identify this cache.
143  * @size: The size of objects to be created in this cache.
144  * @align: The required alignment for the objects.
145  * @flags: SLAB flags
146  * @ctor: A constructor for the objects.
147  *
148  * Returns a ptr to the cache on success, NULL on failure.
149  * Cannot be called within a interrupt, but can be interrupted.
150  * The @ctor is run when new pages are allocated by the cache.
151  *
152  * The flags are
153  *
154  * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
155  * to catch references to uninitialised memory.
156  *
157  * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
158  * for buffer overruns.
159  *
160  * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
161  * cacheline.  This can be beneficial if you're counting cycles as closely
162  * as davem.
163  */
164 
165 struct kmem_cache *
166 kmem_cache_create_memcg(struct mem_cgroup *memcg, const char *name, size_t size,
167 			size_t align, unsigned long flags, void (*ctor)(void *),
168 			struct kmem_cache *parent_cache)
169 {
170 	struct kmem_cache *s = NULL;
171 	int err = 0;
172 
173 	get_online_cpus();
174 	mutex_lock(&slab_mutex);
175 
176 	if (!kmem_cache_sanity_check(memcg, name, size) == 0)
177 		goto out_locked;
178 
179 	/*
180 	 * Some allocators will constraint the set of valid flags to a subset
181 	 * of all flags. We expect them to define CACHE_CREATE_MASK in this
182 	 * case, and we'll just provide them with a sanitized version of the
183 	 * passed flags.
184 	 */
185 	flags &= CACHE_CREATE_MASK;
186 
187 	s = __kmem_cache_alias(memcg, name, size, align, flags, ctor);
188 	if (s)
189 		goto out_locked;
190 
191 	s = kmem_cache_zalloc(kmem_cache, GFP_KERNEL);
192 	if (s) {
193 		s->object_size = s->size = size;
194 		s->align = calculate_alignment(flags, align, size);
195 		s->ctor = ctor;
196 
197 		if (memcg_register_cache(memcg, s, parent_cache)) {
198 			kmem_cache_free(kmem_cache, s);
199 			err = -ENOMEM;
200 			goto out_locked;
201 		}
202 
203 		s->name = kstrdup(name, GFP_KERNEL);
204 		if (!s->name) {
205 			kmem_cache_free(kmem_cache, s);
206 			err = -ENOMEM;
207 			goto out_locked;
208 		}
209 
210 		err = __kmem_cache_create(s, flags);
211 		if (!err) {
212 			s->refcount = 1;
213 			list_add(&s->list, &slab_caches);
214 			memcg_cache_list_add(memcg, s);
215 		} else {
216 			kfree(s->name);
217 			kmem_cache_free(kmem_cache, s);
218 		}
219 	} else
220 		err = -ENOMEM;
221 
222 out_locked:
223 	mutex_unlock(&slab_mutex);
224 	put_online_cpus();
225 
226 	if (err) {
227 
228 		if (flags & SLAB_PANIC)
229 			panic("kmem_cache_create: Failed to create slab '%s'. Error %d\n",
230 				name, err);
231 		else {
232 			printk(KERN_WARNING "kmem_cache_create(%s) failed with error %d",
233 				name, err);
234 			dump_stack();
235 		}
236 
237 		return NULL;
238 	}
239 
240 	return s;
241 }
242 
243 struct kmem_cache *
244 kmem_cache_create(const char *name, size_t size, size_t align,
245 		  unsigned long flags, void (*ctor)(void *))
246 {
247 	return kmem_cache_create_memcg(NULL, name, size, align, flags, ctor, NULL);
248 }
249 EXPORT_SYMBOL(kmem_cache_create);
250 
251 void kmem_cache_destroy(struct kmem_cache *s)
252 {
253 	/* Destroy all the children caches if we aren't a memcg cache */
254 	kmem_cache_destroy_memcg_children(s);
255 
256 	get_online_cpus();
257 	mutex_lock(&slab_mutex);
258 	s->refcount--;
259 	if (!s->refcount) {
260 		list_del(&s->list);
261 
262 		if (!__kmem_cache_shutdown(s)) {
263 			mutex_unlock(&slab_mutex);
264 			if (s->flags & SLAB_DESTROY_BY_RCU)
265 				rcu_barrier();
266 
267 			memcg_release_cache(s);
268 			kfree(s->name);
269 			kmem_cache_free(kmem_cache, s);
270 		} else {
271 			list_add(&s->list, &slab_caches);
272 			mutex_unlock(&slab_mutex);
273 			printk(KERN_ERR "kmem_cache_destroy %s: Slab cache still has objects\n",
274 				s->name);
275 			dump_stack();
276 		}
277 	} else {
278 		mutex_unlock(&slab_mutex);
279 	}
280 	put_online_cpus();
281 }
282 EXPORT_SYMBOL(kmem_cache_destroy);
283 
284 int slab_is_available(void)
285 {
286 	return slab_state >= UP;
287 }
288 
289 #ifndef CONFIG_SLOB
290 /* Create a cache during boot when no slab services are available yet */
291 void __init create_boot_cache(struct kmem_cache *s, const char *name, size_t size,
292 		unsigned long flags)
293 {
294 	int err;
295 
296 	s->name = name;
297 	s->size = s->object_size = size;
298 	s->align = calculate_alignment(flags, ARCH_KMALLOC_MINALIGN, size);
299 	err = __kmem_cache_create(s, flags);
300 
301 	if (err)
302 		panic("Creation of kmalloc slab %s size=%zu failed. Reason %d\n",
303 					name, size, err);
304 
305 	s->refcount = -1;	/* Exempt from merging for now */
306 }
307 
308 struct kmem_cache *__init create_kmalloc_cache(const char *name, size_t size,
309 				unsigned long flags)
310 {
311 	struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
312 
313 	if (!s)
314 		panic("Out of memory when creating slab %s\n", name);
315 
316 	create_boot_cache(s, name, size, flags);
317 	list_add(&s->list, &slab_caches);
318 	s->refcount = 1;
319 	return s;
320 }
321 
322 struct kmem_cache *kmalloc_caches[KMALLOC_SHIFT_HIGH + 1];
323 EXPORT_SYMBOL(kmalloc_caches);
324 
325 #ifdef CONFIG_ZONE_DMA
326 struct kmem_cache *kmalloc_dma_caches[KMALLOC_SHIFT_HIGH + 1];
327 EXPORT_SYMBOL(kmalloc_dma_caches);
328 #endif
329 
330 /*
331  * Conversion table for small slabs sizes / 8 to the index in the
332  * kmalloc array. This is necessary for slabs < 192 since we have non power
333  * of two cache sizes there. The size of larger slabs can be determined using
334  * fls.
335  */
336 static s8 size_index[24] = {
337 	3,	/* 8 */
338 	4,	/* 16 */
339 	5,	/* 24 */
340 	5,	/* 32 */
341 	6,	/* 40 */
342 	6,	/* 48 */
343 	6,	/* 56 */
344 	6,	/* 64 */
345 	1,	/* 72 */
346 	1,	/* 80 */
347 	1,	/* 88 */
348 	1,	/* 96 */
349 	7,	/* 104 */
350 	7,	/* 112 */
351 	7,	/* 120 */
352 	7,	/* 128 */
353 	2,	/* 136 */
354 	2,	/* 144 */
355 	2,	/* 152 */
356 	2,	/* 160 */
357 	2,	/* 168 */
358 	2,	/* 176 */
359 	2,	/* 184 */
360 	2	/* 192 */
361 };
362 
363 static inline int size_index_elem(size_t bytes)
364 {
365 	return (bytes - 1) / 8;
366 }
367 
368 /*
369  * Find the kmem_cache structure that serves a given size of
370  * allocation
371  */
372 struct kmem_cache *kmalloc_slab(size_t size, gfp_t flags)
373 {
374 	int index;
375 
376 	if (size > KMALLOC_MAX_SIZE) {
377 		WARN_ON_ONCE(!(flags & __GFP_NOWARN));
378 		return NULL;
379 	}
380 
381 	if (size <= 192) {
382 		if (!size)
383 			return ZERO_SIZE_PTR;
384 
385 		index = size_index[size_index_elem(size)];
386 	} else
387 		index = fls(size - 1);
388 
389 #ifdef CONFIG_ZONE_DMA
390 	if (unlikely((flags & GFP_DMA)))
391 		return kmalloc_dma_caches[index];
392 
393 #endif
394 	return kmalloc_caches[index];
395 }
396 
397 /*
398  * Create the kmalloc array. Some of the regular kmalloc arrays
399  * may already have been created because they were needed to
400  * enable allocations for slab creation.
401  */
402 void __init create_kmalloc_caches(unsigned long flags)
403 {
404 	int i;
405 
406 	/*
407 	 * Patch up the size_index table if we have strange large alignment
408 	 * requirements for the kmalloc array. This is only the case for
409 	 * MIPS it seems. The standard arches will not generate any code here.
410 	 *
411 	 * Largest permitted alignment is 256 bytes due to the way we
412 	 * handle the index determination for the smaller caches.
413 	 *
414 	 * Make sure that nothing crazy happens if someone starts tinkering
415 	 * around with ARCH_KMALLOC_MINALIGN
416 	 */
417 	BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
418 		(KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
419 
420 	for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
421 		int elem = size_index_elem(i);
422 
423 		if (elem >= ARRAY_SIZE(size_index))
424 			break;
425 		size_index[elem] = KMALLOC_SHIFT_LOW;
426 	}
427 
428 	if (KMALLOC_MIN_SIZE >= 64) {
429 		/*
430 		 * The 96 byte size cache is not used if the alignment
431 		 * is 64 byte.
432 		 */
433 		for (i = 64 + 8; i <= 96; i += 8)
434 			size_index[size_index_elem(i)] = 7;
435 
436 	}
437 
438 	if (KMALLOC_MIN_SIZE >= 128) {
439 		/*
440 		 * The 192 byte sized cache is not used if the alignment
441 		 * is 128 byte. Redirect kmalloc to use the 256 byte cache
442 		 * instead.
443 		 */
444 		for (i = 128 + 8; i <= 192; i += 8)
445 			size_index[size_index_elem(i)] = 8;
446 	}
447 	for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++) {
448 		if (!kmalloc_caches[i]) {
449 			kmalloc_caches[i] = create_kmalloc_cache(NULL,
450 							1 << i, flags);
451 		}
452 
453 		/*
454 		 * Caches that are not of the two-to-the-power-of size.
455 		 * These have to be created immediately after the
456 		 * earlier power of two caches
457 		 */
458 		if (KMALLOC_MIN_SIZE <= 32 && !kmalloc_caches[1] && i == 6)
459 			kmalloc_caches[1] = create_kmalloc_cache(NULL, 96, flags);
460 
461 		if (KMALLOC_MIN_SIZE <= 64 && !kmalloc_caches[2] && i == 7)
462 			kmalloc_caches[2] = create_kmalloc_cache(NULL, 192, flags);
463 	}
464 
465 	/* Kmalloc array is now usable */
466 	slab_state = UP;
467 
468 	for (i = 0; i <= KMALLOC_SHIFT_HIGH; i++) {
469 		struct kmem_cache *s = kmalloc_caches[i];
470 		char *n;
471 
472 		if (s) {
473 			n = kasprintf(GFP_NOWAIT, "kmalloc-%d", kmalloc_size(i));
474 
475 			BUG_ON(!n);
476 			s->name = n;
477 		}
478 	}
479 
480 #ifdef CONFIG_ZONE_DMA
481 	for (i = 0; i <= KMALLOC_SHIFT_HIGH; i++) {
482 		struct kmem_cache *s = kmalloc_caches[i];
483 
484 		if (s) {
485 			int size = kmalloc_size(i);
486 			char *n = kasprintf(GFP_NOWAIT,
487 				 "dma-kmalloc-%d", size);
488 
489 			BUG_ON(!n);
490 			kmalloc_dma_caches[i] = create_kmalloc_cache(n,
491 				size, SLAB_CACHE_DMA | flags);
492 		}
493 	}
494 #endif
495 }
496 #endif /* !CONFIG_SLOB */
497 
498 
499 #ifdef CONFIG_SLABINFO
500 
501 #ifdef CONFIG_SLAB
502 #define SLABINFO_RIGHTS (S_IWUSR | S_IRUSR)
503 #else
504 #define SLABINFO_RIGHTS S_IRUSR
505 #endif
506 
507 void print_slabinfo_header(struct seq_file *m)
508 {
509 	/*
510 	 * Output format version, so at least we can change it
511 	 * without _too_ many complaints.
512 	 */
513 #ifdef CONFIG_DEBUG_SLAB
514 	seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
515 #else
516 	seq_puts(m, "slabinfo - version: 2.1\n");
517 #endif
518 	seq_puts(m, "# name            <active_objs> <num_objs> <objsize> "
519 		 "<objperslab> <pagesperslab>");
520 	seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
521 	seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
522 #ifdef CONFIG_DEBUG_SLAB
523 	seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
524 		 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
525 	seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
526 #endif
527 	seq_putc(m, '\n');
528 }
529 
530 static void *s_start(struct seq_file *m, loff_t *pos)
531 {
532 	loff_t n = *pos;
533 
534 	mutex_lock(&slab_mutex);
535 	if (!n)
536 		print_slabinfo_header(m);
537 
538 	return seq_list_start(&slab_caches, *pos);
539 }
540 
541 void *slab_next(struct seq_file *m, void *p, loff_t *pos)
542 {
543 	return seq_list_next(p, &slab_caches, pos);
544 }
545 
546 void slab_stop(struct seq_file *m, void *p)
547 {
548 	mutex_unlock(&slab_mutex);
549 }
550 
551 static void
552 memcg_accumulate_slabinfo(struct kmem_cache *s, struct slabinfo *info)
553 {
554 	struct kmem_cache *c;
555 	struct slabinfo sinfo;
556 	int i;
557 
558 	if (!is_root_cache(s))
559 		return;
560 
561 	for_each_memcg_cache_index(i) {
562 		c = cache_from_memcg(s, i);
563 		if (!c)
564 			continue;
565 
566 		memset(&sinfo, 0, sizeof(sinfo));
567 		get_slabinfo(c, &sinfo);
568 
569 		info->active_slabs += sinfo.active_slabs;
570 		info->num_slabs += sinfo.num_slabs;
571 		info->shared_avail += sinfo.shared_avail;
572 		info->active_objs += sinfo.active_objs;
573 		info->num_objs += sinfo.num_objs;
574 	}
575 }
576 
577 int cache_show(struct kmem_cache *s, struct seq_file *m)
578 {
579 	struct slabinfo sinfo;
580 
581 	memset(&sinfo, 0, sizeof(sinfo));
582 	get_slabinfo(s, &sinfo);
583 
584 	memcg_accumulate_slabinfo(s, &sinfo);
585 
586 	seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
587 		   cache_name(s), sinfo.active_objs, sinfo.num_objs, s->size,
588 		   sinfo.objects_per_slab, (1 << sinfo.cache_order));
589 
590 	seq_printf(m, " : tunables %4u %4u %4u",
591 		   sinfo.limit, sinfo.batchcount, sinfo.shared);
592 	seq_printf(m, " : slabdata %6lu %6lu %6lu",
593 		   sinfo.active_slabs, sinfo.num_slabs, sinfo.shared_avail);
594 	slabinfo_show_stats(m, s);
595 	seq_putc(m, '\n');
596 	return 0;
597 }
598 
599 static int s_show(struct seq_file *m, void *p)
600 {
601 	struct kmem_cache *s = list_entry(p, struct kmem_cache, list);
602 
603 	if (!is_root_cache(s))
604 		return 0;
605 	return cache_show(s, m);
606 }
607 
608 /*
609  * slabinfo_op - iterator that generates /proc/slabinfo
610  *
611  * Output layout:
612  * cache-name
613  * num-active-objs
614  * total-objs
615  * object size
616  * num-active-slabs
617  * total-slabs
618  * num-pages-per-slab
619  * + further values on SMP and with statistics enabled
620  */
621 static const struct seq_operations slabinfo_op = {
622 	.start = s_start,
623 	.next = slab_next,
624 	.stop = slab_stop,
625 	.show = s_show,
626 };
627 
628 static int slabinfo_open(struct inode *inode, struct file *file)
629 {
630 	return seq_open(file, &slabinfo_op);
631 }
632 
633 static const struct file_operations proc_slabinfo_operations = {
634 	.open		= slabinfo_open,
635 	.read		= seq_read,
636 	.write          = slabinfo_write,
637 	.llseek		= seq_lseek,
638 	.release	= seq_release,
639 };
640 
641 static int __init slab_proc_init(void)
642 {
643 	proc_create("slabinfo", SLABINFO_RIGHTS, NULL,
644 						&proc_slabinfo_operations);
645 	return 0;
646 }
647 module_init(slab_proc_init);
648 #endif /* CONFIG_SLABINFO */
649