mm/slab_common.c

   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
  26 enum slab_state slab_state;
  27 LIST_HEAD(slab_caches);
  28 DEFINE_MUTEX(slab_mutex);
  29 struct kmem_cache *kmem_cache;
  30
  31 #ifdef CONFIG_DEBUG_VM
  32 static int kmem_cache_sanity_check(struct mem_cgroup *memcg, const char *name,
  33                                    size_t size)
  34 {
  35         struct kmem_cache *s = NULL;
  36
  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);
  40                 return -EINVAL;
  41         }
  42
  43         list_for_each_entry(s, &slab_caches, list) {
  44                 char tmp;
  45                 int res;
  46
  47                 /*
  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.
  51                  */
  52                 res = probe_kernel_address(s->name, tmp);
  53                 if (res) {
  54                         pr_err("Slab cache with size %d has lost its name\n",
  55                                s->object_size);
  56                         continue;
  57                 }
  58
  59 #if !defined(CONFIG_SLUB) || !defined(CONFIG_SLUB_DEBUG_ON)
  60                 /*
  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.
  65                  */
  66                 if (!memcg && !strcmp(s->name, name)) {
  67                         pr_err("%s (%s): Cache name already exists.\n",
  68                                __func__, name);
  69                         dump_stack();
  70                         s = NULL;
  71                         return -EINVAL;
  72                 }
  73 #endif
  74         }
  75
  76         WARN_ON(strchr(name, ' '));     /* It confuses parsers */
  77         return 0;
  78 }
  79 #else
  80 static inline int kmem_cache_sanity_check(struct mem_cgroup *memcg,
  81                                           const char *name, size_t size)
  82 {
  83         return 0;
  84 }
  85 #endif
  86
  87 #ifdef CONFIG_MEMCG_KMEM
  88 int memcg_update_all_caches(int num_memcgs)
  89 {
  90         struct kmem_cache *s;
  91         int ret = 0;
  92         mutex_lock(&slab_mutex);
  93
  94         list_for_each_entry(s, &slab_caches, list) {
  95                 if (!is_root_cache(s))
  96                         continue;
  97
  98                 ret = memcg_update_cache_size(s, num_memcgs);
  99                 /*
 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.
 103                  */
 104                 if (ret)
 105                         goto out;
 106         }
 107
 108         memcg_update_array_size(num_memcgs);
 109 out:
 110         mutex_unlock(&slab_mutex);
 111         return ret;
 112 }
 113 #endif
 114
 115 /*
 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.
 118  */
 119 unsigned long calculate_alignment(unsigned long flags,
 120                 unsigned long align, unsigned long size)
 121 {
 122         /*
 123          * If the user wants hardware cache aligned objects then follow that
 124          * suggestion if the object is sufficiently large.
 125          *
 126          * The hardware cache alignment cannot override the specified
 127          * alignment though. If that is greater then use it.
 128          */
 129         if (flags & SLAB_HWCACHE_ALIGN) {
 130                 unsigned long ralign = cache_line_size();
 131                 while (size <= ralign / 2)
 132                         ralign /= 2;
 133                 align = max(align, ralign);
 134         }
 135
 136         if (align < ARCH_SLAB_MINALIGN)
 137                 align = ARCH_SLAB_MINALIGN;
 138
 139         return ALIGN(align, sizeof(void *));
 140 }
 141
 142
 143 /*
 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.
 148  * @flags: SLAB flags
 149  * @ctor: A constructor for the objects.
 150  *
 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.
 154  *
 155  * The flags are
 156  *
 157  * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
 158  * to catch references to uninitialised memory.
 159  *
 160  * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
 161  * for buffer overruns.
 162  *
 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
 165  * as davem.
 166  */
 167
 168 struct kmem_cache *
 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)
 172 {
 173         struct kmem_cache *s = NULL;
 174         int err;
 175
 176         get_online_cpus();
 177         mutex_lock(&slab_mutex);
 178
 179         err = kmem_cache_sanity_check(memcg, name, size);
 180         if (err)
 181                 goto out_unlock;
 182
 183         if (memcg) {
 184                 /*
 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.
 190                  */
 191                 if (cache_from_memcg_idx(parent_cache, memcg_cache_id(memcg)))
 192                         goto out_unlock;
 193         }
 194
 195         /*
 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
 199          * passed flags.
 200          */
 201         flags &= CACHE_CREATE_MASK;
 202
 203         s = __kmem_cache_alias(memcg, name, size, align, flags, ctor);
 204         if (s)
 205                 goto out_unlock;
 206
 207         err = -ENOMEM;
 208         s = kmem_cache_zalloc(kmem_cache, GFP_KERNEL);
 209         if (!s)
 210                 goto out_unlock;
 211
 212         s->object_size = s->size = size;
 213         s->align = calculate_alignment(flags, align, size);
 214         s->ctor = ctor;
 215
 216         s->name = kstrdup(name, GFP_KERNEL);
 217         if (!s->name)
 218                 goto out_free_cache;
 219
 220         err = memcg_alloc_cache_params(memcg, s, parent_cache);
 221         if (err)
 222                 goto out_free_cache;
 223
 224         err = __kmem_cache_create(s, flags);
 225         if (err)
 226                 goto out_free_cache;
 227
 228         s->refcount = 1;
 229         list_add(&s->list, &slab_caches);
 230         memcg_register_cache(s);
 231
 232 out_unlock:
 233         mutex_unlock(&slab_mutex);
 234         put_online_cpus();
 235
 236         if (err) {
 237                 /*
 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.
 243                  */
 244                 if (!memcg)
 245                         return NULL;
 246
 247                 if (flags & SLAB_PANIC)
 248                         panic("kmem_cache_create: Failed to create slab '%s'. Error %d\n",
 249                                 name, err);
 250                 else {
 251                         printk(KERN_WARNING "kmem_cache_create(%s) failed with error %d",
 252                                 name, err);
 253                         dump_stack();
 254                 }
 255                 return NULL;
 256         }
 257         return s;
 258
 259 out_free_cache:
 260         memcg_free_cache_params(s);
 261         kfree(s->name);
 262         kmem_cache_free(kmem_cache, s);
 263         goto out_unlock;
 264 }
 265
 266 struct kmem_cache *
 267 kmem_cache_create(const char *name, size_t size, size_t align,
 268                   unsigned long flags, void (*ctor)(void *))
 269 {
 270         return kmem_cache_create_memcg(NULL, name, size, align, flags, ctor, NULL);
 271 }
 272 EXPORT_SYMBOL(kmem_cache_create);
 273
 274 void kmem_cache_destroy(struct kmem_cache *s)
 275 {
 276         /* Destroy all the children caches if we aren't a memcg cache */
 277         kmem_cache_destroy_memcg_children(s);
 278
 279         get_online_cpus();
 280         mutex_lock(&slab_mutex);
 281         s->refcount--;
 282         if (!s->refcount) {
 283                 list_del(&s->list);
 284
 285                 if (!__kmem_cache_shutdown(s)) {
 286                         memcg_unregister_cache(s);
 287                         mutex_unlock(&slab_mutex);
 288                         if (s->flags & SLAB_DESTROY_BY_RCU)
 289                                 rcu_barrier();
 290
 291                         memcg_free_cache_params(s);
 292                         kfree(s->name);
 293                         kmem_cache_free(kmem_cache, s);
 294                 } else {
 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",
 298                                 s->name);
 299                         dump_stack();
 300                 }
 301         } else {
 302                 mutex_unlock(&slab_mutex);
 303         }
 304         put_online_cpus();
 305 }
 306 EXPORT_SYMBOL(kmem_cache_destroy);
 307
 308 int slab_is_available(void)
 309 {
 310         return slab_state >= UP;
 311 }
 312
 313 #ifndef CONFIG_SLOB
 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,
 316                 unsigned long flags)
 317 {
 318         int err;
 319
 320         s->name = name;
 321         s->size = s->object_size = size;
 322         s->align = calculate_alignment(flags, ARCH_KMALLOC_MINALIGN, size);
 323         err = __kmem_cache_create(s, flags);
 324
 325         if (err)
 326                 panic("Creation of kmalloc slab %s size=%zu failed. Reason %d\n",
 327                                         name, size, err);
 328
 329         s->refcount = -1;       /* Exempt from merging for now */
 330 }
 331
 332 struct kmem_cache *__init create_kmalloc_cache(const char *name, size_t size,
 333                                 unsigned long flags)
 334 {
 335         struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
 336
 337         if (!s)
 338                 panic("Out of memory when creating slab %s\n", name);
 339
 340         create_boot_cache(s, name, size, flags);
 341         list_add(&s->list, &slab_caches);
 342         s->refcount = 1;
 343         return s;
 344 }
 345
 346 struct kmem_cache *kmalloc_caches[KMALLOC_SHIFT_HIGH + 1];
 347 EXPORT_SYMBOL(kmalloc_caches);
 348
 349 #ifdef CONFIG_ZONE_DMA
 350 struct kmem_cache *kmalloc_dma_caches[KMALLOC_SHIFT_HIGH + 1];
 351 EXPORT_SYMBOL(kmalloc_dma_caches);
 352 #endif
 353
 354 /*
 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
 358  * fls.
 359  */
 360 static s8 size_index[24] = {
 361         3,      /* 8 */
 362         4,      /* 16 */
 363         5,      /* 24 */
 364         5,      /* 32 */
 365         6,      /* 40 */
 366         6,      /* 48 */
 367         6,      /* 56 */
 368         6,      /* 64 */
 369         1,      /* 72 */
 370         1,      /* 80 */
 371         1,      /* 88 */
 372         1,      /* 96 */
 373         7,      /* 104 */
 374         7,      /* 112 */
 375         7,      /* 120 */
 376         7,      /* 128 */
 377         2,      /* 136 */
 378         2,      /* 144 */
 379         2,      /* 152 */
 380         2,      /* 160 */
 381         2,      /* 168 */
 382         2,      /* 176 */
 383         2,      /* 184 */
 384         2       /* 192 */
 385 };
 386
 387 static inline int size_index_elem(size_t bytes)
 388 {
 389         return (bytes - 1) / 8;
 390 }
 391
 392 /*
 393  * Find the kmem_cache structure that serves a given size of
 394  * allocation
 395  */
 396 struct kmem_cache *kmalloc_slab(size_t size, gfp_t flags)
 397 {
 398         int index;
 399
 400         if (unlikely(size > KMALLOC_MAX_SIZE)) {
 401                 WARN_ON_ONCE(!(flags & __GFP_NOWARN));
 402                 return NULL;
 403         }
 404
 405         if (size <= 192) {
 406                 if (!size)
 407                         return ZERO_SIZE_PTR;
 408
 409                 index = size_index[size_index_elem(size)];
 410         } else
 411                 index = fls(size - 1);
 412
 413 #ifdef CONFIG_ZONE_DMA
 414         if (unlikely((flags & GFP_DMA)))
 415                 return kmalloc_dma_caches[index];
 416
 417 #endif
 418         return kmalloc_caches[index];
 419 }
 420
 421 /*
 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.
 425  */
 426 void __init create_kmalloc_caches(unsigned long flags)
 427 {
 428         int i;
 429
 430         /*
 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.
 434          *
 435          * Largest permitted alignment is 256 bytes due to the way we
 436          * handle the index determination for the smaller caches.
 437          *
 438          * Make sure that nothing crazy happens if someone starts tinkering
 439          * around with ARCH_KMALLOC_MINALIGN
 440          */
 441         BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
 442                 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
 443
 444         for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
 445                 int elem = size_index_elem(i);
 446
 447                 if (elem >= ARRAY_SIZE(size_index))
 448                         break;
 449                 size_index[elem] = KMALLOC_SHIFT_LOW;
 450         }
 451
 452         if (KMALLOC_MIN_SIZE >= 64) {
 453                 /*
 454                  * The 96 byte size cache is not used if the alignment
 455                  * is 64 byte.
 456                  */
 457                 for (i = 64 + 8; i <= 96; i += 8)
 458                         size_index[size_index_elem(i)] = 7;
 459
 460         }
 461
 462         if (KMALLOC_MIN_SIZE >= 128) {
 463                 /*
 464                  * The 192 byte sized cache is not used if the alignment
 465                  * is 128 byte. Redirect kmalloc to use the 256 byte cache
 466                  * instead.
 467                  */
 468                 for (i = 128 + 8; i <= 192; i += 8)
 469                         size_index[size_index_elem(i)] = 8;
 470         }
 471         for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++) {
 472                 if (!kmalloc_caches[i]) {
 473                         kmalloc_caches[i] = create_kmalloc_cache(NULL,
 474                                                         1 << i, flags);
 475                 }
 476
 477                 /*
 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
 481                  */
 482                 if (KMALLOC_MIN_SIZE <= 32 && !kmalloc_caches[1] && i == 6)
 483                         kmalloc_caches[1] = create_kmalloc_cache(NULL, 96, flags);
 484
 485                 if (KMALLOC_MIN_SIZE <= 64 && !kmalloc_caches[2] && i == 7)
 486                         kmalloc_caches[2] = create_kmalloc_cache(NULL, 192, flags);
 487         }
 488
 489         /* Kmalloc array is now usable */
 490         slab_state = UP;
 491
 492         for (i = 0; i <= KMALLOC_SHIFT_HIGH; i++) {
 493                 struct kmem_cache *s = kmalloc_caches[i];
 494                 char *n;
 495
 496                 if (s) {
 497                         n = kasprintf(GFP_NOWAIT, "kmalloc-%d", kmalloc_size(i));
 498
 499                         BUG_ON(!n);
 500                         s->name = n;
 501                 }
 502         }
 503
 504 #ifdef CONFIG_ZONE_DMA
 505         for (i = 0; i <= KMALLOC_SHIFT_HIGH; i++) {
 506                 struct kmem_cache *s = kmalloc_caches[i];
 507
 508                 if (s) {
 509                         int size = kmalloc_size(i);
 510                         char *n = kasprintf(GFP_NOWAIT,
 511                                  "dma-kmalloc-%d", size);
 512
 513                         BUG_ON(!n);
 514                         kmalloc_dma_caches[i] = create_kmalloc_cache(n,
 515                                 size, SLAB_CACHE_DMA | flags);
 516                 }
 517         }
 518 #endif
 519 }
 520 #endif /* !CONFIG_SLOB */
 521
 522 #ifdef CONFIG_TRACING
 523 void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
 524 {
 525         void *ret = kmalloc_order(size, flags, order);
 526         trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
 527         return ret;
 528 }
 529 EXPORT_SYMBOL(kmalloc_order_trace);
 530 #endif
 531
 532 #ifdef CONFIG_SLABINFO
 533
 534 #ifdef CONFIG_SLAB
 535 #define SLABINFO_RIGHTS (S_IWUSR | S_IRUSR)
 536 #else
 537 #define SLABINFO_RIGHTS S_IRUSR
 538 #endif
 539
 540 void print_slabinfo_header(struct seq_file *m)
 541 {
 542         /*
 543          * Output format version, so at least we can change it
 544          * without _too_ many complaints.
 545          */
 546 #ifdef CONFIG_DEBUG_SLAB
 547         seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
 548 #else
 549         seq_puts(m, "slabinfo - version: 2.1\n");
 550 #endif
 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>");
 559 #endif
 560         seq_putc(m, '\n');
 561 }
 562
 563 static void *s_start(struct seq_file *m, loff_t *pos)
 564 {
 565         loff_t n = *pos;
 566
 567         mutex_lock(&slab_mutex);
 568         if (!n)
 569                 print_slabinfo_header(m);
 570
 571         return seq_list_start(&slab_caches, *pos);
 572 }
 573
 574 void *slab_next(struct seq_file *m, void *p, loff_t *pos)
 575 {
 576         return seq_list_next(p, &slab_caches, pos);
 577 }
 578
 579 void slab_stop(struct seq_file *m, void *p)
 580 {
 581         mutex_unlock(&slab_mutex);
 582 }
 583
 584 static void
 585 memcg_accumulate_slabinfo(struct kmem_cache *s, struct slabinfo *info)
 586 {
 587         struct kmem_cache *c;
 588         struct slabinfo sinfo;
 589         int i;
 590
 591         if (!is_root_cache(s))
 592                 return;
 593
 594         for_each_memcg_cache_index(i) {
 595                 c = cache_from_memcg_idx(s, i);
 596                 if (!c)
 597                         continue;
 598
 599                 memset(&sinfo, 0, sizeof(sinfo));
 600                 get_slabinfo(c, &sinfo);
 601
 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;
 607         }
 608 }
 609
 610 int cache_show(struct kmem_cache *s, struct seq_file *m)
 611 {
 612         struct slabinfo sinfo;
 613
 614         memset(&sinfo, 0, sizeof(sinfo));
 615         get_slabinfo(s, &sinfo);
 616
 617         memcg_accumulate_slabinfo(s, &sinfo);
 618
 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));
 622
 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);
 628         seq_putc(m, '\n');
 629         return 0;
 630 }
 631
 632 static int s_show(struct seq_file *m, void *p)
 633 {
 634         struct kmem_cache *s = list_entry(p, struct kmem_cache, list);
 635
 636         if (!is_root_cache(s))
 637                 return 0;
 638         return cache_show(s, m);
 639 }
 640
 641 /*
 642  * slabinfo_op - iterator that generates /proc/slabinfo
 643  *
 644  * Output layout:
 645  * cache-name
 646  * num-active-objs
 647  * total-objs
 648  * object size
 649  * num-active-slabs
 650  * total-slabs
 651  * num-pages-per-slab
 652  * + further values on SMP and with statistics enabled
 653  */
 654 static const struct seq_operations slabinfo_op = {
 655         .start = s_start,
 656         .next = slab_next,
 657         .stop = slab_stop,
 658         .show = s_show,
 659 };
 660
 661 static int slabinfo_open(struct inode *inode, struct file *file)
 662 {
 663         return seq_open(file, &slabinfo_op);
 664 }
 665
 666 static const struct file_operations proc_slabinfo_operations = {
 667         .open           = slabinfo_open,
 668         .read           = seq_read,
 669         .write          = slabinfo_write,
 670         .llseek         = seq_lseek,
 671         .release        = seq_release,
 672 };
 673
 674 static int __init slab_proc_init(void)
 675 {
 676         proc_create("slabinfo", SLABINFO_RIGHTS, NULL,
 677                                                 &proc_slabinfo_operations);
 678         return 0;
 679 }
 680 module_init(slab_proc_init);
 681 #endif /* CONFIG_SLABINFO */