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         /*
 237          * There is no point in flooding logs with warnings or especially
 238          * crashing the system if we fail to create a cache for a memcg. In
 239          * this case we will be accounting the memcg allocation to the root
 240          * cgroup until we succeed to create its own cache, but it isn't that
 241          * critical.
 242          */
 243         if (err && !memcg) {
 244                 if (flags & SLAB_PANIC)
 245                         panic("kmem_cache_create: Failed to create slab '%s'. Error %d\n",
 246                                 name, err);
 247                 else {
 248                         printk(KERN_WARNING "kmem_cache_create(%s) failed with error %d",
 249                                 name, err);
 250                         dump_stack();
 251                 }
 252                 return NULL;
 253         }
 254         return s;
 255
 256 out_free_cache:
 257         memcg_free_cache_params(s);
 258         kfree(s->name);
 259         kmem_cache_free(kmem_cache, s);
 260         goto out_unlock;
 261 }
 262
 263 struct kmem_cache *
 264 kmem_cache_create(const char *name, size_t size, size_t align,
 265                   unsigned long flags, void (*ctor)(void *))
 266 {
 267         return kmem_cache_create_memcg(NULL, name, size, align, flags, ctor, NULL);
 268 }
 269 EXPORT_SYMBOL(kmem_cache_create);
 270
 271 void kmem_cache_destroy(struct kmem_cache *s)
 272 {
 273         /* Destroy all the children caches if we aren't a memcg cache */
 274         kmem_cache_destroy_memcg_children(s);
 275
 276         get_online_cpus();
 277         mutex_lock(&slab_mutex);
 278         s->refcount--;
 279         if (!s->refcount) {
 280                 list_del(&s->list);
 281
 282                 if (!__kmem_cache_shutdown(s)) {
 283                         memcg_unregister_cache(s);
 284                         mutex_unlock(&slab_mutex);
 285                         if (s->flags & SLAB_DESTROY_BY_RCU)
 286                                 rcu_barrier();
 287
 288                         memcg_free_cache_params(s);
 289                         kfree(s->name);
 290                         kmem_cache_free(kmem_cache, s);
 291                 } else {
 292                         list_add(&s->list, &slab_caches);
 293                         mutex_unlock(&slab_mutex);
 294                         printk(KERN_ERR "kmem_cache_destroy %s: Slab cache still has objects\n",
 295                                 s->name);
 296                         dump_stack();
 297                 }
 298         } else {
 299                 mutex_unlock(&slab_mutex);
 300         }
 301         put_online_cpus();
 302 }
 303 EXPORT_SYMBOL(kmem_cache_destroy);
 304
 305 int slab_is_available(void)
 306 {
 307         return slab_state >= UP;
 308 }
 309
 310 #ifndef CONFIG_SLOB
 311 /* Create a cache during boot when no slab services are available yet */
 312 void __init create_boot_cache(struct kmem_cache *s, const char *name, size_t size,
 313                 unsigned long flags)
 314 {
 315         int err;
 316
 317         s->name = name;
 318         s->size = s->object_size = size;
 319         s->align = calculate_alignment(flags, ARCH_KMALLOC_MINALIGN, size);
 320         err = __kmem_cache_create(s, flags);
 321
 322         if (err)
 323                 panic("Creation of kmalloc slab %s size=%zu failed. Reason %d\n",
 324                                         name, size, err);
 325
 326         s->refcount = -1;       /* Exempt from merging for now */
 327 }
 328
 329 struct kmem_cache *__init create_kmalloc_cache(const char *name, size_t size,
 330                                 unsigned long flags)
 331 {
 332         struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
 333
 334         if (!s)
 335                 panic("Out of memory when creating slab %s\n", name);
 336
 337         create_boot_cache(s, name, size, flags);
 338         list_add(&s->list, &slab_caches);
 339         s->refcount = 1;
 340         return s;
 341 }
 342
 343 struct kmem_cache *kmalloc_caches[KMALLOC_SHIFT_HIGH + 1];
 344 EXPORT_SYMBOL(kmalloc_caches);
 345
 346 #ifdef CONFIG_ZONE_DMA
 347 struct kmem_cache *kmalloc_dma_caches[KMALLOC_SHIFT_HIGH + 1];
 348 EXPORT_SYMBOL(kmalloc_dma_caches);
 349 #endif
 350
 351 /*
 352  * Conversion table for small slabs sizes / 8 to the index in the
 353  * kmalloc array. This is necessary for slabs < 192 since we have non power
 354  * of two cache sizes there. The size of larger slabs can be determined using
 355  * fls.
 356  */
 357 static s8 size_index[24] = {
 358         3,      /* 8 */
 359         4,      /* 16 */
 360         5,      /* 24 */
 361         5,      /* 32 */
 362         6,      /* 40 */
 363         6,      /* 48 */
 364         6,      /* 56 */
 365         6,      /* 64 */
 366         1,      /* 72 */
 367         1,      /* 80 */
 368         1,      /* 88 */
 369         1,      /* 96 */
 370         7,      /* 104 */
 371         7,      /* 112 */
 372         7,      /* 120 */
 373         7,      /* 128 */
 374         2,      /* 136 */
 375         2,      /* 144 */
 376         2,      /* 152 */
 377         2,      /* 160 */
 378         2,      /* 168 */
 379         2,      /* 176 */
 380         2,      /* 184 */
 381         2       /* 192 */
 382 };
 383
 384 static inline int size_index_elem(size_t bytes)
 385 {
 386         return (bytes - 1) / 8;
 387 }
 388
 389 /*
 390  * Find the kmem_cache structure that serves a given size of
 391  * allocation
 392  */
 393 struct kmem_cache *kmalloc_slab(size_t size, gfp_t flags)
 394 {
 395         int index;
 396
 397         if (unlikely(size > KMALLOC_MAX_SIZE)) {
 398                 WARN_ON_ONCE(!(flags & __GFP_NOWARN));
 399                 return NULL;
 400         }
 401
 402         if (size <= 192) {
 403                 if (!size)
 404                         return ZERO_SIZE_PTR;
 405
 406                 index = size_index[size_index_elem(size)];
 407         } else
 408                 index = fls(size - 1);
 409
 410 #ifdef CONFIG_ZONE_DMA
 411         if (unlikely((flags & GFP_DMA)))
 412                 return kmalloc_dma_caches[index];
 413
 414 #endif
 415         return kmalloc_caches[index];
 416 }
 417
 418 /*
 419  * Create the kmalloc array. Some of the regular kmalloc arrays
 420  * may already have been created because they were needed to
 421  * enable allocations for slab creation.
 422  */
 423 void __init create_kmalloc_caches(unsigned long flags)
 424 {
 425         int i;
 426
 427         /*
 428          * Patch up the size_index table if we have strange large alignment
 429          * requirements for the kmalloc array. This is only the case for
 430          * MIPS it seems. The standard arches will not generate any code here.
 431          *
 432          * Largest permitted alignment is 256 bytes due to the way we
 433          * handle the index determination for the smaller caches.
 434          *
 435          * Make sure that nothing crazy happens if someone starts tinkering
 436          * around with ARCH_KMALLOC_MINALIGN
 437          */
 438         BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
 439                 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
 440
 441         for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
 442                 int elem = size_index_elem(i);
 443
 444                 if (elem >= ARRAY_SIZE(size_index))
 445                         break;
 446                 size_index[elem] = KMALLOC_SHIFT_LOW;
 447         }
 448
 449         if (KMALLOC_MIN_SIZE >= 64) {
 450                 /*
 451                  * The 96 byte size cache is not used if the alignment
 452                  * is 64 byte.
 453                  */
 454                 for (i = 64 + 8; i <= 96; i += 8)
 455                         size_index[size_index_elem(i)] = 7;
 456
 457         }
 458
 459         if (KMALLOC_MIN_SIZE >= 128) {
 460                 /*
 461                  * The 192 byte sized cache is not used if the alignment
 462                  * is 128 byte. Redirect kmalloc to use the 256 byte cache
 463                  * instead.
 464                  */
 465                 for (i = 128 + 8; i <= 192; i += 8)
 466                         size_index[size_index_elem(i)] = 8;
 467         }
 468         for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++) {
 469                 if (!kmalloc_caches[i]) {
 470                         kmalloc_caches[i] = create_kmalloc_cache(NULL,
 471                                                         1 << i, flags);
 472                 }
 473
 474                 /*
 475                  * Caches that are not of the two-to-the-power-of size.
 476                  * These have to be created immediately after the
 477                  * earlier power of two caches
 478                  */
 479                 if (KMALLOC_MIN_SIZE <= 32 && !kmalloc_caches[1] && i == 6)
 480                         kmalloc_caches[1] = create_kmalloc_cache(NULL, 96, flags);
 481
 482                 if (KMALLOC_MIN_SIZE <= 64 && !kmalloc_caches[2] && i == 7)
 483                         kmalloc_caches[2] = create_kmalloc_cache(NULL, 192, flags);
 484         }
 485
 486         /* Kmalloc array is now usable */
 487         slab_state = UP;
 488
 489         for (i = 0; i <= KMALLOC_SHIFT_HIGH; i++) {
 490                 struct kmem_cache *s = kmalloc_caches[i];
 491                 char *n;
 492
 493                 if (s) {
 494                         n = kasprintf(GFP_NOWAIT, "kmalloc-%d", kmalloc_size(i));
 495
 496                         BUG_ON(!n);
 497                         s->name = n;
 498                 }
 499         }
 500
 501 #ifdef CONFIG_ZONE_DMA
 502         for (i = 0; i <= KMALLOC_SHIFT_HIGH; i++) {
 503                 struct kmem_cache *s = kmalloc_caches[i];
 504
 505                 if (s) {
 506                         int size = kmalloc_size(i);
 507                         char *n = kasprintf(GFP_NOWAIT,
 508                                  "dma-kmalloc-%d", size);
 509
 510                         BUG_ON(!n);
 511                         kmalloc_dma_caches[i] = create_kmalloc_cache(n,
 512                                 size, SLAB_CACHE_DMA | flags);
 513                 }
 514         }
 515 #endif
 516 }
 517 #endif /* !CONFIG_SLOB */
 518
 519 #ifdef CONFIG_TRACING
 520 void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
 521 {
 522         void *ret = kmalloc_order(size, flags, order);
 523         trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
 524         return ret;
 525 }
 526 EXPORT_SYMBOL(kmalloc_order_trace);
 527 #endif
 528
 529 #ifdef CONFIG_SLABINFO
 530
 531 #ifdef CONFIG_SLAB
 532 #define SLABINFO_RIGHTS (S_IWUSR | S_IRUSR)
 533 #else
 534 #define SLABINFO_RIGHTS S_IRUSR
 535 #endif
 536
 537 void print_slabinfo_header(struct seq_file *m)
 538 {
 539         /*
 540          * Output format version, so at least we can change it
 541          * without _too_ many complaints.
 542          */
 543 #ifdef CONFIG_DEBUG_SLAB
 544         seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
 545 #else
 546         seq_puts(m, "slabinfo - version: 2.1\n");
 547 #endif
 548         seq_puts(m, "# name            <active_objs> <num_objs> <objsize> "
 549                  "<objperslab> <pagesperslab>");
 550         seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
 551         seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
 552 #ifdef CONFIG_DEBUG_SLAB
 553         seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
 554                  "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
 555         seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
 556 #endif
 557         seq_putc(m, '\n');
 558 }
 559
 560 static void *s_start(struct seq_file *m, loff_t *pos)
 561 {
 562         loff_t n = *pos;
 563
 564         mutex_lock(&slab_mutex);
 565         if (!n)
 566                 print_slabinfo_header(m);
 567
 568         return seq_list_start(&slab_caches, *pos);
 569 }
 570
 571 void *slab_next(struct seq_file *m, void *p, loff_t *pos)
 572 {
 573         return seq_list_next(p, &slab_caches, pos);
 574 }
 575
 576 void slab_stop(struct seq_file *m, void *p)
 577 {
 578         mutex_unlock(&slab_mutex);
 579 }
 580
 581 static void
 582 memcg_accumulate_slabinfo(struct kmem_cache *s, struct slabinfo *info)
 583 {
 584         struct kmem_cache *c;
 585         struct slabinfo sinfo;
 586         int i;
 587
 588         if (!is_root_cache(s))
 589                 return;
 590
 591         for_each_memcg_cache_index(i) {
 592                 c = cache_from_memcg_idx(s, i);
 593                 if (!c)
 594                         continue;
 595
 596                 memset(&sinfo, 0, sizeof(sinfo));
 597                 get_slabinfo(c, &sinfo);
 598
 599                 info->active_slabs += sinfo.active_slabs;
 600                 info->num_slabs += sinfo.num_slabs;
 601                 info->shared_avail += sinfo.shared_avail;
 602                 info->active_objs += sinfo.active_objs;
 603                 info->num_objs += sinfo.num_objs;
 604         }
 605 }
 606
 607 int cache_show(struct kmem_cache *s, struct seq_file *m)
 608 {
 609         struct slabinfo sinfo;
 610
 611         memset(&sinfo, 0, sizeof(sinfo));
 612         get_slabinfo(s, &sinfo);
 613
 614         memcg_accumulate_slabinfo(s, &sinfo);
 615
 616         seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
 617                    cache_name(s), sinfo.active_objs, sinfo.num_objs, s->size,
 618                    sinfo.objects_per_slab, (1 << sinfo.cache_order));
 619
 620         seq_printf(m, " : tunables %4u %4u %4u",
 621                    sinfo.limit, sinfo.batchcount, sinfo.shared);
 622         seq_printf(m, " : slabdata %6lu %6lu %6lu",
 623                    sinfo.active_slabs, sinfo.num_slabs, sinfo.shared_avail);
 624         slabinfo_show_stats(m, s);
 625         seq_putc(m, '\n');
 626         return 0;
 627 }
 628
 629 static int s_show(struct seq_file *m, void *p)
 630 {
 631         struct kmem_cache *s = list_entry(p, struct kmem_cache, list);
 632
 633         if (!is_root_cache(s))
 634                 return 0;
 635         return cache_show(s, m);
 636 }
 637
 638 /*
 639  * slabinfo_op - iterator that generates /proc/slabinfo
 640  *
 641  * Output layout:
 642  * cache-name
 643  * num-active-objs
 644  * total-objs
 645  * object size
 646  * num-active-slabs
 647  * total-slabs
 648  * num-pages-per-slab
 649  * + further values on SMP and with statistics enabled
 650  */
 651 static const struct seq_operations slabinfo_op = {
 652         .start = s_start,
 653         .next = slab_next,
 654         .stop = slab_stop,
 655         .show = s_show,
 656 };
 657
 658 static int slabinfo_open(struct inode *inode, struct file *file)
 659 {
 660         return seq_open(file, &slabinfo_op);
 661 }
 662
 663 static const struct file_operations proc_slabinfo_operations = {
 664         .open           = slabinfo_open,
 665         .read           = seq_read,
 666         .write          = slabinfo_write,
 667         .llseek         = seq_lseek,
 668         .release        = seq_release,
 669 };
 670
 671 static int __init slab_proc_init(void)
 672 {
 673         proc_create("slabinfo", SLABINFO_RIGHTS, NULL,
 674                                                 &proc_slabinfo_operations);
 675         return 0;
 676 }
 677 module_init(slab_proc_init);
 678 #endif /* CONFIG_SLABINFO */