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                 /*
  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.
  64                  */
  65                 if (!memcg && !strcmp(s->name, name)) {
  66                         pr_err("%s (%s): Cache name already exists.\n",
  67                                __func__, name);
  68                         dump_stack();
  69                         s = NULL;
  70                         return -EINVAL;
  71                 }
  72         }
  73
  74         WARN_ON(strchr(name, ' '));     /* It confuses parsers */
  75         return 0;
  76 }
  77 #else
  78 static inline int kmem_cache_sanity_check(struct mem_cgroup *memcg,
  79                                           const char *name, size_t size)
  80 {
  81         return 0;
  82 }
  83 #endif
  84
  85 #ifdef CONFIG_MEMCG_KMEM
  86 int memcg_update_all_caches(int num_memcgs)
  87 {
  88         struct kmem_cache *s;
  89         int ret = 0;
  90         mutex_lock(&slab_mutex);
  91
  92         list_for_each_entry(s, &slab_caches, list) {
  93                 if (!is_root_cache(s))
  94                         continue;
  95
  96                 ret = memcg_update_cache_size(s, num_memcgs);
  97                 /*
  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.
 101                  */
 102                 if (ret)
 103                         goto out;
 104         }
 105
 106         memcg_update_array_size(num_memcgs);
 107 out:
 108         mutex_unlock(&slab_mutex);
 109         return ret;
 110 }
 111 #endif
 112
 113 /*
 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.
 116  */
 117 unsigned long calculate_alignment(unsigned long flags,
 118                 unsigned long align, unsigned long size)
 119 {
 120         /*
 121          * If the user wants hardware cache aligned objects then follow that
 122          * suggestion if the object is sufficiently large.
 123          *
 124          * The hardware cache alignment cannot override the specified
 125          * alignment though. If that is greater then use it.
 126          */
 127         if (flags & SLAB_HWCACHE_ALIGN) {
 128                 unsigned long ralign = cache_line_size();
 129                 while (size <= ralign / 2)
 130                         ralign /= 2;
 131                 align = max(align, ralign);
 132         }
 133
 134         if (align < ARCH_SLAB_MINALIGN)
 135                 align = ARCH_SLAB_MINALIGN;
 136
 137         return ALIGN(align, sizeof(void *));
 138 }
 139
 140
 141 /*
 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.
 146  * @flags: SLAB flags
 147  * @ctor: A constructor for the objects.
 148  *
 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.
 152  *
 153  * The flags are
 154  *
 155  * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
 156  * to catch references to uninitialised memory.
 157  *
 158  * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
 159  * for buffer overruns.
 160  *
 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
 163  * as davem.
 164  */
 165
 166 struct kmem_cache *
 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)
 170 {
 171         struct kmem_cache *s = NULL;
 172         int err = 0;
 173
 174         get_online_cpus();
 175         mutex_lock(&slab_mutex);
 176
 177         if (!kmem_cache_sanity_check(memcg, name, size) == 0)
 178                 goto out_locked;
 179
 180         /*
 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
 184          * passed flags.
 185          */
 186         flags &= CACHE_CREATE_MASK;
 187
 188         s = __kmem_cache_alias(memcg, name, size, align, flags, ctor);
 189         if (s)
 190                 goto out_locked;
 191
 192         s = kmem_cache_zalloc(kmem_cache, GFP_KERNEL);
 193         if (s) {
 194                 s->object_size = s->size = size;
 195                 s->align = calculate_alignment(flags, align, size);
 196                 s->ctor = ctor;
 197
 198                 if (memcg_register_cache(memcg, s, parent_cache)) {
 199                         kmem_cache_free(kmem_cache, s);
 200                         err = -ENOMEM;
 201                         goto out_locked;
 202                 }
 203
 204                 s->name = kstrdup(name, GFP_KERNEL);
 205                 if (!s->name) {
 206                         kmem_cache_free(kmem_cache, s);
 207                         err = -ENOMEM;
 208                         goto out_locked;
 209                 }
 210
 211                 err = __kmem_cache_create(s, flags);
 212                 if (!err) {
 213                         s->refcount = 1;
 214                         list_add(&s->list, &slab_caches);
 215                         memcg_cache_list_add(memcg, s);
 216                 } else {
 217                         kfree(s->name);
 218                         kmem_cache_free(kmem_cache, s);
 219                 }
 220         } else
 221                 err = -ENOMEM;
 222
 223 out_locked:
 224         mutex_unlock(&slab_mutex);
 225         put_online_cpus();
 226
 227         if (err) {
 228
 229                 if (flags & SLAB_PANIC)
 230                         panic("kmem_cache_create: Failed to create slab '%s'. Error %d\n",
 231                                 name, err);
 232                 else {
 233                         printk(KERN_WARNING "kmem_cache_create(%s) failed with error %d",
 234                                 name, err);
 235                         dump_stack();
 236                 }
 237
 238                 return NULL;
 239         }
 240
 241         return s;
 242 }
 243
 244 struct kmem_cache *
 245 kmem_cache_create(const char *name, size_t size, size_t align,
 246                   unsigned long flags, void (*ctor)(void *))
 247 {
 248         return kmem_cache_create_memcg(NULL, name, size, align, flags, ctor, NULL);
 249 }
 250 EXPORT_SYMBOL(kmem_cache_create);
 251
 252 void kmem_cache_destroy(struct kmem_cache *s)
 253 {
 254         /* Destroy all the children caches if we aren't a memcg cache */
 255         kmem_cache_destroy_memcg_children(s);
 256
 257         get_online_cpus();
 258         mutex_lock(&slab_mutex);
 259         s->refcount--;
 260         if (!s->refcount) {
 261                 list_del(&s->list);
 262
 263                 if (!__kmem_cache_shutdown(s)) {
 264                         mutex_unlock(&slab_mutex);
 265                         if (s->flags & SLAB_DESTROY_BY_RCU)
 266                                 rcu_barrier();
 267
 268                         memcg_release_cache(s);
 269                         kfree(s->name);
 270                         kmem_cache_free(kmem_cache, s);
 271                 } else {
 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",
 275                                 s->name);
 276                         dump_stack();
 277                 }
 278         } else {
 279                 mutex_unlock(&slab_mutex);
 280         }
 281         put_online_cpus();
 282 }
 283 EXPORT_SYMBOL(kmem_cache_destroy);
 284
 285 int slab_is_available(void)
 286 {
 287         return slab_state >= UP;
 288 }
 289
 290 #ifndef CONFIG_SLOB
 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,
 293                 unsigned long flags)
 294 {
 295         int err;
 296
 297         s->name = name;
 298         s->size = s->object_size = size;
 299         s->align = calculate_alignment(flags, ARCH_KMALLOC_MINALIGN, size);
 300         err = __kmem_cache_create(s, flags);
 301
 302         if (err)
 303                 panic("Creation of kmalloc slab %s size=%zu failed. Reason %d\n",
 304                                         name, size, err);
 305
 306         s->refcount = -1;       /* Exempt from merging for now */
 307 }
 308
 309 struct kmem_cache *__init create_kmalloc_cache(const char *name, size_t size,
 310                                 unsigned long flags)
 311 {
 312         struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
 313
 314         if (!s)
 315                 panic("Out of memory when creating slab %s\n", name);
 316
 317         create_boot_cache(s, name, size, flags);
 318         list_add(&s->list, &slab_caches);
 319         s->refcount = 1;
 320         return s;
 321 }
 322
 323 struct kmem_cache *kmalloc_caches[KMALLOC_SHIFT_HIGH + 1];
 324 EXPORT_SYMBOL(kmalloc_caches);
 325
 326 #ifdef CONFIG_ZONE_DMA
 327 struct kmem_cache *kmalloc_dma_caches[KMALLOC_SHIFT_HIGH + 1];
 328 EXPORT_SYMBOL(kmalloc_dma_caches);
 329 #endif
 330
 331 /*
 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
 335  * fls.
 336  */
 337 static s8 size_index[24] = {
 338         3,      /* 8 */
 339         4,      /* 16 */
 340         5,      /* 24 */
 341         5,      /* 32 */
 342         6,      /* 40 */
 343         6,      /* 48 */
 344         6,      /* 56 */
 345         6,      /* 64 */
 346         1,      /* 72 */
 347         1,      /* 80 */
 348         1,      /* 88 */
 349         1,      /* 96 */
 350         7,      /* 104 */
 351         7,      /* 112 */
 352         7,      /* 120 */
 353         7,      /* 128 */
 354         2,      /* 136 */
 355         2,      /* 144 */
 356         2,      /* 152 */
 357         2,      /* 160 */
 358         2,      /* 168 */
 359         2,      /* 176 */
 360         2,      /* 184 */
 361         2       /* 192 */
 362 };
 363
 364 static inline int size_index_elem(size_t bytes)
 365 {
 366         return (bytes - 1) / 8;
 367 }
 368
 369 /*
 370  * Find the kmem_cache structure that serves a given size of
 371  * allocation
 372  */
 373 struct kmem_cache *kmalloc_slab(size_t size, gfp_t flags)
 374 {
 375         int index;
 376
 377         if (unlikely(size > KMALLOC_MAX_SIZE)) {
 378                 WARN_ON_ONCE(!(flags & __GFP_NOWARN));
 379                 return NULL;
 380         }
 381
 382         if (size <= 192) {
 383                 if (!size)
 384                         return ZERO_SIZE_PTR;
 385
 386                 index = size_index[size_index_elem(size)];
 387         } else
 388                 index = fls(size - 1);
 389
 390 #ifdef CONFIG_ZONE_DMA
 391         if (unlikely((flags & GFP_DMA)))
 392                 return kmalloc_dma_caches[index];
 393
 394 #endif
 395         return kmalloc_caches[index];
 396 }
 397
 398 /*
 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.
 402  */
 403 void __init create_kmalloc_caches(unsigned long flags)
 404 {
 405         int i;
 406
 407         /*
 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.
 411          *
 412          * Largest permitted alignment is 256 bytes due to the way we
 413          * handle the index determination for the smaller caches.
 414          *
 415          * Make sure that nothing crazy happens if someone starts tinkering
 416          * around with ARCH_KMALLOC_MINALIGN
 417          */
 418         BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
 419                 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
 420
 421         for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
 422                 int elem = size_index_elem(i);
 423
 424                 if (elem >= ARRAY_SIZE(size_index))
 425                         break;
 426                 size_index[elem] = KMALLOC_SHIFT_LOW;
 427         }
 428
 429         if (KMALLOC_MIN_SIZE >= 64) {
 430                 /*
 431                  * The 96 byte size cache is not used if the alignment
 432                  * is 64 byte.
 433                  */
 434                 for (i = 64 + 8; i <= 96; i += 8)
 435                         size_index[size_index_elem(i)] = 7;
 436
 437         }
 438
 439         if (KMALLOC_MIN_SIZE >= 128) {
 440                 /*
 441                  * The 192 byte sized cache is not used if the alignment
 442                  * is 128 byte. Redirect kmalloc to use the 256 byte cache
 443                  * instead.
 444                  */
 445                 for (i = 128 + 8; i <= 192; i += 8)
 446                         size_index[size_index_elem(i)] = 8;
 447         }
 448         for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++) {
 449                 if (!kmalloc_caches[i]) {
 450                         kmalloc_caches[i] = create_kmalloc_cache(NULL,
 451                                                         1 << i, flags);
 452                 }
 453
 454                 /*
 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
 458                  */
 459                 if (KMALLOC_MIN_SIZE <= 32 && !kmalloc_caches[1] && i == 6)
 460                         kmalloc_caches[1] = create_kmalloc_cache(NULL, 96, flags);
 461
 462                 if (KMALLOC_MIN_SIZE <= 64 && !kmalloc_caches[2] && i == 7)
 463                         kmalloc_caches[2] = create_kmalloc_cache(NULL, 192, flags);
 464         }
 465
 466         /* Kmalloc array is now usable */
 467         slab_state = UP;
 468
 469         for (i = 0; i <= KMALLOC_SHIFT_HIGH; i++) {
 470                 struct kmem_cache *s = kmalloc_caches[i];
 471                 char *n;
 472
 473                 if (s) {
 474                         n = kasprintf(GFP_NOWAIT, "kmalloc-%d", kmalloc_size(i));
 475
 476                         BUG_ON(!n);
 477                         s->name = n;
 478                 }
 479         }
 480
 481 #ifdef CONFIG_ZONE_DMA
 482         for (i = 0; i <= KMALLOC_SHIFT_HIGH; i++) {
 483                 struct kmem_cache *s = kmalloc_caches[i];
 484
 485                 if (s) {
 486                         int size = kmalloc_size(i);
 487                         char *n = kasprintf(GFP_NOWAIT,
 488                                  "dma-kmalloc-%d", size);
 489
 490                         BUG_ON(!n);
 491                         kmalloc_dma_caches[i] = create_kmalloc_cache(n,
 492                                 size, SLAB_CACHE_DMA | flags);
 493                 }
 494         }
 495 #endif
 496 }
 497 #endif /* !CONFIG_SLOB */
 498
 499 #ifdef CONFIG_TRACING
 500 void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
 501 {
 502         void *ret = kmalloc_order(size, flags, order);
 503         trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
 504         return ret;
 505 }
 506 EXPORT_SYMBOL(kmalloc_order_trace);
 507 #endif
 508
 509 #ifdef CONFIG_SLABINFO
 510
 511 #ifdef CONFIG_SLAB
 512 #define SLABINFO_RIGHTS (S_IWUSR | S_IRUSR)
 513 #else
 514 #define SLABINFO_RIGHTS S_IRUSR
 515 #endif
 516
 517 void print_slabinfo_header(struct seq_file *m)
 518 {
 519         /*
 520          * Output format version, so at least we can change it
 521          * without _too_ many complaints.
 522          */
 523 #ifdef CONFIG_DEBUG_SLAB
 524         seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
 525 #else
 526         seq_puts(m, "slabinfo - version: 2.1\n");
 527 #endif
 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>");
 536 #endif
 537         seq_putc(m, '\n');
 538 }
 539
 540 static void *s_start(struct seq_file *m, loff_t *pos)
 541 {
 542         loff_t n = *pos;
 543
 544         mutex_lock(&slab_mutex);
 545         if (!n)
 546                 print_slabinfo_header(m);
 547
 548         return seq_list_start(&slab_caches, *pos);
 549 }
 550
 551 void *slab_next(struct seq_file *m, void *p, loff_t *pos)
 552 {
 553         return seq_list_next(p, &slab_caches, pos);
 554 }
 555
 556 void slab_stop(struct seq_file *m, void *p)
 557 {
 558         mutex_unlock(&slab_mutex);
 559 }
 560
 561 static void
 562 memcg_accumulate_slabinfo(struct kmem_cache *s, struct slabinfo *info)
 563 {
 564         struct kmem_cache *c;
 565         struct slabinfo sinfo;
 566         int i;
 567
 568         if (!is_root_cache(s))
 569                 return;
 570
 571         for_each_memcg_cache_index(i) {
 572                 c = cache_from_memcg(s, i);
 573                 if (!c)
 574                         continue;
 575
 576                 memset(&sinfo, 0, sizeof(sinfo));
 577                 get_slabinfo(c, &sinfo);
 578
 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;
 584         }
 585 }
 586
 587 int cache_show(struct kmem_cache *s, struct seq_file *m)
 588 {
 589         struct slabinfo sinfo;
 590
 591         memset(&sinfo, 0, sizeof(sinfo));
 592         get_slabinfo(s, &sinfo);
 593
 594         memcg_accumulate_slabinfo(s, &sinfo);
 595
 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));
 599
 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);
 605         seq_putc(m, '\n');
 606         return 0;
 607 }
 608
 609 static int s_show(struct seq_file *m, void *p)
 610 {
 611         struct kmem_cache *s = list_entry(p, struct kmem_cache, list);
 612
 613         if (!is_root_cache(s))
 614                 return 0;
 615         return cache_show(s, m);
 616 }
 617
 618 /*
 619  * slabinfo_op - iterator that generates /proc/slabinfo
 620  *
 621  * Output layout:
 622  * cache-name
 623  * num-active-objs
 624  * total-objs
 625  * object size
 626  * num-active-slabs
 627  * total-slabs
 628  * num-pages-per-slab
 629  * + further values on SMP and with statistics enabled
 630  */
 631 static const struct seq_operations slabinfo_op = {
 632         .start = s_start,
 633         .next = slab_next,
 634         .stop = slab_stop,
 635         .show = s_show,
 636 };
 637
 638 static int slabinfo_open(struct inode *inode, struct file *file)
 639 {
 640         return seq_open(file, &slabinfo_op);
 641 }
 642
 643 static const struct file_operations proc_slabinfo_operations = {
 644         .open           = slabinfo_open,
 645         .read           = seq_read,
 646         .write          = slabinfo_write,
 647         .llseek         = seq_lseek,
 648         .release        = seq_release,
 649 };
 650
 651 static int __init slab_proc_init(void)
 652 {
 653         proc_create("slabinfo", SLABINFO_RIGHTS, NULL,
 654                                                 &proc_slabinfo_operations);
 655         return 0;
 656 }
 657 module_init(slab_proc_init);
 658 #endif /* CONFIG_SLABINFO */