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 = 0;
 175
 176         get_online_cpus();
 177         mutex_lock(&slab_mutex);
 178
 179         if (!kmem_cache_sanity_check(memcg, name, size) == 0)
 180                 goto out_locked;
 181
 182         /*
 183          * Some allocators will constraint the set of valid flags to a subset
 184          * of all flags. We expect them to define CACHE_CREATE_MASK in this
 185          * case, and we'll just provide them with a sanitized version of the
 186          * passed flags.
 187          */
 188         flags &= CACHE_CREATE_MASK;
 189
 190         s = __kmem_cache_alias(memcg, name, size, align, flags, ctor);
 191         if (s)
 192                 goto out_locked;
 193
 194         s = kmem_cache_zalloc(kmem_cache, GFP_KERNEL);
 195         if (s) {
 196                 s->object_size = s->size = size;
 197                 s->align = calculate_alignment(flags, align, size);
 198                 s->ctor = ctor;
 199
 200                 if (memcg_register_cache(memcg, s, parent_cache)) {
 201                         kmem_cache_free(kmem_cache, s);
 202                         err = -ENOMEM;
 203                         goto out_locked;
 204                 }
 205
 206                 s->name = kstrdup(name, GFP_KERNEL);
 207                 if (!s->name) {
 208                         kmem_cache_free(kmem_cache, s);
 209                         err = -ENOMEM;
 210                         goto out_locked;
 211                 }
 212
 213                 err = __kmem_cache_create(s, flags);
 214                 if (!err) {
 215                         s->refcount = 1;
 216                         list_add(&s->list, &slab_caches);
 217                         memcg_cache_list_add(memcg, s);
 218                 } else {
 219                         kfree(s->name);
 220                         kmem_cache_free(kmem_cache, s);
 221                 }
 222         } else
 223                 err = -ENOMEM;
 224
 225 out_locked:
 226         mutex_unlock(&slab_mutex);
 227         put_online_cpus();
 228
 229         if (err) {
 230
 231                 if (flags & SLAB_PANIC)
 232                         panic("kmem_cache_create: Failed to create slab '%s'. Error %d\n",
 233                                 name, err);
 234                 else {
 235                         printk(KERN_WARNING "kmem_cache_create(%s) failed with error %d",
 236                                 name, err);
 237                         dump_stack();
 238                 }
 239
 240                 return NULL;
 241         }
 242
 243         return s;
 244 }
 245
 246 struct kmem_cache *
 247 kmem_cache_create(const char *name, size_t size, size_t align,
 248                   unsigned long flags, void (*ctor)(void *))
 249 {
 250         return kmem_cache_create_memcg(NULL, name, size, align, flags, ctor, NULL);
 251 }
 252 EXPORT_SYMBOL(kmem_cache_create);
 253
 254 void kmem_cache_destroy(struct kmem_cache *s)
 255 {
 256         /* Destroy all the children caches if we aren't a memcg cache */
 257         kmem_cache_destroy_memcg_children(s);
 258
 259         get_online_cpus();
 260         mutex_lock(&slab_mutex);
 261         s->refcount--;
 262         if (!s->refcount) {
 263                 list_del(&s->list);
 264
 265                 if (!__kmem_cache_shutdown(s)) {
 266                         mutex_unlock(&slab_mutex);
 267                         if (s->flags & SLAB_DESTROY_BY_RCU)
 268                                 rcu_barrier();
 269
 270                         memcg_release_cache(s);
 271                         kfree(s->name);
 272                         kmem_cache_free(kmem_cache, s);
 273                 } else {
 274                         list_add(&s->list, &slab_caches);
 275                         mutex_unlock(&slab_mutex);
 276                         printk(KERN_ERR "kmem_cache_destroy %s: Slab cache still has objects\n",
 277                                 s->name);
 278                         dump_stack();
 279                 }
 280         } else {
 281                 mutex_unlock(&slab_mutex);
 282         }
 283         put_online_cpus();
 284 }
 285 EXPORT_SYMBOL(kmem_cache_destroy);
 286
 287 int slab_is_available(void)
 288 {
 289         return slab_state >= UP;
 290 }
 291
 292 #ifndef CONFIG_SLOB
 293 /* Create a cache during boot when no slab services are available yet */
 294 void __init create_boot_cache(struct kmem_cache *s, const char *name, size_t size,
 295                 unsigned long flags)
 296 {
 297         int err;
 298
 299         s->name = name;
 300         s->size = s->object_size = size;
 301         s->align = calculate_alignment(flags, ARCH_KMALLOC_MINALIGN, size);
 302         err = __kmem_cache_create(s, flags);
 303
 304         if (err)
 305                 panic("Creation of kmalloc slab %s size=%zu failed. Reason %d\n",
 306                                         name, size, err);
 307
 308         s->refcount = -1;       /* Exempt from merging for now */
 309 }
 310
 311 struct kmem_cache *__init create_kmalloc_cache(const char *name, size_t size,
 312                                 unsigned long flags)
 313 {
 314         struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
 315
 316         if (!s)
 317                 panic("Out of memory when creating slab %s\n", name);
 318
 319         create_boot_cache(s, name, size, flags);
 320         list_add(&s->list, &slab_caches);
 321         s->refcount = 1;
 322         return s;
 323 }
 324
 325 struct kmem_cache *kmalloc_caches[KMALLOC_SHIFT_HIGH + 1];
 326 EXPORT_SYMBOL(kmalloc_caches);
 327
 328 #ifdef CONFIG_ZONE_DMA
 329 struct kmem_cache *kmalloc_dma_caches[KMALLOC_SHIFT_HIGH + 1];
 330 EXPORT_SYMBOL(kmalloc_dma_caches);
 331 #endif
 332
 333 /*
 334  * Conversion table for small slabs sizes / 8 to the index in the
 335  * kmalloc array. This is necessary for slabs < 192 since we have non power
 336  * of two cache sizes there. The size of larger slabs can be determined using
 337  * fls.
 338  */
 339 static s8 size_index[24] = {
 340         3,      /* 8 */
 341         4,      /* 16 */
 342         5,      /* 24 */
 343         5,      /* 32 */
 344         6,      /* 40 */
 345         6,      /* 48 */
 346         6,      /* 56 */
 347         6,      /* 64 */
 348         1,      /* 72 */
 349         1,      /* 80 */
 350         1,      /* 88 */
 351         1,      /* 96 */
 352         7,      /* 104 */
 353         7,      /* 112 */
 354         7,      /* 120 */
 355         7,      /* 128 */
 356         2,      /* 136 */
 357         2,      /* 144 */
 358         2,      /* 152 */
 359         2,      /* 160 */
 360         2,      /* 168 */
 361         2,      /* 176 */
 362         2,      /* 184 */
 363         2       /* 192 */
 364 };
 365
 366 static inline int size_index_elem(size_t bytes)
 367 {
 368         return (bytes - 1) / 8;
 369 }
 370
 371 /*
 372  * Find the kmem_cache structure that serves a given size of
 373  * allocation
 374  */
 375 struct kmem_cache *kmalloc_slab(size_t size, gfp_t flags)
 376 {
 377         int index;
 378
 379         if (unlikely(size > KMALLOC_MAX_SIZE)) {
 380                 WARN_ON_ONCE(!(flags & __GFP_NOWARN));
 381                 return NULL;
 382         }
 383
 384         if (size <= 192) {
 385                 if (!size)
 386                         return ZERO_SIZE_PTR;
 387
 388                 index = size_index[size_index_elem(size)];
 389         } else
 390                 index = fls(size - 1);
 391
 392 #ifdef CONFIG_ZONE_DMA
 393         if (unlikely((flags & GFP_DMA)))
 394                 return kmalloc_dma_caches[index];
 395
 396 #endif
 397         return kmalloc_caches[index];
 398 }
 399
 400 /*
 401  * Create the kmalloc array. Some of the regular kmalloc arrays
 402  * may already have been created because they were needed to
 403  * enable allocations for slab creation.
 404  */
 405 void __init create_kmalloc_caches(unsigned long flags)
 406 {
 407         int i;
 408
 409         /*
 410          * Patch up the size_index table if we have strange large alignment
 411          * requirements for the kmalloc array. This is only the case for
 412          * MIPS it seems. The standard arches will not generate any code here.
 413          *
 414          * Largest permitted alignment is 256 bytes due to the way we
 415          * handle the index determination for the smaller caches.
 416          *
 417          * Make sure that nothing crazy happens if someone starts tinkering
 418          * around with ARCH_KMALLOC_MINALIGN
 419          */
 420         BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
 421                 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
 422
 423         for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
 424                 int elem = size_index_elem(i);
 425
 426                 if (elem >= ARRAY_SIZE(size_index))
 427                         break;
 428                 size_index[elem] = KMALLOC_SHIFT_LOW;
 429         }
 430
 431         if (KMALLOC_MIN_SIZE >= 64) {
 432                 /*
 433                  * The 96 byte size cache is not used if the alignment
 434                  * is 64 byte.
 435                  */
 436                 for (i = 64 + 8; i <= 96; i += 8)
 437                         size_index[size_index_elem(i)] = 7;
 438
 439         }
 440
 441         if (KMALLOC_MIN_SIZE >= 128) {
 442                 /*
 443                  * The 192 byte sized cache is not used if the alignment
 444                  * is 128 byte. Redirect kmalloc to use the 256 byte cache
 445                  * instead.
 446                  */
 447                 for (i = 128 + 8; i <= 192; i += 8)
 448                         size_index[size_index_elem(i)] = 8;
 449         }
 450         for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++) {
 451                 if (!kmalloc_caches[i]) {
 452                         kmalloc_caches[i] = create_kmalloc_cache(NULL,
 453                                                         1 << i, flags);
 454                 }
 455
 456                 /*
 457                  * Caches that are not of the two-to-the-power-of size.
 458                  * These have to be created immediately after the
 459                  * earlier power of two caches
 460                  */
 461                 if (KMALLOC_MIN_SIZE <= 32 && !kmalloc_caches[1] && i == 6)
 462                         kmalloc_caches[1] = create_kmalloc_cache(NULL, 96, flags);
 463
 464                 if (KMALLOC_MIN_SIZE <= 64 && !kmalloc_caches[2] && i == 7)
 465                         kmalloc_caches[2] = create_kmalloc_cache(NULL, 192, flags);
 466         }
 467
 468         /* Kmalloc array is now usable */
 469         slab_state = UP;
 470
 471         for (i = 0; i <= KMALLOC_SHIFT_HIGH; i++) {
 472                 struct kmem_cache *s = kmalloc_caches[i];
 473                 char *n;
 474
 475                 if (s) {
 476                         n = kasprintf(GFP_NOWAIT, "kmalloc-%d", kmalloc_size(i));
 477
 478                         BUG_ON(!n);
 479                         s->name = n;
 480                 }
 481         }
 482
 483 #ifdef CONFIG_ZONE_DMA
 484         for (i = 0; i <= KMALLOC_SHIFT_HIGH; i++) {
 485                 struct kmem_cache *s = kmalloc_caches[i];
 486
 487                 if (s) {
 488                         int size = kmalloc_size(i);
 489                         char *n = kasprintf(GFP_NOWAIT,
 490                                  "dma-kmalloc-%d", size);
 491
 492                         BUG_ON(!n);
 493                         kmalloc_dma_caches[i] = create_kmalloc_cache(n,
 494                                 size, SLAB_CACHE_DMA | flags);
 495                 }
 496         }
 497 #endif
 498 }
 499 #endif /* !CONFIG_SLOB */
 500
 501 #ifdef CONFIG_TRACING
 502 void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
 503 {
 504         void *ret = kmalloc_order(size, flags, order);
 505         trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
 506         return ret;
 507 }
 508 EXPORT_SYMBOL(kmalloc_order_trace);
 509 #endif
 510
 511 #ifdef CONFIG_SLABINFO
 512
 513 #ifdef CONFIG_SLAB
 514 #define SLABINFO_RIGHTS (S_IWUSR | S_IRUSR)
 515 #else
 516 #define SLABINFO_RIGHTS S_IRUSR
 517 #endif
 518
 519 void print_slabinfo_header(struct seq_file *m)
 520 {
 521         /*
 522          * Output format version, so at least we can change it
 523          * without _too_ many complaints.
 524          */
 525 #ifdef CONFIG_DEBUG_SLAB
 526         seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
 527 #else
 528         seq_puts(m, "slabinfo - version: 2.1\n");
 529 #endif
 530         seq_puts(m, "# name            <active_objs> <num_objs> <objsize> "
 531                  "<objperslab> <pagesperslab>");
 532         seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
 533         seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
 534 #ifdef CONFIG_DEBUG_SLAB
 535         seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
 536                  "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
 537         seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
 538 #endif
 539         seq_putc(m, '\n');
 540 }
 541
 542 static void *s_start(struct seq_file *m, loff_t *pos)
 543 {
 544         loff_t n = *pos;
 545
 546         mutex_lock(&slab_mutex);
 547         if (!n)
 548                 print_slabinfo_header(m);
 549
 550         return seq_list_start(&slab_caches, *pos);
 551 }
 552
 553 void *slab_next(struct seq_file *m, void *p, loff_t *pos)
 554 {
 555         return seq_list_next(p, &slab_caches, pos);
 556 }
 557
 558 void slab_stop(struct seq_file *m, void *p)
 559 {
 560         mutex_unlock(&slab_mutex);
 561 }
 562
 563 static void
 564 memcg_accumulate_slabinfo(struct kmem_cache *s, struct slabinfo *info)
 565 {
 566         struct kmem_cache *c;
 567         struct slabinfo sinfo;
 568         int i;
 569
 570         if (!is_root_cache(s))
 571                 return;
 572
 573         for_each_memcg_cache_index(i) {
 574                 c = cache_from_memcg_idx(s, i);
 575                 if (!c)
 576                         continue;
 577
 578                 memset(&sinfo, 0, sizeof(sinfo));
 579                 get_slabinfo(c, &sinfo);
 580
 581                 info->active_slabs += sinfo.active_slabs;
 582                 info->num_slabs += sinfo.num_slabs;
 583                 info->shared_avail += sinfo.shared_avail;
 584                 info->active_objs += sinfo.active_objs;
 585                 info->num_objs += sinfo.num_objs;
 586         }
 587 }
 588
 589 int cache_show(struct kmem_cache *s, struct seq_file *m)
 590 {
 591         struct slabinfo sinfo;
 592
 593         memset(&sinfo, 0, sizeof(sinfo));
 594         get_slabinfo(s, &sinfo);
 595
 596         memcg_accumulate_slabinfo(s, &sinfo);
 597
 598         seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
 599                    cache_name(s), sinfo.active_objs, sinfo.num_objs, s->size,
 600                    sinfo.objects_per_slab, (1 << sinfo.cache_order));
 601
 602         seq_printf(m, " : tunables %4u %4u %4u",
 603                    sinfo.limit, sinfo.batchcount, sinfo.shared);
 604         seq_printf(m, " : slabdata %6lu %6lu %6lu",
 605                    sinfo.active_slabs, sinfo.num_slabs, sinfo.shared_avail);
 606         slabinfo_show_stats(m, s);
 607         seq_putc(m, '\n');
 608         return 0;
 609 }
 610
 611 static int s_show(struct seq_file *m, void *p)
 612 {
 613         struct kmem_cache *s = list_entry(p, struct kmem_cache, list);
 614
 615         if (!is_root_cache(s))
 616                 return 0;
 617         return cache_show(s, m);
 618 }
 619
 620 /*
 621  * slabinfo_op - iterator that generates /proc/slabinfo
 622  *
 623  * Output layout:
 624  * cache-name
 625  * num-active-objs
 626  * total-objs
 627  * object size
 628  * num-active-slabs
 629  * total-slabs
 630  * num-pages-per-slab
 631  * + further values on SMP and with statistics enabled
 632  */
 633 static const struct seq_operations slabinfo_op = {
 634         .start = s_start,
 635         .next = slab_next,
 636         .stop = slab_stop,
 637         .show = s_show,
 638 };
 639
 640 static int slabinfo_open(struct inode *inode, struct file *file)
 641 {
 642         return seq_open(file, &slabinfo_op);
 643 }
 644
 645 static const struct file_operations proc_slabinfo_operations = {
 646         .open           = slabinfo_open,
 647         .read           = seq_read,
 648         .write          = slabinfo_write,
 649         .llseek         = seq_lseek,
 650         .release        = seq_release,
 651 };
 652
 653 static int __init slab_proc_init(void)
 654 {
 655         proc_create("slabinfo", SLABINFO_RIGHTS, NULL,
 656                                                 &proc_slabinfo_operations);
 657         return 0;
 658 }
 659 module_init(slab_proc_init);
 660 #endif /* CONFIG_SLABINFO */