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
  23 #include "slab.h"
  24
  25 enum slab_state slab_state;
  26 LIST_HEAD(slab_caches);
  27 DEFINE_MUTEX(slab_mutex);
  28 struct kmem_cache *kmem_cache;
  29
  30 #ifdef CONFIG_DEBUG_VM
  31 static int kmem_cache_sanity_check(struct mem_cgroup *memcg, const char *name,
  32                                    size_t size)
  33 {
  34         struct kmem_cache *s = NULL;
  35
  36         if (!name || in_interrupt() || size < sizeof(void *) ||
  37                 size > KMALLOC_MAX_SIZE) {
  38                 pr_err("kmem_cache_create(%s) integrity check failed\n", name);
  39                 return -EINVAL;
  40         }
  41
  42         list_for_each_entry(s, &slab_caches, list) {
  43                 char tmp;
  44                 int res;
  45
  46                 /*
  47                  * This happens when the module gets unloaded and doesn't
  48                  * destroy its slab cache and no-one else reuses the vmalloc
  49                  * area of the module.  Print a warning.
  50                  */
  51                 res = probe_kernel_address(s->name, tmp);
  52                 if (res) {
  53                         pr_err("Slab cache with size %d has lost its name\n",
  54                                s->object_size);
  55                         continue;
  56                 }
  57
  58                 /*
  59                  * For simplicity, we won't check this in the list of memcg
  60                  * caches. We have control over memcg naming, and if there
  61                  * aren't duplicates in the global list, there won't be any
  62                  * duplicates in the memcg lists as well.
  63                  */
  64                 if (!memcg && !strcmp(s->name, name)) {
  65                         pr_err("%s (%s): Cache name already exists.\n",
  66                                __func__, name);
  67                         dump_stack();
  68                         s = NULL;
  69                         return -EINVAL;
  70                 }
  71         }
  72
  73         WARN_ON(strchr(name, ' '));     /* It confuses parsers */
  74         return 0;
  75 }
  76 #else
  77 static inline int kmem_cache_sanity_check(struct mem_cgroup *memcg,
  78                                           const char *name, size_t size)
  79 {
  80         return 0;
  81 }
  82 #endif
  83
  84 #ifdef CONFIG_MEMCG_KMEM
  85 int memcg_update_all_caches(int num_memcgs)
  86 {
  87         struct kmem_cache *s;
  88         int ret = 0;
  89         mutex_lock(&slab_mutex);
  90
  91         list_for_each_entry(s, &slab_caches, list) {
  92                 if (!is_root_cache(s))
  93                         continue;
  94
  95                 ret = memcg_update_cache_size(s, num_memcgs);
  96                 /*
  97                  * See comment in memcontrol.c, memcg_update_cache_size:
  98                  * Instead of freeing the memory, we'll just leave the caches
  99                  * up to this point in an updated state.
 100                  */
 101                 if (ret)
 102                         goto out;
 103         }
 104
 105         memcg_update_array_size(num_memcgs);
 106 out:
 107         mutex_unlock(&slab_mutex);
 108         return ret;
 109 }
 110 #endif
 111
 112 /*
 113  * Figure out what the alignment of the objects will be given a set of
 114  * flags, a user specified alignment and the size of the objects.
 115  */
 116 unsigned long calculate_alignment(unsigned long flags,
 117                 unsigned long align, unsigned long size)
 118 {
 119         /*
 120          * If the user wants hardware cache aligned objects then follow that
 121          * suggestion if the object is sufficiently large.
 122          *
 123          * The hardware cache alignment cannot override the specified
 124          * alignment though. If that is greater then use it.
 125          */
 126         if (flags & SLAB_HWCACHE_ALIGN) {
 127                 unsigned long ralign = cache_line_size();
 128                 while (size <= ralign / 2)
 129                         ralign /= 2;
 130                 align = max(align, ralign);
 131         }
 132
 133         if (align < ARCH_SLAB_MINALIGN)
 134                 align = ARCH_SLAB_MINALIGN;
 135
 136         return ALIGN(align, sizeof(void *));
 137 }
 138
 139
 140 /*
 141  * kmem_cache_create - Create a cache.
 142  * @name: A string which is used in /proc/slabinfo to identify this cache.
 143  * @size: The size of objects to be created in this cache.
 144  * @align: The required alignment for the objects.
 145  * @flags: SLAB flags
 146  * @ctor: A constructor for the objects.
 147  *
 148  * Returns a ptr to the cache on success, NULL on failure.
 149  * Cannot be called within a interrupt, but can be interrupted.
 150  * The @ctor is run when new pages are allocated by the cache.
 151  *
 152  * The flags are
 153  *
 154  * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
 155  * to catch references to uninitialised memory.
 156  *
 157  * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
 158  * for buffer overruns.
 159  *
 160  * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
 161  * cacheline.  This can be beneficial if you're counting cycles as closely
 162  * as davem.
 163  */
 164
 165 struct kmem_cache *
 166 kmem_cache_create_memcg(struct mem_cgroup *memcg, const char *name, size_t size,
 167                         size_t align, unsigned long flags, void (*ctor)(void *),
 168                         struct kmem_cache *parent_cache)
 169 {
 170         struct kmem_cache *s = NULL;
 171         int err = 0;
 172
 173         get_online_cpus();
 174         mutex_lock(&slab_mutex);
 175
 176         if (!kmem_cache_sanity_check(memcg, name, size) == 0)
 177                 goto out_locked;
 178
 179         /*
 180          * Some allocators will constraint the set of valid flags to a subset
 181          * of all flags. We expect them to define CACHE_CREATE_MASK in this
 182          * case, and we'll just provide them with a sanitized version of the
 183          * passed flags.
 184          */
 185         flags &= CACHE_CREATE_MASK;
 186
 187         s = __kmem_cache_alias(memcg, name, size, align, flags, ctor);
 188         if (s)
 189                 goto out_locked;
 190
 191         s = kmem_cache_zalloc(kmem_cache, GFP_KERNEL);
 192         if (s) {
 193                 s->object_size = s->size = size;
 194                 s->align = calculate_alignment(flags, align, size);
 195                 s->ctor = ctor;
 196
 197                 if (memcg_register_cache(memcg, s, parent_cache)) {
 198                         kmem_cache_free(kmem_cache, s);
 199                         err = -ENOMEM;
 200                         goto out_locked;
 201                 }
 202
 203                 s->name = kstrdup(name, GFP_KERNEL);
 204                 if (!s->name) {
 205                         kmem_cache_free(kmem_cache, s);
 206                         err = -ENOMEM;
 207                         goto out_locked;
 208                 }
 209
 210                 err = __kmem_cache_create(s, flags);
 211                 if (!err) {
 212                         s->refcount = 1;
 213                         list_add(&s->list, &slab_caches);
 214                         memcg_cache_list_add(memcg, s);
 215                 } else {
 216                         kfree(s->name);
 217                         kmem_cache_free(kmem_cache, s);
 218                 }
 219         } else
 220                 err = -ENOMEM;
 221
 222 out_locked:
 223         mutex_unlock(&slab_mutex);
 224         put_online_cpus();
 225
 226         if (err) {
 227
 228                 if (flags & SLAB_PANIC)
 229                         panic("kmem_cache_create: Failed to create slab '%s'. Error %d\n",
 230                                 name, err);
 231                 else {
 232                         printk(KERN_WARNING "kmem_cache_create(%s) failed with error %d",
 233                                 name, err);
 234                         dump_stack();
 235                 }
 236
 237                 return NULL;
 238         }
 239
 240         return s;
 241 }
 242
 243 struct kmem_cache *
 244 kmem_cache_create(const char *name, size_t size, size_t align,
 245                   unsigned long flags, void (*ctor)(void *))
 246 {
 247         return kmem_cache_create_memcg(NULL, name, size, align, flags, ctor, NULL);
 248 }
 249 EXPORT_SYMBOL(kmem_cache_create);
 250
 251 void kmem_cache_destroy(struct kmem_cache *s)
 252 {
 253         /* Destroy all the children caches if we aren't a memcg cache */
 254         kmem_cache_destroy_memcg_children(s);
 255
 256         get_online_cpus();
 257         mutex_lock(&slab_mutex);
 258         s->refcount--;
 259         if (!s->refcount) {
 260                 list_del(&s->list);
 261
 262                 if (!__kmem_cache_shutdown(s)) {
 263                         mutex_unlock(&slab_mutex);
 264                         if (s->flags & SLAB_DESTROY_BY_RCU)
 265                                 rcu_barrier();
 266
 267                         memcg_release_cache(s);
 268                         kfree(s->name);
 269                         kmem_cache_free(kmem_cache, s);
 270                 } else {
 271                         list_add(&s->list, &slab_caches);
 272                         mutex_unlock(&slab_mutex);
 273                         printk(KERN_ERR "kmem_cache_destroy %s: Slab cache still has objects\n",
 274                                 s->name);
 275                         dump_stack();
 276                 }
 277         } else {
 278                 mutex_unlock(&slab_mutex);
 279         }
 280         put_online_cpus();
 281 }
 282 EXPORT_SYMBOL(kmem_cache_destroy);
 283
 284 int slab_is_available(void)
 285 {
 286         return slab_state >= UP;
 287 }
 288
 289 #ifndef CONFIG_SLOB
 290 /* Create a cache during boot when no slab services are available yet */
 291 void __init create_boot_cache(struct kmem_cache *s, const char *name, size_t size,
 292                 unsigned long flags)
 293 {
 294         int err;
 295
 296         s->name = name;
 297         s->size = s->object_size = size;
 298         s->align = calculate_alignment(flags, ARCH_KMALLOC_MINALIGN, size);
 299         err = __kmem_cache_create(s, flags);
 300
 301         if (err)
 302                 panic("Creation of kmalloc slab %s size=%zu failed. Reason %d\n",
 303                                         name, size, err);
 304
 305         s->refcount = -1;       /* Exempt from merging for now */
 306 }
 307
 308 struct kmem_cache *__init create_kmalloc_cache(const char *name, size_t size,
 309                                 unsigned long flags)
 310 {
 311         struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
 312
 313         if (!s)
 314                 panic("Out of memory when creating slab %s\n", name);
 315
 316         create_boot_cache(s, name, size, flags);
 317         list_add(&s->list, &slab_caches);
 318         s->refcount = 1;
 319         return s;
 320 }
 321
 322 struct kmem_cache *kmalloc_caches[KMALLOC_SHIFT_HIGH + 1];
 323 EXPORT_SYMBOL(kmalloc_caches);
 324
 325 #ifdef CONFIG_ZONE_DMA
 326 struct kmem_cache *kmalloc_dma_caches[KMALLOC_SHIFT_HIGH + 1];
 327 EXPORT_SYMBOL(kmalloc_dma_caches);
 328 #endif
 329
 330 /*
 331  * Conversion table for small slabs sizes / 8 to the index in the
 332  * kmalloc array. This is necessary for slabs < 192 since we have non power
 333  * of two cache sizes there. The size of larger slabs can be determined using
 334  * fls.
 335  */
 336 static s8 size_index[24] = {
 337         3,      /* 8 */
 338         4,      /* 16 */
 339         5,      /* 24 */
 340         5,      /* 32 */
 341         6,      /* 40 */
 342         6,      /* 48 */
 343         6,      /* 56 */
 344         6,      /* 64 */
 345         1,      /* 72 */
 346         1,      /* 80 */
 347         1,      /* 88 */
 348         1,      /* 96 */
 349         7,      /* 104 */
 350         7,      /* 112 */
 351         7,      /* 120 */
 352         7,      /* 128 */
 353         2,      /* 136 */
 354         2,      /* 144 */
 355         2,      /* 152 */
 356         2,      /* 160 */
 357         2,      /* 168 */
 358         2,      /* 176 */
 359         2,      /* 184 */
 360         2       /* 192 */
 361 };
 362
 363 static inline int size_index_elem(size_t bytes)
 364 {
 365         return (bytes - 1) / 8;
 366 }
 367
 368 /*
 369  * Find the kmem_cache structure that serves a given size of
 370  * allocation
 371  */
 372 struct kmem_cache *kmalloc_slab(size_t size, gfp_t flags)
 373 {
 374         int index;
 375
 376         if (WARN_ON_ONCE(size > KMALLOC_MAX_SIZE))
 377                 return NULL;
 378
 379         if (size <= 192) {
 380                 if (!size)
 381                         return ZERO_SIZE_PTR;
 382
 383                 index = size_index[size_index_elem(size)];
 384         } else
 385                 index = fls(size - 1);
 386
 387 #ifdef CONFIG_ZONE_DMA
 388         if (unlikely((flags & GFP_DMA)))
 389                 return kmalloc_dma_caches[index];
 390
 391 #endif
 392         return kmalloc_caches[index];
 393 }
 394
 395 /*
 396  * Create the kmalloc array. Some of the regular kmalloc arrays
 397  * may already have been created because they were needed to
 398  * enable allocations for slab creation.
 399  */
 400 void __init create_kmalloc_caches(unsigned long flags)
 401 {
 402         int i;
 403
 404         /*
 405          * Patch up the size_index table if we have strange large alignment
 406          * requirements for the kmalloc array. This is only the case for
 407          * MIPS it seems. The standard arches will not generate any code here.
 408          *
 409          * Largest permitted alignment is 256 bytes due to the way we
 410          * handle the index determination for the smaller caches.
 411          *
 412          * Make sure that nothing crazy happens if someone starts tinkering
 413          * around with ARCH_KMALLOC_MINALIGN
 414          */
 415         BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
 416                 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
 417
 418         for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
 419                 int elem = size_index_elem(i);
 420
 421                 if (elem >= ARRAY_SIZE(size_index))
 422                         break;
 423                 size_index[elem] = KMALLOC_SHIFT_LOW;
 424         }
 425
 426         if (KMALLOC_MIN_SIZE >= 64) {
 427                 /*
 428                  * The 96 byte size cache is not used if the alignment
 429                  * is 64 byte.
 430                  */
 431                 for (i = 64 + 8; i <= 96; i += 8)
 432                         size_index[size_index_elem(i)] = 7;
 433
 434         }
 435
 436         if (KMALLOC_MIN_SIZE >= 128) {
 437                 /*
 438                  * The 192 byte sized cache is not used if the alignment
 439                  * is 128 byte. Redirect kmalloc to use the 256 byte cache
 440                  * instead.
 441                  */
 442                 for (i = 128 + 8; i <= 192; i += 8)
 443                         size_index[size_index_elem(i)] = 8;
 444         }
 445         for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++) {
 446                 if (!kmalloc_caches[i]) {
 447                         kmalloc_caches[i] = create_kmalloc_cache(NULL,
 448                                                         1 << i, flags);
 449                 }
 450
 451                 /*
 452                  * Caches that are not of the two-to-the-power-of size.
 453                  * These have to be created immediately after the
 454                  * earlier power of two caches
 455                  */
 456                 if (KMALLOC_MIN_SIZE <= 32 && !kmalloc_caches[1] && i == 6)
 457                         kmalloc_caches[1] = create_kmalloc_cache(NULL, 96, flags);
 458
 459                 if (KMALLOC_MIN_SIZE <= 64 && !kmalloc_caches[2] && i == 7)
 460                         kmalloc_caches[2] = create_kmalloc_cache(NULL, 192, flags);
 461         }
 462
 463         /* Kmalloc array is now usable */
 464         slab_state = UP;
 465
 466         for (i = 0; i <= KMALLOC_SHIFT_HIGH; i++) {
 467                 struct kmem_cache *s = kmalloc_caches[i];
 468                 char *n;
 469
 470                 if (s) {
 471                         n = kasprintf(GFP_NOWAIT, "kmalloc-%d", kmalloc_size(i));
 472
 473                         BUG_ON(!n);
 474                         s->name = n;
 475                 }
 476         }
 477
 478 #ifdef CONFIG_ZONE_DMA
 479         for (i = 0; i <= KMALLOC_SHIFT_HIGH; i++) {
 480                 struct kmem_cache *s = kmalloc_caches[i];
 481
 482                 if (s) {
 483                         int size = kmalloc_size(i);
 484                         char *n = kasprintf(GFP_NOWAIT,
 485                                  "dma-kmalloc-%d", size);
 486
 487                         BUG_ON(!n);
 488                         kmalloc_dma_caches[i] = create_kmalloc_cache(n,
 489                                 size, SLAB_CACHE_DMA | flags);
 490                 }
 491         }
 492 #endif
 493 }
 494 #endif /* !CONFIG_SLOB */
 495
 496
 497 #ifdef CONFIG_SLABINFO
 498 void print_slabinfo_header(struct seq_file *m)
 499 {
 500         /*
 501          * Output format version, so at least we can change it
 502          * without _too_ many complaints.
 503          */
 504 #ifdef CONFIG_DEBUG_SLAB
 505         seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
 506 #else
 507         seq_puts(m, "slabinfo - version: 2.1\n");
 508 #endif
 509         seq_puts(m, "# name            <active_objs> <num_objs> <objsize> "
 510                  "<objperslab> <pagesperslab>");
 511         seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
 512         seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
 513 #ifdef CONFIG_DEBUG_SLAB
 514         seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
 515                  "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
 516         seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
 517 #endif
 518         seq_putc(m, '\n');
 519 }
 520
 521 static void *s_start(struct seq_file *m, loff_t *pos)
 522 {
 523         loff_t n = *pos;
 524
 525         mutex_lock(&slab_mutex);
 526         if (!n)
 527                 print_slabinfo_header(m);
 528
 529         return seq_list_start(&slab_caches, *pos);
 530 }
 531
 532 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
 533 {
 534         return seq_list_next(p, &slab_caches, pos);
 535 }
 536
 537 static void s_stop(struct seq_file *m, void *p)
 538 {
 539         mutex_unlock(&slab_mutex);
 540 }
 541
 542 static void
 543 memcg_accumulate_slabinfo(struct kmem_cache *s, struct slabinfo *info)
 544 {
 545         struct kmem_cache *c;
 546         struct slabinfo sinfo;
 547         int i;
 548
 549         if (!is_root_cache(s))
 550                 return;
 551
 552         for_each_memcg_cache_index(i) {
 553                 c = cache_from_memcg(s, i);
 554                 if (!c)
 555                         continue;
 556
 557                 memset(&sinfo, 0, sizeof(sinfo));
 558                 get_slabinfo(c, &sinfo);
 559
 560                 info->active_slabs += sinfo.active_slabs;
 561                 info->num_slabs += sinfo.num_slabs;
 562                 info->shared_avail += sinfo.shared_avail;
 563                 info->active_objs += sinfo.active_objs;
 564                 info->num_objs += sinfo.num_objs;
 565         }
 566 }
 567
 568 int cache_show(struct kmem_cache *s, struct seq_file *m)
 569 {
 570         struct slabinfo sinfo;
 571
 572         memset(&sinfo, 0, sizeof(sinfo));
 573         get_slabinfo(s, &sinfo);
 574
 575         memcg_accumulate_slabinfo(s, &sinfo);
 576
 577         seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
 578                    cache_name(s), sinfo.active_objs, sinfo.num_objs, s->size,
 579                    sinfo.objects_per_slab, (1 << sinfo.cache_order));
 580
 581         seq_printf(m, " : tunables %4u %4u %4u",
 582                    sinfo.limit, sinfo.batchcount, sinfo.shared);
 583         seq_printf(m, " : slabdata %6lu %6lu %6lu",
 584                    sinfo.active_slabs, sinfo.num_slabs, sinfo.shared_avail);
 585         slabinfo_show_stats(m, s);
 586         seq_putc(m, '\n');
 587         return 0;
 588 }
 589
 590 static int s_show(struct seq_file *m, void *p)
 591 {
 592         struct kmem_cache *s = list_entry(p, struct kmem_cache, list);
 593
 594         if (!is_root_cache(s))
 595                 return 0;
 596         return cache_show(s, m);
 597 }
 598
 599 /*
 600  * slabinfo_op - iterator that generates /proc/slabinfo
 601  *
 602  * Output layout:
 603  * cache-name
 604  * num-active-objs
 605  * total-objs
 606  * object size
 607  * num-active-slabs
 608  * total-slabs
 609  * num-pages-per-slab
 610  * + further values on SMP and with statistics enabled
 611  */
 612 static const struct seq_operations slabinfo_op = {
 613         .start = s_start,
 614         .next = s_next,
 615         .stop = s_stop,
 616         .show = s_show,
 617 };
 618
 619 static int slabinfo_open(struct inode *inode, struct file *file)
 620 {
 621         return seq_open(file, &slabinfo_op);
 622 }
 623
 624 static const struct file_operations proc_slabinfo_operations = {
 625         .open           = slabinfo_open,
 626         .read           = seq_read,
 627         .write          = slabinfo_write,
 628         .llseek         = seq_lseek,
 629         .release        = seq_release,
 630 };
 631
 632 static int __init slab_proc_init(void)
 633 {
 634         proc_create("slabinfo", S_IRUSR, NULL, &proc_slabinfo_operations);
 635         return 0;
 636 }
 637 module_init(slab_proc_init);
 638 #endif /* CONFIG_SLABINFO */