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 #define CREATE_TRACE_POINTS
  24 #include <trace/events/kmem.h>
  25
  26 #include "slab.h"
  27
  28 enum slab_state slab_state;
  29 LIST_HEAD(slab_caches);
  30 DEFINE_MUTEX(slab_mutex);
  31 struct kmem_cache *kmem_cache;
  32
  33 /*
  34  * Set of flags that will prevent slab merging
  35  */
  36 #define SLAB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
  37                 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
  38                 SLAB_FAILSLAB | SLAB_KASAN)
  39
  40 #define SLAB_MERGE_SAME (SLAB_RECLAIM_ACCOUNT | SLAB_CACHE_DMA | SLAB_NOTRACK)
  41
  42 /*
  43  * Merge control. If this is set then no merging of slab caches will occur.
  44  * (Could be removed. This was introduced to pacify the merge skeptics.)
  45  */
  46 static int slab_nomerge;
  47
  48 static int __init setup_slab_nomerge(char *str)
  49 {
  50         slab_nomerge = 1;
  51         return 1;
  52 }
  53
  54 #ifdef CONFIG_SLUB
  55 __setup_param("slub_nomerge", slub_nomerge, setup_slab_nomerge, 0);
  56 #endif
  57
  58 __setup("slab_nomerge", setup_slab_nomerge);
  59
  60 /*
  61  * Determine the size of a slab object
  62  */
  63 unsigned int kmem_cache_size(struct kmem_cache *s)
  64 {
  65         return s->object_size;
  66 }
  67 EXPORT_SYMBOL(kmem_cache_size);
  68
  69 #ifdef CONFIG_DEBUG_VM
  70 static int kmem_cache_sanity_check(const char *name, size_t size)
  71 {
  72         struct kmem_cache *s = NULL;
  73
  74         if (!name || in_interrupt() || size < sizeof(void *) ||
  75                 size > KMALLOC_MAX_SIZE) {
  76                 pr_err("kmem_cache_create(%s) integrity check failed\n", name);
  77                 return -EINVAL;
  78         }
  79
  80         list_for_each_entry(s, &slab_caches, list) {
  81                 char tmp;
  82                 int res;
  83
  84                 /*
  85                  * This happens when the module gets unloaded and doesn't
  86                  * destroy its slab cache and no-one else reuses the vmalloc
  87                  * area of the module.  Print a warning.
  88                  */
  89                 res = probe_kernel_address(s->name, tmp);
  90                 if (res) {
  91                         pr_err("Slab cache with size %d has lost its name\n",
  92                                s->object_size);
  93                         continue;
  94                 }
  95         }
  96
  97         WARN_ON(strchr(name, ' '));     /* It confuses parsers */
  98         return 0;
  99 }
 100 #else
 101 static inline int kmem_cache_sanity_check(const char *name, size_t size)
 102 {
 103         return 0;
 104 }
 105 #endif
 106
 107 void __kmem_cache_free_bulk(struct kmem_cache *s, size_t nr, void **p)
 108 {
 109         size_t i;
 110
 111         for (i = 0; i < nr; i++)
 112                 kmem_cache_free(s, p[i]);
 113 }
 114
 115 int __kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t nr,
 116                                                                 void **p)
 117 {
 118         size_t i;
 119
 120         for (i = 0; i < nr; i++) {
 121                 void *x = p[i] = kmem_cache_alloc(s, flags);
 122                 if (!x) {
 123                         __kmem_cache_free_bulk(s, i, p);
 124                         return 0;
 125                 }
 126         }
 127         return i;
 128 }
 129
 130 #ifdef CONFIG_MEMCG_KMEM
 131 void slab_init_memcg_params(struct kmem_cache *s)
 132 {
 133         s->memcg_params.is_root_cache = true;
 134         INIT_LIST_HEAD(&s->memcg_params.list);
 135         RCU_INIT_POINTER(s->memcg_params.memcg_caches, NULL);
 136 }
 137
 138 static int init_memcg_params(struct kmem_cache *s,
 139                 struct mem_cgroup *memcg, struct kmem_cache *root_cache)
 140 {
 141         struct memcg_cache_array *arr;
 142
 143         if (memcg) {
 144                 s->memcg_params.is_root_cache = false;
 145                 s->memcg_params.memcg = memcg;
 146                 s->memcg_params.root_cache = root_cache;
 147                 return 0;
 148         }
 149
 150         slab_init_memcg_params(s);
 151
 152         if (!memcg_nr_cache_ids)
 153                 return 0;
 154
 155         arr = kzalloc(sizeof(struct memcg_cache_array) +
 156                       memcg_nr_cache_ids * sizeof(void *),
 157                       GFP_KERNEL);
 158         if (!arr)
 159                 return -ENOMEM;
 160
 161         RCU_INIT_POINTER(s->memcg_params.memcg_caches, arr);
 162         return 0;
 163 }
 164
 165 static void destroy_memcg_params(struct kmem_cache *s)
 166 {
 167         if (is_root_cache(s))
 168                 kfree(rcu_access_pointer(s->memcg_params.memcg_caches));
 169 }
 170
 171 static int update_memcg_params(struct kmem_cache *s, int new_array_size)
 172 {
 173         struct memcg_cache_array *old, *new;
 174
 175         if (!is_root_cache(s))
 176                 return 0;
 177
 178         new = kzalloc(sizeof(struct memcg_cache_array) +
 179                       new_array_size * sizeof(void *), GFP_KERNEL);
 180         if (!new)
 181                 return -ENOMEM;
 182
 183         old = rcu_dereference_protected(s->memcg_params.memcg_caches,
 184                                         lockdep_is_held(&slab_mutex));
 185         if (old)
 186                 memcpy(new->entries, old->entries,
 187                        memcg_nr_cache_ids * sizeof(void *));
 188
 189         rcu_assign_pointer(s->memcg_params.memcg_caches, new);
 190         if (old)
 191                 kfree_rcu(old, rcu);
 192         return 0;
 193 }
 194
 195 int memcg_update_all_caches(int num_memcgs)
 196 {
 197         struct kmem_cache *s;
 198         int ret = 0;
 199
 200         mutex_lock(&slab_mutex);
 201         list_for_each_entry(s, &slab_caches, list) {
 202                 ret = update_memcg_params(s, num_memcgs);
 203                 /*
 204                  * Instead of freeing the memory, we'll just leave the caches
 205                  * up to this point in an updated state.
 206                  */
 207                 if (ret)
 208                         break;
 209         }
 210         mutex_unlock(&slab_mutex);
 211         return ret;
 212 }
 213 #else
 214 static inline int init_memcg_params(struct kmem_cache *s,
 215                 struct mem_cgroup *memcg, struct kmem_cache *root_cache)
 216 {
 217         return 0;
 218 }
 219
 220 static inline void destroy_memcg_params(struct kmem_cache *s)
 221 {
 222 }
 223 #endif /* CONFIG_MEMCG_KMEM */
 224
 225 /*
 226  * Find a mergeable slab cache
 227  */
 228 int slab_unmergeable(struct kmem_cache *s)
 229 {
 230         if (slab_nomerge || (s->flags & SLAB_NEVER_MERGE))
 231                 return 1;
 232
 233         if (!is_root_cache(s))
 234                 return 1;
 235
 236         if (s->ctor)
 237                 return 1;
 238
 239         /*
 240          * We may have set a slab to be unmergeable during bootstrap.
 241          */
 242         if (s->refcount < 0)
 243                 return 1;
 244
 245         return 0;
 246 }
 247
 248 struct kmem_cache *find_mergeable(size_t size, size_t align,
 249                 unsigned long flags, const char *name, void (*ctor)(void *))
 250 {
 251         struct kmem_cache *s;
 252
 253         if (slab_nomerge)
 254                 return NULL;
 255
 256         if (ctor)
 257                 return NULL;
 258
 259         size = ALIGN(size, sizeof(void *));
 260         align = calculate_alignment(flags, align, size);
 261         size = ALIGN(size, align);
 262         flags = kmem_cache_flags(size, flags, name, NULL);
 263
 264         if (flags & SLAB_NEVER_MERGE)
 265                 return NULL;
 266
 267         list_for_each_entry_reverse(s, &slab_caches, list) {
 268                 if (slab_unmergeable(s))
 269                         continue;
 270
 271                 if (size > s->size)
 272                         continue;
 273
 274                 if ((flags & SLAB_MERGE_SAME) != (s->flags & SLAB_MERGE_SAME))
 275                         continue;
 276                 /*
 277                  * Check if alignment is compatible.
 278                  * Courtesy of Adrian Drzewiecki
 279                  */
 280                 if ((s->size & ~(align - 1)) != s->size)
 281                         continue;
 282
 283                 if (s->size - size >= sizeof(void *))
 284                         continue;
 285
 286                 if (IS_ENABLED(CONFIG_SLAB) && align &&
 287                         (align > s->align || s->align % align))
 288                         continue;
 289
 290                 return s;
 291         }
 292         return NULL;
 293 }
 294
 295 /*
 296  * Figure out what the alignment of the objects will be given a set of
 297  * flags, a user specified alignment and the size of the objects.
 298  */
 299 unsigned long calculate_alignment(unsigned long flags,
 300                 unsigned long align, unsigned long size)
 301 {
 302         /*
 303          * If the user wants hardware cache aligned objects then follow that
 304          * suggestion if the object is sufficiently large.
 305          *
 306          * The hardware cache alignment cannot override the specified
 307          * alignment though. If that is greater then use it.
 308          */
 309         if (flags & SLAB_HWCACHE_ALIGN) {
 310                 unsigned long ralign = cache_line_size();
 311                 while (size <= ralign / 2)
 312                         ralign /= 2;
 313                 align = max(align, ralign);
 314         }
 315
 316         if (align < ARCH_SLAB_MINALIGN)
 317                 align = ARCH_SLAB_MINALIGN;
 318
 319         return ALIGN(align, sizeof(void *));
 320 }
 321
 322 static struct kmem_cache *create_cache(const char *name,
 323                 size_t object_size, size_t size, size_t align,
 324                 unsigned long flags, void (*ctor)(void *),
 325                 struct mem_cgroup *memcg, struct kmem_cache *root_cache)
 326 {
 327         struct kmem_cache *s;
 328         int err;
 329
 330         err = -ENOMEM;
 331         s = kmem_cache_zalloc(kmem_cache, GFP_KERNEL);
 332         if (!s)
 333                 goto out;
 334
 335         s->name = name;
 336         s->object_size = object_size;
 337         s->size = size;
 338         s->align = align;
 339         s->ctor = ctor;
 340
 341         err = init_memcg_params(s, memcg, root_cache);
 342         if (err)
 343                 goto out_free_cache;
 344
 345         err = __kmem_cache_create(s, flags);
 346         if (err)
 347                 goto out_free_cache;
 348
 349         s->refcount = 1;
 350         list_add(&s->list, &slab_caches);
 351 out:
 352         if (err)
 353                 return ERR_PTR(err);
 354         return s;
 355
 356 out_free_cache:
 357         destroy_memcg_params(s);
 358         kmem_cache_free(kmem_cache, s);
 359         goto out;
 360 }
 361
 362 /*
 363  * kmem_cache_create - Create a cache.
 364  * @name: A string which is used in /proc/slabinfo to identify this cache.
 365  * @size: The size of objects to be created in this cache.
 366  * @align: The required alignment for the objects.
 367  * @flags: SLAB flags
 368  * @ctor: A constructor for the objects.
 369  *
 370  * Returns a ptr to the cache on success, NULL on failure.
 371  * Cannot be called within a interrupt, but can be interrupted.
 372  * The @ctor is run when new pages are allocated by the cache.
 373  *
 374  * The flags are
 375  *
 376  * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
 377  * to catch references to uninitialised memory.
 378  *
 379  * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
 380  * for buffer overruns.
 381  *
 382  * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
 383  * cacheline.  This can be beneficial if you're counting cycles as closely
 384  * as davem.
 385  */
 386 struct kmem_cache *
 387 kmem_cache_create(const char *name, size_t size, size_t align,
 388                   unsigned long flags, void (*ctor)(void *))
 389 {
 390         struct kmem_cache *s = NULL;
 391         const char *cache_name;
 392         int err;
 393
 394         get_online_cpus();
 395         get_online_mems();
 396         memcg_get_cache_ids();
 397
 398         mutex_lock(&slab_mutex);
 399
 400         err = kmem_cache_sanity_check(name, size);
 401         if (err) {
 402                 goto out_unlock;
 403         }
 404
 405         /*
 406          * Some allocators will constraint the set of valid flags to a subset
 407          * of all flags. We expect them to define CACHE_CREATE_MASK in this
 408          * case, and we'll just provide them with a sanitized version of the
 409          * passed flags.
 410          */
 411         flags &= CACHE_CREATE_MASK;
 412
 413         s = __kmem_cache_alias(name, size, align, flags, ctor);
 414         if (s)
 415                 goto out_unlock;
 416
 417         cache_name = kstrdup_const(name, GFP_KERNEL);
 418         if (!cache_name) {
 419                 err = -ENOMEM;
 420                 goto out_unlock;
 421         }
 422
 423         s = create_cache(cache_name, size, size,
 424                          calculate_alignment(flags, align, size),
 425                          flags, ctor, NULL, NULL);
 426         if (IS_ERR(s)) {
 427                 err = PTR_ERR(s);
 428                 kfree_const(cache_name);
 429         }
 430
 431 out_unlock:
 432         mutex_unlock(&slab_mutex);
 433
 434         memcg_put_cache_ids();
 435         put_online_mems();
 436         put_online_cpus();
 437
 438         if (err) {
 439                 if (flags & SLAB_PANIC)
 440                         panic("kmem_cache_create: Failed to create slab '%s'. Error %d\n",
 441                                 name, err);
 442                 else {
 443                         printk(KERN_WARNING "kmem_cache_create(%s) failed with error %d",
 444                                 name, err);
 445                         dump_stack();
 446                 }
 447                 return NULL;
 448         }
 449         return s;
 450 }
 451 EXPORT_SYMBOL(kmem_cache_create);
 452
 453 static int shutdown_cache(struct kmem_cache *s,
 454                 struct list_head *release, bool *need_rcu_barrier)
 455 {
 456         if (__kmem_cache_shutdown(s) != 0)
 457                 return -EBUSY;
 458
 459         if (s->flags & SLAB_DESTROY_BY_RCU)
 460                 *need_rcu_barrier = true;
 461
 462         list_move(&s->list, release);
 463         return 0;
 464 }
 465
 466 static void release_caches(struct list_head *release, bool need_rcu_barrier)
 467 {
 468         struct kmem_cache *s, *s2;
 469
 470         if (need_rcu_barrier)
 471                 rcu_barrier();
 472
 473         list_for_each_entry_safe(s, s2, release, list) {
 474 #ifdef SLAB_SUPPORTS_SYSFS
 475                 sysfs_slab_remove(s);
 476 #else
 477                 slab_kmem_cache_release(s);
 478 #endif
 479         }
 480 }
 481
 482 #ifdef CONFIG_MEMCG_KMEM
 483 /*
 484  * memcg_create_kmem_cache - Create a cache for a memory cgroup.
 485  * @memcg: The memory cgroup the new cache is for.
 486  * @root_cache: The parent of the new cache.
 487  *
 488  * This function attempts to create a kmem cache that will serve allocation
 489  * requests going from @memcg to @root_cache. The new cache inherits properties
 490  * from its parent.
 491  */
 492 void memcg_create_kmem_cache(struct mem_cgroup *memcg,
 493                              struct kmem_cache *root_cache)
 494 {
 495         static char memcg_name_buf[NAME_MAX + 1]; /* protected by slab_mutex */
 496         struct cgroup_subsys_state *css = &memcg->css;
 497         struct memcg_cache_array *arr;
 498         struct kmem_cache *s = NULL;
 499         char *cache_name;
 500         int idx;
 501
 502         get_online_cpus();
 503         get_online_mems();
 504
 505         mutex_lock(&slab_mutex);
 506
 507         /*
 508          * The memory cgroup could have been deactivated while the cache
 509          * creation work was pending.
 510          */
 511         if (!memcg_kmem_is_active(memcg))
 512                 goto out_unlock;
 513
 514         idx = memcg_cache_id(memcg);
 515         arr = rcu_dereference_protected(root_cache->memcg_params.memcg_caches,
 516                                         lockdep_is_held(&slab_mutex));
 517
 518         /*
 519          * Since per-memcg caches are created asynchronously on first
 520          * allocation (see memcg_kmem_get_cache()), several threads can try to
 521          * create the same cache, but only one of them may succeed.
 522          */
 523         if (arr->entries[idx])
 524                 goto out_unlock;
 525
 526         cgroup_name(css->cgroup, memcg_name_buf, sizeof(memcg_name_buf));
 527         cache_name = kasprintf(GFP_KERNEL, "%s(%llu:%s)", root_cache->name,
 528                                css->serial_nr, memcg_name_buf);
 529         if (!cache_name)
 530                 goto out_unlock;
 531
 532         s = create_cache(cache_name, root_cache->object_size,
 533                          root_cache->size, root_cache->align,
 534                          root_cache->flags, root_cache->ctor,
 535                          memcg, root_cache);
 536         /*
 537          * If we could not create a memcg cache, do not complain, because
 538          * that's not critical at all as we can always proceed with the root
 539          * cache.
 540          */
 541         if (IS_ERR(s)) {
 542                 kfree(cache_name);
 543                 goto out_unlock;
 544         }
 545
 546         list_add(&s->memcg_params.list, &root_cache->memcg_params.list);
 547
 548         /*
 549          * Since readers won't lock (see cache_from_memcg_idx()), we need a
 550          * barrier here to ensure nobody will see the kmem_cache partially
 551          * initialized.
 552          */
 553         smp_wmb();
 554         arr->entries[idx] = s;
 555
 556 out_unlock:
 557         mutex_unlock(&slab_mutex);
 558
 559         put_online_mems();
 560         put_online_cpus();
 561 }
 562
 563 void memcg_deactivate_kmem_caches(struct mem_cgroup *memcg)
 564 {
 565         int idx;
 566         struct memcg_cache_array *arr;
 567         struct kmem_cache *s, *c;
 568
 569         idx = memcg_cache_id(memcg);
 570
 571         get_online_cpus();
 572         get_online_mems();
 573
 574         mutex_lock(&slab_mutex);
 575         list_for_each_entry(s, &slab_caches, list) {
 576                 if (!is_root_cache(s))
 577                         continue;
 578
 579                 arr = rcu_dereference_protected(s->memcg_params.memcg_caches,
 580                                                 lockdep_is_held(&slab_mutex));
 581                 c = arr->entries[idx];
 582                 if (!c)
 583                         continue;
 584
 585                 __kmem_cache_shrink(c, true);
 586                 arr->entries[idx] = NULL;
 587         }
 588         mutex_unlock(&slab_mutex);
 589
 590         put_online_mems();
 591         put_online_cpus();
 592 }
 593
 594 static int __shutdown_memcg_cache(struct kmem_cache *s,
 595                 struct list_head *release, bool *need_rcu_barrier)
 596 {
 597         BUG_ON(is_root_cache(s));
 598
 599         if (shutdown_cache(s, release, need_rcu_barrier))
 600                 return -EBUSY;
 601
 602         list_del(&s->memcg_params.list);
 603         return 0;
 604 }
 605
 606 void memcg_destroy_kmem_caches(struct mem_cgroup *memcg)
 607 {
 608         LIST_HEAD(release);
 609         bool need_rcu_barrier = false;
 610         struct kmem_cache *s, *s2;
 611
 612         get_online_cpus();
 613         get_online_mems();
 614
 615         mutex_lock(&slab_mutex);
 616         list_for_each_entry_safe(s, s2, &slab_caches, list) {
 617                 if (is_root_cache(s) || s->memcg_params.memcg != memcg)
 618                         continue;
 619                 /*
 620                  * The cgroup is about to be freed and therefore has no charges
 621                  * left. Hence, all its caches must be empty by now.
 622                  */
 623                 BUG_ON(__shutdown_memcg_cache(s, &release, &need_rcu_barrier));
 624         }
 625         mutex_unlock(&slab_mutex);
 626
 627         put_online_mems();
 628         put_online_cpus();
 629
 630         release_caches(&release, need_rcu_barrier);
 631 }
 632
 633 static int shutdown_memcg_caches(struct kmem_cache *s,
 634                 struct list_head *release, bool *need_rcu_barrier)
 635 {
 636         struct memcg_cache_array *arr;
 637         struct kmem_cache *c, *c2;
 638         LIST_HEAD(busy);
 639         int i;
 640
 641         BUG_ON(!is_root_cache(s));
 642
 643         /*
 644          * First, shutdown active caches, i.e. caches that belong to online
 645          * memory cgroups.
 646          */
 647         arr = rcu_dereference_protected(s->memcg_params.memcg_caches,
 648                                         lockdep_is_held(&slab_mutex));
 649         for_each_memcg_cache_index(i) {
 650                 c = arr->entries[i];
 651                 if (!c)
 652                         continue;
 653                 if (__shutdown_memcg_cache(c, release, need_rcu_barrier))
 654                         /*
 655                          * The cache still has objects. Move it to a temporary
 656                          * list so as not to try to destroy it for a second
 657                          * time while iterating over inactive caches below.
 658                          */
 659                         list_move(&c->memcg_params.list, &busy);
 660                 else
 661                         /*
 662                          * The cache is empty and will be destroyed soon. Clear
 663                          * the pointer to it in the memcg_caches array so that
 664                          * it will never be accessed even if the root cache
 665                          * stays alive.
 666                          */
 667                         arr->entries[i] = NULL;
 668         }
 669
 670         /*
 671          * Second, shutdown all caches left from memory cgroups that are now
 672          * offline.
 673          */
 674         list_for_each_entry_safe(c, c2, &s->memcg_params.list,
 675                                  memcg_params.list)
 676                 __shutdown_memcg_cache(c, release, need_rcu_barrier);
 677
 678         list_splice(&busy, &s->memcg_params.list);
 679
 680         /*
 681          * A cache being destroyed must be empty. In particular, this means
 682          * that all per memcg caches attached to it must be empty too.
 683          */
 684         if (!list_empty(&s->memcg_params.list))
 685                 return -EBUSY;
 686         return 0;
 687 }
 688 #else
 689 static inline int shutdown_memcg_caches(struct kmem_cache *s,
 690                 struct list_head *release, bool *need_rcu_barrier)
 691 {
 692         return 0;
 693 }
 694 #endif /* CONFIG_MEMCG_KMEM */
 695
 696 void slab_kmem_cache_release(struct kmem_cache *s)
 697 {
 698         destroy_memcg_params(s);
 699         kfree_const(s->name);
 700         kmem_cache_free(kmem_cache, s);
 701 }
 702
 703 void kmem_cache_destroy(struct kmem_cache *s)
 704 {
 705         LIST_HEAD(release);
 706         bool need_rcu_barrier = false;
 707         int err;
 708
 709         if (unlikely(!s))
 710                 return;
 711
 712         get_online_cpus();
 713         get_online_mems();
 714
 715         kasan_cache_destroy(s);
 716         mutex_lock(&slab_mutex);
 717
 718         s->refcount--;
 719         if (s->refcount)
 720                 goto out_unlock;
 721
 722         err = shutdown_memcg_caches(s, &release, &need_rcu_barrier);
 723         if (!err)
 724                 err = shutdown_cache(s, &release, &need_rcu_barrier);
 725
 726         if (err) {
 727                 pr_err("kmem_cache_destroy %s: Slab cache still has objects\n",
 728                        s->name);
 729                 dump_stack();
 730         }
 731 out_unlock:
 732         mutex_unlock(&slab_mutex);
 733
 734         put_online_mems();
 735         put_online_cpus();
 736
 737         release_caches(&release, need_rcu_barrier);
 738 }
 739 EXPORT_SYMBOL(kmem_cache_destroy);
 740
 741 /**
 742  * kmem_cache_shrink - Shrink a cache.
 743  * @cachep: The cache to shrink.
 744  *
 745  * Releases as many slabs as possible for a cache.
 746  * To help debugging, a zero exit status indicates all slabs were released.
 747  */
 748 int kmem_cache_shrink(struct kmem_cache *cachep)
 749 {
 750         int ret;
 751
 752         get_online_cpus();
 753         get_online_mems();
 754         kasan_cache_shrink(cachep);
 755         ret = __kmem_cache_shrink(cachep, false);
 756         put_online_mems();
 757         put_online_cpus();
 758         return ret;
 759 }
 760 EXPORT_SYMBOL(kmem_cache_shrink);
 761
 762 bool slab_is_available(void)
 763 {
 764         return slab_state >= UP;
 765 }
 766
 767 #ifndef CONFIG_SLOB
 768 /* Create a cache during boot when no slab services are available yet */
 769 void __init create_boot_cache(struct kmem_cache *s, const char *name, size_t size,
 770                 unsigned long flags)
 771 {
 772         int err;
 773
 774         s->name = name;
 775         s->size = s->object_size = size;
 776         s->align = calculate_alignment(flags, ARCH_KMALLOC_MINALIGN, size);
 777
 778         slab_init_memcg_params(s);
 779
 780         err = __kmem_cache_create(s, flags);
 781
 782         if (err)
 783                 panic("Creation of kmalloc slab %s size=%zu failed. Reason %d\n",
 784                                         name, size, err);
 785
 786         s->refcount = -1;       /* Exempt from merging for now */
 787 }
 788
 789 struct kmem_cache *__init create_kmalloc_cache(const char *name, size_t size,
 790                                 unsigned long flags)
 791 {
 792         struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
 793
 794         if (!s)
 795                 panic("Out of memory when creating slab %s\n", name);
 796
 797         create_boot_cache(s, name, size, flags);
 798         list_add(&s->list, &slab_caches);
 799         s->refcount = 1;
 800         return s;
 801 }
 802
 803 struct kmem_cache *kmalloc_caches[KMALLOC_SHIFT_HIGH + 1];
 804 EXPORT_SYMBOL(kmalloc_caches);
 805
 806 #ifdef CONFIG_ZONE_DMA
 807 struct kmem_cache *kmalloc_dma_caches[KMALLOC_SHIFT_HIGH + 1];
 808 EXPORT_SYMBOL(kmalloc_dma_caches);
 809 #endif
 810
 811 /*
 812  * Conversion table for small slabs sizes / 8 to the index in the
 813  * kmalloc array. This is necessary for slabs < 192 since we have non power
 814  * of two cache sizes there. The size of larger slabs can be determined using
 815  * fls.
 816  */
 817 static s8 size_index[24] = {
 818         3,      /* 8 */
 819         4,      /* 16 */
 820         5,      /* 24 */
 821         5,      /* 32 */
 822         6,      /* 40 */
 823         6,      /* 48 */
 824         6,      /* 56 */
 825         6,      /* 64 */
 826         1,      /* 72 */
 827         1,      /* 80 */
 828         1,      /* 88 */
 829         1,      /* 96 */
 830         7,      /* 104 */
 831         7,      /* 112 */
 832         7,      /* 120 */
 833         7,      /* 128 */
 834         2,      /* 136 */
 835         2,      /* 144 */
 836         2,      /* 152 */
 837         2,      /* 160 */
 838         2,      /* 168 */
 839         2,      /* 176 */
 840         2,      /* 184 */
 841         2       /* 192 */
 842 };
 843
 844 static inline int size_index_elem(size_t bytes)
 845 {
 846         return (bytes - 1) / 8;
 847 }
 848
 849 /*
 850  * Find the kmem_cache structure that serves a given size of
 851  * allocation
 852  */
 853 struct kmem_cache *kmalloc_slab(size_t size, gfp_t flags)
 854 {
 855         int index;
 856
 857         if (unlikely(size > KMALLOC_MAX_SIZE)) {
 858                 WARN_ON_ONCE(!(flags & __GFP_NOWARN));
 859                 return NULL;
 860         }
 861
 862         if (size <= 192) {
 863                 if (!size)
 864                         return ZERO_SIZE_PTR;
 865
 866                 index = size_index[size_index_elem(size)];
 867         } else
 868                 index = fls(size - 1);
 869
 870 #ifdef CONFIG_ZONE_DMA
 871         if (unlikely((flags & GFP_DMA)))
 872                 return kmalloc_dma_caches[index];
 873
 874 #endif
 875         return kmalloc_caches[index];
 876 }
 877
 878 /*
 879  * kmalloc_info[] is to make slub_debug=,kmalloc-xx option work at boot time.
 880  * kmalloc_index() supports up to 2^26=64MB, so the final entry of the table is
 881  * kmalloc-67108864.
 882  */
 883 static struct {
 884         const char *name;
 885         unsigned long size;
 886 } const kmalloc_info[] __initconst = {
 887         {NULL,                      0},         {"kmalloc-96",             96},
 888         {"kmalloc-192",           192},         {"kmalloc-8",               8},
 889         {"kmalloc-16",             16},         {"kmalloc-32",             32},
 890         {"kmalloc-64",             64},         {"kmalloc-128",           128},
 891         {"kmalloc-256",           256},         {"kmalloc-512",           512},
 892         {"kmalloc-1024",         1024},         {"kmalloc-2048",         2048},
 893         {"kmalloc-4096",         4096},         {"kmalloc-8192",         8192},
 894         {"kmalloc-16384",       16384},         {"kmalloc-32768",       32768},
 895         {"kmalloc-65536",       65536},         {"kmalloc-131072",     131072},
 896         {"kmalloc-262144",     262144},         {"kmalloc-524288",     524288},
 897         {"kmalloc-1048576",   1048576},         {"kmalloc-2097152",   2097152},
 898         {"kmalloc-4194304",   4194304},         {"kmalloc-8388608",   8388608},
 899         {"kmalloc-16777216", 16777216},         {"kmalloc-33554432", 33554432},
 900         {"kmalloc-67108864", 67108864}
 901 };
 902
 903 /*
 904  * Patch up the size_index table if we have strange large alignment
 905  * requirements for the kmalloc array. This is only the case for
 906  * MIPS it seems. The standard arches will not generate any code here.
 907  *
 908  * Largest permitted alignment is 256 bytes due to the way we
 909  * handle the index determination for the smaller caches.
 910  *
 911  * Make sure that nothing crazy happens if someone starts tinkering
 912  * around with ARCH_KMALLOC_MINALIGN
 913  */
 914 void __init setup_kmalloc_cache_index_table(void)
 915 {
 916         int i;
 917
 918         BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
 919                 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
 920
 921         for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
 922                 int elem = size_index_elem(i);
 923
 924                 if (elem >= ARRAY_SIZE(size_index))
 925                         break;
 926                 size_index[elem] = KMALLOC_SHIFT_LOW;
 927         }
 928
 929         if (KMALLOC_MIN_SIZE >= 64) {
 930                 /*
 931                  * The 96 byte size cache is not used if the alignment
 932                  * is 64 byte.
 933                  */
 934                 for (i = 64 + 8; i <= 96; i += 8)
 935                         size_index[size_index_elem(i)] = 7;
 936
 937         }
 938
 939         if (KMALLOC_MIN_SIZE >= 128) {
 940                 /*
 941                  * The 192 byte sized cache is not used if the alignment
 942                  * is 128 byte. Redirect kmalloc to use the 256 byte cache
 943                  * instead.
 944                  */
 945                 for (i = 128 + 8; i <= 192; i += 8)
 946                         size_index[size_index_elem(i)] = 8;
 947         }
 948 }
 949
 950 static void __init new_kmalloc_cache(int idx, unsigned long flags)
 951 {
 952         kmalloc_caches[idx] = create_kmalloc_cache(kmalloc_info[idx].name,
 953                                         kmalloc_info[idx].size, flags);
 954 }
 955
 956 /*
 957  * Create the kmalloc array. Some of the regular kmalloc arrays
 958  * may already have been created because they were needed to
 959  * enable allocations for slab creation.
 960  */
 961 void __init create_kmalloc_caches(unsigned long flags)
 962 {
 963         int i;
 964
 965         for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++) {
 966                 if (!kmalloc_caches[i])
 967                         new_kmalloc_cache(i, flags);
 968
 969                 /*
 970                  * Caches that are not of the two-to-the-power-of size.
 971                  * These have to be created immediately after the
 972                  * earlier power of two caches
 973                  */
 974                 if (KMALLOC_MIN_SIZE <= 32 && !kmalloc_caches[1] && i == 6)
 975                         new_kmalloc_cache(1, flags);
 976                 if (KMALLOC_MIN_SIZE <= 64 && !kmalloc_caches[2] && i == 7)
 977                         new_kmalloc_cache(2, flags);
 978         }
 979
 980         /* Kmalloc array is now usable */
 981         slab_state = UP;
 982
 983 #ifdef CONFIG_ZONE_DMA
 984         for (i = 0; i <= KMALLOC_SHIFT_HIGH; i++) {
 985                 struct kmem_cache *s = kmalloc_caches[i];
 986
 987                 if (s) {
 988                         int size = kmalloc_size(i);
 989                         char *n = kasprintf(GFP_NOWAIT,
 990                                  "dma-kmalloc-%d", size);
 991
 992                         BUG_ON(!n);
 993                         kmalloc_dma_caches[i] = create_kmalloc_cache(n,
 994                                 size, SLAB_CACHE_DMA | flags);
 995                 }
 996         }
 997 #endif
 998 }
 999 #endif /* !CONFIG_SLOB */
1000
1001 /*
1002  * To avoid unnecessary overhead, we pass through large allocation requests
1003  * directly to the page allocator. We use __GFP_COMP, because we will need to
1004  * know the allocation order to free the pages properly in kfree.
1005  */
1006 void *kmalloc_order(size_t size, gfp_t flags, unsigned int order)
1007 {
1008         void *ret;
1009         struct page *page;
1010
1011         flags |= __GFP_COMP;
1012         page = alloc_kmem_pages(flags, order);
1013         ret = page ? page_address(page) : NULL;
1014         kmemleak_alloc(ret, size, 1, flags);
1015         kasan_kmalloc_large(ret, size, flags);
1016         return ret;
1017 }
1018 EXPORT_SYMBOL(kmalloc_order);
1019
1020 #ifdef CONFIG_TRACING
1021 void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
1022 {
1023         void *ret = kmalloc_order(size, flags, order);
1024         trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
1025         return ret;
1026 }
1027 EXPORT_SYMBOL(kmalloc_order_trace);
1028 #endif
1029
1030 #ifdef CONFIG_SLABINFO
1031
1032 #ifdef CONFIG_SLAB
1033 #define SLABINFO_RIGHTS (S_IWUSR | S_IRUSR)
1034 #else
1035 #define SLABINFO_RIGHTS S_IRUSR
1036 #endif
1037
1038 static void print_slabinfo_header(struct seq_file *m)
1039 {
1040         /*
1041          * Output format version, so at least we can change it
1042          * without _too_ many complaints.
1043          */
1044 #ifdef CONFIG_DEBUG_SLAB
1045         seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
1046 #else
1047         seq_puts(m, "slabinfo - version: 2.1\n");
1048 #endif
1049         seq_puts(m, "# name            <active_objs> <num_objs> <objsize> <objperslab> <pagesperslab>");
1050         seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
1051         seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
1052 #ifdef CONFIG_DEBUG_SLAB
1053         seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> <error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
1054         seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
1055 #endif
1056         seq_putc(m, '\n');
1057 }
1058
1059 void *slab_start(struct seq_file *m, loff_t *pos)
1060 {
1061         mutex_lock(&slab_mutex);
1062         return seq_list_start(&slab_caches, *pos);
1063 }
1064
1065 void *slab_next(struct seq_file *m, void *p, loff_t *pos)
1066 {
1067         return seq_list_next(p, &slab_caches, pos);
1068 }
1069
1070 void slab_stop(struct seq_file *m, void *p)
1071 {
1072         mutex_unlock(&slab_mutex);
1073 }
1074
1075 static void
1076 memcg_accumulate_slabinfo(struct kmem_cache *s, struct slabinfo *info)
1077 {
1078         struct kmem_cache *c;
1079         struct slabinfo sinfo;
1080
1081         if (!is_root_cache(s))
1082                 return;
1083
1084         for_each_memcg_cache(c, s) {
1085                 memset(&sinfo, 0, sizeof(sinfo));
1086                 get_slabinfo(c, &sinfo);
1087
1088                 info->active_slabs += sinfo.active_slabs;
1089                 info->num_slabs += sinfo.num_slabs;
1090                 info->shared_avail += sinfo.shared_avail;
1091                 info->active_objs += sinfo.active_objs;
1092                 info->num_objs += sinfo.num_objs;
1093         }
1094 }
1095
1096 static void cache_show(struct kmem_cache *s, struct seq_file *m)
1097 {
1098         struct slabinfo sinfo;
1099
1100         memset(&sinfo, 0, sizeof(sinfo));
1101         get_slabinfo(s, &sinfo);
1102
1103         memcg_accumulate_slabinfo(s, &sinfo);
1104
1105         seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
1106                    cache_name(s), sinfo.active_objs, sinfo.num_objs, s->size,
1107                    sinfo.objects_per_slab, (1 << sinfo.cache_order));
1108
1109         seq_printf(m, " : tunables %4u %4u %4u",
1110                    sinfo.limit, sinfo.batchcount, sinfo.shared);
1111         seq_printf(m, " : slabdata %6lu %6lu %6lu",
1112                    sinfo.active_slabs, sinfo.num_slabs, sinfo.shared_avail);
1113         slabinfo_show_stats(m, s);
1114         seq_putc(m, '\n');
1115 }
1116
1117 static int slab_show(struct seq_file *m, void *p)
1118 {
1119         struct kmem_cache *s = list_entry(p, struct kmem_cache, list);
1120
1121         if (p == slab_caches.next)
1122                 print_slabinfo_header(m);
1123         if (is_root_cache(s))
1124                 cache_show(s, m);
1125         return 0;
1126 }
1127
1128 #ifdef CONFIG_MEMCG_KMEM
1129 int memcg_slab_show(struct seq_file *m, void *p)
1130 {
1131         struct kmem_cache *s = list_entry(p, struct kmem_cache, list);
1132         struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
1133
1134         if (p == slab_caches.next)
1135                 print_slabinfo_header(m);
1136         if (!is_root_cache(s) && s->memcg_params.memcg == memcg)
1137                 cache_show(s, m);
1138         return 0;
1139 }
1140 #endif
1141
1142 /*
1143  * slabinfo_op - iterator that generates /proc/slabinfo
1144  *
1145  * Output layout:
1146  * cache-name
1147  * num-active-objs
1148  * total-objs
1149  * object size
1150  * num-active-slabs
1151  * total-slabs
1152  * num-pages-per-slab
1153  * + further values on SMP and with statistics enabled
1154  */
1155 static const struct seq_operations slabinfo_op = {
1156         .start = slab_start,
1157         .next = slab_next,
1158         .stop = slab_stop,
1159         .show = slab_show,
1160 };
1161
1162 static int slabinfo_open(struct inode *inode, struct file *file)
1163 {
1164         return seq_open(file, &slabinfo_op);
1165 }
1166
1167 static const struct file_operations proc_slabinfo_operations = {
1168         .open           = slabinfo_open,
1169         .read           = seq_read,
1170         .write          = slabinfo_write,
1171         .llseek         = seq_lseek,
1172         .release        = seq_release,
1173 };
1174
1175 static int __init slab_proc_init(void)
1176 {
1177         proc_create("slabinfo", SLABINFO_RIGHTS, NULL,
1178                                                 &proc_slabinfo_operations);
1179         return 0;
1180 }
1181 module_init(slab_proc_init);
1182 #endif /* CONFIG_SLABINFO */
1183
1184 static __always_inline void *__do_krealloc(const void *p, size_t new_size,
1185                                            gfp_t flags)
1186 {
1187         void *ret;
1188         size_t ks = 0;
1189
1190         if (p)
1191                 ks = ksize(p);
1192
1193         if (ks >= new_size) {
1194                 kasan_krealloc((void *)p, new_size, flags);
1195                 return (void *)p;
1196         }
1197
1198         ret = kmalloc_track_caller(new_size, flags);
1199         if (ret && p)
1200                 memcpy(ret, p, ks);
1201
1202         return ret;
1203 }
1204
1205 /**
1206  * __krealloc - like krealloc() but don't free @p.
1207  * @p: object to reallocate memory for.
1208  * @new_size: how many bytes of memory are required.
1209  * @flags: the type of memory to allocate.
1210  *
1211  * This function is like krealloc() except it never frees the originally
1212  * allocated buffer. Use this if you don't want to free the buffer immediately
1213  * like, for example, with RCU.
1214  */
1215 void *__krealloc(const void *p, size_t new_size, gfp_t flags)
1216 {
1217         if (unlikely(!new_size))
1218                 return ZERO_SIZE_PTR;
1219
1220         return __do_krealloc(p, new_size, flags);
1221
1222 }
1223 EXPORT_SYMBOL(__krealloc);
1224
1225 /**
1226  * krealloc - reallocate memory. The contents will remain unchanged.
1227  * @p: object to reallocate memory for.
1228  * @new_size: how many bytes of memory are required.
1229  * @flags: the type of memory to allocate.
1230  *
1231  * The contents of the object pointed to are preserved up to the
1232  * lesser of the new and old sizes.  If @p is %NULL, krealloc()
1233  * behaves exactly like kmalloc().  If @new_size is 0 and @p is not a
1234  * %NULL pointer, the object pointed to is freed.
1235  */
1236 void *krealloc(const void *p, size_t new_size, gfp_t flags)
1237 {
1238         void *ret;
1239
1240         if (unlikely(!new_size)) {
1241                 kfree(p);
1242                 return ZERO_SIZE_PTR;
1243         }
1244
1245         ret = __do_krealloc(p, new_size, flags);
1246         if (ret && p != ret)
1247                 kfree(p);
1248
1249         return ret;
1250 }
1251 EXPORT_SYMBOL(krealloc);
1252
1253 /**
1254  * kzfree - like kfree but zero memory
1255  * @p: object to free memory of
1256  *
1257  * The memory of the object @p points to is zeroed before freed.
1258  * If @p is %NULL, kzfree() does nothing.
1259  *
1260  * Note: this function zeroes the whole allocated buffer which can be a good
1261  * deal bigger than the requested buffer size passed to kmalloc(). So be
1262  * careful when using this function in performance sensitive code.
1263  */
1264 void kzfree(const void *p)
1265 {
1266         size_t ks;
1267         void *mem = (void *)p;
1268
1269         if (unlikely(ZERO_OR_NULL_PTR(mem)))
1270                 return;
1271         ks = ksize(mem);
1272         memset(mem, 0, ks);
1273         kfree(mem);
1274 }
1275 EXPORT_SYMBOL(kzfree);
1276
1277 /* Tracepoints definitions. */
1278 EXPORT_TRACEPOINT_SYMBOL(kmalloc);
1279 EXPORT_TRACEPOINT_SYMBOL(kmem_cache_alloc);
1280 EXPORT_TRACEPOINT_SYMBOL(kmalloc_node);
1281 EXPORT_TRACEPOINT_SYMBOL(kmem_cache_alloc_node);
1282 EXPORT_TRACEPOINT_SYMBOL(kfree);
1283 EXPORT_TRACEPOINT_SYMBOL(kmem_cache_free);