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 {
 169         struct kmem_cache *s = NULL;
 170         int err = 0;
 171
 172         get_online_cpus();
 173         mutex_lock(&slab_mutex);
 174
 175         if (!kmem_cache_sanity_check(memcg, name, size) == 0)
 176                 goto out_locked;
 177
 178         /*
 179          * Some allocators will constraint the set of valid flags to a subset
 180          * of all flags. We expect them to define CACHE_CREATE_MASK in this
 181          * case, and we'll just provide them with a sanitized version of the
 182          * passed flags.
 183          */
 184         flags &= CACHE_CREATE_MASK;
 185
 186         s = __kmem_cache_alias(memcg, name, size, align, flags, ctor);
 187         if (s)
 188                 goto out_locked;
 189
 190         s = kmem_cache_zalloc(kmem_cache, GFP_KERNEL);
 191         if (s) {
 192                 s->object_size = s->size = size;
 193                 s->align = calculate_alignment(flags, align, size);
 194                 s->ctor = ctor;
 195
 196                 if (memcg_register_cache(memcg, s)) {
 197                         kmem_cache_free(kmem_cache, s);
 198                         err = -ENOMEM;
 199                         goto out_locked;
 200                 }
 201
 202                 s->name = kstrdup(name, GFP_KERNEL);
 203                 if (!s->name) {
 204                         kmem_cache_free(kmem_cache, s);
 205                         err = -ENOMEM;
 206                         goto out_locked;
 207                 }
 208
 209                 err = __kmem_cache_create(s, flags);
 210                 if (!err) {
 211                         s->refcount = 1;
 212                         list_add(&s->list, &slab_caches);
 213                         memcg_cache_list_add(memcg, s);
 214                 } else {
 215                         kfree(s->name);
 216                         kmem_cache_free(kmem_cache, s);
 217                 }
 218         } else
 219                 err = -ENOMEM;
 220
 221 out_locked:
 222         mutex_unlock(&slab_mutex);
 223         put_online_cpus();
 224
 225         if (err) {
 226
 227                 if (flags & SLAB_PANIC)
 228                         panic("kmem_cache_create: Failed to create slab '%s'. Error %d\n",
 229                                 name, err);
 230                 else {
 231                         printk(KERN_WARNING "kmem_cache_create(%s) failed with error %d",
 232                                 name, err);
 233                         dump_stack();
 234                 }
 235
 236                 return NULL;
 237         }
 238
 239         return s;
 240 }
 241
 242 struct kmem_cache *
 243 kmem_cache_create(const char *name, size_t size, size_t align,
 244                   unsigned long flags, void (*ctor)(void *))
 245 {
 246         return kmem_cache_create_memcg(NULL, name, size, align, flags, ctor);
 247 }
 248 EXPORT_SYMBOL(kmem_cache_create);
 249
 250 void kmem_cache_destroy(struct kmem_cache *s)
 251 {
 252         /* Destroy all the children caches if we aren't a memcg cache */
 253         kmem_cache_destroy_memcg_children(s);
 254
 255         get_online_cpus();
 256         mutex_lock(&slab_mutex);
 257         s->refcount--;
 258         if (!s->refcount) {
 259                 list_del(&s->list);
 260
 261                 if (!__kmem_cache_shutdown(s)) {
 262                         mutex_unlock(&slab_mutex);
 263                         if (s->flags & SLAB_DESTROY_BY_RCU)
 264                                 rcu_barrier();
 265
 266                         memcg_release_cache(s);
 267                         kfree(s->name);
 268                         kmem_cache_free(kmem_cache, s);
 269                 } else {
 270                         list_add(&s->list, &slab_caches);
 271                         mutex_unlock(&slab_mutex);
 272                         printk(KERN_ERR "kmem_cache_destroy %s: Slab cache still has objects\n",
 273                                 s->name);
 274                         dump_stack();
 275                 }
 276         } else {
 277                 mutex_unlock(&slab_mutex);
 278         }
 279         put_online_cpus();
 280 }
 281 EXPORT_SYMBOL(kmem_cache_destroy);
 282
 283 int slab_is_available(void)
 284 {
 285         return slab_state >= UP;
 286 }
 287
 288 #ifndef CONFIG_SLOB
 289 /* Create a cache during boot when no slab services are available yet */
 290 void __init create_boot_cache(struct kmem_cache *s, const char *name, size_t size,
 291                 unsigned long flags)
 292 {
 293         int err;
 294
 295         s->name = name;
 296         s->size = s->object_size = size;
 297         s->align = calculate_alignment(flags, ARCH_KMALLOC_MINALIGN, size);
 298         err = __kmem_cache_create(s, flags);
 299
 300         if (err)
 301                 panic("Creation of kmalloc slab %s size=%zd failed. Reason %d\n",
 302                                         name, size, err);
 303
 304         s->refcount = -1;       /* Exempt from merging for now */
 305 }
 306
 307 struct kmem_cache *__init create_kmalloc_cache(const char *name, size_t size,
 308                                 unsigned long flags)
 309 {
 310         struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
 311
 312         if (!s)
 313                 panic("Out of memory when creating slab %s\n", name);
 314
 315         create_boot_cache(s, name, size, flags);
 316         list_add(&s->list, &slab_caches);
 317         s->refcount = 1;
 318         return s;
 319 }
 320
 321 #endif /* !CONFIG_SLOB */
 322
 323
 324 #ifdef CONFIG_SLABINFO
 325 void print_slabinfo_header(struct seq_file *m)
 326 {
 327         /*
 328          * Output format version, so at least we can change it
 329          * without _too_ many complaints.
 330          */
 331 #ifdef CONFIG_DEBUG_SLAB
 332         seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
 333 #else
 334         seq_puts(m, "slabinfo - version: 2.1\n");
 335 #endif
 336         seq_puts(m, "# name            <active_objs> <num_objs> <objsize> "
 337                  "<objperslab> <pagesperslab>");
 338         seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
 339         seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
 340 #ifdef CONFIG_DEBUG_SLAB
 341         seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
 342                  "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
 343         seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
 344 #endif
 345         seq_putc(m, '\n');
 346 }
 347
 348 static void *s_start(struct seq_file *m, loff_t *pos)
 349 {
 350         loff_t n = *pos;
 351
 352         mutex_lock(&slab_mutex);
 353         if (!n)
 354                 print_slabinfo_header(m);
 355
 356         return seq_list_start(&slab_caches, *pos);
 357 }
 358
 359 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
 360 {
 361         return seq_list_next(p, &slab_caches, pos);
 362 }
 363
 364 static void s_stop(struct seq_file *m, void *p)
 365 {
 366         mutex_unlock(&slab_mutex);
 367 }
 368
 369 static void
 370 memcg_accumulate_slabinfo(struct kmem_cache *s, struct slabinfo *info)
 371 {
 372         struct kmem_cache *c;
 373         struct slabinfo sinfo;
 374         int i;
 375
 376         if (!is_root_cache(s))
 377                 return;
 378
 379         for_each_memcg_cache_index(i) {
 380                 c = cache_from_memcg(s, i);
 381                 if (!c)
 382                         continue;
 383
 384                 memset(&sinfo, 0, sizeof(sinfo));
 385                 get_slabinfo(c, &sinfo);
 386
 387                 info->active_slabs += sinfo.active_slabs;
 388                 info->num_slabs += sinfo.num_slabs;
 389                 info->shared_avail += sinfo.shared_avail;
 390                 info->active_objs += sinfo.active_objs;
 391                 info->num_objs += sinfo.num_objs;
 392         }
 393 }
 394
 395 int cache_show(struct kmem_cache *s, struct seq_file *m)
 396 {
 397         struct slabinfo sinfo;
 398
 399         memset(&sinfo, 0, sizeof(sinfo));
 400         get_slabinfo(s, &sinfo);
 401
 402         memcg_accumulate_slabinfo(s, &sinfo);
 403
 404         seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
 405                    cache_name(s), sinfo.active_objs, sinfo.num_objs, s->size,
 406                    sinfo.objects_per_slab, (1 << sinfo.cache_order));
 407
 408         seq_printf(m, " : tunables %4u %4u %4u",
 409                    sinfo.limit, sinfo.batchcount, sinfo.shared);
 410         seq_printf(m, " : slabdata %6lu %6lu %6lu",
 411                    sinfo.active_slabs, sinfo.num_slabs, sinfo.shared_avail);
 412         slabinfo_show_stats(m, s);
 413         seq_putc(m, '\n');
 414         return 0;
 415 }
 416
 417 static int s_show(struct seq_file *m, void *p)
 418 {
 419         struct kmem_cache *s = list_entry(p, struct kmem_cache, list);
 420
 421         if (!is_root_cache(s))
 422                 return 0;
 423         return cache_show(s, m);
 424 }
 425
 426 /*
 427  * slabinfo_op - iterator that generates /proc/slabinfo
 428  *
 429  * Output layout:
 430  * cache-name
 431  * num-active-objs
 432  * total-objs
 433  * object size
 434  * num-active-slabs
 435  * total-slabs
 436  * num-pages-per-slab
 437  * + further values on SMP and with statistics enabled
 438  */
 439 static const struct seq_operations slabinfo_op = {
 440         .start = s_start,
 441         .next = s_next,
 442         .stop = s_stop,
 443         .show = s_show,
 444 };
 445
 446 static int slabinfo_open(struct inode *inode, struct file *file)
 447 {
 448         return seq_open(file, &slabinfo_op);
 449 }
 450
 451 static const struct file_operations proc_slabinfo_operations = {
 452         .open           = slabinfo_open,
 453         .read           = seq_read,
 454         .write          = slabinfo_write,
 455         .llseek         = seq_lseek,
 456         .release        = seq_release,
 457 };
 458
 459 static int __init slab_proc_init(void)
 460 {
 461         proc_create("slabinfo", S_IRUSR, NULL, &proc_slabinfo_operations);
 462         return 0;
 463 }
 464 module_init(slab_proc_init);
 465 #endif /* CONFIG_SLABINFO */