~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/mm/slab.c

Version: ~ [ linux-5.13-rc1 ] ~ [ linux-5.12.2 ] ~ [ linux-5.11.19 ] ~ [ linux-5.10.35 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.117 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.190 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.232 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.268 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.268 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.18.140 ] ~ [ linux-3.16.85 ] ~ [ linux-3.14.79 ] ~ [ linux-3.12.74 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  1 /*
  2  * linux/mm/slab.c
  3  * Written by Mark Hemment, 1996/97.
  4  * (markhe@nextd.demon.co.uk)
  5  *
  6  * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
  7  *
  8  * Major cleanup, different bufctl logic, per-cpu arrays
  9  *      (c) 2000 Manfred Spraul
 10  *
 11  * Cleanup, make the head arrays unconditional, preparation for NUMA
 12  *      (c) 2002 Manfred Spraul
 13  *
 14  * An implementation of the Slab Allocator as described in outline in;
 15  *      UNIX Internals: The New Frontiers by Uresh Vahalia
 16  *      Pub: Prentice Hall      ISBN 0-13-101908-2
 17  * or with a little more detail in;
 18  *      The Slab Allocator: An Object-Caching Kernel Memory Allocator
 19  *      Jeff Bonwick (Sun Microsystems).
 20  *      Presented at: USENIX Summer 1994 Technical Conference
 21  *
 22  * The memory is organized in caches, one cache for each object type.
 23  * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
 24  * Each cache consists out of many slabs (they are small (usually one
 25  * page long) and always contiguous), and each slab contains multiple
 26  * initialized objects.
 27  *
 28  * This means, that your constructor is used only for newly allocated
 29  * slabs and you must pass objects with the same initializations to
 30  * kmem_cache_free.
 31  *
 32  * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
 33  * normal). If you need a special memory type, then must create a new
 34  * cache for that memory type.
 35  *
 36  * In order to reduce fragmentation, the slabs are sorted in 3 groups:
 37  *   full slabs with 0 free objects
 38  *   partial slabs
 39  *   empty slabs with no allocated objects
 40  *
 41  * If partial slabs exist, then new allocations come from these slabs,
 42  * otherwise from empty slabs or new slabs are allocated.
 43  *
 44  * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
 45  * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
 46  *
 47  * Each cache has a short per-cpu head array, most allocs
 48  * and frees go into that array, and if that array overflows, then 1/2
 49  * of the entries in the array are given back into the global cache.
 50  * The head array is strictly LIFO and should improve the cache hit rates.
 51  * On SMP, it additionally reduces the spinlock operations.
 52  *
 53  * The c_cpuarray may not be read with enabled local interrupts -
 54  * it's changed with a smp_call_function().
 55  *
 56  * SMP synchronization:
 57  *  constructors and destructors are called without any locking.
 58  *  Several members in struct kmem_cache and struct slab never change, they
 59  *      are accessed without any locking.
 60  *  The per-cpu arrays are never accessed from the wrong cpu, no locking,
 61  *      and local interrupts are disabled so slab code is preempt-safe.
 62  *  The non-constant members are protected with a per-cache irq spinlock.
 63  *
 64  * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
 65  * in 2000 - many ideas in the current implementation are derived from
 66  * his patch.
 67  *
 68  * Further notes from the original documentation:
 69  *
 70  * 11 April '97.  Started multi-threading - markhe
 71  *      The global cache-chain is protected by the mutex 'slab_mutex'.
 72  *      The sem is only needed when accessing/extending the cache-chain, which
 73  *      can never happen inside an interrupt (kmem_cache_create(),
 74  *      kmem_cache_shrink() and kmem_cache_reap()).
 75  *
 76  *      At present, each engine can be growing a cache.  This should be blocked.
 77  *
 78  * 15 March 2005. NUMA slab allocator.
 79  *      Shai Fultheim <shai@scalex86.org>.
 80  *      Shobhit Dayal <shobhit@calsoftinc.com>
 81  *      Alok N Kataria <alokk@calsoftinc.com>
 82  *      Christoph Lameter <christoph@lameter.com>
 83  *
 84  *      Modified the slab allocator to be node aware on NUMA systems.
 85  *      Each node has its own list of partial, free and full slabs.
 86  *      All object allocations for a node occur from node specific slab lists.
 87  */
 88 
 89 #include        <linux/slab.h>
 90 #include        <linux/mm.h>
 91 #include        <linux/poison.h>
 92 #include        <linux/swap.h>
 93 #include        <linux/cache.h>
 94 #include        <linux/interrupt.h>
 95 #include        <linux/init.h>
 96 #include        <linux/compiler.h>
 97 #include        <linux/cpuset.h>
 98 #include        <linux/proc_fs.h>
 99 #include        <linux/seq_file.h>
100 #include        <linux/notifier.h>
101 #include        <linux/kallsyms.h>
102 #include        <linux/cpu.h>
103 #include        <linux/sysctl.h>
104 #include        <linux/module.h>
105 #include        <linux/rcupdate.h>
106 #include        <linux/string.h>
107 #include        <linux/uaccess.h>
108 #include        <linux/nodemask.h>
109 #include        <linux/kmemleak.h>
110 #include        <linux/mempolicy.h>
111 #include        <linux/mutex.h>
112 #include        <linux/fault-inject.h>
113 #include        <linux/rtmutex.h>
114 #include        <linux/reciprocal_div.h>
115 #include        <linux/debugobjects.h>
116 #include        <linux/kmemcheck.h>
117 #include        <linux/memory.h>
118 #include        <linux/prefetch.h>
119 #include        <linux/sched/task_stack.h>
120 
121 #include        <net/sock.h>
122 
123 #include        <asm/cacheflush.h>
124 #include        <asm/tlbflush.h>
125 #include        <asm/page.h>
126 
127 #include <trace/events/kmem.h>
128 
129 #include        "internal.h"
130 
131 #include        "slab.h"
132 
133 /*
134  * DEBUG        - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
135  *                0 for faster, smaller code (especially in the critical paths).
136  *
137  * STATS        - 1 to collect stats for /proc/slabinfo.
138  *                0 for faster, smaller code (especially in the critical paths).
139  *
140  * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
141  */
142 
143 #ifdef CONFIG_DEBUG_SLAB
144 #define DEBUG           1
145 #define STATS           1
146 #define FORCED_DEBUG    1
147 #else
148 #define DEBUG           0
149 #define STATS           0
150 #define FORCED_DEBUG    0
151 #endif
152 
153 /* Shouldn't this be in a header file somewhere? */
154 #define BYTES_PER_WORD          sizeof(void *)
155 #define REDZONE_ALIGN           max(BYTES_PER_WORD, __alignof__(unsigned long long))
156 
157 #ifndef ARCH_KMALLOC_FLAGS
158 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
159 #endif
160 
161 #define FREELIST_BYTE_INDEX (((PAGE_SIZE >> BITS_PER_BYTE) \
162                                 <= SLAB_OBJ_MIN_SIZE) ? 1 : 0)
163 
164 #if FREELIST_BYTE_INDEX
165 typedef unsigned char freelist_idx_t;
166 #else
167 typedef unsigned short freelist_idx_t;
168 #endif
169 
170 #define SLAB_OBJ_MAX_NUM ((1 << sizeof(freelist_idx_t) * BITS_PER_BYTE) - 1)
171 
172 /*
173  * struct array_cache
174  *
175  * Purpose:
176  * - LIFO ordering, to hand out cache-warm objects from _alloc
177  * - reduce the number of linked list operations
178  * - reduce spinlock operations
179  *
180  * The limit is stored in the per-cpu structure to reduce the data cache
181  * footprint.
182  *
183  */
184 struct array_cache {
185         unsigned int avail;
186         unsigned int limit;
187         unsigned int batchcount;
188         unsigned int touched;
189         void *entry[];  /*
190                          * Must have this definition in here for the proper
191                          * alignment of array_cache. Also simplifies accessing
192                          * the entries.
193                          */
194 };
195 
196 struct alien_cache {
197         spinlock_t lock;
198         struct array_cache ac;
199 };
200 
201 /*
202  * Need this for bootstrapping a per node allocator.
203  */
204 #define NUM_INIT_LISTS (2 * MAX_NUMNODES)
205 static struct kmem_cache_node __initdata init_kmem_cache_node[NUM_INIT_LISTS];
206 #define CACHE_CACHE 0
207 #define SIZE_NODE (MAX_NUMNODES)
208 
209 static int drain_freelist(struct kmem_cache *cache,
210                         struct kmem_cache_node *n, int tofree);
211 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
212                         int node, struct list_head *list);
213 static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list);
214 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
215 static void cache_reap(struct work_struct *unused);
216 
217 static inline void fixup_objfreelist_debug(struct kmem_cache *cachep,
218                                                 void **list);
219 static inline void fixup_slab_list(struct kmem_cache *cachep,
220                                 struct kmem_cache_node *n, struct page *page,
221                                 void **list);
222 static int slab_early_init = 1;
223 
224 #define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
225 
226 static void kmem_cache_node_init(struct kmem_cache_node *parent)
227 {
228         INIT_LIST_HEAD(&parent->slabs_full);
229         INIT_LIST_HEAD(&parent->slabs_partial);
230         INIT_LIST_HEAD(&parent->slabs_free);
231         parent->total_slabs = 0;
232         parent->free_slabs = 0;
233         parent->shared = NULL;
234         parent->alien = NULL;
235         parent->colour_next = 0;
236         spin_lock_init(&parent->list_lock);
237         parent->free_objects = 0;
238         parent->free_touched = 0;
239 }
240 
241 #define MAKE_LIST(cachep, listp, slab, nodeid)                          \
242         do {                                                            \
243                 INIT_LIST_HEAD(listp);                                  \
244                 list_splice(&get_node(cachep, nodeid)->slab, listp);    \
245         } while (0)
246 
247 #define MAKE_ALL_LISTS(cachep, ptr, nodeid)                             \
248         do {                                                            \
249         MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid);  \
250         MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
251         MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid);  \
252         } while (0)
253 
254 #define CFLGS_OBJFREELIST_SLAB  (0x40000000UL)
255 #define CFLGS_OFF_SLAB          (0x80000000UL)
256 #define OBJFREELIST_SLAB(x)     ((x)->flags & CFLGS_OBJFREELIST_SLAB)
257 #define OFF_SLAB(x)     ((x)->flags & CFLGS_OFF_SLAB)
258 
259 #define BATCHREFILL_LIMIT       16
260 /*
261  * Optimization question: fewer reaps means less probability for unnessary
262  * cpucache drain/refill cycles.
263  *
264  * OTOH the cpuarrays can contain lots of objects,
265  * which could lock up otherwise freeable slabs.
266  */
267 #define REAPTIMEOUT_AC          (2*HZ)
268 #define REAPTIMEOUT_NODE        (4*HZ)
269 
270 #if STATS
271 #define STATS_INC_ACTIVE(x)     ((x)->num_active++)
272 #define STATS_DEC_ACTIVE(x)     ((x)->num_active--)
273 #define STATS_INC_ALLOCED(x)    ((x)->num_allocations++)
274 #define STATS_INC_GROWN(x)      ((x)->grown++)
275 #define STATS_ADD_REAPED(x,y)   ((x)->reaped += (y))
276 #define STATS_SET_HIGH(x)                                               \
277         do {                                                            \
278                 if ((x)->num_active > (x)->high_mark)                   \
279                         (x)->high_mark = (x)->num_active;               \
280         } while (0)
281 #define STATS_INC_ERR(x)        ((x)->errors++)
282 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
283 #define STATS_INC_NODEFREES(x)  ((x)->node_frees++)
284 #define STATS_INC_ACOVERFLOW(x)   ((x)->node_overflow++)
285 #define STATS_SET_FREEABLE(x, i)                                        \
286         do {                                                            \
287                 if ((x)->max_freeable < i)                              \
288                         (x)->max_freeable = i;                          \
289         } while (0)
290 #define STATS_INC_ALLOCHIT(x)   atomic_inc(&(x)->allochit)
291 #define STATS_INC_ALLOCMISS(x)  atomic_inc(&(x)->allocmiss)
292 #define STATS_INC_FREEHIT(x)    atomic_inc(&(x)->freehit)
293 #define STATS_INC_FREEMISS(x)   atomic_inc(&(x)->freemiss)
294 #else
295 #define STATS_INC_ACTIVE(x)     do { } while (0)
296 #define STATS_DEC_ACTIVE(x)     do { } while (0)
297 #define STATS_INC_ALLOCED(x)    do { } while (0)
298 #define STATS_INC_GROWN(x)      do { } while (0)
299 #define STATS_ADD_REAPED(x,y)   do { (void)(y); } while (0)
300 #define STATS_SET_HIGH(x)       do { } while (0)
301 #define STATS_INC_ERR(x)        do { } while (0)
302 #define STATS_INC_NODEALLOCS(x) do { } while (0)
303 #define STATS_INC_NODEFREES(x)  do { } while (0)
304 #define STATS_INC_ACOVERFLOW(x)   do { } while (0)
305 #define STATS_SET_FREEABLE(x, i) do { } while (0)
306 #define STATS_INC_ALLOCHIT(x)   do { } while (0)
307 #define STATS_INC_ALLOCMISS(x)  do { } while (0)
308 #define STATS_INC_FREEHIT(x)    do { } while (0)
309 #define STATS_INC_FREEMISS(x)   do { } while (0)
310 #endif
311 
312 #if DEBUG
313 
314 /*
315  * memory layout of objects:
316  * 0            : objp
317  * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
318  *              the end of an object is aligned with the end of the real
319  *              allocation. Catches writes behind the end of the allocation.
320  * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
321  *              redzone word.
322  * cachep->obj_offset: The real object.
323  * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
324  * cachep->size - 1* BYTES_PER_WORD: last caller address
325  *                                      [BYTES_PER_WORD long]
326  */
327 static int obj_offset(struct kmem_cache *cachep)
328 {
329         return cachep->obj_offset;
330 }
331 
332 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
333 {
334         BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
335         return (unsigned long long*) (objp + obj_offset(cachep) -
336                                       sizeof(unsigned long long));
337 }
338 
339 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
340 {
341         BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
342         if (cachep->flags & SLAB_STORE_USER)
343                 return (unsigned long long *)(objp + cachep->size -
344                                               sizeof(unsigned long long) -
345                                               REDZONE_ALIGN);
346         return (unsigned long long *) (objp + cachep->size -
347                                        sizeof(unsigned long long));
348 }
349 
350 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
351 {
352         BUG_ON(!(cachep->flags & SLAB_STORE_USER));
353         return (void **)(objp + cachep->size - BYTES_PER_WORD);
354 }
355 
356 #else
357 
358 #define obj_offset(x)                   0
359 #define dbg_redzone1(cachep, objp)      ({BUG(); (unsigned long long *)NULL;})
360 #define dbg_redzone2(cachep, objp)      ({BUG(); (unsigned long long *)NULL;})
361 #define dbg_userword(cachep, objp)      ({BUG(); (void **)NULL;})
362 
363 #endif
364 
365 #ifdef CONFIG_DEBUG_SLAB_LEAK
366 
367 static inline bool is_store_user_clean(struct kmem_cache *cachep)
368 {
369         return atomic_read(&cachep->store_user_clean) == 1;
370 }
371 
372 static inline void set_store_user_clean(struct kmem_cache *cachep)
373 {
374         atomic_set(&cachep->store_user_clean, 1);
375 }
376 
377 static inline void set_store_user_dirty(struct kmem_cache *cachep)
378 {
379         if (is_store_user_clean(cachep))
380                 atomic_set(&cachep->store_user_clean, 0);
381 }
382 
383 #else
384 static inline void set_store_user_dirty(struct kmem_cache *cachep) {}
385 
386 #endif
387 
388 /*
389  * Do not go above this order unless 0 objects fit into the slab or
390  * overridden on the command line.
391  */
392 #define SLAB_MAX_ORDER_HI       1
393 #define SLAB_MAX_ORDER_LO       0
394 static int slab_max_order = SLAB_MAX_ORDER_LO;
395 static bool slab_max_order_set __initdata;
396 
397 static inline struct kmem_cache *virt_to_cache(const void *obj)
398 {
399         struct page *page = virt_to_head_page(obj);
400         return page->slab_cache;
401 }
402 
403 static inline void *index_to_obj(struct kmem_cache *cache, struct page *page,
404                                  unsigned int idx)
405 {
406         return page->s_mem + cache->size * idx;
407 }
408 
409 /*
410  * We want to avoid an expensive divide : (offset / cache->size)
411  *   Using the fact that size is a constant for a particular cache,
412  *   we can replace (offset / cache->size) by
413  *   reciprocal_divide(offset, cache->reciprocal_buffer_size)
414  */
415 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
416                                         const struct page *page, void *obj)
417 {
418         u32 offset = (obj - page->s_mem);
419         return reciprocal_divide(offset, cache->reciprocal_buffer_size);
420 }
421 
422 #define BOOT_CPUCACHE_ENTRIES   1
423 /* internal cache of cache description objs */
424 static struct kmem_cache kmem_cache_boot = {
425         .batchcount = 1,
426         .limit = BOOT_CPUCACHE_ENTRIES,
427         .shared = 1,
428         .size = sizeof(struct kmem_cache),
429         .name = "kmem_cache",
430 };
431 
432 static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
433 
434 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
435 {
436         return this_cpu_ptr(cachep->cpu_cache);
437 }
438 
439 /*
440  * Calculate the number of objects and left-over bytes for a given buffer size.
441  */
442 static unsigned int cache_estimate(unsigned long gfporder, size_t buffer_size,
443                 unsigned long flags, size_t *left_over)
444 {
445         unsigned int num;
446         size_t slab_size = PAGE_SIZE << gfporder;
447 
448         /*
449          * The slab management structure can be either off the slab or
450          * on it. For the latter case, the memory allocated for a
451          * slab is used for:
452          *
453          * - @buffer_size bytes for each object
454          * - One freelist_idx_t for each object
455          *
456          * We don't need to consider alignment of freelist because
457          * freelist will be at the end of slab page. The objects will be
458          * at the correct alignment.
459          *
460          * If the slab management structure is off the slab, then the
461          * alignment will already be calculated into the size. Because
462          * the slabs are all pages aligned, the objects will be at the
463          * correct alignment when allocated.
464          */
465         if (flags & (CFLGS_OBJFREELIST_SLAB | CFLGS_OFF_SLAB)) {
466                 num = slab_size / buffer_size;
467                 *left_over = slab_size % buffer_size;
468         } else {
469                 num = slab_size / (buffer_size + sizeof(freelist_idx_t));
470                 *left_over = slab_size %
471                         (buffer_size + sizeof(freelist_idx_t));
472         }
473 
474         return num;
475 }
476 
477 #if DEBUG
478 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
479 
480 static void __slab_error(const char *function, struct kmem_cache *cachep,
481                         char *msg)
482 {
483         pr_err("slab error in %s(): cache `%s': %s\n",
484                function, cachep->name, msg);
485         dump_stack();
486         add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
487 }
488 #endif
489 
490 /*
491  * By default on NUMA we use alien caches to stage the freeing of
492  * objects allocated from other nodes. This causes massive memory
493  * inefficiencies when using fake NUMA setup to split memory into a
494  * large number of small nodes, so it can be disabled on the command
495  * line
496   */
497 
498 static int use_alien_caches __read_mostly = 1;
499 static int __init noaliencache_setup(char *s)
500 {
501         use_alien_caches = 0;
502         return 1;
503 }
504 __setup("noaliencache", noaliencache_setup);
505 
506 static int __init slab_max_order_setup(char *str)
507 {
508         get_option(&str, &slab_max_order);
509         slab_max_order = slab_max_order < 0 ? 0 :
510                                 min(slab_max_order, MAX_ORDER - 1);
511         slab_max_order_set = true;
512 
513         return 1;
514 }
515 __setup("slab_max_order=", slab_max_order_setup);
516 
517 #ifdef CONFIG_NUMA
518 /*
519  * Special reaping functions for NUMA systems called from cache_reap().
520  * These take care of doing round robin flushing of alien caches (containing
521  * objects freed on different nodes from which they were allocated) and the
522  * flushing of remote pcps by calling drain_node_pages.
523  */
524 static DEFINE_PER_CPU(unsigned long, slab_reap_node);
525 
526 static void init_reap_node(int cpu)
527 {
528         per_cpu(slab_reap_node, cpu) = next_node_in(cpu_to_mem(cpu),
529                                                     node_online_map);
530 }
531 
532 static void next_reap_node(void)
533 {
534         int node = __this_cpu_read(slab_reap_node);
535 
536         node = next_node_in(node, node_online_map);
537         __this_cpu_write(slab_reap_node, node);
538 }
539 
540 #else
541 #define init_reap_node(cpu) do { } while (0)
542 #define next_reap_node(void) do { } while (0)
543 #endif
544 
545 /*
546  * Initiate the reap timer running on the target CPU.  We run at around 1 to 2Hz
547  * via the workqueue/eventd.
548  * Add the CPU number into the expiration time to minimize the possibility of
549  * the CPUs getting into lockstep and contending for the global cache chain
550  * lock.
551  */
552 static void start_cpu_timer(int cpu)
553 {
554         struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
555 
556         if (reap_work->work.func == NULL) {
557                 init_reap_node(cpu);
558                 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
559                 schedule_delayed_work_on(cpu, reap_work,
560                                         __round_jiffies_relative(HZ, cpu));
561         }
562 }
563 
564 static void init_arraycache(struct array_cache *ac, int limit, int batch)
565 {
566         /*
567          * The array_cache structures contain pointers to free object.
568          * However, when such objects are allocated or transferred to another
569          * cache the pointers are not cleared and they could be counted as
570          * valid references during a kmemleak scan. Therefore, kmemleak must
571          * not scan such objects.
572          */
573         kmemleak_no_scan(ac);
574         if (ac) {
575                 ac->avail = 0;
576                 ac->limit = limit;
577                 ac->batchcount = batch;
578                 ac->touched = 0;
579         }
580 }
581 
582 static struct array_cache *alloc_arraycache(int node, int entries,
583                                             int batchcount, gfp_t gfp)
584 {
585         size_t memsize = sizeof(void *) * entries + sizeof(struct array_cache);
586         struct array_cache *ac = NULL;
587 
588         ac = kmalloc_node(memsize, gfp, node);
589         init_arraycache(ac, entries, batchcount);
590         return ac;
591 }
592 
593 static noinline void cache_free_pfmemalloc(struct kmem_cache *cachep,
594                                         struct page *page, void *objp)
595 {
596         struct kmem_cache_node *n;
597         int page_node;
598         LIST_HEAD(list);
599 
600         page_node = page_to_nid(page);
601         n = get_node(cachep, page_node);
602 
603         spin_lock(&n->list_lock);
604         free_block(cachep, &objp, 1, page_node, &list);
605         spin_unlock(&n->list_lock);
606 
607         slabs_destroy(cachep, &list);
608 }
609 
610 /*
611  * Transfer objects in one arraycache to another.
612  * Locking must be handled by the caller.
613  *
614  * Return the number of entries transferred.
615  */
616 static int transfer_objects(struct array_cache *to,
617                 struct array_cache *from, unsigned int max)
618 {
619         /* Figure out how many entries to transfer */
620         int nr = min3(from->avail, max, to->limit - to->avail);
621 
622         if (!nr)
623                 return 0;
624 
625         memcpy(to->entry + to->avail, from->entry + from->avail -nr,
626                         sizeof(void *) *nr);
627 
628         from->avail -= nr;
629         to->avail += nr;
630         return nr;
631 }
632 
633 #ifndef CONFIG_NUMA
634 
635 #define drain_alien_cache(cachep, alien) do { } while (0)
636 #define reap_alien(cachep, n) do { } while (0)
637 
638 static inline struct alien_cache **alloc_alien_cache(int node,
639                                                 int limit, gfp_t gfp)
640 {
641         return NULL;
642 }
643 
644 static inline void free_alien_cache(struct alien_cache **ac_ptr)
645 {
646 }
647 
648 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
649 {
650         return 0;
651 }
652 
653 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
654                 gfp_t flags)
655 {
656         return NULL;
657 }
658 
659 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
660                  gfp_t flags, int nodeid)
661 {
662         return NULL;
663 }
664 
665 static inline gfp_t gfp_exact_node(gfp_t flags)
666 {
667         return flags & ~__GFP_NOFAIL;
668 }
669 
670 #else   /* CONFIG_NUMA */
671 
672 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
673 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
674 
675 static struct alien_cache *__alloc_alien_cache(int node, int entries,
676                                                 int batch, gfp_t gfp)
677 {
678         size_t memsize = sizeof(void *) * entries + sizeof(struct alien_cache);
679         struct alien_cache *alc = NULL;
680 
681         alc = kmalloc_node(memsize, gfp, node);
682         init_arraycache(&alc->ac, entries, batch);
683         spin_lock_init(&alc->lock);
684         return alc;
685 }
686 
687 static struct alien_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
688 {
689         struct alien_cache **alc_ptr;
690         size_t memsize = sizeof(void *) * nr_node_ids;
691         int i;
692 
693         if (limit > 1)
694                 limit = 12;
695         alc_ptr = kzalloc_node(memsize, gfp, node);
696         if (!alc_ptr)
697                 return NULL;
698 
699         for_each_node(i) {
700                 if (i == node || !node_online(i))
701                         continue;
702                 alc_ptr[i] = __alloc_alien_cache(node, limit, 0xbaadf00d, gfp);
703                 if (!alc_ptr[i]) {
704                         for (i--; i >= 0; i--)
705                                 kfree(alc_ptr[i]);
706                         kfree(alc_ptr);
707                         return NULL;
708                 }
709         }
710         return alc_ptr;
711 }
712 
713 static void free_alien_cache(struct alien_cache **alc_ptr)
714 {
715         int i;
716 
717         if (!alc_ptr)
718                 return;
719         for_each_node(i)
720             kfree(alc_ptr[i]);
721         kfree(alc_ptr);
722 }
723 
724 static void __drain_alien_cache(struct kmem_cache *cachep,
725                                 struct array_cache *ac, int node,
726                                 struct list_head *list)
727 {
728         struct kmem_cache_node *n = get_node(cachep, node);
729 
730         if (ac->avail) {
731                 spin_lock(&n->list_lock);
732                 /*
733                  * Stuff objects into the remote nodes shared array first.
734                  * That way we could avoid the overhead of putting the objects
735                  * into the free lists and getting them back later.
736                  */
737                 if (n->shared)
738                         transfer_objects(n->shared, ac, ac->limit);
739 
740                 free_block(cachep, ac->entry, ac->avail, node, list);
741                 ac->avail = 0;
742                 spin_unlock(&n->list_lock);
743         }
744 }
745 
746 /*
747  * Called from cache_reap() to regularly drain alien caches round robin.
748  */
749 static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
750 {
751         int node = __this_cpu_read(slab_reap_node);
752 
753         if (n->alien) {
754                 struct alien_cache *alc = n->alien[node];
755                 struct array_cache *ac;
756 
757                 if (alc) {
758                         ac = &alc->ac;
759                         if (ac->avail && spin_trylock_irq(&alc->lock)) {
760                                 LIST_HEAD(list);
761 
762                                 __drain_alien_cache(cachep, ac, node, &list);
763                                 spin_unlock_irq(&alc->lock);
764                                 slabs_destroy(cachep, &list);
765                         }
766                 }
767         }
768 }
769 
770 static void drain_alien_cache(struct kmem_cache *cachep,
771                                 struct alien_cache **alien)
772 {
773         int i = 0;
774         struct alien_cache *alc;
775         struct array_cache *ac;
776         unsigned long flags;
777 
778         for_each_online_node(i) {
779                 alc = alien[i];
780                 if (alc) {
781                         LIST_HEAD(list);
782 
783                         ac = &alc->ac;
784                         spin_lock_irqsave(&alc->lock, flags);
785                         __drain_alien_cache(cachep, ac, i, &list);
786                         spin_unlock_irqrestore(&alc->lock, flags);
787                         slabs_destroy(cachep, &list);
788                 }
789         }
790 }
791 
792 static int __cache_free_alien(struct kmem_cache *cachep, void *objp,
793                                 int node, int page_node)
794 {
795         struct kmem_cache_node *n;
796         struct alien_cache *alien = NULL;
797         struct array_cache *ac;
798         LIST_HEAD(list);
799 
800         n = get_node(cachep, node);
801         STATS_INC_NODEFREES(cachep);
802         if (n->alien && n->alien[page_node]) {
803                 alien = n->alien[page_node];
804                 ac = &alien->ac;
805                 spin_lock(&alien->lock);
806                 if (unlikely(ac->avail == ac->limit)) {
807                         STATS_INC_ACOVERFLOW(cachep);
808                         __drain_alien_cache(cachep, ac, page_node, &list);
809                 }
810                 ac->entry[ac->avail++] = objp;
811                 spin_unlock(&alien->lock);
812                 slabs_destroy(cachep, &list);
813         } else {
814                 n = get_node(cachep, page_node);
815                 spin_lock(&n->list_lock);
816                 free_block(cachep, &objp, 1, page_node, &list);
817                 spin_unlock(&n->list_lock);
818                 slabs_destroy(cachep, &list);
819         }
820         return 1;
821 }
822 
823 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
824 {
825         int page_node = page_to_nid(virt_to_page(objp));
826         int node = numa_mem_id();
827         /*
828          * Make sure we are not freeing a object from another node to the array
829          * cache on this cpu.
830          */
831         if (likely(node == page_node))
832                 return 0;
833 
834         return __cache_free_alien(cachep, objp, node, page_node);
835 }
836 
837 /*
838  * Construct gfp mask to allocate from a specific node but do not reclaim or
839  * warn about failures.
840  */
841 static inline gfp_t gfp_exact_node(gfp_t flags)
842 {
843         return (flags | __GFP_THISNODE | __GFP_NOWARN) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
844 }
845 #endif
846 
847 static int init_cache_node(struct kmem_cache *cachep, int node, gfp_t gfp)
848 {
849         struct kmem_cache_node *n;
850 
851         /*
852          * Set up the kmem_cache_node for cpu before we can
853          * begin anything. Make sure some other cpu on this
854          * node has not already allocated this
855          */
856         n = get_node(cachep, node);
857         if (n) {
858                 spin_lock_irq(&n->list_lock);
859                 n->free_limit = (1 + nr_cpus_node(node)) * cachep->batchcount +
860                                 cachep->num;
861                 spin_unlock_irq(&n->list_lock);
862 
863                 return 0;
864         }
865 
866         n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
867         if (!n)
868                 return -ENOMEM;
869 
870         kmem_cache_node_init(n);
871         n->next_reap = jiffies + REAPTIMEOUT_NODE +
872                     ((unsigned long)cachep) % REAPTIMEOUT_NODE;
873 
874         n->free_limit =
875                 (1 + nr_cpus_node(node)) * cachep->batchcount + cachep->num;
876 
877         /*
878          * The kmem_cache_nodes don't come and go as CPUs
879          * come and go.  slab_mutex is sufficient
880          * protection here.
881          */
882         cachep->node[node] = n;
883 
884         return 0;
885 }
886 
887 #if (defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)) || defined(CONFIG_SMP)
888 /*
889  * Allocates and initializes node for a node on each slab cache, used for
890  * either memory or cpu hotplug.  If memory is being hot-added, the kmem_cache_node
891  * will be allocated off-node since memory is not yet online for the new node.
892  * When hotplugging memory or a cpu, existing node are not replaced if
893  * already in use.
894  *
895  * Must hold slab_mutex.
896  */
897 static int init_cache_node_node(int node)
898 {
899         int ret;
900         struct kmem_cache *cachep;
901 
902         list_for_each_entry(cachep, &slab_caches, list) {
903                 ret = init_cache_node(cachep, node, GFP_KERNEL);
904                 if (ret)
905                         return ret;
906         }
907 
908         return 0;
909 }
910 #endif
911 
912 static int setup_kmem_cache_node(struct kmem_cache *cachep,
913                                 int node, gfp_t gfp, bool force_change)
914 {
915         int ret = -ENOMEM;
916         struct kmem_cache_node *n;
917         struct array_cache *old_shared = NULL;
918         struct array_cache *new_shared = NULL;
919         struct alien_cache **new_alien = NULL;
920         LIST_HEAD(list);
921 
922         if (use_alien_caches) {
923                 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
924                 if (!new_alien)
925                         goto fail;
926         }
927 
928         if (cachep->shared) {
929                 new_shared = alloc_arraycache(node,
930                         cachep->shared * cachep->batchcount, 0xbaadf00d, gfp);
931                 if (!new_shared)
932                         goto fail;
933         }
934 
935         ret = init_cache_node(cachep, node, gfp);
936         if (ret)
937                 goto fail;
938 
939         n = get_node(cachep, node);
940         spin_lock_irq(&n->list_lock);
941         if (n->shared && force_change) {
942                 free_block(cachep, n->shared->entry,
943                                 n->shared->avail, node, &list);
944                 n->shared->avail = 0;
945         }
946 
947         if (!n->shared || force_change) {
948                 old_shared = n->shared;
949                 n->shared = new_shared;
950                 new_shared = NULL;
951         }
952 
953         if (!n->alien) {
954                 n->alien = new_alien;
955                 new_alien = NULL;
956         }
957 
958         spin_unlock_irq(&n->list_lock);
959         slabs_destroy(cachep, &list);
960 
961         /*
962          * To protect lockless access to n->shared during irq disabled context.
963          * If n->shared isn't NULL in irq disabled context, accessing to it is
964          * guaranteed to be valid until irq is re-enabled, because it will be
965          * freed after synchronize_sched().
966          */
967         if (old_shared && force_change)
968                 synchronize_sched();
969 
970 fail:
971         kfree(old_shared);
972         kfree(new_shared);
973         free_alien_cache(new_alien);
974 
975         return ret;
976 }
977 
978 #ifdef CONFIG_SMP
979 
980 static void cpuup_canceled(long cpu)
981 {
982         struct kmem_cache *cachep;
983         struct kmem_cache_node *n = NULL;
984         int node = cpu_to_mem(cpu);
985         const struct cpumask *mask = cpumask_of_node(node);
986 
987         list_for_each_entry(cachep, &slab_caches, list) {
988                 struct array_cache *nc;
989                 struct array_cache *shared;
990                 struct alien_cache **alien;
991                 LIST_HEAD(list);
992 
993                 n = get_node(cachep, node);
994                 if (!n)
995                         continue;
996 
997                 spin_lock_irq(&n->list_lock);
998 
999                 /* Free limit for this kmem_cache_node */
1000                 n->free_limit -= cachep->batchcount;
1001 
1002                 /* cpu is dead; no one can alloc from it. */
1003                 nc = per_cpu_ptr(cachep->cpu_cache, cpu);
1004                 if (nc) {
1005                         free_block(cachep, nc->entry, nc->avail, node, &list);
1006                         nc->avail = 0;
1007                 }
1008 
1009                 if (!cpumask_empty(mask)) {
1010                         spin_unlock_irq(&n->list_lock);
1011                         goto free_slab;
1012                 }
1013 
1014                 shared = n->shared;
1015                 if (shared) {
1016                         free_block(cachep, shared->entry,
1017                                    shared->avail, node, &list);
1018                         n->shared = NULL;
1019                 }
1020 
1021                 alien = n->alien;
1022                 n->alien = NULL;
1023 
1024                 spin_unlock_irq(&n->list_lock);
1025 
1026                 kfree(shared);
1027                 if (alien) {
1028                         drain_alien_cache(cachep, alien);
1029                         free_alien_cache(alien);
1030                 }
1031 
1032 free_slab:
1033                 slabs_destroy(cachep, &list);
1034         }
1035         /*
1036          * In the previous loop, all the objects were freed to
1037          * the respective cache's slabs,  now we can go ahead and
1038          * shrink each nodelist to its limit.
1039          */
1040         list_for_each_entry(cachep, &slab_caches, list) {
1041                 n = get_node(cachep, node);
1042                 if (!n)
1043                         continue;
1044                 drain_freelist(cachep, n, INT_MAX);
1045         }
1046 }
1047 
1048 static int cpuup_prepare(long cpu)
1049 {
1050         struct kmem_cache *cachep;
1051         int node = cpu_to_mem(cpu);
1052         int err;
1053 
1054         /*
1055          * We need to do this right in the beginning since
1056          * alloc_arraycache's are going to use this list.
1057          * kmalloc_node allows us to add the slab to the right
1058          * kmem_cache_node and not this cpu's kmem_cache_node
1059          */
1060         err = init_cache_node_node(node);
1061         if (err < 0)
1062                 goto bad;
1063 
1064         /*
1065          * Now we can go ahead with allocating the shared arrays and
1066          * array caches
1067          */
1068         list_for_each_entry(cachep, &slab_caches, list) {
1069                 err = setup_kmem_cache_node(cachep, node, GFP_KERNEL, false);
1070                 if (err)
1071                         goto bad;
1072         }
1073 
1074         return 0;
1075 bad:
1076         cpuup_canceled(cpu);
1077         return -ENOMEM;
1078 }
1079 
1080 int slab_prepare_cpu(unsigned int cpu)
1081 {
1082         int err;
1083 
1084         mutex_lock(&slab_mutex);
1085         err = cpuup_prepare(cpu);
1086         mutex_unlock(&slab_mutex);
1087         return err;
1088 }
1089 
1090 /*
1091  * This is called for a failed online attempt and for a successful
1092  * offline.
1093  *
1094  * Even if all the cpus of a node are down, we don't free the
1095  * kmem_list3 of any cache. This to avoid a race between cpu_down, and
1096  * a kmalloc allocation from another cpu for memory from the node of
1097  * the cpu going down.  The list3 structure is usually allocated from
1098  * kmem_cache_create() and gets destroyed at kmem_cache_destroy().
1099  */
1100 int slab_dead_cpu(unsigned int cpu)
1101 {
1102         mutex_lock(&slab_mutex);
1103         cpuup_canceled(cpu);
1104         mutex_unlock(&slab_mutex);
1105         return 0;
1106 }
1107 #endif
1108 
1109 static int slab_online_cpu(unsigned int cpu)
1110 {
1111         start_cpu_timer(cpu);
1112         return 0;
1113 }
1114 
1115 static int slab_offline_cpu(unsigned int cpu)
1116 {
1117         /*
1118          * Shutdown cache reaper. Note that the slab_mutex is held so
1119          * that if cache_reap() is invoked it cannot do anything
1120          * expensive but will only modify reap_work and reschedule the
1121          * timer.
1122          */
1123         cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1124         /* Now the cache_reaper is guaranteed to be not running. */
1125         per_cpu(slab_reap_work, cpu).work.func = NULL;
1126         return 0;
1127 }
1128 
1129 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1130 /*
1131  * Drains freelist for a node on each slab cache, used for memory hot-remove.
1132  * Returns -EBUSY if all objects cannot be drained so that the node is not
1133  * removed.
1134  *
1135  * Must hold slab_mutex.
1136  */
1137 static int __meminit drain_cache_node_node(int node)
1138 {
1139         struct kmem_cache *cachep;
1140         int ret = 0;
1141 
1142         list_for_each_entry(cachep, &slab_caches, list) {
1143                 struct kmem_cache_node *n;
1144 
1145                 n = get_node(cachep, node);
1146                 if (!n)
1147                         continue;
1148 
1149                 drain_freelist(cachep, n, INT_MAX);
1150 
1151                 if (!list_empty(&n->slabs_full) ||
1152                     !list_empty(&n->slabs_partial)) {
1153                         ret = -EBUSY;
1154                         break;
1155                 }
1156         }
1157         return ret;
1158 }
1159 
1160 static int __meminit slab_memory_callback(struct notifier_block *self,
1161                                         unsigned long action, void *arg)
1162 {
1163         struct memory_notify *mnb = arg;
1164         int ret = 0;
1165         int nid;
1166 
1167         nid = mnb->status_change_nid;
1168         if (nid < 0)
1169                 goto out;
1170 
1171         switch (action) {
1172         case MEM_GOING_ONLINE:
1173                 mutex_lock(&slab_mutex);
1174                 ret = init_cache_node_node(nid);
1175                 mutex_unlock(&slab_mutex);
1176                 break;
1177         case MEM_GOING_OFFLINE:
1178                 mutex_lock(&slab_mutex);
1179                 ret = drain_cache_node_node(nid);
1180                 mutex_unlock(&slab_mutex);
1181                 break;
1182         case MEM_ONLINE:
1183         case MEM_OFFLINE:
1184         case MEM_CANCEL_ONLINE:
1185         case MEM_CANCEL_OFFLINE:
1186                 break;
1187         }
1188 out:
1189         return notifier_from_errno(ret);
1190 }
1191 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1192 
1193 /*
1194  * swap the static kmem_cache_node with kmalloced memory
1195  */
1196 static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
1197                                 int nodeid)
1198 {
1199         struct kmem_cache_node *ptr;
1200 
1201         ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
1202         BUG_ON(!ptr);
1203 
1204         memcpy(ptr, list, sizeof(struct kmem_cache_node));
1205         /*
1206          * Do not assume that spinlocks can be initialized via memcpy:
1207          */
1208         spin_lock_init(&ptr->list_lock);
1209 
1210         MAKE_ALL_LISTS(cachep, ptr, nodeid);
1211         cachep->node[nodeid] = ptr;
1212 }
1213 
1214 /*
1215  * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1216  * size of kmem_cache_node.
1217  */
1218 static void __init set_up_node(struct kmem_cache *cachep, int index)
1219 {
1220         int node;
1221 
1222         for_each_online_node(node) {
1223                 cachep->node[node] = &init_kmem_cache_node[index + node];
1224                 cachep->node[node]->next_reap = jiffies +
1225                     REAPTIMEOUT_NODE +
1226                     ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1227         }
1228 }
1229 
1230 /*
1231  * Initialisation.  Called after the page allocator have been initialised and
1232  * before smp_init().
1233  */
1234 void __init kmem_cache_init(void)
1235 {
1236         int i;
1237 
1238         BUILD_BUG_ON(sizeof(((struct page *)NULL)->lru) <
1239                                         sizeof(struct rcu_head));
1240         kmem_cache = &kmem_cache_boot;
1241 
1242         if (!IS_ENABLED(CONFIG_NUMA) || num_possible_nodes() == 1)
1243                 use_alien_caches = 0;
1244 
1245         for (i = 0; i < NUM_INIT_LISTS; i++)
1246                 kmem_cache_node_init(&init_kmem_cache_node[i]);
1247 
1248         /*
1249          * Fragmentation resistance on low memory - only use bigger
1250          * page orders on machines with more than 32MB of memory if
1251          * not overridden on the command line.
1252          */
1253         if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
1254                 slab_max_order = SLAB_MAX_ORDER_HI;
1255 
1256         /* Bootstrap is tricky, because several objects are allocated
1257          * from caches that do not exist yet:
1258          * 1) initialize the kmem_cache cache: it contains the struct
1259          *    kmem_cache structures of all caches, except kmem_cache itself:
1260          *    kmem_cache is statically allocated.
1261          *    Initially an __init data area is used for the head array and the
1262          *    kmem_cache_node structures, it's replaced with a kmalloc allocated
1263          *    array at the end of the bootstrap.
1264          * 2) Create the first kmalloc cache.
1265          *    The struct kmem_cache for the new cache is allocated normally.
1266          *    An __init data area is used for the head array.
1267          * 3) Create the remaining kmalloc caches, with minimally sized
1268          *    head arrays.
1269          * 4) Replace the __init data head arrays for kmem_cache and the first
1270          *    kmalloc cache with kmalloc allocated arrays.
1271          * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1272          *    the other cache's with kmalloc allocated memory.
1273          * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1274          */
1275 
1276         /* 1) create the kmem_cache */
1277 
1278         /*
1279          * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1280          */
1281         create_boot_cache(kmem_cache, "kmem_cache",
1282                 offsetof(struct kmem_cache, node) +
1283                                   nr_node_ids * sizeof(struct kmem_cache_node *),
1284                                   SLAB_HWCACHE_ALIGN);
1285         list_add(&kmem_cache->list, &slab_caches);
1286         slab_state = PARTIAL;
1287 
1288         /*
1289          * Initialize the caches that provide memory for the  kmem_cache_node
1290          * structures first.  Without this, further allocations will bug.
1291          */
1292         kmalloc_caches[INDEX_NODE] = create_kmalloc_cache(
1293                                 kmalloc_info[INDEX_NODE].name,
1294                                 kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS);
1295         slab_state = PARTIAL_NODE;
1296         setup_kmalloc_cache_index_table();
1297 
1298         slab_early_init = 0;
1299 
1300         /* 5) Replace the bootstrap kmem_cache_node */
1301         {
1302                 int nid;
1303 
1304                 for_each_online_node(nid) {
1305                         init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
1306 
1307                         init_list(kmalloc_caches[INDEX_NODE],
1308                                           &init_kmem_cache_node[SIZE_NODE + nid], nid);
1309                 }
1310         }
1311 
1312         create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
1313 }
1314 
1315 void __init kmem_cache_init_late(void)
1316 {
1317         struct kmem_cache *cachep;
1318 
1319         slab_state = UP;
1320 
1321         /* 6) resize the head arrays to their final sizes */
1322         mutex_lock(&slab_mutex);
1323         list_for_each_entry(cachep, &slab_caches, list)
1324                 if (enable_cpucache(cachep, GFP_NOWAIT))
1325                         BUG();
1326         mutex_unlock(&slab_mutex);
1327 
1328         /* Done! */
1329         slab_state = FULL;
1330 
1331 #ifdef CONFIG_NUMA
1332         /*
1333          * Register a memory hotplug callback that initializes and frees
1334          * node.
1335          */
1336         hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1337 #endif
1338 
1339         /*
1340          * The reap timers are started later, with a module init call: That part
1341          * of the kernel is not yet operational.
1342          */
1343 }
1344 
1345 static int __init cpucache_init(void)
1346 {
1347         int ret;
1348 
1349         /*
1350          * Register the timers that return unneeded pages to the page allocator
1351          */
1352         ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "SLAB online",
1353                                 slab_online_cpu, slab_offline_cpu);
1354         WARN_ON(ret < 0);
1355 
1356         /* Done! */
1357         slab_state = FULL;
1358         return 0;
1359 }
1360 __initcall(cpucache_init);
1361 
1362 static noinline void
1363 slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1364 {
1365 #if DEBUG
1366         struct kmem_cache_node *n;
1367         unsigned long flags;
1368         int node;
1369         static DEFINE_RATELIMIT_STATE(slab_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1370                                       DEFAULT_RATELIMIT_BURST);
1371 
1372         if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slab_oom_rs))
1373                 return;
1374 
1375         pr_warn("SLAB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
1376                 nodeid, gfpflags, &gfpflags);
1377         pr_warn("  cache: %s, object size: %d, order: %d\n",
1378                 cachep->name, cachep->size, cachep->gfporder);
1379 
1380         for_each_kmem_cache_node(cachep, node, n) {
1381                 unsigned long total_slabs, free_slabs, free_objs;
1382 
1383                 spin_lock_irqsave(&n->list_lock, flags);
1384                 total_slabs = n->total_slabs;
1385                 free_slabs = n->free_slabs;
1386                 free_objs = n->free_objects;
1387                 spin_unlock_irqrestore(&n->list_lock, flags);
1388 
1389                 pr_warn("  node %d: slabs: %ld/%ld, objs: %ld/%ld\n",
1390                         node, total_slabs - free_slabs, total_slabs,
1391                         (total_slabs * cachep->num) - free_objs,
1392                         total_slabs * cachep->num);
1393         }
1394 #endif
1395 }
1396 
1397 /*
1398  * Interface to system's page allocator. No need to hold the
1399  * kmem_cache_node ->list_lock.
1400  *
1401  * If we requested dmaable memory, we will get it. Even if we
1402  * did not request dmaable memory, we might get it, but that
1403  * would be relatively rare and ignorable.
1404  */
1405 static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1406                                                                 int nodeid)
1407 {
1408         struct page *page;
1409         int nr_pages;
1410 
1411         flags |= cachep->allocflags;
1412         if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1413                 flags |= __GFP_RECLAIMABLE;
1414 
1415         page = __alloc_pages_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1416         if (!page) {
1417                 slab_out_of_memory(cachep, flags, nodeid);
1418                 return NULL;
1419         }
1420 
1421         if (memcg_charge_slab(page, flags, cachep->gfporder, cachep)) {
1422                 __free_pages(page, cachep->gfporder);
1423                 return NULL;
1424         }
1425 
1426         nr_pages = (1 << cachep->gfporder);
1427         if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1428                 mod_lruvec_page_state(page, NR_SLAB_RECLAIMABLE, nr_pages);
1429         else
1430                 mod_lruvec_page_state(page, NR_SLAB_UNRECLAIMABLE, nr_pages);
1431 
1432         __SetPageSlab(page);
1433         /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1434         if (sk_memalloc_socks() && page_is_pfmemalloc(page))
1435                 SetPageSlabPfmemalloc(page);
1436 
1437         if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1438                 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1439 
1440                 if (cachep->ctor)
1441                         kmemcheck_mark_uninitialized_pages(page, nr_pages);
1442                 else
1443                         kmemcheck_mark_unallocated_pages(page, nr_pages);
1444         }
1445 
1446         return page;
1447 }
1448 
1449 /*
1450  * Interface to system's page release.
1451  */
1452 static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
1453 {
1454         int order = cachep->gfporder;
1455         unsigned long nr_freed = (1 << order);
1456 
1457         kmemcheck_free_shadow(page, order);
1458 
1459         if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1460                 mod_lruvec_page_state(page, NR_SLAB_RECLAIMABLE, -nr_freed);
1461         else
1462                 mod_lruvec_page_state(page, NR_SLAB_UNRECLAIMABLE, -nr_freed);
1463 
1464         BUG_ON(!PageSlab(page));
1465         __ClearPageSlabPfmemalloc(page);
1466         __ClearPageSlab(page);
1467         page_mapcount_reset(page);
1468         page->mapping = NULL;
1469 
1470         if (current->reclaim_state)
1471                 current->reclaim_state->reclaimed_slab += nr_freed;
1472         memcg_uncharge_slab(page, order, cachep);
1473         __free_pages(page, order);
1474 }
1475 
1476 static void kmem_rcu_free(struct rcu_head *head)
1477 {
1478         struct kmem_cache *cachep;
1479         struct page *page;
1480 
1481         page = container_of(head, struct page, rcu_head);
1482         cachep = page->slab_cache;
1483 
1484         kmem_freepages(cachep, page);
1485 }
1486 
1487 #if DEBUG
1488 static bool is_debug_pagealloc_cache(struct kmem_cache *cachep)
1489 {
1490         if (debug_pagealloc_enabled() && OFF_SLAB(cachep) &&
1491                 (cachep->size % PAGE_SIZE) == 0)
1492                 return true;
1493 
1494         return false;
1495 }
1496 
1497 #ifdef CONFIG_DEBUG_PAGEALLOC
1498 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1499                             unsigned long caller)
1500 {
1501         int size = cachep->object_size;
1502 
1503         addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1504 
1505         if (size < 5 * sizeof(unsigned long))
1506                 return;
1507 
1508         *addr++ = 0x12345678;
1509         *addr++ = caller;
1510         *addr++ = smp_processor_id();
1511         size -= 3 * sizeof(unsigned long);
1512         {
1513                 unsigned long *sptr = &caller;
1514                 unsigned long svalue;
1515 
1516                 while (!kstack_end(sptr)) {
1517                         svalue = *sptr++;
1518                         if (kernel_text_address(svalue)) {
1519                                 *addr++ = svalue;
1520                                 size -= sizeof(unsigned long);
1521                                 if (size <= sizeof(unsigned long))
1522                                         break;
1523                         }
1524                 }
1525 
1526         }
1527         *addr++ = 0x87654321;
1528 }
1529 
1530 static void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1531                                 int map, unsigned long caller)
1532 {
1533         if (!is_debug_pagealloc_cache(cachep))
1534                 return;
1535 
1536         if (caller)
1537                 store_stackinfo(cachep, objp, caller);
1538 
1539         kernel_map_pages(virt_to_page(objp), cachep->size / PAGE_SIZE, map);
1540 }
1541 
1542 #else
1543 static inline void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1544                                 int map, unsigned long caller) {}
1545 
1546 #endif
1547 
1548 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1549 {
1550         int size = cachep->object_size;
1551         addr = &((char *)addr)[obj_offset(cachep)];
1552 
1553         memset(addr, val, size);
1554         *(unsigned char *)(addr + size - 1) = POISON_END;
1555 }
1556 
1557 static void dump_line(char *data, int offset, int limit)
1558 {
1559         int i;
1560         unsigned char error = 0;
1561         int bad_count = 0;
1562 
1563         pr_err("%03x: ", offset);
1564         for (i = 0; i < limit; i++) {
1565                 if (data[offset + i] != POISON_FREE) {
1566                         error = data[offset + i];
1567                         bad_count++;
1568                 }
1569         }
1570         print_hex_dump(KERN_CONT, "", 0, 16, 1,
1571                         &data[offset], limit, 1);
1572 
1573         if (bad_count == 1) {
1574                 error ^= POISON_FREE;
1575                 if (!(error & (error - 1))) {
1576                         pr_err("Single bit error detected. Probably bad RAM.\n");
1577 #ifdef CONFIG_X86
1578                         pr_err("Run memtest86+ or a similar memory test tool.\n");
1579 #else
1580                         pr_err("Run a memory test tool.\n");
1581 #endif
1582                 }
1583         }
1584 }
1585 #endif
1586 
1587 #if DEBUG
1588 
1589 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1590 {
1591         int i, size;
1592         char *realobj;
1593 
1594         if (cachep->flags & SLAB_RED_ZONE) {
1595                 pr_err("Redzone: 0x%llx/0x%llx\n",
1596                        *dbg_redzone1(cachep, objp),
1597                        *dbg_redzone2(cachep, objp));
1598         }
1599 
1600         if (cachep->flags & SLAB_STORE_USER) {
1601                 pr_err("Last user: [<%p>](%pSR)\n",
1602                        *dbg_userword(cachep, objp),
1603                        *dbg_userword(cachep, objp));
1604         }
1605         realobj = (char *)objp + obj_offset(cachep);
1606         size = cachep->object_size;
1607         for (i = 0; i < size && lines; i += 16, lines--) {
1608                 int limit;
1609                 limit = 16;
1610                 if (i + limit > size)
1611                         limit = size - i;
1612                 dump_line(realobj, i, limit);
1613         }
1614 }
1615 
1616 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1617 {
1618         char *realobj;
1619         int size, i;
1620         int lines = 0;
1621 
1622         if (is_debug_pagealloc_cache(cachep))
1623                 return;
1624 
1625         realobj = (char *)objp + obj_offset(cachep);
1626         size = cachep->object_size;
1627 
1628         for (i = 0; i < size; i++) {
1629                 char exp = POISON_FREE;
1630                 if (i == size - 1)
1631                         exp = POISON_END;
1632                 if (realobj[i] != exp) {
1633                         int limit;
1634                         /* Mismatch ! */
1635                         /* Print header */
1636                         if (lines == 0) {
1637                                 pr_err("Slab corruption (%s): %s start=%p, len=%d\n",
1638                                        print_tainted(), cachep->name,
1639                                        realobj, size);
1640                                 print_objinfo(cachep, objp, 0);
1641                         }
1642                         /* Hexdump the affected line */
1643                         i = (i / 16) * 16;
1644                         limit = 16;
1645                         if (i + limit > size)
1646                                 limit = size - i;
1647                         dump_line(realobj, i, limit);
1648                         i += 16;
1649                         lines++;
1650                         /* Limit to 5 lines */
1651                         if (lines > 5)
1652                                 break;
1653                 }
1654         }
1655         if (lines != 0) {
1656                 /* Print some data about the neighboring objects, if they
1657                  * exist:
1658                  */
1659                 struct page *page = virt_to_head_page(objp);
1660                 unsigned int objnr;
1661 
1662                 objnr = obj_to_index(cachep, page, objp);
1663                 if (objnr) {
1664                         objp = index_to_obj(cachep, page, objnr - 1);
1665                         realobj = (char *)objp + obj_offset(cachep);
1666                         pr_err("Prev obj: start=%p, len=%d\n", realobj, size);
1667                         print_objinfo(cachep, objp, 2);
1668                 }
1669                 if (objnr + 1 < cachep->num) {
1670                         objp = index_to_obj(cachep, page, objnr + 1);
1671                         realobj = (char *)objp + obj_offset(cachep);
1672                         pr_err("Next obj: start=%p, len=%d\n", realobj, size);
1673                         print_objinfo(cachep, objp, 2);
1674                 }
1675         }
1676 }
1677 #endif
1678 
1679 #if DEBUG
1680 static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1681                                                 struct page *page)
1682 {
1683         int i;
1684 
1685         if (OBJFREELIST_SLAB(cachep) && cachep->flags & SLAB_POISON) {
1686                 poison_obj(cachep, page->freelist - obj_offset(cachep),
1687                         POISON_FREE);
1688         }
1689 
1690         for (i = 0; i < cachep->num; i++) {
1691                 void *objp = index_to_obj(cachep, page, i);
1692 
1693                 if (cachep->flags & SLAB_POISON) {
1694                         check_poison_obj(cachep, objp);
1695                         slab_kernel_map(cachep, objp, 1, 0);
1696                 }
1697                 if (cachep->flags & SLAB_RED_ZONE) {
1698                         if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1699                                 slab_error(cachep, "start of a freed object was overwritten");
1700                         if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1701                                 slab_error(cachep, "end of a freed object was overwritten");
1702                 }
1703         }
1704 }
1705 #else
1706 static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1707                                                 struct page *page)
1708 {
1709 }
1710 #endif
1711 
1712 /**
1713  * slab_destroy - destroy and release all objects in a slab
1714  * @cachep: cache pointer being destroyed
1715  * @page: page pointer being destroyed
1716  *
1717  * Destroy all the objs in a slab page, and release the mem back to the system.
1718  * Before calling the slab page must have been unlinked from the cache. The
1719  * kmem_cache_node ->list_lock is not held/needed.
1720  */
1721 static void slab_destroy(struct kmem_cache *cachep, struct page *page)
1722 {
1723         void *freelist;
1724 
1725         freelist = page->freelist;
1726         slab_destroy_debugcheck(cachep, page);
1727         if (unlikely(cachep->flags & SLAB_TYPESAFE_BY_RCU))
1728                 call_rcu(&page->rcu_head, kmem_rcu_free);
1729         else
1730                 kmem_freepages(cachep, page);
1731 
1732         /*
1733          * From now on, we don't use freelist
1734          * although actual page can be freed in rcu context
1735          */
1736         if (OFF_SLAB(cachep))
1737                 kmem_cache_free(cachep->freelist_cache, freelist);
1738 }
1739 
1740 static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list)
1741 {
1742         struct page *page, *n;
1743 
1744         list_for_each_entry_safe(page, n, list, lru) {
1745                 list_del(&page->lru);
1746                 slab_destroy(cachep, page);
1747         }
1748 }
1749 
1750 /**
1751  * calculate_slab_order - calculate size (page order) of slabs
1752  * @cachep: pointer to the cache that is being created
1753  * @size: size of objects to be created in this cache.
1754  * @flags: slab allocation flags
1755  *
1756  * Also calculates the number of objects per slab.
1757  *
1758  * This could be made much more intelligent.  For now, try to avoid using
1759  * high order pages for slabs.  When the gfp() functions are more friendly
1760  * towards high-order requests, this should be changed.
1761  */
1762 static size_t calculate_slab_order(struct kmem_cache *cachep,
1763                                 size_t size, unsigned long flags)
1764 {
1765         size_t left_over = 0;
1766         int gfporder;
1767 
1768         for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
1769                 unsigned int num;
1770                 size_t remainder;
1771 
1772                 num = cache_estimate(gfporder, size, flags, &remainder);
1773                 if (!num)
1774                         continue;
1775 
1776                 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1777                 if (num > SLAB_OBJ_MAX_NUM)
1778                         break;
1779 
1780                 if (flags & CFLGS_OFF_SLAB) {
1781                         struct kmem_cache *freelist_cache;
1782                         size_t freelist_size;
1783 
1784                         freelist_size = num * sizeof(freelist_idx_t);
1785                         freelist_cache = kmalloc_slab(freelist_size, 0u);
1786                         if (!freelist_cache)
1787                                 continue;
1788 
1789                         /*
1790                          * Needed to avoid possible looping condition
1791                          * in cache_grow_begin()
1792                          */
1793                         if (OFF_SLAB(freelist_cache))
1794                                 continue;
1795 
1796                         /* check if off slab has enough benefit */
1797                         if (freelist_cache->size > cachep->size / 2)
1798                                 continue;
1799                 }
1800 
1801                 /* Found something acceptable - save it away */
1802                 cachep->num = num;
1803                 cachep->gfporder = gfporder;
1804                 left_over = remainder;
1805 
1806                 /*
1807                  * A VFS-reclaimable slab tends to have most allocations
1808                  * as GFP_NOFS and we really don't want to have to be allocating
1809                  * higher-order pages when we are unable to shrink dcache.
1810                  */
1811                 if (flags & SLAB_RECLAIM_ACCOUNT)
1812                         break;
1813 
1814                 /*
1815                  * Large number of objects is good, but very large slabs are
1816                  * currently bad for the gfp()s.
1817                  */
1818                 if (gfporder >= slab_max_order)
1819                         break;
1820 
1821                 /*
1822                  * Acceptable internal fragmentation?
1823                  */
1824                 if (left_over * 8 <= (PAGE_SIZE << gfporder))
1825                         break;
1826         }
1827         return left_over;
1828 }
1829 
1830 static struct array_cache __percpu *alloc_kmem_cache_cpus(
1831                 struct kmem_cache *cachep, int entries, int batchcount)
1832 {
1833         int cpu;
1834         size_t size;
1835         struct array_cache __percpu *cpu_cache;
1836 
1837         size = sizeof(void *) * entries + sizeof(struct array_cache);
1838         cpu_cache = __alloc_percpu(size, sizeof(void *));
1839 
1840         if (!cpu_cache)
1841                 return NULL;
1842 
1843         for_each_possible_cpu(cpu) {
1844                 init_arraycache(per_cpu_ptr(cpu_cache, cpu),
1845                                 entries, batchcount);
1846         }
1847 
1848         return cpu_cache;
1849 }
1850 
1851 static int __ref setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
1852 {
1853         if (slab_state >= FULL)
1854                 return enable_cpucache(cachep, gfp);
1855 
1856         cachep->cpu_cache = alloc_kmem_cache_cpus(cachep, 1, 1);
1857         if (!cachep->cpu_cache)
1858                 return 1;
1859 
1860         if (slab_state == DOWN) {
1861                 /* Creation of first cache (kmem_cache). */
1862                 set_up_node(kmem_cache, CACHE_CACHE);
1863         } else if (slab_state == PARTIAL) {
1864                 /* For kmem_cache_node */
1865                 set_up_node(cachep, SIZE_NODE);
1866         } else {
1867                 int node;
1868 
1869                 for_each_online_node(node) {
1870                         cachep->node[node] = kmalloc_node(
1871                                 sizeof(struct kmem_cache_node), gfp, node);
1872                         BUG_ON(!cachep->node[node]);
1873                         kmem_cache_node_init(cachep->node[node]);
1874                 }
1875         }
1876 
1877         cachep->node[numa_mem_id()]->next_reap =
1878                         jiffies + REAPTIMEOUT_NODE +
1879                         ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1880 
1881         cpu_cache_get(cachep)->avail = 0;
1882         cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1883         cpu_cache_get(cachep)->batchcount = 1;
1884         cpu_cache_get(cachep)->touched = 0;
1885         cachep->batchcount = 1;
1886         cachep->limit = BOOT_CPUCACHE_ENTRIES;
1887         return 0;
1888 }
1889 
1890 unsigned long kmem_cache_flags(unsigned long object_size,
1891         unsigned long flags, const char *name,
1892         void (*ctor)(void *))
1893 {
1894         return flags;
1895 }
1896 
1897 struct kmem_cache *
1898 __kmem_cache_alias(const char *name, size_t size, size_t align,
1899                    unsigned long flags, void (*ctor)(void *))
1900 {
1901         struct kmem_cache *cachep;
1902 
1903         cachep = find_mergeable(size, align, flags, name, ctor);
1904         if (cachep) {
1905                 cachep->refcount++;
1906 
1907                 /*
1908                  * Adjust the object sizes so that we clear
1909                  * the complete object on kzalloc.
1910                  */
1911                 cachep->object_size = max_t(int, cachep->object_size, size);
1912         }
1913         return cachep;
1914 }
1915 
1916 static bool set_objfreelist_slab_cache(struct kmem_cache *cachep,
1917                         size_t size, unsigned long flags)
1918 {
1919         size_t left;
1920 
1921         cachep->num = 0;
1922 
1923         if (cachep->ctor || flags & SLAB_TYPESAFE_BY_RCU)
1924                 return false;
1925 
1926         left = calculate_slab_order(cachep, size,
1927                         flags | CFLGS_OBJFREELIST_SLAB);
1928         if (!cachep->num)
1929                 return false;
1930 
1931         if (cachep->num * sizeof(freelist_idx_t) > cachep->object_size)
1932                 return false;
1933 
1934         cachep->colour = left / cachep->colour_off;
1935 
1936         return true;
1937 }
1938 
1939 static bool set_off_slab_cache(struct kmem_cache *cachep,
1940                         size_t size, unsigned long flags)
1941 {
1942         size_t left;
1943 
1944         cachep->num = 0;
1945 
1946         /*
1947          * Always use on-slab management when SLAB_NOLEAKTRACE
1948          * to avoid recursive calls into kmemleak.
1949          */
1950         if (flags & SLAB_NOLEAKTRACE)
1951                 return false;
1952 
1953         /*
1954          * Size is large, assume best to place the slab management obj
1955          * off-slab (should allow better packing of objs).
1956          */
1957         left = calculate_slab_order(cachep, size, flags | CFLGS_OFF_SLAB);
1958         if (!cachep->num)
1959                 return false;
1960 
1961         /*
1962          * If the slab has been placed off-slab, and we have enough space then
1963          * move it on-slab. This is at the expense of any extra colouring.
1964          */
1965         if (left >= cachep->num * sizeof(freelist_idx_t))
1966                 return false;
1967 
1968         cachep->colour = left / cachep->colour_off;
1969 
1970         return true;
1971 }
1972 
1973 static bool set_on_slab_cache(struct kmem_cache *cachep,
1974                         size_t size, unsigned long flags)
1975 {
1976         size_t left;
1977 
1978         cachep->num = 0;
1979 
1980         left = calculate_slab_order(cachep, size, flags);
1981         if (!cachep->num)
1982                 return false;
1983 
1984         cachep->colour = left / cachep->colour_off;
1985 
1986         return true;
1987 }
1988 
1989 /**
1990  * __kmem_cache_create - Create a cache.
1991  * @cachep: cache management descriptor
1992  * @flags: SLAB flags
1993  *
1994  * Returns a ptr to the cache on success, NULL on failure.
1995  * Cannot be called within a int, but can be interrupted.
1996  * The @ctor is run when new pages are allocated by the cache.
1997  *
1998  * The flags are
1999  *
2000  * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2001  * to catch references to uninitialised memory.
2002  *
2003  * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2004  * for buffer overruns.
2005  *
2006  * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2007  * cacheline.  This can be beneficial if you're counting cycles as closely
2008  * as davem.
2009  */
2010 int
2011 __kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
2012 {
2013         size_t ralign = BYTES_PER_WORD;
2014         gfp_t gfp;
2015         int err;
2016         size_t size = cachep->size;
2017 
2018 #if DEBUG
2019 #if FORCED_DEBUG
2020         /*
2021          * Enable redzoning and last user accounting, except for caches with
2022          * large objects, if the increased size would increase the object size
2023          * above the next power of two: caches with object sizes just above a
2024          * power of two have a significant amount of internal fragmentation.
2025          */
2026         if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2027                                                 2 * sizeof(unsigned long long)))
2028                 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2029         if (!(flags & SLAB_TYPESAFE_BY_RCU))
2030                 flags |= SLAB_POISON;
2031 #endif
2032 #endif
2033 
2034         /*
2035          * Check that size is in terms of words.  This is needed to avoid
2036          * unaligned accesses for some archs when redzoning is used, and makes
2037          * sure any on-slab bufctl's are also correctly aligned.
2038          */
2039         size = ALIGN(size, BYTES_PER_WORD);
2040 
2041         if (flags & SLAB_RED_ZONE) {
2042                 ralign = REDZONE_ALIGN;
2043                 /* If redzoning, ensure that the second redzone is suitably
2044                  * aligned, by adjusting the object size accordingly. */
2045                 size = ALIGN(size, REDZONE_ALIGN);
2046         }
2047 
2048         /* 3) caller mandated alignment */
2049         if (ralign < cachep->align) {
2050                 ralign = cachep->align;
2051         }
2052         /* disable debug if necessary */
2053         if (ralign > __alignof__(unsigned long long))
2054                 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2055         /*
2056          * 4) Store it.
2057          */
2058         cachep->align = ralign;
2059         cachep->colour_off = cache_line_size();
2060         /* Offset must be a multiple of the alignment. */
2061         if (cachep->colour_off < cachep->align)
2062                 cachep->colour_off = cachep->align;
2063 
2064         if (slab_is_available())
2065                 gfp = GFP_KERNEL;
2066         else
2067                 gfp = GFP_NOWAIT;
2068 
2069 #if DEBUG
2070 
2071         /*
2072          * Both debugging options require word-alignment which is calculated
2073          * into align above.
2074          */
2075         if (flags & SLAB_RED_ZONE) {
2076                 /* add space for red zone words */
2077                 cachep->obj_offset += sizeof(unsigned long long);
2078                 size += 2 * sizeof(unsigned long long);
2079         }
2080         if (flags & SLAB_STORE_USER) {
2081                 /* user store requires one word storage behind the end of
2082                  * the real object. But if the second red zone needs to be
2083                  * aligned to 64 bits, we must allow that much space.
2084                  */
2085                 if (flags & SLAB_RED_ZONE)
2086                         size += REDZONE_ALIGN;
2087                 else
2088                         size += BYTES_PER_WORD;
2089         }
2090 #endif
2091 
2092         kasan_cache_create(cachep, &size, &flags);
2093 
2094         size = ALIGN(size, cachep->align);
2095         /*
2096          * We should restrict the number of objects in a slab to implement
2097          * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2098          */
2099         if (FREELIST_BYTE_INDEX && size < SLAB_OBJ_MIN_SIZE)
2100                 size = ALIGN(SLAB_OBJ_MIN_SIZE, cachep->align);
2101 
2102 #if DEBUG
2103         /*
2104          * To activate debug pagealloc, off-slab management is necessary
2105          * requirement. In early phase of initialization, small sized slab
2106          * doesn't get initialized so it would not be possible. So, we need
2107          * to check size >= 256. It guarantees that all necessary small
2108          * sized slab is initialized in current slab initialization sequence.
2109          */
2110         if (debug_pagealloc_enabled() && (flags & SLAB_POISON) &&
2111                 size >= 256 && cachep->object_size > cache_line_size()) {
2112                 if (size < PAGE_SIZE || size % PAGE_SIZE == 0) {
2113                         size_t tmp_size = ALIGN(size, PAGE_SIZE);
2114 
2115                         if (set_off_slab_cache(cachep, tmp_size, flags)) {
2116                                 flags |= CFLGS_OFF_SLAB;
2117                                 cachep->obj_offset += tmp_size - size;
2118                                 size = tmp_size;
2119                                 goto done;
2120                         }
2121                 }
2122         }
2123 #endif
2124 
2125         if (set_objfreelist_slab_cache(cachep, size, flags)) {
2126                 flags |= CFLGS_OBJFREELIST_SLAB;
2127                 goto done;
2128         }
2129 
2130         if (set_off_slab_cache(cachep, size, flags)) {
2131                 flags |= CFLGS_OFF_SLAB;
2132                 goto done;
2133         }
2134 
2135         if (set_on_slab_cache(cachep, size, flags))
2136                 goto done;
2137 
2138         return -E2BIG;
2139 
2140 done:
2141         cachep->freelist_size = cachep->num * sizeof(freelist_idx_t);
2142         cachep->flags = flags;
2143         cachep->allocflags = __GFP_COMP;
2144         if (flags & SLAB_CACHE_DMA)
2145                 cachep->allocflags |= GFP_DMA;
2146         cachep->size = size;
2147         cachep->reciprocal_buffer_size = reciprocal_value(size);
2148 
2149 #if DEBUG
2150         /*
2151          * If we're going to use the generic kernel_map_pages()
2152          * poisoning, then it's going to smash the contents of
2153          * the redzone and userword anyhow, so switch them off.
2154          */
2155         if (IS_ENABLED(CONFIG_PAGE_POISONING) &&
2156                 (cachep->flags & SLAB_POISON) &&
2157                 is_debug_pagealloc_cache(cachep))
2158                 cachep->flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2159 #endif
2160 
2161         if (OFF_SLAB(cachep)) {
2162                 cachep->freelist_cache =
2163                         kmalloc_slab(cachep->freelist_size, 0u);
2164         }
2165 
2166         err = setup_cpu_cache(cachep, gfp);
2167         if (err) {
2168                 __kmem_cache_release(cachep);
2169                 return err;
2170         }
2171 
2172         return 0;
2173 }
2174 
2175 #if DEBUG
2176 static void check_irq_off(void)
2177 {
2178         BUG_ON(!irqs_disabled());
2179 }
2180 
2181 static void check_irq_on(void)
2182 {
2183         BUG_ON(irqs_disabled());
2184 }
2185 
2186 static void check_mutex_acquired(void)
2187 {
2188         BUG_ON(!mutex_is_locked(&slab_mutex));
2189 }
2190 
2191 static void check_spinlock_acquired(struct kmem_cache *cachep)
2192 {
2193 #ifdef CONFIG_SMP
2194         check_irq_off();
2195         assert_spin_locked(&get_node(cachep, numa_mem_id())->list_lock);
2196 #endif
2197 }
2198 
2199 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2200 {
2201 #ifdef CONFIG_SMP
2202         check_irq_off();
2203         assert_spin_locked(&get_node(cachep, node)->list_lock);
2204 #endif
2205 }
2206 
2207 #else
2208 #define check_irq_off() do { } while(0)
2209 #define check_irq_on()  do { } while(0)
2210 #define check_mutex_acquired()  do { } while(0)
2211 #define check_spinlock_acquired(x) do { } while(0)
2212 #define check_spinlock_acquired_node(x, y) do { } while(0)
2213 #endif
2214 
2215 static void drain_array_locked(struct kmem_cache *cachep, struct array_cache *ac,
2216                                 int node, bool free_all, struct list_head *list)
2217 {
2218         int tofree;
2219 
2220         if (!ac || !ac->avail)
2221                 return;
2222 
2223         tofree = free_all ? ac->avail : (ac->limit + 4) / 5;
2224         if (tofree > ac->avail)
2225                 tofree = (ac->avail + 1) / 2;
2226 
2227         free_block(cachep, ac->entry, tofree, node, list);
2228         ac->avail -= tofree;
2229         memmove(ac->entry, &(ac->entry[tofree]), sizeof(void *) * ac->avail);
2230 }
2231 
2232 static void do_drain(void *arg)
2233 {
2234         struct kmem_cache *cachep = arg;
2235         struct array_cache *ac;
2236         int node = numa_mem_id();
2237         struct kmem_cache_node *n;
2238         LIST_HEAD(list);
2239 
2240         check_irq_off();
2241         ac = cpu_cache_get(cachep);
2242         n = get_node(cachep, node);
2243         spin_lock(&n->list_lock);
2244         free_block(cachep, ac->entry, ac->avail, node, &list);
2245         spin_unlock(&n->list_lock);
2246         slabs_destroy(cachep, &list);
2247         ac->avail = 0;
2248 }
2249 
2250 static void drain_cpu_caches(struct kmem_cache *cachep)
2251 {
2252         struct kmem_cache_node *n;
2253         int node;
2254         LIST_HEAD(list);
2255 
2256         on_each_cpu(do_drain, cachep, 1);
2257         check_irq_on();
2258         for_each_kmem_cache_node(cachep, node, n)
2259                 if (n->alien)
2260                         drain_alien_cache(cachep, n->alien);
2261 
2262         for_each_kmem_cache_node(cachep, node, n) {
2263                 spin_lock_irq(&n->list_lock);
2264                 drain_array_locked(cachep, n->shared, node, true, &list);
2265                 spin_unlock_irq(&n->list_lock);
2266 
2267                 slabs_destroy(cachep, &list);
2268         }
2269 }
2270 
2271 /*
2272  * Remove slabs from the list of free slabs.
2273  * Specify the number of slabs to drain in tofree.
2274  *
2275  * Returns the actual number of slabs released.
2276  */
2277 static int drain_freelist(struct kmem_cache *cache,
2278                         struct kmem_cache_node *n, int tofree)
2279 {
2280         struct list_head *p;
2281         int nr_freed;
2282         struct page *page;
2283 
2284         nr_freed = 0;
2285         while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
2286 
2287                 spin_lock_irq(&n->list_lock);
2288                 p = n->slabs_free.prev;
2289                 if (p == &n->slabs_free) {
2290                         spin_unlock_irq(&n->list_lock);
2291                         goto out;
2292                 }
2293 
2294                 page = list_entry(p, struct page, lru);
2295                 list_del(&page->lru);
2296                 n->free_slabs--;
2297                 n->total_slabs--;
2298                 /*
2299                  * Safe to drop the lock. The slab is no longer linked
2300                  * to the cache.
2301                  */
2302                 n->free_objects -= cache->num;
2303                 spin_unlock_irq(&n->list_lock);
2304                 slab_destroy(cache, page);
2305                 nr_freed++;
2306         }
2307 out:
2308         return nr_freed;
2309 }
2310 
2311 int __kmem_cache_shrink(struct kmem_cache *cachep)
2312 {
2313         int ret = 0;
2314         int node;
2315         struct kmem_cache_node *n;
2316 
2317         drain_cpu_caches(cachep);
2318 
2319         check_irq_on();
2320         for_each_kmem_cache_node(cachep, node, n) {
2321                 drain_freelist(cachep, n, INT_MAX);
2322 
2323                 ret += !list_empty(&n->slabs_full) ||
2324                         !list_empty(&n->slabs_partial);
2325         }
2326         return (ret ? 1 : 0);
2327 }
2328 
2329 #ifdef CONFIG_MEMCG
2330 void __kmemcg_cache_deactivate(struct kmem_cache *cachep)
2331 {
2332         __kmem_cache_shrink(cachep);
2333 }
2334 #endif
2335 
2336 int __kmem_cache_shutdown(struct kmem_cache *cachep)
2337 {
2338         return __kmem_cache_shrink(cachep);
2339 }
2340 
2341 void __kmem_cache_release(struct kmem_cache *cachep)
2342 {
2343         int i;
2344         struct kmem_cache_node *n;
2345 
2346         cache_random_seq_destroy(cachep);
2347 
2348         free_percpu(cachep->cpu_cache);
2349 
2350         /* NUMA: free the node structures */
2351         for_each_kmem_cache_node(cachep, i, n) {
2352                 kfree(n->shared);
2353                 free_alien_cache(n->alien);
2354                 kfree(n);
2355                 cachep->node[i] = NULL;
2356         }
2357 }
2358 
2359 /*
2360  * Get the memory for a slab management obj.
2361  *
2362  * For a slab cache when the slab descriptor is off-slab, the
2363  * slab descriptor can't come from the same cache which is being created,
2364  * Because if it is the case, that means we defer the creation of
2365  * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2366  * And we eventually call down to __kmem_cache_create(), which
2367  * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2368  * This is a "chicken-and-egg" problem.
2369  *
2370  * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2371  * which are all initialized during kmem_cache_init().
2372  */
2373 static void *alloc_slabmgmt(struct kmem_cache *cachep,
2374                                    struct page *page, int colour_off,
2375                                    gfp_t local_flags, int nodeid)
2376 {
2377         void *freelist;
2378         void *addr = page_address(page);
2379 
2380         page->s_mem = addr + colour_off;
2381         page->active = 0;
2382 
2383         if (OBJFREELIST_SLAB(cachep))
2384                 freelist = NULL;
2385         else if (OFF_SLAB(cachep)) {
2386                 /* Slab management obj is off-slab. */
2387                 freelist = kmem_cache_alloc_node(cachep->freelist_cache,
2388                                               local_flags, nodeid);
2389                 if (!freelist)
2390                         return NULL;
2391         } else {
2392                 /* We will use last bytes at the slab for freelist */
2393                 freelist = addr + (PAGE_SIZE << cachep->gfporder) -
2394                                 cachep->freelist_size;
2395         }
2396 
2397         return freelist;
2398 }
2399 
2400 static inline freelist_idx_t get_free_obj(struct page *page, unsigned int idx)
2401 {
2402         return ((freelist_idx_t *)page->freelist)[idx];
2403 }
2404 
2405 static inline void set_free_obj(struct page *page,
2406                                         unsigned int idx, freelist_idx_t val)
2407 {
2408         ((freelist_idx_t *)(page->freelist))[idx] = val;
2409 }
2410 
2411 static void cache_init_objs_debug(struct kmem_cache *cachep, struct page *page)
2412 {
2413 #if DEBUG
2414         int i;
2415 
2416         for (i = 0; i < cachep->num; i++) {
2417                 void *objp = index_to_obj(cachep, page, i);
2418 
2419                 if (cachep->flags & SLAB_STORE_USER)
2420                         *dbg_userword(cachep, objp) = NULL;
2421 
2422                 if (cachep->flags & SLAB_RED_ZONE) {
2423                         *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2424                         *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2425                 }
2426                 /*
2427                  * Constructors are not allowed to allocate memory from the same
2428                  * cache which they are a constructor for.  Otherwise, deadlock.
2429                  * They must also be threaded.
2430                  */
2431                 if (cachep->ctor && !(cachep->flags & SLAB_POISON)) {
2432                         kasan_unpoison_object_data(cachep,
2433                                                    objp + obj_offset(cachep));
2434                         cachep->ctor(objp + obj_offset(cachep));
2435                         kasan_poison_object_data(
2436                                 cachep, objp + obj_offset(cachep));
2437                 }
2438 
2439                 if (cachep->flags & SLAB_RED_ZONE) {
2440                         if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2441                                 slab_error(cachep, "constructor overwrote the end of an object");
2442                         if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2443                                 slab_error(cachep, "constructor overwrote the start of an object");
2444                 }
2445                 /* need to poison the objs? */
2446                 if (cachep->flags & SLAB_POISON) {
2447                         poison_obj(cachep, objp, POISON_FREE);
2448                         slab_kernel_map(cachep, objp, 0, 0);
2449                 }
2450         }
2451 #endif
2452 }
2453 
2454 #ifdef CONFIG_SLAB_FREELIST_RANDOM
2455 /* Hold information during a freelist initialization */
2456 union freelist_init_state {
2457         struct {
2458                 unsigned int pos;
2459                 unsigned int *list;
2460                 unsigned int count;
2461         };
2462         struct rnd_state rnd_state;
2463 };
2464 
2465 /*
2466  * Initialize the state based on the randomization methode available.
2467  * return true if the pre-computed list is available, false otherwize.
2468  */
2469 static bool freelist_state_initialize(union freelist_init_state *state,
2470                                 struct kmem_cache *cachep,
2471                                 unsigned int count)
2472 {
2473         bool ret;
2474         unsigned int rand;
2475 
2476         /* Use best entropy available to define a random shift */
2477         rand = get_random_int();
2478 
2479         /* Use a random state if the pre-computed list is not available */
2480         if (!cachep->random_seq) {
2481                 prandom_seed_state(&state->rnd_state, rand);
2482                 ret = false;
2483         } else {
2484                 state->list = cachep->random_seq;
2485                 state->count = count;
2486                 state->pos = rand % count;
2487                 ret = true;
2488         }
2489         return ret;
2490 }
2491 
2492 /* Get the next entry on the list and randomize it using a random shift */
2493 static freelist_idx_t next_random_slot(union freelist_init_state *state)
2494 {
2495         if (state->pos >= state->count)
2496                 state->pos = 0;
2497         return state->list[state->pos++];
2498 }
2499 
2500 /* Swap two freelist entries */
2501 static void swap_free_obj(struct page *page, unsigned int a, unsigned int b)
2502 {
2503         swap(((freelist_idx_t *)page->freelist)[a],
2504                 ((freelist_idx_t *)page->freelist)[b]);
2505 }
2506 
2507 /*
2508  * Shuffle the freelist initialization state based on pre-computed lists.
2509  * return true if the list was successfully shuffled, false otherwise.
2510  */
2511 static bool shuffle_freelist(struct kmem_cache *cachep, struct page *page)
2512 {
2513         unsigned int objfreelist = 0, i, rand, count = cachep->num;
2514         union freelist_init_state state;
2515         bool precomputed;
2516 
2517         if (count < 2)
2518                 return false;
2519 
2520         precomputed = freelist_state_initialize(&state, cachep, count);
2521 
2522         /* Take a random entry as the objfreelist */
2523         if (OBJFREELIST_SLAB(cachep)) {
2524                 if (!precomputed)
2525                         objfreelist = count - 1;
2526                 else
2527                         objfreelist = next_random_slot(&state);
2528                 page->freelist = index_to_obj(cachep, page, objfreelist) +
2529                                                 obj_offset(cachep);
2530                 count--;
2531         }
2532 
2533         /*
2534          * On early boot, generate the list dynamically.
2535          * Later use a pre-computed list for speed.
2536          */
2537         if (!precomputed) {
2538                 for (i = 0; i < count; i++)
2539                         set_free_obj(page, i, i);
2540 
2541                 /* Fisher-Yates shuffle */
2542                 for (i = count - 1; i > 0; i--) {
2543                         rand = prandom_u32_state(&state.rnd_state);
2544                         rand %= (i + 1);
2545                         swap_free_obj(page, i, rand);
2546                 }
2547         } else {
2548                 for (i = 0; i < count; i++)
2549                         set_free_obj(page, i, next_random_slot(&state));
2550         }
2551 
2552         if (OBJFREELIST_SLAB(cachep))
2553                 set_free_obj(page, cachep->num - 1, objfreelist);
2554 
2555         return true;
2556 }
2557 #else
2558 static inline bool shuffle_freelist(struct kmem_cache *cachep,
2559                                 struct page *page)
2560 {
2561         return false;
2562 }
2563 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
2564 
2565 static void cache_init_objs(struct kmem_cache *cachep,
2566                             struct page *page)
2567 {
2568         int i;
2569         void *objp;
2570         bool shuffled;
2571 
2572         cache_init_objs_debug(cachep, page);
2573 
2574         /* Try to randomize the freelist if enabled */
2575         shuffled = shuffle_freelist(cachep, page);
2576 
2577         if (!shuffled && OBJFREELIST_SLAB(cachep)) {
2578                 page->freelist = index_to_obj(cachep, page, cachep->num - 1) +
2579                                                 obj_offset(cachep);
2580         }
2581 
2582         for (i = 0; i < cachep->num; i++) {
2583                 objp = index_to_obj(cachep, page, i);
2584                 kasan_init_slab_obj(cachep, objp);
2585 
2586                 /* constructor could break poison info */
2587                 if (DEBUG == 0 && cachep->ctor) {
2588                         kasan_unpoison_object_data(cachep, objp);
2589                         cachep->ctor(objp);
2590                         kasan_poison_object_data(cachep, objp);
2591                 }
2592 
2593                 if (!shuffled)
2594                         set_free_obj(page, i, i);
2595         }
2596 }
2597 
2598 static void *slab_get_obj(struct kmem_cache *cachep, struct page *page)
2599 {
2600         void *objp;
2601 
2602         objp = index_to_obj(cachep, page, get_free_obj(page, page->active));
2603         page->active++;
2604 
2605 #if DEBUG
2606         if (cachep->flags & SLAB_STORE_USER)
2607                 set_store_user_dirty(cachep);
2608 #endif
2609 
2610         return objp;
2611 }
2612 
2613 static void slab_put_obj(struct kmem_cache *cachep,
2614                         struct page *page, void *objp)
2615 {
2616         unsigned int objnr = obj_to_index(cachep, page, objp);
2617 #if DEBUG
2618         unsigned int i;
2619 
2620         /* Verify double free bug */
2621         for (i = page->active; i < cachep->num; i++) {
2622                 if (get_free_obj(page, i) == objnr) {
2623                         pr_err("slab: double free detected in cache '%s', objp %p\n",
2624                                cachep->name, objp);
2625                         BUG();
2626                 }
2627         }
2628 #endif
2629         page->active--;
2630         if (!page->freelist)
2631                 page->freelist = objp + obj_offset(cachep);
2632 
2633         set_free_obj(page, page->active, objnr);
2634 }
2635 
2636 /*
2637  * Map pages beginning at addr to the given cache and slab. This is required
2638  * for the slab allocator to be able to lookup the cache and slab of a
2639  * virtual address for kfree, ksize, and slab debugging.
2640  */
2641 static void slab_map_pages(struct kmem_cache *cache, struct page *page,
2642                            void *freelist)
2643 {
2644         page->slab_cache = cache;
2645         page->freelist = freelist;
2646 }
2647 
2648 /*
2649  * Grow (by 1) the number of slabs within a cache.  This is called by
2650  * kmem_cache_alloc() when there are no active objs left in a cache.
2651  */
2652 static struct page *cache_grow_begin(struct kmem_cache *cachep,
2653                                 gfp_t flags, int nodeid)
2654 {
2655         void *freelist;
2656         size_t offset;
2657         gfp_t local_flags;
2658         int page_node;
2659         struct kmem_cache_node *n;
2660         struct page *page;
2661 
2662         /*
2663          * Be lazy and only check for valid flags here,  keeping it out of the
2664          * critical path in kmem_cache_alloc().
2665          */
2666         if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
2667                 gfp_t invalid_mask = flags & GFP_SLAB_BUG_MASK;
2668                 flags &= ~GFP_SLAB_BUG_MASK;
2669                 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n",
2670                                 invalid_mask, &invalid_mask, flags, &flags);
2671                 dump_stack();
2672         }
2673         local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2674 
2675         check_irq_off();
2676         if (gfpflags_allow_blocking(local_flags))
2677                 local_irq_enable();
2678 
2679         /*
2680          * Get mem for the objs.  Attempt to allocate a physical page from
2681          * 'nodeid'.
2682          */
2683         page = kmem_getpages(cachep, local_flags, nodeid);
2684         if (!page)
2685                 goto failed;
2686 
2687         page_node = page_to_nid(page);
2688         n = get_node(cachep, page_node);
2689 
2690         /* Get colour for the slab, and cal the next value. */
2691         n->colour_next++;
2692         if (n->colour_next >= cachep->colour)
2693                 n->colour_next = 0;
2694 
2695         offset = n->colour_next;
2696         if (offset >= cachep->colour)
2697                 offset = 0;
2698 
2699         offset *= cachep->colour_off;
2700 
2701         /* Get slab management. */
2702         freelist = alloc_slabmgmt(cachep, page, offset,
2703                         local_flags & ~GFP_CONSTRAINT_MASK, page_node);
2704         if (OFF_SLAB(cachep) && !freelist)
2705                 goto opps1;
2706 
2707         slab_map_pages(cachep, page, freelist);
2708 
2709         kasan_poison_slab(page);
2710         cache_init_objs(cachep, page);
2711 
2712         if (gfpflags_allow_blocking(local_flags))
2713                 local_irq_disable();
2714 
2715         return page;
2716 
2717 opps1:
2718         kmem_freepages(cachep, page);
2719 failed:
2720         if (gfpflags_allow_blocking(local_flags))
2721                 local_irq_disable();
2722         return NULL;
2723 }
2724 
2725 static void cache_grow_end(struct kmem_cache *cachep, struct page *page)
2726 {
2727         struct kmem_cache_node *n;
2728         void *list = NULL;
2729 
2730         check_irq_off();
2731 
2732         if (!page)
2733                 return;
2734 
2735         INIT_LIST_HEAD(&page->lru);
2736         n = get_node(cachep, page_to_nid(page));
2737 
2738         spin_lock(&n->list_lock);
2739         n->total_slabs++;
2740         if (!page->active) {
2741                 list_add_tail(&page->lru, &(n->slabs_free));
2742                 n->free_slabs++;
2743         } else
2744                 fixup_slab_list(cachep, n, page, &list);
2745 
2746         STATS_INC_GROWN(cachep);
2747         n->free_objects += cachep->num - page->active;
2748         spin_unlock(&n->list_lock);
2749 
2750         fixup_objfreelist_debug(cachep, &list);
2751 }
2752 
2753 #if DEBUG
2754 
2755 /*
2756  * Perform extra freeing checks:
2757  * - detect bad pointers.
2758  * - POISON/RED_ZONE checking
2759  */
2760 static void kfree_debugcheck(const void *objp)
2761 {
2762         if (!virt_addr_valid(objp)) {
2763                 pr_err("kfree_debugcheck: out of range ptr %lxh\n",
2764                        (unsigned long)objp);
2765                 BUG();
2766         }
2767 }
2768 
2769 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2770 {
2771         unsigned long long redzone1, redzone2;
2772 
2773         redzone1 = *dbg_redzone1(cache, obj);
2774         redzone2 = *dbg_redzone2(cache, obj);
2775 
2776         /*
2777          * Redzone is ok.
2778          */
2779         if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2780                 return;
2781 
2782         if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2783                 slab_error(cache, "double free detected");
2784         else
2785                 slab_error(cache, "memory outside object was overwritten");
2786 
2787         pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2788                obj, redzone1, redzone2);
2789 }
2790 
2791 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2792                                    unsigned long caller)
2793 {
2794         unsigned int objnr;
2795         struct page *page;
2796 
2797         BUG_ON(virt_to_cache(objp) != cachep);
2798 
2799         objp -= obj_offset(cachep);
2800         kfree_debugcheck(objp);
2801         page = virt_to_head_page(objp);
2802 
2803         if (cachep->flags & SLAB_RED_ZONE) {
2804                 verify_redzone_free(cachep, objp);
2805                 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2806                 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2807         }
2808         if (cachep->flags & SLAB_STORE_USER) {
2809                 set_store_user_dirty(cachep);
2810                 *dbg_userword(cachep, objp) = (void *)caller;
2811         }
2812 
2813         objnr = obj_to_index(cachep, page, objp);
2814 
2815         BUG_ON(objnr >= cachep->num);
2816         BUG_ON(objp != index_to_obj(cachep, page, objnr));
2817 
2818         if (cachep->flags & SLAB_POISON) {
2819                 poison_obj(cachep, objp, POISON_FREE);
2820                 slab_kernel_map(cachep, objp, 0, caller);
2821         }
2822         return objp;
2823 }
2824 
2825 #else
2826 #define kfree_debugcheck(x) do { } while(0)
2827 #define cache_free_debugcheck(x,objp,z) (objp)
2828 #endif
2829 
2830 static inline void fixup_objfreelist_debug(struct kmem_cache *cachep,
2831                                                 void **list)
2832 {
2833 #if DEBUG
2834         void *next = *list;
2835         void *objp;
2836 
2837         while (next) {
2838                 objp = next - obj_offset(cachep);
2839                 next = *(void **)next;
2840                 poison_obj(cachep, objp, POISON_FREE);
2841         }
2842 #endif
2843 }
2844 
2845 static inline void fixup_slab_list(struct kmem_cache *cachep,
2846                                 struct kmem_cache_node *n, struct page *page,
2847                                 void **list)
2848 {
2849         /* move slabp to correct slabp list: */
2850         list_del(&page->lru);
2851         if (page->active == cachep->num) {
2852                 list_add(&page->lru, &n->slabs_full);
2853                 if (OBJFREELIST_SLAB(cachep)) {
2854 #if DEBUG
2855                         /* Poisoning will be done without holding the lock */
2856                         if (cachep->flags & SLAB_POISON) {
2857                                 void **objp = page->freelist;
2858 
2859                                 *objp = *list;
2860                                 *list = objp;
2861                         }
2862 #endif
2863                         page->freelist = NULL;
2864                 }
2865         } else
2866                 list_add(&page->lru, &n->slabs_partial);
2867 }
2868 
2869 /* Try to find non-pfmemalloc slab if needed */
2870 static noinline struct page *get_valid_first_slab(struct kmem_cache_node *n,
2871                                         struct page *page, bool pfmemalloc)
2872 {
2873         if (!page)
2874                 return NULL;
2875 
2876         if (pfmemalloc)
2877                 return page;
2878 
2879         if (!PageSlabPfmemalloc(page))
2880                 return page;
2881 
2882         /* No need to keep pfmemalloc slab if we have enough free objects */
2883         if (n->free_objects > n->free_limit) {
2884                 ClearPageSlabPfmemalloc(page);
2885                 return page;
2886         }
2887 
2888         /* Move pfmemalloc slab to the end of list to speed up next search */
2889         list_del(&page->lru);
2890         if (!page->active) {
2891                 list_add_tail(&page->lru, &n->slabs_free);
2892                 n->free_slabs++;
2893         } else
2894                 list_add_tail(&page->lru, &n->slabs_partial);
2895 
2896         list_for_each_entry(page, &n->slabs_partial, lru) {
2897                 if (!PageSlabPfmemalloc(page))
2898                         return page;
2899         }
2900 
2901         n->free_touched = 1;
2902         list_for_each_entry(page, &n->slabs_free, lru) {
2903                 if (!PageSlabPfmemalloc(page)) {
2904                         n->free_slabs--;
2905                         return page;
2906                 }
2907         }
2908 
2909         return NULL;
2910 }
2911 
2912 static struct page *get_first_slab(struct kmem_cache_node *n, bool pfmemalloc)
2913 {
2914         struct page *page;
2915 
2916         assert_spin_locked(&n->list_lock);
2917         page = list_first_entry_or_null(&n->slabs_partial, struct page, lru);
2918         if (!page) {
2919                 n->free_touched = 1;
2920                 page = list_first_entry_or_null(&n->slabs_free, struct page,
2921                                                 lru);
2922                 if (page)
2923                         n->free_slabs--;
2924         }
2925 
2926         if (sk_memalloc_socks())
2927                 page = get_valid_first_slab(n, page, pfmemalloc);
2928 
2929         return page;
2930 }
2931 
2932 static noinline void *cache_alloc_pfmemalloc(struct kmem_cache *cachep,
2933                                 struct kmem_cache_node *n, gfp_t flags)
2934 {
2935         struct page *page;
2936         void *obj;
2937         void *list = NULL;
2938 
2939         if (!gfp_pfmemalloc_allowed(flags))
2940                 return NULL;
2941 
2942         spin_lock(&n->list_lock);
2943         page = get_first_slab(n, true);
2944         if (!page) {
2945                 spin_unlock(&n->list_lock);
2946                 return NULL;
2947         }
2948 
2949         obj = slab_get_obj(cachep, page);
2950         n->free_objects--;
2951 
2952         fixup_slab_list(cachep, n, page, &list);
2953 
2954         spin_unlock(&n->list_lock);
2955         fixup_objfreelist_debug(cachep, &list);
2956 
2957         return obj;
2958 }
2959 
2960 /*
2961  * Slab list should be fixed up by fixup_slab_list() for existing slab
2962  * or cache_grow_end() for new slab
2963  */
2964 static __always_inline int alloc_block(struct kmem_cache *cachep,
2965                 struct array_cache *ac, struct page *page, int batchcount)
2966 {
2967         /*
2968          * There must be at least one object available for
2969          * allocation.
2970          */
2971         BUG_ON(page->active >= cachep->num);
2972 
2973         while (page->active < cachep->num && batchcount--) {
2974                 STATS_INC_ALLOCED(cachep);
2975                 STATS_INC_ACTIVE(cachep);
2976                 STATS_SET_HIGH(cachep);
2977 
2978                 ac->entry[ac->avail++] = slab_get_obj(cachep, page);
2979         }
2980 
2981         return batchcount;
2982 }
2983 
2984 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
2985 {
2986         int batchcount;
2987         struct kmem_cache_node *n;
2988         struct array_cache *ac, *shared;
2989         int node;
2990         void *list = NULL;
2991         struct page *page;
2992 
2993         check_irq_off();
2994         node = numa_mem_id();
2995 
2996         ac = cpu_cache_get(cachep);
2997         batchcount = ac->batchcount;
2998         if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2999                 /*
3000                  * If there was little recent activity on this cache, then
3001                  * perform only a partial refill.  Otherwise we could generate
3002                  * refill bouncing.
3003                  */
3004                 batchcount = BATCHREFILL_LIMIT;
3005         }
3006         n = get_node(cachep, node);
3007 
3008         BUG_ON(ac->avail > 0 || !n);
3009         shared = READ_ONCE(n->shared);
3010         if (!n->free_objects && (!shared || !shared->avail))
3011                 goto direct_grow;
3012 
3013         spin_lock(&n->list_lock);
3014         shared = READ_ONCE(n->shared);
3015 
3016         /* See if we can refill from the shared array */
3017         if (shared && transfer_objects(ac, shared, batchcount)) {
3018                 shared->touched = 1;
3019                 goto alloc_done;
3020         }
3021 
3022         while (batchcount > 0) {
3023                 /* Get slab alloc is to come from. */
3024                 page = get_first_slab(n, false);
3025                 if (!page)
3026                         goto must_grow;
3027 
3028                 check_spinlock_acquired(cachep);
3029 
3030                 batchcount = alloc_block(cachep, ac, page, batchcount);
3031                 fixup_slab_list(cachep, n, page, &list);
3032         }
3033 
3034 must_grow:
3035         n->free_objects -= ac->avail;
3036 alloc_done:
3037         spin_unlock(&n->list_lock);
3038         fixup_objfreelist_debug(cachep, &list);
3039 
3040 direct_grow:
3041         if (unlikely(!ac->avail)) {
3042                 /* Check if we can use obj in pfmemalloc slab */
3043                 if (sk_memalloc_socks()) {
3044                         void *obj = cache_alloc_pfmemalloc(cachep, n, flags);
3045 
3046                         if (obj)
3047                                 return obj;
3048                 }
3049 
3050                 page = cache_grow_begin(cachep, gfp_exact_node(flags), node);
3051 
3052                 /*
3053                  * cache_grow_begin() can reenable interrupts,
3054                  * then ac could change.
3055                  */
3056                 ac = cpu_cache_get(cachep);
3057                 if (!ac->avail && page)
3058                         alloc_block(cachep, ac, page, batchcount);
3059                 cache_grow_end(cachep, page);
3060 
3061                 if (!ac->avail)
3062                         return NULL;
3063         }
3064         ac->touched = 1;
3065 
3066         return ac->entry[--ac->avail];
3067 }
3068 
3069 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3070                                                 gfp_t flags)
3071 {
3072         might_sleep_if(gfpflags_allow_blocking(flags));
3073 }
3074 
3075 #if DEBUG
3076 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3077                                 gfp_t flags, void *objp, unsigned long caller)
3078 {
3079         if (!objp)
3080                 return objp;
3081         if (cachep->flags & SLAB_POISON) {
3082                 check_poison_obj(cachep, objp);
3083                 slab_kernel_map(cachep, objp, 1, 0);
3084                 poison_obj(cachep, objp, POISON_INUSE);
3085         }
3086         if (cachep->flags & SLAB_STORE_USER)
3087                 *dbg_userword(cachep, objp) = (void *)caller;
3088 
3089         if (cachep->flags & SLAB_RED_ZONE) {
3090                 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3091                                 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3092                         slab_error(cachep, "double free, or memory outside object was overwritten");
3093                         pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3094                                objp, *dbg_redzone1(cachep, objp),
3095                                *dbg_redzone2(cachep, objp));
3096                 }
3097                 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3098                 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3099         }
3100 
3101         objp += obj_offset(cachep);
3102         if (cachep->ctor && cachep->flags & SLAB_POISON)
3103                 cachep->ctor(objp);
3104         if (ARCH_SLAB_MINALIGN &&
3105             ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
3106                 pr_err("0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3107                        objp, (int)ARCH_SLAB_MINALIGN);
3108         }
3109         return objp;
3110 }
3111 #else
3112 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3113 #endif
3114 
3115 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3116 {
3117         void *objp;
3118         struct array_cache *ac;
3119 
3120         check_irq_off();
3121 
3122         ac = cpu_cache_get(cachep);
3123         if (likely(ac->avail)) {
3124                 ac->touched = 1;
3125                 objp = ac->entry[--ac->avail];
3126 
3127                 STATS_INC_ALLOCHIT(cachep);
3128                 goto out;
3129         }
3130 
3131         STATS_INC_ALLOCMISS(cachep);
3132         objp = cache_alloc_refill(cachep, flags);
3133         /*
3134          * the 'ac' may be updated by cache_alloc_refill(),
3135          * and kmemleak_erase() requires its correct value.
3136          */
3137         ac = cpu_cache_get(cachep);
3138 
3139 out:
3140         /*
3141          * To avoid a false negative, if an object that is in one of the
3142          * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3143          * treat the array pointers as a reference to the object.
3144          */
3145         if (objp)
3146                 kmemleak_erase(&ac->entry[ac->avail]);
3147         return objp;
3148 }
3149 
3150 #ifdef CONFIG_NUMA
3151 /*
3152  * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
3153  *
3154  * If we are in_interrupt, then process context, including cpusets and
3155  * mempolicy, may not apply and should not be used for allocation policy.
3156  */
3157 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3158 {
3159         int nid_alloc, nid_here;
3160 
3161         if (in_interrupt() || (flags & __GFP_THISNODE))
3162                 return NULL;
3163         nid_alloc = nid_here = numa_mem_id();
3164         if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3165                 nid_alloc = cpuset_slab_spread_node();
3166         else if (current->mempolicy)
3167                 nid_alloc = mempolicy_slab_node();
3168         if (nid_alloc != nid_here)
3169                 return ____cache_alloc_node(cachep, flags, nid_alloc);
3170         return NULL;
3171 }
3172 
3173 /*
3174  * Fallback function if there was no memory available and no objects on a
3175  * certain node and fall back is permitted. First we scan all the
3176  * available node for available objects. If that fails then we
3177  * perform an allocation without specifying a node. This allows the page
3178  * allocator to do its reclaim / fallback magic. We then insert the
3179  * slab into the proper nodelist and then allocate from it.
3180  */
3181 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3182 {
3183         struct zonelist *zonelist;
3184         struct zoneref *z;
3185         struct zone *zone;
3186         enum zone_type high_zoneidx = gfp_zone(flags);
3187         void *obj = NULL;
3188         struct page *page;
3189         int nid;
3190         unsigned int cpuset_mems_cookie;
3191 
3192         if (flags & __GFP_THISNODE)
3193                 return NULL;
3194 
3195 retry_cpuset:
3196         cpuset_mems_cookie = read_mems_allowed_begin();
3197         zonelist = node_zonelist(mempolicy_slab_node(), flags);
3198 
3199 retry:
3200         /*
3201          * Look through allowed nodes for objects available
3202          * from existing per node queues.
3203          */
3204         for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3205                 nid = zone_to_nid(zone);
3206 
3207                 if (cpuset_zone_allowed(zone, flags) &&
3208                         get_node(cache, nid) &&
3209                         get_node(cache, nid)->free_objects) {
3210                                 obj = ____cache_alloc_node(cache,
3211                                         gfp_exact_node(flags), nid);
3212                                 if (obj)
3213                                         break;
3214                 }
3215         }
3216 
3217         if (!obj) {
3218                 /*
3219                  * This allocation will be performed within the constraints
3220                  * of the current cpuset / memory policy requirements.
3221                  * We may trigger various forms of reclaim on the allowed
3222                  * set and go into memory reserves if necessary.
3223                  */
3224                 page = cache_grow_begin(cache, flags, numa_mem_id());
3225                 cache_grow_end(cache, page);
3226                 if (page) {
3227                         nid = page_to_nid(page);
3228                         obj = ____cache_alloc_node(cache,
3229                                 gfp_exact_node(flags), nid);
3230 
3231                         /*
3232                          * Another processor may allocate the objects in
3233                          * the slab since we are not holding any locks.
3234                          */
3235                         if (!obj)
3236                                 goto retry;
3237                 }
3238         }
3239 
3240         if (unlikely(!obj && read_mems_allowed_retry(cpuset_mems_cookie)))
3241                 goto retry_cpuset;
3242         return obj;
3243 }
3244 
3245 /*
3246  * A interface to enable slab creation on nodeid
3247  */
3248 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3249                                 int nodeid)
3250 {
3251         struct page *page;
3252         struct kmem_cache_node *n;
3253         void *obj = NULL;
3254         void *list = NULL;
3255 
3256         VM_BUG_ON(nodeid < 0 || nodeid >= MAX_NUMNODES);
3257         n = get_node(cachep, nodeid);
3258         BUG_ON(!n);
3259 
3260         check_irq_off();
3261         spin_lock(&n->list_lock);
3262         page = get_first_slab(n, false);
3263         if (!page)
3264                 goto must_grow;
3265 
3266         check_spinlock_acquired_node(cachep, nodeid);
3267 
3268         STATS_INC_NODEALLOCS(cachep);
3269         STATS_INC_ACTIVE(cachep);
3270         STATS_SET_HIGH(cachep);
3271 
3272         BUG_ON(page->active == cachep->num);
3273 
3274         obj = slab_get_obj(cachep, page);
3275         n->free_objects--;
3276 
3277         fixup_slab_list(cachep, n, page, &list);
3278 
3279         spin_unlock(&n->list_lock);
3280         fixup_objfreelist_debug(cachep, &list);
3281         return obj;
3282 
3283 must_grow:
3284         spin_unlock(&n->list_lock);
3285         page = cache_grow_begin(cachep, gfp_exact_node(flags), nodeid);
3286         if (page) {
3287                 /* This slab isn't counted yet so don't update free_objects */
3288                 obj = slab_get_obj(cachep, page);
3289         }
3290         cache_grow_end(cachep, page);
3291 
3292         return obj ? obj : fallback_alloc(cachep, flags);
3293 }
3294 
3295 static __always_inline void *
3296 slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3297                    unsigned long caller)
3298 {
3299         unsigned long save_flags;
3300         void *ptr;
3301         int slab_node = numa_mem_id();
3302 
3303         flags &= gfp_allowed_mask;
3304         cachep = slab_pre_alloc_hook(cachep, flags);
3305         if (unlikely(!cachep))
3306                 return NULL;
3307 
3308         cache_alloc_debugcheck_before(cachep, flags);
3309         local_irq_save(save_flags);
3310 
3311         if (nodeid == NUMA_NO_NODE)
3312                 nodeid = slab_node;
3313 
3314         if (unlikely(!get_node(cachep, nodeid))) {
3315                 /* Node not bootstrapped yet */
3316                 ptr = fallback_alloc(cachep, flags);
3317                 goto out;
3318         }
3319 
3320         if (nodeid == slab_node) {
3321                 /*
3322                  * Use the locally cached objects if possible.
3323                  * However ____cache_alloc does not allow fallback
3324                  * to other nodes. It may fail while we still have
3325                  * objects on other nodes available.
3326                  */
3327                 ptr = ____cache_alloc(cachep, flags);
3328                 if (ptr)
3329                         goto out;
3330         }
3331         /* ___cache_alloc_node can fall back to other nodes */
3332         ptr = ____cache_alloc_node(cachep, flags, nodeid);
3333   out:
3334         local_irq_restore(save_flags);
3335         ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3336 
3337         if (unlikely(flags & __GFP_ZERO) && ptr)
3338                 memset(ptr, 0, cachep->object_size);
3339 
3340         slab_post_alloc_hook(cachep, flags, 1, &ptr);
3341         return ptr;
3342 }
3343 
3344 static __always_inline void *
3345 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3346 {
3347         void *objp;
3348 
3349         if (current->mempolicy || cpuset_do_slab_mem_spread()) {
3350                 objp = alternate_node_alloc(cache, flags);
3351                 if (objp)
3352                         goto out;
3353         }
3354         objp = ____cache_alloc(cache, flags);
3355 
3356         /*
3357          * We may just have run out of memory on the local node.
3358          * ____cache_alloc_node() knows how to locate memory on other nodes
3359          */
3360         if (!objp)
3361                 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3362 
3363   out:
3364         return objp;
3365 }
3366 #else
3367 
3368 static __always_inline void *
3369 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3370 {
3371         return ____cache_alloc(cachep, flags);
3372 }
3373 
3374 #endif /* CONFIG_NUMA */
3375 
3376 static __always_inline void *
3377 slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
3378 {
3379         unsigned long save_flags;
3380         void *objp;
3381 
3382         flags &= gfp_allowed_mask;
3383         cachep = slab_pre_alloc_hook(cachep, flags);
3384         if (unlikely(!cachep))
3385                 return NULL;
3386 
3387         cache_alloc_debugcheck_before(cachep, flags);
3388         local_irq_save(save_flags);
3389         objp = __do_cache_alloc(cachep, flags);
3390         local_irq_restore(save_flags);
3391         objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3392         prefetchw(objp);
3393 
3394         if (unlikely(flags & __GFP_ZERO) && objp)
3395                 memset(objp, 0, cachep->object_size);
3396 
3397         slab_post_alloc_hook(cachep, flags, 1, &objp);
3398         return objp;
3399 }
3400 
3401 /*
3402  * Caller needs to acquire correct kmem_cache_node's list_lock
3403  * @list: List of detached free slabs should be freed by caller
3404  */
3405 static void free_block(struct kmem_cache *cachep, void **objpp,
3406                         int nr_objects, int node, struct list_head *list)
3407 {
3408         int i;
3409         struct kmem_cache_node *n = get_node(cachep, node);
3410         struct page *page;
3411 
3412         n->free_objects += nr_objects;
3413 
3414         for (i = 0; i < nr_objects; i++) {
3415                 void *objp;
3416                 struct page *page;
3417 
3418                 objp = objpp[i];
3419 
3420                 page = virt_to_head_page(objp);
3421                 list_del(&page->lru);
3422                 check_spinlock_acquired_node(cachep, node);
3423                 slab_put_obj(cachep, page, objp);
3424                 STATS_DEC_ACTIVE(cachep);
3425 
3426                 /* fixup slab chains */
3427                 if (page->active == 0) {
3428                         list_add(&page->lru, &n->slabs_free);
3429                         n->free_slabs++;
3430                 } else {
3431                         /* Unconditionally move a slab to the end of the
3432                          * partial list on free - maximum time for the
3433                          * other objects to be freed, too.
3434                          */
3435                         list_add_tail(&page->lru, &n->slabs_partial);
3436                 }
3437         }
3438 
3439         while (n->free_objects > n->free_limit && !list_empty(&n->slabs_free)) {
3440                 n->free_objects -= cachep->num;
3441 
3442                 page = list_last_entry(&n->slabs_free, struct page, lru);
3443                 list_move(&page->lru, list);
3444                 n->free_slabs--;
3445                 n->total_slabs--;
3446         }
3447 }
3448 
3449 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3450 {
3451         int batchcount;
3452         struct kmem_cache_node *n;
3453         int node = numa_mem_id();
3454         LIST_HEAD(list);
3455 
3456         batchcount = ac->batchcount;
3457 
3458         check_irq_off();
3459         n = get_node(cachep, node);
3460         spin_lock(&n->list_lock);
3461         if (n->shared) {
3462                 struct array_cache *shared_array = n->shared;
3463                 int max = shared_array->limit - shared_array->avail;
3464                 if (max) {
3465                         if (batchcount > max)
3466                                 batchcount = max;
3467                         memcpy(&(shared_array->entry[shared_array->avail]),
3468                                ac->entry, sizeof(void *) * batchcount);
3469                         shared_array->avail += batchcount;
3470                         goto free_done;
3471                 }
3472         }
3473 
3474         free_block(cachep, ac->entry, batchcount, node, &list);
3475 free_done:
3476 #if STATS
3477         {
3478                 int i = 0;
3479                 struct page *page;
3480 
3481                 list_for_each_entry(page, &n->slabs_free, lru) {
3482                         BUG_ON(page->active);
3483 
3484                         i++;
3485                 }
3486                 STATS_SET_FREEABLE(cachep, i);
3487         }
3488 #endif
3489         spin_unlock(&n->list_lock);
3490         slabs_destroy(cachep, &list);
3491         ac->avail -= batchcount;
3492         memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3493 }
3494 
3495 /*
3496  * Release an obj back to its cache. If the obj has a constructed state, it must
3497  * be in this state _before_ it is released.  Called with disabled ints.
3498  */
3499 static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3500                                 unsigned long caller)
3501 {
3502         /* Put the object into the quarantine, don't touch it for now. */
3503         if (kasan_slab_free(cachep, objp))
3504                 return;
3505 
3506         ___cache_free(cachep, objp, caller);
3507 }
3508 
3509 void ___cache_free(struct kmem_cache *cachep, void *objp,
3510                 unsigned long caller)
3511 {
3512         struct array_cache *ac = cpu_cache_get(cachep);
3513 
3514         check_irq_off();
3515         kmemleak_free_recursive(objp, cachep->flags);
3516         objp = cache_free_debugcheck(cachep, objp, caller);
3517 
3518         kmemcheck_slab_free(cachep, objp, cachep->object_size);
3519 
3520         /*
3521          * Skip calling cache_free_alien() when the platform is not numa.
3522          * This will avoid cache misses that happen while accessing slabp (which
3523          * is per page memory  reference) to get nodeid. Instead use a global
3524          * variable to skip the call, which is mostly likely to be present in
3525          * the cache.
3526          */
3527         if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3528                 return;
3529 
3530         if (ac->avail < ac->limit) {
3531                 STATS_INC_FREEHIT(cachep);
3532         } else {
3533                 STATS_INC_FREEMISS(cachep);
3534                 cache_flusharray(cachep, ac);
3535         }
3536 
3537         if (sk_memalloc_socks()) {
3538                 struct page *page = virt_to_head_page(objp);
3539 
3540                 if (unlikely(PageSlabPfmemalloc(page))) {
3541                         cache_free_pfmemalloc(cachep, page, objp);
3542                         return;
3543                 }
3544         }
3545 
3546         ac->entry[ac->avail++] = objp;
3547 }
3548 
3549 /**
3550  * kmem_cache_alloc - Allocate an object
3551  * @cachep: The cache to allocate from.
3552  * @flags: See kmalloc().
3553  *
3554  * Allocate an object from this cache.  The flags are only relevant
3555  * if the cache has no available objects.
3556  */
3557 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3558 {
3559         void *ret = slab_alloc(cachep, flags, _RET_IP_);
3560 
3561         kasan_slab_alloc(cachep, ret, flags);
3562         trace_kmem_cache_alloc(_RET_IP_, ret,
3563                                cachep->object_size, cachep->size, flags);
3564 
3565         return ret;
3566 }
3567 EXPORT_SYMBOL(kmem_cache_alloc);
3568 
3569 static __always_inline void
3570 cache_alloc_debugcheck_after_bulk(struct kmem_cache *s, gfp_t flags,
3571                                   size_t size, void **p, unsigned long caller)
3572 {
3573         size_t i;
3574 
3575         for (i = 0; i < size; i++)
3576                 p[i] = cache_alloc_debugcheck_after(s, flags, p[i], caller);
3577 }
3578 
3579 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3580                           void **p)
3581 {
3582         size_t i;
3583 
3584         s = slab_pre_alloc_hook(s, flags);
3585         if (!s)
3586                 return 0;
3587 
3588         cache_alloc_debugcheck_before(s, flags);
3589 
3590         local_irq_disable();
3591         for (i = 0; i < size; i++) {
3592                 void *objp = __do_cache_alloc(s, flags);
3593 
3594                 if (unlikely(!objp))
3595                         goto error;
3596                 p[i] = objp;
3597         }
3598         local_irq_enable();
3599 
3600         cache_alloc_debugcheck_after_bulk(s, flags, size, p, _RET_IP_);
3601 
3602         /* Clear memory outside IRQ disabled section */
3603         if (unlikely(flags & __GFP_ZERO))
3604                 for (i = 0; i < size; i++)
3605                         memset(p[i], 0, s->object_size);
3606 
3607         slab_post_alloc_hook(s, flags, size, p);
3608         /* FIXME: Trace call missing. Christoph would like a bulk variant */
3609         return size;
3610 error:
3611         local_irq_enable();
3612         cache_alloc_debugcheck_after_bulk(s, flags, i, p, _RET_IP_);
3613         slab_post_alloc_hook(s, flags, i, p);
3614         __kmem_cache_free_bulk(s, i, p);
3615         return 0;
3616 }
3617 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3618 
3619 #ifdef CONFIG_TRACING
3620 void *
3621 kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
3622 {
3623         void *ret;
3624 
3625         ret = slab_alloc(cachep, flags, _RET_IP_);
3626 
3627         kasan_kmalloc(cachep, ret, size, flags);
3628         trace_kmalloc(_RET_IP_, ret,
3629                       size, cachep->size, flags);
3630         return ret;
3631 }
3632 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3633 #endif
3634 
3635 #ifdef CONFIG_NUMA
3636 /**
3637  * kmem_cache_alloc_node - Allocate an object on the specified node
3638  * @cachep: The cache to allocate from.
3639  * @flags: See kmalloc().
3640  * @nodeid: node number of the target node.
3641  *
3642  * Identical to kmem_cache_alloc but it will allocate memory on the given
3643  * node, which can improve the performance for cpu bound structures.
3644  *
3645  * Fallback to other node is possible if __GFP_THISNODE is not set.
3646  */
3647 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3648 {
3649         void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3650 
3651         kasan_slab_alloc(cachep, ret, flags);
3652         trace_kmem_cache_alloc_node(_RET_IP_, ret,
3653                                     cachep->object_size, cachep->size,
3654                                     flags, nodeid);
3655 
3656         return ret;
3657 }
3658 EXPORT_SYMBOL(kmem_cache_alloc_node);
3659 
3660 #ifdef CONFIG_TRACING
3661 void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
3662                                   gfp_t flags,
3663                                   int nodeid,
3664                                   size_t size)
3665 {
3666         void *ret;
3667 
3668         ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3669 
3670         kasan_kmalloc(cachep, ret, size, flags);
3671         trace_kmalloc_node(_RET_IP_, ret,
3672                            size, cachep->size,
3673                            flags, nodeid);
3674         return ret;
3675 }
3676 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3677 #endif
3678 
3679 static __always_inline void *
3680 __do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
3681 {
3682         struct kmem_cache *cachep;
3683         void *ret;
3684 
3685         cachep = kmalloc_slab(size, flags);
3686         if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3687                 return cachep;
3688         ret = kmem_cache_alloc_node_trace(cachep, flags, node, size);
3689         kasan_kmalloc(cachep, ret, size, flags);
3690 
3691         return ret;
3692 }
3693 
3694 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3695 {
3696         return __do_kmalloc_node(size, flags, node, _RET_IP_);
3697 }
3698 EXPORT_SYMBOL(__kmalloc_node);
3699 
3700 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3701                 int node, unsigned long caller)
3702 {
3703         return __do_kmalloc_node(size, flags, node, caller);
3704 }
3705 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3706 #endif /* CONFIG_NUMA */
3707 
3708 /**
3709  * __do_kmalloc - allocate memory
3710  * @size: how many bytes of memory are required.
3711  * @flags: the type of memory to allocate (see kmalloc).
3712  * @caller: function caller for debug tracking of the caller
3713  */
3714 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3715                                           unsigned long caller)
3716 {
3717         struct kmem_cache *cachep;
3718         void *ret;
3719 
3720         cachep = kmalloc_slab(size, flags);
3721         if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3722                 return cachep;
3723         ret = slab_alloc(cachep, flags, caller);
3724 
3725         kasan_kmalloc(cachep, ret, size, flags);
3726         trace_kmalloc(caller, ret,
3727                       size, cachep->size, flags);
3728 
3729         return ret;
3730 }
3731 
3732 void *__kmalloc(size_t size, gfp_t flags)
3733 {
3734         return __do_kmalloc(size, flags, _RET_IP_);
3735 }
3736 EXPORT_SYMBOL(__kmalloc);
3737 
3738 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3739 {
3740         return __do_kmalloc(size, flags, caller);
3741 }
3742 EXPORT_SYMBOL(__kmalloc_track_caller);
3743 
3744 /**
3745  * kmem_cache_free - Deallocate an object
3746  * @cachep: The cache the allocation was from.
3747  * @objp: The previously allocated object.
3748  *
3749  * Free an object which was previously allocated from this
3750  * cache.
3751  */
3752 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3753 {
3754         unsigned long flags;
3755         cachep = cache_from_obj(cachep, objp);
3756         if (!cachep)
3757                 return;
3758 
3759         local_irq_save(flags);
3760         debug_check_no_locks_freed(objp, cachep->object_size);
3761         if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3762                 debug_check_no_obj_freed(objp, cachep->object_size);
3763         __cache_free(cachep, objp, _RET_IP_);
3764         local_irq_restore(flags);
3765 
3766         trace_kmem_cache_free(_RET_IP_, objp);
3767 }
3768 EXPORT_SYMBOL(kmem_cache_free);
3769 
3770 void kmem_cache_free_bulk(struct kmem_cache *orig_s, size_t size, void **p)
3771 {
3772         struct kmem_cache *s;
3773         size_t i;
3774 
3775         local_irq_disable();
3776         for (i = 0; i < size; i++) {
3777                 void *objp = p[i];
3778 
3779                 if (!orig_s) /* called via kfree_bulk */
3780                         s = virt_to_cache(objp);
3781                 else
3782                         s = cache_from_obj(orig_s, objp);
3783 
3784                 debug_check_no_locks_freed(objp, s->object_size);
3785                 if (!(s->flags & SLAB_DEBUG_OBJECTS))
3786                         debug_check_no_obj_freed(objp, s->object_size);
3787 
3788                 __cache_free(s, objp, _RET_IP_);
3789         }
3790         local_irq_enable();
3791 
3792         /* FIXME: add tracing */
3793 }
3794 EXPORT_SYMBOL(kmem_cache_free_bulk);
3795 
3796 /**
3797  * kfree - free previously allocated memory
3798  * @objp: pointer returned by kmalloc.
3799  *
3800  * If @objp is NULL, no operation is performed.
3801  *
3802  * Don't free memory not originally allocated by kmalloc()
3803  * or you will run into trouble.
3804  */
3805 void kfree(const void *objp)
3806 {
3807         struct kmem_cache *c;
3808         unsigned long flags;
3809 
3810         trace_kfree(_RET_IP_, objp);
3811 
3812         if (unlikely(ZERO_OR_NULL_PTR(objp)))
3813                 return;
3814         local_irq_save(flags);
3815         kfree_debugcheck(objp);
3816         c = virt_to_cache(objp);
3817         debug_check_no_locks_freed(objp, c->object_size);
3818 
3819         debug_check_no_obj_freed(objp, c->object_size);
3820         __cache_free(c, (void *)objp, _RET_IP_);
3821         local_irq_restore(flags);
3822 }
3823 EXPORT_SYMBOL(kfree);
3824 
3825 /*
3826  * This initializes kmem_cache_node or resizes various caches for all nodes.
3827  */
3828 static int setup_kmem_cache_nodes(struct kmem_cache *cachep, gfp_t gfp)
3829 {
3830         int ret;
3831         int node;
3832         struct kmem_cache_node *n;
3833 
3834         for_each_online_node(node) {
3835                 ret = setup_kmem_cache_node(cachep, node, gfp, true);
3836                 if (ret)
3837                         goto fail;
3838 
3839         }
3840 
3841         return 0;
3842 
3843 fail:
3844         if (!cachep->list.next) {
3845                 /* Cache is not active yet. Roll back what we did */
3846                 node--;
3847                 while (node >= 0) {
3848                         n = get_node(cachep, node);
3849                         if (n) {
3850                                 kfree(n->shared);
3851                                 free_alien_cache(n->alien);
3852                                 kfree(n);
3853                                 cachep->node[node] = NULL;
3854                         }
3855                         node--;
3856                 }
3857         }
3858         return -ENOMEM;
3859 }
3860 
3861 /* Always called with the slab_mutex held */
3862 static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
3863                                 int batchcount, int shared, gfp_t gfp)
3864 {
3865         struct array_cache __percpu *cpu_cache, *prev;
3866         int cpu;
3867 
3868         cpu_cache = alloc_kmem_cache_cpus(cachep, limit, batchcount);
3869         if (!cpu_cache)
3870                 return -ENOMEM;
3871 
3872         prev = cachep->cpu_cache;
3873         cachep->cpu_cache = cpu_cache;
3874         /*
3875          * Without a previous cpu_cache there's no need to synchronize remote
3876          * cpus, so skip the IPIs.
3877          */
3878         if (prev)
3879                 kick_all_cpus_sync();
3880 
3881         check_irq_on();
3882         cachep->batchcount = batchcount;
3883         cachep->limit = limit;
3884         cachep->shared = shared;
3885 
3886         if (!prev)
3887                 goto setup_node;
3888 
3889         for_each_online_cpu(cpu) {
3890                 LIST_HEAD(list);
3891                 int node;
3892                 struct kmem_cache_node *n;
3893                 struct array_cache *ac = per_cpu_ptr(prev, cpu);
3894 
3895                 node = cpu_to_mem(cpu);
3896                 n = get_node(cachep, node);
3897                 spin_lock_irq(&n->list_lock);
3898                 free_block(cachep, ac->entry, ac->avail, node, &list);
3899                 spin_unlock_irq(&n->list_lock);
3900                 slabs_destroy(cachep, &list);
3901         }
3902         free_percpu(prev);
3903 
3904 setup_node:
3905         return setup_kmem_cache_nodes(cachep, gfp);
3906 }
3907 
3908 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3909                                 int batchcount, int shared, gfp_t gfp)
3910 {
3911         int ret;
3912         struct kmem_cache *c;
3913 
3914         ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3915 
3916         if (slab_state < FULL)
3917                 return ret;
3918 
3919         if ((ret < 0) || !is_root_cache(cachep))
3920                 return ret;
3921 
3922         lockdep_assert_held(&slab_mutex);
3923         for_each_memcg_cache(c, cachep) {
3924                 /* return value determined by the root cache only */
3925                 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
3926         }
3927 
3928         return ret;
3929 }
3930 
3931 /* Called with slab_mutex held always */
3932 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
3933 {
3934         int err;
3935         int limit = 0;
3936         int shared = 0;
3937         int batchcount = 0;
3938 
3939         err = cache_random_seq_create(cachep, cachep->num, gfp);
3940         if (err)
3941                 goto end;
3942 
3943         if (!is_root_cache(cachep)) {
3944                 struct kmem_cache *root = memcg_root_cache(cachep);
3945                 limit = root->limit;
3946                 shared = root->shared;
3947                 batchcount = root->batchcount;
3948         }
3949 
3950         if (limit && shared && batchcount)
3951                 goto skip_setup;
3952         /*
3953          * The head array serves three purposes:
3954          * - create a LIFO ordering, i.e. return objects that are cache-warm
3955          * - reduce the number of spinlock operations.
3956          * - reduce the number of linked list operations on the slab and
3957          *   bufctl chains: array operations are cheaper.
3958          * The numbers are guessed, we should auto-tune as described by
3959          * Bonwick.
3960          */
3961         if (cachep->size > 131072)
3962                 limit = 1;
3963         else if (cachep->size > PAGE_SIZE)
3964                 limit = 8;
3965         else if (cachep->size > 1024)
3966                 limit = 24;
3967         else if (cachep->size > 256)
3968                 limit = 54;
3969         else
3970                 limit = 120;
3971 
3972         /*
3973          * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3974          * allocation behaviour: Most allocs on one cpu, most free operations
3975          * on another cpu. For these cases, an efficient object passing between
3976          * cpus is necessary. This is provided by a shared array. The array
3977          * replaces Bonwick's magazine layer.
3978          * On uniprocessor, it's functionally equivalent (but less efficient)
3979          * to a larger limit. Thus disabled by default.
3980          */
3981         shared = 0;
3982         if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
3983                 shared = 8;
3984 
3985 #if DEBUG
3986         /*
3987          * With debugging enabled, large batchcount lead to excessively long
3988          * periods with disabled local interrupts. Limit the batchcount
3989          */
3990         if (limit > 32)
3991                 limit = 32;
3992 #endif
3993         batchcount = (limit + 1) / 2;
3994 skip_setup:
3995         err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3996 end:
3997         if (err)
3998                 pr_err("enable_cpucache failed for %s, error %d\n",
3999                        cachep->name, -err);
4000         return err;
4001 }
4002 
4003 /*
4004  * Drain an array if it contains any elements taking the node lock only if
4005  * necessary. Note that the node listlock also protects the array_cache
4006  * if drain_array() is used on the shared array.
4007  */
4008 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
4009                          struct array_cache *ac, int node)
4010 {
4011         LIST_HEAD(list);
4012 
4013         /* ac from n->shared can be freed if we don't hold the slab_mutex. */
4014         check_mutex_acquired();
4015 
4016         if (!ac || !ac->avail)
4017                 return;
4018 
4019         if (ac->touched) {
4020                 ac->touched = 0;
4021                 return;
4022         }
4023 
4024         spin_lock_irq(&n->list_lock);
4025         drain_array_locked(cachep, ac, node, false, &list);
4026         spin_unlock_irq(&n->list_lock);
4027 
4028         slabs_destroy(cachep, &list);
4029 }
4030 
4031 /**
4032  * cache_reap - Reclaim memory from caches.
4033  * @w: work descriptor
4034  *
4035  * Called from workqueue/eventd every few seconds.
4036  * Purpose:
4037  * - clear the per-cpu caches for this CPU.
4038  * - return freeable pages to the main free memory pool.
4039  *
4040  * If we cannot acquire the cache chain mutex then just give up - we'll try
4041  * again on the next iteration.
4042  */
4043 static void cache_reap(struct work_struct *w)
4044 {
4045         struct kmem_cache *searchp;
4046         struct kmem_cache_node *n;
4047         int node = numa_mem_id();
4048         struct delayed_work *work = to_delayed_work(w);
4049 
4050         if (!mutex_trylock(&slab_mutex))
4051                 /* Give up. Setup the next iteration. */
4052                 goto out;
4053 
4054         list_for_each_entry(searchp, &slab_caches, list) {
4055                 check_irq_on();
4056 
4057                 /*
4058                  * We only take the node lock if absolutely necessary and we
4059                  * have established with reasonable certainty that
4060                  * we can do some work if the lock was obtained.
4061                  */
4062                 n = get_node(searchp, node);
4063 
4064                 reap_alien(searchp, n);
4065 
4066                 drain_array(searchp, n, cpu_cache_get(searchp), node);
4067 
4068                 /*
4069                  * These are racy checks but it does not matter
4070                  * if we skip one check or scan twice.
4071                  */
4072                 if (time_after(n->next_reap, jiffies))
4073                         goto next;
4074 
4075                 n->next_reap = jiffies + REAPTIMEOUT_NODE;
4076 
4077                 drain_array(searchp, n, n->shared, node);
4078 
4079                 if (n->free_touched)
4080                         n->free_touched = 0;
4081                 else {
4082                         int freed;
4083 
4084                         freed = drain_freelist(searchp, n, (n->free_limit +
4085                                 5 * searchp->num - 1) / (5 * searchp->num));
4086                         STATS_ADD_REAPED(searchp, freed);
4087                 }
4088 next:
4089                 cond_resched();
4090         }
4091         check_irq_on();
4092         mutex_unlock(&slab_mutex);
4093         next_reap_node();
4094 out:
4095         /* Set up the next iteration */
4096         schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_AC));
4097 }
4098 
4099 #ifdef CONFIG_SLABINFO
4100 void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
4101 {
4102         unsigned long active_objs, num_objs, active_slabs;
4103         unsigned long total_slabs = 0, free_objs = 0, shared_avail = 0;
4104         unsigned long free_slabs = 0;
4105         int node;
4106         struct kmem_cache_node *n;
4107 
4108         for_each_kmem_cache_node(cachep, node, n) {
4109                 check_irq_on();
4110                 spin_lock_irq(&n->list_lock);
4111 
4112                 total_slabs += n->total_slabs;
4113                 free_slabs += n->free_slabs;
4114                 free_objs += n->free_objects;
4115 
4116                 if (n->shared)
4117                         shared_avail += n->shared->avail;
4118 
4119                 spin_unlock_irq(&n->list_lock);
4120         }
4121         num_objs = total_slabs * cachep->num;
4122         active_slabs = total_slabs - free_slabs;
4123         active_objs = num_objs - free_objs;
4124 
4125         sinfo->active_objs = active_objs;
4126         sinfo->num_objs = num_objs;
4127         sinfo->active_slabs = active_slabs;
4128         sinfo->num_slabs = total_slabs;
4129         sinfo->shared_avail = shared_avail;
4130         sinfo->limit = cachep->limit;
4131         sinfo->batchcount = cachep->batchcount;
4132         sinfo->shared = cachep->shared;
4133         sinfo->objects_per_slab = cachep->num;
4134         sinfo->cache_order = cachep->gfporder;
4135 }
4136 
4137 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4138 {
4139 #if STATS
4140         {                       /* node stats */
4141                 unsigned long high = cachep->high_mark;
4142                 unsigned long allocs = cachep->num_allocations;
4143                 unsigned long grown = cachep->grown;
4144                 unsigned long reaped = cachep->reaped;
4145                 unsigned long errors = cachep->errors;
4146                 unsigned long max_freeable = cachep->max_freeable;
4147                 unsigned long node_allocs = cachep->node_allocs;
4148                 unsigned long node_frees = cachep->node_frees;
4149                 unsigned long overflows = cachep->node_overflow;
4150 
4151                 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu %4lu %4lu",
4152                            allocs, high, grown,
4153                            reaped, errors, max_freeable, node_allocs,
4154                            node_frees, overflows);
4155         }
4156         /* cpu stats */
4157         {
4158                 unsigned long allochit = atomic_read(&cachep->allochit);
4159                 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4160                 unsigned long freehit = atomic_read(&cachep->freehit);
4161                 unsigned long freemiss = atomic_read(&cachep->freemiss);
4162 
4163                 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4164                            allochit, allocmiss, freehit, freemiss);
4165         }
4166 #endif
4167 }
4168 
4169 #define MAX_SLABINFO_WRITE 128
4170 /**
4171  * slabinfo_write - Tuning for the slab allocator
4172  * @file: unused
4173  * @buffer: user buffer
4174  * @count: data length
4175  * @ppos: unused
4176  */
4177 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4178                        size_t count, loff_t *ppos)
4179 {
4180         char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4181         int limit, batchcount, shared, res;
4182         struct kmem_cache *cachep;
4183 
4184         if (count > MAX_SLABINFO_WRITE)
4185                 return -EINVAL;
4186         if (copy_from_user(&kbuf, buffer, count))
4187                 return -EFAULT;
4188         kbuf[MAX_SLABINFO_WRITE] = '\0';
4189 
4190         tmp = strchr(kbuf, ' ');
4191         if (!tmp)
4192                 return -EINVAL;
4193         *tmp = '\0';
4194         tmp++;
4195         if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4196                 return -EINVAL;
4197 
4198         /* Find the cache in the chain of caches. */
4199         mutex_lock(&slab_mutex);
4200         res = -EINVAL;
4201         list_for_each_entry(cachep, &slab_caches, list) {
4202                 if (!strcmp(cachep->name, kbuf)) {
4203                         if (limit < 1 || batchcount < 1 ||
4204                                         batchcount > limit || shared < 0) {
4205                                 res = 0;
4206                         } else {
4207                                 res = do_tune_cpucache(cachep, limit,
4208                                                        batchcount, shared,
4209                                                        GFP_KERNEL);
4210                         }
4211                         break;
4212                 }
4213         }
4214         mutex_unlock(&slab_mutex);
4215         if (res >= 0)
4216                 res = count;
4217         return res;
4218 }
4219 
4220 #ifdef CONFIG_DEBUG_SLAB_LEAK
4221 
4222 static inline int add_caller(unsigned long *n, unsigned long v)
4223 {
4224         unsigned long *p;
4225         int l;
4226         if (!v)
4227                 return 1;
4228         l = n[1];
4229         p = n + 2;
4230         while (l) {
4231                 int i = l/2;
4232                 unsigned long *q = p + 2 * i;
4233                 if (*q == v) {
4234                         q[1]++;
4235                         return 1;
4236                 }
4237                 if (*q > v) {
4238                         l = i;
4239                 } else {
4240                         p = q + 2;
4241                         l -= i + 1;
4242                 }
4243         }
4244         if (++n[1] == n[0])
4245                 return 0;
4246         memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4247         p[0] = v;
4248         p[1] = 1;
4249         return 1;
4250 }
4251 
4252 static void handle_slab(unsigned long *n, struct kmem_cache *c,
4253                                                 struct page *page)
4254 {
4255         void *p;
4256         int i, j;
4257         unsigned long v;
4258 
4259         if (n[0] == n[1])
4260                 return;
4261         for (i = 0, p = page->s_mem; i < c->num; i++, p += c->size) {
4262                 bool active = true;
4263 
4264                 for (j = page->active; j < c->num; j++) {
4265                         if (get_free_obj(page, j) == i) {
4266                                 active = false;
4267                                 break;
4268                         }
4269                 }
4270 
4271                 if (!active)
4272                         continue;
4273 
4274                 /*
4275                  * probe_kernel_read() is used for DEBUG_PAGEALLOC. page table
4276                  * mapping is established when actual object allocation and
4277                  * we could mistakenly access the unmapped object in the cpu
4278                  * cache.
4279                  */
4280                 if (probe_kernel_read(&v, dbg_userword(c, p), sizeof(v)))
4281                         continue;
4282 
4283                 if (!add_caller(n, v))
4284                         return;
4285         }
4286 }
4287 
4288 static void show_symbol(struct seq_file *m, unsigned long address)
4289 {
4290 #ifdef CONFIG_KALLSYMS
4291         unsigned long offset, size;
4292         char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4293 
4294         if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4295                 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4296                 if (modname[0])
4297                         seq_printf(m, " [%s]", modname);
4298                 return;
4299         }
4300 #endif
4301         seq_printf(m, "%p", (void *)address);
4302 }
4303 
4304 static int leaks_show(struct seq_file *m, void *p)
4305 {
4306         struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4307         struct page *page;
4308         struct kmem_cache_node *n;
4309         const char *name;
4310         unsigned long *x = m->private;
4311         int node;
4312         int i;
4313 
4314         if (!(cachep->flags & SLAB_STORE_USER))
4315                 return 0;
4316         if (!(cachep->flags & SLAB_RED_ZONE))
4317                 return 0;
4318 
4319         /*
4320          * Set store_user_clean and start to grab stored user information
4321          * for all objects on this cache. If some alloc/free requests comes
4322          * during the processing, information would be wrong so restart
4323          * whole processing.
4324          */
4325         do {
4326                 set_store_user_clean(cachep);
4327                 drain_cpu_caches(cachep);
4328 
4329                 x[1] = 0;
4330 
4331                 for_each_kmem_cache_node(cachep, node, n) {
4332 
4333                         check_irq_on();
4334                         spin_lock_irq(&n->list_lock);
4335 
4336                         list_for_each_entry(page, &n->slabs_full, lru)
4337                                 handle_slab(x, cachep, page);
4338                         list_for_each_entry(page, &n->slabs_partial, lru)
4339                                 handle_slab(x, cachep, page);
4340                         spin_unlock_irq(&n->list_lock);
4341                 }
4342         } while (!is_store_user_clean(cachep));
4343 
4344         name = cachep->name;
4345         if (x[0] == x[1]) {
4346                 /* Increase the buffer size */
4347                 mutex_unlock(&slab_mutex);
4348                 m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4349                 if (!m->private) {
4350                         /* Too bad, we are really out */
4351                         m->private = x;
4352                         mutex_lock(&slab_mutex);
4353                         return -ENOMEM;
4354                 }
4355                 *(unsigned long *)m->private = x[0] * 2;
4356                 kfree(x);
4357                 mutex_lock(&slab_mutex);
4358                 /* Now make sure this entry will be retried */
4359                 m->count = m->size;
4360                 return 0;
4361         }
4362         for (i = 0; i < x[1]; i++) {
4363                 seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4364                 show_symbol(m, x[2*i+2]);
4365                 seq_putc(m, '\n');
4366         }
4367 
4368         return 0;
4369 }
4370 
4371 static const struct seq_operations slabstats_op = {
4372         .start = slab_start,
4373         .next = slab_next,
4374         .stop = slab_stop,
4375         .show = leaks_show,
4376 };
4377 
4378 static int slabstats_open(struct inode *inode, struct file *file)
4379 {
4380         unsigned long *n;
4381 
4382         n = __seq_open_private(file, &slabstats_op, PAGE_SIZE);
4383         if (!n)
4384                 return -ENOMEM;
4385 
4386         *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4387 
4388         return 0;
4389 }
4390 
4391 static const struct file_operations proc_slabstats_operations = {
4392         .open           = slabstats_open,
4393         .read           = seq_read,
4394         .llseek         = seq_lseek,
4395         .release        = seq_release_private,
4396 };
4397 #endif
4398 
4399 static int __init slab_proc_init(void)
4400 {
4401 #ifdef CONFIG_DEBUG_SLAB_LEAK
4402         proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4403 #endif
4404         return 0;
4405 }
4406 module_init(slab_proc_init);
4407 #endif
4408 
4409 #ifdef CONFIG_HARDENED_USERCOPY
4410 /*
4411  * Rejects objects that are incorrectly sized.
4412  *
4413  * Returns NULL if check passes, otherwise const char * to name of cache
4414  * to indicate an error.
4415  */
4416 const char *__check_heap_object(const void *ptr, unsigned long n,
4417                                 struct page *page)
4418 {
4419         struct kmem_cache *cachep;
4420         unsigned int objnr;
4421         unsigned long offset;
4422 
4423         /* Find and validate object. */
4424         cachep = page->slab_cache;
4425         objnr = obj_to_index(cachep, page, (void *)ptr);
4426         BUG_ON(objnr >= cachep->num);
4427 
4428         /* Find offset within object. */
4429         offset = ptr - index_to_obj(cachep, page, objnr) - obj_offset(cachep);
4430 
4431         /* Allow address range falling entirely within object size. */
4432         if (offset <= cachep->object_size && n <= cachep->object_size - offset)
4433                 return NULL;
4434 
4435         return cachep->name;
4436 }
4437 #endif /* CONFIG_HARDENED_USERCOPY */
4438 
4439 /**
4440  * ksize - get the actual amount of memory allocated for a given object
4441  * @objp: Pointer to the object
4442  *
4443  * kmalloc may internally round up allocations and return more memory
4444  * than requested. ksize() can be used to determine the actual amount of
4445  * memory allocated. The caller may use this additional memory, even though
4446  * a smaller amount of memory was initially specified with the kmalloc call.
4447  * The caller must guarantee that objp points to a valid object previously
4448  * allocated with either kmalloc() or kmem_cache_alloc(). The object
4449  * must not be freed during the duration of the call.
4450  */
4451 size_t ksize(const void *objp)
4452 {
4453         size_t size;
4454 
4455         BUG_ON(!objp);
4456         if (unlikely(objp == ZERO_SIZE_PTR))
4457                 return 0;
4458 
4459         size = virt_to_cache(objp)->object_size;
4460         /* We assume that ksize callers could use the whole allocated area,
4461          * so we need to unpoison this area.
4462          */
4463         kasan_unpoison_shadow(objp, size);
4464 
4465         return size;
4466 }
4467 EXPORT_SYMBOL(ksize);
4468 

~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

kernel.org | git.kernel.org | LWN.net | Project Home | Wiki (Japanese) | Wiki (English) | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

osdn.jp