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

TOMOYO Linux Cross Reference
Linux/mm/slab.c

Version: ~ [ linux-4.18 ] ~ [ linux-4.17.14 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.62 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.119 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.147 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.19.8 ] ~ [ linux-3.18.118 ] ~ [ linux-3.17.8 ] ~ [ linux-3.16.57 ] ~ [ linux-3.15.10 ] ~ [ linux-3.14.79 ] ~ [ linux-3.13.11 ] ~ [ linux-3.12.74 ] ~ [ linux-3.11.10 ] ~ [ linux-3.10.108 ] ~ [ linux-3.9.11 ] ~ [ linux-3.8.13 ] ~ [ linux-3.7.10 ] ~ [ linux-3.6.11 ] ~ [ linux-3.5.7 ] ~ [ linux-3.4.113 ] ~ [ linux-3.3.8 ] ~ [ linux-3.2.102 ] ~ [ linux-3.1.10 ] ~ [ linux-3.0.101 ] ~ [ linux-2.6.39.4 ] ~ [ linux-2.6.38.8 ] ~ [ linux-2.6.37.6 ] ~ [ linux-2.6.36.4 ] ~ [ linux-2.6.35.14 ] ~ [ linux-2.6.34.15 ] ~ [ linux-2.6.33.20 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.27.62 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~ [ linux-next-20180810 ] ~ [ linux-next-20180813 ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  1 // SPDX-License-Identifier: GPL-2.0
  2 /*
  3  * linux/mm/slab.c
  4  * Written by Mark Hemment, 1996/97.
  5  * (markhe@nextd.demon.co.uk)
  6  *
  7  * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
  8  *
  9  * Major cleanup, different bufctl logic, per-cpu arrays
 10  *      (c) 2000 Manfred Spraul
 11  *
 12  * Cleanup, make the head arrays unconditional, preparation for NUMA
 13  *      (c) 2002 Manfred Spraul
 14  *
 15  * An implementation of the Slab Allocator as described in outline in;
 16  *      UNIX Internals: The New Frontiers by Uresh Vahalia
 17  *      Pub: Prentice Hall      ISBN 0-13-101908-2
 18  * or with a little more detail in;
 19  *      The Slab Allocator: An Object-Caching Kernel Memory Allocator
 20  *      Jeff Bonwick (Sun Microsystems).
 21  *      Presented at: USENIX Summer 1994 Technical Conference
 22  *
 23  * The memory is organized in caches, one cache for each object type.
 24  * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
 25  * Each cache consists out of many slabs (they are small (usually one
 26  * page long) and always contiguous), and each slab contains multiple
 27  * initialized objects.
 28  *
 29  * This means, that your constructor is used only for newly allocated
 30  * slabs and you must pass objects with the same initializations to
 31  * kmem_cache_free.
 32  *
 33  * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
 34  * normal). If you need a special memory type, then must create a new
 35  * cache for that memory type.
 36  *
 37  * In order to reduce fragmentation, the slabs are sorted in 3 groups:
 38  *   full slabs with 0 free objects
 39  *   partial slabs
 40  *   empty slabs with no allocated objects
 41  *
 42  * If partial slabs exist, then new allocations come from these slabs,
 43  * otherwise from empty slabs or new slabs are allocated.
 44  *
 45  * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
 46  * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
 47  *
 48  * Each cache has a short per-cpu head array, most allocs
 49  * and frees go into that array, and if that array overflows, then 1/2
 50  * of the entries in the array are given back into the global cache.
 51  * The head array is strictly LIFO and should improve the cache hit rates.
 52  * On SMP, it additionally reduces the spinlock operations.
 53  *
 54  * The c_cpuarray may not be read with enabled local interrupts -
 55  * it's changed with a smp_call_function().
 56  *
 57  * SMP synchronization:
 58  *  constructors and destructors are called without any locking.
 59  *  Several members in struct kmem_cache and struct slab never change, they
 60  *      are accessed without any locking.
 61  *  The per-cpu arrays are never accessed from the wrong cpu, no locking,
 62  *      and local interrupts are disabled so slab code is preempt-safe.
 63  *  The non-constant members are protected with a per-cache irq spinlock.
 64  *
 65  * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
 66  * in 2000 - many ideas in the current implementation are derived from
 67  * his patch.
 68  *
 69  * Further notes from the original documentation:
 70  *
 71  * 11 April '97.  Started multi-threading - markhe
 72  *      The global cache-chain is protected by the mutex 'slab_mutex'.
 73  *      The sem is only needed when accessing/extending the cache-chain, which
 74  *      can never happen inside an interrupt (kmem_cache_create(),
 75  *      kmem_cache_shrink() and kmem_cache_reap()).
 76  *
 77  *      At present, each engine can be growing a cache.  This should be blocked.
 78  *
 79  * 15 March 2005. NUMA slab allocator.
 80  *      Shai Fultheim <shai@scalex86.org>.
 81  *      Shobhit Dayal <shobhit@calsoftinc.com>
 82  *      Alok N Kataria <alokk@calsoftinc.com>
 83  *      Christoph Lameter <christoph@lameter.com>
 84  *
 85  *      Modified the slab allocator to be node aware on NUMA systems.
 86  *      Each node has its own list of partial, free and full slabs.
 87  *      All object allocations for a node occur from node specific slab lists.
 88  */
 89 
 90 #include        <linux/slab.h>
 91 #include        <linux/mm.h>
 92 #include        <linux/poison.h>
 93 #include        <linux/swap.h>
 94 #include        <linux/cache.h>
 95 #include        <linux/interrupt.h>
 96 #include        <linux/init.h>
 97 #include        <linux/compiler.h>
 98 #include        <linux/cpuset.h>
 99 #include        <linux/proc_fs.h>
100 #include        <linux/seq_file.h>
101 #include        <linux/notifier.h>
102 #include        <linux/kallsyms.h>
103 #include        <linux/cpu.h>
104 #include        <linux/sysctl.h>
105 #include        <linux/module.h>
106 #include        <linux/rcupdate.h>
107 #include        <linux/string.h>
108 #include        <linux/uaccess.h>
109 #include        <linux/nodemask.h>
110 #include        <linux/kmemleak.h>
111 #include        <linux/mempolicy.h>
112 #include        <linux/mutex.h>
113 #include        <linux/fault-inject.h>
114 #include        <linux/rtmutex.h>
115 #include        <linux/reciprocal_div.h>
116 #include        <linux/debugobjects.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  ((slab_flags_t __force)0x40000000U)
255 #define CFLGS_OFF_SLAB          ((slab_flags_t __force)0x80000000U)
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                 slab_flags_t 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         kmem_cache = &kmem_cache_boot;
1239 
1240         if (!IS_ENABLED(CONFIG_NUMA) || num_possible_nodes() == 1)
1241                 use_alien_caches = 0;
1242 
1243         for (i = 0; i < NUM_INIT_LISTS; i++)
1244                 kmem_cache_node_init(&init_kmem_cache_node[i]);
1245 
1246         /*
1247          * Fragmentation resistance on low memory - only use bigger
1248          * page orders on machines with more than 32MB of memory if
1249          * not overridden on the command line.
1250          */
1251         if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
1252                 slab_max_order = SLAB_MAX_ORDER_HI;
1253 
1254         /* Bootstrap is tricky, because several objects are allocated
1255          * from caches that do not exist yet:
1256          * 1) initialize the kmem_cache cache: it contains the struct
1257          *    kmem_cache structures of all caches, except kmem_cache itself:
1258          *    kmem_cache is statically allocated.
1259          *    Initially an __init data area is used for the head array and the
1260          *    kmem_cache_node structures, it's replaced with a kmalloc allocated
1261          *    array at the end of the bootstrap.
1262          * 2) Create the first kmalloc cache.
1263          *    The struct kmem_cache for the new cache is allocated normally.
1264          *    An __init data area is used for the head array.
1265          * 3) Create the remaining kmalloc caches, with minimally sized
1266          *    head arrays.
1267          * 4) Replace the __init data head arrays for kmem_cache and the first
1268          *    kmalloc cache with kmalloc allocated arrays.
1269          * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1270          *    the other cache's with kmalloc allocated memory.
1271          * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1272          */
1273 
1274         /* 1) create the kmem_cache */
1275 
1276         /*
1277          * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1278          */
1279         create_boot_cache(kmem_cache, "kmem_cache",
1280                 offsetof(struct kmem_cache, node) +
1281                                   nr_node_ids * sizeof(struct kmem_cache_node *),
1282                                   SLAB_HWCACHE_ALIGN, 0, 0);
1283         list_add(&kmem_cache->list, &slab_caches);
1284         memcg_link_cache(kmem_cache);
1285         slab_state = PARTIAL;
1286 
1287         /*
1288          * Initialize the caches that provide memory for the  kmem_cache_node
1289          * structures first.  Without this, further allocations will bug.
1290          */
1291         kmalloc_caches[INDEX_NODE] = create_kmalloc_cache(
1292                                 kmalloc_info[INDEX_NODE].name,
1293                                 kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS,
1294                                 0, kmalloc_size(INDEX_NODE));
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         /* 6) resize the head arrays to their final sizes */
1320         mutex_lock(&slab_mutex);
1321         list_for_each_entry(cachep, &slab_caches, list)
1322                 if (enable_cpucache(cachep, GFP_NOWAIT))
1323                         BUG();
1324         mutex_unlock(&slab_mutex);
1325 
1326         /* Done! */
1327         slab_state = FULL;
1328 
1329 #ifdef CONFIG_NUMA
1330         /*
1331          * Register a memory hotplug callback that initializes and frees
1332          * node.
1333          */
1334         hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1335 #endif
1336 
1337         /*
1338          * The reap timers are started later, with a module init call: That part
1339          * of the kernel is not yet operational.
1340          */
1341 }
1342 
1343 static int __init cpucache_init(void)
1344 {
1345         int ret;
1346 
1347         /*
1348          * Register the timers that return unneeded pages to the page allocator
1349          */
1350         ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "SLAB online",
1351                                 slab_online_cpu, slab_offline_cpu);
1352         WARN_ON(ret < 0);
1353 
1354         return 0;
1355 }
1356 __initcall(cpucache_init);
1357 
1358 static noinline void
1359 slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1360 {
1361 #if DEBUG
1362         struct kmem_cache_node *n;
1363         unsigned long flags;
1364         int node;
1365         static DEFINE_RATELIMIT_STATE(slab_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1366                                       DEFAULT_RATELIMIT_BURST);
1367 
1368         if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slab_oom_rs))
1369                 return;
1370 
1371         pr_warn("SLAB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
1372                 nodeid, gfpflags, &gfpflags);
1373         pr_warn("  cache: %s, object size: %d, order: %d\n",
1374                 cachep->name, cachep->size, cachep->gfporder);
1375 
1376         for_each_kmem_cache_node(cachep, node, n) {
1377                 unsigned long total_slabs, free_slabs, free_objs;
1378 
1379                 spin_lock_irqsave(&n->list_lock, flags);
1380                 total_slabs = n->total_slabs;
1381                 free_slabs = n->free_slabs;
1382                 free_objs = n->free_objects;
1383                 spin_unlock_irqrestore(&n->list_lock, flags);
1384 
1385                 pr_warn("  node %d: slabs: %ld/%ld, objs: %ld/%ld\n",
1386                         node, total_slabs - free_slabs, total_slabs,
1387                         (total_slabs * cachep->num) - free_objs,
1388                         total_slabs * cachep->num);
1389         }
1390 #endif
1391 }
1392 
1393 /*
1394  * Interface to system's page allocator. No need to hold the
1395  * kmem_cache_node ->list_lock.
1396  *
1397  * If we requested dmaable memory, we will get it. Even if we
1398  * did not request dmaable memory, we might get it, but that
1399  * would be relatively rare and ignorable.
1400  */
1401 static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1402                                                                 int nodeid)
1403 {
1404         struct page *page;
1405         int nr_pages;
1406 
1407         flags |= cachep->allocflags;
1408 
1409         page = __alloc_pages_node(nodeid, flags, cachep->gfporder);
1410         if (!page) {
1411                 slab_out_of_memory(cachep, flags, nodeid);
1412                 return NULL;
1413         }
1414 
1415         if (memcg_charge_slab(page, flags, cachep->gfporder, cachep)) {
1416                 __free_pages(page, cachep->gfporder);
1417                 return NULL;
1418         }
1419 
1420         nr_pages = (1 << cachep->gfporder);
1421         if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1422                 mod_lruvec_page_state(page, NR_SLAB_RECLAIMABLE, nr_pages);
1423         else
1424                 mod_lruvec_page_state(page, NR_SLAB_UNRECLAIMABLE, nr_pages);
1425 
1426         __SetPageSlab(page);
1427         /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1428         if (sk_memalloc_socks() && page_is_pfmemalloc(page))
1429                 SetPageSlabPfmemalloc(page);
1430 
1431         return page;
1432 }
1433 
1434 /*
1435  * Interface to system's page release.
1436  */
1437 static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
1438 {
1439         int order = cachep->gfporder;
1440         unsigned long nr_freed = (1 << order);
1441 
1442         if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1443                 mod_lruvec_page_state(page, NR_SLAB_RECLAIMABLE, -nr_freed);
1444         else
1445                 mod_lruvec_page_state(page, NR_SLAB_UNRECLAIMABLE, -nr_freed);
1446 
1447         BUG_ON(!PageSlab(page));
1448         __ClearPageSlabPfmemalloc(page);
1449         __ClearPageSlab(page);
1450         page_mapcount_reset(page);
1451         page->mapping = NULL;
1452 
1453         if (current->reclaim_state)
1454                 current->reclaim_state->reclaimed_slab += nr_freed;
1455         memcg_uncharge_slab(page, order, cachep);
1456         __free_pages(page, order);
1457 }
1458 
1459 static void kmem_rcu_free(struct rcu_head *head)
1460 {
1461         struct kmem_cache *cachep;
1462         struct page *page;
1463 
1464         page = container_of(head, struct page, rcu_head);
1465         cachep = page->slab_cache;
1466 
1467         kmem_freepages(cachep, page);
1468 }
1469 
1470 #if DEBUG
1471 static bool is_debug_pagealloc_cache(struct kmem_cache *cachep)
1472 {
1473         if (debug_pagealloc_enabled() && OFF_SLAB(cachep) &&
1474                 (cachep->size % PAGE_SIZE) == 0)
1475                 return true;
1476 
1477         return false;
1478 }
1479 
1480 #ifdef CONFIG_DEBUG_PAGEALLOC
1481 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1482                             unsigned long caller)
1483 {
1484         int size = cachep->object_size;
1485 
1486         addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1487 
1488         if (size < 5 * sizeof(unsigned long))
1489                 return;
1490 
1491         *addr++ = 0x12345678;
1492         *addr++ = caller;
1493         *addr++ = smp_processor_id();
1494         size -= 3 * sizeof(unsigned long);
1495         {
1496                 unsigned long *sptr = &caller;
1497                 unsigned long svalue;
1498 
1499                 while (!kstack_end(sptr)) {
1500                         svalue = *sptr++;
1501                         if (kernel_text_address(svalue)) {
1502                                 *addr++ = svalue;
1503                                 size -= sizeof(unsigned long);
1504                                 if (size <= sizeof(unsigned long))
1505                                         break;
1506                         }
1507                 }
1508 
1509         }
1510         *addr++ = 0x87654321;
1511 }
1512 
1513 static void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1514                                 int map, unsigned long caller)
1515 {
1516         if (!is_debug_pagealloc_cache(cachep))
1517                 return;
1518 
1519         if (caller)
1520                 store_stackinfo(cachep, objp, caller);
1521 
1522         kernel_map_pages(virt_to_page(objp), cachep->size / PAGE_SIZE, map);
1523 }
1524 
1525 #else
1526 static inline void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1527                                 int map, unsigned long caller) {}
1528 
1529 #endif
1530 
1531 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1532 {
1533         int size = cachep->object_size;
1534         addr = &((char *)addr)[obj_offset(cachep)];
1535 
1536         memset(addr, val, size);
1537         *(unsigned char *)(addr + size - 1) = POISON_END;
1538 }
1539 
1540 static void dump_line(char *data, int offset, int limit)
1541 {
1542         int i;
1543         unsigned char error = 0;
1544         int bad_count = 0;
1545 
1546         pr_err("%03x: ", offset);
1547         for (i = 0; i < limit; i++) {
1548                 if (data[offset + i] != POISON_FREE) {
1549                         error = data[offset + i];
1550                         bad_count++;
1551                 }
1552         }
1553         print_hex_dump(KERN_CONT, "", 0, 16, 1,
1554                         &data[offset], limit, 1);
1555 
1556         if (bad_count == 1) {
1557                 error ^= POISON_FREE;
1558                 if (!(error & (error - 1))) {
1559                         pr_err("Single bit error detected. Probably bad RAM.\n");
1560 #ifdef CONFIG_X86
1561                         pr_err("Run memtest86+ or a similar memory test tool.\n");
1562 #else
1563                         pr_err("Run a memory test tool.\n");
1564 #endif
1565                 }
1566         }
1567 }
1568 #endif
1569 
1570 #if DEBUG
1571 
1572 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1573 {
1574         int i, size;
1575         char *realobj;
1576 
1577         if (cachep->flags & SLAB_RED_ZONE) {
1578                 pr_err("Redzone: 0x%llx/0x%llx\n",
1579                        *dbg_redzone1(cachep, objp),
1580                        *dbg_redzone2(cachep, objp));
1581         }
1582 
1583         if (cachep->flags & SLAB_STORE_USER)
1584                 pr_err("Last user: (%pSR)\n", *dbg_userword(cachep, objp));
1585         realobj = (char *)objp + obj_offset(cachep);
1586         size = cachep->object_size;
1587         for (i = 0; i < size && lines; i += 16, lines--) {
1588                 int limit;
1589                 limit = 16;
1590                 if (i + limit > size)
1591                         limit = size - i;
1592                 dump_line(realobj, i, limit);
1593         }
1594 }
1595 
1596 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1597 {
1598         char *realobj;
1599         int size, i;
1600         int lines = 0;
1601 
1602         if (is_debug_pagealloc_cache(cachep))
1603                 return;
1604 
1605         realobj = (char *)objp + obj_offset(cachep);
1606         size = cachep->object_size;
1607 
1608         for (i = 0; i < size; i++) {
1609                 char exp = POISON_FREE;
1610                 if (i == size - 1)
1611                         exp = POISON_END;
1612                 if (realobj[i] != exp) {
1613                         int limit;
1614                         /* Mismatch ! */
1615                         /* Print header */
1616                         if (lines == 0) {
1617                                 pr_err("Slab corruption (%s): %s start=%px, len=%d\n",
1618                                        print_tainted(), cachep->name,
1619                                        realobj, size);
1620                                 print_objinfo(cachep, objp, 0);
1621                         }
1622                         /* Hexdump the affected line */
1623                         i = (i / 16) * 16;
1624                         limit = 16;
1625                         if (i + limit > size)
1626                                 limit = size - i;
1627                         dump_line(realobj, i, limit);
1628                         i += 16;
1629                         lines++;
1630                         /* Limit to 5 lines */
1631                         if (lines > 5)
1632                                 break;
1633                 }
1634         }
1635         if (lines != 0) {
1636                 /* Print some data about the neighboring objects, if they
1637                  * exist:
1638                  */
1639                 struct page *page = virt_to_head_page(objp);
1640                 unsigned int objnr;
1641 
1642                 objnr = obj_to_index(cachep, page, objp);
1643                 if (objnr) {
1644                         objp = index_to_obj(cachep, page, objnr - 1);
1645                         realobj = (char *)objp + obj_offset(cachep);
1646                         pr_err("Prev obj: start=%px, len=%d\n", realobj, size);
1647                         print_objinfo(cachep, objp, 2);
1648                 }
1649                 if (objnr + 1 < cachep->num) {
1650                         objp = index_to_obj(cachep, page, objnr + 1);
1651                         realobj = (char *)objp + obj_offset(cachep);
1652                         pr_err("Next obj: start=%px, len=%d\n", realobj, size);
1653                         print_objinfo(cachep, objp, 2);
1654                 }
1655         }
1656 }
1657 #endif
1658 
1659 #if DEBUG
1660 static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1661                                                 struct page *page)
1662 {
1663         int i;
1664 
1665         if (OBJFREELIST_SLAB(cachep) && cachep->flags & SLAB_POISON) {
1666                 poison_obj(cachep, page->freelist - obj_offset(cachep),
1667                         POISON_FREE);
1668         }
1669 
1670         for (i = 0; i < cachep->num; i++) {
1671                 void *objp = index_to_obj(cachep, page, i);
1672 
1673                 if (cachep->flags & SLAB_POISON) {
1674                         check_poison_obj(cachep, objp);
1675                         slab_kernel_map(cachep, objp, 1, 0);
1676                 }
1677                 if (cachep->flags & SLAB_RED_ZONE) {
1678                         if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1679                                 slab_error(cachep, "start of a freed object was overwritten");
1680                         if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1681                                 slab_error(cachep, "end of a freed object was overwritten");
1682                 }
1683         }
1684 }
1685 #else
1686 static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1687                                                 struct page *page)
1688 {
1689 }
1690 #endif
1691 
1692 /**
1693  * slab_destroy - destroy and release all objects in a slab
1694  * @cachep: cache pointer being destroyed
1695  * @page: page pointer being destroyed
1696  *
1697  * Destroy all the objs in a slab page, and release the mem back to the system.
1698  * Before calling the slab page must have been unlinked from the cache. The
1699  * kmem_cache_node ->list_lock is not held/needed.
1700  */
1701 static void slab_destroy(struct kmem_cache *cachep, struct page *page)
1702 {
1703         void *freelist;
1704 
1705         freelist = page->freelist;
1706         slab_destroy_debugcheck(cachep, page);
1707         if (unlikely(cachep->flags & SLAB_TYPESAFE_BY_RCU))
1708                 call_rcu(&page->rcu_head, kmem_rcu_free);
1709         else
1710                 kmem_freepages(cachep, page);
1711 
1712         /*
1713          * From now on, we don't use freelist
1714          * although actual page can be freed in rcu context
1715          */
1716         if (OFF_SLAB(cachep))
1717                 kmem_cache_free(cachep->freelist_cache, freelist);
1718 }
1719 
1720 static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list)
1721 {
1722         struct page *page, *n;
1723 
1724         list_for_each_entry_safe(page, n, list, lru) {
1725                 list_del(&page->lru);
1726                 slab_destroy(cachep, page);
1727         }
1728 }
1729 
1730 /**
1731  * calculate_slab_order - calculate size (page order) of slabs
1732  * @cachep: pointer to the cache that is being created
1733  * @size: size of objects to be created in this cache.
1734  * @flags: slab allocation flags
1735  *
1736  * Also calculates the number of objects per slab.
1737  *
1738  * This could be made much more intelligent.  For now, try to avoid using
1739  * high order pages for slabs.  When the gfp() functions are more friendly
1740  * towards high-order requests, this should be changed.
1741  */
1742 static size_t calculate_slab_order(struct kmem_cache *cachep,
1743                                 size_t size, slab_flags_t flags)
1744 {
1745         size_t left_over = 0;
1746         int gfporder;
1747 
1748         for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
1749                 unsigned int num;
1750                 size_t remainder;
1751 
1752                 num = cache_estimate(gfporder, size, flags, &remainder);
1753                 if (!num)
1754                         continue;
1755 
1756                 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1757                 if (num > SLAB_OBJ_MAX_NUM)
1758                         break;
1759 
1760                 if (flags & CFLGS_OFF_SLAB) {
1761                         struct kmem_cache *freelist_cache;
1762                         size_t freelist_size;
1763 
1764                         freelist_size = num * sizeof(freelist_idx_t);
1765                         freelist_cache = kmalloc_slab(freelist_size, 0u);
1766                         if (!freelist_cache)
1767                                 continue;
1768 
1769                         /*
1770                          * Needed to avoid possible looping condition
1771                          * in cache_grow_begin()
1772                          */
1773                         if (OFF_SLAB(freelist_cache))
1774                                 continue;
1775 
1776                         /* check if off slab has enough benefit */
1777                         if (freelist_cache->size > cachep->size / 2)
1778                                 continue;
1779                 }
1780 
1781                 /* Found something acceptable - save it away */
1782                 cachep->num = num;
1783                 cachep->gfporder = gfporder;
1784                 left_over = remainder;
1785 
1786                 /*
1787                  * A VFS-reclaimable slab tends to have most allocations
1788                  * as GFP_NOFS and we really don't want to have to be allocating
1789                  * higher-order pages when we are unable to shrink dcache.
1790                  */
1791                 if (flags & SLAB_RECLAIM_ACCOUNT)
1792                         break;
1793 
1794                 /*
1795                  * Large number of objects is good, but very large slabs are
1796                  * currently bad for the gfp()s.
1797                  */
1798                 if (gfporder >= slab_max_order)
1799                         break;
1800 
1801                 /*
1802                  * Acceptable internal fragmentation?
1803                  */
1804                 if (left_over * 8 <= (PAGE_SIZE << gfporder))
1805                         break;
1806         }
1807         return left_over;
1808 }
1809 
1810 static struct array_cache __percpu *alloc_kmem_cache_cpus(
1811                 struct kmem_cache *cachep, int entries, int batchcount)
1812 {
1813         int cpu;
1814         size_t size;
1815         struct array_cache __percpu *cpu_cache;
1816 
1817         size = sizeof(void *) * entries + sizeof(struct array_cache);
1818         cpu_cache = __alloc_percpu(size, sizeof(void *));
1819 
1820         if (!cpu_cache)
1821                 return NULL;
1822 
1823         for_each_possible_cpu(cpu) {
1824                 init_arraycache(per_cpu_ptr(cpu_cache, cpu),
1825                                 entries, batchcount);
1826         }
1827 
1828         return cpu_cache;
1829 }
1830 
1831 static int __ref setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
1832 {
1833         if (slab_state >= FULL)
1834                 return enable_cpucache(cachep, gfp);
1835 
1836         cachep->cpu_cache = alloc_kmem_cache_cpus(cachep, 1, 1);
1837         if (!cachep->cpu_cache)
1838                 return 1;
1839 
1840         if (slab_state == DOWN) {
1841                 /* Creation of first cache (kmem_cache). */
1842                 set_up_node(kmem_cache, CACHE_CACHE);
1843         } else if (slab_state == PARTIAL) {
1844                 /* For kmem_cache_node */
1845                 set_up_node(cachep, SIZE_NODE);
1846         } else {
1847                 int node;
1848 
1849                 for_each_online_node(node) {
1850                         cachep->node[node] = kmalloc_node(
1851                                 sizeof(struct kmem_cache_node), gfp, node);
1852                         BUG_ON(!cachep->node[node]);
1853                         kmem_cache_node_init(cachep->node[node]);
1854                 }
1855         }
1856 
1857         cachep->node[numa_mem_id()]->next_reap =
1858                         jiffies + REAPTIMEOUT_NODE +
1859                         ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1860 
1861         cpu_cache_get(cachep)->avail = 0;
1862         cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1863         cpu_cache_get(cachep)->batchcount = 1;
1864         cpu_cache_get(cachep)->touched = 0;
1865         cachep->batchcount = 1;
1866         cachep->limit = BOOT_CPUCACHE_ENTRIES;
1867         return 0;
1868 }
1869 
1870 slab_flags_t kmem_cache_flags(unsigned int object_size,
1871         slab_flags_t flags, const char *name,
1872         void (*ctor)(void *))
1873 {
1874         return flags;
1875 }
1876 
1877 struct kmem_cache *
1878 __kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
1879                    slab_flags_t flags, void (*ctor)(void *))
1880 {
1881         struct kmem_cache *cachep;
1882 
1883         cachep = find_mergeable(size, align, flags, name, ctor);
1884         if (cachep) {
1885                 cachep->refcount++;
1886 
1887                 /*
1888                  * Adjust the object sizes so that we clear
1889                  * the complete object on kzalloc.
1890                  */
1891                 cachep->object_size = max_t(int, cachep->object_size, size);
1892         }
1893         return cachep;
1894 }
1895 
1896 static bool set_objfreelist_slab_cache(struct kmem_cache *cachep,
1897                         size_t size, slab_flags_t flags)
1898 {
1899         size_t left;
1900 
1901         cachep->num = 0;
1902 
1903         if (cachep->ctor || flags & SLAB_TYPESAFE_BY_RCU)
1904                 return false;
1905 
1906         left = calculate_slab_order(cachep, size,
1907                         flags | CFLGS_OBJFREELIST_SLAB);
1908         if (!cachep->num)
1909                 return false;
1910 
1911         if (cachep->num * sizeof(freelist_idx_t) > cachep->object_size)
1912                 return false;
1913 
1914         cachep->colour = left / cachep->colour_off;
1915 
1916         return true;
1917 }
1918 
1919 static bool set_off_slab_cache(struct kmem_cache *cachep,
1920                         size_t size, slab_flags_t flags)
1921 {
1922         size_t left;
1923 
1924         cachep->num = 0;
1925 
1926         /*
1927          * Always use on-slab management when SLAB_NOLEAKTRACE
1928          * to avoid recursive calls into kmemleak.
1929          */
1930         if (flags & SLAB_NOLEAKTRACE)
1931                 return false;
1932 
1933         /*
1934          * Size is large, assume best to place the slab management obj
1935          * off-slab (should allow better packing of objs).
1936          */
1937         left = calculate_slab_order(cachep, size, flags | CFLGS_OFF_SLAB);
1938         if (!cachep->num)
1939                 return false;
1940 
1941         /*
1942          * If the slab has been placed off-slab, and we have enough space then
1943          * move it on-slab. This is at the expense of any extra colouring.
1944          */
1945         if (left >= cachep->num * sizeof(freelist_idx_t))
1946                 return false;
1947 
1948         cachep->colour = left / cachep->colour_off;
1949 
1950         return true;
1951 }
1952 
1953 static bool set_on_slab_cache(struct kmem_cache *cachep,
1954                         size_t size, slab_flags_t flags)
1955 {
1956         size_t left;
1957 
1958         cachep->num = 0;
1959 
1960         left = calculate_slab_order(cachep, size, flags);
1961         if (!cachep->num)
1962                 return false;
1963 
1964         cachep->colour = left / cachep->colour_off;
1965 
1966         return true;
1967 }
1968 
1969 /**
1970  * __kmem_cache_create - Create a cache.
1971  * @cachep: cache management descriptor
1972  * @flags: SLAB flags
1973  *
1974  * Returns a ptr to the cache on success, NULL on failure.
1975  * Cannot be called within a int, but can be interrupted.
1976  * The @ctor is run when new pages are allocated by the cache.
1977  *
1978  * The flags are
1979  *
1980  * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1981  * to catch references to uninitialised memory.
1982  *
1983  * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1984  * for buffer overruns.
1985  *
1986  * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1987  * cacheline.  This can be beneficial if you're counting cycles as closely
1988  * as davem.
1989  */
1990 int __kmem_cache_create(struct kmem_cache *cachep, slab_flags_t flags)
1991 {
1992         size_t ralign = BYTES_PER_WORD;
1993         gfp_t gfp;
1994         int err;
1995         unsigned int size = cachep->size;
1996 
1997 #if DEBUG
1998 #if FORCED_DEBUG
1999         /*
2000          * Enable redzoning and last user accounting, except for caches with
2001          * large objects, if the increased size would increase the object size
2002          * above the next power of two: caches with object sizes just above a
2003          * power of two have a significant amount of internal fragmentation.
2004          */
2005         if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2006                                                 2 * sizeof(unsigned long long)))
2007                 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2008         if (!(flags & SLAB_TYPESAFE_BY_RCU))
2009                 flags |= SLAB_POISON;
2010 #endif
2011 #endif
2012 
2013         /*
2014          * Check that size is in terms of words.  This is needed to avoid
2015          * unaligned accesses for some archs when redzoning is used, and makes
2016          * sure any on-slab bufctl's are also correctly aligned.
2017          */
2018         size = ALIGN(size, BYTES_PER_WORD);
2019 
2020         if (flags & SLAB_RED_ZONE) {
2021                 ralign = REDZONE_ALIGN;
2022                 /* If redzoning, ensure that the second redzone is suitably
2023                  * aligned, by adjusting the object size accordingly. */
2024                 size = ALIGN(size, REDZONE_ALIGN);
2025         }
2026 
2027         /* 3) caller mandated alignment */
2028         if (ralign < cachep->align) {
2029                 ralign = cachep->align;
2030         }
2031         /* disable debug if necessary */
2032         if (ralign > __alignof__(unsigned long long))
2033                 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2034         /*
2035          * 4) Store it.
2036          */
2037         cachep->align = ralign;
2038         cachep->colour_off = cache_line_size();
2039         /* Offset must be a multiple of the alignment. */
2040         if (cachep->colour_off < cachep->align)
2041                 cachep->colour_off = cachep->align;
2042 
2043         if (slab_is_available())
2044                 gfp = GFP_KERNEL;
2045         else
2046                 gfp = GFP_NOWAIT;
2047 
2048 #if DEBUG
2049 
2050         /*
2051          * Both debugging options require word-alignment which is calculated
2052          * into align above.
2053          */
2054         if (flags & SLAB_RED_ZONE) {
2055                 /* add space for red zone words */
2056                 cachep->obj_offset += sizeof(unsigned long long);
2057                 size += 2 * sizeof(unsigned long long);
2058         }
2059         if (flags & SLAB_STORE_USER) {
2060                 /* user store requires one word storage behind the end of
2061                  * the real object. But if the second red zone needs to be
2062                  * aligned to 64 bits, we must allow that much space.
2063                  */
2064                 if (flags & SLAB_RED_ZONE)
2065                         size += REDZONE_ALIGN;
2066                 else
2067                         size += BYTES_PER_WORD;
2068         }
2069 #endif
2070 
2071         kasan_cache_create(cachep, &size, &flags);
2072 
2073         size = ALIGN(size, cachep->align);
2074         /*
2075          * We should restrict the number of objects in a slab to implement
2076          * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2077          */
2078         if (FREELIST_BYTE_INDEX && size < SLAB_OBJ_MIN_SIZE)
2079                 size = ALIGN(SLAB_OBJ_MIN_SIZE, cachep->align);
2080 
2081 #if DEBUG
2082         /*
2083          * To activate debug pagealloc, off-slab management is necessary
2084          * requirement. In early phase of initialization, small sized slab
2085          * doesn't get initialized so it would not be possible. So, we need
2086          * to check size >= 256. It guarantees that all necessary small
2087          * sized slab is initialized in current slab initialization sequence.
2088          */
2089         if (debug_pagealloc_enabled() && (flags & SLAB_POISON) &&
2090                 size >= 256 && cachep->object_size > cache_line_size()) {
2091                 if (size < PAGE_SIZE || size % PAGE_SIZE == 0) {
2092                         size_t tmp_size = ALIGN(size, PAGE_SIZE);
2093 
2094                         if (set_off_slab_cache(cachep, tmp_size, flags)) {
2095                                 flags |= CFLGS_OFF_SLAB;
2096                                 cachep->obj_offset += tmp_size - size;
2097                                 size = tmp_size;
2098                                 goto done;
2099                         }
2100                 }
2101         }
2102 #endif
2103 
2104         if (set_objfreelist_slab_cache(cachep, size, flags)) {
2105                 flags |= CFLGS_OBJFREELIST_SLAB;
2106                 goto done;
2107         }
2108 
2109         if (set_off_slab_cache(cachep, size, flags)) {
2110                 flags |= CFLGS_OFF_SLAB;
2111                 goto done;
2112         }
2113 
2114         if (set_on_slab_cache(cachep, size, flags))
2115                 goto done;
2116 
2117         return -E2BIG;
2118 
2119 done:
2120         cachep->freelist_size = cachep->num * sizeof(freelist_idx_t);
2121         cachep->flags = flags;
2122         cachep->allocflags = __GFP_COMP;
2123         if (flags & SLAB_CACHE_DMA)
2124                 cachep->allocflags |= GFP_DMA;
2125         if (flags & SLAB_RECLAIM_ACCOUNT)
2126                 cachep->allocflags |= __GFP_RECLAIMABLE;
2127         cachep->size = size;
2128         cachep->reciprocal_buffer_size = reciprocal_value(size);
2129 
2130 #if DEBUG
2131         /*
2132          * If we're going to use the generic kernel_map_pages()
2133          * poisoning, then it's going to smash the contents of
2134          * the redzone and userword anyhow, so switch them off.
2135          */
2136         if (IS_ENABLED(CONFIG_PAGE_POISONING) &&
2137                 (cachep->flags & SLAB_POISON) &&
2138                 is_debug_pagealloc_cache(cachep))
2139                 cachep->flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2140 #endif
2141 
2142         if (OFF_SLAB(cachep)) {
2143                 cachep->freelist_cache =
2144                         kmalloc_slab(cachep->freelist_size, 0u);
2145         }
2146 
2147         err = setup_cpu_cache(cachep, gfp);
2148         if (err) {
2149                 __kmem_cache_release(cachep);
2150                 return err;
2151         }
2152 
2153         return 0;
2154 }
2155 
2156 #if DEBUG
2157 static void check_irq_off(void)
2158 {
2159         BUG_ON(!irqs_disabled());
2160 }
2161 
2162 static void check_irq_on(void)
2163 {
2164         BUG_ON(irqs_disabled());
2165 }
2166 
2167 static void check_mutex_acquired(void)
2168 {
2169         BUG_ON(!mutex_is_locked(&slab_mutex));
2170 }
2171 
2172 static void check_spinlock_acquired(struct kmem_cache *cachep)
2173 {
2174 #ifdef CONFIG_SMP
2175         check_irq_off();
2176         assert_spin_locked(&get_node(cachep, numa_mem_id())->list_lock);
2177 #endif
2178 }
2179 
2180 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2181 {
2182 #ifdef CONFIG_SMP
2183         check_irq_off();
2184         assert_spin_locked(&get_node(cachep, node)->list_lock);
2185 #endif
2186 }
2187 
2188 #else
2189 #define check_irq_off() do { } while(0)
2190 #define check_irq_on()  do { } while(0)
2191 #define check_mutex_acquired()  do { } while(0)
2192 #define check_spinlock_acquired(x) do { } while(0)
2193 #define check_spinlock_acquired_node(x, y) do { } while(0)
2194 #endif
2195 
2196 static void drain_array_locked(struct kmem_cache *cachep, struct array_cache *ac,
2197                                 int node, bool free_all, struct list_head *list)
2198 {
2199         int tofree;
2200 
2201         if (!ac || !ac->avail)
2202                 return;
2203 
2204         tofree = free_all ? ac->avail : (ac->limit + 4) / 5;
2205         if (tofree > ac->avail)
2206                 tofree = (ac->avail + 1) / 2;
2207 
2208         free_block(cachep, ac->entry, tofree, node, list);
2209         ac->avail -= tofree;
2210         memmove(ac->entry, &(ac->entry[tofree]), sizeof(void *) * ac->avail);
2211 }
2212 
2213 static void do_drain(void *arg)
2214 {
2215         struct kmem_cache *cachep = arg;
2216         struct array_cache *ac;
2217         int node = numa_mem_id();
2218         struct kmem_cache_node *n;
2219         LIST_HEAD(list);
2220 
2221         check_irq_off();
2222         ac = cpu_cache_get(cachep);
2223         n = get_node(cachep, node);
2224         spin_lock(&n->list_lock);
2225         free_block(cachep, ac->entry, ac->avail, node, &list);
2226         spin_unlock(&n->list_lock);
2227         slabs_destroy(cachep, &list);
2228         ac->avail = 0;
2229 }
2230 
2231 static void drain_cpu_caches(struct kmem_cache *cachep)
2232 {
2233         struct kmem_cache_node *n;
2234         int node;
2235         LIST_HEAD(list);
2236 
2237         on_each_cpu(do_drain, cachep, 1);
2238         check_irq_on();
2239         for_each_kmem_cache_node(cachep, node, n)
2240                 if (n->alien)
2241                         drain_alien_cache(cachep, n->alien);
2242 
2243         for_each_kmem_cache_node(cachep, node, n) {
2244                 spin_lock_irq(&n->list_lock);
2245                 drain_array_locked(cachep, n->shared, node, true, &list);
2246                 spin_unlock_irq(&n->list_lock);
2247 
2248                 slabs_destroy(cachep, &list);
2249         }
2250 }
2251 
2252 /*
2253  * Remove slabs from the list of free slabs.
2254  * Specify the number of slabs to drain in tofree.
2255  *
2256  * Returns the actual number of slabs released.
2257  */
2258 static int drain_freelist(struct kmem_cache *cache,
2259                         struct kmem_cache_node *n, int tofree)
2260 {
2261         struct list_head *p;
2262         int nr_freed;
2263         struct page *page;
2264 
2265         nr_freed = 0;
2266         while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
2267 
2268                 spin_lock_irq(&n->list_lock);
2269                 p = n->slabs_free.prev;
2270                 if (p == &n->slabs_free) {
2271                         spin_unlock_irq(&n->list_lock);
2272                         goto out;
2273                 }
2274 
2275                 page = list_entry(p, struct page, lru);
2276                 list_del(&page->lru);
2277                 n->free_slabs--;
2278                 n->total_slabs--;
2279                 /*
2280                  * Safe to drop the lock. The slab is no longer linked
2281                  * to the cache.
2282                  */
2283                 n->free_objects -= cache->num;
2284                 spin_unlock_irq(&n->list_lock);
2285                 slab_destroy(cache, page);
2286                 nr_freed++;
2287         }
2288 out:
2289         return nr_freed;
2290 }
2291 
2292 bool __kmem_cache_empty(struct kmem_cache *s)
2293 {
2294         int node;
2295         struct kmem_cache_node *n;
2296 
2297         for_each_kmem_cache_node(s, node, n)
2298                 if (!list_empty(&n->slabs_full) ||
2299                     !list_empty(&n->slabs_partial))
2300                         return false;
2301         return true;
2302 }
2303 
2304 int __kmem_cache_shrink(struct kmem_cache *cachep)
2305 {
2306         int ret = 0;
2307         int node;
2308         struct kmem_cache_node *n;
2309 
2310         drain_cpu_caches(cachep);
2311 
2312         check_irq_on();
2313         for_each_kmem_cache_node(cachep, node, n) {
2314                 drain_freelist(cachep, n, INT_MAX);
2315 
2316                 ret += !list_empty(&n->slabs_full) ||
2317                         !list_empty(&n->slabs_partial);
2318         }
2319         return (ret ? 1 : 0);
2320 }
2321 
2322 #ifdef CONFIG_MEMCG
2323 void __kmemcg_cache_deactivate(struct kmem_cache *cachep)
2324 {
2325         __kmem_cache_shrink(cachep);
2326 }
2327 #endif
2328 
2329 int __kmem_cache_shutdown(struct kmem_cache *cachep)
2330 {
2331         return __kmem_cache_shrink(cachep);
2332 }
2333 
2334 void __kmem_cache_release(struct kmem_cache *cachep)
2335 {
2336         int i;
2337         struct kmem_cache_node *n;
2338 
2339         cache_random_seq_destroy(cachep);
2340 
2341         free_percpu(cachep->cpu_cache);
2342 
2343         /* NUMA: free the node structures */
2344         for_each_kmem_cache_node(cachep, i, n) {
2345                 kfree(n->shared);
2346                 free_alien_cache(n->alien);
2347                 kfree(n);
2348                 cachep->node[i] = NULL;
2349         }
2350 }
2351 
2352 /*
2353  * Get the memory for a slab management obj.
2354  *
2355  * For a slab cache when the slab descriptor is off-slab, the
2356  * slab descriptor can't come from the same cache which is being created,
2357  * Because if it is the case, that means we defer the creation of
2358  * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2359  * And we eventually call down to __kmem_cache_create(), which
2360  * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2361  * This is a "chicken-and-egg" problem.
2362  *
2363  * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2364  * which are all initialized during kmem_cache_init().
2365  */
2366 static void *alloc_slabmgmt(struct kmem_cache *cachep,
2367                                    struct page *page, int colour_off,
2368                                    gfp_t local_flags, int nodeid)
2369 {
2370         void *freelist;
2371         void *addr = page_address(page);
2372 
2373         page->s_mem = addr + colour_off;
2374         page->active = 0;
2375 
2376         if (OBJFREELIST_SLAB(cachep))
2377                 freelist = NULL;
2378         else if (OFF_SLAB(cachep)) {
2379                 /* Slab management obj is off-slab. */
2380                 freelist = kmem_cache_alloc_node(cachep->freelist_cache,
2381                                               local_flags, nodeid);
2382                 if (!freelist)
2383                         return NULL;
2384         } else {
2385                 /* We will use last bytes at the slab for freelist */
2386                 freelist = addr + (PAGE_SIZE << cachep->gfporder) -
2387                                 cachep->freelist_size;
2388         }
2389 
2390         return freelist;
2391 }
2392 
2393 static inline freelist_idx_t get_free_obj(struct page *page, unsigned int idx)
2394 {
2395         return ((freelist_idx_t *)page->freelist)[idx];
2396 }
2397 
2398 static inline void set_free_obj(struct page *page,
2399                                         unsigned int idx, freelist_idx_t val)
2400 {
2401         ((freelist_idx_t *)(page->freelist))[idx] = val;
2402 }
2403 
2404 static void cache_init_objs_debug(struct kmem_cache *cachep, struct page *page)
2405 {
2406 #if DEBUG
2407         int i;
2408 
2409         for (i = 0; i < cachep->num; i++) {
2410                 void *objp = index_to_obj(cachep, page, i);
2411 
2412                 if (cachep->flags & SLAB_STORE_USER)
2413                         *dbg_userword(cachep, objp) = NULL;
2414 
2415                 if (cachep->flags & SLAB_RED_ZONE) {
2416                         *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2417                         *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2418                 }
2419                 /*
2420                  * Constructors are not allowed to allocate memory from the same
2421                  * cache which they are a constructor for.  Otherwise, deadlock.
2422                  * They must also be threaded.
2423                  */
2424                 if (cachep->ctor && !(cachep->flags & SLAB_POISON)) {
2425                         kasan_unpoison_object_data(cachep,
2426                                                    objp + obj_offset(cachep));
2427                         cachep->ctor(objp + obj_offset(cachep));
2428                         kasan_poison_object_data(
2429                                 cachep, objp + obj_offset(cachep));
2430                 }
2431 
2432                 if (cachep->flags & SLAB_RED_ZONE) {
2433                         if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2434                                 slab_error(cachep, "constructor overwrote the end of an object");
2435                         if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2436                                 slab_error(cachep, "constructor overwrote the start of an object");
2437                 }
2438                 /* need to poison the objs? */
2439                 if (cachep->flags & SLAB_POISON) {
2440                         poison_obj(cachep, objp, POISON_FREE);
2441                         slab_kernel_map(cachep, objp, 0, 0);
2442                 }
2443         }
2444 #endif
2445 }
2446 
2447 #ifdef CONFIG_SLAB_FREELIST_RANDOM
2448 /* Hold information during a freelist initialization */
2449 union freelist_init_state {
2450         struct {
2451                 unsigned int pos;
2452                 unsigned int *list;
2453                 unsigned int count;
2454         };
2455         struct rnd_state rnd_state;
2456 };
2457 
2458 /*
2459  * Initialize the state based on the randomization methode available.
2460  * return true if the pre-computed list is available, false otherwize.
2461  */
2462 static bool freelist_state_initialize(union freelist_init_state *state,
2463                                 struct kmem_cache *cachep,
2464                                 unsigned int count)
2465 {
2466         bool ret;
2467         unsigned int rand;
2468 
2469         /* Use best entropy available to define a random shift */
2470         rand = get_random_int();
2471 
2472         /* Use a random state if the pre-computed list is not available */
2473         if (!cachep->random_seq) {
2474                 prandom_seed_state(&state->rnd_state, rand);
2475                 ret = false;
2476         } else {
2477                 state->list = cachep->random_seq;
2478                 state->count = count;
2479                 state->pos = rand % count;
2480                 ret = true;
2481         }
2482         return ret;
2483 }
2484 
2485 /* Get the next entry on the list and randomize it using a random shift */
2486 static freelist_idx_t next_random_slot(union freelist_init_state *state)
2487 {
2488         if (state->pos >= state->count)
2489                 state->pos = 0;
2490         return state->list[state->pos++];
2491 }
2492 
2493 /* Swap two freelist entries */
2494 static void swap_free_obj(struct page *page, unsigned int a, unsigned int b)
2495 {
2496         swap(((freelist_idx_t *)page->freelist)[a],
2497                 ((freelist_idx_t *)page->freelist)[b]);
2498 }
2499 
2500 /*
2501  * Shuffle the freelist initialization state based on pre-computed lists.
2502  * return true if the list was successfully shuffled, false otherwise.
2503  */
2504 static bool shuffle_freelist(struct kmem_cache *cachep, struct page *page)
2505 {
2506         unsigned int objfreelist = 0, i, rand, count = cachep->num;
2507         union freelist_init_state state;
2508         bool precomputed;
2509 
2510         if (count < 2)
2511                 return false;
2512 
2513         precomputed = freelist_state_initialize(&state, cachep, count);
2514 
2515         /* Take a random entry as the objfreelist */
2516         if (OBJFREELIST_SLAB(cachep)) {
2517                 if (!precomputed)
2518                         objfreelist = count - 1;
2519                 else
2520                         objfreelist = next_random_slot(&state);
2521                 page->freelist = index_to_obj(cachep, page, objfreelist) +
2522                                                 obj_offset(cachep);
2523                 count--;
2524         }
2525 
2526         /*
2527          * On early boot, generate the list dynamically.
2528          * Later use a pre-computed list for speed.
2529          */
2530         if (!precomputed) {
2531                 for (i = 0; i < count; i++)
2532                         set_free_obj(page, i, i);
2533 
2534                 /* Fisher-Yates shuffle */
2535                 for (i = count - 1; i > 0; i--) {
2536                         rand = prandom_u32_state(&state.rnd_state);
2537                         rand %= (i + 1);
2538                         swap_free_obj(page, i, rand);
2539                 }
2540         } else {
2541                 for (i = 0; i < count; i++)
2542                         set_free_obj(page, i, next_random_slot(&state));
2543         }
2544 
2545         if (OBJFREELIST_SLAB(cachep))
2546                 set_free_obj(page, cachep->num - 1, objfreelist);
2547 
2548         return true;
2549 }
2550 #else
2551 static inline bool shuffle_freelist(struct kmem_cache *cachep,
2552                                 struct page *page)
2553 {
2554         return false;
2555 }
2556 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
2557 
2558 static void cache_init_objs(struct kmem_cache *cachep,
2559                             struct page *page)
2560 {
2561         int i;
2562         void *objp;
2563         bool shuffled;
2564 
2565         cache_init_objs_debug(cachep, page);
2566 
2567         /* Try to randomize the freelist if enabled */
2568         shuffled = shuffle_freelist(cachep, page);
2569 
2570         if (!shuffled && OBJFREELIST_SLAB(cachep)) {
2571                 page->freelist = index_to_obj(cachep, page, cachep->num - 1) +
2572                                                 obj_offset(cachep);
2573         }
2574 
2575         for (i = 0; i < cachep->num; i++) {
2576                 objp = index_to_obj(cachep, page, i);
2577                 kasan_init_slab_obj(cachep, objp);
2578 
2579                 /* constructor could break poison info */
2580                 if (DEBUG == 0 && cachep->ctor) {
2581                         kasan_unpoison_object_data(cachep, objp);
2582                         cachep->ctor(objp);
2583                         kasan_poison_object_data(cachep, objp);
2584                 }
2585 
2586                 if (!shuffled)
2587                         set_free_obj(page, i, i);
2588         }
2589 }
2590 
2591 static void *slab_get_obj(struct kmem_cache *cachep, struct page *page)
2592 {
2593         void *objp;
2594 
2595         objp = index_to_obj(cachep, page, get_free_obj(page, page->active));
2596         page->active++;
2597 
2598 #if DEBUG
2599         if (cachep->flags & SLAB_STORE_USER)
2600                 set_store_user_dirty(cachep);
2601 #endif
2602 
2603         return objp;
2604 }
2605 
2606 static void slab_put_obj(struct kmem_cache *cachep,
2607                         struct page *page, void *objp)
2608 {
2609         unsigned int objnr = obj_to_index(cachep, page, objp);
2610 #if DEBUG
2611         unsigned int i;
2612 
2613         /* Verify double free bug */
2614         for (i = page->active; i < cachep->num; i++) {
2615                 if (get_free_obj(page, i) == objnr) {
2616                         pr_err("slab: double free detected in cache '%s', objp %px\n",
2617                                cachep->name, objp);
2618                         BUG();
2619                 }
2620         }
2621 #endif
2622         page->active--;
2623         if (!page->freelist)
2624                 page->freelist = objp + obj_offset(cachep);
2625 
2626         set_free_obj(page, page->active, objnr);
2627 }
2628 
2629 /*
2630  * Map pages beginning at addr to the given cache and slab. This is required
2631  * for the slab allocator to be able to lookup the cache and slab of a
2632  * virtual address for kfree, ksize, and slab debugging.
2633  */
2634 static void slab_map_pages(struct kmem_cache *cache, struct page *page,
2635                            void *freelist)
2636 {
2637         page->slab_cache = cache;
2638         page->freelist = freelist;
2639 }
2640 
2641 /*
2642  * Grow (by 1) the number of slabs within a cache.  This is called by
2643  * kmem_cache_alloc() when there are no active objs left in a cache.
2644  */
2645 static struct page *cache_grow_begin(struct kmem_cache *cachep,
2646                                 gfp_t flags, int nodeid)
2647 {
2648         void *freelist;
2649         size_t offset;
2650         gfp_t local_flags;
2651         int page_node;
2652         struct kmem_cache_node *n;
2653         struct page *page;
2654 
2655         /*
2656          * Be lazy and only check for valid flags here,  keeping it out of the
2657          * critical path in kmem_cache_alloc().
2658          */
2659         if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
2660                 gfp_t invalid_mask = flags & GFP_SLAB_BUG_MASK;
2661                 flags &= ~GFP_SLAB_BUG_MASK;
2662                 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n",
2663                                 invalid_mask, &invalid_mask, flags, &flags);
2664                 dump_stack();
2665         }
2666         WARN_ON_ONCE(cachep->ctor && (flags & __GFP_ZERO));
2667         local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2668 
2669         check_irq_off();
2670         if (gfpflags_allow_blocking(local_flags))
2671                 local_irq_enable();
2672 
2673         /*
2674          * Get mem for the objs.  Attempt to allocate a physical page from
2675          * 'nodeid'.
2676          */
2677         page = kmem_getpages(cachep, local_flags, nodeid);
2678         if (!page)
2679                 goto failed;
2680 
2681         page_node = page_to_nid(page);
2682         n = get_node(cachep, page_node);
2683 
2684         /* Get colour for the slab, and cal the next value. */
2685         n->colour_next++;
2686         if (n->colour_next >= cachep->colour)
2687                 n->colour_next = 0;
2688 
2689         offset = n->colour_next;
2690         if (offset >= cachep->colour)
2691                 offset = 0;
2692 
2693         offset *= cachep->colour_off;
2694 
2695         /* Get slab management. */
2696         freelist = alloc_slabmgmt(cachep, page, offset,
2697                         local_flags & ~GFP_CONSTRAINT_MASK, page_node);
2698         if (OFF_SLAB(cachep) && !freelist)
2699                 goto opps1;
2700 
2701         slab_map_pages(cachep, page, freelist);
2702 
2703         kasan_poison_slab(page);
2704         cache_init_objs(cachep, page);
2705 
2706         if (gfpflags_allow_blocking(local_flags))
2707                 local_irq_disable();
2708 
2709         return page;
2710 
2711 opps1:
2712         kmem_freepages(cachep, page);
2713 failed:
2714         if (gfpflags_allow_blocking(local_flags))
2715                 local_irq_disable();
2716         return NULL;
2717 }
2718 
2719 static void cache_grow_end(struct kmem_cache *cachep, struct page *page)
2720 {
2721         struct kmem_cache_node *n;
2722         void *list = NULL;
2723 
2724         check_irq_off();
2725 
2726         if (!page)
2727                 return;
2728 
2729         INIT_LIST_HEAD(&page->lru);
2730         n = get_node(cachep, page_to_nid(page));
2731 
2732         spin_lock(&n->list_lock);
2733         n->total_slabs++;
2734         if (!page->active) {
2735                 list_add_tail(&page->lru, &(n->slabs_free));
2736                 n->free_slabs++;
2737         } else
2738                 fixup_slab_list(cachep, n, page, &list);
2739 
2740         STATS_INC_GROWN(cachep);
2741         n->free_objects += cachep->num - page->active;
2742         spin_unlock(&n->list_lock);
2743 
2744         fixup_objfreelist_debug(cachep, &list);
2745 }
2746 
2747 #if DEBUG
2748 
2749 /*
2750  * Perform extra freeing checks:
2751  * - detect bad pointers.
2752  * - POISON/RED_ZONE checking
2753  */
2754 static void kfree_debugcheck(const void *objp)
2755 {
2756         if (!virt_addr_valid(objp)) {
2757                 pr_err("kfree_debugcheck: out of range ptr %lxh\n",
2758                        (unsigned long)objp);
2759                 BUG();
2760         }
2761 }
2762 
2763 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2764 {
2765         unsigned long long redzone1, redzone2;
2766 
2767         redzone1 = *dbg_redzone1(cache, obj);
2768         redzone2 = *dbg_redzone2(cache, obj);
2769 
2770         /*
2771          * Redzone is ok.
2772          */
2773         if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2774                 return;
2775 
2776         if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2777                 slab_error(cache, "double free detected");
2778         else
2779                 slab_error(cache, "memory outside object was overwritten");
2780 
2781         pr_err("%px: redzone 1:0x%llx, redzone 2:0x%llx\n",
2782                obj, redzone1, redzone2);
2783 }
2784 
2785 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2786                                    unsigned long caller)
2787 {
2788         unsigned int objnr;
2789         struct page *page;
2790 
2791         BUG_ON(virt_to_cache(objp) != cachep);
2792 
2793         objp -= obj_offset(cachep);
2794         kfree_debugcheck(objp);
2795         page = virt_to_head_page(objp);
2796 
2797         if (cachep->flags & SLAB_RED_ZONE) {
2798                 verify_redzone_free(cachep, objp);
2799                 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2800                 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2801         }
2802         if (cachep->flags & SLAB_STORE_USER) {
2803                 set_store_user_dirty(cachep);
2804                 *dbg_userword(cachep, objp) = (void *)caller;
2805         }
2806 
2807         objnr = obj_to_index(cachep, page, objp);
2808 
2809         BUG_ON(objnr >= cachep->num);
2810         BUG_ON(objp != index_to_obj(cachep, page, objnr));
2811 
2812         if (cachep->flags & SLAB_POISON) {
2813                 poison_obj(cachep, objp, POISON_FREE);
2814                 slab_kernel_map(cachep, objp, 0, caller);
2815         }
2816         return objp;
2817 }
2818 
2819 #else
2820 #define kfree_debugcheck(x) do { } while(0)
2821 #define cache_free_debugcheck(x,objp,z) (objp)
2822 #endif
2823 
2824 static inline void fixup_objfreelist_debug(struct kmem_cache *cachep,
2825                                                 void **list)
2826 {
2827 #if DEBUG
2828         void *next = *list;
2829         void *objp;
2830 
2831         while (next) {
2832                 objp = next - obj_offset(cachep);
2833                 next = *(void **)next;
2834                 poison_obj(cachep, objp, POISON_FREE);
2835         }
2836 #endif
2837 }
2838 
2839 static inline void fixup_slab_list(struct kmem_cache *cachep,
2840                                 struct kmem_cache_node *n, struct page *page,
2841                                 void **list)
2842 {
2843         /* move slabp to correct slabp list: */
2844         list_del(&page->lru);
2845         if (page->active == cachep->num) {
2846                 list_add(&page->lru, &n->slabs_full);
2847                 if (OBJFREELIST_SLAB(cachep)) {
2848 #if DEBUG
2849                         /* Poisoning will be done without holding the lock */
2850                         if (cachep->flags & SLAB_POISON) {
2851                                 void **objp = page->freelist;
2852 
2853                                 *objp = *list;
2854                                 *list = objp;
2855                         }
2856 #endif
2857                         page->freelist = NULL;
2858                 }
2859         } else
2860                 list_add(&page->lru, &n->slabs_partial);
2861 }
2862 
2863 /* Try to find non-pfmemalloc slab if needed */
2864 static noinline struct page *get_valid_first_slab(struct kmem_cache_node *n,
2865                                         struct page *page, bool pfmemalloc)
2866 {
2867         if (!page)
2868                 return NULL;
2869 
2870         if (pfmemalloc)
2871                 return page;
2872 
2873         if (!PageSlabPfmemalloc(page))
2874                 return page;
2875 
2876         /* No need to keep pfmemalloc slab if we have enough free objects */
2877         if (n->free_objects > n->free_limit) {
2878                 ClearPageSlabPfmemalloc(page);
2879                 return page;
2880         }
2881 
2882         /* Move pfmemalloc slab to the end of list to speed up next search */
2883         list_del(&page->lru);
2884         if (!page->active) {
2885                 list_add_tail(&page->lru, &n->slabs_free);
2886                 n->free_slabs++;
2887         } else
2888                 list_add_tail(&page->lru, &n->slabs_partial);
2889 
2890         list_for_each_entry(page, &n->slabs_partial, lru) {
2891                 if (!PageSlabPfmemalloc(page))
2892                         return page;
2893         }
2894 
2895         n->free_touched = 1;
2896         list_for_each_entry(page, &n->slabs_free, lru) {
2897                 if (!PageSlabPfmemalloc(page)) {
2898                         n->free_slabs--;
2899                         return page;
2900                 }
2901         }
2902 
2903         return NULL;
2904 }
2905 
2906 static struct page *get_first_slab(struct kmem_cache_node *n, bool pfmemalloc)
2907 {
2908         struct page *page;
2909 
2910         assert_spin_locked(&n->list_lock);
2911         page = list_first_entry_or_null(&n->slabs_partial, struct page, lru);
2912         if (!page) {
2913                 n->free_touched = 1;
2914                 page = list_first_entry_or_null(&n->slabs_free, struct page,
2915                                                 lru);
2916                 if (page)
2917                         n->free_slabs--;
2918         }
2919 
2920         if (sk_memalloc_socks())
2921                 page = get_valid_first_slab(n, page, pfmemalloc);
2922 
2923         return page;
2924 }
2925 
2926 static noinline void *cache_alloc_pfmemalloc(struct kmem_cache *cachep,
2927                                 struct kmem_cache_node *n, gfp_t flags)
2928 {
2929         struct page *page;
2930         void *obj;
2931         void *list = NULL;
2932 
2933         if (!gfp_pfmemalloc_allowed(flags))
2934                 return NULL;
2935 
2936         spin_lock(&n->list_lock);
2937         page = get_first_slab(n, true);
2938         if (!page) {
2939                 spin_unlock(&n->list_lock);
2940                 return NULL;
2941         }
2942 
2943         obj = slab_get_obj(cachep, page);
2944         n->free_objects--;
2945 
2946         fixup_slab_list(cachep, n, page, &list);
2947 
2948         spin_unlock(&n->list_lock);
2949         fixup_objfreelist_debug(cachep, &list);
2950 
2951         return obj;
2952 }
2953 
2954 /*
2955  * Slab list should be fixed up by fixup_slab_list() for existing slab
2956  * or cache_grow_end() for new slab
2957  */
2958 static __always_inline int alloc_block(struct kmem_cache *cachep,
2959                 struct array_cache *ac, struct page *page, int batchcount)
2960 {
2961         /*
2962          * There must be at least one object available for
2963          * allocation.
2964          */
2965         BUG_ON(page->active >= cachep->num);
2966 
2967         while (page->active < cachep->num && batchcount--) {
2968                 STATS_INC_ALLOCED(cachep);
2969                 STATS_INC_ACTIVE(cachep);
2970                 STATS_SET_HIGH(cachep);
2971 
2972                 ac->entry[ac->avail++] = slab_get_obj(cachep, page);
2973         }
2974 
2975         return batchcount;
2976 }
2977 
2978 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
2979 {
2980         int batchcount;
2981         struct kmem_cache_node *n;
2982         struct array_cache *ac, *shared;
2983         int node;
2984         void *list = NULL;
2985         struct page *page;
2986 
2987         check_irq_off();
2988         node = numa_mem_id();
2989 
2990         ac = cpu_cache_get(cachep);
2991         batchcount = ac->batchcount;
2992         if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2993                 /*
2994                  * If there was little recent activity on this cache, then
2995                  * perform only a partial refill.  Otherwise we could generate
2996                  * refill bouncing.
2997                  */
2998                 batchcount = BATCHREFILL_LIMIT;
2999         }
3000         n = get_node(cachep, node);
3001 
3002         BUG_ON(ac->avail > 0 || !n);
3003         shared = READ_ONCE(n->shared);
3004         if (!n->free_objects && (!shared || !shared->avail))
3005                 goto direct_grow;
3006 
3007         spin_lock(&n->list_lock);
3008         shared = READ_ONCE(n->shared);
3009 
3010         /* See if we can refill from the shared array */
3011         if (shared && transfer_objects(ac, shared, batchcount)) {
3012                 shared->touched = 1;
3013                 goto alloc_done;
3014         }
3015 
3016         while (batchcount > 0) {
3017                 /* Get slab alloc is to come from. */
3018                 page = get_first_slab(n, false);
3019                 if (!page)
3020                         goto must_grow;
3021 
3022                 check_spinlock_acquired(cachep);
3023 
3024                 batchcount = alloc_block(cachep, ac, page, batchcount);
3025                 fixup_slab_list(cachep, n, page, &list);
3026         }
3027 
3028 must_grow:
3029         n->free_objects -= ac->avail;
3030 alloc_done:
3031         spin_unlock(&n->list_lock);
3032         fixup_objfreelist_debug(cachep, &list);
3033 
3034 direct_grow:
3035         if (unlikely(!ac->avail)) {
3036                 /* Check if we can use obj in pfmemalloc slab */
3037                 if (sk_memalloc_socks()) {
3038                         void *obj = cache_alloc_pfmemalloc(cachep, n, flags);
3039 
3040                         if (obj)
3041                                 return obj;
3042                 }
3043 
3044                 page = cache_grow_begin(cachep, gfp_exact_node(flags), node);
3045 
3046                 /*
3047                  * cache_grow_begin() can reenable interrupts,
3048                  * then ac could change.
3049                  */
3050                 ac = cpu_cache_get(cachep);
3051                 if (!ac->avail && page)
3052                         alloc_block(cachep, ac, page, batchcount);
3053                 cache_grow_end(cachep, page);
3054 
3055                 if (!ac->avail)
3056                         return NULL;
3057         }
3058         ac->touched = 1;
3059 
3060         return ac->entry[--ac->avail];
3061 }
3062 
3063 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3064                                                 gfp_t flags)
3065 {
3066         might_sleep_if(gfpflags_allow_blocking(flags));
3067 }
3068 
3069 #if DEBUG
3070 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3071                                 gfp_t flags, void *objp, unsigned long caller)
3072 {
3073         WARN_ON_ONCE(cachep->ctor && (flags & __GFP_ZERO));
3074         if (!objp)
3075                 return objp;
3076         if (cachep->flags & SLAB_POISON) {
3077                 check_poison_obj(cachep, objp);
3078                 slab_kernel_map(cachep, objp, 1, 0);
3079                 poison_obj(cachep, objp, POISON_INUSE);
3080         }
3081         if (cachep->flags & SLAB_STORE_USER)
3082                 *dbg_userword(cachep, objp) = (void *)caller;
3083 
3084         if (cachep->flags & SLAB_RED_ZONE) {
3085                 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3086                                 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3087                         slab_error(cachep, "double free, or memory outside object was overwritten");
3088                         pr_err("%px: redzone 1:0x%llx, redzone 2:0x%llx\n",
3089                                objp, *dbg_redzone1(cachep, objp),
3090                                *dbg_redzone2(cachep, objp));
3091                 }
3092                 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3093                 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3094         }
3095 
3096         objp += obj_offset(cachep);
3097         if (cachep->ctor && cachep->flags & SLAB_POISON)
3098                 cachep->ctor(objp);
3099         if (ARCH_SLAB_MINALIGN &&
3100             ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
3101                 pr_err("0x%px: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3102                        objp, (int)ARCH_SLAB_MINALIGN);
3103         }
3104         return objp;
3105 }
3106 #else
3107 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3108 #endif
3109 
3110 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3111 {
3112         void *objp;
3113         struct array_cache *ac;
3114 
3115         check_irq_off();
3116 
3117         ac = cpu_cache_get(cachep);
3118         if (likely(ac->avail)) {
3119                 ac->touched = 1;
3120                 objp = ac->entry[--ac->avail];
3121 
3122                 STATS_INC_ALLOCHIT(cachep);
3123                 goto out;
3124         }
3125 
3126         STATS_INC_ALLOCMISS(cachep);
3127         objp = cache_alloc_refill(cachep, flags);
3128         /*
3129          * the 'ac' may be updated by cache_alloc_refill(),
3130          * and kmemleak_erase() requires its correct value.
3131          */
3132         ac = cpu_cache_get(cachep);
3133 
3134 out:
3135         /*
3136          * To avoid a false negative, if an object that is in one of the
3137          * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3138          * treat the array pointers as a reference to the object.
3139          */
3140         if (objp)
3141                 kmemleak_erase(&ac->entry[ac->avail]);
3142         return objp;
3143 }
3144 
3145 #ifdef CONFIG_NUMA
3146 /*
3147  * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
3148  *
3149  * If we are in_interrupt, then process context, including cpusets and
3150  * mempolicy, may not apply and should not be used for allocation policy.
3151  */
3152 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3153 {
3154         int nid_alloc, nid_here;
3155 
3156         if (in_interrupt() || (flags & __GFP_THISNODE))
3157                 return NULL;
3158         nid_alloc = nid_here = numa_mem_id();
3159         if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3160                 nid_alloc = cpuset_slab_spread_node();
3161         else if (current->mempolicy)
3162                 nid_alloc = mempolicy_slab_node();
3163         if (nid_alloc != nid_here)
3164                 return ____cache_alloc_node(cachep, flags, nid_alloc);
3165         return NULL;
3166 }
3167 
3168 /*
3169  * Fallback function if there was no memory available and no objects on a
3170  * certain node and fall back is permitted. First we scan all the
3171  * available node for available objects. If that fails then we
3172  * perform an allocation without specifying a node. This allows the page
3173  * allocator to do its reclaim / fallback magic. We then insert the
3174  * slab into the proper nodelist and then allocate from it.
3175  */
3176 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3177 {
3178         struct zonelist *zonelist;
3179         struct zoneref *z;
3180         struct zone *zone;
3181         enum zone_type high_zoneidx = gfp_zone(flags);
3182         void *obj = NULL;
3183         struct page *page;
3184         int nid;
3185         unsigned int cpuset_mems_cookie;
3186 
3187         if (flags & __GFP_THISNODE)
3188                 return NULL;
3189 
3190 retry_cpuset:
3191         cpuset_mems_cookie = read_mems_allowed_begin();
3192         zonelist = node_zonelist(mempolicy_slab_node(), flags);
3193 
3194 retry:
3195         /*
3196          * Look through allowed nodes for objects available
3197          * from existing per node queues.
3198          */
3199         for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3200                 nid = zone_to_nid(zone);
3201 
3202                 if (cpuset_zone_allowed(zone, flags) &&
3203                         get_node(cache, nid) &&
3204                         get_node(cache, nid)->free_objects) {
3205                                 obj = ____cache_alloc_node(cache,
3206                                         gfp_exact_node(flags), nid);
3207                                 if (obj)
3208                                         break;
3209                 }
3210         }
3211 
3212         if (!obj) {
3213                 /*
3214                  * This allocation will be performed within the constraints
3215                  * of the current cpuset / memory policy requirements.
3216                  * We may trigger various forms of reclaim on the allowed
3217                  * set and go into memory reserves if necessary.
3218                  */
3219                 page = cache_grow_begin(cache, flags, numa_mem_id());
3220                 cache_grow_end(cache, page);
3221                 if (page) {
3222                         nid = page_to_nid(page);
3223                         obj = ____cache_alloc_node(cache,
3224                                 gfp_exact_node(flags), nid);
3225 
3226                         /*
3227                          * Another processor may allocate the objects in
3228                          * the slab since we are not holding any locks.
3229                          */
3230                         if (!obj)
3231                                 goto retry;
3232                 }
3233         }
3234 
3235         if (unlikely(!obj && read_mems_allowed_retry(cpuset_mems_cookie)))
3236                 goto retry_cpuset;
3237         return obj;
3238 }
3239 
3240 /*
3241  * A interface to enable slab creation on nodeid
3242  */
3243 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3244                                 int nodeid)
3245 {
3246         struct page *page;
3247         struct kmem_cache_node *n;
3248         void *obj = NULL;
3249         void *list = NULL;
3250 
3251         VM_BUG_ON(nodeid < 0 || nodeid >= MAX_NUMNODES);
3252         n = get_node(cachep, nodeid);
3253         BUG_ON(!n);
3254 
3255         check_irq_off();
3256         spin_lock(&n->list_lock);
3257         page = get_first_slab(n, false);
3258         if (!page)
3259                 goto must_grow;
3260 
3261         check_spinlock_acquired_node(cachep, nodeid);
3262 
3263         STATS_INC_NODEALLOCS(cachep);
3264         STATS_INC_ACTIVE(cachep);
3265         STATS_SET_HIGH(cachep);
3266 
3267         BUG_ON(page->active == cachep->num);
3268 
3269         obj = slab_get_obj(cachep, page);
3270         n->free_objects--;
3271 
3272         fixup_slab_list(cachep, n, page, &list);
3273 
3274         spin_unlock(&n->list_lock);
3275         fixup_objfreelist_debug(cachep, &list);
3276         return obj;
3277 
3278 must_grow:
3279         spin_unlock(&n->list_lock);
3280         page = cache_grow_begin(cachep, gfp_exact_node(flags), nodeid);
3281         if (page) {
3282                 /* This slab isn't counted yet so don't update free_objects */
3283                 obj = slab_get_obj(cachep, page);
3284         }
3285         cache_grow_end(cachep, page);
3286 
3287         return obj ? obj : fallback_alloc(cachep, flags);
3288 }
3289 
3290 static __always_inline void *
3291 slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3292                    unsigned long caller)
3293 {
3294         unsigned long save_flags;
3295         void *ptr;
3296         int slab_node = numa_mem_id();
3297 
3298         flags &= gfp_allowed_mask;
3299         cachep = slab_pre_alloc_hook(cachep, flags);
3300         if (unlikely(!cachep))
3301                 return NULL;
3302 
3303         cache_alloc_debugcheck_before(cachep, flags);
3304         local_irq_save(save_flags);
3305 
3306         if (nodeid == NUMA_NO_NODE)
3307                 nodeid = slab_node;
3308 
3309         if (unlikely(!get_node(cachep, nodeid))) {
3310                 /* Node not bootstrapped yet */
3311                 ptr = fallback_alloc(cachep, flags);
3312                 goto out;
3313         }
3314 
3315         if (nodeid == slab_node) {
3316                 /*
3317                  * Use the locally cached objects if possible.
3318                  * However ____cache_alloc does not allow fallback
3319                  * to other nodes. It may fail while we still have
3320                  * objects on other nodes available.
3321                  */
3322                 ptr = ____cache_alloc(cachep, flags);
3323                 if (ptr)
3324                         goto out;
3325         }
3326         /* ___cache_alloc_node can fall back to other nodes */
3327         ptr = ____cache_alloc_node(cachep, flags, nodeid);
3328   out:
3329         local_irq_restore(save_flags);
3330         ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3331 
3332         if (unlikely(flags & __GFP_ZERO) && ptr)
3333                 memset(ptr, 0, cachep->object_size);
3334 
3335         slab_post_alloc_hook(cachep, flags, 1, &ptr);
3336         return ptr;
3337 }
3338 
3339 static __always_inline void *
3340 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3341 {
3342         void *objp;
3343 
3344         if (current->mempolicy || cpuset_do_slab_mem_spread()) {
3345                 objp = alternate_node_alloc(cache, flags);
3346                 if (objp)
3347                         goto out;
3348         }
3349         objp = ____cache_alloc(cache, flags);
3350 
3351         /*
3352          * We may just have run out of memory on the local node.
3353          * ____cache_alloc_node() knows how to locate memory on other nodes
3354          */
3355         if (!objp)
3356                 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3357 
3358   out:
3359         return objp;
3360 }
3361 #else
3362 
3363 static __always_inline void *
3364 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3365 {
3366         return ____cache_alloc(cachep, flags);
3367 }
3368 
3369 #endif /* CONFIG_NUMA */
3370 
3371 static __always_inline void *
3372 slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
3373 {
3374         unsigned long save_flags;
3375         void *objp;
3376 
3377         flags &= gfp_allowed_mask;
3378         cachep = slab_pre_alloc_hook(cachep, flags);
3379         if (unlikely(!cachep))
3380                 return NULL;
3381 
3382         cache_alloc_debugcheck_before(cachep, flags);
3383         local_irq_save(save_flags);
3384         objp = __do_cache_alloc(cachep, flags);
3385         local_irq_restore(save_flags);
3386         objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3387         prefetchw(objp);
3388 
3389         if (unlikely(flags & __GFP_ZERO) && objp)
3390                 memset(objp, 0, cachep->object_size);
3391 
3392         slab_post_alloc_hook(cachep, flags, 1, &objp);
3393         return objp;
3394 }
3395 
3396 /*
3397  * Caller needs to acquire correct kmem_cache_node's list_lock
3398  * @list: List of detached free slabs should be freed by caller
3399  */
3400 static void free_block(struct kmem_cache *cachep, void **objpp,
3401                         int nr_objects, int node, struct list_head *list)
3402 {
3403         int i;
3404         struct kmem_cache_node *n = get_node(cachep, node);
3405         struct page *page;
3406 
3407         n->free_objects += nr_objects;
3408 
3409         for (i = 0; i < nr_objects; i++) {
3410                 void *objp;
3411                 struct page *page;
3412 
3413                 objp = objpp[i];
3414 
3415                 page = virt_to_head_page(objp);
3416                 list_del(&page->lru);
3417                 check_spinlock_acquired_node(cachep, node);
3418                 slab_put_obj(cachep, page, objp);
3419                 STATS_DEC_ACTIVE(cachep);
3420 
3421                 /* fixup slab chains */
3422                 if (page->active == 0) {
3423                         list_add(&page->lru, &n->slabs_free);
3424                         n->free_slabs++;
3425                 } else {
3426                         /* Unconditionally move a slab to the end of the
3427                          * partial list on free - maximum time for the
3428                          * other objects to be freed, too.
3429                          */
3430                         list_add_tail(&page->lru, &n->slabs_partial);
3431                 }
3432         }
3433 
3434         while (n->free_objects > n->free_limit && !list_empty(&n->slabs_free)) {
3435                 n->free_objects -= cachep->num;
3436 
3437                 page = list_last_entry(&n->slabs_free, struct page, lru);
3438                 list_move(&page->lru, list);
3439                 n->free_slabs--;
3440                 n->total_slabs--;
3441         }
3442 }
3443 
3444 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3445 {
3446         int batchcount;
3447         struct kmem_cache_node *n;
3448         int node = numa_mem_id();
3449         LIST_HEAD(list);
3450 
3451         batchcount = ac->batchcount;
3452 
3453         check_irq_off();
3454         n = get_node(cachep, node);
3455         spin_lock(&n->list_lock);
3456         if (n->shared) {
3457                 struct array_cache *shared_array = n->shared;
3458                 int max = shared_array->limit - shared_array->avail;
3459                 if (max) {
3460                         if (batchcount > max)
3461                                 batchcount = max;
3462                         memcpy(&(shared_array->entry[shared_array->avail]),
3463                                ac->entry, sizeof(void *) * batchcount);
3464                         shared_array->avail += batchcount;
3465                         goto free_done;
3466                 }
3467         }
3468 
3469         free_block(cachep, ac->entry, batchcount, node, &list);
3470 free_done:
3471 #if STATS
3472         {
3473                 int i = 0;
3474                 struct page *page;
3475 
3476                 list_for_each_entry(page, &n->slabs_free, lru) {
3477                         BUG_ON(page->active);
3478 
3479                         i++;
3480                 }
3481                 STATS_SET_FREEABLE(cachep, i);
3482         }
3483 #endif
3484         spin_unlock(&n->list_lock);
3485         slabs_destroy(cachep, &list);
3486         ac->avail -= batchcount;
3487         memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3488 }
3489 
3490 /*
3491  * Release an obj back to its cache. If the obj has a constructed state, it must
3492  * be in this state _before_ it is released.  Called with disabled ints.
3493  */
3494 static __always_inline void __cache_free(struct kmem_cache *cachep, void *objp,
3495                                          unsigned long caller)
3496 {
3497         /* Put the object into the quarantine, don't touch it for now. */
3498         if (kasan_slab_free(cachep, objp, _RET_IP_))
3499                 return;
3500 
3501         ___cache_free(cachep, objp, caller);
3502 }
3503 
3504 void ___cache_free(struct kmem_cache *cachep, void *objp,
3505                 unsigned long caller)
3506 {
3507         struct array_cache *ac = cpu_cache_get(cachep);
3508 
3509         check_irq_off();
3510         kmemleak_free_recursive(objp, cachep->flags);
3511         objp = cache_free_debugcheck(cachep, objp, caller);
3512 
3513         /*
3514          * Skip calling cache_free_alien() when the platform is not numa.
3515          * This will avoid cache misses that happen while accessing slabp (which
3516          * is per page memory  reference) to get nodeid. Instead use a global
3517          * variable to skip the call, which is mostly likely to be present in
3518          * the cache.
3519          */
3520         if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3521                 return;
3522 
3523         if (ac->avail < ac->limit) {
3524                 STATS_INC_FREEHIT(cachep);
3525         } else {
3526                 STATS_INC_FREEMISS(cachep);
3527                 cache_flusharray(cachep, ac);
3528         }
3529 
3530         if (sk_memalloc_socks()) {
3531                 struct page *page = virt_to_head_page(objp);
3532 
3533                 if (unlikely(PageSlabPfmemalloc(page))) {
3534                         cache_free_pfmemalloc(cachep, page, objp);
3535                         return;
3536                 }
3537         }
3538 
3539         ac->entry[ac->avail++] = objp;
3540 }
3541 
3542 /**
3543  * kmem_cache_alloc - Allocate an object
3544  * @cachep: The cache to allocate from.
3545  * @flags: See kmalloc().
3546  *
3547  * Allocate an object from this cache.  The flags are only relevant
3548  * if the cache has no available objects.
3549  */
3550 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3551 {
3552         void *ret = slab_alloc(cachep, flags, _RET_IP_);
3553 
3554         kasan_slab_alloc(cachep, ret, flags);
3555         trace_kmem_cache_alloc(_RET_IP_, ret,
3556                                cachep->object_size, cachep->size, flags);
3557 
3558         return ret;
3559 }
3560 EXPORT_SYMBOL(kmem_cache_alloc);
3561 
3562 static __always_inline void
3563 cache_alloc_debugcheck_after_bulk(struct kmem_cache *s, gfp_t flags,
3564                                   size_t size, void **p, unsigned long caller)
3565 {
3566         size_t i;
3567 
3568         for (i = 0; i < size; i++)
3569                 p[i] = cache_alloc_debugcheck_after(s, flags, p[i], caller);
3570 }
3571 
3572 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3573                           void **p)
3574 {
3575         size_t i;
3576 
3577         s = slab_pre_alloc_hook(s, flags);
3578         if (!s)
3579                 return 0;
3580 
3581         cache_alloc_debugcheck_before(s, flags);
3582 
3583         local_irq_disable();
3584         for (i = 0; i < size; i++) {
3585                 void *objp = __do_cache_alloc(s, flags);
3586 
3587                 if (unlikely(!objp))
3588                         goto error;
3589                 p[i] = objp;
3590         }
3591         local_irq_enable();
3592 
3593         cache_alloc_debugcheck_after_bulk(s, flags, size, p, _RET_IP_);
3594 
3595         /* Clear memory outside IRQ disabled section */
3596         if (unlikely(flags & __GFP_ZERO))
3597                 for (i = 0; i < size; i++)
3598                         memset(p[i], 0, s->object_size);
3599 
3600         slab_post_alloc_hook(s, flags, size, p);
3601         /* FIXME: Trace call missing. Christoph would like a bulk variant */
3602         return size;
3603 error:
3604         local_irq_enable();
3605         cache_alloc_debugcheck_after_bulk(s, flags, i, p, _RET_IP_);
3606         slab_post_alloc_hook(s, flags, i, p);
3607         __kmem_cache_free_bulk(s, i, p);
3608         return 0;
3609 }
3610 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3611 
3612 #ifdef CONFIG_TRACING
3613 void *
3614 kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
3615 {
3616         void *ret;
3617 
3618         ret = slab_alloc(cachep, flags, _RET_IP_);
3619 
3620         kasan_kmalloc(cachep, ret, size, flags);
3621         trace_kmalloc(_RET_IP_, ret,
3622                       size, cachep->size, flags);
3623         return ret;
3624 }
3625 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3626 #endif
3627 
3628 #ifdef CONFIG_NUMA
3629 /**
3630  * kmem_cache_alloc_node - Allocate an object on the specified node
3631  * @cachep: The cache to allocate from.
3632  * @flags: See kmalloc().
3633  * @nodeid: node number of the target node.
3634  *
3635  * Identical to kmem_cache_alloc but it will allocate memory on the given
3636  * node, which can improve the performance for cpu bound structures.
3637  *
3638  * Fallback to other node is possible if __GFP_THISNODE is not set.
3639  */
3640 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3641 {
3642         void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3643 
3644         kasan_slab_alloc(cachep, ret, flags);
3645         trace_kmem_cache_alloc_node(_RET_IP_, ret,
3646                                     cachep->object_size, cachep->size,
3647                                     flags, nodeid);
3648 
3649         return ret;
3650 }
3651 EXPORT_SYMBOL(kmem_cache_alloc_node);
3652 
3653 #ifdef CONFIG_TRACING
3654 void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
3655                                   gfp_t flags,
3656                                   int nodeid,
3657                                   size_t size)
3658 {
3659         void *ret;
3660 
3661         ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3662 
3663         kasan_kmalloc(cachep, ret, size, flags);
3664         trace_kmalloc_node(_RET_IP_, ret,
3665                            size, cachep->size,
3666                            flags, nodeid);
3667         return ret;
3668 }
3669 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3670 #endif
3671 
3672 static __always_inline void *
3673 __do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
3674 {
3675         struct kmem_cache *cachep;
3676         void *ret;
3677 
3678         cachep = kmalloc_slab(size, flags);
3679         if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3680                 return cachep;
3681         ret = kmem_cache_alloc_node_trace(cachep, flags, node, size);
3682         kasan_kmalloc(cachep, ret, size, flags);
3683 
3684         return ret;
3685 }
3686 
3687 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3688 {
3689         return __do_kmalloc_node(size, flags, node, _RET_IP_);
3690 }
3691 EXPORT_SYMBOL(__kmalloc_node);
3692 
3693 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3694                 int node, unsigned long caller)
3695 {
3696         return __do_kmalloc_node(size, flags, node, caller);
3697 }
3698 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3699 #endif /* CONFIG_NUMA */
3700 
3701 /**
3702  * __do_kmalloc - allocate memory
3703  * @size: how many bytes of memory are required.
3704  * @flags: the type of memory to allocate (see kmalloc).
3705  * @caller: function caller for debug tracking of the caller
3706  */
3707 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3708                                           unsigned long caller)
3709 {
3710         struct kmem_cache *cachep;
3711         void *ret;
3712 
3713         cachep = kmalloc_slab(size, flags);
3714         if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3715                 return cachep;
3716         ret = slab_alloc(cachep, flags, caller);
3717 
3718         kasan_kmalloc(cachep, ret, size, flags);
3719         trace_kmalloc(caller, ret,
3720                       size, cachep->size, flags);
3721 
3722         return ret;
3723 }
3724 
3725 void *__kmalloc(size_t size, gfp_t flags)
3726 {
3727         return __do_kmalloc(size, flags, _RET_IP_);
3728 }
3729 EXPORT_SYMBOL(__kmalloc);
3730 
3731 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3732 {
3733         return __do_kmalloc(size, flags, caller);
3734 }
3735 EXPORT_SYMBOL(__kmalloc_track_caller);
3736 
3737 /**
3738  * kmem_cache_free - Deallocate an object
3739  * @cachep: The cache the allocation was from.
3740  * @objp: The previously allocated object.
3741  *
3742  * Free an object which was previously allocated from this
3743  * cache.
3744  */
3745 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3746 {
3747         unsigned long flags;
3748         cachep = cache_from_obj(cachep, objp);
3749         if (!cachep)
3750                 return;
3751 
3752         local_irq_save(flags);
3753         debug_check_no_locks_freed(objp, cachep->object_size);
3754         if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3755                 debug_check_no_obj_freed(objp, cachep->object_size);
3756         __cache_free(cachep, objp, _RET_IP_);
3757         local_irq_restore(flags);
3758 
3759         trace_kmem_cache_free(_RET_IP_, objp);
3760 }
3761 EXPORT_SYMBOL(kmem_cache_free);
3762 
3763 void kmem_cache_free_bulk(struct kmem_cache *orig_s, size_t size, void **p)
3764 {
3765         struct kmem_cache *s;
3766         size_t i;
3767 
3768         local_irq_disable();
3769         for (i = 0; i < size; i++) {
3770                 void *objp = p[i];
3771 
3772                 if (!orig_s) /* called via kfree_bulk */
3773                         s = virt_to_cache(objp);
3774                 else
3775                         s = cache_from_obj(orig_s, objp);
3776 
3777                 debug_check_no_locks_freed(objp, s->object_size);
3778                 if (!(s->flags & SLAB_DEBUG_OBJECTS))
3779                         debug_check_no_obj_freed(objp, s->object_size);
3780 
3781                 __cache_free(s, objp, _RET_IP_);
3782         }
3783         local_irq_enable();
3784 
3785         /* FIXME: add tracing */
3786 }
3787 EXPORT_SYMBOL(kmem_cache_free_bulk);
3788 
3789 /**
3790  * kfree - free previously allocated memory
3791  * @objp: pointer returned by kmalloc.
3792  *
3793  * If @objp is NULL, no operation is performed.
3794  *
3795  * Don't free memory not originally allocated by kmalloc()
3796  * or you will run into trouble.
3797  */
3798 void kfree(const void *objp)
3799 {
3800         struct kmem_cache *c;
3801         unsigned long flags;
3802 
3803         trace_kfree(_RET_IP_, objp);
3804 
3805         if (unlikely(ZERO_OR_NULL_PTR(objp)))
3806                 return;
3807         local_irq_save(flags);
3808         kfree_debugcheck(objp);
3809         c = virt_to_cache(objp);
3810         debug_check_no_locks_freed(objp, c->object_size);
3811 
3812         debug_check_no_obj_freed(objp, c->object_size);
3813         __cache_free(c, (void *)objp, _RET_IP_);
3814         local_irq_restore(flags);
3815 }
3816 EXPORT_SYMBOL(kfree);
3817 
3818 /*
3819  * This initializes kmem_cache_node or resizes various caches for all nodes.
3820  */
3821 static int setup_kmem_cache_nodes(struct kmem_cache *cachep, gfp_t gfp)
3822 {
3823         int ret;
3824         int node;
3825         struct kmem_cache_node *n;
3826 
3827         for_each_online_node(node) {
3828                 ret = setup_kmem_cache_node(cachep, node, gfp, true);
3829                 if (ret)
3830                         goto fail;
3831 
3832         }
3833 
3834         return 0;
3835 
3836 fail:
3837         if (!cachep->list.next) {
3838                 /* Cache is not active yet. Roll back what we did */
3839                 node--;
3840                 while (node >= 0) {
3841                         n = get_node(cachep, node);
3842                         if (n) {
3843                                 kfree(n->shared);
3844                                 free_alien_cache(n->alien);
3845                                 kfree(n);
3846                                 cachep->node[node] = NULL;
3847                         }
3848                         node--;
3849                 }
3850         }
3851         return -ENOMEM;
3852 }
3853 
3854 /* Always called with the slab_mutex held */
3855 static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
3856                                 int batchcount, int shared, gfp_t gfp)
3857 {
3858         struct array_cache __percpu *cpu_cache, *prev;
3859         int cpu;
3860 
3861         cpu_cache = alloc_kmem_cache_cpus(cachep, limit, batchcount);
3862         if (!cpu_cache)
3863                 return -ENOMEM;
3864 
3865         prev = cachep->cpu_cache;
3866         cachep->cpu_cache = cpu_cache;
3867         /*
3868          * Without a previous cpu_cache there's no need to synchronize remote
3869          * cpus, so skip the IPIs.
3870          */
3871         if (prev)
3872                 kick_all_cpus_sync();
3873 
3874         check_irq_on();
3875         cachep->batchcount = batchcount;
3876         cachep->limit = limit;
3877         cachep->shared = shared;
3878 
3879         if (!prev)
3880                 goto setup_node;
3881 
3882         for_each_online_cpu(cpu) {
3883                 LIST_HEAD(list);
3884                 int node;
3885                 struct kmem_cache_node *n;
3886                 struct array_cache *ac = per_cpu_ptr(prev, cpu);
3887 
3888                 node = cpu_to_mem(cpu);
3889                 n = get_node(cachep, node);
3890                 spin_lock_irq(&n->list_lock);
3891                 free_block(cachep, ac->entry, ac->avail, node, &list);
3892                 spin_unlock_irq(&n->list_lock);
3893                 slabs_destroy(cachep, &list);
3894         }
3895         free_percpu(prev);
3896 
3897 setup_node:
3898         return setup_kmem_cache_nodes(cachep, gfp);
3899 }
3900 
3901 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3902                                 int batchcount, int shared, gfp_t gfp)
3903 {
3904         int ret;
3905         struct kmem_cache *c;
3906 
3907         ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3908 
3909         if (slab_state < FULL)
3910                 return ret;
3911 
3912         if ((ret < 0) || !is_root_cache(cachep))
3913                 return ret;
3914 
3915         lockdep_assert_held(&slab_mutex);
3916         for_each_memcg_cache(c, cachep) {
3917                 /* return value determined by the root cache only */
3918                 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
3919         }
3920 
3921         return ret;
3922 }
3923 
3924 /* Called with slab_mutex held always */
3925 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
3926 {
3927         int err;
3928         int limit = 0;
3929         int shared = 0;
3930         int batchcount = 0;
3931 
3932         err = cache_random_seq_create(cachep, cachep->num, gfp);
3933         if (err)
3934                 goto end;
3935 
3936         if (!is_root_cache(cachep)) {
3937                 struct kmem_cache *root = memcg_root_cache(cachep);
3938                 limit = root->limit;
3939                 shared = root->shared;
3940                 batchcount = root->batchcount;
3941         }
3942 
3943         if (limit && shared && batchcount)
3944                 goto skip_setup;
3945         /*
3946          * The head array serves three purposes:
3947          * - create a LIFO ordering, i.e. return objects that are cache-warm
3948          * - reduce the number of spinlock operations.
3949          * - reduce the number of linked list operations on the slab and
3950          *   bufctl chains: array operations are cheaper.
3951          * The numbers are guessed, we should auto-tune as described by
3952          * Bonwick.
3953          */
3954         if (cachep->size > 131072)
3955                 limit = 1;
3956         else if (cachep->size > PAGE_SIZE)
3957                 limit = 8;
3958         else if (cachep->size > 1024)
3959                 limit = 24;
3960         else if (cachep->size > 256)
3961                 limit = 54;
3962         else
3963                 limit = 120;
3964 
3965         /*
3966          * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3967          * allocation behaviour: Most allocs on one cpu, most free operations
3968          * on another cpu. For these cases, an efficient object passing between
3969          * cpus is necessary. This is provided by a shared array. The array
3970          * replaces Bonwick's magazine layer.
3971          * On uniprocessor, it's functionally equivalent (but less efficient)
3972          * to a larger limit. Thus disabled by default.
3973          */
3974         shared = 0;
3975         if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
3976                 shared = 8;
3977 
3978 #if DEBUG
3979         /*
3980          * With debugging enabled, large batchcount lead to excessively long
3981          * periods with disabled local interrupts. Limit the batchcount
3982          */
3983         if (limit > 32)
3984                 limit = 32;
3985 #endif
3986         batchcount = (limit + 1) / 2;
3987 skip_setup:
3988         err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3989 end:
3990         if (err)
3991                 pr_err("enable_cpucache failed for %s, error %d\n",
3992                        cachep->name, -err);
3993         return err;
3994 }
3995 
3996 /*
3997  * Drain an array if it contains any elements taking the node lock only if
3998  * necessary. Note that the node listlock also protects the array_cache
3999  * if drain_array() is used on the shared array.
4000  */
4001 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
4002                          struct array_cache *ac, int node)
4003 {
4004         LIST_HEAD(list);
4005 
4006         /* ac from n->shared can be freed if we don't hold the slab_mutex. */
4007         check_mutex_acquired();
4008 
4009         if (!ac || !ac->avail)
4010                 return;
4011 
4012         if (ac->touched) {
4013                 ac->touched = 0;
4014                 return;
4015         }
4016 
4017         spin_lock_irq(&n->list_lock);
4018         drain_array_locked(cachep, ac, node, false, &list);
4019         spin_unlock_irq(&n->list_lock);
4020 
4021         slabs_destroy(cachep, &list);
4022 }
4023 
4024 /**
4025  * cache_reap - Reclaim memory from caches.
4026  * @w: work descriptor
4027  *
4028  * Called from workqueue/eventd every few seconds.
4029  * Purpose:
4030  * - clear the per-cpu caches for this CPU.
4031  * - return freeable pages to the main free memory pool.
4032  *
4033  * If we cannot acquire the cache chain mutex then just give up - we'll try
4034  * again on the next iteration.
4035  */
4036 static void cache_reap(struct work_struct *w)
4037 {
4038         struct kmem_cache *searchp;
4039         struct kmem_cache_node *n;
4040         int node = numa_mem_id();
4041         struct delayed_work *work = to_delayed_work(w);
4042 
4043         if (!mutex_trylock(&slab_mutex))
4044                 /* Give up. Setup the next iteration. */
4045                 goto out;
4046 
4047         list_for_each_entry(searchp, &slab_caches, list) {
4048                 check_irq_on();
4049 
4050                 /*
4051                  * We only take the node lock if absolutely necessary and we
4052                  * have established with reasonable certainty that
4053                  * we can do some work if the lock was obtained.
4054                  */
4055                 n = get_node(searchp, node);
4056 
4057                 reap_alien(searchp, n);
4058 
4059                 drain_array(searchp, n, cpu_cache_get(searchp), node);
4060 
4061                 /*
4062                  * These are racy checks but it does not matter
4063                  * if we skip one check or scan twice.
4064                  */
4065                 if (time_after(n->next_reap, jiffies))
4066                         goto next;
4067 
4068                 n->next_reap = jiffies + REAPTIMEOUT_NODE;
4069 
4070                 drain_array(searchp, n, n->shared, node);
4071 
4072                 if (n->free_touched)
4073                         n->free_touched = 0;
4074                 else {
4075                         int freed;
4076 
4077                         freed = drain_freelist(searchp, n, (n->free_limit +
4078                                 5 * searchp->num - 1) / (5 * searchp->num));
4079                         STATS_ADD_REAPED(searchp, freed);
4080                 }
4081 next:
4082                 cond_resched();
4083         }
4084         check_irq_on();
4085         mutex_unlock(&slab_mutex);
4086         next_reap_node();
4087 out:
4088         /* Set up the next iteration */
4089         schedule_delayed_work_on(smp_processor_id(), work,
4090                                 round_jiffies_relative(REAPTIMEOUT_AC));
4091 }
4092 
4093 void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
4094 {
4095         unsigned long active_objs, num_objs, active_slabs;
4096         unsigned long total_slabs = 0, free_objs = 0, shared_avail = 0;
4097         unsigned long free_slabs = 0;
4098         int node;
4099         struct kmem_cache_node *n;
4100 
4101         for_each_kmem_cache_node(cachep, node, n) {
4102                 check_irq_on();
4103                 spin_lock_irq(&n->list_lock);
4104 
4105                 total_slabs += n->total_slabs;
4106                 free_slabs += n->free_slabs;
4107                 free_objs += n->free_objects;
4108 
4109                 if (n->shared)
4110                         shared_avail += n->shared->avail;
4111 
4112                 spin_unlock_irq(&n->list_lock);
4113         }
4114         num_objs = total_slabs * cachep->num;
4115         active_slabs = total_slabs - free_slabs;
4116         active_objs = num_objs - free_objs;
4117 
4118         sinfo->active_objs = active_objs;
4119         sinfo->num_objs = num_objs;
4120         sinfo->active_slabs = active_slabs;
4121         sinfo->num_slabs = total_slabs;
4122         sinfo->shared_avail = shared_avail;
4123         sinfo->limit = cachep->limit;
4124         sinfo->batchcount = cachep->batchcount;
4125         sinfo->shared = cachep->shared;
4126         sinfo->objects_per_slab = cachep->num;
4127         sinfo->cache_order = cachep->gfporder;
4128 }
4129 
4130 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4131 {
4132 #if STATS
4133         {                       /* node stats */
4134                 unsigned long high = cachep->high_mark;
4135                 unsigned long allocs = cachep->num_allocations;
4136                 unsigned long grown = cachep->grown;
4137                 unsigned long reaped = cachep->reaped;
4138                 unsigned long errors = cachep->errors;
4139                 unsigned long max_freeable = cachep->max_freeable;
4140                 unsigned long node_allocs = cachep->node_allocs;
4141                 unsigned long node_frees = cachep->node_frees;
4142                 unsigned long overflows = cachep->node_overflow;
4143 
4144                 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu %4lu %4lu",
4145                            allocs, high, grown,
4146                            reaped, errors, max_freeable, node_allocs,
4147                            node_frees, overflows);
4148         }
4149         /* cpu stats */
4150         {
4151                 unsigned long allochit = atomic_read(&cachep->allochit);
4152                 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4153                 unsigned long freehit = atomic_read(&cachep->freehit);
4154                 unsigned long freemiss = atomic_read(&cachep->freemiss);
4155 
4156                 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4157                            allochit, allocmiss, freehit, freemiss);
4158         }
4159 #endif
4160 }
4161 
4162 #define MAX_SLABINFO_WRITE 128
4163 /**
4164  * slabinfo_write - Tuning for the slab allocator
4165  * @file: unused
4166  * @buffer: user buffer
4167  * @count: data length
4168  * @ppos: unused
4169  */
4170 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4171                        size_t count, loff_t *ppos)
4172 {
4173         char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4174         int limit, batchcount, shared, res;
4175         struct kmem_cache *cachep;
4176 
4177         if (count > MAX_SLABINFO_WRITE)
4178                 return -EINVAL;
4179         if (copy_from_user(&kbuf, buffer, count))
4180                 return -EFAULT;
4181         kbuf[MAX_SLABINFO_WRITE] = '\0';
4182 
4183         tmp = strchr(kbuf, ' ');
4184         if (!tmp)
4185                 return -EINVAL;
4186         *tmp = '\0';
4187         tmp++;
4188         if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4189                 return -EINVAL;
4190 
4191         /* Find the cache in the chain of caches. */
4192         mutex_lock(&slab_mutex);
4193         res = -EINVAL;
4194         list_for_each_entry(cachep, &slab_caches, list) {
4195                 if (!strcmp(cachep->name, kbuf)) {
4196                         if (limit < 1 || batchcount < 1 ||
4197                                         batchcount > limit || shared < 0) {
4198                                 res = 0;
4199                         } else {
4200                                 res = do_tune_cpucache(cachep, limit,
4201                                                        batchcount, shared,
4202                                                        GFP_KERNEL);
4203                         }
4204                         break;
4205                 }
4206         }
4207         mutex_unlock(&slab_mutex);
4208         if (res >= 0)
4209                 res = count;
4210         return res;
4211 }
4212 
4213 #ifdef CONFIG_DEBUG_SLAB_LEAK
4214 
4215 static inline int add_caller(unsigned long *n, unsigned long v)
4216 {
4217         unsigned long *p;
4218         int l;
4219         if (!v)
4220                 return 1;
4221         l = n[1];
4222         p = n + 2;
4223         while (l) {
4224                 int i = l/2;
4225                 unsigned long *q = p + 2 * i;
4226                 if (*q == v) {
4227                         q[1]++;
4228                         return 1;
4229                 }
4230                 if (*q > v) {
4231                         l = i;
4232                 } else {
4233                         p = q + 2;
4234                         l -= i + 1;
4235                 }
4236         }
4237         if (++n[1] == n[0])
4238                 return 0;
4239         memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4240         p[0] = v;
4241         p[1] = 1;
4242         return 1;
4243 }
4244 
4245 static void handle_slab(unsigned long *n, struct kmem_cache *c,
4246                                                 struct page *page)
4247 {
4248         void *p;
4249         int i, j;
4250         unsigned long v;
4251 
4252         if (n[0] == n[1])
4253                 return;
4254         for (i = 0, p = page->s_mem; i < c->num; i++, p += c->size) {
4255                 bool active = true;
4256 
4257                 for (j = page->active; j < c->num; j++) {
4258                         if (get_free_obj(page, j) == i) {
4259                                 active = false;
4260                                 break;
4261                         }
4262                 }
4263 
4264                 if (!active)
4265                         continue;
4266 
4267                 /*
4268                  * probe_kernel_read() is used for DEBUG_PAGEALLOC. page table
4269                  * mapping is established when actual object allocation and
4270                  * we could mistakenly access the unmapped object in the cpu
4271                  * cache.
4272                  */
4273                 if (probe_kernel_read(&v, dbg_userword(c, p), sizeof(v)))
4274                         continue;
4275 
4276                 if (!add_caller(n, v))
4277                         return;
4278         }
4279 }
4280 
4281 static void show_symbol(struct seq_file *m, unsigned long address)
4282 {
4283 #ifdef CONFIG_KALLSYMS
4284         unsigned long offset, size;
4285         char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4286 
4287         if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4288                 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4289                 if (modname[0])
4290                         seq_printf(m, " [%s]", modname);
4291                 return;
4292         }
4293 #endif
4294         seq_printf(m, "%px", (void *)address);
4295 }
4296 
4297 static int leaks_show(struct seq_file *m, void *p)
4298 {
4299         struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4300         struct page *page;
4301         struct kmem_cache_node *n;
4302         const char *name;
4303         unsigned long *x = m->private;
4304         int node;
4305         int i;
4306 
4307         if (!(cachep->flags & SLAB_STORE_USER))
4308                 return 0;
4309         if (!(cachep->flags & SLAB_RED_ZONE))
4310                 return 0;
4311 
4312         /*
4313          * Set store_user_clean and start to grab stored user information
4314          * for all objects on this cache. If some alloc/free requests comes
4315          * during the processing, information would be wrong so restart
4316          * whole processing.
4317          */
4318         do {
4319                 set_store_user_clean(cachep);
4320                 drain_cpu_caches(cachep);
4321 
4322                 x[1] = 0;
4323 
4324                 for_each_kmem_cache_node(cachep, node, n) {
4325 
4326                         check_irq_on();
4327                         spin_lock_irq(&n->list_lock);
4328 
4329                         list_for_each_entry(page, &n->slabs_full, lru)
4330                                 handle_slab(x, cachep, page);
4331                         list_for_each_entry(page, &n->slabs_partial, lru)
4332                                 handle_slab(x, cachep, page);
4333                         spin_unlock_irq(&n->list_lock);
4334                 }
4335         } while (!is_store_user_clean(cachep));
4336 
4337         name = cachep->name;
4338         if (x[0] == x[1]) {
4339                 /* Increase the buffer size */
4340                 mutex_unlock(&slab_mutex);
4341                 m->private = kcalloc(x[0] * 4, sizeof(unsigned long),
4342                                      GFP_KERNEL);
4343                 if (!m->private) {
4344                         /* Too bad, we are really out */
4345                         m->private = x;
4346                         mutex_lock(&slab_mutex);
4347                         return -ENOMEM;
4348                 }
4349                 *(unsigned long *)m->private = x[0] * 2;
4350                 kfree(x);
4351                 mutex_lock(&slab_mutex);
4352                 /* Now make sure this entry will be retried */
4353                 m->count = m->size;
4354                 return 0;
4355         }
4356         for (i = 0; i < x[1]; i++) {
4357                 seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4358                 show_symbol(m, x[2*i+2]);
4359                 seq_putc(m, '\n');
4360         }
4361 
4362         return 0;
4363 }
4364 
4365 static const struct seq_operations slabstats_op = {
4366         .start = slab_start,
4367         .next = slab_next,
4368         .stop = slab_stop,
4369         .show = leaks_show,
4370 };
4371 
4372 static int slabstats_open(struct inode *inode, struct file *file)
4373 {
4374         unsigned long *n;
4375 
4376         n = __seq_open_private(file, &slabstats_op, PAGE_SIZE);
4377         if (!n)
4378                 return -ENOMEM;
4379 
4380         *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4381 
4382         return 0;
4383 }
4384 
4385 static const struct file_operations proc_slabstats_operations = {
4386         .open           = slabstats_open,
4387         .read           = seq_read,
4388         .llseek         = seq_lseek,
4389         .release        = seq_release_private,
4390 };
4391 #endif
4392 
4393 static int __init slab_proc_init(void)
4394 {
4395 #ifdef CONFIG_DEBUG_SLAB_LEAK
4396         proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4397 #endif
4398         return 0;
4399 }
4400 module_init(slab_proc_init);
4401 
4402 #ifdef CONFIG_HARDENED_USERCOPY
4403 /*
4404  * Rejects incorrectly sized objects and objects that are to be copied
4405  * to/from userspace but do not fall entirely within the containing slab
4406  * cache's usercopy region.
4407  *
4408  * Returns NULL if check passes, otherwise const char * to name of cache
4409  * to indicate an error.
4410  */
4411 void __check_heap_object(const void *ptr, unsigned long n, struct page *page,
4412                          bool to_user)
4413 {
4414         struct kmem_cache *cachep;
4415         unsigned int objnr;
4416         unsigned long offset;
4417 
4418         /* Find and validate object. */
4419         cachep = page->slab_cache;
4420         objnr = obj_to_index(cachep, page, (void *)ptr);
4421         BUG_ON(objnr >= cachep->num);
4422 
4423         /* Find offset within object. */
4424         offset = ptr - index_to_obj(cachep, page, objnr) - obj_offset(cachep);
4425 
4426         /* Allow address range falling entirely within usercopy region. */
4427         if (offset >= cachep->useroffset &&
4428             offset - cachep->useroffset <= cachep->usersize &&
4429             n <= cachep->useroffset - offset + cachep->usersize)
4430                 return;
4431 
4432         /*
4433          * If the copy is still within the allocated object, produce
4434          * a warning instead of rejecting the copy. This is intended
4435          * to be a temporary method to find any missing usercopy
4436          * whitelists.
4437          */
4438         if (usercopy_fallback &&
4439             offset <= cachep->object_size &&
4440             n <= cachep->object_size - offset) {
4441                 usercopy_warn("SLAB object", cachep->name, to_user, offset, n);
4442                 return;
4443         }
4444 
4445         usercopy_abort("SLAB object", cachep->name, to_user, offset, n);
4446 }
4447 #endif /* CONFIG_HARDENED_USERCOPY */
4448 
4449 /**
4450  * ksize - get the actual amount of memory allocated for a given object
4451  * @objp: Pointer to the object
4452  *
4453  * kmalloc may internally round up allocations and return more memory
4454  * than requested. ksize() can be used to determine the actual amount of
4455  * memory allocated. The caller may use this additional memory, even though
4456  * a smaller amount of memory was initially specified with the kmalloc call.
4457  * The caller must guarantee that objp points to a valid object previously
4458  * allocated with either kmalloc() or kmem_cache_alloc(). The object
4459  * must not be freed during the duration of the call.
4460  */
4461 size_t ksize(const void *objp)
4462 {
4463         size_t size;
4464 
4465         BUG_ON(!objp);
4466         if (unlikely(objp == ZERO_SIZE_PTR))
4467                 return 0;
4468 
4469         size = virt_to_cache(objp)->object_size;
4470         /* We assume that ksize callers could use the whole allocated area,
4471          * so we need to unpoison this area.
4472          */
4473         kasan_unpoison_shadow(objp, size);
4474 
4475         return size;
4476 }
4477 EXPORT_SYMBOL(ksize);
4478 

~ [ 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