개념 설명 전체
mm/slub.c · v6.18.37
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * SLUB: A slab allocator that limits cache line use instead of queuing
4 * objects in per cpu and per node lists.
5 *
6 * The allocator synchronizes using per slab locks or atomic operations
7 * and only uses a centralized lock to manage a pool of partial slabs.
8 *
9 * (C) 2007 SGI, Christoph Lameter
10 * (C) 2011 Linux Foundation, Christoph Lameter
11 */
12
13 #include <linux/mm.h>
14 #include <linux/swap.h> /* mm_account_reclaimed_pages() */
15 #include <linux/module.h>
16 #include <linux/bit_spinlock.h>
17 #include <linux/interrupt.h>
18 #include <linux/swab.h>
19 #include <linux/bitops.h>
20 #include <linux/slab.h>
21 #include "slab.h"
22 #include <linux/vmalloc.h>
23 #include <linux/proc_fs.h>
24 #include <linux/seq_file.h>
25 #include <linux/kasan.h>
26 #include <linux/node.h>
27 #include <linux/kmsan.h>
28 #include <linux/cpu.h>
29 #include <linux/cpuset.h>
30 #include <linux/mempolicy.h>
31 #include <linux/ctype.h>
32 #include <linux/stackdepot.h>
33 #include <linux/debugobjects.h>
34 #include <linux/kallsyms.h>
35 #include <linux/kfence.h>
36 #include <linux/memory.h>
37 #include <linux/math64.h>
38 #include <linux/fault-inject.h>
39 #include <linux/kmemleak.h>
40 #include <linux/stacktrace.h>
41 #include <linux/prefetch.h>
42 #include <linux/memcontrol.h>
43 #include <linux/random.h>
44 #include <linux/prandom.h>
45 #include <kunit/test.h>
46 #include <kunit/test-bug.h>
47 #include <linux/sort.h>
48 #include <linux/irq_work.h>
49 #include <linux/kprobes.h>
50 #include <linux/debugfs.h>
51 #include <trace/events/kmem.h>
52
53 #include "internal.h"
54
55 /*
56 * Lock order:
57 * 1. slab_mutex (Global Mutex)
58 * 2. node->list_lock (Spinlock)
59 * 3. kmem_cache->cpu_slab->lock (Local lock)
60 * 4. slab_lock(slab) (Only on some arches)
61 * 5. object_map_lock (Only for debugging)
62 *
63 * slab_mutex
64 *
65 * The role of the slab_mutex is to protect the list of all the slabs
66 * and to synchronize major metadata changes to slab cache structures.
67 * Also synchronizes memory hotplug callbacks.
68 *
69 * slab_lock
70 *
71 * The slab_lock is a wrapper around the page lock, thus it is a bit
72 * spinlock.
73 *
74 * The slab_lock is only used on arches that do not have the ability
75 * to do a cmpxchg_double. It only protects:
76 *
77 * A. slab->freelist -> List of free objects in a slab
78 * B. slab->inuse -> Number of objects in use
79 * C. slab->objects -> Number of objects in slab
80 * D. slab->frozen -> frozen state
81 *
82 * Frozen slabs
83 *
84 * If a slab is frozen then it is exempt from list management. It is
85 * the cpu slab which is actively allocated from by the processor that
86 * froze it and it is not on any list. The processor that froze the
87 * slab is the one who can perform list operations on the slab. Other
88 * processors may put objects onto the freelist but the processor that
89 * froze the slab is the only one that can retrieve the objects from the
90 * slab's freelist.
91 *
92 * CPU partial slabs
93 *
94 * The partially empty slabs cached on the CPU partial list are used
95 * for performance reasons, which speeds up the allocation process.
96 * These slabs are not frozen, but are also exempt from list management,
97 * by clearing the SL_partial flag when moving out of the node
98 * partial list. Please see __slab_free() for more details.
99 *
100 * To sum up, the current scheme is:
101 * - node partial slab: SL_partial && !frozen
102 * - cpu partial slab: !SL_partial && !frozen
103 * - cpu slab: !SL_partial && frozen
104 * - full slab: !SL_partial && !frozen
105 *
106 * list_lock
107 *
108 * The list_lock protects the partial and full list on each node and
109 * the partial slab counter. If taken then no new slabs may be added or
110 * removed from the lists nor make the number of partial slabs be modified.
111 * (Note that the total number of slabs is an atomic value that may be
112 * modified without taking the list lock).
113 *
114 * The list_lock is a centralized lock and thus we avoid taking it as
115 * much as possible. As long as SLUB does not have to handle partial
116 * slabs, operations can continue without any centralized lock. F.e.
117 * allocating a long series of objects that fill up slabs does not require
118 * the list lock.
119 *
120 * For debug caches, all allocations are forced to go through a list_lock
121 * protected region to serialize against concurrent validation.
122 *
123 * cpu_slab->lock local lock
124 *
125 * This locks protect slowpath manipulation of all kmem_cache_cpu fields
126 * except the stat counters. This is a percpu structure manipulated only by
127 * the local cpu, so the lock protects against being preempted or interrupted
128 * by an irq. Fast path operations rely on lockless operations instead.
129 *
130 * On PREEMPT_RT, the local lock neither disables interrupts nor preemption
131 * which means the lockless fastpath cannot be used as it might interfere with
132 * an in-progress slow path operations. In this case the local lock is always
133 * taken but it still utilizes the freelist for the common operations.
134 *
135 * lockless fastpaths
136 *
137 * The fast path allocation (slab_alloc_node()) and freeing (do_slab_free())
138 * are fully lockless when satisfied from the percpu slab (and when
139 * cmpxchg_double is possible to use, otherwise slab_lock is taken).
140 * They also don't disable preemption or migration or irqs. They rely on
141 * the transaction id (tid) field to detect being preempted or moved to
142 * another cpu.
143 *
144 * irq, preemption, migration considerations
145 *
146 * Interrupts are disabled as part of list_lock or local_lock operations, or
147 * around the slab_lock operation, in order to make the slab allocator safe
148 * to use in the context of an irq.
149 *
150 * In addition, preemption (or migration on PREEMPT_RT) is disabled in the
151 * allocation slowpath, bulk allocation, and put_cpu_partial(), so that the
152 * local cpu doesn't change in the process and e.g. the kmem_cache_cpu pointer
153 * doesn't have to be revalidated in each section protected by the local lock.
154 *
155 * SLUB assigns one slab for allocation to each processor.
156 * Allocations only occur from these slabs called cpu slabs.
157 *
158 * Slabs with free elements are kept on a partial list and during regular
159 * operations no list for full slabs is used. If an object in a full slab is
160 * freed then the slab will show up again on the partial lists.
161 * We track full slabs for debugging purposes though because otherwise we
162 * cannot scan all objects.
163 *
164 * Slabs are freed when they become empty. Teardown and setup is
165 * minimal so we rely on the page allocators per cpu caches for
166 * fast frees and allocs.
167 *
168 * slab->frozen The slab is frozen and exempt from list processing.
169 * This means that the slab is dedicated to a purpose
170 * such as satisfying allocations for a specific
171 * processor. Objects may be freed in the slab while
172 * it is frozen but slab_free will then skip the usual
173 * list operations. It is up to the processor holding
174 * the slab to integrate the slab into the slab lists
175 * when the slab is no longer needed.
176 *
177 * One use of this flag is to mark slabs that are
178 * used for allocations. Then such a slab becomes a cpu
179 * slab. The cpu slab may be equipped with an additional
180 * freelist that allows lockless access to
181 * free objects in addition to the regular freelist
182 * that requires the slab lock.
183 *
184 * SLAB_DEBUG_FLAGS Slab requires special handling due to debug
185 * options set. This moves slab handling out of
186 * the fast path and disables lockless freelists.
187 */
188
189 /**
190 * enum slab_flags - How the slab flags bits are used.
191 * @SL_locked: Is locked with slab_lock()
192 * @SL_partial: On the per-node partial list
193 * @SL_pfmemalloc: Was allocated from PF_MEMALLOC reserves
194 *
195 * The slab flags share space with the page flags but some bits have
196 * different interpretations. The high bits are used for information
197 * like zone/node/section.
198 */
199 enum slab_flags {
200 SL_locked = PG_locked,
201 SL_partial = PG_workingset, /* Historical reasons for this bit */
202 SL_pfmemalloc = PG_active, /* Historical reasons for this bit */
203 };
204
205 /*
206 * We could simply use migrate_disable()/enable() but as long as it's a
207 * function call even on !PREEMPT_RT, use inline preempt_disable() there.
208 */
209 #ifndef CONFIG_PREEMPT_RT
210 #define slub_get_cpu_ptr(var) get_cpu_ptr(var)
211 #define slub_put_cpu_ptr(var) put_cpu_ptr(var)
212 #define USE_LOCKLESS_FAST_PATH() (true)
213 #else
214 #define slub_get_cpu_ptr(var) \
215 ({ \
216 migrate_disable(); \
217 this_cpu_ptr(var); \
218 })
219 #define slub_put_cpu_ptr(var) \
220 do { \
221 (void)(var); \
222 migrate_enable(); \
223 } while (0)
224 #define USE_LOCKLESS_FAST_PATH() (false)
225 #endif
226
227 #ifndef CONFIG_SLUB_TINY
228 #define __fastpath_inline __always_inline
229 #else
230 #define __fastpath_inline
231 #endif
232
233 #ifdef CONFIG_SLUB_DEBUG
234 #ifdef CONFIG_SLUB_DEBUG_ON
235 DEFINE_STATIC_KEY_TRUE(slub_debug_enabled);
236 #else
237 DEFINE_STATIC_KEY_FALSE(slub_debug_enabled);
238 #endif
239 #endif /* CONFIG_SLUB_DEBUG */
240
241 #ifdef CONFIG_NUMA
242 static DEFINE_STATIC_KEY_FALSE(strict_numa);
243 #endif
244
245 /* Structure holding parameters for get_partial() call chain */
246 struct partial_context {
247 gfp_t flags;
248 unsigned int orig_size;
249 void *object;
250 };
251
252 static inline bool kmem_cache_debug(struct kmem_cache *s)
253 {
254 return kmem_cache_debug_flags(s, SLAB_DEBUG_FLAGS);
255 }
256
257 void *fixup_red_left(struct kmem_cache *s, void *p)
258 {
259 if (kmem_cache_debug_flags(s, SLAB_RED_ZONE))
260 p += s->red_left_pad;
261
262 return p;
263 }
264
265 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
266 {
267 #ifdef CONFIG_SLUB_CPU_PARTIAL
268 return !kmem_cache_debug(s);
269 #else
270 return false;
271 #endif
272 }
273
274 /*
275 * Issues still to be resolved:
276 *
277 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
278 *
279 * - Variable sizing of the per node arrays
280 */
281
282 /* Enable to log cmpxchg failures */
283 #undef SLUB_DEBUG_CMPXCHG
284
285 #ifndef CONFIG_SLUB_TINY
286 /*
287 * Minimum number of partial slabs. These will be left on the partial
288 * lists even if they are empty. kmem_cache_shrink may reclaim them.
289 */
290 #define MIN_PARTIAL 5
291
292 /*
293 * Maximum number of desirable partial slabs.
294 * The existence of more partial slabs makes kmem_cache_shrink
295 * sort the partial list by the number of objects in use.
296 */
297 #define MAX_PARTIAL 10
298 #else
299 #define MIN_PARTIAL 0
300 #define MAX_PARTIAL 0
301 #endif
302
303 #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
304 SLAB_POISON | SLAB_STORE_USER)
305
306 /*
307 * These debug flags cannot use CMPXCHG because there might be consistency
308 * issues when checking or reading debug information
309 */
310 #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
311 SLAB_TRACE)
312
313
314 /*
315 * Debugging flags that require metadata to be stored in the slab. These get
316 * disabled when slab_debug=O is used and a cache's min order increases with
317 * metadata.
318 */
319 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
320
321 #define OO_SHIFT 16
322 #define OO_MASK ((1 << OO_SHIFT) - 1)
323 #define MAX_OBJS_PER_PAGE 32767 /* since slab.objects is u15 */
324
325 /* Internal SLUB flags */
326 /* Poison object */
327 #define __OBJECT_POISON __SLAB_FLAG_BIT(_SLAB_OBJECT_POISON)
328 /* Use cmpxchg_double */
329
330 #ifdef system_has_freelist_aba
331 #define __CMPXCHG_DOUBLE __SLAB_FLAG_BIT(_SLAB_CMPXCHG_DOUBLE)
332 #else
333 #define __CMPXCHG_DOUBLE __SLAB_FLAG_UNUSED
334 #endif
335
336 /*
337 * Tracking user of a slab.
338 */
339 #define TRACK_ADDRS_COUNT 16
340 struct track {
341 unsigned long addr; /* Called from address */
342 #ifdef CONFIG_STACKDEPOT
343 depot_stack_handle_t handle;
344 #endif
345 int cpu; /* Was running on cpu */
346 int pid; /* Pid context */
347 unsigned long when; /* When did the operation occur */
348 };
349
350 enum track_item { TRACK_ALLOC, TRACK_FREE };
351
352 #ifdef SLAB_SUPPORTS_SYSFS
353 static int sysfs_slab_add(struct kmem_cache *);
354 static int sysfs_slab_alias(struct kmem_cache *, const char *);
355 #else
356 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
357 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
358 { return 0; }
359 #endif
360
361 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_SLUB_DEBUG)
362 static void debugfs_slab_add(struct kmem_cache *);
363 #else
364 static inline void debugfs_slab_add(struct kmem_cache *s) { }
365 #endif
366
367 enum stat_item {
368 ALLOC_PCS, /* Allocation from percpu sheaf */
369 ALLOC_FASTPATH, /* Allocation from cpu slab */
370 ALLOC_SLOWPATH, /* Allocation by getting a new cpu slab */
371 FREE_PCS, /* Free to percpu sheaf */
372 FREE_RCU_SHEAF, /* Free to rcu_free sheaf */
373 FREE_RCU_SHEAF_FAIL, /* Failed to free to a rcu_free sheaf */
374 FREE_FASTPATH, /* Free to cpu slab */
375 FREE_SLOWPATH, /* Freeing not to cpu slab */
376 FREE_FROZEN, /* Freeing to frozen slab */
377 FREE_ADD_PARTIAL, /* Freeing moves slab to partial list */
378 FREE_REMOVE_PARTIAL, /* Freeing removes last object */
379 ALLOC_FROM_PARTIAL, /* Cpu slab acquired from node partial list */
380 ALLOC_SLAB, /* Cpu slab acquired from page allocator */
381 ALLOC_REFILL, /* Refill cpu slab from slab freelist */
382 ALLOC_NODE_MISMATCH, /* Switching cpu slab */
383 FREE_SLAB, /* Slab freed to the page allocator */
384 CPUSLAB_FLUSH, /* Abandoning of the cpu slab */
385 DEACTIVATE_FULL, /* Cpu slab was full when deactivated */
386 DEACTIVATE_EMPTY, /* Cpu slab was empty when deactivated */
387 DEACTIVATE_TO_HEAD, /* Cpu slab was moved to the head of partials */
388 DEACTIVATE_TO_TAIL, /* Cpu slab was moved to the tail of partials */
389 DEACTIVATE_REMOTE_FREES,/* Slab contained remotely freed objects */
390 DEACTIVATE_BYPASS, /* Implicit deactivation */
391 ORDER_FALLBACK, /* Number of times fallback was necessary */
392 CMPXCHG_DOUBLE_CPU_FAIL,/* Failures of this_cpu_cmpxchg_double */
393 CMPXCHG_DOUBLE_FAIL, /* Failures of slab freelist update */
394 CPU_PARTIAL_ALLOC, /* Used cpu partial on alloc */
395 CPU_PARTIAL_FREE, /* Refill cpu partial on free */
396 CPU_PARTIAL_NODE, /* Refill cpu partial from node partial */
397 CPU_PARTIAL_DRAIN, /* Drain cpu partial to node partial */
398 SHEAF_FLUSH, /* Objects flushed from a sheaf */
399 SHEAF_REFILL, /* Objects refilled to a sheaf */
400 SHEAF_ALLOC, /* Allocation of an empty sheaf */
401 SHEAF_FREE, /* Freeing of an empty sheaf */
402 BARN_GET, /* Got full sheaf from barn */
403 BARN_GET_FAIL, /* Failed to get full sheaf from barn */
404 BARN_PUT, /* Put full sheaf to barn */
405 BARN_PUT_FAIL, /* Failed to put full sheaf to barn */
406 SHEAF_PREFILL_FAST, /* Sheaf prefill grabbed the spare sheaf */
407 SHEAF_PREFILL_SLOW, /* Sheaf prefill found no spare sheaf */
408 SHEAF_PREFILL_OVERSIZE, /* Allocation of oversize sheaf for prefill */
409 SHEAF_RETURN_FAST, /* Sheaf return reattached spare sheaf */
410 SHEAF_RETURN_SLOW, /* Sheaf return could not reattach spare */
411 NR_SLUB_STAT_ITEMS
412 };
413
414 /*
415 * When changing the layout, make sure freelist and tid are still compatible
416 * with this_cpu_cmpxchg_double() alignment requirements.
417 */
418 struct kmem_cache_cpu {
419 union {
420 struct {
421 void **freelist; /* Pointer to next available object */
422 unsigned long tid; /* Globally unique transaction id */
423 };
424 freelist_aba_t freelist_tid;
425 };
426 struct slab *slab; /* The slab from which we are allocating */
427 #ifdef CONFIG_SLUB_CPU_PARTIAL
428 struct slab *partial; /* Partially allocated slabs */
429 #endif
430 local_trylock_t lock; /* Protects the fields above */
431 #ifdef CONFIG_SLUB_STATS
432 unsigned int stat[NR_SLUB_STAT_ITEMS];
433 #endif
434 };
435
436 static inline void stat(const struct kmem_cache *s, enum stat_item si)
437 {
438 #ifdef CONFIG_SLUB_STATS
439 /*
440 * The rmw is racy on a preemptible kernel but this is acceptable, so
441 * avoid this_cpu_add()'s irq-disable overhead.
442 */
443 raw_cpu_inc(s->cpu_slab->stat[si]);
444 #endif
445 }
446
447 static inline
448 void stat_add(const struct kmem_cache *s, enum stat_item si, int v)
449 {
450 #ifdef CONFIG_SLUB_STATS
451 raw_cpu_add(s->cpu_slab->stat[si], v);
452 #endif
453 }
454
455 #define MAX_FULL_SHEAVES 10
456 #define MAX_EMPTY_SHEAVES 10
457
458 struct node_barn {
459 spinlock_t lock;
460 struct list_head sheaves_full;
461 struct list_head sheaves_empty;
462 unsigned int nr_full;
463 unsigned int nr_empty;
464 };
465
466 struct slab_sheaf {
467 union {
468 struct rcu_head rcu_head;
469 struct list_head barn_list;
470 /* only used for prefilled sheafs */
471 unsigned int capacity;
472 };
473 struct kmem_cache *cache;
474 unsigned int size;
475 int node; /* only used for rcu_sheaf */
476 void *objects[];
477 };
478
479 struct slub_percpu_sheaves {
480 local_trylock_t lock;
481 struct slab_sheaf *main; /* never NULL when unlocked */
482 struct slab_sheaf *spare; /* empty or full, may be NULL */
483 struct slab_sheaf *rcu_free; /* for batching kfree_rcu() */
484 };
485
486 /*
487 * The slab lists for all objects.
488 */
489 struct kmem_cache_node {
490 spinlock_t list_lock;
491 unsigned long nr_partial;
492 struct list_head partial;
493 #ifdef CONFIG_SLUB_DEBUG
494 atomic_long_t nr_slabs;
495 atomic_long_t total_objects;
496 struct list_head full;
497 #endif
498 struct node_barn *barn;
499 };
500
501 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
502 {
503 return s->node[node];
504 }
505
506 /*
507 * Get the barn of the current cpu's closest memory node. It may not exist on
508 * systems with memoryless nodes but without CONFIG_HAVE_MEMORYLESS_NODES
509 */
510 static inline struct node_barn *get_barn(struct kmem_cache *s)
511 {
512 struct kmem_cache_node *n = get_node(s, numa_mem_id());
513
514 if (!n)
515 return NULL;
516
517 return n->barn;
518 }
519
520 /*
521 * Iterator over all nodes. The body will be executed for each node that has
522 * a kmem_cache_node structure allocated (which is true for all online nodes)
523 */
524 #define for_each_kmem_cache_node(__s, __node, __n) \
525 for (__node = 0; __node < nr_node_ids; __node++) \
526 if ((__n = get_node(__s, __node)))
527
528 /*
529 * Tracks for which NUMA nodes we have kmem_cache_nodes allocated.
530 * Corresponds to node_state[N_MEMORY], but can temporarily
531 * differ during memory hotplug/hotremove operations.
532 * Protected by slab_mutex.
533 */
534 static nodemask_t slab_nodes;
535
536 /*
537 * Workqueue used for flush_cpu_slab().
538 */
539 static struct workqueue_struct *flushwq;
540
541 struct slub_flush_work {
542 struct work_struct work;
543 struct kmem_cache *s;
544 bool skip;
545 };
546
547 static DEFINE_MUTEX(flush_lock);
548 static DEFINE_PER_CPU(struct slub_flush_work, slub_flush);
549
550 /********************************************************************
551 * Core slab cache functions
552 *******************************************************************/
553
554 /*
555 * Returns freelist pointer (ptr). With hardening, this is obfuscated
556 * with an XOR of the address where the pointer is held and a per-cache
557 * random number.
558 */
559 static inline freeptr_t freelist_ptr_encode(const struct kmem_cache *s,
560 void *ptr, unsigned long ptr_addr)
561 {
562 unsigned long encoded;
563
564 #ifdef CONFIG_SLAB_FREELIST_HARDENED
565 encoded = (unsigned long)ptr ^ s->random ^ swab(ptr_addr);
566 #else
567 encoded = (unsigned long)ptr;
568 #endif
569 return (freeptr_t){.v = encoded};
570 }
571
572 static inline void *freelist_ptr_decode(const struct kmem_cache *s,
573 freeptr_t ptr, unsigned long ptr_addr)
574 {
575 void *decoded;
576
577 #ifdef CONFIG_SLAB_FREELIST_HARDENED
578 decoded = (void *)(ptr.v ^ s->random ^ swab(ptr_addr));
579 #else
580 decoded = (void *)ptr.v;
581 #endif
582 return decoded;
583 }
584
585 static inline void *get_freepointer(struct kmem_cache *s, void *object)
586 {
587 unsigned long ptr_addr;
588 freeptr_t p;
589
590 object = kasan_reset_tag(object);
591 ptr_addr = (unsigned long)object + s->offset;
592 p = *(freeptr_t *)(ptr_addr);
593 return freelist_ptr_decode(s, p, ptr_addr);
594 }
595
596 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
597 {
598 prefetchw(object + s->offset);
599 }
600
601 /*
602 * When running under KMSAN, get_freepointer_safe() may return an uninitialized
603 * pointer value in the case the current thread loses the race for the next
604 * memory chunk in the freelist. In that case this_cpu_cmpxchg_double() in
605 * slab_alloc_node() will fail, so the uninitialized value won't be used, but
606 * KMSAN will still check all arguments of cmpxchg because of imperfect
607 * handling of inline assembly.
608 * To work around this problem, we apply __no_kmsan_checks to ensure that
609 * get_freepointer_safe() returns initialized memory.
610 */
611 __no_kmsan_checks
612 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
613 {
614 unsigned long freepointer_addr;
615 freeptr_t p;
616
617 if (!debug_pagealloc_enabled_static())
618 return get_freepointer(s, object);
619
620 object = kasan_reset_tag(object);
621 freepointer_addr = (unsigned long)object + s->offset;
622 copy_from_kernel_nofault(&p, (freeptr_t *)freepointer_addr, sizeof(p));
623 return freelist_ptr_decode(s, p, freepointer_addr);
624 }
625
626 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
627 {
628 unsigned long freeptr_addr = (unsigned long)object + s->offset;
629
630 #ifdef CONFIG_SLAB_FREELIST_HARDENED
631 BUG_ON(object == fp); /* naive detection of double free or corruption */
632 #endif
633
634 freeptr_addr = (unsigned long)kasan_reset_tag((void *)freeptr_addr);
635 *(freeptr_t *)freeptr_addr = freelist_ptr_encode(s, fp, freeptr_addr);
636 }
637
638 /*
639 * See comment in calculate_sizes().
640 */
641 static inline bool freeptr_outside_object(struct kmem_cache *s)
642 {
643 return s->offset >= s->inuse;
644 }
645
646 /*
647 * Return offset of the end of info block which is inuse + free pointer if
648 * not overlapping with object.
649 */
650 static inline unsigned int get_info_end(struct kmem_cache *s)
651 {
652 if (freeptr_outside_object(s))
653 return s->inuse + sizeof(void *);
654 else
655 return s->inuse;
656 }
657
658 /* Loop over all objects in a slab */
659 #define for_each_object(__p, __s, __addr, __objects) \
660 for (__p = fixup_red_left(__s, __addr); \
661 __p < (__addr) + (__objects) * (__s)->size; \
662 __p += (__s)->size)
663
664 static inline unsigned int order_objects(unsigned int order, unsigned int size)
665 {
666 return ((unsigned int)PAGE_SIZE << order) / size;
667 }
668
669 static inline struct kmem_cache_order_objects oo_make(unsigned int order,
670 unsigned int size)
671 {
672 struct kmem_cache_order_objects x = {
673 (order << OO_SHIFT) + order_objects(order, size)
674 };
675
676 return x;
677 }
678
679 static inline unsigned int oo_order(struct kmem_cache_order_objects x)
680 {
681 return x.x >> OO_SHIFT;
682 }
683
684 static inline unsigned int oo_objects(struct kmem_cache_order_objects x)
685 {
686 return x.x & OO_MASK;
687 }
688
689 #ifdef CONFIG_SLUB_CPU_PARTIAL
690 static void slub_set_cpu_partial(struct kmem_cache *s, unsigned int nr_objects)
691 {
692 unsigned int nr_slabs;
693
694 s->cpu_partial = nr_objects;
695
696 /*
697 * We take the number of objects but actually limit the number of
698 * slabs on the per cpu partial list, in order to limit excessive
699 * growth of the list. For simplicity we assume that the slabs will
700 * be half-full.
701 */
702 nr_slabs = DIV_ROUND_UP(nr_objects * 2, oo_objects(s->oo));
703 s->cpu_partial_slabs = nr_slabs;
704 }
705
706 static inline unsigned int slub_get_cpu_partial(struct kmem_cache *s)
707 {
708 return s->cpu_partial_slabs;
709 }
710 #else
711 #ifdef SLAB_SUPPORTS_SYSFS
712 static inline void
713 slub_set_cpu_partial(struct kmem_cache *s, unsigned int nr_objects)
714 {
715 }
716 #endif
717
718 static inline unsigned int slub_get_cpu_partial(struct kmem_cache *s)
719 {
720 return 0;
721 }
722 #endif /* CONFIG_SLUB_CPU_PARTIAL */
723
724 /*
725 * If network-based swap is enabled, slub must keep track of whether memory
726 * were allocated from pfmemalloc reserves.
727 */
728 static inline bool slab_test_pfmemalloc(const struct slab *slab)
729 {
730 return test_bit(SL_pfmemalloc, &slab->flags.f);
731 }
732
733 static inline void slab_set_pfmemalloc(struct slab *slab)
734 {
735 set_bit(SL_pfmemalloc, &slab->flags.f);
736 }
737
738 static inline void __slab_clear_pfmemalloc(struct slab *slab)
739 {
740 __clear_bit(SL_pfmemalloc, &slab->flags.f);
741 }
742
743 /*
744 * Per slab locking using the pagelock
745 */
746 static __always_inline void slab_lock(struct slab *slab)
747 {
748 bit_spin_lock(SL_locked, &slab->flags.f);
749 }
750
751 static __always_inline void slab_unlock(struct slab *slab)
752 {
753 bit_spin_unlock(SL_locked, &slab->flags.f);
754 }
755
756 static inline bool
757 __update_freelist_fast(struct slab *slab,
758 void *freelist_old, unsigned long counters_old,
759 void *freelist_new, unsigned long counters_new)
760 {
761 #ifdef system_has_freelist_aba
762 freelist_aba_t old = { .freelist = freelist_old, .counter = counters_old };
763 freelist_aba_t new = { .freelist = freelist_new, .counter = counters_new };
764
765 return try_cmpxchg_freelist(&slab->freelist_counter.full, &old.full, new.full);
766 #else
767 return false;
768 #endif
769 }
770
771 static inline bool
772 __update_freelist_slow(struct slab *slab,
773 void *freelist_old, unsigned long counters_old,
774 void *freelist_new, unsigned long counters_new)
775 {
776 bool ret = false;
777
778 slab_lock(slab);
779 if (slab->freelist == freelist_old &&
780 slab->counters == counters_old) {
781 slab->freelist = freelist_new;
782 slab->counters = counters_new;
783 ret = true;
784 }
785 slab_unlock(slab);
786
787 return ret;
788 }
789
790 /*
791 * Interrupts must be disabled (for the fallback code to work right), typically
792 * by an _irqsave() lock variant. On PREEMPT_RT the preempt_disable(), which is
793 * part of bit_spin_lock(), is sufficient because the policy is not to allow any
794 * allocation/ free operation in hardirq context. Therefore nothing can
795 * interrupt the operation.
796 */
797 static inline bool __slab_update_freelist(struct kmem_cache *s, struct slab *slab,
798 void *freelist_old, unsigned long counters_old,
799 void *freelist_new, unsigned long counters_new,
800 const char *n)
801 {
802 bool ret;
803
804 if (USE_LOCKLESS_FAST_PATH())
805 lockdep_assert_irqs_disabled();
806
807 if (s->flags & __CMPXCHG_DOUBLE) {
808 ret = __update_freelist_fast(slab, freelist_old, counters_old,
809 freelist_new, counters_new);
810 } else {
811 ret = __update_freelist_slow(slab, freelist_old, counters_old,
812 freelist_new, counters_new);
813 }
814 if (likely(ret))
815 return true;
816
817 cpu_relax();
818 stat(s, CMPXCHG_DOUBLE_FAIL);
819
820 #ifdef SLUB_DEBUG_CMPXCHG
821 pr_info("%s %s: cmpxchg double redo ", n, s->name);
822 #endif
823
824 return false;
825 }
826
827 static inline bool slab_update_freelist(struct kmem_cache *s, struct slab *slab,
828 void *freelist_old, unsigned long counters_old,
829 void *freelist_new, unsigned long counters_new,
830 const char *n)
831 {
832 bool ret;
833
834 if (s->flags & __CMPXCHG_DOUBLE) {
835 ret = __update_freelist_fast(slab, freelist_old, counters_old,
836 freelist_new, counters_new);
837 } else {
838 unsigned long flags;
839
840 local_irq_save(flags);
841 ret = __update_freelist_slow(slab, freelist_old, counters_old,
842 freelist_new, counters_new);
843 local_irq_restore(flags);
844 }
845 if (likely(ret))
846 return true;
847
848 cpu_relax();
849 stat(s, CMPXCHG_DOUBLE_FAIL);
850
851 #ifdef SLUB_DEBUG_CMPXCHG
852 pr_info("%s %s: cmpxchg double redo ", n, s->name);
853 #endif
854
855 return false;
856 }
857
858 /*
859 * kmalloc caches has fixed sizes (mostly power of 2), and kmalloc() API
860 * family will round up the real request size to these fixed ones, so
861 * there could be an extra area than what is requested. Save the original
862 * request size in the meta data area, for better debug and sanity check.
863 */
864 static inline void set_orig_size(struct kmem_cache *s,
865 void *object, unsigned long orig_size)
866 {
867 void *p = kasan_reset_tag(object);
868
869 if (!slub_debug_orig_size(s))
870 return;
871
872 p += get_info_end(s);
873 p += sizeof(struct track) * 2;
874
875 *(unsigned long *)p = orig_size;
876 }
877
878 static inline unsigned long get_orig_size(struct kmem_cache *s, void *object)
879 {
880 void *p = kasan_reset_tag(object);
881
882 if (is_kfence_address(object))
883 return kfence_ksize(object);
884
885 if (!slub_debug_orig_size(s))
886 return s->object_size;
887
888 p += get_info_end(s);
889 p += sizeof(struct track) * 2;
890
891 return *(unsigned long *)p;
892 }
893
894 #ifdef CONFIG_SLUB_DEBUG
895
896 /*
897 * For debugging context when we want to check if the struct slab pointer
898 * appears to be valid.
899 */
900 static inline bool validate_slab_ptr(struct slab *slab)
901 {
902 return PageSlab(slab_page(slab));
903 }
904
905 static unsigned long object_map[BITS_TO_LONGS(MAX_OBJS_PER_PAGE)];
906 static DEFINE_SPINLOCK(object_map_lock);
907
908 static void __fill_map(unsigned long *obj_map, struct kmem_cache *s,
909 struct slab *slab)
910 {
911 void *addr = slab_address(slab);
912 void *p;
913
914 bitmap_zero(obj_map, slab->objects);
915
916 for (p = slab->freelist; p; p = get_freepointer(s, p))
917 set_bit(__obj_to_index(s, addr, p), obj_map);
918 }
919
920 #if IS_ENABLED(CONFIG_KUNIT)
921 static bool slab_add_kunit_errors(void)
922 {
923 struct kunit_resource *resource;
924
925 if (!kunit_get_current_test())
926 return false;
927
928 resource = kunit_find_named_resource(current->kunit_test, "slab_errors");
929 if (!resource)
930 return false;
931
932 (*(int *)resource->data)++;
933 kunit_put_resource(resource);
934 return true;
935 }
936
937 bool slab_in_kunit_test(void)
938 {
939 struct kunit_resource *resource;
940
941 if (!kunit_get_current_test())
942 return false;
943
944 resource = kunit_find_named_resource(current->kunit_test, "slab_errors");
945 if (!resource)
946 return false;
947
948 kunit_put_resource(resource);
949 return true;
950 }
951 #else
952 static inline bool slab_add_kunit_errors(void) { return false; }
953 #endif
954
955 static inline unsigned int size_from_object(struct kmem_cache *s)
956 {
957 if (s->flags & SLAB_RED_ZONE)
958 return s->size - s->red_left_pad;
959
960 return s->size;
961 }
962
963 static inline void *restore_red_left(struct kmem_cache *s, void *p)
964 {
965 if (s->flags & SLAB_RED_ZONE)
966 p -= s->red_left_pad;
967
968 return p;
969 }
970
971 /*
972 * Debug settings:
973 */
974 #if defined(CONFIG_SLUB_DEBUG_ON)
975 static slab_flags_t slub_debug = DEBUG_DEFAULT_FLAGS;
976 #else
977 static slab_flags_t slub_debug;
978 #endif
979
980 static char *slub_debug_string;
981 static int disable_higher_order_debug;
982
983 /*
984 * slub is about to manipulate internal object metadata. This memory lies
985 * outside the range of the allocated object, so accessing it would normally
986 * be reported by kasan as a bounds error. metadata_access_enable() is used
987 * to tell kasan that these accesses are OK.
988 */
989 static inline void metadata_access_enable(void)
990 {
991 kasan_disable_current();
992 kmsan_disable_current();
993 }
994
995 static inline void metadata_access_disable(void)
996 {
997 kmsan_enable_current();
998 kasan_enable_current();
999 }
1000
1001 /*
1002 * Object debugging
1003 */
1004
1005 /* Verify that a pointer has an address that is valid within a slab page */
1006 static inline int check_valid_pointer(struct kmem_cache *s,
1007 struct slab *slab, void *object)
1008 {
1009 void *base;
1010
1011 if (!object)
1012 return 1;
1013
1014 base = slab_address(slab);
1015 object = kasan_reset_tag(object);
1016 object = restore_red_left(s, object);
1017 if (object < base || object >= base + slab->objects * s->size ||
1018 (object - base) % s->size) {
1019 return 0;
1020 }
1021
1022 return 1;
1023 }
1024
1025 static void print_section(char *level, char *text, u8 *addr,
1026 unsigned int length)
1027 {
1028 metadata_access_enable();
1029 print_hex_dump(level, text, DUMP_PREFIX_ADDRESS,
1030 16, 1, kasan_reset_tag((void *)addr), length, 1);
1031 metadata_access_disable();
1032 }
1033
1034 static struct track *get_track(struct kmem_cache *s, void *object,
1035 enum track_item alloc)
1036 {
1037 struct track *p;
1038
1039 p = object + get_info_end(s);
1040
1041 return kasan_reset_tag(p + alloc);
1042 }
1043
1044 #ifdef CONFIG_STACKDEPOT
1045 static noinline depot_stack_handle_t set_track_prepare(gfp_t gfp_flags)
1046 {
1047 depot_stack_handle_t handle;
1048 unsigned long entries[TRACK_ADDRS_COUNT];
1049 unsigned int nr_entries;
1050
1051 nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 3);
1052 handle = stack_depot_save(entries, nr_entries, gfp_flags);
1053
1054 return handle;
1055 }
1056 #else
1057 static inline depot_stack_handle_t set_track_prepare(gfp_t gfp_flags)
1058 {
1059 return 0;
1060 }
1061 #endif
1062
1063 static void set_track_update(struct kmem_cache *s, void *object,
1064 enum track_item alloc, unsigned long addr,
1065 depot_stack_handle_t handle)
1066 {
1067 struct track *p = get_track(s, object, alloc);
1068
1069 #ifdef CONFIG_STACKDEPOT
1070 p->handle = handle;
1071 #endif
1072 p->addr = addr;
1073 p->cpu = smp_processor_id();
1074 p->pid = current->pid;
1075 p->when = jiffies;
1076 }
1077
1078 static __always_inline void set_track(struct kmem_cache *s, void *object,
1079 enum track_item alloc, unsigned long addr, gfp_t gfp_flags)
1080 {
1081 depot_stack_handle_t handle = set_track_prepare(gfp_flags);
1082
1083 set_track_update(s, object, alloc, addr, handle);
1084 }
1085
1086 static void init_tracking(struct kmem_cache *s, void *object)
1087 {
1088 struct track *p;
1089
1090 if (!(s->flags & SLAB_STORE_USER))
1091 return;
1092
1093 p = get_track(s, object, TRACK_ALLOC);
1094 memset(p, 0, 2*sizeof(struct track));
1095 }
1096
1097 static void print_track(const char *s, struct track *t, unsigned long pr_time)
1098 {
1099 depot_stack_handle_t handle __maybe_unused;
1100
1101 if (!t->addr)
1102 return;
1103
1104 pr_err("%s in %pS age=%lu cpu=%u pid=%d\n",
1105 s, (void *)t->addr, pr_time - t->when, t->cpu, t->pid);
1106 #ifdef CONFIG_STACKDEPOT
1107 handle = READ_ONCE(t->handle);
1108 if (handle)
1109 stack_depot_print(handle);
1110 else
1111 pr_err("object allocation/free stack trace missing\n");
1112 #endif
1113 }
1114
1115 void print_tracking(struct kmem_cache *s, void *object)
1116 {
1117 unsigned long pr_time = jiffies;
1118 if (!(s->flags & SLAB_STORE_USER))
1119 return;
1120
1121 print_track("Allocated", get_track(s, object, TRACK_ALLOC), pr_time);
1122 print_track("Freed", get_track(s, object, TRACK_FREE), pr_time);
1123 }
1124
1125 static void print_slab_info(const struct slab *slab)
1126 {
1127 pr_err("Slab 0x%p objects=%u used=%u fp=0x%p flags=%pGp\n",
1128 slab, slab->objects, slab->inuse, slab->freelist,
1129 &slab->flags.f);
1130 }
1131
1132 void skip_orig_size_check(struct kmem_cache *s, const void *object)
1133 {
1134 set_orig_size(s, (void *)object, s->object_size);
1135 }
1136
1137 static void __slab_bug(struct kmem_cache *s, const char *fmt, va_list argsp)
1138 {
1139 struct va_format vaf;
1140 va_list args;
1141
1142 va_copy(args, argsp);
1143 vaf.fmt = fmt;
1144 vaf.va = &args;
1145 pr_err("=============================================================================\n");
1146 pr_err("BUG %s (%s): %pV\n", s ? s->name : "<unknown>", print_tainted(), &vaf);
1147 pr_err("-----------------------------------------------------------------------------\n\n");
1148 va_end(args);
1149 }
1150
1151 static void slab_bug(struct kmem_cache *s, const char *fmt, ...)
1152 {
1153 va_list args;
1154
1155 va_start(args, fmt);
1156 __slab_bug(s, fmt, args);
1157 va_end(args);
1158 }
1159
1160 __printf(2, 3)
1161 static void slab_fix(struct kmem_cache *s, const char *fmt, ...)
1162 {
1163 struct va_format vaf;
1164 va_list args;
1165
1166 if (slab_add_kunit_errors())
1167 return;
1168
1169 va_start(args, fmt);
1170 vaf.fmt = fmt;
1171 vaf.va = &args;
1172 pr_err("FIX %s: %pV\n", s->name, &vaf);
1173 va_end(args);
1174 }
1175
1176 static void print_trailer(struct kmem_cache *s, struct slab *slab, u8 *p)
1177 {
1178 unsigned int off; /* Offset of last byte */
1179 u8 *addr = slab_address(slab);
1180
1181 print_tracking(s, p);
1182
1183 print_slab_info(slab);
1184
1185 pr_err("Object 0x%p @offset=%tu fp=0x%p\n\n",
1186 p, p - addr, get_freepointer(s, p));
1187
1188 if (s->flags & SLAB_RED_ZONE)
1189 print_section(KERN_ERR, "Redzone ", p - s->red_left_pad,
1190 s->red_left_pad);
1191 else if (p > addr + 16)
1192 print_section(KERN_ERR, "Bytes b4 ", p - 16, 16);
1193
1194 print_section(KERN_ERR, "Object ", p,
1195 min_t(unsigned int, s->object_size, PAGE_SIZE));
1196 if (s->flags & SLAB_RED_ZONE)
1197 print_section(KERN_ERR, "Redzone ", p + s->object_size,
1198 s->inuse - s->object_size);
1199
1200 off = get_info_end(s);
1201
1202 if (s->flags & SLAB_STORE_USER)
1203 off += 2 * sizeof(struct track);
1204
1205 if (slub_debug_orig_size(s))
1206 off += sizeof(unsigned long);
1207
1208 off += kasan_metadata_size(s, false);
1209
1210 if (off != size_from_object(s))
1211 /* Beginning of the filler is the free pointer */
1212 print_section(KERN_ERR, "Padding ", p + off,
1213 size_from_object(s) - off);
1214 }
1215
1216 static void object_err(struct kmem_cache *s, struct slab *slab,
1217 u8 *object, const char *reason)
1218 {
1219 if (slab_add_kunit_errors())
1220 return;
1221
1222 slab_bug(s, reason);
1223 if (!object || !check_valid_pointer(s, slab, object)) {
1224 print_slab_info(slab);
1225 pr_err("Invalid pointer 0x%p\n", object);
1226 } else {
1227 print_trailer(s, slab, object);
1228 }
1229 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
1230
1231 WARN_ON(1);
1232 }
1233
1234 static bool freelist_corrupted(struct kmem_cache *s, struct slab *slab,
1235 void **freelist, void *nextfree)
1236 {
1237 if ((s->flags & SLAB_CONSISTENCY_CHECKS) &&
1238 !check_valid_pointer(s, slab, nextfree) && freelist) {
1239 object_err(s, slab, *freelist, "Freechain corrupt");
1240 *freelist = NULL;
1241 slab_fix(s, "Isolate corrupted freechain");
1242 return true;
1243 }
1244
1245 return false;
1246 }
1247
1248 static void __slab_err(struct slab *slab)
1249 {
1250 if (slab_in_kunit_test())
1251 return;
1252
1253 print_slab_info(slab);
1254 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
1255
1256 WARN_ON(1);
1257 }
1258
1259 static __printf(3, 4) void slab_err(struct kmem_cache *s, struct slab *slab,
1260 const char *fmt, ...)
1261 {
1262 va_list args;
1263
1264 if (slab_add_kunit_errors())
1265 return;
1266
1267 va_start(args, fmt);
1268 __slab_bug(s, fmt, args);
1269 va_end(args);
1270
1271 __slab_err(slab);
1272 }
1273
1274 static void init_object(struct kmem_cache *s, void *object, u8 val)
1275 {
1276 u8 *p = kasan_reset_tag(object);
1277 unsigned int poison_size = s->object_size;
1278
1279 if (s->flags & SLAB_RED_ZONE) {
1280 /*
1281 * Here and below, avoid overwriting the KMSAN shadow. Keeping
1282 * the shadow makes it possible to distinguish uninit-value
1283 * from use-after-free.
1284 */
1285 memset_no_sanitize_memory(p - s->red_left_pad, val,
1286 s->red_left_pad);
1287
1288 if (slub_debug_orig_size(s) && val == SLUB_RED_ACTIVE) {
1289 /*
1290 * Redzone the extra allocated space by kmalloc than
1291 * requested, and the poison size will be limited to
1292 * the original request size accordingly.
1293 */
1294 poison_size = get_orig_size(s, object);
1295 }
1296 }
1297
1298 if (s->flags & __OBJECT_POISON) {
1299 memset_no_sanitize_memory(p, POISON_FREE, poison_size - 1);
1300 memset_no_sanitize_memory(p + poison_size - 1, POISON_END, 1);
1301 }
1302
1303 if (s->flags & SLAB_RED_ZONE)
1304 memset_no_sanitize_memory(p + poison_size, val,
1305 s->inuse - poison_size);
1306 }
1307
1308 static void restore_bytes(struct kmem_cache *s, const char *message, u8 data,
1309 void *from, void *to)
1310 {
1311 slab_fix(s, "Restoring %s 0x%p-0x%p=0x%x", message, from, to - 1, data);
1312 memset(from, data, to - from);
1313 }
1314
1315 #ifdef CONFIG_KMSAN
1316 #define pad_check_attributes noinline __no_kmsan_checks
1317 #else
1318 #define pad_check_attributes
1319 #endif
1320
1321 static pad_check_attributes int
1322 check_bytes_and_report(struct kmem_cache *s, struct slab *slab,
1323 u8 *object, const char *what, u8 *start, unsigned int value,
1324 unsigned int bytes, bool slab_obj_print)
1325 {
1326 u8 *fault;
1327 u8 *end;
1328 u8 *addr = slab_address(slab);
1329
1330 metadata_access_enable();
1331 fault = memchr_inv(kasan_reset_tag(start), value, bytes);
1332 metadata_access_disable();
1333 if (!fault)
1334 return 1;
1335
1336 end = start + bytes;
1337 while (end > fault && end[-1] == value)
1338 end--;
1339
1340 if (slab_add_kunit_errors())
1341 goto skip_bug_print;
1342
1343 pr_err("[%s overwritten] 0x%p-0x%p @offset=%tu. First byte 0x%x instead of 0x%x\n",
1344 what, fault, end - 1, fault - addr, fault[0], value);
1345
1346 if (slab_obj_print)
1347 object_err(s, slab, object, "Object corrupt");
1348
1349 skip_bug_print:
1350 restore_bytes(s, what, value, fault, end);
1351 return 0;
1352 }
1353
1354 /*
1355 * Object layout:
1356 *
1357 * object address
1358 * Bytes of the object to be managed.
1359 * If the freepointer may overlay the object then the free
1360 * pointer is at the middle of the object.
1361 *
1362 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
1363 * 0xa5 (POISON_END)
1364 *
1365 * object + s->object_size
1366 * Padding to reach word boundary. This is also used for Redzoning.
1367 * Padding is extended by another word if Redzoning is enabled and
1368 * object_size == inuse.
1369 *
1370 * We fill with 0xbb (SLUB_RED_INACTIVE) for inactive objects and with
1371 * 0xcc (SLUB_RED_ACTIVE) for objects in use.
1372 *
1373 * object + s->inuse
1374 * Meta data starts here.
1375 *
1376 * A. Free pointer (if we cannot overwrite object on free)
1377 * B. Tracking data for SLAB_STORE_USER
1378 * C. Original request size for kmalloc object (SLAB_STORE_USER enabled)
1379 * D. Padding to reach required alignment boundary or at minimum
1380 * one word if debugging is on to be able to detect writes
1381 * before the word boundary.
1382 *
1383 * Padding is done using 0x5a (POISON_INUSE)
1384 *
1385 * object + s->size
1386 * Nothing is used beyond s->size.
1387 *
1388 * If slabcaches are merged then the object_size and inuse boundaries are mostly
1389 * ignored. And therefore no slab options that rely on these boundaries
1390 * may be used with merged slabcaches.
1391 */
1392
1393 static int check_pad_bytes(struct kmem_cache *s, struct slab *slab, u8 *p)
1394 {
1395 unsigned long off = get_info_end(s); /* The end of info */
1396
1397 if (s->flags & SLAB_STORE_USER) {
1398 /* We also have user information there */
1399 off += 2 * sizeof(struct track);
1400
1401 if (s->flags & SLAB_KMALLOC)
1402 off += sizeof(unsigned long);
1403 }
1404
1405 off += kasan_metadata_size(s, false);
1406
1407 if (size_from_object(s) == off)
1408 return 1;
1409
1410 return check_bytes_and_report(s, slab, p, "Object padding",
1411 p + off, POISON_INUSE, size_from_object(s) - off, true);
1412 }
1413
1414 /* Check the pad bytes at the end of a slab page */
1415 static pad_check_attributes void
1416 slab_pad_check(struct kmem_cache *s, struct slab *slab)
1417 {
1418 u8 *start;
1419 u8 *fault;
1420 u8 *end;
1421 u8 *pad;
1422 int length;
1423 int remainder;
1424
1425 if (!(s->flags & SLAB_POISON))
1426 return;
1427
1428 start = slab_address(slab);
1429 length = slab_size(slab);
1430 end = start + length;
1431 remainder = length % s->size;
1432 if (!remainder)
1433 return;
1434
1435 pad = end - remainder;
1436 metadata_access_enable();
1437 fault = memchr_inv(kasan_reset_tag(pad), POISON_INUSE, remainder);
1438 metadata_access_disable();
1439 if (!fault)
1440 return;
1441 while (end > fault && end[-1] == POISON_INUSE)
1442 end--;
1443
1444 slab_bug(s, "Padding overwritten. 0x%p-0x%p @offset=%tu",
1445 fault, end - 1, fault - start);
1446 print_section(KERN_ERR, "Padding ", pad, remainder);
1447 __slab_err(slab);
1448
1449 restore_bytes(s, "slab padding", POISON_INUSE, fault, end);
1450 }
1451
1452 static int check_object(struct kmem_cache *s, struct slab *slab,
1453 void *object, u8 val)
1454 {
1455 u8 *p = object;
1456 u8 *endobject = object + s->object_size;
1457 unsigned int orig_size, kasan_meta_size;
1458 int ret = 1;
1459
1460 if (s->flags & SLAB_RED_ZONE) {
1461 if (!check_bytes_and_report(s, slab, object, "Left Redzone",
1462 object - s->red_left_pad, val, s->red_left_pad, ret))
1463 ret = 0;
1464
1465 if (!check_bytes_and_report(s, slab, object, "Right Redzone",
1466 endobject, val, s->inuse - s->object_size, ret))
1467 ret = 0;
1468
1469 if (slub_debug_orig_size(s) && val == SLUB_RED_ACTIVE) {
1470 orig_size = get_orig_size(s, object);
1471
1472 if (s->object_size > orig_size &&
1473 !check_bytes_and_report(s, slab, object,
1474 "kmalloc Redzone", p + orig_size,
1475 val, s->object_size - orig_size, ret)) {
1476 ret = 0;
1477 }
1478 }
1479 } else {
1480 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
1481 if (!check_bytes_and_report(s, slab, p, "Alignment padding",
1482 endobject, POISON_INUSE,
1483 s->inuse - s->object_size, ret))
1484 ret = 0;
1485 }
1486 }
1487
1488 if (s->flags & SLAB_POISON) {
1489 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON)) {
1490 /*
1491 * KASAN can save its free meta data inside of the
1492 * object at offset 0. Thus, skip checking the part of
1493 * the redzone that overlaps with the meta data.
1494 */
1495 kasan_meta_size = kasan_metadata_size(s, true);
1496 if (kasan_meta_size < s->object_size - 1 &&
1497 !check_bytes_and_report(s, slab, p, "Poison",
1498 p + kasan_meta_size, POISON_FREE,
1499 s->object_size - kasan_meta_size - 1, ret))
1500 ret = 0;
1501 if (kasan_meta_size < s->object_size &&
1502 !check_bytes_and_report(s, slab, p, "End Poison",
1503 p + s->object_size - 1, POISON_END, 1, ret))
1504 ret = 0;
1505 }
1506 /*
1507 * check_pad_bytes cleans up on its own.
1508 */
1509 if (!check_pad_bytes(s, slab, p))
1510 ret = 0;
1511 }
1512
1513 /*
1514 * Cannot check freepointer while object is allocated if
1515 * object and freepointer overlap.
1516 */
1517 if ((freeptr_outside_object(s) || val != SLUB_RED_ACTIVE) &&
1518 !check_valid_pointer(s, slab, get_freepointer(s, p))) {
1519 object_err(s, slab, p, "Freepointer corrupt");
1520 /*
1521 * No choice but to zap it and thus lose the remainder
1522 * of the free objects in this slab. May cause
1523 * another error because the object count is now wrong.
1524 */
1525 set_freepointer(s, p, NULL);
1526 ret = 0;
1527 }
1528
1529 return ret;
1530 }
1531
1532 /*
1533 * Checks if the slab state looks sane. Assumes the struct slab pointer
1534 * was either obtained in a way that ensures it's valid, or validated
1535 * by validate_slab_ptr()
1536 */
1537 static int check_slab(struct kmem_cache *s, struct slab *slab)
1538 {
1539 int maxobj;
1540
1541 maxobj = order_objects(slab_order(slab), s->size);
1542 if (slab->objects > maxobj) {
1543 slab_err(s, slab, "objects %u > max %u",
1544 slab->objects, maxobj);
1545 return 0;
1546 }
1547 if (slab->inuse > slab->objects) {
1548 slab_err(s, slab, "inuse %u > max %u",
1549 slab->inuse, slab->objects);
1550 return 0;
1551 }
1552 if (slab->frozen) {
1553 slab_err(s, slab, "Slab disabled since SLUB metadata consistency check failed");
1554 return 0;
1555 }
1556
1557 /* Slab_pad_check fixes things up after itself */
1558 slab_pad_check(s, slab);
1559 return 1;
1560 }
1561
1562 /*
1563 * Determine if a certain object in a slab is on the freelist. Must hold the
1564 * slab lock to guarantee that the chains are in a consistent state.
1565 */
1566 static bool on_freelist(struct kmem_cache *s, struct slab *slab, void *search)
1567 {
1568 int nr = 0;
1569 void *fp;
1570 void *object = NULL;
1571 int max_objects;
1572
1573 fp = slab->freelist;
1574 while (fp && nr <= slab->objects) {
1575 if (fp == search)
1576 return true;
1577 if (!check_valid_pointer(s, slab, fp)) {
1578 if (object) {
1579 object_err(s, slab, object,
1580 "Freechain corrupt");
1581 set_freepointer(s, object, NULL);
1582 break;
1583 } else {
1584 slab_err(s, slab, "Freepointer corrupt");
1585 slab->freelist = NULL;
1586 slab->inuse = slab->objects;
1587 slab_fix(s, "Freelist cleared");
1588 return false;
1589 }
1590 }
1591 object = fp;
1592 fp = get_freepointer(s, object);
1593 nr++;
1594 }
1595
1596 if (nr > slab->objects) {
1597 slab_err(s, slab, "Freelist cycle detected");
1598 slab->freelist = NULL;
1599 slab->inuse = slab->objects;
1600 slab_fix(s, "Freelist cleared");
1601 return false;
1602 }
1603
1604 max_objects = order_objects(slab_order(slab), s->size);
1605 if (max_objects > MAX_OBJS_PER_PAGE)
1606 max_objects = MAX_OBJS_PER_PAGE;
1607
1608 if (slab->objects != max_objects) {
1609 slab_err(s, slab, "Wrong number of objects. Found %d but should be %d",
1610 slab->objects, max_objects);
1611 slab->objects = max_objects;
1612 slab_fix(s, "Number of objects adjusted");
1613 }
1614 if (slab->inuse != slab->objects - nr) {
1615 slab_err(s, slab, "Wrong object count. Counter is %d but counted were %d",
1616 slab->inuse, slab->objects - nr);
1617 slab->inuse = slab->objects - nr;
1618 slab_fix(s, "Object count adjusted");
1619 }
1620 return search == NULL;
1621 }
1622
1623 static void trace(struct kmem_cache *s, struct slab *slab, void *object,
1624 int alloc)
1625 {
1626 if (s->flags & SLAB_TRACE) {
1627 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
1628 s->name,
1629 alloc ? "alloc" : "free",
1630 object, slab->inuse,
1631 slab->freelist);
1632
1633 if (!alloc)
1634 print_section(KERN_INFO, "Object ", (void *)object,
1635 s->object_size);
1636
1637 dump_stack();
1638 }
1639 }
1640
1641 /*
1642 * Tracking of fully allocated slabs for debugging purposes.
1643 */
1644 static void add_full(struct kmem_cache *s,
1645 struct kmem_cache_node *n, struct slab *slab)
1646 {
1647 if (!(s->flags & SLAB_STORE_USER))
1648 return;
1649
1650 lockdep_assert_held(&n->list_lock);
1651 list_add(&slab->slab_list, &n->full);
1652 }
1653
1654 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct slab *slab)
1655 {
1656 if (!(s->flags & SLAB_STORE_USER))
1657 return;
1658
1659 lockdep_assert_held(&n->list_lock);
1660 list_del(&slab->slab_list);
1661 }
1662
1663 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1664 {
1665 return atomic_long_read(&n->nr_slabs);
1666 }
1667
1668 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1669 {
1670 struct kmem_cache_node *n = get_node(s, node);
1671
1672 atomic_long_inc(&n->nr_slabs);
1673 atomic_long_add(objects, &n->total_objects);
1674 }
1675 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1676 {
1677 struct kmem_cache_node *n = get_node(s, node);
1678
1679 atomic_long_dec(&n->nr_slabs);
1680 atomic_long_sub(objects, &n->total_objects);
1681 }
1682
1683 /* Object debug checks for alloc/free paths */
1684 static void setup_object_debug(struct kmem_cache *s, void *object)
1685 {
1686 if (!kmem_cache_debug_flags(s, SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))
1687 return;
1688
1689 init_object(s, object, SLUB_RED_INACTIVE);
1690 init_tracking(s, object);
1691 }
1692
1693 static
1694 void setup_slab_debug(struct kmem_cache *s, struct slab *slab, void *addr)
1695 {
1696 if (!kmem_cache_debug_flags(s, SLAB_POISON))
1697 return;
1698
1699 metadata_access_enable();
1700 memset(kasan_reset_tag(addr), POISON_INUSE, slab_size(slab));
1701 metadata_access_disable();
1702 }
1703
1704 static inline int alloc_consistency_checks(struct kmem_cache *s,
1705 struct slab *slab, void *object)
1706 {
1707 if (!check_slab(s, slab))
1708 return 0;
1709
1710 if (!check_valid_pointer(s, slab, object)) {
1711 object_err(s, slab, object, "Freelist Pointer check fails");
1712 return 0;
1713 }
1714
1715 if (!check_object(s, slab, object, SLUB_RED_INACTIVE))
1716 return 0;
1717
1718 return 1;
1719 }
1720
1721 static noinline bool alloc_debug_processing(struct kmem_cache *s,
1722 struct slab *slab, void *object, int orig_size)
1723 {
1724 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1725 if (!alloc_consistency_checks(s, slab, object))
1726 goto bad;
1727 }
1728
1729 /* Success. Perform special debug activities for allocs */
1730 trace(s, slab, object, 1);
1731 set_orig_size(s, object, orig_size);
1732 init_object(s, object, SLUB_RED_ACTIVE);
1733 return true;
1734
1735 bad:
1736 /*
1737 * Let's do the best we can to avoid issues in the future. Marking all
1738 * objects as used avoids touching the remaining objects.
1739 */
1740 slab_fix(s, "Marking all objects used");
1741 slab->inuse = slab->objects;
1742 slab->freelist = NULL;
1743 slab->frozen = 1; /* mark consistency-failed slab as frozen */
1744
1745 return false;
1746 }
1747
1748 static inline int free_consistency_checks(struct kmem_cache *s,
1749 struct slab *slab, void *object, unsigned long addr)
1750 {
1751 if (!check_valid_pointer(s, slab, object)) {
1752 slab_err(s, slab, "Invalid object pointer 0x%p", object);
1753 return 0;
1754 }
1755
1756 if (on_freelist(s, slab, object)) {
1757 object_err(s, slab, object, "Object already free");
1758 return 0;
1759 }
1760
1761 if (!check_object(s, slab, object, SLUB_RED_ACTIVE))
1762 return 0;
1763
1764 if (unlikely(s != slab->slab_cache)) {
1765 if (!slab->slab_cache) {
1766 slab_err(NULL, slab, "No slab cache for object 0x%p",
1767 object);
1768 } else {
1769 object_err(s, slab, object,
1770 "page slab pointer corrupt.");
1771 }
1772 return 0;
1773 }
1774 return 1;
1775 }
1776
1777 /*
1778 * Parse a block of slab_debug options. Blocks are delimited by ';'
1779 *
1780 * @str: start of block
1781 * @flags: returns parsed flags, or DEBUG_DEFAULT_FLAGS if none specified
1782 * @slabs: return start of list of slabs, or NULL when there's no list
1783 * @init: assume this is initial parsing and not per-kmem-create parsing
1784 *
1785 * returns the start of next block if there's any, or NULL
1786 */
1787 static char *
1788 parse_slub_debug_flags(char *str, slab_flags_t *flags, char **slabs, bool init)
1789 {
1790 bool higher_order_disable = false;
1791
1792 /* Skip any completely empty blocks */
1793 while (*str && *str == ';')
1794 str++;
1795
1796 if (*str == ',') {
1797 /*
1798 * No options but restriction on slabs. This means full
1799 * debugging for slabs matching a pattern.
1800 */
1801 *flags = DEBUG_DEFAULT_FLAGS;
1802 goto check_slabs;
1803 }
1804 *flags = 0;
1805
1806 /* Determine which debug features should be switched on */
1807 for (; *str && *str != ',' && *str != ';'; str++) {
1808 switch (tolower(*str)) {
1809 case '-':
1810 *flags = 0;
1811 break;
1812 case 'f':
1813 *flags |= SLAB_CONSISTENCY_CHECKS;
1814 break;
1815 case 'z':
1816 *flags |= SLAB_RED_ZONE;
1817 break;
1818 case 'p':
1819 *flags |= SLAB_POISON;
1820 break;
1821 case 'u':
1822 *flags |= SLAB_STORE_USER;
1823 break;
1824 case 't':
1825 *flags |= SLAB_TRACE;
1826 break;
1827 case 'a':
1828 *flags |= SLAB_FAILSLAB;
1829 break;
1830 case 'o':
1831 /*
1832 * Avoid enabling debugging on caches if its minimum
1833 * order would increase as a result.
1834 */
1835 higher_order_disable = true;
1836 break;
1837 default:
1838 if (init)
1839 pr_err("slab_debug option '%c' unknown. skipped\n", *str);
1840 }
1841 }
1842 check_slabs:
1843 if (*str == ',')
1844 *slabs = ++str;
1845 else
1846 *slabs = NULL;
1847
1848 /* Skip over the slab list */
1849 while (*str && *str != ';')
1850 str++;
1851
1852 /* Skip any completely empty blocks */
1853 while (*str && *str == ';')
1854 str++;
1855
1856 if (init && higher_order_disable)
1857 disable_higher_order_debug = 1;
1858
1859 if (*str)
1860 return str;
1861 else
1862 return NULL;
1863 }
1864
1865 static int __init setup_slub_debug(char *str)
1866 {
1867 slab_flags_t flags;
1868 slab_flags_t global_flags;
1869 char *saved_str;
1870 char *slab_list;
1871 bool global_slub_debug_changed = false;
1872 bool slab_list_specified = false;
1873
1874 global_flags = DEBUG_DEFAULT_FLAGS;
1875 if (*str++ != '=' || !*str)
1876 /*
1877 * No options specified. Switch on full debugging.
1878 */
1879 goto out;
1880
1881 saved_str = str;
1882 while (str) {
1883 str = parse_slub_debug_flags(str, &flags, &slab_list, true);
1884
1885 if (!slab_list) {
1886 global_flags = flags;
1887 global_slub_debug_changed = true;
1888 } else {
1889 slab_list_specified = true;
1890 if (flags & SLAB_STORE_USER)
1891 stack_depot_request_early_init();
1892 }
1893 }
1894
1895 /*
1896 * For backwards compatibility, a single list of flags with list of
1897 * slabs means debugging is only changed for those slabs, so the global
1898 * slab_debug should be unchanged (0 or DEBUG_DEFAULT_FLAGS, depending
1899 * on CONFIG_SLUB_DEBUG_ON). We can extended that to multiple lists as
1900 * long as there is no option specifying flags without a slab list.
1901 */
1902 if (slab_list_specified) {
1903 if (!global_slub_debug_changed)
1904 global_flags = slub_debug;
1905 slub_debug_string = saved_str;
1906 }
1907 out:
1908 slub_debug = global_flags;
1909 if (slub_debug & SLAB_STORE_USER)
1910 stack_depot_request_early_init();
1911 if (slub_debug != 0 || slub_debug_string)
1912 static_branch_enable(&slub_debug_enabled);
1913 else
1914 static_branch_disable(&slub_debug_enabled);
1915 if ((static_branch_unlikely(&init_on_alloc) ||
1916 static_branch_unlikely(&init_on_free)) &&
1917 (slub_debug & SLAB_POISON))
1918 pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n");
1919 return 1;
1920 }
1921
1922 __setup("slab_debug", setup_slub_debug);
1923 __setup_param("slub_debug", slub_debug, setup_slub_debug, 0);
1924
1925 /*
1926 * kmem_cache_flags - apply debugging options to the cache
1927 * @flags: flags to set
1928 * @name: name of the cache
1929 *
1930 * Debug option(s) are applied to @flags. In addition to the debug
1931 * option(s), if a slab name (or multiple) is specified i.e.
1932 * slab_debug=<Debug-Options>,<slab name1>,<slab name2> ...
1933 * then only the select slabs will receive the debug option(s).
1934 */
1935 slab_flags_t kmem_cache_flags(slab_flags_t flags, const char *name)
1936 {
1937 char *iter;
1938 size_t len;
1939 char *next_block;
1940 slab_flags_t block_flags;
1941 slab_flags_t slub_debug_local = slub_debug;
1942
1943 if (flags & SLAB_NO_USER_FLAGS)
1944 return flags;
1945
1946 /*
1947 * If the slab cache is for debugging (e.g. kmemleak) then
1948 * don't store user (stack trace) information by default,
1949 * but let the user enable it via the command line below.
1950 */
1951 if (flags & SLAB_NOLEAKTRACE)
1952 slub_debug_local &= ~SLAB_STORE_USER;
1953
1954 len = strlen(name);
1955 next_block = slub_debug_string;
1956 /* Go through all blocks of debug options, see if any matches our slab's name */
1957 while (next_block) {
1958 next_block = parse_slub_debug_flags(next_block, &block_flags, &iter, false);
1959 if (!iter)
1960 continue;
1961 /* Found a block that has a slab list, search it */
1962 while (*iter) {
1963 char *end, *glob;
1964 size_t cmplen;
1965
1966 end = strchrnul(iter, ',');
1967 if (next_block && next_block < end)
1968 end = next_block - 1;
1969
1970 glob = strnchr(iter, end - iter, '*');
1971 if (glob)
1972 cmplen = glob - iter;
1973 else
1974 cmplen = max_t(size_t, len, (end - iter));
1975
1976 if (!strncmp(name, iter, cmplen)) {
1977 flags |= block_flags;
1978 return flags;
1979 }
1980
1981 if (!*end || *end == ';')
1982 break;
1983 iter = end + 1;
1984 }
1985 }
1986
1987 return flags | slub_debug_local;
1988 }
1989 #else /* !CONFIG_SLUB_DEBUG */
1990 static inline void setup_object_debug(struct kmem_cache *s, void *object) {}
1991 static inline
1992 void setup_slab_debug(struct kmem_cache *s, struct slab *slab, void *addr) {}
1993
1994 static inline bool alloc_debug_processing(struct kmem_cache *s,
1995 struct slab *slab, void *object, int orig_size) { return true; }
1996
1997 static inline bool free_debug_processing(struct kmem_cache *s,
1998 struct slab *slab, void *head, void *tail, int *bulk_cnt,
1999 unsigned long addr, depot_stack_handle_t handle) { return true; }
2000
2001 static inline void slab_pad_check(struct kmem_cache *s, struct slab *slab) {}
2002 static inline int check_object(struct kmem_cache *s, struct slab *slab,
2003 void *object, u8 val) { return 1; }
2004 static inline depot_stack_handle_t set_track_prepare(gfp_t gfp_flags) { return 0; }
2005 static inline void set_track(struct kmem_cache *s, void *object,
2006 enum track_item alloc, unsigned long addr, gfp_t gfp_flags) {}
2007 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
2008 struct slab *slab) {}
2009 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
2010 struct slab *slab) {}
2011 slab_flags_t kmem_cache_flags(slab_flags_t flags, const char *name)
2012 {
2013 return flags;
2014 }
2015 #define slub_debug 0
2016
2017 #define disable_higher_order_debug 0
2018
2019 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
2020 { return 0; }
2021 static inline void inc_slabs_node(struct kmem_cache *s, int node,
2022 int objects) {}
2023 static inline void dec_slabs_node(struct kmem_cache *s, int node,
2024 int objects) {}
2025 static bool freelist_corrupted(struct kmem_cache *s, struct slab *slab,
2026 void **freelist, void *nextfree)
2027 {
2028 return false;
2029 }
2030 #endif /* CONFIG_SLUB_DEBUG */
2031
2032 #ifdef CONFIG_SLAB_OBJ_EXT
2033
2034 #ifdef CONFIG_MEM_ALLOC_PROFILING_DEBUG
2035
2036 static inline void mark_objexts_empty(struct slabobj_ext *obj_exts)
2037 {
2038 struct slabobj_ext *slab_exts;
2039 struct slab *obj_exts_slab;
2040
2041 obj_exts_slab = virt_to_slab(obj_exts);
2042 slab_exts = slab_obj_exts(obj_exts_slab);
2043 if (slab_exts) {
2044 unsigned int offs = obj_to_index(obj_exts_slab->slab_cache,
2045 obj_exts_slab, obj_exts);
2046
2047 if (unlikely(is_codetag_empty(&slab_exts[offs].ref)))
2048 return;
2049
2050 /* codetag should be NULL here */
2051 WARN_ON(slab_exts[offs].ref.ct);
2052 set_codetag_empty(&slab_exts[offs].ref);
2053 }
2054 }
2055
2056 static inline bool mark_failed_objexts_alloc(struct slab *slab)
2057 {
2058 return cmpxchg(&slab->obj_exts, 0, OBJEXTS_ALLOC_FAIL) == 0;
2059 }
2060
2061 static inline void handle_failed_objexts_alloc(unsigned long obj_exts,
2062 struct slabobj_ext *vec, unsigned int objects)
2063 {
2064 /*
2065 * If vector previously failed to allocate then we have live
2066 * objects with no tag reference. Mark all references in this
2067 * vector as empty to avoid warnings later on.
2068 */
2069 if (obj_exts == OBJEXTS_ALLOC_FAIL) {
2070 unsigned int i;
2071
2072 for (i = 0; i < objects; i++)
2073 set_codetag_empty(&vec[i].ref);
2074 }
2075 }
2076
2077 #else /* CONFIG_MEM_ALLOC_PROFILING_DEBUG */
2078
2079 static inline void mark_objexts_empty(struct slabobj_ext *obj_exts) {}
2080 static inline bool mark_failed_objexts_alloc(struct slab *slab) { return false; }
2081 static inline void handle_failed_objexts_alloc(unsigned long obj_exts,
2082 struct slabobj_ext *vec, unsigned int objects) {}
2083
2084 #endif /* CONFIG_MEM_ALLOC_PROFILING_DEBUG */
2085
2086 /*
2087 * The allocated objcg pointers array is not accounted directly.
2088 * Moreover, it should not come from DMA buffer and is not readily
2089 * reclaimable. So those GFP bits should be masked off.
2090 */
2091 #define OBJCGS_CLEAR_MASK (__GFP_DMA | __GFP_RECLAIMABLE | \
2092 __GFP_ACCOUNT | __GFP_NOFAIL)
2093
2094 static inline void init_slab_obj_exts(struct slab *slab)
2095 {
2096 slab->obj_exts = 0;
2097 }
2098
2099 /*
2100 * Calculate the allocation size for slabobj_ext array.
2101 *
2102 * When memory allocation profiling is enabled, the obj_exts array
2103 * could be allocated from the same slab cache it's being allocated for.
2104 * This would prevent the slab from ever being freed because it would
2105 * always contain at least one allocated object (its own obj_exts array).
2106 *
2107 * To avoid this, increase the allocation size when we detect the array
2108 * may come from the same cache, forcing it to use a different cache.
2109 */
2110 static inline size_t obj_exts_alloc_size(struct kmem_cache *s,
2111 struct slab *slab, gfp_t gfp)
2112 {
2113 size_t sz = sizeof(struct slabobj_ext) * slab->objects;
2114 struct kmem_cache *obj_exts_cache;
2115
2116 if (sz > KMALLOC_MAX_CACHE_SIZE)
2117 return sz;
2118
2119 if (!is_kmalloc_normal(s))
2120 return sz;
2121
2122 obj_exts_cache = kmalloc_slab(sz, NULL, gfp, 0);
2123 /*
2124 * We can't simply compare s with obj_exts_cache, because random kmalloc
2125 * caches have multiple caches per size, selected by caller address.
2126 * Since caller address may differ between kmalloc_slab() and actual
2127 * allocation, bump size when sizes are equal.
2128 */
2129 if (s->object_size == obj_exts_cache->object_size)
2130 return obj_exts_cache->object_size + 1;
2131
2132 return sz;
2133 }
2134
2135 int alloc_slab_obj_exts(struct slab *slab, struct kmem_cache *s,
2136 gfp_t gfp, bool new_slab)
2137 {
2138 bool allow_spin = gfpflags_allow_spinning(gfp);
2139 unsigned int objects = objs_per_slab(s, slab);
2140 unsigned long new_exts;
2141 unsigned long old_exts;
2142 struct slabobj_ext *vec;
2143 size_t sz;
2144
2145 gfp &= ~OBJCGS_CLEAR_MASK;
2146 /* Prevent recursive extension vector allocation */
2147 gfp |= __GFP_NO_OBJ_EXT;
2148
2149 sz = obj_exts_alloc_size(s, slab, gfp);
2150
2151 /*
2152 * Note that allow_spin may be false during early boot and its
2153 * restricted GFP_BOOT_MASK. Due to kmalloc_nolock() only supporting
2154 * architectures with cmpxchg16b, early obj_exts will be missing for
2155 * very early allocations on those.
2156 */
2157 if (unlikely(!allow_spin))
2158 vec = kmalloc_nolock(sz, __GFP_ZERO | __GFP_NO_OBJ_EXT,
2159 slab_nid(slab));
2160 else
2161 vec = kmalloc_node(sz, gfp | __GFP_ZERO, slab_nid(slab));
2162
2163 if (!vec) {
2164 /*
2165 * Try to mark vectors which failed to allocate.
2166 * If this operation fails, there may be a racing process
2167 * that has already completed the allocation.
2168 */
2169 if (!mark_failed_objexts_alloc(slab) &&
2170 slab_obj_exts(slab))
2171 return 0;
2172
2173 return -ENOMEM;
2174 }
2175
2176 VM_WARN_ON_ONCE(virt_to_slab(vec) != NULL &&
2177 virt_to_slab(vec)->slab_cache == s);
2178
2179 new_exts = (unsigned long)vec;
2180 if (unlikely(!allow_spin))
2181 new_exts |= OBJEXTS_NOSPIN_ALLOC;
2182 #ifdef CONFIG_MEMCG
2183 new_exts |= MEMCG_DATA_OBJEXTS;
2184 #endif
2185 retry:
2186 old_exts = READ_ONCE(slab->obj_exts);
2187 handle_failed_objexts_alloc(old_exts, vec, objects);
2188 if (new_slab) {
2189 /*
2190 * If the slab is brand new and nobody can yet access its
2191 * obj_exts, no synchronization is required and obj_exts can
2192 * be simply assigned.
2193 */
2194 slab->obj_exts = new_exts;
2195 } else if (old_exts & ~OBJEXTS_FLAGS_MASK) {
2196 /*
2197 * If the slab is already in use, somebody can allocate and
2198 * assign slabobj_exts in parallel. In this case the existing
2199 * objcg vector should be reused.
2200 */
2201 mark_objexts_empty(vec);
2202 if (unlikely(!allow_spin))
2203 kfree_nolock(vec);
2204 else
2205 kfree(vec);
2206 return 0;
2207 } else if (cmpxchg(&slab->obj_exts, old_exts, new_exts) != old_exts) {
2208 /* Retry if a racing thread changed slab->obj_exts from under us. */
2209 goto retry;
2210 }
2211
2212 if (allow_spin)
2213 kmemleak_not_leak(vec);
2214 return 0;
2215 }
2216
2217 static inline void free_slab_obj_exts(struct slab *slab)
2218 {
2219 struct slabobj_ext *obj_exts;
2220
2221 obj_exts = slab_obj_exts(slab);
2222 if (!obj_exts) {
2223 /*
2224 * If obj_exts allocation failed, slab->obj_exts is set to
2225 * OBJEXTS_ALLOC_FAIL. In this case, we end up here and should
2226 * clear the flag.
2227 */
2228 slab->obj_exts = 0;
2229 return;
2230 }
2231
2232 /*
2233 * obj_exts was created with __GFP_NO_OBJ_EXT flag, therefore its
2234 * corresponding extension will be NULL. alloc_tag_sub() will throw a
2235 * warning if slab has extensions but the extension of an object is
2236 * NULL, therefore replace NULL with CODETAG_EMPTY to indicate that
2237 * the extension for obj_exts is expected to be NULL.
2238 */
2239 mark_objexts_empty(obj_exts);
2240 if (unlikely(READ_ONCE(slab->obj_exts) & OBJEXTS_NOSPIN_ALLOC))
2241 kfree_nolock(obj_exts);
2242 else
2243 kfree(obj_exts);
2244 slab->obj_exts = 0;
2245 }
2246
2247 #else /* CONFIG_SLAB_OBJ_EXT */
2248
2249 static inline void init_slab_obj_exts(struct slab *slab)
2250 {
2251 }
2252
2253 static int alloc_slab_obj_exts(struct slab *slab, struct kmem_cache *s,
2254 gfp_t gfp, bool new_slab)
2255 {
2256 return 0;
2257 }
2258
2259 static inline void free_slab_obj_exts(struct slab *slab)
2260 {
2261 }
2262
2263 #endif /* CONFIG_SLAB_OBJ_EXT */
2264
2265 #ifdef CONFIG_MEM_ALLOC_PROFILING
2266
2267 static inline struct slabobj_ext *
2268 prepare_slab_obj_exts_hook(struct kmem_cache *s, gfp_t flags, void *p)
2269 {
2270 struct slab *slab;
2271
2272 slab = virt_to_slab(p);
2273 if (!slab_obj_exts(slab) &&
2274 alloc_slab_obj_exts(slab, s, flags, false)) {
2275 pr_warn_once("%s, %s: Failed to create slab extension vector!\n",
2276 __func__, s->name);
2277 return NULL;
2278 }
2279
2280 return slab_obj_exts(slab) + obj_to_index(s, slab, p);
2281 }
2282
2283 /* Should be called only if mem_alloc_profiling_enabled() */
2284 static noinline void
2285 __alloc_tagging_slab_alloc_hook(struct kmem_cache *s, void *object, gfp_t flags)
2286 {
2287 struct slabobj_ext *obj_exts;
2288
2289 if (!object)
2290 return;
2291
2292 if (s->flags & (SLAB_NO_OBJ_EXT | SLAB_NOLEAKTRACE))
2293 return;
2294
2295 if (flags & __GFP_NO_OBJ_EXT)
2296 return;
2297
2298 obj_exts = prepare_slab_obj_exts_hook(s, flags, object);
2299 /*
2300 * Currently obj_exts is used only for allocation profiling.
2301 * If other users appear then mem_alloc_profiling_enabled()
2302 * check should be added before alloc_tag_add().
2303 */
2304 if (likely(obj_exts))
2305 alloc_tag_add(&obj_exts->ref, current->alloc_tag, s->size);
2306 else
2307 alloc_tag_set_inaccurate(current->alloc_tag);
2308 }
2309
2310 static inline void
2311 alloc_tagging_slab_alloc_hook(struct kmem_cache *s, void *object, gfp_t flags)
2312 {
2313 if (mem_alloc_profiling_enabled())
2314 __alloc_tagging_slab_alloc_hook(s, object, flags);
2315 }
2316
2317 /* Should be called only if mem_alloc_profiling_enabled() */
2318 static noinline void
2319 __alloc_tagging_slab_free_hook(struct kmem_cache *s, struct slab *slab, void **p,
2320 int objects)
2321 {
2322 struct slabobj_ext *obj_exts;
2323 int i;
2324
2325 /* slab->obj_exts might not be NULL if it was created for MEMCG accounting. */
2326 if (s->flags & (SLAB_NO_OBJ_EXT | SLAB_NOLEAKTRACE))
2327 return;
2328
2329 obj_exts = slab_obj_exts(slab);
2330 if (!obj_exts)
2331 return;
2332
2333 for (i = 0; i < objects; i++) {
2334 unsigned int off = obj_to_index(s, slab, p[i]);
2335
2336 alloc_tag_sub(&obj_exts[off].ref, s->size);
2337 }
2338 }
2339
2340 static inline void
2341 alloc_tagging_slab_free_hook(struct kmem_cache *s, struct slab *slab, void **p,
2342 int objects)
2343 {
2344 if (mem_alloc_profiling_enabled())
2345 __alloc_tagging_slab_free_hook(s, slab, p, objects);
2346 }
2347
2348 #else /* CONFIG_MEM_ALLOC_PROFILING */
2349
2350 static inline void
2351 alloc_tagging_slab_alloc_hook(struct kmem_cache *s, void *object, gfp_t flags)
2352 {
2353 }
2354
2355 static inline void
2356 alloc_tagging_slab_free_hook(struct kmem_cache *s, struct slab *slab, void **p,
2357 int objects)
2358 {
2359 }
2360
2361 #endif /* CONFIG_MEM_ALLOC_PROFILING */
2362
2363
2364 #ifdef CONFIG_MEMCG
2365
2366 static void memcg_alloc_abort_single(struct kmem_cache *s, void *object);
2367
2368 static __fastpath_inline
2369 bool memcg_slab_post_alloc_hook(struct kmem_cache *s, struct list_lru *lru,
2370 gfp_t flags, size_t size, void **p)
2371 {
2372 if (likely(!memcg_kmem_online()))
2373 return true;
2374
2375 if (likely(!(flags & __GFP_ACCOUNT) && !(s->flags & SLAB_ACCOUNT)))
2376 return true;
2377
2378 if (likely(__memcg_slab_post_alloc_hook(s, lru, flags, size, p)))
2379 return true;
2380
2381 if (likely(size == 1)) {
2382 memcg_alloc_abort_single(s, *p);
2383 *p = NULL;
2384 } else {
2385 kmem_cache_free_bulk(s, size, p);
2386 }
2387
2388 return false;
2389 }
2390
2391 static __fastpath_inline
2392 void memcg_slab_free_hook(struct kmem_cache *s, struct slab *slab, void **p,
2393 int objects)
2394 {
2395 struct slabobj_ext *obj_exts;
2396
2397 if (!memcg_kmem_online())
2398 return;
2399
2400 obj_exts = slab_obj_exts(slab);
2401 if (likely(!obj_exts))
2402 return;
2403
2404 __memcg_slab_free_hook(s, slab, p, objects, obj_exts);
2405 }
2406
2407 static __fastpath_inline
2408 bool memcg_slab_post_charge(void *p, gfp_t flags)
2409 {
2410 struct slabobj_ext *slab_exts;
2411 struct kmem_cache *s;
2412 struct folio *folio;
2413 struct slab *slab;
2414 unsigned long off;
2415
2416 folio = virt_to_folio(p);
2417 if (!folio_test_slab(folio)) {
2418 int size;
2419
2420 if (folio_memcg_kmem(folio))
2421 return true;
2422
2423 if (__memcg_kmem_charge_page(folio_page(folio, 0), flags,
2424 folio_order(folio)))
2425 return false;
2426
2427 /*
2428 * This folio has already been accounted in the global stats but
2429 * not in the memcg stats. So, subtract from the global and use
2430 * the interface which adds to both global and memcg stats.
2431 */
2432 size = folio_size(folio);
2433 node_stat_mod_folio(folio, NR_SLAB_UNRECLAIMABLE_B, -size);
2434 lruvec_stat_mod_folio(folio, NR_SLAB_UNRECLAIMABLE_B, size);
2435 return true;
2436 }
2437
2438 slab = folio_slab(folio);
2439 s = slab->slab_cache;
2440
2441 /*
2442 * Ignore KMALLOC_NORMAL cache to avoid possible circular dependency
2443 * of slab_obj_exts being allocated from the same slab and thus the slab
2444 * becoming effectively unfreeable.
2445 */
2446 if (is_kmalloc_normal(s))
2447 return true;
2448
2449 /* Ignore already charged objects. */
2450 slab_exts = slab_obj_exts(slab);
2451 if (slab_exts) {
2452 off = obj_to_index(s, slab, p);
2453 if (unlikely(slab_exts[off].objcg))
2454 return true;
2455 }
2456
2457 return __memcg_slab_post_alloc_hook(s, NULL, flags, 1, &p);
2458 }
2459
2460 #else /* CONFIG_MEMCG */
2461 static inline bool memcg_slab_post_alloc_hook(struct kmem_cache *s,
2462 struct list_lru *lru,
2463 gfp_t flags, size_t size,
2464 void **p)
2465 {
2466 return true;
2467 }
2468
2469 static inline void memcg_slab_free_hook(struct kmem_cache *s, struct slab *slab,
2470 void **p, int objects)
2471 {
2472 }
2473
2474 static inline bool memcg_slab_post_charge(void *p, gfp_t flags)
2475 {
2476 return true;
2477 }
2478 #endif /* CONFIG_MEMCG */
2479
2480 #ifdef CONFIG_SLUB_RCU_DEBUG
2481 static void slab_free_after_rcu_debug(struct rcu_head *rcu_head);
2482
2483 struct rcu_delayed_free {
2484 struct rcu_head head;
2485 void *object;
2486 };
2487 #endif
2488
2489 /*
2490 * Hooks for other subsystems that check memory allocations. In a typical
2491 * production configuration these hooks all should produce no code at all.
2492 *
2493 * Returns true if freeing of the object can proceed, false if its reuse
2494 * was delayed by CONFIG_SLUB_RCU_DEBUG or KASAN quarantine, or it was returned
2495 * to KFENCE.
2496 */
2497 static __always_inline
2498 bool slab_free_hook(struct kmem_cache *s, void *x, bool init,
2499 bool after_rcu_delay)
2500 {
2501 /* Are the object contents still accessible? */
2502 bool still_accessible = (s->flags & SLAB_TYPESAFE_BY_RCU) && !after_rcu_delay;
2503
2504 kmemleak_free_recursive(x, s->flags);
2505 kmsan_slab_free(s, x);
2506
2507 debug_check_no_locks_freed(x, s->object_size);
2508
2509 if (!(s->flags & SLAB_DEBUG_OBJECTS))
2510 debug_check_no_obj_freed(x, s->object_size);
2511
2512 /* Use KCSAN to help debug racy use-after-free. */
2513 if (!still_accessible)
2514 __kcsan_check_access(x, s->object_size,
2515 KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT);
2516
2517 if (kfence_free(x))
2518 return false;
2519
2520 /*
2521 * Give KASAN a chance to notice an invalid free operation before we
2522 * modify the object.
2523 */
2524 if (kasan_slab_pre_free(s, x))
2525 return false;
2526
2527 #ifdef CONFIG_SLUB_RCU_DEBUG
2528 if (still_accessible) {
2529 struct rcu_delayed_free *delayed_free;
2530
2531 delayed_free = kmalloc(sizeof(*delayed_free), GFP_NOWAIT);
2532 if (delayed_free) {
2533 /*
2534 * Let KASAN track our call stack as a "related work
2535 * creation", just like if the object had been freed
2536 * normally via kfree_rcu().
2537 * We have to do this manually because the rcu_head is
2538 * not located inside the object.
2539 */
2540 kasan_record_aux_stack(x);
2541
2542 delayed_free->object = x;
2543 call_rcu(&delayed_free->head, slab_free_after_rcu_debug);
2544 return false;
2545 }
2546 }
2547 #endif /* CONFIG_SLUB_RCU_DEBUG */
2548
2549 /*
2550 * As memory initialization might be integrated into KASAN,
2551 * kasan_slab_free and initialization memset's must be
2552 * kept together to avoid discrepancies in behavior.
2553 *
2554 * The initialization memset's clear the object and the metadata,
2555 * but don't touch the SLAB redzone.
2556 *
2557 * The object's freepointer is also avoided if stored outside the
2558 * object.
2559 */
2560 if (unlikely(init)) {
2561 int rsize;
2562 unsigned int inuse, orig_size;
2563
2564 inuse = get_info_end(s);
2565 orig_size = get_orig_size(s, x);
2566 if (!kasan_has_integrated_init())
2567 memset(kasan_reset_tag(x), 0, orig_size);
2568 rsize = (s->flags & SLAB_RED_ZONE) ? s->red_left_pad : 0;
2569 memset((char *)kasan_reset_tag(x) + inuse, 0,
2570 s->size - inuse - rsize);
2571 /*
2572 * Restore orig_size, otherwise kmalloc redzone overwritten
2573 * would be reported
2574 */
2575 set_orig_size(s, x, orig_size);
2576
2577 }
2578 /* KASAN might put x into memory quarantine, delaying its reuse. */
2579 return !kasan_slab_free(s, x, init, still_accessible, false);
2580 }
2581
2582 static __fastpath_inline
2583 bool slab_free_freelist_hook(struct kmem_cache *s, void **head, void **tail,
2584 int *cnt)
2585 {
2586
2587 void *object;
2588 void *next = *head;
2589 void *old_tail = *tail;
2590 bool init;
2591
2592 if (is_kfence_address(next)) {
2593 slab_free_hook(s, next, false, false);
2594 return false;
2595 }
2596
2597 /* Head and tail of the reconstructed freelist */
2598 *head = NULL;
2599 *tail = NULL;
2600
2601 init = slab_want_init_on_free(s);
2602
2603 do {
2604 object = next;
2605 next = get_freepointer(s, object);
2606
2607 /* If object's reuse doesn't have to be delayed */
2608 if (likely(slab_free_hook(s, object, init, false))) {
2609 /* Move object to the new freelist */
2610 set_freepointer(s, object, *head);
2611 *head = object;
2612 if (!*tail)
2613 *tail = object;
2614 } else {
2615 /*
2616 * Adjust the reconstructed freelist depth
2617 * accordingly if object's reuse is delayed.
2618 */
2619 --(*cnt);
2620 }
2621 } while (object != old_tail);
2622
2623 return *head != NULL;
2624 }
2625
2626 static void *setup_object(struct kmem_cache *s, void *object)
2627 {
2628 setup_object_debug(s, object);
2629 object = kasan_init_slab_obj(s, object);
2630 if (unlikely(s->ctor)) {
2631 kasan_unpoison_new_object(s, object);
2632 s->ctor(object);
2633 kasan_poison_new_object(s, object);
2634 }
2635 return object;
2636 }
2637
2638 static struct slab_sheaf *alloc_empty_sheaf(struct kmem_cache *s, gfp_t gfp)
2639 {
2640 struct slab_sheaf *sheaf = kzalloc(struct_size(sheaf, objects,
2641 s->sheaf_capacity), gfp);
2642
2643 if (unlikely(!sheaf))
2644 return NULL;
2645
2646 sheaf->cache = s;
2647
2648 stat(s, SHEAF_ALLOC);
2649
2650 return sheaf;
2651 }
2652
2653 static void free_empty_sheaf(struct kmem_cache *s, struct slab_sheaf *sheaf)
2654 {
2655 kfree(sheaf);
2656
2657 stat(s, SHEAF_FREE);
2658 }
2659
2660 static int __kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags,
2661 size_t size, void **p);
2662
2663
2664 static int refill_sheaf(struct kmem_cache *s, struct slab_sheaf *sheaf,
2665 gfp_t gfp)
2666 {
2667 int to_fill = s->sheaf_capacity - sheaf->size;
2668 int filled;
2669
2670 if (!to_fill)
2671 return 0;
2672
2673 filled = __kmem_cache_alloc_bulk(s, gfp, to_fill,
2674 &sheaf->objects[sheaf->size]);
2675
2676 sheaf->size += filled;
2677
2678 stat_add(s, SHEAF_REFILL, filled);
2679
2680 if (filled < to_fill)
2681 return -ENOMEM;
2682
2683 return 0;
2684 }
2685
2686
2687 static struct slab_sheaf *alloc_full_sheaf(struct kmem_cache *s, gfp_t gfp)
2688 {
2689 struct slab_sheaf *sheaf = alloc_empty_sheaf(s, gfp);
2690
2691 if (!sheaf)
2692 return NULL;
2693
2694 if (refill_sheaf(s, sheaf, gfp)) {
2695 free_empty_sheaf(s, sheaf);
2696 return NULL;
2697 }
2698
2699 return sheaf;
2700 }
2701
2702 /*
2703 * Maximum number of objects freed during a single flush of main pcs sheaf.
2704 * Translates directly to an on-stack array size.
2705 */
2706 #define PCS_BATCH_MAX 32U
2707
2708 static void __kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p);
2709
2710 /*
2711 * Free all objects from the main sheaf. In order to perform
2712 * __kmem_cache_free_bulk() outside of cpu_sheaves->lock, work in batches where
2713 * object pointers are moved to a on-stack array under the lock. To bound the
2714 * stack usage, limit each batch to PCS_BATCH_MAX.
2715 *
2716 * Must be called with s->cpu_sheaves->lock locked, returns with the lock
2717 * unlocked.
2718 *
2719 * Returns how many objects are remaining to be flushed
2720 */
2721 static unsigned int __sheaf_flush_main_batch(struct kmem_cache *s)
2722 {
2723 struct slub_percpu_sheaves *pcs;
2724 unsigned int batch, remaining;
2725 void *objects[PCS_BATCH_MAX];
2726 struct slab_sheaf *sheaf;
2727
2728 lockdep_assert_held(this_cpu_ptr(&s->cpu_sheaves->lock));
2729
2730 pcs = this_cpu_ptr(s->cpu_sheaves);
2731 sheaf = pcs->main;
2732
2733 batch = min(PCS_BATCH_MAX, sheaf->size);
2734
2735 sheaf->size -= batch;
2736 memcpy(objects, sheaf->objects + sheaf->size, batch * sizeof(void *));
2737
2738 remaining = sheaf->size;
2739
2740 local_unlock(&s->cpu_sheaves->lock);
2741
2742 __kmem_cache_free_bulk(s, batch, &objects[0]);
2743
2744 stat_add(s, SHEAF_FLUSH, batch);
2745
2746 return remaining;
2747 }
2748
2749 static void sheaf_flush_main(struct kmem_cache *s)
2750 {
2751 unsigned int remaining;
2752
2753 do {
2754 local_lock(&s->cpu_sheaves->lock);
2755
2756 remaining = __sheaf_flush_main_batch(s);
2757
2758 } while (remaining);
2759 }
2760
2761 /*
2762 * Returns true if the main sheaf was at least partially flushed.
2763 */
2764 static bool sheaf_try_flush_main(struct kmem_cache *s)
2765 {
2766 unsigned int remaining;
2767 bool ret = false;
2768
2769 do {
2770 if (!local_trylock(&s->cpu_sheaves->lock))
2771 return ret;
2772
2773 ret = true;
2774 remaining = __sheaf_flush_main_batch(s);
2775
2776 } while (remaining);
2777
2778 return ret;
2779 }
2780
2781 /*
2782 * Free all objects from a sheaf that's unused, i.e. not linked to any
2783 * cpu_sheaves, so we need no locking and batching. The locking is also not
2784 * necessary when flushing cpu's sheaves (both spare and main) during cpu
2785 * hotremove as the cpu is not executing anymore.
2786 */
2787 static void sheaf_flush_unused(struct kmem_cache *s, struct slab_sheaf *sheaf)
2788 {
2789 if (!sheaf->size)
2790 return;
2791
2792 stat_add(s, SHEAF_FLUSH, sheaf->size);
2793
2794 __kmem_cache_free_bulk(s, sheaf->size, &sheaf->objects[0]);
2795
2796 sheaf->size = 0;
2797 }
2798
2799 static void __rcu_free_sheaf_prepare(struct kmem_cache *s,
2800 struct slab_sheaf *sheaf)
2801 {
2802 bool init = slab_want_init_on_free(s);
2803 void **p = &sheaf->objects[0];
2804 unsigned int i = 0;
2805
2806 while (i < sheaf->size) {
2807 struct slab *slab = virt_to_slab(p[i]);
2808
2809 memcg_slab_free_hook(s, slab, p + i, 1);
2810 alloc_tagging_slab_free_hook(s, slab, p + i, 1);
2811
2812 if (unlikely(!slab_free_hook(s, p[i], init, true))) {
2813 p[i] = p[--sheaf->size];
2814 continue;
2815 }
2816
2817 i++;
2818 }
2819 }
2820
2821 static void rcu_free_sheaf_nobarn(struct rcu_head *head)
2822 {
2823 struct slab_sheaf *sheaf;
2824 struct kmem_cache *s;
2825
2826 sheaf = container_of(head, struct slab_sheaf, rcu_head);
2827 s = sheaf->cache;
2828
2829 __rcu_free_sheaf_prepare(s, sheaf);
2830
2831 sheaf_flush_unused(s, sheaf);
2832
2833 free_empty_sheaf(s, sheaf);
2834 }
2835
2836 /*
2837 * Caller needs to make sure migration is disabled in order to fully flush
2838 * single cpu's sheaves
2839 *
2840 * must not be called from an irq
2841 *
2842 * flushing operations are rare so let's keep it simple and flush to slabs
2843 * directly, skipping the barn
2844 */
2845 static void pcs_flush_all(struct kmem_cache *s)
2846 {
2847 struct slub_percpu_sheaves *pcs;
2848 struct slab_sheaf *spare, *rcu_free;
2849
2850 local_lock(&s->cpu_sheaves->lock);
2851 pcs = this_cpu_ptr(s->cpu_sheaves);
2852
2853 spare = pcs->spare;
2854 pcs->spare = NULL;
2855
2856 rcu_free = pcs->rcu_free;
2857 pcs->rcu_free = NULL;
2858
2859 local_unlock(&s->cpu_sheaves->lock);
2860
2861 if (spare) {
2862 sheaf_flush_unused(s, spare);
2863 free_empty_sheaf(s, spare);
2864 }
2865
2866 if (rcu_free)
2867 call_rcu(&rcu_free->rcu_head, rcu_free_sheaf_nobarn);
2868
2869 sheaf_flush_main(s);
2870 }
2871
2872 static void __pcs_flush_all_cpu(struct kmem_cache *s, unsigned int cpu)
2873 {
2874 struct slub_percpu_sheaves *pcs;
2875
2876 pcs = per_cpu_ptr(s->cpu_sheaves, cpu);
2877
2878 /* The cpu is not executing anymore so we don't need pcs->lock */
2879 sheaf_flush_unused(s, pcs->main);
2880 if (pcs->spare) {
2881 sheaf_flush_unused(s, pcs->spare);
2882 free_empty_sheaf(s, pcs->spare);
2883 pcs->spare = NULL;
2884 }
2885
2886 if (pcs->rcu_free) {
2887 call_rcu(&pcs->rcu_free->rcu_head, rcu_free_sheaf_nobarn);
2888 pcs->rcu_free = NULL;
2889 }
2890 }
2891
2892 static void pcs_destroy(struct kmem_cache *s)
2893 {
2894 int cpu;
2895
2896 for_each_possible_cpu(cpu) {
2897 struct slub_percpu_sheaves *pcs;
2898
2899 pcs = per_cpu_ptr(s->cpu_sheaves, cpu);
2900
2901 /* can happen when unwinding failed create */
2902 if (!pcs->main)
2903 continue;
2904
2905 /*
2906 * We have already passed __kmem_cache_shutdown() so everything
2907 * was flushed and there should be no objects allocated from
2908 * slabs, otherwise kmem_cache_destroy() would have aborted.
2909 * Therefore something would have to be really wrong if the
2910 * warnings here trigger, and we should rather leave objects and
2911 * sheaves to leak in that case.
2912 */
2913
2914 WARN_ON(pcs->spare);
2915 WARN_ON(pcs->rcu_free);
2916
2917 if (!WARN_ON(pcs->main->size)) {
2918 free_empty_sheaf(s, pcs->main);
2919 pcs->main = NULL;
2920 }
2921 }
2922
2923 free_percpu(s->cpu_sheaves);
2924 s->cpu_sheaves = NULL;
2925 }
2926
2927 static struct slab_sheaf *barn_get_empty_sheaf(struct node_barn *barn)
2928 {
2929 struct slab_sheaf *empty = NULL;
2930 unsigned long flags;
2931
2932 if (!data_race(barn->nr_empty))
2933 return NULL;
2934
2935 spin_lock_irqsave(&barn->lock, flags);
2936
2937 if (likely(barn->nr_empty)) {
2938 empty = list_first_entry(&barn->sheaves_empty,
2939 struct slab_sheaf, barn_list);
2940 list_del(&empty->barn_list);
2941 barn->nr_empty--;
2942 }
2943
2944 spin_unlock_irqrestore(&barn->lock, flags);
2945
2946 return empty;
2947 }
2948
2949 /*
2950 * The following two functions are used mainly in cases where we have to undo an
2951 * intended action due to a race or cpu migration. Thus they do not check the
2952 * empty or full sheaf limits for simplicity.
2953 */
2954
2955 static void barn_put_empty_sheaf(struct node_barn *barn, struct slab_sheaf *sheaf)
2956 {
2957 unsigned long flags;
2958
2959 spin_lock_irqsave(&barn->lock, flags);
2960
2961 list_add(&sheaf->barn_list, &barn->sheaves_empty);
2962 barn->nr_empty++;
2963
2964 spin_unlock_irqrestore(&barn->lock, flags);
2965 }
2966
2967 static void barn_put_full_sheaf(struct node_barn *barn, struct slab_sheaf *sheaf)
2968 {
2969 unsigned long flags;
2970
2971 spin_lock_irqsave(&barn->lock, flags);
2972
2973 list_add(&sheaf->barn_list, &barn->sheaves_full);
2974 barn->nr_full++;
2975
2976 spin_unlock_irqrestore(&barn->lock, flags);
2977 }
2978
2979 static struct slab_sheaf *barn_get_full_or_empty_sheaf(struct node_barn *barn)
2980 {
2981 struct slab_sheaf *sheaf = NULL;
2982 unsigned long flags;
2983
2984 if (!data_race(barn->nr_full) && !data_race(barn->nr_empty))
2985 return NULL;
2986
2987 spin_lock_irqsave(&barn->lock, flags);
2988
2989 if (barn->nr_full) {
2990 sheaf = list_first_entry(&barn->sheaves_full, struct slab_sheaf,
2991 barn_list);
2992 list_del(&sheaf->barn_list);
2993 barn->nr_full--;
2994 } else if (barn->nr_empty) {
2995 sheaf = list_first_entry(&barn->sheaves_empty,
2996 struct slab_sheaf, barn_list);
2997 list_del(&sheaf->barn_list);
2998 barn->nr_empty--;
2999 }
3000
3001 spin_unlock_irqrestore(&barn->lock, flags);
3002
3003 return sheaf;
3004 }
3005
3006 /*
3007 * If a full sheaf is available, return it and put the supplied empty one to
3008 * barn. We ignore the limit on empty sheaves as the number of sheaves doesn't
3009 * change.
3010 */
3011 static struct slab_sheaf *
3012 barn_replace_empty_sheaf(struct node_barn *barn, struct slab_sheaf *empty)
3013 {
3014 struct slab_sheaf *full = NULL;
3015 unsigned long flags;
3016
3017 if (!data_race(barn->nr_full))
3018 return NULL;
3019
3020 spin_lock_irqsave(&barn->lock, flags);
3021
3022 if (likely(barn->nr_full)) {
3023 full = list_first_entry(&barn->sheaves_full, struct slab_sheaf,
3024 barn_list);
3025 list_del(&full->barn_list);
3026 list_add(&empty->barn_list, &barn->sheaves_empty);
3027 barn->nr_full--;
3028 barn->nr_empty++;
3029 }
3030
3031 spin_unlock_irqrestore(&barn->lock, flags);
3032
3033 return full;
3034 }
3035
3036 /*
3037 * If an empty sheaf is available, return it and put the supplied full one to
3038 * barn. But if there are too many full sheaves, reject this with -E2BIG.
3039 */
3040 static struct slab_sheaf *
3041 barn_replace_full_sheaf(struct node_barn *barn, struct slab_sheaf *full)
3042 {
3043 struct slab_sheaf *empty;
3044 unsigned long flags;
3045
3046 /* we don't repeat this check under barn->lock as it's not critical */
3047 if (data_race(barn->nr_full) >= MAX_FULL_SHEAVES)
3048 return ERR_PTR(-E2BIG);
3049 if (!data_race(barn->nr_empty))
3050 return ERR_PTR(-ENOMEM);
3051
3052 spin_lock_irqsave(&barn->lock, flags);
3053
3054 if (likely(barn->nr_empty)) {
3055 empty = list_first_entry(&barn->sheaves_empty, struct slab_sheaf,
3056 barn_list);
3057 list_del(&empty->barn_list);
3058 list_add(&full->barn_list, &barn->sheaves_full);
3059 barn->nr_empty--;
3060 barn->nr_full++;
3061 } else {
3062 empty = ERR_PTR(-ENOMEM);
3063 }
3064
3065 spin_unlock_irqrestore(&barn->lock, flags);
3066
3067 return empty;
3068 }
3069
3070 static void barn_init(struct node_barn *barn)
3071 {
3072 spin_lock_init(&barn->lock);
3073 INIT_LIST_HEAD(&barn->sheaves_full);
3074 INIT_LIST_HEAD(&barn->sheaves_empty);
3075 barn->nr_full = 0;
3076 barn->nr_empty = 0;
3077 }
3078
3079 static void barn_shrink(struct kmem_cache *s, struct node_barn *barn)
3080 {
3081 struct list_head empty_list;
3082 struct list_head full_list;
3083 struct slab_sheaf *sheaf, *sheaf2;
3084 unsigned long flags;
3085
3086 INIT_LIST_HEAD(&empty_list);
3087 INIT_LIST_HEAD(&full_list);
3088
3089 spin_lock_irqsave(&barn->lock, flags);
3090
3091 list_splice_init(&barn->sheaves_full, &full_list);
3092 barn->nr_full = 0;
3093 list_splice_init(&barn->sheaves_empty, &empty_list);
3094 barn->nr_empty = 0;
3095
3096 spin_unlock_irqrestore(&barn->lock, flags);
3097
3098 list_for_each_entry_safe(sheaf, sheaf2, &full_list, barn_list) {
3099 sheaf_flush_unused(s, sheaf);
3100 free_empty_sheaf(s, sheaf);
3101 }
3102
3103 list_for_each_entry_safe(sheaf, sheaf2, &empty_list, barn_list)
3104 free_empty_sheaf(s, sheaf);
3105 }
3106
3107 /*
3108 * Slab allocation and freeing
3109 */
3110 static inline struct slab *alloc_slab_page(gfp_t flags, int node,
3111 struct kmem_cache_order_objects oo,
3112 bool allow_spin)
3113 {
3114 struct folio *folio;
3115 struct slab *slab;
3116 unsigned int order = oo_order(oo);
3117
3118 if (unlikely(!allow_spin))
3119 folio = (struct folio *)alloc_frozen_pages_nolock(0/* __GFP_COMP is implied */,
3120 node, order);
3121 else if (node == NUMA_NO_NODE)
3122 folio = (struct folio *)alloc_frozen_pages(flags, order);
3123 else
3124 folio = (struct folio *)__alloc_frozen_pages(flags, order, node, NULL);
3125
3126 if (!folio)
3127 return NULL;
3128
3129 slab = folio_slab(folio);
3130 __folio_set_slab(folio);
3131 if (folio_is_pfmemalloc(folio))
3132 slab_set_pfmemalloc(slab);
3133
3134 return slab;
3135 }
3136
3137 #ifdef CONFIG_SLAB_FREELIST_RANDOM
3138 /* Pre-initialize the random sequence cache */
3139 static int init_cache_random_seq(struct kmem_cache *s)
3140 {
3141 unsigned int count = oo_objects(s->oo);
3142 int err;
3143
3144 /* Bailout if already initialised */
3145 if (s->random_seq)
3146 return 0;
3147
3148 err = cache_random_seq_create(s, count, GFP_KERNEL);
3149 if (err) {
3150 pr_err("SLUB: Unable to initialize free list for %s\n",
3151 s->name);
3152 return err;
3153 }
3154
3155 /* Transform to an offset on the set of pages */
3156 if (s->random_seq) {
3157 unsigned int i;
3158
3159 for (i = 0; i < count; i++)
3160 s->random_seq[i] *= s->size;
3161 }
3162 return 0;
3163 }
3164
3165 /* Initialize each random sequence freelist per cache */
3166 static void __init init_freelist_randomization(void)
3167 {
3168 struct kmem_cache *s;
3169
3170 mutex_lock(&slab_mutex);
3171
3172 list_for_each_entry(s, &slab_caches, list)
3173 init_cache_random_seq(s);
3174
3175 mutex_unlock(&slab_mutex);
3176 }
3177
3178 /* Get the next entry on the pre-computed freelist randomized */
3179 static void *next_freelist_entry(struct kmem_cache *s,
3180 unsigned long *pos, void *start,
3181 unsigned long page_limit,
3182 unsigned long freelist_count)
3183 {
3184 unsigned int idx;
3185
3186 /*
3187 * If the target page allocation failed, the number of objects on the
3188 * page might be smaller than the usual size defined by the cache.
3189 */
3190 do {
3191 idx = s->random_seq[*pos];
3192 *pos += 1;
3193 if (*pos >= freelist_count)
3194 *pos = 0;
3195 } while (unlikely(idx >= page_limit));
3196
3197 return (char *)start + idx;
3198 }
3199
3200 static DEFINE_PER_CPU(struct rnd_state, slab_rnd_state);
3201
3202 /* Shuffle the single linked freelist based on a random pre-computed sequence */
3203 static bool shuffle_freelist(struct kmem_cache *s, struct slab *slab,
3204 bool allow_spin)
3205 {
3206 void *start;
3207 void *cur;
3208 void *next;
3209 unsigned long idx, pos, page_limit, freelist_count;
3210
3211 if (slab->objects < 2 || !s->random_seq)
3212 return false;
3213
3214 freelist_count = oo_objects(s->oo);
3215 if (allow_spin) {
3216 pos = get_random_u32_below(freelist_count);
3217 } else {
3218 struct rnd_state *state;
3219
3220 /*
3221 * An interrupt or NMI handler might interrupt and change
3222 * the state in the middle, but that's safe.
3223 */
3224 state = &get_cpu_var(slab_rnd_state);
3225 pos = prandom_u32_state(state) % freelist_count;
3226 put_cpu_var(slab_rnd_state);
3227 }
3228
3229 page_limit = slab->objects * s->size;
3230 start = fixup_red_left(s, slab_address(slab));
3231
3232 /* First entry is used as the base of the freelist */
3233 cur = next_freelist_entry(s, &pos, start, page_limit, freelist_count);
3234 cur = setup_object(s, cur);
3235 slab->freelist = cur;
3236
3237 for (idx = 1; idx < slab->objects; idx++) {
3238 next = next_freelist_entry(s, &pos, start, page_limit,
3239 freelist_count);
3240 next = setup_object(s, next);
3241 set_freepointer(s, cur, next);
3242 cur = next;
3243 }
3244 set_freepointer(s, cur, NULL);
3245
3246 return true;
3247 }
3248 #else
3249 static inline int init_cache_random_seq(struct kmem_cache *s)
3250 {
3251 return 0;
3252 }
3253 static inline void init_freelist_randomization(void) { }
3254 static inline bool shuffle_freelist(struct kmem_cache *s, struct slab *slab,
3255 bool allow_spin)
3256 {
3257 return false;
3258 }
3259 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
3260
3261 static __always_inline void account_slab(struct slab *slab, int order,
3262 struct kmem_cache *s, gfp_t gfp)
3263 {
3264 if (memcg_kmem_online() && (s->flags & SLAB_ACCOUNT))
3265 alloc_slab_obj_exts(slab, s, gfp, true);
3266
3267 mod_node_page_state(slab_pgdat(slab), cache_vmstat_idx(s),
3268 PAGE_SIZE << order);
3269 }
3270
3271 static __always_inline void unaccount_slab(struct slab *slab, int order,
3272 struct kmem_cache *s)
3273 {
3274 /*
3275 * The slab object extensions should now be freed regardless of
3276 * whether mem_alloc_profiling_enabled() or not because profiling
3277 * might have been disabled after slab->obj_exts got allocated.
3278 */
3279 free_slab_obj_exts(slab);
3280
3281 mod_node_page_state(slab_pgdat(slab), cache_vmstat_idx(s),
3282 -(PAGE_SIZE << order));
3283 }
3284
3285 static struct slab *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
3286 {
3287 bool allow_spin = gfpflags_allow_spinning(flags);
3288 struct slab *slab;
3289 struct kmem_cache_order_objects oo = s->oo;
3290 gfp_t alloc_gfp;
3291 void *start, *p, *next;
3292 int idx;
3293 bool shuffle;
3294
3295 flags &= gfp_allowed_mask;
3296
3297 flags |= s->allocflags;
3298
3299 /*
3300 * Let the initial higher-order allocation fail under memory pressure
3301 * so we fall-back to the minimum order allocation.
3302 */
3303 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
3304 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
3305 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~__GFP_RECLAIM;
3306
3307 /*
3308 * __GFP_RECLAIM could be cleared on the first allocation attempt,
3309 * so pass allow_spin flag directly.
3310 */
3311 slab = alloc_slab_page(alloc_gfp, node, oo, allow_spin);
3312 if (unlikely(!slab)) {
3313 oo = s->min;
3314 alloc_gfp = flags;
3315 /*
3316 * Allocation may have failed due to fragmentation.
3317 * Try a lower order alloc if possible
3318 */
3319 slab = alloc_slab_page(alloc_gfp, node, oo, allow_spin);
3320 if (unlikely(!slab))
3321 return NULL;
3322 stat(s, ORDER_FALLBACK);
3323 }
3324
3325 slab->objects = oo_objects(oo);
3326 slab->inuse = 0;
3327 slab->frozen = 0;
3328 init_slab_obj_exts(slab);
3329
3330 account_slab(slab, oo_order(oo), s, flags);
3331
3332 slab->slab_cache = s;
3333
3334 kasan_poison_slab(slab);
3335
3336 start = slab_address(slab);
3337
3338 setup_slab_debug(s, slab, start);
3339
3340 shuffle = shuffle_freelist(s, slab, allow_spin);
3341
3342 if (!shuffle) {
3343 start = fixup_red_left(s, start);
3344 start = setup_object(s, start);
3345 slab->freelist = start;
3346 for (idx = 0, p = start; idx < slab->objects - 1; idx++) {
3347 next = p + s->size;
3348 next = setup_object(s, next);
3349 set_freepointer(s, p, next);
3350 p = next;
3351 }
3352 set_freepointer(s, p, NULL);
3353 }
3354
3355 return slab;
3356 }
3357
3358 static struct slab *new_slab(struct kmem_cache *s, gfp_t flags, int node)
3359 {
3360 if (unlikely(flags & GFP_SLAB_BUG_MASK))
3361 flags = kmalloc_fix_flags(flags);
3362
3363 WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO));
3364
3365 return allocate_slab(s,
3366 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
3367 }
3368
3369 static void __free_slab(struct kmem_cache *s, struct slab *slab)
3370 {
3371 struct folio *folio = slab_folio(slab);
3372 int order = folio_order(folio);
3373 int pages = 1 << order;
3374
3375 __slab_clear_pfmemalloc(slab);
3376 folio->mapping = NULL;
3377 __folio_clear_slab(folio);
3378 mm_account_reclaimed_pages(pages);
3379 unaccount_slab(slab, order, s);
3380 free_frozen_pages(&folio->page, order);
3381 }
3382
3383 static void rcu_free_slab(struct rcu_head *h)
3384 {
3385 struct slab *slab = container_of(h, struct slab, rcu_head);
3386
3387 __free_slab(slab->slab_cache, slab);
3388 }
3389
3390 static void free_slab(struct kmem_cache *s, struct slab *slab)
3391 {
3392 if (kmem_cache_debug_flags(s, SLAB_CONSISTENCY_CHECKS)) {
3393 void *p;
3394
3395 slab_pad_check(s, slab);
3396 for_each_object(p, s, slab_address(slab), slab->objects)
3397 check_object(s, slab, p, SLUB_RED_INACTIVE);
3398 }
3399
3400 if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU))
3401 call_rcu(&slab->rcu_head, rcu_free_slab);
3402 else
3403 __free_slab(s, slab);
3404 }
3405
3406 static void discard_slab(struct kmem_cache *s, struct slab *slab)
3407 {
3408 dec_slabs_node(s, slab_nid(slab), slab->objects);
3409 free_slab(s, slab);
3410 }
3411
3412 static inline bool slab_test_node_partial(const struct slab *slab)
3413 {
3414 return test_bit(SL_partial, &slab->flags.f);
3415 }
3416
3417 static inline void slab_set_node_partial(struct slab *slab)
3418 {
3419 set_bit(SL_partial, &slab->flags.f);
3420 }
3421
3422 static inline void slab_clear_node_partial(struct slab *slab)
3423 {
3424 clear_bit(SL_partial, &slab->flags.f);
3425 }
3426
3427 /*
3428 * Management of partially allocated slabs.
3429 */
3430 static inline void
3431 __add_partial(struct kmem_cache_node *n, struct slab *slab, int tail)
3432 {
3433 n->nr_partial++;
3434 if (tail == DEACTIVATE_TO_TAIL)
3435 list_add_tail(&slab->slab_list, &n->partial);
3436 else
3437 list_add(&slab->slab_list, &n->partial);
3438 slab_set_node_partial(slab);
3439 }
3440
3441 static inline void add_partial(struct kmem_cache_node *n,
3442 struct slab *slab, int tail)
3443 {
3444 lockdep_assert_held(&n->list_lock);
3445 __add_partial(n, slab, tail);
3446 }
3447
3448 static inline void remove_partial(struct kmem_cache_node *n,
3449 struct slab *slab)
3450 {
3451 lockdep_assert_held(&n->list_lock);
3452 list_del(&slab->slab_list);
3453 slab_clear_node_partial(slab);
3454 n->nr_partial--;
3455 }
3456
3457 /*
3458 * Called only for kmem_cache_debug() caches instead of remove_partial(), with a
3459 * slab from the n->partial list. Remove only a single object from the slab, do
3460 * the alloc_debug_processing() checks and leave the slab on the list, or move
3461 * it to full list if it was the last free object.
3462 */
3463 static void *alloc_single_from_partial(struct kmem_cache *s,
3464 struct kmem_cache_node *n, struct slab *slab, int orig_size)
3465 {
3466 void *object;
3467
3468 lockdep_assert_held(&n->list_lock);
3469
3470 #ifdef CONFIG_SLUB_DEBUG
3471 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
3472 if (!validate_slab_ptr(slab)) {
3473 slab_err(s, slab, "Not a valid slab page");
3474 return NULL;
3475 }
3476 }
3477 #endif
3478
3479 object = slab->freelist;
3480 slab->freelist = get_freepointer(s, object);
3481 slab->inuse++;
3482
3483 if (!alloc_debug_processing(s, slab, object, orig_size)) {
3484 remove_partial(n, slab);
3485 return NULL;
3486 }
3487
3488 if (slab->inuse == slab->objects) {
3489 remove_partial(n, slab);
3490 add_full(s, n, slab);
3491 }
3492
3493 return object;
3494 }
3495
3496 static void defer_deactivate_slab(struct slab *slab, void *flush_freelist);
3497
3498 /*
3499 * Called only for kmem_cache_debug() caches to allocate from a freshly
3500 * allocated slab. Allocate a single object instead of whole freelist
3501 * and put the slab to the partial (or full) list.
3502 */
3503 static void *alloc_single_from_new_slab(struct kmem_cache *s, struct slab *slab,
3504 int orig_size, gfp_t gfpflags)
3505 {
3506 bool allow_spin = gfpflags_allow_spinning(gfpflags);
3507 int nid = slab_nid(slab);
3508 struct kmem_cache_node *n = get_node(s, nid);
3509 unsigned long flags;
3510 void *object;
3511
3512 if (!allow_spin && !spin_trylock_irqsave(&n->list_lock, flags)) {
3513 /* Unlucky, discard newly allocated slab */
3514 defer_deactivate_slab(slab, NULL);
3515 return NULL;
3516 }
3517
3518 object = slab->freelist;
3519 slab->freelist = get_freepointer(s, object);
3520 slab->inuse = 1;
3521
3522 if (!alloc_debug_processing(s, slab, object, orig_size)) {
3523 /*
3524 * It's not really expected that this would fail on a
3525 * freshly allocated slab, but a concurrent memory
3526 * corruption in theory could cause that.
3527 * Leak memory of allocated slab.
3528 */
3529 if (!allow_spin)
3530 spin_unlock_irqrestore(&n->list_lock, flags);
3531 return NULL;
3532 }
3533
3534 if (allow_spin)
3535 spin_lock_irqsave(&n->list_lock, flags);
3536
3537 if (slab->inuse == slab->objects)
3538 add_full(s, n, slab);
3539 else
3540 add_partial(n, slab, DEACTIVATE_TO_HEAD);
3541
3542 inc_slabs_node(s, nid, slab->objects);
3543 spin_unlock_irqrestore(&n->list_lock, flags);
3544
3545 return object;
3546 }
3547
3548 #ifdef CONFIG_SLUB_CPU_PARTIAL
3549 static void put_cpu_partial(struct kmem_cache *s, struct slab *slab, int drain);
3550 #else
3551 static inline void put_cpu_partial(struct kmem_cache *s, struct slab *slab,
3552 int drain) { }
3553 #endif
3554 static inline bool pfmemalloc_match(struct slab *slab, gfp_t gfpflags);
3555
3556 /*
3557 * Try to allocate a partial slab from a specific node.
3558 */
3559 static struct slab *get_partial_node(struct kmem_cache *s,
3560 struct kmem_cache_node *n,
3561 struct partial_context *pc)
3562 {
3563 struct slab *slab, *slab2, *partial = NULL;
3564 unsigned long flags;
3565 unsigned int partial_slabs = 0;
3566
3567 /*
3568 * Racy check. If we mistakenly see no partial slabs then we
3569 * just allocate an empty slab. If we mistakenly try to get a
3570 * partial slab and there is none available then get_partial()
3571 * will return NULL.
3572 */
3573 if (!n || !n->nr_partial)
3574 return NULL;
3575
3576 if (gfpflags_allow_spinning(pc->flags))
3577 spin_lock_irqsave(&n->list_lock, flags);
3578 else if (!spin_trylock_irqsave(&n->list_lock, flags))
3579 return NULL;
3580 list_for_each_entry_safe(slab, slab2, &n->partial, slab_list) {
3581 if (!pfmemalloc_match(slab, pc->flags))
3582 continue;
3583
3584 if (IS_ENABLED(CONFIG_SLUB_TINY) || kmem_cache_debug(s)) {
3585 void *object = alloc_single_from_partial(s, n, slab,
3586 pc->orig_size);
3587 if (object) {
3588 partial = slab;
3589 pc->object = object;
3590 break;
3591 }
3592 continue;
3593 }
3594
3595 remove_partial(n, slab);
3596
3597 if (!partial) {
3598 partial = slab;
3599 stat(s, ALLOC_FROM_PARTIAL);
3600
3601 if ((slub_get_cpu_partial(s) == 0)) {
3602 break;
3603 }
3604 } else {
3605 put_cpu_partial(s, slab, 0);
3606 stat(s, CPU_PARTIAL_NODE);
3607
3608 if (++partial_slabs > slub_get_cpu_partial(s) / 2) {
3609 break;
3610 }
3611 }
3612 }
3613 spin_unlock_irqrestore(&n->list_lock, flags);
3614 return partial;
3615 }
3616
3617 /*
3618 * Get a slab from somewhere. Search in increasing NUMA distances.
3619 */
3620 static struct slab *get_any_partial(struct kmem_cache *s,
3621 struct partial_context *pc)
3622 {
3623 #ifdef CONFIG_NUMA
3624 struct zonelist *zonelist;
3625 struct zoneref *z;
3626 struct zone *zone;
3627 enum zone_type highest_zoneidx = gfp_zone(pc->flags);
3628 struct slab *slab;
3629 unsigned int cpuset_mems_cookie;
3630 bool allow_spin = gfpflags_allow_spinning(pc->flags);
3631
3632 /*
3633 * The defrag ratio allows a configuration of the tradeoffs between
3634 * inter node defragmentation and node local allocations. A lower
3635 * defrag_ratio increases the tendency to do local allocations
3636 * instead of attempting to obtain partial slabs from other nodes.
3637 *
3638 * If the defrag_ratio is set to 0 then kmalloc() always
3639 * returns node local objects. If the ratio is higher then kmalloc()
3640 * may return off node objects because partial slabs are obtained
3641 * from other nodes and filled up.
3642 *
3643 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
3644 * (which makes defrag_ratio = 1000) then every (well almost)
3645 * allocation will first attempt to defrag slab caches on other nodes.
3646 * This means scanning over all nodes to look for partial slabs which
3647 * may be expensive if we do it every time we are trying to find a slab
3648 * with available objects.
3649 */
3650 if (!s->remote_node_defrag_ratio ||
3651 get_cycles() % 1024 > s->remote_node_defrag_ratio)
3652 return NULL;
3653
3654 do {
3655 /*
3656 * read_mems_allowed_begin() accesses current->mems_allowed_seq,
3657 * a seqcount_spinlock_t that is not NMI-safe. Do not access
3658 * current->mems_allowed_seq and avoid retry when GFP flags
3659 * indicate spinning is not allowed.
3660 */
3661 if (allow_spin)
3662 cpuset_mems_cookie = read_mems_allowed_begin();
3663
3664 zonelist = node_zonelist(mempolicy_slab_node(), pc->flags);
3665 for_each_zone_zonelist(zone, z, zonelist, highest_zoneidx) {
3666 struct kmem_cache_node *n;
3667
3668 n = get_node(s, zone_to_nid(zone));
3669
3670 if (n && cpuset_zone_allowed(zone, pc->flags) &&
3671 n->nr_partial > s->min_partial) {
3672 slab = get_partial_node(s, n, pc);
3673 if (slab) {
3674 /*
3675 * Don't check read_mems_allowed_retry()
3676 * here - if mems_allowed was updated in
3677 * parallel, that was a harmless race
3678 * between allocation and the cpuset
3679 * update
3680 */
3681 return slab;
3682 }
3683 }
3684 }
3685 } while (allow_spin && read_mems_allowed_retry(cpuset_mems_cookie));
3686 #endif /* CONFIG_NUMA */
3687 return NULL;
3688 }
3689
3690 /*
3691 * Get a partial slab, lock it and return it.
3692 */
3693 static struct slab *get_partial(struct kmem_cache *s, int node,
3694 struct partial_context *pc)
3695 {
3696 struct slab *slab;
3697 int searchnode = node;
3698
3699 if (node == NUMA_NO_NODE)
3700 searchnode = numa_mem_id();
3701
3702 slab = get_partial_node(s, get_node(s, searchnode), pc);
3703 if (slab || (node != NUMA_NO_NODE && (pc->flags & __GFP_THISNODE)))
3704 return slab;
3705
3706 return get_any_partial(s, pc);
3707 }
3708
3709 #ifdef CONFIG_PREEMPTION
3710 /*
3711 * Calculate the next globally unique transaction for disambiguation
3712 * during cmpxchg. The transactions start with the cpu number and are then
3713 * incremented by CONFIG_NR_CPUS.
3714 */
3715 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
3716 #else
3717 /*
3718 * No preemption supported therefore also no need to check for
3719 * different cpus.
3720 */
3721 #define TID_STEP 1
3722 #endif /* CONFIG_PREEMPTION */
3723
3724 static inline unsigned long next_tid(unsigned long tid)
3725 {
3726 return tid + TID_STEP;
3727 }
3728
3729 #ifdef SLUB_DEBUG_CMPXCHG
3730 static inline unsigned int tid_to_cpu(unsigned long tid)
3731 {
3732 return tid % TID_STEP;
3733 }
3734
3735 static inline unsigned long tid_to_event(unsigned long tid)
3736 {
3737 return tid / TID_STEP;
3738 }
3739 #endif
3740
3741 static inline unsigned int init_tid(int cpu)
3742 {
3743 return cpu;
3744 }
3745
3746 static inline void note_cmpxchg_failure(const char *n,
3747 const struct kmem_cache *s, unsigned long tid)
3748 {
3749 #ifdef SLUB_DEBUG_CMPXCHG
3750 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
3751
3752 pr_info("%s %s: cmpxchg redo ", n, s->name);
3753
3754 if (IS_ENABLED(CONFIG_PREEMPTION) &&
3755 tid_to_cpu(tid) != tid_to_cpu(actual_tid)) {
3756 pr_warn("due to cpu change %d -> %d\n",
3757 tid_to_cpu(tid), tid_to_cpu(actual_tid));
3758 } else if (tid_to_event(tid) != tid_to_event(actual_tid)) {
3759 pr_warn("due to cpu running other code. Event %ld->%ld\n",
3760 tid_to_event(tid), tid_to_event(actual_tid));
3761 } else {
3762 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
3763 actual_tid, tid, next_tid(tid));
3764 }
3765 #endif
3766 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
3767 }
3768
3769 static void init_kmem_cache_cpus(struct kmem_cache *s)
3770 {
3771 #ifdef CONFIG_PREEMPT_RT
3772 /*
3773 * Register lockdep key for non-boot kmem caches to avoid
3774 * WARN_ON_ONCE(static_obj(key))) in lockdep_register_key()
3775 */
3776 bool finegrain_lockdep = !init_section_contains(s, 1);
3777 #else
3778 /*
3779 * Don't bother with different lockdep classes for each
3780 * kmem_cache, since we only use local_trylock_irqsave().
3781 */
3782 bool finegrain_lockdep = false;
3783 #endif
3784 int cpu;
3785 struct kmem_cache_cpu *c;
3786
3787 if (finegrain_lockdep)
3788 lockdep_register_key(&s->lock_key);
3789 for_each_possible_cpu(cpu) {
3790 c = per_cpu_ptr(s->cpu_slab, cpu);
3791 local_trylock_init(&c->lock);
3792 if (finegrain_lockdep)
3793 lockdep_set_class(&c->lock, &s->lock_key);
3794 c->tid = init_tid(cpu);
3795 }
3796 }
3797
3798 /*
3799 * Finishes removing the cpu slab. Merges cpu's freelist with slab's freelist,
3800 * unfreezes the slabs and puts it on the proper list.
3801 * Assumes the slab has been already safely taken away from kmem_cache_cpu
3802 * by the caller.
3803 */
3804 static void deactivate_slab(struct kmem_cache *s, struct slab *slab,
3805 void *freelist)
3806 {
3807 struct kmem_cache_node *n = get_node(s, slab_nid(slab));
3808 int free_delta = 0;
3809 void *nextfree, *freelist_iter, *freelist_tail;
3810 int tail = DEACTIVATE_TO_HEAD;
3811 unsigned long flags = 0;
3812 struct slab new;
3813 struct slab old;
3814
3815 if (READ_ONCE(slab->freelist)) {
3816 stat(s, DEACTIVATE_REMOTE_FREES);
3817 tail = DEACTIVATE_TO_TAIL;
3818 }
3819
3820 /*
3821 * Stage one: Count the objects on cpu's freelist as free_delta and
3822 * remember the last object in freelist_tail for later splicing.
3823 */
3824 freelist_tail = NULL;
3825 freelist_iter = freelist;
3826 while (freelist_iter) {
3827 nextfree = get_freepointer(s, freelist_iter);
3828
3829 /*
3830 * If 'nextfree' is invalid, it is possible that the object at
3831 * 'freelist_iter' is already corrupted. So isolate all objects
3832 * starting at 'freelist_iter' by skipping them.
3833 */
3834 if (freelist_corrupted(s, slab, &freelist_iter, nextfree))
3835 break;
3836
3837 freelist_tail = freelist_iter;
3838 free_delta++;
3839
3840 freelist_iter = nextfree;
3841 }
3842
3843 /*
3844 * Stage two: Unfreeze the slab while splicing the per-cpu
3845 * freelist to the head of slab's freelist.
3846 */
3847 do {
3848 old.freelist = READ_ONCE(slab->freelist);
3849 old.counters = READ_ONCE(slab->counters);
3850 VM_BUG_ON(!old.frozen);
3851
3852 /* Determine target state of the slab */
3853 new.counters = old.counters;
3854 new.frozen = 0;
3855 if (freelist_tail) {
3856 new.inuse -= free_delta;
3857 set_freepointer(s, freelist_tail, old.freelist);
3858 new.freelist = freelist;
3859 } else {
3860 new.freelist = old.freelist;
3861 }
3862 } while (!slab_update_freelist(s, slab,
3863 old.freelist, old.counters,
3864 new.freelist, new.counters,
3865 "unfreezing slab"));
3866
3867 /*
3868 * Stage three: Manipulate the slab list based on the updated state.
3869 */
3870 if (!new.inuse && n->nr_partial >= s->min_partial) {
3871 stat(s, DEACTIVATE_EMPTY);
3872 discard_slab(s, slab);
3873 stat(s, FREE_SLAB);
3874 } else if (new.freelist) {
3875 spin_lock_irqsave(&n->list_lock, flags);
3876 add_partial(n, slab, tail);
3877 spin_unlock_irqrestore(&n->list_lock, flags);
3878 stat(s, tail);
3879 } else {
3880 stat(s, DEACTIVATE_FULL);
3881 }
3882 }
3883
3884 /*
3885 * ___slab_alloc()'s caller is supposed to check if kmem_cache::kmem_cache_cpu::lock
3886 * can be acquired without a deadlock before invoking the function.
3887 *
3888 * Without LOCKDEP we trust the code to be correct. kmalloc_nolock() is
3889 * using local_lock_is_locked() properly before calling local_lock_cpu_slab(),
3890 * and kmalloc() is not used in an unsupported context.
3891 *
3892 * With LOCKDEP, on PREEMPT_RT lockdep does its checking in local_lock_irqsave().
3893 * On !PREEMPT_RT we use trylock to avoid false positives in NMI, but
3894 * lockdep_assert() will catch a bug in case:
3895 * #1
3896 * kmalloc() -> ___slab_alloc() -> irqsave -> NMI -> bpf -> kmalloc_nolock()
3897 * or
3898 * #2
3899 * kmalloc() -> ___slab_alloc() -> irqsave -> tracepoint/kprobe -> bpf -> kmalloc_nolock()
3900 *
3901 * On PREEMPT_RT an invocation is not possible from IRQ-off or preempt
3902 * disabled context. The lock will always be acquired and if needed it
3903 * block and sleep until the lock is available.
3904 * #1 is possible in !PREEMPT_RT only.
3905 * #2 is possible in both with a twist that irqsave is replaced with rt_spinlock:
3906 * kmalloc() -> ___slab_alloc() -> rt_spin_lock(kmem_cache_A) ->
3907 * tracepoint/kprobe -> bpf -> kmalloc_nolock() -> rt_spin_lock(kmem_cache_B)
3908 *
3909 * local_lock_is_locked() prevents the case kmem_cache_A == kmem_cache_B
3910 */
3911 #if defined(CONFIG_PREEMPT_RT) || !defined(CONFIG_LOCKDEP)
3912 #define local_lock_cpu_slab(s, flags) \
3913 local_lock_irqsave(&(s)->cpu_slab->lock, flags)
3914 #else
3915 #define local_lock_cpu_slab(s, flags) \
3916 do { \
3917 bool __l = local_trylock_irqsave(&(s)->cpu_slab->lock, flags); \
3918 lockdep_assert(__l); \
3919 } while (0)
3920 #endif
3921
3922 #define local_unlock_cpu_slab(s, flags) \
3923 local_unlock_irqrestore(&(s)->cpu_slab->lock, flags)
3924
3925 #ifdef CONFIG_SLUB_CPU_PARTIAL
3926 static void __put_partials(struct kmem_cache *s, struct slab *partial_slab)
3927 {
3928 struct kmem_cache_node *n = NULL, *n2 = NULL;
3929 struct slab *slab, *slab_to_discard = NULL;
3930 unsigned long flags = 0;
3931
3932 while (partial_slab) {
3933 slab = partial_slab;
3934 partial_slab = slab->next;
3935
3936 n2 = get_node(s, slab_nid(slab));
3937 if (n != n2) {
3938 if (n)
3939 spin_unlock_irqrestore(&n->list_lock, flags);
3940
3941 n = n2;
3942 spin_lock_irqsave(&n->list_lock, flags);
3943 }
3944
3945 if (unlikely(!slab->inuse && n->nr_partial >= s->min_partial)) {
3946 slab->next = slab_to_discard;
3947 slab_to_discard = slab;
3948 } else {
3949 add_partial(n, slab, DEACTIVATE_TO_TAIL);
3950 stat(s, FREE_ADD_PARTIAL);
3951 }
3952 }
3953
3954 if (n)
3955 spin_unlock_irqrestore(&n->list_lock, flags);
3956
3957 while (slab_to_discard) {
3958 slab = slab_to_discard;
3959 slab_to_discard = slab_to_discard->next;
3960
3961 stat(s, DEACTIVATE_EMPTY);
3962 discard_slab(s, slab);
3963 stat(s, FREE_SLAB);
3964 }
3965 }
3966
3967 /*
3968 * Put all the cpu partial slabs to the node partial list.
3969 */
3970 static void put_partials(struct kmem_cache *s)
3971 {
3972 struct slab *partial_slab;
3973 unsigned long flags;
3974
3975 local_lock_irqsave(&s->cpu_slab->lock, flags);
3976 partial_slab = this_cpu_read(s->cpu_slab->partial);
3977 this_cpu_write(s->cpu_slab->partial, NULL);
3978 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3979
3980 if (partial_slab)
3981 __put_partials(s, partial_slab);
3982 }
3983
3984 static void put_partials_cpu(struct kmem_cache *s,
3985 struct kmem_cache_cpu *c)
3986 {
3987 struct slab *partial_slab;
3988
3989 partial_slab = slub_percpu_partial(c);
3990 c->partial = NULL;
3991
3992 if (partial_slab)
3993 __put_partials(s, partial_slab);
3994 }
3995
3996 /*
3997 * Put a slab into a partial slab slot if available.
3998 *
3999 * If we did not find a slot then simply move all the partials to the
4000 * per node partial list.
4001 */
4002 static void put_cpu_partial(struct kmem_cache *s, struct slab *slab, int drain)
4003 {
4004 struct slab *oldslab;
4005 struct slab *slab_to_put = NULL;
4006 unsigned long flags;
4007 int slabs = 0;
4008
4009 local_lock_cpu_slab(s, flags);
4010
4011 oldslab = this_cpu_read(s->cpu_slab->partial);
4012
4013 if (oldslab) {
4014 if (drain && oldslab->slabs >= s->cpu_partial_slabs) {
4015 /*
4016 * Partial array is full. Move the existing set to the
4017 * per node partial list. Postpone the actual unfreezing
4018 * outside of the critical section.
4019 */
4020 slab_to_put = oldslab;
4021 oldslab = NULL;
4022 } else {
4023 slabs = oldslab->slabs;
4024 }
4025 }
4026
4027 slabs++;
4028
4029 slab->slabs = slabs;
4030 slab->next = oldslab;
4031
4032 this_cpu_write(s->cpu_slab->partial, slab);
4033
4034 local_unlock_cpu_slab(s, flags);
4035
4036 if (slab_to_put) {
4037 __put_partials(s, slab_to_put);
4038 stat(s, CPU_PARTIAL_DRAIN);
4039 }
4040 }
4041
4042 #else /* CONFIG_SLUB_CPU_PARTIAL */
4043
4044 static inline void put_partials(struct kmem_cache *s) { }
4045 static inline void put_partials_cpu(struct kmem_cache *s,
4046 struct kmem_cache_cpu *c) { }
4047
4048 #endif /* CONFIG_SLUB_CPU_PARTIAL */
4049
4050 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
4051 {
4052 unsigned long flags;
4053 struct slab *slab;
4054 void *freelist;
4055
4056 local_lock_irqsave(&s->cpu_slab->lock, flags);
4057
4058 slab = c->slab;
4059 freelist = c->freelist;
4060
4061 c->slab = NULL;
4062 c->freelist = NULL;
4063 c->tid = next_tid(c->tid);
4064
4065 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
4066
4067 if (slab) {
4068 deactivate_slab(s, slab, freelist);
4069 stat(s, CPUSLAB_FLUSH);
4070 }
4071 }
4072
4073 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
4074 {
4075 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
4076 void *freelist = c->freelist;
4077 struct slab *slab = c->slab;
4078
4079 c->slab = NULL;
4080 c->freelist = NULL;
4081 c->tid = next_tid(c->tid);
4082
4083 if (slab) {
4084 deactivate_slab(s, slab, freelist);
4085 stat(s, CPUSLAB_FLUSH);
4086 }
4087
4088 put_partials_cpu(s, c);
4089 }
4090
4091 static inline void flush_this_cpu_slab(struct kmem_cache *s)
4092 {
4093 struct kmem_cache_cpu *c = this_cpu_ptr(s->cpu_slab);
4094
4095 if (c->slab)
4096 flush_slab(s, c);
4097
4098 put_partials(s);
4099 }
4100
4101 static bool has_cpu_slab(int cpu, struct kmem_cache *s)
4102 {
4103 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
4104
4105 return c->slab || slub_percpu_partial(c);
4106 }
4107
4108 static bool has_pcs_used(int cpu, struct kmem_cache *s)
4109 {
4110 struct slub_percpu_sheaves *pcs;
4111
4112 if (!s->cpu_sheaves)
4113 return false;
4114
4115 pcs = per_cpu_ptr(s->cpu_sheaves, cpu);
4116
4117 return (pcs->spare || pcs->rcu_free || pcs->main->size);
4118 }
4119
4120 /*
4121 * Flush cpu slab.
4122 *
4123 * Called from CPU work handler with migration disabled.
4124 */
4125 static void flush_cpu_slab(struct work_struct *w)
4126 {
4127 struct kmem_cache *s;
4128 struct slub_flush_work *sfw;
4129
4130 sfw = container_of(w, struct slub_flush_work, work);
4131
4132 s = sfw->s;
4133
4134 if (s->cpu_sheaves)
4135 pcs_flush_all(s);
4136
4137 flush_this_cpu_slab(s);
4138 }
4139
4140 static void flush_all_cpus_locked(struct kmem_cache *s)
4141 {
4142 struct slub_flush_work *sfw;
4143 unsigned int cpu;
4144
4145 lockdep_assert_cpus_held();
4146 mutex_lock(&flush_lock);
4147
4148 for_each_online_cpu(cpu) {
4149 sfw = &per_cpu(slub_flush, cpu);
4150 if (!has_cpu_slab(cpu, s) && !has_pcs_used(cpu, s)) {
4151 sfw->skip = true;
4152 continue;
4153 }
4154 INIT_WORK(&sfw->work, flush_cpu_slab);
4155 sfw->skip = false;
4156 sfw->s = s;
4157 queue_work_on(cpu, flushwq, &sfw->work);
4158 }
4159
4160 for_each_online_cpu(cpu) {
4161 sfw = &per_cpu(slub_flush, cpu);
4162 if (sfw->skip)
4163 continue;
4164 flush_work(&sfw->work);
4165 }
4166
4167 mutex_unlock(&flush_lock);
4168 }
4169
4170 static void flush_all(struct kmem_cache *s)
4171 {
4172 cpus_read_lock();
4173 flush_all_cpus_locked(s);
4174 cpus_read_unlock();
4175 }
4176
4177 static void flush_rcu_sheaf(struct work_struct *w)
4178 {
4179 struct slub_percpu_sheaves *pcs;
4180 struct slab_sheaf *rcu_free;
4181 struct slub_flush_work *sfw;
4182 struct kmem_cache *s;
4183
4184 sfw = container_of(w, struct slub_flush_work, work);
4185 s = sfw->s;
4186
4187 local_lock(&s->cpu_sheaves->lock);
4188 pcs = this_cpu_ptr(s->cpu_sheaves);
4189
4190 rcu_free = pcs->rcu_free;
4191 pcs->rcu_free = NULL;
4192
4193 local_unlock(&s->cpu_sheaves->lock);
4194
4195 if (rcu_free)
4196 call_rcu(&rcu_free->rcu_head, rcu_free_sheaf_nobarn);
4197 }
4198
4199
4200 /* needed for kvfree_rcu_barrier() */
4201 void flush_rcu_sheaves_on_cache(struct kmem_cache *s)
4202 {
4203 struct slub_flush_work *sfw;
4204 unsigned int cpu;
4205
4206 lockdep_assert_cpus_held();
4207 mutex_lock(&flush_lock);
4208
4209 for_each_online_cpu(cpu) {
4210 sfw = &per_cpu(slub_flush, cpu);
4211
4212 /*
4213 * we don't check if rcu_free sheaf exists - racing
4214 * __kfree_rcu_sheaf() might have just removed it.
4215 * by executing flush_rcu_sheaf() on the cpu we make
4216 * sure the __kfree_rcu_sheaf() finished its call_rcu()
4217 */
4218
4219 INIT_WORK(&sfw->work, flush_rcu_sheaf);
4220 sfw->s = s;
4221 queue_work_on(cpu, flushwq, &sfw->work);
4222 }
4223
4224 for_each_online_cpu(cpu) {
4225 sfw = &per_cpu(slub_flush, cpu);
4226 flush_work(&sfw->work);
4227 }
4228
4229 mutex_unlock(&flush_lock);
4230 }
4231
4232 void flush_all_rcu_sheaves(void)
4233 {
4234 struct kmem_cache *s;
4235
4236 cpus_read_lock();
4237 mutex_lock(&slab_mutex);
4238
4239 list_for_each_entry(s, &slab_caches, list) {
4240 if (!s->cpu_sheaves)
4241 continue;
4242 flush_rcu_sheaves_on_cache(s);
4243 }
4244
4245 mutex_unlock(&slab_mutex);
4246 cpus_read_unlock();
4247
4248 rcu_barrier();
4249 }
4250
4251 /*
4252 * Use the cpu notifier to insure that the cpu slabs are flushed when
4253 * necessary.
4254 */
4255 static int slub_cpu_dead(unsigned int cpu)
4256 {
4257 struct kmem_cache *s;
4258
4259 mutex_lock(&slab_mutex);
4260 list_for_each_entry(s, &slab_caches, list) {
4261 __flush_cpu_slab(s, cpu);
4262 if (s->cpu_sheaves)
4263 __pcs_flush_all_cpu(s, cpu);
4264 }
4265 mutex_unlock(&slab_mutex);
4266 return 0;
4267 }
4268
4269 /*
4270 * Check if the objects in a per cpu structure fit numa
4271 * locality expectations.
4272 */
4273 static inline int node_match(struct slab *slab, int node)
4274 {
4275 #ifdef CONFIG_NUMA
4276 if (node != NUMA_NO_NODE && slab_nid(slab) != node)
4277 return 0;
4278 #endif
4279 return 1;
4280 }
4281
4282 #ifdef CONFIG_SLUB_DEBUG
4283 static int count_free(struct slab *slab)
4284 {
4285 return slab->objects - slab->inuse;
4286 }
4287
4288 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
4289 {
4290 return atomic_long_read(&n->total_objects);
4291 }
4292
4293 /* Supports checking bulk free of a constructed freelist */
4294 static inline bool free_debug_processing(struct kmem_cache *s,
4295 struct slab *slab, void *head, void *tail, int *bulk_cnt,
4296 unsigned long addr, depot_stack_handle_t handle)
4297 {
4298 bool checks_ok = false;
4299 void *object = head;
4300 int cnt = 0;
4301
4302 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
4303 if (!check_slab(s, slab))
4304 goto out;
4305 }
4306
4307 if (slab->inuse < *bulk_cnt) {
4308 slab_err(s, slab, "Slab has %d allocated objects but %d are to be freed\n",
4309 slab->inuse, *bulk_cnt);
4310 goto out;
4311 }
4312
4313 next_object:
4314
4315 if (++cnt > *bulk_cnt)
4316 goto out_cnt;
4317
4318 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
4319 if (!free_consistency_checks(s, slab, object, addr))
4320 goto out;
4321 }
4322
4323 if (s->flags & SLAB_STORE_USER)
4324 set_track_update(s, object, TRACK_FREE, addr, handle);
4325 trace(s, slab, object, 0);
4326 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
4327 init_object(s, object, SLUB_RED_INACTIVE);
4328
4329 /* Reached end of constructed freelist yet? */
4330 if (object != tail) {
4331 object = get_freepointer(s, object);
4332 goto next_object;
4333 }
4334 checks_ok = true;
4335
4336 out_cnt:
4337 if (cnt != *bulk_cnt) {
4338 slab_err(s, slab, "Bulk free expected %d objects but found %d\n",
4339 *bulk_cnt, cnt);
4340 *bulk_cnt = cnt;
4341 }
4342
4343 out:
4344
4345 if (!checks_ok)
4346 slab_fix(s, "Object at 0x%p not freed", object);
4347
4348 return checks_ok;
4349 }
4350 #endif /* CONFIG_SLUB_DEBUG */
4351
4352 #if defined(CONFIG_SLUB_DEBUG) || defined(SLAB_SUPPORTS_SYSFS)
4353 static unsigned long count_partial(struct kmem_cache_node *n,
4354 int (*get_count)(struct slab *))
4355 {
4356 unsigned long flags;
4357 unsigned long x = 0;
4358 struct slab *slab;
4359
4360 spin_lock_irqsave(&n->list_lock, flags);
4361 list_for_each_entry(slab, &n->partial, slab_list)
4362 x += get_count(slab);
4363 spin_unlock_irqrestore(&n->list_lock, flags);
4364 return x;
4365 }
4366 #endif /* CONFIG_SLUB_DEBUG || SLAB_SUPPORTS_SYSFS */
4367
4368 #ifdef CONFIG_SLUB_DEBUG
4369 #define MAX_PARTIAL_TO_SCAN 10000
4370
4371 static unsigned long count_partial_free_approx(struct kmem_cache_node *n)
4372 {
4373 unsigned long flags;
4374 unsigned long x = 0;
4375 struct slab *slab;
4376
4377 spin_lock_irqsave(&n->list_lock, flags);
4378 if (n->nr_partial <= MAX_PARTIAL_TO_SCAN) {
4379 list_for_each_entry(slab, &n->partial, slab_list)
4380 x += slab->objects - slab->inuse;
4381 } else {
4382 /*
4383 * For a long list, approximate the total count of objects in
4384 * it to meet the limit on the number of slabs to scan.
4385 * Scan from both the list's head and tail for better accuracy.
4386 */
4387 unsigned long scanned = 0;
4388
4389 list_for_each_entry(slab, &n->partial, slab_list) {
4390 x += slab->objects - slab->inuse;
4391 if (++scanned == MAX_PARTIAL_TO_SCAN / 2)
4392 break;
4393 }
4394 list_for_each_entry_reverse(slab, &n->partial, slab_list) {
4395 x += slab->objects - slab->inuse;
4396 if (++scanned == MAX_PARTIAL_TO_SCAN)
4397 break;
4398 }
4399 x = mult_frac(x, n->nr_partial, scanned);
4400 x = min(x, node_nr_objs(n));
4401 }
4402 spin_unlock_irqrestore(&n->list_lock, flags);
4403 return x;
4404 }
4405
4406 static noinline void
4407 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
4408 {
4409 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
4410 DEFAULT_RATELIMIT_BURST);
4411 int cpu = raw_smp_processor_id();
4412 int node;
4413 struct kmem_cache_node *n;
4414
4415 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
4416 return;
4417
4418 pr_warn("SLUB: Unable to allocate memory on CPU %u (of node %d) on node %d, gfp=%#x(%pGg)\n",
4419 cpu, cpu_to_node(cpu), nid, gfpflags, &gfpflags);
4420 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
4421 s->name, s->object_size, s->size, oo_order(s->oo),
4422 oo_order(s->min));
4423
4424 if (oo_order(s->min) > get_order(s->object_size))
4425 pr_warn(" %s debugging increased min order, use slab_debug=O to disable.\n",
4426 s->name);
4427
4428 for_each_kmem_cache_node(s, node, n) {
4429 unsigned long nr_slabs;
4430 unsigned long nr_objs;
4431 unsigned long nr_free;
4432
4433 nr_free = count_partial_free_approx(n);
4434 nr_slabs = node_nr_slabs(n);
4435 nr_objs = node_nr_objs(n);
4436
4437 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
4438 node, nr_slabs, nr_objs, nr_free);
4439 }
4440 }
4441 #else /* CONFIG_SLUB_DEBUG */
4442 static inline void
4443 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid) { }
4444 #endif
4445
4446 static inline bool pfmemalloc_match(struct slab *slab, gfp_t gfpflags)
4447 {
4448 if (unlikely(slab_test_pfmemalloc(slab)))
4449 return gfp_pfmemalloc_allowed(gfpflags);
4450
4451 return true;
4452 }
4453
4454 static inline bool
4455 __update_cpu_freelist_fast(struct kmem_cache *s,
4456 void *freelist_old, void *freelist_new,
4457 unsigned long tid)
4458 {
4459 freelist_aba_t old = { .freelist = freelist_old, .counter = tid };
4460 freelist_aba_t new = { .freelist = freelist_new, .counter = next_tid(tid) };
4461
4462 return this_cpu_try_cmpxchg_freelist(s->cpu_slab->freelist_tid.full,
4463 &old.full, new.full);
4464 }
4465
4466 /*
4467 * Check the slab->freelist and either transfer the freelist to the
4468 * per cpu freelist or deactivate the slab.
4469 *
4470 * The slab is still frozen if the return value is not NULL.
4471 *
4472 * If this function returns NULL then the slab has been unfrozen.
4473 */
4474 static inline void *get_freelist(struct kmem_cache *s, struct slab *slab)
4475 {
4476 struct slab new;
4477 unsigned long counters;
4478 void *freelist;
4479
4480 lockdep_assert_held(this_cpu_ptr(&s->cpu_slab->lock));
4481
4482 do {
4483 freelist = slab->freelist;
4484 counters = slab->counters;
4485
4486 new.counters = counters;
4487
4488 new.inuse = slab->objects;
4489 new.frozen = freelist != NULL;
4490
4491 } while (!__slab_update_freelist(s, slab,
4492 freelist, counters,
4493 NULL, new.counters,
4494 "get_freelist"));
4495
4496 return freelist;
4497 }
4498
4499 /*
4500 * Freeze the partial slab and return the pointer to the freelist.
4501 */
4502 static inline void *freeze_slab(struct kmem_cache *s, struct slab *slab)
4503 {
4504 struct slab new;
4505 unsigned long counters;
4506 void *freelist;
4507
4508 do {
4509 freelist = slab->freelist;
4510 counters = slab->counters;
4511
4512 new.counters = counters;
4513 VM_BUG_ON(new.frozen);
4514
4515 new.inuse = slab->objects;
4516 new.frozen = 1;
4517
4518 } while (!slab_update_freelist(s, slab,
4519 freelist, counters,
4520 NULL, new.counters,
4521 "freeze_slab"));
4522
4523 return freelist;
4524 }
4525
4526 /*
4527 * Slow path. The lockless freelist is empty or we need to perform
4528 * debugging duties.
4529 *
4530 * Processing is still very fast if new objects have been freed to the
4531 * regular freelist. In that case we simply take over the regular freelist
4532 * as the lockless freelist and zap the regular freelist.
4533 *
4534 * If that is not working then we fall back to the partial lists. We take the
4535 * first element of the freelist as the object to allocate now and move the
4536 * rest of the freelist to the lockless freelist.
4537 *
4538 * And if we were unable to get a new slab from the partial slab lists then
4539 * we need to allocate a new slab. This is the slowest path since it involves
4540 * a call to the page allocator and the setup of a new slab.
4541 *
4542 * Version of __slab_alloc to use when we know that preemption is
4543 * already disabled (which is the case for bulk allocation).
4544 */
4545 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
4546 unsigned long addr, struct kmem_cache_cpu *c, unsigned int orig_size)
4547 {
4548 bool allow_spin = gfpflags_allow_spinning(gfpflags);
4549 void *freelist;
4550 struct slab *slab;
4551 unsigned long flags;
4552 struct partial_context pc;
4553 bool try_thisnode = true;
4554
4555 stat(s, ALLOC_SLOWPATH);
4556
4557 reread_slab:
4558
4559 slab = READ_ONCE(c->slab);
4560 if (!slab) {
4561 /*
4562 * if the node is not online or has no normal memory, just
4563 * ignore the node constraint
4564 */
4565 if (unlikely(node != NUMA_NO_NODE &&
4566 !node_isset(node, slab_nodes)))
4567 node = NUMA_NO_NODE;
4568 goto new_slab;
4569 }
4570
4571 if (unlikely(!node_match(slab, node))) {
4572 /*
4573 * same as above but node_match() being false already
4574 * implies node != NUMA_NO_NODE.
4575 *
4576 * We don't strictly honor pfmemalloc and NUMA preferences
4577 * when !allow_spin because:
4578 *
4579 * 1. Most kmalloc() users allocate objects on the local node,
4580 * so kmalloc_nolock() tries not to interfere with them by
4581 * deactivating the cpu slab.
4582 *
4583 * 2. Deactivating due to NUMA or pfmemalloc mismatch may cause
4584 * unnecessary slab allocations even when n->partial list
4585 * is not empty.
4586 */
4587 if (!node_isset(node, slab_nodes) ||
4588 !allow_spin) {
4589 node = NUMA_NO_NODE;
4590 } else {
4591 stat(s, ALLOC_NODE_MISMATCH);
4592 goto deactivate_slab;
4593 }
4594 }
4595
4596 /*
4597 * By rights, we should be searching for a slab page that was
4598 * PFMEMALLOC but right now, we are losing the pfmemalloc
4599 * information when the page leaves the per-cpu allocator
4600 */
4601 if (unlikely(!pfmemalloc_match(slab, gfpflags) && allow_spin))
4602 goto deactivate_slab;
4603
4604 /* must check again c->slab in case we got preempted and it changed */
4605 local_lock_cpu_slab(s, flags);
4606
4607 if (unlikely(slab != c->slab)) {
4608 local_unlock_cpu_slab(s, flags);
4609 goto reread_slab;
4610 }
4611 freelist = c->freelist;
4612 if (freelist)
4613 goto load_freelist;
4614
4615 freelist = get_freelist(s, slab);
4616
4617 if (!freelist) {
4618 c->slab = NULL;
4619 c->tid = next_tid(c->tid);
4620 local_unlock_cpu_slab(s, flags);
4621 stat(s, DEACTIVATE_BYPASS);
4622 goto new_slab;
4623 }
4624
4625 stat(s, ALLOC_REFILL);
4626
4627 load_freelist:
4628
4629 lockdep_assert_held(this_cpu_ptr(&s->cpu_slab->lock));
4630
4631 /*
4632 * freelist is pointing to the list of objects to be used.
4633 * slab is pointing to the slab from which the objects are obtained.
4634 * That slab must be frozen for per cpu allocations to work.
4635 */
4636 VM_BUG_ON(!c->slab->frozen);
4637 c->freelist = get_freepointer(s, freelist);
4638 c->tid = next_tid(c->tid);
4639 local_unlock_cpu_slab(s, flags);
4640 return freelist;
4641
4642 deactivate_slab:
4643
4644 local_lock_cpu_slab(s, flags);
4645 if (slab != c->slab) {
4646 local_unlock_cpu_slab(s, flags);
4647 goto reread_slab;
4648 }
4649 freelist = c->freelist;
4650 c->slab = NULL;
4651 c->freelist = NULL;
4652 c->tid = next_tid(c->tid);
4653 local_unlock_cpu_slab(s, flags);
4654 deactivate_slab(s, slab, freelist);
4655
4656 new_slab:
4657
4658 #ifdef CONFIG_SLUB_CPU_PARTIAL
4659 while (slub_percpu_partial(c)) {
4660 local_lock_cpu_slab(s, flags);
4661 if (unlikely(c->slab)) {
4662 local_unlock_cpu_slab(s, flags);
4663 goto reread_slab;
4664 }
4665 if (unlikely(!slub_percpu_partial(c))) {
4666 local_unlock_cpu_slab(s, flags);
4667 /* we were preempted and partial list got empty */
4668 goto new_objects;
4669 }
4670
4671 slab = slub_percpu_partial(c);
4672 slub_set_percpu_partial(c, slab);
4673
4674 if (likely(node_match(slab, node) &&
4675 pfmemalloc_match(slab, gfpflags)) ||
4676 !allow_spin) {
4677 c->slab = slab;
4678 freelist = get_freelist(s, slab);
4679 VM_BUG_ON(!freelist);
4680 stat(s, CPU_PARTIAL_ALLOC);
4681 goto load_freelist;
4682 }
4683
4684 local_unlock_cpu_slab(s, flags);
4685
4686 slab->next = NULL;
4687 __put_partials(s, slab);
4688 }
4689 #endif
4690
4691 new_objects:
4692
4693 pc.flags = gfpflags;
4694 /*
4695 * When a preferred node is indicated but no __GFP_THISNODE
4696 *
4697 * 1) try to get a partial slab from target node only by having
4698 * __GFP_THISNODE in pc.flags for get_partial()
4699 * 2) if 1) failed, try to allocate a new slab from target node with
4700 * GPF_NOWAIT | __GFP_THISNODE opportunistically
4701 * 3) if 2) failed, retry with original gfpflags which will allow
4702 * get_partial() try partial lists of other nodes before potentially
4703 * allocating new page from other nodes
4704 */
4705 if (unlikely(node != NUMA_NO_NODE && !(gfpflags & __GFP_THISNODE)
4706 && try_thisnode)) {
4707 if (unlikely(!allow_spin))
4708 /* Do not upgrade gfp to NOWAIT from more restrictive mode */
4709 pc.flags = gfpflags | __GFP_THISNODE;
4710 else
4711 pc.flags = GFP_NOWAIT | __GFP_THISNODE;
4712 }
4713
4714 pc.orig_size = orig_size;
4715 slab = get_partial(s, node, &pc);
4716 if (slab) {
4717 if (IS_ENABLED(CONFIG_SLUB_TINY) || kmem_cache_debug(s)) {
4718 freelist = pc.object;
4719 /*
4720 * For debug caches here we had to go through
4721 * alloc_single_from_partial() so just store the
4722 * tracking info and return the object.
4723 *
4724 * Due to disabled preemption we need to disallow
4725 * blocking. The flags are further adjusted by
4726 * gfp_nested_mask() in stack_depot itself.
4727 */
4728 if (s->flags & SLAB_STORE_USER)
4729 set_track(s, freelist, TRACK_ALLOC, addr,
4730 gfpflags & ~(__GFP_DIRECT_RECLAIM));
4731
4732 return freelist;
4733 }
4734
4735 freelist = freeze_slab(s, slab);
4736 goto retry_load_slab;
4737 }
4738
4739 slub_put_cpu_ptr(s->cpu_slab);
4740 slab = new_slab(s, pc.flags, node);
4741 c = slub_get_cpu_ptr(s->cpu_slab);
4742
4743 if (unlikely(!slab)) {
4744 if (node != NUMA_NO_NODE && !(gfpflags & __GFP_THISNODE)
4745 && try_thisnode) {
4746 try_thisnode = false;
4747 goto new_objects;
4748 }
4749 slab_out_of_memory(s, gfpflags, node);
4750 return NULL;
4751 }
4752
4753 stat(s, ALLOC_SLAB);
4754
4755 if (IS_ENABLED(CONFIG_SLUB_TINY) || kmem_cache_debug(s)) {
4756 freelist = alloc_single_from_new_slab(s, slab, orig_size, gfpflags);
4757
4758 if (unlikely(!freelist)) {
4759 /* This could cause an endless loop. Fail instead. */
4760 if (!allow_spin)
4761 return NULL;
4762 goto new_objects;
4763 }
4764
4765 if (s->flags & SLAB_STORE_USER)
4766 set_track(s, freelist, TRACK_ALLOC, addr,
4767 gfpflags & ~(__GFP_DIRECT_RECLAIM));
4768
4769 return freelist;
4770 }
4771
4772 /*
4773 * No other reference to the slab yet so we can
4774 * muck around with it freely without cmpxchg
4775 */
4776 freelist = slab->freelist;
4777 slab->freelist = NULL;
4778 slab->inuse = slab->objects;
4779 slab->frozen = 1;
4780
4781 inc_slabs_node(s, slab_nid(slab), slab->objects);
4782
4783 if (unlikely(!pfmemalloc_match(slab, gfpflags) && allow_spin)) {
4784 /*
4785 * For !pfmemalloc_match() case we don't load freelist so that
4786 * we don't make further mismatched allocations easier.
4787 */
4788 deactivate_slab(s, slab, get_freepointer(s, freelist));
4789 return freelist;
4790 }
4791
4792 retry_load_slab:
4793
4794 local_lock_cpu_slab(s, flags);
4795 if (unlikely(c->slab)) {
4796 void *flush_freelist = c->freelist;
4797 struct slab *flush_slab = c->slab;
4798
4799 c->slab = NULL;
4800 c->freelist = NULL;
4801 c->tid = next_tid(c->tid);
4802
4803 local_unlock_cpu_slab(s, flags);
4804
4805 if (unlikely(!allow_spin)) {
4806 /* Reentrant slub cannot take locks, defer */
4807 defer_deactivate_slab(flush_slab, flush_freelist);
4808 } else {
4809 deactivate_slab(s, flush_slab, flush_freelist);
4810 }
4811
4812 stat(s, CPUSLAB_FLUSH);
4813
4814 goto retry_load_slab;
4815 }
4816 c->slab = slab;
4817
4818 goto load_freelist;
4819 }
4820 /*
4821 * We disallow kprobes in ___slab_alloc() to prevent reentrance
4822 *
4823 * kmalloc() -> ___slab_alloc() -> local_lock_cpu_slab() protected part of
4824 * ___slab_alloc() manipulating c->freelist -> kprobe -> bpf ->
4825 * kmalloc_nolock() or kfree_nolock() -> __update_cpu_freelist_fast()
4826 * manipulating c->freelist without lock.
4827 *
4828 * This does not prevent kprobe in functions called from ___slab_alloc() such as
4829 * local_lock_irqsave() itself, and that is fine, we only need to protect the
4830 * c->freelist manipulation in ___slab_alloc() itself.
4831 */
4832 NOKPROBE_SYMBOL(___slab_alloc);
4833
4834 /*
4835 * A wrapper for ___slab_alloc() for contexts where preemption is not yet
4836 * disabled. Compensates for possible cpu changes by refetching the per cpu area
4837 * pointer.
4838 */
4839 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
4840 unsigned long addr, struct kmem_cache_cpu *c, unsigned int orig_size)
4841 {
4842 void *p;
4843
4844 #ifdef CONFIG_PREEMPT_COUNT
4845 /*
4846 * We may have been preempted and rescheduled on a different
4847 * cpu before disabling preemption. Need to reload cpu area
4848 * pointer.
4849 */
4850 c = slub_get_cpu_ptr(s->cpu_slab);
4851 #endif
4852 if (unlikely(!gfpflags_allow_spinning(gfpflags))) {
4853 if (local_lock_is_locked(&s->cpu_slab->lock)) {
4854 /*
4855 * EBUSY is an internal signal to kmalloc_nolock() to
4856 * retry a different bucket. It's not propagated
4857 * to the caller.
4858 */
4859 p = ERR_PTR(-EBUSY);
4860 goto out;
4861 }
4862 }
4863 p = ___slab_alloc(s, gfpflags, node, addr, c, orig_size);
4864 out:
4865 #ifdef CONFIG_PREEMPT_COUNT
4866 slub_put_cpu_ptr(s->cpu_slab);
4867 #endif
4868 return p;
4869 }
4870
4871 static __always_inline void *__slab_alloc_node(struct kmem_cache *s,
4872 gfp_t gfpflags, int node, unsigned long addr, size_t orig_size)
4873 {
4874 struct kmem_cache_cpu *c;
4875 struct slab *slab;
4876 unsigned long tid;
4877 void *object;
4878
4879 redo:
4880 /*
4881 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
4882 * enabled. We may switch back and forth between cpus while
4883 * reading from one cpu area. That does not matter as long
4884 * as we end up on the original cpu again when doing the cmpxchg.
4885 *
4886 * We must guarantee that tid and kmem_cache_cpu are retrieved on the
4887 * same cpu. We read first the kmem_cache_cpu pointer and use it to read
4888 * the tid. If we are preempted and switched to another cpu between the
4889 * two reads, it's OK as the two are still associated with the same cpu
4890 * and cmpxchg later will validate the cpu.
4891 */
4892 c = raw_cpu_ptr(s->cpu_slab);
4893 tid = READ_ONCE(c->tid);
4894
4895 /*
4896 * Irqless object alloc/free algorithm used here depends on sequence
4897 * of fetching cpu_slab's data. tid should be fetched before anything
4898 * on c to guarantee that object and slab associated with previous tid
4899 * won't be used with current tid. If we fetch tid first, object and
4900 * slab could be one associated with next tid and our alloc/free
4901 * request will be failed. In this case, we will retry. So, no problem.
4902 */
4903 barrier();
4904
4905 /*
4906 * The transaction ids are globally unique per cpu and per operation on
4907 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
4908 * occurs on the right processor and that there was no operation on the
4909 * linked list in between.
4910 */
4911
4912 object = c->freelist;
4913 slab = c->slab;
4914
4915 #ifdef CONFIG_NUMA
4916 if (static_branch_unlikely(&strict_numa) &&
4917 node == NUMA_NO_NODE) {
4918
4919 struct mempolicy *mpol = current->mempolicy;
4920
4921 if (mpol) {
4922 /*
4923 * Special BIND rule support. If existing slab
4924 * is in permitted set then do not redirect
4925 * to a particular node.
4926 * Otherwise we apply the memory policy to get
4927 * the node we need to allocate on.
4928 */
4929 if (mpol->mode != MPOL_BIND || !slab ||
4930 !node_isset(slab_nid(slab), mpol->nodes))
4931
4932 node = mempolicy_slab_node();
4933 }
4934 }
4935 #endif
4936
4937 if (!USE_LOCKLESS_FAST_PATH() ||
4938 unlikely(!object || !slab || !node_match(slab, node))) {
4939 object = __slab_alloc(s, gfpflags, node, addr, c, orig_size);
4940 } else {
4941 void *next_object = get_freepointer_safe(s, object);
4942
4943 /*
4944 * The cmpxchg will only match if there was no additional
4945 * operation and if we are on the right processor.
4946 *
4947 * The cmpxchg does the following atomically (without lock
4948 * semantics!)
4949 * 1. Relocate first pointer to the current per cpu area.
4950 * 2. Verify that tid and freelist have not been changed
4951 * 3. If they were not changed replace tid and freelist
4952 *
4953 * Since this is without lock semantics the protection is only
4954 * against code executing on this cpu *not* from access by
4955 * other cpus.
4956 */
4957 if (unlikely(!__update_cpu_freelist_fast(s, object, next_object, tid))) {
4958 note_cmpxchg_failure("slab_alloc", s, tid);
4959 goto redo;
4960 }
4961 prefetch_freepointer(s, next_object);
4962 stat(s, ALLOC_FASTPATH);
4963 }
4964
4965 return object;
4966 }
4967
4968 /*
4969 * If the object has been wiped upon free, make sure it's fully initialized by
4970 * zeroing out freelist pointer.
4971 *
4972 * Note that we also wipe custom freelist pointers.
4973 */
4974 static __always_inline void maybe_wipe_obj_freeptr(struct kmem_cache *s,
4975 void *obj)
4976 {
4977 if (unlikely(slab_want_init_on_free(s)) && obj &&
4978 !freeptr_outside_object(s))
4979 memset((void *)((char *)kasan_reset_tag(obj) + s->offset),
4980 0, sizeof(void *));
4981 }
4982
4983 static __fastpath_inline
4984 struct kmem_cache *slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
4985 {
4986 flags &= gfp_allowed_mask;
4987
4988 might_alloc(flags);
4989
4990 if (unlikely(should_failslab(s, flags)))
4991 return NULL;
4992
4993 return s;
4994 }
4995
4996 static __fastpath_inline
4997 bool slab_post_alloc_hook(struct kmem_cache *s, struct list_lru *lru,
4998 gfp_t flags, size_t size, void **p, bool init,
4999 unsigned int orig_size)
5000 {
5001 unsigned int zero_size = s->object_size;
5002 bool kasan_init = init;
5003 size_t i;
5004 gfp_t init_flags = flags & gfp_allowed_mask;
5005
5006 /*
5007 * For kmalloc object, the allocated memory size(object_size) is likely
5008 * larger than the requested size(orig_size). If redzone check is
5009 * enabled for the extra space, don't zero it, as it will be redzoned
5010 * soon. The redzone operation for this extra space could be seen as a
5011 * replacement of current poisoning under certain debug option, and
5012 * won't break other sanity checks.
5013 */
5014 if (kmem_cache_debug_flags(s, SLAB_STORE_USER | SLAB_RED_ZONE) &&
5015 (s->flags & SLAB_KMALLOC))
5016 zero_size = orig_size;
5017
5018 /*
5019 * When slab_debug is enabled, avoid memory initialization integrated
5020 * into KASAN and instead zero out the memory via the memset below with
5021 * the proper size. Otherwise, KASAN might overwrite SLUB redzones and
5022 * cause false-positive reports. This does not lead to a performance
5023 * penalty on production builds, as slab_debug is not intended to be
5024 * enabled there.
5025 */
5026 if (__slub_debug_enabled())
5027 kasan_init = false;
5028
5029 /*
5030 * As memory initialization might be integrated into KASAN,
5031 * kasan_slab_alloc and initialization memset must be
5032 * kept together to avoid discrepancies in behavior.
5033 *
5034 * As p[i] might get tagged, memset and kmemleak hook come after KASAN.
5035 */
5036 for (i = 0; i < size; i++) {
5037 p[i] = kasan_slab_alloc(s, p[i], init_flags, kasan_init);
5038 if (p[i] && init && (!kasan_init ||
5039 !kasan_has_integrated_init()))
5040 memset(p[i], 0, zero_size);
5041 if (gfpflags_allow_spinning(flags))
5042 kmemleak_alloc_recursive(p[i], s->object_size, 1,
5043 s->flags, init_flags);
5044 kmsan_slab_alloc(s, p[i], init_flags);
5045 alloc_tagging_slab_alloc_hook(s, p[i], flags);
5046 }
5047
5048 return memcg_slab_post_alloc_hook(s, lru, flags, size, p);
5049 }
5050
5051 /*
5052 * Replace the empty main sheaf with a (at least partially) full sheaf.
5053 *
5054 * Must be called with the cpu_sheaves local lock locked. If successful, returns
5055 * the pcs pointer and the local lock locked (possibly on a different cpu than
5056 * initially called). If not successful, returns NULL and the local lock
5057 * unlocked.
5058 */
5059 static struct slub_percpu_sheaves *
5060 __pcs_replace_empty_main(struct kmem_cache *s, struct slub_percpu_sheaves *pcs, gfp_t gfp)
5061 {
5062 struct slab_sheaf *empty = NULL;
5063 struct slab_sheaf *full;
5064 struct node_barn *barn;
5065 bool can_alloc;
5066
5067 lockdep_assert_held(this_cpu_ptr(&s->cpu_sheaves->lock));
5068
5069 if (pcs->spare && pcs->spare->size > 0) {
5070 swap(pcs->main, pcs->spare);
5071 return pcs;
5072 }
5073
5074 barn = get_barn(s);
5075 if (!barn) {
5076 local_unlock(&s->cpu_sheaves->lock);
5077 return NULL;
5078 }
5079
5080 full = barn_replace_empty_sheaf(barn, pcs->main);
5081
5082 if (full) {
5083 stat(s, BARN_GET);
5084 pcs->main = full;
5085 return pcs;
5086 }
5087
5088 stat(s, BARN_GET_FAIL);
5089
5090 can_alloc = gfpflags_allow_blocking(gfp);
5091
5092 if (can_alloc) {
5093 if (pcs->spare) {
5094 empty = pcs->spare;
5095 pcs->spare = NULL;
5096 } else {
5097 empty = barn_get_empty_sheaf(barn);
5098 }
5099 }
5100
5101 local_unlock(&s->cpu_sheaves->lock);
5102
5103 if (!can_alloc)
5104 return NULL;
5105
5106 if (empty) {
5107 if (!refill_sheaf(s, empty, gfp)) {
5108 full = empty;
5109 } else {
5110 /*
5111 * we must be very low on memory so don't bother
5112 * with the barn
5113 */
5114 free_empty_sheaf(s, empty);
5115 }
5116 } else {
5117 full = alloc_full_sheaf(s, gfp);
5118 }
5119
5120 if (!full)
5121 return NULL;
5122
5123 /*
5124 * we can reach here only when gfpflags_allow_blocking
5125 * so this must not be an irq
5126 */
5127 local_lock(&s->cpu_sheaves->lock);
5128 pcs = this_cpu_ptr(s->cpu_sheaves);
5129
5130 /*
5131 * If we are returning empty sheaf, we either got it from the
5132 * barn or had to allocate one. If we are returning a full
5133 * sheaf, it's due to racing or being migrated to a different
5134 * cpu. Breaching the barn's sheaf limits should be thus rare
5135 * enough so just ignore them to simplify the recovery.
5136 */
5137
5138 if (pcs->main->size == 0) {
5139 barn_put_empty_sheaf(barn, pcs->main);
5140 pcs->main = full;
5141 return pcs;
5142 }
5143
5144 if (!pcs->spare) {
5145 pcs->spare = full;
5146 return pcs;
5147 }
5148
5149 if (pcs->spare->size == 0) {
5150 barn_put_empty_sheaf(barn, pcs->spare);
5151 pcs->spare = full;
5152 return pcs;
5153 }
5154
5155 barn_put_full_sheaf(barn, full);
5156 stat(s, BARN_PUT);
5157
5158 return pcs;
5159 }
5160
5161 static __fastpath_inline
5162 void *alloc_from_pcs(struct kmem_cache *s, gfp_t gfp, int node)
5163 {
5164 struct slub_percpu_sheaves *pcs;
5165 bool node_requested;
5166 void *object;
5167
5168 #ifdef CONFIG_NUMA
5169 if (static_branch_unlikely(&strict_numa) &&
5170 node == NUMA_NO_NODE) {
5171
5172 struct mempolicy *mpol = current->mempolicy;
5173
5174 if (mpol) {
5175 /*
5176 * Special BIND rule support. If the local node
5177 * is in permitted set then do not redirect
5178 * to a particular node.
5179 * Otherwise we apply the memory policy to get
5180 * the node we need to allocate on.
5181 */
5182 if (mpol->mode != MPOL_BIND ||
5183 !node_isset(numa_mem_id(), mpol->nodes))
5184
5185 node = mempolicy_slab_node();
5186 }
5187 }
5188 #endif
5189
5190 node_requested = IS_ENABLED(CONFIG_NUMA) && node != NUMA_NO_NODE;
5191
5192 /*
5193 * We assume the percpu sheaves contain only local objects although it's
5194 * not completely guaranteed, so we verify later.
5195 */
5196 if (unlikely(node_requested && node != numa_mem_id()))
5197 return NULL;
5198
5199 if (!local_trylock(&s->cpu_sheaves->lock))
5200 return NULL;
5201
5202 pcs = this_cpu_ptr(s->cpu_sheaves);
5203
5204 if (unlikely(pcs->main->size == 0)) {
5205 pcs = __pcs_replace_empty_main(s, pcs, gfp);
5206 if (unlikely(!pcs))
5207 return NULL;
5208 }
5209
5210 object = pcs->main->objects[pcs->main->size - 1];
5211
5212 if (unlikely(node_requested)) {
5213 /*
5214 * Verify that the object was from the node we want. This could
5215 * be false because of cpu migration during an unlocked part of
5216 * the current allocation or previous freeing process.
5217 */
5218 if (folio_nid(virt_to_folio(object)) != node) {
5219 local_unlock(&s->cpu_sheaves->lock);
5220 return NULL;
5221 }
5222 }
5223
5224 pcs->main->size--;
5225
5226 local_unlock(&s->cpu_sheaves->lock);
5227
5228 stat(s, ALLOC_PCS);
5229
5230 return object;
5231 }
5232
5233 static __fastpath_inline
5234 unsigned int alloc_from_pcs_bulk(struct kmem_cache *s, size_t size, void **p)
5235 {
5236 struct slub_percpu_sheaves *pcs;
5237 struct slab_sheaf *main;
5238 unsigned int allocated = 0;
5239 unsigned int batch;
5240
5241 next_batch:
5242 if (!local_trylock(&s->cpu_sheaves->lock))
5243 return allocated;
5244
5245 pcs = this_cpu_ptr(s->cpu_sheaves);
5246
5247 if (unlikely(pcs->main->size == 0)) {
5248
5249 struct slab_sheaf *full;
5250 struct node_barn *barn;
5251
5252 if (pcs->spare && pcs->spare->size > 0) {
5253 swap(pcs->main, pcs->spare);
5254 goto do_alloc;
5255 }
5256
5257 barn = get_barn(s);
5258 if (!barn) {
5259 local_unlock(&s->cpu_sheaves->lock);
5260 return allocated;
5261 }
5262
5263 full = barn_replace_empty_sheaf(barn, pcs->main);
5264
5265 if (full) {
5266 stat(s, BARN_GET);
5267 pcs->main = full;
5268 goto do_alloc;
5269 }
5270
5271 stat(s, BARN_GET_FAIL);
5272
5273 local_unlock(&s->cpu_sheaves->lock);
5274
5275 /*
5276 * Once full sheaves in barn are depleted, let the bulk
5277 * allocation continue from slab pages, otherwise we would just
5278 * be copying arrays of pointers twice.
5279 */
5280 return allocated;
5281 }
5282
5283 do_alloc:
5284
5285 main = pcs->main;
5286 batch = min(size, main->size);
5287
5288 main->size -= batch;
5289 memcpy(p, main->objects + main->size, batch * sizeof(void *));
5290
5291 local_unlock(&s->cpu_sheaves->lock);
5292
5293 stat_add(s, ALLOC_PCS, batch);
5294
5295 allocated += batch;
5296
5297 if (batch < size) {
5298 p += batch;
5299 size -= batch;
5300 goto next_batch;
5301 }
5302
5303 return allocated;
5304 }
5305
5306
5307 /*
5308 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
5309 * have the fastpath folded into their functions. So no function call
5310 * overhead for requests that can be satisfied on the fastpath.
5311 *
5312 * The fastpath works by first checking if the lockless freelist can be used.
5313 * If not then __slab_alloc is called for slow processing.
5314 *
5315 * Otherwise we can simply pick the next object from the lockless free list.
5316 */
5317 static __fastpath_inline void *slab_alloc_node(struct kmem_cache *s, struct list_lru *lru,
5318 gfp_t gfpflags, int node, unsigned long addr, size_t orig_size)
5319 {
5320 void *object;
5321 bool init = false;
5322
5323 s = slab_pre_alloc_hook(s, gfpflags);
5324 if (unlikely(!s))
5325 return NULL;
5326
5327 object = kfence_alloc(s, orig_size, gfpflags);
5328 if (unlikely(object))
5329 goto out;
5330
5331 if (s->cpu_sheaves)
5332 object = alloc_from_pcs(s, gfpflags, node);
5333
5334 if (!object)
5335 object = __slab_alloc_node(s, gfpflags, node, addr, orig_size);
5336
5337 maybe_wipe_obj_freeptr(s, object);
5338 init = slab_want_init_on_alloc(gfpflags, s);
5339
5340 out:
5341 /*
5342 * When init equals 'true', like for kzalloc() family, only
5343 * @orig_size bytes might be zeroed instead of s->object_size
5344 * In case this fails due to memcg_slab_post_alloc_hook(),
5345 * object is set to NULL
5346 */
5347 slab_post_alloc_hook(s, lru, gfpflags, 1, &object, init, orig_size);
5348
5349 return object;
5350 }
5351
5352 void *kmem_cache_alloc_noprof(struct kmem_cache *s, gfp_t gfpflags)
5353 {
5354 void *ret = slab_alloc_node(s, NULL, gfpflags, NUMA_NO_NODE, _RET_IP_,
5355 s->object_size);
5356
5357 trace_kmem_cache_alloc(_RET_IP_, ret, s, gfpflags, NUMA_NO_NODE);
5358
5359 return ret;
5360 }
5361 EXPORT_SYMBOL(kmem_cache_alloc_noprof);
5362
5363 void *kmem_cache_alloc_lru_noprof(struct kmem_cache *s, struct list_lru *lru,
5364 gfp_t gfpflags)
5365 {
5366 void *ret = slab_alloc_node(s, lru, gfpflags, NUMA_NO_NODE, _RET_IP_,
5367 s->object_size);
5368
5369 trace_kmem_cache_alloc(_RET_IP_, ret, s, gfpflags, NUMA_NO_NODE);
5370
5371 return ret;
5372 }
5373 EXPORT_SYMBOL(kmem_cache_alloc_lru_noprof);
5374
5375 bool kmem_cache_charge(void *objp, gfp_t gfpflags)
5376 {
5377 if (!memcg_kmem_online())
5378 return true;
5379
5380 return memcg_slab_post_charge(objp, gfpflags);
5381 }
5382 EXPORT_SYMBOL(kmem_cache_charge);
5383
5384 /**
5385 * kmem_cache_alloc_node - Allocate an object on the specified node
5386 * @s: The cache to allocate from.
5387 * @gfpflags: See kmalloc().
5388 * @node: node number of the target node.
5389 *
5390 * Identical to kmem_cache_alloc but it will allocate memory on the given
5391 * node, which can improve the performance for cpu bound structures.
5392 *
5393 * Fallback to other node is possible if __GFP_THISNODE is not set.
5394 *
5395 * Return: pointer to the new object or %NULL in case of error
5396 */
5397 void *kmem_cache_alloc_node_noprof(struct kmem_cache *s, gfp_t gfpflags, int node)
5398 {
5399 void *ret = slab_alloc_node(s, NULL, gfpflags, node, _RET_IP_, s->object_size);
5400
5401 trace_kmem_cache_alloc(_RET_IP_, ret, s, gfpflags, node);
5402
5403 return ret;
5404 }
5405 EXPORT_SYMBOL(kmem_cache_alloc_node_noprof);
5406
5407 /*
5408 * returns a sheaf that has at least the requested size
5409 * when prefilling is needed, do so with given gfp flags
5410 *
5411 * return NULL if sheaf allocation or prefilling failed
5412 */
5413 struct slab_sheaf *
5414 kmem_cache_prefill_sheaf(struct kmem_cache *s, gfp_t gfp, unsigned int size)
5415 {
5416 struct slub_percpu_sheaves *pcs;
5417 struct slab_sheaf *sheaf = NULL;
5418 struct node_barn *barn;
5419
5420 if (unlikely(size > s->sheaf_capacity)) {
5421
5422 /*
5423 * slab_debug disables cpu sheaves intentionally so all
5424 * prefilled sheaves become "oversize" and we give up on
5425 * performance for the debugging. Same with SLUB_TINY.
5426 * Creating a cache without sheaves and then requesting a
5427 * prefilled sheaf is however not expected, so warn.
5428 */
5429 WARN_ON_ONCE(s->sheaf_capacity == 0 &&
5430 !IS_ENABLED(CONFIG_SLUB_TINY) &&
5431 !(s->flags & SLAB_DEBUG_FLAGS));
5432
5433 sheaf = kzalloc(struct_size(sheaf, objects, size), gfp);
5434 if (!sheaf)
5435 return NULL;
5436
5437 stat(s, SHEAF_PREFILL_OVERSIZE);
5438 sheaf->cache = s;
5439 sheaf->capacity = size;
5440
5441 if (!__kmem_cache_alloc_bulk(s, gfp, size,
5442 &sheaf->objects[0])) {
5443 kfree(sheaf);
5444 return NULL;
5445 }
5446
5447 sheaf->size = size;
5448
5449 return sheaf;
5450 }
5451
5452 local_lock(&s->cpu_sheaves->lock);
5453 pcs = this_cpu_ptr(s->cpu_sheaves);
5454
5455 if (pcs->spare) {
5456 sheaf = pcs->spare;
5457 pcs->spare = NULL;
5458 stat(s, SHEAF_PREFILL_FAST);
5459 } else {
5460 barn = get_barn(s);
5461
5462 stat(s, SHEAF_PREFILL_SLOW);
5463 if (barn)
5464 sheaf = barn_get_full_or_empty_sheaf(barn);
5465 if (sheaf && sheaf->size)
5466 stat(s, BARN_GET);
5467 else
5468 stat(s, BARN_GET_FAIL);
5469 }
5470
5471 local_unlock(&s->cpu_sheaves->lock);
5472
5473
5474 if (!sheaf)
5475 sheaf = alloc_empty_sheaf(s, gfp);
5476
5477 if (sheaf && sheaf->size < size) {
5478 if (refill_sheaf(s, sheaf, gfp)) {
5479 sheaf_flush_unused(s, sheaf);
5480 free_empty_sheaf(s, sheaf);
5481 sheaf = NULL;
5482 }
5483 }
5484
5485 if (sheaf)
5486 sheaf->capacity = s->sheaf_capacity;
5487
5488 return sheaf;
5489 }
5490
5491 /*
5492 * Use this to return a sheaf obtained by kmem_cache_prefill_sheaf()
5493 *
5494 * If the sheaf cannot simply become the percpu spare sheaf, but there's space
5495 * for a full sheaf in the barn, we try to refill the sheaf back to the cache's
5496 * sheaf_capacity to avoid handling partially full sheaves.
5497 *
5498 * If the refill fails because gfp is e.g. GFP_NOWAIT, or the barn is full, the
5499 * sheaf is instead flushed and freed.
5500 */
5501 void kmem_cache_return_sheaf(struct kmem_cache *s, gfp_t gfp,
5502 struct slab_sheaf *sheaf)
5503 {
5504 struct slub_percpu_sheaves *pcs;
5505 struct node_barn *barn;
5506
5507 if (unlikely(sheaf->capacity != s->sheaf_capacity)) {
5508 sheaf_flush_unused(s, sheaf);
5509 kfree(sheaf);
5510 return;
5511 }
5512
5513 local_lock(&s->cpu_sheaves->lock);
5514 pcs = this_cpu_ptr(s->cpu_sheaves);
5515 barn = get_barn(s);
5516
5517 if (!pcs->spare) {
5518 pcs->spare = sheaf;
5519 sheaf = NULL;
5520 stat(s, SHEAF_RETURN_FAST);
5521 }
5522
5523 local_unlock(&s->cpu_sheaves->lock);
5524
5525 if (!sheaf)
5526 return;
5527
5528 stat(s, SHEAF_RETURN_SLOW);
5529
5530 /*
5531 * If the barn has too many full sheaves or we fail to refill the sheaf,
5532 * simply flush and free it.
5533 */
5534 if (!barn || data_race(barn->nr_full) >= MAX_FULL_SHEAVES ||
5535 refill_sheaf(s, sheaf, gfp)) {
5536 sheaf_flush_unused(s, sheaf);
5537 free_empty_sheaf(s, sheaf);
5538 return;
5539 }
5540
5541 barn_put_full_sheaf(barn, sheaf);
5542 stat(s, BARN_PUT);
5543 }
5544
5545 /*
5546 * refill a sheaf previously returned by kmem_cache_prefill_sheaf to at least
5547 * the given size
5548 *
5549 * the sheaf might be replaced by a new one when requesting more than
5550 * s->sheaf_capacity objects if such replacement is necessary, but the refill
5551 * fails (returning -ENOMEM), the existing sheaf is left intact
5552 *
5553 * In practice we always refill to full sheaf's capacity.
5554 */
5555 int kmem_cache_refill_sheaf(struct kmem_cache *s, gfp_t gfp,
5556 struct slab_sheaf **sheafp, unsigned int size)
5557 {
5558 struct slab_sheaf *sheaf;
5559
5560 /*
5561 * TODO: do we want to support *sheaf == NULL to be equivalent of
5562 * kmem_cache_prefill_sheaf() ?
5563 */
5564 if (!sheafp || !(*sheafp))
5565 return -EINVAL;
5566
5567 sheaf = *sheafp;
5568 if (sheaf->size >= size)
5569 return 0;
5570
5571 if (likely(sheaf->capacity >= size)) {
5572 if (likely(sheaf->capacity == s->sheaf_capacity))
5573 return refill_sheaf(s, sheaf, gfp);
5574
5575 if (!__kmem_cache_alloc_bulk(s, gfp, sheaf->capacity - sheaf->size,
5576 &sheaf->objects[sheaf->size])) {
5577 return -ENOMEM;
5578 }
5579 sheaf->size = sheaf->capacity;
5580
5581 return 0;
5582 }
5583
5584 /*
5585 * We had a regular sized sheaf and need an oversize one, or we had an
5586 * oversize one already but need a larger one now.
5587 * This should be a very rare path so let's not complicate it.
5588 */
5589 sheaf = kmem_cache_prefill_sheaf(s, gfp, size);
5590 if (!sheaf)
5591 return -ENOMEM;
5592
5593 kmem_cache_return_sheaf(s, gfp, *sheafp);
5594 *sheafp = sheaf;
5595 return 0;
5596 }
5597
5598 /*
5599 * Allocate from a sheaf obtained by kmem_cache_prefill_sheaf()
5600 *
5601 * Guaranteed not to fail as many allocations as was the requested size.
5602 * After the sheaf is emptied, it fails - no fallback to the slab cache itself.
5603 *
5604 * The gfp parameter is meant only to specify __GFP_ZERO or __GFP_ACCOUNT
5605 * memcg charging is forced over limit if necessary, to avoid failure.
5606 */
5607 void *
5608 kmem_cache_alloc_from_sheaf_noprof(struct kmem_cache *s, gfp_t gfp,
5609 struct slab_sheaf *sheaf)
5610 {
5611 void *ret = NULL;
5612 bool init;
5613
5614 if (sheaf->size == 0)
5615 goto out;
5616
5617 ret = sheaf->objects[--sheaf->size];
5618
5619 init = slab_want_init_on_alloc(gfp, s);
5620
5621 /* add __GFP_NOFAIL to force successful memcg charging */
5622 slab_post_alloc_hook(s, NULL, gfp | __GFP_NOFAIL, 1, &ret, init, s->object_size);
5623 out:
5624 trace_kmem_cache_alloc(_RET_IP_, ret, s, gfp, NUMA_NO_NODE);
5625
5626 return ret;
5627 }
5628
5629 unsigned int kmem_cache_sheaf_size(struct slab_sheaf *sheaf)
5630 {
5631 return sheaf->size;
5632 }
5633 /*
5634 * To avoid unnecessary overhead, we pass through large allocation requests
5635 * directly to the page allocator. We use __GFP_COMP, because we will need to
5636 * know the allocation order to free the pages properly in kfree.
5637 */
5638 static void *___kmalloc_large_node(size_t size, gfp_t flags, int node)
5639 {
5640 struct folio *folio;
5641 void *ptr = NULL;
5642 unsigned int order = get_order(size);
5643
5644 if (unlikely(flags & GFP_SLAB_BUG_MASK))
5645 flags = kmalloc_fix_flags(flags);
5646
5647 flags |= __GFP_COMP;
5648
5649 if (node == NUMA_NO_NODE)
5650 folio = (struct folio *)alloc_frozen_pages_noprof(flags, order);
5651 else
5652 folio = (struct folio *)__alloc_frozen_pages_noprof(flags, order, node, NULL);
5653
5654 if (folio) {
5655 ptr = folio_address(folio);
5656 lruvec_stat_mod_folio(folio, NR_SLAB_UNRECLAIMABLE_B,
5657 PAGE_SIZE << order);
5658 __folio_set_large_kmalloc(folio);
5659 }
5660
5661 ptr = kasan_kmalloc_large(ptr, size, flags);
5662 /* As ptr might get tagged, call kmemleak hook after KASAN. */
5663 kmemleak_alloc(ptr, size, 1, flags);
5664 kmsan_kmalloc_large(ptr, size, flags);
5665
5666 return ptr;
5667 }
5668
5669 void *__kmalloc_large_noprof(size_t size, gfp_t flags)
5670 {
5671 void *ret = ___kmalloc_large_node(size, flags, NUMA_NO_NODE);
5672
5673 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << get_order(size),
5674 flags, NUMA_NO_NODE);
5675 return ret;
5676 }
5677 EXPORT_SYMBOL(__kmalloc_large_noprof);
5678
5679 void *__kmalloc_large_node_noprof(size_t size, gfp_t flags, int node)
5680 {
5681 void *ret = ___kmalloc_large_node(size, flags, node);
5682
5683 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << get_order(size),
5684 flags, node);
5685 return ret;
5686 }
5687 EXPORT_SYMBOL(__kmalloc_large_node_noprof);
5688
5689 static __always_inline
5690 void *__do_kmalloc_node(size_t size, kmem_buckets *b, gfp_t flags, int node,
5691 unsigned long caller)
5692 {
5693 struct kmem_cache *s;
5694 void *ret;
5695
5696 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
5697 ret = __kmalloc_large_node_noprof(size, flags, node);
5698 trace_kmalloc(caller, ret, size,
5699 PAGE_SIZE << get_order(size), flags, node);
5700 return ret;
5701 }
5702
5703 if (unlikely(!size))
5704 return ZERO_SIZE_PTR;
5705
5706 s = kmalloc_slab(size, b, flags, caller);
5707
5708 ret = slab_alloc_node(s, NULL, flags, node, caller, size);
5709 ret = kasan_kmalloc(s, ret, size, flags);
5710 trace_kmalloc(caller, ret, size, s->size, flags, node);
5711 return ret;
5712 }
5713 void *__kmalloc_node_noprof(DECL_BUCKET_PARAMS(size, b), gfp_t flags, int node)
5714 {
5715 return __do_kmalloc_node(size, PASS_BUCKET_PARAM(b), flags, node, _RET_IP_);
5716 }
5717 EXPORT_SYMBOL(__kmalloc_node_noprof);
5718
5719 void *__kmalloc_noprof(size_t size, gfp_t flags)
5720 {
5721 return __do_kmalloc_node(size, NULL, flags, NUMA_NO_NODE, _RET_IP_);
5722 }
5723 EXPORT_SYMBOL(__kmalloc_noprof);
5724
5725 /**
5726 * kmalloc_nolock - Allocate an object of given size from any context.
5727 * @size: size to allocate
5728 * @gfp_flags: GFP flags. Only __GFP_ACCOUNT, __GFP_ZERO, __GFP_NO_OBJ_EXT
5729 * allowed.
5730 * @node: node number of the target node.
5731 *
5732 * Return: pointer to the new object or NULL in case of error.
5733 * NULL does not mean EBUSY or EAGAIN. It means ENOMEM.
5734 * There is no reason to call it again and expect !NULL.
5735 */
5736 void *kmalloc_nolock_noprof(size_t size, gfp_t gfp_flags, int node)
5737 {
5738 gfp_t alloc_gfp = __GFP_NOWARN | __GFP_NOMEMALLOC | gfp_flags;
5739 struct kmem_cache *s;
5740 bool can_retry = true;
5741 void *ret = ERR_PTR(-EBUSY);
5742
5743 VM_WARN_ON_ONCE(gfp_flags & ~(__GFP_ACCOUNT | __GFP_ZERO |
5744 __GFP_NO_OBJ_EXT));
5745
5746 if (unlikely(!size))
5747 return ZERO_SIZE_PTR;
5748
5749 if (IS_ENABLED(CONFIG_PREEMPT_RT) && !preemptible())
5750 /*
5751 * kmalloc_nolock() in PREEMPT_RT is not supported from
5752 * non-preemptible context because local_lock becomes a
5753 * sleeping lock on RT.
5754 */
5755 return NULL;
5756
5757 /* On UP, spin_trylock() always succeeds even when it is locked */
5758 if (!IS_ENABLED(CONFIG_SMP) && in_nmi())
5759 return NULL;
5760
5761 retry:
5762 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
5763 return NULL;
5764 s = kmalloc_slab(size, NULL, alloc_gfp, _RET_IP_);
5765
5766 if (!(s->flags & __CMPXCHG_DOUBLE) && !kmem_cache_debug(s))
5767 /*
5768 * kmalloc_nolock() is not supported on architectures that
5769 * don't implement cmpxchg16b, but debug caches don't use
5770 * per-cpu slab and per-cpu partial slabs. They rely on
5771 * kmem_cache_node->list_lock, so kmalloc_nolock() can
5772 * attempt to allocate from debug caches by
5773 * spin_trylock_irqsave(&n->list_lock, ...)
5774 */
5775 return NULL;
5776
5777 /*
5778 * Do not call slab_alloc_node(), since trylock mode isn't
5779 * compatible with slab_pre_alloc_hook/should_failslab and
5780 * kfence_alloc. Hence call __slab_alloc_node() (at most twice)
5781 * and slab_post_alloc_hook() directly.
5782 *
5783 * In !PREEMPT_RT ___slab_alloc() manipulates (freelist,tid) pair
5784 * in irq saved region. It assumes that the same cpu will not
5785 * __update_cpu_freelist_fast() into the same (freelist,tid) pair.
5786 * Therefore use in_nmi() to check whether particular bucket is in
5787 * irq protected section.
5788 *
5789 * If in_nmi() && local_lock_is_locked(s->cpu_slab) then it means that
5790 * this cpu was interrupted somewhere inside ___slab_alloc() after
5791 * it did local_lock_irqsave(&s->cpu_slab->lock, flags).
5792 * In this case fast path with __update_cpu_freelist_fast() is not safe.
5793 */
5794 if (!in_nmi() || !local_lock_is_locked(&s->cpu_slab->lock))
5795 ret = __slab_alloc_node(s, alloc_gfp, node, _RET_IP_, size);
5796
5797 if (PTR_ERR(ret) == -EBUSY) {
5798 if (can_retry) {
5799 /* pick the next kmalloc bucket */
5800 size = s->object_size + 1;
5801 /*
5802 * Another alternative is to
5803 * if (memcg) alloc_gfp &= ~__GFP_ACCOUNT;
5804 * else if (!memcg) alloc_gfp |= __GFP_ACCOUNT;
5805 * to retry from bucket of the same size.
5806 */
5807 can_retry = false;
5808 goto retry;
5809 }
5810 ret = NULL;
5811 }
5812
5813 maybe_wipe_obj_freeptr(s, ret);
5814 slab_post_alloc_hook(s, NULL, alloc_gfp, 1, &ret,
5815 slab_want_init_on_alloc(alloc_gfp, s), size);
5816
5817 ret = kasan_kmalloc(s, ret, size, alloc_gfp);
5818 return ret;
5819 }
5820 EXPORT_SYMBOL_GPL(kmalloc_nolock_noprof);
5821
5822 void *__kmalloc_node_track_caller_noprof(DECL_BUCKET_PARAMS(size, b), gfp_t flags,
5823 int node, unsigned long caller)
5824 {
5825 return __do_kmalloc_node(size, PASS_BUCKET_PARAM(b), flags, node, caller);
5826
5827 }
5828 EXPORT_SYMBOL(__kmalloc_node_track_caller_noprof);
5829
5830 void *__kmalloc_cache_noprof(struct kmem_cache *s, gfp_t gfpflags, size_t size)
5831 {
5832 void *ret = slab_alloc_node(s, NULL, gfpflags, NUMA_NO_NODE,
5833 _RET_IP_, size);
5834
5835 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags, NUMA_NO_NODE);
5836
5837 ret = kasan_kmalloc(s, ret, size, gfpflags);
5838 return ret;
5839 }
5840 EXPORT_SYMBOL(__kmalloc_cache_noprof);
5841
5842 void *__kmalloc_cache_node_noprof(struct kmem_cache *s, gfp_t gfpflags,
5843 int node, size_t size)
5844 {
5845 void *ret = slab_alloc_node(s, NULL, gfpflags, node, _RET_IP_, size);
5846
5847 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags, node);
5848
5849 ret = kasan_kmalloc(s, ret, size, gfpflags);
5850 return ret;
5851 }
5852 EXPORT_SYMBOL(__kmalloc_cache_node_noprof);
5853
5854 static noinline void free_to_partial_list(
5855 struct kmem_cache *s, struct slab *slab,
5856 void *head, void *tail, int bulk_cnt,
5857 unsigned long addr)
5858 {
5859 struct kmem_cache_node *n = get_node(s, slab_nid(slab));
5860 struct slab *slab_free = NULL;
5861 int cnt = bulk_cnt;
5862 unsigned long flags;
5863 depot_stack_handle_t handle = 0;
5864
5865 /*
5866 * We cannot use GFP_NOWAIT as there are callsites where waking up
5867 * kswapd could deadlock
5868 */
5869 if (s->flags & SLAB_STORE_USER)
5870 handle = set_track_prepare(__GFP_NOWARN);
5871
5872 spin_lock_irqsave(&n->list_lock, flags);
5873
5874 if (free_debug_processing(s, slab, head, tail, &cnt, addr, handle)) {
5875 void *prior = slab->freelist;
5876
5877 /* Perform the actual freeing while we still hold the locks */
5878 slab->inuse -= cnt;
5879 set_freepointer(s, tail, prior);
5880 slab->freelist = head;
5881
5882 /*
5883 * If the slab is empty, and node's partial list is full,
5884 * it should be discarded anyway no matter it's on full or
5885 * partial list.
5886 */
5887 if (slab->inuse == 0 && n->nr_partial >= s->min_partial)
5888 slab_free = slab;
5889
5890 if (!prior) {
5891 /* was on full list */
5892 remove_full(s, n, slab);
5893 if (!slab_free) {
5894 add_partial(n, slab, DEACTIVATE_TO_TAIL);
5895 stat(s, FREE_ADD_PARTIAL);
5896 }
5897 } else if (slab_free) {
5898 remove_partial(n, slab);
5899 stat(s, FREE_REMOVE_PARTIAL);
5900 }
5901 }
5902
5903 if (slab_free) {
5904 /*
5905 * Update the counters while still holding n->list_lock to
5906 * prevent spurious validation warnings
5907 */
5908 dec_slabs_node(s, slab_nid(slab_free), slab_free->objects);
5909 }
5910
5911 spin_unlock_irqrestore(&n->list_lock, flags);
5912
5913 if (slab_free) {
5914 stat(s, FREE_SLAB);
5915 free_slab(s, slab_free);
5916 }
5917 }
5918
5919 /*
5920 * Slow path handling. This may still be called frequently since objects
5921 * have a longer lifetime than the cpu slabs in most processing loads.
5922 *
5923 * So we still attempt to reduce cache line usage. Just take the slab
5924 * lock and free the item. If there is no additional partial slab
5925 * handling required then we can return immediately.
5926 */
5927 static void __slab_free(struct kmem_cache *s, struct slab *slab,
5928 void *head, void *tail, int cnt,
5929 unsigned long addr)
5930
5931 {
5932 void *prior;
5933 int was_frozen;
5934 struct slab new;
5935 unsigned long counters;
5936 struct kmem_cache_node *n = NULL;
5937 unsigned long flags;
5938 bool on_node_partial;
5939
5940 stat(s, FREE_SLOWPATH);
5941
5942 if (IS_ENABLED(CONFIG_SLUB_TINY) || kmem_cache_debug(s)) {
5943 free_to_partial_list(s, slab, head, tail, cnt, addr);
5944 return;
5945 }
5946
5947 do {
5948 if (unlikely(n)) {
5949 spin_unlock_irqrestore(&n->list_lock, flags);
5950 n = NULL;
5951 }
5952 prior = slab->freelist;
5953 counters = slab->counters;
5954 set_freepointer(s, tail, prior);
5955 new.counters = counters;
5956 was_frozen = new.frozen;
5957 new.inuse -= cnt;
5958 if ((!new.inuse || !prior) && !was_frozen) {
5959 /* Needs to be taken off a list */
5960 if (!kmem_cache_has_cpu_partial(s) || prior) {
5961
5962 n = get_node(s, slab_nid(slab));
5963 /*
5964 * Speculatively acquire the list_lock.
5965 * If the cmpxchg does not succeed then we may
5966 * drop the list_lock without any processing.
5967 *
5968 * Otherwise the list_lock will synchronize with
5969 * other processors updating the list of slabs.
5970 */
5971 spin_lock_irqsave(&n->list_lock, flags);
5972
5973 on_node_partial = slab_test_node_partial(slab);
5974 }
5975 }
5976
5977 } while (!slab_update_freelist(s, slab,
5978 prior, counters,
5979 head, new.counters,
5980 "__slab_free"));
5981
5982 if (likely(!n)) {
5983
5984 if (likely(was_frozen)) {
5985 /*
5986 * The list lock was not taken therefore no list
5987 * activity can be necessary.
5988 */
5989 stat(s, FREE_FROZEN);
5990 } else if (kmem_cache_has_cpu_partial(s) && !prior) {
5991 /*
5992 * If we started with a full slab then put it onto the
5993 * per cpu partial list.
5994 */
5995 put_cpu_partial(s, slab, 1);
5996 stat(s, CPU_PARTIAL_FREE);
5997 }
5998
5999 return;
6000 }
6001
6002 /*
6003 * This slab was partially empty but not on the per-node partial list,
6004 * in which case we shouldn't manipulate its list, just return.
6005 */
6006 if (prior && !on_node_partial) {
6007 spin_unlock_irqrestore(&n->list_lock, flags);
6008 return;
6009 }
6010
6011 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
6012 goto slab_empty;
6013
6014 /*
6015 * Objects left in the slab. If it was not on the partial list before
6016 * then add it.
6017 */
6018 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
6019 add_partial(n, slab, DEACTIVATE_TO_TAIL);
6020 stat(s, FREE_ADD_PARTIAL);
6021 }
6022 spin_unlock_irqrestore(&n->list_lock, flags);
6023 return;
6024
6025 slab_empty:
6026 if (prior) {
6027 /*
6028 * Slab on the partial list.
6029 */
6030 remove_partial(n, slab);
6031 stat(s, FREE_REMOVE_PARTIAL);
6032 }
6033
6034 spin_unlock_irqrestore(&n->list_lock, flags);
6035 stat(s, FREE_SLAB);
6036 discard_slab(s, slab);
6037 }
6038
6039 /*
6040 * pcs is locked. We should have get rid of the spare sheaf and obtained an
6041 * empty sheaf, while the main sheaf is full. We want to install the empty sheaf
6042 * as a main sheaf, and make the current main sheaf a spare sheaf.
6043 *
6044 * However due to having relinquished the cpu_sheaves lock when obtaining
6045 * the empty sheaf, we need to handle some unlikely but possible cases.
6046 *
6047 * If we put any sheaf to barn here, it's because we were interrupted or have
6048 * been migrated to a different cpu, which should be rare enough so just ignore
6049 * the barn's limits to simplify the handling.
6050 *
6051 * An alternative scenario that gets us here is when we fail
6052 * barn_replace_full_sheaf(), because there's no empty sheaf available in the
6053 * barn, so we had to allocate it by alloc_empty_sheaf(). But because we saw the
6054 * limit on full sheaves was not exceeded, we assume it didn't change and just
6055 * put the full sheaf there.
6056 */
6057 static void __pcs_install_empty_sheaf(struct kmem_cache *s,
6058 struct slub_percpu_sheaves *pcs, struct slab_sheaf *empty,
6059 struct node_barn *barn)
6060 {
6061 lockdep_assert_held(this_cpu_ptr(&s->cpu_sheaves->lock));
6062
6063 /* This is what we expect to find if nobody interrupted us. */
6064 if (likely(!pcs->spare)) {
6065 pcs->spare = pcs->main;
6066 pcs->main = empty;
6067 return;
6068 }
6069
6070 /*
6071 * Unlikely because if the main sheaf had space, we would have just
6072 * freed to it. Get rid of our empty sheaf.
6073 */
6074 if (pcs->main->size < s->sheaf_capacity) {
6075 barn_put_empty_sheaf(barn, empty);
6076 return;
6077 }
6078
6079 /* Also unlikely for the same reason */
6080 if (pcs->spare->size < s->sheaf_capacity) {
6081 swap(pcs->main, pcs->spare);
6082 barn_put_empty_sheaf(barn, empty);
6083 return;
6084 }
6085
6086 /*
6087 * We probably failed barn_replace_full_sheaf() due to no empty sheaf
6088 * available there, but we allocated one, so finish the job.
6089 */
6090 barn_put_full_sheaf(barn, pcs->main);
6091 stat(s, BARN_PUT);
6092 pcs->main = empty;
6093 }
6094
6095 /*
6096 * Replace the full main sheaf with a (at least partially) empty sheaf.
6097 *
6098 * Must be called with the cpu_sheaves local lock locked. If successful, returns
6099 * the pcs pointer and the local lock locked (possibly on a different cpu than
6100 * initially called). If not successful, returns NULL and the local lock
6101 * unlocked.
6102 */
6103 static struct slub_percpu_sheaves *
6104 __pcs_replace_full_main(struct kmem_cache *s, struct slub_percpu_sheaves *pcs)
6105 {
6106 struct slab_sheaf *empty;
6107 struct node_barn *barn;
6108 bool put_fail;
6109
6110 restart:
6111 lockdep_assert_held(this_cpu_ptr(&s->cpu_sheaves->lock));
6112
6113 barn = get_barn(s);
6114 if (!barn) {
6115 local_unlock(&s->cpu_sheaves->lock);
6116 return NULL;
6117 }
6118
6119 put_fail = false;
6120
6121 if (!pcs->spare) {
6122 empty = barn_get_empty_sheaf(barn);
6123 if (empty) {
6124 pcs->spare = pcs->main;
6125 pcs->main = empty;
6126 return pcs;
6127 }
6128 goto alloc_empty;
6129 }
6130
6131 if (pcs->spare->size < s->sheaf_capacity) {
6132 swap(pcs->main, pcs->spare);
6133 return pcs;
6134 }
6135
6136 empty = barn_replace_full_sheaf(barn, pcs->main);
6137
6138 if (!IS_ERR(empty)) {
6139 stat(s, BARN_PUT);
6140 pcs->main = empty;
6141 return pcs;
6142 }
6143
6144 if (PTR_ERR(empty) == -E2BIG) {
6145 /* Since we got here, spare exists and is full */
6146 struct slab_sheaf *to_flush = pcs->spare;
6147
6148 stat(s, BARN_PUT_FAIL);
6149
6150 pcs->spare = NULL;
6151 local_unlock(&s->cpu_sheaves->lock);
6152
6153 sheaf_flush_unused(s, to_flush);
6154 empty = to_flush;
6155 goto got_empty;
6156 }
6157
6158 /*
6159 * We could not replace full sheaf because barn had no empty
6160 * sheaves. We can still allocate it and put the full sheaf in
6161 * __pcs_install_empty_sheaf(), but if we fail to allocate it,
6162 * make sure to count the fail.
6163 */
6164 put_fail = true;
6165
6166 alloc_empty:
6167 local_unlock(&s->cpu_sheaves->lock);
6168
6169 empty = alloc_empty_sheaf(s, GFP_NOWAIT);
6170 if (empty)
6171 goto got_empty;
6172
6173 if (put_fail)
6174 stat(s, BARN_PUT_FAIL);
6175
6176 if (!sheaf_try_flush_main(s))
6177 return NULL;
6178
6179 if (!local_trylock(&s->cpu_sheaves->lock))
6180 return NULL;
6181
6182 pcs = this_cpu_ptr(s->cpu_sheaves);
6183
6184 /*
6185 * we flushed the main sheaf so it should be empty now,
6186 * but in case we got preempted or migrated, we need to
6187 * check again
6188 */
6189 if (pcs->main->size == s->sheaf_capacity)
6190 goto restart;
6191
6192 return pcs;
6193
6194 got_empty:
6195 if (!local_trylock(&s->cpu_sheaves->lock)) {
6196 barn_put_empty_sheaf(barn, empty);
6197 return NULL;
6198 }
6199
6200 pcs = this_cpu_ptr(s->cpu_sheaves);
6201 __pcs_install_empty_sheaf(s, pcs, empty, barn);
6202
6203 return pcs;
6204 }
6205
6206 /*
6207 * Free an object to the percpu sheaves.
6208 * The object is expected to have passed slab_free_hook() already.
6209 */
6210 static __fastpath_inline
6211 bool free_to_pcs(struct kmem_cache *s, void *object)
6212 {
6213 struct slub_percpu_sheaves *pcs;
6214
6215 if (!local_trylock(&s->cpu_sheaves->lock))
6216 return false;
6217
6218 pcs = this_cpu_ptr(s->cpu_sheaves);
6219
6220 if (unlikely(pcs->main->size == s->sheaf_capacity)) {
6221
6222 pcs = __pcs_replace_full_main(s, pcs);
6223 if (unlikely(!pcs))
6224 return false;
6225 }
6226
6227 pcs->main->objects[pcs->main->size++] = object;
6228
6229 local_unlock(&s->cpu_sheaves->lock);
6230
6231 stat(s, FREE_PCS);
6232
6233 return true;
6234 }
6235
6236 static void rcu_free_sheaf(struct rcu_head *head)
6237 {
6238 struct kmem_cache_node *n;
6239 struct slab_sheaf *sheaf;
6240 struct node_barn *barn = NULL;
6241 struct kmem_cache *s;
6242
6243 sheaf = container_of(head, struct slab_sheaf, rcu_head);
6244
6245 s = sheaf->cache;
6246
6247 /*
6248 * This may remove some objects due to slab_free_hook() returning false,
6249 * so that the sheaf might no longer be completely full. But it's easier
6250 * to handle it as full (unless it became completely empty), as the code
6251 * handles it fine. The only downside is that sheaf will serve fewer
6252 * allocations when reused. It only happens due to debugging, which is a
6253 * performance hit anyway.
6254 */
6255 __rcu_free_sheaf_prepare(s, sheaf);
6256
6257 n = get_node(s, sheaf->node);
6258 if (!n)
6259 goto flush;
6260
6261 barn = n->barn;
6262
6263 /* due to slab_free_hook() */
6264 if (unlikely(sheaf->size == 0))
6265 goto empty;
6266
6267 /*
6268 * Checking nr_full/nr_empty outside lock avoids contention in case the
6269 * barn is at the respective limit. Due to the race we might go over the
6270 * limit but that should be rare and harmless.
6271 */
6272
6273 if (data_race(barn->nr_full) < MAX_FULL_SHEAVES) {
6274 stat(s, BARN_PUT);
6275 barn_put_full_sheaf(barn, sheaf);
6276 return;
6277 }
6278
6279 flush:
6280 stat(s, BARN_PUT_FAIL);
6281 sheaf_flush_unused(s, sheaf);
6282
6283 empty:
6284 if (barn && data_race(barn->nr_empty) < MAX_EMPTY_SHEAVES) {
6285 barn_put_empty_sheaf(barn, sheaf);
6286 return;
6287 }
6288
6289 free_empty_sheaf(s, sheaf);
6290 }
6291
6292 /*
6293 * kvfree_call_rcu() can be called while holding a raw_spinlock_t. Since
6294 * __kfree_rcu_sheaf() may acquire a spinlock_t (sleeping lock on PREEMPT_RT),
6295 * this would violate lock nesting rules. Therefore, kvfree_call_rcu() avoids
6296 * this problem by bypassing the sheaves layer entirely on PREEMPT_RT.
6297 *
6298 * However, lockdep still complains that it is invalid to acquire spinlock_t
6299 * while holding raw_spinlock_t, even on !PREEMPT_RT where spinlock_t is a
6300 * spinning lock. Tell lockdep that acquiring spinlock_t is valid here
6301 * by temporarily raising the wait-type to LD_WAIT_CONFIG.
6302 */
6303 static DEFINE_WAIT_OVERRIDE_MAP(kfree_rcu_sheaf_map, LD_WAIT_CONFIG);
6304
6305 bool __kfree_rcu_sheaf(struct kmem_cache *s, void *obj)
6306 {
6307 struct slub_percpu_sheaves *pcs;
6308 struct slab_sheaf *rcu_sheaf;
6309
6310 if (WARN_ON_ONCE(IS_ENABLED(CONFIG_PREEMPT_RT)))
6311 return false;
6312
6313 lock_map_acquire_try(&kfree_rcu_sheaf_map);
6314
6315 if (!local_trylock(&s->cpu_sheaves->lock))
6316 goto fail;
6317
6318 pcs = this_cpu_ptr(s->cpu_sheaves);
6319
6320 if (unlikely(!pcs->rcu_free)) {
6321
6322 struct slab_sheaf *empty;
6323 struct node_barn *barn;
6324
6325 if (pcs->spare && pcs->spare->size == 0) {
6326 pcs->rcu_free = pcs->spare;
6327 pcs->spare = NULL;
6328 goto do_free;
6329 }
6330
6331 barn = get_barn(s);
6332 if (!barn) {
6333 local_unlock(&s->cpu_sheaves->lock);
6334 goto fail;
6335 }
6336
6337 empty = barn_get_empty_sheaf(barn);
6338
6339 if (empty) {
6340 pcs->rcu_free = empty;
6341 goto do_free;
6342 }
6343
6344 local_unlock(&s->cpu_sheaves->lock);
6345
6346 empty = alloc_empty_sheaf(s, GFP_NOWAIT);
6347
6348 if (!empty)
6349 goto fail;
6350
6351 if (!local_trylock(&s->cpu_sheaves->lock)) {
6352 barn_put_empty_sheaf(barn, empty);
6353 goto fail;
6354 }
6355
6356 pcs = this_cpu_ptr(s->cpu_sheaves);
6357
6358 if (unlikely(pcs->rcu_free))
6359 barn_put_empty_sheaf(barn, empty);
6360 else
6361 pcs->rcu_free = empty;
6362 }
6363
6364 do_free:
6365
6366 rcu_sheaf = pcs->rcu_free;
6367
6368 /*
6369 * Since we flush immediately when size reaches capacity, we never reach
6370 * this with size already at capacity, so no OOB write is possible.
6371 */
6372 rcu_sheaf->objects[rcu_sheaf->size++] = obj;
6373
6374 if (likely(rcu_sheaf->size < s->sheaf_capacity)) {
6375 rcu_sheaf = NULL;
6376 } else {
6377 pcs->rcu_free = NULL;
6378 rcu_sheaf->node = numa_mem_id();
6379 }
6380
6381 /*
6382 * we flush before local_unlock to make sure a racing
6383 * flush_all_rcu_sheaves() doesn't miss this sheaf
6384 */
6385 if (rcu_sheaf)
6386 call_rcu(&rcu_sheaf->rcu_head, rcu_free_sheaf);
6387
6388 local_unlock(&s->cpu_sheaves->lock);
6389
6390 stat(s, FREE_RCU_SHEAF);
6391 lock_map_release(&kfree_rcu_sheaf_map);
6392 return true;
6393
6394 fail:
6395 stat(s, FREE_RCU_SHEAF_FAIL);
6396 lock_map_release(&kfree_rcu_sheaf_map);
6397 return false;
6398 }
6399
6400 /*
6401 * Bulk free objects to the percpu sheaves.
6402 * Unlike free_to_pcs() this includes the calls to all necessary hooks
6403 * and the fallback to freeing to slab pages.
6404 */
6405 static void free_to_pcs_bulk(struct kmem_cache *s, size_t size, void **p)
6406 {
6407 struct slub_percpu_sheaves *pcs;
6408 struct slab_sheaf *main, *empty;
6409 bool init = slab_want_init_on_free(s);
6410 unsigned int batch, i = 0;
6411 struct node_barn *barn;
6412 void *remote_objects[PCS_BATCH_MAX];
6413 unsigned int remote_nr = 0;
6414 int node = numa_mem_id();
6415
6416 next_remote_batch:
6417 while (i < size) {
6418 struct slab *slab = virt_to_slab(p[i]);
6419
6420 memcg_slab_free_hook(s, slab, p + i, 1);
6421 alloc_tagging_slab_free_hook(s, slab, p + i, 1);
6422
6423 if (unlikely(!slab_free_hook(s, p[i], init, false))) {
6424 p[i] = p[--size];
6425 continue;
6426 }
6427
6428 if (unlikely(IS_ENABLED(CONFIG_NUMA) && slab_nid(slab) != node)) {
6429 remote_objects[remote_nr] = p[i];
6430 p[i] = p[--size];
6431 if (++remote_nr >= PCS_BATCH_MAX)
6432 goto flush_remote;
6433 continue;
6434 }
6435
6436 i++;
6437 }
6438
6439 if (!size)
6440 goto flush_remote;
6441
6442 next_batch:
6443 if (!local_trylock(&s->cpu_sheaves->lock))
6444 goto fallback;
6445
6446 pcs = this_cpu_ptr(s->cpu_sheaves);
6447
6448 if (likely(pcs->main->size < s->sheaf_capacity))
6449 goto do_free;
6450
6451 barn = get_barn(s);
6452 if (!barn)
6453 goto no_empty;
6454
6455 if (!pcs->spare) {
6456 empty = barn_get_empty_sheaf(barn);
6457 if (!empty)
6458 goto no_empty;
6459
6460 pcs->spare = pcs->main;
6461 pcs->main = empty;
6462 goto do_free;
6463 }
6464
6465 if (pcs->spare->size < s->sheaf_capacity) {
6466 swap(pcs->main, pcs->spare);
6467 goto do_free;
6468 }
6469
6470 empty = barn_replace_full_sheaf(barn, pcs->main);
6471 if (IS_ERR(empty)) {
6472 stat(s, BARN_PUT_FAIL);
6473 goto no_empty;
6474 }
6475
6476 stat(s, BARN_PUT);
6477 pcs->main = empty;
6478
6479 do_free:
6480 main = pcs->main;
6481 batch = min(size, s->sheaf_capacity - main->size);
6482
6483 memcpy(main->objects + main->size, p, batch * sizeof(void *));
6484 main->size += batch;
6485
6486 local_unlock(&s->cpu_sheaves->lock);
6487
6488 stat_add(s, FREE_PCS, batch);
6489
6490 if (batch < size) {
6491 p += batch;
6492 size -= batch;
6493 goto next_batch;
6494 }
6495
6496 if (remote_nr)
6497 goto flush_remote;
6498
6499 return;
6500
6501 no_empty:
6502 local_unlock(&s->cpu_sheaves->lock);
6503
6504 /*
6505 * if we depleted all empty sheaves in the barn or there are too
6506 * many full sheaves, free the rest to slab pages
6507 */
6508 fallback:
6509 __kmem_cache_free_bulk(s, size, p);
6510
6511 flush_remote:
6512 if (remote_nr) {
6513 __kmem_cache_free_bulk(s, remote_nr, &remote_objects[0]);
6514 if (i < size) {
6515 remote_nr = 0;
6516 goto next_remote_batch;
6517 }
6518 }
6519 }
6520
6521 struct defer_free {
6522 struct llist_head objects;
6523 struct llist_head slabs;
6524 struct irq_work work;
6525 };
6526
6527 static void free_deferred_objects(struct irq_work *work);
6528
6529 static DEFINE_PER_CPU(struct defer_free, defer_free_objects) = {
6530 .objects = LLIST_HEAD_INIT(objects),
6531 .slabs = LLIST_HEAD_INIT(slabs),
6532 .work = IRQ_WORK_INIT(free_deferred_objects),
6533 };
6534
6535 /*
6536 * In PREEMPT_RT irq_work runs in per-cpu kthread, so it's safe
6537 * to take sleeping spin_locks from __slab_free() and deactivate_slab().
6538 * In !PREEMPT_RT irq_work will run after local_unlock_irqrestore().
6539 */
6540 static void free_deferred_objects(struct irq_work *work)
6541 {
6542 struct defer_free *df = container_of(work, struct defer_free, work);
6543 struct llist_head *objs = &df->objects;
6544 struct llist_head *slabs = &df->slabs;
6545 struct llist_node *llnode, *pos, *t;
6546
6547 if (llist_empty(objs) && llist_empty(slabs))
6548 return;
6549
6550 llnode = llist_del_all(objs);
6551 llist_for_each_safe(pos, t, llnode) {
6552 struct kmem_cache *s;
6553 struct slab *slab;
6554 void *x = pos;
6555
6556 slab = virt_to_slab(x);
6557 s = slab->slab_cache;
6558
6559 /* Point 'x' back to the beginning of allocated object */
6560 x -= s->offset;
6561
6562 /*
6563 * We used freepointer in 'x' to link 'x' into df->objects.
6564 * Clear it to NULL to avoid false positive detection
6565 * of "Freepointer corruption".
6566 */
6567 set_freepointer(s, x, NULL);
6568
6569 __slab_free(s, slab, x, x, 1, _THIS_IP_);
6570 }
6571
6572 llnode = llist_del_all(slabs);
6573 llist_for_each_safe(pos, t, llnode) {
6574 struct slab *slab = container_of(pos, struct slab, llnode);
6575
6576 if (slab->frozen)
6577 deactivate_slab(slab->slab_cache, slab, slab->flush_freelist);
6578 else
6579 free_slab(slab->slab_cache, slab);
6580 }
6581 }
6582
6583 static void defer_free(struct kmem_cache *s, void *head)
6584 {
6585 struct defer_free *df;
6586
6587 guard(preempt)();
6588
6589 head = kasan_reset_tag(head);
6590
6591 df = this_cpu_ptr(&defer_free_objects);
6592 if (llist_add(head + s->offset, &df->objects))
6593 irq_work_queue(&df->work);
6594 }
6595
6596 static void defer_deactivate_slab(struct slab *slab, void *flush_freelist)
6597 {
6598 struct defer_free *df;
6599
6600 slab->flush_freelist = flush_freelist;
6601
6602 guard(preempt)();
6603
6604 df = this_cpu_ptr(&defer_free_objects);
6605 if (llist_add(&slab->llnode, &df->slabs))
6606 irq_work_queue(&df->work);
6607 }
6608
6609 void defer_free_barrier(void)
6610 {
6611 int cpu;
6612
6613 for_each_possible_cpu(cpu)
6614 irq_work_sync(&per_cpu_ptr(&defer_free_objects, cpu)->work);
6615 }
6616
6617 /*
6618 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
6619 * can perform fastpath freeing without additional function calls.
6620 *
6621 * The fastpath is only possible if we are freeing to the current cpu slab
6622 * of this processor. This typically the case if we have just allocated
6623 * the item before.
6624 *
6625 * If fastpath is not possible then fall back to __slab_free where we deal
6626 * with all sorts of special processing.
6627 *
6628 * Bulk free of a freelist with several objects (all pointing to the
6629 * same slab) possible by specifying head and tail ptr, plus objects
6630 * count (cnt). Bulk free indicated by tail pointer being set.
6631 */
6632 static __always_inline void do_slab_free(struct kmem_cache *s,
6633 struct slab *slab, void *head, void *tail,
6634 int cnt, unsigned long addr)
6635 {
6636 /* cnt == 0 signals that it's called from kfree_nolock() */
6637 bool allow_spin = cnt;
6638 struct kmem_cache_cpu *c;
6639 unsigned long tid;
6640 void **freelist;
6641
6642 redo:
6643 /*
6644 * Determine the currently cpus per cpu slab.
6645 * The cpu may change afterward. However that does not matter since
6646 * data is retrieved via this pointer. If we are on the same cpu
6647 * during the cmpxchg then the free will succeed.
6648 */
6649 c = raw_cpu_ptr(s->cpu_slab);
6650 tid = READ_ONCE(c->tid);
6651
6652 /* Same with comment on barrier() in __slab_alloc_node() */
6653 barrier();
6654
6655 if (unlikely(slab != c->slab)) {
6656 if (unlikely(!allow_spin)) {
6657 /*
6658 * __slab_free() can locklessly cmpxchg16 into a slab,
6659 * but then it might need to take spin_lock or local_lock
6660 * in put_cpu_partial() for further processing.
6661 * Avoid the complexity and simply add to a deferred list.
6662 */
6663 defer_free(s, head);
6664 } else {
6665 __slab_free(s, slab, head, tail, cnt, addr);
6666 }
6667 return;
6668 }
6669
6670 if (unlikely(!allow_spin)) {
6671 if ((in_nmi() || !USE_LOCKLESS_FAST_PATH()) &&
6672 local_lock_is_locked(&s->cpu_slab->lock)) {
6673 defer_free(s, head);
6674 return;
6675 }
6676 cnt = 1; /* restore cnt. kfree_nolock() frees one object at a time */
6677 }
6678
6679 if (USE_LOCKLESS_FAST_PATH()) {
6680 freelist = READ_ONCE(c->freelist);
6681
6682 set_freepointer(s, tail, freelist);
6683
6684 if (unlikely(!__update_cpu_freelist_fast(s, freelist, head, tid))) {
6685 note_cmpxchg_failure("slab_free", s, tid);
6686 goto redo;
6687 }
6688 } else {
6689 __maybe_unused unsigned long flags = 0;
6690
6691 /* Update the free list under the local lock */
6692 local_lock_cpu_slab(s, flags);
6693 c = this_cpu_ptr(s->cpu_slab);
6694 if (unlikely(slab != c->slab)) {
6695 local_unlock_cpu_slab(s, flags);
6696 goto redo;
6697 }
6698 tid = c->tid;
6699 freelist = c->freelist;
6700
6701 set_freepointer(s, tail, freelist);
6702 c->freelist = head;
6703 c->tid = next_tid(tid);
6704
6705 local_unlock_cpu_slab(s, flags);
6706 }
6707 stat_add(s, FREE_FASTPATH, cnt);
6708 }
6709
6710 static __fastpath_inline
6711 void slab_free(struct kmem_cache *s, struct slab *slab, void *object,
6712 unsigned long addr)
6713 {
6714 memcg_slab_free_hook(s, slab, &object, 1);
6715 alloc_tagging_slab_free_hook(s, slab, &object, 1);
6716
6717 if (unlikely(!slab_free_hook(s, object, slab_want_init_on_free(s), false)))
6718 return;
6719
6720 if (s->cpu_sheaves && likely(!IS_ENABLED(CONFIG_NUMA) ||
6721 slab_nid(slab) == numa_mem_id())) {
6722 if (likely(free_to_pcs(s, object)))
6723 return;
6724 }
6725
6726 do_slab_free(s, slab, object, object, 1, addr);
6727 }
6728
6729 #ifdef CONFIG_MEMCG
6730 /* Do not inline the rare memcg charging failed path into the allocation path */
6731 static noinline
6732 void memcg_alloc_abort_single(struct kmem_cache *s, void *object)
6733 {
6734 struct slab *slab = virt_to_slab(object);
6735
6736 alloc_tagging_slab_free_hook(s, slab, &object, 1);
6737
6738 if (likely(slab_free_hook(s, object, slab_want_init_on_free(s), false)))
6739 do_slab_free(s, slab, object, object, 1, _RET_IP_);
6740 }
6741 #endif
6742
6743 static __fastpath_inline
6744 void slab_free_bulk(struct kmem_cache *s, struct slab *slab, void *head,
6745 void *tail, void **p, int cnt, unsigned long addr)
6746 {
6747 memcg_slab_free_hook(s, slab, p, cnt);
6748 alloc_tagging_slab_free_hook(s, slab, p, cnt);
6749 /*
6750 * With KASAN enabled slab_free_freelist_hook modifies the freelist
6751 * to remove objects, whose reuse must be delayed.
6752 */
6753 if (likely(slab_free_freelist_hook(s, &head, &tail, &cnt)))
6754 do_slab_free(s, slab, head, tail, cnt, addr);
6755 }
6756
6757 #ifdef CONFIG_SLUB_RCU_DEBUG
6758 static void slab_free_after_rcu_debug(struct rcu_head *rcu_head)
6759 {
6760 struct rcu_delayed_free *delayed_free =
6761 container_of(rcu_head, struct rcu_delayed_free, head);
6762 void *object = delayed_free->object;
6763 struct slab *slab = virt_to_slab(object);
6764 struct kmem_cache *s;
6765
6766 kfree(delayed_free);
6767
6768 if (WARN_ON(is_kfence_address(object)))
6769 return;
6770
6771 /* find the object and the cache again */
6772 if (WARN_ON(!slab))
6773 return;
6774 s = slab->slab_cache;
6775 if (WARN_ON(!(s->flags & SLAB_TYPESAFE_BY_RCU)))
6776 return;
6777
6778 /* resume freeing */
6779 if (slab_free_hook(s, object, slab_want_init_on_free(s), true))
6780 do_slab_free(s, slab, object, object, 1, _THIS_IP_);
6781 }
6782 #endif /* CONFIG_SLUB_RCU_DEBUG */
6783
6784 #ifdef CONFIG_KASAN_GENERIC
6785 void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
6786 {
6787 do_slab_free(cache, virt_to_slab(x), x, x, 1, addr);
6788 }
6789 #endif
6790
6791 static inline struct kmem_cache *virt_to_cache(const void *obj)
6792 {
6793 struct slab *slab;
6794
6795 slab = virt_to_slab(obj);
6796 if (WARN_ONCE(!slab, "%s: Object is not a Slab page!\n", __func__))
6797 return NULL;
6798 return slab->slab_cache;
6799 }
6800
6801 static inline struct kmem_cache *cache_from_obj(struct kmem_cache *s, void *x)
6802 {
6803 struct kmem_cache *cachep;
6804
6805 if (!IS_ENABLED(CONFIG_SLAB_FREELIST_HARDENED) &&
6806 !kmem_cache_debug_flags(s, SLAB_CONSISTENCY_CHECKS))
6807 return s;
6808
6809 cachep = virt_to_cache(x);
6810 if (WARN(cachep && cachep != s,
6811 "%s: Wrong slab cache. %s but object is from %s\n",
6812 __func__, s->name, cachep->name))
6813 print_tracking(cachep, x);
6814 return cachep;
6815 }
6816
6817 /**
6818 * kmem_cache_free - Deallocate an object
6819 * @s: The cache the allocation was from.
6820 * @x: The previously allocated object.
6821 *
6822 * Free an object which was previously allocated from this
6823 * cache.
6824 */
6825 void kmem_cache_free(struct kmem_cache *s, void *x)
6826 {
6827 s = cache_from_obj(s, x);
6828 if (!s)
6829 return;
6830 trace_kmem_cache_free(_RET_IP_, x, s);
6831 slab_free(s, virt_to_slab(x), x, _RET_IP_);
6832 }
6833 EXPORT_SYMBOL(kmem_cache_free);
6834
6835 static void free_large_kmalloc(struct folio *folio, void *object)
6836 {
6837 unsigned int order = folio_order(folio);
6838
6839 if (WARN_ON_ONCE(!folio_test_large_kmalloc(folio))) {
6840 dump_page(&folio->page, "Not a kmalloc allocation");
6841 return;
6842 }
6843
6844 if (WARN_ON_ONCE(order == 0))
6845 pr_warn_once("object pointer: 0x%p\n", object);
6846
6847 kmemleak_free(object);
6848 kasan_kfree_large(object);
6849 kmsan_kfree_large(object);
6850
6851 lruvec_stat_mod_folio(folio, NR_SLAB_UNRECLAIMABLE_B,
6852 -(PAGE_SIZE << order));
6853 __folio_clear_large_kmalloc(folio);
6854 free_frozen_pages(&folio->page, order);
6855 }
6856
6857 /*
6858 * Given an rcu_head embedded within an object obtained from kvmalloc at an
6859 * offset < 4k, free the object in question.
6860 */
6861 void kvfree_rcu_cb(struct rcu_head *head)
6862 {
6863 void *obj = head;
6864 struct folio *folio;
6865 struct slab *slab;
6866 struct kmem_cache *s;
6867 void *slab_addr;
6868
6869 if (is_vmalloc_addr(obj)) {
6870 obj = (void *) PAGE_ALIGN_DOWN((unsigned long)obj);
6871 vfree(obj);
6872 return;
6873 }
6874
6875 folio = virt_to_folio(obj);
6876 if (!folio_test_slab(folio)) {
6877 /*
6878 * rcu_head offset can be only less than page size so no need to
6879 * consider folio order
6880 */
6881 obj = (void *) PAGE_ALIGN_DOWN((unsigned long)obj);
6882 free_large_kmalloc(folio, obj);
6883 return;
6884 }
6885
6886 slab = folio_slab(folio);
6887 s = slab->slab_cache;
6888 slab_addr = folio_address(folio);
6889
6890 if (is_kfence_address(obj)) {
6891 obj = kfence_object_start(obj);
6892 } else {
6893 unsigned int idx = __obj_to_index(s, slab_addr, obj);
6894
6895 obj = slab_addr + s->size * idx;
6896 obj = fixup_red_left(s, obj);
6897 }
6898
6899 slab_free(s, slab, obj, _RET_IP_);
6900 }
6901
6902 /**
6903 * kfree - free previously allocated memory
6904 * @object: pointer returned by kmalloc() or kmem_cache_alloc()
6905 *
6906 * If @object is NULL, no operation is performed.
6907 */
6908 void kfree(const void *object)
6909 {
6910 struct folio *folio;
6911 struct slab *slab;
6912 struct kmem_cache *s;
6913 void *x = (void *)object;
6914
6915 trace_kfree(_RET_IP_, object);
6916
6917 if (unlikely(ZERO_OR_NULL_PTR(object)))
6918 return;
6919
6920 folio = virt_to_folio(object);
6921 if (unlikely(!folio_test_slab(folio))) {
6922 free_large_kmalloc(folio, (void *)object);
6923 return;
6924 }
6925
6926 slab = folio_slab(folio);
6927 s = slab->slab_cache;
6928 slab_free(s, slab, x, _RET_IP_);
6929 }
6930 EXPORT_SYMBOL(kfree);
6931
6932 /*
6933 * Can be called while holding raw_spinlock_t or from IRQ and NMI,
6934 * but ONLY for objects allocated by kmalloc_nolock().
6935 * Debug checks (like kmemleak and kfence) were skipped on allocation,
6936 * hence
6937 * obj = kmalloc(); kfree_nolock(obj);
6938 * will miss kmemleak/kfence book keeping and will cause false positives.
6939 * large_kmalloc is not supported either.
6940 */
6941 void kfree_nolock(const void *object)
6942 {
6943 struct folio *folio;
6944 struct slab *slab;
6945 struct kmem_cache *s;
6946 void *x = (void *)object;
6947
6948 if (unlikely(ZERO_OR_NULL_PTR(object)))
6949 return;
6950
6951 folio = virt_to_folio(object);
6952 if (unlikely(!folio_test_slab(folio))) {
6953 WARN_ONCE(1, "large_kmalloc is not supported by kfree_nolock()");
6954 return;
6955 }
6956
6957 slab = folio_slab(folio);
6958 s = slab->slab_cache;
6959
6960 memcg_slab_free_hook(s, slab, &x, 1);
6961 alloc_tagging_slab_free_hook(s, slab, &x, 1);
6962 /*
6963 * Unlike slab_free() do NOT call the following:
6964 * kmemleak_free_recursive(x, s->flags);
6965 * debug_check_no_locks_freed(x, s->object_size);
6966 * debug_check_no_obj_freed(x, s->object_size);
6967 * __kcsan_check_access(x, s->object_size, ..);
6968 * kfence_free(x);
6969 * since they take spinlocks or not safe from any context.
6970 */
6971 kmsan_slab_free(s, x);
6972 /*
6973 * If KASAN finds a kernel bug it will do kasan_report_invalid_free()
6974 * which will call raw_spin_lock_irqsave() which is technically
6975 * unsafe from NMI, but take chance and report kernel bug.
6976 * The sequence of
6977 * kasan_report_invalid_free() -> raw_spin_lock_irqsave() -> NMI
6978 * -> kfree_nolock() -> kasan_report_invalid_free() on the same CPU
6979 * is double buggy and deserves to deadlock.
6980 */
6981 if (kasan_slab_pre_free(s, x))
6982 return;
6983 /*
6984 * memcg, kasan_slab_pre_free are done for 'x'.
6985 * The only thing left is kasan_poison without quarantine,
6986 * since kasan quarantine takes locks and not supported from NMI.
6987 */
6988 kasan_slab_free(s, x, false, false, /* skip quarantine */true);
6989 do_slab_free(s, slab, x, x, 0, _RET_IP_);
6990 }
6991 EXPORT_SYMBOL_GPL(kfree_nolock);
6992
6993 static __always_inline __realloc_size(2) void *
6994 __do_krealloc(const void *p, size_t new_size, unsigned long align, gfp_t flags, int nid)
6995 {
6996 void *ret;
6997 size_t ks = 0;
6998 int orig_size = 0;
6999 struct kmem_cache *s = NULL;
7000
7001 if (unlikely(ZERO_OR_NULL_PTR(p)))
7002 goto alloc_new;
7003
7004 /* Check for double-free. */
7005 if (!kasan_check_byte(p))
7006 return NULL;
7007
7008 if (is_kfence_address(p)) {
7009 ks = orig_size = kfence_ksize(p);
7010 } else {
7011 struct folio *folio;
7012
7013 folio = virt_to_folio(p);
7014 if (unlikely(!folio_test_slab(folio))) {
7015 /* Big kmalloc object */
7016 WARN_ON(folio_size(folio) <= KMALLOC_MAX_CACHE_SIZE);
7017 WARN_ON(p != folio_address(folio));
7018 ks = folio_size(folio);
7019 } else {
7020 s = folio_slab(folio)->slab_cache;
7021 orig_size = get_orig_size(s, (void *)p);
7022 ks = s->object_size;
7023 }
7024 }
7025
7026 /*
7027 * If reallocation is not necessary (e. g. the new size is less
7028 * than the current allocated size), the current allocation will be
7029 * preserved unless __GFP_THISNODE is set. In the latter case a new
7030 * allocation on the requested node will be attempted.
7031 */
7032 if (unlikely(flags & __GFP_THISNODE) && nid != NUMA_NO_NODE &&
7033 nid != page_to_nid(virt_to_page(p)))
7034 goto alloc_new;
7035
7036 /* If the old object doesn't fit, allocate a bigger one */
7037 if (new_size > ks)
7038 goto alloc_new;
7039
7040 /* If the old object doesn't satisfy the new alignment, allocate a new one */
7041 if (!IS_ALIGNED((unsigned long)p, align))
7042 goto alloc_new;
7043
7044 /* Zero out spare memory. */
7045 if (want_init_on_alloc(flags)) {
7046 kasan_disable_current();
7047 if (orig_size && orig_size < new_size)
7048 memset(kasan_reset_tag(p) + orig_size, 0, new_size - orig_size);
7049 else
7050 memset(kasan_reset_tag(p) + new_size, 0, ks - new_size);
7051 kasan_enable_current();
7052 }
7053
7054 /* Setup kmalloc redzone when needed */
7055 if (s && slub_debug_orig_size(s)) {
7056 set_orig_size(s, (void *)p, new_size);
7057 if (s->flags & SLAB_RED_ZONE && new_size < ks)
7058 memset_no_sanitize_memory(kasan_reset_tag(p) + new_size,
7059 SLUB_RED_ACTIVE, ks - new_size);
7060 }
7061
7062 p = kasan_krealloc(p, new_size, flags);
7063 return (void *)p;
7064
7065 alloc_new:
7066 ret = kmalloc_node_track_caller_noprof(new_size, flags, nid, _RET_IP_);
7067 if (ret && p) {
7068 /* Disable KASAN checks as the object's redzone is accessed. */
7069 kasan_disable_current();
7070 memcpy(ret, kasan_reset_tag(p), min(new_size, (size_t)(orig_size ?: ks)));
7071 kasan_enable_current();
7072 }
7073
7074 return ret;
7075 }
7076
7077 /**
7078 * krealloc_node_align - reallocate memory. The contents will remain unchanged.
7079 * @p: object to reallocate memory for.
7080 * @new_size: how many bytes of memory are required.
7081 * @align: desired alignment.
7082 * @flags: the type of memory to allocate.
7083 * @nid: NUMA node or NUMA_NO_NODE
7084 *
7085 * If @p is %NULL, krealloc() behaves exactly like kmalloc(). If @new_size
7086 * is 0 and @p is not a %NULL pointer, the object pointed to is freed.
7087 *
7088 * Only alignments up to those guaranteed by kmalloc() will be honored. Please see
7089 * Documentation/core-api/memory-allocation.rst for more details.
7090 *
7091 * If __GFP_ZERO logic is requested, callers must ensure that, starting with the
7092 * initial memory allocation, every subsequent call to this API for the same
7093 * memory allocation is flagged with __GFP_ZERO. Otherwise, it is possible that
7094 * __GFP_ZERO is not fully honored by this API.
7095 *
7096 * When slub_debug_orig_size() is off, krealloc() only knows about the bucket
7097 * size of an allocation (but not the exact size it was allocated with) and
7098 * hence implements the following semantics for shrinking and growing buffers
7099 * with __GFP_ZERO::
7100 *
7101 * new bucket
7102 * 0 size size
7103 * |--------|----------------|
7104 * | keep | zero |
7105 *
7106 * Otherwise, the original allocation size 'orig_size' could be used to
7107 * precisely clear the requested size, and the new size will also be stored
7108 * as the new 'orig_size'.
7109 *
7110 * In any case, the contents of the object pointed to are preserved up to the
7111 * lesser of the new and old sizes.
7112 *
7113 * Return: pointer to the allocated memory or %NULL in case of error
7114 */
7115 void *krealloc_node_align_noprof(const void *p, size_t new_size, unsigned long align,
7116 gfp_t flags, int nid)
7117 {
7118 void *ret;
7119
7120 if (unlikely(!new_size)) {
7121 kfree(p);
7122 return ZERO_SIZE_PTR;
7123 }
7124
7125 ret = __do_krealloc(p, new_size, align, flags, nid);
7126 if (ret && kasan_reset_tag(p) != kasan_reset_tag(ret))
7127 kfree(p);
7128
7129 return ret;
7130 }
7131 EXPORT_SYMBOL(krealloc_node_align_noprof);
7132
7133 static gfp_t kmalloc_gfp_adjust(gfp_t flags, size_t size)
7134 {
7135 /*
7136 * We want to attempt a large physically contiguous block first because
7137 * it is less likely to fragment multiple larger blocks and therefore
7138 * contribute to a long term fragmentation less than vmalloc fallback.
7139 * However make sure that larger requests are not too disruptive - i.e.
7140 * do not direct reclaim unless physically continuous memory is preferred
7141 * (__GFP_RETRY_MAYFAIL mode). We still kick in kswapd/kcompactd to
7142 * start working in the background
7143 */
7144 if (size > PAGE_SIZE) {
7145 flags |= __GFP_NOWARN;
7146
7147 if (!(flags & __GFP_RETRY_MAYFAIL))
7148 flags &= ~__GFP_DIRECT_RECLAIM;
7149
7150 /* nofail semantic is implemented by the vmalloc fallback */
7151 flags &= ~__GFP_NOFAIL;
7152 }
7153
7154 return flags;
7155 }
7156
7157 /**
7158 * __kvmalloc_node - attempt to allocate physically contiguous memory, but upon
7159 * failure, fall back to non-contiguous (vmalloc) allocation.
7160 * @size: size of the request.
7161 * @b: which set of kmalloc buckets to allocate from.
7162 * @align: desired alignment.
7163 * @flags: gfp mask for the allocation - must be compatible (superset) with GFP_KERNEL.
7164 * @node: numa node to allocate from
7165 *
7166 * Only alignments up to those guaranteed by kmalloc() will be honored. Please see
7167 * Documentation/core-api/memory-allocation.rst for more details.
7168 *
7169 * Uses kmalloc to get the memory but if the allocation fails then falls back
7170 * to the vmalloc allocator. Use kvfree for freeing the memory.
7171 *
7172 * GFP_NOWAIT and GFP_ATOMIC are not supported, neither is the __GFP_NORETRY modifier.
7173 * __GFP_RETRY_MAYFAIL is supported, and it should be used only if kmalloc is
7174 * preferable to the vmalloc fallback, due to visible performance drawbacks.
7175 *
7176 * Return: pointer to the allocated memory of %NULL in case of failure
7177 */
7178 void *__kvmalloc_node_noprof(DECL_BUCKET_PARAMS(size, b), unsigned long align,
7179 gfp_t flags, int node)
7180 {
7181 void *ret;
7182
7183 /*
7184 * It doesn't really make sense to fallback to vmalloc for sub page
7185 * requests
7186 */
7187 ret = __do_kmalloc_node(size, PASS_BUCKET_PARAM(b),
7188 kmalloc_gfp_adjust(flags, size),
7189 node, _RET_IP_);
7190 if (ret || size <= PAGE_SIZE)
7191 return ret;
7192
7193 /* non-sleeping allocations are not supported by vmalloc */
7194 if (!gfpflags_allow_blocking(flags))
7195 return NULL;
7196
7197 /* Don't even allow crazy sizes */
7198 if (unlikely(size > INT_MAX)) {
7199 WARN_ON_ONCE(!(flags & __GFP_NOWARN));
7200 return NULL;
7201 }
7202
7203 /*
7204 * kvmalloc() can always use VM_ALLOW_HUGE_VMAP,
7205 * since the callers already cannot assume anything
7206 * about the resulting pointer, and cannot play
7207 * protection games.
7208 */
7209 return __vmalloc_node_range_noprof(size, align, VMALLOC_START, VMALLOC_END,
7210 flags, PAGE_KERNEL, VM_ALLOW_HUGE_VMAP,
7211 node, __builtin_return_address(0));
7212 }
7213 EXPORT_SYMBOL(__kvmalloc_node_noprof);
7214
7215 /**
7216 * kvfree() - Free memory.
7217 * @addr: Pointer to allocated memory.
7218 *
7219 * kvfree frees memory allocated by any of vmalloc(), kmalloc() or kvmalloc().
7220 * It is slightly more efficient to use kfree() or vfree() if you are certain
7221 * that you know which one to use.
7222 *
7223 * Context: Either preemptible task context or not-NMI interrupt.
7224 */
7225 void kvfree(const void *addr)
7226 {
7227 if (is_vmalloc_addr(addr))
7228 vfree(addr);
7229 else
7230 kfree(addr);
7231 }
7232 EXPORT_SYMBOL(kvfree);
7233
7234 /**
7235 * kvfree_sensitive - Free a data object containing sensitive information.
7236 * @addr: address of the data object to be freed.
7237 * @len: length of the data object.
7238 *
7239 * Use the special memzero_explicit() function to clear the content of a
7240 * kvmalloc'ed object containing sensitive data to make sure that the
7241 * compiler won't optimize out the data clearing.
7242 */
7243 void kvfree_sensitive(const void *addr, size_t len)
7244 {
7245 if (likely(!ZERO_OR_NULL_PTR(addr))) {
7246 memzero_explicit((void *)addr, len);
7247 kvfree(addr);
7248 }
7249 }
7250 EXPORT_SYMBOL(kvfree_sensitive);
7251
7252 /**
7253 * kvrealloc_node_align - reallocate memory; contents remain unchanged
7254 * @p: object to reallocate memory for
7255 * @size: the size to reallocate
7256 * @align: desired alignment
7257 * @flags: the flags for the page level allocator
7258 * @nid: NUMA node id
7259 *
7260 * If @p is %NULL, kvrealloc() behaves exactly like kvmalloc(). If @size is 0
7261 * and @p is not a %NULL pointer, the object pointed to is freed.
7262 *
7263 * Only alignments up to those guaranteed by kmalloc() will be honored. Please see
7264 * Documentation/core-api/memory-allocation.rst for more details.
7265 *
7266 * If __GFP_ZERO logic is requested, callers must ensure that, starting with the
7267 * initial memory allocation, every subsequent call to this API for the same
7268 * memory allocation is flagged with __GFP_ZERO. Otherwise, it is possible that
7269 * __GFP_ZERO is not fully honored by this API.
7270 *
7271 * In any case, the contents of the object pointed to are preserved up to the
7272 * lesser of the new and old sizes.
7273 *
7274 * This function must not be called concurrently with itself or kvfree() for the
7275 * same memory allocation.
7276 *
7277 * Return: pointer to the allocated memory or %NULL in case of error
7278 */
7279 void *kvrealloc_node_align_noprof(const void *p, size_t size, unsigned long align,
7280 gfp_t flags, int nid)
7281 {
7282 void *n;
7283
7284 if (is_vmalloc_addr(p))
7285 return vrealloc_node_align_noprof(p, size, align, flags, nid);
7286
7287 n = krealloc_node_align_noprof(p, size, align, kmalloc_gfp_adjust(flags, size), nid);
7288 if (!n) {
7289 /* We failed to krealloc(), fall back to kvmalloc(). */
7290 n = kvmalloc_node_align_noprof(size, align, flags, nid);
7291 if (!n)
7292 return NULL;
7293
7294 if (p) {
7295 /* We already know that `p` is not a vmalloc address. */
7296 kasan_disable_current();
7297 memcpy(n, kasan_reset_tag(p), min(size, ksize(p)));
7298 kasan_enable_current();
7299
7300 kfree(p);
7301 }
7302 }
7303
7304 return n;
7305 }
7306 EXPORT_SYMBOL(kvrealloc_node_align_noprof);
7307
7308 struct detached_freelist {
7309 struct slab *slab;
7310 void *tail;
7311 void *freelist;
7312 int cnt;
7313 struct kmem_cache *s;
7314 };
7315
7316 /*
7317 * This function progressively scans the array with free objects (with
7318 * a limited look ahead) and extract objects belonging to the same
7319 * slab. It builds a detached freelist directly within the given
7320 * slab/objects. This can happen without any need for
7321 * synchronization, because the objects are owned by running process.
7322 * The freelist is build up as a single linked list in the objects.
7323 * The idea is, that this detached freelist can then be bulk
7324 * transferred to the real freelist(s), but only requiring a single
7325 * synchronization primitive. Look ahead in the array is limited due
7326 * to performance reasons.
7327 */
7328 static inline
7329 int build_detached_freelist(struct kmem_cache *s, size_t size,
7330 void **p, struct detached_freelist *df)
7331 {
7332 int lookahead = 3;
7333 void *object;
7334 struct folio *folio;
7335 size_t same;
7336
7337 object = p[--size];
7338 folio = virt_to_folio(object);
7339 if (!s) {
7340 /* Handle kalloc'ed objects */
7341 if (unlikely(!folio_test_slab(folio))) {
7342 free_large_kmalloc(folio, object);
7343 df->slab = NULL;
7344 return size;
7345 }
7346 /* Derive kmem_cache from object */
7347 df->slab = folio_slab(folio);
7348 df->s = df->slab->slab_cache;
7349 } else {
7350 df->slab = folio_slab(folio);
7351 df->s = cache_from_obj(s, object); /* Support for memcg */
7352 }
7353
7354 /* Start new detached freelist */
7355 df->tail = object;
7356 df->freelist = object;
7357 df->cnt = 1;
7358
7359 if (is_kfence_address(object))
7360 return size;
7361
7362 set_freepointer(df->s, object, NULL);
7363
7364 same = size;
7365 while (size) {
7366 object = p[--size];
7367 /* df->slab is always set at this point */
7368 if (df->slab == virt_to_slab(object)) {
7369 /* Opportunity build freelist */
7370 set_freepointer(df->s, object, df->freelist);
7371 df->freelist = object;
7372 df->cnt++;
7373 same--;
7374 if (size != same)
7375 swap(p[size], p[same]);
7376 continue;
7377 }
7378
7379 /* Limit look ahead search */
7380 if (!--lookahead)
7381 break;
7382 }
7383
7384 return same;
7385 }
7386
7387 /*
7388 * Internal bulk free of objects that were not initialised by the post alloc
7389 * hooks and thus should not be processed by the free hooks
7390 */
7391 static void __kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
7392 {
7393 if (!size)
7394 return;
7395
7396 do {
7397 struct detached_freelist df;
7398
7399 size = build_detached_freelist(s, size, p, &df);
7400 if (!df.slab)
7401 continue;
7402
7403 if (kfence_free(df.freelist))
7404 continue;
7405
7406 do_slab_free(df.s, df.slab, df.freelist, df.tail, df.cnt,
7407 _RET_IP_);
7408 } while (likely(size));
7409 }
7410
7411 /* Note that interrupts must be enabled when calling this function. */
7412 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
7413 {
7414 if (!size)
7415 return;
7416
7417 /*
7418 * freeing to sheaves is so incompatible with the detached freelist so
7419 * once we go that way, we have to do everything differently
7420 */
7421 if (s && s->cpu_sheaves) {
7422 free_to_pcs_bulk(s, size, p);
7423 return;
7424 }
7425
7426 do {
7427 struct detached_freelist df;
7428
7429 size = build_detached_freelist(s, size, p, &df);
7430 if (!df.slab)
7431 continue;
7432
7433 slab_free_bulk(df.s, df.slab, df.freelist, df.tail, &p[size],
7434 df.cnt, _RET_IP_);
7435 } while (likely(size));
7436 }
7437 EXPORT_SYMBOL(kmem_cache_free_bulk);
7438
7439 static inline
7440 int __kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
7441 void **p)
7442 {
7443 struct kmem_cache_cpu *c;
7444 unsigned long irqflags;
7445 int i;
7446
7447 /*
7448 * Drain objects in the per cpu slab, while disabling local
7449 * IRQs, which protects against PREEMPT and interrupts
7450 * handlers invoking normal fastpath.
7451 */
7452 c = slub_get_cpu_ptr(s->cpu_slab);
7453 local_lock_irqsave(&s->cpu_slab->lock, irqflags);
7454
7455 for (i = 0; i < size; i++) {
7456 void *object = kfence_alloc(s, s->object_size, flags);
7457
7458 if (unlikely(object)) {
7459 p[i] = object;
7460 continue;
7461 }
7462
7463 object = c->freelist;
7464 if (unlikely(!object)) {
7465 /*
7466 * We may have removed an object from c->freelist using
7467 * the fastpath in the previous iteration; in that case,
7468 * c->tid has not been bumped yet.
7469 * Since ___slab_alloc() may reenable interrupts while
7470 * allocating memory, we should bump c->tid now.
7471 */
7472 c->tid = next_tid(c->tid);
7473
7474 local_unlock_irqrestore(&s->cpu_slab->lock, irqflags);
7475
7476 /*
7477 * Invoking slow path likely have side-effect
7478 * of re-populating per CPU c->freelist
7479 */
7480 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
7481 _RET_IP_, c, s->object_size);
7482 if (unlikely(!p[i]))
7483 goto error;
7484
7485 c = this_cpu_ptr(s->cpu_slab);
7486 maybe_wipe_obj_freeptr(s, p[i]);
7487
7488 local_lock_irqsave(&s->cpu_slab->lock, irqflags);
7489
7490 continue; /* goto for-loop */
7491 }
7492 c->freelist = get_freepointer(s, object);
7493 p[i] = object;
7494 maybe_wipe_obj_freeptr(s, p[i]);
7495 stat(s, ALLOC_FASTPATH);
7496 }
7497 c->tid = next_tid(c->tid);
7498 local_unlock_irqrestore(&s->cpu_slab->lock, irqflags);
7499 slub_put_cpu_ptr(s->cpu_slab);
7500
7501 return i;
7502
7503 error:
7504 slub_put_cpu_ptr(s->cpu_slab);
7505 __kmem_cache_free_bulk(s, i, p);
7506 return 0;
7507
7508 }
7509
7510 /* Note that interrupts must be enabled when calling this function. */
7511 int kmem_cache_alloc_bulk_noprof(struct kmem_cache *s, gfp_t flags, size_t size,
7512 void **p)
7513 {
7514 unsigned int i = 0;
7515
7516 if (!size)
7517 return 0;
7518
7519 s = slab_pre_alloc_hook(s, flags);
7520 if (unlikely(!s))
7521 return 0;
7522
7523 if (s->cpu_sheaves)
7524 i = alloc_from_pcs_bulk(s, size, p);
7525
7526 if (i < size) {
7527 /*
7528 * If we ran out of memory, don't bother with freeing back to
7529 * the percpu sheaves, we have bigger problems.
7530 */
7531 if (unlikely(__kmem_cache_alloc_bulk(s, flags, size - i, p + i) == 0)) {
7532 if (i > 0)
7533 __kmem_cache_free_bulk(s, i, p);
7534 return 0;
7535 }
7536 }
7537
7538 /*
7539 * memcg and kmem_cache debug support and memory initialization.
7540 * Done outside of the IRQ disabled fastpath loop.
7541 */
7542 if (unlikely(!slab_post_alloc_hook(s, NULL, flags, size, p,
7543 slab_want_init_on_alloc(flags, s), s->object_size))) {
7544 return 0;
7545 }
7546
7547 return size;
7548 }
7549 EXPORT_SYMBOL(kmem_cache_alloc_bulk_noprof);
7550
7551 /*
7552 * Object placement in a slab is made very easy because we always start at
7553 * offset 0. If we tune the size of the object to the alignment then we can
7554 * get the required alignment by putting one properly sized object after
7555 * another.
7556 *
7557 * Notice that the allocation order determines the sizes of the per cpu
7558 * caches. Each processor has always one slab available for allocations.
7559 * Increasing the allocation order reduces the number of times that slabs
7560 * must be moved on and off the partial lists and is therefore a factor in
7561 * locking overhead.
7562 */
7563
7564 /*
7565 * Minimum / Maximum order of slab pages. This influences locking overhead
7566 * and slab fragmentation. A higher order reduces the number of partial slabs
7567 * and increases the number of allocations possible without having to
7568 * take the list_lock.
7569 */
7570 static unsigned int slub_min_order;
7571 static unsigned int slub_max_order =
7572 IS_ENABLED(CONFIG_SLUB_TINY) ? 1 : PAGE_ALLOC_COSTLY_ORDER;
7573 static unsigned int slub_min_objects;
7574
7575 /*
7576 * Calculate the order of allocation given an slab object size.
7577 *
7578 * The order of allocation has significant impact on performance and other
7579 * system components. Generally order 0 allocations should be preferred since
7580 * order 0 does not cause fragmentation in the page allocator. Larger objects
7581 * be problematic to put into order 0 slabs because there may be too much
7582 * unused space left. We go to a higher order if more than 1/16th of the slab
7583 * would be wasted.
7584 *
7585 * In order to reach satisfactory performance we must ensure that a minimum
7586 * number of objects is in one slab. Otherwise we may generate too much
7587 * activity on the partial lists which requires taking the list_lock. This is
7588 * less a concern for large slabs though which are rarely used.
7589 *
7590 * slab_max_order specifies the order where we begin to stop considering the
7591 * number of objects in a slab as critical. If we reach slab_max_order then
7592 * we try to keep the page order as low as possible. So we accept more waste
7593 * of space in favor of a small page order.
7594 *
7595 * Higher order allocations also allow the placement of more objects in a
7596 * slab and thereby reduce object handling overhead. If the user has
7597 * requested a higher minimum order then we start with that one instead of
7598 * the smallest order which will fit the object.
7599 */
7600 static inline unsigned int calc_slab_order(unsigned int size,
7601 unsigned int min_order, unsigned int max_order,
7602 unsigned int fract_leftover)
7603 {
7604 unsigned int order;
7605
7606 for (order = min_order; order <= max_order; order++) {
7607
7608 unsigned int slab_size = (unsigned int)PAGE_SIZE << order;
7609 unsigned int rem;
7610
7611 rem = slab_size % size;
7612
7613 if (rem <= slab_size / fract_leftover)
7614 break;
7615 }
7616
7617 return order;
7618 }
7619
7620 static inline int calculate_order(unsigned int size)
7621 {
7622 unsigned int order;
7623 unsigned int min_objects;
7624 unsigned int max_objects;
7625 unsigned int min_order;
7626
7627 min_objects = slub_min_objects;
7628 if (!min_objects) {
7629 /*
7630 * Some architectures will only update present cpus when
7631 * onlining them, so don't trust the number if it's just 1. But
7632 * we also don't want to use nr_cpu_ids always, as on some other
7633 * architectures, there can be many possible cpus, but never
7634 * onlined. Here we compromise between trying to avoid too high
7635 * order on systems that appear larger than they are, and too
7636 * low order on systems that appear smaller than they are.
7637 */
7638 unsigned int nr_cpus = num_present_cpus();
7639 if (nr_cpus <= 1)
7640 nr_cpus = nr_cpu_ids;
7641 min_objects = 4 * (fls(nr_cpus) + 1);
7642 }
7643 /* min_objects can't be 0 because get_order(0) is undefined */
7644 max_objects = max(order_objects(slub_max_order, size), 1U);
7645 min_objects = min(min_objects, max_objects);
7646
7647 min_order = max_t(unsigned int, slub_min_order,
7648 get_order(min_objects * size));
7649 if (order_objects(min_order, size) > MAX_OBJS_PER_PAGE)
7650 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
7651
7652 /*
7653 * Attempt to find best configuration for a slab. This works by first
7654 * attempting to generate a layout with the best possible configuration
7655 * and backing off gradually.
7656 *
7657 * We start with accepting at most 1/16 waste and try to find the
7658 * smallest order from min_objects-derived/slab_min_order up to
7659 * slab_max_order that will satisfy the constraint. Note that increasing
7660 * the order can only result in same or less fractional waste, not more.
7661 *
7662 * If that fails, we increase the acceptable fraction of waste and try
7663 * again. The last iteration with fraction of 1/2 would effectively
7664 * accept any waste and give us the order determined by min_objects, as
7665 * long as at least single object fits within slab_max_order.
7666 */
7667 for (unsigned int fraction = 16; fraction > 1; fraction /= 2) {
7668 order = calc_slab_order(size, min_order, slub_max_order,
7669 fraction);
7670 if (order <= slub_max_order)
7671 return order;
7672 }
7673
7674 /*
7675 * Doh this slab cannot be placed using slab_max_order.
7676 */
7677 order = get_order(size);
7678 if (order <= MAX_PAGE_ORDER)
7679 return order;
7680 return -ENOSYS;
7681 }
7682
7683 static void
7684 init_kmem_cache_node(struct kmem_cache_node *n, struct node_barn *barn)
7685 {
7686 n->nr_partial = 0;
7687 spin_lock_init(&n->list_lock);
7688 INIT_LIST_HEAD(&n->partial);
7689 #ifdef CONFIG_SLUB_DEBUG
7690 atomic_long_set(&n->nr_slabs, 0);
7691 atomic_long_set(&n->total_objects, 0);
7692 INIT_LIST_HEAD(&n->full);
7693 #endif
7694 n->barn = barn;
7695 if (barn)
7696 barn_init(barn);
7697 }
7698
7699 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
7700 {
7701 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
7702 NR_KMALLOC_TYPES * KMALLOC_SHIFT_HIGH *
7703 sizeof(struct kmem_cache_cpu));
7704
7705 /*
7706 * Must align to double word boundary for the double cmpxchg
7707 * instructions to work; see __pcpu_double_call_return_bool().
7708 */
7709 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
7710 2 * sizeof(void *));
7711
7712 if (!s->cpu_slab)
7713 return 0;
7714
7715 init_kmem_cache_cpus(s);
7716
7717 return 1;
7718 }
7719
7720 static int init_percpu_sheaves(struct kmem_cache *s)
7721 {
7722 int cpu;
7723
7724 for_each_possible_cpu(cpu) {
7725 struct slub_percpu_sheaves *pcs;
7726
7727 pcs = per_cpu_ptr(s->cpu_sheaves, cpu);
7728
7729 local_trylock_init(&pcs->lock);
7730
7731 pcs->main = alloc_empty_sheaf(s, GFP_KERNEL);
7732
7733 if (!pcs->main)
7734 return -ENOMEM;
7735 }
7736
7737 return 0;
7738 }
7739
7740 static struct kmem_cache *kmem_cache_node;
7741
7742 /*
7743 * No kmalloc_node yet so do it by hand. We know that this is the first
7744 * slab on the node for this slabcache. There are no concurrent accesses
7745 * possible.
7746 *
7747 * Note that this function only works on the kmem_cache_node
7748 * when allocating for the kmem_cache_node. This is used for bootstrapping
7749 * memory on a fresh node that has no slab structures yet.
7750 */
7751 static void early_kmem_cache_node_alloc(int node)
7752 {
7753 struct slab *slab;
7754 struct kmem_cache_node *n;
7755
7756 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
7757
7758 slab = new_slab(kmem_cache_node, GFP_NOWAIT, node);
7759
7760 BUG_ON(!slab);
7761 if (slab_nid(slab) != node) {
7762 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
7763 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
7764 }
7765
7766 n = slab->freelist;
7767 BUG_ON(!n);
7768 #ifdef CONFIG_SLUB_DEBUG
7769 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
7770 #endif
7771 n = kasan_slab_alloc(kmem_cache_node, n, GFP_KERNEL, false);
7772 slab->freelist = get_freepointer(kmem_cache_node, n);
7773 slab->inuse = 1;
7774 kmem_cache_node->node[node] = n;
7775 init_kmem_cache_node(n, NULL);
7776 inc_slabs_node(kmem_cache_node, node, slab->objects);
7777
7778 /*
7779 * No locks need to be taken here as it has just been
7780 * initialized and there is no concurrent access.
7781 */
7782 __add_partial(n, slab, DEACTIVATE_TO_HEAD);
7783 }
7784
7785 static void free_kmem_cache_nodes(struct kmem_cache *s)
7786 {
7787 int node;
7788 struct kmem_cache_node *n;
7789
7790 for_each_kmem_cache_node(s, node, n) {
7791 if (n->barn) {
7792 WARN_ON(n->barn->nr_full);
7793 WARN_ON(n->barn->nr_empty);
7794 kfree(n->barn);
7795 n->barn = NULL;
7796 }
7797
7798 s->node[node] = NULL;
7799 kmem_cache_free(kmem_cache_node, n);
7800 }
7801 }
7802
7803 void __kmem_cache_release(struct kmem_cache *s)
7804 {
7805 cache_random_seq_destroy(s);
7806 if (s->cpu_sheaves)
7807 pcs_destroy(s);
7808 #ifdef CONFIG_PREEMPT_RT
7809 if (s->cpu_slab)
7810 lockdep_unregister_key(&s->lock_key);
7811 #endif
7812 free_percpu(s->cpu_slab);
7813 free_kmem_cache_nodes(s);
7814 }
7815
7816 static int init_kmem_cache_nodes(struct kmem_cache *s)
7817 {
7818 int node;
7819
7820 for_each_node_mask(node, slab_nodes) {
7821 struct kmem_cache_node *n;
7822 struct node_barn *barn = NULL;
7823
7824 if (slab_state == DOWN) {
7825 early_kmem_cache_node_alloc(node);
7826 continue;
7827 }
7828
7829 if (s->cpu_sheaves) {
7830 barn = kmalloc_node(sizeof(*barn), GFP_KERNEL, node);
7831
7832 if (!barn)
7833 return 0;
7834 }
7835
7836 n = kmem_cache_alloc_node(kmem_cache_node,
7837 GFP_KERNEL, node);
7838 if (!n) {
7839 kfree(barn);
7840 return 0;
7841 }
7842
7843 init_kmem_cache_node(n, barn);
7844
7845 s->node[node] = n;
7846 }
7847 return 1;
7848 }
7849
7850 static void set_cpu_partial(struct kmem_cache *s)
7851 {
7852 #ifdef CONFIG_SLUB_CPU_PARTIAL
7853 unsigned int nr_objects;
7854
7855 /*
7856 * cpu_partial determined the maximum number of objects kept in the
7857 * per cpu partial lists of a processor.
7858 *
7859 * Per cpu partial lists mainly contain slabs that just have one
7860 * object freed. If they are used for allocation then they can be
7861 * filled up again with minimal effort. The slab will never hit the
7862 * per node partial lists and therefore no locking will be required.
7863 *
7864 * For backwards compatibility reasons, this is determined as number
7865 * of objects, even though we now limit maximum number of pages, see
7866 * slub_set_cpu_partial()
7867 */
7868 if (!kmem_cache_has_cpu_partial(s))
7869 nr_objects = 0;
7870 else if (s->size >= PAGE_SIZE)
7871 nr_objects = 6;
7872 else if (s->size >= 1024)
7873 nr_objects = 24;
7874 else if (s->size >= 256)
7875 nr_objects = 52;
7876 else
7877 nr_objects = 120;
7878
7879 slub_set_cpu_partial(s, nr_objects);
7880 #endif
7881 }
7882
7883 /*
7884 * calculate_sizes() determines the order and the distribution of data within
7885 * a slab object.
7886 */
7887 static int calculate_sizes(struct kmem_cache_args *args, struct kmem_cache *s)
7888 {
7889 slab_flags_t flags = s->flags;
7890 unsigned int size = s->object_size;
7891 unsigned int order;
7892
7893 /*
7894 * Round up object size to the next word boundary. We can only
7895 * place the free pointer at word boundaries and this determines
7896 * the possible location of the free pointer.
7897 */
7898 size = ALIGN(size, sizeof(void *));
7899
7900 #ifdef CONFIG_SLUB_DEBUG
7901 /*
7902 * Determine if we can poison the object itself. If the user of
7903 * the slab may touch the object after free or before allocation
7904 * then we should never poison the object itself.
7905 */
7906 if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) &&
7907 !s->ctor)
7908 s->flags |= __OBJECT_POISON;
7909 else
7910 s->flags &= ~__OBJECT_POISON;
7911
7912
7913 /*
7914 * If we are Redzoning then check if there is some space between the
7915 * end of the object and the free pointer. If not then add an
7916 * additional word to have some bytes to store Redzone information.
7917 */
7918 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
7919 size += sizeof(void *);
7920 #endif
7921
7922 /*
7923 * With that we have determined the number of bytes in actual use
7924 * by the object and redzoning.
7925 */
7926 s->inuse = size;
7927
7928 if (((flags & SLAB_TYPESAFE_BY_RCU) && !args->use_freeptr_offset) ||
7929 (flags & SLAB_POISON) || s->ctor ||
7930 ((flags & SLAB_RED_ZONE) &&
7931 (s->object_size < sizeof(void *) || slub_debug_orig_size(s)))) {
7932 /*
7933 * Relocate free pointer after the object if it is not
7934 * permitted to overwrite the first word of the object on
7935 * kmem_cache_free.
7936 *
7937 * This is the case if we do RCU, have a constructor or
7938 * destructor, are poisoning the objects, or are
7939 * redzoning an object smaller than sizeof(void *) or are
7940 * redzoning an object with slub_debug_orig_size() enabled,
7941 * in which case the right redzone may be extended.
7942 *
7943 * The assumption that s->offset >= s->inuse means free
7944 * pointer is outside of the object is used in the
7945 * freeptr_outside_object() function. If that is no
7946 * longer true, the function needs to be modified.
7947 */
7948 s->offset = size;
7949 size += sizeof(void *);
7950 } else if ((flags & SLAB_TYPESAFE_BY_RCU) && args->use_freeptr_offset) {
7951 s->offset = args->freeptr_offset;
7952 } else {
7953 /*
7954 * Store freelist pointer near middle of object to keep
7955 * it away from the edges of the object to avoid small
7956 * sized over/underflows from neighboring allocations.
7957 */
7958 s->offset = ALIGN_DOWN(s->object_size / 2, sizeof(void *));
7959 }
7960
7961 #ifdef CONFIG_SLUB_DEBUG
7962 if (flags & SLAB_STORE_USER) {
7963 /*
7964 * Need to store information about allocs and frees after
7965 * the object.
7966 */
7967 size += 2 * sizeof(struct track);
7968
7969 /* Save the original kmalloc request size */
7970 if (flags & SLAB_KMALLOC)
7971 size += sizeof(unsigned long);
7972 }
7973 #endif
7974
7975 kasan_cache_create(s, &size, &s->flags);
7976 #ifdef CONFIG_SLUB_DEBUG
7977 if (flags & SLAB_RED_ZONE) {
7978 /*
7979 * Add some empty padding so that we can catch
7980 * overwrites from earlier objects rather than let
7981 * tracking information or the free pointer be
7982 * corrupted if a user writes before the start
7983 * of the object.
7984 */
7985 size += sizeof(void *);
7986
7987 s->red_left_pad = sizeof(void *);
7988 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
7989 size += s->red_left_pad;
7990 }
7991 #endif
7992
7993 /*
7994 * SLUB stores one object immediately after another beginning from
7995 * offset 0. In order to align the objects we have to simply size
7996 * each object to conform to the alignment.
7997 */
7998 size = ALIGN(size, s->align);
7999 s->size = size;
8000 s->reciprocal_size = reciprocal_value(size);
8001 order = calculate_order(size);
8002
8003 if ((int)order < 0)
8004 return 0;
8005
8006 s->allocflags = __GFP_COMP;
8007
8008 if (s->flags & SLAB_CACHE_DMA)
8009 s->allocflags |= GFP_DMA;
8010
8011 if (s->flags & SLAB_CACHE_DMA32)
8012 s->allocflags |= GFP_DMA32;
8013
8014 if (s->flags & SLAB_RECLAIM_ACCOUNT)
8015 s->allocflags |= __GFP_RECLAIMABLE;
8016
8017 /*
8018 * Determine the number of objects per slab
8019 */
8020 s->oo = oo_make(order, size);
8021 s->min = oo_make(get_order(size), size);
8022
8023 return !!oo_objects(s->oo);
8024 }
8025
8026 static void list_slab_objects(struct kmem_cache *s, struct slab *slab)
8027 {
8028 #ifdef CONFIG_SLUB_DEBUG
8029 void *addr = slab_address(slab);
8030 void *p;
8031
8032 if (!slab_add_kunit_errors())
8033 slab_bug(s, "Objects remaining on __kmem_cache_shutdown()");
8034
8035 spin_lock(&object_map_lock);
8036 __fill_map(object_map, s, slab);
8037
8038 for_each_object(p, s, addr, slab->objects) {
8039
8040 if (!test_bit(__obj_to_index(s, addr, p), object_map)) {
8041 if (slab_add_kunit_errors())
8042 continue;
8043 pr_err("Object 0x%p @offset=%tu\n", p, p - addr);
8044 print_tracking(s, p);
8045 }
8046 }
8047 spin_unlock(&object_map_lock);
8048
8049 __slab_err(slab);
8050 #endif
8051 }
8052
8053 /*
8054 * Attempt to free all partial slabs on a node.
8055 * This is called from __kmem_cache_shutdown(). We must take list_lock
8056 * because sysfs file might still access partial list after the shutdowning.
8057 */
8058 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
8059 {
8060 LIST_HEAD(discard);
8061 struct slab *slab, *h;
8062
8063 BUG_ON(irqs_disabled());
8064 spin_lock_irq(&n->list_lock);
8065 list_for_each_entry_safe(slab, h, &n->partial, slab_list) {
8066 if (!slab->inuse) {
8067 remove_partial(n, slab);
8068 list_add(&slab->slab_list, &discard);
8069 } else {
8070 list_slab_objects(s, slab);
8071 }
8072 }
8073 spin_unlock_irq(&n->list_lock);
8074
8075 list_for_each_entry_safe(slab, h, &discard, slab_list)
8076 discard_slab(s, slab);
8077 }
8078
8079 bool __kmem_cache_empty(struct kmem_cache *s)
8080 {
8081 int node;
8082 struct kmem_cache_node *n;
8083
8084 for_each_kmem_cache_node(s, node, n)
8085 if (n->nr_partial || node_nr_slabs(n))
8086 return false;
8087 return true;
8088 }
8089
8090 /*
8091 * Release all resources used by a slab cache.
8092 */
8093 int __kmem_cache_shutdown(struct kmem_cache *s)
8094 {
8095 int node;
8096 struct kmem_cache_node *n;
8097
8098 flush_all_cpus_locked(s);
8099
8100 /* we might have rcu sheaves in flight */
8101 if (s->cpu_sheaves)
8102 rcu_barrier();
8103
8104 /* Attempt to free all objects */
8105 for_each_kmem_cache_node(s, node, n) {
8106 if (n->barn)
8107 barn_shrink(s, n->barn);
8108 free_partial(s, n);
8109 if (n->nr_partial || node_nr_slabs(n))
8110 return 1;
8111 }
8112 return 0;
8113 }
8114
8115 #ifdef CONFIG_PRINTK
8116 void __kmem_obj_info(struct kmem_obj_info *kpp, void *object, struct slab *slab)
8117 {
8118 void *base;
8119 int __maybe_unused i;
8120 unsigned int objnr;
8121 void *objp;
8122 void *objp0;
8123 struct kmem_cache *s = slab->slab_cache;
8124 struct track __maybe_unused *trackp;
8125
8126 kpp->kp_ptr = object;
8127 kpp->kp_slab = slab;
8128 kpp->kp_slab_cache = s;
8129 base = slab_address(slab);
8130 objp0 = kasan_reset_tag(object);
8131 #ifdef CONFIG_SLUB_DEBUG
8132 objp = restore_red_left(s, objp0);
8133 #else
8134 objp = objp0;
8135 #endif
8136 objnr = obj_to_index(s, slab, objp);
8137 kpp->kp_data_offset = (unsigned long)((char *)objp0 - (char *)objp);
8138 objp = base + s->size * objnr;
8139 kpp->kp_objp = objp;
8140 if (WARN_ON_ONCE(objp < base || objp >= base + slab->objects * s->size
8141 || (objp - base) % s->size) ||
8142 !(s->flags & SLAB_STORE_USER))
8143 return;
8144 #ifdef CONFIG_SLUB_DEBUG
8145 objp = fixup_red_left(s, objp);
8146 trackp = get_track(s, objp, TRACK_ALLOC);
8147 kpp->kp_ret = (void *)trackp->addr;
8148 #ifdef CONFIG_STACKDEPOT
8149 {
8150 depot_stack_handle_t handle;
8151 unsigned long *entries;
8152 unsigned int nr_entries;
8153
8154 handle = READ_ONCE(trackp->handle);
8155 if (handle) {
8156 nr_entries = stack_depot_fetch(handle, &entries);
8157 for (i = 0; i < KS_ADDRS_COUNT && i < nr_entries; i++)
8158 kpp->kp_stack[i] = (void *)entries[i];
8159 }
8160
8161 trackp = get_track(s, objp, TRACK_FREE);
8162 handle = READ_ONCE(trackp->handle);
8163 if (handle) {
8164 nr_entries = stack_depot_fetch(handle, &entries);
8165 for (i = 0; i < KS_ADDRS_COUNT && i < nr_entries; i++)
8166 kpp->kp_free_stack[i] = (void *)entries[i];
8167 }
8168 }
8169 #endif
8170 #endif
8171 }
8172 #endif
8173
8174 /********************************************************************
8175 * Kmalloc subsystem
8176 *******************************************************************/
8177
8178 static int __init setup_slub_min_order(char *str)
8179 {
8180 get_option(&str, (int *)&slub_min_order);
8181
8182 if (slub_min_order > slub_max_order)
8183 slub_max_order = slub_min_order;
8184
8185 return 1;
8186 }
8187
8188 __setup("slab_min_order=", setup_slub_min_order);
8189 __setup_param("slub_min_order=", slub_min_order, setup_slub_min_order, 0);
8190
8191
8192 static int __init setup_slub_max_order(char *str)
8193 {
8194 get_option(&str, (int *)&slub_max_order);
8195 slub_max_order = min_t(unsigned int, slub_max_order, MAX_PAGE_ORDER);
8196
8197 if (slub_min_order > slub_max_order)
8198 slub_min_order = slub_max_order;
8199
8200 return 1;
8201 }
8202
8203 __setup("slab_max_order=", setup_slub_max_order);
8204 __setup_param("slub_max_order=", slub_max_order, setup_slub_max_order, 0);
8205
8206 static int __init setup_slub_min_objects(char *str)
8207 {
8208 get_option(&str, (int *)&slub_min_objects);
8209
8210 return 1;
8211 }
8212
8213 __setup("slab_min_objects=", setup_slub_min_objects);
8214 __setup_param("slub_min_objects=", slub_min_objects, setup_slub_min_objects, 0);
8215
8216 #ifdef CONFIG_NUMA
8217 static int __init setup_slab_strict_numa(char *str)
8218 {
8219 if (nr_node_ids > 1) {
8220 static_branch_enable(&strict_numa);
8221 pr_info("SLUB: Strict NUMA enabled.\n");
8222 } else {
8223 pr_warn("slab_strict_numa parameter set on non NUMA system.\n");
8224 }
8225
8226 return 1;
8227 }
8228
8229 __setup("slab_strict_numa", setup_slab_strict_numa);
8230 #endif
8231
8232
8233 #ifdef CONFIG_HARDENED_USERCOPY
8234 /*
8235 * Rejects incorrectly sized objects and objects that are to be copied
8236 * to/from userspace but do not fall entirely within the containing slab
8237 * cache's usercopy region.
8238 *
8239 * Returns NULL if check passes, otherwise const char * to name of cache
8240 * to indicate an error.
8241 */
8242 void __check_heap_object(const void *ptr, unsigned long n,
8243 const struct slab *slab, bool to_user)
8244 {
8245 struct kmem_cache *s;
8246 unsigned int offset;
8247 bool is_kfence = is_kfence_address(ptr);
8248
8249 ptr = kasan_reset_tag(ptr);
8250
8251 /* Find object and usable object size. */
8252 s = slab->slab_cache;
8253
8254 /* Reject impossible pointers. */
8255 if (ptr < slab_address(slab))
8256 usercopy_abort("SLUB object not in SLUB page?!", NULL,
8257 to_user, 0, n);
8258
8259 /* Find offset within object. */
8260 if (is_kfence)
8261 offset = ptr - kfence_object_start(ptr);
8262 else
8263 offset = (ptr - slab_address(slab)) % s->size;
8264
8265 /* Adjust for redzone and reject if within the redzone. */
8266 if (!is_kfence && kmem_cache_debug_flags(s, SLAB_RED_ZONE)) {
8267 if (offset < s->red_left_pad)
8268 usercopy_abort("SLUB object in left red zone",
8269 s->name, to_user, offset, n);
8270 offset -= s->red_left_pad;
8271 }
8272
8273 /* Allow address range falling entirely within usercopy region. */
8274 if (offset >= s->useroffset &&
8275 offset - s->useroffset <= s->usersize &&
8276 n <= s->useroffset - offset + s->usersize)
8277 return;
8278
8279 usercopy_abort("SLUB object", s->name, to_user, offset, n);
8280 }
8281 #endif /* CONFIG_HARDENED_USERCOPY */
8282
8283 #define SHRINK_PROMOTE_MAX 32
8284
8285 /*
8286 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
8287 * up most to the head of the partial lists. New allocations will then
8288 * fill those up and thus they can be removed from the partial lists.
8289 *
8290 * The slabs with the least items are placed last. This results in them
8291 * being allocated from last increasing the chance that the last objects
8292 * are freed in them.
8293 */
8294 static int __kmem_cache_do_shrink(struct kmem_cache *s)
8295 {
8296 int node;
8297 int i;
8298 struct kmem_cache_node *n;
8299 struct slab *slab;
8300 struct slab *t;
8301 struct list_head discard;
8302 struct list_head promote[SHRINK_PROMOTE_MAX];
8303 unsigned long flags;
8304 int ret = 0;
8305
8306 for_each_kmem_cache_node(s, node, n) {
8307 INIT_LIST_HEAD(&discard);
8308 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
8309 INIT_LIST_HEAD(promote + i);
8310
8311 if (n->barn)
8312 barn_shrink(s, n->barn);
8313
8314 spin_lock_irqsave(&n->list_lock, flags);
8315
8316 /*
8317 * Build lists of slabs to discard or promote.
8318 *
8319 * Note that concurrent frees may occur while we hold the
8320 * list_lock. slab->inuse here is the upper limit.
8321 */
8322 list_for_each_entry_safe(slab, t, &n->partial, slab_list) {
8323 int free = slab->objects - slab->inuse;
8324
8325 /* Do not reread slab->inuse */
8326 barrier();
8327
8328 /* We do not keep full slabs on the list */
8329 BUG_ON(free <= 0);
8330
8331 if (free == slab->objects) {
8332 list_move(&slab->slab_list, &discard);
8333 slab_clear_node_partial(slab);
8334 n->nr_partial--;
8335 dec_slabs_node(s, node, slab->objects);
8336 } else if (free <= SHRINK_PROMOTE_MAX)
8337 list_move(&slab->slab_list, promote + free - 1);
8338 }
8339
8340 /*
8341 * Promote the slabs filled up most to the head of the
8342 * partial list.
8343 */
8344 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
8345 list_splice(promote + i, &n->partial);
8346
8347 spin_unlock_irqrestore(&n->list_lock, flags);
8348
8349 /* Release empty slabs */
8350 list_for_each_entry_safe(slab, t, &discard, slab_list)
8351 free_slab(s, slab);
8352
8353 if (node_nr_slabs(n))
8354 ret = 1;
8355 }
8356
8357 return ret;
8358 }
8359
8360 int __kmem_cache_shrink(struct kmem_cache *s)
8361 {
8362 flush_all(s);
8363 return __kmem_cache_do_shrink(s);
8364 }
8365
8366 static int slab_mem_going_offline_callback(void)
8367 {
8368 struct kmem_cache *s;
8369
8370 mutex_lock(&slab_mutex);
8371 list_for_each_entry(s, &slab_caches, list) {
8372 flush_all_cpus_locked(s);
8373 __kmem_cache_do_shrink(s);
8374 }
8375 mutex_unlock(&slab_mutex);
8376
8377 return 0;
8378 }
8379
8380 static int slab_mem_going_online_callback(int nid)
8381 {
8382 struct kmem_cache_node *n;
8383 struct kmem_cache *s;
8384 int ret = 0;
8385
8386 /*
8387 * We are bringing a node online. No memory is available yet. We must
8388 * allocate a kmem_cache_node structure in order to bring the node
8389 * online.
8390 */
8391 mutex_lock(&slab_mutex);
8392 list_for_each_entry(s, &slab_caches, list) {
8393 struct node_barn *barn = NULL;
8394
8395 /*
8396 * The structure may already exist if the node was previously
8397 * onlined and offlined.
8398 */
8399 if (get_node(s, nid))
8400 continue;
8401
8402 if (s->cpu_sheaves) {
8403 barn = kmalloc_node(sizeof(*barn), GFP_KERNEL, nid);
8404
8405 if (!barn) {
8406 ret = -ENOMEM;
8407 goto out;
8408 }
8409 }
8410
8411 /*
8412 * XXX: kmem_cache_alloc_node will fallback to other nodes
8413 * since memory is not yet available from the node that
8414 * is brought up.
8415 */
8416 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
8417 if (!n) {
8418 kfree(barn);
8419 ret = -ENOMEM;
8420 goto out;
8421 }
8422
8423 init_kmem_cache_node(n, barn);
8424
8425 s->node[nid] = n;
8426 }
8427 /*
8428 * Any cache created after this point will also have kmem_cache_node
8429 * initialized for the new node.
8430 */
8431 node_set(nid, slab_nodes);
8432 out:
8433 mutex_unlock(&slab_mutex);
8434 return ret;
8435 }
8436
8437 static int slab_memory_callback(struct notifier_block *self,
8438 unsigned long action, void *arg)
8439 {
8440 struct node_notify *nn = arg;
8441 int nid = nn->nid;
8442 int ret = 0;
8443
8444 switch (action) {
8445 case NODE_ADDING_FIRST_MEMORY:
8446 ret = slab_mem_going_online_callback(nid);
8447 break;
8448 case NODE_REMOVING_LAST_MEMORY:
8449 ret = slab_mem_going_offline_callback();
8450 break;
8451 }
8452 if (ret)
8453 ret = notifier_from_errno(ret);
8454 else
8455 ret = NOTIFY_OK;
8456 return ret;
8457 }
8458
8459 /********************************************************************
8460 * Basic setup of slabs
8461 *******************************************************************/
8462
8463 /*
8464 * Used for early kmem_cache structures that were allocated using
8465 * the page allocator. Allocate them properly then fix up the pointers
8466 * that may be pointing to the wrong kmem_cache structure.
8467 */
8468
8469 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
8470 {
8471 int node;
8472 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
8473 struct kmem_cache_node *n;
8474
8475 memcpy(s, static_cache, kmem_cache->object_size);
8476
8477 /*
8478 * This runs very early, and only the boot processor is supposed to be
8479 * up. Even if it weren't true, IRQs are not up so we couldn't fire
8480 * IPIs around.
8481 */
8482 __flush_cpu_slab(s, smp_processor_id());
8483 for_each_kmem_cache_node(s, node, n) {
8484 struct slab *p;
8485
8486 list_for_each_entry(p, &n->partial, slab_list)
8487 p->slab_cache = s;
8488
8489 #ifdef CONFIG_SLUB_DEBUG
8490 list_for_each_entry(p, &n->full, slab_list)
8491 p->slab_cache = s;
8492 #endif
8493 }
8494 list_add(&s->list, &slab_caches);
8495 return s;
8496 }
8497
8498 void __init kmem_cache_init(void)
8499 {
8500 static __initdata struct kmem_cache boot_kmem_cache,
8501 boot_kmem_cache_node;
8502 int node;
8503
8504 if (debug_guardpage_minorder())
8505 slub_max_order = 0;
8506
8507 /* Inform pointer hashing choice about slub debugging state. */
8508 hash_pointers_finalize(__slub_debug_enabled());
8509
8510 kmem_cache_node = &boot_kmem_cache_node;
8511 kmem_cache = &boot_kmem_cache;
8512
8513 /*
8514 * Initialize the nodemask for which we will allocate per node
8515 * structures. Here we don't need taking slab_mutex yet.
8516 */
8517 for_each_node_state(node, N_MEMORY)
8518 node_set(node, slab_nodes);
8519
8520 create_boot_cache(kmem_cache_node, "kmem_cache_node",
8521 sizeof(struct kmem_cache_node),
8522 SLAB_HWCACHE_ALIGN | SLAB_NO_OBJ_EXT, 0, 0);
8523
8524 hotplug_node_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
8525
8526 /* Able to allocate the per node structures */
8527 slab_state = PARTIAL;
8528
8529 create_boot_cache(kmem_cache, "kmem_cache",
8530 offsetof(struct kmem_cache, node) +
8531 nr_node_ids * sizeof(struct kmem_cache_node *),
8532 SLAB_HWCACHE_ALIGN | SLAB_NO_OBJ_EXT, 0, 0);
8533
8534 kmem_cache = bootstrap(&boot_kmem_cache);
8535 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
8536
8537 /* Now we can use the kmem_cache to allocate kmalloc slabs */
8538 setup_kmalloc_cache_index_table();
8539 create_kmalloc_caches();
8540
8541 /* Setup random freelists for each cache */
8542 init_freelist_randomization();
8543
8544 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL,
8545 slub_cpu_dead);
8546
8547 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n",
8548 cache_line_size(),
8549 slub_min_order, slub_max_order, slub_min_objects,
8550 nr_cpu_ids, nr_node_ids);
8551 }
8552
8553 void __init kmem_cache_init_late(void)
8554 {
8555 flushwq = alloc_workqueue("slub_flushwq", WQ_MEM_RECLAIM, 0);
8556 WARN_ON(!flushwq);
8557 #ifdef CONFIG_SLAB_FREELIST_RANDOM
8558 prandom_init_once(&slab_rnd_state);
8559 #endif
8560 }
8561
8562 struct kmem_cache *
8563 __kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
8564 slab_flags_t flags, void (*ctor)(void *))
8565 {
8566 struct kmem_cache *s;
8567
8568 s = find_mergeable(size, align, flags, name, ctor);
8569 if (s) {
8570 if (sysfs_slab_alias(s, name))
8571 pr_err("SLUB: Unable to add cache alias %s to sysfs\n",
8572 name);
8573
8574 s->refcount++;
8575
8576 /*
8577 * Adjust the object sizes so that we clear
8578 * the complete object on kzalloc.
8579 */
8580 s->object_size = max(s->object_size, size);
8581 s->inuse = max(s->inuse, ALIGN(size, sizeof(void *)));
8582 }
8583
8584 return s;
8585 }
8586
8587 int do_kmem_cache_create(struct kmem_cache *s, const char *name,
8588 unsigned int size, struct kmem_cache_args *args,
8589 slab_flags_t flags)
8590 {
8591 int err = -EINVAL;
8592
8593 s->name = name;
8594 s->size = s->object_size = size;
8595
8596 s->flags = kmem_cache_flags(flags, s->name);
8597 #ifdef CONFIG_SLAB_FREELIST_HARDENED
8598 s->random = get_random_long();
8599 #endif
8600 s->align = args->align;
8601 s->ctor = args->ctor;
8602 #ifdef CONFIG_HARDENED_USERCOPY
8603 s->useroffset = args->useroffset;
8604 s->usersize = args->usersize;
8605 #endif
8606
8607 if (!calculate_sizes(args, s))
8608 goto out;
8609 if (disable_higher_order_debug) {
8610 /*
8611 * Disable debugging flags that store metadata if the min slab
8612 * order increased.
8613 */
8614 if (get_order(s->size) > get_order(s->object_size)) {
8615 s->flags &= ~DEBUG_METADATA_FLAGS;
8616 s->offset = 0;
8617 if (!calculate_sizes(args, s))
8618 goto out;
8619 }
8620 }
8621
8622 #ifdef system_has_freelist_aba
8623 if (system_has_freelist_aba() && !(s->flags & SLAB_NO_CMPXCHG)) {
8624 /* Enable fast mode */
8625 s->flags |= __CMPXCHG_DOUBLE;
8626 }
8627 #endif
8628
8629 /*
8630 * The larger the object size is, the more slabs we want on the partial
8631 * list to avoid pounding the page allocator excessively.
8632 */
8633 s->min_partial = min_t(unsigned long, MAX_PARTIAL, ilog2(s->size) / 2);
8634 s->min_partial = max_t(unsigned long, MIN_PARTIAL, s->min_partial);
8635
8636 set_cpu_partial(s);
8637
8638 if (args->sheaf_capacity && !IS_ENABLED(CONFIG_SLUB_TINY)
8639 && !(s->flags & SLAB_DEBUG_FLAGS)) {
8640 s->cpu_sheaves = alloc_percpu(struct slub_percpu_sheaves);
8641 if (!s->cpu_sheaves) {
8642 err = -ENOMEM;
8643 goto out;
8644 }
8645 // TODO: increase capacity to grow slab_sheaf up to next kmalloc size?
8646 s->sheaf_capacity = args->sheaf_capacity;
8647 }
8648
8649 #ifdef CONFIG_NUMA
8650 s->remote_node_defrag_ratio = 1000;
8651 #endif
8652
8653 /* Initialize the pre-computed randomized freelist if slab is up */
8654 if (slab_state >= UP) {
8655 if (init_cache_random_seq(s))
8656 goto out;
8657 }
8658
8659 if (!init_kmem_cache_nodes(s))
8660 goto out;
8661
8662 if (!alloc_kmem_cache_cpus(s))
8663 goto out;
8664
8665 if (s->cpu_sheaves) {
8666 err = init_percpu_sheaves(s);
8667 if (err)
8668 goto out;
8669 }
8670
8671 err = 0;
8672
8673 /* Mutex is not taken during early boot */
8674 if (slab_state <= UP)
8675 goto out;
8676
8677 /*
8678 * Failing to create sysfs files is not critical to SLUB functionality.
8679 * If it fails, proceed with cache creation without these files.
8680 */
8681 if (sysfs_slab_add(s))
8682 pr_err("SLUB: Unable to add cache %s to sysfs\n", s->name);
8683
8684 if (s->flags & SLAB_STORE_USER)
8685 debugfs_slab_add(s);
8686
8687 out:
8688 if (err)
8689 __kmem_cache_release(s);
8690 return err;
8691 }
8692
8693 #ifdef SLAB_SUPPORTS_SYSFS
8694 static int count_inuse(struct slab *slab)
8695 {
8696 return slab->inuse;
8697 }
8698
8699 static int count_total(struct slab *slab)
8700 {
8701 return slab->objects;
8702 }
8703 #endif
8704
8705 #ifdef CONFIG_SLUB_DEBUG
8706 static void validate_slab(struct kmem_cache *s, struct slab *slab,
8707 unsigned long *obj_map)
8708 {
8709 void *p;
8710 void *addr = slab_address(slab);
8711
8712 if (!validate_slab_ptr(slab)) {
8713 slab_err(s, slab, "Not a valid slab page");
8714 return;
8715 }
8716
8717 if (!check_slab(s, slab) || !on_freelist(s, slab, NULL))
8718 return;
8719
8720 /* Now we know that a valid freelist exists */
8721 __fill_map(obj_map, s, slab);
8722 for_each_object(p, s, addr, slab->objects) {
8723 u8 val = test_bit(__obj_to_index(s, addr, p), obj_map) ?
8724 SLUB_RED_INACTIVE : SLUB_RED_ACTIVE;
8725
8726 if (!check_object(s, slab, p, val))
8727 break;
8728 }
8729 }
8730
8731 static int validate_slab_node(struct kmem_cache *s,
8732 struct kmem_cache_node *n, unsigned long *obj_map)
8733 {
8734 unsigned long count = 0;
8735 struct slab *slab;
8736 unsigned long flags;
8737
8738 spin_lock_irqsave(&n->list_lock, flags);
8739
8740 list_for_each_entry(slab, &n->partial, slab_list) {
8741 validate_slab(s, slab, obj_map);
8742 count++;
8743 }
8744 if (count != n->nr_partial) {
8745 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
8746 s->name, count, n->nr_partial);
8747 slab_add_kunit_errors();
8748 }
8749
8750 if (!(s->flags & SLAB_STORE_USER))
8751 goto out;
8752
8753 list_for_each_entry(slab, &n->full, slab_list) {
8754 validate_slab(s, slab, obj_map);
8755 count++;
8756 }
8757 if (count != node_nr_slabs(n)) {
8758 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
8759 s->name, count, node_nr_slabs(n));
8760 slab_add_kunit_errors();
8761 }
8762
8763 out:
8764 spin_unlock_irqrestore(&n->list_lock, flags);
8765 return count;
8766 }
8767
8768 long validate_slab_cache(struct kmem_cache *s)
8769 {
8770 int node;
8771 unsigned long count = 0;
8772 struct kmem_cache_node *n;
8773 unsigned long *obj_map;
8774
8775 obj_map = bitmap_alloc(oo_objects(s->oo), GFP_KERNEL);
8776 if (!obj_map)
8777 return -ENOMEM;
8778
8779 flush_all(s);
8780 for_each_kmem_cache_node(s, node, n)
8781 count += validate_slab_node(s, n, obj_map);
8782
8783 bitmap_free(obj_map);
8784
8785 return count;
8786 }
8787 EXPORT_SYMBOL(validate_slab_cache);
8788
8789 #ifdef CONFIG_DEBUG_FS
8790 /*
8791 * Generate lists of code addresses where slabcache objects are allocated
8792 * and freed.
8793 */
8794
8795 struct location {
8796 depot_stack_handle_t handle;
8797 unsigned long count;
8798 unsigned long addr;
8799 unsigned long waste;
8800 long long sum_time;
8801 long min_time;
8802 long max_time;
8803 long min_pid;
8804 long max_pid;
8805 DECLARE_BITMAP(cpus, NR_CPUS);
8806 nodemask_t nodes;
8807 };
8808
8809 struct loc_track {
8810 unsigned long max;
8811 unsigned long count;
8812 struct location *loc;
8813 loff_t idx;
8814 };
8815
8816 static struct dentry *slab_debugfs_root;
8817
8818 static void free_loc_track(struct loc_track *t)
8819 {
8820 if (t->max)
8821 free_pages((unsigned long)t->loc,
8822 get_order(sizeof(struct location) * t->max));
8823 }
8824
8825 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
8826 {
8827 struct location *l;
8828 int order;
8829
8830 order = get_order(sizeof(struct location) * max);
8831
8832 l = (void *)__get_free_pages(flags, order);
8833 if (!l)
8834 return 0;
8835
8836 if (t->count) {
8837 memcpy(l, t->loc, sizeof(struct location) * t->count);
8838 free_loc_track(t);
8839 }
8840 t->max = max;
8841 t->loc = l;
8842 return 1;
8843 }
8844
8845 static int add_location(struct loc_track *t, struct kmem_cache *s,
8846 const struct track *track,
8847 unsigned int orig_size)
8848 {
8849 long start, end, pos;
8850 struct location *l;
8851 unsigned long caddr, chandle, cwaste;
8852 unsigned long age = jiffies - track->when;
8853 depot_stack_handle_t handle = 0;
8854 unsigned int waste = s->object_size - orig_size;
8855
8856 #ifdef CONFIG_STACKDEPOT
8857 handle = READ_ONCE(track->handle);
8858 #endif
8859 start = -1;
8860 end = t->count;
8861
8862 for ( ; ; ) {
8863 pos = start + (end - start + 1) / 2;
8864
8865 /*
8866 * There is nothing at "end". If we end up there
8867 * we need to add something to before end.
8868 */
8869 if (pos == end)
8870 break;
8871
8872 l = &t->loc[pos];
8873 caddr = l->addr;
8874 chandle = l->handle;
8875 cwaste = l->waste;
8876 if ((track->addr == caddr) && (handle == chandle) &&
8877 (waste == cwaste)) {
8878
8879 l->count++;
8880 if (track->when) {
8881 l->sum_time += age;
8882 if (age < l->min_time)
8883 l->min_time = age;
8884 if (age > l->max_time)
8885 l->max_time = age;
8886
8887 if (track->pid < l->min_pid)
8888 l->min_pid = track->pid;
8889 if (track->pid > l->max_pid)
8890 l->max_pid = track->pid;
8891
8892 cpumask_set_cpu(track->cpu,
8893 to_cpumask(l->cpus));
8894 }
8895 node_set(page_to_nid(virt_to_page(track)), l->nodes);
8896 return 1;
8897 }
8898
8899 if (track->addr < caddr)
8900 end = pos;
8901 else if (track->addr == caddr && handle < chandle)
8902 end = pos;
8903 else if (track->addr == caddr && handle == chandle &&
8904 waste < cwaste)
8905 end = pos;
8906 else
8907 start = pos;
8908 }
8909
8910 /*
8911 * Not found. Insert new tracking element.
8912 */
8913 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
8914 return 0;
8915
8916 l = t->loc + pos;
8917 if (pos < t->count)
8918 memmove(l + 1, l,
8919 (t->count - pos) * sizeof(struct location));
8920 t->count++;
8921 l->count = 1;
8922 l->addr = track->addr;
8923 l->sum_time = age;
8924 l->min_time = age;
8925 l->max_time = age;
8926 l->min_pid = track->pid;
8927 l->max_pid = track->pid;
8928 l->handle = handle;
8929 l->waste = waste;
8930 cpumask_clear(to_cpumask(l->cpus));
8931 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
8932 nodes_clear(l->nodes);
8933 node_set(page_to_nid(virt_to_page(track)), l->nodes);
8934 return 1;
8935 }
8936
8937 static void process_slab(struct loc_track *t, struct kmem_cache *s,
8938 struct slab *slab, enum track_item alloc,
8939 unsigned long *obj_map)
8940 {
8941 void *addr = slab_address(slab);
8942 bool is_alloc = (alloc == TRACK_ALLOC);
8943 void *p;
8944
8945 __fill_map(obj_map, s, slab);
8946
8947 for_each_object(p, s, addr, slab->objects)
8948 if (!test_bit(__obj_to_index(s, addr, p), obj_map))
8949 add_location(t, s, get_track(s, p, alloc),
8950 is_alloc ? get_orig_size(s, p) :
8951 s->object_size);
8952 }
8953 #endif /* CONFIG_DEBUG_FS */
8954 #endif /* CONFIG_SLUB_DEBUG */
8955
8956 #ifdef SLAB_SUPPORTS_SYSFS
8957 enum slab_stat_type {
8958 SL_ALL, /* All slabs */
8959 SL_PARTIAL, /* Only partially allocated slabs */
8960 SL_CPU, /* Only slabs used for cpu caches */
8961 SL_OBJECTS, /* Determine allocated objects not slabs */
8962 SL_TOTAL /* Determine object capacity not slabs */
8963 };
8964
8965 #define SO_ALL (1 << SL_ALL)
8966 #define SO_PARTIAL (1 << SL_PARTIAL)
8967 #define SO_CPU (1 << SL_CPU)
8968 #define SO_OBJECTS (1 << SL_OBJECTS)
8969 #define SO_TOTAL (1 << SL_TOTAL)
8970
8971 static ssize_t show_slab_objects(struct kmem_cache *s,
8972 char *buf, unsigned long flags)
8973 {
8974 unsigned long total = 0;
8975 int node;
8976 int x;
8977 unsigned long *nodes;
8978 int len = 0;
8979
8980 nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL);
8981 if (!nodes)
8982 return -ENOMEM;
8983
8984 if (flags & SO_CPU) {
8985 int cpu;
8986
8987 for_each_possible_cpu(cpu) {
8988 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
8989 cpu);
8990 int node;
8991 struct slab *slab;
8992
8993 slab = READ_ONCE(c->slab);
8994 if (!slab)
8995 continue;
8996
8997 node = slab_nid(slab);
8998 if (flags & SO_TOTAL)
8999 x = slab->objects;
9000 else if (flags & SO_OBJECTS)
9001 x = slab->inuse;
9002 else
9003 x = 1;
9004
9005 total += x;
9006 nodes[node] += x;
9007
9008 #ifdef CONFIG_SLUB_CPU_PARTIAL
9009 slab = slub_percpu_partial_read_once(c);
9010 if (slab) {
9011 node = slab_nid(slab);
9012 if (flags & SO_TOTAL)
9013 WARN_ON_ONCE(1);
9014 else if (flags & SO_OBJECTS)
9015 WARN_ON_ONCE(1);
9016 else
9017 x = data_race(slab->slabs);
9018 total += x;
9019 nodes[node] += x;
9020 }
9021 #endif
9022 }
9023 }
9024
9025 /*
9026 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
9027 * already held which will conflict with an existing lock order:
9028 *
9029 * mem_hotplug_lock->slab_mutex->kernfs_mutex
9030 *
9031 * We don't really need mem_hotplug_lock (to hold off
9032 * slab_mem_going_offline_callback) here because slab's memory hot
9033 * unplug code doesn't destroy the kmem_cache->node[] data.
9034 */
9035
9036 #ifdef CONFIG_SLUB_DEBUG
9037 if (flags & SO_ALL) {
9038 struct kmem_cache_node *n;
9039
9040 for_each_kmem_cache_node(s, node, n) {
9041
9042 if (flags & SO_TOTAL)
9043 x = node_nr_objs(n);
9044 else if (flags & SO_OBJECTS)
9045 x = node_nr_objs(n) - count_partial(n, count_free);
9046 else
9047 x = node_nr_slabs(n);
9048 total += x;
9049 nodes[node] += x;
9050 }
9051
9052 } else
9053 #endif
9054 if (flags & SO_PARTIAL) {
9055 struct kmem_cache_node *n;
9056
9057 for_each_kmem_cache_node(s, node, n) {
9058 if (flags & SO_TOTAL)
9059 x = count_partial(n, count_total);
9060 else if (flags & SO_OBJECTS)
9061 x = count_partial(n, count_inuse);
9062 else
9063 x = n->nr_partial;
9064 total += x;
9065 nodes[node] += x;
9066 }
9067 }
9068
9069 len += sysfs_emit_at(buf, len, "%lu", total);
9070 #ifdef CONFIG_NUMA
9071 for (node = 0; node < nr_node_ids; node++) {
9072 if (nodes[node])
9073 len += sysfs_emit_at(buf, len, " N%d=%lu",
9074 node, nodes[node]);
9075 }
9076 #endif
9077 len += sysfs_emit_at(buf, len, "\n");
9078 kfree(nodes);
9079
9080 return len;
9081 }
9082
9083 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
9084 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
9085
9086 struct slab_attribute {
9087 struct attribute attr;
9088 ssize_t (*show)(struct kmem_cache *s, char *buf);
9089 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
9090 };
9091
9092 #define SLAB_ATTR_RO(_name) \
9093 static struct slab_attribute _name##_attr = __ATTR_RO_MODE(_name, 0400)
9094
9095 #define SLAB_ATTR(_name) \
9096 static struct slab_attribute _name##_attr = __ATTR_RW_MODE(_name, 0600)
9097
9098 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
9099 {
9100 return sysfs_emit(buf, "%u\n", s->size);
9101 }
9102 SLAB_ATTR_RO(slab_size);
9103
9104 static ssize_t align_show(struct kmem_cache *s, char *buf)
9105 {
9106 return sysfs_emit(buf, "%u\n", s->align);
9107 }
9108 SLAB_ATTR_RO(align);
9109
9110 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
9111 {
9112 return sysfs_emit(buf, "%u\n", s->object_size);
9113 }
9114 SLAB_ATTR_RO(object_size);
9115
9116 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
9117 {
9118 return sysfs_emit(buf, "%u\n", oo_objects(s->oo));
9119 }
9120 SLAB_ATTR_RO(objs_per_slab);
9121
9122 static ssize_t order_show(struct kmem_cache *s, char *buf)
9123 {
9124 return sysfs_emit(buf, "%u\n", oo_order(s->oo));
9125 }
9126 SLAB_ATTR_RO(order);
9127
9128 static ssize_t sheaf_capacity_show(struct kmem_cache *s, char *buf)
9129 {
9130 return sysfs_emit(buf, "%u\n", s->sheaf_capacity);
9131 }
9132 SLAB_ATTR_RO(sheaf_capacity);
9133
9134 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
9135 {
9136 return sysfs_emit(buf, "%lu\n", s->min_partial);
9137 }
9138
9139 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
9140 size_t length)
9141 {
9142 unsigned long min;
9143 int err;
9144
9145 err = kstrtoul(buf, 10, &min);
9146 if (err)
9147 return err;
9148
9149 s->min_partial = min;
9150 return length;
9151 }
9152 SLAB_ATTR(min_partial);
9153
9154 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
9155 {
9156 unsigned int nr_partial = 0;
9157 #ifdef CONFIG_SLUB_CPU_PARTIAL
9158 nr_partial = s->cpu_partial;
9159 #endif
9160
9161 return sysfs_emit(buf, "%u\n", nr_partial);
9162 }
9163
9164 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
9165 size_t length)
9166 {
9167 unsigned int objects;
9168 int err;
9169
9170 err = kstrtouint(buf, 10, &objects);
9171 if (err)
9172 return err;
9173 if (objects && !kmem_cache_has_cpu_partial(s))
9174 return -EINVAL;
9175
9176 slub_set_cpu_partial(s, objects);
9177 flush_all(s);
9178 return length;
9179 }
9180 SLAB_ATTR(cpu_partial);
9181
9182 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
9183 {
9184 if (!s->ctor)
9185 return 0;
9186 return sysfs_emit(buf, "%pS\n", s->ctor);
9187 }
9188 SLAB_ATTR_RO(ctor);
9189
9190 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
9191 {
9192 return sysfs_emit(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
9193 }
9194 SLAB_ATTR_RO(aliases);
9195
9196 static ssize_t partial_show(struct kmem_cache *s, char *buf)
9197 {
9198 return show_slab_objects(s, buf, SO_PARTIAL);
9199 }
9200 SLAB_ATTR_RO(partial);
9201
9202 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
9203 {
9204 return show_slab_objects(s, buf, SO_CPU);
9205 }
9206 SLAB_ATTR_RO(cpu_slabs);
9207
9208 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
9209 {
9210 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
9211 }
9212 SLAB_ATTR_RO(objects_partial);
9213
9214 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
9215 {
9216 int objects = 0;
9217 int slabs = 0;
9218 int cpu __maybe_unused;
9219 int len = 0;
9220
9221 #ifdef CONFIG_SLUB_CPU_PARTIAL
9222 for_each_online_cpu(cpu) {
9223 struct slab *slab;
9224
9225 slab = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
9226
9227 if (slab)
9228 slabs += data_race(slab->slabs);
9229 }
9230 #endif
9231
9232 /* Approximate half-full slabs, see slub_set_cpu_partial() */
9233 objects = (slabs * oo_objects(s->oo)) / 2;
9234 len += sysfs_emit_at(buf, len, "%d(%d)", objects, slabs);
9235
9236 #ifdef CONFIG_SLUB_CPU_PARTIAL
9237 for_each_online_cpu(cpu) {
9238 struct slab *slab;
9239
9240 slab = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
9241 if (slab) {
9242 slabs = data_race(slab->slabs);
9243 objects = (slabs * oo_objects(s->oo)) / 2;
9244 len += sysfs_emit_at(buf, len, " C%d=%d(%d)",
9245 cpu, objects, slabs);
9246 }
9247 }
9248 #endif
9249 len += sysfs_emit_at(buf, len, "\n");
9250
9251 return len;
9252 }
9253 SLAB_ATTR_RO(slabs_cpu_partial);
9254
9255 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
9256 {
9257 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
9258 }
9259 SLAB_ATTR_RO(reclaim_account);
9260
9261 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
9262 {
9263 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
9264 }
9265 SLAB_ATTR_RO(hwcache_align);
9266
9267 #ifdef CONFIG_ZONE_DMA
9268 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
9269 {
9270 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
9271 }
9272 SLAB_ATTR_RO(cache_dma);
9273 #endif
9274
9275 #ifdef CONFIG_HARDENED_USERCOPY
9276 static ssize_t usersize_show(struct kmem_cache *s, char *buf)
9277 {
9278 return sysfs_emit(buf, "%u\n", s->usersize);
9279 }
9280 SLAB_ATTR_RO(usersize);
9281 #endif
9282
9283 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
9284 {
9285 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU));
9286 }
9287 SLAB_ATTR_RO(destroy_by_rcu);
9288
9289 #ifdef CONFIG_SLUB_DEBUG
9290 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
9291 {
9292 return show_slab_objects(s, buf, SO_ALL);
9293 }
9294 SLAB_ATTR_RO(slabs);
9295
9296 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
9297 {
9298 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
9299 }
9300 SLAB_ATTR_RO(total_objects);
9301
9302 static ssize_t objects_show(struct kmem_cache *s, char *buf)
9303 {
9304 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
9305 }
9306 SLAB_ATTR_RO(objects);
9307
9308 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
9309 {
9310 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
9311 }
9312 SLAB_ATTR_RO(sanity_checks);
9313
9314 static ssize_t trace_show(struct kmem_cache *s, char *buf)
9315 {
9316 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TRACE));
9317 }
9318 SLAB_ATTR_RO(trace);
9319
9320 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
9321 {
9322 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
9323 }
9324
9325 SLAB_ATTR_RO(red_zone);
9326
9327 static ssize_t poison_show(struct kmem_cache *s, char *buf)
9328 {
9329 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_POISON));
9330 }
9331
9332 SLAB_ATTR_RO(poison);
9333
9334 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
9335 {
9336 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
9337 }
9338
9339 SLAB_ATTR_RO(store_user);
9340
9341 static ssize_t validate_show(struct kmem_cache *s, char *buf)
9342 {
9343 return 0;
9344 }
9345
9346 static ssize_t validate_store(struct kmem_cache *s,
9347 const char *buf, size_t length)
9348 {
9349 int ret = -EINVAL;
9350
9351 if (buf[0] == '1' && kmem_cache_debug(s)) {
9352 ret = validate_slab_cache(s);
9353 if (ret >= 0)
9354 ret = length;
9355 }
9356 return ret;
9357 }
9358 SLAB_ATTR(validate);
9359
9360 #endif /* CONFIG_SLUB_DEBUG */
9361
9362 #ifdef CONFIG_FAILSLAB
9363 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
9364 {
9365 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
9366 }
9367
9368 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
9369 size_t length)
9370 {
9371 if (s->refcount > 1)
9372 return -EINVAL;
9373
9374 if (buf[0] == '1')
9375 WRITE_ONCE(s->flags, s->flags | SLAB_FAILSLAB);
9376 else
9377 WRITE_ONCE(s->flags, s->flags & ~SLAB_FAILSLAB);
9378
9379 return length;
9380 }
9381 SLAB_ATTR(failslab);
9382 #endif
9383
9384 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
9385 {
9386 return 0;
9387 }
9388
9389 static ssize_t shrink_store(struct kmem_cache *s,
9390 const char *buf, size_t length)
9391 {
9392 if (buf[0] == '1')
9393 kmem_cache_shrink(s);
9394 else
9395 return -EINVAL;
9396 return length;
9397 }
9398 SLAB_ATTR(shrink);
9399
9400 #ifdef CONFIG_NUMA
9401 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
9402 {
9403 return sysfs_emit(buf, "%u\n", s->remote_node_defrag_ratio / 10);
9404 }
9405
9406 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
9407 const char *buf, size_t length)
9408 {
9409 unsigned int ratio;
9410 int err;
9411
9412 err = kstrtouint(buf, 10, &ratio);
9413 if (err)
9414 return err;
9415 if (ratio > 100)
9416 return -ERANGE;
9417
9418 s->remote_node_defrag_ratio = ratio * 10;
9419
9420 return length;
9421 }
9422 SLAB_ATTR(remote_node_defrag_ratio);
9423 #endif
9424
9425 #ifdef CONFIG_SLUB_STATS
9426 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
9427 {
9428 unsigned long sum = 0;
9429 int cpu;
9430 int len = 0;
9431 int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL);
9432
9433 if (!data)
9434 return -ENOMEM;
9435
9436 for_each_online_cpu(cpu) {
9437 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
9438
9439 data[cpu] = x;
9440 sum += x;
9441 }
9442
9443 len += sysfs_emit_at(buf, len, "%lu", sum);
9444
9445 #ifdef CONFIG_SMP
9446 for_each_online_cpu(cpu) {
9447 if (data[cpu])
9448 len += sysfs_emit_at(buf, len, " C%d=%u",
9449 cpu, data[cpu]);
9450 }
9451 #endif
9452 kfree(data);
9453 len += sysfs_emit_at(buf, len, "\n");
9454
9455 return len;
9456 }
9457
9458 static void clear_stat(struct kmem_cache *s, enum stat_item si)
9459 {
9460 int cpu;
9461
9462 for_each_online_cpu(cpu)
9463 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
9464 }
9465
9466 #define STAT_ATTR(si, text) \
9467 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
9468 { \
9469 return show_stat(s, buf, si); \
9470 } \
9471 static ssize_t text##_store(struct kmem_cache *s, \
9472 const char *buf, size_t length) \
9473 { \
9474 if (buf[0] != '0') \
9475 return -EINVAL; \
9476 clear_stat(s, si); \
9477 return length; \
9478 } \
9479 SLAB_ATTR(text); \
9480
9481 STAT_ATTR(ALLOC_PCS, alloc_cpu_sheaf);
9482 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
9483 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
9484 STAT_ATTR(FREE_PCS, free_cpu_sheaf);
9485 STAT_ATTR(FREE_RCU_SHEAF, free_rcu_sheaf);
9486 STAT_ATTR(FREE_RCU_SHEAF_FAIL, free_rcu_sheaf_fail);
9487 STAT_ATTR(FREE_FASTPATH, free_fastpath);
9488 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
9489 STAT_ATTR(FREE_FROZEN, free_frozen);
9490 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
9491 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
9492 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
9493 STAT_ATTR(ALLOC_SLAB, alloc_slab);
9494 STAT_ATTR(ALLOC_REFILL, alloc_refill);
9495 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
9496 STAT_ATTR(FREE_SLAB, free_slab);
9497 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
9498 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
9499 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
9500 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
9501 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
9502 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
9503 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
9504 STAT_ATTR(ORDER_FALLBACK, order_fallback);
9505 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
9506 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
9507 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
9508 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
9509 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
9510 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
9511 STAT_ATTR(SHEAF_FLUSH, sheaf_flush);
9512 STAT_ATTR(SHEAF_REFILL, sheaf_refill);
9513 STAT_ATTR(SHEAF_ALLOC, sheaf_alloc);
9514 STAT_ATTR(SHEAF_FREE, sheaf_free);
9515 STAT_ATTR(BARN_GET, barn_get);
9516 STAT_ATTR(BARN_GET_FAIL, barn_get_fail);
9517 STAT_ATTR(BARN_PUT, barn_put);
9518 STAT_ATTR(BARN_PUT_FAIL, barn_put_fail);
9519 STAT_ATTR(SHEAF_PREFILL_FAST, sheaf_prefill_fast);
9520 STAT_ATTR(SHEAF_PREFILL_SLOW, sheaf_prefill_slow);
9521 STAT_ATTR(SHEAF_PREFILL_OVERSIZE, sheaf_prefill_oversize);
9522 STAT_ATTR(SHEAF_RETURN_FAST, sheaf_return_fast);
9523 STAT_ATTR(SHEAF_RETURN_SLOW, sheaf_return_slow);
9524 #endif /* CONFIG_SLUB_STATS */
9525
9526 #ifdef CONFIG_KFENCE
9527 static ssize_t skip_kfence_show(struct kmem_cache *s, char *buf)
9528 {
9529 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_SKIP_KFENCE));
9530 }
9531
9532 static ssize_t skip_kfence_store(struct kmem_cache *s,
9533 const char *buf, size_t length)
9534 {
9535 int ret = length;
9536
9537 if (buf[0] == '0')
9538 s->flags &= ~SLAB_SKIP_KFENCE;
9539 else if (buf[0] == '1')
9540 s->flags |= SLAB_SKIP_KFENCE;
9541 else
9542 ret = -EINVAL;
9543
9544 return ret;
9545 }
9546 SLAB_ATTR(skip_kfence);
9547 #endif
9548
9549 static struct attribute *slab_attrs[] = {
9550 &slab_size_attr.attr,
9551 &object_size_attr.attr,
9552 &objs_per_slab_attr.attr,
9553 &order_attr.attr,
9554 &sheaf_capacity_attr.attr,
9555 &min_partial_attr.attr,
9556 &cpu_partial_attr.attr,
9557 &objects_partial_attr.attr,
9558 &partial_attr.attr,
9559 &cpu_slabs_attr.attr,
9560 &ctor_attr.attr,
9561 &aliases_attr.attr,
9562 &align_attr.attr,
9563 &hwcache_align_attr.attr,
9564 &reclaim_account_attr.attr,
9565 &destroy_by_rcu_attr.attr,
9566 &shrink_attr.attr,
9567 &slabs_cpu_partial_attr.attr,
9568 #ifdef CONFIG_SLUB_DEBUG
9569 &total_objects_attr.attr,
9570 &objects_attr.attr,
9571 &slabs_attr.attr,
9572 &sanity_checks_attr.attr,
9573 &trace_attr.attr,
9574 &red_zone_attr.attr,
9575 &poison_attr.attr,
9576 &store_user_attr.attr,
9577 &validate_attr.attr,
9578 #endif
9579 #ifdef CONFIG_ZONE_DMA
9580 &cache_dma_attr.attr,
9581 #endif
9582 #ifdef CONFIG_NUMA
9583 &remote_node_defrag_ratio_attr.attr,
9584 #endif
9585 #ifdef CONFIG_SLUB_STATS
9586 &alloc_cpu_sheaf_attr.attr,
9587 &alloc_fastpath_attr.attr,
9588 &alloc_slowpath_attr.attr,
9589 &free_cpu_sheaf_attr.attr,
9590 &free_rcu_sheaf_attr.attr,
9591 &free_rcu_sheaf_fail_attr.attr,
9592 &free_fastpath_attr.attr,
9593 &free_slowpath_attr.attr,
9594 &free_frozen_attr.attr,
9595 &free_add_partial_attr.attr,
9596 &free_remove_partial_attr.attr,
9597 &alloc_from_partial_attr.attr,
9598 &alloc_slab_attr.attr,
9599 &alloc_refill_attr.attr,
9600 &alloc_node_mismatch_attr.attr,
9601 &free_slab_attr.attr,
9602 &cpuslab_flush_attr.attr,
9603 &deactivate_full_attr.attr,
9604 &deactivate_empty_attr.attr,
9605 &deactivate_to_head_attr.attr,
9606 &deactivate_to_tail_attr.attr,
9607 &deactivate_remote_frees_attr.attr,
9608 &deactivate_bypass_attr.attr,
9609 &order_fallback_attr.attr,
9610 &cmpxchg_double_fail_attr.attr,
9611 &cmpxchg_double_cpu_fail_attr.attr,
9612 &cpu_partial_alloc_attr.attr,
9613 &cpu_partial_free_attr.attr,
9614 &cpu_partial_node_attr.attr,
9615 &cpu_partial_drain_attr.attr,
9616 &sheaf_flush_attr.attr,
9617 &sheaf_refill_attr.attr,
9618 &sheaf_alloc_attr.attr,
9619 &sheaf_free_attr.attr,
9620 &barn_get_attr.attr,
9621 &barn_get_fail_attr.attr,
9622 &barn_put_attr.attr,
9623 &barn_put_fail_attr.attr,
9624 &sheaf_prefill_fast_attr.attr,
9625 &sheaf_prefill_slow_attr.attr,
9626 &sheaf_prefill_oversize_attr.attr,
9627 &sheaf_return_fast_attr.attr,
9628 &sheaf_return_slow_attr.attr,
9629 #endif
9630 #ifdef CONFIG_FAILSLAB
9631 &failslab_attr.attr,
9632 #endif
9633 #ifdef CONFIG_HARDENED_USERCOPY
9634 &usersize_attr.attr,
9635 #endif
9636 #ifdef CONFIG_KFENCE
9637 &skip_kfence_attr.attr,
9638 #endif
9639
9640 NULL
9641 };
9642
9643 static const struct attribute_group slab_attr_group = {
9644 .attrs = slab_attrs,
9645 };
9646
9647 static ssize_t slab_attr_show(struct kobject *kobj,
9648 struct attribute *attr,
9649 char *buf)
9650 {
9651 struct slab_attribute *attribute;
9652 struct kmem_cache *s;
9653
9654 attribute = to_slab_attr(attr);
9655 s = to_slab(kobj);
9656
9657 if (!attribute->show)
9658 return -EIO;
9659
9660 return attribute->show(s, buf);
9661 }
9662
9663 static ssize_t slab_attr_store(struct kobject *kobj,
9664 struct attribute *attr,
9665 const char *buf, size_t len)
9666 {
9667 struct slab_attribute *attribute;
9668 struct kmem_cache *s;
9669
9670 attribute = to_slab_attr(attr);
9671 s = to_slab(kobj);
9672
9673 if (!attribute->store)
9674 return -EIO;
9675
9676 return attribute->store(s, buf, len);
9677 }
9678
9679 static void kmem_cache_release(struct kobject *k)
9680 {
9681 slab_kmem_cache_release(to_slab(k));
9682 }
9683
9684 static const struct sysfs_ops slab_sysfs_ops = {
9685 .show = slab_attr_show,
9686 .store = slab_attr_store,
9687 };
9688
9689 static const struct kobj_type slab_ktype = {
9690 .sysfs_ops = &slab_sysfs_ops,
9691 .release = kmem_cache_release,
9692 };
9693
9694 static struct kset *slab_kset;
9695
9696 static inline struct kset *cache_kset(struct kmem_cache *s)
9697 {
9698 return slab_kset;
9699 }
9700
9701 #define ID_STR_LENGTH 32
9702
9703 /* Create a unique string id for a slab cache:
9704 *
9705 * Format :[flags-]size
9706 */
9707 static char *create_unique_id(struct kmem_cache *s)
9708 {
9709 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
9710 char *p = name;
9711
9712 if (!name)
9713 return ERR_PTR(-ENOMEM);
9714
9715 *p++ = ':';
9716 /*
9717 * First flags affecting slabcache operations. We will only
9718 * get here for aliasable slabs so we do not need to support
9719 * too many flags. The flags here must cover all flags that
9720 * are matched during merging to guarantee that the id is
9721 * unique.
9722 */
9723 if (s->flags & SLAB_CACHE_DMA)
9724 *p++ = 'd';
9725 if (s->flags & SLAB_CACHE_DMA32)
9726 *p++ = 'D';
9727 if (s->flags & SLAB_RECLAIM_ACCOUNT)
9728 *p++ = 'a';
9729 if (s->flags & SLAB_CONSISTENCY_CHECKS)
9730 *p++ = 'F';
9731 if (s->flags & SLAB_ACCOUNT)
9732 *p++ = 'A';
9733 if (p != name + 1)
9734 *p++ = '-';
9735 p += snprintf(p, ID_STR_LENGTH - (p - name), "%07u", s->size);
9736
9737 if (WARN_ON(p > name + ID_STR_LENGTH - 1)) {
9738 kfree(name);
9739 return ERR_PTR(-EINVAL);
9740 }
9741 kmsan_unpoison_memory(name, p - name);
9742 return name;
9743 }
9744
9745 static int sysfs_slab_add(struct kmem_cache *s)
9746 {
9747 int err;
9748 const char *name;
9749 struct kset *kset = cache_kset(s);
9750 int unmergeable = slab_unmergeable(s);
9751
9752 if (!unmergeable && disable_higher_order_debug &&
9753 (slub_debug & DEBUG_METADATA_FLAGS))
9754 unmergeable = 1;
9755
9756 if (unmergeable) {
9757 /*
9758 * Slabcache can never be merged so we can use the name proper.
9759 * This is typically the case for debug situations. In that
9760 * case we can catch duplicate names easily.
9761 */
9762 sysfs_remove_link(&slab_kset->kobj, s->name);
9763 name = s->name;
9764 } else {
9765 /*
9766 * Create a unique name for the slab as a target
9767 * for the symlinks.
9768 */
9769 name = create_unique_id(s);
9770 if (IS_ERR(name))
9771 return PTR_ERR(name);
9772 }
9773
9774 s->kobj.kset = kset;
9775 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
9776 if (err)
9777 goto out;
9778
9779 err = sysfs_create_group(&s->kobj, &slab_attr_group);
9780 if (err)
9781 goto out_del_kobj;
9782
9783 if (!unmergeable) {
9784 /* Setup first alias */
9785 sysfs_slab_alias(s, s->name);
9786 }
9787 out:
9788 if (!unmergeable)
9789 kfree(name);
9790 return err;
9791 out_del_kobj:
9792 kobject_del(&s->kobj);
9793 goto out;
9794 }
9795
9796 void sysfs_slab_unlink(struct kmem_cache *s)
9797 {
9798 if (s->kobj.state_in_sysfs)
9799 kobject_del(&s->kobj);
9800 }
9801
9802 void sysfs_slab_release(struct kmem_cache *s)
9803 {
9804 kobject_put(&s->kobj);
9805 }
9806
9807 /*
9808 * Need to buffer aliases during bootup until sysfs becomes
9809 * available lest we lose that information.
9810 */
9811 struct saved_alias {
9812 struct kmem_cache *s;
9813 const char *name;
9814 struct saved_alias *next;
9815 };
9816
9817 static struct saved_alias *alias_list;
9818
9819 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
9820 {
9821 struct saved_alias *al;
9822
9823 if (slab_state == FULL) {
9824 /*
9825 * If we have a leftover link then remove it.
9826 */
9827 sysfs_remove_link(&slab_kset->kobj, name);
9828 /*
9829 * The original cache may have failed to generate sysfs file.
9830 * In that case, sysfs_create_link() returns -ENOENT and
9831 * symbolic link creation is skipped.
9832 */
9833 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
9834 }
9835
9836 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
9837 if (!al)
9838 return -ENOMEM;
9839
9840 al->s = s;
9841 al->name = name;
9842 al->next = alias_list;
9843 alias_list = al;
9844 kmsan_unpoison_memory(al, sizeof(*al));
9845 return 0;
9846 }
9847
9848 static int __init slab_sysfs_init(void)
9849 {
9850 struct kmem_cache *s;
9851 int err;
9852
9853 mutex_lock(&slab_mutex);
9854
9855 slab_kset = kset_create_and_add("slab", NULL, kernel_kobj);
9856 if (!slab_kset) {
9857 mutex_unlock(&slab_mutex);
9858 pr_err("Cannot register slab subsystem.\n");
9859 return -ENOMEM;
9860 }
9861
9862 slab_state = FULL;
9863
9864 list_for_each_entry(s, &slab_caches, list) {
9865 err = sysfs_slab_add(s);
9866 if (err)
9867 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
9868 s->name);
9869 }
9870
9871 while (alias_list) {
9872 struct saved_alias *al = alias_list;
9873
9874 alias_list = alias_list->next;
9875 err = sysfs_slab_alias(al->s, al->name);
9876 if (err)
9877 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
9878 al->name);
9879 kfree(al);
9880 }
9881
9882 mutex_unlock(&slab_mutex);
9883 return 0;
9884 }
9885 late_initcall(slab_sysfs_init);
9886 #endif /* SLAB_SUPPORTS_SYSFS */
9887
9888 #if defined(CONFIG_SLUB_DEBUG) && defined(CONFIG_DEBUG_FS)
9889 static int slab_debugfs_show(struct seq_file *seq, void *v)
9890 {
9891 struct loc_track *t = seq->private;
9892 struct location *l;
9893 unsigned long idx;
9894
9895 idx = (unsigned long) t->idx;
9896 if (idx < t->count) {
9897 l = &t->loc[idx];
9898
9899 seq_printf(seq, "%7ld ", l->count);
9900
9901 if (l->addr)
9902 seq_printf(seq, "%pS", (void *)l->addr);
9903 else
9904 seq_puts(seq, "<not-available>");
9905
9906 if (l->waste)
9907 seq_printf(seq, " waste=%lu/%lu",
9908 l->count * l->waste, l->waste);
9909
9910 if (l->sum_time != l->min_time) {
9911 seq_printf(seq, " age=%ld/%llu/%ld",
9912 l->min_time, div_u64(l->sum_time, l->count),
9913 l->max_time);
9914 } else
9915 seq_printf(seq, " age=%ld", l->min_time);
9916
9917 if (l->min_pid != l->max_pid)
9918 seq_printf(seq, " pid=%ld-%ld", l->min_pid, l->max_pid);
9919 else
9920 seq_printf(seq, " pid=%ld",
9921 l->min_pid);
9922
9923 if (num_online_cpus() > 1 && !cpumask_empty(to_cpumask(l->cpus)))
9924 seq_printf(seq, " cpus=%*pbl",
9925 cpumask_pr_args(to_cpumask(l->cpus)));
9926
9927 if (nr_online_nodes > 1 && !nodes_empty(l->nodes))
9928 seq_printf(seq, " nodes=%*pbl",
9929 nodemask_pr_args(&l->nodes));
9930
9931 #ifdef CONFIG_STACKDEPOT
9932 {
9933 depot_stack_handle_t handle;
9934 unsigned long *entries;
9935 unsigned int nr_entries, j;
9936
9937 handle = READ_ONCE(l->handle);
9938 if (handle) {
9939 nr_entries = stack_depot_fetch(handle, &entries);
9940 seq_puts(seq, "\n");
9941 for (j = 0; j < nr_entries; j++)
9942 seq_printf(seq, " %pS\n", (void *)entries[j]);
9943 }
9944 }
9945 #endif
9946 seq_puts(seq, "\n");
9947 }
9948
9949 if (!idx && !t->count)
9950 seq_puts(seq, "No data\n");
9951
9952 return 0;
9953 }
9954
9955 static void slab_debugfs_stop(struct seq_file *seq, void *v)
9956 {
9957 }
9958
9959 static void *slab_debugfs_next(struct seq_file *seq, void *v, loff_t *ppos)
9960 {
9961 struct loc_track *t = seq->private;
9962
9963 t->idx = ++(*ppos);
9964 if (*ppos <= t->count)
9965 return ppos;
9966
9967 return NULL;
9968 }
9969
9970 static int cmp_loc_by_count(const void *a, const void *b)
9971 {
9972 struct location *loc1 = (struct location *)a;
9973 struct location *loc2 = (struct location *)b;
9974
9975 return cmp_int(loc2->count, loc1->count);
9976 }
9977
9978 static void *slab_debugfs_start(struct seq_file *seq, loff_t *ppos)
9979 {
9980 struct loc_track *t = seq->private;
9981
9982 t->idx = *ppos;
9983 return ppos;
9984 }
9985
9986 static const struct seq_operations slab_debugfs_sops = {
9987 .start = slab_debugfs_start,
9988 .next = slab_debugfs_next,
9989 .stop = slab_debugfs_stop,
9990 .show = slab_debugfs_show,
9991 };
9992
9993 static int slab_debug_trace_open(struct inode *inode, struct file *filep)
9994 {
9995
9996 struct kmem_cache_node *n;
9997 enum track_item alloc;
9998 int node;
9999 struct loc_track *t = __seq_open_private(filep, &slab_debugfs_sops,
10000 sizeof(struct loc_track));
10001 struct kmem_cache *s = file_inode(filep)->i_private;
10002 unsigned long *obj_map;
10003
10004 if (!t)
10005 return -ENOMEM;
10006
10007 obj_map = bitmap_alloc(oo_objects(s->oo), GFP_KERNEL);
10008 if (!obj_map) {
10009 seq_release_private(inode, filep);
10010 return -ENOMEM;
10011 }
10012
10013 alloc = debugfs_get_aux_num(filep);
10014
10015 if (!alloc_loc_track(t, PAGE_SIZE / sizeof(struct location), GFP_KERNEL)) {
10016 bitmap_free(obj_map);
10017 seq_release_private(inode, filep);
10018 return -ENOMEM;
10019 }
10020
10021 for_each_kmem_cache_node(s, node, n) {
10022 unsigned long flags;
10023 struct slab *slab;
10024
10025 if (!node_nr_slabs(n))
10026 continue;
10027
10028 spin_lock_irqsave(&n->list_lock, flags);
10029 list_for_each_entry(slab, &n->partial, slab_list)
10030 process_slab(t, s, slab, alloc, obj_map);
10031 list_for_each_entry(slab, &n->full, slab_list)
10032 process_slab(t, s, slab, alloc, obj_map);
10033 spin_unlock_irqrestore(&n->list_lock, flags);
10034 }
10035
10036 /* Sort locations by count */
10037 sort(t->loc, t->count, sizeof(struct location),
10038 cmp_loc_by_count, NULL);
10039
10040 bitmap_free(obj_map);
10041 return 0;
10042 }
10043
10044 static int slab_debug_trace_release(struct inode *inode, struct file *file)
10045 {
10046 struct seq_file *seq = file->private_data;
10047 struct loc_track *t = seq->private;
10048
10049 free_loc_track(t);
10050 return seq_release_private(inode, file);
10051 }
10052
10053 static const struct file_operations slab_debugfs_fops = {
10054 .open = slab_debug_trace_open,
10055 .read = seq_read,
10056 .llseek = seq_lseek,
10057 .release = slab_debug_trace_release,
10058 };
10059
10060 static void debugfs_slab_add(struct kmem_cache *s)
10061 {
10062 struct dentry *slab_cache_dir;
10063
10064 if (unlikely(!slab_debugfs_root))
10065 return;
10066
10067 slab_cache_dir = debugfs_create_dir(s->name, slab_debugfs_root);
10068
10069 debugfs_create_file_aux_num("alloc_traces", 0400, slab_cache_dir, s,
10070 TRACK_ALLOC, &slab_debugfs_fops);
10071
10072 debugfs_create_file_aux_num("free_traces", 0400, slab_cache_dir, s,
10073 TRACK_FREE, &slab_debugfs_fops);
10074 }
10075
10076 void debugfs_slab_release(struct kmem_cache *s)
10077 {
10078 debugfs_lookup_and_remove(s->name, slab_debugfs_root);
10079 }
10080
10081 static int __init slab_debugfs_init(void)
10082 {
10083 struct kmem_cache *s;
10084
10085 slab_debugfs_root = debugfs_create_dir("slab", NULL);
10086
10087 list_for_each_entry(s, &slab_caches, list)
10088 if (s->flags & SLAB_STORE_USER)
10089 debugfs_slab_add(s);
10090
10091 return 0;
10092
10093 }
10094 __initcall(slab_debugfs_init);
10095 #endif
10096 /*
10097 * The /proc/slabinfo ABI
10098 */
10099 #ifdef CONFIG_SLUB_DEBUG
10100 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
10101 {
10102 unsigned long nr_slabs = 0;
10103 unsigned long nr_objs = 0;
10104 unsigned long nr_free = 0;
10105 int node;
10106 struct kmem_cache_node *n;
10107
10108 for_each_kmem_cache_node(s, node, n) {
10109 nr_slabs += node_nr_slabs(n);
10110 nr_objs += node_nr_objs(n);
10111 nr_free += count_partial_free_approx(n);
10112 }
10113
10114 sinfo->active_objs = nr_objs - nr_free;
10115 sinfo->num_objs = nr_objs;
10116 sinfo->active_slabs = nr_slabs;
10117 sinfo->num_slabs = nr_slabs;
10118 sinfo->objects_per_slab = oo_objects(s->oo);
10119 sinfo->cache_order = oo_order(s->oo);
10120 }
10121 #endif /* CONFIG_SLUB_DEBUG */