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