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/proc_fs.h>
23 #include <linux/seq_file.h>
24 #include <linux/kasan.h>
25 #include <linux/kmsan.h>
26 #include <linux/cpu.h>
27 #include <linux/cpuset.h>
28 #include <linux/mempolicy.h>
29 #include <linux/ctype.h>
30 #include <linux/stackdepot.h>
31 #include <linux/debugobjects.h>
32 #include <linux/kallsyms.h>
33 #include <linux/kfence.h>
34 #include <linux/memory.h>
35 #include <linux/math64.h>
36 #include <linux/fault-inject.h>
37 #include <linux/kmemleak.h>
38 #include <linux/stacktrace.h>
39 #include <linux/prefetch.h>
40 #include <linux/memcontrol.h>
41 #include <linux/random.h>
42 #include <kunit/test.h>
43 #include <kunit/test-bug.h>
44 #include <linux/sort.h>
45
46 #include <linux/debugfs.h>
47 #include <trace/events/kmem.h>
48
49 #include "internal.h"
50
51 /*
52 * Lock order:
53 * 1. slab_mutex (Global Mutex)
54 * 2. node->list_lock (Spinlock)
55 * 3. kmem_cache->cpu_slab->lock (Local lock)
56 * 4. slab_lock(slab) (Only on some arches)
57 * 5. object_map_lock (Only for debugging)
58 *
59 * slab_mutex
60 *
61 * The role of the slab_mutex is to protect the list of all the slabs
62 * and to synchronize major metadata changes to slab cache structures.
63 * Also synchronizes memory hotplug callbacks.
64 *
65 * slab_lock
66 *
67 * The slab_lock is a wrapper around the page lock, thus it is a bit
68 * spinlock.
69 *
70 * The slab_lock is only used on arches that do not have the ability
71 * to do a cmpxchg_double. It only protects:
72 *
73 * A. slab->freelist -> List of free objects in a slab
74 * B. slab->inuse -> Number of objects in use
75 * C. slab->objects -> Number of objects in slab
76 * D. slab->frozen -> frozen state
77 *
78 * Frozen slabs
79 *
80 * If a slab is frozen then it is exempt from list management. It is
81 * the cpu slab which is actively allocated from by the processor that
82 * froze it and it is not on any list. The processor that froze the
83 * slab is the one who can perform list operations on the slab. Other
84 * processors may put objects onto the freelist but the processor that
85 * froze the slab is the only one that can retrieve the objects from the
86 * slab's freelist.
87 *
88 * CPU partial slabs
89 *
90 * The partially empty slabs cached on the CPU partial list are used
91 * for performance reasons, which speeds up the allocation process.
92 * These slabs are not frozen, but are also exempt from list management,
93 * by clearing the PG_workingset flag when moving out of the node
94 * partial list. Please see __slab_free() for more details.
95 *
96 * To sum up, the current scheme is:
97 * - node partial slab: PG_Workingset && !frozen
98 * - cpu partial slab: !PG_Workingset && !frozen
99 * - cpu slab: !PG_Workingset && frozen
100 * - full slab: !PG_Workingset && !frozen
101 *
102 * list_lock
103 *
104 * The list_lock protects the partial and full list on each node and
105 * the partial slab counter. If taken then no new slabs may be added or
106 * removed from the lists nor make the number of partial slabs be modified.
107 * (Note that the total number of slabs is an atomic value that may be
108 * modified without taking the list lock).
109 *
110 * The list_lock is a centralized lock and thus we avoid taking it as
111 * much as possible. As long as SLUB does not have to handle partial
112 * slabs, operations can continue without any centralized lock. F.e.
113 * allocating a long series of objects that fill up slabs does not require
114 * the list lock.
115 *
116 * For debug caches, all allocations are forced to go through a list_lock
117 * protected region to serialize against concurrent validation.
118 *
119 * cpu_slab->lock local lock
120 *
121 * This locks protect slowpath manipulation of all kmem_cache_cpu fields
122 * except the stat counters. This is a percpu structure manipulated only by
123 * the local cpu, so the lock protects against being preempted or interrupted
124 * by an irq. Fast path operations rely on lockless operations instead.
125 *
126 * On PREEMPT_RT, the local lock neither disables interrupts nor preemption
127 * which means the lockless fastpath cannot be used as it might interfere with
128 * an in-progress slow path operations. In this case the local lock is always
129 * taken but it still utilizes the freelist for the common operations.
130 *
131 * lockless fastpaths
132 *
133 * The fast path allocation (slab_alloc_node()) and freeing (do_slab_free())
134 * are fully lockless when satisfied from the percpu slab (and when
135 * cmpxchg_double is possible to use, otherwise slab_lock is taken).
136 * They also don't disable preemption or migration or irqs. They rely on
137 * the transaction id (tid) field to detect being preempted or moved to
138 * another cpu.
139 *
140 * irq, preemption, migration considerations
141 *
142 * Interrupts are disabled as part of list_lock or local_lock operations, or
143 * around the slab_lock operation, in order to make the slab allocator safe
144 * to use in the context of an irq.
145 *
146 * In addition, preemption (or migration on PREEMPT_RT) is disabled in the
147 * allocation slowpath, bulk allocation, and put_cpu_partial(), so that the
148 * local cpu doesn't change in the process and e.g. the kmem_cache_cpu pointer
149 * doesn't have to be revalidated in each section protected by the local lock.
150 *
151 * SLUB assigns one slab for allocation to each processor.
152 * Allocations only occur from these slabs called cpu slabs.
153 *
154 * Slabs with free elements are kept on a partial list and during regular
155 * operations no list for full slabs is used. If an object in a full slab is
156 * freed then the slab will show up again on the partial lists.
157 * We track full slabs for debugging purposes though because otherwise we
158 * cannot scan all objects.
159 *
160 * Slabs are freed when they become empty. Teardown and setup is
161 * minimal so we rely on the page allocators per cpu caches for
162 * fast frees and allocs.
163 *
164 * slab->frozen The slab is frozen and exempt from list processing.
165 * This means that the slab is dedicated to a purpose
166 * such as satisfying allocations for a specific
167 * processor. Objects may be freed in the slab while
168 * it is frozen but slab_free will then skip the usual
169 * list operations. It is up to the processor holding
170 * the slab to integrate the slab into the slab lists
171 * when the slab is no longer needed.
172 *
173 * One use of this flag is to mark slabs that are
174 * used for allocations. Then such a slab becomes a cpu
175 * slab. The cpu slab may be equipped with an additional
176 * freelist that allows lockless access to
177 * free objects in addition to the regular freelist
178 * that requires the slab lock.
179 *
180 * SLAB_DEBUG_FLAGS Slab requires special handling due to debug
181 * options set. This moves slab handling out of
182 * the fast path and disables lockless freelists.
183 */
184
185 /*
186 * We could simply use migrate_disable()/enable() but as long as it's a
187 * function call even on !PREEMPT_RT, use inline preempt_disable() there.
188 */
189 #ifndef CONFIG_PREEMPT_RT
190 #define slub_get_cpu_ptr(var) get_cpu_ptr(var)
191 #define slub_put_cpu_ptr(var) put_cpu_ptr(var)
192 #define USE_LOCKLESS_FAST_PATH() (true)
193 #else
194 #define slub_get_cpu_ptr(var) \
195 ({ \
196 migrate_disable(); \
197 this_cpu_ptr(var); \
198 })
199 #define slub_put_cpu_ptr(var) \
200 do { \
201 (void)(var); \
202 migrate_enable(); \
203 } while (0)
204 #define USE_LOCKLESS_FAST_PATH() (false)
205 #endif
206
207 #ifndef CONFIG_SLUB_TINY
208 #define __fastpath_inline __always_inline
209 #else
210 #define __fastpath_inline
211 #endif
212
213 #ifdef CONFIG_SLUB_DEBUG
214 #ifdef CONFIG_SLUB_DEBUG_ON
215 DEFINE_STATIC_KEY_TRUE(slub_debug_enabled);
216 #else
217 DEFINE_STATIC_KEY_FALSE(slub_debug_enabled);
218 #endif
219 #endif /* CONFIG_SLUB_DEBUG */
220
221 /* Structure holding parameters for get_partial() call chain */
222 struct partial_context {
223 gfp_t flags;
224 unsigned int orig_size;
225 void *object;
226 };
227
kmem_cache_debug(struct kmem_cache * s)228 static inline bool kmem_cache_debug(struct kmem_cache *s)
229 {
230 return kmem_cache_debug_flags(s, SLAB_DEBUG_FLAGS);
231 }
232
slub_debug_orig_size(struct kmem_cache * s)233 static inline bool slub_debug_orig_size(struct kmem_cache *s)
234 {
235 return (kmem_cache_debug_flags(s, SLAB_STORE_USER) &&
236 (s->flags & SLAB_KMALLOC));
237 }
238
fixup_red_left(struct kmem_cache * s,void * p)239 void *fixup_red_left(struct kmem_cache *s, void *p)
240 {
241 if (kmem_cache_debug_flags(s, SLAB_RED_ZONE))
242 p += s->red_left_pad;
243
244 return p;
245 }
246
kmem_cache_has_cpu_partial(struct kmem_cache * s)247 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
248 {
249 #ifdef CONFIG_SLUB_CPU_PARTIAL
250 return !kmem_cache_debug(s);
251 #else
252 return false;
253 #endif
254 }
255
256 /*
257 * Issues still to be resolved:
258 *
259 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
260 *
261 * - Variable sizing of the per node arrays
262 */
263
264 /* Enable to log cmpxchg failures */
265 #undef SLUB_DEBUG_CMPXCHG
266
267 #ifndef CONFIG_SLUB_TINY
268 /*
269 * Minimum number of partial slabs. These will be left on the partial
270 * lists even if they are empty. kmem_cache_shrink may reclaim them.
271 */
272 #define MIN_PARTIAL 5
273
274 /*
275 * Maximum number of desirable partial slabs.
276 * The existence of more partial slabs makes kmem_cache_shrink
277 * sort the partial list by the number of objects in use.
278 */
279 #define MAX_PARTIAL 10
280 #else
281 #define MIN_PARTIAL 0
282 #define MAX_PARTIAL 0
283 #endif
284
285 #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
286 SLAB_POISON | SLAB_STORE_USER)
287
288 /*
289 * These debug flags cannot use CMPXCHG because there might be consistency
290 * issues when checking or reading debug information
291 */
292 #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
293 SLAB_TRACE)
294
295
296 /*
297 * Debugging flags that require metadata to be stored in the slab. These get
298 * disabled when slab_debug=O is used and a cache's min order increases with
299 * metadata.
300 */
301 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
302
303 #define OO_SHIFT 16
304 #define OO_MASK ((1 << OO_SHIFT) - 1)
305 #define MAX_OBJS_PER_PAGE 32767 /* since slab.objects is u15 */
306
307 /* Internal SLUB flags */
308 /* Poison object */
309 #define __OBJECT_POISON __SLAB_FLAG_BIT(_SLAB_OBJECT_POISON)
310 /* Use cmpxchg_double */
311
312 #ifdef system_has_freelist_aba
313 #define __CMPXCHG_DOUBLE __SLAB_FLAG_BIT(_SLAB_CMPXCHG_DOUBLE)
314 #else
315 #define __CMPXCHG_DOUBLE __SLAB_FLAG_UNUSED
316 #endif
317
318 /*
319 * Tracking user of a slab.
320 */
321 #define TRACK_ADDRS_COUNT 16
322 struct track {
323 unsigned long addr; /* Called from address */
324 #ifdef CONFIG_STACKDEPOT
325 depot_stack_handle_t handle;
326 #endif
327 int cpu; /* Was running on cpu */
328 int pid; /* Pid context */
329 unsigned long when; /* When did the operation occur */
330 };
331
332 enum track_item { TRACK_ALLOC, TRACK_FREE };
333
334 #ifdef SLAB_SUPPORTS_SYSFS
335 static int sysfs_slab_add(struct kmem_cache *);
336 static int sysfs_slab_alias(struct kmem_cache *, const char *);
337 #else
sysfs_slab_add(struct kmem_cache * s)338 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
sysfs_slab_alias(struct kmem_cache * s,const char * p)339 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
340 { return 0; }
341 #endif
342
343 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_SLUB_DEBUG)
344 static void debugfs_slab_add(struct kmem_cache *);
345 #else
debugfs_slab_add(struct kmem_cache * s)346 static inline void debugfs_slab_add(struct kmem_cache *s) { }
347 #endif
348
349 enum stat_item {
350 ALLOC_FASTPATH, /* Allocation from cpu slab */
351 ALLOC_SLOWPATH, /* Allocation by getting a new cpu slab */
352 FREE_FASTPATH, /* Free to cpu slab */
353 FREE_SLOWPATH, /* Freeing not to cpu slab */
354 FREE_FROZEN, /* Freeing to frozen slab */
355 FREE_ADD_PARTIAL, /* Freeing moves slab to partial list */
356 FREE_REMOVE_PARTIAL, /* Freeing removes last object */
357 ALLOC_FROM_PARTIAL, /* Cpu slab acquired from node partial list */
358 ALLOC_SLAB, /* Cpu slab acquired from page allocator */
359 ALLOC_REFILL, /* Refill cpu slab from slab freelist */
360 ALLOC_NODE_MISMATCH, /* Switching cpu slab */
361 FREE_SLAB, /* Slab freed to the page allocator */
362 CPUSLAB_FLUSH, /* Abandoning of the cpu slab */
363 DEACTIVATE_FULL, /* Cpu slab was full when deactivated */
364 DEACTIVATE_EMPTY, /* Cpu slab was empty when deactivated */
365 DEACTIVATE_TO_HEAD, /* Cpu slab was moved to the head of partials */
366 DEACTIVATE_TO_TAIL, /* Cpu slab was moved to the tail of partials */
367 DEACTIVATE_REMOTE_FREES,/* Slab contained remotely freed objects */
368 DEACTIVATE_BYPASS, /* Implicit deactivation */
369 ORDER_FALLBACK, /* Number of times fallback was necessary */
370 CMPXCHG_DOUBLE_CPU_FAIL,/* Failures of this_cpu_cmpxchg_double */
371 CMPXCHG_DOUBLE_FAIL, /* Failures of slab freelist update */
372 CPU_PARTIAL_ALLOC, /* Used cpu partial on alloc */
373 CPU_PARTIAL_FREE, /* Refill cpu partial on free */
374 CPU_PARTIAL_NODE, /* Refill cpu partial from node partial */
375 CPU_PARTIAL_DRAIN, /* Drain cpu partial to node partial */
376 NR_SLUB_STAT_ITEMS
377 };
378
379 #ifndef CONFIG_SLUB_TINY
380 /*
381 * When changing the layout, make sure freelist and tid are still compatible
382 * with this_cpu_cmpxchg_double() alignment requirements.
383 */
384 struct kmem_cache_cpu {
385 union {
386 struct {
387 void **freelist; /* Pointer to next available object */
388 unsigned long tid; /* Globally unique transaction id */
389 };
390 freelist_aba_t freelist_tid;
391 };
392 struct slab *slab; /* The slab from which we are allocating */
393 #ifdef CONFIG_SLUB_CPU_PARTIAL
394 struct slab *partial; /* Partially allocated slabs */
395 #endif
396 local_lock_t lock; /* Protects the fields above */
397 #ifdef CONFIG_SLUB_STATS
398 unsigned int stat[NR_SLUB_STAT_ITEMS];
399 #endif
400 };
401 #endif /* CONFIG_SLUB_TINY */
402
stat(const struct kmem_cache * s,enum stat_item si)403 static inline void stat(const struct kmem_cache *s, enum stat_item si)
404 {
405 #ifdef CONFIG_SLUB_STATS
406 /*
407 * The rmw is racy on a preemptible kernel but this is acceptable, so
408 * avoid this_cpu_add()'s irq-disable overhead.
409 */
410 raw_cpu_inc(s->cpu_slab->stat[si]);
411 #endif
412 }
413
414 static inline
stat_add(const struct kmem_cache * s,enum stat_item si,int v)415 void stat_add(const struct kmem_cache *s, enum stat_item si, int v)
416 {
417 #ifdef CONFIG_SLUB_STATS
418 raw_cpu_add(s->cpu_slab->stat[si], v);
419 #endif
420 }
421
422 /*
423 * The slab lists for all objects.
424 */
425 struct kmem_cache_node {
426 spinlock_t list_lock;
427 unsigned long nr_partial;
428 struct list_head partial;
429 #ifdef CONFIG_SLUB_DEBUG
430 atomic_long_t nr_slabs;
431 atomic_long_t total_objects;
432 struct list_head full;
433 #endif
434 };
435
get_node(struct kmem_cache * s,int node)436 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
437 {
438 return s->node[node];
439 }
440
441 /*
442 * Iterator over all nodes. The body will be executed for each node that has
443 * a kmem_cache_node structure allocated (which is true for all online nodes)
444 */
445 #define for_each_kmem_cache_node(__s, __node, __n) \
446 for (__node = 0; __node < nr_node_ids; __node++) \
447 if ((__n = get_node(__s, __node)))
448
449 /*
450 * Tracks for which NUMA nodes we have kmem_cache_nodes allocated.
451 * Corresponds to node_state[N_NORMAL_MEMORY], but can temporarily
452 * differ during memory hotplug/hotremove operations.
453 * Protected by slab_mutex.
454 */
455 static nodemask_t slab_nodes;
456
457 #ifndef CONFIG_SLUB_TINY
458 /*
459 * Workqueue used for flush_cpu_slab().
460 */
461 static struct workqueue_struct *flushwq;
462 #endif
463
464 /********************************************************************
465 * Core slab cache functions
466 *******************************************************************/
467
468 /*
469 * Returns freelist pointer (ptr). With hardening, this is obfuscated
470 * with an XOR of the address where the pointer is held and a per-cache
471 * random number.
472 */
freelist_ptr_encode(const struct kmem_cache * s,void * ptr,unsigned long ptr_addr)473 static inline freeptr_t freelist_ptr_encode(const struct kmem_cache *s,
474 void *ptr, unsigned long ptr_addr)
475 {
476 unsigned long encoded;
477
478 #ifdef CONFIG_SLAB_FREELIST_HARDENED
479 encoded = (unsigned long)ptr ^ s->random ^ swab(ptr_addr);
480 #else
481 encoded = (unsigned long)ptr;
482 #endif
483 return (freeptr_t){.v = encoded};
484 }
485
freelist_ptr_decode(const struct kmem_cache * s,freeptr_t ptr,unsigned long ptr_addr)486 static inline void *freelist_ptr_decode(const struct kmem_cache *s,
487 freeptr_t ptr, unsigned long ptr_addr)
488 {
489 void *decoded;
490
491 #ifdef CONFIG_SLAB_FREELIST_HARDENED
492 decoded = (void *)(ptr.v ^ s->random ^ swab(ptr_addr));
493 #else
494 decoded = (void *)ptr.v;
495 #endif
496 return decoded;
497 }
498
get_freepointer(struct kmem_cache * s,void * object)499 static inline void *get_freepointer(struct kmem_cache *s, void *object)
500 {
501 unsigned long ptr_addr;
502 freeptr_t p;
503
504 object = kasan_reset_tag(object);
505 ptr_addr = (unsigned long)object + s->offset;
506 p = *(freeptr_t *)(ptr_addr);
507 return freelist_ptr_decode(s, p, ptr_addr);
508 }
509
510 #ifndef CONFIG_SLUB_TINY
prefetch_freepointer(const struct kmem_cache * s,void * object)511 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
512 {
513 prefetchw(object + s->offset);
514 }
515 #endif
516
517 /*
518 * When running under KMSAN, get_freepointer_safe() may return an uninitialized
519 * pointer value in the case the current thread loses the race for the next
520 * memory chunk in the freelist. In that case this_cpu_cmpxchg_double() in
521 * slab_alloc_node() will fail, so the uninitialized value won't be used, but
522 * KMSAN will still check all arguments of cmpxchg because of imperfect
523 * handling of inline assembly.
524 * To work around this problem, we apply __no_kmsan_checks to ensure that
525 * get_freepointer_safe() returns initialized memory.
526 */
527 __no_kmsan_checks
get_freepointer_safe(struct kmem_cache * s,void * object)528 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
529 {
530 unsigned long freepointer_addr;
531 freeptr_t p;
532
533 if (!debug_pagealloc_enabled_static())
534 return get_freepointer(s, object);
535
536 object = kasan_reset_tag(object);
537 freepointer_addr = (unsigned long)object + s->offset;
538 copy_from_kernel_nofault(&p, (freeptr_t *)freepointer_addr, sizeof(p));
539 return freelist_ptr_decode(s, p, freepointer_addr);
540 }
541
set_freepointer(struct kmem_cache * s,void * object,void * fp)542 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
543 {
544 unsigned long freeptr_addr = (unsigned long)object + s->offset;
545
546 #ifdef CONFIG_SLAB_FREELIST_HARDENED
547 BUG_ON(object == fp); /* naive detection of double free or corruption */
548 #endif
549
550 freeptr_addr = (unsigned long)kasan_reset_tag((void *)freeptr_addr);
551 *(freeptr_t *)freeptr_addr = freelist_ptr_encode(s, fp, freeptr_addr);
552 }
553
554 /*
555 * See comment in calculate_sizes().
556 */
freeptr_outside_object(struct kmem_cache * s)557 static inline bool freeptr_outside_object(struct kmem_cache *s)
558 {
559 return s->offset >= s->inuse;
560 }
561
562 /*
563 * Return offset of the end of info block which is inuse + free pointer if
564 * not overlapping with object.
565 */
get_info_end(struct kmem_cache * s)566 static inline unsigned int get_info_end(struct kmem_cache *s)
567 {
568 if (freeptr_outside_object(s))
569 return s->inuse + sizeof(void *);
570 else
571 return s->inuse;
572 }
573
574 /* Loop over all objects in a slab */
575 #define for_each_object(__p, __s, __addr, __objects) \
576 for (__p = fixup_red_left(__s, __addr); \
577 __p < (__addr) + (__objects) * (__s)->size; \
578 __p += (__s)->size)
579
order_objects(unsigned int order,unsigned int size)580 static inline unsigned int order_objects(unsigned int order, unsigned int size)
581 {
582 return ((unsigned int)PAGE_SIZE << order) / size;
583 }
584
oo_make(unsigned int order,unsigned int size)585 static inline struct kmem_cache_order_objects oo_make(unsigned int order,
586 unsigned int size)
587 {
588 struct kmem_cache_order_objects x = {
589 (order << OO_SHIFT) + order_objects(order, size)
590 };
591
592 return x;
593 }
594
oo_order(struct kmem_cache_order_objects x)595 static inline unsigned int oo_order(struct kmem_cache_order_objects x)
596 {
597 return x.x >> OO_SHIFT;
598 }
599
oo_objects(struct kmem_cache_order_objects x)600 static inline unsigned int oo_objects(struct kmem_cache_order_objects x)
601 {
602 return x.x & OO_MASK;
603 }
604
605 #ifdef CONFIG_SLUB_CPU_PARTIAL
slub_set_cpu_partial(struct kmem_cache * s,unsigned int nr_objects)606 static void slub_set_cpu_partial(struct kmem_cache *s, unsigned int nr_objects)
607 {
608 unsigned int nr_slabs;
609
610 s->cpu_partial = nr_objects;
611
612 /*
613 * We take the number of objects but actually limit the number of
614 * slabs on the per cpu partial list, in order to limit excessive
615 * growth of the list. For simplicity we assume that the slabs will
616 * be half-full.
617 */
618 nr_slabs = DIV_ROUND_UP(nr_objects * 2, oo_objects(s->oo));
619 s->cpu_partial_slabs = nr_slabs;
620 }
621
slub_get_cpu_partial(struct kmem_cache * s)622 static inline unsigned int slub_get_cpu_partial(struct kmem_cache *s)
623 {
624 return s->cpu_partial_slabs;
625 }
626 #else
627 static inline void
slub_set_cpu_partial(struct kmem_cache * s,unsigned int nr_objects)628 slub_set_cpu_partial(struct kmem_cache *s, unsigned int nr_objects)
629 {
630 }
631
slub_get_cpu_partial(struct kmem_cache * s)632 static inline unsigned int slub_get_cpu_partial(struct kmem_cache *s)
633 {
634 return 0;
635 }
636 #endif /* CONFIG_SLUB_CPU_PARTIAL */
637
638 /*
639 * Per slab locking using the pagelock
640 */
slab_lock(struct slab * slab)641 static __always_inline void slab_lock(struct slab *slab)
642 {
643 bit_spin_lock(PG_locked, &slab->__page_flags);
644 }
645
slab_unlock(struct slab * slab)646 static __always_inline void slab_unlock(struct slab *slab)
647 {
648 bit_spin_unlock(PG_locked, &slab->__page_flags);
649 }
650
651 static inline bool
__update_freelist_fast(struct slab * slab,void * freelist_old,unsigned long counters_old,void * freelist_new,unsigned long counters_new)652 __update_freelist_fast(struct slab *slab,
653 void *freelist_old, unsigned long counters_old,
654 void *freelist_new, unsigned long counters_new)
655 {
656 #ifdef system_has_freelist_aba
657 freelist_aba_t old = { .freelist = freelist_old, .counter = counters_old };
658 freelist_aba_t new = { .freelist = freelist_new, .counter = counters_new };
659
660 return try_cmpxchg_freelist(&slab->freelist_counter.full, &old.full, new.full);
661 #else
662 return false;
663 #endif
664 }
665
666 static inline bool
__update_freelist_slow(struct slab * slab,void * freelist_old,unsigned long counters_old,void * freelist_new,unsigned long counters_new)667 __update_freelist_slow(struct slab *slab,
668 void *freelist_old, unsigned long counters_old,
669 void *freelist_new, unsigned long counters_new)
670 {
671 bool ret = false;
672
673 slab_lock(slab);
674 if (slab->freelist == freelist_old &&
675 slab->counters == counters_old) {
676 slab->freelist = freelist_new;
677 slab->counters = counters_new;
678 ret = true;
679 }
680 slab_unlock(slab);
681
682 return ret;
683 }
684
685 /*
686 * Interrupts must be disabled (for the fallback code to work right), typically
687 * by an _irqsave() lock variant. On PREEMPT_RT the preempt_disable(), which is
688 * part of bit_spin_lock(), is sufficient because the policy is not to allow any
689 * allocation/ free operation in hardirq context. Therefore nothing can
690 * interrupt the operation.
691 */
__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)692 static inline bool __slab_update_freelist(struct kmem_cache *s, struct slab *slab,
693 void *freelist_old, unsigned long counters_old,
694 void *freelist_new, unsigned long counters_new,
695 const char *n)
696 {
697 bool ret;
698
699 if (USE_LOCKLESS_FAST_PATH())
700 lockdep_assert_irqs_disabled();
701
702 if (s->flags & __CMPXCHG_DOUBLE) {
703 ret = __update_freelist_fast(slab, freelist_old, counters_old,
704 freelist_new, counters_new);
705 } else {
706 ret = __update_freelist_slow(slab, freelist_old, counters_old,
707 freelist_new, counters_new);
708 }
709 if (likely(ret))
710 return true;
711
712 cpu_relax();
713 stat(s, CMPXCHG_DOUBLE_FAIL);
714
715 #ifdef SLUB_DEBUG_CMPXCHG
716 pr_info("%s %s: cmpxchg double redo ", n, s->name);
717 #endif
718
719 return false;
720 }
721
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)722 static inline bool slab_update_freelist(struct kmem_cache *s, struct slab *slab,
723 void *freelist_old, unsigned long counters_old,
724 void *freelist_new, unsigned long counters_new,
725 const char *n)
726 {
727 bool ret;
728
729 if (s->flags & __CMPXCHG_DOUBLE) {
730 ret = __update_freelist_fast(slab, freelist_old, counters_old,
731 freelist_new, counters_new);
732 } else {
733 unsigned long flags;
734
735 local_irq_save(flags);
736 ret = __update_freelist_slow(slab, freelist_old, counters_old,
737 freelist_new, counters_new);
738 local_irq_restore(flags);
739 }
740 if (likely(ret))
741 return true;
742
743 cpu_relax();
744 stat(s, CMPXCHG_DOUBLE_FAIL);
745
746 #ifdef SLUB_DEBUG_CMPXCHG
747 pr_info("%s %s: cmpxchg double redo ", n, s->name);
748 #endif
749
750 return false;
751 }
752
753 /*
754 * kmalloc caches has fixed sizes (mostly power of 2), and kmalloc() API
755 * family will round up the real request size to these fixed ones, so
756 * there could be an extra area than what is requested. Save the original
757 * request size in the meta data area, for better debug and sanity check.
758 */
set_orig_size(struct kmem_cache * s,void * object,unsigned int orig_size)759 static inline void set_orig_size(struct kmem_cache *s,
760 void *object, unsigned int orig_size)
761 {
762 void *p = kasan_reset_tag(object);
763 unsigned int kasan_meta_size;
764
765 if (!slub_debug_orig_size(s))
766 return;
767
768 /*
769 * KASAN can save its free meta data inside of the object at offset 0.
770 * If this meta data size is larger than 'orig_size', it will overlap
771 * the data redzone in [orig_size+1, object_size]. Thus, we adjust
772 * 'orig_size' to be as at least as big as KASAN's meta data.
773 */
774 kasan_meta_size = kasan_metadata_size(s, true);
775 if (kasan_meta_size > orig_size)
776 orig_size = kasan_meta_size;
777
778 p += get_info_end(s);
779 p += sizeof(struct track) * 2;
780
781 *(unsigned int *)p = orig_size;
782 }
783
get_orig_size(struct kmem_cache * s,void * object)784 static inline unsigned int get_orig_size(struct kmem_cache *s, void *object)
785 {
786 void *p = kasan_reset_tag(object);
787
788 if (!slub_debug_orig_size(s))
789 return s->object_size;
790
791 p += get_info_end(s);
792 p += sizeof(struct track) * 2;
793
794 return *(unsigned int *)p;
795 }
796
797 #ifdef CONFIG_SLUB_DEBUG
798 static unsigned long object_map[BITS_TO_LONGS(MAX_OBJS_PER_PAGE)];
799 static DEFINE_SPINLOCK(object_map_lock);
800
__fill_map(unsigned long * obj_map,struct kmem_cache * s,struct slab * slab)801 static void __fill_map(unsigned long *obj_map, struct kmem_cache *s,
802 struct slab *slab)
803 {
804 void *addr = slab_address(slab);
805 void *p;
806
807 bitmap_zero(obj_map, slab->objects);
808
809 for (p = slab->freelist; p; p = get_freepointer(s, p))
810 set_bit(__obj_to_index(s, addr, p), obj_map);
811 }
812
813 #if IS_ENABLED(CONFIG_KUNIT)
slab_add_kunit_errors(void)814 static bool slab_add_kunit_errors(void)
815 {
816 struct kunit_resource *resource;
817
818 if (!kunit_get_current_test())
819 return false;
820
821 resource = kunit_find_named_resource(current->kunit_test, "slab_errors");
822 if (!resource)
823 return false;
824
825 (*(int *)resource->data)++;
826 kunit_put_resource(resource);
827 return true;
828 }
829
slab_in_kunit_test(void)830 static bool slab_in_kunit_test(void)
831 {
832 struct kunit_resource *resource;
833
834 if (!kunit_get_current_test())
835 return false;
836
837 resource = kunit_find_named_resource(current->kunit_test, "slab_errors");
838 if (!resource)
839 return false;
840
841 kunit_put_resource(resource);
842 return true;
843 }
844 #else
slab_add_kunit_errors(void)845 static inline bool slab_add_kunit_errors(void) { return false; }
slab_in_kunit_test(void)846 static inline bool slab_in_kunit_test(void) { return false; }
847 #endif
848
size_from_object(struct kmem_cache * s)849 static inline unsigned int size_from_object(struct kmem_cache *s)
850 {
851 if (s->flags & SLAB_RED_ZONE)
852 return s->size - s->red_left_pad;
853
854 return s->size;
855 }
856
restore_red_left(struct kmem_cache * s,void * p)857 static inline void *restore_red_left(struct kmem_cache *s, void *p)
858 {
859 if (s->flags & SLAB_RED_ZONE)
860 p -= s->red_left_pad;
861
862 return p;
863 }
864
865 /*
866 * Debug settings:
867 */
868 #if defined(CONFIG_SLUB_DEBUG_ON)
869 static slab_flags_t slub_debug = DEBUG_DEFAULT_FLAGS;
870 #else
871 static slab_flags_t slub_debug;
872 #endif
873
874 static char *slub_debug_string;
875 static int disable_higher_order_debug;
876
877 /*
878 * slub is about to manipulate internal object metadata. This memory lies
879 * outside the range of the allocated object, so accessing it would normally
880 * be reported by kasan as a bounds error. metadata_access_enable() is used
881 * to tell kasan that these accesses are OK.
882 */
metadata_access_enable(void)883 static inline void metadata_access_enable(void)
884 {
885 kasan_disable_current();
886 kmsan_disable_current();
887 }
888
metadata_access_disable(void)889 static inline void metadata_access_disable(void)
890 {
891 kmsan_enable_current();
892 kasan_enable_current();
893 }
894
895 /*
896 * Object debugging
897 */
898
899 /* 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)900 static inline int check_valid_pointer(struct kmem_cache *s,
901 struct slab *slab, void *object)
902 {
903 void *base;
904
905 if (!object)
906 return 1;
907
908 base = slab_address(slab);
909 object = kasan_reset_tag(object);
910 object = restore_red_left(s, object);
911 if (object < base || object >= base + slab->objects * s->size ||
912 (object - base) % s->size) {
913 return 0;
914 }
915
916 return 1;
917 }
918
print_section(char * level,char * text,u8 * addr,unsigned int length)919 static void print_section(char *level, char *text, u8 *addr,
920 unsigned int length)
921 {
922 metadata_access_enable();
923 print_hex_dump(level, text, DUMP_PREFIX_ADDRESS,
924 16, 1, kasan_reset_tag((void *)addr), length, 1);
925 metadata_access_disable();
926 }
927
get_track(struct kmem_cache * s,void * object,enum track_item alloc)928 static struct track *get_track(struct kmem_cache *s, void *object,
929 enum track_item alloc)
930 {
931 struct track *p;
932
933 p = object + get_info_end(s);
934
935 return kasan_reset_tag(p + alloc);
936 }
937
938 #ifdef CONFIG_STACKDEPOT
set_track_prepare(void)939 static noinline depot_stack_handle_t set_track_prepare(void)
940 {
941 depot_stack_handle_t handle;
942 unsigned long entries[TRACK_ADDRS_COUNT];
943 unsigned int nr_entries;
944
945 nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 3);
946 handle = stack_depot_save(entries, nr_entries, GFP_NOWAIT);
947
948 return handle;
949 }
950 #else
set_track_prepare(void)951 static inline depot_stack_handle_t set_track_prepare(void)
952 {
953 return 0;
954 }
955 #endif
956
set_track_update(struct kmem_cache * s,void * object,enum track_item alloc,unsigned long addr,depot_stack_handle_t handle)957 static void set_track_update(struct kmem_cache *s, void *object,
958 enum track_item alloc, unsigned long addr,
959 depot_stack_handle_t handle)
960 {
961 struct track *p = get_track(s, object, alloc);
962
963 #ifdef CONFIG_STACKDEPOT
964 p->handle = handle;
965 #endif
966 p->addr = addr;
967 p->cpu = smp_processor_id();
968 p->pid = current->pid;
969 p->when = jiffies;
970 }
971
set_track(struct kmem_cache * s,void * object,enum track_item alloc,unsigned long addr)972 static __always_inline void set_track(struct kmem_cache *s, void *object,
973 enum track_item alloc, unsigned long addr)
974 {
975 depot_stack_handle_t handle = set_track_prepare();
976
977 set_track_update(s, object, alloc, addr, handle);
978 }
979
init_tracking(struct kmem_cache * s,void * object)980 static void init_tracking(struct kmem_cache *s, void *object)
981 {
982 struct track *p;
983
984 if (!(s->flags & SLAB_STORE_USER))
985 return;
986
987 p = get_track(s, object, TRACK_ALLOC);
988 memset(p, 0, 2*sizeof(struct track));
989 }
990
print_track(const char * s,struct track * t,unsigned long pr_time)991 static void print_track(const char *s, struct track *t, unsigned long pr_time)
992 {
993 depot_stack_handle_t handle __maybe_unused;
994
995 if (!t->addr)
996 return;
997
998 pr_err("%s in %pS age=%lu cpu=%u pid=%d\n",
999 s, (void *)t->addr, pr_time - t->when, t->cpu, t->pid);
1000 #ifdef CONFIG_STACKDEPOT
1001 handle = READ_ONCE(t->handle);
1002 if (handle)
1003 stack_depot_print(handle);
1004 else
1005 pr_err("object allocation/free stack trace missing\n");
1006 #endif
1007 }
1008
print_tracking(struct kmem_cache * s,void * object)1009 void print_tracking(struct kmem_cache *s, void *object)
1010 {
1011 unsigned long pr_time = jiffies;
1012 if (!(s->flags & SLAB_STORE_USER))
1013 return;
1014
1015 print_track("Allocated", get_track(s, object, TRACK_ALLOC), pr_time);
1016 print_track("Freed", get_track(s, object, TRACK_FREE), pr_time);
1017 }
1018
print_slab_info(const struct slab * slab)1019 static void print_slab_info(const struct slab *slab)
1020 {
1021 pr_err("Slab 0x%p objects=%u used=%u fp=0x%p flags=%pGp\n",
1022 slab, slab->objects, slab->inuse, slab->freelist,
1023 &slab->__page_flags);
1024 }
1025
skip_orig_size_check(struct kmem_cache * s,const void * object)1026 void skip_orig_size_check(struct kmem_cache *s, const void *object)
1027 {
1028 set_orig_size(s, (void *)object, s->object_size);
1029 }
1030
slab_bug(struct kmem_cache * s,char * fmt,...)1031 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
1032 {
1033 struct va_format vaf;
1034 va_list args;
1035
1036 va_start(args, fmt);
1037 vaf.fmt = fmt;
1038 vaf.va = &args;
1039 pr_err("=============================================================================\n");
1040 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
1041 pr_err("-----------------------------------------------------------------------------\n\n");
1042 va_end(args);
1043 }
1044
1045 __printf(2, 3)
slab_fix(struct kmem_cache * s,char * fmt,...)1046 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
1047 {
1048 struct va_format vaf;
1049 va_list args;
1050
1051 if (slab_add_kunit_errors())
1052 return;
1053
1054 va_start(args, fmt);
1055 vaf.fmt = fmt;
1056 vaf.va = &args;
1057 pr_err("FIX %s: %pV\n", s->name, &vaf);
1058 va_end(args);
1059 }
1060
print_trailer(struct kmem_cache * s,struct slab * slab,u8 * p)1061 static void print_trailer(struct kmem_cache *s, struct slab *slab, u8 *p)
1062 {
1063 unsigned int off; /* Offset of last byte */
1064 u8 *addr = slab_address(slab);
1065
1066 print_tracking(s, p);
1067
1068 print_slab_info(slab);
1069
1070 pr_err("Object 0x%p @offset=%tu fp=0x%p\n\n",
1071 p, p - addr, get_freepointer(s, p));
1072
1073 if (s->flags & SLAB_RED_ZONE)
1074 print_section(KERN_ERR, "Redzone ", p - s->red_left_pad,
1075 s->red_left_pad);
1076 else if (p > addr + 16)
1077 print_section(KERN_ERR, "Bytes b4 ", p - 16, 16);
1078
1079 print_section(KERN_ERR, "Object ", p,
1080 min_t(unsigned int, s->object_size, PAGE_SIZE));
1081 if (s->flags & SLAB_RED_ZONE)
1082 print_section(KERN_ERR, "Redzone ", p + s->object_size,
1083 s->inuse - s->object_size);
1084
1085 off = get_info_end(s);
1086
1087 if (s->flags & SLAB_STORE_USER)
1088 off += 2 * sizeof(struct track);
1089
1090 if (slub_debug_orig_size(s))
1091 off += sizeof(unsigned int);
1092
1093 off += kasan_metadata_size(s, false);
1094
1095 if (off != size_from_object(s))
1096 /* Beginning of the filler is the free pointer */
1097 print_section(KERN_ERR, "Padding ", p + off,
1098 size_from_object(s) - off);
1099
1100 dump_stack();
1101 }
1102
object_err(struct kmem_cache * s,struct slab * slab,u8 * object,char * reason)1103 static void object_err(struct kmem_cache *s, struct slab *slab,
1104 u8 *object, char *reason)
1105 {
1106 if (slab_add_kunit_errors())
1107 return;
1108
1109 slab_bug(s, "%s", reason);
1110 print_trailer(s, slab, object);
1111 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
1112 }
1113
freelist_corrupted(struct kmem_cache * s,struct slab * slab,void ** freelist,void * nextfree)1114 static bool freelist_corrupted(struct kmem_cache *s, struct slab *slab,
1115 void **freelist, void *nextfree)
1116 {
1117 if ((s->flags & SLAB_CONSISTENCY_CHECKS) &&
1118 !check_valid_pointer(s, slab, nextfree) && freelist) {
1119 object_err(s, slab, *freelist, "Freechain corrupt");
1120 *freelist = NULL;
1121 slab_fix(s, "Isolate corrupted freechain");
1122 return true;
1123 }
1124
1125 return false;
1126 }
1127
slab_err(struct kmem_cache * s,struct slab * slab,const char * fmt,...)1128 static __printf(3, 4) void slab_err(struct kmem_cache *s, struct slab *slab,
1129 const char *fmt, ...)
1130 {
1131 va_list args;
1132 char buf[100];
1133
1134 if (slab_add_kunit_errors())
1135 return;
1136
1137 va_start(args, fmt);
1138 vsnprintf(buf, sizeof(buf), fmt, args);
1139 va_end(args);
1140 slab_bug(s, "%s", buf);
1141 print_slab_info(slab);
1142 dump_stack();
1143 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
1144 }
1145
init_object(struct kmem_cache * s,void * object,u8 val)1146 static void init_object(struct kmem_cache *s, void *object, u8 val)
1147 {
1148 u8 *p = kasan_reset_tag(object);
1149 unsigned int poison_size = s->object_size;
1150
1151 if (s->flags & SLAB_RED_ZONE) {
1152 /*
1153 * Here and below, avoid overwriting the KMSAN shadow. Keeping
1154 * the shadow makes it possible to distinguish uninit-value
1155 * from use-after-free.
1156 */
1157 memset_no_sanitize_memory(p - s->red_left_pad, val,
1158 s->red_left_pad);
1159
1160 if (slub_debug_orig_size(s) && val == SLUB_RED_ACTIVE) {
1161 /*
1162 * Redzone the extra allocated space by kmalloc than
1163 * requested, and the poison size will be limited to
1164 * the original request size accordingly.
1165 */
1166 poison_size = get_orig_size(s, object);
1167 }
1168 }
1169
1170 if (s->flags & __OBJECT_POISON) {
1171 memset_no_sanitize_memory(p, POISON_FREE, poison_size - 1);
1172 memset_no_sanitize_memory(p + poison_size - 1, POISON_END, 1);
1173 }
1174
1175 if (s->flags & SLAB_RED_ZONE)
1176 memset_no_sanitize_memory(p + poison_size, val,
1177 s->inuse - poison_size);
1178 }
1179
restore_bytes(struct kmem_cache * s,char * message,u8 data,void * from,void * to)1180 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
1181 void *from, void *to)
1182 {
1183 slab_fix(s, "Restoring %s 0x%p-0x%p=0x%x", message, from, to - 1, data);
1184 memset(from, data, to - from);
1185 }
1186
1187 #ifdef CONFIG_KMSAN
1188 #define pad_check_attributes noinline __no_kmsan_checks
1189 #else
1190 #define pad_check_attributes
1191 #endif
1192
1193 static pad_check_attributes int
check_bytes_and_report(struct kmem_cache * s,struct slab * slab,u8 * object,char * what,u8 * start,unsigned int value,unsigned int bytes)1194 check_bytes_and_report(struct kmem_cache *s, struct slab *slab,
1195 u8 *object, char *what,
1196 u8 *start, unsigned int value, unsigned int bytes)
1197 {
1198 u8 *fault;
1199 u8 *end;
1200 u8 *addr = slab_address(slab);
1201
1202 metadata_access_enable();
1203 fault = memchr_inv(kasan_reset_tag(start), value, bytes);
1204 metadata_access_disable();
1205 if (!fault)
1206 return 1;
1207
1208 end = start + bytes;
1209 while (end > fault && end[-1] == value)
1210 end--;
1211
1212 if (slab_add_kunit_errors())
1213 goto skip_bug_print;
1214
1215 slab_bug(s, "%s overwritten", what);
1216 pr_err("0x%p-0x%p @offset=%tu. First byte 0x%x instead of 0x%x\n",
1217 fault, end - 1, fault - addr,
1218 fault[0], value);
1219
1220 skip_bug_print:
1221 restore_bytes(s, what, value, fault, end);
1222 return 0;
1223 }
1224
1225 /*
1226 * Object layout:
1227 *
1228 * object address
1229 * Bytes of the object to be managed.
1230 * If the freepointer may overlay the object then the free
1231 * pointer is at the middle of the object.
1232 *
1233 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
1234 * 0xa5 (POISON_END)
1235 *
1236 * object + s->object_size
1237 * Padding to reach word boundary. This is also used for Redzoning.
1238 * Padding is extended by another word if Redzoning is enabled and
1239 * object_size == inuse.
1240 *
1241 * We fill with 0xbb (SLUB_RED_INACTIVE) for inactive objects and with
1242 * 0xcc (SLUB_RED_ACTIVE) for objects in use.
1243 *
1244 * object + s->inuse
1245 * Meta data starts here.
1246 *
1247 * A. Free pointer (if we cannot overwrite object on free)
1248 * B. Tracking data for SLAB_STORE_USER
1249 * C. Original request size for kmalloc object (SLAB_STORE_USER enabled)
1250 * D. Padding to reach required alignment boundary or at minimum
1251 * one word if debugging is on to be able to detect writes
1252 * before the word boundary.
1253 *
1254 * Padding is done using 0x5a (POISON_INUSE)
1255 *
1256 * object + s->size
1257 * Nothing is used beyond s->size.
1258 *
1259 * If slabcaches are merged then the object_size and inuse boundaries are mostly
1260 * ignored. And therefore no slab options that rely on these boundaries
1261 * may be used with merged slabcaches.
1262 */
1263
check_pad_bytes(struct kmem_cache * s,struct slab * slab,u8 * p)1264 static int check_pad_bytes(struct kmem_cache *s, struct slab *slab, u8 *p)
1265 {
1266 unsigned long off = get_info_end(s); /* The end of info */
1267
1268 if (s->flags & SLAB_STORE_USER) {
1269 /* We also have user information there */
1270 off += 2 * sizeof(struct track);
1271
1272 if (s->flags & SLAB_KMALLOC)
1273 off += sizeof(unsigned int);
1274 }
1275
1276 off += kasan_metadata_size(s, false);
1277
1278 if (size_from_object(s) == off)
1279 return 1;
1280
1281 return check_bytes_and_report(s, slab, p, "Object padding",
1282 p + off, POISON_INUSE, size_from_object(s) - off);
1283 }
1284
1285 /* Check the pad bytes at the end of a slab page */
1286 static pad_check_attributes void
slab_pad_check(struct kmem_cache * s,struct slab * slab)1287 slab_pad_check(struct kmem_cache *s, struct slab *slab)
1288 {
1289 u8 *start;
1290 u8 *fault;
1291 u8 *end;
1292 u8 *pad;
1293 int length;
1294 int remainder;
1295
1296 if (!(s->flags & SLAB_POISON))
1297 return;
1298
1299 start = slab_address(slab);
1300 length = slab_size(slab);
1301 end = start + length;
1302 remainder = length % s->size;
1303 if (!remainder)
1304 return;
1305
1306 pad = end - remainder;
1307 metadata_access_enable();
1308 fault = memchr_inv(kasan_reset_tag(pad), POISON_INUSE, remainder);
1309 metadata_access_disable();
1310 if (!fault)
1311 return;
1312 while (end > fault && end[-1] == POISON_INUSE)
1313 end--;
1314
1315 slab_err(s, slab, "Padding overwritten. 0x%p-0x%p @offset=%tu",
1316 fault, end - 1, fault - start);
1317 print_section(KERN_ERR, "Padding ", pad, remainder);
1318
1319 restore_bytes(s, "slab padding", POISON_INUSE, fault, end);
1320 }
1321
check_object(struct kmem_cache * s,struct slab * slab,void * object,u8 val)1322 static int check_object(struct kmem_cache *s, struct slab *slab,
1323 void *object, u8 val)
1324 {
1325 u8 *p = object;
1326 u8 *endobject = object + s->object_size;
1327 unsigned int orig_size, kasan_meta_size;
1328 int ret = 1;
1329
1330 if (s->flags & SLAB_RED_ZONE) {
1331 if (!check_bytes_and_report(s, slab, object, "Left Redzone",
1332 object - s->red_left_pad, val, s->red_left_pad))
1333 ret = 0;
1334
1335 if (!check_bytes_and_report(s, slab, object, "Right Redzone",
1336 endobject, val, s->inuse - s->object_size))
1337 ret = 0;
1338
1339 if (slub_debug_orig_size(s) && val == SLUB_RED_ACTIVE) {
1340 orig_size = get_orig_size(s, object);
1341
1342 if (s->object_size > orig_size &&
1343 !check_bytes_and_report(s, slab, object,
1344 "kmalloc Redzone", p + orig_size,
1345 val, s->object_size - orig_size)) {
1346 ret = 0;
1347 }
1348 }
1349 } else {
1350 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
1351 if (!check_bytes_and_report(s, slab, p, "Alignment padding",
1352 endobject, POISON_INUSE,
1353 s->inuse - s->object_size))
1354 ret = 0;
1355 }
1356 }
1357
1358 if (s->flags & SLAB_POISON) {
1359 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON)) {
1360 /*
1361 * KASAN can save its free meta data inside of the
1362 * object at offset 0. Thus, skip checking the part of
1363 * the redzone that overlaps with the meta data.
1364 */
1365 kasan_meta_size = kasan_metadata_size(s, true);
1366 if (kasan_meta_size < s->object_size - 1 &&
1367 !check_bytes_and_report(s, slab, p, "Poison",
1368 p + kasan_meta_size, POISON_FREE,
1369 s->object_size - kasan_meta_size - 1))
1370 ret = 0;
1371 if (kasan_meta_size < s->object_size &&
1372 !check_bytes_and_report(s, slab, p, "End Poison",
1373 p + s->object_size - 1, POISON_END, 1))
1374 ret = 0;
1375 }
1376 /*
1377 * check_pad_bytes cleans up on its own.
1378 */
1379 if (!check_pad_bytes(s, slab, p))
1380 ret = 0;
1381 }
1382
1383 /*
1384 * Cannot check freepointer while object is allocated if
1385 * object and freepointer overlap.
1386 */
1387 if ((freeptr_outside_object(s) || val != SLUB_RED_ACTIVE) &&
1388 !check_valid_pointer(s, slab, get_freepointer(s, p))) {
1389 object_err(s, slab, p, "Freepointer corrupt");
1390 /*
1391 * No choice but to zap it and thus lose the remainder
1392 * of the free objects in this slab. May cause
1393 * another error because the object count is now wrong.
1394 */
1395 set_freepointer(s, p, NULL);
1396 ret = 0;
1397 }
1398
1399 if (!ret && !slab_in_kunit_test()) {
1400 print_trailer(s, slab, object);
1401 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
1402 }
1403
1404 return ret;
1405 }
1406
check_slab(struct kmem_cache * s,struct slab * slab)1407 static int check_slab(struct kmem_cache *s, struct slab *slab)
1408 {
1409 int maxobj;
1410
1411 if (!folio_test_slab(slab_folio(slab))) {
1412 slab_err(s, slab, "Not a valid slab page");
1413 return 0;
1414 }
1415
1416 maxobj = order_objects(slab_order(slab), s->size);
1417 if (slab->objects > maxobj) {
1418 slab_err(s, slab, "objects %u > max %u",
1419 slab->objects, maxobj);
1420 return 0;
1421 }
1422 if (slab->inuse > slab->objects) {
1423 slab_err(s, slab, "inuse %u > max %u",
1424 slab->inuse, slab->objects);
1425 return 0;
1426 }
1427 /* Slab_pad_check fixes things up after itself */
1428 slab_pad_check(s, slab);
1429 return 1;
1430 }
1431
1432 /*
1433 * Determine if a certain object in a slab is on the freelist. Must hold the
1434 * slab lock to guarantee that the chains are in a consistent state.
1435 */
on_freelist(struct kmem_cache * s,struct slab * slab,void * search)1436 static int on_freelist(struct kmem_cache *s, struct slab *slab, void *search)
1437 {
1438 int nr = 0;
1439 void *fp;
1440 void *object = NULL;
1441 int max_objects;
1442
1443 fp = slab->freelist;
1444 while (fp && nr <= slab->objects) {
1445 if (fp == search)
1446 return 1;
1447 if (!check_valid_pointer(s, slab, fp)) {
1448 if (object) {
1449 object_err(s, slab, object,
1450 "Freechain corrupt");
1451 set_freepointer(s, object, NULL);
1452 } else {
1453 slab_err(s, slab, "Freepointer corrupt");
1454 slab->freelist = NULL;
1455 slab->inuse = slab->objects;
1456 slab_fix(s, "Freelist cleared");
1457 return 0;
1458 }
1459 break;
1460 }
1461 object = fp;
1462 fp = get_freepointer(s, object);
1463 nr++;
1464 }
1465
1466 max_objects = order_objects(slab_order(slab), s->size);
1467 if (max_objects > MAX_OBJS_PER_PAGE)
1468 max_objects = MAX_OBJS_PER_PAGE;
1469
1470 if (slab->objects != max_objects) {
1471 slab_err(s, slab, "Wrong number of objects. Found %d but should be %d",
1472 slab->objects, max_objects);
1473 slab->objects = max_objects;
1474 slab_fix(s, "Number of objects adjusted");
1475 }
1476 if (slab->inuse != slab->objects - nr) {
1477 slab_err(s, slab, "Wrong object count. Counter is %d but counted were %d",
1478 slab->inuse, slab->objects - nr);
1479 slab->inuse = slab->objects - nr;
1480 slab_fix(s, "Object count adjusted");
1481 }
1482 return search == NULL;
1483 }
1484
trace(struct kmem_cache * s,struct slab * slab,void * object,int alloc)1485 static void trace(struct kmem_cache *s, struct slab *slab, void *object,
1486 int alloc)
1487 {
1488 if (s->flags & SLAB_TRACE) {
1489 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
1490 s->name,
1491 alloc ? "alloc" : "free",
1492 object, slab->inuse,
1493 slab->freelist);
1494
1495 if (!alloc)
1496 print_section(KERN_INFO, "Object ", (void *)object,
1497 s->object_size);
1498
1499 dump_stack();
1500 }
1501 }
1502
1503 /*
1504 * Tracking of fully allocated slabs for debugging purposes.
1505 */
add_full(struct kmem_cache * s,struct kmem_cache_node * n,struct slab * slab)1506 static void add_full(struct kmem_cache *s,
1507 struct kmem_cache_node *n, struct slab *slab)
1508 {
1509 if (!(s->flags & SLAB_STORE_USER))
1510 return;
1511
1512 lockdep_assert_held(&n->list_lock);
1513 list_add(&slab->slab_list, &n->full);
1514 }
1515
remove_full(struct kmem_cache * s,struct kmem_cache_node * n,struct slab * slab)1516 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct slab *slab)
1517 {
1518 if (!(s->flags & SLAB_STORE_USER))
1519 return;
1520
1521 lockdep_assert_held(&n->list_lock);
1522 list_del(&slab->slab_list);
1523 }
1524
node_nr_slabs(struct kmem_cache_node * n)1525 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1526 {
1527 return atomic_long_read(&n->nr_slabs);
1528 }
1529
inc_slabs_node(struct kmem_cache * s,int node,int objects)1530 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1531 {
1532 struct kmem_cache_node *n = get_node(s, node);
1533
1534 atomic_long_inc(&n->nr_slabs);
1535 atomic_long_add(objects, &n->total_objects);
1536 }
dec_slabs_node(struct kmem_cache * s,int node,int objects)1537 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1538 {
1539 struct kmem_cache_node *n = get_node(s, node);
1540
1541 atomic_long_dec(&n->nr_slabs);
1542 atomic_long_sub(objects, &n->total_objects);
1543 }
1544
1545 /* Object debug checks for alloc/free paths */
setup_object_debug(struct kmem_cache * s,void * object)1546 static void setup_object_debug(struct kmem_cache *s, void *object)
1547 {
1548 if (!kmem_cache_debug_flags(s, SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))
1549 return;
1550
1551 init_object(s, object, SLUB_RED_INACTIVE);
1552 init_tracking(s, object);
1553 }
1554
1555 static
setup_slab_debug(struct kmem_cache * s,struct slab * slab,void * addr)1556 void setup_slab_debug(struct kmem_cache *s, struct slab *slab, void *addr)
1557 {
1558 if (!kmem_cache_debug_flags(s, SLAB_POISON))
1559 return;
1560
1561 metadata_access_enable();
1562 memset(kasan_reset_tag(addr), POISON_INUSE, slab_size(slab));
1563 metadata_access_disable();
1564 }
1565
alloc_consistency_checks(struct kmem_cache * s,struct slab * slab,void * object)1566 static inline int alloc_consistency_checks(struct kmem_cache *s,
1567 struct slab *slab, void *object)
1568 {
1569 if (!check_slab(s, slab))
1570 return 0;
1571
1572 if (!check_valid_pointer(s, slab, object)) {
1573 object_err(s, slab, object, "Freelist Pointer check fails");
1574 return 0;
1575 }
1576
1577 if (!check_object(s, slab, object, SLUB_RED_INACTIVE))
1578 return 0;
1579
1580 return 1;
1581 }
1582
alloc_debug_processing(struct kmem_cache * s,struct slab * slab,void * object,int orig_size)1583 static noinline bool alloc_debug_processing(struct kmem_cache *s,
1584 struct slab *slab, void *object, int orig_size)
1585 {
1586 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1587 if (!alloc_consistency_checks(s, slab, object))
1588 goto bad;
1589 }
1590
1591 /* Success. Perform special debug activities for allocs */
1592 trace(s, slab, object, 1);
1593 set_orig_size(s, object, orig_size);
1594 init_object(s, object, SLUB_RED_ACTIVE);
1595 return true;
1596
1597 bad:
1598 if (folio_test_slab(slab_folio(slab))) {
1599 /*
1600 * If this is a slab page then lets do the best we can
1601 * to avoid issues in the future. Marking all objects
1602 * as used avoids touching the remaining objects.
1603 */
1604 slab_fix(s, "Marking all objects used");
1605 slab->inuse = slab->objects;
1606 slab->freelist = NULL;
1607 }
1608 return false;
1609 }
1610
free_consistency_checks(struct kmem_cache * s,struct slab * slab,void * object,unsigned long addr)1611 static inline int free_consistency_checks(struct kmem_cache *s,
1612 struct slab *slab, void *object, unsigned long addr)
1613 {
1614 if (!check_valid_pointer(s, slab, object)) {
1615 slab_err(s, slab, "Invalid object pointer 0x%p", object);
1616 return 0;
1617 }
1618
1619 if (on_freelist(s, slab, object)) {
1620 object_err(s, slab, object, "Object already free");
1621 return 0;
1622 }
1623
1624 if (!check_object(s, slab, object, SLUB_RED_ACTIVE))
1625 return 0;
1626
1627 if (unlikely(s != slab->slab_cache)) {
1628 if (!folio_test_slab(slab_folio(slab))) {
1629 slab_err(s, slab, "Attempt to free object(0x%p) outside of slab",
1630 object);
1631 } else if (!slab->slab_cache) {
1632 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1633 object);
1634 dump_stack();
1635 } else
1636 object_err(s, slab, object,
1637 "page slab pointer corrupt.");
1638 return 0;
1639 }
1640 return 1;
1641 }
1642
1643 /*
1644 * Parse a block of slab_debug options. Blocks are delimited by ';'
1645 *
1646 * @str: start of block
1647 * @flags: returns parsed flags, or DEBUG_DEFAULT_FLAGS if none specified
1648 * @slabs: return start of list of slabs, or NULL when there's no list
1649 * @init: assume this is initial parsing and not per-kmem-create parsing
1650 *
1651 * returns the start of next block if there's any, or NULL
1652 */
1653 static char *
parse_slub_debug_flags(char * str,slab_flags_t * flags,char ** slabs,bool init)1654 parse_slub_debug_flags(char *str, slab_flags_t *flags, char **slabs, bool init)
1655 {
1656 bool higher_order_disable = false;
1657
1658 /* Skip any completely empty blocks */
1659 while (*str && *str == ';')
1660 str++;
1661
1662 if (*str == ',') {
1663 /*
1664 * No options but restriction on slabs. This means full
1665 * debugging for slabs matching a pattern.
1666 */
1667 *flags = DEBUG_DEFAULT_FLAGS;
1668 goto check_slabs;
1669 }
1670 *flags = 0;
1671
1672 /* Determine which debug features should be switched on */
1673 for (; *str && *str != ',' && *str != ';'; str++) {
1674 switch (tolower(*str)) {
1675 case '-':
1676 *flags = 0;
1677 break;
1678 case 'f':
1679 *flags |= SLAB_CONSISTENCY_CHECKS;
1680 break;
1681 case 'z':
1682 *flags |= SLAB_RED_ZONE;
1683 break;
1684 case 'p':
1685 *flags |= SLAB_POISON;
1686 break;
1687 case 'u':
1688 *flags |= SLAB_STORE_USER;
1689 break;
1690 case 't':
1691 *flags |= SLAB_TRACE;
1692 break;
1693 case 'a':
1694 *flags |= SLAB_FAILSLAB;
1695 break;
1696 case 'o':
1697 /*
1698 * Avoid enabling debugging on caches if its minimum
1699 * order would increase as a result.
1700 */
1701 higher_order_disable = true;
1702 break;
1703 default:
1704 if (init)
1705 pr_err("slab_debug option '%c' unknown. skipped\n", *str);
1706 }
1707 }
1708 check_slabs:
1709 if (*str == ',')
1710 *slabs = ++str;
1711 else
1712 *slabs = NULL;
1713
1714 /* Skip over the slab list */
1715 while (*str && *str != ';')
1716 str++;
1717
1718 /* Skip any completely empty blocks */
1719 while (*str && *str == ';')
1720 str++;
1721
1722 if (init && higher_order_disable)
1723 disable_higher_order_debug = 1;
1724
1725 if (*str)
1726 return str;
1727 else
1728 return NULL;
1729 }
1730
setup_slub_debug(char * str)1731 static int __init setup_slub_debug(char *str)
1732 {
1733 slab_flags_t flags;
1734 slab_flags_t global_flags;
1735 char *saved_str;
1736 char *slab_list;
1737 bool global_slub_debug_changed = false;
1738 bool slab_list_specified = false;
1739
1740 global_flags = DEBUG_DEFAULT_FLAGS;
1741 if (*str++ != '=' || !*str)
1742 /*
1743 * No options specified. Switch on full debugging.
1744 */
1745 goto out;
1746
1747 saved_str = str;
1748 while (str) {
1749 str = parse_slub_debug_flags(str, &flags, &slab_list, true);
1750
1751 if (!slab_list) {
1752 global_flags = flags;
1753 global_slub_debug_changed = true;
1754 } else {
1755 slab_list_specified = true;
1756 if (flags & SLAB_STORE_USER)
1757 stack_depot_request_early_init();
1758 }
1759 }
1760
1761 /*
1762 * For backwards compatibility, a single list of flags with list of
1763 * slabs means debugging is only changed for those slabs, so the global
1764 * slab_debug should be unchanged (0 or DEBUG_DEFAULT_FLAGS, depending
1765 * on CONFIG_SLUB_DEBUG_ON). We can extended that to multiple lists as
1766 * long as there is no option specifying flags without a slab list.
1767 */
1768 if (slab_list_specified) {
1769 if (!global_slub_debug_changed)
1770 global_flags = slub_debug;
1771 slub_debug_string = saved_str;
1772 }
1773 out:
1774 slub_debug = global_flags;
1775 if (slub_debug & SLAB_STORE_USER)
1776 stack_depot_request_early_init();
1777 if (slub_debug != 0 || slub_debug_string)
1778 static_branch_enable(&slub_debug_enabled);
1779 else
1780 static_branch_disable(&slub_debug_enabled);
1781 if ((static_branch_unlikely(&init_on_alloc) ||
1782 static_branch_unlikely(&init_on_free)) &&
1783 (slub_debug & SLAB_POISON))
1784 pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n");
1785 return 1;
1786 }
1787
1788 __setup("slab_debug", setup_slub_debug);
1789 __setup_param("slub_debug", slub_debug, setup_slub_debug, 0);
1790
1791 /*
1792 * kmem_cache_flags - apply debugging options to the cache
1793 * @flags: flags to set
1794 * @name: name of the cache
1795 *
1796 * Debug option(s) are applied to @flags. In addition to the debug
1797 * option(s), if a slab name (or multiple) is specified i.e.
1798 * slab_debug=<Debug-Options>,<slab name1>,<slab name2> ...
1799 * then only the select slabs will receive the debug option(s).
1800 */
kmem_cache_flags(slab_flags_t flags,const char * name)1801 slab_flags_t kmem_cache_flags(slab_flags_t flags, const char *name)
1802 {
1803 char *iter;
1804 size_t len;
1805 char *next_block;
1806 slab_flags_t block_flags;
1807 slab_flags_t slub_debug_local = slub_debug;
1808
1809 if (flags & SLAB_NO_USER_FLAGS)
1810 return flags;
1811
1812 /*
1813 * If the slab cache is for debugging (e.g. kmemleak) then
1814 * don't store user (stack trace) information by default,
1815 * but let the user enable it via the command line below.
1816 */
1817 if (flags & SLAB_NOLEAKTRACE)
1818 slub_debug_local &= ~SLAB_STORE_USER;
1819
1820 len = strlen(name);
1821 next_block = slub_debug_string;
1822 /* Go through all blocks of debug options, see if any matches our slab's name */
1823 while (next_block) {
1824 next_block = parse_slub_debug_flags(next_block, &block_flags, &iter, false);
1825 if (!iter)
1826 continue;
1827 /* Found a block that has a slab list, search it */
1828 while (*iter) {
1829 char *end, *glob;
1830 size_t cmplen;
1831
1832 end = strchrnul(iter, ',');
1833 if (next_block && next_block < end)
1834 end = next_block - 1;
1835
1836 glob = strnchr(iter, end - iter, '*');
1837 if (glob)
1838 cmplen = glob - iter;
1839 else
1840 cmplen = max_t(size_t, len, (end - iter));
1841
1842 if (!strncmp(name, iter, cmplen)) {
1843 flags |= block_flags;
1844 return flags;
1845 }
1846
1847 if (!*end || *end == ';')
1848 break;
1849 iter = end + 1;
1850 }
1851 }
1852
1853 return flags | slub_debug_local;
1854 }
1855 #else /* !CONFIG_SLUB_DEBUG */
setup_object_debug(struct kmem_cache * s,void * object)1856 static inline void setup_object_debug(struct kmem_cache *s, void *object) {}
1857 static inline
setup_slab_debug(struct kmem_cache * s,struct slab * slab,void * addr)1858 void setup_slab_debug(struct kmem_cache *s, struct slab *slab, void *addr) {}
1859
alloc_debug_processing(struct kmem_cache * s,struct slab * slab,void * object,int orig_size)1860 static inline bool alloc_debug_processing(struct kmem_cache *s,
1861 struct slab *slab, void *object, int orig_size) { return true; }
1862
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)1863 static inline bool free_debug_processing(struct kmem_cache *s,
1864 struct slab *slab, void *head, void *tail, int *bulk_cnt,
1865 unsigned long addr, depot_stack_handle_t handle) { return true; }
1866
slab_pad_check(struct kmem_cache * s,struct slab * slab)1867 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)1868 static inline int check_object(struct kmem_cache *s, struct slab *slab,
1869 void *object, u8 val) { return 1; }
set_track_prepare(void)1870 static inline depot_stack_handle_t set_track_prepare(void) { return 0; }
set_track(struct kmem_cache * s,void * object,enum track_item alloc,unsigned long addr)1871 static inline void set_track(struct kmem_cache *s, void *object,
1872 enum track_item alloc, unsigned long addr) {}
add_full(struct kmem_cache * s,struct kmem_cache_node * n,struct slab * slab)1873 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1874 struct slab *slab) {}
remove_full(struct kmem_cache * s,struct kmem_cache_node * n,struct slab * slab)1875 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1876 struct slab *slab) {}
kmem_cache_flags(slab_flags_t flags,const char * name)1877 slab_flags_t kmem_cache_flags(slab_flags_t flags, const char *name)
1878 {
1879 return flags;
1880 }
1881 #define slub_debug 0
1882
1883 #define disable_higher_order_debug 0
1884
node_nr_slabs(struct kmem_cache_node * n)1885 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1886 { return 0; }
inc_slabs_node(struct kmem_cache * s,int node,int objects)1887 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1888 int objects) {}
dec_slabs_node(struct kmem_cache * s,int node,int objects)1889 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1890 int objects) {}
1891 #ifndef CONFIG_SLUB_TINY
freelist_corrupted(struct kmem_cache * s,struct slab * slab,void ** freelist,void * nextfree)1892 static bool freelist_corrupted(struct kmem_cache *s, struct slab *slab,
1893 void **freelist, void *nextfree)
1894 {
1895 return false;
1896 }
1897 #endif
1898 #endif /* CONFIG_SLUB_DEBUG */
1899
1900 #ifdef CONFIG_SLAB_OBJ_EXT
1901
1902 #ifdef CONFIG_MEM_ALLOC_PROFILING_DEBUG
1903
mark_objexts_empty(struct slabobj_ext * obj_exts)1904 static inline void mark_objexts_empty(struct slabobj_ext *obj_exts)
1905 {
1906 struct slabobj_ext *slab_exts;
1907 struct slab *obj_exts_slab;
1908
1909 obj_exts_slab = virt_to_slab(obj_exts);
1910 slab_exts = slab_obj_exts(obj_exts_slab);
1911 if (slab_exts) {
1912 unsigned int offs = obj_to_index(obj_exts_slab->slab_cache,
1913 obj_exts_slab, obj_exts);
1914 /* codetag should be NULL */
1915 WARN_ON(slab_exts[offs].ref.ct);
1916 set_codetag_empty(&slab_exts[offs].ref);
1917 }
1918 }
1919
mark_failed_objexts_alloc(struct slab * slab)1920 static inline void mark_failed_objexts_alloc(struct slab *slab)
1921 {
1922 slab->obj_exts = OBJEXTS_ALLOC_FAIL;
1923 }
1924
handle_failed_objexts_alloc(unsigned long obj_exts,struct slabobj_ext * vec,unsigned int objects)1925 static inline void handle_failed_objexts_alloc(unsigned long obj_exts,
1926 struct slabobj_ext *vec, unsigned int objects)
1927 {
1928 /*
1929 * If vector previously failed to allocate then we have live
1930 * objects with no tag reference. Mark all references in this
1931 * vector as empty to avoid warnings later on.
1932 */
1933 if (obj_exts & OBJEXTS_ALLOC_FAIL) {
1934 unsigned int i;
1935
1936 for (i = 0; i < objects; i++)
1937 set_codetag_empty(&vec[i].ref);
1938 }
1939 }
1940
1941 #else /* CONFIG_MEM_ALLOC_PROFILING_DEBUG */
1942
mark_objexts_empty(struct slabobj_ext * obj_exts)1943 static inline void mark_objexts_empty(struct slabobj_ext *obj_exts) {}
mark_failed_objexts_alloc(struct slab * slab)1944 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)1945 static inline void handle_failed_objexts_alloc(unsigned long obj_exts,
1946 struct slabobj_ext *vec, unsigned int objects) {}
1947
1948 #endif /* CONFIG_MEM_ALLOC_PROFILING_DEBUG */
1949
1950 /*
1951 * The allocated objcg pointers array is not accounted directly.
1952 * Moreover, it should not come from DMA buffer and is not readily
1953 * reclaimable. So those GFP bits should be masked off.
1954 */
1955 #define OBJCGS_CLEAR_MASK (__GFP_DMA | __GFP_RECLAIMABLE | \
1956 __GFP_ACCOUNT | __GFP_NOFAIL)
1957
alloc_slab_obj_exts(struct slab * slab,struct kmem_cache * s,gfp_t gfp,bool new_slab)1958 int alloc_slab_obj_exts(struct slab *slab, struct kmem_cache *s,
1959 gfp_t gfp, bool new_slab)
1960 {
1961 unsigned int objects = objs_per_slab(s, slab);
1962 unsigned long new_exts;
1963 unsigned long old_exts;
1964 struct slabobj_ext *vec;
1965
1966 gfp &= ~OBJCGS_CLEAR_MASK;
1967 /* Prevent recursive extension vector allocation */
1968 gfp |= __GFP_NO_OBJ_EXT;
1969 vec = kcalloc_node(objects, sizeof(struct slabobj_ext), gfp,
1970 slab_nid(slab));
1971 if (!vec) {
1972 /* Mark vectors which failed to allocate */
1973 if (new_slab)
1974 mark_failed_objexts_alloc(slab);
1975
1976 return -ENOMEM;
1977 }
1978
1979 new_exts = (unsigned long)vec;
1980 #ifdef CONFIG_MEMCG
1981 new_exts |= MEMCG_DATA_OBJEXTS;
1982 #endif
1983 old_exts = READ_ONCE(slab->obj_exts);
1984 handle_failed_objexts_alloc(old_exts, vec, objects);
1985 if (new_slab) {
1986 /*
1987 * If the slab is brand new and nobody can yet access its
1988 * obj_exts, no synchronization is required and obj_exts can
1989 * be simply assigned.
1990 */
1991 slab->obj_exts = new_exts;
1992 } else if ((old_exts & ~OBJEXTS_FLAGS_MASK) ||
1993 cmpxchg(&slab->obj_exts, old_exts, new_exts) != old_exts) {
1994 /*
1995 * If the slab is already in use, somebody can allocate and
1996 * assign slabobj_exts in parallel. In this case the existing
1997 * objcg vector should be reused.
1998 */
1999 mark_objexts_empty(vec);
2000 kfree(vec);
2001 return 0;
2002 }
2003
2004 kmemleak_not_leak(vec);
2005 return 0;
2006 }
2007
free_slab_obj_exts(struct slab * slab)2008 static inline void free_slab_obj_exts(struct slab *slab)
2009 {
2010 struct slabobj_ext *obj_exts;
2011
2012 obj_exts = slab_obj_exts(slab);
2013 if (!obj_exts)
2014 return;
2015
2016 /*
2017 * obj_exts was created with __GFP_NO_OBJ_EXT flag, therefore its
2018 * corresponding extension will be NULL. alloc_tag_sub() will throw a
2019 * warning if slab has extensions but the extension of an object is
2020 * NULL, therefore replace NULL with CODETAG_EMPTY to indicate that
2021 * the extension for obj_exts is expected to be NULL.
2022 */
2023 mark_objexts_empty(obj_exts);
2024 kfree(obj_exts);
2025 slab->obj_exts = 0;
2026 }
2027
need_slab_obj_ext(void)2028 static inline bool need_slab_obj_ext(void)
2029 {
2030 if (mem_alloc_profiling_enabled())
2031 return true;
2032
2033 /*
2034 * CONFIG_MEMCG creates vector of obj_cgroup objects conditionally
2035 * inside memcg_slab_post_alloc_hook. No other users for now.
2036 */
2037 return false;
2038 }
2039
2040 #else /* CONFIG_SLAB_OBJ_EXT */
2041
alloc_slab_obj_exts(struct slab * slab,struct kmem_cache * s,gfp_t gfp,bool new_slab)2042 static int alloc_slab_obj_exts(struct slab *slab, struct kmem_cache *s,
2043 gfp_t gfp, bool new_slab)
2044 {
2045 return 0;
2046 }
2047
free_slab_obj_exts(struct slab * slab)2048 static inline void free_slab_obj_exts(struct slab *slab)
2049 {
2050 }
2051
need_slab_obj_ext(void)2052 static inline bool need_slab_obj_ext(void)
2053 {
2054 return false;
2055 }
2056
2057 #endif /* CONFIG_SLAB_OBJ_EXT */
2058
2059 #ifdef CONFIG_MEM_ALLOC_PROFILING
2060
2061 static inline struct slabobj_ext *
prepare_slab_obj_exts_hook(struct kmem_cache * s,gfp_t flags,void * p)2062 prepare_slab_obj_exts_hook(struct kmem_cache *s, gfp_t flags, void *p)
2063 {
2064 struct slab *slab;
2065
2066 if (!p)
2067 return NULL;
2068
2069 if (s->flags & (SLAB_NO_OBJ_EXT | SLAB_NOLEAKTRACE))
2070 return NULL;
2071
2072 if (flags & __GFP_NO_OBJ_EXT)
2073 return NULL;
2074
2075 slab = virt_to_slab(p);
2076 if (!slab_obj_exts(slab) &&
2077 WARN(alloc_slab_obj_exts(slab, s, flags, false),
2078 "%s, %s: Failed to create slab extension vector!\n",
2079 __func__, s->name))
2080 return NULL;
2081
2082 return slab_obj_exts(slab) + obj_to_index(s, slab, p);
2083 }
2084
2085 static inline void
alloc_tagging_slab_alloc_hook(struct kmem_cache * s,void * object,gfp_t flags)2086 alloc_tagging_slab_alloc_hook(struct kmem_cache *s, void *object, gfp_t flags)
2087 {
2088 if (need_slab_obj_ext()) {
2089 struct slabobj_ext *obj_exts;
2090
2091 obj_exts = prepare_slab_obj_exts_hook(s, flags, object);
2092 /*
2093 * Currently obj_exts is used only for allocation profiling.
2094 * If other users appear then mem_alloc_profiling_enabled()
2095 * check should be added before alloc_tag_add().
2096 */
2097 if (likely(obj_exts))
2098 alloc_tag_add(&obj_exts->ref, current->alloc_tag, s->size);
2099 }
2100 }
2101
2102 static inline void
alloc_tagging_slab_free_hook(struct kmem_cache * s,struct slab * slab,void ** p,int objects)2103 alloc_tagging_slab_free_hook(struct kmem_cache *s, struct slab *slab, void **p,
2104 int objects)
2105 {
2106 struct slabobj_ext *obj_exts;
2107 int i;
2108
2109 if (!mem_alloc_profiling_enabled())
2110 return;
2111
2112 /* slab->obj_exts might not be NULL if it was created for MEMCG accounting. */
2113 if (s->flags & (SLAB_NO_OBJ_EXT | SLAB_NOLEAKTRACE))
2114 return;
2115
2116 obj_exts = slab_obj_exts(slab);
2117 if (!obj_exts)
2118 return;
2119
2120 for (i = 0; i < objects; i++) {
2121 unsigned int off = obj_to_index(s, slab, p[i]);
2122
2123 alloc_tag_sub(&obj_exts[off].ref, s->size);
2124 }
2125 }
2126
2127 #else /* CONFIG_MEM_ALLOC_PROFILING */
2128
2129 static inline void
alloc_tagging_slab_alloc_hook(struct kmem_cache * s,void * object,gfp_t flags)2130 alloc_tagging_slab_alloc_hook(struct kmem_cache *s, void *object, gfp_t flags)
2131 {
2132 }
2133
2134 static inline void
alloc_tagging_slab_free_hook(struct kmem_cache * s,struct slab * slab,void ** p,int objects)2135 alloc_tagging_slab_free_hook(struct kmem_cache *s, struct slab *slab, void **p,
2136 int objects)
2137 {
2138 }
2139
2140 #endif /* CONFIG_MEM_ALLOC_PROFILING */
2141
2142
2143 #ifdef CONFIG_MEMCG
2144
2145 static void memcg_alloc_abort_single(struct kmem_cache *s, void *object);
2146
2147 static __fastpath_inline
memcg_slab_post_alloc_hook(struct kmem_cache * s,struct list_lru * lru,gfp_t flags,size_t size,void ** p)2148 bool memcg_slab_post_alloc_hook(struct kmem_cache *s, struct list_lru *lru,
2149 gfp_t flags, size_t size, void **p)
2150 {
2151 if (likely(!memcg_kmem_online()))
2152 return true;
2153
2154 if (likely(!(flags & __GFP_ACCOUNT) && !(s->flags & SLAB_ACCOUNT)))
2155 return true;
2156
2157 if (likely(__memcg_slab_post_alloc_hook(s, lru, flags, size, p)))
2158 return true;
2159
2160 if (likely(size == 1)) {
2161 memcg_alloc_abort_single(s, *p);
2162 *p = NULL;
2163 } else {
2164 kmem_cache_free_bulk(s, size, p);
2165 }
2166
2167 return false;
2168 }
2169
2170 static __fastpath_inline
memcg_slab_free_hook(struct kmem_cache * s,struct slab * slab,void ** p,int objects)2171 void memcg_slab_free_hook(struct kmem_cache *s, struct slab *slab, void **p,
2172 int objects)
2173 {
2174 struct slabobj_ext *obj_exts;
2175
2176 if (!memcg_kmem_online())
2177 return;
2178
2179 obj_exts = slab_obj_exts(slab);
2180 if (likely(!obj_exts))
2181 return;
2182
2183 __memcg_slab_free_hook(s, slab, p, objects, obj_exts);
2184 }
2185
2186 static __fastpath_inline
memcg_slab_post_charge(void * p,gfp_t flags)2187 bool memcg_slab_post_charge(void *p, gfp_t flags)
2188 {
2189 struct slabobj_ext *slab_exts;
2190 struct kmem_cache *s;
2191 struct folio *folio;
2192 struct slab *slab;
2193 unsigned long off;
2194
2195 folio = virt_to_folio(p);
2196 if (!folio_test_slab(folio)) {
2197 return folio_memcg_kmem(folio) ||
2198 (__memcg_kmem_charge_page(folio_page(folio, 0), flags,
2199 folio_order(folio)) == 0);
2200 }
2201
2202 slab = folio_slab(folio);
2203 s = slab->slab_cache;
2204
2205 /*
2206 * Ignore KMALLOC_NORMAL cache to avoid possible circular dependency
2207 * of slab_obj_exts being allocated from the same slab and thus the slab
2208 * becoming effectively unfreeable.
2209 */
2210 if (is_kmalloc_normal(s))
2211 return true;
2212
2213 /* Ignore already charged objects. */
2214 slab_exts = slab_obj_exts(slab);
2215 if (slab_exts) {
2216 off = obj_to_index(s, slab, p);
2217 if (unlikely(slab_exts[off].objcg))
2218 return true;
2219 }
2220
2221 return __memcg_slab_post_alloc_hook(s, NULL, flags, 1, &p);
2222 }
2223
2224 #else /* CONFIG_MEMCG */
memcg_slab_post_alloc_hook(struct kmem_cache * s,struct list_lru * lru,gfp_t flags,size_t size,void ** p)2225 static inline bool memcg_slab_post_alloc_hook(struct kmem_cache *s,
2226 struct list_lru *lru,
2227 gfp_t flags, size_t size,
2228 void **p)
2229 {
2230 return true;
2231 }
2232
memcg_slab_free_hook(struct kmem_cache * s,struct slab * slab,void ** p,int objects)2233 static inline void memcg_slab_free_hook(struct kmem_cache *s, struct slab *slab,
2234 void **p, int objects)
2235 {
2236 }
2237
memcg_slab_post_charge(void * p,gfp_t flags)2238 static inline bool memcg_slab_post_charge(void *p, gfp_t flags)
2239 {
2240 return true;
2241 }
2242 #endif /* CONFIG_MEMCG */
2243
2244 #ifdef CONFIG_SLUB_RCU_DEBUG
2245 static void slab_free_after_rcu_debug(struct rcu_head *rcu_head);
2246
2247 struct rcu_delayed_free {
2248 struct rcu_head head;
2249 void *object;
2250 };
2251 #endif
2252
2253 /*
2254 * Hooks for other subsystems that check memory allocations. In a typical
2255 * production configuration these hooks all should produce no code at all.
2256 *
2257 * Returns true if freeing of the object can proceed, false if its reuse
2258 * was delayed by CONFIG_SLUB_RCU_DEBUG or KASAN quarantine, or it was returned
2259 * to KFENCE.
2260 */
2261 static __always_inline
slab_free_hook(struct kmem_cache * s,void * x,bool init,bool after_rcu_delay)2262 bool slab_free_hook(struct kmem_cache *s, void *x, bool init,
2263 bool after_rcu_delay)
2264 {
2265 /* Are the object contents still accessible? */
2266 bool still_accessible = (s->flags & SLAB_TYPESAFE_BY_RCU) && !after_rcu_delay;
2267
2268 kmemleak_free_recursive(x, s->flags);
2269 kmsan_slab_free(s, x);
2270
2271 debug_check_no_locks_freed(x, s->object_size);
2272
2273 if (!(s->flags & SLAB_DEBUG_OBJECTS))
2274 debug_check_no_obj_freed(x, s->object_size);
2275
2276 /* Use KCSAN to help debug racy use-after-free. */
2277 if (!still_accessible)
2278 __kcsan_check_access(x, s->object_size,
2279 KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT);
2280
2281 if (kfence_free(x))
2282 return false;
2283
2284 /*
2285 * Give KASAN a chance to notice an invalid free operation before we
2286 * modify the object.
2287 */
2288 if (kasan_slab_pre_free(s, x))
2289 return false;
2290
2291 #ifdef CONFIG_SLUB_RCU_DEBUG
2292 if (still_accessible) {
2293 struct rcu_delayed_free *delayed_free;
2294
2295 delayed_free = kmalloc(sizeof(*delayed_free), GFP_NOWAIT);
2296 if (delayed_free) {
2297 /*
2298 * Let KASAN track our call stack as a "related work
2299 * creation", just like if the object had been freed
2300 * normally via kfree_rcu().
2301 * We have to do this manually because the rcu_head is
2302 * not located inside the object.
2303 */
2304 kasan_record_aux_stack_noalloc(x);
2305
2306 delayed_free->object = x;
2307 call_rcu(&delayed_free->head, slab_free_after_rcu_debug);
2308 return false;
2309 }
2310 }
2311 #endif /* CONFIG_SLUB_RCU_DEBUG */
2312
2313 /*
2314 * As memory initialization might be integrated into KASAN,
2315 * kasan_slab_free and initialization memset's must be
2316 * kept together to avoid discrepancies in behavior.
2317 *
2318 * The initialization memset's clear the object and the metadata,
2319 * but don't touch the SLAB redzone.
2320 *
2321 * The object's freepointer is also avoided if stored outside the
2322 * object.
2323 */
2324 if (unlikely(init)) {
2325 int rsize;
2326 unsigned int inuse, orig_size;
2327
2328 inuse = get_info_end(s);
2329 orig_size = get_orig_size(s, x);
2330 if (!kasan_has_integrated_init())
2331 memset(kasan_reset_tag(x), 0, orig_size);
2332 rsize = (s->flags & SLAB_RED_ZONE) ? s->red_left_pad : 0;
2333 memset((char *)kasan_reset_tag(x) + inuse, 0,
2334 s->size - inuse - rsize);
2335 /*
2336 * Restore orig_size, otherwize kmalloc redzone overwritten
2337 * would be reported
2338 */
2339 set_orig_size(s, x, orig_size);
2340
2341 }
2342 /* KASAN might put x into memory quarantine, delaying its reuse. */
2343 return !kasan_slab_free(s, x, init, still_accessible);
2344 }
2345
2346 static __fastpath_inline
slab_free_freelist_hook(struct kmem_cache * s,void ** head,void ** tail,int * cnt)2347 bool slab_free_freelist_hook(struct kmem_cache *s, void **head, void **tail,
2348 int *cnt)
2349 {
2350
2351 void *object;
2352 void *next = *head;
2353 void *old_tail = *tail;
2354 bool init;
2355
2356 if (is_kfence_address(next)) {
2357 slab_free_hook(s, next, false, false);
2358 return false;
2359 }
2360
2361 /* Head and tail of the reconstructed freelist */
2362 *head = NULL;
2363 *tail = NULL;
2364
2365 init = slab_want_init_on_free(s);
2366
2367 do {
2368 object = next;
2369 next = get_freepointer(s, object);
2370
2371 /* If object's reuse doesn't have to be delayed */
2372 if (likely(slab_free_hook(s, object, init, false))) {
2373 /* Move object to the new freelist */
2374 set_freepointer(s, object, *head);
2375 *head = object;
2376 if (!*tail)
2377 *tail = object;
2378 } else {
2379 /*
2380 * Adjust the reconstructed freelist depth
2381 * accordingly if object's reuse is delayed.
2382 */
2383 --(*cnt);
2384 }
2385 } while (object != old_tail);
2386
2387 return *head != NULL;
2388 }
2389
setup_object(struct kmem_cache * s,void * object)2390 static void *setup_object(struct kmem_cache *s, void *object)
2391 {
2392 setup_object_debug(s, object);
2393 object = kasan_init_slab_obj(s, object);
2394 if (unlikely(s->ctor)) {
2395 kasan_unpoison_new_object(s, object);
2396 s->ctor(object);
2397 kasan_poison_new_object(s, object);
2398 }
2399 return object;
2400 }
2401
2402 /*
2403 * Slab allocation and freeing
2404 */
alloc_slab_page(gfp_t flags,int node,struct kmem_cache_order_objects oo)2405 static inline struct slab *alloc_slab_page(gfp_t flags, int node,
2406 struct kmem_cache_order_objects oo)
2407 {
2408 struct folio *folio;
2409 struct slab *slab;
2410 unsigned int order = oo_order(oo);
2411
2412 if (node == NUMA_NO_NODE)
2413 folio = (struct folio *)alloc_pages(flags, order);
2414 else
2415 folio = (struct folio *)__alloc_pages_node(node, flags, order);
2416
2417 if (!folio)
2418 return NULL;
2419
2420 slab = folio_slab(folio);
2421 __folio_set_slab(folio);
2422 /* Make the flag visible before any changes to folio->mapping */
2423 smp_wmb();
2424 if (folio_is_pfmemalloc(folio))
2425 slab_set_pfmemalloc(slab);
2426
2427 return slab;
2428 }
2429
2430 #ifdef CONFIG_SLAB_FREELIST_RANDOM
2431 /* Pre-initialize the random sequence cache */
init_cache_random_seq(struct kmem_cache * s)2432 static int init_cache_random_seq(struct kmem_cache *s)
2433 {
2434 unsigned int count = oo_objects(s->oo);
2435 int err;
2436
2437 /* Bailout if already initialised */
2438 if (s->random_seq)
2439 return 0;
2440
2441 err = cache_random_seq_create(s, count, GFP_KERNEL);
2442 if (err) {
2443 pr_err("SLUB: Unable to initialize free list for %s\n",
2444 s->name);
2445 return err;
2446 }
2447
2448 /* Transform to an offset on the set of pages */
2449 if (s->random_seq) {
2450 unsigned int i;
2451
2452 for (i = 0; i < count; i++)
2453 s->random_seq[i] *= s->size;
2454 }
2455 return 0;
2456 }
2457
2458 /* Initialize each random sequence freelist per cache */
init_freelist_randomization(void)2459 static void __init init_freelist_randomization(void)
2460 {
2461 struct kmem_cache *s;
2462
2463 mutex_lock(&slab_mutex);
2464
2465 list_for_each_entry(s, &slab_caches, list)
2466 init_cache_random_seq(s);
2467
2468 mutex_unlock(&slab_mutex);
2469 }
2470
2471 /* 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)2472 static void *next_freelist_entry(struct kmem_cache *s,
2473 unsigned long *pos, void *start,
2474 unsigned long page_limit,
2475 unsigned long freelist_count)
2476 {
2477 unsigned int idx;
2478
2479 /*
2480 * If the target page allocation failed, the number of objects on the
2481 * page might be smaller than the usual size defined by the cache.
2482 */
2483 do {
2484 idx = s->random_seq[*pos];
2485 *pos += 1;
2486 if (*pos >= freelist_count)
2487 *pos = 0;
2488 } while (unlikely(idx >= page_limit));
2489
2490 return (char *)start + idx;
2491 }
2492
2493 /* Shuffle the single linked freelist based on a random pre-computed sequence */
shuffle_freelist(struct kmem_cache * s,struct slab * slab)2494 static bool shuffle_freelist(struct kmem_cache *s, struct slab *slab)
2495 {
2496 void *start;
2497 void *cur;
2498 void *next;
2499 unsigned long idx, pos, page_limit, freelist_count;
2500
2501 if (slab->objects < 2 || !s->random_seq)
2502 return false;
2503
2504 freelist_count = oo_objects(s->oo);
2505 pos = get_random_u32_below(freelist_count);
2506
2507 page_limit = slab->objects * s->size;
2508 start = fixup_red_left(s, slab_address(slab));
2509
2510 /* First entry is used as the base of the freelist */
2511 cur = next_freelist_entry(s, &pos, start, page_limit, freelist_count);
2512 cur = setup_object(s, cur);
2513 slab->freelist = cur;
2514
2515 for (idx = 1; idx < slab->objects; idx++) {
2516 next = next_freelist_entry(s, &pos, start, page_limit,
2517 freelist_count);
2518 next = setup_object(s, next);
2519 set_freepointer(s, cur, next);
2520 cur = next;
2521 }
2522 set_freepointer(s, cur, NULL);
2523
2524 return true;
2525 }
2526 #else
init_cache_random_seq(struct kmem_cache * s)2527 static inline int init_cache_random_seq(struct kmem_cache *s)
2528 {
2529 return 0;
2530 }
init_freelist_randomization(void)2531 static inline void init_freelist_randomization(void) { }
shuffle_freelist(struct kmem_cache * s,struct slab * slab)2532 static inline bool shuffle_freelist(struct kmem_cache *s, struct slab *slab)
2533 {
2534 return false;
2535 }
2536 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
2537
account_slab(struct slab * slab,int order,struct kmem_cache * s,gfp_t gfp)2538 static __always_inline void account_slab(struct slab *slab, int order,
2539 struct kmem_cache *s, gfp_t gfp)
2540 {
2541 if (memcg_kmem_online() && (s->flags & SLAB_ACCOUNT))
2542 alloc_slab_obj_exts(slab, s, gfp, true);
2543
2544 mod_node_page_state(slab_pgdat(slab), cache_vmstat_idx(s),
2545 PAGE_SIZE << order);
2546 }
2547
unaccount_slab(struct slab * slab,int order,struct kmem_cache * s)2548 static __always_inline void unaccount_slab(struct slab *slab, int order,
2549 struct kmem_cache *s)
2550 {
2551 if (memcg_kmem_online() || need_slab_obj_ext())
2552 free_slab_obj_exts(slab);
2553
2554 mod_node_page_state(slab_pgdat(slab), cache_vmstat_idx(s),
2555 -(PAGE_SIZE << order));
2556 }
2557
allocate_slab(struct kmem_cache * s,gfp_t flags,int node)2558 static struct slab *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
2559 {
2560 struct slab *slab;
2561 struct kmem_cache_order_objects oo = s->oo;
2562 gfp_t alloc_gfp;
2563 void *start, *p, *next;
2564 int idx;
2565 bool shuffle;
2566
2567 flags &= gfp_allowed_mask;
2568
2569 flags |= s->allocflags;
2570
2571 /*
2572 * Let the initial higher-order allocation fail under memory pressure
2573 * so we fall-back to the minimum order allocation.
2574 */
2575 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
2576 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
2577 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~__GFP_RECLAIM;
2578
2579 slab = alloc_slab_page(alloc_gfp, node, oo);
2580 if (unlikely(!slab)) {
2581 oo = s->min;
2582 alloc_gfp = flags;
2583 /*
2584 * Allocation may have failed due to fragmentation.
2585 * Try a lower order alloc if possible
2586 */
2587 slab = alloc_slab_page(alloc_gfp, node, oo);
2588 if (unlikely(!slab))
2589 return NULL;
2590 stat(s, ORDER_FALLBACK);
2591 }
2592
2593 slab->objects = oo_objects(oo);
2594 slab->inuse = 0;
2595 slab->frozen = 0;
2596
2597 account_slab(slab, oo_order(oo), s, flags);
2598
2599 slab->slab_cache = s;
2600
2601 kasan_poison_slab(slab);
2602
2603 start = slab_address(slab);
2604
2605 setup_slab_debug(s, slab, start);
2606
2607 shuffle = shuffle_freelist(s, slab);
2608
2609 if (!shuffle) {
2610 start = fixup_red_left(s, start);
2611 start = setup_object(s, start);
2612 slab->freelist = start;
2613 for (idx = 0, p = start; idx < slab->objects - 1; idx++) {
2614 next = p + s->size;
2615 next = setup_object(s, next);
2616 set_freepointer(s, p, next);
2617 p = next;
2618 }
2619 set_freepointer(s, p, NULL);
2620 }
2621
2622 return slab;
2623 }
2624
new_slab(struct kmem_cache * s,gfp_t flags,int node)2625 static struct slab *new_slab(struct kmem_cache *s, gfp_t flags, int node)
2626 {
2627 if (unlikely(flags & GFP_SLAB_BUG_MASK))
2628 flags = kmalloc_fix_flags(flags);
2629
2630 WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO));
2631
2632 return allocate_slab(s,
2633 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
2634 }
2635
__free_slab(struct kmem_cache * s,struct slab * slab)2636 static void __free_slab(struct kmem_cache *s, struct slab *slab)
2637 {
2638 struct folio *folio = slab_folio(slab);
2639 int order = folio_order(folio);
2640 int pages = 1 << order;
2641
2642 __slab_clear_pfmemalloc(slab);
2643 folio->mapping = NULL;
2644 /* Make the mapping reset visible before clearing the flag */
2645 smp_wmb();
2646 __folio_clear_slab(folio);
2647 mm_account_reclaimed_pages(pages);
2648 unaccount_slab(slab, order, s);
2649 __free_pages(&folio->page, order);
2650 }
2651
rcu_free_slab(struct rcu_head * h)2652 static void rcu_free_slab(struct rcu_head *h)
2653 {
2654 struct slab *slab = container_of(h, struct slab, rcu_head);
2655
2656 __free_slab(slab->slab_cache, slab);
2657 }
2658
free_slab(struct kmem_cache * s,struct slab * slab)2659 static void free_slab(struct kmem_cache *s, struct slab *slab)
2660 {
2661 if (kmem_cache_debug_flags(s, SLAB_CONSISTENCY_CHECKS)) {
2662 void *p;
2663
2664 slab_pad_check(s, slab);
2665 for_each_object(p, s, slab_address(slab), slab->objects)
2666 check_object(s, slab, p, SLUB_RED_INACTIVE);
2667 }
2668
2669 if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU))
2670 call_rcu(&slab->rcu_head, rcu_free_slab);
2671 else
2672 __free_slab(s, slab);
2673 }
2674
discard_slab(struct kmem_cache * s,struct slab * slab)2675 static void discard_slab(struct kmem_cache *s, struct slab *slab)
2676 {
2677 dec_slabs_node(s, slab_nid(slab), slab->objects);
2678 free_slab(s, slab);
2679 }
2680
2681 /*
2682 * SLUB reuses PG_workingset bit to keep track of whether it's on
2683 * the per-node partial list.
2684 */
slab_test_node_partial(const struct slab * slab)2685 static inline bool slab_test_node_partial(const struct slab *slab)
2686 {
2687 return folio_test_workingset(slab_folio(slab));
2688 }
2689
slab_set_node_partial(struct slab * slab)2690 static inline void slab_set_node_partial(struct slab *slab)
2691 {
2692 set_bit(PG_workingset, folio_flags(slab_folio(slab), 0));
2693 }
2694
slab_clear_node_partial(struct slab * slab)2695 static inline void slab_clear_node_partial(struct slab *slab)
2696 {
2697 clear_bit(PG_workingset, folio_flags(slab_folio(slab), 0));
2698 }
2699
2700 /*
2701 * Management of partially allocated slabs.
2702 */
2703 static inline void
__add_partial(struct kmem_cache_node * n,struct slab * slab,int tail)2704 __add_partial(struct kmem_cache_node *n, struct slab *slab, int tail)
2705 {
2706 n->nr_partial++;
2707 if (tail == DEACTIVATE_TO_TAIL)
2708 list_add_tail(&slab->slab_list, &n->partial);
2709 else
2710 list_add(&slab->slab_list, &n->partial);
2711 slab_set_node_partial(slab);
2712 }
2713
add_partial(struct kmem_cache_node * n,struct slab * slab,int tail)2714 static inline void add_partial(struct kmem_cache_node *n,
2715 struct slab *slab, int tail)
2716 {
2717 lockdep_assert_held(&n->list_lock);
2718 __add_partial(n, slab, tail);
2719 }
2720
remove_partial(struct kmem_cache_node * n,struct slab * slab)2721 static inline void remove_partial(struct kmem_cache_node *n,
2722 struct slab *slab)
2723 {
2724 lockdep_assert_held(&n->list_lock);
2725 list_del(&slab->slab_list);
2726 slab_clear_node_partial(slab);
2727 n->nr_partial--;
2728 }
2729
2730 /*
2731 * Called only for kmem_cache_debug() caches instead of remove_partial(), with a
2732 * slab from the n->partial list. Remove only a single object from the slab, do
2733 * the alloc_debug_processing() checks and leave the slab on the list, or move
2734 * it to full list if it was the last free object.
2735 */
alloc_single_from_partial(struct kmem_cache * s,struct kmem_cache_node * n,struct slab * slab,int orig_size)2736 static void *alloc_single_from_partial(struct kmem_cache *s,
2737 struct kmem_cache_node *n, struct slab *slab, int orig_size)
2738 {
2739 void *object;
2740
2741 lockdep_assert_held(&n->list_lock);
2742
2743 object = slab->freelist;
2744 slab->freelist = get_freepointer(s, object);
2745 slab->inuse++;
2746
2747 if (!alloc_debug_processing(s, slab, object, orig_size)) {
2748 remove_partial(n, slab);
2749 return NULL;
2750 }
2751
2752 if (slab->inuse == slab->objects) {
2753 remove_partial(n, slab);
2754 add_full(s, n, slab);
2755 }
2756
2757 return object;
2758 }
2759
2760 /*
2761 * Called only for kmem_cache_debug() caches to allocate from a freshly
2762 * allocated slab. Allocate a single object instead of whole freelist
2763 * and put the slab to the partial (or full) list.
2764 */
alloc_single_from_new_slab(struct kmem_cache * s,struct slab * slab,int orig_size)2765 static void *alloc_single_from_new_slab(struct kmem_cache *s,
2766 struct slab *slab, int orig_size)
2767 {
2768 int nid = slab_nid(slab);
2769 struct kmem_cache_node *n = get_node(s, nid);
2770 unsigned long flags;
2771 void *object;
2772
2773
2774 object = slab->freelist;
2775 slab->freelist = get_freepointer(s, object);
2776 slab->inuse = 1;
2777
2778 if (!alloc_debug_processing(s, slab, object, orig_size))
2779 /*
2780 * It's not really expected that this would fail on a
2781 * freshly allocated slab, but a concurrent memory
2782 * corruption in theory could cause that.
2783 */
2784 return NULL;
2785
2786 spin_lock_irqsave(&n->list_lock, flags);
2787
2788 if (slab->inuse == slab->objects)
2789 add_full(s, n, slab);
2790 else
2791 add_partial(n, slab, DEACTIVATE_TO_HEAD);
2792
2793 inc_slabs_node(s, nid, slab->objects);
2794 spin_unlock_irqrestore(&n->list_lock, flags);
2795
2796 return object;
2797 }
2798
2799 #ifdef CONFIG_SLUB_CPU_PARTIAL
2800 static void put_cpu_partial(struct kmem_cache *s, struct slab *slab, int drain);
2801 #else
put_cpu_partial(struct kmem_cache * s,struct slab * slab,int drain)2802 static inline void put_cpu_partial(struct kmem_cache *s, struct slab *slab,
2803 int drain) { }
2804 #endif
2805 static inline bool pfmemalloc_match(struct slab *slab, gfp_t gfpflags);
2806
2807 /*
2808 * Try to allocate a partial slab from a specific node.
2809 */
get_partial_node(struct kmem_cache * s,struct kmem_cache_node * n,struct partial_context * pc)2810 static struct slab *get_partial_node(struct kmem_cache *s,
2811 struct kmem_cache_node *n,
2812 struct partial_context *pc)
2813 {
2814 struct slab *slab, *slab2, *partial = NULL;
2815 unsigned long flags;
2816 unsigned int partial_slabs = 0;
2817
2818 /*
2819 * Racy check. If we mistakenly see no partial slabs then we
2820 * just allocate an empty slab. If we mistakenly try to get a
2821 * partial slab and there is none available then get_partial()
2822 * will return NULL.
2823 */
2824 if (!n || !n->nr_partial)
2825 return NULL;
2826
2827 spin_lock_irqsave(&n->list_lock, flags);
2828 list_for_each_entry_safe(slab, slab2, &n->partial, slab_list) {
2829 if (!pfmemalloc_match(slab, pc->flags))
2830 continue;
2831
2832 if (IS_ENABLED(CONFIG_SLUB_TINY) || kmem_cache_debug(s)) {
2833 void *object = alloc_single_from_partial(s, n, slab,
2834 pc->orig_size);
2835 if (object) {
2836 partial = slab;
2837 pc->object = object;
2838 break;
2839 }
2840 continue;
2841 }
2842
2843 remove_partial(n, slab);
2844
2845 if (!partial) {
2846 partial = slab;
2847 stat(s, ALLOC_FROM_PARTIAL);
2848
2849 if ((slub_get_cpu_partial(s) == 0)) {
2850 break;
2851 }
2852 } else {
2853 put_cpu_partial(s, slab, 0);
2854 stat(s, CPU_PARTIAL_NODE);
2855
2856 if (++partial_slabs > slub_get_cpu_partial(s) / 2) {
2857 break;
2858 }
2859 }
2860 }
2861 spin_unlock_irqrestore(&n->list_lock, flags);
2862 return partial;
2863 }
2864
2865 /*
2866 * Get a slab from somewhere. Search in increasing NUMA distances.
2867 */
get_any_partial(struct kmem_cache * s,struct partial_context * pc)2868 static struct slab *get_any_partial(struct kmem_cache *s,
2869 struct partial_context *pc)
2870 {
2871 #ifdef CONFIG_NUMA
2872 struct zonelist *zonelist;
2873 struct zoneref *z;
2874 struct zone *zone;
2875 enum zone_type highest_zoneidx = gfp_zone(pc->flags);
2876 struct slab *slab;
2877 unsigned int cpuset_mems_cookie;
2878
2879 /*
2880 * The defrag ratio allows a configuration of the tradeoffs between
2881 * inter node defragmentation and node local allocations. A lower
2882 * defrag_ratio increases the tendency to do local allocations
2883 * instead of attempting to obtain partial slabs from other nodes.
2884 *
2885 * If the defrag_ratio is set to 0 then kmalloc() always
2886 * returns node local objects. If the ratio is higher then kmalloc()
2887 * may return off node objects because partial slabs are obtained
2888 * from other nodes and filled up.
2889 *
2890 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
2891 * (which makes defrag_ratio = 1000) then every (well almost)
2892 * allocation will first attempt to defrag slab caches on other nodes.
2893 * This means scanning over all nodes to look for partial slabs which
2894 * may be expensive if we do it every time we are trying to find a slab
2895 * with available objects.
2896 */
2897 if (!s->remote_node_defrag_ratio ||
2898 get_cycles() % 1024 > s->remote_node_defrag_ratio)
2899 return NULL;
2900
2901 do {
2902 cpuset_mems_cookie = read_mems_allowed_begin();
2903 zonelist = node_zonelist(mempolicy_slab_node(), pc->flags);
2904 for_each_zone_zonelist(zone, z, zonelist, highest_zoneidx) {
2905 struct kmem_cache_node *n;
2906
2907 n = get_node(s, zone_to_nid(zone));
2908
2909 if (n && cpuset_zone_allowed(zone, pc->flags) &&
2910 n->nr_partial > s->min_partial) {
2911 slab = get_partial_node(s, n, pc);
2912 if (slab) {
2913 /*
2914 * Don't check read_mems_allowed_retry()
2915 * here - if mems_allowed was updated in
2916 * parallel, that was a harmless race
2917 * between allocation and the cpuset
2918 * update
2919 */
2920 return slab;
2921 }
2922 }
2923 }
2924 } while (read_mems_allowed_retry(cpuset_mems_cookie));
2925 #endif /* CONFIG_NUMA */
2926 return NULL;
2927 }
2928
2929 /*
2930 * Get a partial slab, lock it and return it.
2931 */
get_partial(struct kmem_cache * s,int node,struct partial_context * pc)2932 static struct slab *get_partial(struct kmem_cache *s, int node,
2933 struct partial_context *pc)
2934 {
2935 struct slab *slab;
2936 int searchnode = node;
2937
2938 if (node == NUMA_NO_NODE)
2939 searchnode = numa_mem_id();
2940
2941 slab = get_partial_node(s, get_node(s, searchnode), pc);
2942 if (slab || (node != NUMA_NO_NODE && (pc->flags & __GFP_THISNODE)))
2943 return slab;
2944
2945 return get_any_partial(s, pc);
2946 }
2947
2948 #ifndef CONFIG_SLUB_TINY
2949
2950 #ifdef CONFIG_PREEMPTION
2951 /*
2952 * Calculate the next globally unique transaction for disambiguation
2953 * during cmpxchg. The transactions start with the cpu number and are then
2954 * incremented by CONFIG_NR_CPUS.
2955 */
2956 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
2957 #else
2958 /*
2959 * No preemption supported therefore also no need to check for
2960 * different cpus.
2961 */
2962 #define TID_STEP 1
2963 #endif /* CONFIG_PREEMPTION */
2964
next_tid(unsigned long tid)2965 static inline unsigned long next_tid(unsigned long tid)
2966 {
2967 return tid + TID_STEP;
2968 }
2969
2970 #ifdef SLUB_DEBUG_CMPXCHG
tid_to_cpu(unsigned long tid)2971 static inline unsigned int tid_to_cpu(unsigned long tid)
2972 {
2973 return tid % TID_STEP;
2974 }
2975
tid_to_event(unsigned long tid)2976 static inline unsigned long tid_to_event(unsigned long tid)
2977 {
2978 return tid / TID_STEP;
2979 }
2980 #endif
2981
init_tid(int cpu)2982 static inline unsigned int init_tid(int cpu)
2983 {
2984 return cpu;
2985 }
2986
note_cmpxchg_failure(const char * n,const struct kmem_cache * s,unsigned long tid)2987 static inline void note_cmpxchg_failure(const char *n,
2988 const struct kmem_cache *s, unsigned long tid)
2989 {
2990 #ifdef SLUB_DEBUG_CMPXCHG
2991 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
2992
2993 pr_info("%s %s: cmpxchg redo ", n, s->name);
2994
2995 #ifdef CONFIG_PREEMPTION
2996 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
2997 pr_warn("due to cpu change %d -> %d\n",
2998 tid_to_cpu(tid), tid_to_cpu(actual_tid));
2999 else
3000 #endif
3001 if (tid_to_event(tid) != tid_to_event(actual_tid))
3002 pr_warn("due to cpu running other code. Event %ld->%ld\n",
3003 tid_to_event(tid), tid_to_event(actual_tid));
3004 else
3005 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
3006 actual_tid, tid, next_tid(tid));
3007 #endif
3008 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
3009 }
3010
init_kmem_cache_cpus(struct kmem_cache * s)3011 static void init_kmem_cache_cpus(struct kmem_cache *s)
3012 {
3013 int cpu;
3014 struct kmem_cache_cpu *c;
3015
3016 for_each_possible_cpu(cpu) {
3017 c = per_cpu_ptr(s->cpu_slab, cpu);
3018 local_lock_init(&c->lock);
3019 c->tid = init_tid(cpu);
3020 }
3021 }
3022
3023 /*
3024 * Finishes removing the cpu slab. Merges cpu's freelist with slab's freelist,
3025 * unfreezes the slabs and puts it on the proper list.
3026 * Assumes the slab has been already safely taken away from kmem_cache_cpu
3027 * by the caller.
3028 */
deactivate_slab(struct kmem_cache * s,struct slab * slab,void * freelist)3029 static void deactivate_slab(struct kmem_cache *s, struct slab *slab,
3030 void *freelist)
3031 {
3032 struct kmem_cache_node *n = get_node(s, slab_nid(slab));
3033 int free_delta = 0;
3034 void *nextfree, *freelist_iter, *freelist_tail;
3035 int tail = DEACTIVATE_TO_HEAD;
3036 unsigned long flags = 0;
3037 struct slab new;
3038 struct slab old;
3039
3040 if (READ_ONCE(slab->freelist)) {
3041 stat(s, DEACTIVATE_REMOTE_FREES);
3042 tail = DEACTIVATE_TO_TAIL;
3043 }
3044
3045 /*
3046 * Stage one: Count the objects on cpu's freelist as free_delta and
3047 * remember the last object in freelist_tail for later splicing.
3048 */
3049 freelist_tail = NULL;
3050 freelist_iter = freelist;
3051 while (freelist_iter) {
3052 nextfree = get_freepointer(s, freelist_iter);
3053
3054 /*
3055 * If 'nextfree' is invalid, it is possible that the object at
3056 * 'freelist_iter' is already corrupted. So isolate all objects
3057 * starting at 'freelist_iter' by skipping them.
3058 */
3059 if (freelist_corrupted(s, slab, &freelist_iter, nextfree))
3060 break;
3061
3062 freelist_tail = freelist_iter;
3063 free_delta++;
3064
3065 freelist_iter = nextfree;
3066 }
3067
3068 /*
3069 * Stage two: Unfreeze the slab while splicing the per-cpu
3070 * freelist to the head of slab's freelist.
3071 */
3072 do {
3073 old.freelist = READ_ONCE(slab->freelist);
3074 old.counters = READ_ONCE(slab->counters);
3075 VM_BUG_ON(!old.frozen);
3076
3077 /* Determine target state of the slab */
3078 new.counters = old.counters;
3079 new.frozen = 0;
3080 if (freelist_tail) {
3081 new.inuse -= free_delta;
3082 set_freepointer(s, freelist_tail, old.freelist);
3083 new.freelist = freelist;
3084 } else {
3085 new.freelist = old.freelist;
3086 }
3087 } while (!slab_update_freelist(s, slab,
3088 old.freelist, old.counters,
3089 new.freelist, new.counters,
3090 "unfreezing slab"));
3091
3092 /*
3093 * Stage three: Manipulate the slab list based on the updated state.
3094 */
3095 if (!new.inuse && n->nr_partial >= s->min_partial) {
3096 stat(s, DEACTIVATE_EMPTY);
3097 discard_slab(s, slab);
3098 stat(s, FREE_SLAB);
3099 } else if (new.freelist) {
3100 spin_lock_irqsave(&n->list_lock, flags);
3101 add_partial(n, slab, tail);
3102 spin_unlock_irqrestore(&n->list_lock, flags);
3103 stat(s, tail);
3104 } else {
3105 stat(s, DEACTIVATE_FULL);
3106 }
3107 }
3108
3109 #ifdef CONFIG_SLUB_CPU_PARTIAL
__put_partials(struct kmem_cache * s,struct slab * partial_slab)3110 static void __put_partials(struct kmem_cache *s, struct slab *partial_slab)
3111 {
3112 struct kmem_cache_node *n = NULL, *n2 = NULL;
3113 struct slab *slab, *slab_to_discard = NULL;
3114 unsigned long flags = 0;
3115
3116 while (partial_slab) {
3117 slab = partial_slab;
3118 partial_slab = slab->next;
3119
3120 n2 = get_node(s, slab_nid(slab));
3121 if (n != n2) {
3122 if (n)
3123 spin_unlock_irqrestore(&n->list_lock, flags);
3124
3125 n = n2;
3126 spin_lock_irqsave(&n->list_lock, flags);
3127 }
3128
3129 if (unlikely(!slab->inuse && n->nr_partial >= s->min_partial)) {
3130 slab->next = slab_to_discard;
3131 slab_to_discard = slab;
3132 } else {
3133 add_partial(n, slab, DEACTIVATE_TO_TAIL);
3134 stat(s, FREE_ADD_PARTIAL);
3135 }
3136 }
3137
3138 if (n)
3139 spin_unlock_irqrestore(&n->list_lock, flags);
3140
3141 while (slab_to_discard) {
3142 slab = slab_to_discard;
3143 slab_to_discard = slab_to_discard->next;
3144
3145 stat(s, DEACTIVATE_EMPTY);
3146 discard_slab(s, slab);
3147 stat(s, FREE_SLAB);
3148 }
3149 }
3150
3151 /*
3152 * Put all the cpu partial slabs to the node partial list.
3153 */
put_partials(struct kmem_cache * s)3154 static void put_partials(struct kmem_cache *s)
3155 {
3156 struct slab *partial_slab;
3157 unsigned long flags;
3158
3159 local_lock_irqsave(&s->cpu_slab->lock, flags);
3160 partial_slab = this_cpu_read(s->cpu_slab->partial);
3161 this_cpu_write(s->cpu_slab->partial, NULL);
3162 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3163
3164 if (partial_slab)
3165 __put_partials(s, partial_slab);
3166 }
3167
put_partials_cpu(struct kmem_cache * s,struct kmem_cache_cpu * c)3168 static void put_partials_cpu(struct kmem_cache *s,
3169 struct kmem_cache_cpu *c)
3170 {
3171 struct slab *partial_slab;
3172
3173 partial_slab = slub_percpu_partial(c);
3174 c->partial = NULL;
3175
3176 if (partial_slab)
3177 __put_partials(s, partial_slab);
3178 }
3179
3180 /*
3181 * Put a slab into a partial slab slot if available.
3182 *
3183 * If we did not find a slot then simply move all the partials to the
3184 * per node partial list.
3185 */
put_cpu_partial(struct kmem_cache * s,struct slab * slab,int drain)3186 static void put_cpu_partial(struct kmem_cache *s, struct slab *slab, int drain)
3187 {
3188 struct slab *oldslab;
3189 struct slab *slab_to_put = NULL;
3190 unsigned long flags;
3191 int slabs = 0;
3192
3193 local_lock_irqsave(&s->cpu_slab->lock, flags);
3194
3195 oldslab = this_cpu_read(s->cpu_slab->partial);
3196
3197 if (oldslab) {
3198 if (drain && oldslab->slabs >= s->cpu_partial_slabs) {
3199 /*
3200 * Partial array is full. Move the existing set to the
3201 * per node partial list. Postpone the actual unfreezing
3202 * outside of the critical section.
3203 */
3204 slab_to_put = oldslab;
3205 oldslab = NULL;
3206 } else {
3207 slabs = oldslab->slabs;
3208 }
3209 }
3210
3211 slabs++;
3212
3213 slab->slabs = slabs;
3214 slab->next = oldslab;
3215
3216 this_cpu_write(s->cpu_slab->partial, slab);
3217
3218 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3219
3220 if (slab_to_put) {
3221 __put_partials(s, slab_to_put);
3222 stat(s, CPU_PARTIAL_DRAIN);
3223 }
3224 }
3225
3226 #else /* CONFIG_SLUB_CPU_PARTIAL */
3227
put_partials(struct kmem_cache * s)3228 static inline void put_partials(struct kmem_cache *s) { }
put_partials_cpu(struct kmem_cache * s,struct kmem_cache_cpu * c)3229 static inline void put_partials_cpu(struct kmem_cache *s,
3230 struct kmem_cache_cpu *c) { }
3231
3232 #endif /* CONFIG_SLUB_CPU_PARTIAL */
3233
flush_slab(struct kmem_cache * s,struct kmem_cache_cpu * c)3234 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
3235 {
3236 unsigned long flags;
3237 struct slab *slab;
3238 void *freelist;
3239
3240 local_lock_irqsave(&s->cpu_slab->lock, flags);
3241
3242 slab = c->slab;
3243 freelist = c->freelist;
3244
3245 c->slab = NULL;
3246 c->freelist = NULL;
3247 c->tid = next_tid(c->tid);
3248
3249 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3250
3251 if (slab) {
3252 deactivate_slab(s, slab, freelist);
3253 stat(s, CPUSLAB_FLUSH);
3254 }
3255 }
3256
__flush_cpu_slab(struct kmem_cache * s,int cpu)3257 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
3258 {
3259 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
3260 void *freelist = c->freelist;
3261 struct slab *slab = c->slab;
3262
3263 c->slab = NULL;
3264 c->freelist = NULL;
3265 c->tid = next_tid(c->tid);
3266
3267 if (slab) {
3268 deactivate_slab(s, slab, freelist);
3269 stat(s, CPUSLAB_FLUSH);
3270 }
3271
3272 put_partials_cpu(s, c);
3273 }
3274
3275 struct slub_flush_work {
3276 struct work_struct work;
3277 struct kmem_cache *s;
3278 bool skip;
3279 };
3280
3281 /*
3282 * Flush cpu slab.
3283 *
3284 * Called from CPU work handler with migration disabled.
3285 */
flush_cpu_slab(struct work_struct * w)3286 static void flush_cpu_slab(struct work_struct *w)
3287 {
3288 struct kmem_cache *s;
3289 struct kmem_cache_cpu *c;
3290 struct slub_flush_work *sfw;
3291
3292 sfw = container_of(w, struct slub_flush_work, work);
3293
3294 s = sfw->s;
3295 c = this_cpu_ptr(s->cpu_slab);
3296
3297 if (c->slab)
3298 flush_slab(s, c);
3299
3300 put_partials(s);
3301 }
3302
has_cpu_slab(int cpu,struct kmem_cache * s)3303 static bool has_cpu_slab(int cpu, struct kmem_cache *s)
3304 {
3305 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
3306
3307 return c->slab || slub_percpu_partial(c);
3308 }
3309
3310 static DEFINE_MUTEX(flush_lock);
3311 static DEFINE_PER_CPU(struct slub_flush_work, slub_flush);
3312
flush_all_cpus_locked(struct kmem_cache * s)3313 static void flush_all_cpus_locked(struct kmem_cache *s)
3314 {
3315 struct slub_flush_work *sfw;
3316 unsigned int cpu;
3317
3318 lockdep_assert_cpus_held();
3319 mutex_lock(&flush_lock);
3320
3321 for_each_online_cpu(cpu) {
3322 sfw = &per_cpu(slub_flush, cpu);
3323 if (!has_cpu_slab(cpu, s)) {
3324 sfw->skip = true;
3325 continue;
3326 }
3327 INIT_WORK(&sfw->work, flush_cpu_slab);
3328 sfw->skip = false;
3329 sfw->s = s;
3330 queue_work_on(cpu, flushwq, &sfw->work);
3331 }
3332
3333 for_each_online_cpu(cpu) {
3334 sfw = &per_cpu(slub_flush, cpu);
3335 if (sfw->skip)
3336 continue;
3337 flush_work(&sfw->work);
3338 }
3339
3340 mutex_unlock(&flush_lock);
3341 }
3342
flush_all(struct kmem_cache * s)3343 static void flush_all(struct kmem_cache *s)
3344 {
3345 cpus_read_lock();
3346 flush_all_cpus_locked(s);
3347 cpus_read_unlock();
3348 }
3349
3350 /*
3351 * Use the cpu notifier to insure that the cpu slabs are flushed when
3352 * necessary.
3353 */
slub_cpu_dead(unsigned int cpu)3354 static int slub_cpu_dead(unsigned int cpu)
3355 {
3356 struct kmem_cache *s;
3357
3358 mutex_lock(&slab_mutex);
3359 list_for_each_entry(s, &slab_caches, list)
3360 __flush_cpu_slab(s, cpu);
3361 mutex_unlock(&slab_mutex);
3362 return 0;
3363 }
3364
3365 #else /* CONFIG_SLUB_TINY */
flush_all_cpus_locked(struct kmem_cache * s)3366 static inline void flush_all_cpus_locked(struct kmem_cache *s) { }
flush_all(struct kmem_cache * s)3367 static inline void flush_all(struct kmem_cache *s) { }
__flush_cpu_slab(struct kmem_cache * s,int cpu)3368 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu) { }
slub_cpu_dead(unsigned int cpu)3369 static inline int slub_cpu_dead(unsigned int cpu) { return 0; }
3370 #endif /* CONFIG_SLUB_TINY */
3371
3372 /*
3373 * Check if the objects in a per cpu structure fit numa
3374 * locality expectations.
3375 */
node_match(struct slab * slab,int node)3376 static inline int node_match(struct slab *slab, int node)
3377 {
3378 #ifdef CONFIG_NUMA
3379 if (node != NUMA_NO_NODE && slab_nid(slab) != node)
3380 return 0;
3381 #endif
3382 return 1;
3383 }
3384
3385 #ifdef CONFIG_SLUB_DEBUG
count_free(struct slab * slab)3386 static int count_free(struct slab *slab)
3387 {
3388 return slab->objects - slab->inuse;
3389 }
3390
node_nr_objs(struct kmem_cache_node * n)3391 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
3392 {
3393 return atomic_long_read(&n->total_objects);
3394 }
3395
3396 /* 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)3397 static inline bool free_debug_processing(struct kmem_cache *s,
3398 struct slab *slab, void *head, void *tail, int *bulk_cnt,
3399 unsigned long addr, depot_stack_handle_t handle)
3400 {
3401 bool checks_ok = false;
3402 void *object = head;
3403 int cnt = 0;
3404
3405 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
3406 if (!check_slab(s, slab))
3407 goto out;
3408 }
3409
3410 if (slab->inuse < *bulk_cnt) {
3411 slab_err(s, slab, "Slab has %d allocated objects but %d are to be freed\n",
3412 slab->inuse, *bulk_cnt);
3413 goto out;
3414 }
3415
3416 next_object:
3417
3418 if (++cnt > *bulk_cnt)
3419 goto out_cnt;
3420
3421 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
3422 if (!free_consistency_checks(s, slab, object, addr))
3423 goto out;
3424 }
3425
3426 if (s->flags & SLAB_STORE_USER)
3427 set_track_update(s, object, TRACK_FREE, addr, handle);
3428 trace(s, slab, object, 0);
3429 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
3430 init_object(s, object, SLUB_RED_INACTIVE);
3431
3432 /* Reached end of constructed freelist yet? */
3433 if (object != tail) {
3434 object = get_freepointer(s, object);
3435 goto next_object;
3436 }
3437 checks_ok = true;
3438
3439 out_cnt:
3440 if (cnt != *bulk_cnt) {
3441 slab_err(s, slab, "Bulk free expected %d objects but found %d\n",
3442 *bulk_cnt, cnt);
3443 *bulk_cnt = cnt;
3444 }
3445
3446 out:
3447
3448 if (!checks_ok)
3449 slab_fix(s, "Object at 0x%p not freed", object);
3450
3451 return checks_ok;
3452 }
3453 #endif /* CONFIG_SLUB_DEBUG */
3454
3455 #if defined(CONFIG_SLUB_DEBUG) || defined(SLAB_SUPPORTS_SYSFS)
count_partial(struct kmem_cache_node * n,int (* get_count)(struct slab *))3456 static unsigned long count_partial(struct kmem_cache_node *n,
3457 int (*get_count)(struct slab *))
3458 {
3459 unsigned long flags;
3460 unsigned long x = 0;
3461 struct slab *slab;
3462
3463 spin_lock_irqsave(&n->list_lock, flags);
3464 list_for_each_entry(slab, &n->partial, slab_list)
3465 x += get_count(slab);
3466 spin_unlock_irqrestore(&n->list_lock, flags);
3467 return x;
3468 }
3469 #endif /* CONFIG_SLUB_DEBUG || SLAB_SUPPORTS_SYSFS */
3470
3471 #ifdef CONFIG_SLUB_DEBUG
3472 #define MAX_PARTIAL_TO_SCAN 10000
3473
count_partial_free_approx(struct kmem_cache_node * n)3474 static unsigned long count_partial_free_approx(struct kmem_cache_node *n)
3475 {
3476 unsigned long flags;
3477 unsigned long x = 0;
3478 struct slab *slab;
3479
3480 spin_lock_irqsave(&n->list_lock, flags);
3481 if (n->nr_partial <= MAX_PARTIAL_TO_SCAN) {
3482 list_for_each_entry(slab, &n->partial, slab_list)
3483 x += slab->objects - slab->inuse;
3484 } else {
3485 /*
3486 * For a long list, approximate the total count of objects in
3487 * it to meet the limit on the number of slabs to scan.
3488 * Scan from both the list's head and tail for better accuracy.
3489 */
3490 unsigned long scanned = 0;
3491
3492 list_for_each_entry(slab, &n->partial, slab_list) {
3493 x += slab->objects - slab->inuse;
3494 if (++scanned == MAX_PARTIAL_TO_SCAN / 2)
3495 break;
3496 }
3497 list_for_each_entry_reverse(slab, &n->partial, slab_list) {
3498 x += slab->objects - slab->inuse;
3499 if (++scanned == MAX_PARTIAL_TO_SCAN)
3500 break;
3501 }
3502 x = mult_frac(x, n->nr_partial, scanned);
3503 x = min(x, node_nr_objs(n));
3504 }
3505 spin_unlock_irqrestore(&n->list_lock, flags);
3506 return x;
3507 }
3508
3509 static noinline void
slab_out_of_memory(struct kmem_cache * s,gfp_t gfpflags,int nid)3510 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
3511 {
3512 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
3513 DEFAULT_RATELIMIT_BURST);
3514 int cpu = raw_smp_processor_id();
3515 int node;
3516 struct kmem_cache_node *n;
3517
3518 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
3519 return;
3520
3521 pr_warn("SLUB: Unable to allocate memory on CPU %u (of node %d) on node %d, gfp=%#x(%pGg)\n",
3522 cpu, cpu_to_node(cpu), nid, gfpflags, &gfpflags);
3523 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
3524 s->name, s->object_size, s->size, oo_order(s->oo),
3525 oo_order(s->min));
3526
3527 if (oo_order(s->min) > get_order(s->object_size))
3528 pr_warn(" %s debugging increased min order, use slab_debug=O to disable.\n",
3529 s->name);
3530
3531 for_each_kmem_cache_node(s, node, n) {
3532 unsigned long nr_slabs;
3533 unsigned long nr_objs;
3534 unsigned long nr_free;
3535
3536 nr_free = count_partial_free_approx(n);
3537 nr_slabs = node_nr_slabs(n);
3538 nr_objs = node_nr_objs(n);
3539
3540 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
3541 node, nr_slabs, nr_objs, nr_free);
3542 }
3543 }
3544 #else /* CONFIG_SLUB_DEBUG */
3545 static inline void
slab_out_of_memory(struct kmem_cache * s,gfp_t gfpflags,int nid)3546 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid) { }
3547 #endif
3548
pfmemalloc_match(struct slab * slab,gfp_t gfpflags)3549 static inline bool pfmemalloc_match(struct slab *slab, gfp_t gfpflags)
3550 {
3551 if (unlikely(slab_test_pfmemalloc(slab)))
3552 return gfp_pfmemalloc_allowed(gfpflags);
3553
3554 return true;
3555 }
3556
3557 #ifndef CONFIG_SLUB_TINY
3558 static inline bool
__update_cpu_freelist_fast(struct kmem_cache * s,void * freelist_old,void * freelist_new,unsigned long tid)3559 __update_cpu_freelist_fast(struct kmem_cache *s,
3560 void *freelist_old, void *freelist_new,
3561 unsigned long tid)
3562 {
3563 freelist_aba_t old = { .freelist = freelist_old, .counter = tid };
3564 freelist_aba_t new = { .freelist = freelist_new, .counter = next_tid(tid) };
3565
3566 return this_cpu_try_cmpxchg_freelist(s->cpu_slab->freelist_tid.full,
3567 &old.full, new.full);
3568 }
3569
3570 /*
3571 * Check the slab->freelist and either transfer the freelist to the
3572 * per cpu freelist or deactivate the slab.
3573 *
3574 * The slab is still frozen if the return value is not NULL.
3575 *
3576 * If this function returns NULL then the slab has been unfrozen.
3577 */
get_freelist(struct kmem_cache * s,struct slab * slab)3578 static inline void *get_freelist(struct kmem_cache *s, struct slab *slab)
3579 {
3580 struct slab new;
3581 unsigned long counters;
3582 void *freelist;
3583
3584 lockdep_assert_held(this_cpu_ptr(&s->cpu_slab->lock));
3585
3586 do {
3587 freelist = slab->freelist;
3588 counters = slab->counters;
3589
3590 new.counters = counters;
3591
3592 new.inuse = slab->objects;
3593 new.frozen = freelist != NULL;
3594
3595 } while (!__slab_update_freelist(s, slab,
3596 freelist, counters,
3597 NULL, new.counters,
3598 "get_freelist"));
3599
3600 return freelist;
3601 }
3602
3603 /*
3604 * Freeze the partial slab and return the pointer to the freelist.
3605 */
freeze_slab(struct kmem_cache * s,struct slab * slab)3606 static inline void *freeze_slab(struct kmem_cache *s, struct slab *slab)
3607 {
3608 struct slab new;
3609 unsigned long counters;
3610 void *freelist;
3611
3612 do {
3613 freelist = slab->freelist;
3614 counters = slab->counters;
3615
3616 new.counters = counters;
3617 VM_BUG_ON(new.frozen);
3618
3619 new.inuse = slab->objects;
3620 new.frozen = 1;
3621
3622 } while (!slab_update_freelist(s, slab,
3623 freelist, counters,
3624 NULL, new.counters,
3625 "freeze_slab"));
3626
3627 return freelist;
3628 }
3629
3630 /*
3631 * Slow path. The lockless freelist is empty or we need to perform
3632 * debugging duties.
3633 *
3634 * Processing is still very fast if new objects have been freed to the
3635 * regular freelist. In that case we simply take over the regular freelist
3636 * as the lockless freelist and zap the regular freelist.
3637 *
3638 * If that is not working then we fall back to the partial lists. We take the
3639 * first element of the freelist as the object to allocate now and move the
3640 * rest of the freelist to the lockless freelist.
3641 *
3642 * And if we were unable to get a new slab from the partial slab lists then
3643 * we need to allocate a new slab. This is the slowest path since it involves
3644 * a call to the page allocator and the setup of a new slab.
3645 *
3646 * Version of __slab_alloc to use when we know that preemption is
3647 * already disabled (which is the case for bulk allocation).
3648 */
___slab_alloc(struct kmem_cache * s,gfp_t gfpflags,int node,unsigned long addr,struct kmem_cache_cpu * c,unsigned int orig_size)3649 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
3650 unsigned long addr, struct kmem_cache_cpu *c, unsigned int orig_size)
3651 {
3652 void *freelist;
3653 struct slab *slab;
3654 unsigned long flags;
3655 struct partial_context pc;
3656 bool try_thisnode = true;
3657
3658 stat(s, ALLOC_SLOWPATH);
3659
3660 reread_slab:
3661
3662 slab = READ_ONCE(c->slab);
3663 if (!slab) {
3664 /*
3665 * if the node is not online or has no normal memory, just
3666 * ignore the node constraint
3667 */
3668 if (unlikely(node != NUMA_NO_NODE &&
3669 !node_isset(node, slab_nodes)))
3670 node = NUMA_NO_NODE;
3671 goto new_slab;
3672 }
3673
3674 if (unlikely(!node_match(slab, node))) {
3675 /*
3676 * same as above but node_match() being false already
3677 * implies node != NUMA_NO_NODE
3678 */
3679 if (!node_isset(node, slab_nodes)) {
3680 node = NUMA_NO_NODE;
3681 } else {
3682 stat(s, ALLOC_NODE_MISMATCH);
3683 goto deactivate_slab;
3684 }
3685 }
3686
3687 /*
3688 * By rights, we should be searching for a slab page that was
3689 * PFMEMALLOC but right now, we are losing the pfmemalloc
3690 * information when the page leaves the per-cpu allocator
3691 */
3692 if (unlikely(!pfmemalloc_match(slab, gfpflags)))
3693 goto deactivate_slab;
3694
3695 /* must check again c->slab in case we got preempted and it changed */
3696 local_lock_irqsave(&s->cpu_slab->lock, flags);
3697 if (unlikely(slab != c->slab)) {
3698 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3699 goto reread_slab;
3700 }
3701 freelist = c->freelist;
3702 if (freelist)
3703 goto load_freelist;
3704
3705 freelist = get_freelist(s, slab);
3706
3707 if (!freelist) {
3708 c->slab = NULL;
3709 c->tid = next_tid(c->tid);
3710 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3711 stat(s, DEACTIVATE_BYPASS);
3712 goto new_slab;
3713 }
3714
3715 stat(s, ALLOC_REFILL);
3716
3717 load_freelist:
3718
3719 lockdep_assert_held(this_cpu_ptr(&s->cpu_slab->lock));
3720
3721 /*
3722 * freelist is pointing to the list of objects to be used.
3723 * slab is pointing to the slab from which the objects are obtained.
3724 * That slab must be frozen for per cpu allocations to work.
3725 */
3726 VM_BUG_ON(!c->slab->frozen);
3727 c->freelist = get_freepointer(s, freelist);
3728 c->tid = next_tid(c->tid);
3729 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3730 return freelist;
3731
3732 deactivate_slab:
3733
3734 local_lock_irqsave(&s->cpu_slab->lock, flags);
3735 if (slab != c->slab) {
3736 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3737 goto reread_slab;
3738 }
3739 freelist = c->freelist;
3740 c->slab = NULL;
3741 c->freelist = NULL;
3742 c->tid = next_tid(c->tid);
3743 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3744 deactivate_slab(s, slab, freelist);
3745
3746 new_slab:
3747
3748 #ifdef CONFIG_SLUB_CPU_PARTIAL
3749 while (slub_percpu_partial(c)) {
3750 local_lock_irqsave(&s->cpu_slab->lock, flags);
3751 if (unlikely(c->slab)) {
3752 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3753 goto reread_slab;
3754 }
3755 if (unlikely(!slub_percpu_partial(c))) {
3756 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3757 /* we were preempted and partial list got empty */
3758 goto new_objects;
3759 }
3760
3761 slab = slub_percpu_partial(c);
3762 slub_set_percpu_partial(c, slab);
3763
3764 if (likely(node_match(slab, node) &&
3765 pfmemalloc_match(slab, gfpflags))) {
3766 c->slab = slab;
3767 freelist = get_freelist(s, slab);
3768 VM_BUG_ON(!freelist);
3769 stat(s, CPU_PARTIAL_ALLOC);
3770 goto load_freelist;
3771 }
3772
3773 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3774
3775 slab->next = NULL;
3776 __put_partials(s, slab);
3777 }
3778 #endif
3779
3780 new_objects:
3781
3782 pc.flags = gfpflags;
3783 /*
3784 * When a preferred node is indicated but no __GFP_THISNODE
3785 *
3786 * 1) try to get a partial slab from target node only by having
3787 * __GFP_THISNODE in pc.flags for get_partial()
3788 * 2) if 1) failed, try to allocate a new slab from target node with
3789 * GPF_NOWAIT | __GFP_THISNODE opportunistically
3790 * 3) if 2) failed, retry with original gfpflags which will allow
3791 * get_partial() try partial lists of other nodes before potentially
3792 * allocating new page from other nodes
3793 */
3794 if (unlikely(node != NUMA_NO_NODE && !(gfpflags & __GFP_THISNODE)
3795 && try_thisnode))
3796 pc.flags = GFP_NOWAIT | __GFP_THISNODE;
3797
3798 pc.orig_size = orig_size;
3799 slab = get_partial(s, node, &pc);
3800 if (slab) {
3801 if (kmem_cache_debug(s)) {
3802 freelist = pc.object;
3803 /*
3804 * For debug caches here we had to go through
3805 * alloc_single_from_partial() so just store the
3806 * tracking info and return the object.
3807 */
3808 if (s->flags & SLAB_STORE_USER)
3809 set_track(s, freelist, TRACK_ALLOC, addr);
3810
3811 return freelist;
3812 }
3813
3814 freelist = freeze_slab(s, slab);
3815 goto retry_load_slab;
3816 }
3817
3818 slub_put_cpu_ptr(s->cpu_slab);
3819 slab = new_slab(s, pc.flags, node);
3820 c = slub_get_cpu_ptr(s->cpu_slab);
3821
3822 if (unlikely(!slab)) {
3823 if (node != NUMA_NO_NODE && !(gfpflags & __GFP_THISNODE)
3824 && try_thisnode) {
3825 try_thisnode = false;
3826 goto new_objects;
3827 }
3828 slab_out_of_memory(s, gfpflags, node);
3829 return NULL;
3830 }
3831
3832 stat(s, ALLOC_SLAB);
3833
3834 if (kmem_cache_debug(s)) {
3835 freelist = alloc_single_from_new_slab(s, slab, orig_size);
3836
3837 if (unlikely(!freelist))
3838 goto new_objects;
3839
3840 if (s->flags & SLAB_STORE_USER)
3841 set_track(s, freelist, TRACK_ALLOC, addr);
3842
3843 return freelist;
3844 }
3845
3846 /*
3847 * No other reference to the slab yet so we can
3848 * muck around with it freely without cmpxchg
3849 */
3850 freelist = slab->freelist;
3851 slab->freelist = NULL;
3852 slab->inuse = slab->objects;
3853 slab->frozen = 1;
3854
3855 inc_slabs_node(s, slab_nid(slab), slab->objects);
3856
3857 if (unlikely(!pfmemalloc_match(slab, gfpflags))) {
3858 /*
3859 * For !pfmemalloc_match() case we don't load freelist so that
3860 * we don't make further mismatched allocations easier.
3861 */
3862 deactivate_slab(s, slab, get_freepointer(s, freelist));
3863 return freelist;
3864 }
3865
3866 retry_load_slab:
3867
3868 local_lock_irqsave(&s->cpu_slab->lock, flags);
3869 if (unlikely(c->slab)) {
3870 void *flush_freelist = c->freelist;
3871 struct slab *flush_slab = c->slab;
3872
3873 c->slab = NULL;
3874 c->freelist = NULL;
3875 c->tid = next_tid(c->tid);
3876
3877 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3878
3879 deactivate_slab(s, flush_slab, flush_freelist);
3880
3881 stat(s, CPUSLAB_FLUSH);
3882
3883 goto retry_load_slab;
3884 }
3885 c->slab = slab;
3886
3887 goto load_freelist;
3888 }
3889
3890 /*
3891 * A wrapper for ___slab_alloc() for contexts where preemption is not yet
3892 * disabled. Compensates for possible cpu changes by refetching the per cpu area
3893 * pointer.
3894 */
__slab_alloc(struct kmem_cache * s,gfp_t gfpflags,int node,unsigned long addr,struct kmem_cache_cpu * c,unsigned int orig_size)3895 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
3896 unsigned long addr, struct kmem_cache_cpu *c, unsigned int orig_size)
3897 {
3898 void *p;
3899
3900 #ifdef CONFIG_PREEMPT_COUNT
3901 /*
3902 * We may have been preempted and rescheduled on a different
3903 * cpu before disabling preemption. Need to reload cpu area
3904 * pointer.
3905 */
3906 c = slub_get_cpu_ptr(s->cpu_slab);
3907 #endif
3908
3909 p = ___slab_alloc(s, gfpflags, node, addr, c, orig_size);
3910 #ifdef CONFIG_PREEMPT_COUNT
3911 slub_put_cpu_ptr(s->cpu_slab);
3912 #endif
3913 return p;
3914 }
3915
__slab_alloc_node(struct kmem_cache * s,gfp_t gfpflags,int node,unsigned long addr,size_t orig_size)3916 static __always_inline void *__slab_alloc_node(struct kmem_cache *s,
3917 gfp_t gfpflags, int node, unsigned long addr, size_t orig_size)
3918 {
3919 struct kmem_cache_cpu *c;
3920 struct slab *slab;
3921 unsigned long tid;
3922 void *object;
3923
3924 redo:
3925 /*
3926 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
3927 * enabled. We may switch back and forth between cpus while
3928 * reading from one cpu area. That does not matter as long
3929 * as we end up on the original cpu again when doing the cmpxchg.
3930 *
3931 * We must guarantee that tid and kmem_cache_cpu are retrieved on the
3932 * same cpu. We read first the kmem_cache_cpu pointer and use it to read
3933 * the tid. If we are preempted and switched to another cpu between the
3934 * two reads, it's OK as the two are still associated with the same cpu
3935 * and cmpxchg later will validate the cpu.
3936 */
3937 c = raw_cpu_ptr(s->cpu_slab);
3938 tid = READ_ONCE(c->tid);
3939
3940 /*
3941 * Irqless object alloc/free algorithm used here depends on sequence
3942 * of fetching cpu_slab's data. tid should be fetched before anything
3943 * on c to guarantee that object and slab associated with previous tid
3944 * won't be used with current tid. If we fetch tid first, object and
3945 * slab could be one associated with next tid and our alloc/free
3946 * request will be failed. In this case, we will retry. So, no problem.
3947 */
3948 barrier();
3949
3950 /*
3951 * The transaction ids are globally unique per cpu and per operation on
3952 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
3953 * occurs on the right processor and that there was no operation on the
3954 * linked list in between.
3955 */
3956
3957 object = c->freelist;
3958 slab = c->slab;
3959
3960 if (!USE_LOCKLESS_FAST_PATH() ||
3961 unlikely(!object || !slab || !node_match(slab, node))) {
3962 object = __slab_alloc(s, gfpflags, node, addr, c, orig_size);
3963 } else {
3964 void *next_object = get_freepointer_safe(s, object);
3965
3966 /*
3967 * The cmpxchg will only match if there was no additional
3968 * operation and if we are on the right processor.
3969 *
3970 * The cmpxchg does the following atomically (without lock
3971 * semantics!)
3972 * 1. Relocate first pointer to the current per cpu area.
3973 * 2. Verify that tid and freelist have not been changed
3974 * 3. If they were not changed replace tid and freelist
3975 *
3976 * Since this is without lock semantics the protection is only
3977 * against code executing on this cpu *not* from access by
3978 * other cpus.
3979 */
3980 if (unlikely(!__update_cpu_freelist_fast(s, object, next_object, tid))) {
3981 note_cmpxchg_failure("slab_alloc", s, tid);
3982 goto redo;
3983 }
3984 prefetch_freepointer(s, next_object);
3985 stat(s, ALLOC_FASTPATH);
3986 }
3987
3988 return object;
3989 }
3990 #else /* CONFIG_SLUB_TINY */
__slab_alloc_node(struct kmem_cache * s,gfp_t gfpflags,int node,unsigned long addr,size_t orig_size)3991 static void *__slab_alloc_node(struct kmem_cache *s,
3992 gfp_t gfpflags, int node, unsigned long addr, size_t orig_size)
3993 {
3994 struct partial_context pc;
3995 struct slab *slab;
3996 void *object;
3997
3998 pc.flags = gfpflags;
3999 pc.orig_size = orig_size;
4000 slab = get_partial(s, node, &pc);
4001
4002 if (slab)
4003 return pc.object;
4004
4005 slab = new_slab(s, gfpflags, node);
4006 if (unlikely(!slab)) {
4007 slab_out_of_memory(s, gfpflags, node);
4008 return NULL;
4009 }
4010
4011 object = alloc_single_from_new_slab(s, slab, orig_size);
4012
4013 return object;
4014 }
4015 #endif /* CONFIG_SLUB_TINY */
4016
4017 /*
4018 * If the object has been wiped upon free, make sure it's fully initialized by
4019 * zeroing out freelist pointer.
4020 *
4021 * Note that we also wipe custom freelist pointers.
4022 */
maybe_wipe_obj_freeptr(struct kmem_cache * s,void * obj)4023 static __always_inline void maybe_wipe_obj_freeptr(struct kmem_cache *s,
4024 void *obj)
4025 {
4026 if (unlikely(slab_want_init_on_free(s)) && obj &&
4027 !freeptr_outside_object(s))
4028 memset((void *)((char *)kasan_reset_tag(obj) + s->offset),
4029 0, sizeof(void *));
4030 }
4031
4032 static __fastpath_inline
slab_pre_alloc_hook(struct kmem_cache * s,gfp_t flags)4033 struct kmem_cache *slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
4034 {
4035 flags &= gfp_allowed_mask;
4036
4037 might_alloc(flags);
4038
4039 if (unlikely(should_failslab(s, flags)))
4040 return NULL;
4041
4042 return s;
4043 }
4044
4045 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)4046 bool slab_post_alloc_hook(struct kmem_cache *s, struct list_lru *lru,
4047 gfp_t flags, size_t size, void **p, bool init,
4048 unsigned int orig_size)
4049 {
4050 unsigned int zero_size = s->object_size;
4051 bool kasan_init = init;
4052 size_t i;
4053 gfp_t init_flags = flags & gfp_allowed_mask;
4054
4055 /*
4056 * For kmalloc object, the allocated memory size(object_size) is likely
4057 * larger than the requested size(orig_size). If redzone check is
4058 * enabled for the extra space, don't zero it, as it will be redzoned
4059 * soon. The redzone operation for this extra space could be seen as a
4060 * replacement of current poisoning under certain debug option, and
4061 * won't break other sanity checks.
4062 */
4063 if (kmem_cache_debug_flags(s, SLAB_STORE_USER | SLAB_RED_ZONE) &&
4064 (s->flags & SLAB_KMALLOC))
4065 zero_size = orig_size;
4066
4067 /*
4068 * When slab_debug is enabled, avoid memory initialization integrated
4069 * into KASAN and instead zero out the memory via the memset below with
4070 * the proper size. Otherwise, KASAN might overwrite SLUB redzones and
4071 * cause false-positive reports. This does not lead to a performance
4072 * penalty on production builds, as slab_debug is not intended to be
4073 * enabled there.
4074 */
4075 if (__slub_debug_enabled())
4076 kasan_init = false;
4077
4078 /*
4079 * As memory initialization might be integrated into KASAN,
4080 * kasan_slab_alloc and initialization memset must be
4081 * kept together to avoid discrepancies in behavior.
4082 *
4083 * As p[i] might get tagged, memset and kmemleak hook come after KASAN.
4084 */
4085 for (i = 0; i < size; i++) {
4086 p[i] = kasan_slab_alloc(s, p[i], init_flags, kasan_init);
4087 if (p[i] && init && (!kasan_init ||
4088 !kasan_has_integrated_init()))
4089 memset(p[i], 0, zero_size);
4090 kmemleak_alloc_recursive(p[i], s->object_size, 1,
4091 s->flags, init_flags);
4092 kmsan_slab_alloc(s, p[i], init_flags);
4093 alloc_tagging_slab_alloc_hook(s, p[i], flags);
4094 }
4095
4096 return memcg_slab_post_alloc_hook(s, lru, flags, size, p);
4097 }
4098
4099 /*
4100 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
4101 * have the fastpath folded into their functions. So no function call
4102 * overhead for requests that can be satisfied on the fastpath.
4103 *
4104 * The fastpath works by first checking if the lockless freelist can be used.
4105 * If not then __slab_alloc is called for slow processing.
4106 *
4107 * Otherwise we can simply pick the next object from the lockless free list.
4108 */
slab_alloc_node(struct kmem_cache * s,struct list_lru * lru,gfp_t gfpflags,int node,unsigned long addr,size_t orig_size)4109 static __fastpath_inline void *slab_alloc_node(struct kmem_cache *s, struct list_lru *lru,
4110 gfp_t gfpflags, int node, unsigned long addr, size_t orig_size)
4111 {
4112 void *object;
4113 bool init = false;
4114
4115 s = slab_pre_alloc_hook(s, gfpflags);
4116 if (unlikely(!s))
4117 return NULL;
4118
4119 object = kfence_alloc(s, orig_size, gfpflags);
4120 if (unlikely(object))
4121 goto out;
4122
4123 object = __slab_alloc_node(s, gfpflags, node, addr, orig_size);
4124
4125 maybe_wipe_obj_freeptr(s, object);
4126 init = slab_want_init_on_alloc(gfpflags, s);
4127
4128 out:
4129 /*
4130 * When init equals 'true', like for kzalloc() family, only
4131 * @orig_size bytes might be zeroed instead of s->object_size
4132 * In case this fails due to memcg_slab_post_alloc_hook(),
4133 * object is set to NULL
4134 */
4135 slab_post_alloc_hook(s, lru, gfpflags, 1, &object, init, orig_size);
4136
4137 return object;
4138 }
4139
kmem_cache_alloc_noprof(struct kmem_cache * s,gfp_t gfpflags)4140 void *kmem_cache_alloc_noprof(struct kmem_cache *s, gfp_t gfpflags)
4141 {
4142 void *ret = slab_alloc_node(s, NULL, gfpflags, NUMA_NO_NODE, _RET_IP_,
4143 s->object_size);
4144
4145 trace_kmem_cache_alloc(_RET_IP_, ret, s, gfpflags, NUMA_NO_NODE);
4146
4147 return ret;
4148 }
4149 EXPORT_SYMBOL(kmem_cache_alloc_noprof);
4150
kmem_cache_alloc_lru_noprof(struct kmem_cache * s,struct list_lru * lru,gfp_t gfpflags)4151 void *kmem_cache_alloc_lru_noprof(struct kmem_cache *s, struct list_lru *lru,
4152 gfp_t gfpflags)
4153 {
4154 void *ret = slab_alloc_node(s, lru, gfpflags, NUMA_NO_NODE, _RET_IP_,
4155 s->object_size);
4156
4157 trace_kmem_cache_alloc(_RET_IP_, ret, s, gfpflags, NUMA_NO_NODE);
4158
4159 return ret;
4160 }
4161 EXPORT_SYMBOL(kmem_cache_alloc_lru_noprof);
4162
kmem_cache_charge(void * objp,gfp_t gfpflags)4163 bool kmem_cache_charge(void *objp, gfp_t gfpflags)
4164 {
4165 if (!memcg_kmem_online())
4166 return true;
4167
4168 return memcg_slab_post_charge(objp, gfpflags);
4169 }
4170 EXPORT_SYMBOL(kmem_cache_charge);
4171
4172 /**
4173 * kmem_cache_alloc_node - Allocate an object on the specified node
4174 * @s: The cache to allocate from.
4175 * @gfpflags: See kmalloc().
4176 * @node: node number of the target node.
4177 *
4178 * Identical to kmem_cache_alloc but it will allocate memory on the given
4179 * node, which can improve the performance for cpu bound structures.
4180 *
4181 * Fallback to other node is possible if __GFP_THISNODE is not set.
4182 *
4183 * Return: pointer to the new object or %NULL in case of error
4184 */
kmem_cache_alloc_node_noprof(struct kmem_cache * s,gfp_t gfpflags,int node)4185 void *kmem_cache_alloc_node_noprof(struct kmem_cache *s, gfp_t gfpflags, int node)
4186 {
4187 void *ret = slab_alloc_node(s, NULL, gfpflags, node, _RET_IP_, s->object_size);
4188
4189 trace_kmem_cache_alloc(_RET_IP_, ret, s, gfpflags, node);
4190
4191 return ret;
4192 }
4193 EXPORT_SYMBOL(kmem_cache_alloc_node_noprof);
4194
4195 /*
4196 * To avoid unnecessary overhead, we pass through large allocation requests
4197 * directly to the page allocator. We use __GFP_COMP, because we will need to
4198 * know the allocation order to free the pages properly in kfree.
4199 */
___kmalloc_large_node(size_t size,gfp_t flags,int node)4200 static void *___kmalloc_large_node(size_t size, gfp_t flags, int node)
4201 {
4202 struct folio *folio;
4203 void *ptr = NULL;
4204 unsigned int order = get_order(size);
4205
4206 if (unlikely(flags & GFP_SLAB_BUG_MASK))
4207 flags = kmalloc_fix_flags(flags);
4208
4209 flags |= __GFP_COMP;
4210 folio = (struct folio *)alloc_pages_node_noprof(node, flags, order);
4211 if (folio) {
4212 ptr = folio_address(folio);
4213 lruvec_stat_mod_folio(folio, NR_SLAB_UNRECLAIMABLE_B,
4214 PAGE_SIZE << order);
4215 }
4216
4217 ptr = kasan_kmalloc_large(ptr, size, flags);
4218 /* As ptr might get tagged, call kmemleak hook after KASAN. */
4219 kmemleak_alloc(ptr, size, 1, flags);
4220 kmsan_kmalloc_large(ptr, size, flags);
4221
4222 return ptr;
4223 }
4224
__kmalloc_large_noprof(size_t size,gfp_t flags)4225 void *__kmalloc_large_noprof(size_t size, gfp_t flags)
4226 {
4227 void *ret = ___kmalloc_large_node(size, flags, NUMA_NO_NODE);
4228
4229 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << get_order(size),
4230 flags, NUMA_NO_NODE);
4231 return ret;
4232 }
4233 EXPORT_SYMBOL(__kmalloc_large_noprof);
4234
__kmalloc_large_node_noprof(size_t size,gfp_t flags,int node)4235 void *__kmalloc_large_node_noprof(size_t size, gfp_t flags, int node)
4236 {
4237 void *ret = ___kmalloc_large_node(size, flags, node);
4238
4239 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << get_order(size),
4240 flags, node);
4241 return ret;
4242 }
4243 EXPORT_SYMBOL(__kmalloc_large_node_noprof);
4244
4245 static __always_inline
__do_kmalloc_node(size_t size,kmem_buckets * b,gfp_t flags,int node,unsigned long caller)4246 void *__do_kmalloc_node(size_t size, kmem_buckets *b, gfp_t flags, int node,
4247 unsigned long caller)
4248 {
4249 struct kmem_cache *s;
4250 void *ret;
4251
4252 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4253 ret = __kmalloc_large_node_noprof(size, flags, node);
4254 trace_kmalloc(caller, ret, size,
4255 PAGE_SIZE << get_order(size), flags, node);
4256 return ret;
4257 }
4258
4259 if (unlikely(!size))
4260 return ZERO_SIZE_PTR;
4261
4262 s = kmalloc_slab(size, b, flags, caller);
4263
4264 ret = slab_alloc_node(s, NULL, flags, node, caller, size);
4265 ret = kasan_kmalloc(s, ret, size, flags);
4266 trace_kmalloc(caller, ret, size, s->size, flags, node);
4267 return ret;
4268 }
__kmalloc_node_noprof(DECL_BUCKET_PARAMS (size,b),gfp_t flags,int node)4269 void *__kmalloc_node_noprof(DECL_BUCKET_PARAMS(size, b), gfp_t flags, int node)
4270 {
4271 return __do_kmalloc_node(size, PASS_BUCKET_PARAM(b), flags, node, _RET_IP_);
4272 }
4273 EXPORT_SYMBOL(__kmalloc_node_noprof);
4274
__kmalloc_noprof(size_t size,gfp_t flags)4275 void *__kmalloc_noprof(size_t size, gfp_t flags)
4276 {
4277 return __do_kmalloc_node(size, NULL, flags, NUMA_NO_NODE, _RET_IP_);
4278 }
4279 EXPORT_SYMBOL(__kmalloc_noprof);
4280
__kmalloc_node_track_caller_noprof(DECL_BUCKET_PARAMS (size,b),gfp_t flags,int node,unsigned long caller)4281 void *__kmalloc_node_track_caller_noprof(DECL_BUCKET_PARAMS(size, b), gfp_t flags,
4282 int node, unsigned long caller)
4283 {
4284 return __do_kmalloc_node(size, PASS_BUCKET_PARAM(b), flags, node, caller);
4285
4286 }
4287 EXPORT_SYMBOL(__kmalloc_node_track_caller_noprof);
4288
__kmalloc_cache_noprof(struct kmem_cache * s,gfp_t gfpflags,size_t size)4289 void *__kmalloc_cache_noprof(struct kmem_cache *s, gfp_t gfpflags, size_t size)
4290 {
4291 void *ret = slab_alloc_node(s, NULL, gfpflags, NUMA_NO_NODE,
4292 _RET_IP_, size);
4293
4294 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags, NUMA_NO_NODE);
4295
4296 ret = kasan_kmalloc(s, ret, size, gfpflags);
4297 return ret;
4298 }
4299 EXPORT_SYMBOL(__kmalloc_cache_noprof);
4300
__kmalloc_cache_node_noprof(struct kmem_cache * s,gfp_t gfpflags,int node,size_t size)4301 void *__kmalloc_cache_node_noprof(struct kmem_cache *s, gfp_t gfpflags,
4302 int node, size_t size)
4303 {
4304 void *ret = slab_alloc_node(s, NULL, gfpflags, node, _RET_IP_, size);
4305
4306 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags, node);
4307
4308 ret = kasan_kmalloc(s, ret, size, gfpflags);
4309 return ret;
4310 }
4311 EXPORT_SYMBOL(__kmalloc_cache_node_noprof);
4312
free_to_partial_list(struct kmem_cache * s,struct slab * slab,void * head,void * tail,int bulk_cnt,unsigned long addr)4313 static noinline void free_to_partial_list(
4314 struct kmem_cache *s, struct slab *slab,
4315 void *head, void *tail, int bulk_cnt,
4316 unsigned long addr)
4317 {
4318 struct kmem_cache_node *n = get_node(s, slab_nid(slab));
4319 struct slab *slab_free = NULL;
4320 int cnt = bulk_cnt;
4321 unsigned long flags;
4322 depot_stack_handle_t handle = 0;
4323
4324 if (s->flags & SLAB_STORE_USER)
4325 handle = set_track_prepare();
4326
4327 spin_lock_irqsave(&n->list_lock, flags);
4328
4329 if (free_debug_processing(s, slab, head, tail, &cnt, addr, handle)) {
4330 void *prior = slab->freelist;
4331
4332 /* Perform the actual freeing while we still hold the locks */
4333 slab->inuse -= cnt;
4334 set_freepointer(s, tail, prior);
4335 slab->freelist = head;
4336
4337 /*
4338 * If the slab is empty, and node's partial list is full,
4339 * it should be discarded anyway no matter it's on full or
4340 * partial list.
4341 */
4342 if (slab->inuse == 0 && n->nr_partial >= s->min_partial)
4343 slab_free = slab;
4344
4345 if (!prior) {
4346 /* was on full list */
4347 remove_full(s, n, slab);
4348 if (!slab_free) {
4349 add_partial(n, slab, DEACTIVATE_TO_TAIL);
4350 stat(s, FREE_ADD_PARTIAL);
4351 }
4352 } else if (slab_free) {
4353 remove_partial(n, slab);
4354 stat(s, FREE_REMOVE_PARTIAL);
4355 }
4356 }
4357
4358 if (slab_free) {
4359 /*
4360 * Update the counters while still holding n->list_lock to
4361 * prevent spurious validation warnings
4362 */
4363 dec_slabs_node(s, slab_nid(slab_free), slab_free->objects);
4364 }
4365
4366 spin_unlock_irqrestore(&n->list_lock, flags);
4367
4368 if (slab_free) {
4369 stat(s, FREE_SLAB);
4370 free_slab(s, slab_free);
4371 }
4372 }
4373
4374 /*
4375 * Slow path handling. This may still be called frequently since objects
4376 * have a longer lifetime than the cpu slabs in most processing loads.
4377 *
4378 * So we still attempt to reduce cache line usage. Just take the slab
4379 * lock and free the item. If there is no additional partial slab
4380 * handling required then we can return immediately.
4381 */
__slab_free(struct kmem_cache * s,struct slab * slab,void * head,void * tail,int cnt,unsigned long addr)4382 static void __slab_free(struct kmem_cache *s, struct slab *slab,
4383 void *head, void *tail, int cnt,
4384 unsigned long addr)
4385
4386 {
4387 void *prior;
4388 int was_frozen;
4389 struct slab new;
4390 unsigned long counters;
4391 struct kmem_cache_node *n = NULL;
4392 unsigned long flags;
4393 bool on_node_partial;
4394
4395 stat(s, FREE_SLOWPATH);
4396
4397 if (IS_ENABLED(CONFIG_SLUB_TINY) || kmem_cache_debug(s)) {
4398 free_to_partial_list(s, slab, head, tail, cnt, addr);
4399 return;
4400 }
4401
4402 do {
4403 if (unlikely(n)) {
4404 spin_unlock_irqrestore(&n->list_lock, flags);
4405 n = NULL;
4406 }
4407 prior = slab->freelist;
4408 counters = slab->counters;
4409 set_freepointer(s, tail, prior);
4410 new.counters = counters;
4411 was_frozen = new.frozen;
4412 new.inuse -= cnt;
4413 if ((!new.inuse || !prior) && !was_frozen) {
4414 /* Needs to be taken off a list */
4415 if (!kmem_cache_has_cpu_partial(s) || prior) {
4416
4417 n = get_node(s, slab_nid(slab));
4418 /*
4419 * Speculatively acquire the list_lock.
4420 * If the cmpxchg does not succeed then we may
4421 * drop the list_lock without any processing.
4422 *
4423 * Otherwise the list_lock will synchronize with
4424 * other processors updating the list of slabs.
4425 */
4426 spin_lock_irqsave(&n->list_lock, flags);
4427
4428 on_node_partial = slab_test_node_partial(slab);
4429 }
4430 }
4431
4432 } while (!slab_update_freelist(s, slab,
4433 prior, counters,
4434 head, new.counters,
4435 "__slab_free"));
4436
4437 if (likely(!n)) {
4438
4439 if (likely(was_frozen)) {
4440 /*
4441 * The list lock was not taken therefore no list
4442 * activity can be necessary.
4443 */
4444 stat(s, FREE_FROZEN);
4445 } else if (kmem_cache_has_cpu_partial(s) && !prior) {
4446 /*
4447 * If we started with a full slab then put it onto the
4448 * per cpu partial list.
4449 */
4450 put_cpu_partial(s, slab, 1);
4451 stat(s, CPU_PARTIAL_FREE);
4452 }
4453
4454 return;
4455 }
4456
4457 /*
4458 * This slab was partially empty but not on the per-node partial list,
4459 * in which case we shouldn't manipulate its list, just return.
4460 */
4461 if (prior && !on_node_partial) {
4462 spin_unlock_irqrestore(&n->list_lock, flags);
4463 return;
4464 }
4465
4466 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
4467 goto slab_empty;
4468
4469 /*
4470 * Objects left in the slab. If it was not on the partial list before
4471 * then add it.
4472 */
4473 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
4474 add_partial(n, slab, DEACTIVATE_TO_TAIL);
4475 stat(s, FREE_ADD_PARTIAL);
4476 }
4477 spin_unlock_irqrestore(&n->list_lock, flags);
4478 return;
4479
4480 slab_empty:
4481 if (prior) {
4482 /*
4483 * Slab on the partial list.
4484 */
4485 remove_partial(n, slab);
4486 stat(s, FREE_REMOVE_PARTIAL);
4487 }
4488
4489 spin_unlock_irqrestore(&n->list_lock, flags);
4490 stat(s, FREE_SLAB);
4491 discard_slab(s, slab);
4492 }
4493
4494 #ifndef CONFIG_SLUB_TINY
4495 /*
4496 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
4497 * can perform fastpath freeing without additional function calls.
4498 *
4499 * The fastpath is only possible if we are freeing to the current cpu slab
4500 * of this processor. This typically the case if we have just allocated
4501 * the item before.
4502 *
4503 * If fastpath is not possible then fall back to __slab_free where we deal
4504 * with all sorts of special processing.
4505 *
4506 * Bulk free of a freelist with several objects (all pointing to the
4507 * same slab) possible by specifying head and tail ptr, plus objects
4508 * count (cnt). Bulk free indicated by tail pointer being set.
4509 */
do_slab_free(struct kmem_cache * s,struct slab * slab,void * head,void * tail,int cnt,unsigned long addr)4510 static __always_inline void do_slab_free(struct kmem_cache *s,
4511 struct slab *slab, void *head, void *tail,
4512 int cnt, unsigned long addr)
4513 {
4514 struct kmem_cache_cpu *c;
4515 unsigned long tid;
4516 void **freelist;
4517
4518 redo:
4519 /*
4520 * Determine the currently cpus per cpu slab.
4521 * The cpu may change afterward. However that does not matter since
4522 * data is retrieved via this pointer. If we are on the same cpu
4523 * during the cmpxchg then the free will succeed.
4524 */
4525 c = raw_cpu_ptr(s->cpu_slab);
4526 tid = READ_ONCE(c->tid);
4527
4528 /* Same with comment on barrier() in __slab_alloc_node() */
4529 barrier();
4530
4531 if (unlikely(slab != c->slab)) {
4532 __slab_free(s, slab, head, tail, cnt, addr);
4533 return;
4534 }
4535
4536 if (USE_LOCKLESS_FAST_PATH()) {
4537 freelist = READ_ONCE(c->freelist);
4538
4539 set_freepointer(s, tail, freelist);
4540
4541 if (unlikely(!__update_cpu_freelist_fast(s, freelist, head, tid))) {
4542 note_cmpxchg_failure("slab_free", s, tid);
4543 goto redo;
4544 }
4545 } else {
4546 /* Update the free list under the local lock */
4547 local_lock(&s->cpu_slab->lock);
4548 c = this_cpu_ptr(s->cpu_slab);
4549 if (unlikely(slab != c->slab)) {
4550 local_unlock(&s->cpu_slab->lock);
4551 goto redo;
4552 }
4553 tid = c->tid;
4554 freelist = c->freelist;
4555
4556 set_freepointer(s, tail, freelist);
4557 c->freelist = head;
4558 c->tid = next_tid(tid);
4559
4560 local_unlock(&s->cpu_slab->lock);
4561 }
4562 stat_add(s, FREE_FASTPATH, cnt);
4563 }
4564 #else /* CONFIG_SLUB_TINY */
do_slab_free(struct kmem_cache * s,struct slab * slab,void * head,void * tail,int cnt,unsigned long addr)4565 static void do_slab_free(struct kmem_cache *s,
4566 struct slab *slab, void *head, void *tail,
4567 int cnt, unsigned long addr)
4568 {
4569 __slab_free(s, slab, head, tail, cnt, addr);
4570 }
4571 #endif /* CONFIG_SLUB_TINY */
4572
4573 static __fastpath_inline
slab_free(struct kmem_cache * s,struct slab * slab,void * object,unsigned long addr)4574 void slab_free(struct kmem_cache *s, struct slab *slab, void *object,
4575 unsigned long addr)
4576 {
4577 memcg_slab_free_hook(s, slab, &object, 1);
4578 alloc_tagging_slab_free_hook(s, slab, &object, 1);
4579
4580 if (likely(slab_free_hook(s, object, slab_want_init_on_free(s), false)))
4581 do_slab_free(s, slab, object, object, 1, addr);
4582 }
4583
4584 #ifdef CONFIG_MEMCG
4585 /* Do not inline the rare memcg charging failed path into the allocation path */
4586 static noinline
memcg_alloc_abort_single(struct kmem_cache * s,void * object)4587 void memcg_alloc_abort_single(struct kmem_cache *s, void *object)
4588 {
4589 if (likely(slab_free_hook(s, object, slab_want_init_on_free(s), false)))
4590 do_slab_free(s, virt_to_slab(object), object, object, 1, _RET_IP_);
4591 }
4592 #endif
4593
4594 static __fastpath_inline
slab_free_bulk(struct kmem_cache * s,struct slab * slab,void * head,void * tail,void ** p,int cnt,unsigned long addr)4595 void slab_free_bulk(struct kmem_cache *s, struct slab *slab, void *head,
4596 void *tail, void **p, int cnt, unsigned long addr)
4597 {
4598 memcg_slab_free_hook(s, slab, p, cnt);
4599 alloc_tagging_slab_free_hook(s, slab, p, cnt);
4600 /*
4601 * With KASAN enabled slab_free_freelist_hook modifies the freelist
4602 * to remove objects, whose reuse must be delayed.
4603 */
4604 if (likely(slab_free_freelist_hook(s, &head, &tail, &cnt)))
4605 do_slab_free(s, slab, head, tail, cnt, addr);
4606 }
4607
4608 #ifdef CONFIG_SLUB_RCU_DEBUG
slab_free_after_rcu_debug(struct rcu_head * rcu_head)4609 static void slab_free_after_rcu_debug(struct rcu_head *rcu_head)
4610 {
4611 struct rcu_delayed_free *delayed_free =
4612 container_of(rcu_head, struct rcu_delayed_free, head);
4613 void *object = delayed_free->object;
4614 struct slab *slab = virt_to_slab(object);
4615 struct kmem_cache *s;
4616
4617 kfree(delayed_free);
4618
4619 if (WARN_ON(is_kfence_address(object)))
4620 return;
4621
4622 /* find the object and the cache again */
4623 if (WARN_ON(!slab))
4624 return;
4625 s = slab->slab_cache;
4626 if (WARN_ON(!(s->flags & SLAB_TYPESAFE_BY_RCU)))
4627 return;
4628
4629 /* resume freeing */
4630 if (slab_free_hook(s, object, slab_want_init_on_free(s), true))
4631 do_slab_free(s, slab, object, object, 1, _THIS_IP_);
4632 }
4633 #endif /* CONFIG_SLUB_RCU_DEBUG */
4634
4635 #ifdef CONFIG_KASAN_GENERIC
___cache_free(struct kmem_cache * cache,void * x,unsigned long addr)4636 void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
4637 {
4638 do_slab_free(cache, virt_to_slab(x), x, x, 1, addr);
4639 }
4640 #endif
4641
virt_to_cache(const void * obj)4642 static inline struct kmem_cache *virt_to_cache(const void *obj)
4643 {
4644 struct slab *slab;
4645
4646 slab = virt_to_slab(obj);
4647 if (WARN_ONCE(!slab, "%s: Object is not a Slab page!\n", __func__))
4648 return NULL;
4649 return slab->slab_cache;
4650 }
4651
cache_from_obj(struct kmem_cache * s,void * x)4652 static inline struct kmem_cache *cache_from_obj(struct kmem_cache *s, void *x)
4653 {
4654 struct kmem_cache *cachep;
4655
4656 if (!IS_ENABLED(CONFIG_SLAB_FREELIST_HARDENED) &&
4657 !kmem_cache_debug_flags(s, SLAB_CONSISTENCY_CHECKS))
4658 return s;
4659
4660 cachep = virt_to_cache(x);
4661 if (WARN(cachep && cachep != s,
4662 "%s: Wrong slab cache. %s but object is from %s\n",
4663 __func__, s->name, cachep->name))
4664 print_tracking(cachep, x);
4665 return cachep;
4666 }
4667
4668 /**
4669 * kmem_cache_free - Deallocate an object
4670 * @s: The cache the allocation was from.
4671 * @x: The previously allocated object.
4672 *
4673 * Free an object which was previously allocated from this
4674 * cache.
4675 */
kmem_cache_free(struct kmem_cache * s,void * x)4676 void kmem_cache_free(struct kmem_cache *s, void *x)
4677 {
4678 s = cache_from_obj(s, x);
4679 if (!s)
4680 return;
4681 trace_kmem_cache_free(_RET_IP_, x, s);
4682 slab_free(s, virt_to_slab(x), x, _RET_IP_);
4683 }
4684 EXPORT_SYMBOL(kmem_cache_free);
4685
free_large_kmalloc(struct folio * folio,void * object)4686 static void free_large_kmalloc(struct folio *folio, void *object)
4687 {
4688 unsigned int order = folio_order(folio);
4689
4690 if (WARN_ON_ONCE(order == 0))
4691 pr_warn_once("object pointer: 0x%p\n", object);
4692
4693 kmemleak_free(object);
4694 kasan_kfree_large(object);
4695 kmsan_kfree_large(object);
4696
4697 lruvec_stat_mod_folio(folio, NR_SLAB_UNRECLAIMABLE_B,
4698 -(PAGE_SIZE << order));
4699 folio_put(folio);
4700 }
4701
4702 /**
4703 * kfree - free previously allocated memory
4704 * @object: pointer returned by kmalloc() or kmem_cache_alloc()
4705 *
4706 * If @object is NULL, no operation is performed.
4707 */
kfree(const void * object)4708 void kfree(const void *object)
4709 {
4710 struct folio *folio;
4711 struct slab *slab;
4712 struct kmem_cache *s;
4713 void *x = (void *)object;
4714
4715 trace_kfree(_RET_IP_, object);
4716
4717 if (unlikely(ZERO_OR_NULL_PTR(object)))
4718 return;
4719
4720 folio = virt_to_folio(object);
4721 if (unlikely(!folio_test_slab(folio))) {
4722 free_large_kmalloc(folio, (void *)object);
4723 return;
4724 }
4725
4726 slab = folio_slab(folio);
4727 s = slab->slab_cache;
4728 slab_free(s, slab, x, _RET_IP_);
4729 }
4730 EXPORT_SYMBOL(kfree);
4731
4732 struct detached_freelist {
4733 struct slab *slab;
4734 void *tail;
4735 void *freelist;
4736 int cnt;
4737 struct kmem_cache *s;
4738 };
4739
4740 /*
4741 * This function progressively scans the array with free objects (with
4742 * a limited look ahead) and extract objects belonging to the same
4743 * slab. It builds a detached freelist directly within the given
4744 * slab/objects. This can happen without any need for
4745 * synchronization, because the objects are owned by running process.
4746 * The freelist is build up as a single linked list in the objects.
4747 * The idea is, that this detached freelist can then be bulk
4748 * transferred to the real freelist(s), but only requiring a single
4749 * synchronization primitive. Look ahead in the array is limited due
4750 * to performance reasons.
4751 */
4752 static inline
build_detached_freelist(struct kmem_cache * s,size_t size,void ** p,struct detached_freelist * df)4753 int build_detached_freelist(struct kmem_cache *s, size_t size,
4754 void **p, struct detached_freelist *df)
4755 {
4756 int lookahead = 3;
4757 void *object;
4758 struct folio *folio;
4759 size_t same;
4760
4761 object = p[--size];
4762 folio = virt_to_folio(object);
4763 if (!s) {
4764 /* Handle kalloc'ed objects */
4765 if (unlikely(!folio_test_slab(folio))) {
4766 free_large_kmalloc(folio, object);
4767 df->slab = NULL;
4768 return size;
4769 }
4770 /* Derive kmem_cache from object */
4771 df->slab = folio_slab(folio);
4772 df->s = df->slab->slab_cache;
4773 } else {
4774 df->slab = folio_slab(folio);
4775 df->s = cache_from_obj(s, object); /* Support for memcg */
4776 }
4777
4778 /* Start new detached freelist */
4779 df->tail = object;
4780 df->freelist = object;
4781 df->cnt = 1;
4782
4783 if (is_kfence_address(object))
4784 return size;
4785
4786 set_freepointer(df->s, object, NULL);
4787
4788 same = size;
4789 while (size) {
4790 object = p[--size];
4791 /* df->slab is always set at this point */
4792 if (df->slab == virt_to_slab(object)) {
4793 /* Opportunity build freelist */
4794 set_freepointer(df->s, object, df->freelist);
4795 df->freelist = object;
4796 df->cnt++;
4797 same--;
4798 if (size != same)
4799 swap(p[size], p[same]);
4800 continue;
4801 }
4802
4803 /* Limit look ahead search */
4804 if (!--lookahead)
4805 break;
4806 }
4807
4808 return same;
4809 }
4810
4811 /*
4812 * Internal bulk free of objects that were not initialised by the post alloc
4813 * hooks and thus should not be processed by the free hooks
4814 */
__kmem_cache_free_bulk(struct kmem_cache * s,size_t size,void ** p)4815 static void __kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
4816 {
4817 if (!size)
4818 return;
4819
4820 do {
4821 struct detached_freelist df;
4822
4823 size = build_detached_freelist(s, size, p, &df);
4824 if (!df.slab)
4825 continue;
4826
4827 if (kfence_free(df.freelist))
4828 continue;
4829
4830 do_slab_free(df.s, df.slab, df.freelist, df.tail, df.cnt,
4831 _RET_IP_);
4832 } while (likely(size));
4833 }
4834
4835 /* Note that interrupts must be enabled when calling this function. */
kmem_cache_free_bulk(struct kmem_cache * s,size_t size,void ** p)4836 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
4837 {
4838 if (!size)
4839 return;
4840
4841 do {
4842 struct detached_freelist df;
4843
4844 size = build_detached_freelist(s, size, p, &df);
4845 if (!df.slab)
4846 continue;
4847
4848 slab_free_bulk(df.s, df.slab, df.freelist, df.tail, &p[size],
4849 df.cnt, _RET_IP_);
4850 } while (likely(size));
4851 }
4852 EXPORT_SYMBOL(kmem_cache_free_bulk);
4853
4854 #ifndef CONFIG_SLUB_TINY
4855 static inline
__kmem_cache_alloc_bulk(struct kmem_cache * s,gfp_t flags,size_t size,void ** p)4856 int __kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
4857 void **p)
4858 {
4859 struct kmem_cache_cpu *c;
4860 unsigned long irqflags;
4861 int i;
4862
4863 /*
4864 * Drain objects in the per cpu slab, while disabling local
4865 * IRQs, which protects against PREEMPT and interrupts
4866 * handlers invoking normal fastpath.
4867 */
4868 c = slub_get_cpu_ptr(s->cpu_slab);
4869 local_lock_irqsave(&s->cpu_slab->lock, irqflags);
4870
4871 for (i = 0; i < size; i++) {
4872 void *object = kfence_alloc(s, s->object_size, flags);
4873
4874 if (unlikely(object)) {
4875 p[i] = object;
4876 continue;
4877 }
4878
4879 object = c->freelist;
4880 if (unlikely(!object)) {
4881 /*
4882 * We may have removed an object from c->freelist using
4883 * the fastpath in the previous iteration; in that case,
4884 * c->tid has not been bumped yet.
4885 * Since ___slab_alloc() may reenable interrupts while
4886 * allocating memory, we should bump c->tid now.
4887 */
4888 c->tid = next_tid(c->tid);
4889
4890 local_unlock_irqrestore(&s->cpu_slab->lock, irqflags);
4891
4892 /*
4893 * Invoking slow path likely have side-effect
4894 * of re-populating per CPU c->freelist
4895 */
4896 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
4897 _RET_IP_, c, s->object_size);
4898 if (unlikely(!p[i]))
4899 goto error;
4900
4901 c = this_cpu_ptr(s->cpu_slab);
4902 maybe_wipe_obj_freeptr(s, p[i]);
4903
4904 local_lock_irqsave(&s->cpu_slab->lock, irqflags);
4905
4906 continue; /* goto for-loop */
4907 }
4908 c->freelist = get_freepointer(s, object);
4909 p[i] = object;
4910 maybe_wipe_obj_freeptr(s, p[i]);
4911 stat(s, ALLOC_FASTPATH);
4912 }
4913 c->tid = next_tid(c->tid);
4914 local_unlock_irqrestore(&s->cpu_slab->lock, irqflags);
4915 slub_put_cpu_ptr(s->cpu_slab);
4916
4917 return i;
4918
4919 error:
4920 slub_put_cpu_ptr(s->cpu_slab);
4921 __kmem_cache_free_bulk(s, i, p);
4922 return 0;
4923
4924 }
4925 #else /* CONFIG_SLUB_TINY */
__kmem_cache_alloc_bulk(struct kmem_cache * s,gfp_t flags,size_t size,void ** p)4926 static int __kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags,
4927 size_t size, void **p)
4928 {
4929 int i;
4930
4931 for (i = 0; i < size; i++) {
4932 void *object = kfence_alloc(s, s->object_size, flags);
4933
4934 if (unlikely(object)) {
4935 p[i] = object;
4936 continue;
4937 }
4938
4939 p[i] = __slab_alloc_node(s, flags, NUMA_NO_NODE,
4940 _RET_IP_, s->object_size);
4941 if (unlikely(!p[i]))
4942 goto error;
4943
4944 maybe_wipe_obj_freeptr(s, p[i]);
4945 }
4946
4947 return i;
4948
4949 error:
4950 __kmem_cache_free_bulk(s, i, p);
4951 return 0;
4952 }
4953 #endif /* CONFIG_SLUB_TINY */
4954
4955 /* 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)4956 int kmem_cache_alloc_bulk_noprof(struct kmem_cache *s, gfp_t flags, size_t size,
4957 void **p)
4958 {
4959 int i;
4960
4961 if (!size)
4962 return 0;
4963
4964 s = slab_pre_alloc_hook(s, flags);
4965 if (unlikely(!s))
4966 return 0;
4967
4968 i = __kmem_cache_alloc_bulk(s, flags, size, p);
4969 if (unlikely(i == 0))
4970 return 0;
4971
4972 /*
4973 * memcg and kmem_cache debug support and memory initialization.
4974 * Done outside of the IRQ disabled fastpath loop.
4975 */
4976 if (unlikely(!slab_post_alloc_hook(s, NULL, flags, size, p,
4977 slab_want_init_on_alloc(flags, s), s->object_size))) {
4978 return 0;
4979 }
4980 return i;
4981 }
4982 EXPORT_SYMBOL(kmem_cache_alloc_bulk_noprof);
4983
4984
4985 /*
4986 * Object placement in a slab is made very easy because we always start at
4987 * offset 0. If we tune the size of the object to the alignment then we can
4988 * get the required alignment by putting one properly sized object after
4989 * another.
4990 *
4991 * Notice that the allocation order determines the sizes of the per cpu
4992 * caches. Each processor has always one slab available for allocations.
4993 * Increasing the allocation order reduces the number of times that slabs
4994 * must be moved on and off the partial lists and is therefore a factor in
4995 * locking overhead.
4996 */
4997
4998 /*
4999 * Minimum / Maximum order of slab pages. This influences locking overhead
5000 * and slab fragmentation. A higher order reduces the number of partial slabs
5001 * and increases the number of allocations possible without having to
5002 * take the list_lock.
5003 */
5004 static unsigned int slub_min_order;
5005 static unsigned int slub_max_order =
5006 IS_ENABLED(CONFIG_SLUB_TINY) ? 1 : PAGE_ALLOC_COSTLY_ORDER;
5007 static unsigned int slub_min_objects;
5008
5009 /*
5010 * Calculate the order of allocation given an slab object size.
5011 *
5012 * The order of allocation has significant impact on performance and other
5013 * system components. Generally order 0 allocations should be preferred since
5014 * order 0 does not cause fragmentation in the page allocator. Larger objects
5015 * be problematic to put into order 0 slabs because there may be too much
5016 * unused space left. We go to a higher order if more than 1/16th of the slab
5017 * would be wasted.
5018 *
5019 * In order to reach satisfactory performance we must ensure that a minimum
5020 * number of objects is in one slab. Otherwise we may generate too much
5021 * activity on the partial lists which requires taking the list_lock. This is
5022 * less a concern for large slabs though which are rarely used.
5023 *
5024 * slab_max_order specifies the order where we begin to stop considering the
5025 * number of objects in a slab as critical. If we reach slab_max_order then
5026 * we try to keep the page order as low as possible. So we accept more waste
5027 * of space in favor of a small page order.
5028 *
5029 * Higher order allocations also allow the placement of more objects in a
5030 * slab and thereby reduce object handling overhead. If the user has
5031 * requested a higher minimum order then we start with that one instead of
5032 * the smallest order which will fit the object.
5033 */
calc_slab_order(unsigned int size,unsigned int min_order,unsigned int max_order,unsigned int fract_leftover)5034 static inline unsigned int calc_slab_order(unsigned int size,
5035 unsigned int min_order, unsigned int max_order,
5036 unsigned int fract_leftover)
5037 {
5038 unsigned int order;
5039
5040 for (order = min_order; order <= max_order; order++) {
5041
5042 unsigned int slab_size = (unsigned int)PAGE_SIZE << order;
5043 unsigned int rem;
5044
5045 rem = slab_size % size;
5046
5047 if (rem <= slab_size / fract_leftover)
5048 break;
5049 }
5050
5051 return order;
5052 }
5053
calculate_order(unsigned int size)5054 static inline int calculate_order(unsigned int size)
5055 {
5056 unsigned int order;
5057 unsigned int min_objects;
5058 unsigned int max_objects;
5059 unsigned int min_order;
5060
5061 min_objects = slub_min_objects;
5062 if (!min_objects) {
5063 /*
5064 * Some architectures will only update present cpus when
5065 * onlining them, so don't trust the number if it's just 1. But
5066 * we also don't want to use nr_cpu_ids always, as on some other
5067 * architectures, there can be many possible cpus, but never
5068 * onlined. Here we compromise between trying to avoid too high
5069 * order on systems that appear larger than they are, and too
5070 * low order on systems that appear smaller than they are.
5071 */
5072 unsigned int nr_cpus = num_present_cpus();
5073 if (nr_cpus <= 1)
5074 nr_cpus = nr_cpu_ids;
5075 min_objects = 4 * (fls(nr_cpus) + 1);
5076 }
5077 /* min_objects can't be 0 because get_order(0) is undefined */
5078 max_objects = max(order_objects(slub_max_order, size), 1U);
5079 min_objects = min(min_objects, max_objects);
5080
5081 min_order = max_t(unsigned int, slub_min_order,
5082 get_order(min_objects * size));
5083 if (order_objects(min_order, size) > MAX_OBJS_PER_PAGE)
5084 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
5085
5086 /*
5087 * Attempt to find best configuration for a slab. This works by first
5088 * attempting to generate a layout with the best possible configuration
5089 * and backing off gradually.
5090 *
5091 * We start with accepting at most 1/16 waste and try to find the
5092 * smallest order from min_objects-derived/slab_min_order up to
5093 * slab_max_order that will satisfy the constraint. Note that increasing
5094 * the order can only result in same or less fractional waste, not more.
5095 *
5096 * If that fails, we increase the acceptable fraction of waste and try
5097 * again. The last iteration with fraction of 1/2 would effectively
5098 * accept any waste and give us the order determined by min_objects, as
5099 * long as at least single object fits within slab_max_order.
5100 */
5101 for (unsigned int fraction = 16; fraction > 1; fraction /= 2) {
5102 order = calc_slab_order(size, min_order, slub_max_order,
5103 fraction);
5104 if (order <= slub_max_order)
5105 return order;
5106 }
5107
5108 /*
5109 * Doh this slab cannot be placed using slab_max_order.
5110 */
5111 order = get_order(size);
5112 if (order <= MAX_PAGE_ORDER)
5113 return order;
5114 return -ENOSYS;
5115 }
5116
5117 static void
init_kmem_cache_node(struct kmem_cache_node * n)5118 init_kmem_cache_node(struct kmem_cache_node *n)
5119 {
5120 n->nr_partial = 0;
5121 spin_lock_init(&n->list_lock);
5122 INIT_LIST_HEAD(&n->partial);
5123 #ifdef CONFIG_SLUB_DEBUG
5124 atomic_long_set(&n->nr_slabs, 0);
5125 atomic_long_set(&n->total_objects, 0);
5126 INIT_LIST_HEAD(&n->full);
5127 #endif
5128 }
5129
5130 #ifndef CONFIG_SLUB_TINY
alloc_kmem_cache_cpus(struct kmem_cache * s)5131 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
5132 {
5133 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
5134 NR_KMALLOC_TYPES * KMALLOC_SHIFT_HIGH *
5135 sizeof(struct kmem_cache_cpu));
5136
5137 /*
5138 * Must align to double word boundary for the double cmpxchg
5139 * instructions to work; see __pcpu_double_call_return_bool().
5140 */
5141 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
5142 2 * sizeof(void *));
5143
5144 if (!s->cpu_slab)
5145 return 0;
5146
5147 init_kmem_cache_cpus(s);
5148
5149 return 1;
5150 }
5151 #else
alloc_kmem_cache_cpus(struct kmem_cache * s)5152 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
5153 {
5154 return 1;
5155 }
5156 #endif /* CONFIG_SLUB_TINY */
5157
5158 static struct kmem_cache *kmem_cache_node;
5159
5160 /*
5161 * No kmalloc_node yet so do it by hand. We know that this is the first
5162 * slab on the node for this slabcache. There are no concurrent accesses
5163 * possible.
5164 *
5165 * Note that this function only works on the kmem_cache_node
5166 * when allocating for the kmem_cache_node. This is used for bootstrapping
5167 * memory on a fresh node that has no slab structures yet.
5168 */
early_kmem_cache_node_alloc(int node)5169 static void early_kmem_cache_node_alloc(int node)
5170 {
5171 struct slab *slab;
5172 struct kmem_cache_node *n;
5173
5174 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
5175
5176 slab = new_slab(kmem_cache_node, GFP_NOWAIT, node);
5177
5178 BUG_ON(!slab);
5179 if (slab_nid(slab) != node) {
5180 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
5181 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
5182 }
5183
5184 n = slab->freelist;
5185 BUG_ON(!n);
5186 #ifdef CONFIG_SLUB_DEBUG
5187 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
5188 #endif
5189 n = kasan_slab_alloc(kmem_cache_node, n, GFP_KERNEL, false);
5190 slab->freelist = get_freepointer(kmem_cache_node, n);
5191 slab->inuse = 1;
5192 kmem_cache_node->node[node] = n;
5193 init_kmem_cache_node(n);
5194 inc_slabs_node(kmem_cache_node, node, slab->objects);
5195
5196 /*
5197 * No locks need to be taken here as it has just been
5198 * initialized and there is no concurrent access.
5199 */
5200 __add_partial(n, slab, DEACTIVATE_TO_HEAD);
5201 }
5202
free_kmem_cache_nodes(struct kmem_cache * s)5203 static void free_kmem_cache_nodes(struct kmem_cache *s)
5204 {
5205 int node;
5206 struct kmem_cache_node *n;
5207
5208 for_each_kmem_cache_node(s, node, n) {
5209 s->node[node] = NULL;
5210 kmem_cache_free(kmem_cache_node, n);
5211 }
5212 }
5213
__kmem_cache_release(struct kmem_cache * s)5214 void __kmem_cache_release(struct kmem_cache *s)
5215 {
5216 cache_random_seq_destroy(s);
5217 #ifndef CONFIG_SLUB_TINY
5218 free_percpu(s->cpu_slab);
5219 #endif
5220 free_kmem_cache_nodes(s);
5221 }
5222
init_kmem_cache_nodes(struct kmem_cache * s)5223 static int init_kmem_cache_nodes(struct kmem_cache *s)
5224 {
5225 int node;
5226
5227 for_each_node_mask(node, slab_nodes) {
5228 struct kmem_cache_node *n;
5229
5230 if (slab_state == DOWN) {
5231 early_kmem_cache_node_alloc(node);
5232 continue;
5233 }
5234 n = kmem_cache_alloc_node(kmem_cache_node,
5235 GFP_KERNEL, node);
5236
5237 if (!n) {
5238 free_kmem_cache_nodes(s);
5239 return 0;
5240 }
5241
5242 init_kmem_cache_node(n);
5243 s->node[node] = n;
5244 }
5245 return 1;
5246 }
5247
set_cpu_partial(struct kmem_cache * s)5248 static void set_cpu_partial(struct kmem_cache *s)
5249 {
5250 #ifdef CONFIG_SLUB_CPU_PARTIAL
5251 unsigned int nr_objects;
5252
5253 /*
5254 * cpu_partial determined the maximum number of objects kept in the
5255 * per cpu partial lists of a processor.
5256 *
5257 * Per cpu partial lists mainly contain slabs that just have one
5258 * object freed. If they are used for allocation then they can be
5259 * filled up again with minimal effort. The slab will never hit the
5260 * per node partial lists and therefore no locking will be required.
5261 *
5262 * For backwards compatibility reasons, this is determined as number
5263 * of objects, even though we now limit maximum number of pages, see
5264 * slub_set_cpu_partial()
5265 */
5266 if (!kmem_cache_has_cpu_partial(s))
5267 nr_objects = 0;
5268 else if (s->size >= PAGE_SIZE)
5269 nr_objects = 6;
5270 else if (s->size >= 1024)
5271 nr_objects = 24;
5272 else if (s->size >= 256)
5273 nr_objects = 52;
5274 else
5275 nr_objects = 120;
5276
5277 slub_set_cpu_partial(s, nr_objects);
5278 #endif
5279 }
5280
5281 /*
5282 * calculate_sizes() determines the order and the distribution of data within
5283 * a slab object.
5284 */
calculate_sizes(struct kmem_cache_args * args,struct kmem_cache * s)5285 static int calculate_sizes(struct kmem_cache_args *args, struct kmem_cache *s)
5286 {
5287 slab_flags_t flags = s->flags;
5288 unsigned int size = s->object_size;
5289 unsigned int order;
5290
5291 /*
5292 * Round up object size to the next word boundary. We can only
5293 * place the free pointer at word boundaries and this determines
5294 * the possible location of the free pointer.
5295 */
5296 size = ALIGN(size, sizeof(void *));
5297
5298 #ifdef CONFIG_SLUB_DEBUG
5299 /*
5300 * Determine if we can poison the object itself. If the user of
5301 * the slab may touch the object after free or before allocation
5302 * then we should never poison the object itself.
5303 */
5304 if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) &&
5305 !s->ctor)
5306 s->flags |= __OBJECT_POISON;
5307 else
5308 s->flags &= ~__OBJECT_POISON;
5309
5310
5311 /*
5312 * If we are Redzoning then check if there is some space between the
5313 * end of the object and the free pointer. If not then add an
5314 * additional word to have some bytes to store Redzone information.
5315 */
5316 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
5317 size += sizeof(void *);
5318 #endif
5319
5320 /*
5321 * With that we have determined the number of bytes in actual use
5322 * by the object and redzoning.
5323 */
5324 s->inuse = size;
5325
5326 if (((flags & SLAB_TYPESAFE_BY_RCU) && !args->use_freeptr_offset) ||
5327 (flags & SLAB_POISON) || s->ctor ||
5328 ((flags & SLAB_RED_ZONE) &&
5329 (s->object_size < sizeof(void *) || slub_debug_orig_size(s)))) {
5330 /*
5331 * Relocate free pointer after the object if it is not
5332 * permitted to overwrite the first word of the object on
5333 * kmem_cache_free.
5334 *
5335 * This is the case if we do RCU, have a constructor or
5336 * destructor, are poisoning the objects, or are
5337 * redzoning an object smaller than sizeof(void *) or are
5338 * redzoning an object with slub_debug_orig_size() enabled,
5339 * in which case the right redzone may be extended.
5340 *
5341 * The assumption that s->offset >= s->inuse means free
5342 * pointer is outside of the object is used in the
5343 * freeptr_outside_object() function. If that is no
5344 * longer true, the function needs to be modified.
5345 */
5346 s->offset = size;
5347 size += sizeof(void *);
5348 } else if ((flags & SLAB_TYPESAFE_BY_RCU) && args->use_freeptr_offset) {
5349 s->offset = args->freeptr_offset;
5350 } else {
5351 /*
5352 * Store freelist pointer near middle of object to keep
5353 * it away from the edges of the object to avoid small
5354 * sized over/underflows from neighboring allocations.
5355 */
5356 s->offset = ALIGN_DOWN(s->object_size / 2, sizeof(void *));
5357 }
5358
5359 #ifdef CONFIG_SLUB_DEBUG
5360 if (flags & SLAB_STORE_USER) {
5361 /*
5362 * Need to store information about allocs and frees after
5363 * the object.
5364 */
5365 size += 2 * sizeof(struct track);
5366
5367 /* Save the original kmalloc request size */
5368 if (flags & SLAB_KMALLOC)
5369 size += sizeof(unsigned int);
5370 }
5371 #endif
5372
5373 kasan_cache_create(s, &size, &s->flags);
5374 #ifdef CONFIG_SLUB_DEBUG
5375 if (flags & SLAB_RED_ZONE) {
5376 /*
5377 * Add some empty padding so that we can catch
5378 * overwrites from earlier objects rather than let
5379 * tracking information or the free pointer be
5380 * corrupted if a user writes before the start
5381 * of the object.
5382 */
5383 size += sizeof(void *);
5384
5385 s->red_left_pad = sizeof(void *);
5386 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
5387 size += s->red_left_pad;
5388 }
5389 #endif
5390
5391 /*
5392 * SLUB stores one object immediately after another beginning from
5393 * offset 0. In order to align the objects we have to simply size
5394 * each object to conform to the alignment.
5395 */
5396 size = ALIGN(size, s->align);
5397 s->size = size;
5398 s->reciprocal_size = reciprocal_value(size);
5399 order = calculate_order(size);
5400
5401 if ((int)order < 0)
5402 return 0;
5403
5404 s->allocflags = __GFP_COMP;
5405
5406 if (s->flags & SLAB_CACHE_DMA)
5407 s->allocflags |= GFP_DMA;
5408
5409 if (s->flags & SLAB_CACHE_DMA32)
5410 s->allocflags |= GFP_DMA32;
5411
5412 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5413 s->allocflags |= __GFP_RECLAIMABLE;
5414
5415 /*
5416 * Determine the number of objects per slab
5417 */
5418 s->oo = oo_make(order, size);
5419 s->min = oo_make(get_order(size), size);
5420
5421 return !!oo_objects(s->oo);
5422 }
5423
list_slab_objects(struct kmem_cache * s,struct slab * slab,const char * text)5424 static void list_slab_objects(struct kmem_cache *s, struct slab *slab,
5425 const char *text)
5426 {
5427 #ifdef CONFIG_SLUB_DEBUG
5428 void *addr = slab_address(slab);
5429 void *p;
5430
5431 slab_err(s, slab, text, s->name);
5432
5433 spin_lock(&object_map_lock);
5434 __fill_map(object_map, s, slab);
5435
5436 for_each_object(p, s, addr, slab->objects) {
5437
5438 if (!test_bit(__obj_to_index(s, addr, p), object_map)) {
5439 pr_err("Object 0x%p @offset=%tu\n", p, p - addr);
5440 print_tracking(s, p);
5441 }
5442 }
5443 spin_unlock(&object_map_lock);
5444 #endif
5445 }
5446
5447 /*
5448 * Attempt to free all partial slabs on a node.
5449 * This is called from __kmem_cache_shutdown(). We must take list_lock
5450 * because sysfs file might still access partial list after the shutdowning.
5451 */
free_partial(struct kmem_cache * s,struct kmem_cache_node * n)5452 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
5453 {
5454 LIST_HEAD(discard);
5455 struct slab *slab, *h;
5456
5457 BUG_ON(irqs_disabled());
5458 spin_lock_irq(&n->list_lock);
5459 list_for_each_entry_safe(slab, h, &n->partial, slab_list) {
5460 if (!slab->inuse) {
5461 remove_partial(n, slab);
5462 list_add(&slab->slab_list, &discard);
5463 } else {
5464 list_slab_objects(s, slab,
5465 "Objects remaining in %s on __kmem_cache_shutdown()");
5466 }
5467 }
5468 spin_unlock_irq(&n->list_lock);
5469
5470 list_for_each_entry_safe(slab, h, &discard, slab_list)
5471 discard_slab(s, slab);
5472 }
5473
__kmem_cache_empty(struct kmem_cache * s)5474 bool __kmem_cache_empty(struct kmem_cache *s)
5475 {
5476 int node;
5477 struct kmem_cache_node *n;
5478
5479 for_each_kmem_cache_node(s, node, n)
5480 if (n->nr_partial || node_nr_slabs(n))
5481 return false;
5482 return true;
5483 }
5484
5485 /*
5486 * Release all resources used by a slab cache.
5487 */
__kmem_cache_shutdown(struct kmem_cache * s)5488 int __kmem_cache_shutdown(struct kmem_cache *s)
5489 {
5490 int node;
5491 struct kmem_cache_node *n;
5492
5493 flush_all_cpus_locked(s);
5494 /* Attempt to free all objects */
5495 for_each_kmem_cache_node(s, node, n) {
5496 free_partial(s, n);
5497 if (n->nr_partial || node_nr_slabs(n))
5498 return 1;
5499 }
5500 return 0;
5501 }
5502
5503 #ifdef CONFIG_PRINTK
__kmem_obj_info(struct kmem_obj_info * kpp,void * object,struct slab * slab)5504 void __kmem_obj_info(struct kmem_obj_info *kpp, void *object, struct slab *slab)
5505 {
5506 void *base;
5507 int __maybe_unused i;
5508 unsigned int objnr;
5509 void *objp;
5510 void *objp0;
5511 struct kmem_cache *s = slab->slab_cache;
5512 struct track __maybe_unused *trackp;
5513
5514 kpp->kp_ptr = object;
5515 kpp->kp_slab = slab;
5516 kpp->kp_slab_cache = s;
5517 base = slab_address(slab);
5518 objp0 = kasan_reset_tag(object);
5519 #ifdef CONFIG_SLUB_DEBUG
5520 objp = restore_red_left(s, objp0);
5521 #else
5522 objp = objp0;
5523 #endif
5524 objnr = obj_to_index(s, slab, objp);
5525 kpp->kp_data_offset = (unsigned long)((char *)objp0 - (char *)objp);
5526 objp = base + s->size * objnr;
5527 kpp->kp_objp = objp;
5528 if (WARN_ON_ONCE(objp < base || objp >= base + slab->objects * s->size
5529 || (objp - base) % s->size) ||
5530 !(s->flags & SLAB_STORE_USER))
5531 return;
5532 #ifdef CONFIG_SLUB_DEBUG
5533 objp = fixup_red_left(s, objp);
5534 trackp = get_track(s, objp, TRACK_ALLOC);
5535 kpp->kp_ret = (void *)trackp->addr;
5536 #ifdef CONFIG_STACKDEPOT
5537 {
5538 depot_stack_handle_t handle;
5539 unsigned long *entries;
5540 unsigned int nr_entries;
5541
5542 handle = READ_ONCE(trackp->handle);
5543 if (handle) {
5544 nr_entries = stack_depot_fetch(handle, &entries);
5545 for (i = 0; i < KS_ADDRS_COUNT && i < nr_entries; i++)
5546 kpp->kp_stack[i] = (void *)entries[i];
5547 }
5548
5549 trackp = get_track(s, objp, TRACK_FREE);
5550 handle = READ_ONCE(trackp->handle);
5551 if (handle) {
5552 nr_entries = stack_depot_fetch(handle, &entries);
5553 for (i = 0; i < KS_ADDRS_COUNT && i < nr_entries; i++)
5554 kpp->kp_free_stack[i] = (void *)entries[i];
5555 }
5556 }
5557 #endif
5558 #endif
5559 }
5560 #endif
5561
5562 /********************************************************************
5563 * Kmalloc subsystem
5564 *******************************************************************/
5565
setup_slub_min_order(char * str)5566 static int __init setup_slub_min_order(char *str)
5567 {
5568 get_option(&str, (int *)&slub_min_order);
5569
5570 if (slub_min_order > slub_max_order)
5571 slub_max_order = slub_min_order;
5572
5573 return 1;
5574 }
5575
5576 __setup("slab_min_order=", setup_slub_min_order);
5577 __setup_param("slub_min_order=", slub_min_order, setup_slub_min_order, 0);
5578
5579
setup_slub_max_order(char * str)5580 static int __init setup_slub_max_order(char *str)
5581 {
5582 get_option(&str, (int *)&slub_max_order);
5583 slub_max_order = min_t(unsigned int, slub_max_order, MAX_PAGE_ORDER);
5584
5585 if (slub_min_order > slub_max_order)
5586 slub_min_order = slub_max_order;
5587
5588 return 1;
5589 }
5590
5591 __setup("slab_max_order=", setup_slub_max_order);
5592 __setup_param("slub_max_order=", slub_max_order, setup_slub_max_order, 0);
5593
setup_slub_min_objects(char * str)5594 static int __init setup_slub_min_objects(char *str)
5595 {
5596 get_option(&str, (int *)&slub_min_objects);
5597
5598 return 1;
5599 }
5600
5601 __setup("slab_min_objects=", setup_slub_min_objects);
5602 __setup_param("slub_min_objects=", slub_min_objects, setup_slub_min_objects, 0);
5603
5604 #ifdef CONFIG_HARDENED_USERCOPY
5605 /*
5606 * Rejects incorrectly sized objects and objects that are to be copied
5607 * to/from userspace but do not fall entirely within the containing slab
5608 * cache's usercopy region.
5609 *
5610 * Returns NULL if check passes, otherwise const char * to name of cache
5611 * to indicate an error.
5612 */
__check_heap_object(const void * ptr,unsigned long n,const struct slab * slab,bool to_user)5613 void __check_heap_object(const void *ptr, unsigned long n,
5614 const struct slab *slab, bool to_user)
5615 {
5616 struct kmem_cache *s;
5617 unsigned int offset;
5618 bool is_kfence = is_kfence_address(ptr);
5619
5620 ptr = kasan_reset_tag(ptr);
5621
5622 /* Find object and usable object size. */
5623 s = slab->slab_cache;
5624
5625 /* Reject impossible pointers. */
5626 if (ptr < slab_address(slab))
5627 usercopy_abort("SLUB object not in SLUB page?!", NULL,
5628 to_user, 0, n);
5629
5630 /* Find offset within object. */
5631 if (is_kfence)
5632 offset = ptr - kfence_object_start(ptr);
5633 else
5634 offset = (ptr - slab_address(slab)) % s->size;
5635
5636 /* Adjust for redzone and reject if within the redzone. */
5637 if (!is_kfence && kmem_cache_debug_flags(s, SLAB_RED_ZONE)) {
5638 if (offset < s->red_left_pad)
5639 usercopy_abort("SLUB object in left red zone",
5640 s->name, to_user, offset, n);
5641 offset -= s->red_left_pad;
5642 }
5643
5644 /* Allow address range falling entirely within usercopy region. */
5645 if (offset >= s->useroffset &&
5646 offset - s->useroffset <= s->usersize &&
5647 n <= s->useroffset - offset + s->usersize)
5648 return;
5649
5650 usercopy_abort("SLUB object", s->name, to_user, offset, n);
5651 }
5652 #endif /* CONFIG_HARDENED_USERCOPY */
5653
5654 #define SHRINK_PROMOTE_MAX 32
5655
5656 /*
5657 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
5658 * up most to the head of the partial lists. New allocations will then
5659 * fill those up and thus they can be removed from the partial lists.
5660 *
5661 * The slabs with the least items are placed last. This results in them
5662 * being allocated from last increasing the chance that the last objects
5663 * are freed in them.
5664 */
__kmem_cache_do_shrink(struct kmem_cache * s)5665 static int __kmem_cache_do_shrink(struct kmem_cache *s)
5666 {
5667 int node;
5668 int i;
5669 struct kmem_cache_node *n;
5670 struct slab *slab;
5671 struct slab *t;
5672 struct list_head discard;
5673 struct list_head promote[SHRINK_PROMOTE_MAX];
5674 unsigned long flags;
5675 int ret = 0;
5676
5677 for_each_kmem_cache_node(s, node, n) {
5678 INIT_LIST_HEAD(&discard);
5679 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
5680 INIT_LIST_HEAD(promote + i);
5681
5682 spin_lock_irqsave(&n->list_lock, flags);
5683
5684 /*
5685 * Build lists of slabs to discard or promote.
5686 *
5687 * Note that concurrent frees may occur while we hold the
5688 * list_lock. slab->inuse here is the upper limit.
5689 */
5690 list_for_each_entry_safe(slab, t, &n->partial, slab_list) {
5691 int free = slab->objects - slab->inuse;
5692
5693 /* Do not reread slab->inuse */
5694 barrier();
5695
5696 /* We do not keep full slabs on the list */
5697 BUG_ON(free <= 0);
5698
5699 if (free == slab->objects) {
5700 list_move(&slab->slab_list, &discard);
5701 slab_clear_node_partial(slab);
5702 n->nr_partial--;
5703 dec_slabs_node(s, node, slab->objects);
5704 } else if (free <= SHRINK_PROMOTE_MAX)
5705 list_move(&slab->slab_list, promote + free - 1);
5706 }
5707
5708 /*
5709 * Promote the slabs filled up most to the head of the
5710 * partial list.
5711 */
5712 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
5713 list_splice(promote + i, &n->partial);
5714
5715 spin_unlock_irqrestore(&n->list_lock, flags);
5716
5717 /* Release empty slabs */
5718 list_for_each_entry_safe(slab, t, &discard, slab_list)
5719 free_slab(s, slab);
5720
5721 if (node_nr_slabs(n))
5722 ret = 1;
5723 }
5724
5725 return ret;
5726 }
5727
__kmem_cache_shrink(struct kmem_cache * s)5728 int __kmem_cache_shrink(struct kmem_cache *s)
5729 {
5730 flush_all(s);
5731 return __kmem_cache_do_shrink(s);
5732 }
5733
slab_mem_going_offline_callback(void * arg)5734 static int slab_mem_going_offline_callback(void *arg)
5735 {
5736 struct kmem_cache *s;
5737
5738 mutex_lock(&slab_mutex);
5739 list_for_each_entry(s, &slab_caches, list) {
5740 flush_all_cpus_locked(s);
5741 __kmem_cache_do_shrink(s);
5742 }
5743 mutex_unlock(&slab_mutex);
5744
5745 return 0;
5746 }
5747
slab_mem_offline_callback(void * arg)5748 static void slab_mem_offline_callback(void *arg)
5749 {
5750 struct memory_notify *marg = arg;
5751 int offline_node;
5752
5753 offline_node = marg->status_change_nid_normal;
5754
5755 /*
5756 * If the node still has available memory. we need kmem_cache_node
5757 * for it yet.
5758 */
5759 if (offline_node < 0)
5760 return;
5761
5762 mutex_lock(&slab_mutex);
5763 node_clear(offline_node, slab_nodes);
5764 /*
5765 * We no longer free kmem_cache_node structures here, as it would be
5766 * racy with all get_node() users, and infeasible to protect them with
5767 * slab_mutex.
5768 */
5769 mutex_unlock(&slab_mutex);
5770 }
5771
slab_mem_going_online_callback(void * arg)5772 static int slab_mem_going_online_callback(void *arg)
5773 {
5774 struct kmem_cache_node *n;
5775 struct kmem_cache *s;
5776 struct memory_notify *marg = arg;
5777 int nid = marg->status_change_nid_normal;
5778 int ret = 0;
5779
5780 /*
5781 * If the node's memory is already available, then kmem_cache_node is
5782 * already created. Nothing to do.
5783 */
5784 if (nid < 0)
5785 return 0;
5786
5787 /*
5788 * We are bringing a node online. No memory is available yet. We must
5789 * allocate a kmem_cache_node structure in order to bring the node
5790 * online.
5791 */
5792 mutex_lock(&slab_mutex);
5793 list_for_each_entry(s, &slab_caches, list) {
5794 /*
5795 * The structure may already exist if the node was previously
5796 * onlined and offlined.
5797 */
5798 if (get_node(s, nid))
5799 continue;
5800 /*
5801 * XXX: kmem_cache_alloc_node will fallback to other nodes
5802 * since memory is not yet available from the node that
5803 * is brought up.
5804 */
5805 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
5806 if (!n) {
5807 ret = -ENOMEM;
5808 goto out;
5809 }
5810 init_kmem_cache_node(n);
5811 s->node[nid] = n;
5812 }
5813 /*
5814 * Any cache created after this point will also have kmem_cache_node
5815 * initialized for the new node.
5816 */
5817 node_set(nid, slab_nodes);
5818 out:
5819 mutex_unlock(&slab_mutex);
5820 return ret;
5821 }
5822
slab_memory_callback(struct notifier_block * self,unsigned long action,void * arg)5823 static int slab_memory_callback(struct notifier_block *self,
5824 unsigned long action, void *arg)
5825 {
5826 int ret = 0;
5827
5828 switch (action) {
5829 case MEM_GOING_ONLINE:
5830 ret = slab_mem_going_online_callback(arg);
5831 break;
5832 case MEM_GOING_OFFLINE:
5833 ret = slab_mem_going_offline_callback(arg);
5834 break;
5835 case MEM_OFFLINE:
5836 case MEM_CANCEL_ONLINE:
5837 slab_mem_offline_callback(arg);
5838 break;
5839 case MEM_ONLINE:
5840 case MEM_CANCEL_OFFLINE:
5841 break;
5842 }
5843 if (ret)
5844 ret = notifier_from_errno(ret);
5845 else
5846 ret = NOTIFY_OK;
5847 return ret;
5848 }
5849
5850 /********************************************************************
5851 * Basic setup of slabs
5852 *******************************************************************/
5853
5854 /*
5855 * Used for early kmem_cache structures that were allocated using
5856 * the page allocator. Allocate them properly then fix up the pointers
5857 * that may be pointing to the wrong kmem_cache structure.
5858 */
5859
bootstrap(struct kmem_cache * static_cache)5860 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
5861 {
5862 int node;
5863 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
5864 struct kmem_cache_node *n;
5865
5866 memcpy(s, static_cache, kmem_cache->object_size);
5867
5868 /*
5869 * This runs very early, and only the boot processor is supposed to be
5870 * up. Even if it weren't true, IRQs are not up so we couldn't fire
5871 * IPIs around.
5872 */
5873 __flush_cpu_slab(s, smp_processor_id());
5874 for_each_kmem_cache_node(s, node, n) {
5875 struct slab *p;
5876
5877 list_for_each_entry(p, &n->partial, slab_list)
5878 p->slab_cache = s;
5879
5880 #ifdef CONFIG_SLUB_DEBUG
5881 list_for_each_entry(p, &n->full, slab_list)
5882 p->slab_cache = s;
5883 #endif
5884 }
5885 list_add(&s->list, &slab_caches);
5886 return s;
5887 }
5888
kmem_cache_init(void)5889 void __init kmem_cache_init(void)
5890 {
5891 static __initdata struct kmem_cache boot_kmem_cache,
5892 boot_kmem_cache_node;
5893 int node;
5894
5895 if (debug_guardpage_minorder())
5896 slub_max_order = 0;
5897
5898 /* Print slub debugging pointers without hashing */
5899 if (__slub_debug_enabled())
5900 no_hash_pointers_enable(NULL);
5901
5902 kmem_cache_node = &boot_kmem_cache_node;
5903 kmem_cache = &boot_kmem_cache;
5904
5905 /*
5906 * Initialize the nodemask for which we will allocate per node
5907 * structures. Here we don't need taking slab_mutex yet.
5908 */
5909 for_each_node_state(node, N_NORMAL_MEMORY)
5910 node_set(node, slab_nodes);
5911
5912 create_boot_cache(kmem_cache_node, "kmem_cache_node",
5913 sizeof(struct kmem_cache_node),
5914 SLAB_HWCACHE_ALIGN | SLAB_NO_OBJ_EXT, 0, 0);
5915
5916 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
5917
5918 /* Able to allocate the per node structures */
5919 slab_state = PARTIAL;
5920
5921 create_boot_cache(kmem_cache, "kmem_cache",
5922 offsetof(struct kmem_cache, node) +
5923 nr_node_ids * sizeof(struct kmem_cache_node *),
5924 SLAB_HWCACHE_ALIGN | SLAB_NO_OBJ_EXT, 0, 0);
5925
5926 kmem_cache = bootstrap(&boot_kmem_cache);
5927 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
5928
5929 /* Now we can use the kmem_cache to allocate kmalloc slabs */
5930 setup_kmalloc_cache_index_table();
5931 create_kmalloc_caches();
5932
5933 /* Setup random freelists for each cache */
5934 init_freelist_randomization();
5935
5936 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL,
5937 slub_cpu_dead);
5938
5939 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n",
5940 cache_line_size(),
5941 slub_min_order, slub_max_order, slub_min_objects,
5942 nr_cpu_ids, nr_node_ids);
5943 }
5944
kmem_cache_init_late(void)5945 void __init kmem_cache_init_late(void)
5946 {
5947 #ifndef CONFIG_SLUB_TINY
5948 flushwq = alloc_workqueue("slub_flushwq", WQ_MEM_RECLAIM, 0);
5949 WARN_ON(!flushwq);
5950 #endif
5951 }
5952
5953 struct kmem_cache *
__kmem_cache_alias(const char * name,unsigned int size,unsigned int align,slab_flags_t flags,void (* ctor)(void *))5954 __kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
5955 slab_flags_t flags, void (*ctor)(void *))
5956 {
5957 struct kmem_cache *s;
5958
5959 s = find_mergeable(size, align, flags, name, ctor);
5960 if (s) {
5961 if (sysfs_slab_alias(s, name))
5962 return NULL;
5963
5964 s->refcount++;
5965
5966 /*
5967 * Adjust the object sizes so that we clear
5968 * the complete object on kzalloc.
5969 */
5970 s->object_size = max(s->object_size, size);
5971 s->inuse = max(s->inuse, ALIGN(size, sizeof(void *)));
5972 }
5973
5974 return s;
5975 }
5976
do_kmem_cache_create(struct kmem_cache * s,const char * name,unsigned int size,struct kmem_cache_args * args,slab_flags_t flags)5977 int do_kmem_cache_create(struct kmem_cache *s, const char *name,
5978 unsigned int size, struct kmem_cache_args *args,
5979 slab_flags_t flags)
5980 {
5981 int err = -EINVAL;
5982
5983 s->name = name;
5984 s->size = s->object_size = size;
5985
5986 s->flags = kmem_cache_flags(flags, s->name);
5987 #ifdef CONFIG_SLAB_FREELIST_HARDENED
5988 s->random = get_random_long();
5989 #endif
5990 s->align = args->align;
5991 s->ctor = args->ctor;
5992 #ifdef CONFIG_HARDENED_USERCOPY
5993 s->useroffset = args->useroffset;
5994 s->usersize = args->usersize;
5995 #endif
5996
5997 if (!calculate_sizes(args, s))
5998 goto out;
5999 if (disable_higher_order_debug) {
6000 /*
6001 * Disable debugging flags that store metadata if the min slab
6002 * order increased.
6003 */
6004 if (get_order(s->size) > get_order(s->object_size)) {
6005 s->flags &= ~DEBUG_METADATA_FLAGS;
6006 s->offset = 0;
6007 if (!calculate_sizes(args, s))
6008 goto out;
6009 }
6010 }
6011
6012 #ifdef system_has_freelist_aba
6013 if (system_has_freelist_aba() && !(s->flags & SLAB_NO_CMPXCHG)) {
6014 /* Enable fast mode */
6015 s->flags |= __CMPXCHG_DOUBLE;
6016 }
6017 #endif
6018
6019 /*
6020 * The larger the object size is, the more slabs we want on the partial
6021 * list to avoid pounding the page allocator excessively.
6022 */
6023 s->min_partial = min_t(unsigned long, MAX_PARTIAL, ilog2(s->size) / 2);
6024 s->min_partial = max_t(unsigned long, MIN_PARTIAL, s->min_partial);
6025
6026 set_cpu_partial(s);
6027
6028 #ifdef CONFIG_NUMA
6029 s->remote_node_defrag_ratio = 1000;
6030 #endif
6031
6032 /* Initialize the pre-computed randomized freelist if slab is up */
6033 if (slab_state >= UP) {
6034 if (init_cache_random_seq(s))
6035 goto out;
6036 }
6037
6038 if (!init_kmem_cache_nodes(s))
6039 goto out;
6040
6041 if (!alloc_kmem_cache_cpus(s))
6042 goto out;
6043
6044 /* Mutex is not taken during early boot */
6045 if (slab_state <= UP) {
6046 err = 0;
6047 goto out;
6048 }
6049
6050 err = sysfs_slab_add(s);
6051 if (err)
6052 goto out;
6053
6054 if (s->flags & SLAB_STORE_USER)
6055 debugfs_slab_add(s);
6056
6057 out:
6058 if (err)
6059 __kmem_cache_release(s);
6060 return err;
6061 }
6062
6063 #ifdef SLAB_SUPPORTS_SYSFS
count_inuse(struct slab * slab)6064 static int count_inuse(struct slab *slab)
6065 {
6066 return slab->inuse;
6067 }
6068
count_total(struct slab * slab)6069 static int count_total(struct slab *slab)
6070 {
6071 return slab->objects;
6072 }
6073 #endif
6074
6075 #ifdef CONFIG_SLUB_DEBUG
validate_slab(struct kmem_cache * s,struct slab * slab,unsigned long * obj_map)6076 static void validate_slab(struct kmem_cache *s, struct slab *slab,
6077 unsigned long *obj_map)
6078 {
6079 void *p;
6080 void *addr = slab_address(slab);
6081
6082 if (!check_slab(s, slab) || !on_freelist(s, slab, NULL))
6083 return;
6084
6085 /* Now we know that a valid freelist exists */
6086 __fill_map(obj_map, s, slab);
6087 for_each_object(p, s, addr, slab->objects) {
6088 u8 val = test_bit(__obj_to_index(s, addr, p), obj_map) ?
6089 SLUB_RED_INACTIVE : SLUB_RED_ACTIVE;
6090
6091 if (!check_object(s, slab, p, val))
6092 break;
6093 }
6094 }
6095
validate_slab_node(struct kmem_cache * s,struct kmem_cache_node * n,unsigned long * obj_map)6096 static int validate_slab_node(struct kmem_cache *s,
6097 struct kmem_cache_node *n, unsigned long *obj_map)
6098 {
6099 unsigned long count = 0;
6100 struct slab *slab;
6101 unsigned long flags;
6102
6103 spin_lock_irqsave(&n->list_lock, flags);
6104
6105 list_for_each_entry(slab, &n->partial, slab_list) {
6106 validate_slab(s, slab, obj_map);
6107 count++;
6108 }
6109 if (count != n->nr_partial) {
6110 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
6111 s->name, count, n->nr_partial);
6112 slab_add_kunit_errors();
6113 }
6114
6115 if (!(s->flags & SLAB_STORE_USER))
6116 goto out;
6117
6118 list_for_each_entry(slab, &n->full, slab_list) {
6119 validate_slab(s, slab, obj_map);
6120 count++;
6121 }
6122 if (count != node_nr_slabs(n)) {
6123 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
6124 s->name, count, node_nr_slabs(n));
6125 slab_add_kunit_errors();
6126 }
6127
6128 out:
6129 spin_unlock_irqrestore(&n->list_lock, flags);
6130 return count;
6131 }
6132
validate_slab_cache(struct kmem_cache * s)6133 long validate_slab_cache(struct kmem_cache *s)
6134 {
6135 int node;
6136 unsigned long count = 0;
6137 struct kmem_cache_node *n;
6138 unsigned long *obj_map;
6139
6140 obj_map = bitmap_alloc(oo_objects(s->oo), GFP_KERNEL);
6141 if (!obj_map)
6142 return -ENOMEM;
6143
6144 flush_all(s);
6145 for_each_kmem_cache_node(s, node, n)
6146 count += validate_slab_node(s, n, obj_map);
6147
6148 bitmap_free(obj_map);
6149
6150 return count;
6151 }
6152 EXPORT_SYMBOL(validate_slab_cache);
6153
6154 #ifdef CONFIG_DEBUG_FS
6155 /*
6156 * Generate lists of code addresses where slabcache objects are allocated
6157 * and freed.
6158 */
6159
6160 struct location {
6161 depot_stack_handle_t handle;
6162 unsigned long count;
6163 unsigned long addr;
6164 unsigned long waste;
6165 long long sum_time;
6166 long min_time;
6167 long max_time;
6168 long min_pid;
6169 long max_pid;
6170 DECLARE_BITMAP(cpus, NR_CPUS);
6171 nodemask_t nodes;
6172 };
6173
6174 struct loc_track {
6175 unsigned long max;
6176 unsigned long count;
6177 struct location *loc;
6178 loff_t idx;
6179 };
6180
6181 static struct dentry *slab_debugfs_root;
6182
free_loc_track(struct loc_track * t)6183 static void free_loc_track(struct loc_track *t)
6184 {
6185 if (t->max)
6186 free_pages((unsigned long)t->loc,
6187 get_order(sizeof(struct location) * t->max));
6188 }
6189
alloc_loc_track(struct loc_track * t,unsigned long max,gfp_t flags)6190 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
6191 {
6192 struct location *l;
6193 int order;
6194
6195 order = get_order(sizeof(struct location) * max);
6196
6197 l = (void *)__get_free_pages(flags, order);
6198 if (!l)
6199 return 0;
6200
6201 if (t->count) {
6202 memcpy(l, t->loc, sizeof(struct location) * t->count);
6203 free_loc_track(t);
6204 }
6205 t->max = max;
6206 t->loc = l;
6207 return 1;
6208 }
6209
add_location(struct loc_track * t,struct kmem_cache * s,const struct track * track,unsigned int orig_size)6210 static int add_location(struct loc_track *t, struct kmem_cache *s,
6211 const struct track *track,
6212 unsigned int orig_size)
6213 {
6214 long start, end, pos;
6215 struct location *l;
6216 unsigned long caddr, chandle, cwaste;
6217 unsigned long age = jiffies - track->when;
6218 depot_stack_handle_t handle = 0;
6219 unsigned int waste = s->object_size - orig_size;
6220
6221 #ifdef CONFIG_STACKDEPOT
6222 handle = READ_ONCE(track->handle);
6223 #endif
6224 start = -1;
6225 end = t->count;
6226
6227 for ( ; ; ) {
6228 pos = start + (end - start + 1) / 2;
6229
6230 /*
6231 * There is nothing at "end". If we end up there
6232 * we need to add something to before end.
6233 */
6234 if (pos == end)
6235 break;
6236
6237 l = &t->loc[pos];
6238 caddr = l->addr;
6239 chandle = l->handle;
6240 cwaste = l->waste;
6241 if ((track->addr == caddr) && (handle == chandle) &&
6242 (waste == cwaste)) {
6243
6244 l->count++;
6245 if (track->when) {
6246 l->sum_time += age;
6247 if (age < l->min_time)
6248 l->min_time = age;
6249 if (age > l->max_time)
6250 l->max_time = age;
6251
6252 if (track->pid < l->min_pid)
6253 l->min_pid = track->pid;
6254 if (track->pid > l->max_pid)
6255 l->max_pid = track->pid;
6256
6257 cpumask_set_cpu(track->cpu,
6258 to_cpumask(l->cpus));
6259 }
6260 node_set(page_to_nid(virt_to_page(track)), l->nodes);
6261 return 1;
6262 }
6263
6264 if (track->addr < caddr)
6265 end = pos;
6266 else if (track->addr == caddr && handle < chandle)
6267 end = pos;
6268 else if (track->addr == caddr && handle == chandle &&
6269 waste < cwaste)
6270 end = pos;
6271 else
6272 start = pos;
6273 }
6274
6275 /*
6276 * Not found. Insert new tracking element.
6277 */
6278 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
6279 return 0;
6280
6281 l = t->loc + pos;
6282 if (pos < t->count)
6283 memmove(l + 1, l,
6284 (t->count - pos) * sizeof(struct location));
6285 t->count++;
6286 l->count = 1;
6287 l->addr = track->addr;
6288 l->sum_time = age;
6289 l->min_time = age;
6290 l->max_time = age;
6291 l->min_pid = track->pid;
6292 l->max_pid = track->pid;
6293 l->handle = handle;
6294 l->waste = waste;
6295 cpumask_clear(to_cpumask(l->cpus));
6296 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
6297 nodes_clear(l->nodes);
6298 node_set(page_to_nid(virt_to_page(track)), l->nodes);
6299 return 1;
6300 }
6301
process_slab(struct loc_track * t,struct kmem_cache * s,struct slab * slab,enum track_item alloc,unsigned long * obj_map)6302 static void process_slab(struct loc_track *t, struct kmem_cache *s,
6303 struct slab *slab, enum track_item alloc,
6304 unsigned long *obj_map)
6305 {
6306 void *addr = slab_address(slab);
6307 bool is_alloc = (alloc == TRACK_ALLOC);
6308 void *p;
6309
6310 __fill_map(obj_map, s, slab);
6311
6312 for_each_object(p, s, addr, slab->objects)
6313 if (!test_bit(__obj_to_index(s, addr, p), obj_map))
6314 add_location(t, s, get_track(s, p, alloc),
6315 is_alloc ? get_orig_size(s, p) :
6316 s->object_size);
6317 }
6318 #endif /* CONFIG_DEBUG_FS */
6319 #endif /* CONFIG_SLUB_DEBUG */
6320
6321 #ifdef SLAB_SUPPORTS_SYSFS
6322 enum slab_stat_type {
6323 SL_ALL, /* All slabs */
6324 SL_PARTIAL, /* Only partially allocated slabs */
6325 SL_CPU, /* Only slabs used for cpu caches */
6326 SL_OBJECTS, /* Determine allocated objects not slabs */
6327 SL_TOTAL /* Determine object capacity not slabs */
6328 };
6329
6330 #define SO_ALL (1 << SL_ALL)
6331 #define SO_PARTIAL (1 << SL_PARTIAL)
6332 #define SO_CPU (1 << SL_CPU)
6333 #define SO_OBJECTS (1 << SL_OBJECTS)
6334 #define SO_TOTAL (1 << SL_TOTAL)
6335
show_slab_objects(struct kmem_cache * s,char * buf,unsigned long flags)6336 static ssize_t show_slab_objects(struct kmem_cache *s,
6337 char *buf, unsigned long flags)
6338 {
6339 unsigned long total = 0;
6340 int node;
6341 int x;
6342 unsigned long *nodes;
6343 int len = 0;
6344
6345 nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL);
6346 if (!nodes)
6347 return -ENOMEM;
6348
6349 if (flags & SO_CPU) {
6350 int cpu;
6351
6352 for_each_possible_cpu(cpu) {
6353 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
6354 cpu);
6355 int node;
6356 struct slab *slab;
6357
6358 slab = READ_ONCE(c->slab);
6359 if (!slab)
6360 continue;
6361
6362 node = slab_nid(slab);
6363 if (flags & SO_TOTAL)
6364 x = slab->objects;
6365 else if (flags & SO_OBJECTS)
6366 x = slab->inuse;
6367 else
6368 x = 1;
6369
6370 total += x;
6371 nodes[node] += x;
6372
6373 #ifdef CONFIG_SLUB_CPU_PARTIAL
6374 slab = slub_percpu_partial_read_once(c);
6375 if (slab) {
6376 node = slab_nid(slab);
6377 if (flags & SO_TOTAL)
6378 WARN_ON_ONCE(1);
6379 else if (flags & SO_OBJECTS)
6380 WARN_ON_ONCE(1);
6381 else
6382 x = data_race(slab->slabs);
6383 total += x;
6384 nodes[node] += x;
6385 }
6386 #endif
6387 }
6388 }
6389
6390 /*
6391 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
6392 * already held which will conflict with an existing lock order:
6393 *
6394 * mem_hotplug_lock->slab_mutex->kernfs_mutex
6395 *
6396 * We don't really need mem_hotplug_lock (to hold off
6397 * slab_mem_going_offline_callback) here because slab's memory hot
6398 * unplug code doesn't destroy the kmem_cache->node[] data.
6399 */
6400
6401 #ifdef CONFIG_SLUB_DEBUG
6402 if (flags & SO_ALL) {
6403 struct kmem_cache_node *n;
6404
6405 for_each_kmem_cache_node(s, node, n) {
6406
6407 if (flags & SO_TOTAL)
6408 x = node_nr_objs(n);
6409 else if (flags & SO_OBJECTS)
6410 x = node_nr_objs(n) - count_partial(n, count_free);
6411 else
6412 x = node_nr_slabs(n);
6413 total += x;
6414 nodes[node] += x;
6415 }
6416
6417 } else
6418 #endif
6419 if (flags & SO_PARTIAL) {
6420 struct kmem_cache_node *n;
6421
6422 for_each_kmem_cache_node(s, node, n) {
6423 if (flags & SO_TOTAL)
6424 x = count_partial(n, count_total);
6425 else if (flags & SO_OBJECTS)
6426 x = count_partial(n, count_inuse);
6427 else
6428 x = n->nr_partial;
6429 total += x;
6430 nodes[node] += x;
6431 }
6432 }
6433
6434 len += sysfs_emit_at(buf, len, "%lu", total);
6435 #ifdef CONFIG_NUMA
6436 for (node = 0; node < nr_node_ids; node++) {
6437 if (nodes[node])
6438 len += sysfs_emit_at(buf, len, " N%d=%lu",
6439 node, nodes[node]);
6440 }
6441 #endif
6442 len += sysfs_emit_at(buf, len, "\n");
6443 kfree(nodes);
6444
6445 return len;
6446 }
6447
6448 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
6449 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
6450
6451 struct slab_attribute {
6452 struct attribute attr;
6453 ssize_t (*show)(struct kmem_cache *s, char *buf);
6454 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
6455 };
6456
6457 #define SLAB_ATTR_RO(_name) \
6458 static struct slab_attribute _name##_attr = __ATTR_RO_MODE(_name, 0400)
6459
6460 #define SLAB_ATTR(_name) \
6461 static struct slab_attribute _name##_attr = __ATTR_RW_MODE(_name, 0600)
6462
slab_size_show(struct kmem_cache * s,char * buf)6463 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
6464 {
6465 return sysfs_emit(buf, "%u\n", s->size);
6466 }
6467 SLAB_ATTR_RO(slab_size);
6468
align_show(struct kmem_cache * s,char * buf)6469 static ssize_t align_show(struct kmem_cache *s, char *buf)
6470 {
6471 return sysfs_emit(buf, "%u\n", s->align);
6472 }
6473 SLAB_ATTR_RO(align);
6474
object_size_show(struct kmem_cache * s,char * buf)6475 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
6476 {
6477 return sysfs_emit(buf, "%u\n", s->object_size);
6478 }
6479 SLAB_ATTR_RO(object_size);
6480
objs_per_slab_show(struct kmem_cache * s,char * buf)6481 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
6482 {
6483 return sysfs_emit(buf, "%u\n", oo_objects(s->oo));
6484 }
6485 SLAB_ATTR_RO(objs_per_slab);
6486
order_show(struct kmem_cache * s,char * buf)6487 static ssize_t order_show(struct kmem_cache *s, char *buf)
6488 {
6489 return sysfs_emit(buf, "%u\n", oo_order(s->oo));
6490 }
6491 SLAB_ATTR_RO(order);
6492
min_partial_show(struct kmem_cache * s,char * buf)6493 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
6494 {
6495 return sysfs_emit(buf, "%lu\n", s->min_partial);
6496 }
6497
min_partial_store(struct kmem_cache * s,const char * buf,size_t length)6498 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
6499 size_t length)
6500 {
6501 unsigned long min;
6502 int err;
6503
6504 err = kstrtoul(buf, 10, &min);
6505 if (err)
6506 return err;
6507
6508 s->min_partial = min;
6509 return length;
6510 }
6511 SLAB_ATTR(min_partial);
6512
cpu_partial_show(struct kmem_cache * s,char * buf)6513 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
6514 {
6515 unsigned int nr_partial = 0;
6516 #ifdef CONFIG_SLUB_CPU_PARTIAL
6517 nr_partial = s->cpu_partial;
6518 #endif
6519
6520 return sysfs_emit(buf, "%u\n", nr_partial);
6521 }
6522
cpu_partial_store(struct kmem_cache * s,const char * buf,size_t length)6523 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
6524 size_t length)
6525 {
6526 unsigned int objects;
6527 int err;
6528
6529 err = kstrtouint(buf, 10, &objects);
6530 if (err)
6531 return err;
6532 if (objects && !kmem_cache_has_cpu_partial(s))
6533 return -EINVAL;
6534
6535 slub_set_cpu_partial(s, objects);
6536 flush_all(s);
6537 return length;
6538 }
6539 SLAB_ATTR(cpu_partial);
6540
ctor_show(struct kmem_cache * s,char * buf)6541 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
6542 {
6543 if (!s->ctor)
6544 return 0;
6545 return sysfs_emit(buf, "%pS\n", s->ctor);
6546 }
6547 SLAB_ATTR_RO(ctor);
6548
aliases_show(struct kmem_cache * s,char * buf)6549 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
6550 {
6551 return sysfs_emit(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
6552 }
6553 SLAB_ATTR_RO(aliases);
6554
partial_show(struct kmem_cache * s,char * buf)6555 static ssize_t partial_show(struct kmem_cache *s, char *buf)
6556 {
6557 return show_slab_objects(s, buf, SO_PARTIAL);
6558 }
6559 SLAB_ATTR_RO(partial);
6560
cpu_slabs_show(struct kmem_cache * s,char * buf)6561 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
6562 {
6563 return show_slab_objects(s, buf, SO_CPU);
6564 }
6565 SLAB_ATTR_RO(cpu_slabs);
6566
objects_partial_show(struct kmem_cache * s,char * buf)6567 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
6568 {
6569 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
6570 }
6571 SLAB_ATTR_RO(objects_partial);
6572
slabs_cpu_partial_show(struct kmem_cache * s,char * buf)6573 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
6574 {
6575 int objects = 0;
6576 int slabs = 0;
6577 int cpu __maybe_unused;
6578 int len = 0;
6579
6580 #ifdef CONFIG_SLUB_CPU_PARTIAL
6581 for_each_online_cpu(cpu) {
6582 struct slab *slab;
6583
6584 slab = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
6585
6586 if (slab)
6587 slabs += data_race(slab->slabs);
6588 }
6589 #endif
6590
6591 /* Approximate half-full slabs, see slub_set_cpu_partial() */
6592 objects = (slabs * oo_objects(s->oo)) / 2;
6593 len += sysfs_emit_at(buf, len, "%d(%d)", objects, slabs);
6594
6595 #ifdef CONFIG_SLUB_CPU_PARTIAL
6596 for_each_online_cpu(cpu) {
6597 struct slab *slab;
6598
6599 slab = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
6600 if (slab) {
6601 slabs = data_race(slab->slabs);
6602 objects = (slabs * oo_objects(s->oo)) / 2;
6603 len += sysfs_emit_at(buf, len, " C%d=%d(%d)",
6604 cpu, objects, slabs);
6605 }
6606 }
6607 #endif
6608 len += sysfs_emit_at(buf, len, "\n");
6609
6610 return len;
6611 }
6612 SLAB_ATTR_RO(slabs_cpu_partial);
6613
reclaim_account_show(struct kmem_cache * s,char * buf)6614 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
6615 {
6616 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
6617 }
6618 SLAB_ATTR_RO(reclaim_account);
6619
hwcache_align_show(struct kmem_cache * s,char * buf)6620 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
6621 {
6622 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
6623 }
6624 SLAB_ATTR_RO(hwcache_align);
6625
6626 #ifdef CONFIG_ZONE_DMA
cache_dma_show(struct kmem_cache * s,char * buf)6627 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
6628 {
6629 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
6630 }
6631 SLAB_ATTR_RO(cache_dma);
6632 #endif
6633
6634 #ifdef CONFIG_HARDENED_USERCOPY
usersize_show(struct kmem_cache * s,char * buf)6635 static ssize_t usersize_show(struct kmem_cache *s, char *buf)
6636 {
6637 return sysfs_emit(buf, "%u\n", s->usersize);
6638 }
6639 SLAB_ATTR_RO(usersize);
6640 #endif
6641
destroy_by_rcu_show(struct kmem_cache * s,char * buf)6642 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
6643 {
6644 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU));
6645 }
6646 SLAB_ATTR_RO(destroy_by_rcu);
6647
6648 #ifdef CONFIG_SLUB_DEBUG
slabs_show(struct kmem_cache * s,char * buf)6649 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
6650 {
6651 return show_slab_objects(s, buf, SO_ALL);
6652 }
6653 SLAB_ATTR_RO(slabs);
6654
total_objects_show(struct kmem_cache * s,char * buf)6655 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
6656 {
6657 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
6658 }
6659 SLAB_ATTR_RO(total_objects);
6660
objects_show(struct kmem_cache * s,char * buf)6661 static ssize_t objects_show(struct kmem_cache *s, char *buf)
6662 {
6663 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
6664 }
6665 SLAB_ATTR_RO(objects);
6666
sanity_checks_show(struct kmem_cache * s,char * buf)6667 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
6668 {
6669 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
6670 }
6671 SLAB_ATTR_RO(sanity_checks);
6672
trace_show(struct kmem_cache * s,char * buf)6673 static ssize_t trace_show(struct kmem_cache *s, char *buf)
6674 {
6675 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TRACE));
6676 }
6677 SLAB_ATTR_RO(trace);
6678
red_zone_show(struct kmem_cache * s,char * buf)6679 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
6680 {
6681 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
6682 }
6683
6684 SLAB_ATTR_RO(red_zone);
6685
poison_show(struct kmem_cache * s,char * buf)6686 static ssize_t poison_show(struct kmem_cache *s, char *buf)
6687 {
6688 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_POISON));
6689 }
6690
6691 SLAB_ATTR_RO(poison);
6692
store_user_show(struct kmem_cache * s,char * buf)6693 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
6694 {
6695 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
6696 }
6697
6698 SLAB_ATTR_RO(store_user);
6699
validate_show(struct kmem_cache * s,char * buf)6700 static ssize_t validate_show(struct kmem_cache *s, char *buf)
6701 {
6702 return 0;
6703 }
6704
validate_store(struct kmem_cache * s,const char * buf,size_t length)6705 static ssize_t validate_store(struct kmem_cache *s,
6706 const char *buf, size_t length)
6707 {
6708 int ret = -EINVAL;
6709
6710 if (buf[0] == '1' && kmem_cache_debug(s)) {
6711 ret = validate_slab_cache(s);
6712 if (ret >= 0)
6713 ret = length;
6714 }
6715 return ret;
6716 }
6717 SLAB_ATTR(validate);
6718
6719 #endif /* CONFIG_SLUB_DEBUG */
6720
6721 #ifdef CONFIG_FAILSLAB
failslab_show(struct kmem_cache * s,char * buf)6722 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
6723 {
6724 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
6725 }
6726
failslab_store(struct kmem_cache * s,const char * buf,size_t length)6727 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
6728 size_t length)
6729 {
6730 if (s->refcount > 1)
6731 return -EINVAL;
6732
6733 if (buf[0] == '1')
6734 WRITE_ONCE(s->flags, s->flags | SLAB_FAILSLAB);
6735 else
6736 WRITE_ONCE(s->flags, s->flags & ~SLAB_FAILSLAB);
6737
6738 return length;
6739 }
6740 SLAB_ATTR(failslab);
6741 #endif
6742
shrink_show(struct kmem_cache * s,char * buf)6743 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
6744 {
6745 return 0;
6746 }
6747
shrink_store(struct kmem_cache * s,const char * buf,size_t length)6748 static ssize_t shrink_store(struct kmem_cache *s,
6749 const char *buf, size_t length)
6750 {
6751 if (buf[0] == '1')
6752 kmem_cache_shrink(s);
6753 else
6754 return -EINVAL;
6755 return length;
6756 }
6757 SLAB_ATTR(shrink);
6758
6759 #ifdef CONFIG_NUMA
remote_node_defrag_ratio_show(struct kmem_cache * s,char * buf)6760 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
6761 {
6762 return sysfs_emit(buf, "%u\n", s->remote_node_defrag_ratio / 10);
6763 }
6764
remote_node_defrag_ratio_store(struct kmem_cache * s,const char * buf,size_t length)6765 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
6766 const char *buf, size_t length)
6767 {
6768 unsigned int ratio;
6769 int err;
6770
6771 err = kstrtouint(buf, 10, &ratio);
6772 if (err)
6773 return err;
6774 if (ratio > 100)
6775 return -ERANGE;
6776
6777 s->remote_node_defrag_ratio = ratio * 10;
6778
6779 return length;
6780 }
6781 SLAB_ATTR(remote_node_defrag_ratio);
6782 #endif
6783
6784 #ifdef CONFIG_SLUB_STATS
show_stat(struct kmem_cache * s,char * buf,enum stat_item si)6785 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
6786 {
6787 unsigned long sum = 0;
6788 int cpu;
6789 int len = 0;
6790 int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL);
6791
6792 if (!data)
6793 return -ENOMEM;
6794
6795 for_each_online_cpu(cpu) {
6796 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
6797
6798 data[cpu] = x;
6799 sum += x;
6800 }
6801
6802 len += sysfs_emit_at(buf, len, "%lu", sum);
6803
6804 #ifdef CONFIG_SMP
6805 for_each_online_cpu(cpu) {
6806 if (data[cpu])
6807 len += sysfs_emit_at(buf, len, " C%d=%u",
6808 cpu, data[cpu]);
6809 }
6810 #endif
6811 kfree(data);
6812 len += sysfs_emit_at(buf, len, "\n");
6813
6814 return len;
6815 }
6816
clear_stat(struct kmem_cache * s,enum stat_item si)6817 static void clear_stat(struct kmem_cache *s, enum stat_item si)
6818 {
6819 int cpu;
6820
6821 for_each_online_cpu(cpu)
6822 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
6823 }
6824
6825 #define STAT_ATTR(si, text) \
6826 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
6827 { \
6828 return show_stat(s, buf, si); \
6829 } \
6830 static ssize_t text##_store(struct kmem_cache *s, \
6831 const char *buf, size_t length) \
6832 { \
6833 if (buf[0] != '0') \
6834 return -EINVAL; \
6835 clear_stat(s, si); \
6836 return length; \
6837 } \
6838 SLAB_ATTR(text); \
6839
6840 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
6841 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
6842 STAT_ATTR(FREE_FASTPATH, free_fastpath);
6843 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
6844 STAT_ATTR(FREE_FROZEN, free_frozen);
6845 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
6846 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
6847 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
6848 STAT_ATTR(ALLOC_SLAB, alloc_slab);
6849 STAT_ATTR(ALLOC_REFILL, alloc_refill);
6850 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
6851 STAT_ATTR(FREE_SLAB, free_slab);
6852 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
6853 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
6854 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
6855 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
6856 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
6857 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
6858 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
6859 STAT_ATTR(ORDER_FALLBACK, order_fallback);
6860 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
6861 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
6862 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
6863 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
6864 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
6865 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
6866 #endif /* CONFIG_SLUB_STATS */
6867
6868 #ifdef CONFIG_KFENCE
skip_kfence_show(struct kmem_cache * s,char * buf)6869 static ssize_t skip_kfence_show(struct kmem_cache *s, char *buf)
6870 {
6871 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_SKIP_KFENCE));
6872 }
6873
skip_kfence_store(struct kmem_cache * s,const char * buf,size_t length)6874 static ssize_t skip_kfence_store(struct kmem_cache *s,
6875 const char *buf, size_t length)
6876 {
6877 int ret = length;
6878
6879 if (buf[0] == '0')
6880 s->flags &= ~SLAB_SKIP_KFENCE;
6881 else if (buf[0] == '1')
6882 s->flags |= SLAB_SKIP_KFENCE;
6883 else
6884 ret = -EINVAL;
6885
6886 return ret;
6887 }
6888 SLAB_ATTR(skip_kfence);
6889 #endif
6890
6891 static struct attribute *slab_attrs[] = {
6892 &slab_size_attr.attr,
6893 &object_size_attr.attr,
6894 &objs_per_slab_attr.attr,
6895 &order_attr.attr,
6896 &min_partial_attr.attr,
6897 &cpu_partial_attr.attr,
6898 &objects_partial_attr.attr,
6899 &partial_attr.attr,
6900 &cpu_slabs_attr.attr,
6901 &ctor_attr.attr,
6902 &aliases_attr.attr,
6903 &align_attr.attr,
6904 &hwcache_align_attr.attr,
6905 &reclaim_account_attr.attr,
6906 &destroy_by_rcu_attr.attr,
6907 &shrink_attr.attr,
6908 &slabs_cpu_partial_attr.attr,
6909 #ifdef CONFIG_SLUB_DEBUG
6910 &total_objects_attr.attr,
6911 &objects_attr.attr,
6912 &slabs_attr.attr,
6913 &sanity_checks_attr.attr,
6914 &trace_attr.attr,
6915 &red_zone_attr.attr,
6916 &poison_attr.attr,
6917 &store_user_attr.attr,
6918 &validate_attr.attr,
6919 #endif
6920 #ifdef CONFIG_ZONE_DMA
6921 &cache_dma_attr.attr,
6922 #endif
6923 #ifdef CONFIG_NUMA
6924 &remote_node_defrag_ratio_attr.attr,
6925 #endif
6926 #ifdef CONFIG_SLUB_STATS
6927 &alloc_fastpath_attr.attr,
6928 &alloc_slowpath_attr.attr,
6929 &free_fastpath_attr.attr,
6930 &free_slowpath_attr.attr,
6931 &free_frozen_attr.attr,
6932 &free_add_partial_attr.attr,
6933 &free_remove_partial_attr.attr,
6934 &alloc_from_partial_attr.attr,
6935 &alloc_slab_attr.attr,
6936 &alloc_refill_attr.attr,
6937 &alloc_node_mismatch_attr.attr,
6938 &free_slab_attr.attr,
6939 &cpuslab_flush_attr.attr,
6940 &deactivate_full_attr.attr,
6941 &deactivate_empty_attr.attr,
6942 &deactivate_to_head_attr.attr,
6943 &deactivate_to_tail_attr.attr,
6944 &deactivate_remote_frees_attr.attr,
6945 &deactivate_bypass_attr.attr,
6946 &order_fallback_attr.attr,
6947 &cmpxchg_double_fail_attr.attr,
6948 &cmpxchg_double_cpu_fail_attr.attr,
6949 &cpu_partial_alloc_attr.attr,
6950 &cpu_partial_free_attr.attr,
6951 &cpu_partial_node_attr.attr,
6952 &cpu_partial_drain_attr.attr,
6953 #endif
6954 #ifdef CONFIG_FAILSLAB
6955 &failslab_attr.attr,
6956 #endif
6957 #ifdef CONFIG_HARDENED_USERCOPY
6958 &usersize_attr.attr,
6959 #endif
6960 #ifdef CONFIG_KFENCE
6961 &skip_kfence_attr.attr,
6962 #endif
6963
6964 NULL
6965 };
6966
6967 static const struct attribute_group slab_attr_group = {
6968 .attrs = slab_attrs,
6969 };
6970
slab_attr_show(struct kobject * kobj,struct attribute * attr,char * buf)6971 static ssize_t slab_attr_show(struct kobject *kobj,
6972 struct attribute *attr,
6973 char *buf)
6974 {
6975 struct slab_attribute *attribute;
6976 struct kmem_cache *s;
6977
6978 attribute = to_slab_attr(attr);
6979 s = to_slab(kobj);
6980
6981 if (!attribute->show)
6982 return -EIO;
6983
6984 return attribute->show(s, buf);
6985 }
6986
slab_attr_store(struct kobject * kobj,struct attribute * attr,const char * buf,size_t len)6987 static ssize_t slab_attr_store(struct kobject *kobj,
6988 struct attribute *attr,
6989 const char *buf, size_t len)
6990 {
6991 struct slab_attribute *attribute;
6992 struct kmem_cache *s;
6993
6994 attribute = to_slab_attr(attr);
6995 s = to_slab(kobj);
6996
6997 if (!attribute->store)
6998 return -EIO;
6999
7000 return attribute->store(s, buf, len);
7001 }
7002
kmem_cache_release(struct kobject * k)7003 static void kmem_cache_release(struct kobject *k)
7004 {
7005 slab_kmem_cache_release(to_slab(k));
7006 }
7007
7008 static const struct sysfs_ops slab_sysfs_ops = {
7009 .show = slab_attr_show,
7010 .store = slab_attr_store,
7011 };
7012
7013 static const struct kobj_type slab_ktype = {
7014 .sysfs_ops = &slab_sysfs_ops,
7015 .release = kmem_cache_release,
7016 };
7017
7018 static struct kset *slab_kset;
7019
cache_kset(struct kmem_cache * s)7020 static inline struct kset *cache_kset(struct kmem_cache *s)
7021 {
7022 return slab_kset;
7023 }
7024
7025 #define ID_STR_LENGTH 32
7026
7027 /* Create a unique string id for a slab cache:
7028 *
7029 * Format :[flags-]size
7030 */
create_unique_id(struct kmem_cache * s)7031 static char *create_unique_id(struct kmem_cache *s)
7032 {
7033 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
7034 char *p = name;
7035
7036 if (!name)
7037 return ERR_PTR(-ENOMEM);
7038
7039 *p++ = ':';
7040 /*
7041 * First flags affecting slabcache operations. We will only
7042 * get here for aliasable slabs so we do not need to support
7043 * too many flags. The flags here must cover all flags that
7044 * are matched during merging to guarantee that the id is
7045 * unique.
7046 */
7047 if (s->flags & SLAB_CACHE_DMA)
7048 *p++ = 'd';
7049 if (s->flags & SLAB_CACHE_DMA32)
7050 *p++ = 'D';
7051 if (s->flags & SLAB_RECLAIM_ACCOUNT)
7052 *p++ = 'a';
7053 if (s->flags & SLAB_CONSISTENCY_CHECKS)
7054 *p++ = 'F';
7055 if (s->flags & SLAB_ACCOUNT)
7056 *p++ = 'A';
7057 if (p != name + 1)
7058 *p++ = '-';
7059 p += snprintf(p, ID_STR_LENGTH - (p - name), "%07u", s->size);
7060
7061 if (WARN_ON(p > name + ID_STR_LENGTH - 1)) {
7062 kfree(name);
7063 return ERR_PTR(-EINVAL);
7064 }
7065 kmsan_unpoison_memory(name, p - name);
7066 return name;
7067 }
7068
sysfs_slab_add(struct kmem_cache * s)7069 static int sysfs_slab_add(struct kmem_cache *s)
7070 {
7071 int err;
7072 const char *name;
7073 struct kset *kset = cache_kset(s);
7074 int unmergeable = slab_unmergeable(s);
7075
7076 if (!unmergeable && disable_higher_order_debug &&
7077 (slub_debug & DEBUG_METADATA_FLAGS))
7078 unmergeable = 1;
7079
7080 if (unmergeable) {
7081 /*
7082 * Slabcache can never be merged so we can use the name proper.
7083 * This is typically the case for debug situations. In that
7084 * case we can catch duplicate names easily.
7085 */
7086 sysfs_remove_link(&slab_kset->kobj, s->name);
7087 name = s->name;
7088 } else {
7089 /*
7090 * Create a unique name for the slab as a target
7091 * for the symlinks.
7092 */
7093 name = create_unique_id(s);
7094 if (IS_ERR(name))
7095 return PTR_ERR(name);
7096 }
7097
7098 s->kobj.kset = kset;
7099 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
7100 if (err)
7101 goto out;
7102
7103 err = sysfs_create_group(&s->kobj, &slab_attr_group);
7104 if (err)
7105 goto out_del_kobj;
7106
7107 if (!unmergeable) {
7108 /* Setup first alias */
7109 sysfs_slab_alias(s, s->name);
7110 }
7111 out:
7112 if (!unmergeable)
7113 kfree(name);
7114 return err;
7115 out_del_kobj:
7116 kobject_del(&s->kobj);
7117 goto out;
7118 }
7119
sysfs_slab_unlink(struct kmem_cache * s)7120 void sysfs_slab_unlink(struct kmem_cache *s)
7121 {
7122 kobject_del(&s->kobj);
7123 }
7124
sysfs_slab_release(struct kmem_cache * s)7125 void sysfs_slab_release(struct kmem_cache *s)
7126 {
7127 kobject_put(&s->kobj);
7128 }
7129
7130 /*
7131 * Need to buffer aliases during bootup until sysfs becomes
7132 * available lest we lose that information.
7133 */
7134 struct saved_alias {
7135 struct kmem_cache *s;
7136 const char *name;
7137 struct saved_alias *next;
7138 };
7139
7140 static struct saved_alias *alias_list;
7141
sysfs_slab_alias(struct kmem_cache * s,const char * name)7142 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
7143 {
7144 struct saved_alias *al;
7145
7146 if (slab_state == FULL) {
7147 /*
7148 * If we have a leftover link then remove it.
7149 */
7150 sysfs_remove_link(&slab_kset->kobj, name);
7151 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
7152 }
7153
7154 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
7155 if (!al)
7156 return -ENOMEM;
7157
7158 al->s = s;
7159 al->name = name;
7160 al->next = alias_list;
7161 alias_list = al;
7162 kmsan_unpoison_memory(al, sizeof(*al));
7163 return 0;
7164 }
7165
slab_sysfs_init(void)7166 static int __init slab_sysfs_init(void)
7167 {
7168 struct kmem_cache *s;
7169 int err;
7170
7171 mutex_lock(&slab_mutex);
7172
7173 slab_kset = kset_create_and_add("slab", NULL, kernel_kobj);
7174 if (!slab_kset) {
7175 mutex_unlock(&slab_mutex);
7176 pr_err("Cannot register slab subsystem.\n");
7177 return -ENOMEM;
7178 }
7179
7180 slab_state = FULL;
7181
7182 list_for_each_entry(s, &slab_caches, list) {
7183 err = sysfs_slab_add(s);
7184 if (err)
7185 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
7186 s->name);
7187 }
7188
7189 while (alias_list) {
7190 struct saved_alias *al = alias_list;
7191
7192 alias_list = alias_list->next;
7193 err = sysfs_slab_alias(al->s, al->name);
7194 if (err)
7195 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
7196 al->name);
7197 kfree(al);
7198 }
7199
7200 mutex_unlock(&slab_mutex);
7201 return 0;
7202 }
7203 late_initcall(slab_sysfs_init);
7204 #endif /* SLAB_SUPPORTS_SYSFS */
7205
7206 #if defined(CONFIG_SLUB_DEBUG) && defined(CONFIG_DEBUG_FS)
slab_debugfs_show(struct seq_file * seq,void * v)7207 static int slab_debugfs_show(struct seq_file *seq, void *v)
7208 {
7209 struct loc_track *t = seq->private;
7210 struct location *l;
7211 unsigned long idx;
7212
7213 idx = (unsigned long) t->idx;
7214 if (idx < t->count) {
7215 l = &t->loc[idx];
7216
7217 seq_printf(seq, "%7ld ", l->count);
7218
7219 if (l->addr)
7220 seq_printf(seq, "%pS", (void *)l->addr);
7221 else
7222 seq_puts(seq, "<not-available>");
7223
7224 if (l->waste)
7225 seq_printf(seq, " waste=%lu/%lu",
7226 l->count * l->waste, l->waste);
7227
7228 if (l->sum_time != l->min_time) {
7229 seq_printf(seq, " age=%ld/%llu/%ld",
7230 l->min_time, div_u64(l->sum_time, l->count),
7231 l->max_time);
7232 } else
7233 seq_printf(seq, " age=%ld", l->min_time);
7234
7235 if (l->min_pid != l->max_pid)
7236 seq_printf(seq, " pid=%ld-%ld", l->min_pid, l->max_pid);
7237 else
7238 seq_printf(seq, " pid=%ld",
7239 l->min_pid);
7240
7241 if (num_online_cpus() > 1 && !cpumask_empty(to_cpumask(l->cpus)))
7242 seq_printf(seq, " cpus=%*pbl",
7243 cpumask_pr_args(to_cpumask(l->cpus)));
7244
7245 if (nr_online_nodes > 1 && !nodes_empty(l->nodes))
7246 seq_printf(seq, " nodes=%*pbl",
7247 nodemask_pr_args(&l->nodes));
7248
7249 #ifdef CONFIG_STACKDEPOT
7250 {
7251 depot_stack_handle_t handle;
7252 unsigned long *entries;
7253 unsigned int nr_entries, j;
7254
7255 handle = READ_ONCE(l->handle);
7256 if (handle) {
7257 nr_entries = stack_depot_fetch(handle, &entries);
7258 seq_puts(seq, "\n");
7259 for (j = 0; j < nr_entries; j++)
7260 seq_printf(seq, " %pS\n", (void *)entries[j]);
7261 }
7262 }
7263 #endif
7264 seq_puts(seq, "\n");
7265 }
7266
7267 if (!idx && !t->count)
7268 seq_puts(seq, "No data\n");
7269
7270 return 0;
7271 }
7272
slab_debugfs_stop(struct seq_file * seq,void * v)7273 static void slab_debugfs_stop(struct seq_file *seq, void *v)
7274 {
7275 }
7276
slab_debugfs_next(struct seq_file * seq,void * v,loff_t * ppos)7277 static void *slab_debugfs_next(struct seq_file *seq, void *v, loff_t *ppos)
7278 {
7279 struct loc_track *t = seq->private;
7280
7281 t->idx = ++(*ppos);
7282 if (*ppos <= t->count)
7283 return ppos;
7284
7285 return NULL;
7286 }
7287
cmp_loc_by_count(const void * a,const void * b,const void * data)7288 static int cmp_loc_by_count(const void *a, const void *b, const void *data)
7289 {
7290 struct location *loc1 = (struct location *)a;
7291 struct location *loc2 = (struct location *)b;
7292
7293 if (loc1->count > loc2->count)
7294 return -1;
7295 else
7296 return 1;
7297 }
7298
slab_debugfs_start(struct seq_file * seq,loff_t * ppos)7299 static void *slab_debugfs_start(struct seq_file *seq, loff_t *ppos)
7300 {
7301 struct loc_track *t = seq->private;
7302
7303 t->idx = *ppos;
7304 return ppos;
7305 }
7306
7307 static const struct seq_operations slab_debugfs_sops = {
7308 .start = slab_debugfs_start,
7309 .next = slab_debugfs_next,
7310 .stop = slab_debugfs_stop,
7311 .show = slab_debugfs_show,
7312 };
7313
slab_debug_trace_open(struct inode * inode,struct file * filep)7314 static int slab_debug_trace_open(struct inode *inode, struct file *filep)
7315 {
7316
7317 struct kmem_cache_node *n;
7318 enum track_item alloc;
7319 int node;
7320 struct loc_track *t = __seq_open_private(filep, &slab_debugfs_sops,
7321 sizeof(struct loc_track));
7322 struct kmem_cache *s = file_inode(filep)->i_private;
7323 unsigned long *obj_map;
7324
7325 if (!t)
7326 return -ENOMEM;
7327
7328 obj_map = bitmap_alloc(oo_objects(s->oo), GFP_KERNEL);
7329 if (!obj_map) {
7330 seq_release_private(inode, filep);
7331 return -ENOMEM;
7332 }
7333
7334 if (strcmp(filep->f_path.dentry->d_name.name, "alloc_traces") == 0)
7335 alloc = TRACK_ALLOC;
7336 else
7337 alloc = TRACK_FREE;
7338
7339 if (!alloc_loc_track(t, PAGE_SIZE / sizeof(struct location), GFP_KERNEL)) {
7340 bitmap_free(obj_map);
7341 seq_release_private(inode, filep);
7342 return -ENOMEM;
7343 }
7344
7345 for_each_kmem_cache_node(s, node, n) {
7346 unsigned long flags;
7347 struct slab *slab;
7348
7349 if (!node_nr_slabs(n))
7350 continue;
7351
7352 spin_lock_irqsave(&n->list_lock, flags);
7353 list_for_each_entry(slab, &n->partial, slab_list)
7354 process_slab(t, s, slab, alloc, obj_map);
7355 list_for_each_entry(slab, &n->full, slab_list)
7356 process_slab(t, s, slab, alloc, obj_map);
7357 spin_unlock_irqrestore(&n->list_lock, flags);
7358 }
7359
7360 /* Sort locations by count */
7361 sort_r(t->loc, t->count, sizeof(struct location),
7362 cmp_loc_by_count, NULL, NULL);
7363
7364 bitmap_free(obj_map);
7365 return 0;
7366 }
7367
slab_debug_trace_release(struct inode * inode,struct file * file)7368 static int slab_debug_trace_release(struct inode *inode, struct file *file)
7369 {
7370 struct seq_file *seq = file->private_data;
7371 struct loc_track *t = seq->private;
7372
7373 free_loc_track(t);
7374 return seq_release_private(inode, file);
7375 }
7376
7377 static const struct file_operations slab_debugfs_fops = {
7378 .open = slab_debug_trace_open,
7379 .read = seq_read,
7380 .llseek = seq_lseek,
7381 .release = slab_debug_trace_release,
7382 };
7383
debugfs_slab_add(struct kmem_cache * s)7384 static void debugfs_slab_add(struct kmem_cache *s)
7385 {
7386 struct dentry *slab_cache_dir;
7387
7388 if (unlikely(!slab_debugfs_root))
7389 return;
7390
7391 slab_cache_dir = debugfs_create_dir(s->name, slab_debugfs_root);
7392
7393 debugfs_create_file("alloc_traces", 0400,
7394 slab_cache_dir, s, &slab_debugfs_fops);
7395
7396 debugfs_create_file("free_traces", 0400,
7397 slab_cache_dir, s, &slab_debugfs_fops);
7398 }
7399
debugfs_slab_release(struct kmem_cache * s)7400 void debugfs_slab_release(struct kmem_cache *s)
7401 {
7402 debugfs_lookup_and_remove(s->name, slab_debugfs_root);
7403 }
7404
slab_debugfs_init(void)7405 static int __init slab_debugfs_init(void)
7406 {
7407 struct kmem_cache *s;
7408
7409 slab_debugfs_root = debugfs_create_dir("slab", NULL);
7410
7411 list_for_each_entry(s, &slab_caches, list)
7412 if (s->flags & SLAB_STORE_USER)
7413 debugfs_slab_add(s);
7414
7415 return 0;
7416
7417 }
7418 __initcall(slab_debugfs_init);
7419 #endif
7420 /*
7421 * The /proc/slabinfo ABI
7422 */
7423 #ifdef CONFIG_SLUB_DEBUG
get_slabinfo(struct kmem_cache * s,struct slabinfo * sinfo)7424 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
7425 {
7426 unsigned long nr_slabs = 0;
7427 unsigned long nr_objs = 0;
7428 unsigned long nr_free = 0;
7429 int node;
7430 struct kmem_cache_node *n;
7431
7432 for_each_kmem_cache_node(s, node, n) {
7433 nr_slabs += node_nr_slabs(n);
7434 nr_objs += node_nr_objs(n);
7435 nr_free += count_partial_free_approx(n);
7436 }
7437
7438 sinfo->active_objs = nr_objs - nr_free;
7439 sinfo->num_objs = nr_objs;
7440 sinfo->active_slabs = nr_slabs;
7441 sinfo->num_slabs = nr_slabs;
7442 sinfo->objects_per_slab = oo_objects(s->oo);
7443 sinfo->cache_order = oo_order(s->oo);
7444 }
7445 #endif /* CONFIG_SLUB_DEBUG */
7446