Line data Source code
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> /* struct reclaim_state */
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/cpu.h>
26 : #include <linux/cpuset.h>
27 : #include <linux/mempolicy.h>
28 : #include <linux/ctype.h>
29 : #include <linux/debugobjects.h>
30 : #include <linux/kallsyms.h>
31 : #include <linux/kfence.h>
32 : #include <linux/memory.h>
33 : #include <linux/math64.h>
34 : #include <linux/fault-inject.h>
35 : #include <linux/stacktrace.h>
36 : #include <linux/prefetch.h>
37 : #include <linux/memcontrol.h>
38 : #include <linux/random.h>
39 : #include <kunit/test.h>
40 :
41 : #include <linux/debugfs.h>
42 : #include <trace/events/kmem.h>
43 :
44 : #include "internal.h"
45 :
46 : /*
47 : * Lock order:
48 : * 1. slab_mutex (Global Mutex)
49 : * 2. node->list_lock (Spinlock)
50 : * 3. kmem_cache->cpu_slab->lock (Local lock)
51 : * 4. slab_lock(slab) (Only on some arches or for debugging)
52 : * 5. object_map_lock (Only for debugging)
53 : *
54 : * slab_mutex
55 : *
56 : * The role of the slab_mutex is to protect the list of all the slabs
57 : * and to synchronize major metadata changes to slab cache structures.
58 : * Also synchronizes memory hotplug callbacks.
59 : *
60 : * slab_lock
61 : *
62 : * The slab_lock is a wrapper around the page lock, thus it is a bit
63 : * spinlock.
64 : *
65 : * The slab_lock is only used for debugging and on arches that do not
66 : * have the ability to do a cmpxchg_double. It only protects:
67 : * A. slab->freelist -> List of free objects in a slab
68 : * B. slab->inuse -> Number of objects in use
69 : * C. slab->objects -> Number of objects in slab
70 : * D. slab->frozen -> frozen state
71 : *
72 : * Frozen slabs
73 : *
74 : * If a slab is frozen then it is exempt from list management. It is not
75 : * on any list except per cpu partial list. The processor that froze the
76 : * slab is the one who can perform list operations on the slab. Other
77 : * processors may put objects onto the freelist but the processor that
78 : * froze the slab is the only one that can retrieve the objects from the
79 : * slab's freelist.
80 : *
81 : * list_lock
82 : *
83 : * The list_lock protects the partial and full list on each node and
84 : * the partial slab counter. If taken then no new slabs may be added or
85 : * removed from the lists nor make the number of partial slabs be modified.
86 : * (Note that the total number of slabs is an atomic value that may be
87 : * modified without taking the list lock).
88 : *
89 : * The list_lock is a centralized lock and thus we avoid taking it as
90 : * much as possible. As long as SLUB does not have to handle partial
91 : * slabs, operations can continue without any centralized lock. F.e.
92 : * allocating a long series of objects that fill up slabs does not require
93 : * the list lock.
94 : *
95 : * cpu_slab->lock local lock
96 : *
97 : * This locks protect slowpath manipulation of all kmem_cache_cpu fields
98 : * except the stat counters. This is a percpu structure manipulated only by
99 : * the local cpu, so the lock protects against being preempted or interrupted
100 : * by an irq. Fast path operations rely on lockless operations instead.
101 : * On PREEMPT_RT, the local lock does not actually disable irqs (and thus
102 : * prevent the lockless operations), so fastpath operations also need to take
103 : * the lock and are no longer lockless.
104 : *
105 : * lockless fastpaths
106 : *
107 : * The fast path allocation (slab_alloc_node()) and freeing (do_slab_free())
108 : * are fully lockless when satisfied from the percpu slab (and when
109 : * cmpxchg_double is possible to use, otherwise slab_lock is taken).
110 : * They also don't disable preemption or migration or irqs. They rely on
111 : * the transaction id (tid) field to detect being preempted or moved to
112 : * another cpu.
113 : *
114 : * irq, preemption, migration considerations
115 : *
116 : * Interrupts are disabled as part of list_lock or local_lock operations, or
117 : * around the slab_lock operation, in order to make the slab allocator safe
118 : * to use in the context of an irq.
119 : *
120 : * In addition, preemption (or migration on PREEMPT_RT) is disabled in the
121 : * allocation slowpath, bulk allocation, and put_cpu_partial(), so that the
122 : * local cpu doesn't change in the process and e.g. the kmem_cache_cpu pointer
123 : * doesn't have to be revalidated in each section protected by the local lock.
124 : *
125 : * SLUB assigns one slab for allocation to each processor.
126 : * Allocations only occur from these slabs called cpu slabs.
127 : *
128 : * Slabs with free elements are kept on a partial list and during regular
129 : * operations no list for full slabs is used. If an object in a full slab is
130 : * freed then the slab will show up again on the partial lists.
131 : * We track full slabs for debugging purposes though because otherwise we
132 : * cannot scan all objects.
133 : *
134 : * Slabs are freed when they become empty. Teardown and setup is
135 : * minimal so we rely on the page allocators per cpu caches for
136 : * fast frees and allocs.
137 : *
138 : * slab->frozen The slab is frozen and exempt from list processing.
139 : * This means that the slab is dedicated to a purpose
140 : * such as satisfying allocations for a specific
141 : * processor. Objects may be freed in the slab while
142 : * it is frozen but slab_free will then skip the usual
143 : * list operations. It is up to the processor holding
144 : * the slab to integrate the slab into the slab lists
145 : * when the slab is no longer needed.
146 : *
147 : * One use of this flag is to mark slabs that are
148 : * used for allocations. Then such a slab becomes a cpu
149 : * slab. The cpu slab may be equipped with an additional
150 : * freelist that allows lockless access to
151 : * free objects in addition to the regular freelist
152 : * that requires the slab lock.
153 : *
154 : * SLAB_DEBUG_FLAGS Slab requires special handling due to debug
155 : * options set. This moves slab handling out of
156 : * the fast path and disables lockless freelists.
157 : */
158 :
159 : /*
160 : * We could simply use migrate_disable()/enable() but as long as it's a
161 : * function call even on !PREEMPT_RT, use inline preempt_disable() there.
162 : */
163 : #ifndef CONFIG_PREEMPT_RT
164 : #define slub_get_cpu_ptr(var) get_cpu_ptr(var)
165 : #define slub_put_cpu_ptr(var) put_cpu_ptr(var)
166 : #else
167 : #define slub_get_cpu_ptr(var) \
168 : ({ \
169 : migrate_disable(); \
170 : this_cpu_ptr(var); \
171 : })
172 : #define slub_put_cpu_ptr(var) \
173 : do { \
174 : (void)(var); \
175 : migrate_enable(); \
176 : } while (0)
177 : #endif
178 :
179 : #ifdef CONFIG_SLUB_DEBUG
180 : #ifdef CONFIG_SLUB_DEBUG_ON
181 : DEFINE_STATIC_KEY_TRUE(slub_debug_enabled);
182 : #else
183 : DEFINE_STATIC_KEY_FALSE(slub_debug_enabled);
184 : #endif
185 : #endif /* CONFIG_SLUB_DEBUG */
186 :
187 : static inline bool kmem_cache_debug(struct kmem_cache *s)
188 : {
189 3554 : return kmem_cache_debug_flags(s, SLAB_DEBUG_FLAGS);
190 : }
191 :
192 0 : void *fixup_red_left(struct kmem_cache *s, void *p)
193 : {
194 908 : if (kmem_cache_debug_flags(s, SLAB_RED_ZONE))
195 0 : p += s->red_left_pad;
196 :
197 0 : return p;
198 : }
199 :
200 : static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
201 : {
202 : #ifdef CONFIG_SLUB_CPU_PARTIAL
203 : return !kmem_cache_debug(s);
204 : #else
205 : return false;
206 : #endif
207 : }
208 :
209 : /*
210 : * Issues still to be resolved:
211 : *
212 : * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
213 : *
214 : * - Variable sizing of the per node arrays
215 : */
216 :
217 : /* Enable to log cmpxchg failures */
218 : #undef SLUB_DEBUG_CMPXCHG
219 :
220 : /*
221 : * Minimum number of partial slabs. These will be left on the partial
222 : * lists even if they are empty. kmem_cache_shrink may reclaim them.
223 : */
224 : #define MIN_PARTIAL 5
225 :
226 : /*
227 : * Maximum number of desirable partial slabs.
228 : * The existence of more partial slabs makes kmem_cache_shrink
229 : * sort the partial list by the number of objects in use.
230 : */
231 : #define MAX_PARTIAL 10
232 :
233 : #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
234 : SLAB_POISON | SLAB_STORE_USER)
235 :
236 : /*
237 : * These debug flags cannot use CMPXCHG because there might be consistency
238 : * issues when checking or reading debug information
239 : */
240 : #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
241 : SLAB_TRACE)
242 :
243 :
244 : /*
245 : * Debugging flags that require metadata to be stored in the slab. These get
246 : * disabled when slub_debug=O is used and a cache's min order increases with
247 : * metadata.
248 : */
249 : #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
250 :
251 : #define OO_SHIFT 16
252 : #define OO_MASK ((1 << OO_SHIFT) - 1)
253 : #define MAX_OBJS_PER_PAGE 32767 /* since slab.objects is u15 */
254 :
255 : /* Internal SLUB flags */
256 : /* Poison object */
257 : #define __OBJECT_POISON ((slab_flags_t __force)0x80000000U)
258 : /* Use cmpxchg_double */
259 : #define __CMPXCHG_DOUBLE ((slab_flags_t __force)0x40000000U)
260 :
261 : /*
262 : * Tracking user of a slab.
263 : */
264 : #define TRACK_ADDRS_COUNT 16
265 : struct track {
266 : unsigned long addr; /* Called from address */
267 : #ifdef CONFIG_STACKTRACE
268 : unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
269 : #endif
270 : int cpu; /* Was running on cpu */
271 : int pid; /* Pid context */
272 : unsigned long when; /* When did the operation occur */
273 : };
274 :
275 : enum track_item { TRACK_ALLOC, TRACK_FREE };
276 :
277 : #ifdef CONFIG_SYSFS
278 : static int sysfs_slab_add(struct kmem_cache *);
279 : static int sysfs_slab_alias(struct kmem_cache *, const char *);
280 : #else
281 : static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
282 : static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
283 : { return 0; }
284 : #endif
285 :
286 : #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_SLUB_DEBUG)
287 : static void debugfs_slab_add(struct kmem_cache *);
288 : #else
289 : static inline void debugfs_slab_add(struct kmem_cache *s) { }
290 : #endif
291 :
292 : static inline void stat(const struct kmem_cache *s, enum stat_item si)
293 : {
294 : #ifdef CONFIG_SLUB_STATS
295 : /*
296 : * The rmw is racy on a preemptible kernel but this is acceptable, so
297 : * avoid this_cpu_add()'s irq-disable overhead.
298 : */
299 : raw_cpu_inc(s->cpu_slab->stat[si]);
300 : #endif
301 : }
302 :
303 : /*
304 : * Tracks for which NUMA nodes we have kmem_cache_nodes allocated.
305 : * Corresponds to node_state[N_NORMAL_MEMORY], but can temporarily
306 : * differ during memory hotplug/hotremove operations.
307 : * Protected by slab_mutex.
308 : */
309 : static nodemask_t slab_nodes;
310 :
311 : /********************************************************************
312 : * Core slab cache functions
313 : *******************************************************************/
314 :
315 : /*
316 : * Returns freelist pointer (ptr). With hardening, this is obfuscated
317 : * with an XOR of the address where the pointer is held and a per-cache
318 : * random number.
319 : */
320 : static inline void *freelist_ptr(const struct kmem_cache *s, void *ptr,
321 : unsigned long ptr_addr)
322 : {
323 : #ifdef CONFIG_SLAB_FREELIST_HARDENED
324 : /*
325 : * When CONFIG_KASAN_SW/HW_TAGS is enabled, ptr_addr might be tagged.
326 : * Normally, this doesn't cause any issues, as both set_freepointer()
327 : * and get_freepointer() are called with a pointer with the same tag.
328 : * However, there are some issues with CONFIG_SLUB_DEBUG code. For
329 : * example, when __free_slub() iterates over objects in a cache, it
330 : * passes untagged pointers to check_object(). check_object() in turns
331 : * calls get_freepointer() with an untagged pointer, which causes the
332 : * freepointer to be restored incorrectly.
333 : */
334 : return (void *)((unsigned long)ptr ^ s->random ^
335 : swab((unsigned long)kasan_reset_tag((void *)ptr_addr)));
336 : #else
337 : return ptr;
338 : #endif
339 : }
340 :
341 : /* Returns the freelist pointer recorded at location ptr_addr. */
342 : static inline void *freelist_dereference(const struct kmem_cache *s,
343 : void *ptr_addr)
344 : {
345 23721 : return freelist_ptr(s, (void *)*(unsigned long *)(ptr_addr),
346 : (unsigned long)ptr_addr);
347 : }
348 :
349 : static inline void *get_freepointer(struct kmem_cache *s, void *object)
350 : {
351 23721 : object = kasan_reset_tag(object);
352 47442 : return freelist_dereference(s, object + s->offset);
353 : }
354 :
355 : static void prefetch_freepointer(const struct kmem_cache *s, void *object)
356 : {
357 17828 : prefetchw(object + s->offset);
358 : }
359 :
360 : static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
361 : {
362 : unsigned long freepointer_addr;
363 : void *p;
364 :
365 : if (!debug_pagealloc_enabled_static())
366 35656 : return get_freepointer(s, object);
367 :
368 : object = kasan_reset_tag(object);
369 : freepointer_addr = (unsigned long)object + s->offset;
370 : copy_from_kernel_nofault(&p, (void **)freepointer_addr, sizeof(p));
371 : return freelist_ptr(s, p, freepointer_addr);
372 : }
373 :
374 : static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
375 : {
376 24810 : unsigned long freeptr_addr = (unsigned long)object + s->offset;
377 :
378 : #ifdef CONFIG_SLAB_FREELIST_HARDENED
379 : BUG_ON(object == fp); /* naive detection of double free or corruption */
380 : #endif
381 :
382 24810 : freeptr_addr = (unsigned long)kasan_reset_tag((void *)freeptr_addr);
383 24810 : *(void **)freeptr_addr = freelist_ptr(s, fp, freeptr_addr);
384 : }
385 :
386 : /* Loop over all objects in a slab */
387 : #define for_each_object(__p, __s, __addr, __objects) \
388 : for (__p = fixup_red_left(__s, __addr); \
389 : __p < (__addr) + (__objects) * (__s)->size; \
390 : __p += (__s)->size)
391 :
392 : static inline unsigned int order_objects(unsigned int order, unsigned int size)
393 : {
394 269 : return ((unsigned int)PAGE_SIZE << order) / size;
395 : }
396 :
397 : static inline struct kmem_cache_order_objects oo_make(unsigned int order,
398 : unsigned int size)
399 : {
400 134 : struct kmem_cache_order_objects x = {
401 268 : (order << OO_SHIFT) + order_objects(order, size)
402 : };
403 :
404 : return x;
405 : }
406 :
407 : static inline unsigned int oo_order(struct kmem_cache_order_objects x)
408 : {
409 1318 : return x.x >> OO_SHIFT;
410 : }
411 :
412 : static inline unsigned int oo_objects(struct kmem_cache_order_objects x)
413 : {
414 67 : return x.x & OO_MASK;
415 : }
416 :
417 : #ifdef CONFIG_SLUB_CPU_PARTIAL
418 : static void slub_set_cpu_partial(struct kmem_cache *s, unsigned int nr_objects)
419 : {
420 : unsigned int nr_slabs;
421 :
422 : s->cpu_partial = nr_objects;
423 :
424 : /*
425 : * We take the number of objects but actually limit the number of
426 : * slabs on the per cpu partial list, in order to limit excessive
427 : * growth of the list. For simplicity we assume that the slabs will
428 : * be half-full.
429 : */
430 : nr_slabs = DIV_ROUND_UP(nr_objects * 2, oo_objects(s->oo));
431 : s->cpu_partial_slabs = nr_slabs;
432 : }
433 : #else
434 : static inline void
435 : slub_set_cpu_partial(struct kmem_cache *s, unsigned int nr_objects)
436 : {
437 : }
438 : #endif /* CONFIG_SLUB_CPU_PARTIAL */
439 :
440 : /*
441 : * Per slab locking using the pagelock
442 : */
443 : static __always_inline void __slab_lock(struct slab *slab)
444 : {
445 1789 : struct page *page = slab_page(slab);
446 :
447 : VM_BUG_ON_PAGE(PageTail(page), page);
448 1789 : bit_spin_lock(PG_locked, &page->flags);
449 : }
450 :
451 : static __always_inline void __slab_unlock(struct slab *slab)
452 : {
453 1789 : struct page *page = slab_page(slab);
454 :
455 : VM_BUG_ON_PAGE(PageTail(page), page);
456 1789 : __bit_spin_unlock(PG_locked, &page->flags);
457 : }
458 :
459 : static __always_inline void slab_lock(struct slab *slab, unsigned long *flags)
460 : {
461 : if (IS_ENABLED(CONFIG_PREEMPT_RT))
462 : local_irq_save(*flags);
463 509 : __slab_lock(slab);
464 : }
465 :
466 : static __always_inline void slab_unlock(struct slab *slab, unsigned long *flags)
467 : {
468 509 : __slab_unlock(slab);
469 : if (IS_ENABLED(CONFIG_PREEMPT_RT))
470 : local_irq_restore(*flags);
471 : }
472 :
473 : /*
474 : * Interrupts must be disabled (for the fallback code to work right), typically
475 : * by an _irqsave() lock variant. Except on PREEMPT_RT where locks are different
476 : * so we disable interrupts as part of slab_[un]lock().
477 : */
478 : static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct slab *slab,
479 : void *freelist_old, unsigned long counters_old,
480 : void *freelist_new, unsigned long counters_new,
481 : const char *n)
482 : {
483 : if (!IS_ENABLED(CONFIG_PREEMPT_RT))
484 : lockdep_assert_irqs_disabled();
485 : #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
486 : defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
487 : if (s->flags & __CMPXCHG_DOUBLE) {
488 : if (cmpxchg_double(&slab->freelist, &slab->counters,
489 : freelist_old, counters_old,
490 : freelist_new, counters_new))
491 : return true;
492 : } else
493 : #endif
494 : {
495 : /* init to 0 to prevent spurious warnings */
496 509 : unsigned long flags = 0;
497 :
498 509 : slab_lock(slab, &flags);
499 1018 : if (slab->freelist == freelist_old &&
500 509 : slab->counters == counters_old) {
501 509 : slab->freelist = freelist_new;
502 509 : slab->counters = counters_new;
503 509 : slab_unlock(slab, &flags);
504 509 : return true;
505 : }
506 0 : slab_unlock(slab, &flags);
507 : }
508 :
509 : cpu_relax();
510 0 : stat(s, CMPXCHG_DOUBLE_FAIL);
511 :
512 : #ifdef SLUB_DEBUG_CMPXCHG
513 : pr_info("%s %s: cmpxchg double redo ", n, s->name);
514 : #endif
515 :
516 : return false;
517 : }
518 :
519 1280 : static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct slab *slab,
520 : void *freelist_old, unsigned long counters_old,
521 : void *freelist_new, unsigned long counters_new,
522 : const char *n)
523 : {
524 : #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
525 : defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
526 : if (s->flags & __CMPXCHG_DOUBLE) {
527 : if (cmpxchg_double(&slab->freelist, &slab->counters,
528 : freelist_old, counters_old,
529 : freelist_new, counters_new))
530 : return true;
531 : } else
532 : #endif
533 : {
534 : unsigned long flags;
535 :
536 1280 : local_irq_save(flags);
537 1280 : __slab_lock(slab);
538 2560 : if (slab->freelist == freelist_old &&
539 1280 : slab->counters == counters_old) {
540 1280 : slab->freelist = freelist_new;
541 1280 : slab->counters = counters_new;
542 1280 : __slab_unlock(slab);
543 2560 : local_irq_restore(flags);
544 : return true;
545 : }
546 0 : __slab_unlock(slab);
547 0 : local_irq_restore(flags);
548 : }
549 :
550 : cpu_relax();
551 0 : stat(s, CMPXCHG_DOUBLE_FAIL);
552 :
553 : #ifdef SLUB_DEBUG_CMPXCHG
554 : pr_info("%s %s: cmpxchg double redo ", n, s->name);
555 : #endif
556 :
557 : return false;
558 : }
559 :
560 : #ifdef CONFIG_SLUB_DEBUG
561 : static unsigned long object_map[BITS_TO_LONGS(MAX_OBJS_PER_PAGE)];
562 : static DEFINE_RAW_SPINLOCK(object_map_lock);
563 :
564 0 : static void __fill_map(unsigned long *obj_map, struct kmem_cache *s,
565 : struct slab *slab)
566 : {
567 0 : void *addr = slab_address(slab);
568 : void *p;
569 :
570 0 : bitmap_zero(obj_map, slab->objects);
571 :
572 0 : for (p = slab->freelist; p; p = get_freepointer(s, p))
573 0 : set_bit(__obj_to_index(s, addr, p), obj_map);
574 0 : }
575 :
576 : #if IS_ENABLED(CONFIG_KUNIT)
577 0 : static bool slab_add_kunit_errors(void)
578 : {
579 : struct kunit_resource *resource;
580 :
581 0 : if (likely(!current->kunit_test))
582 : return false;
583 :
584 0 : resource = kunit_find_named_resource(current->kunit_test, "slab_errors");
585 0 : if (!resource)
586 : return false;
587 :
588 0 : (*(int *)resource->data)++;
589 0 : kunit_put_resource(resource);
590 0 : return true;
591 : }
592 : #else
593 : static inline bool slab_add_kunit_errors(void) { return false; }
594 : #endif
595 :
596 : /*
597 : * Determine a map of objects in use in a slab.
598 : *
599 : * Node listlock must be held to guarantee that the slab does
600 : * not vanish from under us.
601 : */
602 : static unsigned long *get_map(struct kmem_cache *s, struct slab *slab)
603 : __acquires(&object_map_lock)
604 : {
605 : VM_BUG_ON(!irqs_disabled());
606 :
607 0 : raw_spin_lock(&object_map_lock);
608 :
609 0 : __fill_map(object_map, s, slab);
610 :
611 : return object_map;
612 : }
613 :
614 : static void put_map(unsigned long *map) __releases(&object_map_lock)
615 : {
616 : VM_BUG_ON(map != object_map);
617 0 : raw_spin_unlock(&object_map_lock);
618 : }
619 :
620 : static inline unsigned int size_from_object(struct kmem_cache *s)
621 : {
622 0 : if (s->flags & SLAB_RED_ZONE)
623 0 : return s->size - s->red_left_pad;
624 :
625 : return s->size;
626 : }
627 :
628 : static inline void *restore_red_left(struct kmem_cache *s, void *p)
629 : {
630 0 : if (s->flags & SLAB_RED_ZONE)
631 0 : p -= s->red_left_pad;
632 :
633 : return p;
634 : }
635 :
636 : /*
637 : * Debug settings:
638 : */
639 : #if defined(CONFIG_SLUB_DEBUG_ON)
640 : static slab_flags_t slub_debug = DEBUG_DEFAULT_FLAGS;
641 : #else
642 : static slab_flags_t slub_debug;
643 : #endif
644 :
645 : static char *slub_debug_string;
646 : static int disable_higher_order_debug;
647 :
648 : /*
649 : * slub is about to manipulate internal object metadata. This memory lies
650 : * outside the range of the allocated object, so accessing it would normally
651 : * be reported by kasan as a bounds error. metadata_access_enable() is used
652 : * to tell kasan that these accesses are OK.
653 : */
654 : static inline void metadata_access_enable(void)
655 : {
656 : kasan_disable_current();
657 : }
658 :
659 : static inline void metadata_access_disable(void)
660 : {
661 : kasan_enable_current();
662 : }
663 :
664 : /*
665 : * Object debugging
666 : */
667 :
668 : /* Verify that a pointer has an address that is valid within a slab page */
669 0 : static inline int check_valid_pointer(struct kmem_cache *s,
670 : struct slab *slab, void *object)
671 : {
672 : void *base;
673 :
674 0 : if (!object)
675 : return 1;
676 :
677 0 : base = slab_address(slab);
678 0 : object = kasan_reset_tag(object);
679 0 : object = restore_red_left(s, object);
680 0 : if (object < base || object >= base + slab->objects * s->size ||
681 0 : (object - base) % s->size) {
682 : return 0;
683 : }
684 :
685 0 : return 1;
686 : }
687 :
688 : static void print_section(char *level, char *text, u8 *addr,
689 : unsigned int length)
690 : {
691 : metadata_access_enable();
692 0 : print_hex_dump(level, text, DUMP_PREFIX_ADDRESS,
693 0 : 16, 1, kasan_reset_tag((void *)addr), length, 1);
694 : metadata_access_disable();
695 : }
696 :
697 : /*
698 : * See comment in calculate_sizes().
699 : */
700 : static inline bool freeptr_outside_object(struct kmem_cache *s)
701 : {
702 : return s->offset >= s->inuse;
703 : }
704 :
705 : /*
706 : * Return offset of the end of info block which is inuse + free pointer if
707 : * not overlapping with object.
708 : */
709 : static inline unsigned int get_info_end(struct kmem_cache *s)
710 : {
711 0 : if (freeptr_outside_object(s))
712 0 : return s->inuse + sizeof(void *);
713 : else
714 : return s->inuse;
715 : }
716 :
717 : static struct track *get_track(struct kmem_cache *s, void *object,
718 : enum track_item alloc)
719 : {
720 : struct track *p;
721 :
722 0 : p = object + get_info_end(s);
723 :
724 0 : return kasan_reset_tag(p + alloc);
725 : }
726 :
727 0 : static void set_track(struct kmem_cache *s, void *object,
728 : enum track_item alloc, unsigned long addr)
729 : {
730 0 : struct track *p = get_track(s, object, alloc);
731 :
732 0 : if (addr) {
733 : #ifdef CONFIG_STACKTRACE
734 : unsigned int nr_entries;
735 :
736 : metadata_access_enable();
737 0 : nr_entries = stack_trace_save(kasan_reset_tag(p->addrs),
738 : TRACK_ADDRS_COUNT, 3);
739 : metadata_access_disable();
740 :
741 0 : if (nr_entries < TRACK_ADDRS_COUNT)
742 0 : p->addrs[nr_entries] = 0;
743 : #endif
744 0 : p->addr = addr;
745 0 : p->cpu = smp_processor_id();
746 0 : p->pid = current->pid;
747 0 : p->when = jiffies;
748 : } else {
749 0 : memset(p, 0, sizeof(struct track));
750 : }
751 0 : }
752 :
753 1 : static void init_tracking(struct kmem_cache *s, void *object)
754 : {
755 1 : if (!(s->flags & SLAB_STORE_USER))
756 : return;
757 :
758 0 : set_track(s, object, TRACK_FREE, 0UL);
759 0 : set_track(s, object, TRACK_ALLOC, 0UL);
760 : }
761 :
762 0 : static void print_track(const char *s, struct track *t, unsigned long pr_time)
763 : {
764 0 : if (!t->addr)
765 : return;
766 :
767 0 : pr_err("%s in %pS age=%lu cpu=%u pid=%d\n",
768 : s, (void *)t->addr, pr_time - t->when, t->cpu, t->pid);
769 : #ifdef CONFIG_STACKTRACE
770 : {
771 : int i;
772 0 : for (i = 0; i < TRACK_ADDRS_COUNT; i++)
773 0 : if (t->addrs[i])
774 0 : pr_err("\t%pS\n", (void *)t->addrs[i]);
775 : else
776 : break;
777 : }
778 : #endif
779 : }
780 :
781 0 : void print_tracking(struct kmem_cache *s, void *object)
782 : {
783 0 : unsigned long pr_time = jiffies;
784 0 : if (!(s->flags & SLAB_STORE_USER))
785 : return;
786 :
787 0 : print_track("Allocated", get_track(s, object, TRACK_ALLOC), pr_time);
788 0 : print_track("Freed", get_track(s, object, TRACK_FREE), pr_time);
789 : }
790 :
791 : static void print_slab_info(const struct slab *slab)
792 : {
793 0 : struct folio *folio = (struct folio *)slab_folio(slab);
794 :
795 0 : pr_err("Slab 0x%p objects=%u used=%u fp=0x%p flags=%pGp\n",
796 : slab, slab->objects, slab->inuse, slab->freelist,
797 : folio_flags(folio, 0));
798 : }
799 :
800 0 : static void slab_bug(struct kmem_cache *s, char *fmt, ...)
801 : {
802 : struct va_format vaf;
803 : va_list args;
804 :
805 0 : va_start(args, fmt);
806 0 : vaf.fmt = fmt;
807 0 : vaf.va = &args;
808 0 : pr_err("=============================================================================\n");
809 0 : pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
810 0 : pr_err("-----------------------------------------------------------------------------\n\n");
811 0 : va_end(args);
812 0 : }
813 :
814 : __printf(2, 3)
815 0 : static void slab_fix(struct kmem_cache *s, char *fmt, ...)
816 : {
817 : struct va_format vaf;
818 : va_list args;
819 :
820 0 : if (slab_add_kunit_errors())
821 0 : return;
822 :
823 0 : va_start(args, fmt);
824 0 : vaf.fmt = fmt;
825 0 : vaf.va = &args;
826 0 : pr_err("FIX %s: %pV\n", s->name, &vaf);
827 0 : va_end(args);
828 : }
829 :
830 0 : static void print_trailer(struct kmem_cache *s, struct slab *slab, u8 *p)
831 : {
832 : unsigned int off; /* Offset of last byte */
833 0 : u8 *addr = slab_address(slab);
834 :
835 0 : print_tracking(s, p);
836 :
837 0 : print_slab_info(slab);
838 :
839 0 : pr_err("Object 0x%p @offset=%tu fp=0x%p\n\n",
840 : p, p - addr, get_freepointer(s, p));
841 :
842 0 : if (s->flags & SLAB_RED_ZONE)
843 0 : print_section(KERN_ERR, "Redzone ", p - s->red_left_pad,
844 : s->red_left_pad);
845 0 : else if (p > addr + 16)
846 0 : print_section(KERN_ERR, "Bytes b4 ", p - 16, 16);
847 :
848 0 : print_section(KERN_ERR, "Object ", p,
849 0 : min_t(unsigned int, s->object_size, PAGE_SIZE));
850 0 : if (s->flags & SLAB_RED_ZONE)
851 0 : print_section(KERN_ERR, "Redzone ", p + s->object_size,
852 0 : s->inuse - s->object_size);
853 :
854 0 : off = get_info_end(s);
855 :
856 0 : if (s->flags & SLAB_STORE_USER)
857 0 : off += 2 * sizeof(struct track);
858 :
859 0 : off += kasan_metadata_size(s);
860 :
861 0 : if (off != size_from_object(s))
862 : /* Beginning of the filler is the free pointer */
863 0 : print_section(KERN_ERR, "Padding ", p + off,
864 0 : size_from_object(s) - off);
865 :
866 0 : dump_stack();
867 0 : }
868 :
869 0 : static void object_err(struct kmem_cache *s, struct slab *slab,
870 : u8 *object, char *reason)
871 : {
872 0 : if (slab_add_kunit_errors())
873 : return;
874 :
875 0 : slab_bug(s, "%s", reason);
876 0 : print_trailer(s, slab, object);
877 0 : add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
878 : }
879 :
880 82 : static bool freelist_corrupted(struct kmem_cache *s, struct slab *slab,
881 : void **freelist, void *nextfree)
882 : {
883 82 : if ((s->flags & SLAB_CONSISTENCY_CHECKS) &&
884 0 : !check_valid_pointer(s, slab, nextfree) && freelist) {
885 0 : object_err(s, slab, *freelist, "Freechain corrupt");
886 0 : *freelist = NULL;
887 0 : slab_fix(s, "Isolate corrupted freechain");
888 0 : return true;
889 : }
890 :
891 : return false;
892 : }
893 :
894 0 : static __printf(3, 4) void slab_err(struct kmem_cache *s, struct slab *slab,
895 : const char *fmt, ...)
896 : {
897 : va_list args;
898 : char buf[100];
899 :
900 0 : if (slab_add_kunit_errors())
901 0 : return;
902 :
903 0 : va_start(args, fmt);
904 0 : vsnprintf(buf, sizeof(buf), fmt, args);
905 0 : va_end(args);
906 0 : slab_bug(s, "%s", buf);
907 0 : print_slab_info(slab);
908 0 : dump_stack();
909 0 : add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
910 : }
911 :
912 1 : static void init_object(struct kmem_cache *s, void *object, u8 val)
913 : {
914 1 : u8 *p = kasan_reset_tag(object);
915 :
916 1 : if (s->flags & SLAB_RED_ZONE)
917 0 : memset(p - s->red_left_pad, val, s->red_left_pad);
918 :
919 1 : if (s->flags & __OBJECT_POISON) {
920 0 : memset(p, POISON_FREE, s->object_size - 1);
921 0 : p[s->object_size - 1] = POISON_END;
922 : }
923 :
924 1 : if (s->flags & SLAB_RED_ZONE)
925 0 : memset(p + s->object_size, val, s->inuse - s->object_size);
926 1 : }
927 :
928 0 : static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
929 : void *from, void *to)
930 : {
931 0 : slab_fix(s, "Restoring %s 0x%p-0x%p=0x%x", message, from, to - 1, data);
932 0 : memset(from, data, to - from);
933 0 : }
934 :
935 0 : static int check_bytes_and_report(struct kmem_cache *s, struct slab *slab,
936 : u8 *object, char *what,
937 : u8 *start, unsigned int value, unsigned int bytes)
938 : {
939 : u8 *fault;
940 : u8 *end;
941 0 : u8 *addr = slab_address(slab);
942 :
943 : metadata_access_enable();
944 0 : fault = memchr_inv(kasan_reset_tag(start), value, bytes);
945 : metadata_access_disable();
946 0 : if (!fault)
947 : return 1;
948 :
949 0 : end = start + bytes;
950 0 : while (end > fault && end[-1] == value)
951 0 : end--;
952 :
953 0 : if (slab_add_kunit_errors())
954 : goto skip_bug_print;
955 :
956 0 : slab_bug(s, "%s overwritten", what);
957 0 : pr_err("0x%p-0x%p @offset=%tu. First byte 0x%x instead of 0x%x\n",
958 : fault, end - 1, fault - addr,
959 : fault[0], value);
960 0 : print_trailer(s, slab, object);
961 0 : add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
962 :
963 : skip_bug_print:
964 0 : restore_bytes(s, what, value, fault, end);
965 0 : return 0;
966 : }
967 :
968 : /*
969 : * Object layout:
970 : *
971 : * object address
972 : * Bytes of the object to be managed.
973 : * If the freepointer may overlay the object then the free
974 : * pointer is at the middle of the object.
975 : *
976 : * Poisoning uses 0x6b (POISON_FREE) and the last byte is
977 : * 0xa5 (POISON_END)
978 : *
979 : * object + s->object_size
980 : * Padding to reach word boundary. This is also used for Redzoning.
981 : * Padding is extended by another word if Redzoning is enabled and
982 : * object_size == inuse.
983 : *
984 : * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
985 : * 0xcc (RED_ACTIVE) for objects in use.
986 : *
987 : * object + s->inuse
988 : * Meta data starts here.
989 : *
990 : * A. Free pointer (if we cannot overwrite object on free)
991 : * B. Tracking data for SLAB_STORE_USER
992 : * C. Padding to reach required alignment boundary or at minimum
993 : * one word if debugging is on to be able to detect writes
994 : * before the word boundary.
995 : *
996 : * Padding is done using 0x5a (POISON_INUSE)
997 : *
998 : * object + s->size
999 : * Nothing is used beyond s->size.
1000 : *
1001 : * If slabcaches are merged then the object_size and inuse boundaries are mostly
1002 : * ignored. And therefore no slab options that rely on these boundaries
1003 : * may be used with merged slabcaches.
1004 : */
1005 :
1006 0 : static int check_pad_bytes(struct kmem_cache *s, struct slab *slab, u8 *p)
1007 : {
1008 0 : unsigned long off = get_info_end(s); /* The end of info */
1009 :
1010 0 : if (s->flags & SLAB_STORE_USER)
1011 : /* We also have user information there */
1012 0 : off += 2 * sizeof(struct track);
1013 :
1014 0 : off += kasan_metadata_size(s);
1015 :
1016 0 : if (size_from_object(s) == off)
1017 : return 1;
1018 :
1019 0 : return check_bytes_and_report(s, slab, p, "Object padding",
1020 0 : p + off, POISON_INUSE, size_from_object(s) - off);
1021 : }
1022 :
1023 : /* Check the pad bytes at the end of a slab page */
1024 0 : static int slab_pad_check(struct kmem_cache *s, struct slab *slab)
1025 : {
1026 : u8 *start;
1027 : u8 *fault;
1028 : u8 *end;
1029 : u8 *pad;
1030 : int length;
1031 : int remainder;
1032 :
1033 0 : if (!(s->flags & SLAB_POISON))
1034 : return 1;
1035 :
1036 0 : start = slab_address(slab);
1037 0 : length = slab_size(slab);
1038 0 : end = start + length;
1039 0 : remainder = length % s->size;
1040 0 : if (!remainder)
1041 : return 1;
1042 :
1043 0 : pad = end - remainder;
1044 : metadata_access_enable();
1045 0 : fault = memchr_inv(kasan_reset_tag(pad), POISON_INUSE, remainder);
1046 : metadata_access_disable();
1047 0 : if (!fault)
1048 : return 1;
1049 0 : while (end > fault && end[-1] == POISON_INUSE)
1050 0 : end--;
1051 :
1052 0 : slab_err(s, slab, "Padding overwritten. 0x%p-0x%p @offset=%tu",
1053 : fault, end - 1, fault - start);
1054 0 : print_section(KERN_ERR, "Padding ", pad, remainder);
1055 :
1056 0 : restore_bytes(s, "slab padding", POISON_INUSE, fault, end);
1057 0 : return 0;
1058 : }
1059 :
1060 0 : static int check_object(struct kmem_cache *s, struct slab *slab,
1061 : void *object, u8 val)
1062 : {
1063 0 : u8 *p = object;
1064 0 : u8 *endobject = object + s->object_size;
1065 :
1066 0 : if (s->flags & SLAB_RED_ZONE) {
1067 0 : if (!check_bytes_and_report(s, slab, object, "Left Redzone",
1068 0 : object - s->red_left_pad, val, s->red_left_pad))
1069 : return 0;
1070 :
1071 0 : if (!check_bytes_and_report(s, slab, object, "Right Redzone",
1072 0 : endobject, val, s->inuse - s->object_size))
1073 : return 0;
1074 : } else {
1075 0 : if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
1076 0 : check_bytes_and_report(s, slab, p, "Alignment padding",
1077 : endobject, POISON_INUSE,
1078 : s->inuse - s->object_size);
1079 : }
1080 : }
1081 :
1082 0 : if (s->flags & SLAB_POISON) {
1083 0 : if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
1084 0 : (!check_bytes_and_report(s, slab, p, "Poison", p,
1085 0 : POISON_FREE, s->object_size - 1) ||
1086 0 : !check_bytes_and_report(s, slab, p, "End Poison",
1087 0 : p + s->object_size - 1, POISON_END, 1)))
1088 : return 0;
1089 : /*
1090 : * check_pad_bytes cleans up on its own.
1091 : */
1092 0 : check_pad_bytes(s, slab, p);
1093 : }
1094 :
1095 0 : if (!freeptr_outside_object(s) && val == SLUB_RED_ACTIVE)
1096 : /*
1097 : * Object and freepointer overlap. Cannot check
1098 : * freepointer while object is allocated.
1099 : */
1100 : return 1;
1101 :
1102 : /* Check free pointer validity */
1103 0 : if (!check_valid_pointer(s, slab, get_freepointer(s, p))) {
1104 0 : object_err(s, slab, p, "Freepointer corrupt");
1105 : /*
1106 : * No choice but to zap it and thus lose the remainder
1107 : * of the free objects in this slab. May cause
1108 : * another error because the object count is now wrong.
1109 : */
1110 0 : set_freepointer(s, p, NULL);
1111 0 : return 0;
1112 : }
1113 : return 1;
1114 : }
1115 :
1116 0 : static int check_slab(struct kmem_cache *s, struct slab *slab)
1117 : {
1118 : int maxobj;
1119 :
1120 0 : if (!folio_test_slab(slab_folio(slab))) {
1121 0 : slab_err(s, slab, "Not a valid slab page");
1122 0 : return 0;
1123 : }
1124 :
1125 0 : maxobj = order_objects(slab_order(slab), s->size);
1126 0 : if (slab->objects > maxobj) {
1127 0 : slab_err(s, slab, "objects %u > max %u",
1128 : slab->objects, maxobj);
1129 0 : return 0;
1130 : }
1131 0 : if (slab->inuse > slab->objects) {
1132 0 : slab_err(s, slab, "inuse %u > max %u",
1133 : slab->inuse, slab->objects);
1134 0 : return 0;
1135 : }
1136 : /* Slab_pad_check fixes things up after itself */
1137 0 : slab_pad_check(s, slab);
1138 0 : return 1;
1139 : }
1140 :
1141 : /*
1142 : * Determine if a certain object in a slab is on the freelist. Must hold the
1143 : * slab lock to guarantee that the chains are in a consistent state.
1144 : */
1145 0 : static int on_freelist(struct kmem_cache *s, struct slab *slab, void *search)
1146 : {
1147 0 : int nr = 0;
1148 : void *fp;
1149 0 : void *object = NULL;
1150 : int max_objects;
1151 :
1152 0 : fp = slab->freelist;
1153 0 : while (fp && nr <= slab->objects) {
1154 0 : if (fp == search)
1155 : return 1;
1156 0 : if (!check_valid_pointer(s, slab, fp)) {
1157 0 : if (object) {
1158 0 : object_err(s, slab, object,
1159 : "Freechain corrupt");
1160 0 : set_freepointer(s, object, NULL);
1161 : } else {
1162 0 : slab_err(s, slab, "Freepointer corrupt");
1163 0 : slab->freelist = NULL;
1164 0 : slab->inuse = slab->objects;
1165 0 : slab_fix(s, "Freelist cleared");
1166 0 : return 0;
1167 : }
1168 : break;
1169 : }
1170 0 : object = fp;
1171 0 : fp = get_freepointer(s, object);
1172 0 : nr++;
1173 : }
1174 :
1175 0 : max_objects = order_objects(slab_order(slab), s->size);
1176 0 : if (max_objects > MAX_OBJS_PER_PAGE)
1177 0 : max_objects = MAX_OBJS_PER_PAGE;
1178 :
1179 0 : if (slab->objects != max_objects) {
1180 0 : slab_err(s, slab, "Wrong number of objects. Found %d but should be %d",
1181 : slab->objects, max_objects);
1182 0 : slab->objects = max_objects;
1183 0 : slab_fix(s, "Number of objects adjusted");
1184 : }
1185 0 : if (slab->inuse != slab->objects - nr) {
1186 0 : slab_err(s, slab, "Wrong object count. Counter is %d but counted were %d",
1187 : slab->inuse, slab->objects - nr);
1188 0 : slab->inuse = slab->objects - nr;
1189 0 : slab_fix(s, "Object count adjusted");
1190 : }
1191 0 : return search == NULL;
1192 : }
1193 :
1194 0 : static void trace(struct kmem_cache *s, struct slab *slab, void *object,
1195 : int alloc)
1196 : {
1197 0 : if (s->flags & SLAB_TRACE) {
1198 0 : pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
1199 : s->name,
1200 : alloc ? "alloc" : "free",
1201 : object, slab->inuse,
1202 : slab->freelist);
1203 :
1204 0 : if (!alloc)
1205 0 : print_section(KERN_INFO, "Object ", (void *)object,
1206 : s->object_size);
1207 :
1208 0 : dump_stack();
1209 : }
1210 0 : }
1211 :
1212 : /*
1213 : * Tracking of fully allocated slabs for debugging purposes.
1214 : */
1215 : static void add_full(struct kmem_cache *s,
1216 : struct kmem_cache_node *n, struct slab *slab)
1217 : {
1218 0 : if (!(s->flags & SLAB_STORE_USER))
1219 : return;
1220 :
1221 : lockdep_assert_held(&n->list_lock);
1222 0 : list_add(&slab->slab_list, &n->full);
1223 : }
1224 :
1225 : static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct slab *slab)
1226 : {
1227 49 : if (!(s->flags & SLAB_STORE_USER))
1228 : return;
1229 :
1230 : lockdep_assert_held(&n->list_lock);
1231 0 : list_del(&slab->slab_list);
1232 : }
1233 :
1234 : /* Tracking of the number of slabs for debugging purposes */
1235 : static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1236 : {
1237 0 : struct kmem_cache_node *n = get_node(s, node);
1238 :
1239 0 : return atomic_long_read(&n->nr_slabs);
1240 : }
1241 :
1242 : static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1243 : {
1244 0 : return atomic_long_read(&n->nr_slabs);
1245 : }
1246 :
1247 : static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1248 : {
1249 455 : struct kmem_cache_node *n = get_node(s, node);
1250 :
1251 : /*
1252 : * May be called early in order to allocate a slab for the
1253 : * kmem_cache_node structure. Solve the chicken-egg
1254 : * dilemma by deferring the increment of the count during
1255 : * bootstrap (see early_kmem_cache_node_alloc).
1256 : */
1257 455 : if (likely(n)) {
1258 908 : atomic_long_inc(&n->nr_slabs);
1259 454 : atomic_long_add(objects, &n->total_objects);
1260 : }
1261 : }
1262 : static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1263 : {
1264 3 : struct kmem_cache_node *n = get_node(s, node);
1265 :
1266 6 : atomic_long_dec(&n->nr_slabs);
1267 6 : atomic_long_sub(objects, &n->total_objects);
1268 : }
1269 :
1270 : /* Object debug checks for alloc/free paths */
1271 14186 : static void setup_object_debug(struct kmem_cache *s, struct slab *slab,
1272 : void *object)
1273 : {
1274 28372 : if (!kmem_cache_debug_flags(s, SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))
1275 : return;
1276 :
1277 0 : init_object(s, object, SLUB_RED_INACTIVE);
1278 0 : init_tracking(s, object);
1279 : }
1280 :
1281 : static
1282 454 : void setup_slab_debug(struct kmem_cache *s, struct slab *slab, void *addr)
1283 : {
1284 908 : if (!kmem_cache_debug_flags(s, SLAB_POISON))
1285 : return;
1286 :
1287 0 : metadata_access_enable();
1288 0 : memset(kasan_reset_tag(addr), POISON_INUSE, slab_size(slab));
1289 : metadata_access_disable();
1290 : }
1291 :
1292 0 : static inline int alloc_consistency_checks(struct kmem_cache *s,
1293 : struct slab *slab, void *object)
1294 : {
1295 0 : if (!check_slab(s, slab))
1296 : return 0;
1297 :
1298 0 : if (!check_valid_pointer(s, slab, object)) {
1299 0 : object_err(s, slab, object, "Freelist Pointer check fails");
1300 0 : return 0;
1301 : }
1302 :
1303 0 : if (!check_object(s, slab, object, SLUB_RED_INACTIVE))
1304 : return 0;
1305 :
1306 0 : return 1;
1307 : }
1308 :
1309 0 : static noinline int alloc_debug_processing(struct kmem_cache *s,
1310 : struct slab *slab,
1311 : void *object, unsigned long addr)
1312 : {
1313 0 : if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1314 0 : if (!alloc_consistency_checks(s, slab, object))
1315 : goto bad;
1316 : }
1317 :
1318 : /* Success perform special debug activities for allocs */
1319 0 : if (s->flags & SLAB_STORE_USER)
1320 0 : set_track(s, object, TRACK_ALLOC, addr);
1321 0 : trace(s, slab, object, 1);
1322 0 : init_object(s, object, SLUB_RED_ACTIVE);
1323 0 : return 1;
1324 :
1325 : bad:
1326 0 : if (folio_test_slab(slab_folio(slab))) {
1327 : /*
1328 : * If this is a slab page then lets do the best we can
1329 : * to avoid issues in the future. Marking all objects
1330 : * as used avoids touching the remaining objects.
1331 : */
1332 0 : slab_fix(s, "Marking all objects used");
1333 0 : slab->inuse = slab->objects;
1334 0 : slab->freelist = NULL;
1335 : }
1336 : return 0;
1337 : }
1338 :
1339 0 : static inline int free_consistency_checks(struct kmem_cache *s,
1340 : struct slab *slab, void *object, unsigned long addr)
1341 : {
1342 0 : if (!check_valid_pointer(s, slab, object)) {
1343 0 : slab_err(s, slab, "Invalid object pointer 0x%p", object);
1344 : return 0;
1345 : }
1346 :
1347 0 : if (on_freelist(s, slab, object)) {
1348 0 : object_err(s, slab, object, "Object already free");
1349 : return 0;
1350 : }
1351 :
1352 0 : if (!check_object(s, slab, object, SLUB_RED_ACTIVE))
1353 : return 0;
1354 :
1355 0 : if (unlikely(s != slab->slab_cache)) {
1356 0 : if (!folio_test_slab(slab_folio(slab))) {
1357 0 : slab_err(s, slab, "Attempt to free object(0x%p) outside of slab",
1358 : object);
1359 0 : } else if (!slab->slab_cache) {
1360 0 : pr_err("SLUB <none>: no slab for object 0x%p.\n",
1361 : object);
1362 0 : dump_stack();
1363 : } else
1364 0 : object_err(s, slab, object,
1365 : "page slab pointer corrupt.");
1366 : return 0;
1367 : }
1368 : return 1;
1369 : }
1370 :
1371 : /* Supports checking bulk free of a constructed freelist */
1372 0 : static noinline int free_debug_processing(
1373 : struct kmem_cache *s, struct slab *slab,
1374 : void *head, void *tail, int bulk_cnt,
1375 : unsigned long addr)
1376 : {
1377 0 : struct kmem_cache_node *n = get_node(s, slab_nid(slab));
1378 0 : void *object = head;
1379 0 : int cnt = 0;
1380 : unsigned long flags, flags2;
1381 0 : int ret = 0;
1382 :
1383 0 : spin_lock_irqsave(&n->list_lock, flags);
1384 0 : slab_lock(slab, &flags2);
1385 :
1386 0 : if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1387 0 : if (!check_slab(s, slab))
1388 : goto out;
1389 : }
1390 :
1391 : next_object:
1392 0 : cnt++;
1393 :
1394 0 : if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1395 0 : if (!free_consistency_checks(s, slab, object, addr))
1396 : goto out;
1397 : }
1398 :
1399 0 : if (s->flags & SLAB_STORE_USER)
1400 0 : set_track(s, object, TRACK_FREE, addr);
1401 0 : trace(s, slab, object, 0);
1402 : /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1403 0 : init_object(s, object, SLUB_RED_INACTIVE);
1404 :
1405 : /* Reached end of constructed freelist yet? */
1406 0 : if (object != tail) {
1407 0 : object = get_freepointer(s, object);
1408 0 : goto next_object;
1409 : }
1410 : ret = 1;
1411 :
1412 : out:
1413 0 : if (cnt != bulk_cnt)
1414 0 : slab_err(s, slab, "Bulk freelist count(%d) invalid(%d)\n",
1415 : bulk_cnt, cnt);
1416 :
1417 0 : slab_unlock(slab, &flags2);
1418 0 : spin_unlock_irqrestore(&n->list_lock, flags);
1419 0 : if (!ret)
1420 0 : slab_fix(s, "Object at 0x%p not freed", object);
1421 0 : return ret;
1422 : }
1423 :
1424 : /*
1425 : * Parse a block of slub_debug options. Blocks are delimited by ';'
1426 : *
1427 : * @str: start of block
1428 : * @flags: returns parsed flags, or DEBUG_DEFAULT_FLAGS if none specified
1429 : * @slabs: return start of list of slabs, or NULL when there's no list
1430 : * @init: assume this is initial parsing and not per-kmem-create parsing
1431 : *
1432 : * returns the start of next block if there's any, or NULL
1433 : */
1434 : static char *
1435 0 : parse_slub_debug_flags(char *str, slab_flags_t *flags, char **slabs, bool init)
1436 : {
1437 0 : bool higher_order_disable = false;
1438 :
1439 : /* Skip any completely empty blocks */
1440 0 : while (*str && *str == ';')
1441 0 : str++;
1442 :
1443 0 : if (*str == ',') {
1444 : /*
1445 : * No options but restriction on slabs. This means full
1446 : * debugging for slabs matching a pattern.
1447 : */
1448 0 : *flags = DEBUG_DEFAULT_FLAGS;
1449 0 : goto check_slabs;
1450 : }
1451 0 : *flags = 0;
1452 :
1453 : /* Determine which debug features should be switched on */
1454 0 : for (; *str && *str != ',' && *str != ';'; str++) {
1455 0 : switch (tolower(*str)) {
1456 : case '-':
1457 0 : *flags = 0;
1458 0 : break;
1459 : case 'f':
1460 0 : *flags |= SLAB_CONSISTENCY_CHECKS;
1461 0 : break;
1462 : case 'z':
1463 0 : *flags |= SLAB_RED_ZONE;
1464 0 : break;
1465 : case 'p':
1466 0 : *flags |= SLAB_POISON;
1467 0 : break;
1468 : case 'u':
1469 0 : *flags |= SLAB_STORE_USER;
1470 0 : break;
1471 : case 't':
1472 0 : *flags |= SLAB_TRACE;
1473 0 : break;
1474 : case 'a':
1475 : *flags |= SLAB_FAILSLAB;
1476 0 : break;
1477 : case 'o':
1478 : /*
1479 : * Avoid enabling debugging on caches if its minimum
1480 : * order would increase as a result.
1481 : */
1482 : higher_order_disable = true;
1483 : break;
1484 : default:
1485 0 : if (init)
1486 0 : pr_err("slub_debug option '%c' unknown. skipped\n", *str);
1487 : }
1488 : }
1489 : check_slabs:
1490 0 : if (*str == ',')
1491 0 : *slabs = ++str;
1492 : else
1493 0 : *slabs = NULL;
1494 :
1495 : /* Skip over the slab list */
1496 0 : while (*str && *str != ';')
1497 0 : str++;
1498 :
1499 : /* Skip any completely empty blocks */
1500 0 : while (*str && *str == ';')
1501 0 : str++;
1502 :
1503 0 : if (init && higher_order_disable)
1504 0 : disable_higher_order_debug = 1;
1505 :
1506 0 : if (*str)
1507 : return str;
1508 : else
1509 0 : return NULL;
1510 : }
1511 :
1512 0 : static int __init setup_slub_debug(char *str)
1513 : {
1514 : slab_flags_t flags;
1515 : slab_flags_t global_flags;
1516 : char *saved_str;
1517 : char *slab_list;
1518 0 : bool global_slub_debug_changed = false;
1519 0 : bool slab_list_specified = false;
1520 :
1521 0 : global_flags = DEBUG_DEFAULT_FLAGS;
1522 0 : if (*str++ != '=' || !*str)
1523 : /*
1524 : * No options specified. Switch on full debugging.
1525 : */
1526 : goto out;
1527 :
1528 : saved_str = str;
1529 0 : while (str) {
1530 0 : str = parse_slub_debug_flags(str, &flags, &slab_list, true);
1531 :
1532 0 : if (!slab_list) {
1533 0 : global_flags = flags;
1534 0 : global_slub_debug_changed = true;
1535 : } else {
1536 : slab_list_specified = true;
1537 : }
1538 : }
1539 :
1540 : /*
1541 : * For backwards compatibility, a single list of flags with list of
1542 : * slabs means debugging is only changed for those slabs, so the global
1543 : * slub_debug should be unchanged (0 or DEBUG_DEFAULT_FLAGS, depending
1544 : * on CONFIG_SLUB_DEBUG_ON). We can extended that to multiple lists as
1545 : * long as there is no option specifying flags without a slab list.
1546 : */
1547 0 : if (slab_list_specified) {
1548 0 : if (!global_slub_debug_changed)
1549 0 : global_flags = slub_debug;
1550 0 : slub_debug_string = saved_str;
1551 : }
1552 : out:
1553 0 : slub_debug = global_flags;
1554 0 : if (slub_debug != 0 || slub_debug_string)
1555 0 : static_branch_enable(&slub_debug_enabled);
1556 : else
1557 0 : static_branch_disable(&slub_debug_enabled);
1558 0 : if ((static_branch_unlikely(&init_on_alloc) ||
1559 0 : static_branch_unlikely(&init_on_free)) &&
1560 0 : (slub_debug & SLAB_POISON))
1561 0 : pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n");
1562 0 : return 1;
1563 : }
1564 :
1565 : __setup("slub_debug", setup_slub_debug);
1566 :
1567 : /*
1568 : * kmem_cache_flags - apply debugging options to the cache
1569 : * @object_size: the size of an object without meta data
1570 : * @flags: flags to set
1571 : * @name: name of the cache
1572 : *
1573 : * Debug option(s) are applied to @flags. In addition to the debug
1574 : * option(s), if a slab name (or multiple) is specified i.e.
1575 : * slub_debug=<Debug-Options>,<slab name1>,<slab name2> ...
1576 : * then only the select slabs will receive the debug option(s).
1577 : */
1578 114 : slab_flags_t kmem_cache_flags(unsigned int object_size,
1579 : slab_flags_t flags, const char *name)
1580 : {
1581 : char *iter;
1582 : size_t len;
1583 : char *next_block;
1584 : slab_flags_t block_flags;
1585 114 : slab_flags_t slub_debug_local = slub_debug;
1586 :
1587 : /*
1588 : * If the slab cache is for debugging (e.g. kmemleak) then
1589 : * don't store user (stack trace) information by default,
1590 : * but let the user enable it via the command line below.
1591 : */
1592 114 : if (flags & SLAB_NOLEAKTRACE)
1593 0 : slub_debug_local &= ~SLAB_STORE_USER;
1594 :
1595 114 : len = strlen(name);
1596 114 : next_block = slub_debug_string;
1597 : /* Go through all blocks of debug options, see if any matches our slab's name */
1598 228 : while (next_block) {
1599 0 : next_block = parse_slub_debug_flags(next_block, &block_flags, &iter, false);
1600 0 : if (!iter)
1601 0 : continue;
1602 : /* Found a block that has a slab list, search it */
1603 0 : while (*iter) {
1604 : char *end, *glob;
1605 : size_t cmplen;
1606 :
1607 0 : end = strchrnul(iter, ',');
1608 0 : if (next_block && next_block < end)
1609 0 : end = next_block - 1;
1610 :
1611 0 : glob = strnchr(iter, end - iter, '*');
1612 0 : if (glob)
1613 0 : cmplen = glob - iter;
1614 : else
1615 0 : cmplen = max_t(size_t, len, (end - iter));
1616 :
1617 0 : if (!strncmp(name, iter, cmplen)) {
1618 0 : flags |= block_flags;
1619 0 : return flags;
1620 : }
1621 :
1622 0 : if (!*end || *end == ';')
1623 : break;
1624 0 : iter = end + 1;
1625 : }
1626 : }
1627 :
1628 114 : return flags | slub_debug_local;
1629 : }
1630 : #else /* !CONFIG_SLUB_DEBUG */
1631 : static inline void setup_object_debug(struct kmem_cache *s,
1632 : struct slab *slab, void *object) {}
1633 : static inline
1634 : void setup_slab_debug(struct kmem_cache *s, struct slab *slab, void *addr) {}
1635 :
1636 : static inline int alloc_debug_processing(struct kmem_cache *s,
1637 : struct slab *slab, void *object, unsigned long addr) { return 0; }
1638 :
1639 : static inline int free_debug_processing(
1640 : struct kmem_cache *s, struct slab *slab,
1641 : void *head, void *tail, int bulk_cnt,
1642 : unsigned long addr) { return 0; }
1643 :
1644 : static inline int slab_pad_check(struct kmem_cache *s, struct slab *slab)
1645 : { return 1; }
1646 : static inline int check_object(struct kmem_cache *s, struct slab *slab,
1647 : void *object, u8 val) { return 1; }
1648 : static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1649 : struct slab *slab) {}
1650 : static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1651 : struct slab *slab) {}
1652 : slab_flags_t kmem_cache_flags(unsigned int object_size,
1653 : slab_flags_t flags, const char *name)
1654 : {
1655 : return flags;
1656 : }
1657 : #define slub_debug 0
1658 :
1659 : #define disable_higher_order_debug 0
1660 :
1661 : static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1662 : { return 0; }
1663 : static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1664 : { return 0; }
1665 : static inline void inc_slabs_node(struct kmem_cache *s, int node,
1666 : int objects) {}
1667 : static inline void dec_slabs_node(struct kmem_cache *s, int node,
1668 : int objects) {}
1669 :
1670 : static bool freelist_corrupted(struct kmem_cache *s, struct slab *slab,
1671 : void **freelist, void *nextfree)
1672 : {
1673 : return false;
1674 : }
1675 : #endif /* CONFIG_SLUB_DEBUG */
1676 :
1677 : /*
1678 : * Hooks for other subsystems that check memory allocations. In a typical
1679 : * production configuration these hooks all should produce no code at all.
1680 : */
1681 : static inline void *kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1682 : {
1683 : ptr = kasan_kmalloc_large(ptr, size, flags);
1684 : /* As ptr might get tagged, call kmemleak hook after KASAN. */
1685 : kmemleak_alloc(ptr, size, 1, flags);
1686 : return ptr;
1687 : }
1688 :
1689 : static __always_inline void kfree_hook(void *x)
1690 : {
1691 8 : kmemleak_free(x);
1692 8 : kasan_kfree_large(x);
1693 : }
1694 :
1695 : static __always_inline bool slab_free_hook(struct kmem_cache *s,
1696 : void *x, bool init)
1697 : {
1698 5311 : kmemleak_free_recursive(x, s->flags);
1699 :
1700 5311 : debug_check_no_locks_freed(x, s->object_size);
1701 :
1702 : if (!(s->flags & SLAB_DEBUG_OBJECTS))
1703 5311 : debug_check_no_obj_freed(x, s->object_size);
1704 :
1705 : /* Use KCSAN to help debug racy use-after-free. */
1706 : if (!(s->flags & SLAB_TYPESAFE_BY_RCU))
1707 : __kcsan_check_access(x, s->object_size,
1708 : KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT);
1709 :
1710 : /*
1711 : * As memory initialization might be integrated into KASAN,
1712 : * kasan_slab_free and initialization memset's must be
1713 : * kept together to avoid discrepancies in behavior.
1714 : *
1715 : * The initialization memset's clear the object and the metadata,
1716 : * but don't touch the SLAB redzone.
1717 : */
1718 5311 : if (init) {
1719 : int rsize;
1720 :
1721 : if (!kasan_has_integrated_init())
1722 0 : memset(kasan_reset_tag(x), 0, s->object_size);
1723 0 : rsize = (s->flags & SLAB_RED_ZONE) ? s->red_left_pad : 0;
1724 0 : memset((char *)kasan_reset_tag(x) + s->inuse, 0,
1725 0 : s->size - s->inuse - rsize);
1726 : }
1727 : /* KASAN might put x into memory quarantine, delaying its reuse. */
1728 5311 : return kasan_slab_free(s, x, init);
1729 : }
1730 :
1731 5311 : static inline bool slab_free_freelist_hook(struct kmem_cache *s,
1732 : void **head, void **tail,
1733 : int *cnt)
1734 : {
1735 :
1736 : void *object;
1737 5311 : void *next = *head;
1738 5311 : void *old_tail = *tail ? *tail : *head;
1739 :
1740 5311 : if (is_kfence_address(next)) {
1741 : slab_free_hook(s, next, false);
1742 : return true;
1743 : }
1744 :
1745 : /* Head and tail of the reconstructed freelist */
1746 5311 : *head = NULL;
1747 5311 : *tail = NULL;
1748 :
1749 : do {
1750 5311 : object = next;
1751 10622 : next = get_freepointer(s, object);
1752 :
1753 : /* If object's reuse doesn't have to be delayed */
1754 15933 : if (!slab_free_hook(s, object, slab_want_init_on_free(s))) {
1755 : /* Move object to the new freelist */
1756 10622 : set_freepointer(s, object, *head);
1757 5311 : *head = object;
1758 5311 : if (!*tail)
1759 5311 : *tail = object;
1760 : } else {
1761 : /*
1762 : * Adjust the reconstructed freelist depth
1763 : * accordingly if object's reuse is delayed.
1764 : */
1765 : --(*cnt);
1766 : }
1767 5311 : } while (object != old_tail);
1768 :
1769 5311 : if (*head == *tail)
1770 5311 : *tail = NULL;
1771 :
1772 5311 : return *head != NULL;
1773 : }
1774 :
1775 : static void *setup_object(struct kmem_cache *s, struct slab *slab,
1776 : void *object)
1777 : {
1778 14186 : setup_object_debug(s, slab, object);
1779 14186 : object = kasan_init_slab_obj(s, object);
1780 14186 : if (unlikely(s->ctor)) {
1781 287 : kasan_unpoison_object_data(s, object);
1782 287 : s->ctor(object);
1783 287 : kasan_poison_object_data(s, object);
1784 : }
1785 : return object;
1786 : }
1787 :
1788 : /*
1789 : * Slab allocation and freeing
1790 : */
1791 454 : static inline struct slab *alloc_slab_page(gfp_t flags, int node,
1792 : struct kmem_cache_order_objects oo)
1793 : {
1794 : struct folio *folio;
1795 : struct slab *slab;
1796 454 : unsigned int order = oo_order(oo);
1797 :
1798 454 : if (node == NUMA_NO_NODE)
1799 453 : folio = (struct folio *)alloc_pages(flags, order);
1800 : else
1801 1 : folio = (struct folio *)__alloc_pages_node(node, flags, order);
1802 :
1803 454 : if (!folio)
1804 : return NULL;
1805 :
1806 454 : slab = folio_slab(folio);
1807 454 : __folio_set_slab(folio);
1808 908 : if (page_is_pfmemalloc(folio_page(folio, 0)))
1809 : slab_set_pfmemalloc(slab);
1810 :
1811 : return slab;
1812 : }
1813 :
1814 : #ifdef CONFIG_SLAB_FREELIST_RANDOM
1815 : /* Pre-initialize the random sequence cache */
1816 : static int init_cache_random_seq(struct kmem_cache *s)
1817 : {
1818 : unsigned int count = oo_objects(s->oo);
1819 : int err;
1820 :
1821 : /* Bailout if already initialised */
1822 : if (s->random_seq)
1823 : return 0;
1824 :
1825 : err = cache_random_seq_create(s, count, GFP_KERNEL);
1826 : if (err) {
1827 : pr_err("SLUB: Unable to initialize free list for %s\n",
1828 : s->name);
1829 : return err;
1830 : }
1831 :
1832 : /* Transform to an offset on the set of pages */
1833 : if (s->random_seq) {
1834 : unsigned int i;
1835 :
1836 : for (i = 0; i < count; i++)
1837 : s->random_seq[i] *= s->size;
1838 : }
1839 : return 0;
1840 : }
1841 :
1842 : /* Initialize each random sequence freelist per cache */
1843 : static void __init init_freelist_randomization(void)
1844 : {
1845 : struct kmem_cache *s;
1846 :
1847 : mutex_lock(&slab_mutex);
1848 :
1849 : list_for_each_entry(s, &slab_caches, list)
1850 : init_cache_random_seq(s);
1851 :
1852 : mutex_unlock(&slab_mutex);
1853 : }
1854 :
1855 : /* Get the next entry on the pre-computed freelist randomized */
1856 : static void *next_freelist_entry(struct kmem_cache *s, struct slab *slab,
1857 : unsigned long *pos, void *start,
1858 : unsigned long page_limit,
1859 : unsigned long freelist_count)
1860 : {
1861 : unsigned int idx;
1862 :
1863 : /*
1864 : * If the target page allocation failed, the number of objects on the
1865 : * page might be smaller than the usual size defined by the cache.
1866 : */
1867 : do {
1868 : idx = s->random_seq[*pos];
1869 : *pos += 1;
1870 : if (*pos >= freelist_count)
1871 : *pos = 0;
1872 : } while (unlikely(idx >= page_limit));
1873 :
1874 : return (char *)start + idx;
1875 : }
1876 :
1877 : /* Shuffle the single linked freelist based on a random pre-computed sequence */
1878 : static bool shuffle_freelist(struct kmem_cache *s, struct slab *slab)
1879 : {
1880 : void *start;
1881 : void *cur;
1882 : void *next;
1883 : unsigned long idx, pos, page_limit, freelist_count;
1884 :
1885 : if (slab->objects < 2 || !s->random_seq)
1886 : return false;
1887 :
1888 : freelist_count = oo_objects(s->oo);
1889 : pos = get_random_int() % freelist_count;
1890 :
1891 : page_limit = slab->objects * s->size;
1892 : start = fixup_red_left(s, slab_address(slab));
1893 :
1894 : /* First entry is used as the base of the freelist */
1895 : cur = next_freelist_entry(s, slab, &pos, start, page_limit,
1896 : freelist_count);
1897 : cur = setup_object(s, slab, cur);
1898 : slab->freelist = cur;
1899 :
1900 : for (idx = 1; idx < slab->objects; idx++) {
1901 : next = next_freelist_entry(s, slab, &pos, start, page_limit,
1902 : freelist_count);
1903 : next = setup_object(s, slab, next);
1904 : set_freepointer(s, cur, next);
1905 : cur = next;
1906 : }
1907 : set_freepointer(s, cur, NULL);
1908 :
1909 : return true;
1910 : }
1911 : #else
1912 : static inline int init_cache_random_seq(struct kmem_cache *s)
1913 : {
1914 : return 0;
1915 : }
1916 : static inline void init_freelist_randomization(void) { }
1917 : static inline bool shuffle_freelist(struct kmem_cache *s, struct slab *slab)
1918 : {
1919 : return false;
1920 : }
1921 : #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1922 :
1923 454 : static struct slab *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1924 : {
1925 : struct slab *slab;
1926 454 : struct kmem_cache_order_objects oo = s->oo;
1927 : gfp_t alloc_gfp;
1928 : void *start, *p, *next;
1929 : int idx;
1930 : bool shuffle;
1931 :
1932 454 : flags &= gfp_allowed_mask;
1933 :
1934 454 : flags |= s->allocflags;
1935 :
1936 : /*
1937 : * Let the initial higher-order allocation fail under memory pressure
1938 : * so we fall-back to the minimum order allocation.
1939 : */
1940 454 : alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1941 864 : if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1942 76 : alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
1943 :
1944 454 : slab = alloc_slab_page(alloc_gfp, node, oo);
1945 454 : if (unlikely(!slab)) {
1946 0 : oo = s->min;
1947 0 : alloc_gfp = flags;
1948 : /*
1949 : * Allocation may have failed due to fragmentation.
1950 : * Try a lower order alloc if possible
1951 : */
1952 0 : slab = alloc_slab_page(alloc_gfp, node, oo);
1953 0 : if (unlikely(!slab))
1954 : goto out;
1955 : stat(s, ORDER_FALLBACK);
1956 : }
1957 :
1958 454 : slab->objects = oo_objects(oo);
1959 :
1960 908 : account_slab(slab, oo_order(oo), s, flags);
1961 :
1962 454 : slab->slab_cache = s;
1963 :
1964 454 : kasan_poison_slab(slab);
1965 :
1966 454 : start = slab_address(slab);
1967 :
1968 454 : setup_slab_debug(s, slab, start);
1969 :
1970 454 : shuffle = shuffle_freelist(s, slab);
1971 :
1972 : if (!shuffle) {
1973 454 : start = fixup_red_left(s, start);
1974 908 : start = setup_object(s, slab, start);
1975 454 : slab->freelist = start;
1976 14186 : for (idx = 0, p = start; idx < slab->objects - 1; idx++) {
1977 13732 : next = p + s->size;
1978 27464 : next = setup_object(s, slab, next);
1979 27464 : set_freepointer(s, p, next);
1980 13732 : p = next;
1981 : }
1982 454 : set_freepointer(s, p, NULL);
1983 : }
1984 :
1985 454 : slab->inuse = slab->objects;
1986 454 : slab->frozen = 1;
1987 :
1988 : out:
1989 454 : if (!slab)
1990 : return NULL;
1991 :
1992 908 : inc_slabs_node(s, slab_nid(slab), slab->objects);
1993 :
1994 : return slab;
1995 : }
1996 :
1997 454 : static struct slab *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1998 : {
1999 454 : if (unlikely(flags & GFP_SLAB_BUG_MASK))
2000 0 : flags = kmalloc_fix_flags(flags);
2001 :
2002 454 : WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO));
2003 :
2004 454 : return allocate_slab(s,
2005 : flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
2006 : }
2007 :
2008 3 : static void __free_slab(struct kmem_cache *s, struct slab *slab)
2009 : {
2010 3 : struct folio *folio = slab_folio(slab);
2011 3 : int order = folio_order(folio);
2012 3 : int pages = 1 << order;
2013 :
2014 6 : if (kmem_cache_debug_flags(s, SLAB_CONSISTENCY_CHECKS)) {
2015 : void *p;
2016 :
2017 0 : slab_pad_check(s, slab);
2018 0 : for_each_object(p, s, slab_address(slab), slab->objects)
2019 0 : check_object(s, slab, p, SLUB_RED_INACTIVE);
2020 : }
2021 :
2022 3 : __slab_clear_pfmemalloc(slab);
2023 3 : __folio_clear_slab(folio);
2024 3 : folio->mapping = NULL;
2025 3 : if (current->reclaim_state)
2026 0 : current->reclaim_state->reclaimed_slab += pages;
2027 3 : unaccount_slab(slab, order, s);
2028 3 : __free_pages(folio_page(folio, 0), order);
2029 3 : }
2030 :
2031 0 : static void rcu_free_slab(struct rcu_head *h)
2032 : {
2033 0 : struct slab *slab = container_of(h, struct slab, rcu_head);
2034 :
2035 0 : __free_slab(slab->slab_cache, slab);
2036 0 : }
2037 :
2038 3 : static void free_slab(struct kmem_cache *s, struct slab *slab)
2039 : {
2040 3 : if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU)) {
2041 0 : call_rcu(&slab->rcu_head, rcu_free_slab);
2042 : } else
2043 3 : __free_slab(s, slab);
2044 3 : }
2045 :
2046 : static void discard_slab(struct kmem_cache *s, struct slab *slab)
2047 : {
2048 9 : dec_slabs_node(s, slab_nid(slab), slab->objects);
2049 3 : free_slab(s, slab);
2050 : }
2051 :
2052 : /*
2053 : * Management of partially allocated slabs.
2054 : */
2055 : static inline void
2056 : __add_partial(struct kmem_cache_node *n, struct slab *slab, int tail)
2057 : {
2058 52 : n->nr_partial++;
2059 2 : if (tail == DEACTIVATE_TO_TAIL)
2060 49 : list_add_tail(&slab->slab_list, &n->partial);
2061 : else
2062 3 : list_add(&slab->slab_list, &n->partial);
2063 : }
2064 :
2065 : static inline void add_partial(struct kmem_cache_node *n,
2066 : struct slab *slab, int tail)
2067 : {
2068 : lockdep_assert_held(&n->list_lock);
2069 2 : __add_partial(n, slab, tail);
2070 : }
2071 :
2072 : static inline void remove_partial(struct kmem_cache_node *n,
2073 : struct slab *slab)
2074 : {
2075 : lockdep_assert_held(&n->list_lock);
2076 98 : list_del(&slab->slab_list);
2077 49 : n->nr_partial--;
2078 : }
2079 :
2080 : /*
2081 : * Remove slab from the partial list, freeze it and
2082 : * return the pointer to the freelist.
2083 : *
2084 : * Returns a list of objects or NULL if it fails.
2085 : */
2086 46 : static inline void *acquire_slab(struct kmem_cache *s,
2087 : struct kmem_cache_node *n, struct slab *slab,
2088 : int mode)
2089 : {
2090 : void *freelist;
2091 : unsigned long counters;
2092 : struct slab new;
2093 :
2094 : lockdep_assert_held(&n->list_lock);
2095 :
2096 : /*
2097 : * Zap the freelist and set the frozen bit.
2098 : * The old freelist is the list of objects for the
2099 : * per cpu allocation list.
2100 : */
2101 46 : freelist = slab->freelist;
2102 46 : counters = slab->counters;
2103 46 : new.counters = counters;
2104 46 : if (mode) {
2105 46 : new.inuse = slab->objects;
2106 46 : new.freelist = NULL;
2107 : } else {
2108 : new.freelist = freelist;
2109 : }
2110 :
2111 : VM_BUG_ON(new.frozen);
2112 46 : new.frozen = 1;
2113 :
2114 92 : if (!__cmpxchg_double_slab(s, slab,
2115 : freelist, counters,
2116 : new.freelist, new.counters,
2117 : "acquire_slab"))
2118 : return NULL;
2119 :
2120 92 : remove_partial(n, slab);
2121 46 : WARN_ON(!freelist);
2122 : return freelist;
2123 : }
2124 :
2125 : #ifdef CONFIG_SLUB_CPU_PARTIAL
2126 : static void put_cpu_partial(struct kmem_cache *s, struct slab *slab, int drain);
2127 : #else
2128 : static inline void put_cpu_partial(struct kmem_cache *s, struct slab *slab,
2129 : int drain) { }
2130 : #endif
2131 : static inline bool pfmemalloc_match(struct slab *slab, gfp_t gfpflags);
2132 :
2133 : /*
2134 : * Try to allocate a partial slab from a specific node.
2135 : */
2136 499 : static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
2137 : struct slab **ret_slab, gfp_t gfpflags)
2138 : {
2139 : struct slab *slab, *slab2;
2140 499 : void *object = NULL;
2141 : unsigned long flags;
2142 499 : unsigned int partial_slabs = 0;
2143 :
2144 : /*
2145 : * Racy check. If we mistakenly see no partial slabs then we
2146 : * just allocate an empty slab. If we mistakenly try to get a
2147 : * partial slab and there is none available then get_partial()
2148 : * will return NULL.
2149 : */
2150 499 : if (!n || !n->nr_partial)
2151 : return NULL;
2152 :
2153 46 : spin_lock_irqsave(&n->list_lock, flags);
2154 92 : list_for_each_entry_safe(slab, slab2, &n->partial, slab_list) {
2155 : void *t;
2156 :
2157 46 : if (!pfmemalloc_match(slab, gfpflags))
2158 0 : continue;
2159 :
2160 46 : t = acquire_slab(s, n, slab, object == NULL);
2161 46 : if (!t)
2162 : break;
2163 :
2164 : if (!object) {
2165 46 : *ret_slab = slab;
2166 46 : stat(s, ALLOC_FROM_PARTIAL);
2167 46 : object = t;
2168 : } else {
2169 : put_cpu_partial(s, slab, 0);
2170 : stat(s, CPU_PARTIAL_NODE);
2171 : partial_slabs++;
2172 : }
2173 : #ifdef CONFIG_SLUB_CPU_PARTIAL
2174 : if (!kmem_cache_has_cpu_partial(s)
2175 : || partial_slabs > s->cpu_partial_slabs / 2)
2176 : break;
2177 : #else
2178 : break;
2179 : #endif
2180 :
2181 : }
2182 92 : spin_unlock_irqrestore(&n->list_lock, flags);
2183 : return object;
2184 : }
2185 :
2186 : /*
2187 : * Get a slab from somewhere. Search in increasing NUMA distances.
2188 : */
2189 : static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
2190 : struct slab **ret_slab)
2191 : {
2192 : #ifdef CONFIG_NUMA
2193 : struct zonelist *zonelist;
2194 : struct zoneref *z;
2195 : struct zone *zone;
2196 : enum zone_type highest_zoneidx = gfp_zone(flags);
2197 : void *object;
2198 : unsigned int cpuset_mems_cookie;
2199 :
2200 : /*
2201 : * The defrag ratio allows a configuration of the tradeoffs between
2202 : * inter node defragmentation and node local allocations. A lower
2203 : * defrag_ratio increases the tendency to do local allocations
2204 : * instead of attempting to obtain partial slabs from other nodes.
2205 : *
2206 : * If the defrag_ratio is set to 0 then kmalloc() always
2207 : * returns node local objects. If the ratio is higher then kmalloc()
2208 : * may return off node objects because partial slabs are obtained
2209 : * from other nodes and filled up.
2210 : *
2211 : * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
2212 : * (which makes defrag_ratio = 1000) then every (well almost)
2213 : * allocation will first attempt to defrag slab caches on other nodes.
2214 : * This means scanning over all nodes to look for partial slabs which
2215 : * may be expensive if we do it every time we are trying to find a slab
2216 : * with available objects.
2217 : */
2218 : if (!s->remote_node_defrag_ratio ||
2219 : get_cycles() % 1024 > s->remote_node_defrag_ratio)
2220 : return NULL;
2221 :
2222 : do {
2223 : cpuset_mems_cookie = read_mems_allowed_begin();
2224 : zonelist = node_zonelist(mempolicy_slab_node(), flags);
2225 : for_each_zone_zonelist(zone, z, zonelist, highest_zoneidx) {
2226 : struct kmem_cache_node *n;
2227 :
2228 : n = get_node(s, zone_to_nid(zone));
2229 :
2230 : if (n && cpuset_zone_allowed(zone, flags) &&
2231 : n->nr_partial > s->min_partial) {
2232 : object = get_partial_node(s, n, ret_slab, flags);
2233 : if (object) {
2234 : /*
2235 : * Don't check read_mems_allowed_retry()
2236 : * here - if mems_allowed was updated in
2237 : * parallel, that was a harmless race
2238 : * between allocation and the cpuset
2239 : * update
2240 : */
2241 : return object;
2242 : }
2243 : }
2244 : }
2245 : } while (read_mems_allowed_retry(cpuset_mems_cookie));
2246 : #endif /* CONFIG_NUMA */
2247 : return NULL;
2248 : }
2249 :
2250 : /*
2251 : * Get a partial slab, lock it and return it.
2252 : */
2253 : static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
2254 : struct slab **ret_slab)
2255 : {
2256 : void *object;
2257 499 : int searchnode = node;
2258 :
2259 499 : if (node == NUMA_NO_NODE)
2260 499 : searchnode = numa_mem_id();
2261 :
2262 499 : object = get_partial_node(s, get_node(s, searchnode), ret_slab, flags);
2263 499 : if (object || node != NUMA_NO_NODE)
2264 : return object;
2265 :
2266 453 : return get_any_partial(s, flags, ret_slab);
2267 : }
2268 :
2269 : #ifdef CONFIG_PREEMPTION
2270 : /*
2271 : * Calculate the next globally unique transaction for disambiguation
2272 : * during cmpxchg. The transactions start with the cpu number and are then
2273 : * incremented by CONFIG_NR_CPUS.
2274 : */
2275 : #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
2276 : #else
2277 : /*
2278 : * No preemption supported therefore also no need to check for
2279 : * different cpus.
2280 : */
2281 : #define TID_STEP 1
2282 : #endif
2283 :
2284 : static inline unsigned long next_tid(unsigned long tid)
2285 : {
2286 22362 : return tid + TID_STEP;
2287 : }
2288 :
2289 : #ifdef SLUB_DEBUG_CMPXCHG
2290 : static inline unsigned int tid_to_cpu(unsigned long tid)
2291 : {
2292 : return tid % TID_STEP;
2293 : }
2294 :
2295 : static inline unsigned long tid_to_event(unsigned long tid)
2296 : {
2297 : return tid / TID_STEP;
2298 : }
2299 : #endif
2300 :
2301 : static inline unsigned int init_tid(int cpu)
2302 : {
2303 67 : return cpu;
2304 : }
2305 :
2306 : static inline void note_cmpxchg_failure(const char *n,
2307 : const struct kmem_cache *s, unsigned long tid)
2308 : {
2309 : #ifdef SLUB_DEBUG_CMPXCHG
2310 : unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
2311 :
2312 : pr_info("%s %s: cmpxchg redo ", n, s->name);
2313 :
2314 : #ifdef CONFIG_PREEMPTION
2315 : if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
2316 : pr_warn("due to cpu change %d -> %d\n",
2317 : tid_to_cpu(tid), tid_to_cpu(actual_tid));
2318 : else
2319 : #endif
2320 : if (tid_to_event(tid) != tid_to_event(actual_tid))
2321 : pr_warn("due to cpu running other code. Event %ld->%ld\n",
2322 : tid_to_event(tid), tid_to_event(actual_tid));
2323 : else
2324 : pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2325 : actual_tid, tid, next_tid(tid));
2326 : #endif
2327 : stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
2328 : }
2329 :
2330 : static void init_kmem_cache_cpus(struct kmem_cache *s)
2331 : {
2332 : int cpu;
2333 : struct kmem_cache_cpu *c;
2334 :
2335 67 : for_each_possible_cpu(cpu) {
2336 67 : c = per_cpu_ptr(s->cpu_slab, cpu);
2337 67 : local_lock_init(&c->lock);
2338 67 : c->tid = init_tid(cpu);
2339 : }
2340 : }
2341 :
2342 : /*
2343 : * Finishes removing the cpu slab. Merges cpu's freelist with slab's freelist,
2344 : * unfreezes the slabs and puts it on the proper list.
2345 : * Assumes the slab has been already safely taken away from kmem_cache_cpu
2346 : * by the caller.
2347 : */
2348 2 : static void deactivate_slab(struct kmem_cache *s, struct slab *slab,
2349 : void *freelist)
2350 : {
2351 : enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE, M_FULL_NOLIST };
2352 6 : struct kmem_cache_node *n = get_node(s, slab_nid(slab));
2353 2 : int free_delta = 0;
2354 2 : enum slab_modes mode = M_NONE;
2355 : void *nextfree, *freelist_iter, *freelist_tail;
2356 2 : int tail = DEACTIVATE_TO_HEAD;
2357 2 : unsigned long flags = 0;
2358 : struct slab new;
2359 : struct slab old;
2360 :
2361 2 : if (slab->freelist) {
2362 0 : stat(s, DEACTIVATE_REMOTE_FREES);
2363 0 : tail = DEACTIVATE_TO_TAIL;
2364 : }
2365 :
2366 : /*
2367 : * Stage one: Count the objects on cpu's freelist as free_delta and
2368 : * remember the last object in freelist_tail for later splicing.
2369 : */
2370 2 : freelist_tail = NULL;
2371 2 : freelist_iter = freelist;
2372 86 : while (freelist_iter) {
2373 164 : nextfree = get_freepointer(s, freelist_iter);
2374 :
2375 : /*
2376 : * If 'nextfree' is invalid, it is possible that the object at
2377 : * 'freelist_iter' is already corrupted. So isolate all objects
2378 : * starting at 'freelist_iter' by skipping them.
2379 : */
2380 82 : if (freelist_corrupted(s, slab, &freelist_iter, nextfree))
2381 : break;
2382 :
2383 82 : freelist_tail = freelist_iter;
2384 82 : free_delta++;
2385 :
2386 82 : freelist_iter = nextfree;
2387 : }
2388 :
2389 : /*
2390 : * Stage two: Unfreeze the slab while splicing the per-cpu
2391 : * freelist to the head of slab's freelist.
2392 : *
2393 : * Ensure that the slab is unfrozen while the list presence
2394 : * reflects the actual number of objects during unfreeze.
2395 : *
2396 : * We first perform cmpxchg holding lock and insert to list
2397 : * when it succeed. If there is mismatch then the slab is not
2398 : * unfrozen and number of objects in the slab may have changed.
2399 : * Then release lock and retry cmpxchg again.
2400 : */
2401 : redo:
2402 :
2403 2 : old.freelist = READ_ONCE(slab->freelist);
2404 2 : old.counters = READ_ONCE(slab->counters);
2405 : VM_BUG_ON(!old.frozen);
2406 :
2407 : /* Determine target state of the slab */
2408 2 : new.counters = old.counters;
2409 2 : if (freelist_tail) {
2410 2 : new.inuse -= free_delta;
2411 4 : set_freepointer(s, freelist_tail, old.freelist);
2412 2 : new.freelist = freelist;
2413 : } else
2414 : new.freelist = old.freelist;
2415 :
2416 2 : new.frozen = 0;
2417 :
2418 2 : if (!new.inuse && n->nr_partial >= s->min_partial) {
2419 : mode = M_FREE;
2420 2 : } else if (new.freelist) {
2421 2 : mode = M_PARTIAL;
2422 : /*
2423 : * Taking the spinlock removes the possibility that
2424 : * acquire_slab() will see a slab that is frozen
2425 : */
2426 2 : spin_lock_irqsave(&n->list_lock, flags);
2427 0 : } else if (kmem_cache_debug_flags(s, SLAB_STORE_USER)) {
2428 0 : mode = M_FULL;
2429 : /*
2430 : * This also ensures that the scanning of full
2431 : * slabs from diagnostic functions will not see
2432 : * any frozen slabs.
2433 : */
2434 0 : spin_lock_irqsave(&n->list_lock, flags);
2435 : } else {
2436 : mode = M_FULL_NOLIST;
2437 : }
2438 :
2439 :
2440 2 : if (!cmpxchg_double_slab(s, slab,
2441 : old.freelist, old.counters,
2442 : new.freelist, new.counters,
2443 : "unfreezing slab")) {
2444 0 : if (mode == M_PARTIAL || mode == M_FULL)
2445 0 : spin_unlock_irqrestore(&n->list_lock, flags);
2446 : goto redo;
2447 : }
2448 :
2449 :
2450 2 : if (mode == M_PARTIAL) {
2451 2 : add_partial(n, slab, tail);
2452 4 : spin_unlock_irqrestore(&n->list_lock, flags);
2453 2 : stat(s, tail);
2454 0 : } else if (mode == M_FREE) {
2455 0 : stat(s, DEACTIVATE_EMPTY);
2456 : discard_slab(s, slab);
2457 : stat(s, FREE_SLAB);
2458 0 : } else if (mode == M_FULL) {
2459 0 : add_full(s, n, slab);
2460 0 : spin_unlock_irqrestore(&n->list_lock, flags);
2461 : stat(s, DEACTIVATE_FULL);
2462 : } else if (mode == M_FULL_NOLIST) {
2463 : stat(s, DEACTIVATE_FULL);
2464 : }
2465 2 : }
2466 :
2467 : #ifdef CONFIG_SLUB_CPU_PARTIAL
2468 : static void __unfreeze_partials(struct kmem_cache *s, struct slab *partial_slab)
2469 : {
2470 : struct kmem_cache_node *n = NULL, *n2 = NULL;
2471 : struct slab *slab, *slab_to_discard = NULL;
2472 : unsigned long flags = 0;
2473 :
2474 : while (partial_slab) {
2475 : struct slab new;
2476 : struct slab old;
2477 :
2478 : slab = partial_slab;
2479 : partial_slab = slab->next;
2480 :
2481 : n2 = get_node(s, slab_nid(slab));
2482 : if (n != n2) {
2483 : if (n)
2484 : spin_unlock_irqrestore(&n->list_lock, flags);
2485 :
2486 : n = n2;
2487 : spin_lock_irqsave(&n->list_lock, flags);
2488 : }
2489 :
2490 : do {
2491 :
2492 : old.freelist = slab->freelist;
2493 : old.counters = slab->counters;
2494 : VM_BUG_ON(!old.frozen);
2495 :
2496 : new.counters = old.counters;
2497 : new.freelist = old.freelist;
2498 :
2499 : new.frozen = 0;
2500 :
2501 : } while (!__cmpxchg_double_slab(s, slab,
2502 : old.freelist, old.counters,
2503 : new.freelist, new.counters,
2504 : "unfreezing slab"));
2505 :
2506 : if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2507 : slab->next = slab_to_discard;
2508 : slab_to_discard = slab;
2509 : } else {
2510 : add_partial(n, slab, DEACTIVATE_TO_TAIL);
2511 : stat(s, FREE_ADD_PARTIAL);
2512 : }
2513 : }
2514 :
2515 : if (n)
2516 : spin_unlock_irqrestore(&n->list_lock, flags);
2517 :
2518 : while (slab_to_discard) {
2519 : slab = slab_to_discard;
2520 : slab_to_discard = slab_to_discard->next;
2521 :
2522 : stat(s, DEACTIVATE_EMPTY);
2523 : discard_slab(s, slab);
2524 : stat(s, FREE_SLAB);
2525 : }
2526 : }
2527 :
2528 : /*
2529 : * Unfreeze all the cpu partial slabs.
2530 : */
2531 : static void unfreeze_partials(struct kmem_cache *s)
2532 : {
2533 : struct slab *partial_slab;
2534 : unsigned long flags;
2535 :
2536 : local_lock_irqsave(&s->cpu_slab->lock, flags);
2537 : partial_slab = this_cpu_read(s->cpu_slab->partial);
2538 : this_cpu_write(s->cpu_slab->partial, NULL);
2539 : local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2540 :
2541 : if (partial_slab)
2542 : __unfreeze_partials(s, partial_slab);
2543 : }
2544 :
2545 : static void unfreeze_partials_cpu(struct kmem_cache *s,
2546 : struct kmem_cache_cpu *c)
2547 : {
2548 : struct slab *partial_slab;
2549 :
2550 : partial_slab = slub_percpu_partial(c);
2551 : c->partial = NULL;
2552 :
2553 : if (partial_slab)
2554 : __unfreeze_partials(s, partial_slab);
2555 : }
2556 :
2557 : /*
2558 : * Put a slab that was just frozen (in __slab_free|get_partial_node) into a
2559 : * partial slab slot if available.
2560 : *
2561 : * If we did not find a slot then simply move all the partials to the
2562 : * per node partial list.
2563 : */
2564 : static void put_cpu_partial(struct kmem_cache *s, struct slab *slab, int drain)
2565 : {
2566 : struct slab *oldslab;
2567 : struct slab *slab_to_unfreeze = NULL;
2568 : unsigned long flags;
2569 : int slabs = 0;
2570 :
2571 : local_lock_irqsave(&s->cpu_slab->lock, flags);
2572 :
2573 : oldslab = this_cpu_read(s->cpu_slab->partial);
2574 :
2575 : if (oldslab) {
2576 : if (drain && oldslab->slabs >= s->cpu_partial_slabs) {
2577 : /*
2578 : * Partial array is full. Move the existing set to the
2579 : * per node partial list. Postpone the actual unfreezing
2580 : * outside of the critical section.
2581 : */
2582 : slab_to_unfreeze = oldslab;
2583 : oldslab = NULL;
2584 : } else {
2585 : slabs = oldslab->slabs;
2586 : }
2587 : }
2588 :
2589 : slabs++;
2590 :
2591 : slab->slabs = slabs;
2592 : slab->next = oldslab;
2593 :
2594 : this_cpu_write(s->cpu_slab->partial, slab);
2595 :
2596 : local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2597 :
2598 : if (slab_to_unfreeze) {
2599 : __unfreeze_partials(s, slab_to_unfreeze);
2600 : stat(s, CPU_PARTIAL_DRAIN);
2601 : }
2602 : }
2603 :
2604 : #else /* CONFIG_SLUB_CPU_PARTIAL */
2605 :
2606 : static inline void unfreeze_partials(struct kmem_cache *s) { }
2607 : static inline void unfreeze_partials_cpu(struct kmem_cache *s,
2608 : struct kmem_cache_cpu *c) { }
2609 :
2610 : #endif /* CONFIG_SLUB_CPU_PARTIAL */
2611 :
2612 0 : static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2613 : {
2614 : unsigned long flags;
2615 : struct slab *slab;
2616 : void *freelist;
2617 :
2618 0 : local_lock_irqsave(&s->cpu_slab->lock, flags);
2619 :
2620 0 : slab = c->slab;
2621 0 : freelist = c->freelist;
2622 :
2623 0 : c->slab = NULL;
2624 0 : c->freelist = NULL;
2625 0 : c->tid = next_tid(c->tid);
2626 :
2627 0 : local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2628 :
2629 0 : if (slab) {
2630 0 : deactivate_slab(s, slab, freelist);
2631 0 : stat(s, CPUSLAB_FLUSH);
2632 : }
2633 0 : }
2634 :
2635 2 : static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2636 : {
2637 2 : struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2638 2 : void *freelist = c->freelist;
2639 2 : struct slab *slab = c->slab;
2640 :
2641 2 : c->slab = NULL;
2642 2 : c->freelist = NULL;
2643 4 : c->tid = next_tid(c->tid);
2644 :
2645 2 : if (slab) {
2646 2 : deactivate_slab(s, slab, freelist);
2647 2 : stat(s, CPUSLAB_FLUSH);
2648 : }
2649 :
2650 2 : unfreeze_partials_cpu(s, c);
2651 2 : }
2652 :
2653 : struct slub_flush_work {
2654 : struct work_struct work;
2655 : struct kmem_cache *s;
2656 : bool skip;
2657 : };
2658 :
2659 : /*
2660 : * Flush cpu slab.
2661 : *
2662 : * Called from CPU work handler with migration disabled.
2663 : */
2664 0 : static void flush_cpu_slab(struct work_struct *w)
2665 : {
2666 : struct kmem_cache *s;
2667 : struct kmem_cache_cpu *c;
2668 : struct slub_flush_work *sfw;
2669 :
2670 0 : sfw = container_of(w, struct slub_flush_work, work);
2671 :
2672 0 : s = sfw->s;
2673 0 : c = this_cpu_ptr(s->cpu_slab);
2674 :
2675 0 : if (c->slab)
2676 0 : flush_slab(s, c);
2677 :
2678 0 : unfreeze_partials(s);
2679 0 : }
2680 :
2681 : static bool has_cpu_slab(int cpu, struct kmem_cache *s)
2682 : {
2683 0 : struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2684 :
2685 0 : return c->slab || slub_percpu_partial(c);
2686 : }
2687 :
2688 : static DEFINE_MUTEX(flush_lock);
2689 : static DEFINE_PER_CPU(struct slub_flush_work, slub_flush);
2690 :
2691 0 : static void flush_all_cpus_locked(struct kmem_cache *s)
2692 : {
2693 : struct slub_flush_work *sfw;
2694 : unsigned int cpu;
2695 :
2696 : lockdep_assert_cpus_held();
2697 0 : mutex_lock(&flush_lock);
2698 :
2699 0 : for_each_online_cpu(cpu) {
2700 0 : sfw = &per_cpu(slub_flush, cpu);
2701 0 : if (!has_cpu_slab(cpu, s)) {
2702 0 : sfw->skip = true;
2703 0 : continue;
2704 : }
2705 0 : INIT_WORK(&sfw->work, flush_cpu_slab);
2706 0 : sfw->skip = false;
2707 0 : sfw->s = s;
2708 0 : schedule_work_on(cpu, &sfw->work);
2709 : }
2710 :
2711 0 : for_each_online_cpu(cpu) {
2712 0 : sfw = &per_cpu(slub_flush, cpu);
2713 0 : if (sfw->skip)
2714 0 : continue;
2715 0 : flush_work(&sfw->work);
2716 : }
2717 :
2718 0 : mutex_unlock(&flush_lock);
2719 0 : }
2720 :
2721 : static void flush_all(struct kmem_cache *s)
2722 : {
2723 : cpus_read_lock();
2724 0 : flush_all_cpus_locked(s);
2725 : cpus_read_unlock();
2726 : }
2727 :
2728 : /*
2729 : * Use the cpu notifier to insure that the cpu slabs are flushed when
2730 : * necessary.
2731 : */
2732 0 : static int slub_cpu_dead(unsigned int cpu)
2733 : {
2734 : struct kmem_cache *s;
2735 :
2736 0 : mutex_lock(&slab_mutex);
2737 0 : list_for_each_entry(s, &slab_caches, list)
2738 0 : __flush_cpu_slab(s, cpu);
2739 0 : mutex_unlock(&slab_mutex);
2740 0 : return 0;
2741 : }
2742 :
2743 : /*
2744 : * Check if the objects in a per cpu structure fit numa
2745 : * locality expectations.
2746 : */
2747 : static inline int node_match(struct slab *slab, int node)
2748 : {
2749 : #ifdef CONFIG_NUMA
2750 : if (node != NUMA_NO_NODE && slab_nid(slab) != node)
2751 : return 0;
2752 : #endif
2753 : return 1;
2754 : }
2755 :
2756 : #ifdef CONFIG_SLUB_DEBUG
2757 0 : static int count_free(struct slab *slab)
2758 : {
2759 0 : return slab->objects - slab->inuse;
2760 : }
2761 :
2762 : static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2763 : {
2764 0 : return atomic_long_read(&n->total_objects);
2765 : }
2766 : #endif /* CONFIG_SLUB_DEBUG */
2767 :
2768 : #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2769 0 : static unsigned long count_partial(struct kmem_cache_node *n,
2770 : int (*get_count)(struct slab *))
2771 : {
2772 : unsigned long flags;
2773 0 : unsigned long x = 0;
2774 : struct slab *slab;
2775 :
2776 0 : spin_lock_irqsave(&n->list_lock, flags);
2777 0 : list_for_each_entry(slab, &n->partial, slab_list)
2778 0 : x += get_count(slab);
2779 0 : spin_unlock_irqrestore(&n->list_lock, flags);
2780 0 : return x;
2781 : }
2782 : #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2783 :
2784 : static noinline void
2785 0 : slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2786 : {
2787 : #ifdef CONFIG_SLUB_DEBUG
2788 : static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2789 : DEFAULT_RATELIMIT_BURST);
2790 : int node;
2791 : struct kmem_cache_node *n;
2792 :
2793 0 : if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2794 : return;
2795 :
2796 0 : pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2797 : nid, gfpflags, &gfpflags);
2798 0 : pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
2799 : s->name, s->object_size, s->size, oo_order(s->oo),
2800 : oo_order(s->min));
2801 :
2802 0 : if (oo_order(s->min) > get_order(s->object_size))
2803 0 : pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2804 : s->name);
2805 :
2806 0 : for_each_kmem_cache_node(s, node, n) {
2807 : unsigned long nr_slabs;
2808 : unsigned long nr_objs;
2809 : unsigned long nr_free;
2810 :
2811 0 : nr_free = count_partial(n, count_free);
2812 0 : nr_slabs = node_nr_slabs(n);
2813 0 : nr_objs = node_nr_objs(n);
2814 :
2815 0 : pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2816 : node, nr_slabs, nr_objs, nr_free);
2817 : }
2818 : #endif
2819 : }
2820 :
2821 : static inline bool pfmemalloc_match(struct slab *slab, gfp_t gfpflags)
2822 : {
2823 1008 : if (unlikely(slab_test_pfmemalloc(slab)))
2824 0 : return gfp_pfmemalloc_allowed(gfpflags);
2825 :
2826 : return true;
2827 : }
2828 :
2829 : /*
2830 : * Check the slab->freelist and either transfer the freelist to the
2831 : * per cpu freelist or deactivate the slab.
2832 : *
2833 : * The slab is still frozen if the return value is not NULL.
2834 : *
2835 : * If this function returns NULL then the slab has been unfrozen.
2836 : */
2837 : static inline void *get_freelist(struct kmem_cache *s, struct slab *slab)
2838 : {
2839 : struct slab new;
2840 : unsigned long counters;
2841 : void *freelist;
2842 :
2843 : lockdep_assert_held(this_cpu_ptr(&s->cpu_slab->lock));
2844 :
2845 : do {
2846 463 : freelist = slab->freelist;
2847 463 : counters = slab->counters;
2848 :
2849 463 : new.counters = counters;
2850 : VM_BUG_ON(!new.frozen);
2851 :
2852 463 : new.inuse = slab->objects;
2853 463 : new.frozen = freelist != NULL;
2854 :
2855 926 : } while (!__cmpxchg_double_slab(s, slab,
2856 : freelist, counters,
2857 : NULL, new.counters,
2858 463 : "get_freelist"));
2859 :
2860 : return freelist;
2861 : }
2862 :
2863 : /*
2864 : * Slow path. The lockless freelist is empty or we need to perform
2865 : * debugging duties.
2866 : *
2867 : * Processing is still very fast if new objects have been freed to the
2868 : * regular freelist. In that case we simply take over the regular freelist
2869 : * as the lockless freelist and zap the regular freelist.
2870 : *
2871 : * If that is not working then we fall back to the partial lists. We take the
2872 : * first element of the freelist as the object to allocate now and move the
2873 : * rest of the freelist to the lockless freelist.
2874 : *
2875 : * And if we were unable to get a new slab from the partial slab lists then
2876 : * we need to allocate a new slab. This is the slowest path since it involves
2877 : * a call to the page allocator and the setup of a new slab.
2878 : *
2879 : * Version of __slab_alloc to use when we know that preemption is
2880 : * already disabled (which is the case for bulk allocation).
2881 : */
2882 499 : static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2883 : unsigned long addr, struct kmem_cache_cpu *c)
2884 : {
2885 : void *freelist;
2886 : struct slab *slab;
2887 : unsigned long flags;
2888 :
2889 499 : stat(s, ALLOC_SLOWPATH);
2890 :
2891 : reread_slab:
2892 :
2893 499 : slab = READ_ONCE(c->slab);
2894 499 : if (!slab) {
2895 : /*
2896 : * if the node is not online or has no normal memory, just
2897 : * ignore the node constraint
2898 : */
2899 36 : if (unlikely(node != NUMA_NO_NODE &&
2900 : !node_isset(node, slab_nodes)))
2901 0 : node = NUMA_NO_NODE;
2902 : goto new_slab;
2903 : }
2904 : redo:
2905 :
2906 463 : if (unlikely(!node_match(slab, node))) {
2907 : /*
2908 : * same as above but node_match() being false already
2909 : * implies node != NUMA_NO_NODE
2910 : */
2911 : if (!node_isset(node, slab_nodes)) {
2912 : node = NUMA_NO_NODE;
2913 : goto redo;
2914 : } else {
2915 : stat(s, ALLOC_NODE_MISMATCH);
2916 : goto deactivate_slab;
2917 : }
2918 : }
2919 :
2920 : /*
2921 : * By rights, we should be searching for a slab page that was
2922 : * PFMEMALLOC but right now, we are losing the pfmemalloc
2923 : * information when the page leaves the per-cpu allocator
2924 : */
2925 926 : if (unlikely(!pfmemalloc_match(slab, gfpflags)))
2926 : goto deactivate_slab;
2927 :
2928 : /* must check again c->slab in case we got preempted and it changed */
2929 463 : local_lock_irqsave(&s->cpu_slab->lock, flags);
2930 463 : if (unlikely(slab != c->slab)) {
2931 0 : local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2932 : goto reread_slab;
2933 : }
2934 463 : freelist = c->freelist;
2935 463 : if (freelist)
2936 : goto load_freelist;
2937 :
2938 463 : freelist = get_freelist(s, slab);
2939 :
2940 463 : if (!freelist) {
2941 463 : c->slab = NULL;
2942 463 : local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2943 : stat(s, DEACTIVATE_BYPASS);
2944 : goto new_slab;
2945 : }
2946 :
2947 : stat(s, ALLOC_REFILL);
2948 :
2949 : load_freelist:
2950 :
2951 499 : lockdep_assert_held(this_cpu_ptr(&s->cpu_slab->lock));
2952 :
2953 : /*
2954 : * freelist is pointing to the list of objects to be used.
2955 : * slab is pointing to the slab from which the objects are obtained.
2956 : * That slab must be frozen for per cpu allocations to work.
2957 : */
2958 : VM_BUG_ON(!c->slab->frozen);
2959 998 : c->freelist = get_freepointer(s, freelist);
2960 998 : c->tid = next_tid(c->tid);
2961 998 : local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2962 499 : return freelist;
2963 :
2964 : deactivate_slab:
2965 :
2966 0 : local_lock_irqsave(&s->cpu_slab->lock, flags);
2967 0 : if (slab != c->slab) {
2968 0 : local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2969 : goto reread_slab;
2970 : }
2971 0 : freelist = c->freelist;
2972 0 : c->slab = NULL;
2973 0 : c->freelist = NULL;
2974 0 : local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2975 0 : deactivate_slab(s, slab, freelist);
2976 :
2977 : new_slab:
2978 :
2979 : if (slub_percpu_partial(c)) {
2980 : local_lock_irqsave(&s->cpu_slab->lock, flags);
2981 : if (unlikely(c->slab)) {
2982 : local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2983 : goto reread_slab;
2984 : }
2985 : if (unlikely(!slub_percpu_partial(c))) {
2986 : local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2987 : /* we were preempted and partial list got empty */
2988 : goto new_objects;
2989 : }
2990 :
2991 : slab = c->slab = slub_percpu_partial(c);
2992 : slub_set_percpu_partial(c, slab);
2993 : local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2994 : stat(s, CPU_PARTIAL_ALLOC);
2995 : goto redo;
2996 : }
2997 :
2998 : new_objects:
2999 :
3000 499 : freelist = get_partial(s, gfpflags, node, &slab);
3001 499 : if (freelist)
3002 : goto check_new_slab;
3003 :
3004 453 : slub_put_cpu_ptr(s->cpu_slab);
3005 453 : slab = new_slab(s, gfpflags, node);
3006 453 : c = slub_get_cpu_ptr(s->cpu_slab);
3007 :
3008 453 : if (unlikely(!slab)) {
3009 0 : slab_out_of_memory(s, gfpflags, node);
3010 0 : return NULL;
3011 : }
3012 :
3013 : /*
3014 : * No other reference to the slab yet so we can
3015 : * muck around with it freely without cmpxchg
3016 : */
3017 453 : freelist = slab->freelist;
3018 453 : slab->freelist = NULL;
3019 :
3020 453 : stat(s, ALLOC_SLAB);
3021 :
3022 : check_new_slab:
3023 :
3024 499 : if (kmem_cache_debug(s)) {
3025 0 : if (!alloc_debug_processing(s, slab, freelist, addr)) {
3026 : /* Slab failed checks. Next slab needed */
3027 : goto new_slab;
3028 : } else {
3029 : /*
3030 : * For debug case, we don't load freelist so that all
3031 : * allocations go through alloc_debug_processing()
3032 : */
3033 : goto return_single;
3034 : }
3035 : }
3036 :
3037 998 : if (unlikely(!pfmemalloc_match(slab, gfpflags)))
3038 : /*
3039 : * For !pfmemalloc_match() case we don't load freelist so that
3040 : * we don't make further mismatched allocations easier.
3041 : */
3042 : goto return_single;
3043 :
3044 : retry_load_slab:
3045 :
3046 499 : local_lock_irqsave(&s->cpu_slab->lock, flags);
3047 499 : if (unlikely(c->slab)) {
3048 0 : void *flush_freelist = c->freelist;
3049 0 : struct slab *flush_slab = c->slab;
3050 :
3051 0 : c->slab = NULL;
3052 0 : c->freelist = NULL;
3053 0 : c->tid = next_tid(c->tid);
3054 :
3055 0 : local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3056 :
3057 0 : deactivate_slab(s, flush_slab, flush_freelist);
3058 :
3059 0 : stat(s, CPUSLAB_FLUSH);
3060 :
3061 : goto retry_load_slab;
3062 : }
3063 499 : c->slab = slab;
3064 :
3065 499 : goto load_freelist;
3066 :
3067 : return_single:
3068 :
3069 0 : deactivate_slab(s, slab, get_freepointer(s, freelist));
3070 0 : return freelist;
3071 : }
3072 :
3073 : /*
3074 : * A wrapper for ___slab_alloc() for contexts where preemption is not yet
3075 : * disabled. Compensates for possible cpu changes by refetching the per cpu area
3076 : * pointer.
3077 : */
3078 : static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
3079 : unsigned long addr, struct kmem_cache_cpu *c)
3080 : {
3081 : void *p;
3082 :
3083 : #ifdef CONFIG_PREEMPT_COUNT
3084 : /*
3085 : * We may have been preempted and rescheduled on a different
3086 : * cpu before disabling preemption. Need to reload cpu area
3087 : * pointer.
3088 : */
3089 : c = slub_get_cpu_ptr(s->cpu_slab);
3090 : #endif
3091 :
3092 499 : p = ___slab_alloc(s, gfpflags, node, addr, c);
3093 : #ifdef CONFIG_PREEMPT_COUNT
3094 : slub_put_cpu_ptr(s->cpu_slab);
3095 : #endif
3096 : return p;
3097 : }
3098 :
3099 : /*
3100 : * If the object has been wiped upon free, make sure it's fully initialized by
3101 : * zeroing out freelist pointer.
3102 : */
3103 : static __always_inline void maybe_wipe_obj_freeptr(struct kmem_cache *s,
3104 : void *obj)
3105 : {
3106 18327 : if (unlikely(slab_want_init_on_free(s)) && obj)
3107 0 : memset((void *)((char *)kasan_reset_tag(obj) + s->offset),
3108 : 0, sizeof(void *));
3109 : }
3110 :
3111 : /*
3112 : * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
3113 : * have the fastpath folded into their functions. So no function call
3114 : * overhead for requests that can be satisfied on the fastpath.
3115 : *
3116 : * The fastpath works by first checking if the lockless freelist can be used.
3117 : * If not then __slab_alloc is called for slow processing.
3118 : *
3119 : * Otherwise we can simply pick the next object from the lockless free list.
3120 : */
3121 : static __always_inline void *slab_alloc_node(struct kmem_cache *s, struct list_lru *lru,
3122 : gfp_t gfpflags, int node, unsigned long addr, size_t orig_size)
3123 : {
3124 : void *object;
3125 : struct kmem_cache_cpu *c;
3126 : struct slab *slab;
3127 : unsigned long tid;
3128 18327 : struct obj_cgroup *objcg = NULL;
3129 18327 : bool init = false;
3130 :
3131 36654 : s = slab_pre_alloc_hook(s, lru, &objcg, 1, gfpflags);
3132 18327 : if (!s)
3133 : return NULL;
3134 :
3135 18327 : object = kfence_alloc(s, orig_size, gfpflags);
3136 : if (unlikely(object))
3137 : goto out;
3138 :
3139 : redo:
3140 : /*
3141 : * Must read kmem_cache cpu data via this cpu ptr. Preemption is
3142 : * enabled. We may switch back and forth between cpus while
3143 : * reading from one cpu area. That does not matter as long
3144 : * as we end up on the original cpu again when doing the cmpxchg.
3145 : *
3146 : * We must guarantee that tid and kmem_cache_cpu are retrieved on the
3147 : * same cpu. We read first the kmem_cache_cpu pointer and use it to read
3148 : * the tid. If we are preempted and switched to another cpu between the
3149 : * two reads, it's OK as the two are still associated with the same cpu
3150 : * and cmpxchg later will validate the cpu.
3151 : */
3152 18327 : c = raw_cpu_ptr(s->cpu_slab);
3153 18327 : tid = READ_ONCE(c->tid);
3154 :
3155 : /*
3156 : * Irqless object alloc/free algorithm used here depends on sequence
3157 : * of fetching cpu_slab's data. tid should be fetched before anything
3158 : * on c to guarantee that object and slab associated with previous tid
3159 : * won't be used with current tid. If we fetch tid first, object and
3160 : * slab could be one associated with next tid and our alloc/free
3161 : * request will be failed. In this case, we will retry. So, no problem.
3162 : */
3163 18327 : barrier();
3164 :
3165 : /*
3166 : * The transaction ids are globally unique per cpu and per operation on
3167 : * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
3168 : * occurs on the right processor and that there was no operation on the
3169 : * linked list in between.
3170 : */
3171 :
3172 18327 : object = c->freelist;
3173 18327 : slab = c->slab;
3174 : /*
3175 : * We cannot use the lockless fastpath on PREEMPT_RT because if a
3176 : * slowpath has taken the local_lock_irqsave(), it is not protected
3177 : * against a fast path operation in an irq handler. So we need to take
3178 : * the slow path which uses local_lock. It is still relatively fast if
3179 : * there is a suitable cpu freelist.
3180 : */
3181 18327 : if (IS_ENABLED(CONFIG_PREEMPT_RT) ||
3182 36155 : unlikely(!object || !slab || !node_match(slab, node))) {
3183 499 : object = __slab_alloc(s, gfpflags, node, addr, c);
3184 : } else {
3185 17828 : void *next_object = get_freepointer_safe(s, object);
3186 :
3187 : /*
3188 : * The cmpxchg will only match if there was no additional
3189 : * operation and if we are on the right processor.
3190 : *
3191 : * The cmpxchg does the following atomically (without lock
3192 : * semantics!)
3193 : * 1. Relocate first pointer to the current per cpu area.
3194 : * 2. Verify that tid and freelist have not been changed
3195 : * 3. If they were not changed replace tid and freelist
3196 : *
3197 : * Since this is without lock semantics the protection is only
3198 : * against code executing on this cpu *not* from access by
3199 : * other cpus.
3200 : */
3201 71312 : if (unlikely(!this_cpu_cmpxchg_double(
3202 : s->cpu_slab->freelist, s->cpu_slab->tid,
3203 : object, tid,
3204 : next_object, next_tid(tid)))) {
3205 :
3206 : note_cmpxchg_failure("slab_alloc", s, tid);
3207 : goto redo;
3208 : }
3209 17828 : prefetch_freepointer(s, next_object);
3210 : stat(s, ALLOC_FASTPATH);
3211 : }
3212 :
3213 36654 : maybe_wipe_obj_freeptr(s, object);
3214 36654 : init = slab_want_init_on_alloc(gfpflags, s);
3215 :
3216 : out:
3217 18327 : slab_post_alloc_hook(s, objcg, gfpflags, 1, &object, init);
3218 :
3219 18327 : return object;
3220 : }
3221 :
3222 : static __always_inline void *slab_alloc(struct kmem_cache *s, struct list_lru *lru,
3223 : gfp_t gfpflags, unsigned long addr, size_t orig_size)
3224 : {
3225 18327 : return slab_alloc_node(s, lru, gfpflags, NUMA_NO_NODE, addr, orig_size);
3226 : }
3227 :
3228 : static __always_inline
3229 : void *__kmem_cache_alloc_lru(struct kmem_cache *s, struct list_lru *lru,
3230 : gfp_t gfpflags)
3231 : {
3232 29858 : void *ret = slab_alloc(s, lru, gfpflags, _RET_IP_, s->object_size);
3233 :
3234 14929 : trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
3235 14929 : s->size, gfpflags);
3236 :
3237 : return ret;
3238 : }
3239 :
3240 14895 : void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
3241 : {
3242 14895 : return __kmem_cache_alloc_lru(s, NULL, gfpflags);
3243 : }
3244 : EXPORT_SYMBOL(kmem_cache_alloc);
3245 :
3246 34 : void *kmem_cache_alloc_lru(struct kmem_cache *s, struct list_lru *lru,
3247 : gfp_t gfpflags)
3248 : {
3249 34 : return __kmem_cache_alloc_lru(s, lru, gfpflags);
3250 : }
3251 : EXPORT_SYMBOL(kmem_cache_alloc_lru);
3252 :
3253 : #ifdef CONFIG_TRACING
3254 : void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
3255 : {
3256 : void *ret = slab_alloc(s, NULL, gfpflags, _RET_IP_, size);
3257 : trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
3258 : ret = kasan_kmalloc(s, ret, size, gfpflags);
3259 : return ret;
3260 : }
3261 : EXPORT_SYMBOL(kmem_cache_alloc_trace);
3262 : #endif
3263 :
3264 : #ifdef CONFIG_NUMA
3265 : void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
3266 : {
3267 : void *ret = slab_alloc_node(s, NULL, gfpflags, node, _RET_IP_, s->object_size);
3268 :
3269 : trace_kmem_cache_alloc_node(_RET_IP_, ret,
3270 : s->object_size, s->size, gfpflags, node);
3271 :
3272 : return ret;
3273 : }
3274 : EXPORT_SYMBOL(kmem_cache_alloc_node);
3275 :
3276 : #ifdef CONFIG_TRACING
3277 : void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
3278 : gfp_t gfpflags,
3279 : int node, size_t size)
3280 : {
3281 : void *ret = slab_alloc_node(s, NULL, gfpflags, node, _RET_IP_, size);
3282 :
3283 : trace_kmalloc_node(_RET_IP_, ret,
3284 : size, s->size, gfpflags, node);
3285 :
3286 : ret = kasan_kmalloc(s, ret, size, gfpflags);
3287 : return ret;
3288 : }
3289 : EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3290 : #endif
3291 : #endif /* CONFIG_NUMA */
3292 :
3293 : /*
3294 : * Slow path handling. This may still be called frequently since objects
3295 : * have a longer lifetime than the cpu slabs in most processing loads.
3296 : *
3297 : * So we still attempt to reduce cache line usage. Just take the slab
3298 : * lock and free the item. If there is no additional partial slab
3299 : * handling required then we can return immediately.
3300 : */
3301 1278 : static void __slab_free(struct kmem_cache *s, struct slab *slab,
3302 : void *head, void *tail, int cnt,
3303 : unsigned long addr)
3304 :
3305 : {
3306 : void *prior;
3307 : int was_frozen;
3308 : struct slab new;
3309 : unsigned long counters;
3310 1278 : struct kmem_cache_node *n = NULL;
3311 : unsigned long flags;
3312 :
3313 1278 : stat(s, FREE_SLOWPATH);
3314 :
3315 1278 : if (kfence_free(head))
3316 1275 : return;
3317 :
3318 1278 : if (kmem_cache_debug(s) &&
3319 0 : !free_debug_processing(s, slab, head, tail, cnt, addr))
3320 : return;
3321 :
3322 : do {
3323 1278 : if (unlikely(n)) {
3324 0 : spin_unlock_irqrestore(&n->list_lock, flags);
3325 0 : n = NULL;
3326 : }
3327 1278 : prior = slab->freelist;
3328 1278 : counters = slab->counters;
3329 2556 : set_freepointer(s, tail, prior);
3330 1278 : new.counters = counters;
3331 1278 : was_frozen = new.frozen;
3332 1278 : new.inuse -= cnt;
3333 1278 : if ((!new.inuse || !prior) && !was_frozen) {
3334 :
3335 91 : if (kmem_cache_has_cpu_partial(s) && !prior) {
3336 :
3337 : /*
3338 : * Slab was on no list before and will be
3339 : * partially empty
3340 : * We can defer the list move and instead
3341 : * freeze it.
3342 : */
3343 : new.frozen = 1;
3344 :
3345 : } else { /* Needs to be taken off a list */
3346 :
3347 273 : n = get_node(s, slab_nid(slab));
3348 : /*
3349 : * Speculatively acquire the list_lock.
3350 : * If the cmpxchg does not succeed then we may
3351 : * drop the list_lock without any processing.
3352 : *
3353 : * Otherwise the list_lock will synchronize with
3354 : * other processors updating the list of slabs.
3355 : */
3356 91 : spin_lock_irqsave(&n->list_lock, flags);
3357 :
3358 : }
3359 : }
3360 :
3361 1278 : } while (!cmpxchg_double_slab(s, slab,
3362 : prior, counters,
3363 : head, new.counters,
3364 1278 : "__slab_free"));
3365 :
3366 1278 : if (likely(!n)) {
3367 :
3368 : if (likely(was_frozen)) {
3369 : /*
3370 : * The list lock was not taken therefore no list
3371 : * activity can be necessary.
3372 : */
3373 : stat(s, FREE_FROZEN);
3374 : } else if (new.frozen) {
3375 : /*
3376 : * If we just froze the slab then put it onto the
3377 : * per cpu partial list.
3378 : */
3379 : put_cpu_partial(s, slab, 1);
3380 : stat(s, CPU_PARTIAL_FREE);
3381 : }
3382 :
3383 : return;
3384 : }
3385 :
3386 91 : if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
3387 : goto slab_empty;
3388 :
3389 : /*
3390 : * Objects left in the slab. If it was not on the partial list before
3391 : * then add it.
3392 : */
3393 88 : if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
3394 98 : remove_full(s, n, slab);
3395 : add_partial(n, slab, DEACTIVATE_TO_TAIL);
3396 : stat(s, FREE_ADD_PARTIAL);
3397 : }
3398 88 : spin_unlock_irqrestore(&n->list_lock, flags);
3399 : return;
3400 :
3401 : slab_empty:
3402 3 : if (prior) {
3403 : /*
3404 : * Slab on the partial list.
3405 : */
3406 3 : remove_partial(n, slab);
3407 : stat(s, FREE_REMOVE_PARTIAL);
3408 : } else {
3409 : /* Slab must be on the full list */
3410 0 : remove_full(s, n, slab);
3411 : }
3412 :
3413 6 : spin_unlock_irqrestore(&n->list_lock, flags);
3414 3 : stat(s, FREE_SLAB);
3415 3 : discard_slab(s, slab);
3416 : }
3417 :
3418 : /*
3419 : * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
3420 : * can perform fastpath freeing without additional function calls.
3421 : *
3422 : * The fastpath is only possible if we are freeing to the current cpu slab
3423 : * of this processor. This typically the case if we have just allocated
3424 : * the item before.
3425 : *
3426 : * If fastpath is not possible then fall back to __slab_free where we deal
3427 : * with all sorts of special processing.
3428 : *
3429 : * Bulk free of a freelist with several objects (all pointing to the
3430 : * same slab) possible by specifying head and tail ptr, plus objects
3431 : * count (cnt). Bulk free indicated by tail pointer being set.
3432 : */
3433 : static __always_inline void do_slab_free(struct kmem_cache *s,
3434 : struct slab *slab, void *head, void *tail,
3435 : int cnt, unsigned long addr)
3436 : {
3437 5311 : void *tail_obj = tail ? : head;
3438 : struct kmem_cache_cpu *c;
3439 : unsigned long tid;
3440 :
3441 : /* memcg_slab_free_hook() is already called for bulk free. */
3442 : if (!tail)
3443 : memcg_slab_free_hook(s, &head, 1);
3444 : redo:
3445 : /*
3446 : * Determine the currently cpus per cpu slab.
3447 : * The cpu may change afterward. However that does not matter since
3448 : * data is retrieved via this pointer. If we are on the same cpu
3449 : * during the cmpxchg then the free will succeed.
3450 : */
3451 5311 : c = raw_cpu_ptr(s->cpu_slab);
3452 5311 : tid = READ_ONCE(c->tid);
3453 :
3454 : /* Same with comment on barrier() in slab_alloc_node() */
3455 5311 : barrier();
3456 :
3457 5311 : if (likely(slab == c->slab)) {
3458 : #ifndef CONFIG_PREEMPT_RT
3459 4033 : void **freelist = READ_ONCE(c->freelist);
3460 :
3461 8066 : set_freepointer(s, tail_obj, freelist);
3462 :
3463 16132 : if (unlikely(!this_cpu_cmpxchg_double(
3464 : s->cpu_slab->freelist, s->cpu_slab->tid,
3465 : freelist, tid,
3466 : head, next_tid(tid)))) {
3467 :
3468 : note_cmpxchg_failure("slab_free", s, tid);
3469 : goto redo;
3470 : }
3471 : #else /* CONFIG_PREEMPT_RT */
3472 : /*
3473 : * We cannot use the lockless fastpath on PREEMPT_RT because if
3474 : * a slowpath has taken the local_lock_irqsave(), it is not
3475 : * protected against a fast path operation in an irq handler. So
3476 : * we need to take the local_lock. We shouldn't simply defer to
3477 : * __slab_free() as that wouldn't use the cpu freelist at all.
3478 : */
3479 : void **freelist;
3480 :
3481 : local_lock(&s->cpu_slab->lock);
3482 : c = this_cpu_ptr(s->cpu_slab);
3483 : if (unlikely(slab != c->slab)) {
3484 : local_unlock(&s->cpu_slab->lock);
3485 : goto redo;
3486 : }
3487 : tid = c->tid;
3488 : freelist = c->freelist;
3489 :
3490 : set_freepointer(s, tail_obj, freelist);
3491 : c->freelist = head;
3492 : c->tid = next_tid(tid);
3493 :
3494 : local_unlock(&s->cpu_slab->lock);
3495 : #endif
3496 : stat(s, FREE_FASTPATH);
3497 : } else
3498 1278 : __slab_free(s, slab, head, tail_obj, cnt, addr);
3499 :
3500 : }
3501 :
3502 : static __always_inline void slab_free(struct kmem_cache *s, struct slab *slab,
3503 : void *head, void *tail, int cnt,
3504 : unsigned long addr)
3505 : {
3506 : /*
3507 : * With KASAN enabled slab_free_freelist_hook modifies the freelist
3508 : * to remove objects, whose reuse must be delayed.
3509 : */
3510 5311 : if (slab_free_freelist_hook(s, &head, &tail, &cnt))
3511 5311 : do_slab_free(s, slab, head, tail, cnt, addr);
3512 : }
3513 :
3514 : #ifdef CONFIG_KASAN_GENERIC
3515 : void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
3516 : {
3517 : do_slab_free(cache, virt_to_slab(x), x, NULL, 1, addr);
3518 : }
3519 : #endif
3520 :
3521 2969 : void kmem_cache_free(struct kmem_cache *s, void *x)
3522 : {
3523 2969 : s = cache_from_obj(s, x);
3524 2969 : if (!s)
3525 : return;
3526 2969 : trace_kmem_cache_free(_RET_IP_, x, s->name);
3527 5938 : slab_free(s, virt_to_slab(x), x, NULL, 1, _RET_IP_);
3528 : }
3529 : EXPORT_SYMBOL(kmem_cache_free);
3530 :
3531 : struct detached_freelist {
3532 : struct slab *slab;
3533 : void *tail;
3534 : void *freelist;
3535 : int cnt;
3536 : struct kmem_cache *s;
3537 : };
3538 :
3539 8 : static inline void free_large_kmalloc(struct folio *folio, void *object)
3540 : {
3541 8 : unsigned int order = folio_order(folio);
3542 :
3543 8 : if (WARN_ON_ONCE(order == 0))
3544 0 : pr_warn_once("object pointer: 0x%p\n", object);
3545 :
3546 8 : kfree_hook(object);
3547 16 : mod_lruvec_page_state(folio_page(folio, 0), NR_SLAB_UNRECLAIMABLE_B,
3548 8 : -(PAGE_SIZE << order));
3549 8 : __free_pages(folio_page(folio, 0), order);
3550 8 : }
3551 :
3552 : /*
3553 : * This function progressively scans the array with free objects (with
3554 : * a limited look ahead) and extract objects belonging to the same
3555 : * slab. It builds a detached freelist directly within the given
3556 : * slab/objects. This can happen without any need for
3557 : * synchronization, because the objects are owned by running process.
3558 : * The freelist is build up as a single linked list in the objects.
3559 : * The idea is, that this detached freelist can then be bulk
3560 : * transferred to the real freelist(s), but only requiring a single
3561 : * synchronization primitive. Look ahead in the array is limited due
3562 : * to performance reasons.
3563 : */
3564 : static inline
3565 0 : int build_detached_freelist(struct kmem_cache *s, size_t size,
3566 : void **p, struct detached_freelist *df)
3567 : {
3568 0 : size_t first_skipped_index = 0;
3569 0 : int lookahead = 3;
3570 : void *object;
3571 : struct folio *folio;
3572 : struct slab *slab;
3573 :
3574 : /* Always re-init detached_freelist */
3575 0 : df->slab = NULL;
3576 :
3577 : do {
3578 0 : object = p[--size];
3579 : /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3580 0 : } while (!object && size);
3581 :
3582 0 : if (!object)
3583 : return 0;
3584 :
3585 0 : folio = virt_to_folio(object);
3586 0 : if (!s) {
3587 : /* Handle kalloc'ed objects */
3588 0 : if (unlikely(!folio_test_slab(folio))) {
3589 0 : free_large_kmalloc(folio, object);
3590 0 : p[size] = NULL; /* mark object processed */
3591 0 : return size;
3592 : }
3593 : /* Derive kmem_cache from object */
3594 0 : slab = folio_slab(folio);
3595 0 : df->s = slab->slab_cache;
3596 : } else {
3597 0 : slab = folio_slab(folio);
3598 0 : df->s = cache_from_obj(s, object); /* Support for memcg */
3599 : }
3600 :
3601 0 : if (is_kfence_address(object)) {
3602 : slab_free_hook(df->s, object, false);
3603 : __kfence_free(object);
3604 : p[size] = NULL; /* mark object processed */
3605 : return size;
3606 : }
3607 :
3608 : /* Start new detached freelist */
3609 0 : df->slab = slab;
3610 0 : set_freepointer(df->s, object, NULL);
3611 0 : df->tail = object;
3612 0 : df->freelist = object;
3613 0 : p[size] = NULL; /* mark object processed */
3614 0 : df->cnt = 1;
3615 :
3616 0 : while (size) {
3617 0 : object = p[--size];
3618 0 : if (!object)
3619 0 : continue; /* Skip processed objects */
3620 :
3621 : /* df->slab is always set at this point */
3622 0 : if (df->slab == virt_to_slab(object)) {
3623 : /* Opportunity build freelist */
3624 0 : set_freepointer(df->s, object, df->freelist);
3625 0 : df->freelist = object;
3626 0 : df->cnt++;
3627 0 : p[size] = NULL; /* mark object processed */
3628 :
3629 0 : continue;
3630 : }
3631 :
3632 : /* Limit look ahead search */
3633 0 : if (!--lookahead)
3634 : break;
3635 :
3636 0 : if (!first_skipped_index)
3637 0 : first_skipped_index = size + 1;
3638 : }
3639 :
3640 0 : return first_skipped_index;
3641 : }
3642 :
3643 : /* Note that interrupts must be enabled when calling this function. */
3644 0 : void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
3645 : {
3646 0 : if (WARN_ON(!size))
3647 : return;
3648 :
3649 : memcg_slab_free_hook(s, p, size);
3650 : do {
3651 : struct detached_freelist df;
3652 :
3653 0 : size = build_detached_freelist(s, size, p, &df);
3654 0 : if (!df.slab)
3655 0 : continue;
3656 :
3657 0 : slab_free(df.s, df.slab, df.freelist, df.tail, df.cnt, _RET_IP_);
3658 0 : } while (likely(size));
3659 : }
3660 : EXPORT_SYMBOL(kmem_cache_free_bulk);
3661 :
3662 : /* Note that interrupts must be enabled when calling this function. */
3663 0 : int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3664 : void **p)
3665 : {
3666 : struct kmem_cache_cpu *c;
3667 : int i;
3668 0 : struct obj_cgroup *objcg = NULL;
3669 :
3670 : /* memcg and kmem_cache debug support */
3671 0 : s = slab_pre_alloc_hook(s, NULL, &objcg, size, flags);
3672 0 : if (unlikely(!s))
3673 : return false;
3674 : /*
3675 : * Drain objects in the per cpu slab, while disabling local
3676 : * IRQs, which protects against PREEMPT and interrupts
3677 : * handlers invoking normal fastpath.
3678 : */
3679 0 : c = slub_get_cpu_ptr(s->cpu_slab);
3680 0 : local_lock_irq(&s->cpu_slab->lock);
3681 :
3682 0 : for (i = 0; i < size; i++) {
3683 0 : void *object = kfence_alloc(s, s->object_size, flags);
3684 :
3685 : if (unlikely(object)) {
3686 : p[i] = object;
3687 : continue;
3688 : }
3689 :
3690 0 : object = c->freelist;
3691 0 : if (unlikely(!object)) {
3692 : /*
3693 : * We may have removed an object from c->freelist using
3694 : * the fastpath in the previous iteration; in that case,
3695 : * c->tid has not been bumped yet.
3696 : * Since ___slab_alloc() may reenable interrupts while
3697 : * allocating memory, we should bump c->tid now.
3698 : */
3699 0 : c->tid = next_tid(c->tid);
3700 :
3701 0 : local_unlock_irq(&s->cpu_slab->lock);
3702 :
3703 : /*
3704 : * Invoking slow path likely have side-effect
3705 : * of re-populating per CPU c->freelist
3706 : */
3707 0 : p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
3708 0 : _RET_IP_, c);
3709 0 : if (unlikely(!p[i]))
3710 : goto error;
3711 :
3712 0 : c = this_cpu_ptr(s->cpu_slab);
3713 0 : maybe_wipe_obj_freeptr(s, p[i]);
3714 :
3715 0 : local_lock_irq(&s->cpu_slab->lock);
3716 :
3717 0 : continue; /* goto for-loop */
3718 : }
3719 0 : c->freelist = get_freepointer(s, object);
3720 0 : p[i] = object;
3721 0 : maybe_wipe_obj_freeptr(s, p[i]);
3722 : }
3723 0 : c->tid = next_tid(c->tid);
3724 0 : local_unlock_irq(&s->cpu_slab->lock);
3725 0 : slub_put_cpu_ptr(s->cpu_slab);
3726 :
3727 : /*
3728 : * memcg and kmem_cache debug support and memory initialization.
3729 : * Done outside of the IRQ disabled fastpath loop.
3730 : */
3731 0 : slab_post_alloc_hook(s, objcg, flags, size, p,
3732 0 : slab_want_init_on_alloc(flags, s));
3733 0 : return i;
3734 : error:
3735 0 : slub_put_cpu_ptr(s->cpu_slab);
3736 0 : slab_post_alloc_hook(s, objcg, flags, i, p, false);
3737 0 : __kmem_cache_free_bulk(s, i, p);
3738 0 : return 0;
3739 : }
3740 : EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3741 :
3742 :
3743 : /*
3744 : * Object placement in a slab is made very easy because we always start at
3745 : * offset 0. If we tune the size of the object to the alignment then we can
3746 : * get the required alignment by putting one properly sized object after
3747 : * another.
3748 : *
3749 : * Notice that the allocation order determines the sizes of the per cpu
3750 : * caches. Each processor has always one slab available for allocations.
3751 : * Increasing the allocation order reduces the number of times that slabs
3752 : * must be moved on and off the partial lists and is therefore a factor in
3753 : * locking overhead.
3754 : */
3755 :
3756 : /*
3757 : * Minimum / Maximum order of slab pages. This influences locking overhead
3758 : * and slab fragmentation. A higher order reduces the number of partial slabs
3759 : * and increases the number of allocations possible without having to
3760 : * take the list_lock.
3761 : */
3762 : static unsigned int slub_min_order;
3763 : static unsigned int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
3764 : static unsigned int slub_min_objects;
3765 :
3766 : /*
3767 : * Calculate the order of allocation given an slab object size.
3768 : *
3769 : * The order of allocation has significant impact on performance and other
3770 : * system components. Generally order 0 allocations should be preferred since
3771 : * order 0 does not cause fragmentation in the page allocator. Larger objects
3772 : * be problematic to put into order 0 slabs because there may be too much
3773 : * unused space left. We go to a higher order if more than 1/16th of the slab
3774 : * would be wasted.
3775 : *
3776 : * In order to reach satisfactory performance we must ensure that a minimum
3777 : * number of objects is in one slab. Otherwise we may generate too much
3778 : * activity on the partial lists which requires taking the list_lock. This is
3779 : * less a concern for large slabs though which are rarely used.
3780 : *
3781 : * slub_max_order specifies the order where we begin to stop considering the
3782 : * number of objects in a slab as critical. If we reach slub_max_order then
3783 : * we try to keep the page order as low as possible. So we accept more waste
3784 : * of space in favor of a small page order.
3785 : *
3786 : * Higher order allocations also allow the placement of more objects in a
3787 : * slab and thereby reduce object handling overhead. If the user has
3788 : * requested a higher minimum order then we start with that one instead of
3789 : * the smallest order which will fit the object.
3790 : */
3791 68 : static inline unsigned int calc_slab_order(unsigned int size,
3792 : unsigned int min_objects, unsigned int max_order,
3793 : unsigned int fract_leftover)
3794 : {
3795 68 : unsigned int min_order = slub_min_order;
3796 : unsigned int order;
3797 :
3798 68 : if (order_objects(min_order, size) > MAX_OBJS_PER_PAGE)
3799 0 : return get_order(size * MAX_OBJS_PER_PAGE) - 1;
3800 :
3801 207 : for (order = max(min_order, (unsigned int)get_order(min_objects * size));
3802 3 : order <= max_order; order++) {
3803 :
3804 70 : unsigned int slab_size = (unsigned int)PAGE_SIZE << order;
3805 : unsigned int rem;
3806 :
3807 70 : rem = slab_size % size;
3808 :
3809 70 : if (rem <= slab_size / fract_leftover)
3810 : break;
3811 : }
3812 :
3813 : return order;
3814 : }
3815 :
3816 67 : static inline int calculate_order(unsigned int size)
3817 : {
3818 : unsigned int order;
3819 : unsigned int min_objects;
3820 : unsigned int max_objects;
3821 : unsigned int nr_cpus;
3822 :
3823 : /*
3824 : * Attempt to find best configuration for a slab. This
3825 : * works by first attempting to generate a layout with
3826 : * the best configuration and backing off gradually.
3827 : *
3828 : * First we increase the acceptable waste in a slab. Then
3829 : * we reduce the minimum objects required in a slab.
3830 : */
3831 67 : min_objects = slub_min_objects;
3832 67 : if (!min_objects) {
3833 : /*
3834 : * Some architectures will only update present cpus when
3835 : * onlining them, so don't trust the number if it's just 1. But
3836 : * we also don't want to use nr_cpu_ids always, as on some other
3837 : * architectures, there can be many possible cpus, but never
3838 : * onlined. Here we compromise between trying to avoid too high
3839 : * order on systems that appear larger than they are, and too
3840 : * low order on systems that appear smaller than they are.
3841 : */
3842 67 : nr_cpus = num_present_cpus();
3843 : if (nr_cpus <= 1)
3844 67 : nr_cpus = nr_cpu_ids;
3845 67 : min_objects = 4 * (fls(nr_cpus) + 1);
3846 : }
3847 134 : max_objects = order_objects(slub_max_order, size);
3848 67 : min_objects = min(min_objects, max_objects);
3849 :
3850 134 : while (min_objects > 1) {
3851 : unsigned int fraction;
3852 :
3853 : fraction = 16;
3854 68 : while (fraction >= 4) {
3855 68 : order = calc_slab_order(size, min_objects,
3856 : slub_max_order, fraction);
3857 68 : if (order <= slub_max_order)
3858 67 : return order;
3859 1 : fraction /= 2;
3860 : }
3861 0 : min_objects--;
3862 : }
3863 :
3864 : /*
3865 : * We were unable to place multiple objects in a slab. Now
3866 : * lets see if we can place a single object there.
3867 : */
3868 0 : order = calc_slab_order(size, 1, slub_max_order, 1);
3869 0 : if (order <= slub_max_order)
3870 0 : return order;
3871 :
3872 : /*
3873 : * Doh this slab cannot be placed using slub_max_order.
3874 : */
3875 0 : order = calc_slab_order(size, 1, MAX_ORDER, 1);
3876 0 : if (order < MAX_ORDER)
3877 0 : return order;
3878 : return -ENOSYS;
3879 : }
3880 :
3881 : static void
3882 : init_kmem_cache_node(struct kmem_cache_node *n)
3883 : {
3884 67 : n->nr_partial = 0;
3885 67 : spin_lock_init(&n->list_lock);
3886 134 : INIT_LIST_HEAD(&n->partial);
3887 : #ifdef CONFIG_SLUB_DEBUG
3888 134 : atomic_long_set(&n->nr_slabs, 0);
3889 134 : atomic_long_set(&n->total_objects, 0);
3890 134 : INIT_LIST_HEAD(&n->full);
3891 : #endif
3892 : }
3893 :
3894 67 : static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3895 : {
3896 : BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3897 : KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3898 :
3899 : /*
3900 : * Must align to double word boundary for the double cmpxchg
3901 : * instructions to work; see __pcpu_double_call_return_bool().
3902 : */
3903 67 : s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3904 : 2 * sizeof(void *));
3905 :
3906 67 : if (!s->cpu_slab)
3907 : return 0;
3908 :
3909 : init_kmem_cache_cpus(s);
3910 :
3911 : return 1;
3912 : }
3913 :
3914 : static struct kmem_cache *kmem_cache_node;
3915 :
3916 : /*
3917 : * No kmalloc_node yet so do it by hand. We know that this is the first
3918 : * slab on the node for this slabcache. There are no concurrent accesses
3919 : * possible.
3920 : *
3921 : * Note that this function only works on the kmem_cache_node
3922 : * when allocating for the kmem_cache_node. This is used for bootstrapping
3923 : * memory on a fresh node that has no slab structures yet.
3924 : */
3925 1 : static void early_kmem_cache_node_alloc(int node)
3926 : {
3927 : struct slab *slab;
3928 : struct kmem_cache_node *n;
3929 :
3930 1 : BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3931 :
3932 1 : slab = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3933 :
3934 1 : BUG_ON(!slab);
3935 2 : if (slab_nid(slab) != node) {
3936 0 : pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3937 0 : pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3938 : }
3939 :
3940 1 : n = slab->freelist;
3941 1 : BUG_ON(!n);
3942 : #ifdef CONFIG_SLUB_DEBUG
3943 1 : init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3944 1 : init_tracking(kmem_cache_node, n);
3945 : #endif
3946 1 : n = kasan_slab_alloc(kmem_cache_node, n, GFP_KERNEL, false);
3947 2 : slab->freelist = get_freepointer(kmem_cache_node, n);
3948 1 : slab->inuse = 1;
3949 1 : slab->frozen = 0;
3950 1 : kmem_cache_node->node[node] = n;
3951 1 : init_kmem_cache_node(n);
3952 2 : inc_slabs_node(kmem_cache_node, node, slab->objects);
3953 :
3954 : /*
3955 : * No locks need to be taken here as it has just been
3956 : * initialized and there is no concurrent access.
3957 : */
3958 1 : __add_partial(n, slab, DEACTIVATE_TO_HEAD);
3959 1 : }
3960 :
3961 0 : static void free_kmem_cache_nodes(struct kmem_cache *s)
3962 : {
3963 : int node;
3964 : struct kmem_cache_node *n;
3965 :
3966 0 : for_each_kmem_cache_node(s, node, n) {
3967 0 : s->node[node] = NULL;
3968 0 : kmem_cache_free(kmem_cache_node, n);
3969 : }
3970 0 : }
3971 :
3972 0 : void __kmem_cache_release(struct kmem_cache *s)
3973 : {
3974 0 : cache_random_seq_destroy(s);
3975 0 : free_percpu(s->cpu_slab);
3976 0 : free_kmem_cache_nodes(s);
3977 0 : }
3978 :
3979 67 : static int init_kmem_cache_nodes(struct kmem_cache *s)
3980 : {
3981 : int node;
3982 :
3983 134 : for_each_node_mask(node, slab_nodes) {
3984 : struct kmem_cache_node *n;
3985 :
3986 67 : if (slab_state == DOWN) {
3987 1 : early_kmem_cache_node_alloc(node);
3988 1 : continue;
3989 : }
3990 132 : n = kmem_cache_alloc_node(kmem_cache_node,
3991 : GFP_KERNEL, node);
3992 :
3993 66 : if (!n) {
3994 0 : free_kmem_cache_nodes(s);
3995 0 : return 0;
3996 : }
3997 :
3998 66 : init_kmem_cache_node(n);
3999 66 : s->node[node] = n;
4000 : }
4001 : return 1;
4002 : }
4003 :
4004 : static void set_cpu_partial(struct kmem_cache *s)
4005 : {
4006 : #ifdef CONFIG_SLUB_CPU_PARTIAL
4007 : unsigned int nr_objects;
4008 :
4009 : /*
4010 : * cpu_partial determined the maximum number of objects kept in the
4011 : * per cpu partial lists of a processor.
4012 : *
4013 : * Per cpu partial lists mainly contain slabs that just have one
4014 : * object freed. If they are used for allocation then they can be
4015 : * filled up again with minimal effort. The slab will never hit the
4016 : * per node partial lists and therefore no locking will be required.
4017 : *
4018 : * For backwards compatibility reasons, this is determined as number
4019 : * of objects, even though we now limit maximum number of pages, see
4020 : * slub_set_cpu_partial()
4021 : */
4022 : if (!kmem_cache_has_cpu_partial(s))
4023 : nr_objects = 0;
4024 : else if (s->size >= PAGE_SIZE)
4025 : nr_objects = 6;
4026 : else if (s->size >= 1024)
4027 : nr_objects = 24;
4028 : else if (s->size >= 256)
4029 : nr_objects = 52;
4030 : else
4031 : nr_objects = 120;
4032 :
4033 : slub_set_cpu_partial(s, nr_objects);
4034 : #endif
4035 : }
4036 :
4037 : /*
4038 : * calculate_sizes() determines the order and the distribution of data within
4039 : * a slab object.
4040 : */
4041 67 : static int calculate_sizes(struct kmem_cache *s)
4042 : {
4043 67 : slab_flags_t flags = s->flags;
4044 67 : unsigned int size = s->object_size;
4045 : unsigned int order;
4046 :
4047 : /*
4048 : * Round up object size to the next word boundary. We can only
4049 : * place the free pointer at word boundaries and this determines
4050 : * the possible location of the free pointer.
4051 : */
4052 67 : size = ALIGN(size, sizeof(void *));
4053 :
4054 : #ifdef CONFIG_SLUB_DEBUG
4055 : /*
4056 : * Determine if we can poison the object itself. If the user of
4057 : * the slab may touch the object after free or before allocation
4058 : * then we should never poison the object itself.
4059 : */
4060 67 : if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) &&
4061 0 : !s->ctor)
4062 0 : s->flags |= __OBJECT_POISON;
4063 : else
4064 67 : s->flags &= ~__OBJECT_POISON;
4065 :
4066 :
4067 : /*
4068 : * If we are Redzoning then check if there is some space between the
4069 : * end of the object and the free pointer. If not then add an
4070 : * additional word to have some bytes to store Redzone information.
4071 : */
4072 67 : if ((flags & SLAB_RED_ZONE) && size == s->object_size)
4073 0 : size += sizeof(void *);
4074 : #endif
4075 :
4076 : /*
4077 : * With that we have determined the number of bytes in actual use
4078 : * by the object and redzoning.
4079 : */
4080 67 : s->inuse = size;
4081 :
4082 67 : if ((flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) ||
4083 63 : ((flags & SLAB_RED_ZONE) && s->object_size < sizeof(void *)) ||
4084 63 : s->ctor) {
4085 : /*
4086 : * Relocate free pointer after the object if it is not
4087 : * permitted to overwrite the first word of the object on
4088 : * kmem_cache_free.
4089 : *
4090 : * This is the case if we do RCU, have a constructor or
4091 : * destructor, are poisoning the objects, or are
4092 : * redzoning an object smaller than sizeof(void *).
4093 : *
4094 : * The assumption that s->offset >= s->inuse means free
4095 : * pointer is outside of the object is used in the
4096 : * freeptr_outside_object() function. If that is no
4097 : * longer true, the function needs to be modified.
4098 : */
4099 9 : s->offset = size;
4100 9 : size += sizeof(void *);
4101 : } else {
4102 : /*
4103 : * Store freelist pointer near middle of object to keep
4104 : * it away from the edges of the object to avoid small
4105 : * sized over/underflows from neighboring allocations.
4106 : */
4107 58 : s->offset = ALIGN_DOWN(s->object_size / 2, sizeof(void *));
4108 : }
4109 :
4110 : #ifdef CONFIG_SLUB_DEBUG
4111 67 : if (flags & SLAB_STORE_USER)
4112 : /*
4113 : * Need to store information about allocs and frees after
4114 : * the object.
4115 : */
4116 0 : size += 2 * sizeof(struct track);
4117 : #endif
4118 :
4119 67 : kasan_cache_create(s, &size, &s->flags);
4120 : #ifdef CONFIG_SLUB_DEBUG
4121 67 : if (flags & SLAB_RED_ZONE) {
4122 : /*
4123 : * Add some empty padding so that we can catch
4124 : * overwrites from earlier objects rather than let
4125 : * tracking information or the free pointer be
4126 : * corrupted if a user writes before the start
4127 : * of the object.
4128 : */
4129 0 : size += sizeof(void *);
4130 :
4131 : s->red_left_pad = sizeof(void *);
4132 0 : s->red_left_pad = ALIGN(s->red_left_pad, s->align);
4133 0 : size += s->red_left_pad;
4134 : }
4135 : #endif
4136 :
4137 : /*
4138 : * SLUB stores one object immediately after another beginning from
4139 : * offset 0. In order to align the objects we have to simply size
4140 : * each object to conform to the alignment.
4141 : */
4142 67 : size = ALIGN(size, s->align);
4143 67 : s->size = size;
4144 67 : s->reciprocal_size = reciprocal_value(size);
4145 67 : order = calculate_order(size);
4146 :
4147 67 : if ((int)order < 0)
4148 : return 0;
4149 :
4150 67 : s->allocflags = 0;
4151 67 : if (order)
4152 25 : s->allocflags |= __GFP_COMP;
4153 :
4154 67 : if (s->flags & SLAB_CACHE_DMA)
4155 0 : s->allocflags |= GFP_DMA;
4156 :
4157 67 : if (s->flags & SLAB_CACHE_DMA32)
4158 0 : s->allocflags |= GFP_DMA32;
4159 :
4160 67 : if (s->flags & SLAB_RECLAIM_ACCOUNT)
4161 19 : s->allocflags |= __GFP_RECLAIMABLE;
4162 :
4163 : /*
4164 : * Determine the number of objects per slab
4165 : */
4166 134 : s->oo = oo_make(order, size);
4167 201 : s->min = oo_make(get_order(size), size);
4168 67 : if (oo_objects(s->oo) > oo_objects(s->max))
4169 67 : s->max = s->oo;
4170 :
4171 67 : return !!oo_objects(s->oo);
4172 : }
4173 :
4174 67 : static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags)
4175 : {
4176 67 : s->flags = kmem_cache_flags(s->size, flags, s->name);
4177 : #ifdef CONFIG_SLAB_FREELIST_HARDENED
4178 : s->random = get_random_long();
4179 : #endif
4180 :
4181 67 : if (!calculate_sizes(s))
4182 : goto error;
4183 67 : if (disable_higher_order_debug) {
4184 : /*
4185 : * Disable debugging flags that store metadata if the min slab
4186 : * order increased.
4187 : */
4188 0 : if (get_order(s->size) > get_order(s->object_size)) {
4189 0 : s->flags &= ~DEBUG_METADATA_FLAGS;
4190 0 : s->offset = 0;
4191 0 : if (!calculate_sizes(s))
4192 : goto error;
4193 : }
4194 : }
4195 :
4196 : #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
4197 : defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
4198 : if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0)
4199 : /* Enable fast mode */
4200 : s->flags |= __CMPXCHG_DOUBLE;
4201 : #endif
4202 :
4203 : /*
4204 : * The larger the object size is, the more slabs we want on the partial
4205 : * list to avoid pounding the page allocator excessively.
4206 : */
4207 134 : s->min_partial = min_t(unsigned long, MAX_PARTIAL, ilog2(s->size) / 2);
4208 67 : s->min_partial = max_t(unsigned long, MIN_PARTIAL, s->min_partial);
4209 :
4210 67 : set_cpu_partial(s);
4211 :
4212 : #ifdef CONFIG_NUMA
4213 : s->remote_node_defrag_ratio = 1000;
4214 : #endif
4215 :
4216 : /* Initialize the pre-computed randomized freelist if slab is up */
4217 : if (slab_state >= UP) {
4218 : if (init_cache_random_seq(s))
4219 : goto error;
4220 : }
4221 :
4222 67 : if (!init_kmem_cache_nodes(s))
4223 : goto error;
4224 :
4225 67 : if (alloc_kmem_cache_cpus(s))
4226 : return 0;
4227 :
4228 : error:
4229 0 : __kmem_cache_release(s);
4230 0 : return -EINVAL;
4231 : }
4232 :
4233 0 : static void list_slab_objects(struct kmem_cache *s, struct slab *slab,
4234 : const char *text)
4235 : {
4236 : #ifdef CONFIG_SLUB_DEBUG
4237 0 : void *addr = slab_address(slab);
4238 : unsigned long flags;
4239 : unsigned long *map;
4240 : void *p;
4241 :
4242 0 : slab_err(s, slab, text, s->name);
4243 0 : slab_lock(slab, &flags);
4244 :
4245 0 : map = get_map(s, slab);
4246 0 : for_each_object(p, s, addr, slab->objects) {
4247 :
4248 0 : if (!test_bit(__obj_to_index(s, addr, p), map)) {
4249 0 : pr_err("Object 0x%p @offset=%tu\n", p, p - addr);
4250 0 : print_tracking(s, p);
4251 : }
4252 : }
4253 0 : put_map(map);
4254 0 : slab_unlock(slab, &flags);
4255 : #endif
4256 0 : }
4257 :
4258 : /*
4259 : * Attempt to free all partial slabs on a node.
4260 : * This is called from __kmem_cache_shutdown(). We must take list_lock
4261 : * because sysfs file might still access partial list after the shutdowning.
4262 : */
4263 0 : static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
4264 : {
4265 0 : LIST_HEAD(discard);
4266 : struct slab *slab, *h;
4267 :
4268 0 : BUG_ON(irqs_disabled());
4269 0 : spin_lock_irq(&n->list_lock);
4270 0 : list_for_each_entry_safe(slab, h, &n->partial, slab_list) {
4271 0 : if (!slab->inuse) {
4272 0 : remove_partial(n, slab);
4273 0 : list_add(&slab->slab_list, &discard);
4274 : } else {
4275 0 : list_slab_objects(s, slab,
4276 : "Objects remaining in %s on __kmem_cache_shutdown()");
4277 : }
4278 : }
4279 0 : spin_unlock_irq(&n->list_lock);
4280 :
4281 0 : list_for_each_entry_safe(slab, h, &discard, slab_list)
4282 0 : discard_slab(s, slab);
4283 0 : }
4284 :
4285 0 : bool __kmem_cache_empty(struct kmem_cache *s)
4286 : {
4287 : int node;
4288 : struct kmem_cache_node *n;
4289 :
4290 0 : for_each_kmem_cache_node(s, node, n)
4291 0 : if (n->nr_partial || slabs_node(s, node))
4292 : return false;
4293 : return true;
4294 : }
4295 :
4296 : /*
4297 : * Release all resources used by a slab cache.
4298 : */
4299 0 : int __kmem_cache_shutdown(struct kmem_cache *s)
4300 : {
4301 : int node;
4302 : struct kmem_cache_node *n;
4303 :
4304 0 : flush_all_cpus_locked(s);
4305 : /* Attempt to free all objects */
4306 0 : for_each_kmem_cache_node(s, node, n) {
4307 0 : free_partial(s, n);
4308 0 : if (n->nr_partial || slabs_node(s, node))
4309 : return 1;
4310 : }
4311 : return 0;
4312 : }
4313 :
4314 : #ifdef CONFIG_PRINTK
4315 0 : void __kmem_obj_info(struct kmem_obj_info *kpp, void *object, struct slab *slab)
4316 : {
4317 : void *base;
4318 : int __maybe_unused i;
4319 : unsigned int objnr;
4320 : void *objp;
4321 : void *objp0;
4322 0 : struct kmem_cache *s = slab->slab_cache;
4323 : struct track __maybe_unused *trackp;
4324 :
4325 0 : kpp->kp_ptr = object;
4326 0 : kpp->kp_slab = slab;
4327 0 : kpp->kp_slab_cache = s;
4328 0 : base = slab_address(slab);
4329 0 : objp0 = kasan_reset_tag(object);
4330 : #ifdef CONFIG_SLUB_DEBUG
4331 0 : objp = restore_red_left(s, objp0);
4332 : #else
4333 : objp = objp0;
4334 : #endif
4335 0 : objnr = obj_to_index(s, slab, objp);
4336 0 : kpp->kp_data_offset = (unsigned long)((char *)objp0 - (char *)objp);
4337 0 : objp = base + s->size * objnr;
4338 0 : kpp->kp_objp = objp;
4339 0 : if (WARN_ON_ONCE(objp < base || objp >= base + slab->objects * s->size
4340 0 : || (objp - base) % s->size) ||
4341 0 : !(s->flags & SLAB_STORE_USER))
4342 : return;
4343 : #ifdef CONFIG_SLUB_DEBUG
4344 0 : objp = fixup_red_left(s, objp);
4345 0 : trackp = get_track(s, objp, TRACK_ALLOC);
4346 0 : kpp->kp_ret = (void *)trackp->addr;
4347 : #ifdef CONFIG_STACKTRACE
4348 0 : for (i = 0; i < KS_ADDRS_COUNT && i < TRACK_ADDRS_COUNT; i++) {
4349 0 : kpp->kp_stack[i] = (void *)trackp->addrs[i];
4350 0 : if (!kpp->kp_stack[i])
4351 : break;
4352 : }
4353 :
4354 0 : trackp = get_track(s, objp, TRACK_FREE);
4355 0 : for (i = 0; i < KS_ADDRS_COUNT && i < TRACK_ADDRS_COUNT; i++) {
4356 0 : kpp->kp_free_stack[i] = (void *)trackp->addrs[i];
4357 0 : if (!kpp->kp_free_stack[i])
4358 : break;
4359 : }
4360 : #endif
4361 : #endif
4362 : }
4363 : #endif
4364 :
4365 : /********************************************************************
4366 : * Kmalloc subsystem
4367 : *******************************************************************/
4368 :
4369 0 : static int __init setup_slub_min_order(char *str)
4370 : {
4371 0 : get_option(&str, (int *)&slub_min_order);
4372 :
4373 0 : return 1;
4374 : }
4375 :
4376 : __setup("slub_min_order=", setup_slub_min_order);
4377 :
4378 0 : static int __init setup_slub_max_order(char *str)
4379 : {
4380 0 : get_option(&str, (int *)&slub_max_order);
4381 0 : slub_max_order = min(slub_max_order, (unsigned int)MAX_ORDER - 1);
4382 :
4383 0 : return 1;
4384 : }
4385 :
4386 : __setup("slub_max_order=", setup_slub_max_order);
4387 :
4388 0 : static int __init setup_slub_min_objects(char *str)
4389 : {
4390 0 : get_option(&str, (int *)&slub_min_objects);
4391 :
4392 0 : return 1;
4393 : }
4394 :
4395 : __setup("slub_min_objects=", setup_slub_min_objects);
4396 :
4397 1011 : void *__kmalloc(size_t size, gfp_t flags)
4398 : {
4399 : struct kmem_cache *s;
4400 : void *ret;
4401 :
4402 1011 : if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4403 8 : return kmalloc_large(size, flags);
4404 :
4405 1003 : s = kmalloc_slab(size, flags);
4406 :
4407 1003 : if (unlikely(ZERO_OR_NULL_PTR(s)))
4408 : return s;
4409 :
4410 2006 : ret = slab_alloc(s, NULL, flags, _RET_IP_, size);
4411 :
4412 1003 : trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
4413 :
4414 1003 : ret = kasan_kmalloc(s, ret, size, flags);
4415 :
4416 1003 : return ret;
4417 : }
4418 : EXPORT_SYMBOL(__kmalloc);
4419 :
4420 : #ifdef CONFIG_NUMA
4421 : static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
4422 : {
4423 : struct page *page;
4424 : void *ptr = NULL;
4425 : unsigned int order = get_order(size);
4426 :
4427 : flags |= __GFP_COMP;
4428 : page = alloc_pages_node(node, flags, order);
4429 : if (page) {
4430 : ptr = page_address(page);
4431 : mod_lruvec_page_state(page, NR_SLAB_UNRECLAIMABLE_B,
4432 : PAGE_SIZE << order);
4433 : }
4434 :
4435 : return kmalloc_large_node_hook(ptr, size, flags);
4436 : }
4437 :
4438 : void *__kmalloc_node(size_t size, gfp_t flags, int node)
4439 : {
4440 : struct kmem_cache *s;
4441 : void *ret;
4442 :
4443 : if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4444 : ret = kmalloc_large_node(size, flags, node);
4445 :
4446 : trace_kmalloc_node(_RET_IP_, ret,
4447 : size, PAGE_SIZE << get_order(size),
4448 : flags, node);
4449 :
4450 : return ret;
4451 : }
4452 :
4453 : s = kmalloc_slab(size, flags);
4454 :
4455 : if (unlikely(ZERO_OR_NULL_PTR(s)))
4456 : return s;
4457 :
4458 : ret = slab_alloc_node(s, NULL, flags, node, _RET_IP_, size);
4459 :
4460 : trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
4461 :
4462 : ret = kasan_kmalloc(s, ret, size, flags);
4463 :
4464 : return ret;
4465 : }
4466 : EXPORT_SYMBOL(__kmalloc_node);
4467 : #endif /* CONFIG_NUMA */
4468 :
4469 : #ifdef CONFIG_HARDENED_USERCOPY
4470 : /*
4471 : * Rejects incorrectly sized objects and objects that are to be copied
4472 : * to/from userspace but do not fall entirely within the containing slab
4473 : * cache's usercopy region.
4474 : *
4475 : * Returns NULL if check passes, otherwise const char * to name of cache
4476 : * to indicate an error.
4477 : */
4478 : void __check_heap_object(const void *ptr, unsigned long n,
4479 : const struct slab *slab, bool to_user)
4480 : {
4481 : struct kmem_cache *s;
4482 : unsigned int offset;
4483 : bool is_kfence = is_kfence_address(ptr);
4484 :
4485 : ptr = kasan_reset_tag(ptr);
4486 :
4487 : /* Find object and usable object size. */
4488 : s = slab->slab_cache;
4489 :
4490 : /* Reject impossible pointers. */
4491 : if (ptr < slab_address(slab))
4492 : usercopy_abort("SLUB object not in SLUB page?!", NULL,
4493 : to_user, 0, n);
4494 :
4495 : /* Find offset within object. */
4496 : if (is_kfence)
4497 : offset = ptr - kfence_object_start(ptr);
4498 : else
4499 : offset = (ptr - slab_address(slab)) % s->size;
4500 :
4501 : /* Adjust for redzone and reject if within the redzone. */
4502 : if (!is_kfence && kmem_cache_debug_flags(s, SLAB_RED_ZONE)) {
4503 : if (offset < s->red_left_pad)
4504 : usercopy_abort("SLUB object in left red zone",
4505 : s->name, to_user, offset, n);
4506 : offset -= s->red_left_pad;
4507 : }
4508 :
4509 : /* Allow address range falling entirely within usercopy region. */
4510 : if (offset >= s->useroffset &&
4511 : offset - s->useroffset <= s->usersize &&
4512 : n <= s->useroffset - offset + s->usersize)
4513 : return;
4514 :
4515 : usercopy_abort("SLUB object", s->name, to_user, offset, n);
4516 : }
4517 : #endif /* CONFIG_HARDENED_USERCOPY */
4518 :
4519 228 : size_t __ksize(const void *object)
4520 : {
4521 : struct folio *folio;
4522 :
4523 228 : if (unlikely(object == ZERO_SIZE_PTR))
4524 : return 0;
4525 :
4526 228 : folio = virt_to_folio(object);
4527 :
4528 228 : if (unlikely(!folio_test_slab(folio)))
4529 0 : return folio_size(folio);
4530 :
4531 228 : return slab_ksize(folio_slab(folio)->slab_cache);
4532 : }
4533 : EXPORT_SYMBOL(__ksize);
4534 :
4535 3807 : void kfree(const void *x)
4536 : {
4537 : struct folio *folio;
4538 : struct slab *slab;
4539 3807 : void *object = (void *)x;
4540 :
4541 3807 : trace_kfree(_RET_IP_, x);
4542 :
4543 3807 : if (unlikely(ZERO_OR_NULL_PTR(x)))
4544 : return;
4545 :
4546 2350 : folio = virt_to_folio(x);
4547 2350 : if (unlikely(!folio_test_slab(folio))) {
4548 8 : free_large_kmalloc(folio, object);
4549 8 : return;
4550 : }
4551 2342 : slab = folio_slab(folio);
4552 2342 : slab_free(slab->slab_cache, slab, object, NULL, 1, _RET_IP_);
4553 : }
4554 : EXPORT_SYMBOL(kfree);
4555 :
4556 : #define SHRINK_PROMOTE_MAX 32
4557 :
4558 : /*
4559 : * kmem_cache_shrink discards empty slabs and promotes the slabs filled
4560 : * up most to the head of the partial lists. New allocations will then
4561 : * fill those up and thus they can be removed from the partial lists.
4562 : *
4563 : * The slabs with the least items are placed last. This results in them
4564 : * being allocated from last increasing the chance that the last objects
4565 : * are freed in them.
4566 : */
4567 0 : static int __kmem_cache_do_shrink(struct kmem_cache *s)
4568 : {
4569 : int node;
4570 : int i;
4571 : struct kmem_cache_node *n;
4572 : struct slab *slab;
4573 : struct slab *t;
4574 : struct list_head discard;
4575 : struct list_head promote[SHRINK_PROMOTE_MAX];
4576 : unsigned long flags;
4577 0 : int ret = 0;
4578 :
4579 0 : for_each_kmem_cache_node(s, node, n) {
4580 0 : INIT_LIST_HEAD(&discard);
4581 0 : for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
4582 0 : INIT_LIST_HEAD(promote + i);
4583 :
4584 0 : spin_lock_irqsave(&n->list_lock, flags);
4585 :
4586 : /*
4587 : * Build lists of slabs to discard or promote.
4588 : *
4589 : * Note that concurrent frees may occur while we hold the
4590 : * list_lock. slab->inuse here is the upper limit.
4591 : */
4592 0 : list_for_each_entry_safe(slab, t, &n->partial, slab_list) {
4593 0 : int free = slab->objects - slab->inuse;
4594 :
4595 : /* Do not reread slab->inuse */
4596 0 : barrier();
4597 :
4598 : /* We do not keep full slabs on the list */
4599 0 : BUG_ON(free <= 0);
4600 :
4601 0 : if (free == slab->objects) {
4602 0 : list_move(&slab->slab_list, &discard);
4603 0 : n->nr_partial--;
4604 0 : } else if (free <= SHRINK_PROMOTE_MAX)
4605 0 : list_move(&slab->slab_list, promote + free - 1);
4606 : }
4607 :
4608 : /*
4609 : * Promote the slabs filled up most to the head of the
4610 : * partial list.
4611 : */
4612 0 : for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
4613 0 : list_splice(promote + i, &n->partial);
4614 :
4615 0 : spin_unlock_irqrestore(&n->list_lock, flags);
4616 :
4617 : /* Release empty slabs */
4618 0 : list_for_each_entry_safe(slab, t, &discard, slab_list)
4619 0 : discard_slab(s, slab);
4620 :
4621 0 : if (slabs_node(s, node))
4622 0 : ret = 1;
4623 : }
4624 :
4625 0 : return ret;
4626 : }
4627 :
4628 0 : int __kmem_cache_shrink(struct kmem_cache *s)
4629 : {
4630 0 : flush_all(s);
4631 0 : return __kmem_cache_do_shrink(s);
4632 : }
4633 :
4634 : static int slab_mem_going_offline_callback(void *arg)
4635 : {
4636 : struct kmem_cache *s;
4637 :
4638 : mutex_lock(&slab_mutex);
4639 : list_for_each_entry(s, &slab_caches, list) {
4640 : flush_all_cpus_locked(s);
4641 : __kmem_cache_do_shrink(s);
4642 : }
4643 : mutex_unlock(&slab_mutex);
4644 :
4645 : return 0;
4646 : }
4647 :
4648 : static void slab_mem_offline_callback(void *arg)
4649 : {
4650 : struct memory_notify *marg = arg;
4651 : int offline_node;
4652 :
4653 : offline_node = marg->status_change_nid_normal;
4654 :
4655 : /*
4656 : * If the node still has available memory. we need kmem_cache_node
4657 : * for it yet.
4658 : */
4659 : if (offline_node < 0)
4660 : return;
4661 :
4662 : mutex_lock(&slab_mutex);
4663 : node_clear(offline_node, slab_nodes);
4664 : /*
4665 : * We no longer free kmem_cache_node structures here, as it would be
4666 : * racy with all get_node() users, and infeasible to protect them with
4667 : * slab_mutex.
4668 : */
4669 : mutex_unlock(&slab_mutex);
4670 : }
4671 :
4672 : static int slab_mem_going_online_callback(void *arg)
4673 : {
4674 : struct kmem_cache_node *n;
4675 : struct kmem_cache *s;
4676 : struct memory_notify *marg = arg;
4677 : int nid = marg->status_change_nid_normal;
4678 : int ret = 0;
4679 :
4680 : /*
4681 : * If the node's memory is already available, then kmem_cache_node is
4682 : * already created. Nothing to do.
4683 : */
4684 : if (nid < 0)
4685 : return 0;
4686 :
4687 : /*
4688 : * We are bringing a node online. No memory is available yet. We must
4689 : * allocate a kmem_cache_node structure in order to bring the node
4690 : * online.
4691 : */
4692 : mutex_lock(&slab_mutex);
4693 : list_for_each_entry(s, &slab_caches, list) {
4694 : /*
4695 : * The structure may already exist if the node was previously
4696 : * onlined and offlined.
4697 : */
4698 : if (get_node(s, nid))
4699 : continue;
4700 : /*
4701 : * XXX: kmem_cache_alloc_node will fallback to other nodes
4702 : * since memory is not yet available from the node that
4703 : * is brought up.
4704 : */
4705 : n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
4706 : if (!n) {
4707 : ret = -ENOMEM;
4708 : goto out;
4709 : }
4710 : init_kmem_cache_node(n);
4711 : s->node[nid] = n;
4712 : }
4713 : /*
4714 : * Any cache created after this point will also have kmem_cache_node
4715 : * initialized for the new node.
4716 : */
4717 : node_set(nid, slab_nodes);
4718 : out:
4719 : mutex_unlock(&slab_mutex);
4720 : return ret;
4721 : }
4722 :
4723 : static int slab_memory_callback(struct notifier_block *self,
4724 : unsigned long action, void *arg)
4725 : {
4726 : int ret = 0;
4727 :
4728 : switch (action) {
4729 : case MEM_GOING_ONLINE:
4730 : ret = slab_mem_going_online_callback(arg);
4731 : break;
4732 : case MEM_GOING_OFFLINE:
4733 : ret = slab_mem_going_offline_callback(arg);
4734 : break;
4735 : case MEM_OFFLINE:
4736 : case MEM_CANCEL_ONLINE:
4737 : slab_mem_offline_callback(arg);
4738 : break;
4739 : case MEM_ONLINE:
4740 : case MEM_CANCEL_OFFLINE:
4741 : break;
4742 : }
4743 : if (ret)
4744 : ret = notifier_from_errno(ret);
4745 : else
4746 : ret = NOTIFY_OK;
4747 : return ret;
4748 : }
4749 :
4750 : static struct notifier_block slab_memory_callback_nb = {
4751 : .notifier_call = slab_memory_callback,
4752 : .priority = SLAB_CALLBACK_PRI,
4753 : };
4754 :
4755 : /********************************************************************
4756 : * Basic setup of slabs
4757 : *******************************************************************/
4758 :
4759 : /*
4760 : * Used for early kmem_cache structures that were allocated using
4761 : * the page allocator. Allocate them properly then fix up the pointers
4762 : * that may be pointing to the wrong kmem_cache structure.
4763 : */
4764 :
4765 2 : static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
4766 : {
4767 : int node;
4768 4 : struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
4769 : struct kmem_cache_node *n;
4770 :
4771 2 : memcpy(s, static_cache, kmem_cache->object_size);
4772 :
4773 : /*
4774 : * This runs very early, and only the boot processor is supposed to be
4775 : * up. Even if it weren't true, IRQs are not up so we couldn't fire
4776 : * IPIs around.
4777 : */
4778 2 : __flush_cpu_slab(s, smp_processor_id());
4779 6 : for_each_kmem_cache_node(s, node, n) {
4780 : struct slab *p;
4781 :
4782 4 : list_for_each_entry(p, &n->partial, slab_list)
4783 2 : p->slab_cache = s;
4784 :
4785 : #ifdef CONFIG_SLUB_DEBUG
4786 2 : list_for_each_entry(p, &n->full, slab_list)
4787 0 : p->slab_cache = s;
4788 : #endif
4789 : }
4790 4 : list_add(&s->list, &slab_caches);
4791 2 : return s;
4792 : }
4793 :
4794 1 : void __init kmem_cache_init(void)
4795 : {
4796 : static __initdata struct kmem_cache boot_kmem_cache,
4797 : boot_kmem_cache_node;
4798 : int node;
4799 :
4800 : if (debug_guardpage_minorder())
4801 : slub_max_order = 0;
4802 :
4803 : /* Print slub debugging pointers without hashing */
4804 1 : if (__slub_debug_enabled())
4805 0 : no_hash_pointers_enable(NULL);
4806 :
4807 1 : kmem_cache_node = &boot_kmem_cache_node;
4808 1 : kmem_cache = &boot_kmem_cache;
4809 :
4810 : /*
4811 : * Initialize the nodemask for which we will allocate per node
4812 : * structures. Here we don't need taking slab_mutex yet.
4813 : */
4814 3 : for_each_node_state(node, N_NORMAL_MEMORY)
4815 : node_set(node, slab_nodes);
4816 :
4817 1 : create_boot_cache(kmem_cache_node, "kmem_cache_node",
4818 : sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN, 0, 0);
4819 :
4820 : register_hotmemory_notifier(&slab_memory_callback_nb);
4821 :
4822 : /* Able to allocate the per node structures */
4823 1 : slab_state = PARTIAL;
4824 :
4825 1 : create_boot_cache(kmem_cache, "kmem_cache",
4826 : offsetof(struct kmem_cache, node) +
4827 : nr_node_ids * sizeof(struct kmem_cache_node *),
4828 : SLAB_HWCACHE_ALIGN, 0, 0);
4829 :
4830 1 : kmem_cache = bootstrap(&boot_kmem_cache);
4831 1 : kmem_cache_node = bootstrap(&boot_kmem_cache_node);
4832 :
4833 : /* Now we can use the kmem_cache to allocate kmalloc slabs */
4834 1 : setup_kmalloc_cache_index_table();
4835 1 : create_kmalloc_caches(0);
4836 :
4837 : /* Setup random freelists for each cache */
4838 1 : init_freelist_randomization();
4839 :
4840 1 : cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL,
4841 : slub_cpu_dead);
4842 :
4843 1 : pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n",
4844 : cache_line_size(),
4845 : slub_min_order, slub_max_order, slub_min_objects,
4846 : nr_cpu_ids, nr_node_ids);
4847 1 : }
4848 :
4849 1 : void __init kmem_cache_init_late(void)
4850 : {
4851 1 : }
4852 :
4853 : struct kmem_cache *
4854 54 : __kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
4855 : slab_flags_t flags, void (*ctor)(void *))
4856 : {
4857 : struct kmem_cache *s;
4858 :
4859 54 : s = find_mergeable(size, align, flags, name, ctor);
4860 54 : if (s) {
4861 21 : s->refcount++;
4862 :
4863 : /*
4864 : * Adjust the object sizes so that we clear
4865 : * the complete object on kzalloc.
4866 : */
4867 21 : s->object_size = max(s->object_size, size);
4868 21 : s->inuse = max(s->inuse, ALIGN(size, sizeof(void *)));
4869 :
4870 21 : if (sysfs_slab_alias(s, name)) {
4871 0 : s->refcount--;
4872 0 : s = NULL;
4873 : }
4874 : }
4875 :
4876 54 : return s;
4877 : }
4878 :
4879 67 : int __kmem_cache_create(struct kmem_cache *s, slab_flags_t flags)
4880 : {
4881 : int err;
4882 :
4883 67 : err = kmem_cache_open(s, flags);
4884 67 : if (err)
4885 : return err;
4886 :
4887 : /* Mutex is not taken during early boot */
4888 67 : if (slab_state <= UP)
4889 : return 0;
4890 :
4891 3 : err = sysfs_slab_add(s);
4892 3 : if (err) {
4893 0 : __kmem_cache_release(s);
4894 0 : return err;
4895 : }
4896 :
4897 : if (s->flags & SLAB_STORE_USER)
4898 : debugfs_slab_add(s);
4899 :
4900 : return 0;
4901 : }
4902 :
4903 2395 : void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4904 : {
4905 : struct kmem_cache *s;
4906 : void *ret;
4907 :
4908 2395 : if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4909 0 : return kmalloc_large(size, gfpflags);
4910 :
4911 2395 : s = kmalloc_slab(size, gfpflags);
4912 :
4913 2395 : if (unlikely(ZERO_OR_NULL_PTR(s)))
4914 : return s;
4915 :
4916 2395 : ret = slab_alloc(s, NULL, gfpflags, caller, size);
4917 :
4918 : /* Honor the call site pointer we received. */
4919 2395 : trace_kmalloc(caller, ret, size, s->size, gfpflags);
4920 :
4921 2395 : return ret;
4922 : }
4923 : EXPORT_SYMBOL(__kmalloc_track_caller);
4924 :
4925 : #ifdef CONFIG_NUMA
4926 : void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4927 : int node, unsigned long caller)
4928 : {
4929 : struct kmem_cache *s;
4930 : void *ret;
4931 :
4932 : if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4933 : ret = kmalloc_large_node(size, gfpflags, node);
4934 :
4935 : trace_kmalloc_node(caller, ret,
4936 : size, PAGE_SIZE << get_order(size),
4937 : gfpflags, node);
4938 :
4939 : return ret;
4940 : }
4941 :
4942 : s = kmalloc_slab(size, gfpflags);
4943 :
4944 : if (unlikely(ZERO_OR_NULL_PTR(s)))
4945 : return s;
4946 :
4947 : ret = slab_alloc_node(s, NULL, gfpflags, node, caller, size);
4948 :
4949 : /* Honor the call site pointer we received. */
4950 : trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4951 :
4952 : return ret;
4953 : }
4954 : EXPORT_SYMBOL(__kmalloc_node_track_caller);
4955 : #endif
4956 :
4957 : #ifdef CONFIG_SYSFS
4958 0 : static int count_inuse(struct slab *slab)
4959 : {
4960 0 : return slab->inuse;
4961 : }
4962 :
4963 0 : static int count_total(struct slab *slab)
4964 : {
4965 0 : return slab->objects;
4966 : }
4967 : #endif
4968 :
4969 : #ifdef CONFIG_SLUB_DEBUG
4970 0 : static void validate_slab(struct kmem_cache *s, struct slab *slab,
4971 : unsigned long *obj_map)
4972 : {
4973 : void *p;
4974 0 : void *addr = slab_address(slab);
4975 : unsigned long flags;
4976 :
4977 0 : slab_lock(slab, &flags);
4978 :
4979 0 : if (!check_slab(s, slab) || !on_freelist(s, slab, NULL))
4980 : goto unlock;
4981 :
4982 : /* Now we know that a valid freelist exists */
4983 0 : __fill_map(obj_map, s, slab);
4984 0 : for_each_object(p, s, addr, slab->objects) {
4985 0 : u8 val = test_bit(__obj_to_index(s, addr, p), obj_map) ?
4986 : SLUB_RED_INACTIVE : SLUB_RED_ACTIVE;
4987 :
4988 0 : if (!check_object(s, slab, p, val))
4989 : break;
4990 : }
4991 : unlock:
4992 0 : slab_unlock(slab, &flags);
4993 0 : }
4994 :
4995 0 : static int validate_slab_node(struct kmem_cache *s,
4996 : struct kmem_cache_node *n, unsigned long *obj_map)
4997 : {
4998 0 : unsigned long count = 0;
4999 : struct slab *slab;
5000 : unsigned long flags;
5001 :
5002 0 : spin_lock_irqsave(&n->list_lock, flags);
5003 :
5004 0 : list_for_each_entry(slab, &n->partial, slab_list) {
5005 0 : validate_slab(s, slab, obj_map);
5006 0 : count++;
5007 : }
5008 0 : if (count != n->nr_partial) {
5009 0 : pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
5010 : s->name, count, n->nr_partial);
5011 0 : slab_add_kunit_errors();
5012 : }
5013 :
5014 0 : if (!(s->flags & SLAB_STORE_USER))
5015 : goto out;
5016 :
5017 0 : list_for_each_entry(slab, &n->full, slab_list) {
5018 0 : validate_slab(s, slab, obj_map);
5019 0 : count++;
5020 : }
5021 0 : if (count != atomic_long_read(&n->nr_slabs)) {
5022 0 : pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
5023 : s->name, count, atomic_long_read(&n->nr_slabs));
5024 0 : slab_add_kunit_errors();
5025 : }
5026 :
5027 : out:
5028 0 : spin_unlock_irqrestore(&n->list_lock, flags);
5029 0 : return count;
5030 : }
5031 :
5032 0 : long validate_slab_cache(struct kmem_cache *s)
5033 : {
5034 : int node;
5035 0 : unsigned long count = 0;
5036 : struct kmem_cache_node *n;
5037 : unsigned long *obj_map;
5038 :
5039 0 : obj_map = bitmap_alloc(oo_objects(s->oo), GFP_KERNEL);
5040 0 : if (!obj_map)
5041 : return -ENOMEM;
5042 :
5043 0 : flush_all(s);
5044 0 : for_each_kmem_cache_node(s, node, n)
5045 0 : count += validate_slab_node(s, n, obj_map);
5046 :
5047 0 : bitmap_free(obj_map);
5048 :
5049 0 : return count;
5050 : }
5051 : EXPORT_SYMBOL(validate_slab_cache);
5052 :
5053 : #ifdef CONFIG_DEBUG_FS
5054 : /*
5055 : * Generate lists of code addresses where slabcache objects are allocated
5056 : * and freed.
5057 : */
5058 :
5059 : struct location {
5060 : unsigned long count;
5061 : unsigned long addr;
5062 : long long sum_time;
5063 : long min_time;
5064 : long max_time;
5065 : long min_pid;
5066 : long max_pid;
5067 : DECLARE_BITMAP(cpus, NR_CPUS);
5068 : nodemask_t nodes;
5069 : };
5070 :
5071 : struct loc_track {
5072 : unsigned long max;
5073 : unsigned long count;
5074 : struct location *loc;
5075 : loff_t idx;
5076 : };
5077 :
5078 : static struct dentry *slab_debugfs_root;
5079 :
5080 : static void free_loc_track(struct loc_track *t)
5081 : {
5082 : if (t->max)
5083 : free_pages((unsigned long)t->loc,
5084 : get_order(sizeof(struct location) * t->max));
5085 : }
5086 :
5087 : static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
5088 : {
5089 : struct location *l;
5090 : int order;
5091 :
5092 : order = get_order(sizeof(struct location) * max);
5093 :
5094 : l = (void *)__get_free_pages(flags, order);
5095 : if (!l)
5096 : return 0;
5097 :
5098 : if (t->count) {
5099 : memcpy(l, t->loc, sizeof(struct location) * t->count);
5100 : free_loc_track(t);
5101 : }
5102 : t->max = max;
5103 : t->loc = l;
5104 : return 1;
5105 : }
5106 :
5107 : static int add_location(struct loc_track *t, struct kmem_cache *s,
5108 : const struct track *track)
5109 : {
5110 : long start, end, pos;
5111 : struct location *l;
5112 : unsigned long caddr;
5113 : unsigned long age = jiffies - track->when;
5114 :
5115 : start = -1;
5116 : end = t->count;
5117 :
5118 : for ( ; ; ) {
5119 : pos = start + (end - start + 1) / 2;
5120 :
5121 : /*
5122 : * There is nothing at "end". If we end up there
5123 : * we need to add something to before end.
5124 : */
5125 : if (pos == end)
5126 : break;
5127 :
5128 : caddr = t->loc[pos].addr;
5129 : if (track->addr == caddr) {
5130 :
5131 : l = &t->loc[pos];
5132 : l->count++;
5133 : if (track->when) {
5134 : l->sum_time += age;
5135 : if (age < l->min_time)
5136 : l->min_time = age;
5137 : if (age > l->max_time)
5138 : l->max_time = age;
5139 :
5140 : if (track->pid < l->min_pid)
5141 : l->min_pid = track->pid;
5142 : if (track->pid > l->max_pid)
5143 : l->max_pid = track->pid;
5144 :
5145 : cpumask_set_cpu(track->cpu,
5146 : to_cpumask(l->cpus));
5147 : }
5148 : node_set(page_to_nid(virt_to_page(track)), l->nodes);
5149 : return 1;
5150 : }
5151 :
5152 : if (track->addr < caddr)
5153 : end = pos;
5154 : else
5155 : start = pos;
5156 : }
5157 :
5158 : /*
5159 : * Not found. Insert new tracking element.
5160 : */
5161 : if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
5162 : return 0;
5163 :
5164 : l = t->loc + pos;
5165 : if (pos < t->count)
5166 : memmove(l + 1, l,
5167 : (t->count - pos) * sizeof(struct location));
5168 : t->count++;
5169 : l->count = 1;
5170 : l->addr = track->addr;
5171 : l->sum_time = age;
5172 : l->min_time = age;
5173 : l->max_time = age;
5174 : l->min_pid = track->pid;
5175 : l->max_pid = track->pid;
5176 : cpumask_clear(to_cpumask(l->cpus));
5177 : cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
5178 : nodes_clear(l->nodes);
5179 : node_set(page_to_nid(virt_to_page(track)), l->nodes);
5180 : return 1;
5181 : }
5182 :
5183 : static void process_slab(struct loc_track *t, struct kmem_cache *s,
5184 : struct slab *slab, enum track_item alloc,
5185 : unsigned long *obj_map)
5186 : {
5187 : void *addr = slab_address(slab);
5188 : void *p;
5189 :
5190 : __fill_map(obj_map, s, slab);
5191 :
5192 : for_each_object(p, s, addr, slab->objects)
5193 : if (!test_bit(__obj_to_index(s, addr, p), obj_map))
5194 : add_location(t, s, get_track(s, p, alloc));
5195 : }
5196 : #endif /* CONFIG_DEBUG_FS */
5197 : #endif /* CONFIG_SLUB_DEBUG */
5198 :
5199 : #ifdef CONFIG_SYSFS
5200 : enum slab_stat_type {
5201 : SL_ALL, /* All slabs */
5202 : SL_PARTIAL, /* Only partially allocated slabs */
5203 : SL_CPU, /* Only slabs used for cpu caches */
5204 : SL_OBJECTS, /* Determine allocated objects not slabs */
5205 : SL_TOTAL /* Determine object capacity not slabs */
5206 : };
5207 :
5208 : #define SO_ALL (1 << SL_ALL)
5209 : #define SO_PARTIAL (1 << SL_PARTIAL)
5210 : #define SO_CPU (1 << SL_CPU)
5211 : #define SO_OBJECTS (1 << SL_OBJECTS)
5212 : #define SO_TOTAL (1 << SL_TOTAL)
5213 :
5214 0 : static ssize_t show_slab_objects(struct kmem_cache *s,
5215 : char *buf, unsigned long flags)
5216 : {
5217 0 : unsigned long total = 0;
5218 : int node;
5219 : int x;
5220 : unsigned long *nodes;
5221 0 : int len = 0;
5222 :
5223 0 : nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL);
5224 0 : if (!nodes)
5225 : return -ENOMEM;
5226 :
5227 0 : if (flags & SO_CPU) {
5228 : int cpu;
5229 :
5230 0 : for_each_possible_cpu(cpu) {
5231 0 : struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
5232 : cpu);
5233 : int node;
5234 : struct slab *slab;
5235 :
5236 0 : slab = READ_ONCE(c->slab);
5237 0 : if (!slab)
5238 0 : continue;
5239 :
5240 0 : node = slab_nid(slab);
5241 0 : if (flags & SO_TOTAL)
5242 0 : x = slab->objects;
5243 0 : else if (flags & SO_OBJECTS)
5244 0 : x = slab->inuse;
5245 : else
5246 : x = 1;
5247 :
5248 0 : total += x;
5249 0 : nodes[node] += x;
5250 :
5251 : #ifdef CONFIG_SLUB_CPU_PARTIAL
5252 : slab = slub_percpu_partial_read_once(c);
5253 : if (slab) {
5254 : node = slab_nid(slab);
5255 : if (flags & SO_TOTAL)
5256 : WARN_ON_ONCE(1);
5257 : else if (flags & SO_OBJECTS)
5258 : WARN_ON_ONCE(1);
5259 : else
5260 : x = slab->slabs;
5261 : total += x;
5262 : nodes[node] += x;
5263 : }
5264 : #endif
5265 : }
5266 : }
5267 :
5268 : /*
5269 : * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
5270 : * already held which will conflict with an existing lock order:
5271 : *
5272 : * mem_hotplug_lock->slab_mutex->kernfs_mutex
5273 : *
5274 : * We don't really need mem_hotplug_lock (to hold off
5275 : * slab_mem_going_offline_callback) here because slab's memory hot
5276 : * unplug code doesn't destroy the kmem_cache->node[] data.
5277 : */
5278 :
5279 : #ifdef CONFIG_SLUB_DEBUG
5280 0 : if (flags & SO_ALL) {
5281 : struct kmem_cache_node *n;
5282 :
5283 0 : for_each_kmem_cache_node(s, node, n) {
5284 :
5285 0 : if (flags & SO_TOTAL)
5286 0 : x = atomic_long_read(&n->total_objects);
5287 0 : else if (flags & SO_OBJECTS)
5288 0 : x = atomic_long_read(&n->total_objects) -
5289 0 : count_partial(n, count_free);
5290 : else
5291 0 : x = atomic_long_read(&n->nr_slabs);
5292 0 : total += x;
5293 0 : nodes[node] += x;
5294 : }
5295 :
5296 : } else
5297 : #endif
5298 0 : if (flags & SO_PARTIAL) {
5299 : struct kmem_cache_node *n;
5300 :
5301 0 : for_each_kmem_cache_node(s, node, n) {
5302 0 : if (flags & SO_TOTAL)
5303 0 : x = count_partial(n, count_total);
5304 0 : else if (flags & SO_OBJECTS)
5305 0 : x = count_partial(n, count_inuse);
5306 : else
5307 0 : x = n->nr_partial;
5308 0 : total += x;
5309 0 : nodes[node] += x;
5310 : }
5311 : }
5312 :
5313 0 : len += sysfs_emit_at(buf, len, "%lu", total);
5314 : #ifdef CONFIG_NUMA
5315 : for (node = 0; node < nr_node_ids; node++) {
5316 : if (nodes[node])
5317 : len += sysfs_emit_at(buf, len, " N%d=%lu",
5318 : node, nodes[node]);
5319 : }
5320 : #endif
5321 0 : len += sysfs_emit_at(buf, len, "\n");
5322 0 : kfree(nodes);
5323 :
5324 0 : return len;
5325 : }
5326 :
5327 : #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
5328 : #define to_slab(n) container_of(n, struct kmem_cache, kobj)
5329 :
5330 : struct slab_attribute {
5331 : struct attribute attr;
5332 : ssize_t (*show)(struct kmem_cache *s, char *buf);
5333 : ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
5334 : };
5335 :
5336 : #define SLAB_ATTR_RO(_name) \
5337 : static struct slab_attribute _name##_attr = __ATTR_RO_MODE(_name, 0400)
5338 :
5339 : #define SLAB_ATTR(_name) \
5340 : static struct slab_attribute _name##_attr = __ATTR_RW_MODE(_name, 0600)
5341 :
5342 0 : static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
5343 : {
5344 0 : return sysfs_emit(buf, "%u\n", s->size);
5345 : }
5346 : SLAB_ATTR_RO(slab_size);
5347 :
5348 0 : static ssize_t align_show(struct kmem_cache *s, char *buf)
5349 : {
5350 0 : return sysfs_emit(buf, "%u\n", s->align);
5351 : }
5352 : SLAB_ATTR_RO(align);
5353 :
5354 0 : static ssize_t object_size_show(struct kmem_cache *s, char *buf)
5355 : {
5356 0 : return sysfs_emit(buf, "%u\n", s->object_size);
5357 : }
5358 : SLAB_ATTR_RO(object_size);
5359 :
5360 0 : static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
5361 : {
5362 0 : return sysfs_emit(buf, "%u\n", oo_objects(s->oo));
5363 : }
5364 : SLAB_ATTR_RO(objs_per_slab);
5365 :
5366 0 : static ssize_t order_show(struct kmem_cache *s, char *buf)
5367 : {
5368 0 : return sysfs_emit(buf, "%u\n", oo_order(s->oo));
5369 : }
5370 : SLAB_ATTR_RO(order);
5371 :
5372 0 : static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
5373 : {
5374 0 : return sysfs_emit(buf, "%lu\n", s->min_partial);
5375 : }
5376 :
5377 0 : static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
5378 : size_t length)
5379 : {
5380 : unsigned long min;
5381 : int err;
5382 :
5383 0 : err = kstrtoul(buf, 10, &min);
5384 0 : if (err)
5385 0 : return err;
5386 :
5387 0 : s->min_partial = min;
5388 0 : return length;
5389 : }
5390 : SLAB_ATTR(min_partial);
5391 :
5392 0 : static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
5393 : {
5394 0 : unsigned int nr_partial = 0;
5395 : #ifdef CONFIG_SLUB_CPU_PARTIAL
5396 : nr_partial = s->cpu_partial;
5397 : #endif
5398 :
5399 0 : return sysfs_emit(buf, "%u\n", nr_partial);
5400 : }
5401 :
5402 0 : static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
5403 : size_t length)
5404 : {
5405 : unsigned int objects;
5406 : int err;
5407 :
5408 0 : err = kstrtouint(buf, 10, &objects);
5409 0 : if (err)
5410 0 : return err;
5411 0 : if (objects && !kmem_cache_has_cpu_partial(s))
5412 : return -EINVAL;
5413 :
5414 0 : slub_set_cpu_partial(s, objects);
5415 0 : flush_all(s);
5416 0 : return length;
5417 : }
5418 : SLAB_ATTR(cpu_partial);
5419 :
5420 0 : static ssize_t ctor_show(struct kmem_cache *s, char *buf)
5421 : {
5422 0 : if (!s->ctor)
5423 : return 0;
5424 0 : return sysfs_emit(buf, "%pS\n", s->ctor);
5425 : }
5426 : SLAB_ATTR_RO(ctor);
5427 :
5428 0 : static ssize_t aliases_show(struct kmem_cache *s, char *buf)
5429 : {
5430 0 : return sysfs_emit(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
5431 : }
5432 : SLAB_ATTR_RO(aliases);
5433 :
5434 0 : static ssize_t partial_show(struct kmem_cache *s, char *buf)
5435 : {
5436 0 : return show_slab_objects(s, buf, SO_PARTIAL);
5437 : }
5438 : SLAB_ATTR_RO(partial);
5439 :
5440 0 : static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
5441 : {
5442 0 : return show_slab_objects(s, buf, SO_CPU);
5443 : }
5444 : SLAB_ATTR_RO(cpu_slabs);
5445 :
5446 0 : static ssize_t objects_show(struct kmem_cache *s, char *buf)
5447 : {
5448 0 : return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
5449 : }
5450 : SLAB_ATTR_RO(objects);
5451 :
5452 0 : static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
5453 : {
5454 0 : return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
5455 : }
5456 : SLAB_ATTR_RO(objects_partial);
5457 :
5458 0 : static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
5459 : {
5460 0 : int objects = 0;
5461 0 : int slabs = 0;
5462 : int cpu __maybe_unused;
5463 0 : int len = 0;
5464 :
5465 : #ifdef CONFIG_SLUB_CPU_PARTIAL
5466 : for_each_online_cpu(cpu) {
5467 : struct slab *slab;
5468 :
5469 : slab = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5470 :
5471 : if (slab)
5472 : slabs += slab->slabs;
5473 : }
5474 : #endif
5475 :
5476 : /* Approximate half-full slabs, see slub_set_cpu_partial() */
5477 0 : objects = (slabs * oo_objects(s->oo)) / 2;
5478 0 : len += sysfs_emit_at(buf, len, "%d(%d)", objects, slabs);
5479 :
5480 : #if defined(CONFIG_SLUB_CPU_PARTIAL) && defined(CONFIG_SMP)
5481 : for_each_online_cpu(cpu) {
5482 : struct slab *slab;
5483 :
5484 : slab = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5485 : if (slab) {
5486 : slabs = READ_ONCE(slab->slabs);
5487 : objects = (slabs * oo_objects(s->oo)) / 2;
5488 : len += sysfs_emit_at(buf, len, " C%d=%d(%d)",
5489 : cpu, objects, slabs);
5490 : }
5491 : }
5492 : #endif
5493 0 : len += sysfs_emit_at(buf, len, "\n");
5494 :
5495 0 : return len;
5496 : }
5497 : SLAB_ATTR_RO(slabs_cpu_partial);
5498 :
5499 0 : static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
5500 : {
5501 0 : return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
5502 : }
5503 : SLAB_ATTR_RO(reclaim_account);
5504 :
5505 0 : static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
5506 : {
5507 0 : return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
5508 : }
5509 : SLAB_ATTR_RO(hwcache_align);
5510 :
5511 : #ifdef CONFIG_ZONE_DMA
5512 : static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
5513 : {
5514 : return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
5515 : }
5516 : SLAB_ATTR_RO(cache_dma);
5517 : #endif
5518 :
5519 0 : static ssize_t usersize_show(struct kmem_cache *s, char *buf)
5520 : {
5521 0 : return sysfs_emit(buf, "%u\n", s->usersize);
5522 : }
5523 : SLAB_ATTR_RO(usersize);
5524 :
5525 0 : static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
5526 : {
5527 0 : return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU));
5528 : }
5529 : SLAB_ATTR_RO(destroy_by_rcu);
5530 :
5531 : #ifdef CONFIG_SLUB_DEBUG
5532 0 : static ssize_t slabs_show(struct kmem_cache *s, char *buf)
5533 : {
5534 0 : return show_slab_objects(s, buf, SO_ALL);
5535 : }
5536 : SLAB_ATTR_RO(slabs);
5537 :
5538 0 : static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
5539 : {
5540 0 : return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
5541 : }
5542 : SLAB_ATTR_RO(total_objects);
5543 :
5544 0 : static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
5545 : {
5546 0 : return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
5547 : }
5548 : SLAB_ATTR_RO(sanity_checks);
5549 :
5550 0 : static ssize_t trace_show(struct kmem_cache *s, char *buf)
5551 : {
5552 0 : return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TRACE));
5553 : }
5554 : SLAB_ATTR_RO(trace);
5555 :
5556 0 : static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
5557 : {
5558 0 : return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
5559 : }
5560 :
5561 : SLAB_ATTR_RO(red_zone);
5562 :
5563 0 : static ssize_t poison_show(struct kmem_cache *s, char *buf)
5564 : {
5565 0 : return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_POISON));
5566 : }
5567 :
5568 : SLAB_ATTR_RO(poison);
5569 :
5570 0 : static ssize_t store_user_show(struct kmem_cache *s, char *buf)
5571 : {
5572 0 : return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
5573 : }
5574 :
5575 : SLAB_ATTR_RO(store_user);
5576 :
5577 0 : static ssize_t validate_show(struct kmem_cache *s, char *buf)
5578 : {
5579 0 : return 0;
5580 : }
5581 :
5582 0 : static ssize_t validate_store(struct kmem_cache *s,
5583 : const char *buf, size_t length)
5584 : {
5585 0 : int ret = -EINVAL;
5586 :
5587 0 : if (buf[0] == '1') {
5588 0 : ret = validate_slab_cache(s);
5589 0 : if (ret >= 0)
5590 0 : ret = length;
5591 : }
5592 0 : return ret;
5593 : }
5594 : SLAB_ATTR(validate);
5595 :
5596 : #endif /* CONFIG_SLUB_DEBUG */
5597 :
5598 : #ifdef CONFIG_FAILSLAB
5599 : static ssize_t failslab_show(struct kmem_cache *s, char *buf)
5600 : {
5601 : return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
5602 : }
5603 : SLAB_ATTR_RO(failslab);
5604 : #endif
5605 :
5606 0 : static ssize_t shrink_show(struct kmem_cache *s, char *buf)
5607 : {
5608 0 : return 0;
5609 : }
5610 :
5611 0 : static ssize_t shrink_store(struct kmem_cache *s,
5612 : const char *buf, size_t length)
5613 : {
5614 0 : if (buf[0] == '1')
5615 0 : kmem_cache_shrink(s);
5616 : else
5617 : return -EINVAL;
5618 0 : return length;
5619 : }
5620 : SLAB_ATTR(shrink);
5621 :
5622 : #ifdef CONFIG_NUMA
5623 : static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5624 : {
5625 : return sysfs_emit(buf, "%u\n", s->remote_node_defrag_ratio / 10);
5626 : }
5627 :
5628 : static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5629 : const char *buf, size_t length)
5630 : {
5631 : unsigned int ratio;
5632 : int err;
5633 :
5634 : err = kstrtouint(buf, 10, &ratio);
5635 : if (err)
5636 : return err;
5637 : if (ratio > 100)
5638 : return -ERANGE;
5639 :
5640 : s->remote_node_defrag_ratio = ratio * 10;
5641 :
5642 : return length;
5643 : }
5644 : SLAB_ATTR(remote_node_defrag_ratio);
5645 : #endif
5646 :
5647 : #ifdef CONFIG_SLUB_STATS
5648 : static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5649 : {
5650 : unsigned long sum = 0;
5651 : int cpu;
5652 : int len = 0;
5653 : int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL);
5654 :
5655 : if (!data)
5656 : return -ENOMEM;
5657 :
5658 : for_each_online_cpu(cpu) {
5659 : unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5660 :
5661 : data[cpu] = x;
5662 : sum += x;
5663 : }
5664 :
5665 : len += sysfs_emit_at(buf, len, "%lu", sum);
5666 :
5667 : #ifdef CONFIG_SMP
5668 : for_each_online_cpu(cpu) {
5669 : if (data[cpu])
5670 : len += sysfs_emit_at(buf, len, " C%d=%u",
5671 : cpu, data[cpu]);
5672 : }
5673 : #endif
5674 : kfree(data);
5675 : len += sysfs_emit_at(buf, len, "\n");
5676 :
5677 : return len;
5678 : }
5679 :
5680 : static void clear_stat(struct kmem_cache *s, enum stat_item si)
5681 : {
5682 : int cpu;
5683 :
5684 : for_each_online_cpu(cpu)
5685 : per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5686 : }
5687 :
5688 : #define STAT_ATTR(si, text) \
5689 : static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5690 : { \
5691 : return show_stat(s, buf, si); \
5692 : } \
5693 : static ssize_t text##_store(struct kmem_cache *s, \
5694 : const char *buf, size_t length) \
5695 : { \
5696 : if (buf[0] != '0') \
5697 : return -EINVAL; \
5698 : clear_stat(s, si); \
5699 : return length; \
5700 : } \
5701 : SLAB_ATTR(text); \
5702 :
5703 : STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5704 : STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5705 : STAT_ATTR(FREE_FASTPATH, free_fastpath);
5706 : STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5707 : STAT_ATTR(FREE_FROZEN, free_frozen);
5708 : STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5709 : STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5710 : STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5711 : STAT_ATTR(ALLOC_SLAB, alloc_slab);
5712 : STAT_ATTR(ALLOC_REFILL, alloc_refill);
5713 : STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5714 : STAT_ATTR(FREE_SLAB, free_slab);
5715 : STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5716 : STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5717 : STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5718 : STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5719 : STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5720 : STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5721 : STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5722 : STAT_ATTR(ORDER_FALLBACK, order_fallback);
5723 : STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5724 : STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5725 : STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5726 : STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5727 : STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5728 : STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5729 : #endif /* CONFIG_SLUB_STATS */
5730 :
5731 : static struct attribute *slab_attrs[] = {
5732 : &slab_size_attr.attr,
5733 : &object_size_attr.attr,
5734 : &objs_per_slab_attr.attr,
5735 : &order_attr.attr,
5736 : &min_partial_attr.attr,
5737 : &cpu_partial_attr.attr,
5738 : &objects_attr.attr,
5739 : &objects_partial_attr.attr,
5740 : &partial_attr.attr,
5741 : &cpu_slabs_attr.attr,
5742 : &ctor_attr.attr,
5743 : &aliases_attr.attr,
5744 : &align_attr.attr,
5745 : &hwcache_align_attr.attr,
5746 : &reclaim_account_attr.attr,
5747 : &destroy_by_rcu_attr.attr,
5748 : &shrink_attr.attr,
5749 : &slabs_cpu_partial_attr.attr,
5750 : #ifdef CONFIG_SLUB_DEBUG
5751 : &total_objects_attr.attr,
5752 : &slabs_attr.attr,
5753 : &sanity_checks_attr.attr,
5754 : &trace_attr.attr,
5755 : &red_zone_attr.attr,
5756 : &poison_attr.attr,
5757 : &store_user_attr.attr,
5758 : &validate_attr.attr,
5759 : #endif
5760 : #ifdef CONFIG_ZONE_DMA
5761 : &cache_dma_attr.attr,
5762 : #endif
5763 : #ifdef CONFIG_NUMA
5764 : &remote_node_defrag_ratio_attr.attr,
5765 : #endif
5766 : #ifdef CONFIG_SLUB_STATS
5767 : &alloc_fastpath_attr.attr,
5768 : &alloc_slowpath_attr.attr,
5769 : &free_fastpath_attr.attr,
5770 : &free_slowpath_attr.attr,
5771 : &free_frozen_attr.attr,
5772 : &free_add_partial_attr.attr,
5773 : &free_remove_partial_attr.attr,
5774 : &alloc_from_partial_attr.attr,
5775 : &alloc_slab_attr.attr,
5776 : &alloc_refill_attr.attr,
5777 : &alloc_node_mismatch_attr.attr,
5778 : &free_slab_attr.attr,
5779 : &cpuslab_flush_attr.attr,
5780 : &deactivate_full_attr.attr,
5781 : &deactivate_empty_attr.attr,
5782 : &deactivate_to_head_attr.attr,
5783 : &deactivate_to_tail_attr.attr,
5784 : &deactivate_remote_frees_attr.attr,
5785 : &deactivate_bypass_attr.attr,
5786 : &order_fallback_attr.attr,
5787 : &cmpxchg_double_fail_attr.attr,
5788 : &cmpxchg_double_cpu_fail_attr.attr,
5789 : &cpu_partial_alloc_attr.attr,
5790 : &cpu_partial_free_attr.attr,
5791 : &cpu_partial_node_attr.attr,
5792 : &cpu_partial_drain_attr.attr,
5793 : #endif
5794 : #ifdef CONFIG_FAILSLAB
5795 : &failslab_attr.attr,
5796 : #endif
5797 : &usersize_attr.attr,
5798 :
5799 : NULL
5800 : };
5801 :
5802 : static const struct attribute_group slab_attr_group = {
5803 : .attrs = slab_attrs,
5804 : };
5805 :
5806 0 : static ssize_t slab_attr_show(struct kobject *kobj,
5807 : struct attribute *attr,
5808 : char *buf)
5809 : {
5810 : struct slab_attribute *attribute;
5811 : struct kmem_cache *s;
5812 : int err;
5813 :
5814 0 : attribute = to_slab_attr(attr);
5815 0 : s = to_slab(kobj);
5816 :
5817 0 : if (!attribute->show)
5818 : return -EIO;
5819 :
5820 0 : err = attribute->show(s, buf);
5821 :
5822 0 : return err;
5823 : }
5824 :
5825 0 : static ssize_t slab_attr_store(struct kobject *kobj,
5826 : struct attribute *attr,
5827 : const char *buf, size_t len)
5828 : {
5829 : struct slab_attribute *attribute;
5830 : struct kmem_cache *s;
5831 : int err;
5832 :
5833 0 : attribute = to_slab_attr(attr);
5834 0 : s = to_slab(kobj);
5835 :
5836 0 : if (!attribute->store)
5837 : return -EIO;
5838 :
5839 0 : err = attribute->store(s, buf, len);
5840 0 : return err;
5841 : }
5842 :
5843 0 : static void kmem_cache_release(struct kobject *k)
5844 : {
5845 0 : slab_kmem_cache_release(to_slab(k));
5846 0 : }
5847 :
5848 : static const struct sysfs_ops slab_sysfs_ops = {
5849 : .show = slab_attr_show,
5850 : .store = slab_attr_store,
5851 : };
5852 :
5853 : static struct kobj_type slab_ktype = {
5854 : .sysfs_ops = &slab_sysfs_ops,
5855 : .release = kmem_cache_release,
5856 : };
5857 :
5858 : static struct kset *slab_kset;
5859 :
5860 : static inline struct kset *cache_kset(struct kmem_cache *s)
5861 : {
5862 67 : return slab_kset;
5863 : }
5864 :
5865 : #define ID_STR_LENGTH 64
5866 :
5867 : /* Create a unique string id for a slab cache:
5868 : *
5869 : * Format :[flags-]size
5870 : */
5871 24 : static char *create_unique_id(struct kmem_cache *s)
5872 : {
5873 24 : char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5874 24 : char *p = name;
5875 :
5876 24 : BUG_ON(!name);
5877 :
5878 24 : *p++ = ':';
5879 : /*
5880 : * First flags affecting slabcache operations. We will only
5881 : * get here for aliasable slabs so we do not need to support
5882 : * too many flags. The flags here must cover all flags that
5883 : * are matched during merging to guarantee that the id is
5884 : * unique.
5885 : */
5886 24 : if (s->flags & SLAB_CACHE_DMA)
5887 0 : *p++ = 'd';
5888 24 : if (s->flags & SLAB_CACHE_DMA32)
5889 0 : *p++ = 'D';
5890 24 : if (s->flags & SLAB_RECLAIM_ACCOUNT)
5891 1 : *p++ = 'a';
5892 24 : if (s->flags & SLAB_CONSISTENCY_CHECKS)
5893 0 : *p++ = 'F';
5894 : if (s->flags & SLAB_ACCOUNT)
5895 : *p++ = 'A';
5896 24 : if (p != name + 1)
5897 1 : *p++ = '-';
5898 24 : p += sprintf(p, "%07u", s->size);
5899 :
5900 24 : BUG_ON(p > name + ID_STR_LENGTH - 1);
5901 24 : return name;
5902 : }
5903 :
5904 67 : static int sysfs_slab_add(struct kmem_cache *s)
5905 : {
5906 : int err;
5907 : const char *name;
5908 134 : struct kset *kset = cache_kset(s);
5909 67 : int unmergeable = slab_unmergeable(s);
5910 :
5911 67 : if (!kset) {
5912 0 : kobject_init(&s->kobj, &slab_ktype);
5913 0 : return 0;
5914 : }
5915 :
5916 67 : if (!unmergeable && disable_higher_order_debug &&
5917 0 : (slub_debug & DEBUG_METADATA_FLAGS))
5918 0 : unmergeable = 1;
5919 :
5920 67 : if (unmergeable) {
5921 : /*
5922 : * Slabcache can never be merged so we can use the name proper.
5923 : * This is typically the case for debug situations. In that
5924 : * case we can catch duplicate names easily.
5925 : */
5926 43 : sysfs_remove_link(&slab_kset->kobj, s->name);
5927 43 : name = s->name;
5928 : } else {
5929 : /*
5930 : * Create a unique name for the slab as a target
5931 : * for the symlinks.
5932 : */
5933 24 : name = create_unique_id(s);
5934 : }
5935 :
5936 67 : s->kobj.kset = kset;
5937 67 : err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5938 67 : if (err)
5939 : goto out;
5940 :
5941 67 : err = sysfs_create_group(&s->kobj, &slab_attr_group);
5942 67 : if (err)
5943 : goto out_del_kobj;
5944 :
5945 67 : if (!unmergeable) {
5946 : /* Setup first alias */
5947 24 : sysfs_slab_alias(s, s->name);
5948 : }
5949 : out:
5950 67 : if (!unmergeable)
5951 24 : kfree(name);
5952 : return err;
5953 : out_del_kobj:
5954 0 : kobject_del(&s->kobj);
5955 0 : goto out;
5956 : }
5957 :
5958 0 : void sysfs_slab_unlink(struct kmem_cache *s)
5959 : {
5960 0 : if (slab_state >= FULL)
5961 0 : kobject_del(&s->kobj);
5962 0 : }
5963 :
5964 0 : void sysfs_slab_release(struct kmem_cache *s)
5965 : {
5966 0 : if (slab_state >= FULL)
5967 0 : kobject_put(&s->kobj);
5968 0 : }
5969 :
5970 : /*
5971 : * Need to buffer aliases during bootup until sysfs becomes
5972 : * available lest we lose that information.
5973 : */
5974 : struct saved_alias {
5975 : struct kmem_cache *s;
5976 : const char *name;
5977 : struct saved_alias *next;
5978 : };
5979 :
5980 : static struct saved_alias *alias_list;
5981 :
5982 58 : static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5983 : {
5984 : struct saved_alias *al;
5985 :
5986 58 : if (slab_state == FULL) {
5987 : /*
5988 : * If we have a leftover link then remove it.
5989 : */
5990 45 : sysfs_remove_link(&slab_kset->kobj, name);
5991 45 : return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5992 : }
5993 :
5994 13 : al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5995 13 : if (!al)
5996 : return -ENOMEM;
5997 :
5998 13 : al->s = s;
5999 13 : al->name = name;
6000 13 : al->next = alias_list;
6001 13 : alias_list = al;
6002 13 : return 0;
6003 : }
6004 :
6005 1 : static int __init slab_sysfs_init(void)
6006 : {
6007 : struct kmem_cache *s;
6008 : int err;
6009 :
6010 1 : mutex_lock(&slab_mutex);
6011 :
6012 1 : slab_kset = kset_create_and_add("slab", NULL, kernel_kobj);
6013 1 : if (!slab_kset) {
6014 0 : mutex_unlock(&slab_mutex);
6015 0 : pr_err("Cannot register slab subsystem.\n");
6016 0 : return -ENOSYS;
6017 : }
6018 :
6019 1 : slab_state = FULL;
6020 :
6021 65 : list_for_each_entry(s, &slab_caches, list) {
6022 64 : err = sysfs_slab_add(s);
6023 64 : if (err)
6024 0 : pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
6025 : s->name);
6026 : }
6027 :
6028 14 : while (alias_list) {
6029 13 : struct saved_alias *al = alias_list;
6030 :
6031 13 : alias_list = alias_list->next;
6032 13 : err = sysfs_slab_alias(al->s, al->name);
6033 13 : if (err)
6034 0 : pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
6035 : al->name);
6036 13 : kfree(al);
6037 : }
6038 :
6039 1 : mutex_unlock(&slab_mutex);
6040 1 : return 0;
6041 : }
6042 :
6043 : __initcall(slab_sysfs_init);
6044 : #endif /* CONFIG_SYSFS */
6045 :
6046 : #if defined(CONFIG_SLUB_DEBUG) && defined(CONFIG_DEBUG_FS)
6047 : static int slab_debugfs_show(struct seq_file *seq, void *v)
6048 : {
6049 : struct loc_track *t = seq->private;
6050 : struct location *l;
6051 : unsigned long idx;
6052 :
6053 : idx = (unsigned long) t->idx;
6054 : if (idx < t->count) {
6055 : l = &t->loc[idx];
6056 :
6057 : seq_printf(seq, "%7ld ", l->count);
6058 :
6059 : if (l->addr)
6060 : seq_printf(seq, "%pS", (void *)l->addr);
6061 : else
6062 : seq_puts(seq, "<not-available>");
6063 :
6064 : if (l->sum_time != l->min_time) {
6065 : seq_printf(seq, " age=%ld/%llu/%ld",
6066 : l->min_time, div_u64(l->sum_time, l->count),
6067 : l->max_time);
6068 : } else
6069 : seq_printf(seq, " age=%ld", l->min_time);
6070 :
6071 : if (l->min_pid != l->max_pid)
6072 : seq_printf(seq, " pid=%ld-%ld", l->min_pid, l->max_pid);
6073 : else
6074 : seq_printf(seq, " pid=%ld",
6075 : l->min_pid);
6076 :
6077 : if (num_online_cpus() > 1 && !cpumask_empty(to_cpumask(l->cpus)))
6078 : seq_printf(seq, " cpus=%*pbl",
6079 : cpumask_pr_args(to_cpumask(l->cpus)));
6080 :
6081 : if (nr_online_nodes > 1 && !nodes_empty(l->nodes))
6082 : seq_printf(seq, " nodes=%*pbl",
6083 : nodemask_pr_args(&l->nodes));
6084 :
6085 : seq_puts(seq, "\n");
6086 : }
6087 :
6088 : if (!idx && !t->count)
6089 : seq_puts(seq, "No data\n");
6090 :
6091 : return 0;
6092 : }
6093 :
6094 : static void slab_debugfs_stop(struct seq_file *seq, void *v)
6095 : {
6096 : }
6097 :
6098 : static void *slab_debugfs_next(struct seq_file *seq, void *v, loff_t *ppos)
6099 : {
6100 : struct loc_track *t = seq->private;
6101 :
6102 : t->idx = ++(*ppos);
6103 : if (*ppos <= t->count)
6104 : return ppos;
6105 :
6106 : return NULL;
6107 : }
6108 :
6109 : static void *slab_debugfs_start(struct seq_file *seq, loff_t *ppos)
6110 : {
6111 : struct loc_track *t = seq->private;
6112 :
6113 : t->idx = *ppos;
6114 : return ppos;
6115 : }
6116 :
6117 : static const struct seq_operations slab_debugfs_sops = {
6118 : .start = slab_debugfs_start,
6119 : .next = slab_debugfs_next,
6120 : .stop = slab_debugfs_stop,
6121 : .show = slab_debugfs_show,
6122 : };
6123 :
6124 : static int slab_debug_trace_open(struct inode *inode, struct file *filep)
6125 : {
6126 :
6127 : struct kmem_cache_node *n;
6128 : enum track_item alloc;
6129 : int node;
6130 : struct loc_track *t = __seq_open_private(filep, &slab_debugfs_sops,
6131 : sizeof(struct loc_track));
6132 : struct kmem_cache *s = file_inode(filep)->i_private;
6133 : unsigned long *obj_map;
6134 :
6135 : if (!t)
6136 : return -ENOMEM;
6137 :
6138 : obj_map = bitmap_alloc(oo_objects(s->oo), GFP_KERNEL);
6139 : if (!obj_map) {
6140 : seq_release_private(inode, filep);
6141 : return -ENOMEM;
6142 : }
6143 :
6144 : if (strcmp(filep->f_path.dentry->d_name.name, "alloc_traces") == 0)
6145 : alloc = TRACK_ALLOC;
6146 : else
6147 : alloc = TRACK_FREE;
6148 :
6149 : if (!alloc_loc_track(t, PAGE_SIZE / sizeof(struct location), GFP_KERNEL)) {
6150 : bitmap_free(obj_map);
6151 : seq_release_private(inode, filep);
6152 : return -ENOMEM;
6153 : }
6154 :
6155 : for_each_kmem_cache_node(s, node, n) {
6156 : unsigned long flags;
6157 : struct slab *slab;
6158 :
6159 : if (!atomic_long_read(&n->nr_slabs))
6160 : continue;
6161 :
6162 : spin_lock_irqsave(&n->list_lock, flags);
6163 : list_for_each_entry(slab, &n->partial, slab_list)
6164 : process_slab(t, s, slab, alloc, obj_map);
6165 : list_for_each_entry(slab, &n->full, slab_list)
6166 : process_slab(t, s, slab, alloc, obj_map);
6167 : spin_unlock_irqrestore(&n->list_lock, flags);
6168 : }
6169 :
6170 : bitmap_free(obj_map);
6171 : return 0;
6172 : }
6173 :
6174 : static int slab_debug_trace_release(struct inode *inode, struct file *file)
6175 : {
6176 : struct seq_file *seq = file->private_data;
6177 : struct loc_track *t = seq->private;
6178 :
6179 : free_loc_track(t);
6180 : return seq_release_private(inode, file);
6181 : }
6182 :
6183 : static const struct file_operations slab_debugfs_fops = {
6184 : .open = slab_debug_trace_open,
6185 : .read = seq_read,
6186 : .llseek = seq_lseek,
6187 : .release = slab_debug_trace_release,
6188 : };
6189 :
6190 : static void debugfs_slab_add(struct kmem_cache *s)
6191 : {
6192 : struct dentry *slab_cache_dir;
6193 :
6194 : if (unlikely(!slab_debugfs_root))
6195 : return;
6196 :
6197 : slab_cache_dir = debugfs_create_dir(s->name, slab_debugfs_root);
6198 :
6199 : debugfs_create_file("alloc_traces", 0400,
6200 : slab_cache_dir, s, &slab_debugfs_fops);
6201 :
6202 : debugfs_create_file("free_traces", 0400,
6203 : slab_cache_dir, s, &slab_debugfs_fops);
6204 : }
6205 :
6206 : void debugfs_slab_release(struct kmem_cache *s)
6207 : {
6208 : debugfs_remove_recursive(debugfs_lookup(s->name, slab_debugfs_root));
6209 : }
6210 :
6211 : static int __init slab_debugfs_init(void)
6212 : {
6213 : struct kmem_cache *s;
6214 :
6215 : slab_debugfs_root = debugfs_create_dir("slab", NULL);
6216 :
6217 : list_for_each_entry(s, &slab_caches, list)
6218 : if (s->flags & SLAB_STORE_USER)
6219 : debugfs_slab_add(s);
6220 :
6221 : return 0;
6222 :
6223 : }
6224 : __initcall(slab_debugfs_init);
6225 : #endif
6226 : /*
6227 : * The /proc/slabinfo ABI
6228 : */
6229 : #ifdef CONFIG_SLUB_DEBUG
6230 0 : void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
6231 : {
6232 0 : unsigned long nr_slabs = 0;
6233 0 : unsigned long nr_objs = 0;
6234 0 : unsigned long nr_free = 0;
6235 : int node;
6236 : struct kmem_cache_node *n;
6237 :
6238 0 : for_each_kmem_cache_node(s, node, n) {
6239 0 : nr_slabs += node_nr_slabs(n);
6240 0 : nr_objs += node_nr_objs(n);
6241 0 : nr_free += count_partial(n, count_free);
6242 : }
6243 :
6244 0 : sinfo->active_objs = nr_objs - nr_free;
6245 0 : sinfo->num_objs = nr_objs;
6246 0 : sinfo->active_slabs = nr_slabs;
6247 0 : sinfo->num_slabs = nr_slabs;
6248 0 : sinfo->objects_per_slab = oo_objects(s->oo);
6249 0 : sinfo->cache_order = oo_order(s->oo);
6250 0 : }
6251 :
6252 0 : void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
6253 : {
6254 0 : }
6255 :
6256 0 : ssize_t slabinfo_write(struct file *file, const char __user *buffer,
6257 : size_t count, loff_t *ppos)
6258 : {
6259 0 : return -EIO;
6260 : }
6261 : #endif /* CONFIG_SLUB_DEBUG */
|
