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/kmsan.h>
26 #include <linux/cpu.h>
27 #include <linux/cpuset.h>
28 #include <linux/mempolicy.h>
29 #include <linux/ctype.h>
30 #include <linux/stackdepot.h>
31 #include <linux/debugobjects.h>
32 #include <linux/kallsyms.h>
33 #include <linux/kfence.h>
34 #include <linux/memory.h>
35 #include <linux/math64.h>
36 #include <linux/fault-inject.h>
37 #include <linux/stacktrace.h>
38 #include <linux/prefetch.h>
39 #include <linux/memcontrol.h>
40 #include <linux/random.h>
41 #include <kunit/test.h>
42 #include <linux/sort.h>
43
44 #include <linux/debugfs.h>
45 #include <trace/events/kmem.h>
46
47 #include "internal.h"
48
49 /*
50 * Lock order:
51 * 1. slab_mutex (Global Mutex)
52 * 2. node->list_lock (Spinlock)
53 * 3. kmem_cache->cpu_slab->lock (Local lock)
54 * 4. slab_lock(slab) (Only on some arches)
55 * 5. object_map_lock (Only for debugging)
56 *
57 * slab_mutex
58 *
59 * The role of the slab_mutex is to protect the list of all the slabs
60 * and to synchronize major metadata changes to slab cache structures.
61 * Also synchronizes memory hotplug callbacks.
62 *
63 * slab_lock
64 *
65 * The slab_lock is a wrapper around the page lock, thus it is a bit
66 * spinlock.
67 *
68 * The slab_lock is only used on arches that do not have the ability
69 * to do a cmpxchg_double. It only protects:
70 *
71 * A. slab->freelist -> List of free objects in a slab
72 * B. slab->inuse -> Number of objects in use
73 * C. slab->objects -> Number of objects in slab
74 * D. slab->frozen -> frozen state
75 *
76 * Frozen slabs
77 *
78 * If a slab is frozen then it is exempt from list management. It is not
79 * on any list except per cpu partial list. The processor that froze the
80 * slab is the one who can perform list operations on the slab. Other
81 * processors may put objects onto the freelist but the processor that
82 * froze the slab is the only one that can retrieve the objects from the
83 * slab's freelist.
84 *
85 * list_lock
86 *
87 * The list_lock protects the partial and full list on each node and
88 * the partial slab counter. If taken then no new slabs may be added or
89 * removed from the lists nor make the number of partial slabs be modified.
90 * (Note that the total number of slabs is an atomic value that may be
91 * modified without taking the list lock).
92 *
93 * The list_lock is a centralized lock and thus we avoid taking it as
94 * much as possible. As long as SLUB does not have to handle partial
95 * slabs, operations can continue without any centralized lock. F.e.
96 * allocating a long series of objects that fill up slabs does not require
97 * the list lock.
98 *
99 * For debug caches, all allocations are forced to go through a list_lock
100 * protected region to serialize against concurrent validation.
101 *
102 * cpu_slab->lock local lock
103 *
104 * This locks protect slowpath manipulation of all kmem_cache_cpu fields
105 * except the stat counters. This is a percpu structure manipulated only by
106 * the local cpu, so the lock protects against being preempted or interrupted
107 * by an irq. Fast path operations rely on lockless operations instead.
108 *
109 * On PREEMPT_RT, the local lock neither disables interrupts nor preemption
110 * which means the lockless fastpath cannot be used as it might interfere with
111 * an in-progress slow path operations. In this case the local lock is always
112 * taken but it still utilizes the freelist for the common operations.
113 *
114 * lockless fastpaths
115 *
116 * The fast path allocation (slab_alloc_node()) and freeing (do_slab_free())
117 * are fully lockless when satisfied from the percpu slab (and when
118 * cmpxchg_double is possible to use, otherwise slab_lock is taken).
119 * They also don't disable preemption or migration or irqs. They rely on
120 * the transaction id (tid) field to detect being preempted or moved to
121 * another cpu.
122 *
123 * irq, preemption, migration considerations
124 *
125 * Interrupts are disabled as part of list_lock or local_lock operations, or
126 * around the slab_lock operation, in order to make the slab allocator safe
127 * to use in the context of an irq.
128 *
129 * In addition, preemption (or migration on PREEMPT_RT) is disabled in the
130 * allocation slowpath, bulk allocation, and put_cpu_partial(), so that the
131 * local cpu doesn't change in the process and e.g. the kmem_cache_cpu pointer
132 * doesn't have to be revalidated in each section protected by the local lock.
133 *
134 * SLUB assigns one slab for allocation to each processor.
135 * Allocations only occur from these slabs called cpu slabs.
136 *
137 * Slabs with free elements are kept on a partial list and during regular
138 * operations no list for full slabs is used. If an object in a full slab is
139 * freed then the slab will show up again on the partial lists.
140 * We track full slabs for debugging purposes though because otherwise we
141 * cannot scan all objects.
142 *
143 * Slabs are freed when they become empty. Teardown and setup is
144 * minimal so we rely on the page allocators per cpu caches for
145 * fast frees and allocs.
146 *
147 * slab->frozen The slab is frozen and exempt from list processing.
148 * This means that the slab is dedicated to a purpose
149 * such as satisfying allocations for a specific
150 * processor. Objects may be freed in the slab while
151 * it is frozen but slab_free will then skip the usual
152 * list operations. It is up to the processor holding
153 * the slab to integrate the slab into the slab lists
154 * when the slab is no longer needed.
155 *
156 * One use of this flag is to mark slabs that are
157 * used for allocations. Then such a slab becomes a cpu
158 * slab. The cpu slab may be equipped with an additional
159 * freelist that allows lockless access to
160 * free objects in addition to the regular freelist
161 * that requires the slab lock.
162 *
163 * SLAB_DEBUG_FLAGS Slab requires special handling due to debug
164 * options set. This moves slab handling out of
165 * the fast path and disables lockless freelists.
166 */
167
168 /*
169 * We could simply use migrate_disable()/enable() but as long as it's a
170 * function call even on !PREEMPT_RT, use inline preempt_disable() there.
171 */
172 #ifndef CONFIG_PREEMPT_RT
173 #define slub_get_cpu_ptr(var) get_cpu_ptr(var)
174 #define slub_put_cpu_ptr(var) put_cpu_ptr(var)
175 #define USE_LOCKLESS_FAST_PATH() (true)
176 #else
177 #define slub_get_cpu_ptr(var) \
178 ({ \
179 migrate_disable(); \
180 this_cpu_ptr(var); \
181 })
182 #define slub_put_cpu_ptr(var) \
183 do { \
184 (void)(var); \
185 migrate_enable(); \
186 } while (0)
187 #define USE_LOCKLESS_FAST_PATH() (false)
188 #endif
189
190 #ifdef CONFIG_SLUB_DEBUG
191 #ifdef CONFIG_SLUB_DEBUG_ON
192 DEFINE_STATIC_KEY_TRUE(slub_debug_enabled);
193 #else
194 DEFINE_STATIC_KEY_FALSE(slub_debug_enabled);
195 #endif
196 #endif /* CONFIG_SLUB_DEBUG */
197
198 /* Structure holding parameters for get_partial() call chain */
199 struct partial_context {
200 struct slab **slab;
201 gfp_t flags;
202 unsigned int orig_size;
203 };
204
kmem_cache_debug(struct kmem_cache * s)205 static inline bool kmem_cache_debug(struct kmem_cache *s)
206 {
207 return kmem_cache_debug_flags(s, SLAB_DEBUG_FLAGS);
208 }
209
slub_debug_orig_size(struct kmem_cache * s)210 static inline bool slub_debug_orig_size(struct kmem_cache *s)
211 {
212 return (kmem_cache_debug_flags(s, SLAB_STORE_USER) &&
213 (s->flags & SLAB_KMALLOC));
214 }
215
fixup_red_left(struct kmem_cache * s,void * p)216 void *fixup_red_left(struct kmem_cache *s, void *p)
217 {
218 if (kmem_cache_debug_flags(s, SLAB_RED_ZONE))
219 p += s->red_left_pad;
220
221 return p;
222 }
223
kmem_cache_has_cpu_partial(struct kmem_cache * s)224 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
225 {
226 #ifdef CONFIG_SLUB_CPU_PARTIAL
227 return !kmem_cache_debug(s);
228 #else
229 return false;
230 #endif
231 }
232
233 /*
234 * Issues still to be resolved:
235 *
236 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
237 *
238 * - Variable sizing of the per node arrays
239 */
240
241 /* Enable to log cmpxchg failures */
242 #undef SLUB_DEBUG_CMPXCHG
243
244 /*
245 * Minimum number of partial slabs. These will be left on the partial
246 * lists even if they are empty. kmem_cache_shrink may reclaim them.
247 */
248 #define MIN_PARTIAL 5
249
250 /*
251 * Maximum number of desirable partial slabs.
252 * The existence of more partial slabs makes kmem_cache_shrink
253 * sort the partial list by the number of objects in use.
254 */
255 #define MAX_PARTIAL 10
256
257 #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
258 SLAB_POISON | SLAB_STORE_USER)
259
260 /*
261 * These debug flags cannot use CMPXCHG because there might be consistency
262 * issues when checking or reading debug information
263 */
264 #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
265 SLAB_TRACE)
266
267
268 /*
269 * Debugging flags that require metadata to be stored in the slab. These get
270 * disabled when slub_debug=O is used and a cache's min order increases with
271 * metadata.
272 */
273 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
274
275 #define OO_SHIFT 16
276 #define OO_MASK ((1 << OO_SHIFT) - 1)
277 #define MAX_OBJS_PER_PAGE 32767 /* since slab.objects is u15 */
278
279 /* Internal SLUB flags */
280 /* Poison object */
281 #define __OBJECT_POISON ((slab_flags_t __force)0x80000000U)
282 /* Use cmpxchg_double */
283 #define __CMPXCHG_DOUBLE ((slab_flags_t __force)0x40000000U)
284
285 /*
286 * Tracking user of a slab.
287 */
288 #define TRACK_ADDRS_COUNT 16
289 struct track {
290 unsigned long addr; /* Called from address */
291 #ifdef CONFIG_STACKDEPOT
292 depot_stack_handle_t handle;
293 #endif
294 int cpu; /* Was running on cpu */
295 int pid; /* Pid context */
296 unsigned long when; /* When did the operation occur */
297 };
298
299 enum track_item { TRACK_ALLOC, TRACK_FREE };
300
301 #ifdef CONFIG_SYSFS
302 static int sysfs_slab_add(struct kmem_cache *);
303 static int sysfs_slab_alias(struct kmem_cache *, const char *);
304 #else
sysfs_slab_add(struct kmem_cache * s)305 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
sysfs_slab_alias(struct kmem_cache * s,const char * p)306 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
307 { return 0; }
308 #endif
309
310 #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_SLUB_DEBUG)
311 static void debugfs_slab_add(struct kmem_cache *);
312 #else
debugfs_slab_add(struct kmem_cache * s)313 static inline void debugfs_slab_add(struct kmem_cache *s) { }
314 #endif
315
stat(const struct kmem_cache * s,enum stat_item si)316 static inline void stat(const struct kmem_cache *s, enum stat_item si)
317 {
318 #ifdef CONFIG_SLUB_STATS
319 /*
320 * The rmw is racy on a preemptible kernel but this is acceptable, so
321 * avoid this_cpu_add()'s irq-disable overhead.
322 */
323 raw_cpu_inc(s->cpu_slab->stat[si]);
324 #endif
325 }
326
327 /*
328 * Tracks for which NUMA nodes we have kmem_cache_nodes allocated.
329 * Corresponds to node_state[N_NORMAL_MEMORY], but can temporarily
330 * differ during memory hotplug/hotremove operations.
331 * Protected by slab_mutex.
332 */
333 static nodemask_t slab_nodes;
334
335 /*
336 * Workqueue used for flush_cpu_slab().
337 */
338 static struct workqueue_struct *flushwq;
339
340 /********************************************************************
341 * Core slab cache functions
342 *******************************************************************/
343
344 /*
345 * Returns freelist pointer (ptr). With hardening, this is obfuscated
346 * with an XOR of the address where the pointer is held and a per-cache
347 * random number.
348 */
freelist_ptr(const struct kmem_cache * s,void * ptr,unsigned long ptr_addr)349 static inline void *freelist_ptr(const struct kmem_cache *s, void *ptr,
350 unsigned long ptr_addr)
351 {
352 #ifdef CONFIG_SLAB_FREELIST_HARDENED
353 /*
354 * When CONFIG_KASAN_SW/HW_TAGS is enabled, ptr_addr might be tagged.
355 * Normally, this doesn't cause any issues, as both set_freepointer()
356 * and get_freepointer() are called with a pointer with the same tag.
357 * However, there are some issues with CONFIG_SLUB_DEBUG code. For
358 * example, when __free_slub() iterates over objects in a cache, it
359 * passes untagged pointers to check_object(). check_object() in turns
360 * calls get_freepointer() with an untagged pointer, which causes the
361 * freepointer to be restored incorrectly.
362 */
363 return (void *)((unsigned long)ptr ^ s->random ^
364 swab((unsigned long)kasan_reset_tag((void *)ptr_addr)));
365 #else
366 return ptr;
367 #endif
368 }
369
370 /* Returns the freelist pointer recorded at location ptr_addr. */
freelist_dereference(const struct kmem_cache * s,void * ptr_addr)371 static inline void *freelist_dereference(const struct kmem_cache *s,
372 void *ptr_addr)
373 {
374 return freelist_ptr(s, (void *)*(unsigned long *)(ptr_addr),
375 (unsigned long)ptr_addr);
376 }
377
get_freepointer(struct kmem_cache * s,void * object)378 static inline void *get_freepointer(struct kmem_cache *s, void *object)
379 {
380 object = kasan_reset_tag(object);
381 return freelist_dereference(s, object + s->offset);
382 }
383
prefetch_freepointer(const struct kmem_cache * s,void * object)384 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
385 {
386 prefetchw(object + s->offset);
387 }
388
389 /*
390 * When running under KMSAN, get_freepointer_safe() may return an uninitialized
391 * pointer value in the case the current thread loses the race for the next
392 * memory chunk in the freelist. In that case this_cpu_cmpxchg_double() in
393 * slab_alloc_node() will fail, so the uninitialized value won't be used, but
394 * KMSAN will still check all arguments of cmpxchg because of imperfect
395 * handling of inline assembly.
396 * To work around this problem, we apply __no_kmsan_checks to ensure that
397 * get_freepointer_safe() returns initialized memory.
398 */
399 __no_kmsan_checks
get_freepointer_safe(struct kmem_cache * s,void * object)400 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
401 {
402 unsigned long freepointer_addr;
403 void *p;
404
405 if (!debug_pagealloc_enabled_static())
406 return get_freepointer(s, object);
407
408 object = kasan_reset_tag(object);
409 freepointer_addr = (unsigned long)object + s->offset;
410 copy_from_kernel_nofault(&p, (void **)freepointer_addr, sizeof(p));
411 return freelist_ptr(s, p, freepointer_addr);
412 }
413
set_freepointer(struct kmem_cache * s,void * object,void * fp)414 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
415 {
416 unsigned long freeptr_addr = (unsigned long)object + s->offset;
417
418 #ifdef CONFIG_SLAB_FREELIST_HARDENED
419 BUG_ON(object == fp); /* naive detection of double free or corruption */
420 #endif
421
422 freeptr_addr = (unsigned long)kasan_reset_tag((void *)freeptr_addr);
423 *(void **)freeptr_addr = freelist_ptr(s, fp, freeptr_addr);
424 }
425
426 /* Loop over all objects in a slab */
427 #define for_each_object(__p, __s, __addr, __objects) \
428 for (__p = fixup_red_left(__s, __addr); \
429 __p < (__addr) + (__objects) * (__s)->size; \
430 __p += (__s)->size)
431
order_objects(unsigned int order,unsigned int size)432 static inline unsigned int order_objects(unsigned int order, unsigned int size)
433 {
434 return ((unsigned int)PAGE_SIZE << order) / size;
435 }
436
oo_make(unsigned int order,unsigned int size)437 static inline struct kmem_cache_order_objects oo_make(unsigned int order,
438 unsigned int size)
439 {
440 struct kmem_cache_order_objects x = {
441 (order << OO_SHIFT) + order_objects(order, size)
442 };
443
444 return x;
445 }
446
oo_order(struct kmem_cache_order_objects x)447 static inline unsigned int oo_order(struct kmem_cache_order_objects x)
448 {
449 return x.x >> OO_SHIFT;
450 }
451
oo_objects(struct kmem_cache_order_objects x)452 static inline unsigned int oo_objects(struct kmem_cache_order_objects x)
453 {
454 return x.x & OO_MASK;
455 }
456
457 #ifdef CONFIG_SLUB_CPU_PARTIAL
slub_set_cpu_partial(struct kmem_cache * s,unsigned int nr_objects)458 static void slub_set_cpu_partial(struct kmem_cache *s, unsigned int nr_objects)
459 {
460 unsigned int nr_slabs;
461
462 s->cpu_partial = nr_objects;
463
464 /*
465 * We take the number of objects but actually limit the number of
466 * slabs on the per cpu partial list, in order to limit excessive
467 * growth of the list. For simplicity we assume that the slabs will
468 * be half-full.
469 */
470 nr_slabs = DIV_ROUND_UP(nr_objects * 2, oo_objects(s->oo));
471 s->cpu_partial_slabs = nr_slabs;
472 }
473 #else
474 static inline void
slub_set_cpu_partial(struct kmem_cache * s,unsigned int nr_objects)475 slub_set_cpu_partial(struct kmem_cache *s, unsigned int nr_objects)
476 {
477 }
478 #endif /* CONFIG_SLUB_CPU_PARTIAL */
479
480 /*
481 * Per slab locking using the pagelock
482 */
slab_lock(struct slab * slab)483 static __always_inline void slab_lock(struct slab *slab)
484 {
485 struct page *page = slab_page(slab);
486
487 VM_BUG_ON_PAGE(PageTail(page), page);
488 bit_spin_lock(PG_locked, &page->flags);
489 }
490
slab_unlock(struct slab * slab)491 static __always_inline void slab_unlock(struct slab *slab)
492 {
493 struct page *page = slab_page(slab);
494
495 VM_BUG_ON_PAGE(PageTail(page), page);
496 __bit_spin_unlock(PG_locked, &page->flags);
497 }
498
499 /*
500 * Interrupts must be disabled (for the fallback code to work right), typically
501 * by an _irqsave() lock variant. On PREEMPT_RT the preempt_disable(), which is
502 * part of bit_spin_lock(), is sufficient because the policy is not to allow any
503 * allocation/ free operation in hardirq context. Therefore nothing can
504 * interrupt the operation.
505 */
__cmpxchg_double_slab(struct kmem_cache * s,struct slab * slab,void * freelist_old,unsigned long counters_old,void * freelist_new,unsigned long counters_new,const char * n)506 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct slab *slab,
507 void *freelist_old, unsigned long counters_old,
508 void *freelist_new, unsigned long counters_new,
509 const char *n)
510 {
511 if (USE_LOCKLESS_FAST_PATH())
512 lockdep_assert_irqs_disabled();
513 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
514 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
515 if (s->flags & __CMPXCHG_DOUBLE) {
516 if (cmpxchg_double(&slab->freelist, &slab->counters,
517 freelist_old, counters_old,
518 freelist_new, counters_new))
519 return true;
520 } else
521 #endif
522 {
523 slab_lock(slab);
524 if (slab->freelist == freelist_old &&
525 slab->counters == counters_old) {
526 slab->freelist = freelist_new;
527 slab->counters = counters_new;
528 slab_unlock(slab);
529 return true;
530 }
531 slab_unlock(slab);
532 }
533
534 cpu_relax();
535 stat(s, CMPXCHG_DOUBLE_FAIL);
536
537 #ifdef SLUB_DEBUG_CMPXCHG
538 pr_info("%s %s: cmpxchg double redo ", n, s->name);
539 #endif
540
541 return false;
542 }
543
cmpxchg_double_slab(struct kmem_cache * s,struct slab * slab,void * freelist_old,unsigned long counters_old,void * freelist_new,unsigned long counters_new,const char * n)544 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct slab *slab,
545 void *freelist_old, unsigned long counters_old,
546 void *freelist_new, unsigned long counters_new,
547 const char *n)
548 {
549 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
550 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
551 if (s->flags & __CMPXCHG_DOUBLE) {
552 if (cmpxchg_double(&slab->freelist, &slab->counters,
553 freelist_old, counters_old,
554 freelist_new, counters_new))
555 return true;
556 } else
557 #endif
558 {
559 unsigned long flags;
560
561 local_irq_save(flags);
562 slab_lock(slab);
563 if (slab->freelist == freelist_old &&
564 slab->counters == counters_old) {
565 slab->freelist = freelist_new;
566 slab->counters = counters_new;
567 slab_unlock(slab);
568 local_irq_restore(flags);
569 return true;
570 }
571 slab_unlock(slab);
572 local_irq_restore(flags);
573 }
574
575 cpu_relax();
576 stat(s, CMPXCHG_DOUBLE_FAIL);
577
578 #ifdef SLUB_DEBUG_CMPXCHG
579 pr_info("%s %s: cmpxchg double redo ", n, s->name);
580 #endif
581
582 return false;
583 }
584
585 #ifdef CONFIG_SLUB_DEBUG
586 static unsigned long object_map[BITS_TO_LONGS(MAX_OBJS_PER_PAGE)];
587 static DEFINE_SPINLOCK(object_map_lock);
588
__fill_map(unsigned long * obj_map,struct kmem_cache * s,struct slab * slab)589 static void __fill_map(unsigned long *obj_map, struct kmem_cache *s,
590 struct slab *slab)
591 {
592 void *addr = slab_address(slab);
593 void *p;
594
595 bitmap_zero(obj_map, slab->objects);
596
597 for (p = slab->freelist; p; p = get_freepointer(s, p))
598 set_bit(__obj_to_index(s, addr, p), obj_map);
599 }
600
601 #if IS_ENABLED(CONFIG_KUNIT)
slab_add_kunit_errors(void)602 static bool slab_add_kunit_errors(void)
603 {
604 struct kunit_resource *resource;
605
606 if (likely(!current->kunit_test))
607 return false;
608
609 resource = kunit_find_named_resource(current->kunit_test, "slab_errors");
610 if (!resource)
611 return false;
612
613 (*(int *)resource->data)++;
614 kunit_put_resource(resource);
615 return true;
616 }
617 #else
slab_add_kunit_errors(void)618 static inline bool slab_add_kunit_errors(void) { return false; }
619 #endif
620
size_from_object(struct kmem_cache * s)621 static inline unsigned int size_from_object(struct kmem_cache *s)
622 {
623 if (s->flags & SLAB_RED_ZONE)
624 return s->size - s->red_left_pad;
625
626 return s->size;
627 }
628
restore_red_left(struct kmem_cache * s,void * p)629 static inline void *restore_red_left(struct kmem_cache *s, void *p)
630 {
631 if (s->flags & SLAB_RED_ZONE)
632 p -= s->red_left_pad;
633
634 return p;
635 }
636
637 /*
638 * Debug settings:
639 */
640 #if defined(CONFIG_SLUB_DEBUG_ON)
641 static slab_flags_t slub_debug = DEBUG_DEFAULT_FLAGS;
642 #else
643 static slab_flags_t slub_debug;
644 #endif
645
646 static char *slub_debug_string;
647 static int disable_higher_order_debug;
648
649 /*
650 * slub is about to manipulate internal object metadata. This memory lies
651 * outside the range of the allocated object, so accessing it would normally
652 * be reported by kasan as a bounds error. metadata_access_enable() is used
653 * to tell kasan that these accesses are OK.
654 */
metadata_access_enable(void)655 static inline void metadata_access_enable(void)
656 {
657 kasan_disable_current();
658 }
659
metadata_access_disable(void)660 static inline void metadata_access_disable(void)
661 {
662 kasan_enable_current();
663 }
664
665 /*
666 * Object debugging
667 */
668
669 /* Verify that a pointer has an address that is valid within a slab page */
check_valid_pointer(struct kmem_cache * s,struct slab * slab,void * object)670 static inline int check_valid_pointer(struct kmem_cache *s,
671 struct slab *slab, void *object)
672 {
673 void *base;
674
675 if (!object)
676 return 1;
677
678 base = slab_address(slab);
679 object = kasan_reset_tag(object);
680 object = restore_red_left(s, object);
681 if (object < base || object >= base + slab->objects * s->size ||
682 (object - base) % s->size) {
683 return 0;
684 }
685
686 return 1;
687 }
688
print_section(char * level,char * text,u8 * addr,unsigned int length)689 static void print_section(char *level, char *text, u8 *addr,
690 unsigned int length)
691 {
692 metadata_access_enable();
693 print_hex_dump(level, text, DUMP_PREFIX_ADDRESS,
694 16, 1, kasan_reset_tag((void *)addr), length, 1);
695 metadata_access_disable();
696 }
697
698 /*
699 * See comment in calculate_sizes().
700 */
freeptr_outside_object(struct kmem_cache * s)701 static inline bool freeptr_outside_object(struct kmem_cache *s)
702 {
703 return s->offset >= s->inuse;
704 }
705
706 /*
707 * Return offset of the end of info block which is inuse + free pointer if
708 * not overlapping with object.
709 */
get_info_end(struct kmem_cache * s)710 static inline unsigned int get_info_end(struct kmem_cache *s)
711 {
712 if (freeptr_outside_object(s))
713 return s->inuse + sizeof(void *);
714 else
715 return s->inuse;
716 }
717
get_track(struct kmem_cache * s,void * object,enum track_item alloc)718 static struct track *get_track(struct kmem_cache *s, void *object,
719 enum track_item alloc)
720 {
721 struct track *p;
722
723 p = object + get_info_end(s);
724
725 return kasan_reset_tag(p + alloc);
726 }
727
728 #ifdef CONFIG_STACKDEPOT
set_track_prepare(void)729 static noinline depot_stack_handle_t set_track_prepare(void)
730 {
731 depot_stack_handle_t handle;
732 unsigned long entries[TRACK_ADDRS_COUNT];
733 unsigned int nr_entries;
734
735 nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 3);
736 handle = stack_depot_save(entries, nr_entries, GFP_NOWAIT);
737
738 return handle;
739 }
740 #else
set_track_prepare(void)741 static inline depot_stack_handle_t set_track_prepare(void)
742 {
743 return 0;
744 }
745 #endif
746
set_track_update(struct kmem_cache * s,void * object,enum track_item alloc,unsigned long addr,depot_stack_handle_t handle)747 static void set_track_update(struct kmem_cache *s, void *object,
748 enum track_item alloc, unsigned long addr,
749 depot_stack_handle_t handle)
750 {
751 struct track *p = get_track(s, object, alloc);
752
753 #ifdef CONFIG_STACKDEPOT
754 p->handle = handle;
755 #endif
756 p->addr = addr;
757 p->cpu = smp_processor_id();
758 p->pid = current->pid;
759 p->when = jiffies;
760 }
761
set_track(struct kmem_cache * s,void * object,enum track_item alloc,unsigned long addr)762 static __always_inline void set_track(struct kmem_cache *s, void *object,
763 enum track_item alloc, unsigned long addr)
764 {
765 depot_stack_handle_t handle = set_track_prepare();
766
767 set_track_update(s, object, alloc, addr, handle);
768 }
769
init_tracking(struct kmem_cache * s,void * object)770 static void init_tracking(struct kmem_cache *s, void *object)
771 {
772 struct track *p;
773
774 if (!(s->flags & SLAB_STORE_USER))
775 return;
776
777 p = get_track(s, object, TRACK_ALLOC);
778 memset(p, 0, 2*sizeof(struct track));
779 }
780
print_track(const char * s,struct track * t,unsigned long pr_time)781 static void print_track(const char *s, struct track *t, unsigned long pr_time)
782 {
783 depot_stack_handle_t handle __maybe_unused;
784
785 if (!t->addr)
786 return;
787
788 pr_err("%s in %pS age=%lu cpu=%u pid=%d\n",
789 s, (void *)t->addr, pr_time - t->when, t->cpu, t->pid);
790 #ifdef CONFIG_STACKDEPOT
791 handle = READ_ONCE(t->handle);
792 if (handle)
793 stack_depot_print(handle);
794 else
795 pr_err("object allocation/free stack trace missing\n");
796 #endif
797 }
798
print_tracking(struct kmem_cache * s,void * object)799 void print_tracking(struct kmem_cache *s, void *object)
800 {
801 unsigned long pr_time = jiffies;
802 if (!(s->flags & SLAB_STORE_USER))
803 return;
804
805 print_track("Allocated", get_track(s, object, TRACK_ALLOC), pr_time);
806 print_track("Freed", get_track(s, object, TRACK_FREE), pr_time);
807 }
808
print_slab_info(const struct slab * slab)809 static void print_slab_info(const struct slab *slab)
810 {
811 struct folio *folio = (struct folio *)slab_folio(slab);
812
813 pr_err("Slab 0x%p objects=%u used=%u fp=0x%p flags=%pGp\n",
814 slab, slab->objects, slab->inuse, slab->freelist,
815 folio_flags(folio, 0));
816 }
817
818 /*
819 * kmalloc caches has fixed sizes (mostly power of 2), and kmalloc() API
820 * family will round up the real request size to these fixed ones, so
821 * there could be an extra area than what is requested. Save the original
822 * request size in the meta data area, for better debug and sanity check.
823 */
set_orig_size(struct kmem_cache * s,void * object,unsigned int orig_size)824 static inline void set_orig_size(struct kmem_cache *s,
825 void *object, unsigned int orig_size)
826 {
827 void *p = kasan_reset_tag(object);
828
829 if (!slub_debug_orig_size(s))
830 return;
831
832 p += get_info_end(s);
833 p += sizeof(struct track) * 2;
834
835 *(unsigned int *)p = orig_size;
836 }
837
get_orig_size(struct kmem_cache * s,void * object)838 static inline unsigned int get_orig_size(struct kmem_cache *s, void *object)
839 {
840 void *p = kasan_reset_tag(object);
841
842 if (!slub_debug_orig_size(s))
843 return s->object_size;
844
845 p += get_info_end(s);
846 p += sizeof(struct track) * 2;
847
848 return *(unsigned int *)p;
849 }
850
slab_bug(struct kmem_cache * s,char * fmt,...)851 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
852 {
853 struct va_format vaf;
854 va_list args;
855
856 va_start(args, fmt);
857 vaf.fmt = fmt;
858 vaf.va = &args;
859 pr_err("=============================================================================\n");
860 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
861 pr_err("-----------------------------------------------------------------------------\n\n");
862 va_end(args);
863 }
864
865 __printf(2, 3)
slab_fix(struct kmem_cache * s,char * fmt,...)866 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
867 {
868 struct va_format vaf;
869 va_list args;
870
871 if (slab_add_kunit_errors())
872 return;
873
874 va_start(args, fmt);
875 vaf.fmt = fmt;
876 vaf.va = &args;
877 pr_err("FIX %s: %pV\n", s->name, &vaf);
878 va_end(args);
879 }
880
print_trailer(struct kmem_cache * s,struct slab * slab,u8 * p)881 static void print_trailer(struct kmem_cache *s, struct slab *slab, u8 *p)
882 {
883 unsigned int off; /* Offset of last byte */
884 u8 *addr = slab_address(slab);
885
886 print_tracking(s, p);
887
888 print_slab_info(slab);
889
890 pr_err("Object 0x%p @offset=%tu fp=0x%p\n\n",
891 p, p - addr, get_freepointer(s, p));
892
893 if (s->flags & SLAB_RED_ZONE)
894 print_section(KERN_ERR, "Redzone ", p - s->red_left_pad,
895 s->red_left_pad);
896 else if (p > addr + 16)
897 print_section(KERN_ERR, "Bytes b4 ", p - 16, 16);
898
899 print_section(KERN_ERR, "Object ", p,
900 min_t(unsigned int, s->object_size, PAGE_SIZE));
901 if (s->flags & SLAB_RED_ZONE)
902 print_section(KERN_ERR, "Redzone ", p + s->object_size,
903 s->inuse - s->object_size);
904
905 off = get_info_end(s);
906
907 if (s->flags & SLAB_STORE_USER)
908 off += 2 * sizeof(struct track);
909
910 if (slub_debug_orig_size(s))
911 off += sizeof(unsigned int);
912
913 off += kasan_metadata_size(s);
914
915 if (off != size_from_object(s))
916 /* Beginning of the filler is the free pointer */
917 print_section(KERN_ERR, "Padding ", p + off,
918 size_from_object(s) - off);
919
920 dump_stack();
921 }
922
object_err(struct kmem_cache * s,struct slab * slab,u8 * object,char * reason)923 static void object_err(struct kmem_cache *s, struct slab *slab,
924 u8 *object, char *reason)
925 {
926 if (slab_add_kunit_errors())
927 return;
928
929 slab_bug(s, "%s", reason);
930 print_trailer(s, slab, object);
931 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
932 }
933
freelist_corrupted(struct kmem_cache * s,struct slab * slab,void ** freelist,void * nextfree)934 static bool freelist_corrupted(struct kmem_cache *s, struct slab *slab,
935 void **freelist, void *nextfree)
936 {
937 if ((s->flags & SLAB_CONSISTENCY_CHECKS) &&
938 !check_valid_pointer(s, slab, nextfree) && freelist) {
939 object_err(s, slab, *freelist, "Freechain corrupt");
940 *freelist = NULL;
941 slab_fix(s, "Isolate corrupted freechain");
942 return true;
943 }
944
945 return false;
946 }
947
slab_err(struct kmem_cache * s,struct slab * slab,const char * fmt,...)948 static __printf(3, 4) void slab_err(struct kmem_cache *s, struct slab *slab,
949 const char *fmt, ...)
950 {
951 va_list args;
952 char buf[100];
953
954 if (slab_add_kunit_errors())
955 return;
956
957 va_start(args, fmt);
958 vsnprintf(buf, sizeof(buf), fmt, args);
959 va_end(args);
960 slab_bug(s, "%s", buf);
961 print_slab_info(slab);
962 dump_stack();
963 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
964 }
965
init_object(struct kmem_cache * s,void * object,u8 val)966 static void init_object(struct kmem_cache *s, void *object, u8 val)
967 {
968 u8 *p = kasan_reset_tag(object);
969
970 if (s->flags & SLAB_RED_ZONE)
971 memset(p - s->red_left_pad, val, s->red_left_pad);
972
973 if (s->flags & __OBJECT_POISON) {
974 memset(p, POISON_FREE, s->object_size - 1);
975 p[s->object_size - 1] = POISON_END;
976 }
977
978 if (s->flags & SLAB_RED_ZONE)
979 memset(p + s->object_size, val, s->inuse - s->object_size);
980 }
981
restore_bytes(struct kmem_cache * s,char * message,u8 data,void * from,void * to)982 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
983 void *from, void *to)
984 {
985 slab_fix(s, "Restoring %s 0x%p-0x%p=0x%x", message, from, to - 1, data);
986 memset(from, data, to - from);
987 }
988
check_bytes_and_report(struct kmem_cache * s,struct slab * slab,u8 * object,char * what,u8 * start,unsigned int value,unsigned int bytes)989 static int check_bytes_and_report(struct kmem_cache *s, struct slab *slab,
990 u8 *object, char *what,
991 u8 *start, unsigned int value, unsigned int bytes)
992 {
993 u8 *fault;
994 u8 *end;
995 u8 *addr = slab_address(slab);
996
997 metadata_access_enable();
998 fault = memchr_inv(kasan_reset_tag(start), value, bytes);
999 metadata_access_disable();
1000 if (!fault)
1001 return 1;
1002
1003 end = start + bytes;
1004 while (end > fault && end[-1] == value)
1005 end--;
1006
1007 if (slab_add_kunit_errors())
1008 goto skip_bug_print;
1009
1010 slab_bug(s, "%s overwritten", what);
1011 pr_err("0x%p-0x%p @offset=%tu. First byte 0x%x instead of 0x%x\n",
1012 fault, end - 1, fault - addr,
1013 fault[0], value);
1014 print_trailer(s, slab, object);
1015 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
1016
1017 skip_bug_print:
1018 restore_bytes(s, what, value, fault, end);
1019 return 0;
1020 }
1021
1022 /*
1023 * Object layout:
1024 *
1025 * object address
1026 * Bytes of the object to be managed.
1027 * If the freepointer may overlay the object then the free
1028 * pointer is at the middle of the object.
1029 *
1030 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
1031 * 0xa5 (POISON_END)
1032 *
1033 * object + s->object_size
1034 * Padding to reach word boundary. This is also used for Redzoning.
1035 * Padding is extended by another word if Redzoning is enabled and
1036 * object_size == inuse.
1037 *
1038 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
1039 * 0xcc (RED_ACTIVE) for objects in use.
1040 *
1041 * object + s->inuse
1042 * Meta data starts here.
1043 *
1044 * A. Free pointer (if we cannot overwrite object on free)
1045 * B. Tracking data for SLAB_STORE_USER
1046 * C. Original request size for kmalloc object (SLAB_STORE_USER enabled)
1047 * D. Padding to reach required alignment boundary or at minimum
1048 * one word if debugging is on to be able to detect writes
1049 * before the word boundary.
1050 *
1051 * Padding is done using 0x5a (POISON_INUSE)
1052 *
1053 * object + s->size
1054 * Nothing is used beyond s->size.
1055 *
1056 * If slabcaches are merged then the object_size and inuse boundaries are mostly
1057 * ignored. And therefore no slab options that rely on these boundaries
1058 * may be used with merged slabcaches.
1059 */
1060
check_pad_bytes(struct kmem_cache * s,struct slab * slab,u8 * p)1061 static int check_pad_bytes(struct kmem_cache *s, struct slab *slab, u8 *p)
1062 {
1063 unsigned long off = get_info_end(s); /* The end of info */
1064
1065 if (s->flags & SLAB_STORE_USER) {
1066 /* We also have user information there */
1067 off += 2 * sizeof(struct track);
1068
1069 if (s->flags & SLAB_KMALLOC)
1070 off += sizeof(unsigned int);
1071 }
1072
1073 off += kasan_metadata_size(s);
1074
1075 if (size_from_object(s) == off)
1076 return 1;
1077
1078 return check_bytes_and_report(s, slab, p, "Object padding",
1079 p + off, POISON_INUSE, size_from_object(s) - off);
1080 }
1081
1082 /* Check the pad bytes at the end of a slab page */
slab_pad_check(struct kmem_cache * s,struct slab * slab)1083 static void slab_pad_check(struct kmem_cache *s, struct slab *slab)
1084 {
1085 u8 *start;
1086 u8 *fault;
1087 u8 *end;
1088 u8 *pad;
1089 int length;
1090 int remainder;
1091
1092 if (!(s->flags & SLAB_POISON))
1093 return;
1094
1095 start = slab_address(slab);
1096 length = slab_size(slab);
1097 end = start + length;
1098 remainder = length % s->size;
1099 if (!remainder)
1100 return;
1101
1102 pad = end - remainder;
1103 metadata_access_enable();
1104 fault = memchr_inv(kasan_reset_tag(pad), POISON_INUSE, remainder);
1105 metadata_access_disable();
1106 if (!fault)
1107 return;
1108 while (end > fault && end[-1] == POISON_INUSE)
1109 end--;
1110
1111 slab_err(s, slab, "Padding overwritten. 0x%p-0x%p @offset=%tu",
1112 fault, end - 1, fault - start);
1113 print_section(KERN_ERR, "Padding ", pad, remainder);
1114
1115 restore_bytes(s, "slab padding", POISON_INUSE, fault, end);
1116 }
1117
check_object(struct kmem_cache * s,struct slab * slab,void * object,u8 val)1118 static int check_object(struct kmem_cache *s, struct slab *slab,
1119 void *object, u8 val)
1120 {
1121 u8 *p = object;
1122 u8 *endobject = object + s->object_size;
1123
1124 if (s->flags & SLAB_RED_ZONE) {
1125 if (!check_bytes_and_report(s, slab, object, "Left Redzone",
1126 object - s->red_left_pad, val, s->red_left_pad))
1127 return 0;
1128
1129 if (!check_bytes_and_report(s, slab, object, "Right Redzone",
1130 endobject, val, s->inuse - s->object_size))
1131 return 0;
1132 } else {
1133 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
1134 check_bytes_and_report(s, slab, p, "Alignment padding",
1135 endobject, POISON_INUSE,
1136 s->inuse - s->object_size);
1137 }
1138 }
1139
1140 if (s->flags & SLAB_POISON) {
1141 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
1142 (!check_bytes_and_report(s, slab, p, "Poison", p,
1143 POISON_FREE, s->object_size - 1) ||
1144 !check_bytes_and_report(s, slab, p, "End Poison",
1145 p + s->object_size - 1, POISON_END, 1)))
1146 return 0;
1147 /*
1148 * check_pad_bytes cleans up on its own.
1149 */
1150 check_pad_bytes(s, slab, p);
1151 }
1152
1153 if (!freeptr_outside_object(s) && val == SLUB_RED_ACTIVE)
1154 /*
1155 * Object and freepointer overlap. Cannot check
1156 * freepointer while object is allocated.
1157 */
1158 return 1;
1159
1160 /* Check free pointer validity */
1161 if (!check_valid_pointer(s, slab, get_freepointer(s, p))) {
1162 object_err(s, slab, p, "Freepointer corrupt");
1163 /*
1164 * No choice but to zap it and thus lose the remainder
1165 * of the free objects in this slab. May cause
1166 * another error because the object count is now wrong.
1167 */
1168 set_freepointer(s, p, NULL);
1169 return 0;
1170 }
1171 return 1;
1172 }
1173
check_slab(struct kmem_cache * s,struct slab * slab)1174 static int check_slab(struct kmem_cache *s, struct slab *slab)
1175 {
1176 int maxobj;
1177
1178 if (!folio_test_slab(slab_folio(slab))) {
1179 slab_err(s, slab, "Not a valid slab page");
1180 return 0;
1181 }
1182
1183 maxobj = order_objects(slab_order(slab), s->size);
1184 if (slab->objects > maxobj) {
1185 slab_err(s, slab, "objects %u > max %u",
1186 slab->objects, maxobj);
1187 return 0;
1188 }
1189 if (slab->inuse > slab->objects) {
1190 slab_err(s, slab, "inuse %u > max %u",
1191 slab->inuse, slab->objects);
1192 return 0;
1193 }
1194 /* Slab_pad_check fixes things up after itself */
1195 slab_pad_check(s, slab);
1196 return 1;
1197 }
1198
1199 /*
1200 * Determine if a certain object in a slab is on the freelist. Must hold the
1201 * slab lock to guarantee that the chains are in a consistent state.
1202 */
on_freelist(struct kmem_cache * s,struct slab * slab,void * search)1203 static int on_freelist(struct kmem_cache *s, struct slab *slab, void *search)
1204 {
1205 int nr = 0;
1206 void *fp;
1207 void *object = NULL;
1208 int max_objects;
1209
1210 fp = slab->freelist;
1211 while (fp && nr <= slab->objects) {
1212 if (fp == search)
1213 return 1;
1214 if (!check_valid_pointer(s, slab, fp)) {
1215 if (object) {
1216 object_err(s, slab, object,
1217 "Freechain corrupt");
1218 set_freepointer(s, object, NULL);
1219 } else {
1220 slab_err(s, slab, "Freepointer corrupt");
1221 slab->freelist = NULL;
1222 slab->inuse = slab->objects;
1223 slab_fix(s, "Freelist cleared");
1224 return 0;
1225 }
1226 break;
1227 }
1228 object = fp;
1229 fp = get_freepointer(s, object);
1230 nr++;
1231 }
1232
1233 max_objects = order_objects(slab_order(slab), s->size);
1234 if (max_objects > MAX_OBJS_PER_PAGE)
1235 max_objects = MAX_OBJS_PER_PAGE;
1236
1237 if (slab->objects != max_objects) {
1238 slab_err(s, slab, "Wrong number of objects. Found %d but should be %d",
1239 slab->objects, max_objects);
1240 slab->objects = max_objects;
1241 slab_fix(s, "Number of objects adjusted");
1242 }
1243 if (slab->inuse != slab->objects - nr) {
1244 slab_err(s, slab, "Wrong object count. Counter is %d but counted were %d",
1245 slab->inuse, slab->objects - nr);
1246 slab->inuse = slab->objects - nr;
1247 slab_fix(s, "Object count adjusted");
1248 }
1249 return search == NULL;
1250 }
1251
trace(struct kmem_cache * s,struct slab * slab,void * object,int alloc)1252 static void trace(struct kmem_cache *s, struct slab *slab, void *object,
1253 int alloc)
1254 {
1255 if (s->flags & SLAB_TRACE) {
1256 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
1257 s->name,
1258 alloc ? "alloc" : "free",
1259 object, slab->inuse,
1260 slab->freelist);
1261
1262 if (!alloc)
1263 print_section(KERN_INFO, "Object ", (void *)object,
1264 s->object_size);
1265
1266 dump_stack();
1267 }
1268 }
1269
1270 /*
1271 * Tracking of fully allocated slabs for debugging purposes.
1272 */
add_full(struct kmem_cache * s,struct kmem_cache_node * n,struct slab * slab)1273 static void add_full(struct kmem_cache *s,
1274 struct kmem_cache_node *n, struct slab *slab)
1275 {
1276 if (!(s->flags & SLAB_STORE_USER))
1277 return;
1278
1279 lockdep_assert_held(&n->list_lock);
1280 list_add(&slab->slab_list, &n->full);
1281 }
1282
remove_full(struct kmem_cache * s,struct kmem_cache_node * n,struct slab * slab)1283 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct slab *slab)
1284 {
1285 if (!(s->flags & SLAB_STORE_USER))
1286 return;
1287
1288 lockdep_assert_held(&n->list_lock);
1289 list_del(&slab->slab_list);
1290 }
1291
1292 /* Tracking of the number of slabs for debugging purposes */
slabs_node(struct kmem_cache * s,int node)1293 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1294 {
1295 struct kmem_cache_node *n = get_node(s, node);
1296
1297 return atomic_long_read(&n->nr_slabs);
1298 }
1299
node_nr_slabs(struct kmem_cache_node * n)1300 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1301 {
1302 return atomic_long_read(&n->nr_slabs);
1303 }
1304
inc_slabs_node(struct kmem_cache * s,int node,int objects)1305 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1306 {
1307 struct kmem_cache_node *n = get_node(s, node);
1308
1309 /*
1310 * May be called early in order to allocate a slab for the
1311 * kmem_cache_node structure. Solve the chicken-egg
1312 * dilemma by deferring the increment of the count during
1313 * bootstrap (see early_kmem_cache_node_alloc).
1314 */
1315 if (likely(n)) {
1316 atomic_long_inc(&n->nr_slabs);
1317 atomic_long_add(objects, &n->total_objects);
1318 }
1319 }
dec_slabs_node(struct kmem_cache * s,int node,int objects)1320 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1321 {
1322 struct kmem_cache_node *n = get_node(s, node);
1323
1324 atomic_long_dec(&n->nr_slabs);
1325 atomic_long_sub(objects, &n->total_objects);
1326 }
1327
1328 /* Object debug checks for alloc/free paths */
setup_object_debug(struct kmem_cache * s,void * object)1329 static void setup_object_debug(struct kmem_cache *s, void *object)
1330 {
1331 if (!kmem_cache_debug_flags(s, SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))
1332 return;
1333
1334 init_object(s, object, SLUB_RED_INACTIVE);
1335 init_tracking(s, object);
1336 }
1337
1338 static
setup_slab_debug(struct kmem_cache * s,struct slab * slab,void * addr)1339 void setup_slab_debug(struct kmem_cache *s, struct slab *slab, void *addr)
1340 {
1341 if (!kmem_cache_debug_flags(s, SLAB_POISON))
1342 return;
1343
1344 metadata_access_enable();
1345 memset(kasan_reset_tag(addr), POISON_INUSE, slab_size(slab));
1346 metadata_access_disable();
1347 }
1348
alloc_consistency_checks(struct kmem_cache * s,struct slab * slab,void * object)1349 static inline int alloc_consistency_checks(struct kmem_cache *s,
1350 struct slab *slab, void *object)
1351 {
1352 if (!check_slab(s, slab))
1353 return 0;
1354
1355 if (!check_valid_pointer(s, slab, object)) {
1356 object_err(s, slab, object, "Freelist Pointer check fails");
1357 return 0;
1358 }
1359
1360 if (!check_object(s, slab, object, SLUB_RED_INACTIVE))
1361 return 0;
1362
1363 return 1;
1364 }
1365
alloc_debug_processing(struct kmem_cache * s,struct slab * slab,void * object,int orig_size)1366 static noinline int alloc_debug_processing(struct kmem_cache *s,
1367 struct slab *slab, void *object, int orig_size)
1368 {
1369 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1370 if (!alloc_consistency_checks(s, slab, object))
1371 goto bad;
1372 }
1373
1374 /* Success. Perform special debug activities for allocs */
1375 trace(s, slab, object, 1);
1376 set_orig_size(s, object, orig_size);
1377 init_object(s, object, SLUB_RED_ACTIVE);
1378 return 1;
1379
1380 bad:
1381 if (folio_test_slab(slab_folio(slab))) {
1382 /*
1383 * If this is a slab page then lets do the best we can
1384 * to avoid issues in the future. Marking all objects
1385 * as used avoids touching the remaining objects.
1386 */
1387 slab_fix(s, "Marking all objects used");
1388 slab->inuse = slab->objects;
1389 slab->freelist = NULL;
1390 }
1391 return 0;
1392 }
1393
free_consistency_checks(struct kmem_cache * s,struct slab * slab,void * object,unsigned long addr)1394 static inline int free_consistency_checks(struct kmem_cache *s,
1395 struct slab *slab, void *object, unsigned long addr)
1396 {
1397 if (!check_valid_pointer(s, slab, object)) {
1398 slab_err(s, slab, "Invalid object pointer 0x%p", object);
1399 return 0;
1400 }
1401
1402 if (on_freelist(s, slab, object)) {
1403 object_err(s, slab, object, "Object already free");
1404 return 0;
1405 }
1406
1407 if (!check_object(s, slab, object, SLUB_RED_ACTIVE))
1408 return 0;
1409
1410 if (unlikely(s != slab->slab_cache)) {
1411 if (!folio_test_slab(slab_folio(slab))) {
1412 slab_err(s, slab, "Attempt to free object(0x%p) outside of slab",
1413 object);
1414 } else if (!slab->slab_cache) {
1415 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1416 object);
1417 dump_stack();
1418 } else
1419 object_err(s, slab, object,
1420 "page slab pointer corrupt.");
1421 return 0;
1422 }
1423 return 1;
1424 }
1425
1426 /*
1427 * Parse a block of slub_debug options. Blocks are delimited by ';'
1428 *
1429 * @str: start of block
1430 * @flags: returns parsed flags, or DEBUG_DEFAULT_FLAGS if none specified
1431 * @slabs: return start of list of slabs, or NULL when there's no list
1432 * @init: assume this is initial parsing and not per-kmem-create parsing
1433 *
1434 * returns the start of next block if there's any, or NULL
1435 */
1436 static char *
parse_slub_debug_flags(char * str,slab_flags_t * flags,char ** slabs,bool init)1437 parse_slub_debug_flags(char *str, slab_flags_t *flags, char **slabs, bool init)
1438 {
1439 bool higher_order_disable = false;
1440
1441 /* Skip any completely empty blocks */
1442 while (*str && *str == ';')
1443 str++;
1444
1445 if (*str == ',') {
1446 /*
1447 * No options but restriction on slabs. This means full
1448 * debugging for slabs matching a pattern.
1449 */
1450 *flags = DEBUG_DEFAULT_FLAGS;
1451 goto check_slabs;
1452 }
1453 *flags = 0;
1454
1455 /* Determine which debug features should be switched on */
1456 for (; *str && *str != ',' && *str != ';'; str++) {
1457 switch (tolower(*str)) {
1458 case '-':
1459 *flags = 0;
1460 break;
1461 case 'f':
1462 *flags |= SLAB_CONSISTENCY_CHECKS;
1463 break;
1464 case 'z':
1465 *flags |= SLAB_RED_ZONE;
1466 break;
1467 case 'p':
1468 *flags |= SLAB_POISON;
1469 break;
1470 case 'u':
1471 *flags |= SLAB_STORE_USER;
1472 break;
1473 case 't':
1474 *flags |= SLAB_TRACE;
1475 break;
1476 case 'a':
1477 *flags |= SLAB_FAILSLAB;
1478 break;
1479 case 'o':
1480 /*
1481 * Avoid enabling debugging on caches if its minimum
1482 * order would increase as a result.
1483 */
1484 higher_order_disable = true;
1485 break;
1486 default:
1487 if (init)
1488 pr_err("slub_debug option '%c' unknown. skipped\n", *str);
1489 }
1490 }
1491 check_slabs:
1492 if (*str == ',')
1493 *slabs = ++str;
1494 else
1495 *slabs = NULL;
1496
1497 /* Skip over the slab list */
1498 while (*str && *str != ';')
1499 str++;
1500
1501 /* Skip any completely empty blocks */
1502 while (*str && *str == ';')
1503 str++;
1504
1505 if (init && higher_order_disable)
1506 disable_higher_order_debug = 1;
1507
1508 if (*str)
1509 return str;
1510 else
1511 return NULL;
1512 }
1513
setup_slub_debug(char * str)1514 static int __init setup_slub_debug(char *str)
1515 {
1516 slab_flags_t flags;
1517 slab_flags_t global_flags;
1518 char *saved_str;
1519 char *slab_list;
1520 bool global_slub_debug_changed = false;
1521 bool slab_list_specified = false;
1522
1523 global_flags = DEBUG_DEFAULT_FLAGS;
1524 if (*str++ != '=' || !*str)
1525 /*
1526 * No options specified. Switch on full debugging.
1527 */
1528 goto out;
1529
1530 saved_str = str;
1531 while (str) {
1532 str = parse_slub_debug_flags(str, &flags, &slab_list, true);
1533
1534 if (!slab_list) {
1535 global_flags = flags;
1536 global_slub_debug_changed = true;
1537 } else {
1538 slab_list_specified = true;
1539 if (flags & SLAB_STORE_USER)
1540 stack_depot_want_early_init();
1541 }
1542 }
1543
1544 /*
1545 * For backwards compatibility, a single list of flags with list of
1546 * slabs means debugging is only changed for those slabs, so the global
1547 * slub_debug should be unchanged (0 or DEBUG_DEFAULT_FLAGS, depending
1548 * on CONFIG_SLUB_DEBUG_ON). We can extended that to multiple lists as
1549 * long as there is no option specifying flags without a slab list.
1550 */
1551 if (slab_list_specified) {
1552 if (!global_slub_debug_changed)
1553 global_flags = slub_debug;
1554 slub_debug_string = saved_str;
1555 }
1556 out:
1557 slub_debug = global_flags;
1558 if (slub_debug & SLAB_STORE_USER)
1559 stack_depot_want_early_init();
1560 if (slub_debug != 0 || slub_debug_string)
1561 static_branch_enable(&slub_debug_enabled);
1562 else
1563 static_branch_disable(&slub_debug_enabled);
1564 if ((static_branch_unlikely(&init_on_alloc) ||
1565 static_branch_unlikely(&init_on_free)) &&
1566 (slub_debug & SLAB_POISON))
1567 pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n");
1568 return 1;
1569 }
1570
1571 __setup("slub_debug", setup_slub_debug);
1572
1573 /*
1574 * kmem_cache_flags - apply debugging options to the cache
1575 * @object_size: the size of an object without meta data
1576 * @flags: flags to set
1577 * @name: name of the cache
1578 *
1579 * Debug option(s) are applied to @flags. In addition to the debug
1580 * option(s), if a slab name (or multiple) is specified i.e.
1581 * slub_debug=<Debug-Options>,<slab name1>,<slab name2> ...
1582 * then only the select slabs will receive the debug option(s).
1583 */
kmem_cache_flags(unsigned int object_size,slab_flags_t flags,const char * name)1584 slab_flags_t kmem_cache_flags(unsigned int object_size,
1585 slab_flags_t flags, const char *name)
1586 {
1587 char *iter;
1588 size_t len;
1589 char *next_block;
1590 slab_flags_t block_flags;
1591 slab_flags_t slub_debug_local = slub_debug;
1592
1593 if (flags & SLAB_NO_USER_FLAGS)
1594 return flags;
1595
1596 /*
1597 * If the slab cache is for debugging (e.g. kmemleak) then
1598 * don't store user (stack trace) information by default,
1599 * but let the user enable it via the command line below.
1600 */
1601 if (flags & SLAB_NOLEAKTRACE)
1602 slub_debug_local &= ~SLAB_STORE_USER;
1603
1604 len = strlen(name);
1605 next_block = slub_debug_string;
1606 /* Go through all blocks of debug options, see if any matches our slab's name */
1607 while (next_block) {
1608 next_block = parse_slub_debug_flags(next_block, &block_flags, &iter, false);
1609 if (!iter)
1610 continue;
1611 /* Found a block that has a slab list, search it */
1612 while (*iter) {
1613 char *end, *glob;
1614 size_t cmplen;
1615
1616 end = strchrnul(iter, ',');
1617 if (next_block && next_block < end)
1618 end = next_block - 1;
1619
1620 glob = strnchr(iter, end - iter, '*');
1621 if (glob)
1622 cmplen = glob - iter;
1623 else
1624 cmplen = max_t(size_t, len, (end - iter));
1625
1626 if (!strncmp(name, iter, cmplen)) {
1627 flags |= block_flags;
1628 return flags;
1629 }
1630
1631 if (!*end || *end == ';')
1632 break;
1633 iter = end + 1;
1634 }
1635 }
1636
1637 return flags | slub_debug_local;
1638 }
1639 #else /* !CONFIG_SLUB_DEBUG */
setup_object_debug(struct kmem_cache * s,void * object)1640 static inline void setup_object_debug(struct kmem_cache *s, void *object) {}
1641 static inline
setup_slab_debug(struct kmem_cache * s,struct slab * slab,void * addr)1642 void setup_slab_debug(struct kmem_cache *s, struct slab *slab, void *addr) {}
1643
alloc_debug_processing(struct kmem_cache * s,struct slab * slab,void * object,int orig_size)1644 static inline int alloc_debug_processing(struct kmem_cache *s,
1645 struct slab *slab, void *object, int orig_size) { return 0; }
1646
free_debug_processing(struct kmem_cache * s,struct slab * slab,void * head,void * tail,int bulk_cnt,unsigned long addr)1647 static inline void free_debug_processing(
1648 struct kmem_cache *s, struct slab *slab,
1649 void *head, void *tail, int bulk_cnt,
1650 unsigned long addr) {}
1651
slab_pad_check(struct kmem_cache * s,struct slab * slab)1652 static inline void slab_pad_check(struct kmem_cache *s, struct slab *slab) {}
check_object(struct kmem_cache * s,struct slab * slab,void * object,u8 val)1653 static inline int check_object(struct kmem_cache *s, struct slab *slab,
1654 void *object, u8 val) { return 1; }
set_track(struct kmem_cache * s,void * object,enum track_item alloc,unsigned long addr)1655 static inline void set_track(struct kmem_cache *s, void *object,
1656 enum track_item alloc, unsigned long addr) {}
add_full(struct kmem_cache * s,struct kmem_cache_node * n,struct slab * slab)1657 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1658 struct slab *slab) {}
remove_full(struct kmem_cache * s,struct kmem_cache_node * n,struct slab * slab)1659 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1660 struct slab *slab) {}
kmem_cache_flags(unsigned int object_size,slab_flags_t flags,const char * name)1661 slab_flags_t kmem_cache_flags(unsigned int object_size,
1662 slab_flags_t flags, const char *name)
1663 {
1664 return flags;
1665 }
1666 #define slub_debug 0
1667
1668 #define disable_higher_order_debug 0
1669
slabs_node(struct kmem_cache * s,int node)1670 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1671 { return 0; }
node_nr_slabs(struct kmem_cache_node * n)1672 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1673 { return 0; }
inc_slabs_node(struct kmem_cache * s,int node,int objects)1674 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1675 int objects) {}
dec_slabs_node(struct kmem_cache * s,int node,int objects)1676 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1677 int objects) {}
1678
freelist_corrupted(struct kmem_cache * s,struct slab * slab,void ** freelist,void * nextfree)1679 static bool freelist_corrupted(struct kmem_cache *s, struct slab *slab,
1680 void **freelist, void *nextfree)
1681 {
1682 return false;
1683 }
1684 #endif /* CONFIG_SLUB_DEBUG */
1685
1686 /*
1687 * Hooks for other subsystems that check memory allocations. In a typical
1688 * production configuration these hooks all should produce no code at all.
1689 */
slab_free_hook(struct kmem_cache * s,void * x,bool init)1690 static __always_inline bool slab_free_hook(struct kmem_cache *s,
1691 void *x, bool init)
1692 {
1693 kmemleak_free_recursive(x, s->flags);
1694 kmsan_slab_free(s, x);
1695
1696 debug_check_no_locks_freed(x, s->object_size);
1697
1698 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1699 debug_check_no_obj_freed(x, s->object_size);
1700
1701 /* Use KCSAN to help debug racy use-after-free. */
1702 if (!(s->flags & SLAB_TYPESAFE_BY_RCU))
1703 __kcsan_check_access(x, s->object_size,
1704 KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT);
1705
1706 /*
1707 * As memory initialization might be integrated into KASAN,
1708 * kasan_slab_free and initialization memset's must be
1709 * kept together to avoid discrepancies in behavior.
1710 *
1711 * The initialization memset's clear the object and the metadata,
1712 * but don't touch the SLAB redzone.
1713 */
1714 if (init) {
1715 int rsize;
1716
1717 if (!kasan_has_integrated_init())
1718 memset(kasan_reset_tag(x), 0, s->object_size);
1719 rsize = (s->flags & SLAB_RED_ZONE) ? s->red_left_pad : 0;
1720 memset((char *)kasan_reset_tag(x) + s->inuse, 0,
1721 s->size - s->inuse - rsize);
1722 }
1723 /* KASAN might put x into memory quarantine, delaying its reuse. */
1724 return kasan_slab_free(s, x, init);
1725 }
1726
slab_free_freelist_hook(struct kmem_cache * s,void ** head,void ** tail,int * cnt)1727 static inline bool slab_free_freelist_hook(struct kmem_cache *s,
1728 void **head, void **tail,
1729 int *cnt)
1730 {
1731
1732 void *object;
1733 void *next = *head;
1734 void *old_tail = *tail ? *tail : *head;
1735
1736 if (is_kfence_address(next)) {
1737 slab_free_hook(s, next, false);
1738 return true;
1739 }
1740
1741 /* Head and tail of the reconstructed freelist */
1742 *head = NULL;
1743 *tail = NULL;
1744
1745 do {
1746 object = next;
1747 next = get_freepointer(s, object);
1748
1749 /* If object's reuse doesn't have to be delayed */
1750 if (!slab_free_hook(s, object, slab_want_init_on_free(s))) {
1751 /* Move object to the new freelist */
1752 set_freepointer(s, object, *head);
1753 *head = object;
1754 if (!*tail)
1755 *tail = object;
1756 } else {
1757 /*
1758 * Adjust the reconstructed freelist depth
1759 * accordingly if object's reuse is delayed.
1760 */
1761 --(*cnt);
1762 }
1763 } while (object != old_tail);
1764
1765 if (*head == *tail)
1766 *tail = NULL;
1767
1768 return *head != NULL;
1769 }
1770
setup_object(struct kmem_cache * s,void * object)1771 static void *setup_object(struct kmem_cache *s, void *object)
1772 {
1773 setup_object_debug(s, object);
1774 object = kasan_init_slab_obj(s, object);
1775 if (unlikely(s->ctor)) {
1776 kasan_unpoison_object_data(s, object);
1777 s->ctor(object);
1778 kasan_poison_object_data(s, object);
1779 }
1780 return object;
1781 }
1782
1783 /*
1784 * Slab allocation and freeing
1785 */
alloc_slab_page(gfp_t flags,int node,struct kmem_cache_order_objects oo)1786 static inline struct slab *alloc_slab_page(gfp_t flags, int node,
1787 struct kmem_cache_order_objects oo)
1788 {
1789 struct folio *folio;
1790 struct slab *slab;
1791 unsigned int order = oo_order(oo);
1792
1793 if (node == NUMA_NO_NODE)
1794 folio = (struct folio *)alloc_pages(flags, order);
1795 else
1796 folio = (struct folio *)__alloc_pages_node(node, flags, order);
1797
1798 if (!folio)
1799 return NULL;
1800
1801 slab = folio_slab(folio);
1802 __folio_set_slab(folio);
1803 if (page_is_pfmemalloc(folio_page(folio, 0)))
1804 slab_set_pfmemalloc(slab);
1805
1806 return slab;
1807 }
1808
1809 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1810 /* Pre-initialize the random sequence cache */
init_cache_random_seq(struct kmem_cache * s)1811 static int init_cache_random_seq(struct kmem_cache *s)
1812 {
1813 unsigned int count = oo_objects(s->oo);
1814 int err;
1815
1816 /* Bailout if already initialised */
1817 if (s->random_seq)
1818 return 0;
1819
1820 err = cache_random_seq_create(s, count, GFP_KERNEL);
1821 if (err) {
1822 pr_err("SLUB: Unable to initialize free list for %s\n",
1823 s->name);
1824 return err;
1825 }
1826
1827 /* Transform to an offset on the set of pages */
1828 if (s->random_seq) {
1829 unsigned int i;
1830
1831 for (i = 0; i < count; i++)
1832 s->random_seq[i] *= s->size;
1833 }
1834 return 0;
1835 }
1836
1837 /* Initialize each random sequence freelist per cache */
init_freelist_randomization(void)1838 static void __init init_freelist_randomization(void)
1839 {
1840 struct kmem_cache *s;
1841
1842 mutex_lock(&slab_mutex);
1843
1844 list_for_each_entry(s, &slab_caches, list)
1845 init_cache_random_seq(s);
1846
1847 mutex_unlock(&slab_mutex);
1848 }
1849
1850 /* Get the next entry on the pre-computed freelist randomized */
next_freelist_entry(struct kmem_cache * s,struct slab * slab,unsigned long * pos,void * start,unsigned long page_limit,unsigned long freelist_count)1851 static void *next_freelist_entry(struct kmem_cache *s, struct slab *slab,
1852 unsigned long *pos, void *start,
1853 unsigned long page_limit,
1854 unsigned long freelist_count)
1855 {
1856 unsigned int idx;
1857
1858 /*
1859 * If the target page allocation failed, the number of objects on the
1860 * page might be smaller than the usual size defined by the cache.
1861 */
1862 do {
1863 idx = s->random_seq[*pos];
1864 *pos += 1;
1865 if (*pos >= freelist_count)
1866 *pos = 0;
1867 } while (unlikely(idx >= page_limit));
1868
1869 return (char *)start + idx;
1870 }
1871
1872 /* Shuffle the single linked freelist based on a random pre-computed sequence */
shuffle_freelist(struct kmem_cache * s,struct slab * slab)1873 static bool shuffle_freelist(struct kmem_cache *s, struct slab *slab)
1874 {
1875 void *start;
1876 void *cur;
1877 void *next;
1878 unsigned long idx, pos, page_limit, freelist_count;
1879
1880 if (slab->objects < 2 || !s->random_seq)
1881 return false;
1882
1883 freelist_count = oo_objects(s->oo);
1884 pos = prandom_u32_max(freelist_count);
1885
1886 page_limit = slab->objects * s->size;
1887 start = fixup_red_left(s, slab_address(slab));
1888
1889 /* First entry is used as the base of the freelist */
1890 cur = next_freelist_entry(s, slab, &pos, start, page_limit,
1891 freelist_count);
1892 cur = setup_object(s, cur);
1893 slab->freelist = cur;
1894
1895 for (idx = 1; idx < slab->objects; idx++) {
1896 next = next_freelist_entry(s, slab, &pos, start, page_limit,
1897 freelist_count);
1898 next = setup_object(s, next);
1899 set_freepointer(s, cur, next);
1900 cur = next;
1901 }
1902 set_freepointer(s, cur, NULL);
1903
1904 return true;
1905 }
1906 #else
init_cache_random_seq(struct kmem_cache * s)1907 static inline int init_cache_random_seq(struct kmem_cache *s)
1908 {
1909 return 0;
1910 }
init_freelist_randomization(void)1911 static inline void init_freelist_randomization(void) { }
shuffle_freelist(struct kmem_cache * s,struct slab * slab)1912 static inline bool shuffle_freelist(struct kmem_cache *s, struct slab *slab)
1913 {
1914 return false;
1915 }
1916 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1917
allocate_slab(struct kmem_cache * s,gfp_t flags,int node)1918 static struct slab *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1919 {
1920 struct slab *slab;
1921 struct kmem_cache_order_objects oo = s->oo;
1922 gfp_t alloc_gfp;
1923 void *start, *p, *next;
1924 int idx;
1925 bool shuffle;
1926
1927 flags &= gfp_allowed_mask;
1928
1929 flags |= s->allocflags;
1930
1931 /*
1932 * Let the initial higher-order allocation fail under memory pressure
1933 * so we fall-back to the minimum order allocation.
1934 */
1935 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1936 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1937 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~__GFP_RECLAIM;
1938
1939 slab = alloc_slab_page(alloc_gfp, node, oo);
1940 if (unlikely(!slab)) {
1941 oo = s->min;
1942 alloc_gfp = flags;
1943 /*
1944 * Allocation may have failed due to fragmentation.
1945 * Try a lower order alloc if possible
1946 */
1947 slab = alloc_slab_page(alloc_gfp, node, oo);
1948 if (unlikely(!slab))
1949 return NULL;
1950 stat(s, ORDER_FALLBACK);
1951 }
1952
1953 slab->objects = oo_objects(oo);
1954 slab->inuse = 0;
1955 slab->frozen = 0;
1956
1957 account_slab(slab, oo_order(oo), s, flags);
1958
1959 slab->slab_cache = s;
1960
1961 kasan_poison_slab(slab);
1962
1963 start = slab_address(slab);
1964
1965 setup_slab_debug(s, slab, start);
1966
1967 shuffle = shuffle_freelist(s, slab);
1968
1969 if (!shuffle) {
1970 start = fixup_red_left(s, start);
1971 start = setup_object(s, start);
1972 slab->freelist = start;
1973 for (idx = 0, p = start; idx < slab->objects - 1; idx++) {
1974 next = p + s->size;
1975 next = setup_object(s, next);
1976 set_freepointer(s, p, next);
1977 p = next;
1978 }
1979 set_freepointer(s, p, NULL);
1980 }
1981
1982 return slab;
1983 }
1984
new_slab(struct kmem_cache * s,gfp_t flags,int node)1985 static struct slab *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1986 {
1987 if (unlikely(flags & GFP_SLAB_BUG_MASK))
1988 flags = kmalloc_fix_flags(flags);
1989
1990 WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO));
1991
1992 return allocate_slab(s,
1993 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1994 }
1995
__free_slab(struct kmem_cache * s,struct slab * slab)1996 static void __free_slab(struct kmem_cache *s, struct slab *slab)
1997 {
1998 struct folio *folio = slab_folio(slab);
1999 int order = folio_order(folio);
2000 int pages = 1 << order;
2001
2002 if (kmem_cache_debug_flags(s, SLAB_CONSISTENCY_CHECKS)) {
2003 void *p;
2004
2005 slab_pad_check(s, slab);
2006 for_each_object(p, s, slab_address(slab), slab->objects)
2007 check_object(s, slab, p, SLUB_RED_INACTIVE);
2008 }
2009
2010 __slab_clear_pfmemalloc(slab);
2011 __folio_clear_slab(folio);
2012 folio->mapping = NULL;
2013 if (current->reclaim_state)
2014 current->reclaim_state->reclaimed_slab += pages;
2015 unaccount_slab(slab, order, s);
2016 __free_pages(folio_page(folio, 0), order);
2017 }
2018
rcu_free_slab(struct rcu_head * h)2019 static void rcu_free_slab(struct rcu_head *h)
2020 {
2021 struct slab *slab = container_of(h, struct slab, rcu_head);
2022
2023 __free_slab(slab->slab_cache, slab);
2024 }
2025
free_slab(struct kmem_cache * s,struct slab * slab)2026 static void free_slab(struct kmem_cache *s, struct slab *slab)
2027 {
2028 if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU)) {
2029 call_rcu(&slab->rcu_head, rcu_free_slab);
2030 } else
2031 __free_slab(s, slab);
2032 }
2033
discard_slab(struct kmem_cache * s,struct slab * slab)2034 static void discard_slab(struct kmem_cache *s, struct slab *slab)
2035 {
2036 dec_slabs_node(s, slab_nid(slab), slab->objects);
2037 free_slab(s, slab);
2038 }
2039
2040 /*
2041 * Management of partially allocated slabs.
2042 */
2043 static inline void
__add_partial(struct kmem_cache_node * n,struct slab * slab,int tail)2044 __add_partial(struct kmem_cache_node *n, struct slab *slab, int tail)
2045 {
2046 n->nr_partial++;
2047 if (tail == DEACTIVATE_TO_TAIL)
2048 list_add_tail(&slab->slab_list, &n->partial);
2049 else
2050 list_add(&slab->slab_list, &n->partial);
2051 }
2052
add_partial(struct kmem_cache_node * n,struct slab * slab,int tail)2053 static inline void add_partial(struct kmem_cache_node *n,
2054 struct slab *slab, int tail)
2055 {
2056 lockdep_assert_held(&n->list_lock);
2057 __add_partial(n, slab, tail);
2058 }
2059
remove_partial(struct kmem_cache_node * n,struct slab * slab)2060 static inline void remove_partial(struct kmem_cache_node *n,
2061 struct slab *slab)
2062 {
2063 lockdep_assert_held(&n->list_lock);
2064 list_del(&slab->slab_list);
2065 n->nr_partial--;
2066 }
2067
2068 /*
2069 * Called only for kmem_cache_debug() caches instead of acquire_slab(), with a
2070 * slab from the n->partial list. Remove only a single object from the slab, do
2071 * the alloc_debug_processing() checks and leave the slab on the list, or move
2072 * it to full list if it was the last free object.
2073 */
alloc_single_from_partial(struct kmem_cache * s,struct kmem_cache_node * n,struct slab * slab,int orig_size)2074 static void *alloc_single_from_partial(struct kmem_cache *s,
2075 struct kmem_cache_node *n, struct slab *slab, int orig_size)
2076 {
2077 void *object;
2078
2079 lockdep_assert_held(&n->list_lock);
2080
2081 object = slab->freelist;
2082 slab->freelist = get_freepointer(s, object);
2083 slab->inuse++;
2084
2085 if (!alloc_debug_processing(s, slab, object, orig_size)) {
2086 remove_partial(n, slab);
2087 return NULL;
2088 }
2089
2090 if (slab->inuse == slab->objects) {
2091 remove_partial(n, slab);
2092 add_full(s, n, slab);
2093 }
2094
2095 return object;
2096 }
2097
2098 /*
2099 * Called only for kmem_cache_debug() caches to allocate from a freshly
2100 * allocated slab. Allocate a single object instead of whole freelist
2101 * and put the slab to the partial (or full) list.
2102 */
alloc_single_from_new_slab(struct kmem_cache * s,struct slab * slab,int orig_size)2103 static void *alloc_single_from_new_slab(struct kmem_cache *s,
2104 struct slab *slab, int orig_size)
2105 {
2106 int nid = slab_nid(slab);
2107 struct kmem_cache_node *n = get_node(s, nid);
2108 unsigned long flags;
2109 void *object;
2110
2111
2112 object = slab->freelist;
2113 slab->freelist = get_freepointer(s, object);
2114 slab->inuse = 1;
2115
2116 if (!alloc_debug_processing(s, slab, object, orig_size))
2117 /*
2118 * It's not really expected that this would fail on a
2119 * freshly allocated slab, but a concurrent memory
2120 * corruption in theory could cause that.
2121 */
2122 return NULL;
2123
2124 spin_lock_irqsave(&n->list_lock, flags);
2125
2126 if (slab->inuse == slab->objects)
2127 add_full(s, n, slab);
2128 else
2129 add_partial(n, slab, DEACTIVATE_TO_HEAD);
2130
2131 inc_slabs_node(s, nid, slab->objects);
2132 spin_unlock_irqrestore(&n->list_lock, flags);
2133
2134 return object;
2135 }
2136
2137 /*
2138 * Remove slab from the partial list, freeze it and
2139 * return the pointer to the freelist.
2140 *
2141 * Returns a list of objects or NULL if it fails.
2142 */
acquire_slab(struct kmem_cache * s,struct kmem_cache_node * n,struct slab * slab,int mode)2143 static inline void *acquire_slab(struct kmem_cache *s,
2144 struct kmem_cache_node *n, struct slab *slab,
2145 int mode)
2146 {
2147 void *freelist;
2148 unsigned long counters;
2149 struct slab new;
2150
2151 lockdep_assert_held(&n->list_lock);
2152
2153 /*
2154 * Zap the freelist and set the frozen bit.
2155 * The old freelist is the list of objects for the
2156 * per cpu allocation list.
2157 */
2158 freelist = slab->freelist;
2159 counters = slab->counters;
2160 new.counters = counters;
2161 if (mode) {
2162 new.inuse = slab->objects;
2163 new.freelist = NULL;
2164 } else {
2165 new.freelist = freelist;
2166 }
2167
2168 VM_BUG_ON(new.frozen);
2169 new.frozen = 1;
2170
2171 if (!__cmpxchg_double_slab(s, slab,
2172 freelist, counters,
2173 new.freelist, new.counters,
2174 "acquire_slab"))
2175 return NULL;
2176
2177 remove_partial(n, slab);
2178 WARN_ON(!freelist);
2179 return freelist;
2180 }
2181
2182 #ifdef CONFIG_SLUB_CPU_PARTIAL
2183 static void put_cpu_partial(struct kmem_cache *s, struct slab *slab, int drain);
2184 #else
put_cpu_partial(struct kmem_cache * s,struct slab * slab,int drain)2185 static inline void put_cpu_partial(struct kmem_cache *s, struct slab *slab,
2186 int drain) { }
2187 #endif
2188 static inline bool pfmemalloc_match(struct slab *slab, gfp_t gfpflags);
2189
2190 /*
2191 * Try to allocate a partial slab from a specific node.
2192 */
get_partial_node(struct kmem_cache * s,struct kmem_cache_node * n,struct partial_context * pc)2193 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
2194 struct partial_context *pc)
2195 {
2196 struct slab *slab, *slab2;
2197 void *object = NULL;
2198 unsigned long flags;
2199 unsigned int partial_slabs = 0;
2200
2201 /*
2202 * Racy check. If we mistakenly see no partial slabs then we
2203 * just allocate an empty slab. If we mistakenly try to get a
2204 * partial slab and there is none available then get_partial()
2205 * will return NULL.
2206 */
2207 if (!n || !n->nr_partial)
2208 return NULL;
2209
2210 spin_lock_irqsave(&n->list_lock, flags);
2211 list_for_each_entry_safe(slab, slab2, &n->partial, slab_list) {
2212 void *t;
2213
2214 if (!pfmemalloc_match(slab, pc->flags))
2215 continue;
2216
2217 if (kmem_cache_debug(s)) {
2218 object = alloc_single_from_partial(s, n, slab,
2219 pc->orig_size);
2220 if (object)
2221 break;
2222 continue;
2223 }
2224
2225 t = acquire_slab(s, n, slab, object == NULL);
2226 if (!t)
2227 break;
2228
2229 if (!object) {
2230 *pc->slab = slab;
2231 stat(s, ALLOC_FROM_PARTIAL);
2232 object = t;
2233 } else {
2234 put_cpu_partial(s, slab, 0);
2235 stat(s, CPU_PARTIAL_NODE);
2236 partial_slabs++;
2237 }
2238 #ifdef CONFIG_SLUB_CPU_PARTIAL
2239 if (!kmem_cache_has_cpu_partial(s)
2240 || partial_slabs > s->cpu_partial_slabs / 2)
2241 break;
2242 #else
2243 break;
2244 #endif
2245
2246 }
2247 spin_unlock_irqrestore(&n->list_lock, flags);
2248 return object;
2249 }
2250
2251 /*
2252 * Get a slab from somewhere. Search in increasing NUMA distances.
2253 */
get_any_partial(struct kmem_cache * s,struct partial_context * pc)2254 static void *get_any_partial(struct kmem_cache *s, struct partial_context *pc)
2255 {
2256 #ifdef CONFIG_NUMA
2257 struct zonelist *zonelist;
2258 struct zoneref *z;
2259 struct zone *zone;
2260 enum zone_type highest_zoneidx = gfp_zone(pc->flags);
2261 void *object;
2262 unsigned int cpuset_mems_cookie;
2263
2264 /*
2265 * The defrag ratio allows a configuration of the tradeoffs between
2266 * inter node defragmentation and node local allocations. A lower
2267 * defrag_ratio increases the tendency to do local allocations
2268 * instead of attempting to obtain partial slabs from other nodes.
2269 *
2270 * If the defrag_ratio is set to 0 then kmalloc() always
2271 * returns node local objects. If the ratio is higher then kmalloc()
2272 * may return off node objects because partial slabs are obtained
2273 * from other nodes and filled up.
2274 *
2275 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
2276 * (which makes defrag_ratio = 1000) then every (well almost)
2277 * allocation will first attempt to defrag slab caches on other nodes.
2278 * This means scanning over all nodes to look for partial slabs which
2279 * may be expensive if we do it every time we are trying to find a slab
2280 * with available objects.
2281 */
2282 if (!s->remote_node_defrag_ratio ||
2283 get_cycles() % 1024 > s->remote_node_defrag_ratio)
2284 return NULL;
2285
2286 do {
2287 cpuset_mems_cookie = read_mems_allowed_begin();
2288 zonelist = node_zonelist(mempolicy_slab_node(), pc->flags);
2289 for_each_zone_zonelist(zone, z, zonelist, highest_zoneidx) {
2290 struct kmem_cache_node *n;
2291
2292 n = get_node(s, zone_to_nid(zone));
2293
2294 if (n && cpuset_zone_allowed(zone, pc->flags) &&
2295 n->nr_partial > s->min_partial) {
2296 object = get_partial_node(s, n, pc);
2297 if (object) {
2298 /*
2299 * Don't check read_mems_allowed_retry()
2300 * here - if mems_allowed was updated in
2301 * parallel, that was a harmless race
2302 * between allocation and the cpuset
2303 * update
2304 */
2305 return object;
2306 }
2307 }
2308 }
2309 } while (read_mems_allowed_retry(cpuset_mems_cookie));
2310 #endif /* CONFIG_NUMA */
2311 return NULL;
2312 }
2313
2314 /*
2315 * Get a partial slab, lock it and return it.
2316 */
get_partial(struct kmem_cache * s,int node,struct partial_context * pc)2317 static void *get_partial(struct kmem_cache *s, int node, struct partial_context *pc)
2318 {
2319 void *object;
2320 int searchnode = node;
2321
2322 if (node == NUMA_NO_NODE)
2323 searchnode = numa_mem_id();
2324
2325 object = get_partial_node(s, get_node(s, searchnode), pc);
2326 if (object || node != NUMA_NO_NODE)
2327 return object;
2328
2329 return get_any_partial(s, pc);
2330 }
2331
2332 #ifdef CONFIG_PREEMPTION
2333 /*
2334 * Calculate the next globally unique transaction for disambiguation
2335 * during cmpxchg. The transactions start with the cpu number and are then
2336 * incremented by CONFIG_NR_CPUS.
2337 */
2338 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
2339 #else
2340 /*
2341 * No preemption supported therefore also no need to check for
2342 * different cpus.
2343 */
2344 #define TID_STEP 1
2345 #endif
2346
next_tid(unsigned long tid)2347 static inline unsigned long next_tid(unsigned long tid)
2348 {
2349 return tid + TID_STEP;
2350 }
2351
2352 #ifdef SLUB_DEBUG_CMPXCHG
tid_to_cpu(unsigned long tid)2353 static inline unsigned int tid_to_cpu(unsigned long tid)
2354 {
2355 return tid % TID_STEP;
2356 }
2357
tid_to_event(unsigned long tid)2358 static inline unsigned long tid_to_event(unsigned long tid)
2359 {
2360 return tid / TID_STEP;
2361 }
2362 #endif
2363
init_tid(int cpu)2364 static inline unsigned int init_tid(int cpu)
2365 {
2366 return cpu;
2367 }
2368
note_cmpxchg_failure(const char * n,const struct kmem_cache * s,unsigned long tid)2369 static inline void note_cmpxchg_failure(const char *n,
2370 const struct kmem_cache *s, unsigned long tid)
2371 {
2372 #ifdef SLUB_DEBUG_CMPXCHG
2373 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
2374
2375 pr_info("%s %s: cmpxchg redo ", n, s->name);
2376
2377 #ifdef CONFIG_PREEMPTION
2378 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
2379 pr_warn("due to cpu change %d -> %d\n",
2380 tid_to_cpu(tid), tid_to_cpu(actual_tid));
2381 else
2382 #endif
2383 if (tid_to_event(tid) != tid_to_event(actual_tid))
2384 pr_warn("due to cpu running other code. Event %ld->%ld\n",
2385 tid_to_event(tid), tid_to_event(actual_tid));
2386 else
2387 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2388 actual_tid, tid, next_tid(tid));
2389 #endif
2390 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
2391 }
2392
init_kmem_cache_cpus(struct kmem_cache * s)2393 static void init_kmem_cache_cpus(struct kmem_cache *s)
2394 {
2395 int cpu;
2396 struct kmem_cache_cpu *c;
2397
2398 for_each_possible_cpu(cpu) {
2399 c = per_cpu_ptr(s->cpu_slab, cpu);
2400 local_lock_init(&c->lock);
2401 c->tid = init_tid(cpu);
2402 }
2403 }
2404
2405 /*
2406 * Finishes removing the cpu slab. Merges cpu's freelist with slab's freelist,
2407 * unfreezes the slabs and puts it on the proper list.
2408 * Assumes the slab has been already safely taken away from kmem_cache_cpu
2409 * by the caller.
2410 */
deactivate_slab(struct kmem_cache * s,struct slab * slab,void * freelist)2411 static void deactivate_slab(struct kmem_cache *s, struct slab *slab,
2412 void *freelist)
2413 {
2414 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE, M_FULL_NOLIST };
2415 struct kmem_cache_node *n = get_node(s, slab_nid(slab));
2416 int free_delta = 0;
2417 enum slab_modes mode = M_NONE;
2418 void *nextfree, *freelist_iter, *freelist_tail;
2419 int tail = DEACTIVATE_TO_HEAD;
2420 unsigned long flags = 0;
2421 struct slab new;
2422 struct slab old;
2423
2424 if (slab->freelist) {
2425 stat(s, DEACTIVATE_REMOTE_FREES);
2426 tail = DEACTIVATE_TO_TAIL;
2427 }
2428
2429 /*
2430 * Stage one: Count the objects on cpu's freelist as free_delta and
2431 * remember the last object in freelist_tail for later splicing.
2432 */
2433 freelist_tail = NULL;
2434 freelist_iter = freelist;
2435 while (freelist_iter) {
2436 nextfree = get_freepointer(s, freelist_iter);
2437
2438 /*
2439 * If 'nextfree' is invalid, it is possible that the object at
2440 * 'freelist_iter' is already corrupted. So isolate all objects
2441 * starting at 'freelist_iter' by skipping them.
2442 */
2443 if (freelist_corrupted(s, slab, &freelist_iter, nextfree))
2444 break;
2445
2446 freelist_tail = freelist_iter;
2447 free_delta++;
2448
2449 freelist_iter = nextfree;
2450 }
2451
2452 /*
2453 * Stage two: Unfreeze the slab while splicing the per-cpu
2454 * freelist to the head of slab's freelist.
2455 *
2456 * Ensure that the slab is unfrozen while the list presence
2457 * reflects the actual number of objects during unfreeze.
2458 *
2459 * We first perform cmpxchg holding lock and insert to list
2460 * when it succeed. If there is mismatch then the slab is not
2461 * unfrozen and number of objects in the slab may have changed.
2462 * Then release lock and retry cmpxchg again.
2463 */
2464 redo:
2465
2466 old.freelist = READ_ONCE(slab->freelist);
2467 old.counters = READ_ONCE(slab->counters);
2468 VM_BUG_ON(!old.frozen);
2469
2470 /* Determine target state of the slab */
2471 new.counters = old.counters;
2472 if (freelist_tail) {
2473 new.inuse -= free_delta;
2474 set_freepointer(s, freelist_tail, old.freelist);
2475 new.freelist = freelist;
2476 } else
2477 new.freelist = old.freelist;
2478
2479 new.frozen = 0;
2480
2481 if (!new.inuse && n->nr_partial >= s->min_partial) {
2482 mode = M_FREE;
2483 } else if (new.freelist) {
2484 mode = M_PARTIAL;
2485 /*
2486 * Taking the spinlock removes the possibility that
2487 * acquire_slab() will see a slab that is frozen
2488 */
2489 spin_lock_irqsave(&n->list_lock, flags);
2490 } else if (kmem_cache_debug_flags(s, SLAB_STORE_USER)) {
2491 mode = M_FULL;
2492 /*
2493 * This also ensures that the scanning of full
2494 * slabs from diagnostic functions will not see
2495 * any frozen slabs.
2496 */
2497 spin_lock_irqsave(&n->list_lock, flags);
2498 } else {
2499 mode = M_FULL_NOLIST;
2500 }
2501
2502
2503 if (!cmpxchg_double_slab(s, slab,
2504 old.freelist, old.counters,
2505 new.freelist, new.counters,
2506 "unfreezing slab")) {
2507 if (mode == M_PARTIAL || mode == M_FULL)
2508 spin_unlock_irqrestore(&n->list_lock, flags);
2509 goto redo;
2510 }
2511
2512
2513 if (mode == M_PARTIAL) {
2514 add_partial(n, slab, tail);
2515 spin_unlock_irqrestore(&n->list_lock, flags);
2516 stat(s, tail);
2517 } else if (mode == M_FREE) {
2518 stat(s, DEACTIVATE_EMPTY);
2519 discard_slab(s, slab);
2520 stat(s, FREE_SLAB);
2521 } else if (mode == M_FULL) {
2522 add_full(s, n, slab);
2523 spin_unlock_irqrestore(&n->list_lock, flags);
2524 stat(s, DEACTIVATE_FULL);
2525 } else if (mode == M_FULL_NOLIST) {
2526 stat(s, DEACTIVATE_FULL);
2527 }
2528 }
2529
2530 #ifdef CONFIG_SLUB_CPU_PARTIAL
__unfreeze_partials(struct kmem_cache * s,struct slab * partial_slab)2531 static void __unfreeze_partials(struct kmem_cache *s, struct slab *partial_slab)
2532 {
2533 struct kmem_cache_node *n = NULL, *n2 = NULL;
2534 struct slab *slab, *slab_to_discard = NULL;
2535 unsigned long flags = 0;
2536
2537 while (partial_slab) {
2538 struct slab new;
2539 struct slab old;
2540
2541 slab = partial_slab;
2542 partial_slab = slab->next;
2543
2544 n2 = get_node(s, slab_nid(slab));
2545 if (n != n2) {
2546 if (n)
2547 spin_unlock_irqrestore(&n->list_lock, flags);
2548
2549 n = n2;
2550 spin_lock_irqsave(&n->list_lock, flags);
2551 }
2552
2553 do {
2554
2555 old.freelist = slab->freelist;
2556 old.counters = slab->counters;
2557 VM_BUG_ON(!old.frozen);
2558
2559 new.counters = old.counters;
2560 new.freelist = old.freelist;
2561
2562 new.frozen = 0;
2563
2564 } while (!__cmpxchg_double_slab(s, slab,
2565 old.freelist, old.counters,
2566 new.freelist, new.counters,
2567 "unfreezing slab"));
2568
2569 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2570 slab->next = slab_to_discard;
2571 slab_to_discard = slab;
2572 } else {
2573 add_partial(n, slab, DEACTIVATE_TO_TAIL);
2574 stat(s, FREE_ADD_PARTIAL);
2575 }
2576 }
2577
2578 if (n)
2579 spin_unlock_irqrestore(&n->list_lock, flags);
2580
2581 while (slab_to_discard) {
2582 slab = slab_to_discard;
2583 slab_to_discard = slab_to_discard->next;
2584
2585 stat(s, DEACTIVATE_EMPTY);
2586 discard_slab(s, slab);
2587 stat(s, FREE_SLAB);
2588 }
2589 }
2590
2591 /*
2592 * Unfreeze all the cpu partial slabs.
2593 */
unfreeze_partials(struct kmem_cache * s)2594 static void unfreeze_partials(struct kmem_cache *s)
2595 {
2596 struct slab *partial_slab;
2597 unsigned long flags;
2598
2599 local_lock_irqsave(&s->cpu_slab->lock, flags);
2600 partial_slab = this_cpu_read(s->cpu_slab->partial);
2601 this_cpu_write(s->cpu_slab->partial, NULL);
2602 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2603
2604 if (partial_slab)
2605 __unfreeze_partials(s, partial_slab);
2606 }
2607
unfreeze_partials_cpu(struct kmem_cache * s,struct kmem_cache_cpu * c)2608 static void unfreeze_partials_cpu(struct kmem_cache *s,
2609 struct kmem_cache_cpu *c)
2610 {
2611 struct slab *partial_slab;
2612
2613 partial_slab = slub_percpu_partial(c);
2614 c->partial = NULL;
2615
2616 if (partial_slab)
2617 __unfreeze_partials(s, partial_slab);
2618 }
2619
2620 /*
2621 * Put a slab that was just frozen (in __slab_free|get_partial_node) into a
2622 * partial slab slot if available.
2623 *
2624 * If we did not find a slot then simply move all the partials to the
2625 * per node partial list.
2626 */
put_cpu_partial(struct kmem_cache * s,struct slab * slab,int drain)2627 static void put_cpu_partial(struct kmem_cache *s, struct slab *slab, int drain)
2628 {
2629 struct slab *oldslab;
2630 struct slab *slab_to_unfreeze = NULL;
2631 unsigned long flags;
2632 int slabs = 0;
2633
2634 local_lock_irqsave(&s->cpu_slab->lock, flags);
2635
2636 oldslab = this_cpu_read(s->cpu_slab->partial);
2637
2638 if (oldslab) {
2639 if (drain && oldslab->slabs >= s->cpu_partial_slabs) {
2640 /*
2641 * Partial array is full. Move the existing set to the
2642 * per node partial list. Postpone the actual unfreezing
2643 * outside of the critical section.
2644 */
2645 slab_to_unfreeze = oldslab;
2646 oldslab = NULL;
2647 } else {
2648 slabs = oldslab->slabs;
2649 }
2650 }
2651
2652 slabs++;
2653
2654 slab->slabs = slabs;
2655 slab->next = oldslab;
2656
2657 this_cpu_write(s->cpu_slab->partial, slab);
2658
2659 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2660
2661 if (slab_to_unfreeze) {
2662 __unfreeze_partials(s, slab_to_unfreeze);
2663 stat(s, CPU_PARTIAL_DRAIN);
2664 }
2665 }
2666
2667 #else /* CONFIG_SLUB_CPU_PARTIAL */
2668
unfreeze_partials(struct kmem_cache * s)2669 static inline void unfreeze_partials(struct kmem_cache *s) { }
unfreeze_partials_cpu(struct kmem_cache * s,struct kmem_cache_cpu * c)2670 static inline void unfreeze_partials_cpu(struct kmem_cache *s,
2671 struct kmem_cache_cpu *c) { }
2672
2673 #endif /* CONFIG_SLUB_CPU_PARTIAL */
2674
flush_slab(struct kmem_cache * s,struct kmem_cache_cpu * c)2675 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2676 {
2677 unsigned long flags;
2678 struct slab *slab;
2679 void *freelist;
2680
2681 local_lock_irqsave(&s->cpu_slab->lock, flags);
2682
2683 slab = c->slab;
2684 freelist = c->freelist;
2685
2686 c->slab = NULL;
2687 c->freelist = NULL;
2688 c->tid = next_tid(c->tid);
2689
2690 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2691
2692 if (slab) {
2693 deactivate_slab(s, slab, freelist);
2694 stat(s, CPUSLAB_FLUSH);
2695 }
2696 }
2697
__flush_cpu_slab(struct kmem_cache * s,int cpu)2698 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2699 {
2700 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2701 void *freelist = c->freelist;
2702 struct slab *slab = c->slab;
2703
2704 c->slab = NULL;
2705 c->freelist = NULL;
2706 c->tid = next_tid(c->tid);
2707
2708 if (slab) {
2709 deactivate_slab(s, slab, freelist);
2710 stat(s, CPUSLAB_FLUSH);
2711 }
2712
2713 unfreeze_partials_cpu(s, c);
2714 }
2715
2716 struct slub_flush_work {
2717 struct work_struct work;
2718 struct kmem_cache *s;
2719 bool skip;
2720 };
2721
2722 /*
2723 * Flush cpu slab.
2724 *
2725 * Called from CPU work handler with migration disabled.
2726 */
flush_cpu_slab(struct work_struct * w)2727 static void flush_cpu_slab(struct work_struct *w)
2728 {
2729 struct kmem_cache *s;
2730 struct kmem_cache_cpu *c;
2731 struct slub_flush_work *sfw;
2732
2733 sfw = container_of(w, struct slub_flush_work, work);
2734
2735 s = sfw->s;
2736 c = this_cpu_ptr(s->cpu_slab);
2737
2738 if (c->slab)
2739 flush_slab(s, c);
2740
2741 unfreeze_partials(s);
2742 }
2743
has_cpu_slab(int cpu,struct kmem_cache * s)2744 static bool has_cpu_slab(int cpu, struct kmem_cache *s)
2745 {
2746 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2747
2748 return c->slab || slub_percpu_partial(c);
2749 }
2750
2751 static DEFINE_MUTEX(flush_lock);
2752 static DEFINE_PER_CPU(struct slub_flush_work, slub_flush);
2753
flush_all_cpus_locked(struct kmem_cache * s)2754 static void flush_all_cpus_locked(struct kmem_cache *s)
2755 {
2756 struct slub_flush_work *sfw;
2757 unsigned int cpu;
2758
2759 lockdep_assert_cpus_held();
2760 mutex_lock(&flush_lock);
2761
2762 for_each_online_cpu(cpu) {
2763 sfw = &per_cpu(slub_flush, cpu);
2764 if (!has_cpu_slab(cpu, s)) {
2765 sfw->skip = true;
2766 continue;
2767 }
2768 INIT_WORK(&sfw->work, flush_cpu_slab);
2769 sfw->skip = false;
2770 sfw->s = s;
2771 queue_work_on(cpu, flushwq, &sfw->work);
2772 }
2773
2774 for_each_online_cpu(cpu) {
2775 sfw = &per_cpu(slub_flush, cpu);
2776 if (sfw->skip)
2777 continue;
2778 flush_work(&sfw->work);
2779 }
2780
2781 mutex_unlock(&flush_lock);
2782 }
2783
flush_all(struct kmem_cache * s)2784 static void flush_all(struct kmem_cache *s)
2785 {
2786 cpus_read_lock();
2787 flush_all_cpus_locked(s);
2788 cpus_read_unlock();
2789 }
2790
2791 /*
2792 * Use the cpu notifier to insure that the cpu slabs are flushed when
2793 * necessary.
2794 */
slub_cpu_dead(unsigned int cpu)2795 static int slub_cpu_dead(unsigned int cpu)
2796 {
2797 struct kmem_cache *s;
2798
2799 mutex_lock(&slab_mutex);
2800 list_for_each_entry(s, &slab_caches, list)
2801 __flush_cpu_slab(s, cpu);
2802 mutex_unlock(&slab_mutex);
2803 return 0;
2804 }
2805
2806 /*
2807 * Check if the objects in a per cpu structure fit numa
2808 * locality expectations.
2809 */
node_match(struct slab * slab,int node)2810 static inline int node_match(struct slab *slab, int node)
2811 {
2812 #ifdef CONFIG_NUMA
2813 if (node != NUMA_NO_NODE && slab_nid(slab) != node)
2814 return 0;
2815 #endif
2816 return 1;
2817 }
2818
2819 #ifdef CONFIG_SLUB_DEBUG
count_free(struct slab * slab)2820 static int count_free(struct slab *slab)
2821 {
2822 return slab->objects - slab->inuse;
2823 }
2824
node_nr_objs(struct kmem_cache_node * n)2825 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2826 {
2827 return atomic_long_read(&n->total_objects);
2828 }
2829
2830 /* Supports checking bulk free of a constructed freelist */
free_debug_processing(struct kmem_cache * s,struct slab * slab,void * head,void * tail,int bulk_cnt,unsigned long addr)2831 static noinline void free_debug_processing(
2832 struct kmem_cache *s, struct slab *slab,
2833 void *head, void *tail, int bulk_cnt,
2834 unsigned long addr)
2835 {
2836 struct kmem_cache_node *n = get_node(s, slab_nid(slab));
2837 struct slab *slab_free = NULL;
2838 void *object = head;
2839 int cnt = 0;
2840 unsigned long flags;
2841 bool checks_ok = false;
2842 depot_stack_handle_t handle = 0;
2843
2844 if (s->flags & SLAB_STORE_USER)
2845 handle = set_track_prepare();
2846
2847 spin_lock_irqsave(&n->list_lock, flags);
2848
2849 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
2850 if (!check_slab(s, slab))
2851 goto out;
2852 }
2853
2854 if (slab->inuse < bulk_cnt) {
2855 slab_err(s, slab, "Slab has %d allocated objects but %d are to be freed\n",
2856 slab->inuse, bulk_cnt);
2857 goto out;
2858 }
2859
2860 next_object:
2861
2862 if (++cnt > bulk_cnt)
2863 goto out_cnt;
2864
2865 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
2866 if (!free_consistency_checks(s, slab, object, addr))
2867 goto out;
2868 }
2869
2870 if (s->flags & SLAB_STORE_USER)
2871 set_track_update(s, object, TRACK_FREE, addr, handle);
2872 trace(s, slab, object, 0);
2873 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
2874 init_object(s, object, SLUB_RED_INACTIVE);
2875
2876 /* Reached end of constructed freelist yet? */
2877 if (object != tail) {
2878 object = get_freepointer(s, object);
2879 goto next_object;
2880 }
2881 checks_ok = true;
2882
2883 out_cnt:
2884 if (cnt != bulk_cnt)
2885 slab_err(s, slab, "Bulk free expected %d objects but found %d\n",
2886 bulk_cnt, cnt);
2887
2888 out:
2889 if (checks_ok) {
2890 void *prior = slab->freelist;
2891
2892 /* Perform the actual freeing while we still hold the locks */
2893 slab->inuse -= cnt;
2894 set_freepointer(s, tail, prior);
2895 slab->freelist = head;
2896
2897 /*
2898 * If the slab is empty, and node's partial list is full,
2899 * it should be discarded anyway no matter it's on full or
2900 * partial list.
2901 */
2902 if (slab->inuse == 0 && n->nr_partial >= s->min_partial)
2903 slab_free = slab;
2904
2905 if (!prior) {
2906 /* was on full list */
2907 remove_full(s, n, slab);
2908 if (!slab_free) {
2909 add_partial(n, slab, DEACTIVATE_TO_TAIL);
2910 stat(s, FREE_ADD_PARTIAL);
2911 }
2912 } else if (slab_free) {
2913 remove_partial(n, slab);
2914 stat(s, FREE_REMOVE_PARTIAL);
2915 }
2916 }
2917
2918 if (slab_free) {
2919 /*
2920 * Update the counters while still holding n->list_lock to
2921 * prevent spurious validation warnings
2922 */
2923 dec_slabs_node(s, slab_nid(slab_free), slab_free->objects);
2924 }
2925
2926 spin_unlock_irqrestore(&n->list_lock, flags);
2927
2928 if (!checks_ok)
2929 slab_fix(s, "Object at 0x%p not freed", object);
2930
2931 if (slab_free) {
2932 stat(s, FREE_SLAB);
2933 free_slab(s, slab_free);
2934 }
2935 }
2936 #endif /* CONFIG_SLUB_DEBUG */
2937
2938 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
count_partial(struct kmem_cache_node * n,int (* get_count)(struct slab *))2939 static unsigned long count_partial(struct kmem_cache_node *n,
2940 int (*get_count)(struct slab *))
2941 {
2942 unsigned long flags;
2943 unsigned long x = 0;
2944 struct slab *slab;
2945
2946 spin_lock_irqsave(&n->list_lock, flags);
2947 list_for_each_entry(slab, &n->partial, slab_list)
2948 x += get_count(slab);
2949 spin_unlock_irqrestore(&n->list_lock, flags);
2950 return x;
2951 }
2952 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2953
2954 static noinline void
slab_out_of_memory(struct kmem_cache * s,gfp_t gfpflags,int nid)2955 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2956 {
2957 #ifdef CONFIG_SLUB_DEBUG
2958 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2959 DEFAULT_RATELIMIT_BURST);
2960 int node;
2961 struct kmem_cache_node *n;
2962
2963 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2964 return;
2965
2966 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2967 nid, gfpflags, &gfpflags);
2968 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
2969 s->name, s->object_size, s->size, oo_order(s->oo),
2970 oo_order(s->min));
2971
2972 if (oo_order(s->min) > get_order(s->object_size))
2973 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2974 s->name);
2975
2976 for_each_kmem_cache_node(s, node, n) {
2977 unsigned long nr_slabs;
2978 unsigned long nr_objs;
2979 unsigned long nr_free;
2980
2981 nr_free = count_partial(n, count_free);
2982 nr_slabs = node_nr_slabs(n);
2983 nr_objs = node_nr_objs(n);
2984
2985 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2986 node, nr_slabs, nr_objs, nr_free);
2987 }
2988 #endif
2989 }
2990
pfmemalloc_match(struct slab * slab,gfp_t gfpflags)2991 static inline bool pfmemalloc_match(struct slab *slab, gfp_t gfpflags)
2992 {
2993 if (unlikely(slab_test_pfmemalloc(slab)))
2994 return gfp_pfmemalloc_allowed(gfpflags);
2995
2996 return true;
2997 }
2998
2999 /*
3000 * Check the slab->freelist and either transfer the freelist to the
3001 * per cpu freelist or deactivate the slab.
3002 *
3003 * The slab is still frozen if the return value is not NULL.
3004 *
3005 * If this function returns NULL then the slab has been unfrozen.
3006 */
get_freelist(struct kmem_cache * s,struct slab * slab)3007 static inline void *get_freelist(struct kmem_cache *s, struct slab *slab)
3008 {
3009 struct slab new;
3010 unsigned long counters;
3011 void *freelist;
3012
3013 lockdep_assert_held(this_cpu_ptr(&s->cpu_slab->lock));
3014
3015 do {
3016 freelist = slab->freelist;
3017 counters = slab->counters;
3018
3019 new.counters = counters;
3020 VM_BUG_ON(!new.frozen);
3021
3022 new.inuse = slab->objects;
3023 new.frozen = freelist != NULL;
3024
3025 } while (!__cmpxchg_double_slab(s, slab,
3026 freelist, counters,
3027 NULL, new.counters,
3028 "get_freelist"));
3029
3030 return freelist;
3031 }
3032
3033 /*
3034 * Slow path. The lockless freelist is empty or we need to perform
3035 * debugging duties.
3036 *
3037 * Processing is still very fast if new objects have been freed to the
3038 * regular freelist. In that case we simply take over the regular freelist
3039 * as the lockless freelist and zap the regular freelist.
3040 *
3041 * If that is not working then we fall back to the partial lists. We take the
3042 * first element of the freelist as the object to allocate now and move the
3043 * rest of the freelist to the lockless freelist.
3044 *
3045 * And if we were unable to get a new slab from the partial slab lists then
3046 * we need to allocate a new slab. This is the slowest path since it involves
3047 * a call to the page allocator and the setup of a new slab.
3048 *
3049 * Version of __slab_alloc to use when we know that preemption is
3050 * already disabled (which is the case for bulk allocation).
3051 */
___slab_alloc(struct kmem_cache * s,gfp_t gfpflags,int node,unsigned long addr,struct kmem_cache_cpu * c,unsigned int orig_size)3052 static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
3053 unsigned long addr, struct kmem_cache_cpu *c, unsigned int orig_size)
3054 {
3055 void *freelist;
3056 struct slab *slab;
3057 unsigned long flags;
3058 struct partial_context pc;
3059
3060 stat(s, ALLOC_SLOWPATH);
3061
3062 reread_slab:
3063
3064 slab = READ_ONCE(c->slab);
3065 if (!slab) {
3066 /*
3067 * if the node is not online or has no normal memory, just
3068 * ignore the node constraint
3069 */
3070 if (unlikely(node != NUMA_NO_NODE &&
3071 !node_isset(node, slab_nodes)))
3072 node = NUMA_NO_NODE;
3073 goto new_slab;
3074 }
3075 redo:
3076
3077 if (unlikely(!node_match(slab, node))) {
3078 /*
3079 * same as above but node_match() being false already
3080 * implies node != NUMA_NO_NODE
3081 */
3082 if (!node_isset(node, slab_nodes)) {
3083 node = NUMA_NO_NODE;
3084 } else {
3085 stat(s, ALLOC_NODE_MISMATCH);
3086 goto deactivate_slab;
3087 }
3088 }
3089
3090 /*
3091 * By rights, we should be searching for a slab page that was
3092 * PFMEMALLOC but right now, we are losing the pfmemalloc
3093 * information when the page leaves the per-cpu allocator
3094 */
3095 if (unlikely(!pfmemalloc_match(slab, gfpflags)))
3096 goto deactivate_slab;
3097
3098 /* must check again c->slab in case we got preempted and it changed */
3099 local_lock_irqsave(&s->cpu_slab->lock, flags);
3100 if (unlikely(slab != c->slab)) {
3101 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3102 goto reread_slab;
3103 }
3104 freelist = c->freelist;
3105 if (freelist)
3106 goto load_freelist;
3107
3108 freelist = get_freelist(s, slab);
3109
3110 if (!freelist) {
3111 c->slab = NULL;
3112 c->tid = next_tid(c->tid);
3113 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3114 stat(s, DEACTIVATE_BYPASS);
3115 goto new_slab;
3116 }
3117
3118 stat(s, ALLOC_REFILL);
3119
3120 load_freelist:
3121
3122 lockdep_assert_held(this_cpu_ptr(&s->cpu_slab->lock));
3123
3124 /*
3125 * freelist is pointing to the list of objects to be used.
3126 * slab is pointing to the slab from which the objects are obtained.
3127 * That slab must be frozen for per cpu allocations to work.
3128 */
3129 VM_BUG_ON(!c->slab->frozen);
3130 c->freelist = get_freepointer(s, freelist);
3131 c->tid = next_tid(c->tid);
3132 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3133 return freelist;
3134
3135 deactivate_slab:
3136
3137 local_lock_irqsave(&s->cpu_slab->lock, flags);
3138 if (slab != c->slab) {
3139 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3140 goto reread_slab;
3141 }
3142 freelist = c->freelist;
3143 c->slab = NULL;
3144 c->freelist = NULL;
3145 c->tid = next_tid(c->tid);
3146 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3147 deactivate_slab(s, slab, freelist);
3148
3149 new_slab:
3150
3151 if (slub_percpu_partial(c)) {
3152 local_lock_irqsave(&s->cpu_slab->lock, flags);
3153 if (unlikely(c->slab)) {
3154 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3155 goto reread_slab;
3156 }
3157 if (unlikely(!slub_percpu_partial(c))) {
3158 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3159 /* we were preempted and partial list got empty */
3160 goto new_objects;
3161 }
3162
3163 slab = c->slab = slub_percpu_partial(c);
3164 slub_set_percpu_partial(c, slab);
3165 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3166 stat(s, CPU_PARTIAL_ALLOC);
3167 goto redo;
3168 }
3169
3170 new_objects:
3171
3172 pc.flags = gfpflags;
3173 pc.slab = &slab;
3174 pc.orig_size = orig_size;
3175 freelist = get_partial(s, node, &pc);
3176 if (freelist)
3177 goto check_new_slab;
3178
3179 slub_put_cpu_ptr(s->cpu_slab);
3180 slab = new_slab(s, gfpflags, node);
3181 c = slub_get_cpu_ptr(s->cpu_slab);
3182
3183 if (unlikely(!slab)) {
3184 slab_out_of_memory(s, gfpflags, node);
3185 return NULL;
3186 }
3187
3188 stat(s, ALLOC_SLAB);
3189
3190 if (kmem_cache_debug(s)) {
3191 freelist = alloc_single_from_new_slab(s, slab, orig_size);
3192
3193 if (unlikely(!freelist))
3194 goto new_objects;
3195
3196 if (s->flags & SLAB_STORE_USER)
3197 set_track(s, freelist, TRACK_ALLOC, addr);
3198
3199 return freelist;
3200 }
3201
3202 /*
3203 * No other reference to the slab yet so we can
3204 * muck around with it freely without cmpxchg
3205 */
3206 freelist = slab->freelist;
3207 slab->freelist = NULL;
3208 slab->inuse = slab->objects;
3209 slab->frozen = 1;
3210
3211 inc_slabs_node(s, slab_nid(slab), slab->objects);
3212
3213 check_new_slab:
3214
3215 if (kmem_cache_debug(s)) {
3216 /*
3217 * For debug caches here we had to go through
3218 * alloc_single_from_partial() so just store the tracking info
3219 * and return the object
3220 */
3221 if (s->flags & SLAB_STORE_USER)
3222 set_track(s, freelist, TRACK_ALLOC, addr);
3223
3224 return freelist;
3225 }
3226
3227 if (unlikely(!pfmemalloc_match(slab, gfpflags))) {
3228 /*
3229 * For !pfmemalloc_match() case we don't load freelist so that
3230 * we don't make further mismatched allocations easier.
3231 */
3232 deactivate_slab(s, slab, get_freepointer(s, freelist));
3233 return freelist;
3234 }
3235
3236 retry_load_slab:
3237
3238 local_lock_irqsave(&s->cpu_slab->lock, flags);
3239 if (unlikely(c->slab)) {
3240 void *flush_freelist = c->freelist;
3241 struct slab *flush_slab = c->slab;
3242
3243 c->slab = NULL;
3244 c->freelist = NULL;
3245 c->tid = next_tid(c->tid);
3246
3247 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3248
3249 deactivate_slab(s, flush_slab, flush_freelist);
3250
3251 stat(s, CPUSLAB_FLUSH);
3252
3253 goto retry_load_slab;
3254 }
3255 c->slab = slab;
3256
3257 goto load_freelist;
3258 }
3259
3260 /*
3261 * A wrapper for ___slab_alloc() for contexts where preemption is not yet
3262 * disabled. Compensates for possible cpu changes by refetching the per cpu area
3263 * pointer.
3264 */
__slab_alloc(struct kmem_cache * s,gfp_t gfpflags,int node,unsigned long addr,struct kmem_cache_cpu * c,unsigned int orig_size)3265 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
3266 unsigned long addr, struct kmem_cache_cpu *c, unsigned int orig_size)
3267 {
3268 void *p;
3269
3270 #ifdef CONFIG_PREEMPT_COUNT
3271 /*
3272 * We may have been preempted and rescheduled on a different
3273 * cpu before disabling preemption. Need to reload cpu area
3274 * pointer.
3275 */
3276 c = slub_get_cpu_ptr(s->cpu_slab);
3277 #endif
3278
3279 p = ___slab_alloc(s, gfpflags, node, addr, c, orig_size);
3280 #ifdef CONFIG_PREEMPT_COUNT
3281 slub_put_cpu_ptr(s->cpu_slab);
3282 #endif
3283 return p;
3284 }
3285
3286 /*
3287 * If the object has been wiped upon free, make sure it's fully initialized by
3288 * zeroing out freelist pointer.
3289 */
maybe_wipe_obj_freeptr(struct kmem_cache * s,void * obj)3290 static __always_inline void maybe_wipe_obj_freeptr(struct kmem_cache *s,
3291 void *obj)
3292 {
3293 if (unlikely(slab_want_init_on_free(s)) && obj)
3294 memset((void *)((char *)kasan_reset_tag(obj) + s->offset),
3295 0, sizeof(void *));
3296 }
3297
3298 /*
3299 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
3300 * have the fastpath folded into their functions. So no function call
3301 * overhead for requests that can be satisfied on the fastpath.
3302 *
3303 * The fastpath works by first checking if the lockless freelist can be used.
3304 * If not then __slab_alloc is called for slow processing.
3305 *
3306 * Otherwise we can simply pick the next object from the lockless free list.
3307 */
slab_alloc_node(struct kmem_cache * s,struct list_lru * lru,gfp_t gfpflags,int node,unsigned long addr,size_t orig_size)3308 static __always_inline void *slab_alloc_node(struct kmem_cache *s, struct list_lru *lru,
3309 gfp_t gfpflags, int node, unsigned long addr, size_t orig_size)
3310 {
3311 void *object;
3312 struct kmem_cache_cpu *c;
3313 struct slab *slab;
3314 unsigned long tid;
3315 struct obj_cgroup *objcg = NULL;
3316 bool init = false;
3317
3318 s = slab_pre_alloc_hook(s, lru, &objcg, 1, gfpflags);
3319 if (!s)
3320 return NULL;
3321
3322 object = kfence_alloc(s, orig_size, gfpflags);
3323 if (unlikely(object))
3324 goto out;
3325
3326 redo:
3327 /*
3328 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
3329 * enabled. We may switch back and forth between cpus while
3330 * reading from one cpu area. That does not matter as long
3331 * as we end up on the original cpu again when doing the cmpxchg.
3332 *
3333 * We must guarantee that tid and kmem_cache_cpu are retrieved on the
3334 * same cpu. We read first the kmem_cache_cpu pointer and use it to read
3335 * the tid. If we are preempted and switched to another cpu between the
3336 * two reads, it's OK as the two are still associated with the same cpu
3337 * and cmpxchg later will validate the cpu.
3338 */
3339 c = raw_cpu_ptr(s->cpu_slab);
3340 tid = READ_ONCE(c->tid);
3341
3342 /*
3343 * Irqless object alloc/free algorithm used here depends on sequence
3344 * of fetching cpu_slab's data. tid should be fetched before anything
3345 * on c to guarantee that object and slab associated with previous tid
3346 * won't be used with current tid. If we fetch tid first, object and
3347 * slab could be one associated with next tid and our alloc/free
3348 * request will be failed. In this case, we will retry. So, no problem.
3349 */
3350 barrier();
3351
3352 /*
3353 * The transaction ids are globally unique per cpu and per operation on
3354 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
3355 * occurs on the right processor and that there was no operation on the
3356 * linked list in between.
3357 */
3358
3359 object = c->freelist;
3360 slab = c->slab;
3361
3362 if (!USE_LOCKLESS_FAST_PATH() ||
3363 unlikely(!object || !slab || !node_match(slab, node))) {
3364 object = __slab_alloc(s, gfpflags, node, addr, c, orig_size);
3365 } else {
3366 void *next_object = get_freepointer_safe(s, object);
3367
3368 /*
3369 * The cmpxchg will only match if there was no additional
3370 * operation and if we are on the right processor.
3371 *
3372 * The cmpxchg does the following atomically (without lock
3373 * semantics!)
3374 * 1. Relocate first pointer to the current per cpu area.
3375 * 2. Verify that tid and freelist have not been changed
3376 * 3. If they were not changed replace tid and freelist
3377 *
3378 * Since this is without lock semantics the protection is only
3379 * against code executing on this cpu *not* from access by
3380 * other cpus.
3381 */
3382 if (unlikely(!this_cpu_cmpxchg_double(
3383 s->cpu_slab->freelist, s->cpu_slab->tid,
3384 object, tid,
3385 next_object, next_tid(tid)))) {
3386
3387 note_cmpxchg_failure("slab_alloc", s, tid);
3388 goto redo;
3389 }
3390 prefetch_freepointer(s, next_object);
3391 stat(s, ALLOC_FASTPATH);
3392 }
3393
3394 maybe_wipe_obj_freeptr(s, object);
3395 init = slab_want_init_on_alloc(gfpflags, s);
3396
3397 out:
3398 slab_post_alloc_hook(s, objcg, gfpflags, 1, &object, init);
3399
3400 return object;
3401 }
3402
slab_alloc(struct kmem_cache * s,struct list_lru * lru,gfp_t gfpflags,unsigned long addr,size_t orig_size)3403 static __always_inline void *slab_alloc(struct kmem_cache *s, struct list_lru *lru,
3404 gfp_t gfpflags, unsigned long addr, size_t orig_size)
3405 {
3406 return slab_alloc_node(s, lru, gfpflags, NUMA_NO_NODE, addr, orig_size);
3407 }
3408
3409 static __always_inline
__kmem_cache_alloc_lru(struct kmem_cache * s,struct list_lru * lru,gfp_t gfpflags)3410 void *__kmem_cache_alloc_lru(struct kmem_cache *s, struct list_lru *lru,
3411 gfp_t gfpflags)
3412 {
3413 void *ret = slab_alloc(s, lru, gfpflags, _RET_IP_, s->object_size);
3414
3415 trace_kmem_cache_alloc(_RET_IP_, ret, s, gfpflags, NUMA_NO_NODE);
3416
3417 return ret;
3418 }
3419
kmem_cache_alloc(struct kmem_cache * s,gfp_t gfpflags)3420 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
3421 {
3422 return __kmem_cache_alloc_lru(s, NULL, gfpflags);
3423 }
3424 EXPORT_SYMBOL(kmem_cache_alloc);
3425
kmem_cache_alloc_lru(struct kmem_cache * s,struct list_lru * lru,gfp_t gfpflags)3426 void *kmem_cache_alloc_lru(struct kmem_cache *s, struct list_lru *lru,
3427 gfp_t gfpflags)
3428 {
3429 return __kmem_cache_alloc_lru(s, lru, gfpflags);
3430 }
3431 EXPORT_SYMBOL(kmem_cache_alloc_lru);
3432
__kmem_cache_alloc_node(struct kmem_cache * s,gfp_t gfpflags,int node,size_t orig_size,unsigned long caller)3433 void *__kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags,
3434 int node, size_t orig_size,
3435 unsigned long caller)
3436 {
3437 return slab_alloc_node(s, NULL, gfpflags, node,
3438 caller, orig_size);
3439 }
3440
kmem_cache_alloc_node(struct kmem_cache * s,gfp_t gfpflags,int node)3441 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
3442 {
3443 void *ret = slab_alloc_node(s, NULL, gfpflags, node, _RET_IP_, s->object_size);
3444
3445 trace_kmem_cache_alloc(_RET_IP_, ret, s, gfpflags, node);
3446
3447 return ret;
3448 }
3449 EXPORT_SYMBOL(kmem_cache_alloc_node);
3450
3451 /*
3452 * Slow path handling. This may still be called frequently since objects
3453 * have a longer lifetime than the cpu slabs in most processing loads.
3454 *
3455 * So we still attempt to reduce cache line usage. Just take the slab
3456 * lock and free the item. If there is no additional partial slab
3457 * handling required then we can return immediately.
3458 */
__slab_free(struct kmem_cache * s,struct slab * slab,void * head,void * tail,int cnt,unsigned long addr)3459 static void __slab_free(struct kmem_cache *s, struct slab *slab,
3460 void *head, void *tail, int cnt,
3461 unsigned long addr)
3462
3463 {
3464 void *prior;
3465 int was_frozen;
3466 struct slab new;
3467 unsigned long counters;
3468 struct kmem_cache_node *n = NULL;
3469 unsigned long flags;
3470
3471 stat(s, FREE_SLOWPATH);
3472
3473 if (kfence_free(head))
3474 return;
3475
3476 if (kmem_cache_debug(s)) {
3477 free_debug_processing(s, slab, head, tail, cnt, addr);
3478 return;
3479 }
3480
3481 do {
3482 if (unlikely(n)) {
3483 spin_unlock_irqrestore(&n->list_lock, flags);
3484 n = NULL;
3485 }
3486 prior = slab->freelist;
3487 counters = slab->counters;
3488 set_freepointer(s, tail, prior);
3489 new.counters = counters;
3490 was_frozen = new.frozen;
3491 new.inuse -= cnt;
3492 if ((!new.inuse || !prior) && !was_frozen) {
3493
3494 if (kmem_cache_has_cpu_partial(s) && !prior) {
3495
3496 /*
3497 * Slab was on no list before and will be
3498 * partially empty
3499 * We can defer the list move and instead
3500 * freeze it.
3501 */
3502 new.frozen = 1;
3503
3504 } else { /* Needs to be taken off a list */
3505
3506 n = get_node(s, slab_nid(slab));
3507 /*
3508 * Speculatively acquire the list_lock.
3509 * If the cmpxchg does not succeed then we may
3510 * drop the list_lock without any processing.
3511 *
3512 * Otherwise the list_lock will synchronize with
3513 * other processors updating the list of slabs.
3514 */
3515 spin_lock_irqsave(&n->list_lock, flags);
3516
3517 }
3518 }
3519
3520 } while (!cmpxchg_double_slab(s, slab,
3521 prior, counters,
3522 head, new.counters,
3523 "__slab_free"));
3524
3525 if (likely(!n)) {
3526
3527 if (likely(was_frozen)) {
3528 /*
3529 * The list lock was not taken therefore no list
3530 * activity can be necessary.
3531 */
3532 stat(s, FREE_FROZEN);
3533 } else if (new.frozen) {
3534 /*
3535 * If we just froze the slab then put it onto the
3536 * per cpu partial list.
3537 */
3538 put_cpu_partial(s, slab, 1);
3539 stat(s, CPU_PARTIAL_FREE);
3540 }
3541
3542 return;
3543 }
3544
3545 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
3546 goto slab_empty;
3547
3548 /*
3549 * Objects left in the slab. If it was not on the partial list before
3550 * then add it.
3551 */
3552 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
3553 remove_full(s, n, slab);
3554 add_partial(n, slab, DEACTIVATE_TO_TAIL);
3555 stat(s, FREE_ADD_PARTIAL);
3556 }
3557 spin_unlock_irqrestore(&n->list_lock, flags);
3558 return;
3559
3560 slab_empty:
3561 if (prior) {
3562 /*
3563 * Slab on the partial list.
3564 */
3565 remove_partial(n, slab);
3566 stat(s, FREE_REMOVE_PARTIAL);
3567 } else {
3568 /* Slab must be on the full list */
3569 remove_full(s, n, slab);
3570 }
3571
3572 spin_unlock_irqrestore(&n->list_lock, flags);
3573 stat(s, FREE_SLAB);
3574 discard_slab(s, slab);
3575 }
3576
3577 /*
3578 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
3579 * can perform fastpath freeing without additional function calls.
3580 *
3581 * The fastpath is only possible if we are freeing to the current cpu slab
3582 * of this processor. This typically the case if we have just allocated
3583 * the item before.
3584 *
3585 * If fastpath is not possible then fall back to __slab_free where we deal
3586 * with all sorts of special processing.
3587 *
3588 * Bulk free of a freelist with several objects (all pointing to the
3589 * same slab) possible by specifying head and tail ptr, plus objects
3590 * count (cnt). Bulk free indicated by tail pointer being set.
3591 */
do_slab_free(struct kmem_cache * s,struct slab * slab,void * head,void * tail,int cnt,unsigned long addr)3592 static __always_inline void do_slab_free(struct kmem_cache *s,
3593 struct slab *slab, void *head, void *tail,
3594 int cnt, unsigned long addr)
3595 {
3596 void *tail_obj = tail ? : head;
3597 struct kmem_cache_cpu *c;
3598 unsigned long tid;
3599 void **freelist;
3600
3601 redo:
3602 /*
3603 * Determine the currently cpus per cpu slab.
3604 * The cpu may change afterward. However that does not matter since
3605 * data is retrieved via this pointer. If we are on the same cpu
3606 * during the cmpxchg then the free will succeed.
3607 */
3608 c = raw_cpu_ptr(s->cpu_slab);
3609 tid = READ_ONCE(c->tid);
3610
3611 /* Same with comment on barrier() in slab_alloc_node() */
3612 barrier();
3613
3614 if (unlikely(slab != c->slab)) {
3615 __slab_free(s, slab, head, tail_obj, cnt, addr);
3616 return;
3617 }
3618
3619 if (USE_LOCKLESS_FAST_PATH()) {
3620 freelist = READ_ONCE(c->freelist);
3621
3622 set_freepointer(s, tail_obj, freelist);
3623
3624 if (unlikely(!this_cpu_cmpxchg_double(
3625 s->cpu_slab->freelist, s->cpu_slab->tid,
3626 freelist, tid,
3627 head, next_tid(tid)))) {
3628
3629 note_cmpxchg_failure("slab_free", s, tid);
3630 goto redo;
3631 }
3632 } else {
3633 /* Update the free list under the local lock */
3634 local_lock(&s->cpu_slab->lock);
3635 c = this_cpu_ptr(s->cpu_slab);
3636 if (unlikely(slab != c->slab)) {
3637 local_unlock(&s->cpu_slab->lock);
3638 goto redo;
3639 }
3640 tid = c->tid;
3641 freelist = c->freelist;
3642
3643 set_freepointer(s, tail_obj, freelist);
3644 c->freelist = head;
3645 c->tid = next_tid(tid);
3646
3647 local_unlock(&s->cpu_slab->lock);
3648 }
3649 stat(s, FREE_FASTPATH);
3650 }
3651
slab_free(struct kmem_cache * s,struct slab * slab,void * head,void * tail,void ** p,int cnt,unsigned long addr)3652 static __always_inline void slab_free(struct kmem_cache *s, struct slab *slab,
3653 void *head, void *tail, void **p, int cnt,
3654 unsigned long addr)
3655 {
3656 memcg_slab_free_hook(s, slab, p, cnt);
3657 /*
3658 * With KASAN enabled slab_free_freelist_hook modifies the freelist
3659 * to remove objects, whose reuse must be delayed.
3660 */
3661 if (slab_free_freelist_hook(s, &head, &tail, &cnt))
3662 do_slab_free(s, slab, head, tail, cnt, addr);
3663 }
3664
3665 #ifdef CONFIG_KASAN_GENERIC
___cache_free(struct kmem_cache * cache,void * x,unsigned long addr)3666 void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
3667 {
3668 do_slab_free(cache, virt_to_slab(x), x, NULL, 1, addr);
3669 }
3670 #endif
3671
__kmem_cache_free(struct kmem_cache * s,void * x,unsigned long caller)3672 void __kmem_cache_free(struct kmem_cache *s, void *x, unsigned long caller)
3673 {
3674 slab_free(s, virt_to_slab(x), x, NULL, &x, 1, caller);
3675 }
3676
kmem_cache_free(struct kmem_cache * s,void * x)3677 void kmem_cache_free(struct kmem_cache *s, void *x)
3678 {
3679 s = cache_from_obj(s, x);
3680 if (!s)
3681 return;
3682 trace_kmem_cache_free(_RET_IP_, x, s);
3683 slab_free(s, virt_to_slab(x), x, NULL, &x, 1, _RET_IP_);
3684 }
3685 EXPORT_SYMBOL(kmem_cache_free);
3686
3687 struct detached_freelist {
3688 struct slab *slab;
3689 void *tail;
3690 void *freelist;
3691 int cnt;
3692 struct kmem_cache *s;
3693 };
3694
3695 /*
3696 * This function progressively scans the array with free objects (with
3697 * a limited look ahead) and extract objects belonging to the same
3698 * slab. It builds a detached freelist directly within the given
3699 * slab/objects. This can happen without any need for
3700 * synchronization, because the objects are owned by running process.
3701 * The freelist is build up as a single linked list in the objects.
3702 * The idea is, that this detached freelist can then be bulk
3703 * transferred to the real freelist(s), but only requiring a single
3704 * synchronization primitive. Look ahead in the array is limited due
3705 * to performance reasons.
3706 */
3707 static inline
build_detached_freelist(struct kmem_cache * s,size_t size,void ** p,struct detached_freelist * df)3708 int build_detached_freelist(struct kmem_cache *s, size_t size,
3709 void **p, struct detached_freelist *df)
3710 {
3711 int lookahead = 3;
3712 void *object;
3713 struct folio *folio;
3714 size_t same;
3715
3716 object = p[--size];
3717 folio = virt_to_folio(object);
3718 if (!s) {
3719 /* Handle kalloc'ed objects */
3720 if (unlikely(!folio_test_slab(folio))) {
3721 free_large_kmalloc(folio, object);
3722 df->slab = NULL;
3723 return size;
3724 }
3725 /* Derive kmem_cache from object */
3726 df->slab = folio_slab(folio);
3727 df->s = df->slab->slab_cache;
3728 } else {
3729 df->slab = folio_slab(folio);
3730 df->s = cache_from_obj(s, object); /* Support for memcg */
3731 }
3732
3733 /* Start new detached freelist */
3734 df->tail = object;
3735 df->freelist = object;
3736 df->cnt = 1;
3737
3738 if (is_kfence_address(object))
3739 return size;
3740
3741 set_freepointer(df->s, object, NULL);
3742
3743 same = size;
3744 while (size) {
3745 object = p[--size];
3746 /* df->slab is always set at this point */
3747 if (df->slab == virt_to_slab(object)) {
3748 /* Opportunity build freelist */
3749 set_freepointer(df->s, object, df->freelist);
3750 df->freelist = object;
3751 df->cnt++;
3752 same--;
3753 if (size != same)
3754 swap(p[size], p[same]);
3755 continue;
3756 }
3757
3758 /* Limit look ahead search */
3759 if (!--lookahead)
3760 break;
3761 }
3762
3763 return same;
3764 }
3765
3766 /* Note that interrupts must be enabled when calling this function. */
kmem_cache_free_bulk(struct kmem_cache * s,size_t size,void ** p)3767 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
3768 {
3769 if (!size)
3770 return;
3771
3772 do {
3773 struct detached_freelist df;
3774
3775 size = build_detached_freelist(s, size, p, &df);
3776 if (!df.slab)
3777 continue;
3778
3779 slab_free(df.s, df.slab, df.freelist, df.tail, &p[size], df.cnt,
3780 _RET_IP_);
3781 } while (likely(size));
3782 }
3783 EXPORT_SYMBOL(kmem_cache_free_bulk);
3784
3785 /* Note that interrupts must be enabled when calling this function. */
kmem_cache_alloc_bulk(struct kmem_cache * s,gfp_t flags,size_t size,void ** p)3786 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3787 void **p)
3788 {
3789 struct kmem_cache_cpu *c;
3790 int i;
3791 struct obj_cgroup *objcg = NULL;
3792
3793 /* memcg and kmem_cache debug support */
3794 s = slab_pre_alloc_hook(s, NULL, &objcg, size, flags);
3795 if (unlikely(!s))
3796 return false;
3797 /*
3798 * Drain objects in the per cpu slab, while disabling local
3799 * IRQs, which protects against PREEMPT and interrupts
3800 * handlers invoking normal fastpath.
3801 */
3802 c = slub_get_cpu_ptr(s->cpu_slab);
3803 local_lock_irq(&s->cpu_slab->lock);
3804
3805 for (i = 0; i < size; i++) {
3806 void *object = kfence_alloc(s, s->object_size, flags);
3807
3808 if (unlikely(object)) {
3809 p[i] = object;
3810 continue;
3811 }
3812
3813 object = c->freelist;
3814 if (unlikely(!object)) {
3815 /*
3816 * We may have removed an object from c->freelist using
3817 * the fastpath in the previous iteration; in that case,
3818 * c->tid has not been bumped yet.
3819 * Since ___slab_alloc() may reenable interrupts while
3820 * allocating memory, we should bump c->tid now.
3821 */
3822 c->tid = next_tid(c->tid);
3823
3824 local_unlock_irq(&s->cpu_slab->lock);
3825
3826 /*
3827 * Invoking slow path likely have side-effect
3828 * of re-populating per CPU c->freelist
3829 */
3830 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
3831 _RET_IP_, c, s->object_size);
3832 if (unlikely(!p[i]))
3833 goto error;
3834
3835 c = this_cpu_ptr(s->cpu_slab);
3836 maybe_wipe_obj_freeptr(s, p[i]);
3837
3838 local_lock_irq(&s->cpu_slab->lock);
3839
3840 continue; /* goto for-loop */
3841 }
3842 c->freelist = get_freepointer(s, object);
3843 p[i] = object;
3844 maybe_wipe_obj_freeptr(s, p[i]);
3845 }
3846 c->tid = next_tid(c->tid);
3847 local_unlock_irq(&s->cpu_slab->lock);
3848 slub_put_cpu_ptr(s->cpu_slab);
3849
3850 /*
3851 * memcg and kmem_cache debug support and memory initialization.
3852 * Done outside of the IRQ disabled fastpath loop.
3853 */
3854 slab_post_alloc_hook(s, objcg, flags, size, p,
3855 slab_want_init_on_alloc(flags, s));
3856 return i;
3857 error:
3858 slub_put_cpu_ptr(s->cpu_slab);
3859 slab_post_alloc_hook(s, objcg, flags, i, p, false);
3860 kmem_cache_free_bulk(s, i, p);
3861 return 0;
3862 }
3863 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3864
3865
3866 /*
3867 * Object placement in a slab is made very easy because we always start at
3868 * offset 0. If we tune the size of the object to the alignment then we can
3869 * get the required alignment by putting one properly sized object after
3870 * another.
3871 *
3872 * Notice that the allocation order determines the sizes of the per cpu
3873 * caches. Each processor has always one slab available for allocations.
3874 * Increasing the allocation order reduces the number of times that slabs
3875 * must be moved on and off the partial lists and is therefore a factor in
3876 * locking overhead.
3877 */
3878
3879 /*
3880 * Minimum / Maximum order of slab pages. This influences locking overhead
3881 * and slab fragmentation. A higher order reduces the number of partial slabs
3882 * and increases the number of allocations possible without having to
3883 * take the list_lock.
3884 */
3885 static unsigned int slub_min_order;
3886 static unsigned int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
3887 static unsigned int slub_min_objects;
3888
3889 /*
3890 * Calculate the order of allocation given an slab object size.
3891 *
3892 * The order of allocation has significant impact on performance and other
3893 * system components. Generally order 0 allocations should be preferred since
3894 * order 0 does not cause fragmentation in the page allocator. Larger objects
3895 * be problematic to put into order 0 slabs because there may be too much
3896 * unused space left. We go to a higher order if more than 1/16th of the slab
3897 * would be wasted.
3898 *
3899 * In order to reach satisfactory performance we must ensure that a minimum
3900 * number of objects is in one slab. Otherwise we may generate too much
3901 * activity on the partial lists which requires taking the list_lock. This is
3902 * less a concern for large slabs though which are rarely used.
3903 *
3904 * slub_max_order specifies the order where we begin to stop considering the
3905 * number of objects in a slab as critical. If we reach slub_max_order then
3906 * we try to keep the page order as low as possible. So we accept more waste
3907 * of space in favor of a small page order.
3908 *
3909 * Higher order allocations also allow the placement of more objects in a
3910 * slab and thereby reduce object handling overhead. If the user has
3911 * requested a higher minimum order then we start with that one instead of
3912 * the smallest order which will fit the object.
3913 */
calc_slab_order(unsigned int size,unsigned int min_objects,unsigned int max_order,unsigned int fract_leftover)3914 static inline unsigned int calc_slab_order(unsigned int size,
3915 unsigned int min_objects, unsigned int max_order,
3916 unsigned int fract_leftover)
3917 {
3918 unsigned int min_order = slub_min_order;
3919 unsigned int order;
3920
3921 if (order_objects(min_order, size) > MAX_OBJS_PER_PAGE)
3922 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
3923
3924 for (order = max(min_order, (unsigned int)get_order(min_objects * size));
3925 order <= max_order; order++) {
3926
3927 unsigned int slab_size = (unsigned int)PAGE_SIZE << order;
3928 unsigned int rem;
3929
3930 rem = slab_size % size;
3931
3932 if (rem <= slab_size / fract_leftover)
3933 break;
3934 }
3935
3936 return order;
3937 }
3938
calculate_order(unsigned int size)3939 static inline int calculate_order(unsigned int size)
3940 {
3941 unsigned int order;
3942 unsigned int min_objects;
3943 unsigned int max_objects;
3944 unsigned int nr_cpus;
3945
3946 /*
3947 * Attempt to find best configuration for a slab. This
3948 * works by first attempting to generate a layout with
3949 * the best configuration and backing off gradually.
3950 *
3951 * First we increase the acceptable waste in a slab. Then
3952 * we reduce the minimum objects required in a slab.
3953 */
3954 min_objects = slub_min_objects;
3955 if (!min_objects) {
3956 /*
3957 * Some architectures will only update present cpus when
3958 * onlining them, so don't trust the number if it's just 1. But
3959 * we also don't want to use nr_cpu_ids always, as on some other
3960 * architectures, there can be many possible cpus, but never
3961 * onlined. Here we compromise between trying to avoid too high
3962 * order on systems that appear larger than they are, and too
3963 * low order on systems that appear smaller than they are.
3964 */
3965 nr_cpus = num_present_cpus();
3966 if (nr_cpus <= 1)
3967 nr_cpus = nr_cpu_ids;
3968 min_objects = 4 * (fls(nr_cpus) + 1);
3969 }
3970 max_objects = order_objects(slub_max_order, size);
3971 min_objects = min(min_objects, max_objects);
3972
3973 while (min_objects > 1) {
3974 unsigned int fraction;
3975
3976 fraction = 16;
3977 while (fraction >= 4) {
3978 order = calc_slab_order(size, min_objects,
3979 slub_max_order, fraction);
3980 if (order <= slub_max_order)
3981 return order;
3982 fraction /= 2;
3983 }
3984 min_objects--;
3985 }
3986
3987 /*
3988 * We were unable to place multiple objects in a slab. Now
3989 * lets see if we can place a single object there.
3990 */
3991 order = calc_slab_order(size, 1, slub_max_order, 1);
3992 if (order <= slub_max_order)
3993 return order;
3994
3995 /*
3996 * Doh this slab cannot be placed using slub_max_order.
3997 */
3998 order = calc_slab_order(size, 1, MAX_ORDER, 1);
3999 if (order < MAX_ORDER)
4000 return order;
4001 return -ENOSYS;
4002 }
4003
4004 static void
init_kmem_cache_node(struct kmem_cache_node * n)4005 init_kmem_cache_node(struct kmem_cache_node *n)
4006 {
4007 n->nr_partial = 0;
4008 spin_lock_init(&n->list_lock);
4009 INIT_LIST_HEAD(&n->partial);
4010 #ifdef CONFIG_SLUB_DEBUG
4011 atomic_long_set(&n->nr_slabs, 0);
4012 atomic_long_set(&n->total_objects, 0);
4013 INIT_LIST_HEAD(&n->full);
4014 #endif
4015 }
4016
alloc_kmem_cache_cpus(struct kmem_cache * s)4017 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
4018 {
4019 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
4020 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
4021
4022 /*
4023 * Must align to double word boundary for the double cmpxchg
4024 * instructions to work; see __pcpu_double_call_return_bool().
4025 */
4026 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
4027 2 * sizeof(void *));
4028
4029 if (!s->cpu_slab)
4030 return 0;
4031
4032 init_kmem_cache_cpus(s);
4033
4034 return 1;
4035 }
4036
4037 static struct kmem_cache *kmem_cache_node;
4038
4039 /*
4040 * No kmalloc_node yet so do it by hand. We know that this is the first
4041 * slab on the node for this slabcache. There are no concurrent accesses
4042 * possible.
4043 *
4044 * Note that this function only works on the kmem_cache_node
4045 * when allocating for the kmem_cache_node. This is used for bootstrapping
4046 * memory on a fresh node that has no slab structures yet.
4047 */
early_kmem_cache_node_alloc(int node)4048 static void early_kmem_cache_node_alloc(int node)
4049 {
4050 struct slab *slab;
4051 struct kmem_cache_node *n;
4052
4053 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
4054
4055 slab = new_slab(kmem_cache_node, GFP_NOWAIT, node);
4056
4057 BUG_ON(!slab);
4058 inc_slabs_node(kmem_cache_node, slab_nid(slab), slab->objects);
4059 if (slab_nid(slab) != node) {
4060 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
4061 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
4062 }
4063
4064 n = slab->freelist;
4065 BUG_ON(!n);
4066 #ifdef CONFIG_SLUB_DEBUG
4067 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
4068 init_tracking(kmem_cache_node, n);
4069 #endif
4070 n = kasan_slab_alloc(kmem_cache_node, n, GFP_KERNEL, false);
4071 slab->freelist = get_freepointer(kmem_cache_node, n);
4072 slab->inuse = 1;
4073 kmem_cache_node->node[node] = n;
4074 init_kmem_cache_node(n);
4075 inc_slabs_node(kmem_cache_node, node, slab->objects);
4076
4077 /*
4078 * No locks need to be taken here as it has just been
4079 * initialized and there is no concurrent access.
4080 */
4081 __add_partial(n, slab, DEACTIVATE_TO_HEAD);
4082 }
4083
free_kmem_cache_nodes(struct kmem_cache * s)4084 static void free_kmem_cache_nodes(struct kmem_cache *s)
4085 {
4086 int node;
4087 struct kmem_cache_node *n;
4088
4089 for_each_kmem_cache_node(s, node, n) {
4090 s->node[node] = NULL;
4091 kmem_cache_free(kmem_cache_node, n);
4092 }
4093 }
4094
__kmem_cache_release(struct kmem_cache * s)4095 void __kmem_cache_release(struct kmem_cache *s)
4096 {
4097 cache_random_seq_destroy(s);
4098 free_percpu(s->cpu_slab);
4099 free_kmem_cache_nodes(s);
4100 }
4101
init_kmem_cache_nodes(struct kmem_cache * s)4102 static int init_kmem_cache_nodes(struct kmem_cache *s)
4103 {
4104 int node;
4105
4106 for_each_node_mask(node, slab_nodes) {
4107 struct kmem_cache_node *n;
4108
4109 if (slab_state == DOWN) {
4110 early_kmem_cache_node_alloc(node);
4111 continue;
4112 }
4113 n = kmem_cache_alloc_node(kmem_cache_node,
4114 GFP_KERNEL, node);
4115
4116 if (!n) {
4117 free_kmem_cache_nodes(s);
4118 return 0;
4119 }
4120
4121 init_kmem_cache_node(n);
4122 s->node[node] = n;
4123 }
4124 return 1;
4125 }
4126
set_cpu_partial(struct kmem_cache * s)4127 static void set_cpu_partial(struct kmem_cache *s)
4128 {
4129 #ifdef CONFIG_SLUB_CPU_PARTIAL
4130 unsigned int nr_objects;
4131
4132 /*
4133 * cpu_partial determined the maximum number of objects kept in the
4134 * per cpu partial lists of a processor.
4135 *
4136 * Per cpu partial lists mainly contain slabs that just have one
4137 * object freed. If they are used for allocation then they can be
4138 * filled up again with minimal effort. The slab will never hit the
4139 * per node partial lists and therefore no locking will be required.
4140 *
4141 * For backwards compatibility reasons, this is determined as number
4142 * of objects, even though we now limit maximum number of pages, see
4143 * slub_set_cpu_partial()
4144 */
4145 if (!kmem_cache_has_cpu_partial(s))
4146 nr_objects = 0;
4147 else if (s->size >= PAGE_SIZE)
4148 nr_objects = 6;
4149 else if (s->size >= 1024)
4150 nr_objects = 24;
4151 else if (s->size >= 256)
4152 nr_objects = 52;
4153 else
4154 nr_objects = 120;
4155
4156 slub_set_cpu_partial(s, nr_objects);
4157 #endif
4158 }
4159
4160 /*
4161 * calculate_sizes() determines the order and the distribution of data within
4162 * a slab object.
4163 */
calculate_sizes(struct kmem_cache * s)4164 static int calculate_sizes(struct kmem_cache *s)
4165 {
4166 slab_flags_t flags = s->flags;
4167 unsigned int size = s->object_size;
4168 unsigned int order;
4169
4170 /*
4171 * Round up object size to the next word boundary. We can only
4172 * place the free pointer at word boundaries and this determines
4173 * the possible location of the free pointer.
4174 */
4175 size = ALIGN(size, sizeof(void *));
4176
4177 #ifdef CONFIG_SLUB_DEBUG
4178 /*
4179 * Determine if we can poison the object itself. If the user of
4180 * the slab may touch the object after free or before allocation
4181 * then we should never poison the object itself.
4182 */
4183 if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) &&
4184 !s->ctor)
4185 s->flags |= __OBJECT_POISON;
4186 else
4187 s->flags &= ~__OBJECT_POISON;
4188
4189
4190 /*
4191 * If we are Redzoning then check if there is some space between the
4192 * end of the object and the free pointer. If not then add an
4193 * additional word to have some bytes to store Redzone information.
4194 */
4195 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
4196 size += sizeof(void *);
4197 #endif
4198
4199 /*
4200 * With that we have determined the number of bytes in actual use
4201 * by the object and redzoning.
4202 */
4203 s->inuse = size;
4204
4205 if ((flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) ||
4206 ((flags & SLAB_RED_ZONE) && s->object_size < sizeof(void *)) ||
4207 s->ctor) {
4208 /*
4209 * Relocate free pointer after the object if it is not
4210 * permitted to overwrite the first word of the object on
4211 * kmem_cache_free.
4212 *
4213 * This is the case if we do RCU, have a constructor or
4214 * destructor, are poisoning the objects, or are
4215 * redzoning an object smaller than sizeof(void *).
4216 *
4217 * The assumption that s->offset >= s->inuse means free
4218 * pointer is outside of the object is used in the
4219 * freeptr_outside_object() function. If that is no
4220 * longer true, the function needs to be modified.
4221 */
4222 s->offset = size;
4223 size += sizeof(void *);
4224 } else {
4225 /*
4226 * Store freelist pointer near middle of object to keep
4227 * it away from the edges of the object to avoid small
4228 * sized over/underflows from neighboring allocations.
4229 */
4230 s->offset = ALIGN_DOWN(s->object_size / 2, sizeof(void *));
4231 }
4232
4233 #ifdef CONFIG_SLUB_DEBUG
4234 if (flags & SLAB_STORE_USER) {
4235 /*
4236 * Need to store information about allocs and frees after
4237 * the object.
4238 */
4239 size += 2 * sizeof(struct track);
4240
4241 /* Save the original kmalloc request size */
4242 if (flags & SLAB_KMALLOC)
4243 size += sizeof(unsigned int);
4244 }
4245 #endif
4246
4247 kasan_cache_create(s, &size, &s->flags);
4248 #ifdef CONFIG_SLUB_DEBUG
4249 if (flags & SLAB_RED_ZONE) {
4250 /*
4251 * Add some empty padding so that we can catch
4252 * overwrites from earlier objects rather than let
4253 * tracking information or the free pointer be
4254 * corrupted if a user writes before the start
4255 * of the object.
4256 */
4257 size += sizeof(void *);
4258
4259 s->red_left_pad = sizeof(void *);
4260 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
4261 size += s->red_left_pad;
4262 }
4263 #endif
4264
4265 /*
4266 * SLUB stores one object immediately after another beginning from
4267 * offset 0. In order to align the objects we have to simply size
4268 * each object to conform to the alignment.
4269 */
4270 size = ALIGN(size, s->align);
4271 s->size = size;
4272 s->reciprocal_size = reciprocal_value(size);
4273 order = calculate_order(size);
4274
4275 if ((int)order < 0)
4276 return 0;
4277
4278 s->allocflags = 0;
4279 if (order)
4280 s->allocflags |= __GFP_COMP;
4281
4282 if (s->flags & SLAB_CACHE_DMA)
4283 s->allocflags |= GFP_DMA;
4284
4285 if (s->flags & SLAB_CACHE_DMA32)
4286 s->allocflags |= GFP_DMA32;
4287
4288 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4289 s->allocflags |= __GFP_RECLAIMABLE;
4290
4291 /*
4292 * Determine the number of objects per slab
4293 */
4294 s->oo = oo_make(order, size);
4295 s->min = oo_make(get_order(size), size);
4296
4297 return !!oo_objects(s->oo);
4298 }
4299
kmem_cache_open(struct kmem_cache * s,slab_flags_t flags)4300 static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags)
4301 {
4302 s->flags = kmem_cache_flags(s->size, flags, s->name);
4303 #ifdef CONFIG_SLAB_FREELIST_HARDENED
4304 s->random = get_random_long();
4305 #endif
4306
4307 if (!calculate_sizes(s))
4308 goto error;
4309 if (disable_higher_order_debug) {
4310 /*
4311 * Disable debugging flags that store metadata if the min slab
4312 * order increased.
4313 */
4314 if (get_order(s->size) > get_order(s->object_size)) {
4315 s->flags &= ~DEBUG_METADATA_FLAGS;
4316 s->offset = 0;
4317 if (!calculate_sizes(s))
4318 goto error;
4319 }
4320 }
4321
4322 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
4323 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
4324 if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0)
4325 /* Enable fast mode */
4326 s->flags |= __CMPXCHG_DOUBLE;
4327 #endif
4328
4329 /*
4330 * The larger the object size is, the more slabs we want on the partial
4331 * list to avoid pounding the page allocator excessively.
4332 */
4333 s->min_partial = min_t(unsigned long, MAX_PARTIAL, ilog2(s->size) / 2);
4334 s->min_partial = max_t(unsigned long, MIN_PARTIAL, s->min_partial);
4335
4336 set_cpu_partial(s);
4337
4338 #ifdef CONFIG_NUMA
4339 s->remote_node_defrag_ratio = 1000;
4340 #endif
4341
4342 /* Initialize the pre-computed randomized freelist if slab is up */
4343 if (slab_state >= UP) {
4344 if (init_cache_random_seq(s))
4345 goto error;
4346 }
4347
4348 if (!init_kmem_cache_nodes(s))
4349 goto error;
4350
4351 if (alloc_kmem_cache_cpus(s))
4352 return 0;
4353
4354 error:
4355 __kmem_cache_release(s);
4356 return -EINVAL;
4357 }
4358
list_slab_objects(struct kmem_cache * s,struct slab * slab,const char * text)4359 static void list_slab_objects(struct kmem_cache *s, struct slab *slab,
4360 const char *text)
4361 {
4362 #ifdef CONFIG_SLUB_DEBUG
4363 void *addr = slab_address(slab);
4364 void *p;
4365
4366 slab_err(s, slab, text, s->name);
4367
4368 spin_lock(&object_map_lock);
4369 __fill_map(object_map, s, slab);
4370
4371 for_each_object(p, s, addr, slab->objects) {
4372
4373 if (!test_bit(__obj_to_index(s, addr, p), object_map)) {
4374 pr_err("Object 0x%p @offset=%tu\n", p, p - addr);
4375 print_tracking(s, p);
4376 }
4377 }
4378 spin_unlock(&object_map_lock);
4379 #endif
4380 }
4381
4382 /*
4383 * Attempt to free all partial slabs on a node.
4384 * This is called from __kmem_cache_shutdown(). We must take list_lock
4385 * because sysfs file might still access partial list after the shutdowning.
4386 */
free_partial(struct kmem_cache * s,struct kmem_cache_node * n)4387 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
4388 {
4389 LIST_HEAD(discard);
4390 struct slab *slab, *h;
4391
4392 BUG_ON(irqs_disabled());
4393 spin_lock_irq(&n->list_lock);
4394 list_for_each_entry_safe(slab, h, &n->partial, slab_list) {
4395 if (!slab->inuse) {
4396 remove_partial(n, slab);
4397 list_add(&slab->slab_list, &discard);
4398 } else {
4399 list_slab_objects(s, slab,
4400 "Objects remaining in %s on __kmem_cache_shutdown()");
4401 }
4402 }
4403 spin_unlock_irq(&n->list_lock);
4404
4405 list_for_each_entry_safe(slab, h, &discard, slab_list)
4406 discard_slab(s, slab);
4407 }
4408
__kmem_cache_empty(struct kmem_cache * s)4409 bool __kmem_cache_empty(struct kmem_cache *s)
4410 {
4411 int node;
4412 struct kmem_cache_node *n;
4413
4414 for_each_kmem_cache_node(s, node, n)
4415 if (n->nr_partial || slabs_node(s, node))
4416 return false;
4417 return true;
4418 }
4419
4420 /*
4421 * Release all resources used by a slab cache.
4422 */
__kmem_cache_shutdown(struct kmem_cache * s)4423 int __kmem_cache_shutdown(struct kmem_cache *s)
4424 {
4425 int node;
4426 struct kmem_cache_node *n;
4427
4428 flush_all_cpus_locked(s);
4429 /* Attempt to free all objects */
4430 for_each_kmem_cache_node(s, node, n) {
4431 free_partial(s, n);
4432 if (n->nr_partial || slabs_node(s, node))
4433 return 1;
4434 }
4435 return 0;
4436 }
4437
4438 #ifdef CONFIG_PRINTK
__kmem_obj_info(struct kmem_obj_info * kpp,void * object,struct slab * slab)4439 void __kmem_obj_info(struct kmem_obj_info *kpp, void *object, struct slab *slab)
4440 {
4441 void *base;
4442 int __maybe_unused i;
4443 unsigned int objnr;
4444 void *objp;
4445 void *objp0;
4446 struct kmem_cache *s = slab->slab_cache;
4447 struct track __maybe_unused *trackp;
4448
4449 kpp->kp_ptr = object;
4450 kpp->kp_slab = slab;
4451 kpp->kp_slab_cache = s;
4452 base = slab_address(slab);
4453 objp0 = kasan_reset_tag(object);
4454 #ifdef CONFIG_SLUB_DEBUG
4455 objp = restore_red_left(s, objp0);
4456 #else
4457 objp = objp0;
4458 #endif
4459 objnr = obj_to_index(s, slab, objp);
4460 kpp->kp_data_offset = (unsigned long)((char *)objp0 - (char *)objp);
4461 objp = base + s->size * objnr;
4462 kpp->kp_objp = objp;
4463 if (WARN_ON_ONCE(objp < base || objp >= base + slab->objects * s->size
4464 || (objp - base) % s->size) ||
4465 !(s->flags & SLAB_STORE_USER))
4466 return;
4467 #ifdef CONFIG_SLUB_DEBUG
4468 objp = fixup_red_left(s, objp);
4469 trackp = get_track(s, objp, TRACK_ALLOC);
4470 kpp->kp_ret = (void *)trackp->addr;
4471 #ifdef CONFIG_STACKDEPOT
4472 {
4473 depot_stack_handle_t handle;
4474 unsigned long *entries;
4475 unsigned int nr_entries;
4476
4477 handle = READ_ONCE(trackp->handle);
4478 if (handle) {
4479 nr_entries = stack_depot_fetch(handle, &entries);
4480 for (i = 0; i < KS_ADDRS_COUNT && i < nr_entries; i++)
4481 kpp->kp_stack[i] = (void *)entries[i];
4482 }
4483
4484 trackp = get_track(s, objp, TRACK_FREE);
4485 handle = READ_ONCE(trackp->handle);
4486 if (handle) {
4487 nr_entries = stack_depot_fetch(handle, &entries);
4488 for (i = 0; i < KS_ADDRS_COUNT && i < nr_entries; i++)
4489 kpp->kp_free_stack[i] = (void *)entries[i];
4490 }
4491 }
4492 #endif
4493 #endif
4494 }
4495 #endif
4496
4497 /********************************************************************
4498 * Kmalloc subsystem
4499 *******************************************************************/
4500
setup_slub_min_order(char * str)4501 static int __init setup_slub_min_order(char *str)
4502 {
4503 get_option(&str, (int *)&slub_min_order);
4504
4505 return 1;
4506 }
4507
4508 __setup("slub_min_order=", setup_slub_min_order);
4509
setup_slub_max_order(char * str)4510 static int __init setup_slub_max_order(char *str)
4511 {
4512 get_option(&str, (int *)&slub_max_order);
4513 slub_max_order = min(slub_max_order, (unsigned int)MAX_ORDER - 1);
4514
4515 return 1;
4516 }
4517
4518 __setup("slub_max_order=", setup_slub_max_order);
4519
setup_slub_min_objects(char * str)4520 static int __init setup_slub_min_objects(char *str)
4521 {
4522 get_option(&str, (int *)&slub_min_objects);
4523
4524 return 1;
4525 }
4526
4527 __setup("slub_min_objects=", setup_slub_min_objects);
4528
4529 #ifdef CONFIG_HARDENED_USERCOPY
4530 /*
4531 * Rejects incorrectly sized objects and objects that are to be copied
4532 * to/from userspace but do not fall entirely within the containing slab
4533 * cache's usercopy region.
4534 *
4535 * Returns NULL if check passes, otherwise const char * to name of cache
4536 * to indicate an error.
4537 */
__check_heap_object(const void * ptr,unsigned long n,const struct slab * slab,bool to_user)4538 void __check_heap_object(const void *ptr, unsigned long n,
4539 const struct slab *slab, bool to_user)
4540 {
4541 struct kmem_cache *s;
4542 unsigned int offset;
4543 bool is_kfence = is_kfence_address(ptr);
4544
4545 ptr = kasan_reset_tag(ptr);
4546
4547 /* Find object and usable object size. */
4548 s = slab->slab_cache;
4549
4550 /* Reject impossible pointers. */
4551 if (ptr < slab_address(slab))
4552 usercopy_abort("SLUB object not in SLUB page?!", NULL,
4553 to_user, 0, n);
4554
4555 /* Find offset within object. */
4556 if (is_kfence)
4557 offset = ptr - kfence_object_start(ptr);
4558 else
4559 offset = (ptr - slab_address(slab)) % s->size;
4560
4561 /* Adjust for redzone and reject if within the redzone. */
4562 if (!is_kfence && kmem_cache_debug_flags(s, SLAB_RED_ZONE)) {
4563 if (offset < s->red_left_pad)
4564 usercopy_abort("SLUB object in left red zone",
4565 s->name, to_user, offset, n);
4566 offset -= s->red_left_pad;
4567 }
4568
4569 /* Allow address range falling entirely within usercopy region. */
4570 if (offset >= s->useroffset &&
4571 offset - s->useroffset <= s->usersize &&
4572 n <= s->useroffset - offset + s->usersize)
4573 return;
4574
4575 usercopy_abort("SLUB object", s->name, to_user, offset, n);
4576 }
4577 #endif /* CONFIG_HARDENED_USERCOPY */
4578
4579 #define SHRINK_PROMOTE_MAX 32
4580
4581 /*
4582 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
4583 * up most to the head of the partial lists. New allocations will then
4584 * fill those up and thus they can be removed from the partial lists.
4585 *
4586 * The slabs with the least items are placed last. This results in them
4587 * being allocated from last increasing the chance that the last objects
4588 * are freed in them.
4589 */
__kmem_cache_do_shrink(struct kmem_cache * s)4590 static int __kmem_cache_do_shrink(struct kmem_cache *s)
4591 {
4592 int node;
4593 int i;
4594 struct kmem_cache_node *n;
4595 struct slab *slab;
4596 struct slab *t;
4597 struct list_head discard;
4598 struct list_head promote[SHRINK_PROMOTE_MAX];
4599 unsigned long flags;
4600 int ret = 0;
4601
4602 for_each_kmem_cache_node(s, node, n) {
4603 INIT_LIST_HEAD(&discard);
4604 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
4605 INIT_LIST_HEAD(promote + i);
4606
4607 spin_lock_irqsave(&n->list_lock, flags);
4608
4609 /*
4610 * Build lists of slabs to discard or promote.
4611 *
4612 * Note that concurrent frees may occur while we hold the
4613 * list_lock. slab->inuse here is the upper limit.
4614 */
4615 list_for_each_entry_safe(slab, t, &n->partial, slab_list) {
4616 int free = slab->objects - slab->inuse;
4617
4618 /* Do not reread slab->inuse */
4619 barrier();
4620
4621 /* We do not keep full slabs on the list */
4622 BUG_ON(free <= 0);
4623
4624 if (free == slab->objects) {
4625 list_move(&slab->slab_list, &discard);
4626 n->nr_partial--;
4627 dec_slabs_node(s, node, slab->objects);
4628 } else if (free <= SHRINK_PROMOTE_MAX)
4629 list_move(&slab->slab_list, promote + free - 1);
4630 }
4631
4632 /*
4633 * Promote the slabs filled up most to the head of the
4634 * partial list.
4635 */
4636 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
4637 list_splice(promote + i, &n->partial);
4638
4639 spin_unlock_irqrestore(&n->list_lock, flags);
4640
4641 /* Release empty slabs */
4642 list_for_each_entry_safe(slab, t, &discard, slab_list)
4643 free_slab(s, slab);
4644
4645 if (slabs_node(s, node))
4646 ret = 1;
4647 }
4648
4649 return ret;
4650 }
4651
__kmem_cache_shrink(struct kmem_cache * s)4652 int __kmem_cache_shrink(struct kmem_cache *s)
4653 {
4654 flush_all(s);
4655 return __kmem_cache_do_shrink(s);
4656 }
4657
slab_mem_going_offline_callback(void * arg)4658 static int slab_mem_going_offline_callback(void *arg)
4659 {
4660 struct kmem_cache *s;
4661
4662 mutex_lock(&slab_mutex);
4663 list_for_each_entry(s, &slab_caches, list) {
4664 flush_all_cpus_locked(s);
4665 __kmem_cache_do_shrink(s);
4666 }
4667 mutex_unlock(&slab_mutex);
4668
4669 return 0;
4670 }
4671
slab_mem_offline_callback(void * arg)4672 static void slab_mem_offline_callback(void *arg)
4673 {
4674 struct memory_notify *marg = arg;
4675 int offline_node;
4676
4677 offline_node = marg->status_change_nid_normal;
4678
4679 /*
4680 * If the node still has available memory. we need kmem_cache_node
4681 * for it yet.
4682 */
4683 if (offline_node < 0)
4684 return;
4685
4686 mutex_lock(&slab_mutex);
4687 node_clear(offline_node, slab_nodes);
4688 /*
4689 * We no longer free kmem_cache_node structures here, as it would be
4690 * racy with all get_node() users, and infeasible to protect them with
4691 * slab_mutex.
4692 */
4693 mutex_unlock(&slab_mutex);
4694 }
4695
slab_mem_going_online_callback(void * arg)4696 static int slab_mem_going_online_callback(void *arg)
4697 {
4698 struct kmem_cache_node *n;
4699 struct kmem_cache *s;
4700 struct memory_notify *marg = arg;
4701 int nid = marg->status_change_nid_normal;
4702 int ret = 0;
4703
4704 /*
4705 * If the node's memory is already available, then kmem_cache_node is
4706 * already created. Nothing to do.
4707 */
4708 if (nid < 0)
4709 return 0;
4710
4711 /*
4712 * We are bringing a node online. No memory is available yet. We must
4713 * allocate a kmem_cache_node structure in order to bring the node
4714 * online.
4715 */
4716 mutex_lock(&slab_mutex);
4717 list_for_each_entry(s, &slab_caches, list) {
4718 /*
4719 * The structure may already exist if the node was previously
4720 * onlined and offlined.
4721 */
4722 if (get_node(s, nid))
4723 continue;
4724 /*
4725 * XXX: kmem_cache_alloc_node will fallback to other nodes
4726 * since memory is not yet available from the node that
4727 * is brought up.
4728 */
4729 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
4730 if (!n) {
4731 ret = -ENOMEM;
4732 goto out;
4733 }
4734 init_kmem_cache_node(n);
4735 s->node[nid] = n;
4736 }
4737 /*
4738 * Any cache created after this point will also have kmem_cache_node
4739 * initialized for the new node.
4740 */
4741 node_set(nid, slab_nodes);
4742 out:
4743 mutex_unlock(&slab_mutex);
4744 return ret;
4745 }
4746
slab_memory_callback(struct notifier_block * self,unsigned long action,void * arg)4747 static int slab_memory_callback(struct notifier_block *self,
4748 unsigned long action, void *arg)
4749 {
4750 int ret = 0;
4751
4752 switch (action) {
4753 case MEM_GOING_ONLINE:
4754 ret = slab_mem_going_online_callback(arg);
4755 break;
4756 case MEM_GOING_OFFLINE:
4757 ret = slab_mem_going_offline_callback(arg);
4758 break;
4759 case MEM_OFFLINE:
4760 case MEM_CANCEL_ONLINE:
4761 slab_mem_offline_callback(arg);
4762 break;
4763 case MEM_ONLINE:
4764 case MEM_CANCEL_OFFLINE:
4765 break;
4766 }
4767 if (ret)
4768 ret = notifier_from_errno(ret);
4769 else
4770 ret = NOTIFY_OK;
4771 return ret;
4772 }
4773
4774 static struct notifier_block slab_memory_callback_nb = {
4775 .notifier_call = slab_memory_callback,
4776 .priority = SLAB_CALLBACK_PRI,
4777 };
4778
4779 /********************************************************************
4780 * Basic setup of slabs
4781 *******************************************************************/
4782
4783 /*
4784 * Used for early kmem_cache structures that were allocated using
4785 * the page allocator. Allocate them properly then fix up the pointers
4786 * that may be pointing to the wrong kmem_cache structure.
4787 */
4788
bootstrap(struct kmem_cache * static_cache)4789 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
4790 {
4791 int node;
4792 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
4793 struct kmem_cache_node *n;
4794
4795 memcpy(s, static_cache, kmem_cache->object_size);
4796
4797 /*
4798 * This runs very early, and only the boot processor is supposed to be
4799 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4800 * IPIs around.
4801 */
4802 __flush_cpu_slab(s, smp_processor_id());
4803 for_each_kmem_cache_node(s, node, n) {
4804 struct slab *p;
4805
4806 list_for_each_entry(p, &n->partial, slab_list)
4807 p->slab_cache = s;
4808
4809 #ifdef CONFIG_SLUB_DEBUG
4810 list_for_each_entry(p, &n->full, slab_list)
4811 p->slab_cache = s;
4812 #endif
4813 }
4814 list_add(&s->list, &slab_caches);
4815 return s;
4816 }
4817
kmem_cache_init(void)4818 void __init kmem_cache_init(void)
4819 {
4820 static __initdata struct kmem_cache boot_kmem_cache,
4821 boot_kmem_cache_node;
4822 int node;
4823
4824 if (debug_guardpage_minorder())
4825 slub_max_order = 0;
4826
4827 /* Print slub debugging pointers without hashing */
4828 if (__slub_debug_enabled())
4829 no_hash_pointers_enable(NULL);
4830
4831 kmem_cache_node = &boot_kmem_cache_node;
4832 kmem_cache = &boot_kmem_cache;
4833
4834 /*
4835 * Initialize the nodemask for which we will allocate per node
4836 * structures. Here we don't need taking slab_mutex yet.
4837 */
4838 for_each_node_state(node, N_NORMAL_MEMORY)
4839 node_set(node, slab_nodes);
4840
4841 create_boot_cache(kmem_cache_node, "kmem_cache_node",
4842 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN, 0, 0);
4843
4844 register_hotmemory_notifier(&slab_memory_callback_nb);
4845
4846 /* Able to allocate the per node structures */
4847 slab_state = PARTIAL;
4848
4849 create_boot_cache(kmem_cache, "kmem_cache",
4850 offsetof(struct kmem_cache, node) +
4851 nr_node_ids * sizeof(struct kmem_cache_node *),
4852 SLAB_HWCACHE_ALIGN, 0, 0);
4853
4854 kmem_cache = bootstrap(&boot_kmem_cache);
4855 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
4856
4857 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4858 setup_kmalloc_cache_index_table();
4859 create_kmalloc_caches(0);
4860
4861 /* Setup random freelists for each cache */
4862 init_freelist_randomization();
4863
4864 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL,
4865 slub_cpu_dead);
4866
4867 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n",
4868 cache_line_size(),
4869 slub_min_order, slub_max_order, slub_min_objects,
4870 nr_cpu_ids, nr_node_ids);
4871 }
4872
kmem_cache_init_late(void)4873 void __init kmem_cache_init_late(void)
4874 {
4875 flushwq = alloc_workqueue("slub_flushwq", WQ_MEM_RECLAIM, 0);
4876 WARN_ON(!flushwq);
4877 }
4878
4879 struct kmem_cache *
__kmem_cache_alias(const char * name,unsigned int size,unsigned int align,slab_flags_t flags,void (* ctor)(void *))4880 __kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
4881 slab_flags_t flags, void (*ctor)(void *))
4882 {
4883 struct kmem_cache *s;
4884
4885 s = find_mergeable(size, align, flags, name, ctor);
4886 if (s) {
4887 if (sysfs_slab_alias(s, name))
4888 return NULL;
4889
4890 s->refcount++;
4891
4892 /*
4893 * Adjust the object sizes so that we clear
4894 * the complete object on kzalloc.
4895 */
4896 s->object_size = max(s->object_size, size);
4897 s->inuse = max(s->inuse, ALIGN(size, sizeof(void *)));
4898 }
4899
4900 return s;
4901 }
4902
__kmem_cache_create(struct kmem_cache * s,slab_flags_t flags)4903 int __kmem_cache_create(struct kmem_cache *s, slab_flags_t flags)
4904 {
4905 int err;
4906
4907 err = kmem_cache_open(s, flags);
4908 if (err)
4909 return err;
4910
4911 /* Mutex is not taken during early boot */
4912 if (slab_state <= UP)
4913 return 0;
4914
4915 err = sysfs_slab_add(s);
4916 if (err) {
4917 __kmem_cache_release(s);
4918 return err;
4919 }
4920
4921 if (s->flags & SLAB_STORE_USER)
4922 debugfs_slab_add(s);
4923
4924 return 0;
4925 }
4926
4927 #ifdef CONFIG_SYSFS
count_inuse(struct slab * slab)4928 static int count_inuse(struct slab *slab)
4929 {
4930 return slab->inuse;
4931 }
4932
count_total(struct slab * slab)4933 static int count_total(struct slab *slab)
4934 {
4935 return slab->objects;
4936 }
4937 #endif
4938
4939 #ifdef CONFIG_SLUB_DEBUG
validate_slab(struct kmem_cache * s,struct slab * slab,unsigned long * obj_map)4940 static void validate_slab(struct kmem_cache *s, struct slab *slab,
4941 unsigned long *obj_map)
4942 {
4943 void *p;
4944 void *addr = slab_address(slab);
4945
4946 if (!check_slab(s, slab) || !on_freelist(s, slab, NULL))
4947 return;
4948
4949 /* Now we know that a valid freelist exists */
4950 __fill_map(obj_map, s, slab);
4951 for_each_object(p, s, addr, slab->objects) {
4952 u8 val = test_bit(__obj_to_index(s, addr, p), obj_map) ?
4953 SLUB_RED_INACTIVE : SLUB_RED_ACTIVE;
4954
4955 if (!check_object(s, slab, p, val))
4956 break;
4957 }
4958 }
4959
validate_slab_node(struct kmem_cache * s,struct kmem_cache_node * n,unsigned long * obj_map)4960 static int validate_slab_node(struct kmem_cache *s,
4961 struct kmem_cache_node *n, unsigned long *obj_map)
4962 {
4963 unsigned long count = 0;
4964 struct slab *slab;
4965 unsigned long flags;
4966
4967 spin_lock_irqsave(&n->list_lock, flags);
4968
4969 list_for_each_entry(slab, &n->partial, slab_list) {
4970 validate_slab(s, slab, obj_map);
4971 count++;
4972 }
4973 if (count != n->nr_partial) {
4974 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4975 s->name, count, n->nr_partial);
4976 slab_add_kunit_errors();
4977 }
4978
4979 if (!(s->flags & SLAB_STORE_USER))
4980 goto out;
4981
4982 list_for_each_entry(slab, &n->full, slab_list) {
4983 validate_slab(s, slab, obj_map);
4984 count++;
4985 }
4986 if (count != atomic_long_read(&n->nr_slabs)) {
4987 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4988 s->name, count, atomic_long_read(&n->nr_slabs));
4989 slab_add_kunit_errors();
4990 }
4991
4992 out:
4993 spin_unlock_irqrestore(&n->list_lock, flags);
4994 return count;
4995 }
4996
validate_slab_cache(struct kmem_cache * s)4997 long validate_slab_cache(struct kmem_cache *s)
4998 {
4999 int node;
5000 unsigned long count = 0;
5001 struct kmem_cache_node *n;
5002 unsigned long *obj_map;
5003
5004 obj_map = bitmap_alloc(oo_objects(s->oo), GFP_KERNEL);
5005 if (!obj_map)
5006 return -ENOMEM;
5007
5008 flush_all(s);
5009 for_each_kmem_cache_node(s, node, n)
5010 count += validate_slab_node(s, n, obj_map);
5011
5012 bitmap_free(obj_map);
5013
5014 return count;
5015 }
5016 EXPORT_SYMBOL(validate_slab_cache);
5017
5018 #ifdef CONFIG_DEBUG_FS
5019 /*
5020 * Generate lists of code addresses where slabcache objects are allocated
5021 * and freed.
5022 */
5023
5024 struct location {
5025 depot_stack_handle_t handle;
5026 unsigned long count;
5027 unsigned long addr;
5028 unsigned long waste;
5029 long long sum_time;
5030 long min_time;
5031 long max_time;
5032 long min_pid;
5033 long max_pid;
5034 DECLARE_BITMAP(cpus, NR_CPUS);
5035 nodemask_t nodes;
5036 };
5037
5038 struct loc_track {
5039 unsigned long max;
5040 unsigned long count;
5041 struct location *loc;
5042 loff_t idx;
5043 };
5044
5045 static struct dentry *slab_debugfs_root;
5046
free_loc_track(struct loc_track * t)5047 static void free_loc_track(struct loc_track *t)
5048 {
5049 if (t->max)
5050 free_pages((unsigned long)t->loc,
5051 get_order(sizeof(struct location) * t->max));
5052 }
5053
alloc_loc_track(struct loc_track * t,unsigned long max,gfp_t flags)5054 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
5055 {
5056 struct location *l;
5057 int order;
5058
5059 order = get_order(sizeof(struct location) * max);
5060
5061 l = (void *)__get_free_pages(flags, order);
5062 if (!l)
5063 return 0;
5064
5065 if (t->count) {
5066 memcpy(l, t->loc, sizeof(struct location) * t->count);
5067 free_loc_track(t);
5068 }
5069 t->max = max;
5070 t->loc = l;
5071 return 1;
5072 }
5073
add_location(struct loc_track * t,struct kmem_cache * s,const struct track * track,unsigned int orig_size)5074 static int add_location(struct loc_track *t, struct kmem_cache *s,
5075 const struct track *track,
5076 unsigned int orig_size)
5077 {
5078 long start, end, pos;
5079 struct location *l;
5080 unsigned long caddr, chandle, cwaste;
5081 unsigned long age = jiffies - track->when;
5082 depot_stack_handle_t handle = 0;
5083 unsigned int waste = s->object_size - orig_size;
5084
5085 #ifdef CONFIG_STACKDEPOT
5086 handle = READ_ONCE(track->handle);
5087 #endif
5088 start = -1;
5089 end = t->count;
5090
5091 for ( ; ; ) {
5092 pos = start + (end - start + 1) / 2;
5093
5094 /*
5095 * There is nothing at "end". If we end up there
5096 * we need to add something to before end.
5097 */
5098 if (pos == end)
5099 break;
5100
5101 l = &t->loc[pos];
5102 caddr = l->addr;
5103 chandle = l->handle;
5104 cwaste = l->waste;
5105 if ((track->addr == caddr) && (handle == chandle) &&
5106 (waste == cwaste)) {
5107
5108 l->count++;
5109 if (track->when) {
5110 l->sum_time += age;
5111 if (age < l->min_time)
5112 l->min_time = age;
5113 if (age > l->max_time)
5114 l->max_time = age;
5115
5116 if (track->pid < l->min_pid)
5117 l->min_pid = track->pid;
5118 if (track->pid > l->max_pid)
5119 l->max_pid = track->pid;
5120
5121 cpumask_set_cpu(track->cpu,
5122 to_cpumask(l->cpus));
5123 }
5124 node_set(page_to_nid(virt_to_page(track)), l->nodes);
5125 return 1;
5126 }
5127
5128 if (track->addr < caddr)
5129 end = pos;
5130 else if (track->addr == caddr && handle < chandle)
5131 end = pos;
5132 else if (track->addr == caddr && handle == chandle &&
5133 waste < cwaste)
5134 end = pos;
5135 else
5136 start = pos;
5137 }
5138
5139 /*
5140 * Not found. Insert new tracking element.
5141 */
5142 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
5143 return 0;
5144
5145 l = t->loc + pos;
5146 if (pos < t->count)
5147 memmove(l + 1, l,
5148 (t->count - pos) * sizeof(struct location));
5149 t->count++;
5150 l->count = 1;
5151 l->addr = track->addr;
5152 l->sum_time = age;
5153 l->min_time = age;
5154 l->max_time = age;
5155 l->min_pid = track->pid;
5156 l->max_pid = track->pid;
5157 l->handle = handle;
5158 l->waste = waste;
5159 cpumask_clear(to_cpumask(l->cpus));
5160 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
5161 nodes_clear(l->nodes);
5162 node_set(page_to_nid(virt_to_page(track)), l->nodes);
5163 return 1;
5164 }
5165
process_slab(struct loc_track * t,struct kmem_cache * s,struct slab * slab,enum track_item alloc,unsigned long * obj_map)5166 static void process_slab(struct loc_track *t, struct kmem_cache *s,
5167 struct slab *slab, enum track_item alloc,
5168 unsigned long *obj_map)
5169 {
5170 void *addr = slab_address(slab);
5171 bool is_alloc = (alloc == TRACK_ALLOC);
5172 void *p;
5173
5174 __fill_map(obj_map, s, slab);
5175
5176 for_each_object(p, s, addr, slab->objects)
5177 if (!test_bit(__obj_to_index(s, addr, p), obj_map))
5178 add_location(t, s, get_track(s, p, alloc),
5179 is_alloc ? get_orig_size(s, p) :
5180 s->object_size);
5181 }
5182 #endif /* CONFIG_DEBUG_FS */
5183 #endif /* CONFIG_SLUB_DEBUG */
5184
5185 #ifdef CONFIG_SYSFS
5186 enum slab_stat_type {
5187 SL_ALL, /* All slabs */
5188 SL_PARTIAL, /* Only partially allocated slabs */
5189 SL_CPU, /* Only slabs used for cpu caches */
5190 SL_OBJECTS, /* Determine allocated objects not slabs */
5191 SL_TOTAL /* Determine object capacity not slabs */
5192 };
5193
5194 #define SO_ALL (1 << SL_ALL)
5195 #define SO_PARTIAL (1 << SL_PARTIAL)
5196 #define SO_CPU (1 << SL_CPU)
5197 #define SO_OBJECTS (1 << SL_OBJECTS)
5198 #define SO_TOTAL (1 << SL_TOTAL)
5199
show_slab_objects(struct kmem_cache * s,char * buf,unsigned long flags)5200 static ssize_t show_slab_objects(struct kmem_cache *s,
5201 char *buf, unsigned long flags)
5202 {
5203 unsigned long total = 0;
5204 int node;
5205 int x;
5206 unsigned long *nodes;
5207 int len = 0;
5208
5209 nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL);
5210 if (!nodes)
5211 return -ENOMEM;
5212
5213 if (flags & SO_CPU) {
5214 int cpu;
5215
5216 for_each_possible_cpu(cpu) {
5217 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
5218 cpu);
5219 int node;
5220 struct slab *slab;
5221
5222 slab = READ_ONCE(c->slab);
5223 if (!slab)
5224 continue;
5225
5226 node = slab_nid(slab);
5227 if (flags & SO_TOTAL)
5228 x = slab->objects;
5229 else if (flags & SO_OBJECTS)
5230 x = slab->inuse;
5231 else
5232 x = 1;
5233
5234 total += x;
5235 nodes[node] += x;
5236
5237 #ifdef CONFIG_SLUB_CPU_PARTIAL
5238 slab = slub_percpu_partial_read_once(c);
5239 if (slab) {
5240 node = slab_nid(slab);
5241 if (flags & SO_TOTAL)
5242 WARN_ON_ONCE(1);
5243 else if (flags & SO_OBJECTS)
5244 WARN_ON_ONCE(1);
5245 else
5246 x = slab->slabs;
5247 total += x;
5248 nodes[node] += x;
5249 }
5250 #endif
5251 }
5252 }
5253
5254 /*
5255 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
5256 * already held which will conflict with an existing lock order:
5257 *
5258 * mem_hotplug_lock->slab_mutex->kernfs_mutex
5259 *
5260 * We don't really need mem_hotplug_lock (to hold off
5261 * slab_mem_going_offline_callback) here because slab's memory hot
5262 * unplug code doesn't destroy the kmem_cache->node[] data.
5263 */
5264
5265 #ifdef CONFIG_SLUB_DEBUG
5266 if (flags & SO_ALL) {
5267 struct kmem_cache_node *n;
5268
5269 for_each_kmem_cache_node(s, node, n) {
5270
5271 if (flags & SO_TOTAL)
5272 x = atomic_long_read(&n->total_objects);
5273 else if (flags & SO_OBJECTS)
5274 x = atomic_long_read(&n->total_objects) -
5275 count_partial(n, count_free);
5276 else
5277 x = atomic_long_read(&n->nr_slabs);
5278 total += x;
5279 nodes[node] += x;
5280 }
5281
5282 } else
5283 #endif
5284 if (flags & SO_PARTIAL) {
5285 struct kmem_cache_node *n;
5286
5287 for_each_kmem_cache_node(s, node, n) {
5288 if (flags & SO_TOTAL)
5289 x = count_partial(n, count_total);
5290 else if (flags & SO_OBJECTS)
5291 x = count_partial(n, count_inuse);
5292 else
5293 x = n->nr_partial;
5294 total += x;
5295 nodes[node] += x;
5296 }
5297 }
5298
5299 len += sysfs_emit_at(buf, len, "%lu", total);
5300 #ifdef CONFIG_NUMA
5301 for (node = 0; node < nr_node_ids; node++) {
5302 if (nodes[node])
5303 len += sysfs_emit_at(buf, len, " N%d=%lu",
5304 node, nodes[node]);
5305 }
5306 #endif
5307 len += sysfs_emit_at(buf, len, "\n");
5308 kfree(nodes);
5309
5310 return len;
5311 }
5312
5313 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
5314 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
5315
5316 struct slab_attribute {
5317 struct attribute attr;
5318 ssize_t (*show)(struct kmem_cache *s, char *buf);
5319 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
5320 };
5321
5322 #define SLAB_ATTR_RO(_name) \
5323 static struct slab_attribute _name##_attr = __ATTR_RO_MODE(_name, 0400)
5324
5325 #define SLAB_ATTR(_name) \
5326 static struct slab_attribute _name##_attr = __ATTR_RW_MODE(_name, 0600)
5327
slab_size_show(struct kmem_cache * s,char * buf)5328 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
5329 {
5330 return sysfs_emit(buf, "%u\n", s->size);
5331 }
5332 SLAB_ATTR_RO(slab_size);
5333
align_show(struct kmem_cache * s,char * buf)5334 static ssize_t align_show(struct kmem_cache *s, char *buf)
5335 {
5336 return sysfs_emit(buf, "%u\n", s->align);
5337 }
5338 SLAB_ATTR_RO(align);
5339
object_size_show(struct kmem_cache * s,char * buf)5340 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
5341 {
5342 return sysfs_emit(buf, "%u\n", s->object_size);
5343 }
5344 SLAB_ATTR_RO(object_size);
5345
objs_per_slab_show(struct kmem_cache * s,char * buf)5346 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
5347 {
5348 return sysfs_emit(buf, "%u\n", oo_objects(s->oo));
5349 }
5350 SLAB_ATTR_RO(objs_per_slab);
5351
order_show(struct kmem_cache * s,char * buf)5352 static ssize_t order_show(struct kmem_cache *s, char *buf)
5353 {
5354 return sysfs_emit(buf, "%u\n", oo_order(s->oo));
5355 }
5356 SLAB_ATTR_RO(order);
5357
min_partial_show(struct kmem_cache * s,char * buf)5358 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
5359 {
5360 return sysfs_emit(buf, "%lu\n", s->min_partial);
5361 }
5362
min_partial_store(struct kmem_cache * s,const char * buf,size_t length)5363 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
5364 size_t length)
5365 {
5366 unsigned long min;
5367 int err;
5368
5369 err = kstrtoul(buf, 10, &min);
5370 if (err)
5371 return err;
5372
5373 s->min_partial = min;
5374 return length;
5375 }
5376 SLAB_ATTR(min_partial);
5377
cpu_partial_show(struct kmem_cache * s,char * buf)5378 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
5379 {
5380 unsigned int nr_partial = 0;
5381 #ifdef CONFIG_SLUB_CPU_PARTIAL
5382 nr_partial = s->cpu_partial;
5383 #endif
5384
5385 return sysfs_emit(buf, "%u\n", nr_partial);
5386 }
5387
cpu_partial_store(struct kmem_cache * s,const char * buf,size_t length)5388 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
5389 size_t length)
5390 {
5391 unsigned int objects;
5392 int err;
5393
5394 err = kstrtouint(buf, 10, &objects);
5395 if (err)
5396 return err;
5397 if (objects && !kmem_cache_has_cpu_partial(s))
5398 return -EINVAL;
5399
5400 slub_set_cpu_partial(s, objects);
5401 flush_all(s);
5402 return length;
5403 }
5404 SLAB_ATTR(cpu_partial);
5405
ctor_show(struct kmem_cache * s,char * buf)5406 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
5407 {
5408 if (!s->ctor)
5409 return 0;
5410 return sysfs_emit(buf, "%pS\n", s->ctor);
5411 }
5412 SLAB_ATTR_RO(ctor);
5413
aliases_show(struct kmem_cache * s,char * buf)5414 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
5415 {
5416 return sysfs_emit(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
5417 }
5418 SLAB_ATTR_RO(aliases);
5419
partial_show(struct kmem_cache * s,char * buf)5420 static ssize_t partial_show(struct kmem_cache *s, char *buf)
5421 {
5422 return show_slab_objects(s, buf, SO_PARTIAL);
5423 }
5424 SLAB_ATTR_RO(partial);
5425
cpu_slabs_show(struct kmem_cache * s,char * buf)5426 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
5427 {
5428 return show_slab_objects(s, buf, SO_CPU);
5429 }
5430 SLAB_ATTR_RO(cpu_slabs);
5431
objects_show(struct kmem_cache * s,char * buf)5432 static ssize_t objects_show(struct kmem_cache *s, char *buf)
5433 {
5434 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
5435 }
5436 SLAB_ATTR_RO(objects);
5437
objects_partial_show(struct kmem_cache * s,char * buf)5438 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
5439 {
5440 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
5441 }
5442 SLAB_ATTR_RO(objects_partial);
5443
slabs_cpu_partial_show(struct kmem_cache * s,char * buf)5444 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
5445 {
5446 int objects = 0;
5447 int slabs = 0;
5448 int cpu __maybe_unused;
5449 int len = 0;
5450
5451 #ifdef CONFIG_SLUB_CPU_PARTIAL
5452 for_each_online_cpu(cpu) {
5453 struct slab *slab;
5454
5455 slab = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5456
5457 if (slab)
5458 slabs += slab->slabs;
5459 }
5460 #endif
5461
5462 /* Approximate half-full slabs, see slub_set_cpu_partial() */
5463 objects = (slabs * oo_objects(s->oo)) / 2;
5464 len += sysfs_emit_at(buf, len, "%d(%d)", objects, slabs);
5465
5466 #if defined(CONFIG_SLUB_CPU_PARTIAL) && defined(CONFIG_SMP)
5467 for_each_online_cpu(cpu) {
5468 struct slab *slab;
5469
5470 slab = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5471 if (slab) {
5472 slabs = READ_ONCE(slab->slabs);
5473 objects = (slabs * oo_objects(s->oo)) / 2;
5474 len += sysfs_emit_at(buf, len, " C%d=%d(%d)",
5475 cpu, objects, slabs);
5476 }
5477 }
5478 #endif
5479 len += sysfs_emit_at(buf, len, "\n");
5480
5481 return len;
5482 }
5483 SLAB_ATTR_RO(slabs_cpu_partial);
5484
reclaim_account_show(struct kmem_cache * s,char * buf)5485 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
5486 {
5487 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
5488 }
5489 SLAB_ATTR_RO(reclaim_account);
5490
hwcache_align_show(struct kmem_cache * s,char * buf)5491 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
5492 {
5493 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
5494 }
5495 SLAB_ATTR_RO(hwcache_align);
5496
5497 #ifdef CONFIG_ZONE_DMA
cache_dma_show(struct kmem_cache * s,char * buf)5498 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
5499 {
5500 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
5501 }
5502 SLAB_ATTR_RO(cache_dma);
5503 #endif
5504
usersize_show(struct kmem_cache * s,char * buf)5505 static ssize_t usersize_show(struct kmem_cache *s, char *buf)
5506 {
5507 return sysfs_emit(buf, "%u\n", s->usersize);
5508 }
5509 SLAB_ATTR_RO(usersize);
5510
destroy_by_rcu_show(struct kmem_cache * s,char * buf)5511 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
5512 {
5513 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU));
5514 }
5515 SLAB_ATTR_RO(destroy_by_rcu);
5516
5517 #ifdef CONFIG_SLUB_DEBUG
slabs_show(struct kmem_cache * s,char * buf)5518 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
5519 {
5520 return show_slab_objects(s, buf, SO_ALL);
5521 }
5522 SLAB_ATTR_RO(slabs);
5523
total_objects_show(struct kmem_cache * s,char * buf)5524 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
5525 {
5526 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
5527 }
5528 SLAB_ATTR_RO(total_objects);
5529
sanity_checks_show(struct kmem_cache * s,char * buf)5530 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
5531 {
5532 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
5533 }
5534 SLAB_ATTR_RO(sanity_checks);
5535
trace_show(struct kmem_cache * s,char * buf)5536 static ssize_t trace_show(struct kmem_cache *s, char *buf)
5537 {
5538 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TRACE));
5539 }
5540 SLAB_ATTR_RO(trace);
5541
red_zone_show(struct kmem_cache * s,char * buf)5542 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
5543 {
5544 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
5545 }
5546
5547 SLAB_ATTR_RO(red_zone);
5548
poison_show(struct kmem_cache * s,char * buf)5549 static ssize_t poison_show(struct kmem_cache *s, char *buf)
5550 {
5551 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_POISON));
5552 }
5553
5554 SLAB_ATTR_RO(poison);
5555
store_user_show(struct kmem_cache * s,char * buf)5556 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
5557 {
5558 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
5559 }
5560
5561 SLAB_ATTR_RO(store_user);
5562
validate_show(struct kmem_cache * s,char * buf)5563 static ssize_t validate_show(struct kmem_cache *s, char *buf)
5564 {
5565 return 0;
5566 }
5567
validate_store(struct kmem_cache * s,const char * buf,size_t length)5568 static ssize_t validate_store(struct kmem_cache *s,
5569 const char *buf, size_t length)
5570 {
5571 int ret = -EINVAL;
5572
5573 if (buf[0] == '1' && kmem_cache_debug(s)) {
5574 ret = validate_slab_cache(s);
5575 if (ret >= 0)
5576 ret = length;
5577 }
5578 return ret;
5579 }
5580 SLAB_ATTR(validate);
5581
5582 #endif /* CONFIG_SLUB_DEBUG */
5583
5584 #ifdef CONFIG_FAILSLAB
failslab_show(struct kmem_cache * s,char * buf)5585 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
5586 {
5587 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
5588 }
5589 SLAB_ATTR_RO(failslab);
5590 #endif
5591
shrink_show(struct kmem_cache * s,char * buf)5592 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
5593 {
5594 return 0;
5595 }
5596
shrink_store(struct kmem_cache * s,const char * buf,size_t length)5597 static ssize_t shrink_store(struct kmem_cache *s,
5598 const char *buf, size_t length)
5599 {
5600 if (buf[0] == '1')
5601 kmem_cache_shrink(s);
5602 else
5603 return -EINVAL;
5604 return length;
5605 }
5606 SLAB_ATTR(shrink);
5607
5608 #ifdef CONFIG_NUMA
remote_node_defrag_ratio_show(struct kmem_cache * s,char * buf)5609 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5610 {
5611 return sysfs_emit(buf, "%u\n", s->remote_node_defrag_ratio / 10);
5612 }
5613
remote_node_defrag_ratio_store(struct kmem_cache * s,const char * buf,size_t length)5614 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5615 const char *buf, size_t length)
5616 {
5617 unsigned int ratio;
5618 int err;
5619
5620 err = kstrtouint(buf, 10, &ratio);
5621 if (err)
5622 return err;
5623 if (ratio > 100)
5624 return -ERANGE;
5625
5626 s->remote_node_defrag_ratio = ratio * 10;
5627
5628 return length;
5629 }
5630 SLAB_ATTR(remote_node_defrag_ratio);
5631 #endif
5632
5633 #ifdef CONFIG_SLUB_STATS
show_stat(struct kmem_cache * s,char * buf,enum stat_item si)5634 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5635 {
5636 unsigned long sum = 0;
5637 int cpu;
5638 int len = 0;
5639 int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL);
5640
5641 if (!data)
5642 return -ENOMEM;
5643
5644 for_each_online_cpu(cpu) {
5645 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5646
5647 data[cpu] = x;
5648 sum += x;
5649 }
5650
5651 len += sysfs_emit_at(buf, len, "%lu", sum);
5652
5653 #ifdef CONFIG_SMP
5654 for_each_online_cpu(cpu) {
5655 if (data[cpu])
5656 len += sysfs_emit_at(buf, len, " C%d=%u",
5657 cpu, data[cpu]);
5658 }
5659 #endif
5660 kfree(data);
5661 len += sysfs_emit_at(buf, len, "\n");
5662
5663 return len;
5664 }
5665
clear_stat(struct kmem_cache * s,enum stat_item si)5666 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5667 {
5668 int cpu;
5669
5670 for_each_online_cpu(cpu)
5671 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5672 }
5673
5674 #define STAT_ATTR(si, text) \
5675 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5676 { \
5677 return show_stat(s, buf, si); \
5678 } \
5679 static ssize_t text##_store(struct kmem_cache *s, \
5680 const char *buf, size_t length) \
5681 { \
5682 if (buf[0] != '0') \
5683 return -EINVAL; \
5684 clear_stat(s, si); \
5685 return length; \
5686 } \
5687 SLAB_ATTR(text); \
5688
5689 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5690 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5691 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5692 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5693 STAT_ATTR(FREE_FROZEN, free_frozen);
5694 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5695 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5696 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5697 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5698 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5699 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5700 STAT_ATTR(FREE_SLAB, free_slab);
5701 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5702 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5703 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5704 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5705 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5706 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5707 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5708 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5709 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5710 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5711 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5712 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5713 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5714 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5715 #endif /* CONFIG_SLUB_STATS */
5716
5717 #ifdef CONFIG_KFENCE
skip_kfence_show(struct kmem_cache * s,char * buf)5718 static ssize_t skip_kfence_show(struct kmem_cache *s, char *buf)
5719 {
5720 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_SKIP_KFENCE));
5721 }
5722
skip_kfence_store(struct kmem_cache * s,const char * buf,size_t length)5723 static ssize_t skip_kfence_store(struct kmem_cache *s,
5724 const char *buf, size_t length)
5725 {
5726 int ret = length;
5727
5728 if (buf[0] == '0')
5729 s->flags &= ~SLAB_SKIP_KFENCE;
5730 else if (buf[0] == '1')
5731 s->flags |= SLAB_SKIP_KFENCE;
5732 else
5733 ret = -EINVAL;
5734
5735 return ret;
5736 }
5737 SLAB_ATTR(skip_kfence);
5738 #endif
5739
5740 static struct attribute *slab_attrs[] = {
5741 &slab_size_attr.attr,
5742 &object_size_attr.attr,
5743 &objs_per_slab_attr.attr,
5744 &order_attr.attr,
5745 &min_partial_attr.attr,
5746 &cpu_partial_attr.attr,
5747 &objects_attr.attr,
5748 &objects_partial_attr.attr,
5749 &partial_attr.attr,
5750 &cpu_slabs_attr.attr,
5751 &ctor_attr.attr,
5752 &aliases_attr.attr,
5753 &align_attr.attr,
5754 &hwcache_align_attr.attr,
5755 &reclaim_account_attr.attr,
5756 &destroy_by_rcu_attr.attr,
5757 &shrink_attr.attr,
5758 &slabs_cpu_partial_attr.attr,
5759 #ifdef CONFIG_SLUB_DEBUG
5760 &total_objects_attr.attr,
5761 &slabs_attr.attr,
5762 &sanity_checks_attr.attr,
5763 &trace_attr.attr,
5764 &red_zone_attr.attr,
5765 &poison_attr.attr,
5766 &store_user_attr.attr,
5767 &validate_attr.attr,
5768 #endif
5769 #ifdef CONFIG_ZONE_DMA
5770 &cache_dma_attr.attr,
5771 #endif
5772 #ifdef CONFIG_NUMA
5773 &remote_node_defrag_ratio_attr.attr,
5774 #endif
5775 #ifdef CONFIG_SLUB_STATS
5776 &alloc_fastpath_attr.attr,
5777 &alloc_slowpath_attr.attr,
5778 &free_fastpath_attr.attr,
5779 &free_slowpath_attr.attr,
5780 &free_frozen_attr.attr,
5781 &free_add_partial_attr.attr,
5782 &free_remove_partial_attr.attr,
5783 &alloc_from_partial_attr.attr,
5784 &alloc_slab_attr.attr,
5785 &alloc_refill_attr.attr,
5786 &alloc_node_mismatch_attr.attr,
5787 &free_slab_attr.attr,
5788 &cpuslab_flush_attr.attr,
5789 &deactivate_full_attr.attr,
5790 &deactivate_empty_attr.attr,
5791 &deactivate_to_head_attr.attr,
5792 &deactivate_to_tail_attr.attr,
5793 &deactivate_remote_frees_attr.attr,
5794 &deactivate_bypass_attr.attr,
5795 &order_fallback_attr.attr,
5796 &cmpxchg_double_fail_attr.attr,
5797 &cmpxchg_double_cpu_fail_attr.attr,
5798 &cpu_partial_alloc_attr.attr,
5799 &cpu_partial_free_attr.attr,
5800 &cpu_partial_node_attr.attr,
5801 &cpu_partial_drain_attr.attr,
5802 #endif
5803 #ifdef CONFIG_FAILSLAB
5804 &failslab_attr.attr,
5805 #endif
5806 &usersize_attr.attr,
5807 #ifdef CONFIG_KFENCE
5808 &skip_kfence_attr.attr,
5809 #endif
5810
5811 NULL
5812 };
5813
5814 static const struct attribute_group slab_attr_group = {
5815 .attrs = slab_attrs,
5816 };
5817
slab_attr_show(struct kobject * kobj,struct attribute * attr,char * buf)5818 static ssize_t slab_attr_show(struct kobject *kobj,
5819 struct attribute *attr,
5820 char *buf)
5821 {
5822 struct slab_attribute *attribute;
5823 struct kmem_cache *s;
5824
5825 attribute = to_slab_attr(attr);
5826 s = to_slab(kobj);
5827
5828 if (!attribute->show)
5829 return -EIO;
5830
5831 return attribute->show(s, buf);
5832 }
5833
slab_attr_store(struct kobject * kobj,struct attribute * attr,const char * buf,size_t len)5834 static ssize_t slab_attr_store(struct kobject *kobj,
5835 struct attribute *attr,
5836 const char *buf, size_t len)
5837 {
5838 struct slab_attribute *attribute;
5839 struct kmem_cache *s;
5840
5841 attribute = to_slab_attr(attr);
5842 s = to_slab(kobj);
5843
5844 if (!attribute->store)
5845 return -EIO;
5846
5847 return attribute->store(s, buf, len);
5848 }
5849
kmem_cache_release(struct kobject * k)5850 static void kmem_cache_release(struct kobject *k)
5851 {
5852 slab_kmem_cache_release(to_slab(k));
5853 }
5854
5855 static const struct sysfs_ops slab_sysfs_ops = {
5856 .show = slab_attr_show,
5857 .store = slab_attr_store,
5858 };
5859
5860 static struct kobj_type slab_ktype = {
5861 .sysfs_ops = &slab_sysfs_ops,
5862 .release = kmem_cache_release,
5863 };
5864
5865 static struct kset *slab_kset;
5866
cache_kset(struct kmem_cache * s)5867 static inline struct kset *cache_kset(struct kmem_cache *s)
5868 {
5869 return slab_kset;
5870 }
5871
5872 #define ID_STR_LENGTH 32
5873
5874 /* Create a unique string id for a slab cache:
5875 *
5876 * Format :[flags-]size
5877 */
create_unique_id(struct kmem_cache * s)5878 static char *create_unique_id(struct kmem_cache *s)
5879 {
5880 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5881 char *p = name;
5882
5883 if (!name)
5884 return ERR_PTR(-ENOMEM);
5885
5886 *p++ = ':';
5887 /*
5888 * First flags affecting slabcache operations. We will only
5889 * get here for aliasable slabs so we do not need to support
5890 * too many flags. The flags here must cover all flags that
5891 * are matched during merging to guarantee that the id is
5892 * unique.
5893 */
5894 if (s->flags & SLAB_CACHE_DMA)
5895 *p++ = 'd';
5896 if (s->flags & SLAB_CACHE_DMA32)
5897 *p++ = 'D';
5898 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5899 *p++ = 'a';
5900 if (s->flags & SLAB_CONSISTENCY_CHECKS)
5901 *p++ = 'F';
5902 if (s->flags & SLAB_ACCOUNT)
5903 *p++ = 'A';
5904 if (p != name + 1)
5905 *p++ = '-';
5906 p += snprintf(p, ID_STR_LENGTH - (p - name), "%07u", s->size);
5907
5908 if (WARN_ON(p > name + ID_STR_LENGTH - 1)) {
5909 kfree(name);
5910 return ERR_PTR(-EINVAL);
5911 }
5912 kmsan_unpoison_memory(name, p - name);
5913 return name;
5914 }
5915
sysfs_slab_add(struct kmem_cache * s)5916 static int sysfs_slab_add(struct kmem_cache *s)
5917 {
5918 int err;
5919 const char *name;
5920 struct kset *kset = cache_kset(s);
5921 int unmergeable = slab_unmergeable(s);
5922
5923 if (!kset) {
5924 kobject_init(&s->kobj, &slab_ktype);
5925 return 0;
5926 }
5927
5928 if (!unmergeable && disable_higher_order_debug &&
5929 (slub_debug & DEBUG_METADATA_FLAGS))
5930 unmergeable = 1;
5931
5932 if (unmergeable) {
5933 /*
5934 * Slabcache can never be merged so we can use the name proper.
5935 * This is typically the case for debug situations. In that
5936 * case we can catch duplicate names easily.
5937 */
5938 sysfs_remove_link(&slab_kset->kobj, s->name);
5939 name = s->name;
5940 } else {
5941 /*
5942 * Create a unique name for the slab as a target
5943 * for the symlinks.
5944 */
5945 name = create_unique_id(s);
5946 if (IS_ERR(name))
5947 return PTR_ERR(name);
5948 }
5949
5950 s->kobj.kset = kset;
5951 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5952 if (err)
5953 goto out;
5954
5955 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5956 if (err)
5957 goto out_del_kobj;
5958
5959 if (!unmergeable) {
5960 /* Setup first alias */
5961 sysfs_slab_alias(s, s->name);
5962 }
5963 out:
5964 if (!unmergeable)
5965 kfree(name);
5966 return err;
5967 out_del_kobj:
5968 kobject_del(&s->kobj);
5969 goto out;
5970 }
5971
sysfs_slab_unlink(struct kmem_cache * s)5972 void sysfs_slab_unlink(struct kmem_cache *s)
5973 {
5974 if (slab_state >= FULL)
5975 kobject_del(&s->kobj);
5976 }
5977
sysfs_slab_release(struct kmem_cache * s)5978 void sysfs_slab_release(struct kmem_cache *s)
5979 {
5980 if (slab_state >= FULL)
5981 kobject_put(&s->kobj);
5982 }
5983
5984 /*
5985 * Need to buffer aliases during bootup until sysfs becomes
5986 * available lest we lose that information.
5987 */
5988 struct saved_alias {
5989 struct kmem_cache *s;
5990 const char *name;
5991 struct saved_alias *next;
5992 };
5993
5994 static struct saved_alias *alias_list;
5995
sysfs_slab_alias(struct kmem_cache * s,const char * name)5996 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5997 {
5998 struct saved_alias *al;
5999
6000 if (slab_state == FULL) {
6001 /*
6002 * If we have a leftover link then remove it.
6003 */
6004 sysfs_remove_link(&slab_kset->kobj, name);
6005 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
6006 }
6007
6008 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
6009 if (!al)
6010 return -ENOMEM;
6011
6012 al->s = s;
6013 al->name = name;
6014 al->next = alias_list;
6015 alias_list = al;
6016 kmsan_unpoison_memory(al, sizeof(*al));
6017 return 0;
6018 }
6019
slab_sysfs_init(void)6020 static int __init slab_sysfs_init(void)
6021 {
6022 struct kmem_cache *s;
6023 int err;
6024
6025 mutex_lock(&slab_mutex);
6026
6027 slab_kset = kset_create_and_add("slab", NULL, kernel_kobj);
6028 if (!slab_kset) {
6029 mutex_unlock(&slab_mutex);
6030 pr_err("Cannot register slab subsystem.\n");
6031 return -ENOSYS;
6032 }
6033
6034 slab_state = FULL;
6035
6036 list_for_each_entry(s, &slab_caches, list) {
6037 err = sysfs_slab_add(s);
6038 if (err)
6039 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
6040 s->name);
6041 }
6042
6043 while (alias_list) {
6044 struct saved_alias *al = alias_list;
6045
6046 alias_list = alias_list->next;
6047 err = sysfs_slab_alias(al->s, al->name);
6048 if (err)
6049 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
6050 al->name);
6051 kfree(al);
6052 }
6053
6054 mutex_unlock(&slab_mutex);
6055 return 0;
6056 }
6057
6058 __initcall(slab_sysfs_init);
6059 #endif /* CONFIG_SYSFS */
6060
6061 #if defined(CONFIG_SLUB_DEBUG) && defined(CONFIG_DEBUG_FS)
slab_debugfs_show(struct seq_file * seq,void * v)6062 static int slab_debugfs_show(struct seq_file *seq, void *v)
6063 {
6064 struct loc_track *t = seq->private;
6065 struct location *l;
6066 unsigned long idx;
6067
6068 idx = (unsigned long) t->idx;
6069 if (idx < t->count) {
6070 l = &t->loc[idx];
6071
6072 seq_printf(seq, "%7ld ", l->count);
6073
6074 if (l->addr)
6075 seq_printf(seq, "%pS", (void *)l->addr);
6076 else
6077 seq_puts(seq, "<not-available>");
6078
6079 if (l->waste)
6080 seq_printf(seq, " waste=%lu/%lu",
6081 l->count * l->waste, l->waste);
6082
6083 if (l->sum_time != l->min_time) {
6084 seq_printf(seq, " age=%ld/%llu/%ld",
6085 l->min_time, div_u64(l->sum_time, l->count),
6086 l->max_time);
6087 } else
6088 seq_printf(seq, " age=%ld", l->min_time);
6089
6090 if (l->min_pid != l->max_pid)
6091 seq_printf(seq, " pid=%ld-%ld", l->min_pid, l->max_pid);
6092 else
6093 seq_printf(seq, " pid=%ld",
6094 l->min_pid);
6095
6096 if (num_online_cpus() > 1 && !cpumask_empty(to_cpumask(l->cpus)))
6097 seq_printf(seq, " cpus=%*pbl",
6098 cpumask_pr_args(to_cpumask(l->cpus)));
6099
6100 if (nr_online_nodes > 1 && !nodes_empty(l->nodes))
6101 seq_printf(seq, " nodes=%*pbl",
6102 nodemask_pr_args(&l->nodes));
6103
6104 #ifdef CONFIG_STACKDEPOT
6105 {
6106 depot_stack_handle_t handle;
6107 unsigned long *entries;
6108 unsigned int nr_entries, j;
6109
6110 handle = READ_ONCE(l->handle);
6111 if (handle) {
6112 nr_entries = stack_depot_fetch(handle, &entries);
6113 seq_puts(seq, "\n");
6114 for (j = 0; j < nr_entries; j++)
6115 seq_printf(seq, " %pS\n", (void *)entries[j]);
6116 }
6117 }
6118 #endif
6119 seq_puts(seq, "\n");
6120 }
6121
6122 if (!idx && !t->count)
6123 seq_puts(seq, "No data\n");
6124
6125 return 0;
6126 }
6127
slab_debugfs_stop(struct seq_file * seq,void * v)6128 static void slab_debugfs_stop(struct seq_file *seq, void *v)
6129 {
6130 }
6131
slab_debugfs_next(struct seq_file * seq,void * v,loff_t * ppos)6132 static void *slab_debugfs_next(struct seq_file *seq, void *v, loff_t *ppos)
6133 {
6134 struct loc_track *t = seq->private;
6135
6136 t->idx = ++(*ppos);
6137 if (*ppos <= t->count)
6138 return ppos;
6139
6140 return NULL;
6141 }
6142
cmp_loc_by_count(const void * a,const void * b,const void * data)6143 static int cmp_loc_by_count(const void *a, const void *b, const void *data)
6144 {
6145 struct location *loc1 = (struct location *)a;
6146 struct location *loc2 = (struct location *)b;
6147
6148 if (loc1->count > loc2->count)
6149 return -1;
6150 else
6151 return 1;
6152 }
6153
slab_debugfs_start(struct seq_file * seq,loff_t * ppos)6154 static void *slab_debugfs_start(struct seq_file *seq, loff_t *ppos)
6155 {
6156 struct loc_track *t = seq->private;
6157
6158 t->idx = *ppos;
6159 return ppos;
6160 }
6161
6162 static const struct seq_operations slab_debugfs_sops = {
6163 .start = slab_debugfs_start,
6164 .next = slab_debugfs_next,
6165 .stop = slab_debugfs_stop,
6166 .show = slab_debugfs_show,
6167 };
6168
slab_debug_trace_open(struct inode * inode,struct file * filep)6169 static int slab_debug_trace_open(struct inode *inode, struct file *filep)
6170 {
6171
6172 struct kmem_cache_node *n;
6173 enum track_item alloc;
6174 int node;
6175 struct loc_track *t = __seq_open_private(filep, &slab_debugfs_sops,
6176 sizeof(struct loc_track));
6177 struct kmem_cache *s = file_inode(filep)->i_private;
6178 unsigned long *obj_map;
6179
6180 if (!t)
6181 return -ENOMEM;
6182
6183 obj_map = bitmap_alloc(oo_objects(s->oo), GFP_KERNEL);
6184 if (!obj_map) {
6185 seq_release_private(inode, filep);
6186 return -ENOMEM;
6187 }
6188
6189 if (strcmp(filep->f_path.dentry->d_name.name, "alloc_traces") == 0)
6190 alloc = TRACK_ALLOC;
6191 else
6192 alloc = TRACK_FREE;
6193
6194 if (!alloc_loc_track(t, PAGE_SIZE / sizeof(struct location), GFP_KERNEL)) {
6195 bitmap_free(obj_map);
6196 seq_release_private(inode, filep);
6197 return -ENOMEM;
6198 }
6199
6200 for_each_kmem_cache_node(s, node, n) {
6201 unsigned long flags;
6202 struct slab *slab;
6203
6204 if (!atomic_long_read(&n->nr_slabs))
6205 continue;
6206
6207 spin_lock_irqsave(&n->list_lock, flags);
6208 list_for_each_entry(slab, &n->partial, slab_list)
6209 process_slab(t, s, slab, alloc, obj_map);
6210 list_for_each_entry(slab, &n->full, slab_list)
6211 process_slab(t, s, slab, alloc, obj_map);
6212 spin_unlock_irqrestore(&n->list_lock, flags);
6213 }
6214
6215 /* Sort locations by count */
6216 sort_r(t->loc, t->count, sizeof(struct location),
6217 cmp_loc_by_count, NULL, NULL);
6218
6219 bitmap_free(obj_map);
6220 return 0;
6221 }
6222
slab_debug_trace_release(struct inode * inode,struct file * file)6223 static int slab_debug_trace_release(struct inode *inode, struct file *file)
6224 {
6225 struct seq_file *seq = file->private_data;
6226 struct loc_track *t = seq->private;
6227
6228 free_loc_track(t);
6229 return seq_release_private(inode, file);
6230 }
6231
6232 static const struct file_operations slab_debugfs_fops = {
6233 .open = slab_debug_trace_open,
6234 .read = seq_read,
6235 .llseek = seq_lseek,
6236 .release = slab_debug_trace_release,
6237 };
6238
debugfs_slab_add(struct kmem_cache * s)6239 static void debugfs_slab_add(struct kmem_cache *s)
6240 {
6241 struct dentry *slab_cache_dir;
6242
6243 if (unlikely(!slab_debugfs_root))
6244 return;
6245
6246 slab_cache_dir = debugfs_create_dir(s->name, slab_debugfs_root);
6247
6248 debugfs_create_file("alloc_traces", 0400,
6249 slab_cache_dir, s, &slab_debugfs_fops);
6250
6251 debugfs_create_file("free_traces", 0400,
6252 slab_cache_dir, s, &slab_debugfs_fops);
6253 }
6254
debugfs_slab_release(struct kmem_cache * s)6255 void debugfs_slab_release(struct kmem_cache *s)
6256 {
6257 debugfs_remove_recursive(debugfs_lookup(s->name, slab_debugfs_root));
6258 }
6259
slab_debugfs_init(void)6260 static int __init slab_debugfs_init(void)
6261 {
6262 struct kmem_cache *s;
6263
6264 slab_debugfs_root = debugfs_create_dir("slab", NULL);
6265
6266 list_for_each_entry(s, &slab_caches, list)
6267 if (s->flags & SLAB_STORE_USER)
6268 debugfs_slab_add(s);
6269
6270 return 0;
6271
6272 }
6273 __initcall(slab_debugfs_init);
6274 #endif
6275 /*
6276 * The /proc/slabinfo ABI
6277 */
6278 #ifdef CONFIG_SLUB_DEBUG
get_slabinfo(struct kmem_cache * s,struct slabinfo * sinfo)6279 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
6280 {
6281 unsigned long nr_slabs = 0;
6282 unsigned long nr_objs = 0;
6283 unsigned long nr_free = 0;
6284 int node;
6285 struct kmem_cache_node *n;
6286
6287 for_each_kmem_cache_node(s, node, n) {
6288 nr_slabs += node_nr_slabs(n);
6289 nr_objs += node_nr_objs(n);
6290 nr_free += count_partial(n, count_free);
6291 }
6292
6293 sinfo->active_objs = nr_objs - nr_free;
6294 sinfo->num_objs = nr_objs;
6295 sinfo->active_slabs = nr_slabs;
6296 sinfo->num_slabs = nr_slabs;
6297 sinfo->objects_per_slab = oo_objects(s->oo);
6298 sinfo->cache_order = oo_order(s->oo);
6299 }
6300
slabinfo_show_stats(struct seq_file * m,struct kmem_cache * s)6301 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
6302 {
6303 }
6304
slabinfo_write(struct file * file,const char __user * buffer,size_t count,loff_t * ppos)6305 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
6306 size_t count, loff_t *ppos)
6307 {
6308 return -EIO;
6309 }
6310 #endif /* CONFIG_SLUB_DEBUG */
6311