1 | // SPDX-License-Identifier: GPL-2.0 |
2 | /* |
3 | * SLUB: A slab allocator that limits cache line use instead of queuing |
4 | * objects in per cpu and per node lists. |
5 | * |
6 | * The allocator synchronizes using per slab locks or atomic operations |
7 | * and only uses a centralized lock to manage a pool of partial slabs. |
8 | * |
9 | * (C) 2007 SGI, Christoph Lameter |
10 | * (C) 2011 Linux Foundation, Christoph Lameter |
11 | */ |
12 | |
13 | #include <linux/mm.h> |
14 | #include <linux/swap.h> /* mm_account_reclaimed_pages() */ |
15 | #include <linux/module.h> |
16 | #include <linux/bit_spinlock.h> |
17 | #include <linux/interrupt.h> |
18 | #include <linux/swab.h> |
19 | #include <linux/bitops.h> |
20 | #include <linux/slab.h> |
21 | #include "slab.h" |
22 | #include <linux/proc_fs.h> |
23 | #include <linux/seq_file.h> |
24 | #include <linux/kasan.h> |
25 | #include <linux/kmsan.h> |
26 | #include <linux/cpu.h> |
27 | #include <linux/cpuset.h> |
28 | #include <linux/mempolicy.h> |
29 | #include <linux/ctype.h> |
30 | #include <linux/stackdepot.h> |
31 | #include <linux/debugobjects.h> |
32 | #include <linux/kallsyms.h> |
33 | #include <linux/kfence.h> |
34 | #include <linux/memory.h> |
35 | #include <linux/math64.h> |
36 | #include <linux/fault-inject.h> |
37 | #include <linux/kmemleak.h> |
38 | #include <linux/stacktrace.h> |
39 | #include <linux/prefetch.h> |
40 | #include <linux/memcontrol.h> |
41 | #include <linux/random.h> |
42 | #include <kunit/test.h> |
43 | #include <kunit/test-bug.h> |
44 | #include <linux/sort.h> |
45 | |
46 | #include <linux/debugfs.h> |
47 | #include <trace/events/kmem.h> |
48 | |
49 | #include "internal.h" |
50 | |
51 | /* |
52 | * Lock order: |
53 | * 1. slab_mutex (Global Mutex) |
54 | * 2. node->list_lock (Spinlock) |
55 | * 3. kmem_cache->cpu_slab->lock (Local lock) |
56 | * 4. slab_lock(slab) (Only on some arches) |
57 | * 5. object_map_lock (Only for debugging) |
58 | * |
59 | * slab_mutex |
60 | * |
61 | * The role of the slab_mutex is to protect the list of all the slabs |
62 | * and to synchronize major metadata changes to slab cache structures. |
63 | * Also synchronizes memory hotplug callbacks. |
64 | * |
65 | * slab_lock |
66 | * |
67 | * The slab_lock is a wrapper around the page lock, thus it is a bit |
68 | * spinlock. |
69 | * |
70 | * The slab_lock is only used on arches that do not have the ability |
71 | * to do a cmpxchg_double. It only protects: |
72 | * |
73 | * A. slab->freelist -> List of free objects in a slab |
74 | * B. slab->inuse -> Number of objects in use |
75 | * C. slab->objects -> Number of objects in slab |
76 | * D. slab->frozen -> frozen state |
77 | * |
78 | * Frozen slabs |
79 | * |
80 | * If a slab is frozen then it is exempt from list management. It is |
81 | * the cpu slab which is actively allocated from by the processor that |
82 | * froze it and it is not on any list. The processor that froze the |
83 | * slab is the one who can perform list operations on the slab. Other |
84 | * processors may put objects onto the freelist but the processor that |
85 | * froze the slab is the only one that can retrieve the objects from the |
86 | * slab's freelist. |
87 | * |
88 | * CPU partial slabs |
89 | * |
90 | * The partially empty slabs cached on the CPU partial list are used |
91 | * for performance reasons, which speeds up the allocation process. |
92 | * These slabs are not frozen, but are also exempt from list management, |
93 | * by clearing the PG_workingset flag when moving out of the node |
94 | * partial list. Please see __slab_free() for more details. |
95 | * |
96 | * To sum up, the current scheme is: |
97 | * - node partial slab: PG_Workingset && !frozen |
98 | * - cpu partial slab: !PG_Workingset && !frozen |
99 | * - cpu slab: !PG_Workingset && frozen |
100 | * - full slab: !PG_Workingset && !frozen |
101 | * |
102 | * list_lock |
103 | * |
104 | * The list_lock protects the partial and full list on each node and |
105 | * the partial slab counter. If taken then no new slabs may be added or |
106 | * removed from the lists nor make the number of partial slabs be modified. |
107 | * (Note that the total number of slabs is an atomic value that may be |
108 | * modified without taking the list lock). |
109 | * |
110 | * The list_lock is a centralized lock and thus we avoid taking it as |
111 | * much as possible. As long as SLUB does not have to handle partial |
112 | * slabs, operations can continue without any centralized lock. F.e. |
113 | * allocating a long series of objects that fill up slabs does not require |
114 | * the list lock. |
115 | * |
116 | * For debug caches, all allocations are forced to go through a list_lock |
117 | * protected region to serialize against concurrent validation. |
118 | * |
119 | * cpu_slab->lock local lock |
120 | * |
121 | * This locks protect slowpath manipulation of all kmem_cache_cpu fields |
122 | * except the stat counters. This is a percpu structure manipulated only by |
123 | * the local cpu, so the lock protects against being preempted or interrupted |
124 | * by an irq. Fast path operations rely on lockless operations instead. |
125 | * |
126 | * On PREEMPT_RT, the local lock neither disables interrupts nor preemption |
127 | * which means the lockless fastpath cannot be used as it might interfere with |
128 | * an in-progress slow path operations. In this case the local lock is always |
129 | * taken but it still utilizes the freelist for the common operations. |
130 | * |
131 | * lockless fastpaths |
132 | * |
133 | * The fast path allocation (slab_alloc_node()) and freeing (do_slab_free()) |
134 | * are fully lockless when satisfied from the percpu slab (and when |
135 | * cmpxchg_double is possible to use, otherwise slab_lock is taken). |
136 | * They also don't disable preemption or migration or irqs. They rely on |
137 | * the transaction id (tid) field to detect being preempted or moved to |
138 | * another cpu. |
139 | * |
140 | * irq, preemption, migration considerations |
141 | * |
142 | * Interrupts are disabled as part of list_lock or local_lock operations, or |
143 | * around the slab_lock operation, in order to make the slab allocator safe |
144 | * to use in the context of an irq. |
145 | * |
146 | * In addition, preemption (or migration on PREEMPT_RT) is disabled in the |
147 | * allocation slowpath, bulk allocation, and put_cpu_partial(), so that the |
148 | * local cpu doesn't change in the process and e.g. the kmem_cache_cpu pointer |
149 | * doesn't have to be revalidated in each section protected by the local lock. |
150 | * |
151 | * SLUB assigns one slab for allocation to each processor. |
152 | * Allocations only occur from these slabs called cpu slabs. |
153 | * |
154 | * Slabs with free elements are kept on a partial list and during regular |
155 | * operations no list for full slabs is used. If an object in a full slab is |
156 | * freed then the slab will show up again on the partial lists. |
157 | * We track full slabs for debugging purposes though because otherwise we |
158 | * cannot scan all objects. |
159 | * |
160 | * Slabs are freed when they become empty. Teardown and setup is |
161 | * minimal so we rely on the page allocators per cpu caches for |
162 | * fast frees and allocs. |
163 | * |
164 | * slab->frozen The slab is frozen and exempt from list processing. |
165 | * This means that the slab is dedicated to a purpose |
166 | * such as satisfying allocations for a specific |
167 | * processor. Objects may be freed in the slab while |
168 | * it is frozen but slab_free will then skip the usual |
169 | * list operations. It is up to the processor holding |
170 | * the slab to integrate the slab into the slab lists |
171 | * when the slab is no longer needed. |
172 | * |
173 | * One use of this flag is to mark slabs that are |
174 | * used for allocations. Then such a slab becomes a cpu |
175 | * slab. The cpu slab may be equipped with an additional |
176 | * freelist that allows lockless access to |
177 | * free objects in addition to the regular freelist |
178 | * that requires the slab lock. |
179 | * |
180 | * SLAB_DEBUG_FLAGS Slab requires special handling due to debug |
181 | * options set. This moves slab handling out of |
182 | * the fast path and disables lockless freelists. |
183 | */ |
184 | |
185 | /* |
186 | * We could simply use migrate_disable()/enable() but as long as it's a |
187 | * function call even on !PREEMPT_RT, use inline preempt_disable() there. |
188 | */ |
189 | #ifndef CONFIG_PREEMPT_RT |
190 | #define slub_get_cpu_ptr(var) get_cpu_ptr(var) |
191 | #define slub_put_cpu_ptr(var) put_cpu_ptr(var) |
192 | #define USE_LOCKLESS_FAST_PATH() (true) |
193 | #else |
194 | #define slub_get_cpu_ptr(var) \ |
195 | ({ \ |
196 | migrate_disable(); \ |
197 | this_cpu_ptr(var); \ |
198 | }) |
199 | #define slub_put_cpu_ptr(var) \ |
200 | do { \ |
201 | (void)(var); \ |
202 | migrate_enable(); \ |
203 | } while (0) |
204 | #define USE_LOCKLESS_FAST_PATH() (false) |
205 | #endif |
206 | |
207 | #ifndef CONFIG_SLUB_TINY |
208 | #define __fastpath_inline __always_inline |
209 | #else |
210 | #define __fastpath_inline |
211 | #endif |
212 | |
213 | #ifdef CONFIG_SLUB_DEBUG |
214 | #ifdef CONFIG_SLUB_DEBUG_ON |
215 | DEFINE_STATIC_KEY_TRUE(slub_debug_enabled); |
216 | #else |
217 | DEFINE_STATIC_KEY_FALSE(slub_debug_enabled); |
218 | #endif |
219 | #endif /* CONFIG_SLUB_DEBUG */ |
220 | |
221 | /* Structure holding parameters for get_partial() call chain */ |
222 | struct partial_context { |
223 | gfp_t flags; |
224 | unsigned int orig_size; |
225 | void *object; |
226 | }; |
227 | |
228 | static inline bool kmem_cache_debug(struct kmem_cache *s) |
229 | { |
230 | return kmem_cache_debug_flags(s, SLAB_DEBUG_FLAGS); |
231 | } |
232 | |
233 | static inline bool slub_debug_orig_size(struct kmem_cache *s) |
234 | { |
235 | return (kmem_cache_debug_flags(s, SLAB_STORE_USER) && |
236 | (s->flags & SLAB_KMALLOC)); |
237 | } |
238 | |
239 | void *fixup_red_left(struct kmem_cache *s, void *p) |
240 | { |
241 | if (kmem_cache_debug_flags(s, SLAB_RED_ZONE)) |
242 | p += s->red_left_pad; |
243 | |
244 | return p; |
245 | } |
246 | |
247 | static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s) |
248 | { |
249 | #ifdef CONFIG_SLUB_CPU_PARTIAL |
250 | return !kmem_cache_debug(s); |
251 | #else |
252 | return false; |
253 | #endif |
254 | } |
255 | |
256 | /* |
257 | * Issues still to be resolved: |
258 | * |
259 | * - Support PAGE_ALLOC_DEBUG. Should be easy to do. |
260 | * |
261 | * - Variable sizing of the per node arrays |
262 | */ |
263 | |
264 | /* Enable to log cmpxchg failures */ |
265 | #undef SLUB_DEBUG_CMPXCHG |
266 | |
267 | #ifndef CONFIG_SLUB_TINY |
268 | /* |
269 | * Minimum number of partial slabs. These will be left on the partial |
270 | * lists even if they are empty. kmem_cache_shrink may reclaim them. |
271 | */ |
272 | #define MIN_PARTIAL 5 |
273 | |
274 | /* |
275 | * Maximum number of desirable partial slabs. |
276 | * The existence of more partial slabs makes kmem_cache_shrink |
277 | * sort the partial list by the number of objects in use. |
278 | */ |
279 | #define MAX_PARTIAL 10 |
280 | #else |
281 | #define MIN_PARTIAL 0 |
282 | #define MAX_PARTIAL 0 |
283 | #endif |
284 | |
285 | #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \ |
286 | SLAB_POISON | SLAB_STORE_USER) |
287 | |
288 | /* |
289 | * These debug flags cannot use CMPXCHG because there might be consistency |
290 | * issues when checking or reading debug information |
291 | */ |
292 | #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \ |
293 | SLAB_TRACE) |
294 | |
295 | |
296 | /* |
297 | * Debugging flags that require metadata to be stored in the slab. These get |
298 | * disabled when slab_debug=O is used and a cache's min order increases with |
299 | * metadata. |
300 | */ |
301 | #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER) |
302 | |
303 | #define OO_SHIFT 16 |
304 | #define OO_MASK ((1 << OO_SHIFT) - 1) |
305 | #define MAX_OBJS_PER_PAGE 32767 /* since slab.objects is u15 */ |
306 | |
307 | /* Internal SLUB flags */ |
308 | /* Poison object */ |
309 | #define __OBJECT_POISON __SLAB_FLAG_BIT(_SLAB_OBJECT_POISON) |
310 | /* Use cmpxchg_double */ |
311 | |
312 | #ifdef system_has_freelist_aba |
313 | #define __CMPXCHG_DOUBLE __SLAB_FLAG_BIT(_SLAB_CMPXCHG_DOUBLE) |
314 | #else |
315 | #define __CMPXCHG_DOUBLE __SLAB_FLAG_UNUSED |
316 | #endif |
317 | |
318 | /* |
319 | * Tracking user of a slab. |
320 | */ |
321 | #define TRACK_ADDRS_COUNT 16 |
322 | struct track { |
323 | unsigned long addr; /* Called from address */ |
324 | #ifdef CONFIG_STACKDEPOT |
325 | depot_stack_handle_t handle; |
326 | #endif |
327 | int cpu; /* Was running on cpu */ |
328 | int pid; /* Pid context */ |
329 | unsigned long when; /* When did the operation occur */ |
330 | }; |
331 | |
332 | enum track_item { TRACK_ALLOC, TRACK_FREE }; |
333 | |
334 | #ifdef SLAB_SUPPORTS_SYSFS |
335 | static int sysfs_slab_add(struct kmem_cache *); |
336 | static int sysfs_slab_alias(struct kmem_cache *, const char *); |
337 | #else |
338 | static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; } |
339 | static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p) |
340 | { return 0; } |
341 | #endif |
342 | |
343 | #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_SLUB_DEBUG) |
344 | static void debugfs_slab_add(struct kmem_cache *); |
345 | #else |
346 | static inline void debugfs_slab_add(struct kmem_cache *s) { } |
347 | #endif |
348 | |
349 | enum stat_item { |
350 | ALLOC_FASTPATH, /* Allocation from cpu slab */ |
351 | ALLOC_SLOWPATH, /* Allocation by getting a new cpu slab */ |
352 | FREE_FASTPATH, /* Free to cpu slab */ |
353 | FREE_SLOWPATH, /* Freeing not to cpu slab */ |
354 | FREE_FROZEN, /* Freeing to frozen slab */ |
355 | FREE_ADD_PARTIAL, /* Freeing moves slab to partial list */ |
356 | FREE_REMOVE_PARTIAL, /* Freeing removes last object */ |
357 | ALLOC_FROM_PARTIAL, /* Cpu slab acquired from node partial list */ |
358 | ALLOC_SLAB, /* Cpu slab acquired from page allocator */ |
359 | ALLOC_REFILL, /* Refill cpu slab from slab freelist */ |
360 | ALLOC_NODE_MISMATCH, /* Switching cpu slab */ |
361 | FREE_SLAB, /* Slab freed to the page allocator */ |
362 | CPUSLAB_FLUSH, /* Abandoning of the cpu slab */ |
363 | DEACTIVATE_FULL, /* Cpu slab was full when deactivated */ |
364 | DEACTIVATE_EMPTY, /* Cpu slab was empty when deactivated */ |
365 | DEACTIVATE_TO_HEAD, /* Cpu slab was moved to the head of partials */ |
366 | DEACTIVATE_TO_TAIL, /* Cpu slab was moved to the tail of partials */ |
367 | DEACTIVATE_REMOTE_FREES,/* Slab contained remotely freed objects */ |
368 | DEACTIVATE_BYPASS, /* Implicit deactivation */ |
369 | ORDER_FALLBACK, /* Number of times fallback was necessary */ |
370 | CMPXCHG_DOUBLE_CPU_FAIL,/* Failures of this_cpu_cmpxchg_double */ |
371 | CMPXCHG_DOUBLE_FAIL, /* Failures of slab freelist update */ |
372 | CPU_PARTIAL_ALLOC, /* Used cpu partial on alloc */ |
373 | CPU_PARTIAL_FREE, /* Refill cpu partial on free */ |
374 | CPU_PARTIAL_NODE, /* Refill cpu partial from node partial */ |
375 | CPU_PARTIAL_DRAIN, /* Drain cpu partial to node partial */ |
376 | NR_SLUB_STAT_ITEMS |
377 | }; |
378 | |
379 | #ifndef CONFIG_SLUB_TINY |
380 | /* |
381 | * When changing the layout, make sure freelist and tid are still compatible |
382 | * with this_cpu_cmpxchg_double() alignment requirements. |
383 | */ |
384 | struct kmem_cache_cpu { |
385 | union { |
386 | struct { |
387 | void **freelist; /* Pointer to next available object */ |
388 | unsigned long tid; /* Globally unique transaction id */ |
389 | }; |
390 | freelist_aba_t freelist_tid; |
391 | }; |
392 | struct slab *slab; /* The slab from which we are allocating */ |
393 | #ifdef CONFIG_SLUB_CPU_PARTIAL |
394 | struct slab *partial; /* Partially allocated slabs */ |
395 | #endif |
396 | local_lock_t lock; /* Protects the fields above */ |
397 | #ifdef CONFIG_SLUB_STATS |
398 | unsigned int stat[NR_SLUB_STAT_ITEMS]; |
399 | #endif |
400 | }; |
401 | #endif /* CONFIG_SLUB_TINY */ |
402 | |
403 | static inline void stat(const struct kmem_cache *s, enum stat_item si) |
404 | { |
405 | #ifdef CONFIG_SLUB_STATS |
406 | /* |
407 | * The rmw is racy on a preemptible kernel but this is acceptable, so |
408 | * avoid this_cpu_add()'s irq-disable overhead. |
409 | */ |
410 | raw_cpu_inc(s->cpu_slab->stat[si]); |
411 | #endif |
412 | } |
413 | |
414 | static inline |
415 | void stat_add(const struct kmem_cache *s, enum stat_item si, int v) |
416 | { |
417 | #ifdef CONFIG_SLUB_STATS |
418 | raw_cpu_add(s->cpu_slab->stat[si], v); |
419 | #endif |
420 | } |
421 | |
422 | /* |
423 | * The slab lists for all objects. |
424 | */ |
425 | struct kmem_cache_node { |
426 | spinlock_t list_lock; |
427 | unsigned long nr_partial; |
428 | struct list_head partial; |
429 | #ifdef CONFIG_SLUB_DEBUG |
430 | atomic_long_t nr_slabs; |
431 | atomic_long_t total_objects; |
432 | struct list_head full; |
433 | #endif |
434 | }; |
435 | |
436 | static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node) |
437 | { |
438 | return s->node[node]; |
439 | } |
440 | |
441 | /* |
442 | * Iterator over all nodes. The body will be executed for each node that has |
443 | * a kmem_cache_node structure allocated (which is true for all online nodes) |
444 | */ |
445 | #define for_each_kmem_cache_node(__s, __node, __n) \ |
446 | for (__node = 0; __node < nr_node_ids; __node++) \ |
447 | if ((__n = get_node(__s, __node))) |
448 | |
449 | /* |
450 | * Tracks for which NUMA nodes we have kmem_cache_nodes allocated. |
451 | * Corresponds to node_state[N_NORMAL_MEMORY], but can temporarily |
452 | * differ during memory hotplug/hotremove operations. |
453 | * Protected by slab_mutex. |
454 | */ |
455 | static nodemask_t slab_nodes; |
456 | |
457 | #ifndef CONFIG_SLUB_TINY |
458 | /* |
459 | * Workqueue used for flush_cpu_slab(). |
460 | */ |
461 | static struct workqueue_struct *flushwq; |
462 | #endif |
463 | |
464 | /******************************************************************** |
465 | * Core slab cache functions |
466 | *******************************************************************/ |
467 | |
468 | /* |
469 | * freeptr_t represents a SLUB freelist pointer, which might be encoded |
470 | * and not dereferenceable if CONFIG_SLAB_FREELIST_HARDENED is enabled. |
471 | */ |
472 | typedef struct { unsigned long v; } freeptr_t; |
473 | |
474 | /* |
475 | * Returns freelist pointer (ptr). With hardening, this is obfuscated |
476 | * with an XOR of the address where the pointer is held and a per-cache |
477 | * random number. |
478 | */ |
479 | static inline freeptr_t freelist_ptr_encode(const struct kmem_cache *s, |
480 | void *ptr, unsigned long ptr_addr) |
481 | { |
482 | unsigned long encoded; |
483 | |
484 | #ifdef CONFIG_SLAB_FREELIST_HARDENED |
485 | encoded = (unsigned long)ptr ^ s->random ^ swab(ptr_addr); |
486 | #else |
487 | encoded = (unsigned long)ptr; |
488 | #endif |
489 | return (freeptr_t){.v = encoded}; |
490 | } |
491 | |
492 | static inline void *freelist_ptr_decode(const struct kmem_cache *s, |
493 | freeptr_t ptr, unsigned long ptr_addr) |
494 | { |
495 | void *decoded; |
496 | |
497 | #ifdef CONFIG_SLAB_FREELIST_HARDENED |
498 | decoded = (void *)(ptr.v ^ s->random ^ swab(ptr_addr)); |
499 | #else |
500 | decoded = (void *)ptr.v; |
501 | #endif |
502 | return decoded; |
503 | } |
504 | |
505 | static inline void *get_freepointer(struct kmem_cache *s, void *object) |
506 | { |
507 | unsigned long ptr_addr; |
508 | freeptr_t p; |
509 | |
510 | object = kasan_reset_tag(addr: object); |
511 | ptr_addr = (unsigned long)object + s->offset; |
512 | p = *(freeptr_t *)(ptr_addr); |
513 | return freelist_ptr_decode(s, ptr: p, ptr_addr); |
514 | } |
515 | |
516 | #ifndef CONFIG_SLUB_TINY |
517 | static void prefetch_freepointer(const struct kmem_cache *s, void *object) |
518 | { |
519 | prefetchw(object + s->offset); |
520 | } |
521 | #endif |
522 | |
523 | /* |
524 | * When running under KMSAN, get_freepointer_safe() may return an uninitialized |
525 | * pointer value in the case the current thread loses the race for the next |
526 | * memory chunk in the freelist. In that case this_cpu_cmpxchg_double() in |
527 | * slab_alloc_node() will fail, so the uninitialized value won't be used, but |
528 | * KMSAN will still check all arguments of cmpxchg because of imperfect |
529 | * handling of inline assembly. |
530 | * To work around this problem, we apply __no_kmsan_checks to ensure that |
531 | * get_freepointer_safe() returns initialized memory. |
532 | */ |
533 | __no_kmsan_checks |
534 | static inline void *get_freepointer_safe(struct kmem_cache *s, void *object) |
535 | { |
536 | unsigned long freepointer_addr; |
537 | freeptr_t p; |
538 | |
539 | if (!debug_pagealloc_enabled_static()) |
540 | return get_freepointer(s, object); |
541 | |
542 | object = kasan_reset_tag(addr: object); |
543 | freepointer_addr = (unsigned long)object + s->offset; |
544 | copy_from_kernel_nofault(dst: &p, src: (freeptr_t *)freepointer_addr, size: sizeof(p)); |
545 | return freelist_ptr_decode(s, ptr: p, ptr_addr: freepointer_addr); |
546 | } |
547 | |
548 | static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp) |
549 | { |
550 | unsigned long freeptr_addr = (unsigned long)object + s->offset; |
551 | |
552 | #ifdef CONFIG_SLAB_FREELIST_HARDENED |
553 | BUG_ON(object == fp); /* naive detection of double free or corruption */ |
554 | #endif |
555 | |
556 | freeptr_addr = (unsigned long)kasan_reset_tag(addr: (void *)freeptr_addr); |
557 | *(freeptr_t *)freeptr_addr = freelist_ptr_encode(s, ptr: fp, ptr_addr: freeptr_addr); |
558 | } |
559 | |
560 | /* Loop over all objects in a slab */ |
561 | #define for_each_object(__p, __s, __addr, __objects) \ |
562 | for (__p = fixup_red_left(__s, __addr); \ |
563 | __p < (__addr) + (__objects) * (__s)->size; \ |
564 | __p += (__s)->size) |
565 | |
566 | static inline unsigned int order_objects(unsigned int order, unsigned int size) |
567 | { |
568 | return ((unsigned int)PAGE_SIZE << order) / size; |
569 | } |
570 | |
571 | static inline struct kmem_cache_order_objects oo_make(unsigned int order, |
572 | unsigned int size) |
573 | { |
574 | struct kmem_cache_order_objects x = { |
575 | (order << OO_SHIFT) + order_objects(order, size) |
576 | }; |
577 | |
578 | return x; |
579 | } |
580 | |
581 | static inline unsigned int oo_order(struct kmem_cache_order_objects x) |
582 | { |
583 | return x.x >> OO_SHIFT; |
584 | } |
585 | |
586 | static inline unsigned int oo_objects(struct kmem_cache_order_objects x) |
587 | { |
588 | return x.x & OO_MASK; |
589 | } |
590 | |
591 | #ifdef CONFIG_SLUB_CPU_PARTIAL |
592 | static void slub_set_cpu_partial(struct kmem_cache *s, unsigned int nr_objects) |
593 | { |
594 | unsigned int nr_slabs; |
595 | |
596 | s->cpu_partial = nr_objects; |
597 | |
598 | /* |
599 | * We take the number of objects but actually limit the number of |
600 | * slabs on the per cpu partial list, in order to limit excessive |
601 | * growth of the list. For simplicity we assume that the slabs will |
602 | * be half-full. |
603 | */ |
604 | nr_slabs = DIV_ROUND_UP(nr_objects * 2, oo_objects(s->oo)); |
605 | s->cpu_partial_slabs = nr_slabs; |
606 | } |
607 | #else |
608 | static inline void |
609 | slub_set_cpu_partial(struct kmem_cache *s, unsigned int nr_objects) |
610 | { |
611 | } |
612 | #endif /* CONFIG_SLUB_CPU_PARTIAL */ |
613 | |
614 | /* |
615 | * Per slab locking using the pagelock |
616 | */ |
617 | static __always_inline void slab_lock(struct slab *slab) |
618 | { |
619 | struct page *page = slab_page(slab); |
620 | |
621 | VM_BUG_ON_PAGE(PageTail(page), page); |
622 | bit_spin_lock(bitnum: PG_locked, addr: &page->flags); |
623 | } |
624 | |
625 | static __always_inline void slab_unlock(struct slab *slab) |
626 | { |
627 | struct page *page = slab_page(slab); |
628 | |
629 | VM_BUG_ON_PAGE(PageTail(page), page); |
630 | bit_spin_unlock(bitnum: PG_locked, addr: &page->flags); |
631 | } |
632 | |
633 | static inline bool |
634 | __update_freelist_fast(struct slab *slab, |
635 | void *freelist_old, unsigned long counters_old, |
636 | void *freelist_new, unsigned long counters_new) |
637 | { |
638 | #ifdef system_has_freelist_aba |
639 | freelist_aba_t old = { .freelist = freelist_old, .counter = counters_old }; |
640 | freelist_aba_t new = { .freelist = freelist_new, .counter = counters_new }; |
641 | |
642 | return try_cmpxchg_freelist(&slab->freelist_counter.full, &old.full, new.full); |
643 | #else |
644 | return false; |
645 | #endif |
646 | } |
647 | |
648 | static inline bool |
649 | __update_freelist_slow(struct slab *slab, |
650 | void *freelist_old, unsigned long counters_old, |
651 | void *freelist_new, unsigned long counters_new) |
652 | { |
653 | bool ret = false; |
654 | |
655 | slab_lock(slab); |
656 | if (slab->freelist == freelist_old && |
657 | slab->counters == counters_old) { |
658 | slab->freelist = freelist_new; |
659 | slab->counters = counters_new; |
660 | ret = true; |
661 | } |
662 | slab_unlock(slab); |
663 | |
664 | return ret; |
665 | } |
666 | |
667 | /* |
668 | * Interrupts must be disabled (for the fallback code to work right), typically |
669 | * by an _irqsave() lock variant. On PREEMPT_RT the preempt_disable(), which is |
670 | * part of bit_spin_lock(), is sufficient because the policy is not to allow any |
671 | * allocation/ free operation in hardirq context. Therefore nothing can |
672 | * interrupt the operation. |
673 | */ |
674 | static inline bool __slab_update_freelist(struct kmem_cache *s, struct slab *slab, |
675 | void *freelist_old, unsigned long counters_old, |
676 | void *freelist_new, unsigned long counters_new, |
677 | const char *n) |
678 | { |
679 | bool ret; |
680 | |
681 | if (USE_LOCKLESS_FAST_PATH()) |
682 | lockdep_assert_irqs_disabled(); |
683 | |
684 | if (s->flags & __CMPXCHG_DOUBLE) { |
685 | ret = __update_freelist_fast(slab, freelist_old, counters_old, |
686 | freelist_new, counters_new); |
687 | } else { |
688 | ret = __update_freelist_slow(slab, freelist_old, counters_old, |
689 | freelist_new, counters_new); |
690 | } |
691 | if (likely(ret)) |
692 | return true; |
693 | |
694 | cpu_relax(); |
695 | stat(s, si: CMPXCHG_DOUBLE_FAIL); |
696 | |
697 | #ifdef SLUB_DEBUG_CMPXCHG |
698 | pr_info("%s %s: cmpxchg double redo " , n, s->name); |
699 | #endif |
700 | |
701 | return false; |
702 | } |
703 | |
704 | static inline bool slab_update_freelist(struct kmem_cache *s, struct slab *slab, |
705 | void *freelist_old, unsigned long counters_old, |
706 | void *freelist_new, unsigned long counters_new, |
707 | const char *n) |
708 | { |
709 | bool ret; |
710 | |
711 | if (s->flags & __CMPXCHG_DOUBLE) { |
712 | ret = __update_freelist_fast(slab, freelist_old, counters_old, |
713 | freelist_new, counters_new); |
714 | } else { |
715 | unsigned long flags; |
716 | |
717 | local_irq_save(flags); |
718 | ret = __update_freelist_slow(slab, freelist_old, counters_old, |
719 | freelist_new, counters_new); |
720 | local_irq_restore(flags); |
721 | } |
722 | if (likely(ret)) |
723 | return true; |
724 | |
725 | cpu_relax(); |
726 | stat(s, si: CMPXCHG_DOUBLE_FAIL); |
727 | |
728 | #ifdef SLUB_DEBUG_CMPXCHG |
729 | pr_info("%s %s: cmpxchg double redo " , n, s->name); |
730 | #endif |
731 | |
732 | return false; |
733 | } |
734 | |
735 | #ifdef CONFIG_SLUB_DEBUG |
736 | static unsigned long object_map[BITS_TO_LONGS(MAX_OBJS_PER_PAGE)]; |
737 | static DEFINE_SPINLOCK(object_map_lock); |
738 | |
739 | static void __fill_map(unsigned long *obj_map, struct kmem_cache *s, |
740 | struct slab *slab) |
741 | { |
742 | void *addr = slab_address(slab); |
743 | void *p; |
744 | |
745 | bitmap_zero(obj_map, slab->objects); |
746 | |
747 | for (p = slab->freelist; p; p = get_freepointer(s, p)) |
748 | set_bit(__obj_to_index(s, addr, p), obj_map); |
749 | } |
750 | |
751 | #if IS_ENABLED(CONFIG_KUNIT) |
752 | static bool slab_add_kunit_errors(void) |
753 | { |
754 | struct kunit_resource *resource; |
755 | |
756 | if (!kunit_get_current_test()) |
757 | return false; |
758 | |
759 | resource = kunit_find_named_resource(current->kunit_test, "slab_errors" ); |
760 | if (!resource) |
761 | return false; |
762 | |
763 | (*(int *)resource->data)++; |
764 | kunit_put_resource(resource); |
765 | return true; |
766 | } |
767 | #else |
768 | static inline bool slab_add_kunit_errors(void) { return false; } |
769 | #endif |
770 | |
771 | static inline unsigned int size_from_object(struct kmem_cache *s) |
772 | { |
773 | if (s->flags & SLAB_RED_ZONE) |
774 | return s->size - s->red_left_pad; |
775 | |
776 | return s->size; |
777 | } |
778 | |
779 | static inline void *restore_red_left(struct kmem_cache *s, void *p) |
780 | { |
781 | if (s->flags & SLAB_RED_ZONE) |
782 | p -= s->red_left_pad; |
783 | |
784 | return p; |
785 | } |
786 | |
787 | /* |
788 | * Debug settings: |
789 | */ |
790 | #if defined(CONFIG_SLUB_DEBUG_ON) |
791 | static slab_flags_t slub_debug = DEBUG_DEFAULT_FLAGS; |
792 | #else |
793 | static slab_flags_t slub_debug; |
794 | #endif |
795 | |
796 | static char *slub_debug_string; |
797 | static int disable_higher_order_debug; |
798 | |
799 | /* |
800 | * slub is about to manipulate internal object metadata. This memory lies |
801 | * outside the range of the allocated object, so accessing it would normally |
802 | * be reported by kasan as a bounds error. metadata_access_enable() is used |
803 | * to tell kasan that these accesses are OK. |
804 | */ |
805 | static inline void metadata_access_enable(void) |
806 | { |
807 | kasan_disable_current(); |
808 | } |
809 | |
810 | static inline void metadata_access_disable(void) |
811 | { |
812 | kasan_enable_current(); |
813 | } |
814 | |
815 | /* |
816 | * Object debugging |
817 | */ |
818 | |
819 | /* Verify that a pointer has an address that is valid within a slab page */ |
820 | static inline int check_valid_pointer(struct kmem_cache *s, |
821 | struct slab *slab, void *object) |
822 | { |
823 | void *base; |
824 | |
825 | if (!object) |
826 | return 1; |
827 | |
828 | base = slab_address(slab); |
829 | object = kasan_reset_tag(object); |
830 | object = restore_red_left(s, object); |
831 | if (object < base || object >= base + slab->objects * s->size || |
832 | (object - base) % s->size) { |
833 | return 0; |
834 | } |
835 | |
836 | return 1; |
837 | } |
838 | |
839 | static void print_section(char *level, char *text, u8 *addr, |
840 | unsigned int length) |
841 | { |
842 | metadata_access_enable(); |
843 | print_hex_dump(level, text, DUMP_PREFIX_ADDRESS, |
844 | 16, 1, kasan_reset_tag((void *)addr), length, 1); |
845 | metadata_access_disable(); |
846 | } |
847 | |
848 | /* |
849 | * See comment in calculate_sizes(). |
850 | */ |
851 | static inline bool freeptr_outside_object(struct kmem_cache *s) |
852 | { |
853 | return s->offset >= s->inuse; |
854 | } |
855 | |
856 | /* |
857 | * Return offset of the end of info block which is inuse + free pointer if |
858 | * not overlapping with object. |
859 | */ |
860 | static inline unsigned int get_info_end(struct kmem_cache *s) |
861 | { |
862 | if (freeptr_outside_object(s)) |
863 | return s->inuse + sizeof(void *); |
864 | else |
865 | return s->inuse; |
866 | } |
867 | |
868 | static struct track *get_track(struct kmem_cache *s, void *object, |
869 | enum track_item alloc) |
870 | { |
871 | struct track *p; |
872 | |
873 | p = object + get_info_end(s); |
874 | |
875 | return kasan_reset_tag(p + alloc); |
876 | } |
877 | |
878 | #ifdef CONFIG_STACKDEPOT |
879 | static noinline depot_stack_handle_t set_track_prepare(void) |
880 | { |
881 | depot_stack_handle_t handle; |
882 | unsigned long entries[TRACK_ADDRS_COUNT]; |
883 | unsigned int nr_entries; |
884 | |
885 | nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 3); |
886 | handle = stack_depot_save(entries, nr_entries, GFP_NOWAIT); |
887 | |
888 | return handle; |
889 | } |
890 | #else |
891 | static inline depot_stack_handle_t set_track_prepare(void) |
892 | { |
893 | return 0; |
894 | } |
895 | #endif |
896 | |
897 | static void set_track_update(struct kmem_cache *s, void *object, |
898 | enum track_item alloc, unsigned long addr, |
899 | depot_stack_handle_t handle) |
900 | { |
901 | struct track *p = get_track(s, object, alloc); |
902 | |
903 | #ifdef CONFIG_STACKDEPOT |
904 | p->handle = handle; |
905 | #endif |
906 | p->addr = addr; |
907 | p->cpu = smp_processor_id(); |
908 | p->pid = current->pid; |
909 | p->when = jiffies; |
910 | } |
911 | |
912 | static __always_inline void set_track(struct kmem_cache *s, void *object, |
913 | enum track_item alloc, unsigned long addr) |
914 | { |
915 | depot_stack_handle_t handle = set_track_prepare(); |
916 | |
917 | set_track_update(s, object, alloc, addr, handle); |
918 | } |
919 | |
920 | static void init_tracking(struct kmem_cache *s, void *object) |
921 | { |
922 | struct track *p; |
923 | |
924 | if (!(s->flags & SLAB_STORE_USER)) |
925 | return; |
926 | |
927 | p = get_track(s, object, TRACK_ALLOC); |
928 | memset(p, 0, 2*sizeof(struct track)); |
929 | } |
930 | |
931 | static void print_track(const char *s, struct track *t, unsigned long pr_time) |
932 | { |
933 | depot_stack_handle_t handle __maybe_unused; |
934 | |
935 | if (!t->addr) |
936 | return; |
937 | |
938 | pr_err("%s in %pS age=%lu cpu=%u pid=%d\n" , |
939 | s, (void *)t->addr, pr_time - t->when, t->cpu, t->pid); |
940 | #ifdef CONFIG_STACKDEPOT |
941 | handle = READ_ONCE(t->handle); |
942 | if (handle) |
943 | stack_depot_print(handle); |
944 | else |
945 | pr_err("object allocation/free stack trace missing\n" ); |
946 | #endif |
947 | } |
948 | |
949 | void print_tracking(struct kmem_cache *s, void *object) |
950 | { |
951 | unsigned long pr_time = jiffies; |
952 | if (!(s->flags & SLAB_STORE_USER)) |
953 | return; |
954 | |
955 | print_track("Allocated" , get_track(s, object, TRACK_ALLOC), pr_time); |
956 | print_track("Freed" , get_track(s, object, TRACK_FREE), pr_time); |
957 | } |
958 | |
959 | static void print_slab_info(const struct slab *slab) |
960 | { |
961 | struct folio *folio = (struct folio *)slab_folio(slab); |
962 | |
963 | pr_err("Slab 0x%p objects=%u used=%u fp=0x%p flags=%pGp\n" , |
964 | slab, slab->objects, slab->inuse, slab->freelist, |
965 | folio_flags(folio, 0)); |
966 | } |
967 | |
968 | /* |
969 | * kmalloc caches has fixed sizes (mostly power of 2), and kmalloc() API |
970 | * family will round up the real request size to these fixed ones, so |
971 | * there could be an extra area than what is requested. Save the original |
972 | * request size in the meta data area, for better debug and sanity check. |
973 | */ |
974 | static inline void set_orig_size(struct kmem_cache *s, |
975 | void *object, unsigned int orig_size) |
976 | { |
977 | void *p = kasan_reset_tag(object); |
978 | unsigned int kasan_meta_size; |
979 | |
980 | if (!slub_debug_orig_size(s)) |
981 | return; |
982 | |
983 | /* |
984 | * KASAN can save its free meta data inside of the object at offset 0. |
985 | * If this meta data size is larger than 'orig_size', it will overlap |
986 | * the data redzone in [orig_size+1, object_size]. Thus, we adjust |
987 | * 'orig_size' to be as at least as big as KASAN's meta data. |
988 | */ |
989 | kasan_meta_size = kasan_metadata_size(s, true); |
990 | if (kasan_meta_size > orig_size) |
991 | orig_size = kasan_meta_size; |
992 | |
993 | p += get_info_end(s); |
994 | p += sizeof(struct track) * 2; |
995 | |
996 | *(unsigned int *)p = orig_size; |
997 | } |
998 | |
999 | static inline unsigned int get_orig_size(struct kmem_cache *s, void *object) |
1000 | { |
1001 | void *p = kasan_reset_tag(object); |
1002 | |
1003 | if (!slub_debug_orig_size(s)) |
1004 | return s->object_size; |
1005 | |
1006 | p += get_info_end(s); |
1007 | p += sizeof(struct track) * 2; |
1008 | |
1009 | return *(unsigned int *)p; |
1010 | } |
1011 | |
1012 | void skip_orig_size_check(struct kmem_cache *s, const void *object) |
1013 | { |
1014 | set_orig_size(s, (void *)object, s->object_size); |
1015 | } |
1016 | |
1017 | static void slab_bug(struct kmem_cache *s, char *fmt, ...) |
1018 | { |
1019 | struct va_format vaf; |
1020 | va_list args; |
1021 | |
1022 | va_start(args, fmt); |
1023 | vaf.fmt = fmt; |
1024 | vaf.va = &args; |
1025 | pr_err("=============================================================================\n" ); |
1026 | pr_err("BUG %s (%s): %pV\n" , s->name, print_tainted(), &vaf); |
1027 | pr_err("-----------------------------------------------------------------------------\n\n" ); |
1028 | va_end(args); |
1029 | } |
1030 | |
1031 | __printf(2, 3) |
1032 | static void slab_fix(struct kmem_cache *s, char *fmt, ...) |
1033 | { |
1034 | struct va_format vaf; |
1035 | va_list args; |
1036 | |
1037 | if (slab_add_kunit_errors()) |
1038 | return; |
1039 | |
1040 | va_start(args, fmt); |
1041 | vaf.fmt = fmt; |
1042 | vaf.va = &args; |
1043 | pr_err("FIX %s: %pV\n" , s->name, &vaf); |
1044 | va_end(args); |
1045 | } |
1046 | |
1047 | static void print_trailer(struct kmem_cache *s, struct slab *slab, u8 *p) |
1048 | { |
1049 | unsigned int off; /* Offset of last byte */ |
1050 | u8 *addr = slab_address(slab); |
1051 | |
1052 | print_tracking(s, p); |
1053 | |
1054 | print_slab_info(slab); |
1055 | |
1056 | pr_err("Object 0x%p @offset=%tu fp=0x%p\n\n" , |
1057 | p, p - addr, get_freepointer(s, p)); |
1058 | |
1059 | if (s->flags & SLAB_RED_ZONE) |
1060 | print_section(KERN_ERR, "Redzone " , p - s->red_left_pad, |
1061 | s->red_left_pad); |
1062 | else if (p > addr + 16) |
1063 | print_section(KERN_ERR, "Bytes b4 " , p - 16, 16); |
1064 | |
1065 | print_section(KERN_ERR, "Object " , p, |
1066 | min_t(unsigned int, s->object_size, PAGE_SIZE)); |
1067 | if (s->flags & SLAB_RED_ZONE) |
1068 | print_section(KERN_ERR, "Redzone " , p + s->object_size, |
1069 | s->inuse - s->object_size); |
1070 | |
1071 | off = get_info_end(s); |
1072 | |
1073 | if (s->flags & SLAB_STORE_USER) |
1074 | off += 2 * sizeof(struct track); |
1075 | |
1076 | if (slub_debug_orig_size(s)) |
1077 | off += sizeof(unsigned int); |
1078 | |
1079 | off += kasan_metadata_size(s, false); |
1080 | |
1081 | if (off != size_from_object(s)) |
1082 | /* Beginning of the filler is the free pointer */ |
1083 | print_section(KERN_ERR, "Padding " , p + off, |
1084 | size_from_object(s) - off); |
1085 | |
1086 | dump_stack(); |
1087 | } |
1088 | |
1089 | static void object_err(struct kmem_cache *s, struct slab *slab, |
1090 | u8 *object, char *reason) |
1091 | { |
1092 | if (slab_add_kunit_errors()) |
1093 | return; |
1094 | |
1095 | slab_bug(s, "%s" , reason); |
1096 | print_trailer(s, slab, object); |
1097 | add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE); |
1098 | } |
1099 | |
1100 | static bool freelist_corrupted(struct kmem_cache *s, struct slab *slab, |
1101 | void **freelist, void *nextfree) |
1102 | { |
1103 | if ((s->flags & SLAB_CONSISTENCY_CHECKS) && |
1104 | !check_valid_pointer(s, slab, nextfree) && freelist) { |
1105 | object_err(s, slab, *freelist, "Freechain corrupt" ); |
1106 | *freelist = NULL; |
1107 | slab_fix(s, "Isolate corrupted freechain" ); |
1108 | return true; |
1109 | } |
1110 | |
1111 | return false; |
1112 | } |
1113 | |
1114 | static __printf(3, 4) void slab_err(struct kmem_cache *s, struct slab *slab, |
1115 | const char *fmt, ...) |
1116 | { |
1117 | va_list args; |
1118 | char buf[100]; |
1119 | |
1120 | if (slab_add_kunit_errors()) |
1121 | return; |
1122 | |
1123 | va_start(args, fmt); |
1124 | vsnprintf(buf, sizeof(buf), fmt, args); |
1125 | va_end(args); |
1126 | slab_bug(s, "%s" , buf); |
1127 | print_slab_info(slab); |
1128 | dump_stack(); |
1129 | add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE); |
1130 | } |
1131 | |
1132 | static void init_object(struct kmem_cache *s, void *object, u8 val) |
1133 | { |
1134 | u8 *p = kasan_reset_tag(object); |
1135 | unsigned int poison_size = s->object_size; |
1136 | |
1137 | if (s->flags & SLAB_RED_ZONE) { |
1138 | memset(p - s->red_left_pad, val, s->red_left_pad); |
1139 | |
1140 | if (slub_debug_orig_size(s) && val == SLUB_RED_ACTIVE) { |
1141 | /* |
1142 | * Redzone the extra allocated space by kmalloc than |
1143 | * requested, and the poison size will be limited to |
1144 | * the original request size accordingly. |
1145 | */ |
1146 | poison_size = get_orig_size(s, object); |
1147 | } |
1148 | } |
1149 | |
1150 | if (s->flags & __OBJECT_POISON) { |
1151 | memset(p, POISON_FREE, poison_size - 1); |
1152 | p[poison_size - 1] = POISON_END; |
1153 | } |
1154 | |
1155 | if (s->flags & SLAB_RED_ZONE) |
1156 | memset(p + poison_size, val, s->inuse - poison_size); |
1157 | } |
1158 | |
1159 | static void restore_bytes(struct kmem_cache *s, char *message, u8 data, |
1160 | void *from, void *to) |
1161 | { |
1162 | slab_fix(s, "Restoring %s 0x%p-0x%p=0x%x" , message, from, to - 1, data); |
1163 | memset(from, data, to - from); |
1164 | } |
1165 | |
1166 | static int check_bytes_and_report(struct kmem_cache *s, struct slab *slab, |
1167 | u8 *object, char *what, |
1168 | u8 *start, unsigned int value, unsigned int bytes) |
1169 | { |
1170 | u8 *fault; |
1171 | u8 *end; |
1172 | u8 *addr = slab_address(slab); |
1173 | |
1174 | metadata_access_enable(); |
1175 | fault = memchr_inv(kasan_reset_tag(start), value, bytes); |
1176 | metadata_access_disable(); |
1177 | if (!fault) |
1178 | return 1; |
1179 | |
1180 | end = start + bytes; |
1181 | while (end > fault && end[-1] == value) |
1182 | end--; |
1183 | |
1184 | if (slab_add_kunit_errors()) |
1185 | goto skip_bug_print; |
1186 | |
1187 | slab_bug(s, "%s overwritten" , what); |
1188 | pr_err("0x%p-0x%p @offset=%tu. First byte 0x%x instead of 0x%x\n" , |
1189 | fault, end - 1, fault - addr, |
1190 | fault[0], value); |
1191 | print_trailer(s, slab, object); |
1192 | add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE); |
1193 | |
1194 | skip_bug_print: |
1195 | restore_bytes(s, what, value, fault, end); |
1196 | return 0; |
1197 | } |
1198 | |
1199 | /* |
1200 | * Object layout: |
1201 | * |
1202 | * object address |
1203 | * Bytes of the object to be managed. |
1204 | * If the freepointer may overlay the object then the free |
1205 | * pointer is at the middle of the object. |
1206 | * |
1207 | * Poisoning uses 0x6b (POISON_FREE) and the last byte is |
1208 | * 0xa5 (POISON_END) |
1209 | * |
1210 | * object + s->object_size |
1211 | * Padding to reach word boundary. This is also used for Redzoning. |
1212 | * Padding is extended by another word if Redzoning is enabled and |
1213 | * object_size == inuse. |
1214 | * |
1215 | * We fill with 0xbb (RED_INACTIVE) for inactive objects and with |
1216 | * 0xcc (RED_ACTIVE) for objects in use. |
1217 | * |
1218 | * object + s->inuse |
1219 | * Meta data starts here. |
1220 | * |
1221 | * A. Free pointer (if we cannot overwrite object on free) |
1222 | * B. Tracking data for SLAB_STORE_USER |
1223 | * C. Original request size for kmalloc object (SLAB_STORE_USER enabled) |
1224 | * D. Padding to reach required alignment boundary or at minimum |
1225 | * one word if debugging is on to be able to detect writes |
1226 | * before the word boundary. |
1227 | * |
1228 | * Padding is done using 0x5a (POISON_INUSE) |
1229 | * |
1230 | * object + s->size |
1231 | * Nothing is used beyond s->size. |
1232 | * |
1233 | * If slabcaches are merged then the object_size and inuse boundaries are mostly |
1234 | * ignored. And therefore no slab options that rely on these boundaries |
1235 | * may be used with merged slabcaches. |
1236 | */ |
1237 | |
1238 | static int check_pad_bytes(struct kmem_cache *s, struct slab *slab, u8 *p) |
1239 | { |
1240 | unsigned long off = get_info_end(s); /* The end of info */ |
1241 | |
1242 | if (s->flags & SLAB_STORE_USER) { |
1243 | /* We also have user information there */ |
1244 | off += 2 * sizeof(struct track); |
1245 | |
1246 | if (s->flags & SLAB_KMALLOC) |
1247 | off += sizeof(unsigned int); |
1248 | } |
1249 | |
1250 | off += kasan_metadata_size(s, false); |
1251 | |
1252 | if (size_from_object(s) == off) |
1253 | return 1; |
1254 | |
1255 | return check_bytes_and_report(s, slab, p, "Object padding" , |
1256 | p + off, POISON_INUSE, size_from_object(s) - off); |
1257 | } |
1258 | |
1259 | /* Check the pad bytes at the end of a slab page */ |
1260 | static void slab_pad_check(struct kmem_cache *s, struct slab *slab) |
1261 | { |
1262 | u8 *start; |
1263 | u8 *fault; |
1264 | u8 *end; |
1265 | u8 *pad; |
1266 | int length; |
1267 | int remainder; |
1268 | |
1269 | if (!(s->flags & SLAB_POISON)) |
1270 | return; |
1271 | |
1272 | start = slab_address(slab); |
1273 | length = slab_size(slab); |
1274 | end = start + length; |
1275 | remainder = length % s->size; |
1276 | if (!remainder) |
1277 | return; |
1278 | |
1279 | pad = end - remainder; |
1280 | metadata_access_enable(); |
1281 | fault = memchr_inv(kasan_reset_tag(pad), POISON_INUSE, remainder); |
1282 | metadata_access_disable(); |
1283 | if (!fault) |
1284 | return; |
1285 | while (end > fault && end[-1] == POISON_INUSE) |
1286 | end--; |
1287 | |
1288 | slab_err(s, slab, "Padding overwritten. 0x%p-0x%p @offset=%tu" , |
1289 | fault, end - 1, fault - start); |
1290 | print_section(KERN_ERR, "Padding " , pad, remainder); |
1291 | |
1292 | restore_bytes(s, "slab padding" , POISON_INUSE, fault, end); |
1293 | } |
1294 | |
1295 | static int check_object(struct kmem_cache *s, struct slab *slab, |
1296 | void *object, u8 val) |
1297 | { |
1298 | u8 *p = object; |
1299 | u8 *endobject = object + s->object_size; |
1300 | unsigned int orig_size, kasan_meta_size; |
1301 | |
1302 | if (s->flags & SLAB_RED_ZONE) { |
1303 | if (!check_bytes_and_report(s, slab, object, "Left Redzone" , |
1304 | object - s->red_left_pad, val, s->red_left_pad)) |
1305 | return 0; |
1306 | |
1307 | if (!check_bytes_and_report(s, slab, object, "Right Redzone" , |
1308 | endobject, val, s->inuse - s->object_size)) |
1309 | return 0; |
1310 | |
1311 | if (slub_debug_orig_size(s) && val == SLUB_RED_ACTIVE) { |
1312 | orig_size = get_orig_size(s, object); |
1313 | |
1314 | if (s->object_size > orig_size && |
1315 | !check_bytes_and_report(s, slab, object, |
1316 | "kmalloc Redzone" , p + orig_size, |
1317 | val, s->object_size - orig_size)) { |
1318 | return 0; |
1319 | } |
1320 | } |
1321 | } else { |
1322 | if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) { |
1323 | check_bytes_and_report(s, slab, p, "Alignment padding" , |
1324 | endobject, POISON_INUSE, |
1325 | s->inuse - s->object_size); |
1326 | } |
1327 | } |
1328 | |
1329 | if (s->flags & SLAB_POISON) { |
1330 | if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON)) { |
1331 | /* |
1332 | * KASAN can save its free meta data inside of the |
1333 | * object at offset 0. Thus, skip checking the part of |
1334 | * the redzone that overlaps with the meta data. |
1335 | */ |
1336 | kasan_meta_size = kasan_metadata_size(s, true); |
1337 | if (kasan_meta_size < s->object_size - 1 && |
1338 | !check_bytes_and_report(s, slab, p, "Poison" , |
1339 | p + kasan_meta_size, POISON_FREE, |
1340 | s->object_size - kasan_meta_size - 1)) |
1341 | return 0; |
1342 | if (kasan_meta_size < s->object_size && |
1343 | !check_bytes_and_report(s, slab, p, "End Poison" , |
1344 | p + s->object_size - 1, POISON_END, 1)) |
1345 | return 0; |
1346 | } |
1347 | /* |
1348 | * check_pad_bytes cleans up on its own. |
1349 | */ |
1350 | check_pad_bytes(s, slab, p); |
1351 | } |
1352 | |
1353 | if (!freeptr_outside_object(s) && val == SLUB_RED_ACTIVE) |
1354 | /* |
1355 | * Object and freepointer overlap. Cannot check |
1356 | * freepointer while object is allocated. |
1357 | */ |
1358 | return 1; |
1359 | |
1360 | /* Check free pointer validity */ |
1361 | if (!check_valid_pointer(s, slab, get_freepointer(s, p))) { |
1362 | object_err(s, slab, p, "Freepointer corrupt" ); |
1363 | /* |
1364 | * No choice but to zap it and thus lose the remainder |
1365 | * of the free objects in this slab. May cause |
1366 | * another error because the object count is now wrong. |
1367 | */ |
1368 | set_freepointer(s, p, NULL); |
1369 | return 0; |
1370 | } |
1371 | return 1; |
1372 | } |
1373 | |
1374 | static int check_slab(struct kmem_cache *s, struct slab *slab) |
1375 | { |
1376 | int maxobj; |
1377 | |
1378 | if (!folio_test_slab(slab_folio(slab))) { |
1379 | slab_err(s, slab, "Not a valid slab page" ); |
1380 | return 0; |
1381 | } |
1382 | |
1383 | maxobj = order_objects(slab_order(slab), s->size); |
1384 | if (slab->objects > maxobj) { |
1385 | slab_err(s, slab, "objects %u > max %u" , |
1386 | slab->objects, maxobj); |
1387 | return 0; |
1388 | } |
1389 | if (slab->inuse > slab->objects) { |
1390 | slab_err(s, slab, "inuse %u > max %u" , |
1391 | slab->inuse, slab->objects); |
1392 | return 0; |
1393 | } |
1394 | /* Slab_pad_check fixes things up after itself */ |
1395 | slab_pad_check(s, slab); |
1396 | return 1; |
1397 | } |
1398 | |
1399 | /* |
1400 | * Determine if a certain object in a slab is on the freelist. Must hold the |
1401 | * slab lock to guarantee that the chains are in a consistent state. |
1402 | */ |
1403 | static int on_freelist(struct kmem_cache *s, struct slab *slab, void *search) |
1404 | { |
1405 | int nr = 0; |
1406 | void *fp; |
1407 | void *object = NULL; |
1408 | int max_objects; |
1409 | |
1410 | fp = slab->freelist; |
1411 | while (fp && nr <= slab->objects) { |
1412 | if (fp == search) |
1413 | return 1; |
1414 | if (!check_valid_pointer(s, slab, fp)) { |
1415 | if (object) { |
1416 | object_err(s, slab, object, |
1417 | "Freechain corrupt" ); |
1418 | set_freepointer(s, object, NULL); |
1419 | } else { |
1420 | slab_err(s, slab, "Freepointer corrupt" ); |
1421 | slab->freelist = NULL; |
1422 | slab->inuse = slab->objects; |
1423 | slab_fix(s, "Freelist cleared" ); |
1424 | return 0; |
1425 | } |
1426 | break; |
1427 | } |
1428 | object = fp; |
1429 | fp = get_freepointer(s, object); |
1430 | nr++; |
1431 | } |
1432 | |
1433 | max_objects = order_objects(slab_order(slab), s->size); |
1434 | if (max_objects > MAX_OBJS_PER_PAGE) |
1435 | max_objects = MAX_OBJS_PER_PAGE; |
1436 | |
1437 | if (slab->objects != max_objects) { |
1438 | slab_err(s, slab, "Wrong number of objects. Found %d but should be %d" , |
1439 | slab->objects, max_objects); |
1440 | slab->objects = max_objects; |
1441 | slab_fix(s, "Number of objects adjusted" ); |
1442 | } |
1443 | if (slab->inuse != slab->objects - nr) { |
1444 | slab_err(s, slab, "Wrong object count. Counter is %d but counted were %d" , |
1445 | slab->inuse, slab->objects - nr); |
1446 | slab->inuse = slab->objects - nr; |
1447 | slab_fix(s, "Object count adjusted" ); |
1448 | } |
1449 | return search == NULL; |
1450 | } |
1451 | |
1452 | static void trace(struct kmem_cache *s, struct slab *slab, void *object, |
1453 | int alloc) |
1454 | { |
1455 | if (s->flags & SLAB_TRACE) { |
1456 | pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n" , |
1457 | s->name, |
1458 | alloc ? "alloc" : "free" , |
1459 | object, slab->inuse, |
1460 | slab->freelist); |
1461 | |
1462 | if (!alloc) |
1463 | print_section(KERN_INFO, "Object " , (void *)object, |
1464 | s->object_size); |
1465 | |
1466 | dump_stack(); |
1467 | } |
1468 | } |
1469 | |
1470 | /* |
1471 | * Tracking of fully allocated slabs for debugging purposes. |
1472 | */ |
1473 | static void add_full(struct kmem_cache *s, |
1474 | struct kmem_cache_node *n, struct slab *slab) |
1475 | { |
1476 | if (!(s->flags & SLAB_STORE_USER)) |
1477 | return; |
1478 | |
1479 | lockdep_assert_held(&n->list_lock); |
1480 | list_add(&slab->slab_list, &n->full); |
1481 | } |
1482 | |
1483 | static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct slab *slab) |
1484 | { |
1485 | if (!(s->flags & SLAB_STORE_USER)) |
1486 | return; |
1487 | |
1488 | lockdep_assert_held(&n->list_lock); |
1489 | list_del(&slab->slab_list); |
1490 | } |
1491 | |
1492 | static inline unsigned long node_nr_slabs(struct kmem_cache_node *n) |
1493 | { |
1494 | return atomic_long_read(&n->nr_slabs); |
1495 | } |
1496 | |
1497 | static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects) |
1498 | { |
1499 | struct kmem_cache_node *n = get_node(s, node); |
1500 | |
1501 | atomic_long_inc(&n->nr_slabs); |
1502 | atomic_long_add(objects, &n->total_objects); |
1503 | } |
1504 | static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects) |
1505 | { |
1506 | struct kmem_cache_node *n = get_node(s, node); |
1507 | |
1508 | atomic_long_dec(&n->nr_slabs); |
1509 | atomic_long_sub(objects, &n->total_objects); |
1510 | } |
1511 | |
1512 | /* Object debug checks for alloc/free paths */ |
1513 | static void setup_object_debug(struct kmem_cache *s, void *object) |
1514 | { |
1515 | if (!kmem_cache_debug_flags(s, SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)) |
1516 | return; |
1517 | |
1518 | init_object(s, object, SLUB_RED_INACTIVE); |
1519 | init_tracking(s, object); |
1520 | } |
1521 | |
1522 | static |
1523 | void setup_slab_debug(struct kmem_cache *s, struct slab *slab, void *addr) |
1524 | { |
1525 | if (!kmem_cache_debug_flags(s, SLAB_POISON)) |
1526 | return; |
1527 | |
1528 | metadata_access_enable(); |
1529 | memset(kasan_reset_tag(addr), POISON_INUSE, slab_size(slab)); |
1530 | metadata_access_disable(); |
1531 | } |
1532 | |
1533 | static inline int alloc_consistency_checks(struct kmem_cache *s, |
1534 | struct slab *slab, void *object) |
1535 | { |
1536 | if (!check_slab(s, slab)) |
1537 | return 0; |
1538 | |
1539 | if (!check_valid_pointer(s, slab, object)) { |
1540 | object_err(s, slab, object, "Freelist Pointer check fails" ); |
1541 | return 0; |
1542 | } |
1543 | |
1544 | if (!check_object(s, slab, object, SLUB_RED_INACTIVE)) |
1545 | return 0; |
1546 | |
1547 | return 1; |
1548 | } |
1549 | |
1550 | static noinline bool alloc_debug_processing(struct kmem_cache *s, |
1551 | struct slab *slab, void *object, int orig_size) |
1552 | { |
1553 | if (s->flags & SLAB_CONSISTENCY_CHECKS) { |
1554 | if (!alloc_consistency_checks(s, slab, object)) |
1555 | goto bad; |
1556 | } |
1557 | |
1558 | /* Success. Perform special debug activities for allocs */ |
1559 | trace(s, slab, object, 1); |
1560 | set_orig_size(s, object, orig_size); |
1561 | init_object(s, object, SLUB_RED_ACTIVE); |
1562 | return true; |
1563 | |
1564 | bad: |
1565 | if (folio_test_slab(slab_folio(slab))) { |
1566 | /* |
1567 | * If this is a slab page then lets do the best we can |
1568 | * to avoid issues in the future. Marking all objects |
1569 | * as used avoids touching the remaining objects. |
1570 | */ |
1571 | slab_fix(s, "Marking all objects used" ); |
1572 | slab->inuse = slab->objects; |
1573 | slab->freelist = NULL; |
1574 | } |
1575 | return false; |
1576 | } |
1577 | |
1578 | static inline int free_consistency_checks(struct kmem_cache *s, |
1579 | struct slab *slab, void *object, unsigned long addr) |
1580 | { |
1581 | if (!check_valid_pointer(s, slab, object)) { |
1582 | slab_err(s, slab, "Invalid object pointer 0x%p" , object); |
1583 | return 0; |
1584 | } |
1585 | |
1586 | if (on_freelist(s, slab, object)) { |
1587 | object_err(s, slab, object, "Object already free" ); |
1588 | return 0; |
1589 | } |
1590 | |
1591 | if (!check_object(s, slab, object, SLUB_RED_ACTIVE)) |
1592 | return 0; |
1593 | |
1594 | if (unlikely(s != slab->slab_cache)) { |
1595 | if (!folio_test_slab(slab_folio(slab))) { |
1596 | slab_err(s, slab, "Attempt to free object(0x%p) outside of slab" , |
1597 | object); |
1598 | } else if (!slab->slab_cache) { |
1599 | pr_err("SLUB <none>: no slab for object 0x%p.\n" , |
1600 | object); |
1601 | dump_stack(); |
1602 | } else |
1603 | object_err(s, slab, object, |
1604 | "page slab pointer corrupt." ); |
1605 | return 0; |
1606 | } |
1607 | return 1; |
1608 | } |
1609 | |
1610 | /* |
1611 | * Parse a block of slab_debug options. Blocks are delimited by ';' |
1612 | * |
1613 | * @str: start of block |
1614 | * @flags: returns parsed flags, or DEBUG_DEFAULT_FLAGS if none specified |
1615 | * @slabs: return start of list of slabs, or NULL when there's no list |
1616 | * @init: assume this is initial parsing and not per-kmem-create parsing |
1617 | * |
1618 | * returns the start of next block if there's any, or NULL |
1619 | */ |
1620 | static char * |
1621 | parse_slub_debug_flags(char *str, slab_flags_t *flags, char **slabs, bool init) |
1622 | { |
1623 | bool higher_order_disable = false; |
1624 | |
1625 | /* Skip any completely empty blocks */ |
1626 | while (*str && *str == ';') |
1627 | str++; |
1628 | |
1629 | if (*str == ',') { |
1630 | /* |
1631 | * No options but restriction on slabs. This means full |
1632 | * debugging for slabs matching a pattern. |
1633 | */ |
1634 | *flags = DEBUG_DEFAULT_FLAGS; |
1635 | goto check_slabs; |
1636 | } |
1637 | *flags = 0; |
1638 | |
1639 | /* Determine which debug features should be switched on */ |
1640 | for (; *str && *str != ',' && *str != ';'; str++) { |
1641 | switch (tolower(*str)) { |
1642 | case '-': |
1643 | *flags = 0; |
1644 | break; |
1645 | case 'f': |
1646 | *flags |= SLAB_CONSISTENCY_CHECKS; |
1647 | break; |
1648 | case 'z': |
1649 | *flags |= SLAB_RED_ZONE; |
1650 | break; |
1651 | case 'p': |
1652 | *flags |= SLAB_POISON; |
1653 | break; |
1654 | case 'u': |
1655 | *flags |= SLAB_STORE_USER; |
1656 | break; |
1657 | case 't': |
1658 | *flags |= SLAB_TRACE; |
1659 | break; |
1660 | case 'a': |
1661 | *flags |= SLAB_FAILSLAB; |
1662 | break; |
1663 | case 'o': |
1664 | /* |
1665 | * Avoid enabling debugging on caches if its minimum |
1666 | * order would increase as a result. |
1667 | */ |
1668 | higher_order_disable = true; |
1669 | break; |
1670 | default: |
1671 | if (init) |
1672 | pr_err("slab_debug option '%c' unknown. skipped\n" , *str); |
1673 | } |
1674 | } |
1675 | check_slabs: |
1676 | if (*str == ',') |
1677 | *slabs = ++str; |
1678 | else |
1679 | *slabs = NULL; |
1680 | |
1681 | /* Skip over the slab list */ |
1682 | while (*str && *str != ';') |
1683 | str++; |
1684 | |
1685 | /* Skip any completely empty blocks */ |
1686 | while (*str && *str == ';') |
1687 | str++; |
1688 | |
1689 | if (init && higher_order_disable) |
1690 | disable_higher_order_debug = 1; |
1691 | |
1692 | if (*str) |
1693 | return str; |
1694 | else |
1695 | return NULL; |
1696 | } |
1697 | |
1698 | static int __init setup_slub_debug(char *str) |
1699 | { |
1700 | slab_flags_t flags; |
1701 | slab_flags_t global_flags; |
1702 | char *saved_str; |
1703 | char *slab_list; |
1704 | bool global_slub_debug_changed = false; |
1705 | bool slab_list_specified = false; |
1706 | |
1707 | global_flags = DEBUG_DEFAULT_FLAGS; |
1708 | if (*str++ != '=' || !*str) |
1709 | /* |
1710 | * No options specified. Switch on full debugging. |
1711 | */ |
1712 | goto out; |
1713 | |
1714 | saved_str = str; |
1715 | while (str) { |
1716 | str = parse_slub_debug_flags(str, &flags, &slab_list, true); |
1717 | |
1718 | if (!slab_list) { |
1719 | global_flags = flags; |
1720 | global_slub_debug_changed = true; |
1721 | } else { |
1722 | slab_list_specified = true; |
1723 | if (flags & SLAB_STORE_USER) |
1724 | stack_depot_request_early_init(); |
1725 | } |
1726 | } |
1727 | |
1728 | /* |
1729 | * For backwards compatibility, a single list of flags with list of |
1730 | * slabs means debugging is only changed for those slabs, so the global |
1731 | * slab_debug should be unchanged (0 or DEBUG_DEFAULT_FLAGS, depending |
1732 | * on CONFIG_SLUB_DEBUG_ON). We can extended that to multiple lists as |
1733 | * long as there is no option specifying flags without a slab list. |
1734 | */ |
1735 | if (slab_list_specified) { |
1736 | if (!global_slub_debug_changed) |
1737 | global_flags = slub_debug; |
1738 | slub_debug_string = saved_str; |
1739 | } |
1740 | out: |
1741 | slub_debug = global_flags; |
1742 | if (slub_debug & SLAB_STORE_USER) |
1743 | stack_depot_request_early_init(); |
1744 | if (slub_debug != 0 || slub_debug_string) |
1745 | static_branch_enable(&slub_debug_enabled); |
1746 | else |
1747 | static_branch_disable(&slub_debug_enabled); |
1748 | if ((static_branch_unlikely(&init_on_alloc) || |
1749 | static_branch_unlikely(&init_on_free)) && |
1750 | (slub_debug & SLAB_POISON)) |
1751 | pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n" ); |
1752 | return 1; |
1753 | } |
1754 | |
1755 | __setup("slab_debug" , setup_slub_debug); |
1756 | __setup_param("slub_debug" , slub_debug, setup_slub_debug, 0); |
1757 | |
1758 | /* |
1759 | * kmem_cache_flags - apply debugging options to the cache |
1760 | * @flags: flags to set |
1761 | * @name: name of the cache |
1762 | * |
1763 | * Debug option(s) are applied to @flags. In addition to the debug |
1764 | * option(s), if a slab name (or multiple) is specified i.e. |
1765 | * slab_debug=<Debug-Options>,<slab name1>,<slab name2> ... |
1766 | * then only the select slabs will receive the debug option(s). |
1767 | */ |
1768 | slab_flags_t kmem_cache_flags(slab_flags_t flags, const char *name) |
1769 | { |
1770 | char *iter; |
1771 | size_t len; |
1772 | char *next_block; |
1773 | slab_flags_t block_flags; |
1774 | slab_flags_t slub_debug_local = slub_debug; |
1775 | |
1776 | if (flags & SLAB_NO_USER_FLAGS) |
1777 | return flags; |
1778 | |
1779 | /* |
1780 | * If the slab cache is for debugging (e.g. kmemleak) then |
1781 | * don't store user (stack trace) information by default, |
1782 | * but let the user enable it via the command line below. |
1783 | */ |
1784 | if (flags & SLAB_NOLEAKTRACE) |
1785 | slub_debug_local &= ~SLAB_STORE_USER; |
1786 | |
1787 | len = strlen(name); |
1788 | next_block = slub_debug_string; |
1789 | /* Go through all blocks of debug options, see if any matches our slab's name */ |
1790 | while (next_block) { |
1791 | next_block = parse_slub_debug_flags(next_block, &block_flags, &iter, false); |
1792 | if (!iter) |
1793 | continue; |
1794 | /* Found a block that has a slab list, search it */ |
1795 | while (*iter) { |
1796 | char *end, *glob; |
1797 | size_t cmplen; |
1798 | |
1799 | end = strchrnul(iter, ','); |
1800 | if (next_block && next_block < end) |
1801 | end = next_block - 1; |
1802 | |
1803 | glob = strnchr(iter, end - iter, '*'); |
1804 | if (glob) |
1805 | cmplen = glob - iter; |
1806 | else |
1807 | cmplen = max_t(size_t, len, (end - iter)); |
1808 | |
1809 | if (!strncmp(name, iter, cmplen)) { |
1810 | flags |= block_flags; |
1811 | return flags; |
1812 | } |
1813 | |
1814 | if (!*end || *end == ';') |
1815 | break; |
1816 | iter = end + 1; |
1817 | } |
1818 | } |
1819 | |
1820 | return flags | slub_debug_local; |
1821 | } |
1822 | #else /* !CONFIG_SLUB_DEBUG */ |
1823 | static inline void setup_object_debug(struct kmem_cache *s, void *object) {} |
1824 | static inline |
1825 | void setup_slab_debug(struct kmem_cache *s, struct slab *slab, void *addr) {} |
1826 | |
1827 | static inline bool alloc_debug_processing(struct kmem_cache *s, |
1828 | struct slab *slab, void *object, int orig_size) { return true; } |
1829 | |
1830 | static inline bool free_debug_processing(struct kmem_cache *s, |
1831 | struct slab *slab, void *head, void *tail, int *bulk_cnt, |
1832 | unsigned long addr, depot_stack_handle_t handle) { return true; } |
1833 | |
1834 | static inline void slab_pad_check(struct kmem_cache *s, struct slab *slab) {} |
1835 | static inline int check_object(struct kmem_cache *s, struct slab *slab, |
1836 | void *object, u8 val) { return 1; } |
1837 | static inline depot_stack_handle_t set_track_prepare(void) { return 0; } |
1838 | static inline void set_track(struct kmem_cache *s, void *object, |
1839 | enum track_item alloc, unsigned long addr) {} |
1840 | static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n, |
1841 | struct slab *slab) {} |
1842 | static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, |
1843 | struct slab *slab) {} |
1844 | slab_flags_t kmem_cache_flags(slab_flags_t flags, const char *name) |
1845 | { |
1846 | return flags; |
1847 | } |
1848 | #define slub_debug 0 |
1849 | |
1850 | #define disable_higher_order_debug 0 |
1851 | |
1852 | static inline unsigned long node_nr_slabs(struct kmem_cache_node *n) |
1853 | { return 0; } |
1854 | static inline void inc_slabs_node(struct kmem_cache *s, int node, |
1855 | int objects) {} |
1856 | static inline void dec_slabs_node(struct kmem_cache *s, int node, |
1857 | int objects) {} |
1858 | |
1859 | #ifndef CONFIG_SLUB_TINY |
1860 | static bool freelist_corrupted(struct kmem_cache *s, struct slab *slab, |
1861 | void **freelist, void *nextfree) |
1862 | { |
1863 | return false; |
1864 | } |
1865 | #endif |
1866 | #endif /* CONFIG_SLUB_DEBUG */ |
1867 | |
1868 | static inline enum node_stat_item cache_vmstat_idx(struct kmem_cache *s) |
1869 | { |
1870 | return (s->flags & SLAB_RECLAIM_ACCOUNT) ? |
1871 | NR_SLAB_RECLAIMABLE_B : NR_SLAB_UNRECLAIMABLE_B; |
1872 | } |
1873 | |
1874 | #ifdef CONFIG_MEMCG_KMEM |
1875 | static inline void memcg_free_slab_cgroups(struct slab *slab) |
1876 | { |
1877 | kfree(objp: slab_objcgs(slab)); |
1878 | slab->memcg_data = 0; |
1879 | } |
1880 | |
1881 | static inline size_t obj_full_size(struct kmem_cache *s) |
1882 | { |
1883 | /* |
1884 | * For each accounted object there is an extra space which is used |
1885 | * to store obj_cgroup membership. Charge it too. |
1886 | */ |
1887 | return s->size + sizeof(struct obj_cgroup *); |
1888 | } |
1889 | |
1890 | /* |
1891 | * Returns false if the allocation should fail. |
1892 | */ |
1893 | static bool __memcg_slab_pre_alloc_hook(struct kmem_cache *s, |
1894 | struct list_lru *lru, |
1895 | struct obj_cgroup **objcgp, |
1896 | size_t objects, gfp_t flags) |
1897 | { |
1898 | /* |
1899 | * The obtained objcg pointer is safe to use within the current scope, |
1900 | * defined by current task or set_active_memcg() pair. |
1901 | * obj_cgroup_get() is used to get a permanent reference. |
1902 | */ |
1903 | struct obj_cgroup *objcg = current_obj_cgroup(); |
1904 | if (!objcg) |
1905 | return true; |
1906 | |
1907 | if (lru) { |
1908 | int ret; |
1909 | struct mem_cgroup *memcg; |
1910 | |
1911 | memcg = get_mem_cgroup_from_objcg(objcg); |
1912 | ret = memcg_list_lru_alloc(memcg, lru, gfp: flags); |
1913 | css_put(css: &memcg->css); |
1914 | |
1915 | if (ret) |
1916 | return false; |
1917 | } |
1918 | |
1919 | if (obj_cgroup_charge(objcg, gfp: flags, size: objects * obj_full_size(s))) |
1920 | return false; |
1921 | |
1922 | *objcgp = objcg; |
1923 | return true; |
1924 | } |
1925 | |
1926 | /* |
1927 | * Returns false if the allocation should fail. |
1928 | */ |
1929 | static __fastpath_inline |
1930 | bool memcg_slab_pre_alloc_hook(struct kmem_cache *s, struct list_lru *lru, |
1931 | struct obj_cgroup **objcgp, size_t objects, |
1932 | gfp_t flags) |
1933 | { |
1934 | if (!memcg_kmem_online()) |
1935 | return true; |
1936 | |
1937 | if (likely(!(flags & __GFP_ACCOUNT) && !(s->flags & SLAB_ACCOUNT))) |
1938 | return true; |
1939 | |
1940 | return likely(__memcg_slab_pre_alloc_hook(s, lru, objcgp, objects, |
1941 | flags)); |
1942 | } |
1943 | |
1944 | static void __memcg_slab_post_alloc_hook(struct kmem_cache *s, |
1945 | struct obj_cgroup *objcg, |
1946 | gfp_t flags, size_t size, |
1947 | void **p) |
1948 | { |
1949 | struct slab *slab; |
1950 | unsigned long off; |
1951 | size_t i; |
1952 | |
1953 | flags &= gfp_allowed_mask; |
1954 | |
1955 | for (i = 0; i < size; i++) { |
1956 | if (likely(p[i])) { |
1957 | slab = virt_to_slab(addr: p[i]); |
1958 | |
1959 | if (!slab_objcgs(slab) && |
1960 | memcg_alloc_slab_cgroups(slab, s, gfp: flags, new_slab: false)) { |
1961 | obj_cgroup_uncharge(objcg, size: obj_full_size(s)); |
1962 | continue; |
1963 | } |
1964 | |
1965 | off = obj_to_index(cache: s, slab, obj: p[i]); |
1966 | obj_cgroup_get(objcg); |
1967 | slab_objcgs(slab)[off] = objcg; |
1968 | mod_objcg_state(objcg, pgdat: slab_pgdat(slab), |
1969 | idx: cache_vmstat_idx(s), nr: obj_full_size(s)); |
1970 | } else { |
1971 | obj_cgroup_uncharge(objcg, size: obj_full_size(s)); |
1972 | } |
1973 | } |
1974 | } |
1975 | |
1976 | static __fastpath_inline |
1977 | void memcg_slab_post_alloc_hook(struct kmem_cache *s, struct obj_cgroup *objcg, |
1978 | gfp_t flags, size_t size, void **p) |
1979 | { |
1980 | if (likely(!memcg_kmem_online() || !objcg)) |
1981 | return; |
1982 | |
1983 | return __memcg_slab_post_alloc_hook(s, objcg, flags, size, p); |
1984 | } |
1985 | |
1986 | static void __memcg_slab_free_hook(struct kmem_cache *s, struct slab *slab, |
1987 | void **p, int objects, |
1988 | struct obj_cgroup **objcgs) |
1989 | { |
1990 | for (int i = 0; i < objects; i++) { |
1991 | struct obj_cgroup *objcg; |
1992 | unsigned int off; |
1993 | |
1994 | off = obj_to_index(cache: s, slab, obj: p[i]); |
1995 | objcg = objcgs[off]; |
1996 | if (!objcg) |
1997 | continue; |
1998 | |
1999 | objcgs[off] = NULL; |
2000 | obj_cgroup_uncharge(objcg, size: obj_full_size(s)); |
2001 | mod_objcg_state(objcg, pgdat: slab_pgdat(slab), idx: cache_vmstat_idx(s), |
2002 | nr: -obj_full_size(s)); |
2003 | obj_cgroup_put(objcg); |
2004 | } |
2005 | } |
2006 | |
2007 | static __fastpath_inline |
2008 | void memcg_slab_free_hook(struct kmem_cache *s, struct slab *slab, void **p, |
2009 | int objects) |
2010 | { |
2011 | struct obj_cgroup **objcgs; |
2012 | |
2013 | if (!memcg_kmem_online()) |
2014 | return; |
2015 | |
2016 | objcgs = slab_objcgs(slab); |
2017 | if (likely(!objcgs)) |
2018 | return; |
2019 | |
2020 | __memcg_slab_free_hook(s, slab, p, objects, objcgs); |
2021 | } |
2022 | |
2023 | static inline |
2024 | void memcg_slab_alloc_error_hook(struct kmem_cache *s, int objects, |
2025 | struct obj_cgroup *objcg) |
2026 | { |
2027 | if (objcg) |
2028 | obj_cgroup_uncharge(objcg, size: objects * obj_full_size(s)); |
2029 | } |
2030 | #else /* CONFIG_MEMCG_KMEM */ |
2031 | static inline void memcg_free_slab_cgroups(struct slab *slab) |
2032 | { |
2033 | } |
2034 | |
2035 | static inline bool memcg_slab_pre_alloc_hook(struct kmem_cache *s, |
2036 | struct list_lru *lru, |
2037 | struct obj_cgroup **objcgp, |
2038 | size_t objects, gfp_t flags) |
2039 | { |
2040 | return true; |
2041 | } |
2042 | |
2043 | static inline void memcg_slab_post_alloc_hook(struct kmem_cache *s, |
2044 | struct obj_cgroup *objcg, |
2045 | gfp_t flags, size_t size, |
2046 | void **p) |
2047 | { |
2048 | } |
2049 | |
2050 | static inline void memcg_slab_free_hook(struct kmem_cache *s, struct slab *slab, |
2051 | void **p, int objects) |
2052 | { |
2053 | } |
2054 | |
2055 | static inline |
2056 | void memcg_slab_alloc_error_hook(struct kmem_cache *s, int objects, |
2057 | struct obj_cgroup *objcg) |
2058 | { |
2059 | } |
2060 | #endif /* CONFIG_MEMCG_KMEM */ |
2061 | |
2062 | /* |
2063 | * Hooks for other subsystems that check memory allocations. In a typical |
2064 | * production configuration these hooks all should produce no code at all. |
2065 | * |
2066 | * Returns true if freeing of the object can proceed, false if its reuse |
2067 | * was delayed by KASAN quarantine, or it was returned to KFENCE. |
2068 | */ |
2069 | static __always_inline |
2070 | bool slab_free_hook(struct kmem_cache *s, void *x, bool init) |
2071 | { |
2072 | kmemleak_free_recursive(ptr: x, flags: s->flags); |
2073 | kmsan_slab_free(s, object: x); |
2074 | |
2075 | debug_check_no_locks_freed(from: x, len: s->object_size); |
2076 | |
2077 | if (!(s->flags & SLAB_DEBUG_OBJECTS)) |
2078 | debug_check_no_obj_freed(address: x, size: s->object_size); |
2079 | |
2080 | /* Use KCSAN to help debug racy use-after-free. */ |
2081 | if (!(s->flags & SLAB_TYPESAFE_BY_RCU)) |
2082 | __kcsan_check_access(ptr: x, size: s->object_size, |
2083 | KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT); |
2084 | |
2085 | if (kfence_free(addr: x)) |
2086 | return false; |
2087 | |
2088 | /* |
2089 | * As memory initialization might be integrated into KASAN, |
2090 | * kasan_slab_free and initialization memset's must be |
2091 | * kept together to avoid discrepancies in behavior. |
2092 | * |
2093 | * The initialization memset's clear the object and the metadata, |
2094 | * but don't touch the SLAB redzone. |
2095 | */ |
2096 | if (unlikely(init)) { |
2097 | int rsize; |
2098 | |
2099 | if (!kasan_has_integrated_init()) |
2100 | memset(kasan_reset_tag(x), 0, s->object_size); |
2101 | rsize = (s->flags & SLAB_RED_ZONE) ? s->red_left_pad : 0; |
2102 | memset((char *)kasan_reset_tag(x) + s->inuse, 0, |
2103 | s->size - s->inuse - rsize); |
2104 | } |
2105 | /* KASAN might put x into memory quarantine, delaying its reuse. */ |
2106 | return !kasan_slab_free(s, object: x, init); |
2107 | } |
2108 | |
2109 | static inline bool slab_free_freelist_hook(struct kmem_cache *s, |
2110 | void **head, void **tail, |
2111 | int *cnt) |
2112 | { |
2113 | |
2114 | void *object; |
2115 | void *next = *head; |
2116 | void *old_tail = *tail; |
2117 | bool init; |
2118 | |
2119 | if (is_kfence_address(addr: next)) { |
2120 | slab_free_hook(s, x: next, init: false); |
2121 | return false; |
2122 | } |
2123 | |
2124 | /* Head and tail of the reconstructed freelist */ |
2125 | *head = NULL; |
2126 | *tail = NULL; |
2127 | |
2128 | init = slab_want_init_on_free(c: s); |
2129 | |
2130 | do { |
2131 | object = next; |
2132 | next = get_freepointer(s, object); |
2133 | |
2134 | /* If object's reuse doesn't have to be delayed */ |
2135 | if (likely(slab_free_hook(s, object, init))) { |
2136 | /* Move object to the new freelist */ |
2137 | set_freepointer(s, object, fp: *head); |
2138 | *head = object; |
2139 | if (!*tail) |
2140 | *tail = object; |
2141 | } else { |
2142 | /* |
2143 | * Adjust the reconstructed freelist depth |
2144 | * accordingly if object's reuse is delayed. |
2145 | */ |
2146 | --(*cnt); |
2147 | } |
2148 | } while (object != old_tail); |
2149 | |
2150 | return *head != NULL; |
2151 | } |
2152 | |
2153 | static void *setup_object(struct kmem_cache *s, void *object) |
2154 | { |
2155 | setup_object_debug(s, object); |
2156 | object = kasan_init_slab_obj(cache: s, object); |
2157 | if (unlikely(s->ctor)) { |
2158 | kasan_unpoison_new_object(cache: s, object); |
2159 | s->ctor(object); |
2160 | kasan_poison_new_object(cache: s, object); |
2161 | } |
2162 | return object; |
2163 | } |
2164 | |
2165 | /* |
2166 | * Slab allocation and freeing |
2167 | */ |
2168 | static inline struct slab *alloc_slab_page(gfp_t flags, int node, |
2169 | struct kmem_cache_order_objects oo) |
2170 | { |
2171 | struct folio *folio; |
2172 | struct slab *slab; |
2173 | unsigned int order = oo_order(x: oo); |
2174 | |
2175 | folio = (struct folio *)alloc_pages_node(nid: node, gfp_mask: flags, order); |
2176 | if (!folio) |
2177 | return NULL; |
2178 | |
2179 | slab = folio_slab(folio); |
2180 | __folio_set_slab(folio); |
2181 | /* Make the flag visible before any changes to folio->mapping */ |
2182 | smp_wmb(); |
2183 | if (folio_is_pfmemalloc(folio)) |
2184 | slab_set_pfmemalloc(slab); |
2185 | |
2186 | return slab; |
2187 | } |
2188 | |
2189 | #ifdef CONFIG_SLAB_FREELIST_RANDOM |
2190 | /* Pre-initialize the random sequence cache */ |
2191 | static int init_cache_random_seq(struct kmem_cache *s) |
2192 | { |
2193 | unsigned int count = oo_objects(s->oo); |
2194 | int err; |
2195 | |
2196 | /* Bailout if already initialised */ |
2197 | if (s->random_seq) |
2198 | return 0; |
2199 | |
2200 | err = cache_random_seq_create(s, count, GFP_KERNEL); |
2201 | if (err) { |
2202 | pr_err("SLUB: Unable to initialize free list for %s\n" , |
2203 | s->name); |
2204 | return err; |
2205 | } |
2206 | |
2207 | /* Transform to an offset on the set of pages */ |
2208 | if (s->random_seq) { |
2209 | unsigned int i; |
2210 | |
2211 | for (i = 0; i < count; i++) |
2212 | s->random_seq[i] *= s->size; |
2213 | } |
2214 | return 0; |
2215 | } |
2216 | |
2217 | /* Initialize each random sequence freelist per cache */ |
2218 | static void __init init_freelist_randomization(void) |
2219 | { |
2220 | struct kmem_cache *s; |
2221 | |
2222 | mutex_lock(&slab_mutex); |
2223 | |
2224 | list_for_each_entry(s, &slab_caches, list) |
2225 | init_cache_random_seq(s); |
2226 | |
2227 | mutex_unlock(&slab_mutex); |
2228 | } |
2229 | |
2230 | /* Get the next entry on the pre-computed freelist randomized */ |
2231 | static void *next_freelist_entry(struct kmem_cache *s, |
2232 | unsigned long *pos, void *start, |
2233 | unsigned long page_limit, |
2234 | unsigned long freelist_count) |
2235 | { |
2236 | unsigned int idx; |
2237 | |
2238 | /* |
2239 | * If the target page allocation failed, the number of objects on the |
2240 | * page might be smaller than the usual size defined by the cache. |
2241 | */ |
2242 | do { |
2243 | idx = s->random_seq[*pos]; |
2244 | *pos += 1; |
2245 | if (*pos >= freelist_count) |
2246 | *pos = 0; |
2247 | } while (unlikely(idx >= page_limit)); |
2248 | |
2249 | return (char *)start + idx; |
2250 | } |
2251 | |
2252 | /* Shuffle the single linked freelist based on a random pre-computed sequence */ |
2253 | static bool shuffle_freelist(struct kmem_cache *s, struct slab *slab) |
2254 | { |
2255 | void *start; |
2256 | void *cur; |
2257 | void *next; |
2258 | unsigned long idx, pos, page_limit, freelist_count; |
2259 | |
2260 | if (slab->objects < 2 || !s->random_seq) |
2261 | return false; |
2262 | |
2263 | freelist_count = oo_objects(s->oo); |
2264 | pos = get_random_u32_below(freelist_count); |
2265 | |
2266 | page_limit = slab->objects * s->size; |
2267 | start = fixup_red_left(s, slab_address(slab)); |
2268 | |
2269 | /* First entry is used as the base of the freelist */ |
2270 | cur = next_freelist_entry(s, &pos, start, page_limit, freelist_count); |
2271 | cur = setup_object(s, cur); |
2272 | slab->freelist = cur; |
2273 | |
2274 | for (idx = 1; idx < slab->objects; idx++) { |
2275 | next = next_freelist_entry(s, &pos, start, page_limit, |
2276 | freelist_count); |
2277 | next = setup_object(s, next); |
2278 | set_freepointer(s, cur, next); |
2279 | cur = next; |
2280 | } |
2281 | set_freepointer(s, cur, NULL); |
2282 | |
2283 | return true; |
2284 | } |
2285 | #else |
2286 | static inline int init_cache_random_seq(struct kmem_cache *s) |
2287 | { |
2288 | return 0; |
2289 | } |
2290 | static inline void init_freelist_randomization(void) { } |
2291 | static inline bool shuffle_freelist(struct kmem_cache *s, struct slab *slab) |
2292 | { |
2293 | return false; |
2294 | } |
2295 | #endif /* CONFIG_SLAB_FREELIST_RANDOM */ |
2296 | |
2297 | static __always_inline void account_slab(struct slab *slab, int order, |
2298 | struct kmem_cache *s, gfp_t gfp) |
2299 | { |
2300 | if (memcg_kmem_online() && (s->flags & SLAB_ACCOUNT)) |
2301 | memcg_alloc_slab_cgroups(slab, s, gfp, new_slab: true); |
2302 | |
2303 | mod_node_page_state(slab_pgdat(slab), cache_vmstat_idx(s), |
2304 | PAGE_SIZE << order); |
2305 | } |
2306 | |
2307 | static __always_inline void unaccount_slab(struct slab *slab, int order, |
2308 | struct kmem_cache *s) |
2309 | { |
2310 | if (memcg_kmem_online()) |
2311 | memcg_free_slab_cgroups(slab); |
2312 | |
2313 | mod_node_page_state(slab_pgdat(slab), cache_vmstat_idx(s), |
2314 | -(PAGE_SIZE << order)); |
2315 | } |
2316 | |
2317 | static struct slab *allocate_slab(struct kmem_cache *s, gfp_t flags, int node) |
2318 | { |
2319 | struct slab *slab; |
2320 | struct kmem_cache_order_objects oo = s->oo; |
2321 | gfp_t alloc_gfp; |
2322 | void *start, *p, *next; |
2323 | int idx; |
2324 | bool shuffle; |
2325 | |
2326 | flags &= gfp_allowed_mask; |
2327 | |
2328 | flags |= s->allocflags; |
2329 | |
2330 | /* |
2331 | * Let the initial higher-order allocation fail under memory pressure |
2332 | * so we fall-back to the minimum order allocation. |
2333 | */ |
2334 | alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL; |
2335 | if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(x: oo) > oo_order(x: s->min)) |
2336 | alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~__GFP_RECLAIM; |
2337 | |
2338 | slab = alloc_slab_page(flags: alloc_gfp, node, oo); |
2339 | if (unlikely(!slab)) { |
2340 | oo = s->min; |
2341 | alloc_gfp = flags; |
2342 | /* |
2343 | * Allocation may have failed due to fragmentation. |
2344 | * Try a lower order alloc if possible |
2345 | */ |
2346 | slab = alloc_slab_page(flags: alloc_gfp, node, oo); |
2347 | if (unlikely(!slab)) |
2348 | return NULL; |
2349 | stat(s, si: ORDER_FALLBACK); |
2350 | } |
2351 | |
2352 | slab->objects = oo_objects(x: oo); |
2353 | slab->inuse = 0; |
2354 | slab->frozen = 0; |
2355 | |
2356 | account_slab(slab, order: oo_order(x: oo), s, gfp: flags); |
2357 | |
2358 | slab->slab_cache = s; |
2359 | |
2360 | kasan_poison_slab(slab); |
2361 | |
2362 | start = slab_address(slab); |
2363 | |
2364 | setup_slab_debug(s, slab, addr: start); |
2365 | |
2366 | shuffle = shuffle_freelist(s, slab); |
2367 | |
2368 | if (!shuffle) { |
2369 | start = fixup_red_left(s, p: start); |
2370 | start = setup_object(s, object: start); |
2371 | slab->freelist = start; |
2372 | for (idx = 0, p = start; idx < slab->objects - 1; idx++) { |
2373 | next = p + s->size; |
2374 | next = setup_object(s, object: next); |
2375 | set_freepointer(s, object: p, fp: next); |
2376 | p = next; |
2377 | } |
2378 | set_freepointer(s, object: p, NULL); |
2379 | } |
2380 | |
2381 | return slab; |
2382 | } |
2383 | |
2384 | static struct slab *new_slab(struct kmem_cache *s, gfp_t flags, int node) |
2385 | { |
2386 | if (unlikely(flags & GFP_SLAB_BUG_MASK)) |
2387 | flags = kmalloc_fix_flags(flags); |
2388 | |
2389 | WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO)); |
2390 | |
2391 | return allocate_slab(s, |
2392 | flags: flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node); |
2393 | } |
2394 | |
2395 | static void __free_slab(struct kmem_cache *s, struct slab *slab) |
2396 | { |
2397 | struct folio *folio = slab_folio(slab); |
2398 | int order = folio_order(folio); |
2399 | int pages = 1 << order; |
2400 | |
2401 | __slab_clear_pfmemalloc(slab); |
2402 | folio->mapping = NULL; |
2403 | /* Make the mapping reset visible before clearing the flag */ |
2404 | smp_wmb(); |
2405 | __folio_clear_slab(folio); |
2406 | mm_account_reclaimed_pages(pages); |
2407 | unaccount_slab(slab, order, s); |
2408 | __free_pages(page: &folio->page, order); |
2409 | } |
2410 | |
2411 | static void rcu_free_slab(struct rcu_head *h) |
2412 | { |
2413 | struct slab *slab = container_of(h, struct slab, rcu_head); |
2414 | |
2415 | __free_slab(s: slab->slab_cache, slab); |
2416 | } |
2417 | |
2418 | static void free_slab(struct kmem_cache *s, struct slab *slab) |
2419 | { |
2420 | if (kmem_cache_debug_flags(s, SLAB_CONSISTENCY_CHECKS)) { |
2421 | void *p; |
2422 | |
2423 | slab_pad_check(s, slab); |
2424 | for_each_object(p, s, slab_address(slab), slab->objects) |
2425 | check_object(s, slab, object: p, SLUB_RED_INACTIVE); |
2426 | } |
2427 | |
2428 | if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU)) |
2429 | call_rcu(head: &slab->rcu_head, func: rcu_free_slab); |
2430 | else |
2431 | __free_slab(s, slab); |
2432 | } |
2433 | |
2434 | static void discard_slab(struct kmem_cache *s, struct slab *slab) |
2435 | { |
2436 | dec_slabs_node(s, node: slab_nid(slab), objects: slab->objects); |
2437 | free_slab(s, slab); |
2438 | } |
2439 | |
2440 | /* |
2441 | * SLUB reuses PG_workingset bit to keep track of whether it's on |
2442 | * the per-node partial list. |
2443 | */ |
2444 | static inline bool slab_test_node_partial(const struct slab *slab) |
2445 | { |
2446 | return folio_test_workingset(folio: (struct folio *)slab_folio(slab)); |
2447 | } |
2448 | |
2449 | static inline void slab_set_node_partial(struct slab *slab) |
2450 | { |
2451 | set_bit(nr: PG_workingset, addr: folio_flags(slab_folio(slab), n: 0)); |
2452 | } |
2453 | |
2454 | static inline void slab_clear_node_partial(struct slab *slab) |
2455 | { |
2456 | clear_bit(nr: PG_workingset, addr: folio_flags(slab_folio(slab), n: 0)); |
2457 | } |
2458 | |
2459 | /* |
2460 | * Management of partially allocated slabs. |
2461 | */ |
2462 | static inline void |
2463 | __add_partial(struct kmem_cache_node *n, struct slab *slab, int tail) |
2464 | { |
2465 | n->nr_partial++; |
2466 | if (tail == DEACTIVATE_TO_TAIL) |
2467 | list_add_tail(new: &slab->slab_list, head: &n->partial); |
2468 | else |
2469 | list_add(new: &slab->slab_list, head: &n->partial); |
2470 | slab_set_node_partial(slab); |
2471 | } |
2472 | |
2473 | static inline void add_partial(struct kmem_cache_node *n, |
2474 | struct slab *slab, int tail) |
2475 | { |
2476 | lockdep_assert_held(&n->list_lock); |
2477 | __add_partial(n, slab, tail); |
2478 | } |
2479 | |
2480 | static inline void remove_partial(struct kmem_cache_node *n, |
2481 | struct slab *slab) |
2482 | { |
2483 | lockdep_assert_held(&n->list_lock); |
2484 | list_del(entry: &slab->slab_list); |
2485 | slab_clear_node_partial(slab); |
2486 | n->nr_partial--; |
2487 | } |
2488 | |
2489 | /* |
2490 | * Called only for kmem_cache_debug() caches instead of remove_partial(), with a |
2491 | * slab from the n->partial list. Remove only a single object from the slab, do |
2492 | * the alloc_debug_processing() checks and leave the slab on the list, or move |
2493 | * it to full list if it was the last free object. |
2494 | */ |
2495 | static void *alloc_single_from_partial(struct kmem_cache *s, |
2496 | struct kmem_cache_node *n, struct slab *slab, int orig_size) |
2497 | { |
2498 | void *object; |
2499 | |
2500 | lockdep_assert_held(&n->list_lock); |
2501 | |
2502 | object = slab->freelist; |
2503 | slab->freelist = get_freepointer(s, object); |
2504 | slab->inuse++; |
2505 | |
2506 | if (!alloc_debug_processing(s, slab, object, orig_size)) { |
2507 | remove_partial(n, slab); |
2508 | return NULL; |
2509 | } |
2510 | |
2511 | if (slab->inuse == slab->objects) { |
2512 | remove_partial(n, slab); |
2513 | add_full(s, n, slab); |
2514 | } |
2515 | |
2516 | return object; |
2517 | } |
2518 | |
2519 | /* |
2520 | * Called only for kmem_cache_debug() caches to allocate from a freshly |
2521 | * allocated slab. Allocate a single object instead of whole freelist |
2522 | * and put the slab to the partial (or full) list. |
2523 | */ |
2524 | static void *alloc_single_from_new_slab(struct kmem_cache *s, |
2525 | struct slab *slab, int orig_size) |
2526 | { |
2527 | int nid = slab_nid(slab); |
2528 | struct kmem_cache_node *n = get_node(s, node: nid); |
2529 | unsigned long flags; |
2530 | void *object; |
2531 | |
2532 | |
2533 | object = slab->freelist; |
2534 | slab->freelist = get_freepointer(s, object); |
2535 | slab->inuse = 1; |
2536 | |
2537 | if (!alloc_debug_processing(s, slab, object, orig_size)) |
2538 | /* |
2539 | * It's not really expected that this would fail on a |
2540 | * freshly allocated slab, but a concurrent memory |
2541 | * corruption in theory could cause that. |
2542 | */ |
2543 | return NULL; |
2544 | |
2545 | spin_lock_irqsave(&n->list_lock, flags); |
2546 | |
2547 | if (slab->inuse == slab->objects) |
2548 | add_full(s, n, slab); |
2549 | else |
2550 | add_partial(n, slab, tail: DEACTIVATE_TO_HEAD); |
2551 | |
2552 | inc_slabs_node(s, node: nid, objects: slab->objects); |
2553 | spin_unlock_irqrestore(lock: &n->list_lock, flags); |
2554 | |
2555 | return object; |
2556 | } |
2557 | |
2558 | #ifdef CONFIG_SLUB_CPU_PARTIAL |
2559 | static void put_cpu_partial(struct kmem_cache *s, struct slab *slab, int drain); |
2560 | #else |
2561 | static inline void put_cpu_partial(struct kmem_cache *s, struct slab *slab, |
2562 | int drain) { } |
2563 | #endif |
2564 | static inline bool pfmemalloc_match(struct slab *slab, gfp_t gfpflags); |
2565 | |
2566 | /* |
2567 | * Try to allocate a partial slab from a specific node. |
2568 | */ |
2569 | static struct slab *get_partial_node(struct kmem_cache *s, |
2570 | struct kmem_cache_node *n, |
2571 | struct partial_context *pc) |
2572 | { |
2573 | struct slab *slab, *slab2, *partial = NULL; |
2574 | unsigned long flags; |
2575 | unsigned int partial_slabs = 0; |
2576 | |
2577 | /* |
2578 | * Racy check. If we mistakenly see no partial slabs then we |
2579 | * just allocate an empty slab. If we mistakenly try to get a |
2580 | * partial slab and there is none available then get_partial() |
2581 | * will return NULL. |
2582 | */ |
2583 | if (!n || !n->nr_partial) |
2584 | return NULL; |
2585 | |
2586 | spin_lock_irqsave(&n->list_lock, flags); |
2587 | list_for_each_entry_safe(slab, slab2, &n->partial, slab_list) { |
2588 | if (!pfmemalloc_match(slab, gfpflags: pc->flags)) |
2589 | continue; |
2590 | |
2591 | if (IS_ENABLED(CONFIG_SLUB_TINY) || kmem_cache_debug(s)) { |
2592 | void *object = alloc_single_from_partial(s, n, slab, |
2593 | orig_size: pc->orig_size); |
2594 | if (object) { |
2595 | partial = slab; |
2596 | pc->object = object; |
2597 | break; |
2598 | } |
2599 | continue; |
2600 | } |
2601 | |
2602 | remove_partial(n, slab); |
2603 | |
2604 | if (!partial) { |
2605 | partial = slab; |
2606 | stat(s, si: ALLOC_FROM_PARTIAL); |
2607 | } else { |
2608 | put_cpu_partial(s, slab, drain: 0); |
2609 | stat(s, si: CPU_PARTIAL_NODE); |
2610 | partial_slabs++; |
2611 | } |
2612 | #ifdef CONFIG_SLUB_CPU_PARTIAL |
2613 | if (!kmem_cache_has_cpu_partial(s) |
2614 | || partial_slabs > s->cpu_partial_slabs / 2) |
2615 | break; |
2616 | #else |
2617 | break; |
2618 | #endif |
2619 | |
2620 | } |
2621 | spin_unlock_irqrestore(lock: &n->list_lock, flags); |
2622 | return partial; |
2623 | } |
2624 | |
2625 | /* |
2626 | * Get a slab from somewhere. Search in increasing NUMA distances. |
2627 | */ |
2628 | static struct slab *get_any_partial(struct kmem_cache *s, |
2629 | struct partial_context *pc) |
2630 | { |
2631 | #ifdef CONFIG_NUMA |
2632 | struct zonelist *zonelist; |
2633 | struct zoneref *z; |
2634 | struct zone *zone; |
2635 | enum zone_type highest_zoneidx = gfp_zone(flags: pc->flags); |
2636 | struct slab *slab; |
2637 | unsigned int cpuset_mems_cookie; |
2638 | |
2639 | /* |
2640 | * The defrag ratio allows a configuration of the tradeoffs between |
2641 | * inter node defragmentation and node local allocations. A lower |
2642 | * defrag_ratio increases the tendency to do local allocations |
2643 | * instead of attempting to obtain partial slabs from other nodes. |
2644 | * |
2645 | * If the defrag_ratio is set to 0 then kmalloc() always |
2646 | * returns node local objects. If the ratio is higher then kmalloc() |
2647 | * may return off node objects because partial slabs are obtained |
2648 | * from other nodes and filled up. |
2649 | * |
2650 | * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100 |
2651 | * (which makes defrag_ratio = 1000) then every (well almost) |
2652 | * allocation will first attempt to defrag slab caches on other nodes. |
2653 | * This means scanning over all nodes to look for partial slabs which |
2654 | * may be expensive if we do it every time we are trying to find a slab |
2655 | * with available objects. |
2656 | */ |
2657 | if (!s->remote_node_defrag_ratio || |
2658 | get_cycles() % 1024 > s->remote_node_defrag_ratio) |
2659 | return NULL; |
2660 | |
2661 | do { |
2662 | cpuset_mems_cookie = read_mems_allowed_begin(); |
2663 | zonelist = node_zonelist(nid: mempolicy_slab_node(), flags: pc->flags); |
2664 | for_each_zone_zonelist(zone, z, zonelist, highest_zoneidx) { |
2665 | struct kmem_cache_node *n; |
2666 | |
2667 | n = get_node(s, node: zone_to_nid(zone)); |
2668 | |
2669 | if (n && cpuset_zone_allowed(z: zone, gfp_mask: pc->flags) && |
2670 | n->nr_partial > s->min_partial) { |
2671 | slab = get_partial_node(s, n, pc); |
2672 | if (slab) { |
2673 | /* |
2674 | * Don't check read_mems_allowed_retry() |
2675 | * here - if mems_allowed was updated in |
2676 | * parallel, that was a harmless race |
2677 | * between allocation and the cpuset |
2678 | * update |
2679 | */ |
2680 | return slab; |
2681 | } |
2682 | } |
2683 | } |
2684 | } while (read_mems_allowed_retry(seq: cpuset_mems_cookie)); |
2685 | #endif /* CONFIG_NUMA */ |
2686 | return NULL; |
2687 | } |
2688 | |
2689 | /* |
2690 | * Get a partial slab, lock it and return it. |
2691 | */ |
2692 | static struct slab *get_partial(struct kmem_cache *s, int node, |
2693 | struct partial_context *pc) |
2694 | { |
2695 | struct slab *slab; |
2696 | int searchnode = node; |
2697 | |
2698 | if (node == NUMA_NO_NODE) |
2699 | searchnode = numa_mem_id(); |
2700 | |
2701 | slab = get_partial_node(s, n: get_node(s, node: searchnode), pc); |
2702 | if (slab || node != NUMA_NO_NODE) |
2703 | return slab; |
2704 | |
2705 | return get_any_partial(s, pc); |
2706 | } |
2707 | |
2708 | #ifndef CONFIG_SLUB_TINY |
2709 | |
2710 | #ifdef CONFIG_PREEMPTION |
2711 | /* |
2712 | * Calculate the next globally unique transaction for disambiguation |
2713 | * during cmpxchg. The transactions start with the cpu number and are then |
2714 | * incremented by CONFIG_NR_CPUS. |
2715 | */ |
2716 | #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS) |
2717 | #else |
2718 | /* |
2719 | * No preemption supported therefore also no need to check for |
2720 | * different cpus. |
2721 | */ |
2722 | #define TID_STEP 1 |
2723 | #endif /* CONFIG_PREEMPTION */ |
2724 | |
2725 | static inline unsigned long next_tid(unsigned long tid) |
2726 | { |
2727 | return tid + TID_STEP; |
2728 | } |
2729 | |
2730 | #ifdef SLUB_DEBUG_CMPXCHG |
2731 | static inline unsigned int tid_to_cpu(unsigned long tid) |
2732 | { |
2733 | return tid % TID_STEP; |
2734 | } |
2735 | |
2736 | static inline unsigned long tid_to_event(unsigned long tid) |
2737 | { |
2738 | return tid / TID_STEP; |
2739 | } |
2740 | #endif |
2741 | |
2742 | static inline unsigned int init_tid(int cpu) |
2743 | { |
2744 | return cpu; |
2745 | } |
2746 | |
2747 | static inline void note_cmpxchg_failure(const char *n, |
2748 | const struct kmem_cache *s, unsigned long tid) |
2749 | { |
2750 | #ifdef SLUB_DEBUG_CMPXCHG |
2751 | unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid); |
2752 | |
2753 | pr_info("%s %s: cmpxchg redo " , n, s->name); |
2754 | |
2755 | #ifdef CONFIG_PREEMPTION |
2756 | if (tid_to_cpu(tid) != tid_to_cpu(actual_tid)) |
2757 | pr_warn("due to cpu change %d -> %d\n" , |
2758 | tid_to_cpu(tid), tid_to_cpu(actual_tid)); |
2759 | else |
2760 | #endif |
2761 | if (tid_to_event(tid) != tid_to_event(actual_tid)) |
2762 | pr_warn("due to cpu running other code. Event %ld->%ld\n" , |
2763 | tid_to_event(tid), tid_to_event(actual_tid)); |
2764 | else |
2765 | pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n" , |
2766 | actual_tid, tid, next_tid(tid)); |
2767 | #endif |
2768 | stat(s, CMPXCHG_DOUBLE_CPU_FAIL); |
2769 | } |
2770 | |
2771 | static void init_kmem_cache_cpus(struct kmem_cache *s) |
2772 | { |
2773 | int cpu; |
2774 | struct kmem_cache_cpu *c; |
2775 | |
2776 | for_each_possible_cpu(cpu) { |
2777 | c = per_cpu_ptr(s->cpu_slab, cpu); |
2778 | local_lock_init(&c->lock); |
2779 | c->tid = init_tid(cpu); |
2780 | } |
2781 | } |
2782 | |
2783 | /* |
2784 | * Finishes removing the cpu slab. Merges cpu's freelist with slab's freelist, |
2785 | * unfreezes the slabs and puts it on the proper list. |
2786 | * Assumes the slab has been already safely taken away from kmem_cache_cpu |
2787 | * by the caller. |
2788 | */ |
2789 | static void deactivate_slab(struct kmem_cache *s, struct slab *slab, |
2790 | void *freelist) |
2791 | { |
2792 | struct kmem_cache_node *n = get_node(s, slab_nid(slab)); |
2793 | int free_delta = 0; |
2794 | void *nextfree, *freelist_iter, *freelist_tail; |
2795 | int tail = DEACTIVATE_TO_HEAD; |
2796 | unsigned long flags = 0; |
2797 | struct slab new; |
2798 | struct slab old; |
2799 | |
2800 | if (slab->freelist) { |
2801 | stat(s, DEACTIVATE_REMOTE_FREES); |
2802 | tail = DEACTIVATE_TO_TAIL; |
2803 | } |
2804 | |
2805 | /* |
2806 | * Stage one: Count the objects on cpu's freelist as free_delta and |
2807 | * remember the last object in freelist_tail for later splicing. |
2808 | */ |
2809 | freelist_tail = NULL; |
2810 | freelist_iter = freelist; |
2811 | while (freelist_iter) { |
2812 | nextfree = get_freepointer(s, freelist_iter); |
2813 | |
2814 | /* |
2815 | * If 'nextfree' is invalid, it is possible that the object at |
2816 | * 'freelist_iter' is already corrupted. So isolate all objects |
2817 | * starting at 'freelist_iter' by skipping them. |
2818 | */ |
2819 | if (freelist_corrupted(s, slab, &freelist_iter, nextfree)) |
2820 | break; |
2821 | |
2822 | freelist_tail = freelist_iter; |
2823 | free_delta++; |
2824 | |
2825 | freelist_iter = nextfree; |
2826 | } |
2827 | |
2828 | /* |
2829 | * Stage two: Unfreeze the slab while splicing the per-cpu |
2830 | * freelist to the head of slab's freelist. |
2831 | */ |
2832 | do { |
2833 | old.freelist = READ_ONCE(slab->freelist); |
2834 | old.counters = READ_ONCE(slab->counters); |
2835 | VM_BUG_ON(!old.frozen); |
2836 | |
2837 | /* Determine target state of the slab */ |
2838 | new.counters = old.counters; |
2839 | new.frozen = 0; |
2840 | if (freelist_tail) { |
2841 | new.inuse -= free_delta; |
2842 | set_freepointer(s, freelist_tail, old.freelist); |
2843 | new.freelist = freelist; |
2844 | } else { |
2845 | new.freelist = old.freelist; |
2846 | } |
2847 | } while (!slab_update_freelist(s, slab, |
2848 | old.freelist, old.counters, |
2849 | new.freelist, new.counters, |
2850 | "unfreezing slab" )); |
2851 | |
2852 | /* |
2853 | * Stage three: Manipulate the slab list based on the updated state. |
2854 | */ |
2855 | if (!new.inuse && n->nr_partial >= s->min_partial) { |
2856 | stat(s, DEACTIVATE_EMPTY); |
2857 | discard_slab(s, slab); |
2858 | stat(s, FREE_SLAB); |
2859 | } else if (new.freelist) { |
2860 | spin_lock_irqsave(&n->list_lock, flags); |
2861 | add_partial(n, slab, tail); |
2862 | spin_unlock_irqrestore(&n->list_lock, flags); |
2863 | stat(s, tail); |
2864 | } else { |
2865 | stat(s, DEACTIVATE_FULL); |
2866 | } |
2867 | } |
2868 | |
2869 | #ifdef CONFIG_SLUB_CPU_PARTIAL |
2870 | static void __put_partials(struct kmem_cache *s, struct slab *partial_slab) |
2871 | { |
2872 | struct kmem_cache_node *n = NULL, *n2 = NULL; |
2873 | struct slab *slab, *slab_to_discard = NULL; |
2874 | unsigned long flags = 0; |
2875 | |
2876 | while (partial_slab) { |
2877 | slab = partial_slab; |
2878 | partial_slab = slab->next; |
2879 | |
2880 | n2 = get_node(s, slab_nid(slab)); |
2881 | if (n != n2) { |
2882 | if (n) |
2883 | spin_unlock_irqrestore(&n->list_lock, flags); |
2884 | |
2885 | n = n2; |
2886 | spin_lock_irqsave(&n->list_lock, flags); |
2887 | } |
2888 | |
2889 | if (unlikely(!slab->inuse && n->nr_partial >= s->min_partial)) { |
2890 | slab->next = slab_to_discard; |
2891 | slab_to_discard = slab; |
2892 | } else { |
2893 | add_partial(n, slab, DEACTIVATE_TO_TAIL); |
2894 | stat(s, FREE_ADD_PARTIAL); |
2895 | } |
2896 | } |
2897 | |
2898 | if (n) |
2899 | spin_unlock_irqrestore(&n->list_lock, flags); |
2900 | |
2901 | while (slab_to_discard) { |
2902 | slab = slab_to_discard; |
2903 | slab_to_discard = slab_to_discard->next; |
2904 | |
2905 | stat(s, DEACTIVATE_EMPTY); |
2906 | discard_slab(s, slab); |
2907 | stat(s, FREE_SLAB); |
2908 | } |
2909 | } |
2910 | |
2911 | /* |
2912 | * Put all the cpu partial slabs to the node partial list. |
2913 | */ |
2914 | static void put_partials(struct kmem_cache *s) |
2915 | { |
2916 | struct slab *partial_slab; |
2917 | unsigned long flags; |
2918 | |
2919 | local_lock_irqsave(&s->cpu_slab->lock, flags); |
2920 | partial_slab = this_cpu_read(s->cpu_slab->partial); |
2921 | this_cpu_write(s->cpu_slab->partial, NULL); |
2922 | local_unlock_irqrestore(&s->cpu_slab->lock, flags); |
2923 | |
2924 | if (partial_slab) |
2925 | __put_partials(s, partial_slab); |
2926 | } |
2927 | |
2928 | static void put_partials_cpu(struct kmem_cache *s, |
2929 | struct kmem_cache_cpu *c) |
2930 | { |
2931 | struct slab *partial_slab; |
2932 | |
2933 | partial_slab = slub_percpu_partial(c); |
2934 | c->partial = NULL; |
2935 | |
2936 | if (partial_slab) |
2937 | __put_partials(s, partial_slab); |
2938 | } |
2939 | |
2940 | /* |
2941 | * Put a slab into a partial slab slot if available. |
2942 | * |
2943 | * If we did not find a slot then simply move all the partials to the |
2944 | * per node partial list. |
2945 | */ |
2946 | static void put_cpu_partial(struct kmem_cache *s, struct slab *slab, int drain) |
2947 | { |
2948 | struct slab *oldslab; |
2949 | struct slab *slab_to_put = NULL; |
2950 | unsigned long flags; |
2951 | int slabs = 0; |
2952 | |
2953 | local_lock_irqsave(&s->cpu_slab->lock, flags); |
2954 | |
2955 | oldslab = this_cpu_read(s->cpu_slab->partial); |
2956 | |
2957 | if (oldslab) { |
2958 | if (drain && oldslab->slabs >= s->cpu_partial_slabs) { |
2959 | /* |
2960 | * Partial array is full. Move the existing set to the |
2961 | * per node partial list. Postpone the actual unfreezing |
2962 | * outside of the critical section. |
2963 | */ |
2964 | slab_to_put = oldslab; |
2965 | oldslab = NULL; |
2966 | } else { |
2967 | slabs = oldslab->slabs; |
2968 | } |
2969 | } |
2970 | |
2971 | slabs++; |
2972 | |
2973 | slab->slabs = slabs; |
2974 | slab->next = oldslab; |
2975 | |
2976 | this_cpu_write(s->cpu_slab->partial, slab); |
2977 | |
2978 | local_unlock_irqrestore(&s->cpu_slab->lock, flags); |
2979 | |
2980 | if (slab_to_put) { |
2981 | __put_partials(s, slab_to_put); |
2982 | stat(s, CPU_PARTIAL_DRAIN); |
2983 | } |
2984 | } |
2985 | |
2986 | #else /* CONFIG_SLUB_CPU_PARTIAL */ |
2987 | |
2988 | static inline void put_partials(struct kmem_cache *s) { } |
2989 | static inline void put_partials_cpu(struct kmem_cache *s, |
2990 | struct kmem_cache_cpu *c) { } |
2991 | |
2992 | #endif /* CONFIG_SLUB_CPU_PARTIAL */ |
2993 | |
2994 | static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c) |
2995 | { |
2996 | unsigned long flags; |
2997 | struct slab *slab; |
2998 | void *freelist; |
2999 | |
3000 | local_lock_irqsave(&s->cpu_slab->lock, flags); |
3001 | |
3002 | slab = c->slab; |
3003 | freelist = c->freelist; |
3004 | |
3005 | c->slab = NULL; |
3006 | c->freelist = NULL; |
3007 | c->tid = next_tid(c->tid); |
3008 | |
3009 | local_unlock_irqrestore(&s->cpu_slab->lock, flags); |
3010 | |
3011 | if (slab) { |
3012 | deactivate_slab(s, slab, freelist); |
3013 | stat(s, CPUSLAB_FLUSH); |
3014 | } |
3015 | } |
3016 | |
3017 | static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu) |
3018 | { |
3019 | struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu); |
3020 | void *freelist = c->freelist; |
3021 | struct slab *slab = c->slab; |
3022 | |
3023 | c->slab = NULL; |
3024 | c->freelist = NULL; |
3025 | c->tid = next_tid(c->tid); |
3026 | |
3027 | if (slab) { |
3028 | deactivate_slab(s, slab, freelist); |
3029 | stat(s, CPUSLAB_FLUSH); |
3030 | } |
3031 | |
3032 | put_partials_cpu(s, c); |
3033 | } |
3034 | |
3035 | struct slub_flush_work { |
3036 | struct work_struct work; |
3037 | struct kmem_cache *s; |
3038 | bool skip; |
3039 | }; |
3040 | |
3041 | /* |
3042 | * Flush cpu slab. |
3043 | * |
3044 | * Called from CPU work handler with migration disabled. |
3045 | */ |
3046 | static void flush_cpu_slab(struct work_struct *w) |
3047 | { |
3048 | struct kmem_cache *s; |
3049 | struct kmem_cache_cpu *c; |
3050 | struct slub_flush_work *sfw; |
3051 | |
3052 | sfw = container_of(w, struct slub_flush_work, work); |
3053 | |
3054 | s = sfw->s; |
3055 | c = this_cpu_ptr(s->cpu_slab); |
3056 | |
3057 | if (c->slab) |
3058 | flush_slab(s, c); |
3059 | |
3060 | put_partials(s); |
3061 | } |
3062 | |
3063 | static bool has_cpu_slab(int cpu, struct kmem_cache *s) |
3064 | { |
3065 | struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu); |
3066 | |
3067 | return c->slab || slub_percpu_partial(c); |
3068 | } |
3069 | |
3070 | static DEFINE_MUTEX(flush_lock); |
3071 | static DEFINE_PER_CPU(struct slub_flush_work, slub_flush); |
3072 | |
3073 | static void flush_all_cpus_locked(struct kmem_cache *s) |
3074 | { |
3075 | struct slub_flush_work *sfw; |
3076 | unsigned int cpu; |
3077 | |
3078 | lockdep_assert_cpus_held(); |
3079 | mutex_lock(&flush_lock); |
3080 | |
3081 | for_each_online_cpu(cpu) { |
3082 | sfw = &per_cpu(slub_flush, cpu); |
3083 | if (!has_cpu_slab(cpu, s)) { |
3084 | sfw->skip = true; |
3085 | continue; |
3086 | } |
3087 | INIT_WORK(&sfw->work, flush_cpu_slab); |
3088 | sfw->skip = false; |
3089 | sfw->s = s; |
3090 | queue_work_on(cpu, flushwq, &sfw->work); |
3091 | } |
3092 | |
3093 | for_each_online_cpu(cpu) { |
3094 | sfw = &per_cpu(slub_flush, cpu); |
3095 | if (sfw->skip) |
3096 | continue; |
3097 | flush_work(&sfw->work); |
3098 | } |
3099 | |
3100 | mutex_unlock(&flush_lock); |
3101 | } |
3102 | |
3103 | static void flush_all(struct kmem_cache *s) |
3104 | { |
3105 | cpus_read_lock(); |
3106 | flush_all_cpus_locked(s); |
3107 | cpus_read_unlock(); |
3108 | } |
3109 | |
3110 | /* |
3111 | * Use the cpu notifier to insure that the cpu slabs are flushed when |
3112 | * necessary. |
3113 | */ |
3114 | static int slub_cpu_dead(unsigned int cpu) |
3115 | { |
3116 | struct kmem_cache *s; |
3117 | |
3118 | mutex_lock(&slab_mutex); |
3119 | list_for_each_entry(s, &slab_caches, list) |
3120 | __flush_cpu_slab(s, cpu); |
3121 | mutex_unlock(&slab_mutex); |
3122 | return 0; |
3123 | } |
3124 | |
3125 | #else /* CONFIG_SLUB_TINY */ |
3126 | static inline void flush_all_cpus_locked(struct kmem_cache *s) { } |
3127 | static inline void flush_all(struct kmem_cache *s) { } |
3128 | static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu) { } |
3129 | static inline int slub_cpu_dead(unsigned int cpu) { return 0; } |
3130 | #endif /* CONFIG_SLUB_TINY */ |
3131 | |
3132 | /* |
3133 | * Check if the objects in a per cpu structure fit numa |
3134 | * locality expectations. |
3135 | */ |
3136 | static inline int node_match(struct slab *slab, int node) |
3137 | { |
3138 | #ifdef CONFIG_NUMA |
3139 | if (node != NUMA_NO_NODE && slab_nid(slab) != node) |
3140 | return 0; |
3141 | #endif |
3142 | return 1; |
3143 | } |
3144 | |
3145 | #ifdef CONFIG_SLUB_DEBUG |
3146 | static int count_free(struct slab *slab) |
3147 | { |
3148 | return slab->objects - slab->inuse; |
3149 | } |
3150 | |
3151 | static inline unsigned long node_nr_objs(struct kmem_cache_node *n) |
3152 | { |
3153 | return atomic_long_read(&n->total_objects); |
3154 | } |
3155 | |
3156 | /* Supports checking bulk free of a constructed freelist */ |
3157 | static inline bool free_debug_processing(struct kmem_cache *s, |
3158 | struct slab *slab, void *head, void *tail, int *bulk_cnt, |
3159 | unsigned long addr, depot_stack_handle_t handle) |
3160 | { |
3161 | bool checks_ok = false; |
3162 | void *object = head; |
3163 | int cnt = 0; |
3164 | |
3165 | if (s->flags & SLAB_CONSISTENCY_CHECKS) { |
3166 | if (!check_slab(s, slab)) |
3167 | goto out; |
3168 | } |
3169 | |
3170 | if (slab->inuse < *bulk_cnt) { |
3171 | slab_err(s, slab, "Slab has %d allocated objects but %d are to be freed\n" , |
3172 | slab->inuse, *bulk_cnt); |
3173 | goto out; |
3174 | } |
3175 | |
3176 | next_object: |
3177 | |
3178 | if (++cnt > *bulk_cnt) |
3179 | goto out_cnt; |
3180 | |
3181 | if (s->flags & SLAB_CONSISTENCY_CHECKS) { |
3182 | if (!free_consistency_checks(s, slab, object, addr)) |
3183 | goto out; |
3184 | } |
3185 | |
3186 | if (s->flags & SLAB_STORE_USER) |
3187 | set_track_update(s, object, TRACK_FREE, addr, handle); |
3188 | trace(s, slab, object, 0); |
3189 | /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */ |
3190 | init_object(s, object, SLUB_RED_INACTIVE); |
3191 | |
3192 | /* Reached end of constructed freelist yet? */ |
3193 | if (object != tail) { |
3194 | object = get_freepointer(s, object); |
3195 | goto next_object; |
3196 | } |
3197 | checks_ok = true; |
3198 | |
3199 | out_cnt: |
3200 | if (cnt != *bulk_cnt) { |
3201 | slab_err(s, slab, "Bulk free expected %d objects but found %d\n" , |
3202 | *bulk_cnt, cnt); |
3203 | *bulk_cnt = cnt; |
3204 | } |
3205 | |
3206 | out: |
3207 | |
3208 | if (!checks_ok) |
3209 | slab_fix(s, "Object at 0x%p not freed" , object); |
3210 | |
3211 | return checks_ok; |
3212 | } |
3213 | #endif /* CONFIG_SLUB_DEBUG */ |
3214 | |
3215 | #if defined(CONFIG_SLUB_DEBUG) || defined(SLAB_SUPPORTS_SYSFS) |
3216 | static unsigned long count_partial(struct kmem_cache_node *n, |
3217 | int (*get_count)(struct slab *)) |
3218 | { |
3219 | unsigned long flags; |
3220 | unsigned long x = 0; |
3221 | struct slab *slab; |
3222 | |
3223 | spin_lock_irqsave(&n->list_lock, flags); |
3224 | list_for_each_entry(slab, &n->partial, slab_list) |
3225 | x += get_count(slab); |
3226 | spin_unlock_irqrestore(&n->list_lock, flags); |
3227 | return x; |
3228 | } |
3229 | #endif /* CONFIG_SLUB_DEBUG || SLAB_SUPPORTS_SYSFS */ |
3230 | |
3231 | #ifdef CONFIG_SLUB_DEBUG |
3232 | static noinline void |
3233 | slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid) |
3234 | { |
3235 | static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL, |
3236 | DEFAULT_RATELIMIT_BURST); |
3237 | int node; |
3238 | struct kmem_cache_node *n; |
3239 | |
3240 | if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs)) |
3241 | return; |
3242 | |
3243 | pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n" , |
3244 | nid, gfpflags, &gfpflags); |
3245 | pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n" , |
3246 | s->name, s->object_size, s->size, oo_order(s->oo), |
3247 | oo_order(s->min)); |
3248 | |
3249 | if (oo_order(s->min) > get_order(s->object_size)) |
3250 | pr_warn(" %s debugging increased min order, use slab_debug=O to disable.\n" , |
3251 | s->name); |
3252 | |
3253 | for_each_kmem_cache_node(s, node, n) { |
3254 | unsigned long nr_slabs; |
3255 | unsigned long nr_objs; |
3256 | unsigned long nr_free; |
3257 | |
3258 | nr_free = count_partial(n, count_free); |
3259 | nr_slabs = node_nr_slabs(n); |
3260 | nr_objs = node_nr_objs(n); |
3261 | |
3262 | pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n" , |
3263 | node, nr_slabs, nr_objs, nr_free); |
3264 | } |
3265 | } |
3266 | #else /* CONFIG_SLUB_DEBUG */ |
3267 | static inline void |
3268 | slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid) { } |
3269 | #endif |
3270 | |
3271 | static inline bool pfmemalloc_match(struct slab *slab, gfp_t gfpflags) |
3272 | { |
3273 | if (unlikely(slab_test_pfmemalloc(slab))) |
3274 | return gfp_pfmemalloc_allowed(gfp_mask: gfpflags); |
3275 | |
3276 | return true; |
3277 | } |
3278 | |
3279 | #ifndef CONFIG_SLUB_TINY |
3280 | static inline bool |
3281 | __update_cpu_freelist_fast(struct kmem_cache *s, |
3282 | void *freelist_old, void *freelist_new, |
3283 | unsigned long tid) |
3284 | { |
3285 | freelist_aba_t old = { .freelist = freelist_old, .counter = tid }; |
3286 | freelist_aba_t new = { .freelist = freelist_new, .counter = next_tid(tid) }; |
3287 | |
3288 | return this_cpu_try_cmpxchg_freelist(s->cpu_slab->freelist_tid.full, |
3289 | &old.full, new.full); |
3290 | } |
3291 | |
3292 | /* |
3293 | * Check the slab->freelist and either transfer the freelist to the |
3294 | * per cpu freelist or deactivate the slab. |
3295 | * |
3296 | * The slab is still frozen if the return value is not NULL. |
3297 | * |
3298 | * If this function returns NULL then the slab has been unfrozen. |
3299 | */ |
3300 | static inline void *get_freelist(struct kmem_cache *s, struct slab *slab) |
3301 | { |
3302 | struct slab new; |
3303 | unsigned long counters; |
3304 | void *freelist; |
3305 | |
3306 | lockdep_assert_held(this_cpu_ptr(&s->cpu_slab->lock)); |
3307 | |
3308 | do { |
3309 | freelist = slab->freelist; |
3310 | counters = slab->counters; |
3311 | |
3312 | new.counters = counters; |
3313 | |
3314 | new.inuse = slab->objects; |
3315 | new.frozen = freelist != NULL; |
3316 | |
3317 | } while (!__slab_update_freelist(s, slab, |
3318 | freelist, counters, |
3319 | NULL, new.counters, |
3320 | "get_freelist" )); |
3321 | |
3322 | return freelist; |
3323 | } |
3324 | |
3325 | /* |
3326 | * Freeze the partial slab and return the pointer to the freelist. |
3327 | */ |
3328 | static inline void *freeze_slab(struct kmem_cache *s, struct slab *slab) |
3329 | { |
3330 | struct slab new; |
3331 | unsigned long counters; |
3332 | void *freelist; |
3333 | |
3334 | do { |
3335 | freelist = slab->freelist; |
3336 | counters = slab->counters; |
3337 | |
3338 | new.counters = counters; |
3339 | VM_BUG_ON(new.frozen); |
3340 | |
3341 | new.inuse = slab->objects; |
3342 | new.frozen = 1; |
3343 | |
3344 | } while (!slab_update_freelist(s, slab, |
3345 | freelist, counters, |
3346 | NULL, new.counters, |
3347 | "freeze_slab" )); |
3348 | |
3349 | return freelist; |
3350 | } |
3351 | |
3352 | /* |
3353 | * Slow path. The lockless freelist is empty or we need to perform |
3354 | * debugging duties. |
3355 | * |
3356 | * Processing is still very fast if new objects have been freed to the |
3357 | * regular freelist. In that case we simply take over the regular freelist |
3358 | * as the lockless freelist and zap the regular freelist. |
3359 | * |
3360 | * If that is not working then we fall back to the partial lists. We take the |
3361 | * first element of the freelist as the object to allocate now and move the |
3362 | * rest of the freelist to the lockless freelist. |
3363 | * |
3364 | * And if we were unable to get a new slab from the partial slab lists then |
3365 | * we need to allocate a new slab. This is the slowest path since it involves |
3366 | * a call to the page allocator and the setup of a new slab. |
3367 | * |
3368 | * Version of __slab_alloc to use when we know that preemption is |
3369 | * already disabled (which is the case for bulk allocation). |
3370 | */ |
3371 | static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node, |
3372 | unsigned long addr, struct kmem_cache_cpu *c, unsigned int orig_size) |
3373 | { |
3374 | void *freelist; |
3375 | struct slab *slab; |
3376 | unsigned long flags; |
3377 | struct partial_context pc; |
3378 | |
3379 | stat(s, ALLOC_SLOWPATH); |
3380 | |
3381 | reread_slab: |
3382 | |
3383 | slab = READ_ONCE(c->slab); |
3384 | if (!slab) { |
3385 | /* |
3386 | * if the node is not online or has no normal memory, just |
3387 | * ignore the node constraint |
3388 | */ |
3389 | if (unlikely(node != NUMA_NO_NODE && |
3390 | !node_isset(node, slab_nodes))) |
3391 | node = NUMA_NO_NODE; |
3392 | goto new_slab; |
3393 | } |
3394 | |
3395 | if (unlikely(!node_match(slab, node))) { |
3396 | /* |
3397 | * same as above but node_match() being false already |
3398 | * implies node != NUMA_NO_NODE |
3399 | */ |
3400 | if (!node_isset(node, slab_nodes)) { |
3401 | node = NUMA_NO_NODE; |
3402 | } else { |
3403 | stat(s, ALLOC_NODE_MISMATCH); |
3404 | goto deactivate_slab; |
3405 | } |
3406 | } |
3407 | |
3408 | /* |
3409 | * By rights, we should be searching for a slab page that was |
3410 | * PFMEMALLOC but right now, we are losing the pfmemalloc |
3411 | * information when the page leaves the per-cpu allocator |
3412 | */ |
3413 | if (unlikely(!pfmemalloc_match(slab, gfpflags))) |
3414 | goto deactivate_slab; |
3415 | |
3416 | /* must check again c->slab in case we got preempted and it changed */ |
3417 | local_lock_irqsave(&s->cpu_slab->lock, flags); |
3418 | if (unlikely(slab != c->slab)) { |
3419 | local_unlock_irqrestore(&s->cpu_slab->lock, flags); |
3420 | goto reread_slab; |
3421 | } |
3422 | freelist = c->freelist; |
3423 | if (freelist) |
3424 | goto load_freelist; |
3425 | |
3426 | freelist = get_freelist(s, slab); |
3427 | |
3428 | if (!freelist) { |
3429 | c->slab = NULL; |
3430 | c->tid = next_tid(c->tid); |
3431 | local_unlock_irqrestore(&s->cpu_slab->lock, flags); |
3432 | stat(s, DEACTIVATE_BYPASS); |
3433 | goto new_slab; |
3434 | } |
3435 | |
3436 | stat(s, ALLOC_REFILL); |
3437 | |
3438 | load_freelist: |
3439 | |
3440 | lockdep_assert_held(this_cpu_ptr(&s->cpu_slab->lock)); |
3441 | |
3442 | /* |
3443 | * freelist is pointing to the list of objects to be used. |
3444 | * slab is pointing to the slab from which the objects are obtained. |
3445 | * That slab must be frozen for per cpu allocations to work. |
3446 | */ |
3447 | VM_BUG_ON(!c->slab->frozen); |
3448 | c->freelist = get_freepointer(s, freelist); |
3449 | c->tid = next_tid(c->tid); |
3450 | local_unlock_irqrestore(&s->cpu_slab->lock, flags); |
3451 | return freelist; |
3452 | |
3453 | deactivate_slab: |
3454 | |
3455 | local_lock_irqsave(&s->cpu_slab->lock, flags); |
3456 | if (slab != c->slab) { |
3457 | local_unlock_irqrestore(&s->cpu_slab->lock, flags); |
3458 | goto reread_slab; |
3459 | } |
3460 | freelist = c->freelist; |
3461 | c->slab = NULL; |
3462 | c->freelist = NULL; |
3463 | c->tid = next_tid(c->tid); |
3464 | local_unlock_irqrestore(&s->cpu_slab->lock, flags); |
3465 | deactivate_slab(s, slab, freelist); |
3466 | |
3467 | new_slab: |
3468 | |
3469 | #ifdef CONFIG_SLUB_CPU_PARTIAL |
3470 | while (slub_percpu_partial(c)) { |
3471 | local_lock_irqsave(&s->cpu_slab->lock, flags); |
3472 | if (unlikely(c->slab)) { |
3473 | local_unlock_irqrestore(&s->cpu_slab->lock, flags); |
3474 | goto reread_slab; |
3475 | } |
3476 | if (unlikely(!slub_percpu_partial(c))) { |
3477 | local_unlock_irqrestore(&s->cpu_slab->lock, flags); |
3478 | /* we were preempted and partial list got empty */ |
3479 | goto new_objects; |
3480 | } |
3481 | |
3482 | slab = slub_percpu_partial(c); |
3483 | slub_set_percpu_partial(c, slab); |
3484 | |
3485 | if (likely(node_match(slab, node) && |
3486 | pfmemalloc_match(slab, gfpflags))) { |
3487 | c->slab = slab; |
3488 | freelist = get_freelist(s, slab); |
3489 | VM_BUG_ON(!freelist); |
3490 | stat(s, CPU_PARTIAL_ALLOC); |
3491 | goto load_freelist; |
3492 | } |
3493 | |
3494 | local_unlock_irqrestore(&s->cpu_slab->lock, flags); |
3495 | |
3496 | slab->next = NULL; |
3497 | __put_partials(s, slab); |
3498 | } |
3499 | #endif |
3500 | |
3501 | new_objects: |
3502 | |
3503 | pc.flags = gfpflags; |
3504 | pc.orig_size = orig_size; |
3505 | slab = get_partial(s, node, &pc); |
3506 | if (slab) { |
3507 | if (kmem_cache_debug(s)) { |
3508 | freelist = pc.object; |
3509 | /* |
3510 | * For debug caches here we had to go through |
3511 | * alloc_single_from_partial() so just store the |
3512 | * tracking info and return the object. |
3513 | */ |
3514 | if (s->flags & SLAB_STORE_USER) |
3515 | set_track(s, freelist, TRACK_ALLOC, addr); |
3516 | |
3517 | return freelist; |
3518 | } |
3519 | |
3520 | freelist = freeze_slab(s, slab); |
3521 | goto retry_load_slab; |
3522 | } |
3523 | |
3524 | slub_put_cpu_ptr(s->cpu_slab); |
3525 | slab = new_slab(s, gfpflags, node); |
3526 | c = slub_get_cpu_ptr(s->cpu_slab); |
3527 | |
3528 | if (unlikely(!slab)) { |
3529 | slab_out_of_memory(s, gfpflags, node); |
3530 | return NULL; |
3531 | } |
3532 | |
3533 | stat(s, ALLOC_SLAB); |
3534 | |
3535 | if (kmem_cache_debug(s)) { |
3536 | freelist = alloc_single_from_new_slab(s, slab, orig_size); |
3537 | |
3538 | if (unlikely(!freelist)) |
3539 | goto new_objects; |
3540 | |
3541 | if (s->flags & SLAB_STORE_USER) |
3542 | set_track(s, freelist, TRACK_ALLOC, addr); |
3543 | |
3544 | return freelist; |
3545 | } |
3546 | |
3547 | /* |
3548 | * No other reference to the slab yet so we can |
3549 | * muck around with it freely without cmpxchg |
3550 | */ |
3551 | freelist = slab->freelist; |
3552 | slab->freelist = NULL; |
3553 | slab->inuse = slab->objects; |
3554 | slab->frozen = 1; |
3555 | |
3556 | inc_slabs_node(s, slab_nid(slab), slab->objects); |
3557 | |
3558 | if (unlikely(!pfmemalloc_match(slab, gfpflags))) { |
3559 | /* |
3560 | * For !pfmemalloc_match() case we don't load freelist so that |
3561 | * we don't make further mismatched allocations easier. |
3562 | */ |
3563 | deactivate_slab(s, slab, get_freepointer(s, freelist)); |
3564 | return freelist; |
3565 | } |
3566 | |
3567 | retry_load_slab: |
3568 | |
3569 | local_lock_irqsave(&s->cpu_slab->lock, flags); |
3570 | if (unlikely(c->slab)) { |
3571 | void *flush_freelist = c->freelist; |
3572 | struct slab *flush_slab = c->slab; |
3573 | |
3574 | c->slab = NULL; |
3575 | c->freelist = NULL; |
3576 | c->tid = next_tid(c->tid); |
3577 | |
3578 | local_unlock_irqrestore(&s->cpu_slab->lock, flags); |
3579 | |
3580 | deactivate_slab(s, flush_slab, flush_freelist); |
3581 | |
3582 | stat(s, CPUSLAB_FLUSH); |
3583 | |
3584 | goto retry_load_slab; |
3585 | } |
3586 | c->slab = slab; |
3587 | |
3588 | goto load_freelist; |
3589 | } |
3590 | |
3591 | /* |
3592 | * A wrapper for ___slab_alloc() for contexts where preemption is not yet |
3593 | * disabled. Compensates for possible cpu changes by refetching the per cpu area |
3594 | * pointer. |
3595 | */ |
3596 | static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node, |
3597 | unsigned long addr, struct kmem_cache_cpu *c, unsigned int orig_size) |
3598 | { |
3599 | void *p; |
3600 | |
3601 | #ifdef CONFIG_PREEMPT_COUNT |
3602 | /* |
3603 | * We may have been preempted and rescheduled on a different |
3604 | * cpu before disabling preemption. Need to reload cpu area |
3605 | * pointer. |
3606 | */ |
3607 | c = slub_get_cpu_ptr(s->cpu_slab); |
3608 | #endif |
3609 | |
3610 | p = ___slab_alloc(s, gfpflags, node, addr, c, orig_size); |
3611 | #ifdef CONFIG_PREEMPT_COUNT |
3612 | slub_put_cpu_ptr(s->cpu_slab); |
3613 | #endif |
3614 | return p; |
3615 | } |
3616 | |
3617 | static __always_inline void *__slab_alloc_node(struct kmem_cache *s, |
3618 | gfp_t gfpflags, int node, unsigned long addr, size_t orig_size) |
3619 | { |
3620 | struct kmem_cache_cpu *c; |
3621 | struct slab *slab; |
3622 | unsigned long tid; |
3623 | void *object; |
3624 | |
3625 | redo: |
3626 | /* |
3627 | * Must read kmem_cache cpu data via this cpu ptr. Preemption is |
3628 | * enabled. We may switch back and forth between cpus while |
3629 | * reading from one cpu area. That does not matter as long |
3630 | * as we end up on the original cpu again when doing the cmpxchg. |
3631 | * |
3632 | * We must guarantee that tid and kmem_cache_cpu are retrieved on the |
3633 | * same cpu. We read first the kmem_cache_cpu pointer and use it to read |
3634 | * the tid. If we are preempted and switched to another cpu between the |
3635 | * two reads, it's OK as the two are still associated with the same cpu |
3636 | * and cmpxchg later will validate the cpu. |
3637 | */ |
3638 | c = raw_cpu_ptr(s->cpu_slab); |
3639 | tid = READ_ONCE(c->tid); |
3640 | |
3641 | /* |
3642 | * Irqless object alloc/free algorithm used here depends on sequence |
3643 | * of fetching cpu_slab's data. tid should be fetched before anything |
3644 | * on c to guarantee that object and slab associated with previous tid |
3645 | * won't be used with current tid. If we fetch tid first, object and |
3646 | * slab could be one associated with next tid and our alloc/free |
3647 | * request will be failed. In this case, we will retry. So, no problem. |
3648 | */ |
3649 | barrier(); |
3650 | |
3651 | /* |
3652 | * The transaction ids are globally unique per cpu and per operation on |
3653 | * a per cpu queue. Thus they can be guarantee that the cmpxchg_double |
3654 | * occurs on the right processor and that there was no operation on the |
3655 | * linked list in between. |
3656 | */ |
3657 | |
3658 | object = c->freelist; |
3659 | slab = c->slab; |
3660 | |
3661 | if (!USE_LOCKLESS_FAST_PATH() || |
3662 | unlikely(!object || !slab || !node_match(slab, node))) { |
3663 | object = __slab_alloc(s, gfpflags, node, addr, c, orig_size); |
3664 | } else { |
3665 | void *next_object = get_freepointer_safe(s, object); |
3666 | |
3667 | /* |
3668 | * The cmpxchg will only match if there was no additional |
3669 | * operation and if we are on the right processor. |
3670 | * |
3671 | * The cmpxchg does the following atomically (without lock |
3672 | * semantics!) |
3673 | * 1. Relocate first pointer to the current per cpu area. |
3674 | * 2. Verify that tid and freelist have not been changed |
3675 | * 3. If they were not changed replace tid and freelist |
3676 | * |
3677 | * Since this is without lock semantics the protection is only |
3678 | * against code executing on this cpu *not* from access by |
3679 | * other cpus. |
3680 | */ |
3681 | if (unlikely(!__update_cpu_freelist_fast(s, object, next_object, tid))) { |
3682 | note_cmpxchg_failure("slab_alloc" , s, tid); |
3683 | goto redo; |
3684 | } |
3685 | prefetch_freepointer(s, next_object); |
3686 | stat(s, ALLOC_FASTPATH); |
3687 | } |
3688 | |
3689 | return object; |
3690 | } |
3691 | #else /* CONFIG_SLUB_TINY */ |
3692 | static void *__slab_alloc_node(struct kmem_cache *s, |
3693 | gfp_t gfpflags, int node, unsigned long addr, size_t orig_size) |
3694 | { |
3695 | struct partial_context pc; |
3696 | struct slab *slab; |
3697 | void *object; |
3698 | |
3699 | pc.flags = gfpflags; |
3700 | pc.orig_size = orig_size; |
3701 | slab = get_partial(s, node, pc: &pc); |
3702 | |
3703 | if (slab) |
3704 | return pc.object; |
3705 | |
3706 | slab = new_slab(s, flags: gfpflags, node); |
3707 | if (unlikely(!slab)) { |
3708 | slab_out_of_memory(s, gfpflags, nid: node); |
3709 | return NULL; |
3710 | } |
3711 | |
3712 | object = alloc_single_from_new_slab(s, slab, orig_size); |
3713 | |
3714 | return object; |
3715 | } |
3716 | #endif /* CONFIG_SLUB_TINY */ |
3717 | |
3718 | /* |
3719 | * If the object has been wiped upon free, make sure it's fully initialized by |
3720 | * zeroing out freelist pointer. |
3721 | */ |
3722 | static __always_inline void maybe_wipe_obj_freeptr(struct kmem_cache *s, |
3723 | void *obj) |
3724 | { |
3725 | if (unlikely(slab_want_init_on_free(s)) && obj) |
3726 | memset((void *)((char *)kasan_reset_tag(obj) + s->offset), |
3727 | 0, sizeof(void *)); |
3728 | } |
3729 | |
3730 | noinline int should_failslab(struct kmem_cache *s, gfp_t gfpflags) |
3731 | { |
3732 | if (__should_failslab(s, gfpflags)) |
3733 | return -ENOMEM; |
3734 | return 0; |
3735 | } |
3736 | ALLOW_ERROR_INJECTION(should_failslab, ERRNO); |
3737 | |
3738 | static __fastpath_inline |
3739 | struct kmem_cache *slab_pre_alloc_hook(struct kmem_cache *s, |
3740 | struct list_lru *lru, |
3741 | struct obj_cgroup **objcgp, |
3742 | size_t size, gfp_t flags) |
3743 | { |
3744 | flags &= gfp_allowed_mask; |
3745 | |
3746 | might_alloc(gfp_mask: flags); |
3747 | |
3748 | if (unlikely(should_failslab(s, flags))) |
3749 | return NULL; |
3750 | |
3751 | if (unlikely(!memcg_slab_pre_alloc_hook(s, lru, objcgp, size, flags))) |
3752 | return NULL; |
3753 | |
3754 | return s; |
3755 | } |
3756 | |
3757 | static __fastpath_inline |
3758 | void slab_post_alloc_hook(struct kmem_cache *s, struct obj_cgroup *objcg, |
3759 | gfp_t flags, size_t size, void **p, bool init, |
3760 | unsigned int orig_size) |
3761 | { |
3762 | unsigned int zero_size = s->object_size; |
3763 | bool kasan_init = init; |
3764 | size_t i; |
3765 | gfp_t init_flags = flags & gfp_allowed_mask; |
3766 | |
3767 | /* |
3768 | * For kmalloc object, the allocated memory size(object_size) is likely |
3769 | * larger than the requested size(orig_size). If redzone check is |
3770 | * enabled for the extra space, don't zero it, as it will be redzoned |
3771 | * soon. The redzone operation for this extra space could be seen as a |
3772 | * replacement of current poisoning under certain debug option, and |
3773 | * won't break other sanity checks. |
3774 | */ |
3775 | if (kmem_cache_debug_flags(s, SLAB_STORE_USER | SLAB_RED_ZONE) && |
3776 | (s->flags & SLAB_KMALLOC)) |
3777 | zero_size = orig_size; |
3778 | |
3779 | /* |
3780 | * When slab_debug is enabled, avoid memory initialization integrated |
3781 | * into KASAN and instead zero out the memory via the memset below with |
3782 | * the proper size. Otherwise, KASAN might overwrite SLUB redzones and |
3783 | * cause false-positive reports. This does not lead to a performance |
3784 | * penalty on production builds, as slab_debug is not intended to be |
3785 | * enabled there. |
3786 | */ |
3787 | if (__slub_debug_enabled()) |
3788 | kasan_init = false; |
3789 | |
3790 | /* |
3791 | * As memory initialization might be integrated into KASAN, |
3792 | * kasan_slab_alloc and initialization memset must be |
3793 | * kept together to avoid discrepancies in behavior. |
3794 | * |
3795 | * As p[i] might get tagged, memset and kmemleak hook come after KASAN. |
3796 | */ |
3797 | for (i = 0; i < size; i++) { |
3798 | p[i] = kasan_slab_alloc(s, object: p[i], flags: init_flags, init: kasan_init); |
3799 | if (p[i] && init && (!kasan_init || |
3800 | !kasan_has_integrated_init())) |
3801 | memset(p[i], 0, zero_size); |
3802 | kmemleak_alloc_recursive(ptr: p[i], size: s->object_size, min_count: 1, |
3803 | flags: s->flags, gfp: init_flags); |
3804 | kmsan_slab_alloc(s, object: p[i], flags: init_flags); |
3805 | } |
3806 | |
3807 | memcg_slab_post_alloc_hook(s, objcg, flags, size, p); |
3808 | } |
3809 | |
3810 | /* |
3811 | * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc) |
3812 | * have the fastpath folded into their functions. So no function call |
3813 | * overhead for requests that can be satisfied on the fastpath. |
3814 | * |
3815 | * The fastpath works by first checking if the lockless freelist can be used. |
3816 | * If not then __slab_alloc is called for slow processing. |
3817 | * |
3818 | * Otherwise we can simply pick the next object from the lockless free list. |
3819 | */ |
3820 | static __fastpath_inline void *slab_alloc_node(struct kmem_cache *s, struct list_lru *lru, |
3821 | gfp_t gfpflags, int node, unsigned long addr, size_t orig_size) |
3822 | { |
3823 | void *object; |
3824 | struct obj_cgroup *objcg = NULL; |
3825 | bool init = false; |
3826 | |
3827 | s = slab_pre_alloc_hook(s, lru, objcgp: &objcg, size: 1, flags: gfpflags); |
3828 | if (unlikely(!s)) |
3829 | return NULL; |
3830 | |
3831 | object = kfence_alloc(s, size: orig_size, flags: gfpflags); |
3832 | if (unlikely(object)) |
3833 | goto out; |
3834 | |
3835 | object = __slab_alloc_node(s, gfpflags, node, addr, orig_size); |
3836 | |
3837 | maybe_wipe_obj_freeptr(s, obj: object); |
3838 | init = slab_want_init_on_alloc(flags: gfpflags, c: s); |
3839 | |
3840 | out: |
3841 | /* |
3842 | * When init equals 'true', like for kzalloc() family, only |
3843 | * @orig_size bytes might be zeroed instead of s->object_size |
3844 | */ |
3845 | slab_post_alloc_hook(s, objcg, flags: gfpflags, size: 1, p: &object, init, orig_size); |
3846 | |
3847 | return object; |
3848 | } |
3849 | |
3850 | void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags) |
3851 | { |
3852 | void *ret = slab_alloc_node(s, NULL, gfpflags, NUMA_NO_NODE, _RET_IP_, |
3853 | orig_size: s->object_size); |
3854 | |
3855 | trace_kmem_cache_alloc(_RET_IP_, ptr: ret, s, gfp_flags: gfpflags, NUMA_NO_NODE); |
3856 | |
3857 | return ret; |
3858 | } |
3859 | EXPORT_SYMBOL(kmem_cache_alloc); |
3860 | |
3861 | void *kmem_cache_alloc_lru(struct kmem_cache *s, struct list_lru *lru, |
3862 | gfp_t gfpflags) |
3863 | { |
3864 | void *ret = slab_alloc_node(s, lru, gfpflags, NUMA_NO_NODE, _RET_IP_, |
3865 | orig_size: s->object_size); |
3866 | |
3867 | trace_kmem_cache_alloc(_RET_IP_, ptr: ret, s, gfp_flags: gfpflags, NUMA_NO_NODE); |
3868 | |
3869 | return ret; |
3870 | } |
3871 | EXPORT_SYMBOL(kmem_cache_alloc_lru); |
3872 | |
3873 | /** |
3874 | * kmem_cache_alloc_node - Allocate an object on the specified node |
3875 | * @s: The cache to allocate from. |
3876 | * @gfpflags: See kmalloc(). |
3877 | * @node: node number of the target node. |
3878 | * |
3879 | * Identical to kmem_cache_alloc but it will allocate memory on the given |
3880 | * node, which can improve the performance for cpu bound structures. |
3881 | * |
3882 | * Fallback to other node is possible if __GFP_THISNODE is not set. |
3883 | * |
3884 | * Return: pointer to the new object or %NULL in case of error |
3885 | */ |
3886 | void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node) |
3887 | { |
3888 | void *ret = slab_alloc_node(s, NULL, gfpflags, node, _RET_IP_, orig_size: s->object_size); |
3889 | |
3890 | trace_kmem_cache_alloc(_RET_IP_, ptr: ret, s, gfp_flags: gfpflags, node); |
3891 | |
3892 | return ret; |
3893 | } |
3894 | EXPORT_SYMBOL(kmem_cache_alloc_node); |
3895 | |
3896 | /* |
3897 | * To avoid unnecessary overhead, we pass through large allocation requests |
3898 | * directly to the page allocator. We use __GFP_COMP, because we will need to |
3899 | * know the allocation order to free the pages properly in kfree. |
3900 | */ |
3901 | static void *__kmalloc_large_node(size_t size, gfp_t flags, int node) |
3902 | { |
3903 | struct folio *folio; |
3904 | void *ptr = NULL; |
3905 | unsigned int order = get_order(size); |
3906 | |
3907 | if (unlikely(flags & GFP_SLAB_BUG_MASK)) |
3908 | flags = kmalloc_fix_flags(flags); |
3909 | |
3910 | flags |= __GFP_COMP; |
3911 | folio = (struct folio *)alloc_pages_node(nid: node, gfp_mask: flags, order); |
3912 | if (folio) { |
3913 | ptr = folio_address(folio); |
3914 | lruvec_stat_mod_folio(folio, idx: NR_SLAB_UNRECLAIMABLE_B, |
3915 | PAGE_SIZE << order); |
3916 | } |
3917 | |
3918 | ptr = kasan_kmalloc_large(ptr, size, flags); |
3919 | /* As ptr might get tagged, call kmemleak hook after KASAN. */ |
3920 | kmemleak_alloc(ptr, size, min_count: 1, gfp: flags); |
3921 | kmsan_kmalloc_large(ptr, size, flags); |
3922 | |
3923 | return ptr; |
3924 | } |
3925 | |
3926 | void *kmalloc_large(size_t size, gfp_t flags) |
3927 | { |
3928 | void *ret = __kmalloc_large_node(size, flags, NUMA_NO_NODE); |
3929 | |
3930 | trace_kmalloc(_RET_IP_, ptr: ret, bytes_req: size, PAGE_SIZE << get_order(size), |
3931 | gfp_flags: flags, NUMA_NO_NODE); |
3932 | return ret; |
3933 | } |
3934 | EXPORT_SYMBOL(kmalloc_large); |
3935 | |
3936 | void *kmalloc_large_node(size_t size, gfp_t flags, int node) |
3937 | { |
3938 | void *ret = __kmalloc_large_node(size, flags, node); |
3939 | |
3940 | trace_kmalloc(_RET_IP_, ptr: ret, bytes_req: size, PAGE_SIZE << get_order(size), |
3941 | gfp_flags: flags, node); |
3942 | return ret; |
3943 | } |
3944 | EXPORT_SYMBOL(kmalloc_large_node); |
3945 | |
3946 | static __always_inline |
3947 | void *__do_kmalloc_node(size_t size, gfp_t flags, int node, |
3948 | unsigned long caller) |
3949 | { |
3950 | struct kmem_cache *s; |
3951 | void *ret; |
3952 | |
3953 | if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) { |
3954 | ret = __kmalloc_large_node(size, flags, node); |
3955 | trace_kmalloc(call_site: caller, ptr: ret, bytes_req: size, |
3956 | PAGE_SIZE << get_order(size), gfp_flags: flags, node); |
3957 | return ret; |
3958 | } |
3959 | |
3960 | if (unlikely(!size)) |
3961 | return ZERO_SIZE_PTR; |
3962 | |
3963 | s = kmalloc_slab(size, flags, caller); |
3964 | |
3965 | ret = slab_alloc_node(s, NULL, gfpflags: flags, node, addr: caller, orig_size: size); |
3966 | ret = kasan_kmalloc(s, object: ret, size, flags); |
3967 | trace_kmalloc(call_site: caller, ptr: ret, bytes_req: size, bytes_alloc: s->size, gfp_flags: flags, node); |
3968 | return ret; |
3969 | } |
3970 | |
3971 | void *__kmalloc_node(size_t size, gfp_t flags, int node) |
3972 | { |
3973 | return __do_kmalloc_node(size, flags, node, _RET_IP_); |
3974 | } |
3975 | EXPORT_SYMBOL(__kmalloc_node); |
3976 | |
3977 | void *__kmalloc(size_t size, gfp_t flags) |
3978 | { |
3979 | return __do_kmalloc_node(size, flags, NUMA_NO_NODE, _RET_IP_); |
3980 | } |
3981 | EXPORT_SYMBOL(__kmalloc); |
3982 | |
3983 | void *__kmalloc_node_track_caller(size_t size, gfp_t flags, |
3984 | int node, unsigned long caller) |
3985 | { |
3986 | return __do_kmalloc_node(size, flags, node, caller); |
3987 | } |
3988 | EXPORT_SYMBOL(__kmalloc_node_track_caller); |
3989 | |
3990 | void *kmalloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size) |
3991 | { |
3992 | void *ret = slab_alloc_node(s, NULL, gfpflags, NUMA_NO_NODE, |
3993 | _RET_IP_, orig_size: size); |
3994 | |
3995 | trace_kmalloc(_RET_IP_, ptr: ret, bytes_req: size, bytes_alloc: s->size, gfp_flags: gfpflags, NUMA_NO_NODE); |
3996 | |
3997 | ret = kasan_kmalloc(s, object: ret, size, flags: gfpflags); |
3998 | return ret; |
3999 | } |
4000 | EXPORT_SYMBOL(kmalloc_trace); |
4001 | |
4002 | void *kmalloc_node_trace(struct kmem_cache *s, gfp_t gfpflags, |
4003 | int node, size_t size) |
4004 | { |
4005 | void *ret = slab_alloc_node(s, NULL, gfpflags, node, _RET_IP_, orig_size: size); |
4006 | |
4007 | trace_kmalloc(_RET_IP_, ptr: ret, bytes_req: size, bytes_alloc: s->size, gfp_flags: gfpflags, node); |
4008 | |
4009 | ret = kasan_kmalloc(s, object: ret, size, flags: gfpflags); |
4010 | return ret; |
4011 | } |
4012 | EXPORT_SYMBOL(kmalloc_node_trace); |
4013 | |
4014 | static noinline void free_to_partial_list( |
4015 | struct kmem_cache *s, struct slab *slab, |
4016 | void *head, void *tail, int bulk_cnt, |
4017 | unsigned long addr) |
4018 | { |
4019 | struct kmem_cache_node *n = get_node(s, node: slab_nid(slab)); |
4020 | struct slab *slab_free = NULL; |
4021 | int cnt = bulk_cnt; |
4022 | unsigned long flags; |
4023 | depot_stack_handle_t handle = 0; |
4024 | |
4025 | if (s->flags & SLAB_STORE_USER) |
4026 | handle = set_track_prepare(); |
4027 | |
4028 | spin_lock_irqsave(&n->list_lock, flags); |
4029 | |
4030 | if (free_debug_processing(s, slab, head, tail, bulk_cnt: &cnt, addr, handle)) { |
4031 | void *prior = slab->freelist; |
4032 | |
4033 | /* Perform the actual freeing while we still hold the locks */ |
4034 | slab->inuse -= cnt; |
4035 | set_freepointer(s, object: tail, fp: prior); |
4036 | slab->freelist = head; |
4037 | |
4038 | /* |
4039 | * If the slab is empty, and node's partial list is full, |
4040 | * it should be discarded anyway no matter it's on full or |
4041 | * partial list. |
4042 | */ |
4043 | if (slab->inuse == 0 && n->nr_partial >= s->min_partial) |
4044 | slab_free = slab; |
4045 | |
4046 | if (!prior) { |
4047 | /* was on full list */ |
4048 | remove_full(s, n, slab); |
4049 | if (!slab_free) { |
4050 | add_partial(n, slab, tail: DEACTIVATE_TO_TAIL); |
4051 | stat(s, si: FREE_ADD_PARTIAL); |
4052 | } |
4053 | } else if (slab_free) { |
4054 | remove_partial(n, slab); |
4055 | stat(s, si: FREE_REMOVE_PARTIAL); |
4056 | } |
4057 | } |
4058 | |
4059 | if (slab_free) { |
4060 | /* |
4061 | * Update the counters while still holding n->list_lock to |
4062 | * prevent spurious validation warnings |
4063 | */ |
4064 | dec_slabs_node(s, node: slab_nid(slab: slab_free), objects: slab_free->objects); |
4065 | } |
4066 | |
4067 | spin_unlock_irqrestore(lock: &n->list_lock, flags); |
4068 | |
4069 | if (slab_free) { |
4070 | stat(s, si: FREE_SLAB); |
4071 | free_slab(s, slab: slab_free); |
4072 | } |
4073 | } |
4074 | |
4075 | /* |
4076 | * Slow path handling. This may still be called frequently since objects |
4077 | * have a longer lifetime than the cpu slabs in most processing loads. |
4078 | * |
4079 | * So we still attempt to reduce cache line usage. Just take the slab |
4080 | * lock and free the item. If there is no additional partial slab |
4081 | * handling required then we can return immediately. |
4082 | */ |
4083 | static void __slab_free(struct kmem_cache *s, struct slab *slab, |
4084 | void *head, void *tail, int cnt, |
4085 | unsigned long addr) |
4086 | |
4087 | { |
4088 | void *prior; |
4089 | int was_frozen; |
4090 | struct slab new; |
4091 | unsigned long counters; |
4092 | struct kmem_cache_node *n = NULL; |
4093 | unsigned long flags; |
4094 | bool on_node_partial; |
4095 | |
4096 | stat(s, si: FREE_SLOWPATH); |
4097 | |
4098 | if (IS_ENABLED(CONFIG_SLUB_TINY) || kmem_cache_debug(s)) { |
4099 | free_to_partial_list(s, slab, head, tail, bulk_cnt: cnt, addr); |
4100 | return; |
4101 | } |
4102 | |
4103 | do { |
4104 | if (unlikely(n)) { |
4105 | spin_unlock_irqrestore(lock: &n->list_lock, flags); |
4106 | n = NULL; |
4107 | } |
4108 | prior = slab->freelist; |
4109 | counters = slab->counters; |
4110 | set_freepointer(s, object: tail, fp: prior); |
4111 | new.counters = counters; |
4112 | was_frozen = new.frozen; |
4113 | new.inuse -= cnt; |
4114 | if ((!new.inuse || !prior) && !was_frozen) { |
4115 | /* Needs to be taken off a list */ |
4116 | if (!kmem_cache_has_cpu_partial(s) || prior) { |
4117 | |
4118 | n = get_node(s, node: slab_nid(slab)); |
4119 | /* |
4120 | * Speculatively acquire the list_lock. |
4121 | * If the cmpxchg does not succeed then we may |
4122 | * drop the list_lock without any processing. |
4123 | * |
4124 | * Otherwise the list_lock will synchronize with |
4125 | * other processors updating the list of slabs. |
4126 | */ |
4127 | spin_lock_irqsave(&n->list_lock, flags); |
4128 | |
4129 | on_node_partial = slab_test_node_partial(slab); |
4130 | } |
4131 | } |
4132 | |
4133 | } while (!slab_update_freelist(s, slab, |
4134 | freelist_old: prior, counters_old: counters, |
4135 | freelist_new: head, counters_new: new.counters, |
4136 | n: "__slab_free" )); |
4137 | |
4138 | if (likely(!n)) { |
4139 | |
4140 | if (likely(was_frozen)) { |
4141 | /* |
4142 | * The list lock was not taken therefore no list |
4143 | * activity can be necessary. |
4144 | */ |
4145 | stat(s, si: FREE_FROZEN); |
4146 | } else if (kmem_cache_has_cpu_partial(s) && !prior) { |
4147 | /* |
4148 | * If we started with a full slab then put it onto the |
4149 | * per cpu partial list. |
4150 | */ |
4151 | put_cpu_partial(s, slab, drain: 1); |
4152 | stat(s, si: CPU_PARTIAL_FREE); |
4153 | } |
4154 | |
4155 | return; |
4156 | } |
4157 | |
4158 | /* |
4159 | * This slab was partially empty but not on the per-node partial list, |
4160 | * in which case we shouldn't manipulate its list, just return. |
4161 | */ |
4162 | if (prior && !on_node_partial) { |
4163 | spin_unlock_irqrestore(lock: &n->list_lock, flags); |
4164 | return; |
4165 | } |
4166 | |
4167 | if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) |
4168 | goto slab_empty; |
4169 | |
4170 | /* |
4171 | * Objects left in the slab. If it was not on the partial list before |
4172 | * then add it. |
4173 | */ |
4174 | if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) { |
4175 | add_partial(n, slab, tail: DEACTIVATE_TO_TAIL); |
4176 | stat(s, si: FREE_ADD_PARTIAL); |
4177 | } |
4178 | spin_unlock_irqrestore(lock: &n->list_lock, flags); |
4179 | return; |
4180 | |
4181 | slab_empty: |
4182 | if (prior) { |
4183 | /* |
4184 | * Slab on the partial list. |
4185 | */ |
4186 | remove_partial(n, slab); |
4187 | stat(s, si: FREE_REMOVE_PARTIAL); |
4188 | } |
4189 | |
4190 | spin_unlock_irqrestore(lock: &n->list_lock, flags); |
4191 | stat(s, si: FREE_SLAB); |
4192 | discard_slab(s, slab); |
4193 | } |
4194 | |
4195 | #ifndef CONFIG_SLUB_TINY |
4196 | /* |
4197 | * Fastpath with forced inlining to produce a kfree and kmem_cache_free that |
4198 | * can perform fastpath freeing without additional function calls. |
4199 | * |
4200 | * The fastpath is only possible if we are freeing to the current cpu slab |
4201 | * of this processor. This typically the case if we have just allocated |
4202 | * the item before. |
4203 | * |
4204 | * If fastpath is not possible then fall back to __slab_free where we deal |
4205 | * with all sorts of special processing. |
4206 | * |
4207 | * Bulk free of a freelist with several objects (all pointing to the |
4208 | * same slab) possible by specifying head and tail ptr, plus objects |
4209 | * count (cnt). Bulk free indicated by tail pointer being set. |
4210 | */ |
4211 | static __always_inline void do_slab_free(struct kmem_cache *s, |
4212 | struct slab *slab, void *head, void *tail, |
4213 | int cnt, unsigned long addr) |
4214 | { |
4215 | struct kmem_cache_cpu *c; |
4216 | unsigned long tid; |
4217 | void **freelist; |
4218 | |
4219 | redo: |
4220 | /* |
4221 | * Determine the currently cpus per cpu slab. |
4222 | * The cpu may change afterward. However that does not matter since |
4223 | * data is retrieved via this pointer. If we are on the same cpu |
4224 | * during the cmpxchg then the free will succeed. |
4225 | */ |
4226 | c = raw_cpu_ptr(s->cpu_slab); |
4227 | tid = READ_ONCE(c->tid); |
4228 | |
4229 | /* Same with comment on barrier() in slab_alloc_node() */ |
4230 | barrier(); |
4231 | |
4232 | if (unlikely(slab != c->slab)) { |
4233 | __slab_free(s, slab, head, tail, cnt, addr); |
4234 | return; |
4235 | } |
4236 | |
4237 | if (USE_LOCKLESS_FAST_PATH()) { |
4238 | freelist = READ_ONCE(c->freelist); |
4239 | |
4240 | set_freepointer(s, tail, freelist); |
4241 | |
4242 | if (unlikely(!__update_cpu_freelist_fast(s, freelist, head, tid))) { |
4243 | note_cmpxchg_failure("slab_free" , s, tid); |
4244 | goto redo; |
4245 | } |
4246 | } else { |
4247 | /* Update the free list under the local lock */ |
4248 | local_lock(&s->cpu_slab->lock); |
4249 | c = this_cpu_ptr(s->cpu_slab); |
4250 | if (unlikely(slab != c->slab)) { |
4251 | local_unlock(&s->cpu_slab->lock); |
4252 | goto redo; |
4253 | } |
4254 | tid = c->tid; |
4255 | freelist = c->freelist; |
4256 | |
4257 | set_freepointer(s, tail, freelist); |
4258 | c->freelist = head; |
4259 | c->tid = next_tid(tid); |
4260 | |
4261 | local_unlock(&s->cpu_slab->lock); |
4262 | } |
4263 | stat_add(s, FREE_FASTPATH, cnt); |
4264 | } |
4265 | #else /* CONFIG_SLUB_TINY */ |
4266 | static void do_slab_free(struct kmem_cache *s, |
4267 | struct slab *slab, void *head, void *tail, |
4268 | int cnt, unsigned long addr) |
4269 | { |
4270 | __slab_free(s, slab, head, tail, cnt, addr); |
4271 | } |
4272 | #endif /* CONFIG_SLUB_TINY */ |
4273 | |
4274 | static __fastpath_inline |
4275 | void slab_free(struct kmem_cache *s, struct slab *slab, void *object, |
4276 | unsigned long addr) |
4277 | { |
4278 | memcg_slab_free_hook(s, slab, p: &object, objects: 1); |
4279 | |
4280 | if (likely(slab_free_hook(s, object, slab_want_init_on_free(s)))) |
4281 | do_slab_free(s, slab, head: object, tail: object, cnt: 1, addr); |
4282 | } |
4283 | |
4284 | static __fastpath_inline |
4285 | void slab_free_bulk(struct kmem_cache *s, struct slab *slab, void *head, |
4286 | void *tail, void **p, int cnt, unsigned long addr) |
4287 | { |
4288 | memcg_slab_free_hook(s, slab, p, objects: cnt); |
4289 | /* |
4290 | * With KASAN enabled slab_free_freelist_hook modifies the freelist |
4291 | * to remove objects, whose reuse must be delayed. |
4292 | */ |
4293 | if (likely(slab_free_freelist_hook(s, &head, &tail, &cnt))) |
4294 | do_slab_free(s, slab, head, tail, cnt, addr); |
4295 | } |
4296 | |
4297 | #ifdef CONFIG_KASAN_GENERIC |
4298 | void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr) |
4299 | { |
4300 | do_slab_free(cache, virt_to_slab(x), x, x, 1, addr); |
4301 | } |
4302 | #endif |
4303 | |
4304 | static inline struct kmem_cache *virt_to_cache(const void *obj) |
4305 | { |
4306 | struct slab *slab; |
4307 | |
4308 | slab = virt_to_slab(addr: obj); |
4309 | if (WARN_ONCE(!slab, "%s: Object is not a Slab page!\n" , __func__)) |
4310 | return NULL; |
4311 | return slab->slab_cache; |
4312 | } |
4313 | |
4314 | static inline struct kmem_cache *cache_from_obj(struct kmem_cache *s, void *x) |
4315 | { |
4316 | struct kmem_cache *cachep; |
4317 | |
4318 | if (!IS_ENABLED(CONFIG_SLAB_FREELIST_HARDENED) && |
4319 | !kmem_cache_debug_flags(s, SLAB_CONSISTENCY_CHECKS)) |
4320 | return s; |
4321 | |
4322 | cachep = virt_to_cache(obj: x); |
4323 | if (WARN(cachep && cachep != s, |
4324 | "%s: Wrong slab cache. %s but object is from %s\n" , |
4325 | __func__, s->name, cachep->name)) |
4326 | print_tracking(s: cachep, object: x); |
4327 | return cachep; |
4328 | } |
4329 | |
4330 | /** |
4331 | * kmem_cache_free - Deallocate an object |
4332 | * @s: The cache the allocation was from. |
4333 | * @x: The previously allocated object. |
4334 | * |
4335 | * Free an object which was previously allocated from this |
4336 | * cache. |
4337 | */ |
4338 | void kmem_cache_free(struct kmem_cache *s, void *x) |
4339 | { |
4340 | s = cache_from_obj(s, x); |
4341 | if (!s) |
4342 | return; |
4343 | trace_kmem_cache_free(_RET_IP_, ptr: x, s); |
4344 | slab_free(s, slab: virt_to_slab(addr: x), object: x, _RET_IP_); |
4345 | } |
4346 | EXPORT_SYMBOL(kmem_cache_free); |
4347 | |
4348 | static void free_large_kmalloc(struct folio *folio, void *object) |
4349 | { |
4350 | unsigned int order = folio_order(folio); |
4351 | |
4352 | if (WARN_ON_ONCE(order == 0)) |
4353 | pr_warn_once("object pointer: 0x%p\n" , object); |
4354 | |
4355 | kmemleak_free(ptr: object); |
4356 | kasan_kfree_large(ptr: object); |
4357 | kmsan_kfree_large(ptr: object); |
4358 | |
4359 | lruvec_stat_mod_folio(folio, idx: NR_SLAB_UNRECLAIMABLE_B, |
4360 | val: -(PAGE_SIZE << order)); |
4361 | folio_put(folio); |
4362 | } |
4363 | |
4364 | /** |
4365 | * kfree - free previously allocated memory |
4366 | * @object: pointer returned by kmalloc() or kmem_cache_alloc() |
4367 | * |
4368 | * If @object is NULL, no operation is performed. |
4369 | */ |
4370 | void kfree(const void *object) |
4371 | { |
4372 | struct folio *folio; |
4373 | struct slab *slab; |
4374 | struct kmem_cache *s; |
4375 | void *x = (void *)object; |
4376 | |
4377 | trace_kfree(_RET_IP_, ptr: object); |
4378 | |
4379 | if (unlikely(ZERO_OR_NULL_PTR(object))) |
4380 | return; |
4381 | |
4382 | folio = virt_to_folio(x: object); |
4383 | if (unlikely(!folio_test_slab(folio))) { |
4384 | free_large_kmalloc(folio, object: (void *)object); |
4385 | return; |
4386 | } |
4387 | |
4388 | slab = folio_slab(folio); |
4389 | s = slab->slab_cache; |
4390 | slab_free(s, slab, object: x, _RET_IP_); |
4391 | } |
4392 | EXPORT_SYMBOL(kfree); |
4393 | |
4394 | struct detached_freelist { |
4395 | struct slab *slab; |
4396 | void *tail; |
4397 | void *freelist; |
4398 | int cnt; |
4399 | struct kmem_cache *s; |
4400 | }; |
4401 | |
4402 | /* |
4403 | * This function progressively scans the array with free objects (with |
4404 | * a limited look ahead) and extract objects belonging to the same |
4405 | * slab. It builds a detached freelist directly within the given |
4406 | * slab/objects. This can happen without any need for |
4407 | * synchronization, because the objects are owned by running process. |
4408 | * The freelist is build up as a single linked list in the objects. |
4409 | * The idea is, that this detached freelist can then be bulk |
4410 | * transferred to the real freelist(s), but only requiring a single |
4411 | * synchronization primitive. Look ahead in the array is limited due |
4412 | * to performance reasons. |
4413 | */ |
4414 | static inline |
4415 | int build_detached_freelist(struct kmem_cache *s, size_t size, |
4416 | void **p, struct detached_freelist *df) |
4417 | { |
4418 | int lookahead = 3; |
4419 | void *object; |
4420 | struct folio *folio; |
4421 | size_t same; |
4422 | |
4423 | object = p[--size]; |
4424 | folio = virt_to_folio(x: object); |
4425 | if (!s) { |
4426 | /* Handle kalloc'ed objects */ |
4427 | if (unlikely(!folio_test_slab(folio))) { |
4428 | free_large_kmalloc(folio, object); |
4429 | df->slab = NULL; |
4430 | return size; |
4431 | } |
4432 | /* Derive kmem_cache from object */ |
4433 | df->slab = folio_slab(folio); |
4434 | df->s = df->slab->slab_cache; |
4435 | } else { |
4436 | df->slab = folio_slab(folio); |
4437 | df->s = cache_from_obj(s, x: object); /* Support for memcg */ |
4438 | } |
4439 | |
4440 | /* Start new detached freelist */ |
4441 | df->tail = object; |
4442 | df->freelist = object; |
4443 | df->cnt = 1; |
4444 | |
4445 | if (is_kfence_address(addr: object)) |
4446 | return size; |
4447 | |
4448 | set_freepointer(s: df->s, object, NULL); |
4449 | |
4450 | same = size; |
4451 | while (size) { |
4452 | object = p[--size]; |
4453 | /* df->slab is always set at this point */ |
4454 | if (df->slab == virt_to_slab(addr: object)) { |
4455 | /* Opportunity build freelist */ |
4456 | set_freepointer(s: df->s, object, fp: df->freelist); |
4457 | df->freelist = object; |
4458 | df->cnt++; |
4459 | same--; |
4460 | if (size != same) |
4461 | swap(p[size], p[same]); |
4462 | continue; |
4463 | } |
4464 | |
4465 | /* Limit look ahead search */ |
4466 | if (!--lookahead) |
4467 | break; |
4468 | } |
4469 | |
4470 | return same; |
4471 | } |
4472 | |
4473 | /* |
4474 | * Internal bulk free of objects that were not initialised by the post alloc |
4475 | * hooks and thus should not be processed by the free hooks |
4476 | */ |
4477 | static void __kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p) |
4478 | { |
4479 | if (!size) |
4480 | return; |
4481 | |
4482 | do { |
4483 | struct detached_freelist df; |
4484 | |
4485 | size = build_detached_freelist(s, size, p, df: &df); |
4486 | if (!df.slab) |
4487 | continue; |
4488 | |
4489 | do_slab_free(s: df.s, slab: df.slab, head: df.freelist, tail: df.tail, cnt: df.cnt, |
4490 | _RET_IP_); |
4491 | } while (likely(size)); |
4492 | } |
4493 | |
4494 | /* Note that interrupts must be enabled when calling this function. */ |
4495 | void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p) |
4496 | { |
4497 | if (!size) |
4498 | return; |
4499 | |
4500 | do { |
4501 | struct detached_freelist df; |
4502 | |
4503 | size = build_detached_freelist(s, size, p, df: &df); |
4504 | if (!df.slab) |
4505 | continue; |
4506 | |
4507 | slab_free_bulk(s: df.s, slab: df.slab, head: df.freelist, tail: df.tail, p: &p[size], |
4508 | cnt: df.cnt, _RET_IP_); |
4509 | } while (likely(size)); |
4510 | } |
4511 | EXPORT_SYMBOL(kmem_cache_free_bulk); |
4512 | |
4513 | #ifndef CONFIG_SLUB_TINY |
4514 | static inline |
4515 | int __kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size, |
4516 | void **p) |
4517 | { |
4518 | struct kmem_cache_cpu *c; |
4519 | unsigned long irqflags; |
4520 | int i; |
4521 | |
4522 | /* |
4523 | * Drain objects in the per cpu slab, while disabling local |
4524 | * IRQs, which protects against PREEMPT and interrupts |
4525 | * handlers invoking normal fastpath. |
4526 | */ |
4527 | c = slub_get_cpu_ptr(s->cpu_slab); |
4528 | local_lock_irqsave(&s->cpu_slab->lock, irqflags); |
4529 | |
4530 | for (i = 0; i < size; i++) { |
4531 | void *object = kfence_alloc(s, s->object_size, flags); |
4532 | |
4533 | if (unlikely(object)) { |
4534 | p[i] = object; |
4535 | continue; |
4536 | } |
4537 | |
4538 | object = c->freelist; |
4539 | if (unlikely(!object)) { |
4540 | /* |
4541 | * We may have removed an object from c->freelist using |
4542 | * the fastpath in the previous iteration; in that case, |
4543 | * c->tid has not been bumped yet. |
4544 | * Since ___slab_alloc() may reenable interrupts while |
4545 | * allocating memory, we should bump c->tid now. |
4546 | */ |
4547 | c->tid = next_tid(c->tid); |
4548 | |
4549 | local_unlock_irqrestore(&s->cpu_slab->lock, irqflags); |
4550 | |
4551 | /* |
4552 | * Invoking slow path likely have side-effect |
4553 | * of re-populating per CPU c->freelist |
4554 | */ |
4555 | p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE, |
4556 | _RET_IP_, c, s->object_size); |
4557 | if (unlikely(!p[i])) |
4558 | goto error; |
4559 | |
4560 | c = this_cpu_ptr(s->cpu_slab); |
4561 | maybe_wipe_obj_freeptr(s, p[i]); |
4562 | |
4563 | local_lock_irqsave(&s->cpu_slab->lock, irqflags); |
4564 | |
4565 | continue; /* goto for-loop */ |
4566 | } |
4567 | c->freelist = get_freepointer(s, object); |
4568 | p[i] = object; |
4569 | maybe_wipe_obj_freeptr(s, p[i]); |
4570 | stat(s, ALLOC_FASTPATH); |
4571 | } |
4572 | c->tid = next_tid(c->tid); |
4573 | local_unlock_irqrestore(&s->cpu_slab->lock, irqflags); |
4574 | slub_put_cpu_ptr(s->cpu_slab); |
4575 | |
4576 | return i; |
4577 | |
4578 | error: |
4579 | slub_put_cpu_ptr(s->cpu_slab); |
4580 | __kmem_cache_free_bulk(s, i, p); |
4581 | return 0; |
4582 | |
4583 | } |
4584 | #else /* CONFIG_SLUB_TINY */ |
4585 | static int __kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, |
4586 | size_t size, void **p) |
4587 | { |
4588 | int i; |
4589 | |
4590 | for (i = 0; i < size; i++) { |
4591 | void *object = kfence_alloc(s, size: s->object_size, flags); |
4592 | |
4593 | if (unlikely(object)) { |
4594 | p[i] = object; |
4595 | continue; |
4596 | } |
4597 | |
4598 | p[i] = __slab_alloc_node(s, gfpflags: flags, NUMA_NO_NODE, |
4599 | _RET_IP_, orig_size: s->object_size); |
4600 | if (unlikely(!p[i])) |
4601 | goto error; |
4602 | |
4603 | maybe_wipe_obj_freeptr(s, obj: p[i]); |
4604 | } |
4605 | |
4606 | return i; |
4607 | |
4608 | error: |
4609 | __kmem_cache_free_bulk(s, size: i, p); |
4610 | return 0; |
4611 | } |
4612 | #endif /* CONFIG_SLUB_TINY */ |
4613 | |
4614 | /* Note that interrupts must be enabled when calling this function. */ |
4615 | int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size, |
4616 | void **p) |
4617 | { |
4618 | int i; |
4619 | struct obj_cgroup *objcg = NULL; |
4620 | |
4621 | if (!size) |
4622 | return 0; |
4623 | |
4624 | /* memcg and kmem_cache debug support */ |
4625 | s = slab_pre_alloc_hook(s, NULL, objcgp: &objcg, size, flags); |
4626 | if (unlikely(!s)) |
4627 | return 0; |
4628 | |
4629 | i = __kmem_cache_alloc_bulk(s, flags, size, p); |
4630 | |
4631 | /* |
4632 | * memcg and kmem_cache debug support and memory initialization. |
4633 | * Done outside of the IRQ disabled fastpath loop. |
4634 | */ |
4635 | if (likely(i != 0)) { |
4636 | slab_post_alloc_hook(s, objcg, flags, size, p, |
4637 | init: slab_want_init_on_alloc(flags, c: s), orig_size: s->object_size); |
4638 | } else { |
4639 | memcg_slab_alloc_error_hook(s, objects: size, objcg); |
4640 | } |
4641 | |
4642 | return i; |
4643 | } |
4644 | EXPORT_SYMBOL(kmem_cache_alloc_bulk); |
4645 | |
4646 | |
4647 | /* |
4648 | * Object placement in a slab is made very easy because we always start at |
4649 | * offset 0. If we tune the size of the object to the alignment then we can |
4650 | * get the required alignment by putting one properly sized object after |
4651 | * another. |
4652 | * |
4653 | * Notice that the allocation order determines the sizes of the per cpu |
4654 | * caches. Each processor has always one slab available for allocations. |
4655 | * Increasing the allocation order reduces the number of times that slabs |
4656 | * must be moved on and off the partial lists and is therefore a factor in |
4657 | * locking overhead. |
4658 | */ |
4659 | |
4660 | /* |
4661 | * Minimum / Maximum order of slab pages. This influences locking overhead |
4662 | * and slab fragmentation. A higher order reduces the number of partial slabs |
4663 | * and increases the number of allocations possible without having to |
4664 | * take the list_lock. |
4665 | */ |
4666 | static unsigned int slub_min_order; |
4667 | static unsigned int slub_max_order = |
4668 | IS_ENABLED(CONFIG_SLUB_TINY) ? 1 : PAGE_ALLOC_COSTLY_ORDER; |
4669 | static unsigned int slub_min_objects; |
4670 | |
4671 | /* |
4672 | * Calculate the order of allocation given an slab object size. |
4673 | * |
4674 | * The order of allocation has significant impact on performance and other |
4675 | * system components. Generally order 0 allocations should be preferred since |
4676 | * order 0 does not cause fragmentation in the page allocator. Larger objects |
4677 | * be problematic to put into order 0 slabs because there may be too much |
4678 | * unused space left. We go to a higher order if more than 1/16th of the slab |
4679 | * would be wasted. |
4680 | * |
4681 | * In order to reach satisfactory performance we must ensure that a minimum |
4682 | * number of objects is in one slab. Otherwise we may generate too much |
4683 | * activity on the partial lists which requires taking the list_lock. This is |
4684 | * less a concern for large slabs though which are rarely used. |
4685 | * |
4686 | * slab_max_order specifies the order where we begin to stop considering the |
4687 | * number of objects in a slab as critical. If we reach slab_max_order then |
4688 | * we try to keep the page order as low as possible. So we accept more waste |
4689 | * of space in favor of a small page order. |
4690 | * |
4691 | * Higher order allocations also allow the placement of more objects in a |
4692 | * slab and thereby reduce object handling overhead. If the user has |
4693 | * requested a higher minimum order then we start with that one instead of |
4694 | * the smallest order which will fit the object. |
4695 | */ |
4696 | static inline unsigned int calc_slab_order(unsigned int size, |
4697 | unsigned int min_order, unsigned int max_order, |
4698 | unsigned int fract_leftover) |
4699 | { |
4700 | unsigned int order; |
4701 | |
4702 | for (order = min_order; order <= max_order; order++) { |
4703 | |
4704 | unsigned int slab_size = (unsigned int)PAGE_SIZE << order; |
4705 | unsigned int rem; |
4706 | |
4707 | rem = slab_size % size; |
4708 | |
4709 | if (rem <= slab_size / fract_leftover) |
4710 | break; |
4711 | } |
4712 | |
4713 | return order; |
4714 | } |
4715 | |
4716 | static inline int calculate_order(unsigned int size) |
4717 | { |
4718 | unsigned int order; |
4719 | unsigned int min_objects; |
4720 | unsigned int max_objects; |
4721 | unsigned int min_order; |
4722 | |
4723 | min_objects = slub_min_objects; |
4724 | if (!min_objects) { |
4725 | /* |
4726 | * Some architectures will only update present cpus when |
4727 | * onlining them, so don't trust the number if it's just 1. But |
4728 | * we also don't want to use nr_cpu_ids always, as on some other |
4729 | * architectures, there can be many possible cpus, but never |
4730 | * onlined. Here we compromise between trying to avoid too high |
4731 | * order on systems that appear larger than they are, and too |
4732 | * low order on systems that appear smaller than they are. |
4733 | */ |
4734 | unsigned int nr_cpus = num_present_cpus(); |
4735 | if (nr_cpus <= 1) |
4736 | nr_cpus = nr_cpu_ids; |
4737 | min_objects = 4 * (fls(x: nr_cpus) + 1); |
4738 | } |
4739 | /* min_objects can't be 0 because get_order(0) is undefined */ |
4740 | max_objects = max(order_objects(slub_max_order, size), 1U); |
4741 | min_objects = min(min_objects, max_objects); |
4742 | |
4743 | min_order = max_t(unsigned int, slub_min_order, |
4744 | get_order(min_objects * size)); |
4745 | if (order_objects(order: min_order, size) > MAX_OBJS_PER_PAGE) |
4746 | return get_order(size: size * MAX_OBJS_PER_PAGE) - 1; |
4747 | |
4748 | /* |
4749 | * Attempt to find best configuration for a slab. This works by first |
4750 | * attempting to generate a layout with the best possible configuration |
4751 | * and backing off gradually. |
4752 | * |
4753 | * We start with accepting at most 1/16 waste and try to find the |
4754 | * smallest order from min_objects-derived/slab_min_order up to |
4755 | * slab_max_order that will satisfy the constraint. Note that increasing |
4756 | * the order can only result in same or less fractional waste, not more. |
4757 | * |
4758 | * If that fails, we increase the acceptable fraction of waste and try |
4759 | * again. The last iteration with fraction of 1/2 would effectively |
4760 | * accept any waste and give us the order determined by min_objects, as |
4761 | * long as at least single object fits within slab_max_order. |
4762 | */ |
4763 | for (unsigned int fraction = 16; fraction > 1; fraction /= 2) { |
4764 | order = calc_slab_order(size, min_order, max_order: slub_max_order, |
4765 | fract_leftover: fraction); |
4766 | if (order <= slub_max_order) |
4767 | return order; |
4768 | } |
4769 | |
4770 | /* |
4771 | * Doh this slab cannot be placed using slab_max_order. |
4772 | */ |
4773 | order = get_order(size); |
4774 | if (order <= MAX_PAGE_ORDER) |
4775 | return order; |
4776 | return -ENOSYS; |
4777 | } |
4778 | |
4779 | static void |
4780 | init_kmem_cache_node(struct kmem_cache_node *n) |
4781 | { |
4782 | n->nr_partial = 0; |
4783 | spin_lock_init(&n->list_lock); |
4784 | INIT_LIST_HEAD(list: &n->partial); |
4785 | #ifdef CONFIG_SLUB_DEBUG |
4786 | atomic_long_set(&n->nr_slabs, 0); |
4787 | atomic_long_set(&n->total_objects, 0); |
4788 | INIT_LIST_HEAD(&n->full); |
4789 | #endif |
4790 | } |
4791 | |
4792 | #ifndef CONFIG_SLUB_TINY |
4793 | static inline int alloc_kmem_cache_cpus(struct kmem_cache *s) |
4794 | { |
4795 | BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE < |
4796 | NR_KMALLOC_TYPES * KMALLOC_SHIFT_HIGH * |
4797 | sizeof(struct kmem_cache_cpu)); |
4798 | |
4799 | /* |
4800 | * Must align to double word boundary for the double cmpxchg |
4801 | * instructions to work; see __pcpu_double_call_return_bool(). |
4802 | */ |
4803 | s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu), |
4804 | 2 * sizeof(void *)); |
4805 | |
4806 | if (!s->cpu_slab) |
4807 | return 0; |
4808 | |
4809 | init_kmem_cache_cpus(s); |
4810 | |
4811 | return 1; |
4812 | } |
4813 | #else |
4814 | static inline int alloc_kmem_cache_cpus(struct kmem_cache *s) |
4815 | { |
4816 | return 1; |
4817 | } |
4818 | #endif /* CONFIG_SLUB_TINY */ |
4819 | |
4820 | static struct kmem_cache *kmem_cache_node; |
4821 | |
4822 | /* |
4823 | * No kmalloc_node yet so do it by hand. We know that this is the first |
4824 | * slab on the node for this slabcache. There are no concurrent accesses |
4825 | * possible. |
4826 | * |
4827 | * Note that this function only works on the kmem_cache_node |
4828 | * when allocating for the kmem_cache_node. This is used for bootstrapping |
4829 | * memory on a fresh node that has no slab structures yet. |
4830 | */ |
4831 | static void early_kmem_cache_node_alloc(int node) |
4832 | { |
4833 | struct slab *slab; |
4834 | struct kmem_cache_node *n; |
4835 | |
4836 | BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node)); |
4837 | |
4838 | slab = new_slab(s: kmem_cache_node, GFP_NOWAIT, node); |
4839 | |
4840 | BUG_ON(!slab); |
4841 | if (slab_nid(slab) != node) { |
4842 | pr_err("SLUB: Unable to allocate memory from node %d\n" , node); |
4843 | pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n" ); |
4844 | } |
4845 | |
4846 | n = slab->freelist; |
4847 | BUG_ON(!n); |
4848 | #ifdef CONFIG_SLUB_DEBUG |
4849 | init_object(kmem_cache_node, n, SLUB_RED_ACTIVE); |
4850 | init_tracking(kmem_cache_node, n); |
4851 | #endif |
4852 | n = kasan_slab_alloc(s: kmem_cache_node, object: n, GFP_KERNEL, init: false); |
4853 | slab->freelist = get_freepointer(s: kmem_cache_node, object: n); |
4854 | slab->inuse = 1; |
4855 | kmem_cache_node->node[node] = n; |
4856 | init_kmem_cache_node(n); |
4857 | inc_slabs_node(s: kmem_cache_node, node, objects: slab->objects); |
4858 | |
4859 | /* |
4860 | * No locks need to be taken here as it has just been |
4861 | * initialized and there is no concurrent access. |
4862 | */ |
4863 | __add_partial(n, slab, tail: DEACTIVATE_TO_HEAD); |
4864 | } |
4865 | |
4866 | static void free_kmem_cache_nodes(struct kmem_cache *s) |
4867 | { |
4868 | int node; |
4869 | struct kmem_cache_node *n; |
4870 | |
4871 | for_each_kmem_cache_node(s, node, n) { |
4872 | s->node[node] = NULL; |
4873 | kmem_cache_free(kmem_cache_node, n); |
4874 | } |
4875 | } |
4876 | |
4877 | void __kmem_cache_release(struct kmem_cache *s) |
4878 | { |
4879 | cache_random_seq_destroy(cachep: s); |
4880 | #ifndef CONFIG_SLUB_TINY |
4881 | free_percpu(s->cpu_slab); |
4882 | #endif |
4883 | free_kmem_cache_nodes(s); |
4884 | } |
4885 | |
4886 | static int init_kmem_cache_nodes(struct kmem_cache *s) |
4887 | { |
4888 | int node; |
4889 | |
4890 | for_each_node_mask(node, slab_nodes) { |
4891 | struct kmem_cache_node *n; |
4892 | |
4893 | if (slab_state == DOWN) { |
4894 | early_kmem_cache_node_alloc(node); |
4895 | continue; |
4896 | } |
4897 | n = kmem_cache_alloc_node(kmem_cache_node, |
4898 | GFP_KERNEL, node); |
4899 | |
4900 | if (!n) { |
4901 | free_kmem_cache_nodes(s); |
4902 | return 0; |
4903 | } |
4904 | |
4905 | init_kmem_cache_node(n); |
4906 | s->node[node] = n; |
4907 | } |
4908 | return 1; |
4909 | } |
4910 | |
4911 | static void set_cpu_partial(struct kmem_cache *s) |
4912 | { |
4913 | #ifdef CONFIG_SLUB_CPU_PARTIAL |
4914 | unsigned int nr_objects; |
4915 | |
4916 | /* |
4917 | * cpu_partial determined the maximum number of objects kept in the |
4918 | * per cpu partial lists of a processor. |
4919 | * |
4920 | * Per cpu partial lists mainly contain slabs that just have one |
4921 | * object freed. If they are used for allocation then they can be |
4922 | * filled up again with minimal effort. The slab will never hit the |
4923 | * per node partial lists and therefore no locking will be required. |
4924 | * |
4925 | * For backwards compatibility reasons, this is determined as number |
4926 | * of objects, even though we now limit maximum number of pages, see |
4927 | * slub_set_cpu_partial() |
4928 | */ |
4929 | if (!kmem_cache_has_cpu_partial(s)) |
4930 | nr_objects = 0; |
4931 | else if (s->size >= PAGE_SIZE) |
4932 | nr_objects = 6; |
4933 | else if (s->size >= 1024) |
4934 | nr_objects = 24; |
4935 | else if (s->size >= 256) |
4936 | nr_objects = 52; |
4937 | else |
4938 | nr_objects = 120; |
4939 | |
4940 | slub_set_cpu_partial(s, nr_objects); |
4941 | #endif |
4942 | } |
4943 | |
4944 | /* |
4945 | * calculate_sizes() determines the order and the distribution of data within |
4946 | * a slab object. |
4947 | */ |
4948 | static int calculate_sizes(struct kmem_cache *s) |
4949 | { |
4950 | slab_flags_t flags = s->flags; |
4951 | unsigned int size = s->object_size; |
4952 | unsigned int order; |
4953 | |
4954 | /* |
4955 | * Round up object size to the next word boundary. We can only |
4956 | * place the free pointer at word boundaries and this determines |
4957 | * the possible location of the free pointer. |
4958 | */ |
4959 | size = ALIGN(size, sizeof(void *)); |
4960 | |
4961 | #ifdef CONFIG_SLUB_DEBUG |
4962 | /* |
4963 | * Determine if we can poison the object itself. If the user of |
4964 | * the slab may touch the object after free or before allocation |
4965 | * then we should never poison the object itself. |
4966 | */ |
4967 | if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) && |
4968 | !s->ctor) |
4969 | s->flags |= __OBJECT_POISON; |
4970 | else |
4971 | s->flags &= ~__OBJECT_POISON; |
4972 | |
4973 | |
4974 | /* |
4975 | * If we are Redzoning then check if there is some space between the |
4976 | * end of the object and the free pointer. If not then add an |
4977 | * additional word to have some bytes to store Redzone information. |
4978 | */ |
4979 | if ((flags & SLAB_RED_ZONE) && size == s->object_size) |
4980 | size += sizeof(void *); |
4981 | #endif |
4982 | |
4983 | /* |
4984 | * With that we have determined the number of bytes in actual use |
4985 | * by the object and redzoning. |
4986 | */ |
4987 | s->inuse = size; |
4988 | |
4989 | if (slub_debug_orig_size(s) || |
4990 | (flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) || |
4991 | ((flags & SLAB_RED_ZONE) && s->object_size < sizeof(void *)) || |
4992 | s->ctor) { |
4993 | /* |
4994 | * Relocate free pointer after the object if it is not |
4995 | * permitted to overwrite the first word of the object on |
4996 | * kmem_cache_free. |
4997 | * |
4998 | * This is the case if we do RCU, have a constructor or |
4999 | * destructor, are poisoning the objects, or are |
5000 | * redzoning an object smaller than sizeof(void *). |
5001 | * |
5002 | * The assumption that s->offset >= s->inuse means free |
5003 | * pointer is outside of the object is used in the |
5004 | * freeptr_outside_object() function. If that is no |
5005 | * longer true, the function needs to be modified. |
5006 | */ |
5007 | s->offset = size; |
5008 | size += sizeof(void *); |
5009 | } else { |
5010 | /* |
5011 | * Store freelist pointer near middle of object to keep |
5012 | * it away from the edges of the object to avoid small |
5013 | * sized over/underflows from neighboring allocations. |
5014 | */ |
5015 | s->offset = ALIGN_DOWN(s->object_size / 2, sizeof(void *)); |
5016 | } |
5017 | |
5018 | #ifdef CONFIG_SLUB_DEBUG |
5019 | if (flags & SLAB_STORE_USER) { |
5020 | /* |
5021 | * Need to store information about allocs and frees after |
5022 | * the object. |
5023 | */ |
5024 | size += 2 * sizeof(struct track); |
5025 | |
5026 | /* Save the original kmalloc request size */ |
5027 | if (flags & SLAB_KMALLOC) |
5028 | size += sizeof(unsigned int); |
5029 | } |
5030 | #endif |
5031 | |
5032 | kasan_cache_create(cache: s, size: &size, flags: &s->flags); |
5033 | #ifdef CONFIG_SLUB_DEBUG |
5034 | if (flags & SLAB_RED_ZONE) { |
5035 | /* |
5036 | * Add some empty padding so that we can catch |
5037 | * overwrites from earlier objects rather than let |
5038 | * tracking information or the free pointer be |
5039 | * corrupted if a user writes before the start |
5040 | * of the object. |
5041 | */ |
5042 | size += sizeof(void *); |
5043 | |
5044 | s->red_left_pad = sizeof(void *); |
5045 | s->red_left_pad = ALIGN(s->red_left_pad, s->align); |
5046 | size += s->red_left_pad; |
5047 | } |
5048 | #endif |
5049 | |
5050 | /* |
5051 | * SLUB stores one object immediately after another beginning from |
5052 | * offset 0. In order to align the objects we have to simply size |
5053 | * each object to conform to the alignment. |
5054 | */ |
5055 | size = ALIGN(size, s->align); |
5056 | s->size = size; |
5057 | s->reciprocal_size = reciprocal_value(d: size); |
5058 | order = calculate_order(size); |
5059 | |
5060 | if ((int)order < 0) |
5061 | return 0; |
5062 | |
5063 | s->allocflags = 0; |
5064 | if (order) |
5065 | s->allocflags |= __GFP_COMP; |
5066 | |
5067 | if (s->flags & SLAB_CACHE_DMA) |
5068 | s->allocflags |= GFP_DMA; |
5069 | |
5070 | if (s->flags & SLAB_CACHE_DMA32) |
5071 | s->allocflags |= GFP_DMA32; |
5072 | |
5073 | if (s->flags & SLAB_RECLAIM_ACCOUNT) |
5074 | s->allocflags |= __GFP_RECLAIMABLE; |
5075 | |
5076 | /* |
5077 | * Determine the number of objects per slab |
5078 | */ |
5079 | s->oo = oo_make(order, size); |
5080 | s->min = oo_make(order: get_order(size), size); |
5081 | |
5082 | return !!oo_objects(x: s->oo); |
5083 | } |
5084 | |
5085 | static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags) |
5086 | { |
5087 | s->flags = kmem_cache_flags(flags, name: s->name); |
5088 | #ifdef CONFIG_SLAB_FREELIST_HARDENED |
5089 | s->random = get_random_long(); |
5090 | #endif |
5091 | |
5092 | if (!calculate_sizes(s)) |
5093 | goto error; |
5094 | if (disable_higher_order_debug) { |
5095 | /* |
5096 | * Disable debugging flags that store metadata if the min slab |
5097 | * order increased. |
5098 | */ |
5099 | if (get_order(size: s->size) > get_order(size: s->object_size)) { |
5100 | s->flags &= ~DEBUG_METADATA_FLAGS; |
5101 | s->offset = 0; |
5102 | if (!calculate_sizes(s)) |
5103 | goto error; |
5104 | } |
5105 | } |
5106 | |
5107 | #ifdef system_has_freelist_aba |
5108 | if (system_has_freelist_aba() && !(s->flags & SLAB_NO_CMPXCHG)) { |
5109 | /* Enable fast mode */ |
5110 | s->flags |= __CMPXCHG_DOUBLE; |
5111 | } |
5112 | #endif |
5113 | |
5114 | /* |
5115 | * The larger the object size is, the more slabs we want on the partial |
5116 | * list to avoid pounding the page allocator excessively. |
5117 | */ |
5118 | s->min_partial = min_t(unsigned long, MAX_PARTIAL, ilog2(s->size) / 2); |
5119 | s->min_partial = max_t(unsigned long, MIN_PARTIAL, s->min_partial); |
5120 | |
5121 | set_cpu_partial(s); |
5122 | |
5123 | #ifdef CONFIG_NUMA |
5124 | s->remote_node_defrag_ratio = 1000; |
5125 | #endif |
5126 | |
5127 | /* Initialize the pre-computed randomized freelist if slab is up */ |
5128 | if (slab_state >= UP) { |
5129 | if (init_cache_random_seq(s)) |
5130 | goto error; |
5131 | } |
5132 | |
5133 | if (!init_kmem_cache_nodes(s)) |
5134 | goto error; |
5135 | |
5136 | if (alloc_kmem_cache_cpus(s)) |
5137 | return 0; |
5138 | |
5139 | error: |
5140 | __kmem_cache_release(s); |
5141 | return -EINVAL; |
5142 | } |
5143 | |
5144 | static void list_slab_objects(struct kmem_cache *s, struct slab *slab, |
5145 | const char *text) |
5146 | { |
5147 | #ifdef CONFIG_SLUB_DEBUG |
5148 | void *addr = slab_address(slab); |
5149 | void *p; |
5150 | |
5151 | slab_err(s, slab, text, s->name); |
5152 | |
5153 | spin_lock(&object_map_lock); |
5154 | __fill_map(object_map, s, slab); |
5155 | |
5156 | for_each_object(p, s, addr, slab->objects) { |
5157 | |
5158 | if (!test_bit(__obj_to_index(s, addr, p), object_map)) { |
5159 | pr_err("Object 0x%p @offset=%tu\n" , p, p - addr); |
5160 | print_tracking(s, p); |
5161 | } |
5162 | } |
5163 | spin_unlock(&object_map_lock); |
5164 | #endif |
5165 | } |
5166 | |
5167 | /* |
5168 | * Attempt to free all partial slabs on a node. |
5169 | * This is called from __kmem_cache_shutdown(). We must take list_lock |
5170 | * because sysfs file might still access partial list after the shutdowning. |
5171 | */ |
5172 | static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n) |
5173 | { |
5174 | LIST_HEAD(discard); |
5175 | struct slab *slab, *h; |
5176 | |
5177 | BUG_ON(irqs_disabled()); |
5178 | spin_lock_irq(lock: &n->list_lock); |
5179 | list_for_each_entry_safe(slab, h, &n->partial, slab_list) { |
5180 | if (!slab->inuse) { |
5181 | remove_partial(n, slab); |
5182 | list_add(new: &slab->slab_list, head: &discard); |
5183 | } else { |
5184 | list_slab_objects(s, slab, |
5185 | text: "Objects remaining in %s on __kmem_cache_shutdown()" ); |
5186 | } |
5187 | } |
5188 | spin_unlock_irq(lock: &n->list_lock); |
5189 | |
5190 | list_for_each_entry_safe(slab, h, &discard, slab_list) |
5191 | discard_slab(s, slab); |
5192 | } |
5193 | |
5194 | bool __kmem_cache_empty(struct kmem_cache *s) |
5195 | { |
5196 | int node; |
5197 | struct kmem_cache_node *n; |
5198 | |
5199 | for_each_kmem_cache_node(s, node, n) |
5200 | if (n->nr_partial || node_nr_slabs(n)) |
5201 | return false; |
5202 | return true; |
5203 | } |
5204 | |
5205 | /* |
5206 | * Release all resources used by a slab cache. |
5207 | */ |
5208 | int __kmem_cache_shutdown(struct kmem_cache *s) |
5209 | { |
5210 | int node; |
5211 | struct kmem_cache_node *n; |
5212 | |
5213 | flush_all_cpus_locked(s); |
5214 | /* Attempt to free all objects */ |
5215 | for_each_kmem_cache_node(s, node, n) { |
5216 | free_partial(s, n); |
5217 | if (n->nr_partial || node_nr_slabs(n)) |
5218 | return 1; |
5219 | } |
5220 | return 0; |
5221 | } |
5222 | |
5223 | #ifdef CONFIG_PRINTK |
5224 | void __kmem_obj_info(struct kmem_obj_info *kpp, void *object, struct slab *slab) |
5225 | { |
5226 | void *base; |
5227 | int __maybe_unused i; |
5228 | unsigned int objnr; |
5229 | void *objp; |
5230 | void *objp0; |
5231 | struct kmem_cache *s = slab->slab_cache; |
5232 | struct track __maybe_unused *trackp; |
5233 | |
5234 | kpp->kp_ptr = object; |
5235 | kpp->kp_slab = slab; |
5236 | kpp->kp_slab_cache = s; |
5237 | base = slab_address(slab); |
5238 | objp0 = kasan_reset_tag(addr: object); |
5239 | #ifdef CONFIG_SLUB_DEBUG |
5240 | objp = restore_red_left(s, objp0); |
5241 | #else |
5242 | objp = objp0; |
5243 | #endif |
5244 | objnr = obj_to_index(cache: s, slab, obj: objp); |
5245 | kpp->kp_data_offset = (unsigned long)((char *)objp0 - (char *)objp); |
5246 | objp = base + s->size * objnr; |
5247 | kpp->kp_objp = objp; |
5248 | if (WARN_ON_ONCE(objp < base || objp >= base + slab->objects * s->size |
5249 | || (objp - base) % s->size) || |
5250 | !(s->flags & SLAB_STORE_USER)) |
5251 | return; |
5252 | #ifdef CONFIG_SLUB_DEBUG |
5253 | objp = fixup_red_left(s, objp); |
5254 | trackp = get_track(s, objp, TRACK_ALLOC); |
5255 | kpp->kp_ret = (void *)trackp->addr; |
5256 | #ifdef CONFIG_STACKDEPOT |
5257 | { |
5258 | depot_stack_handle_t handle; |
5259 | unsigned long *entries; |
5260 | unsigned int nr_entries; |
5261 | |
5262 | handle = READ_ONCE(trackp->handle); |
5263 | if (handle) { |
5264 | nr_entries = stack_depot_fetch(handle, &entries); |
5265 | for (i = 0; i < KS_ADDRS_COUNT && i < nr_entries; i++) |
5266 | kpp->kp_stack[i] = (void *)entries[i]; |
5267 | } |
5268 | |
5269 | trackp = get_track(s, objp, TRACK_FREE); |
5270 | handle = READ_ONCE(trackp->handle); |
5271 | if (handle) { |
5272 | nr_entries = stack_depot_fetch(handle, &entries); |
5273 | for (i = 0; i < KS_ADDRS_COUNT && i < nr_entries; i++) |
5274 | kpp->kp_free_stack[i] = (void *)entries[i]; |
5275 | } |
5276 | } |
5277 | #endif |
5278 | #endif |
5279 | } |
5280 | #endif |
5281 | |
5282 | /******************************************************************** |
5283 | * Kmalloc subsystem |
5284 | *******************************************************************/ |
5285 | |
5286 | static int __init setup_slub_min_order(char *str) |
5287 | { |
5288 | get_option(str: &str, pint: (int *)&slub_min_order); |
5289 | |
5290 | if (slub_min_order > slub_max_order) |
5291 | slub_max_order = slub_min_order; |
5292 | |
5293 | return 1; |
5294 | } |
5295 | |
5296 | __setup("slab_min_order=" , setup_slub_min_order); |
5297 | __setup_param("slub_min_order=" , slub_min_order, setup_slub_min_order, 0); |
5298 | |
5299 | |
5300 | static int __init setup_slub_max_order(char *str) |
5301 | { |
5302 | get_option(str: &str, pint: (int *)&slub_max_order); |
5303 | slub_max_order = min_t(unsigned int, slub_max_order, MAX_PAGE_ORDER); |
5304 | |
5305 | if (slub_min_order > slub_max_order) |
5306 | slub_min_order = slub_max_order; |
5307 | |
5308 | return 1; |
5309 | } |
5310 | |
5311 | __setup("slab_max_order=" , setup_slub_max_order); |
5312 | __setup_param("slub_max_order=" , slub_max_order, setup_slub_max_order, 0); |
5313 | |
5314 | static int __init setup_slub_min_objects(char *str) |
5315 | { |
5316 | get_option(str: &str, pint: (int *)&slub_min_objects); |
5317 | |
5318 | return 1; |
5319 | } |
5320 | |
5321 | __setup("slab_min_objects=" , setup_slub_min_objects); |
5322 | __setup_param("slub_min_objects=" , slub_min_objects, setup_slub_min_objects, 0); |
5323 | |
5324 | #ifdef CONFIG_HARDENED_USERCOPY |
5325 | /* |
5326 | * Rejects incorrectly sized objects and objects that are to be copied |
5327 | * to/from userspace but do not fall entirely within the containing slab |
5328 | * cache's usercopy region. |
5329 | * |
5330 | * Returns NULL if check passes, otherwise const char * to name of cache |
5331 | * to indicate an error. |
5332 | */ |
5333 | void __check_heap_object(const void *ptr, unsigned long n, |
5334 | const struct slab *slab, bool to_user) |
5335 | { |
5336 | struct kmem_cache *s; |
5337 | unsigned int offset; |
5338 | bool is_kfence = is_kfence_address(addr: ptr); |
5339 | |
5340 | ptr = kasan_reset_tag(addr: ptr); |
5341 | |
5342 | /* Find object and usable object size. */ |
5343 | s = slab->slab_cache; |
5344 | |
5345 | /* Reject impossible pointers. */ |
5346 | if (ptr < slab_address(slab)) |
5347 | usercopy_abort(name: "SLUB object not in SLUB page?!" , NULL, |
5348 | to_user, offset: 0, len: n); |
5349 | |
5350 | /* Find offset within object. */ |
5351 | if (is_kfence) |
5352 | offset = ptr - kfence_object_start(addr: ptr); |
5353 | else |
5354 | offset = (ptr - slab_address(slab)) % s->size; |
5355 | |
5356 | /* Adjust for redzone and reject if within the redzone. */ |
5357 | if (!is_kfence && kmem_cache_debug_flags(s, SLAB_RED_ZONE)) { |
5358 | if (offset < s->red_left_pad) |
5359 | usercopy_abort(name: "SLUB object in left red zone" , |
5360 | detail: s->name, to_user, offset, len: n); |
5361 | offset -= s->red_left_pad; |
5362 | } |
5363 | |
5364 | /* Allow address range falling entirely within usercopy region. */ |
5365 | if (offset >= s->useroffset && |
5366 | offset - s->useroffset <= s->usersize && |
5367 | n <= s->useroffset - offset + s->usersize) |
5368 | return; |
5369 | |
5370 | usercopy_abort(name: "SLUB object" , detail: s->name, to_user, offset, len: n); |
5371 | } |
5372 | #endif /* CONFIG_HARDENED_USERCOPY */ |
5373 | |
5374 | #define SHRINK_PROMOTE_MAX 32 |
5375 | |
5376 | /* |
5377 | * kmem_cache_shrink discards empty slabs and promotes the slabs filled |
5378 | * up most to the head of the partial lists. New allocations will then |
5379 | * fill those up and thus they can be removed from the partial lists. |
5380 | * |
5381 | * The slabs with the least items are placed last. This results in them |
5382 | * being allocated from last increasing the chance that the last objects |
5383 | * are freed in them. |
5384 | */ |
5385 | static int __kmem_cache_do_shrink(struct kmem_cache *s) |
5386 | { |
5387 | int node; |
5388 | int i; |
5389 | struct kmem_cache_node *n; |
5390 | struct slab *slab; |
5391 | struct slab *t; |
5392 | struct list_head discard; |
5393 | struct list_head promote[SHRINK_PROMOTE_MAX]; |
5394 | unsigned long flags; |
5395 | int ret = 0; |
5396 | |
5397 | for_each_kmem_cache_node(s, node, n) { |
5398 | INIT_LIST_HEAD(list: &discard); |
5399 | for (i = 0; i < SHRINK_PROMOTE_MAX; i++) |
5400 | INIT_LIST_HEAD(list: promote + i); |
5401 | |
5402 | spin_lock_irqsave(&n->list_lock, flags); |
5403 | |
5404 | /* |
5405 | * Build lists of slabs to discard or promote. |
5406 | * |
5407 | * Note that concurrent frees may occur while we hold the |
5408 | * list_lock. slab->inuse here is the upper limit. |
5409 | */ |
5410 | list_for_each_entry_safe(slab, t, &n->partial, slab_list) { |
5411 | int free = slab->objects - slab->inuse; |
5412 | |
5413 | /* Do not reread slab->inuse */ |
5414 | barrier(); |
5415 | |
5416 | /* We do not keep full slabs on the list */ |
5417 | BUG_ON(free <= 0); |
5418 | |
5419 | if (free == slab->objects) { |
5420 | list_move(list: &slab->slab_list, head: &discard); |
5421 | slab_clear_node_partial(slab); |
5422 | n->nr_partial--; |
5423 | dec_slabs_node(s, node, objects: slab->objects); |
5424 | } else if (free <= SHRINK_PROMOTE_MAX) |
5425 | list_move(list: &slab->slab_list, head: promote + free - 1); |
5426 | } |
5427 | |
5428 | /* |
5429 | * Promote the slabs filled up most to the head of the |
5430 | * partial list. |
5431 | */ |
5432 | for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--) |
5433 | list_splice(list: promote + i, head: &n->partial); |
5434 | |
5435 | spin_unlock_irqrestore(lock: &n->list_lock, flags); |
5436 | |
5437 | /* Release empty slabs */ |
5438 | list_for_each_entry_safe(slab, t, &discard, slab_list) |
5439 | free_slab(s, slab); |
5440 | |
5441 | if (node_nr_slabs(n)) |
5442 | ret = 1; |
5443 | } |
5444 | |
5445 | return ret; |
5446 | } |
5447 | |
5448 | int __kmem_cache_shrink(struct kmem_cache *s) |
5449 | { |
5450 | flush_all(s); |
5451 | return __kmem_cache_do_shrink(s); |
5452 | } |
5453 | |
5454 | static int slab_mem_going_offline_callback(void *arg) |
5455 | { |
5456 | struct kmem_cache *s; |
5457 | |
5458 | mutex_lock(&slab_mutex); |
5459 | list_for_each_entry(s, &slab_caches, list) { |
5460 | flush_all_cpus_locked(s); |
5461 | __kmem_cache_do_shrink(s); |
5462 | } |
5463 | mutex_unlock(lock: &slab_mutex); |
5464 | |
5465 | return 0; |
5466 | } |
5467 | |
5468 | static void slab_mem_offline_callback(void *arg) |
5469 | { |
5470 | struct memory_notify *marg = arg; |
5471 | int offline_node; |
5472 | |
5473 | offline_node = marg->status_change_nid_normal; |
5474 | |
5475 | /* |
5476 | * If the node still has available memory. we need kmem_cache_node |
5477 | * for it yet. |
5478 | */ |
5479 | if (offline_node < 0) |
5480 | return; |
5481 | |
5482 | mutex_lock(&slab_mutex); |
5483 | node_clear(offline_node, slab_nodes); |
5484 | /* |
5485 | * We no longer free kmem_cache_node structures here, as it would be |
5486 | * racy with all get_node() users, and infeasible to protect them with |
5487 | * slab_mutex. |
5488 | */ |
5489 | mutex_unlock(lock: &slab_mutex); |
5490 | } |
5491 | |
5492 | static int slab_mem_going_online_callback(void *arg) |
5493 | { |
5494 | struct kmem_cache_node *n; |
5495 | struct kmem_cache *s; |
5496 | struct memory_notify *marg = arg; |
5497 | int nid = marg->status_change_nid_normal; |
5498 | int ret = 0; |
5499 | |
5500 | /* |
5501 | * If the node's memory is already available, then kmem_cache_node is |
5502 | * already created. Nothing to do. |
5503 | */ |
5504 | if (nid < 0) |
5505 | return 0; |
5506 | |
5507 | /* |
5508 | * We are bringing a node online. No memory is available yet. We must |
5509 | * allocate a kmem_cache_node structure in order to bring the node |
5510 | * online. |
5511 | */ |
5512 | mutex_lock(&slab_mutex); |
5513 | list_for_each_entry(s, &slab_caches, list) { |
5514 | /* |
5515 | * The structure may already exist if the node was previously |
5516 | * onlined and offlined. |
5517 | */ |
5518 | if (get_node(s, node: nid)) |
5519 | continue; |
5520 | /* |
5521 | * XXX: kmem_cache_alloc_node will fallback to other nodes |
5522 | * since memory is not yet available from the node that |
5523 | * is brought up. |
5524 | */ |
5525 | n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL); |
5526 | if (!n) { |
5527 | ret = -ENOMEM; |
5528 | goto out; |
5529 | } |
5530 | init_kmem_cache_node(n); |
5531 | s->node[nid] = n; |
5532 | } |
5533 | /* |
5534 | * Any cache created after this point will also have kmem_cache_node |
5535 | * initialized for the new node. |
5536 | */ |
5537 | node_set(nid, slab_nodes); |
5538 | out: |
5539 | mutex_unlock(lock: &slab_mutex); |
5540 | return ret; |
5541 | } |
5542 | |
5543 | static int slab_memory_callback(struct notifier_block *self, |
5544 | unsigned long action, void *arg) |
5545 | { |
5546 | int ret = 0; |
5547 | |
5548 | switch (action) { |
5549 | case MEM_GOING_ONLINE: |
5550 | ret = slab_mem_going_online_callback(arg); |
5551 | break; |
5552 | case MEM_GOING_OFFLINE: |
5553 | ret = slab_mem_going_offline_callback(arg); |
5554 | break; |
5555 | case MEM_OFFLINE: |
5556 | case MEM_CANCEL_ONLINE: |
5557 | slab_mem_offline_callback(arg); |
5558 | break; |
5559 | case MEM_ONLINE: |
5560 | case MEM_CANCEL_OFFLINE: |
5561 | break; |
5562 | } |
5563 | if (ret) |
5564 | ret = notifier_from_errno(err: ret); |
5565 | else |
5566 | ret = NOTIFY_OK; |
5567 | return ret; |
5568 | } |
5569 | |
5570 | /******************************************************************** |
5571 | * Basic setup of slabs |
5572 | *******************************************************************/ |
5573 | |
5574 | /* |
5575 | * Used for early kmem_cache structures that were allocated using |
5576 | * the page allocator. Allocate them properly then fix up the pointers |
5577 | * that may be pointing to the wrong kmem_cache structure. |
5578 | */ |
5579 | |
5580 | static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache) |
5581 | { |
5582 | int node; |
5583 | struct kmem_cache *s = kmem_cache_zalloc(k: kmem_cache, GFP_NOWAIT); |
5584 | struct kmem_cache_node *n; |
5585 | |
5586 | memcpy(s, static_cache, kmem_cache->object_size); |
5587 | |
5588 | /* |
5589 | * This runs very early, and only the boot processor is supposed to be |
5590 | * up. Even if it weren't true, IRQs are not up so we couldn't fire |
5591 | * IPIs around. |
5592 | */ |
5593 | __flush_cpu_slab(s, smp_processor_id()); |
5594 | for_each_kmem_cache_node(s, node, n) { |
5595 | struct slab *p; |
5596 | |
5597 | list_for_each_entry(p, &n->partial, slab_list) |
5598 | p->slab_cache = s; |
5599 | |
5600 | #ifdef CONFIG_SLUB_DEBUG |
5601 | list_for_each_entry(p, &n->full, slab_list) |
5602 | p->slab_cache = s; |
5603 | #endif |
5604 | } |
5605 | list_add(new: &s->list, head: &slab_caches); |
5606 | return s; |
5607 | } |
5608 | |
5609 | void __init kmem_cache_init(void) |
5610 | { |
5611 | static __initdata struct kmem_cache boot_kmem_cache, |
5612 | boot_kmem_cache_node; |
5613 | int node; |
5614 | |
5615 | if (debug_guardpage_minorder()) |
5616 | slub_max_order = 0; |
5617 | |
5618 | /* Print slub debugging pointers without hashing */ |
5619 | if (__slub_debug_enabled()) |
5620 | no_hash_pointers_enable(NULL); |
5621 | |
5622 | kmem_cache_node = &boot_kmem_cache_node; |
5623 | kmem_cache = &boot_kmem_cache; |
5624 | |
5625 | /* |
5626 | * Initialize the nodemask for which we will allocate per node |
5627 | * structures. Here we don't need taking slab_mutex yet. |
5628 | */ |
5629 | for_each_node_state(node, N_NORMAL_MEMORY) |
5630 | node_set(node, slab_nodes); |
5631 | |
5632 | create_boot_cache(kmem_cache_node, name: "kmem_cache_node" , |
5633 | size: sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN, useroffset: 0, usersize: 0); |
5634 | |
5635 | hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI); |
5636 | |
5637 | /* Able to allocate the per node structures */ |
5638 | slab_state = PARTIAL; |
5639 | |
5640 | create_boot_cache(kmem_cache, name: "kmem_cache" , |
5641 | offsetof(struct kmem_cache, node) + |
5642 | nr_node_ids * sizeof(struct kmem_cache_node *), |
5643 | SLAB_HWCACHE_ALIGN, useroffset: 0, usersize: 0); |
5644 | |
5645 | kmem_cache = bootstrap(static_cache: &boot_kmem_cache); |
5646 | kmem_cache_node = bootstrap(static_cache: &boot_kmem_cache_node); |
5647 | |
5648 | /* Now we can use the kmem_cache to allocate kmalloc slabs */ |
5649 | setup_kmalloc_cache_index_table(); |
5650 | create_kmalloc_caches(); |
5651 | |
5652 | /* Setup random freelists for each cache */ |
5653 | init_freelist_randomization(); |
5654 | |
5655 | cpuhp_setup_state_nocalls(state: CPUHP_SLUB_DEAD, name: "slub:dead" , NULL, |
5656 | teardown: slub_cpu_dead); |
5657 | |
5658 | pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n" , |
5659 | cache_line_size(), |
5660 | slub_min_order, slub_max_order, slub_min_objects, |
5661 | nr_cpu_ids, nr_node_ids); |
5662 | } |
5663 | |
5664 | void __init kmem_cache_init_late(void) |
5665 | { |
5666 | #ifndef CONFIG_SLUB_TINY |
5667 | flushwq = alloc_workqueue("slub_flushwq" , WQ_MEM_RECLAIM, 0); |
5668 | WARN_ON(!flushwq); |
5669 | #endif |
5670 | } |
5671 | |
5672 | struct kmem_cache * |
5673 | __kmem_cache_alias(const char *name, unsigned int size, unsigned int align, |
5674 | slab_flags_t flags, void (*ctor)(void *)) |
5675 | { |
5676 | struct kmem_cache *s; |
5677 | |
5678 | s = find_mergeable(size, align, flags, name, ctor); |
5679 | if (s) { |
5680 | if (sysfs_slab_alias(s, p: name)) |
5681 | return NULL; |
5682 | |
5683 | s->refcount++; |
5684 | |
5685 | /* |
5686 | * Adjust the object sizes so that we clear |
5687 | * the complete object on kzalloc. |
5688 | */ |
5689 | s->object_size = max(s->object_size, size); |
5690 | s->inuse = max(s->inuse, ALIGN(size, sizeof(void *))); |
5691 | } |
5692 | |
5693 | return s; |
5694 | } |
5695 | |
5696 | int __kmem_cache_create(struct kmem_cache *s, slab_flags_t flags) |
5697 | { |
5698 | int err; |
5699 | |
5700 | err = kmem_cache_open(s, flags); |
5701 | if (err) |
5702 | return err; |
5703 | |
5704 | /* Mutex is not taken during early boot */ |
5705 | if (slab_state <= UP) |
5706 | return 0; |
5707 | |
5708 | err = sysfs_slab_add(s); |
5709 | if (err) { |
5710 | __kmem_cache_release(s); |
5711 | return err; |
5712 | } |
5713 | |
5714 | if (s->flags & SLAB_STORE_USER) |
5715 | debugfs_slab_add(s); |
5716 | |
5717 | return 0; |
5718 | } |
5719 | |
5720 | #ifdef SLAB_SUPPORTS_SYSFS |
5721 | static int count_inuse(struct slab *slab) |
5722 | { |
5723 | return slab->inuse; |
5724 | } |
5725 | |
5726 | static int count_total(struct slab *slab) |
5727 | { |
5728 | return slab->objects; |
5729 | } |
5730 | #endif |
5731 | |
5732 | #ifdef CONFIG_SLUB_DEBUG |
5733 | static void validate_slab(struct kmem_cache *s, struct slab *slab, |
5734 | unsigned long *obj_map) |
5735 | { |
5736 | void *p; |
5737 | void *addr = slab_address(slab); |
5738 | |
5739 | if (!check_slab(s, slab) || !on_freelist(s, slab, NULL)) |
5740 | return; |
5741 | |
5742 | /* Now we know that a valid freelist exists */ |
5743 | __fill_map(obj_map, s, slab); |
5744 | for_each_object(p, s, addr, slab->objects) { |
5745 | u8 val = test_bit(__obj_to_index(s, addr, p), obj_map) ? |
5746 | SLUB_RED_INACTIVE : SLUB_RED_ACTIVE; |
5747 | |
5748 | if (!check_object(s, slab, p, val)) |
5749 | break; |
5750 | } |
5751 | } |
5752 | |
5753 | static int validate_slab_node(struct kmem_cache *s, |
5754 | struct kmem_cache_node *n, unsigned long *obj_map) |
5755 | { |
5756 | unsigned long count = 0; |
5757 | struct slab *slab; |
5758 | unsigned long flags; |
5759 | |
5760 | spin_lock_irqsave(&n->list_lock, flags); |
5761 | |
5762 | list_for_each_entry(slab, &n->partial, slab_list) { |
5763 | validate_slab(s, slab, obj_map); |
5764 | count++; |
5765 | } |
5766 | if (count != n->nr_partial) { |
5767 | pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n" , |
5768 | s->name, count, n->nr_partial); |
5769 | slab_add_kunit_errors(); |
5770 | } |
5771 | |
5772 | if (!(s->flags & SLAB_STORE_USER)) |
5773 | goto out; |
5774 | |
5775 | list_for_each_entry(slab, &n->full, slab_list) { |
5776 | validate_slab(s, slab, obj_map); |
5777 | count++; |
5778 | } |
5779 | if (count != node_nr_slabs(n)) { |
5780 | pr_err("SLUB: %s %ld slabs counted but counter=%ld\n" , |
5781 | s->name, count, node_nr_slabs(n)); |
5782 | slab_add_kunit_errors(); |
5783 | } |
5784 | |
5785 | out: |
5786 | spin_unlock_irqrestore(&n->list_lock, flags); |
5787 | return count; |
5788 | } |
5789 | |
5790 | long validate_slab_cache(struct kmem_cache *s) |
5791 | { |
5792 | int node; |
5793 | unsigned long count = 0; |
5794 | struct kmem_cache_node *n; |
5795 | unsigned long *obj_map; |
5796 | |
5797 | obj_map = bitmap_alloc(oo_objects(s->oo), GFP_KERNEL); |
5798 | if (!obj_map) |
5799 | return -ENOMEM; |
5800 | |
5801 | flush_all(s); |
5802 | for_each_kmem_cache_node(s, node, n) |
5803 | count += validate_slab_node(s, n, obj_map); |
5804 | |
5805 | bitmap_free(obj_map); |
5806 | |
5807 | return count; |
5808 | } |
5809 | EXPORT_SYMBOL(validate_slab_cache); |
5810 | |
5811 | #ifdef CONFIG_DEBUG_FS |
5812 | /* |
5813 | * Generate lists of code addresses where slabcache objects are allocated |
5814 | * and freed. |
5815 | */ |
5816 | |
5817 | struct location { |
5818 | depot_stack_handle_t handle; |
5819 | unsigned long count; |
5820 | unsigned long addr; |
5821 | unsigned long waste; |
5822 | long long sum_time; |
5823 | long min_time; |
5824 | long max_time; |
5825 | long min_pid; |
5826 | long max_pid; |
5827 | DECLARE_BITMAP(cpus, NR_CPUS); |
5828 | nodemask_t nodes; |
5829 | }; |
5830 | |
5831 | struct loc_track { |
5832 | unsigned long max; |
5833 | unsigned long count; |
5834 | struct location *loc; |
5835 | loff_t idx; |
5836 | }; |
5837 | |
5838 | static struct dentry *slab_debugfs_root; |
5839 | |
5840 | static void free_loc_track(struct loc_track *t) |
5841 | { |
5842 | if (t->max) |
5843 | free_pages((unsigned long)t->loc, |
5844 | get_order(sizeof(struct location) * t->max)); |
5845 | } |
5846 | |
5847 | static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags) |
5848 | { |
5849 | struct location *l; |
5850 | int order; |
5851 | |
5852 | order = get_order(sizeof(struct location) * max); |
5853 | |
5854 | l = (void *)__get_free_pages(flags, order); |
5855 | if (!l) |
5856 | return 0; |
5857 | |
5858 | if (t->count) { |
5859 | memcpy(l, t->loc, sizeof(struct location) * t->count); |
5860 | free_loc_track(t); |
5861 | } |
5862 | t->max = max; |
5863 | t->loc = l; |
5864 | return 1; |
5865 | } |
5866 | |
5867 | static int add_location(struct loc_track *t, struct kmem_cache *s, |
5868 | const struct track *track, |
5869 | unsigned int orig_size) |
5870 | { |
5871 | long start, end, pos; |
5872 | struct location *l; |
5873 | unsigned long caddr, chandle, cwaste; |
5874 | unsigned long age = jiffies - track->when; |
5875 | depot_stack_handle_t handle = 0; |
5876 | unsigned int waste = s->object_size - orig_size; |
5877 | |
5878 | #ifdef CONFIG_STACKDEPOT |
5879 | handle = READ_ONCE(track->handle); |
5880 | #endif |
5881 | start = -1; |
5882 | end = t->count; |
5883 | |
5884 | for ( ; ; ) { |
5885 | pos = start + (end - start + 1) / 2; |
5886 | |
5887 | /* |
5888 | * There is nothing at "end". If we end up there |
5889 | * we need to add something to before end. |
5890 | */ |
5891 | if (pos == end) |
5892 | break; |
5893 | |
5894 | l = &t->loc[pos]; |
5895 | caddr = l->addr; |
5896 | chandle = l->handle; |
5897 | cwaste = l->waste; |
5898 | if ((track->addr == caddr) && (handle == chandle) && |
5899 | (waste == cwaste)) { |
5900 | |
5901 | l->count++; |
5902 | if (track->when) { |
5903 | l->sum_time += age; |
5904 | if (age < l->min_time) |
5905 | l->min_time = age; |
5906 | if (age > l->max_time) |
5907 | l->max_time = age; |
5908 | |
5909 | if (track->pid < l->min_pid) |
5910 | l->min_pid = track->pid; |
5911 | if (track->pid > l->max_pid) |
5912 | l->max_pid = track->pid; |
5913 | |
5914 | cpumask_set_cpu(track->cpu, |
5915 | to_cpumask(l->cpus)); |
5916 | } |
5917 | node_set(page_to_nid(virt_to_page(track)), l->nodes); |
5918 | return 1; |
5919 | } |
5920 | |
5921 | if (track->addr < caddr) |
5922 | end = pos; |
5923 | else if (track->addr == caddr && handle < chandle) |
5924 | end = pos; |
5925 | else if (track->addr == caddr && handle == chandle && |
5926 | waste < cwaste) |
5927 | end = pos; |
5928 | else |
5929 | start = pos; |
5930 | } |
5931 | |
5932 | /* |
5933 | * Not found. Insert new tracking element. |
5934 | */ |
5935 | if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC)) |
5936 | return 0; |
5937 | |
5938 | l = t->loc + pos; |
5939 | if (pos < t->count) |
5940 | memmove(l + 1, l, |
5941 | (t->count - pos) * sizeof(struct location)); |
5942 | t->count++; |
5943 | l->count = 1; |
5944 | l->addr = track->addr; |
5945 | l->sum_time = age; |
5946 | l->min_time = age; |
5947 | l->max_time = age; |
5948 | l->min_pid = track->pid; |
5949 | l->max_pid = track->pid; |
5950 | l->handle = handle; |
5951 | l->waste = waste; |
5952 | cpumask_clear(to_cpumask(l->cpus)); |
5953 | cpumask_set_cpu(track->cpu, to_cpumask(l->cpus)); |
5954 | nodes_clear(l->nodes); |
5955 | node_set(page_to_nid(virt_to_page(track)), l->nodes); |
5956 | return 1; |
5957 | } |
5958 | |
5959 | static void process_slab(struct loc_track *t, struct kmem_cache *s, |
5960 | struct slab *slab, enum track_item alloc, |
5961 | unsigned long *obj_map) |
5962 | { |
5963 | void *addr = slab_address(slab); |
5964 | bool is_alloc = (alloc == TRACK_ALLOC); |
5965 | void *p; |
5966 | |
5967 | __fill_map(obj_map, s, slab); |
5968 | |
5969 | for_each_object(p, s, addr, slab->objects) |
5970 | if (!test_bit(__obj_to_index(s, addr, p), obj_map)) |
5971 | add_location(t, s, get_track(s, p, alloc), |
5972 | is_alloc ? get_orig_size(s, p) : |
5973 | s->object_size); |
5974 | } |
5975 | #endif /* CONFIG_DEBUG_FS */ |
5976 | #endif /* CONFIG_SLUB_DEBUG */ |
5977 | |
5978 | #ifdef SLAB_SUPPORTS_SYSFS |
5979 | enum slab_stat_type { |
5980 | SL_ALL, /* All slabs */ |
5981 | SL_PARTIAL, /* Only partially allocated slabs */ |
5982 | SL_CPU, /* Only slabs used for cpu caches */ |
5983 | SL_OBJECTS, /* Determine allocated objects not slabs */ |
5984 | SL_TOTAL /* Determine object capacity not slabs */ |
5985 | }; |
5986 | |
5987 | #define SO_ALL (1 << SL_ALL) |
5988 | #define SO_PARTIAL (1 << SL_PARTIAL) |
5989 | #define SO_CPU (1 << SL_CPU) |
5990 | #define SO_OBJECTS (1 << SL_OBJECTS) |
5991 | #define SO_TOTAL (1 << SL_TOTAL) |
5992 | |
5993 | static ssize_t show_slab_objects(struct kmem_cache *s, |
5994 | char *buf, unsigned long flags) |
5995 | { |
5996 | unsigned long total = 0; |
5997 | int node; |
5998 | int x; |
5999 | unsigned long *nodes; |
6000 | int len = 0; |
6001 | |
6002 | nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL); |
6003 | if (!nodes) |
6004 | return -ENOMEM; |
6005 | |
6006 | if (flags & SO_CPU) { |
6007 | int cpu; |
6008 | |
6009 | for_each_possible_cpu(cpu) { |
6010 | struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, |
6011 | cpu); |
6012 | int node; |
6013 | struct slab *slab; |
6014 | |
6015 | slab = READ_ONCE(c->slab); |
6016 | if (!slab) |
6017 | continue; |
6018 | |
6019 | node = slab_nid(slab); |
6020 | if (flags & SO_TOTAL) |
6021 | x = slab->objects; |
6022 | else if (flags & SO_OBJECTS) |
6023 | x = slab->inuse; |
6024 | else |
6025 | x = 1; |
6026 | |
6027 | total += x; |
6028 | nodes[node] += x; |
6029 | |
6030 | #ifdef CONFIG_SLUB_CPU_PARTIAL |
6031 | slab = slub_percpu_partial_read_once(c); |
6032 | if (slab) { |
6033 | node = slab_nid(slab); |
6034 | if (flags & SO_TOTAL) |
6035 | WARN_ON_ONCE(1); |
6036 | else if (flags & SO_OBJECTS) |
6037 | WARN_ON_ONCE(1); |
6038 | else |
6039 | x = slab->slabs; |
6040 | total += x; |
6041 | nodes[node] += x; |
6042 | } |
6043 | #endif |
6044 | } |
6045 | } |
6046 | |
6047 | /* |
6048 | * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex" |
6049 | * already held which will conflict with an existing lock order: |
6050 | * |
6051 | * mem_hotplug_lock->slab_mutex->kernfs_mutex |
6052 | * |
6053 | * We don't really need mem_hotplug_lock (to hold off |
6054 | * slab_mem_going_offline_callback) here because slab's memory hot |
6055 | * unplug code doesn't destroy the kmem_cache->node[] data. |
6056 | */ |
6057 | |
6058 | #ifdef CONFIG_SLUB_DEBUG |
6059 | if (flags & SO_ALL) { |
6060 | struct kmem_cache_node *n; |
6061 | |
6062 | for_each_kmem_cache_node(s, node, n) { |
6063 | |
6064 | if (flags & SO_TOTAL) |
6065 | x = node_nr_objs(n); |
6066 | else if (flags & SO_OBJECTS) |
6067 | x = node_nr_objs(n) - count_partial(n, count_free); |
6068 | else |
6069 | x = node_nr_slabs(n); |
6070 | total += x; |
6071 | nodes[node] += x; |
6072 | } |
6073 | |
6074 | } else |
6075 | #endif |
6076 | if (flags & SO_PARTIAL) { |
6077 | struct kmem_cache_node *n; |
6078 | |
6079 | for_each_kmem_cache_node(s, node, n) { |
6080 | if (flags & SO_TOTAL) |
6081 | x = count_partial(n, count_total); |
6082 | else if (flags & SO_OBJECTS) |
6083 | x = count_partial(n, count_inuse); |
6084 | else |
6085 | x = n->nr_partial; |
6086 | total += x; |
6087 | nodes[node] += x; |
6088 | } |
6089 | } |
6090 | |
6091 | len += sysfs_emit_at(buf, len, "%lu" , total); |
6092 | #ifdef CONFIG_NUMA |
6093 | for (node = 0; node < nr_node_ids; node++) { |
6094 | if (nodes[node]) |
6095 | len += sysfs_emit_at(buf, len, " N%d=%lu" , |
6096 | node, nodes[node]); |
6097 | } |
6098 | #endif |
6099 | len += sysfs_emit_at(buf, len, "\n" ); |
6100 | kfree(nodes); |
6101 | |
6102 | return len; |
6103 | } |
6104 | |
6105 | #define to_slab_attr(n) container_of(n, struct slab_attribute, attr) |
6106 | #define to_slab(n) container_of(n, struct kmem_cache, kobj) |
6107 | |
6108 | struct slab_attribute { |
6109 | struct attribute attr; |
6110 | ssize_t (*show)(struct kmem_cache *s, char *buf); |
6111 | ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count); |
6112 | }; |
6113 | |
6114 | #define SLAB_ATTR_RO(_name) \ |
6115 | static struct slab_attribute _name##_attr = __ATTR_RO_MODE(_name, 0400) |
6116 | |
6117 | #define SLAB_ATTR(_name) \ |
6118 | static struct slab_attribute _name##_attr = __ATTR_RW_MODE(_name, 0600) |
6119 | |
6120 | static ssize_t slab_size_show(struct kmem_cache *s, char *buf) |
6121 | { |
6122 | return sysfs_emit(buf, "%u\n" , s->size); |
6123 | } |
6124 | SLAB_ATTR_RO(slab_size); |
6125 | |
6126 | static ssize_t align_show(struct kmem_cache *s, char *buf) |
6127 | { |
6128 | return sysfs_emit(buf, "%u\n" , s->align); |
6129 | } |
6130 | SLAB_ATTR_RO(align); |
6131 | |
6132 | static ssize_t object_size_show(struct kmem_cache *s, char *buf) |
6133 | { |
6134 | return sysfs_emit(buf, "%u\n" , s->object_size); |
6135 | } |
6136 | SLAB_ATTR_RO(object_size); |
6137 | |
6138 | static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf) |
6139 | { |
6140 | return sysfs_emit(buf, "%u\n" , oo_objects(s->oo)); |
6141 | } |
6142 | SLAB_ATTR_RO(objs_per_slab); |
6143 | |
6144 | static ssize_t order_show(struct kmem_cache *s, char *buf) |
6145 | { |
6146 | return sysfs_emit(buf, "%u\n" , oo_order(s->oo)); |
6147 | } |
6148 | SLAB_ATTR_RO(order); |
6149 | |
6150 | static ssize_t min_partial_show(struct kmem_cache *s, char *buf) |
6151 | { |
6152 | return sysfs_emit(buf, "%lu\n" , s->min_partial); |
6153 | } |
6154 | |
6155 | static ssize_t min_partial_store(struct kmem_cache *s, const char *buf, |
6156 | size_t length) |
6157 | { |
6158 | unsigned long min; |
6159 | int err; |
6160 | |
6161 | err = kstrtoul(buf, 10, &min); |
6162 | if (err) |
6163 | return err; |
6164 | |
6165 | s->min_partial = min; |
6166 | return length; |
6167 | } |
6168 | SLAB_ATTR(min_partial); |
6169 | |
6170 | static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf) |
6171 | { |
6172 | unsigned int nr_partial = 0; |
6173 | #ifdef CONFIG_SLUB_CPU_PARTIAL |
6174 | nr_partial = s->cpu_partial; |
6175 | #endif |
6176 | |
6177 | return sysfs_emit(buf, "%u\n" , nr_partial); |
6178 | } |
6179 | |
6180 | static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf, |
6181 | size_t length) |
6182 | { |
6183 | unsigned int objects; |
6184 | int err; |
6185 | |
6186 | err = kstrtouint(buf, 10, &objects); |
6187 | if (err) |
6188 | return err; |
6189 | if (objects && !kmem_cache_has_cpu_partial(s)) |
6190 | return -EINVAL; |
6191 | |
6192 | slub_set_cpu_partial(s, objects); |
6193 | flush_all(s); |
6194 | return length; |
6195 | } |
6196 | SLAB_ATTR(cpu_partial); |
6197 | |
6198 | static ssize_t ctor_show(struct kmem_cache *s, char *buf) |
6199 | { |
6200 | if (!s->ctor) |
6201 | return 0; |
6202 | return sysfs_emit(buf, "%pS\n" , s->ctor); |
6203 | } |
6204 | SLAB_ATTR_RO(ctor); |
6205 | |
6206 | static ssize_t aliases_show(struct kmem_cache *s, char *buf) |
6207 | { |
6208 | return sysfs_emit(buf, "%d\n" , s->refcount < 0 ? 0 : s->refcount - 1); |
6209 | } |
6210 | SLAB_ATTR_RO(aliases); |
6211 | |
6212 | static ssize_t partial_show(struct kmem_cache *s, char *buf) |
6213 | { |
6214 | return show_slab_objects(s, buf, SO_PARTIAL); |
6215 | } |
6216 | SLAB_ATTR_RO(partial); |
6217 | |
6218 | static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf) |
6219 | { |
6220 | return show_slab_objects(s, buf, SO_CPU); |
6221 | } |
6222 | SLAB_ATTR_RO(cpu_slabs); |
6223 | |
6224 | static ssize_t objects_partial_show(struct kmem_cache *s, char *buf) |
6225 | { |
6226 | return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS); |
6227 | } |
6228 | SLAB_ATTR_RO(objects_partial); |
6229 | |
6230 | static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf) |
6231 | { |
6232 | int objects = 0; |
6233 | int slabs = 0; |
6234 | int cpu __maybe_unused; |
6235 | int len = 0; |
6236 | |
6237 | #ifdef CONFIG_SLUB_CPU_PARTIAL |
6238 | for_each_online_cpu(cpu) { |
6239 | struct slab *slab; |
6240 | |
6241 | slab = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu)); |
6242 | |
6243 | if (slab) |
6244 | slabs += slab->slabs; |
6245 | } |
6246 | #endif |
6247 | |
6248 | /* Approximate half-full slabs, see slub_set_cpu_partial() */ |
6249 | objects = (slabs * oo_objects(s->oo)) / 2; |
6250 | len += sysfs_emit_at(buf, len, "%d(%d)" , objects, slabs); |
6251 | |
6252 | #ifdef CONFIG_SLUB_CPU_PARTIAL |
6253 | for_each_online_cpu(cpu) { |
6254 | struct slab *slab; |
6255 | |
6256 | slab = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu)); |
6257 | if (slab) { |
6258 | slabs = READ_ONCE(slab->slabs); |
6259 | objects = (slabs * oo_objects(s->oo)) / 2; |
6260 | len += sysfs_emit_at(buf, len, " C%d=%d(%d)" , |
6261 | cpu, objects, slabs); |
6262 | } |
6263 | } |
6264 | #endif |
6265 | len += sysfs_emit_at(buf, len, "\n" ); |
6266 | |
6267 | return len; |
6268 | } |
6269 | SLAB_ATTR_RO(slabs_cpu_partial); |
6270 | |
6271 | static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf) |
6272 | { |
6273 | return sysfs_emit(buf, "%d\n" , !!(s->flags & SLAB_RECLAIM_ACCOUNT)); |
6274 | } |
6275 | SLAB_ATTR_RO(reclaim_account); |
6276 | |
6277 | static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf) |
6278 | { |
6279 | return sysfs_emit(buf, "%d\n" , !!(s->flags & SLAB_HWCACHE_ALIGN)); |
6280 | } |
6281 | SLAB_ATTR_RO(hwcache_align); |
6282 | |
6283 | #ifdef CONFIG_ZONE_DMA |
6284 | static ssize_t cache_dma_show(struct kmem_cache *s, char *buf) |
6285 | { |
6286 | return sysfs_emit(buf, "%d\n" , !!(s->flags & SLAB_CACHE_DMA)); |
6287 | } |
6288 | SLAB_ATTR_RO(cache_dma); |
6289 | #endif |
6290 | |
6291 | #ifdef CONFIG_HARDENED_USERCOPY |
6292 | static ssize_t usersize_show(struct kmem_cache *s, char *buf) |
6293 | { |
6294 | return sysfs_emit(buf, "%u\n" , s->usersize); |
6295 | } |
6296 | SLAB_ATTR_RO(usersize); |
6297 | #endif |
6298 | |
6299 | static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf) |
6300 | { |
6301 | return sysfs_emit(buf, "%d\n" , !!(s->flags & SLAB_TYPESAFE_BY_RCU)); |
6302 | } |
6303 | SLAB_ATTR_RO(destroy_by_rcu); |
6304 | |
6305 | #ifdef CONFIG_SLUB_DEBUG |
6306 | static ssize_t slabs_show(struct kmem_cache *s, char *buf) |
6307 | { |
6308 | return show_slab_objects(s, buf, SO_ALL); |
6309 | } |
6310 | SLAB_ATTR_RO(slabs); |
6311 | |
6312 | static ssize_t total_objects_show(struct kmem_cache *s, char *buf) |
6313 | { |
6314 | return show_slab_objects(s, buf, SO_ALL|SO_TOTAL); |
6315 | } |
6316 | SLAB_ATTR_RO(total_objects); |
6317 | |
6318 | static ssize_t objects_show(struct kmem_cache *s, char *buf) |
6319 | { |
6320 | return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS); |
6321 | } |
6322 | SLAB_ATTR_RO(objects); |
6323 | |
6324 | static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf) |
6325 | { |
6326 | return sysfs_emit(buf, "%d\n" , !!(s->flags & SLAB_CONSISTENCY_CHECKS)); |
6327 | } |
6328 | SLAB_ATTR_RO(sanity_checks); |
6329 | |
6330 | static ssize_t trace_show(struct kmem_cache *s, char *buf) |
6331 | { |
6332 | return sysfs_emit(buf, "%d\n" , !!(s->flags & SLAB_TRACE)); |
6333 | } |
6334 | SLAB_ATTR_RO(trace); |
6335 | |
6336 | static ssize_t red_zone_show(struct kmem_cache *s, char *buf) |
6337 | { |
6338 | return sysfs_emit(buf, "%d\n" , !!(s->flags & SLAB_RED_ZONE)); |
6339 | } |
6340 | |
6341 | SLAB_ATTR_RO(red_zone); |
6342 | |
6343 | static ssize_t poison_show(struct kmem_cache *s, char *buf) |
6344 | { |
6345 | return sysfs_emit(buf, "%d\n" , !!(s->flags & SLAB_POISON)); |
6346 | } |
6347 | |
6348 | SLAB_ATTR_RO(poison); |
6349 | |
6350 | static ssize_t store_user_show(struct kmem_cache *s, char *buf) |
6351 | { |
6352 | return sysfs_emit(buf, "%d\n" , !!(s->flags & SLAB_STORE_USER)); |
6353 | } |
6354 | |
6355 | SLAB_ATTR_RO(store_user); |
6356 | |
6357 | static ssize_t validate_show(struct kmem_cache *s, char *buf) |
6358 | { |
6359 | return 0; |
6360 | } |
6361 | |
6362 | static ssize_t validate_store(struct kmem_cache *s, |
6363 | const char *buf, size_t length) |
6364 | { |
6365 | int ret = -EINVAL; |
6366 | |
6367 | if (buf[0] == '1' && kmem_cache_debug(s)) { |
6368 | ret = validate_slab_cache(s); |
6369 | if (ret >= 0) |
6370 | ret = length; |
6371 | } |
6372 | return ret; |
6373 | } |
6374 | SLAB_ATTR(validate); |
6375 | |
6376 | #endif /* CONFIG_SLUB_DEBUG */ |
6377 | |
6378 | #ifdef CONFIG_FAILSLAB |
6379 | static ssize_t failslab_show(struct kmem_cache *s, char *buf) |
6380 | { |
6381 | return sysfs_emit(buf, "%d\n" , !!(s->flags & SLAB_FAILSLAB)); |
6382 | } |
6383 | |
6384 | static ssize_t failslab_store(struct kmem_cache *s, const char *buf, |
6385 | size_t length) |
6386 | { |
6387 | if (s->refcount > 1) |
6388 | return -EINVAL; |
6389 | |
6390 | if (buf[0] == '1') |
6391 | WRITE_ONCE(s->flags, s->flags | SLAB_FAILSLAB); |
6392 | else |
6393 | WRITE_ONCE(s->flags, s->flags & ~SLAB_FAILSLAB); |
6394 | |
6395 | return length; |
6396 | } |
6397 | SLAB_ATTR(failslab); |
6398 | #endif |
6399 | |
6400 | static ssize_t shrink_show(struct kmem_cache *s, char *buf) |
6401 | { |
6402 | return 0; |
6403 | } |
6404 | |
6405 | static ssize_t shrink_store(struct kmem_cache *s, |
6406 | const char *buf, size_t length) |
6407 | { |
6408 | if (buf[0] == '1') |
6409 | kmem_cache_shrink(s); |
6410 | else |
6411 | return -EINVAL; |
6412 | return length; |
6413 | } |
6414 | SLAB_ATTR(shrink); |
6415 | |
6416 | #ifdef CONFIG_NUMA |
6417 | static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf) |
6418 | { |
6419 | return sysfs_emit(buf, "%u\n" , s->remote_node_defrag_ratio / 10); |
6420 | } |
6421 | |
6422 | static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s, |
6423 | const char *buf, size_t length) |
6424 | { |
6425 | unsigned int ratio; |
6426 | int err; |
6427 | |
6428 | err = kstrtouint(buf, 10, &ratio); |
6429 | if (err) |
6430 | return err; |
6431 | if (ratio > 100) |
6432 | return -ERANGE; |
6433 | |
6434 | s->remote_node_defrag_ratio = ratio * 10; |
6435 | |
6436 | return length; |
6437 | } |
6438 | SLAB_ATTR(remote_node_defrag_ratio); |
6439 | #endif |
6440 | |
6441 | #ifdef CONFIG_SLUB_STATS |
6442 | static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si) |
6443 | { |
6444 | unsigned long sum = 0; |
6445 | int cpu; |
6446 | int len = 0; |
6447 | int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL); |
6448 | |
6449 | if (!data) |
6450 | return -ENOMEM; |
6451 | |
6452 | for_each_online_cpu(cpu) { |
6453 | unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si]; |
6454 | |
6455 | data[cpu] = x; |
6456 | sum += x; |
6457 | } |
6458 | |
6459 | len += sysfs_emit_at(buf, len, "%lu" , sum); |
6460 | |
6461 | #ifdef CONFIG_SMP |
6462 | for_each_online_cpu(cpu) { |
6463 | if (data[cpu]) |
6464 | len += sysfs_emit_at(buf, len, " C%d=%u" , |
6465 | cpu, data[cpu]); |
6466 | } |
6467 | #endif |
6468 | kfree(data); |
6469 | len += sysfs_emit_at(buf, len, "\n" ); |
6470 | |
6471 | return len; |
6472 | } |
6473 | |
6474 | static void clear_stat(struct kmem_cache *s, enum stat_item si) |
6475 | { |
6476 | int cpu; |
6477 | |
6478 | for_each_online_cpu(cpu) |
6479 | per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0; |
6480 | } |
6481 | |
6482 | #define STAT_ATTR(si, text) \ |
6483 | static ssize_t text##_show(struct kmem_cache *s, char *buf) \ |
6484 | { \ |
6485 | return show_stat(s, buf, si); \ |
6486 | } \ |
6487 | static ssize_t text##_store(struct kmem_cache *s, \ |
6488 | const char *buf, size_t length) \ |
6489 | { \ |
6490 | if (buf[0] != '0') \ |
6491 | return -EINVAL; \ |
6492 | clear_stat(s, si); \ |
6493 | return length; \ |
6494 | } \ |
6495 | SLAB_ATTR(text); \ |
6496 | |
6497 | STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath); |
6498 | STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath); |
6499 | STAT_ATTR(FREE_FASTPATH, free_fastpath); |
6500 | STAT_ATTR(FREE_SLOWPATH, free_slowpath); |
6501 | STAT_ATTR(FREE_FROZEN, free_frozen); |
6502 | STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial); |
6503 | STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial); |
6504 | STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial); |
6505 | STAT_ATTR(ALLOC_SLAB, alloc_slab); |
6506 | STAT_ATTR(ALLOC_REFILL, alloc_refill); |
6507 | STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch); |
6508 | STAT_ATTR(FREE_SLAB, free_slab); |
6509 | STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush); |
6510 | STAT_ATTR(DEACTIVATE_FULL, deactivate_full); |
6511 | STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty); |
6512 | STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head); |
6513 | STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail); |
6514 | STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees); |
6515 | STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass); |
6516 | STAT_ATTR(ORDER_FALLBACK, order_fallback); |
6517 | STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail); |
6518 | STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail); |
6519 | STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc); |
6520 | STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free); |
6521 | STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node); |
6522 | STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain); |
6523 | #endif /* CONFIG_SLUB_STATS */ |
6524 | |
6525 | #ifdef CONFIG_KFENCE |
6526 | static ssize_t skip_kfence_show(struct kmem_cache *s, char *buf) |
6527 | { |
6528 | return sysfs_emit(buf, "%d\n" , !!(s->flags & SLAB_SKIP_KFENCE)); |
6529 | } |
6530 | |
6531 | static ssize_t skip_kfence_store(struct kmem_cache *s, |
6532 | const char *buf, size_t length) |
6533 | { |
6534 | int ret = length; |
6535 | |
6536 | if (buf[0] == '0') |
6537 | s->flags &= ~SLAB_SKIP_KFENCE; |
6538 | else if (buf[0] == '1') |
6539 | s->flags |= SLAB_SKIP_KFENCE; |
6540 | else |
6541 | ret = -EINVAL; |
6542 | |
6543 | return ret; |
6544 | } |
6545 | SLAB_ATTR(skip_kfence); |
6546 | #endif |
6547 | |
6548 | static struct attribute *slab_attrs[] = { |
6549 | &slab_size_attr.attr, |
6550 | &object_size_attr.attr, |
6551 | &objs_per_slab_attr.attr, |
6552 | &order_attr.attr, |
6553 | &min_partial_attr.attr, |
6554 | &cpu_partial_attr.attr, |
6555 | &objects_partial_attr.attr, |
6556 | &partial_attr.attr, |
6557 | &cpu_slabs_attr.attr, |
6558 | &ctor_attr.attr, |
6559 | &aliases_attr.attr, |
6560 | &align_attr.attr, |
6561 | &hwcache_align_attr.attr, |
6562 | &reclaim_account_attr.attr, |
6563 | &destroy_by_rcu_attr.attr, |
6564 | &shrink_attr.attr, |
6565 | &slabs_cpu_partial_attr.attr, |
6566 | #ifdef CONFIG_SLUB_DEBUG |
6567 | &total_objects_attr.attr, |
6568 | &objects_attr.attr, |
6569 | &slabs_attr.attr, |
6570 | &sanity_checks_attr.attr, |
6571 | &trace_attr.attr, |
6572 | &red_zone_attr.attr, |
6573 | &poison_attr.attr, |
6574 | &store_user_attr.attr, |
6575 | &validate_attr.attr, |
6576 | #endif |
6577 | #ifdef CONFIG_ZONE_DMA |
6578 | &cache_dma_attr.attr, |
6579 | #endif |
6580 | #ifdef CONFIG_NUMA |
6581 | &remote_node_defrag_ratio_attr.attr, |
6582 | #endif |
6583 | #ifdef CONFIG_SLUB_STATS |
6584 | &alloc_fastpath_attr.attr, |
6585 | &alloc_slowpath_attr.attr, |
6586 | &free_fastpath_attr.attr, |
6587 | &free_slowpath_attr.attr, |
6588 | &free_frozen_attr.attr, |
6589 | &free_add_partial_attr.attr, |
6590 | &free_remove_partial_attr.attr, |
6591 | &alloc_from_partial_attr.attr, |
6592 | &alloc_slab_attr.attr, |
6593 | &alloc_refill_attr.attr, |
6594 | &alloc_node_mismatch_attr.attr, |
6595 | &free_slab_attr.attr, |
6596 | &cpuslab_flush_attr.attr, |
6597 | &deactivate_full_attr.attr, |
6598 | &deactivate_empty_attr.attr, |
6599 | &deactivate_to_head_attr.attr, |
6600 | &deactivate_to_tail_attr.attr, |
6601 | &deactivate_remote_frees_attr.attr, |
6602 | &deactivate_bypass_attr.attr, |
6603 | &order_fallback_attr.attr, |
6604 | &cmpxchg_double_fail_attr.attr, |
6605 | &cmpxchg_double_cpu_fail_attr.attr, |
6606 | &cpu_partial_alloc_attr.attr, |
6607 | &cpu_partial_free_attr.attr, |
6608 | &cpu_partial_node_attr.attr, |
6609 | &cpu_partial_drain_attr.attr, |
6610 | #endif |
6611 | #ifdef CONFIG_FAILSLAB |
6612 | &failslab_attr.attr, |
6613 | #endif |
6614 | #ifdef CONFIG_HARDENED_USERCOPY |
6615 | &usersize_attr.attr, |
6616 | #endif |
6617 | #ifdef CONFIG_KFENCE |
6618 | &skip_kfence_attr.attr, |
6619 | #endif |
6620 | |
6621 | NULL |
6622 | }; |
6623 | |
6624 | static const struct attribute_group slab_attr_group = { |
6625 | .attrs = slab_attrs, |
6626 | }; |
6627 | |
6628 | static ssize_t slab_attr_show(struct kobject *kobj, |
6629 | struct attribute *attr, |
6630 | char *buf) |
6631 | { |
6632 | struct slab_attribute *attribute; |
6633 | struct kmem_cache *s; |
6634 | |
6635 | attribute = to_slab_attr(attr); |
6636 | s = to_slab(kobj); |
6637 | |
6638 | if (!attribute->show) |
6639 | return -EIO; |
6640 | |
6641 | return attribute->show(s, buf); |
6642 | } |
6643 | |
6644 | static ssize_t slab_attr_store(struct kobject *kobj, |
6645 | struct attribute *attr, |
6646 | const char *buf, size_t len) |
6647 | { |
6648 | struct slab_attribute *attribute; |
6649 | struct kmem_cache *s; |
6650 | |
6651 | attribute = to_slab_attr(attr); |
6652 | s = to_slab(kobj); |
6653 | |
6654 | if (!attribute->store) |
6655 | return -EIO; |
6656 | |
6657 | return attribute->store(s, buf, len); |
6658 | } |
6659 | |
6660 | static void kmem_cache_release(struct kobject *k) |
6661 | { |
6662 | slab_kmem_cache_release(to_slab(k)); |
6663 | } |
6664 | |
6665 | static const struct sysfs_ops slab_sysfs_ops = { |
6666 | .show = slab_attr_show, |
6667 | .store = slab_attr_store, |
6668 | }; |
6669 | |
6670 | static const struct kobj_type slab_ktype = { |
6671 | .sysfs_ops = &slab_sysfs_ops, |
6672 | .release = kmem_cache_release, |
6673 | }; |
6674 | |
6675 | static struct kset *slab_kset; |
6676 | |
6677 | static inline struct kset *cache_kset(struct kmem_cache *s) |
6678 | { |
6679 | return slab_kset; |
6680 | } |
6681 | |
6682 | #define ID_STR_LENGTH 32 |
6683 | |
6684 | /* Create a unique string id for a slab cache: |
6685 | * |
6686 | * Format :[flags-]size |
6687 | */ |
6688 | static char *create_unique_id(struct kmem_cache *s) |
6689 | { |
6690 | char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL); |
6691 | char *p = name; |
6692 | |
6693 | if (!name) |
6694 | return ERR_PTR(-ENOMEM); |
6695 | |
6696 | *p++ = ':'; |
6697 | /* |
6698 | * First flags affecting slabcache operations. We will only |
6699 | * get here for aliasable slabs so we do not need to support |
6700 | * too many flags. The flags here must cover all flags that |
6701 | * are matched during merging to guarantee that the id is |
6702 | * unique. |
6703 | */ |
6704 | if (s->flags & SLAB_CACHE_DMA) |
6705 | *p++ = 'd'; |
6706 | if (s->flags & SLAB_CACHE_DMA32) |
6707 | *p++ = 'D'; |
6708 | if (s->flags & SLAB_RECLAIM_ACCOUNT) |
6709 | *p++ = 'a'; |
6710 | if (s->flags & SLAB_CONSISTENCY_CHECKS) |
6711 | *p++ = 'F'; |
6712 | if (s->flags & SLAB_ACCOUNT) |
6713 | *p++ = 'A'; |
6714 | if (p != name + 1) |
6715 | *p++ = '-'; |
6716 | p += snprintf(p, ID_STR_LENGTH - (p - name), "%07u" , s->size); |
6717 | |
6718 | if (WARN_ON(p > name + ID_STR_LENGTH - 1)) { |
6719 | kfree(name); |
6720 | return ERR_PTR(-EINVAL); |
6721 | } |
6722 | kmsan_unpoison_memory(name, p - name); |
6723 | return name; |
6724 | } |
6725 | |
6726 | static int sysfs_slab_add(struct kmem_cache *s) |
6727 | { |
6728 | int err; |
6729 | const char *name; |
6730 | struct kset *kset = cache_kset(s); |
6731 | int unmergeable = slab_unmergeable(s); |
6732 | |
6733 | if (!unmergeable && disable_higher_order_debug && |
6734 | (slub_debug & DEBUG_METADATA_FLAGS)) |
6735 | unmergeable = 1; |
6736 | |
6737 | if (unmergeable) { |
6738 | /* |
6739 | * Slabcache can never be merged so we can use the name proper. |
6740 | * This is typically the case for debug situations. In that |
6741 | * case we can catch duplicate names easily. |
6742 | */ |
6743 | sysfs_remove_link(&slab_kset->kobj, s->name); |
6744 | name = s->name; |
6745 | } else { |
6746 | /* |
6747 | * Create a unique name for the slab as a target |
6748 | * for the symlinks. |
6749 | */ |
6750 | name = create_unique_id(s); |
6751 | if (IS_ERR(name)) |
6752 | return PTR_ERR(name); |
6753 | } |
6754 | |
6755 | s->kobj.kset = kset; |
6756 | err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s" , name); |
6757 | if (err) |
6758 | goto out; |
6759 | |
6760 | err = sysfs_create_group(&s->kobj, &slab_attr_group); |
6761 | if (err) |
6762 | goto out_del_kobj; |
6763 | |
6764 | if (!unmergeable) { |
6765 | /* Setup first alias */ |
6766 | sysfs_slab_alias(s, s->name); |
6767 | } |
6768 | out: |
6769 | if (!unmergeable) |
6770 | kfree(name); |
6771 | return err; |
6772 | out_del_kobj: |
6773 | kobject_del(&s->kobj); |
6774 | goto out; |
6775 | } |
6776 | |
6777 | void sysfs_slab_unlink(struct kmem_cache *s) |
6778 | { |
6779 | kobject_del(&s->kobj); |
6780 | } |
6781 | |
6782 | void sysfs_slab_release(struct kmem_cache *s) |
6783 | { |
6784 | kobject_put(&s->kobj); |
6785 | } |
6786 | |
6787 | /* |
6788 | * Need to buffer aliases during bootup until sysfs becomes |
6789 | * available lest we lose that information. |
6790 | */ |
6791 | struct saved_alias { |
6792 | struct kmem_cache *s; |
6793 | const char *name; |
6794 | struct saved_alias *next; |
6795 | }; |
6796 | |
6797 | static struct saved_alias *alias_list; |
6798 | |
6799 | static int sysfs_slab_alias(struct kmem_cache *s, const char *name) |
6800 | { |
6801 | struct saved_alias *al; |
6802 | |
6803 | if (slab_state == FULL) { |
6804 | /* |
6805 | * If we have a leftover link then remove it. |
6806 | */ |
6807 | sysfs_remove_link(&slab_kset->kobj, name); |
6808 | return sysfs_create_link(&slab_kset->kobj, &s->kobj, name); |
6809 | } |
6810 | |
6811 | al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL); |
6812 | if (!al) |
6813 | return -ENOMEM; |
6814 | |
6815 | al->s = s; |
6816 | al->name = name; |
6817 | al->next = alias_list; |
6818 | alias_list = al; |
6819 | kmsan_unpoison_memory(al, sizeof(*al)); |
6820 | return 0; |
6821 | } |
6822 | |
6823 | static int __init slab_sysfs_init(void) |
6824 | { |
6825 | struct kmem_cache *s; |
6826 | int err; |
6827 | |
6828 | mutex_lock(&slab_mutex); |
6829 | |
6830 | slab_kset = kset_create_and_add("slab" , NULL, kernel_kobj); |
6831 | if (!slab_kset) { |
6832 | mutex_unlock(&slab_mutex); |
6833 | pr_err("Cannot register slab subsystem.\n" ); |
6834 | return -ENOMEM; |
6835 | } |
6836 | |
6837 | slab_state = FULL; |
6838 | |
6839 | list_for_each_entry(s, &slab_caches, list) { |
6840 | err = sysfs_slab_add(s); |
6841 | if (err) |
6842 | pr_err("SLUB: Unable to add boot slab %s to sysfs\n" , |
6843 | s->name); |
6844 | } |
6845 | |
6846 | while (alias_list) { |
6847 | struct saved_alias *al = alias_list; |
6848 | |
6849 | alias_list = alias_list->next; |
6850 | err = sysfs_slab_alias(al->s, al->name); |
6851 | if (err) |
6852 | pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n" , |
6853 | al->name); |
6854 | kfree(al); |
6855 | } |
6856 | |
6857 | mutex_unlock(&slab_mutex); |
6858 | return 0; |
6859 | } |
6860 | late_initcall(slab_sysfs_init); |
6861 | #endif /* SLAB_SUPPORTS_SYSFS */ |
6862 | |
6863 | #if defined(CONFIG_SLUB_DEBUG) && defined(CONFIG_DEBUG_FS) |
6864 | static int slab_debugfs_show(struct seq_file *seq, void *v) |
6865 | { |
6866 | struct loc_track *t = seq->private; |
6867 | struct location *l; |
6868 | unsigned long idx; |
6869 | |
6870 | idx = (unsigned long) t->idx; |
6871 | if (idx < t->count) { |
6872 | l = &t->loc[idx]; |
6873 | |
6874 | seq_printf(seq, "%7ld " , l->count); |
6875 | |
6876 | if (l->addr) |
6877 | seq_printf(seq, "%pS" , (void *)l->addr); |
6878 | else |
6879 | seq_puts(seq, "<not-available>" ); |
6880 | |
6881 | if (l->waste) |
6882 | seq_printf(seq, " waste=%lu/%lu" , |
6883 | l->count * l->waste, l->waste); |
6884 | |
6885 | if (l->sum_time != l->min_time) { |
6886 | seq_printf(seq, " age=%ld/%llu/%ld" , |
6887 | l->min_time, div_u64(l->sum_time, l->count), |
6888 | l->max_time); |
6889 | } else |
6890 | seq_printf(seq, " age=%ld" , l->min_time); |
6891 | |
6892 | if (l->min_pid != l->max_pid) |
6893 | seq_printf(seq, " pid=%ld-%ld" , l->min_pid, l->max_pid); |
6894 | else |
6895 | seq_printf(seq, " pid=%ld" , |
6896 | l->min_pid); |
6897 | |
6898 | if (num_online_cpus() > 1 && !cpumask_empty(to_cpumask(l->cpus))) |
6899 | seq_printf(seq, " cpus=%*pbl" , |
6900 | cpumask_pr_args(to_cpumask(l->cpus))); |
6901 | |
6902 | if (nr_online_nodes > 1 && !nodes_empty(l->nodes)) |
6903 | seq_printf(seq, " nodes=%*pbl" , |
6904 | nodemask_pr_args(&l->nodes)); |
6905 | |
6906 | #ifdef CONFIG_STACKDEPOT |
6907 | { |
6908 | depot_stack_handle_t handle; |
6909 | unsigned long *entries; |
6910 | unsigned int nr_entries, j; |
6911 | |
6912 | handle = READ_ONCE(l->handle); |
6913 | if (handle) { |
6914 | nr_entries = stack_depot_fetch(handle, &entries); |
6915 | seq_puts(seq, "\n" ); |
6916 | for (j = 0; j < nr_entries; j++) |
6917 | seq_printf(seq, " %pS\n" , (void *)entries[j]); |
6918 | } |
6919 | } |
6920 | #endif |
6921 | seq_puts(seq, "\n" ); |
6922 | } |
6923 | |
6924 | if (!idx && !t->count) |
6925 | seq_puts(seq, "No data\n" ); |
6926 | |
6927 | return 0; |
6928 | } |
6929 | |
6930 | static void slab_debugfs_stop(struct seq_file *seq, void *v) |
6931 | { |
6932 | } |
6933 | |
6934 | static void *slab_debugfs_next(struct seq_file *seq, void *v, loff_t *ppos) |
6935 | { |
6936 | struct loc_track *t = seq->private; |
6937 | |
6938 | t->idx = ++(*ppos); |
6939 | if (*ppos <= t->count) |
6940 | return ppos; |
6941 | |
6942 | return NULL; |
6943 | } |
6944 | |
6945 | static int cmp_loc_by_count(const void *a, const void *b, const void *data) |
6946 | { |
6947 | struct location *loc1 = (struct location *)a; |
6948 | struct location *loc2 = (struct location *)b; |
6949 | |
6950 | if (loc1->count > loc2->count) |
6951 | return -1; |
6952 | else |
6953 | return 1; |
6954 | } |
6955 | |
6956 | static void *slab_debugfs_start(struct seq_file *seq, loff_t *ppos) |
6957 | { |
6958 | struct loc_track *t = seq->private; |
6959 | |
6960 | t->idx = *ppos; |
6961 | return ppos; |
6962 | } |
6963 | |
6964 | static const struct seq_operations slab_debugfs_sops = { |
6965 | .start = slab_debugfs_start, |
6966 | .next = slab_debugfs_next, |
6967 | .stop = slab_debugfs_stop, |
6968 | .show = slab_debugfs_show, |
6969 | }; |
6970 | |
6971 | static int slab_debug_trace_open(struct inode *inode, struct file *filep) |
6972 | { |
6973 | |
6974 | struct kmem_cache_node *n; |
6975 | enum track_item alloc; |
6976 | int node; |
6977 | struct loc_track *t = __seq_open_private(filep, &slab_debugfs_sops, |
6978 | sizeof(struct loc_track)); |
6979 | struct kmem_cache *s = file_inode(filep)->i_private; |
6980 | unsigned long *obj_map; |
6981 | |
6982 | if (!t) |
6983 | return -ENOMEM; |
6984 | |
6985 | obj_map = bitmap_alloc(oo_objects(s->oo), GFP_KERNEL); |
6986 | if (!obj_map) { |
6987 | seq_release_private(inode, filep); |
6988 | return -ENOMEM; |
6989 | } |
6990 | |
6991 | if (strcmp(filep->f_path.dentry->d_name.name, "alloc_traces" ) == 0) |
6992 | alloc = TRACK_ALLOC; |
6993 | else |
6994 | alloc = TRACK_FREE; |
6995 | |
6996 | if (!alloc_loc_track(t, PAGE_SIZE / sizeof(struct location), GFP_KERNEL)) { |
6997 | bitmap_free(obj_map); |
6998 | seq_release_private(inode, filep); |
6999 | return -ENOMEM; |
7000 | } |
7001 | |
7002 | for_each_kmem_cache_node(s, node, n) { |
7003 | unsigned long flags; |
7004 | struct slab *slab; |
7005 | |
7006 | if (!node_nr_slabs(n)) |
7007 | continue; |
7008 | |
7009 | spin_lock_irqsave(&n->list_lock, flags); |
7010 | list_for_each_entry(slab, &n->partial, slab_list) |
7011 | process_slab(t, s, slab, alloc, obj_map); |
7012 | list_for_each_entry(slab, &n->full, slab_list) |
7013 | process_slab(t, s, slab, alloc, obj_map); |
7014 | spin_unlock_irqrestore(&n->list_lock, flags); |
7015 | } |
7016 | |
7017 | /* Sort locations by count */ |
7018 | sort_r(t->loc, t->count, sizeof(struct location), |
7019 | cmp_loc_by_count, NULL, NULL); |
7020 | |
7021 | bitmap_free(obj_map); |
7022 | return 0; |
7023 | } |
7024 | |
7025 | static int slab_debug_trace_release(struct inode *inode, struct file *file) |
7026 | { |
7027 | struct seq_file *seq = file->private_data; |
7028 | struct loc_track *t = seq->private; |
7029 | |
7030 | free_loc_track(t); |
7031 | return seq_release_private(inode, file); |
7032 | } |
7033 | |
7034 | static const struct file_operations slab_debugfs_fops = { |
7035 | .open = slab_debug_trace_open, |
7036 | .read = seq_read, |
7037 | .llseek = seq_lseek, |
7038 | .release = slab_debug_trace_release, |
7039 | }; |
7040 | |
7041 | static void debugfs_slab_add(struct kmem_cache *s) |
7042 | { |
7043 | struct dentry *slab_cache_dir; |
7044 | |
7045 | if (unlikely(!slab_debugfs_root)) |
7046 | return; |
7047 | |
7048 | slab_cache_dir = debugfs_create_dir(s->name, slab_debugfs_root); |
7049 | |
7050 | debugfs_create_file("alloc_traces" , 0400, |
7051 | slab_cache_dir, s, &slab_debugfs_fops); |
7052 | |
7053 | debugfs_create_file("free_traces" , 0400, |
7054 | slab_cache_dir, s, &slab_debugfs_fops); |
7055 | } |
7056 | |
7057 | void debugfs_slab_release(struct kmem_cache *s) |
7058 | { |
7059 | debugfs_lookup_and_remove(s->name, slab_debugfs_root); |
7060 | } |
7061 | |
7062 | static int __init slab_debugfs_init(void) |
7063 | { |
7064 | struct kmem_cache *s; |
7065 | |
7066 | slab_debugfs_root = debugfs_create_dir("slab" , NULL); |
7067 | |
7068 | list_for_each_entry(s, &slab_caches, list) |
7069 | if (s->flags & SLAB_STORE_USER) |
7070 | debugfs_slab_add(s); |
7071 | |
7072 | return 0; |
7073 | |
7074 | } |
7075 | __initcall(slab_debugfs_init); |
7076 | #endif |
7077 | /* |
7078 | * The /proc/slabinfo ABI |
7079 | */ |
7080 | #ifdef CONFIG_SLUB_DEBUG |
7081 | void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo) |
7082 | { |
7083 | unsigned long nr_slabs = 0; |
7084 | unsigned long nr_objs = 0; |
7085 | unsigned long nr_free = 0; |
7086 | int node; |
7087 | struct kmem_cache_node *n; |
7088 | |
7089 | for_each_kmem_cache_node(s, node, n) { |
7090 | nr_slabs += node_nr_slabs(n); |
7091 | nr_objs += node_nr_objs(n); |
7092 | nr_free += count_partial(n, count_free); |
7093 | } |
7094 | |
7095 | sinfo->active_objs = nr_objs - nr_free; |
7096 | sinfo->num_objs = nr_objs; |
7097 | sinfo->active_slabs = nr_slabs; |
7098 | sinfo->num_slabs = nr_slabs; |
7099 | sinfo->objects_per_slab = oo_objects(s->oo); |
7100 | sinfo->cache_order = oo_order(s->oo); |
7101 | } |
7102 | |
7103 | void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s) |
7104 | { |
7105 | } |
7106 | |
7107 | ssize_t slabinfo_write(struct file *file, const char __user *buffer, |
7108 | size_t count, loff_t *ppos) |
7109 | { |
7110 | return -EIO; |
7111 | } |
7112 | #endif /* CONFIG_SLUB_DEBUG */ |
7113 | |