| 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/stacktrace.h> |
| 38 | #include <linux/prefetch.h> |
| 39 | #include <linux/memcontrol.h> |
| 40 | #include <linux/random.h> |
| 41 | #include <kunit/test.h> |
| 42 | #include <kunit/test-bug.h> |
| 43 | #include <linux/sort.h> |
| 44 | |
| 45 | #include <linux/debugfs.h> |
| 46 | #include <trace/events/kmem.h> |
| 47 | |
| 48 | #include "internal.h" |
| 49 | |
| 50 | /* |
| 51 | * Lock order: |
| 52 | * 1. slab_mutex (Global Mutex) |
| 53 | * 2. node->list_lock (Spinlock) |
| 54 | * 3. kmem_cache->cpu_slab->lock (Local lock) |
| 55 | * 4. slab_lock(slab) (Only on some arches) |
| 56 | * 5. object_map_lock (Only for debugging) |
| 57 | * |
| 58 | * slab_mutex |
| 59 | * |
| 60 | * The role of the slab_mutex is to protect the list of all the slabs |
| 61 | * and to synchronize major metadata changes to slab cache structures. |
| 62 | * Also synchronizes memory hotplug callbacks. |
| 63 | * |
| 64 | * slab_lock |
| 65 | * |
| 66 | * The slab_lock is a wrapper around the page lock, thus it is a bit |
| 67 | * spinlock. |
| 68 | * |
| 69 | * The slab_lock is only used on arches that do not have the ability |
| 70 | * to do a cmpxchg_double. It only protects: |
| 71 | * |
| 72 | * A. slab->freelist -> List of free objects in a slab |
| 73 | * B. slab->inuse -> Number of objects in use |
| 74 | * C. slab->objects -> Number of objects in slab |
| 75 | * D. slab->frozen -> frozen state |
| 76 | * |
| 77 | * Frozen slabs |
| 78 | * |
| 79 | * If a slab is frozen then it is exempt from list management. It is not |
| 80 | * on any list except per cpu partial list. The processor that froze the |
| 81 | * slab is the one who can perform list operations on the slab. Other |
| 82 | * processors may put objects onto the freelist but the processor that |
| 83 | * froze the slab is the only one that can retrieve the objects from the |
| 84 | * slab's freelist. |
| 85 | * |
| 86 | * list_lock |
| 87 | * |
| 88 | * The list_lock protects the partial and full list on each node and |
| 89 | * the partial slab counter. If taken then no new slabs may be added or |
| 90 | * removed from the lists nor make the number of partial slabs be modified. |
| 91 | * (Note that the total number of slabs is an atomic value that may be |
| 92 | * modified without taking the list lock). |
| 93 | * |
| 94 | * The list_lock is a centralized lock and thus we avoid taking it as |
| 95 | * much as possible. As long as SLUB does not have to handle partial |
| 96 | * slabs, operations can continue without any centralized lock. F.e. |
| 97 | * allocating a long series of objects that fill up slabs does not require |
| 98 | * the list lock. |
| 99 | * |
| 100 | * For debug caches, all allocations are forced to go through a list_lock |
| 101 | * protected region to serialize against concurrent validation. |
| 102 | * |
| 103 | * cpu_slab->lock local lock |
| 104 | * |
| 105 | * This locks protect slowpath manipulation of all kmem_cache_cpu fields |
| 106 | * except the stat counters. This is a percpu structure manipulated only by |
| 107 | * the local cpu, so the lock protects against being preempted or interrupted |
| 108 | * by an irq. Fast path operations rely on lockless operations instead. |
| 109 | * |
| 110 | * On PREEMPT_RT, the local lock neither disables interrupts nor preemption |
| 111 | * which means the lockless fastpath cannot be used as it might interfere with |
| 112 | * an in-progress slow path operations. In this case the local lock is always |
| 113 | * taken but it still utilizes the freelist for the common operations. |
| 114 | * |
| 115 | * lockless fastpaths |
| 116 | * |
| 117 | * The fast path allocation (slab_alloc_node()) and freeing (do_slab_free()) |
| 118 | * are fully lockless when satisfied from the percpu slab (and when |
| 119 | * cmpxchg_double is possible to use, otherwise slab_lock is taken). |
| 120 | * They also don't disable preemption or migration or irqs. They rely on |
| 121 | * the transaction id (tid) field to detect being preempted or moved to |
| 122 | * another cpu. |
| 123 | * |
| 124 | * irq, preemption, migration considerations |
| 125 | * |
| 126 | * Interrupts are disabled as part of list_lock or local_lock operations, or |
| 127 | * around the slab_lock operation, in order to make the slab allocator safe |
| 128 | * to use in the context of an irq. |
| 129 | * |
| 130 | * In addition, preemption (or migration on PREEMPT_RT) is disabled in the |
| 131 | * allocation slowpath, bulk allocation, and put_cpu_partial(), so that the |
| 132 | * local cpu doesn't change in the process and e.g. the kmem_cache_cpu pointer |
| 133 | * doesn't have to be revalidated in each section protected by the local lock. |
| 134 | * |
| 135 | * SLUB assigns one slab for allocation to each processor. |
| 136 | * Allocations only occur from these slabs called cpu slabs. |
| 137 | * |
| 138 | * Slabs with free elements are kept on a partial list and during regular |
| 139 | * operations no list for full slabs is used. If an object in a full slab is |
| 140 | * freed then the slab will show up again on the partial lists. |
| 141 | * We track full slabs for debugging purposes though because otherwise we |
| 142 | * cannot scan all objects. |
| 143 | * |
| 144 | * Slabs are freed when they become empty. Teardown and setup is |
| 145 | * minimal so we rely on the page allocators per cpu caches for |
| 146 | * fast frees and allocs. |
| 147 | * |
| 148 | * slab->frozen The slab is frozen and exempt from list processing. |
| 149 | * This means that the slab is dedicated to a purpose |
| 150 | * such as satisfying allocations for a specific |
| 151 | * processor. Objects may be freed in the slab while |
| 152 | * it is frozen but slab_free will then skip the usual |
| 153 | * list operations. It is up to the processor holding |
| 154 | * the slab to integrate the slab into the slab lists |
| 155 | * when the slab is no longer needed. |
| 156 | * |
| 157 | * One use of this flag is to mark slabs that are |
| 158 | * used for allocations. Then such a slab becomes a cpu |
| 159 | * slab. The cpu slab may be equipped with an additional |
| 160 | * freelist that allows lockless access to |
| 161 | * free objects in addition to the regular freelist |
| 162 | * that requires the slab lock. |
| 163 | * |
| 164 | * SLAB_DEBUG_FLAGS Slab requires special handling due to debug |
| 165 | * options set. This moves slab handling out of |
| 166 | * the fast path and disables lockless freelists. |
| 167 | */ |
| 168 | |
| 169 | /* |
| 170 | * We could simply use migrate_disable()/enable() but as long as it's a |
| 171 | * function call even on !PREEMPT_RT, use inline preempt_disable() there. |
| 172 | */ |
| 173 | #ifndef CONFIG_PREEMPT_RT |
| 174 | #define slub_get_cpu_ptr(var) get_cpu_ptr(var) |
| 175 | #define slub_put_cpu_ptr(var) put_cpu_ptr(var) |
| 176 | #define USE_LOCKLESS_FAST_PATH() (true) |
| 177 | #else |
| 178 | #define slub_get_cpu_ptr(var) \ |
| 179 | ({ \ |
| 180 | migrate_disable(); \ |
| 181 | this_cpu_ptr(var); \ |
| 182 | }) |
| 183 | #define slub_put_cpu_ptr(var) \ |
| 184 | do { \ |
| 185 | (void)(var); \ |
| 186 | migrate_enable(); \ |
| 187 | } while (0) |
| 188 | #define USE_LOCKLESS_FAST_PATH() (false) |
| 189 | #endif |
| 190 | |
| 191 | #ifndef CONFIG_SLUB_TINY |
| 192 | #define __fastpath_inline __always_inline |
| 193 | #else |
| 194 | #define __fastpath_inline |
| 195 | #endif |
| 196 | |
| 197 | #ifdef CONFIG_SLUB_DEBUG |
| 198 | #ifdef CONFIG_SLUB_DEBUG_ON |
| 199 | DEFINE_STATIC_KEY_TRUE(slub_debug_enabled); |
| 200 | #else |
| 201 | DEFINE_STATIC_KEY_FALSE(slub_debug_enabled); |
| 202 | #endif |
| 203 | #endif /* CONFIG_SLUB_DEBUG */ |
| 204 | |
| 205 | /* Structure holding parameters for get_partial() call chain */ |
| 206 | struct partial_context { |
| 207 | struct slab **slab; |
| 208 | gfp_t flags; |
| 209 | unsigned int orig_size; |
| 210 | }; |
| 211 | |
| 212 | static inline bool kmem_cache_debug(struct kmem_cache *s) |
| 213 | { |
| 214 | return kmem_cache_debug_flags(s, SLAB_DEBUG_FLAGS); |
| 215 | } |
| 216 | |
| 217 | static inline bool slub_debug_orig_size(struct kmem_cache *s) |
| 218 | { |
| 219 | return (kmem_cache_debug_flags(s, SLAB_STORE_USER) && |
| 220 | (s->flags & SLAB_KMALLOC)); |
| 221 | } |
| 222 | |
| 223 | void *fixup_red_left(struct kmem_cache *s, void *p) |
| 224 | { |
| 225 | if (kmem_cache_debug_flags(s, SLAB_RED_ZONE)) |
| 226 | p += s->red_left_pad; |
| 227 | |
| 228 | return p; |
| 229 | } |
| 230 | |
| 231 | static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s) |
| 232 | { |
| 233 | #ifdef CONFIG_SLUB_CPU_PARTIAL |
| 234 | return !kmem_cache_debug(s); |
| 235 | #else |
| 236 | return false; |
| 237 | #endif |
| 238 | } |
| 239 | |
| 240 | /* |
| 241 | * Issues still to be resolved: |
| 242 | * |
| 243 | * - Support PAGE_ALLOC_DEBUG. Should be easy to do. |
| 244 | * |
| 245 | * - Variable sizing of the per node arrays |
| 246 | */ |
| 247 | |
| 248 | /* Enable to log cmpxchg failures */ |
| 249 | #undef SLUB_DEBUG_CMPXCHG |
| 250 | |
| 251 | #ifndef CONFIG_SLUB_TINY |
| 252 | /* |
| 253 | * Minimum number of partial slabs. These will be left on the partial |
| 254 | * lists even if they are empty. kmem_cache_shrink may reclaim them. |
| 255 | */ |
| 256 | #define MIN_PARTIAL 5 |
| 257 | |
| 258 | /* |
| 259 | * Maximum number of desirable partial slabs. |
| 260 | * The existence of more partial slabs makes kmem_cache_shrink |
| 261 | * sort the partial list by the number of objects in use. |
| 262 | */ |
| 263 | #define MAX_PARTIAL 10 |
| 264 | #else |
| 265 | #define MIN_PARTIAL 0 |
| 266 | #define MAX_PARTIAL 0 |
| 267 | #endif |
| 268 | |
| 269 | #define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \ |
| 270 | SLAB_POISON | SLAB_STORE_USER) |
| 271 | |
| 272 | /* |
| 273 | * These debug flags cannot use CMPXCHG because there might be consistency |
| 274 | * issues when checking or reading debug information |
| 275 | */ |
| 276 | #define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \ |
| 277 | SLAB_TRACE) |
| 278 | |
| 279 | |
| 280 | /* |
| 281 | * Debugging flags that require metadata to be stored in the slab. These get |
| 282 | * disabled when slub_debug=O is used and a cache's min order increases with |
| 283 | * metadata. |
| 284 | */ |
| 285 | #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER) |
| 286 | |
| 287 | #define OO_SHIFT 16 |
| 288 | #define OO_MASK ((1 << OO_SHIFT) - 1) |
| 289 | #define MAX_OBJS_PER_PAGE 32767 /* since slab.objects is u15 */ |
| 290 | |
| 291 | /* Internal SLUB flags */ |
| 292 | /* Poison object */ |
| 293 | #define __OBJECT_POISON ((slab_flags_t __force)0x80000000U) |
| 294 | /* Use cmpxchg_double */ |
| 295 | |
| 296 | #ifdef system_has_freelist_aba |
| 297 | #define __CMPXCHG_DOUBLE ((slab_flags_t __force)0x40000000U) |
| 298 | #else |
| 299 | #define __CMPXCHG_DOUBLE ((slab_flags_t __force)0U) |
| 300 | #endif |
| 301 | |
| 302 | /* |
| 303 | * Tracking user of a slab. |
| 304 | */ |
| 305 | #define TRACK_ADDRS_COUNT 16 |
| 306 | struct track { |
| 307 | unsigned long addr; /* Called from address */ |
| 308 | #ifdef CONFIG_STACKDEPOT |
| 309 | depot_stack_handle_t handle; |
| 310 | #endif |
| 311 | int cpu; /* Was running on cpu */ |
| 312 | int pid; /* Pid context */ |
| 313 | unsigned long when; /* When did the operation occur */ |
| 314 | }; |
| 315 | |
| 316 | enum track_item { TRACK_ALLOC, TRACK_FREE }; |
| 317 | |
| 318 | #ifdef SLAB_SUPPORTS_SYSFS |
| 319 | static int sysfs_slab_add(struct kmem_cache *); |
| 320 | static int sysfs_slab_alias(struct kmem_cache *, const char *); |
| 321 | #else |
| 322 | static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; } |
| 323 | static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p) |
| 324 | { return 0; } |
| 325 | #endif |
| 326 | |
| 327 | #if defined(CONFIG_DEBUG_FS) && defined(CONFIG_SLUB_DEBUG) |
| 328 | static void debugfs_slab_add(struct kmem_cache *); |
| 329 | #else |
| 330 | static inline void debugfs_slab_add(struct kmem_cache *s) { } |
| 331 | #endif |
| 332 | |
| 333 | static inline void stat(const struct kmem_cache *s, enum stat_item si) |
| 334 | { |
| 335 | #ifdef CONFIG_SLUB_STATS |
| 336 | /* |
| 337 | * The rmw is racy on a preemptible kernel but this is acceptable, so |
| 338 | * avoid this_cpu_add()'s irq-disable overhead. |
| 339 | */ |
| 340 | raw_cpu_inc(s->cpu_slab->stat[si]); |
| 341 | #endif |
| 342 | } |
| 343 | |
| 344 | /* |
| 345 | * Tracks for which NUMA nodes we have kmem_cache_nodes allocated. |
| 346 | * Corresponds to node_state[N_NORMAL_MEMORY], but can temporarily |
| 347 | * differ during memory hotplug/hotremove operations. |
| 348 | * Protected by slab_mutex. |
| 349 | */ |
| 350 | static nodemask_t slab_nodes; |
| 351 | |
| 352 | #ifndef CONFIG_SLUB_TINY |
| 353 | /* |
| 354 | * Workqueue used for flush_cpu_slab(). |
| 355 | */ |
| 356 | static struct workqueue_struct *flushwq; |
| 357 | #endif |
| 358 | |
| 359 | /******************************************************************** |
| 360 | * Core slab cache functions |
| 361 | *******************************************************************/ |
| 362 | |
| 363 | /* |
| 364 | * freeptr_t represents a SLUB freelist pointer, which might be encoded |
| 365 | * and not dereferenceable if CONFIG_SLAB_FREELIST_HARDENED is enabled. |
| 366 | */ |
| 367 | typedef struct { unsigned long v; } freeptr_t; |
| 368 | |
| 369 | /* |
| 370 | * Returns freelist pointer (ptr). With hardening, this is obfuscated |
| 371 | * with an XOR of the address where the pointer is held and a per-cache |
| 372 | * random number. |
| 373 | */ |
| 374 | static inline freeptr_t freelist_ptr_encode(const struct kmem_cache *s, |
| 375 | void *ptr, unsigned long ptr_addr) |
| 376 | { |
| 377 | unsigned long encoded; |
| 378 | |
| 379 | #ifdef CONFIG_SLAB_FREELIST_HARDENED |
| 380 | encoded = (unsigned long)ptr ^ s->random ^ swab(ptr_addr); |
| 381 | #else |
| 382 | encoded = (unsigned long)ptr; |
| 383 | #endif |
| 384 | return (freeptr_t){.v = encoded}; |
| 385 | } |
| 386 | |
| 387 | static inline void *freelist_ptr_decode(const struct kmem_cache *s, |
| 388 | freeptr_t ptr, unsigned long ptr_addr) |
| 389 | { |
| 390 | void *decoded; |
| 391 | |
| 392 | #ifdef CONFIG_SLAB_FREELIST_HARDENED |
| 393 | decoded = (void *)(ptr.v ^ s->random ^ swab(ptr_addr)); |
| 394 | #else |
| 395 | decoded = (void *)ptr.v; |
| 396 | #endif |
| 397 | return decoded; |
| 398 | } |
| 399 | |
| 400 | /* Returns the freelist pointer recorded at location ptr_addr. */ |
| 401 | static inline void *freelist_dereference(const struct kmem_cache *s, |
| 402 | void *ptr_addr) |
| 403 | { |
| 404 | return freelist_ptr_decode(s, *(freeptr_t *)(ptr_addr), |
| 405 | (unsigned long)ptr_addr); |
| 406 | } |
| 407 | |
| 408 | static inline void *get_freepointer(struct kmem_cache *s, void *object) |
| 409 | { |
| 410 | object = kasan_reset_tag(object); |
| 411 | return freelist_dereference(s, (freeptr_t *)(object + s->offset)); |
| 412 | } |
| 413 | |
| 414 | #ifndef CONFIG_SLUB_TINY |
| 415 | static void prefetch_freepointer(const struct kmem_cache *s, void *object) |
| 416 | { |
| 417 | prefetchw(object + s->offset); |
| 418 | } |
| 419 | #endif |
| 420 | |
| 421 | /* |
| 422 | * When running under KMSAN, get_freepointer_safe() may return an uninitialized |
| 423 | * pointer value in the case the current thread loses the race for the next |
| 424 | * memory chunk in the freelist. In that case this_cpu_cmpxchg_double() in |
| 425 | * slab_alloc_node() will fail, so the uninitialized value won't be used, but |
| 426 | * KMSAN will still check all arguments of cmpxchg because of imperfect |
| 427 | * handling of inline assembly. |
| 428 | * To work around this problem, we apply __no_kmsan_checks to ensure that |
| 429 | * get_freepointer_safe() returns initialized memory. |
| 430 | */ |
| 431 | __no_kmsan_checks |
| 432 | static inline void *get_freepointer_safe(struct kmem_cache *s, void *object) |
| 433 | { |
| 434 | unsigned long freepointer_addr; |
| 435 | freeptr_t p; |
| 436 | |
| 437 | if (!debug_pagealloc_enabled_static()) |
| 438 | return get_freepointer(s, object); |
| 439 | |
| 440 | object = kasan_reset_tag(object); |
| 441 | freepointer_addr = (unsigned long)object + s->offset; |
| 442 | copy_from_kernel_nofault(&p, (freeptr_t *)freepointer_addr, sizeof(p)); |
| 443 | return freelist_ptr_decode(s, p, freepointer_addr); |
| 444 | } |
| 445 | |
| 446 | static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp) |
| 447 | { |
| 448 | unsigned long freeptr_addr = (unsigned long)object + s->offset; |
| 449 | |
| 450 | #ifdef CONFIG_SLAB_FREELIST_HARDENED |
| 451 | BUG_ON(object == fp); /* naive detection of double free or corruption */ |
| 452 | #endif |
| 453 | |
| 454 | freeptr_addr = (unsigned long)kasan_reset_tag((void *)freeptr_addr); |
| 455 | *(freeptr_t *)freeptr_addr = freelist_ptr_encode(s, fp, freeptr_addr); |
| 456 | } |
| 457 | |
| 458 | /* Loop over all objects in a slab */ |
| 459 | #define for_each_object(__p, __s, __addr, __objects) \ |
| 460 | for (__p = fixup_red_left(__s, __addr); \ |
| 461 | __p < (__addr) + (__objects) * (__s)->size; \ |
| 462 | __p += (__s)->size) |
| 463 | |
| 464 | static inline unsigned int order_objects(unsigned int order, unsigned int size) |
| 465 | { |
| 466 | return ((unsigned int)PAGE_SIZE << order) / size; |
| 467 | } |
| 468 | |
| 469 | static inline struct kmem_cache_order_objects oo_make(unsigned int order, |
| 470 | unsigned int size) |
| 471 | { |
| 472 | struct kmem_cache_order_objects x = { |
| 473 | (order << OO_SHIFT) + order_objects(order, size) |
| 474 | }; |
| 475 | |
| 476 | return x; |
| 477 | } |
| 478 | |
| 479 | static inline unsigned int oo_order(struct kmem_cache_order_objects x) |
| 480 | { |
| 481 | return x.x >> OO_SHIFT; |
| 482 | } |
| 483 | |
| 484 | static inline unsigned int oo_objects(struct kmem_cache_order_objects x) |
| 485 | { |
| 486 | return x.x & OO_MASK; |
| 487 | } |
| 488 | |
| 489 | #ifdef CONFIG_SLUB_CPU_PARTIAL |
| 490 | static void slub_set_cpu_partial(struct kmem_cache *s, unsigned int nr_objects) |
| 491 | { |
| 492 | unsigned int nr_slabs; |
| 493 | |
| 494 | s->cpu_partial = nr_objects; |
| 495 | |
| 496 | /* |
| 497 | * We take the number of objects but actually limit the number of |
| 498 | * slabs on the per cpu partial list, in order to limit excessive |
| 499 | * growth of the list. For simplicity we assume that the slabs will |
| 500 | * be half-full. |
| 501 | */ |
| 502 | nr_slabs = DIV_ROUND_UP(nr_objects * 2, oo_objects(s->oo)); |
| 503 | s->cpu_partial_slabs = nr_slabs; |
| 504 | } |
| 505 | #else |
| 506 | static inline void |
| 507 | slub_set_cpu_partial(struct kmem_cache *s, unsigned int nr_objects) |
| 508 | { |
| 509 | } |
| 510 | #endif /* CONFIG_SLUB_CPU_PARTIAL */ |
| 511 | |
| 512 | /* |
| 513 | * Per slab locking using the pagelock |
| 514 | */ |
| 515 | static __always_inline void slab_lock(struct slab *slab) |
| 516 | { |
| 517 | struct page *page = slab_page(slab); |
| 518 | |
| 519 | VM_BUG_ON_PAGE(PageTail(page), page); |
| 520 | bit_spin_lock(PG_locked, &page->flags); |
| 521 | } |
| 522 | |
| 523 | static __always_inline void slab_unlock(struct slab *slab) |
| 524 | { |
| 525 | struct page *page = slab_page(slab); |
| 526 | |
| 527 | VM_BUG_ON_PAGE(PageTail(page), page); |
| 528 | __bit_spin_unlock(PG_locked, &page->flags); |
| 529 | } |
| 530 | |
| 531 | static inline bool |
| 532 | __update_freelist_fast(struct slab *slab, |
| 533 | void *freelist_old, unsigned long counters_old, |
| 534 | void *freelist_new, unsigned long counters_new) |
| 535 | { |
| 536 | #ifdef system_has_freelist_aba |
| 537 | freelist_aba_t old = { .freelist = freelist_old, .counter = counters_old }; |
| 538 | freelist_aba_t new = { .freelist = freelist_new, .counter = counters_new }; |
| 539 | |
| 540 | return try_cmpxchg_freelist(&slab->freelist_counter.full, &old.full, new.full); |
| 541 | #else |
| 542 | return false; |
| 543 | #endif |
| 544 | } |
| 545 | |
| 546 | static inline bool |
| 547 | __update_freelist_slow(struct slab *slab, |
| 548 | void *freelist_old, unsigned long counters_old, |
| 549 | void *freelist_new, unsigned long counters_new) |
| 550 | { |
| 551 | bool ret = false; |
| 552 | |
| 553 | slab_lock(slab); |
| 554 | if (slab->freelist == freelist_old && |
| 555 | slab->counters == counters_old) { |
| 556 | slab->freelist = freelist_new; |
| 557 | slab->counters = counters_new; |
| 558 | ret = true; |
| 559 | } |
| 560 | slab_unlock(slab); |
| 561 | |
| 562 | return ret; |
| 563 | } |
| 564 | |
| 565 | /* |
| 566 | * Interrupts must be disabled (for the fallback code to work right), typically |
| 567 | * by an _irqsave() lock variant. On PREEMPT_RT the preempt_disable(), which is |
| 568 | * part of bit_spin_lock(), is sufficient because the policy is not to allow any |
| 569 | * allocation/ free operation in hardirq context. Therefore nothing can |
| 570 | * interrupt the operation. |
| 571 | */ |
| 572 | static inline bool __slab_update_freelist(struct kmem_cache *s, struct slab *slab, |
| 573 | void *freelist_old, unsigned long counters_old, |
| 574 | void *freelist_new, unsigned long counters_new, |
| 575 | const char *n) |
| 576 | { |
| 577 | bool ret; |
| 578 | |
| 579 | if (USE_LOCKLESS_FAST_PATH()) |
| 580 | lockdep_assert_irqs_disabled(); |
| 581 | |
| 582 | if (s->flags & __CMPXCHG_DOUBLE) { |
| 583 | ret = __update_freelist_fast(slab, freelist_old, counters_old, |
| 584 | freelist_new, counters_new); |
| 585 | } else { |
| 586 | ret = __update_freelist_slow(slab, freelist_old, counters_old, |
| 587 | freelist_new, counters_new); |
| 588 | } |
| 589 | if (likely(ret)) |
| 590 | return true; |
| 591 | |
| 592 | cpu_relax(); |
| 593 | stat(s, CMPXCHG_DOUBLE_FAIL); |
| 594 | |
| 595 | #ifdef SLUB_DEBUG_CMPXCHG |
| 596 | pr_info("%s %s: cmpxchg double redo ", n, s->name); |
| 597 | #endif |
| 598 | |
| 599 | return false; |
| 600 | } |
| 601 | |
| 602 | static inline bool slab_update_freelist(struct kmem_cache *s, struct slab *slab, |
| 603 | void *freelist_old, unsigned long counters_old, |
| 604 | void *freelist_new, unsigned long counters_new, |
| 605 | const char *n) |
| 606 | { |
| 607 | bool ret; |
| 608 | |
| 609 | if (s->flags & __CMPXCHG_DOUBLE) { |
| 610 | ret = __update_freelist_fast(slab, freelist_old, counters_old, |
| 611 | freelist_new, counters_new); |
| 612 | } else { |
| 613 | unsigned long flags; |
| 614 | |
| 615 | local_irq_save(flags); |
| 616 | ret = __update_freelist_slow(slab, freelist_old, counters_old, |
| 617 | freelist_new, counters_new); |
| 618 | local_irq_restore(flags); |
| 619 | } |
| 620 | if (likely(ret)) |
| 621 | return true; |
| 622 | |
| 623 | cpu_relax(); |
| 624 | stat(s, CMPXCHG_DOUBLE_FAIL); |
| 625 | |
| 626 | #ifdef SLUB_DEBUG_CMPXCHG |
| 627 | pr_info("%s %s: cmpxchg double redo ", n, s->name); |
| 628 | #endif |
| 629 | |
| 630 | return false; |
| 631 | } |
| 632 | |
| 633 | #ifdef CONFIG_SLUB_DEBUG |
| 634 | static unsigned long object_map[BITS_TO_LONGS(MAX_OBJS_PER_PAGE)]; |
| 635 | static DEFINE_SPINLOCK(object_map_lock); |
| 636 | |
| 637 | static void __fill_map(unsigned long *obj_map, struct kmem_cache *s, |
| 638 | struct slab *slab) |
| 639 | { |
| 640 | void *addr = slab_address(slab); |
| 641 | void *p; |
| 642 | |
| 643 | bitmap_zero(obj_map, slab->objects); |
| 644 | |
| 645 | for (p = slab->freelist; p; p = get_freepointer(s, p)) |
| 646 | set_bit(__obj_to_index(s, addr, p), obj_map); |
| 647 | } |
| 648 | |
| 649 | #if IS_ENABLED(CONFIG_KUNIT) |
| 650 | static bool slab_add_kunit_errors(void) |
| 651 | { |
| 652 | struct kunit_resource *resource; |
| 653 | |
| 654 | if (!kunit_get_current_test()) |
| 655 | return false; |
| 656 | |
| 657 | resource = kunit_find_named_resource(current->kunit_test, "slab_errors"); |
| 658 | if (!resource) |
| 659 | return false; |
| 660 | |
| 661 | (*(int *)resource->data)++; |
| 662 | kunit_put_resource(resource); |
| 663 | return true; |
| 664 | } |
| 665 | #else |
| 666 | static inline bool slab_add_kunit_errors(void) { return false; } |
| 667 | #endif |
| 668 | |
| 669 | static inline unsigned int size_from_object(struct kmem_cache *s) |
| 670 | { |
| 671 | if (s->flags & SLAB_RED_ZONE) |
| 672 | return s->size - s->red_left_pad; |
| 673 | |
| 674 | return s->size; |
| 675 | } |
| 676 | |
| 677 | static inline void *restore_red_left(struct kmem_cache *s, void *p) |
| 678 | { |
| 679 | if (s->flags & SLAB_RED_ZONE) |
| 680 | p -= s->red_left_pad; |
| 681 | |
| 682 | return p; |
| 683 | } |
| 684 | |
| 685 | /* |
| 686 | * Debug settings: |
| 687 | */ |
| 688 | #if defined(CONFIG_SLUB_DEBUG_ON) |
| 689 | static slab_flags_t slub_debug = DEBUG_DEFAULT_FLAGS; |
| 690 | #else |
| 691 | static slab_flags_t slub_debug; |
| 692 | #endif |
| 693 | |
| 694 | static char *slub_debug_string; |
| 695 | static int disable_higher_order_debug; |
| 696 | |
| 697 | /* |
| 698 | * slub is about to manipulate internal object metadata. This memory lies |
| 699 | * outside the range of the allocated object, so accessing it would normally |
| 700 | * be reported by kasan as a bounds error. metadata_access_enable() is used |
| 701 | * to tell kasan that these accesses are OK. |
| 702 | */ |
| 703 | static inline void metadata_access_enable(void) |
| 704 | { |
| 705 | kasan_disable_current(); |
| 706 | } |
| 707 | |
| 708 | static inline void metadata_access_disable(void) |
| 709 | { |
| 710 | kasan_enable_current(); |
| 711 | } |
| 712 | |
| 713 | /* |
| 714 | * Object debugging |
| 715 | */ |
| 716 | |
| 717 | /* Verify that a pointer has an address that is valid within a slab page */ |
| 718 | static inline int check_valid_pointer(struct kmem_cache *s, |
| 719 | struct slab *slab, void *object) |
| 720 | { |
| 721 | void *base; |
| 722 | |
| 723 | if (!object) |
| 724 | return 1; |
| 725 | |
| 726 | base = slab_address(slab); |
| 727 | object = kasan_reset_tag(object); |
| 728 | object = restore_red_left(s, object); |
| 729 | if (object < base || object >= base + slab->objects * s->size || |
| 730 | (object - base) % s->size) { |
| 731 | return 0; |
| 732 | } |
| 733 | |
| 734 | return 1; |
| 735 | } |
| 736 | |
| 737 | static void print_section(char *level, char *text, u8 *addr, |
| 738 | unsigned int length) |
| 739 | { |
| 740 | metadata_access_enable(); |
| 741 | print_hex_dump(level, text, DUMP_PREFIX_ADDRESS, |
| 742 | 16, 1, kasan_reset_tag((void *)addr), length, 1); |
| 743 | metadata_access_disable(); |
| 744 | } |
| 745 | |
| 746 | /* |
| 747 | * See comment in calculate_sizes(). |
| 748 | */ |
| 749 | static inline bool freeptr_outside_object(struct kmem_cache *s) |
| 750 | { |
| 751 | return s->offset >= s->inuse; |
| 752 | } |
| 753 | |
| 754 | /* |
| 755 | * Return offset of the end of info block which is inuse + free pointer if |
| 756 | * not overlapping with object. |
| 757 | */ |
| 758 | static inline unsigned int get_info_end(struct kmem_cache *s) |
| 759 | { |
| 760 | if (freeptr_outside_object(s)) |
| 761 | return s->inuse + sizeof(void *); |
| 762 | else |
| 763 | return s->inuse; |
| 764 | } |
| 765 | |
| 766 | static struct track *get_track(struct kmem_cache *s, void *object, |
| 767 | enum track_item alloc) |
| 768 | { |
| 769 | struct track *p; |
| 770 | |
| 771 | p = object + get_info_end(s); |
| 772 | |
| 773 | return kasan_reset_tag(p + alloc); |
| 774 | } |
| 775 | |
| 776 | #ifdef CONFIG_STACKDEPOT |
| 777 | static noinline depot_stack_handle_t set_track_prepare(void) |
| 778 | { |
| 779 | depot_stack_handle_t handle; |
| 780 | unsigned long entries[TRACK_ADDRS_COUNT]; |
| 781 | unsigned int nr_entries; |
| 782 | |
| 783 | nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 3); |
| 784 | handle = stack_depot_save(entries, nr_entries, GFP_NOWAIT); |
| 785 | |
| 786 | return handle; |
| 787 | } |
| 788 | #else |
| 789 | static inline depot_stack_handle_t set_track_prepare(void) |
| 790 | { |
| 791 | return 0; |
| 792 | } |
| 793 | #endif |
| 794 | |
| 795 | static void set_track_update(struct kmem_cache *s, void *object, |
| 796 | enum track_item alloc, unsigned long addr, |
| 797 | depot_stack_handle_t handle) |
| 798 | { |
| 799 | struct track *p = get_track(s, object, alloc); |
| 800 | |
| 801 | #ifdef CONFIG_STACKDEPOT |
| 802 | p->handle = handle; |
| 803 | #endif |
| 804 | p->addr = addr; |
| 805 | p->cpu = smp_processor_id(); |
| 806 | p->pid = current->pid; |
| 807 | p->when = jiffies; |
| 808 | } |
| 809 | |
| 810 | static __always_inline void set_track(struct kmem_cache *s, void *object, |
| 811 | enum track_item alloc, unsigned long addr) |
| 812 | { |
| 813 | depot_stack_handle_t handle = set_track_prepare(); |
| 814 | |
| 815 | set_track_update(s, object, alloc, addr, handle); |
| 816 | } |
| 817 | |
| 818 | static void init_tracking(struct kmem_cache *s, void *object) |
| 819 | { |
| 820 | struct track *p; |
| 821 | |
| 822 | if (!(s->flags & SLAB_STORE_USER)) |
| 823 | return; |
| 824 | |
| 825 | p = get_track(s, object, TRACK_ALLOC); |
| 826 | memset(p, 0, 2*sizeof(struct track)); |
| 827 | } |
| 828 | |
| 829 | static void print_track(const char *s, struct track *t, unsigned long pr_time) |
| 830 | { |
| 831 | depot_stack_handle_t handle __maybe_unused; |
| 832 | |
| 833 | if (!t->addr) |
| 834 | return; |
| 835 | |
| 836 | pr_err("%s in %pS age=%lu cpu=%u pid=%d\n", |
| 837 | s, (void *)t->addr, pr_time - t->when, t->cpu, t->pid); |
| 838 | #ifdef CONFIG_STACKDEPOT |
| 839 | handle = READ_ONCE(t->handle); |
| 840 | if (handle) |
| 841 | stack_depot_print(handle); |
| 842 | else |
| 843 | pr_err("object allocation/free stack trace missing\n"); |
| 844 | #endif |
| 845 | } |
| 846 | |
| 847 | void print_tracking(struct kmem_cache *s, void *object) |
| 848 | { |
| 849 | unsigned long pr_time = jiffies; |
| 850 | if (!(s->flags & SLAB_STORE_USER)) |
| 851 | return; |
| 852 | |
| 853 | print_track("Allocated", get_track(s, object, TRACK_ALLOC), pr_time); |
| 854 | print_track("Freed", get_track(s, object, TRACK_FREE), pr_time); |
| 855 | } |
| 856 | |
| 857 | static void print_slab_info(const struct slab *slab) |
| 858 | { |
| 859 | struct folio *folio = (struct folio *)slab_folio(slab); |
| 860 | |
| 861 | pr_err("Slab 0x%p objects=%u used=%u fp=0x%p flags=%pGp\n", |
| 862 | slab, slab->objects, slab->inuse, slab->freelist, |
| 863 | folio_flags(folio, 0)); |
| 864 | } |
| 865 | |
| 866 | /* |
| 867 | * kmalloc caches has fixed sizes (mostly power of 2), and kmalloc() API |
| 868 | * family will round up the real request size to these fixed ones, so |
| 869 | * there could be an extra area than what is requested. Save the original |
| 870 | * request size in the meta data area, for better debug and sanity check. |
| 871 | */ |
| 872 | static inline void set_orig_size(struct kmem_cache *s, |
| 873 | void *object, unsigned int orig_size) |
| 874 | { |
| 875 | void *p = kasan_reset_tag(object); |
| 876 | |
| 877 | if (!slub_debug_orig_size(s)) |
| 878 | return; |
| 879 | |
| 880 | #ifdef CONFIG_KASAN_GENERIC |
| 881 | /* |
| 882 | * KASAN could save its free meta data in object's data area at |
| 883 | * offset 0, if the size is larger than 'orig_size', it will |
| 884 | * overlap the data redzone in [orig_size+1, object_size], and |
| 885 | * the check should be skipped. |
| 886 | */ |
| 887 | if (kasan_metadata_size(s, true) > orig_size) |
| 888 | orig_size = s->object_size; |
| 889 | #endif |
| 890 | |
| 891 | p += get_info_end(s); |
| 892 | p += sizeof(struct track) * 2; |
| 893 | |
| 894 | *(unsigned int *)p = orig_size; |
| 895 | } |
| 896 | |
| 897 | static inline unsigned int get_orig_size(struct kmem_cache *s, void *object) |
| 898 | { |
| 899 | void *p = kasan_reset_tag(object); |
| 900 | |
| 901 | if (!slub_debug_orig_size(s)) |
| 902 | return s->object_size; |
| 903 | |
| 904 | p += get_info_end(s); |
| 905 | p += sizeof(struct track) * 2; |
| 906 | |
| 907 | return *(unsigned int *)p; |
| 908 | } |
| 909 | |
| 910 | void skip_orig_size_check(struct kmem_cache *s, const void *object) |
| 911 | { |
| 912 | set_orig_size(s, (void *)object, s->object_size); |
| 913 | } |
| 914 | |
| 915 | static void slab_bug(struct kmem_cache *s, char *fmt, ...) |
| 916 | { |
| 917 | struct va_format vaf; |
| 918 | va_list args; |
| 919 | |
| 920 | va_start(args, fmt); |
| 921 | vaf.fmt = fmt; |
| 922 | vaf.va = &args; |
| 923 | pr_err("=============================================================================\n"); |
| 924 | pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf); |
| 925 | pr_err("-----------------------------------------------------------------------------\n\n"); |
| 926 | va_end(args); |
| 927 | } |
| 928 | |
| 929 | __printf(2, 3) |
| 930 | static void slab_fix(struct kmem_cache *s, char *fmt, ...) |
| 931 | { |
| 932 | struct va_format vaf; |
| 933 | va_list args; |
| 934 | |
| 935 | if (slab_add_kunit_errors()) |
| 936 | return; |
| 937 | |
| 938 | va_start(args, fmt); |
| 939 | vaf.fmt = fmt; |
| 940 | vaf.va = &args; |
| 941 | pr_err("FIX %s: %pV\n", s->name, &vaf); |
| 942 | va_end(args); |
| 943 | } |
| 944 | |
| 945 | static void print_trailer(struct kmem_cache *s, struct slab *slab, u8 *p) |
| 946 | { |
| 947 | unsigned int off; /* Offset of last byte */ |
| 948 | u8 *addr = slab_address(slab); |
| 949 | |
| 950 | print_tracking(s, p); |
| 951 | |
| 952 | print_slab_info(slab); |
| 953 | |
| 954 | pr_err("Object 0x%p @offset=%tu fp=0x%p\n\n", |
| 955 | p, p - addr, get_freepointer(s, p)); |
| 956 | |
| 957 | if (s->flags & SLAB_RED_ZONE) |
| 958 | print_section(KERN_ERR, "Redzone ", p - s->red_left_pad, |
| 959 | s->red_left_pad); |
| 960 | else if (p > addr + 16) |
| 961 | print_section(KERN_ERR, "Bytes b4 ", p - 16, 16); |
| 962 | |
| 963 | print_section(KERN_ERR, "Object ", p, |
| 964 | min_t(unsigned int, s->object_size, PAGE_SIZE)); |
| 965 | if (s->flags & SLAB_RED_ZONE) |
| 966 | print_section(KERN_ERR, "Redzone ", p + s->object_size, |
| 967 | s->inuse - s->object_size); |
| 968 | |
| 969 | off = get_info_end(s); |
| 970 | |
| 971 | if (s->flags & SLAB_STORE_USER) |
| 972 | off += 2 * sizeof(struct track); |
| 973 | |
| 974 | if (slub_debug_orig_size(s)) |
| 975 | off += sizeof(unsigned int); |
| 976 | |
| 977 | off += kasan_metadata_size(s, false); |
| 978 | |
| 979 | if (off != size_from_object(s)) |
| 980 | /* Beginning of the filler is the free pointer */ |
| 981 | print_section(KERN_ERR, "Padding ", p + off, |
| 982 | size_from_object(s) - off); |
| 983 | |
| 984 | dump_stack(); |
| 985 | } |
| 986 | |
| 987 | static void object_err(struct kmem_cache *s, struct slab *slab, |
| 988 | u8 *object, char *reason) |
| 989 | { |
| 990 | if (slab_add_kunit_errors()) |
| 991 | return; |
| 992 | |
| 993 | slab_bug(s, "%s", reason); |
| 994 | print_trailer(s, slab, object); |
| 995 | add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE); |
| 996 | } |
| 997 | |
| 998 | static bool freelist_corrupted(struct kmem_cache *s, struct slab *slab, |
| 999 | void **freelist, void *nextfree) |
| 1000 | { |
| 1001 | if ((s->flags & SLAB_CONSISTENCY_CHECKS) && |
| 1002 | !check_valid_pointer(s, slab, nextfree) && freelist) { |
| 1003 | object_err(s, slab, *freelist, "Freechain corrupt"); |
| 1004 | *freelist = NULL; |
| 1005 | slab_fix(s, "Isolate corrupted freechain"); |
| 1006 | return true; |
| 1007 | } |
| 1008 | |
| 1009 | return false; |
| 1010 | } |
| 1011 | |
| 1012 | static __printf(3, 4) void slab_err(struct kmem_cache *s, struct slab *slab, |
| 1013 | const char *fmt, ...) |
| 1014 | { |
| 1015 | va_list args; |
| 1016 | char buf[100]; |
| 1017 | |
| 1018 | if (slab_add_kunit_errors()) |
| 1019 | return; |
| 1020 | |
| 1021 | va_start(args, fmt); |
| 1022 | vsnprintf(buf, sizeof(buf), fmt, args); |
| 1023 | va_end(args); |
| 1024 | slab_bug(s, "%s", buf); |
| 1025 | print_slab_info(slab); |
| 1026 | dump_stack(); |
| 1027 | add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE); |
| 1028 | } |
| 1029 | |
| 1030 | static void init_object(struct kmem_cache *s, void *object, u8 val) |
| 1031 | { |
| 1032 | u8 *p = kasan_reset_tag(object); |
| 1033 | unsigned int poison_size = s->object_size; |
| 1034 | |
| 1035 | if (s->flags & SLAB_RED_ZONE) { |
| 1036 | memset(p - s->red_left_pad, val, s->red_left_pad); |
| 1037 | |
| 1038 | if (slub_debug_orig_size(s) && val == SLUB_RED_ACTIVE) { |
| 1039 | /* |
| 1040 | * Redzone the extra allocated space by kmalloc than |
| 1041 | * requested, and the poison size will be limited to |
| 1042 | * the original request size accordingly. |
| 1043 | */ |
| 1044 | poison_size = get_orig_size(s, object); |
| 1045 | } |
| 1046 | } |
| 1047 | |
| 1048 | if (s->flags & __OBJECT_POISON) { |
| 1049 | memset(p, POISON_FREE, poison_size - 1); |
| 1050 | p[poison_size - 1] = POISON_END; |
| 1051 | } |
| 1052 | |
| 1053 | if (s->flags & SLAB_RED_ZONE) |
| 1054 | memset(p + poison_size, val, s->inuse - poison_size); |
| 1055 | } |
| 1056 | |
| 1057 | static void restore_bytes(struct kmem_cache *s, char *message, u8 data, |
| 1058 | void *from, void *to) |
| 1059 | { |
| 1060 | slab_fix(s, "Restoring %s 0x%p-0x%p=0x%x", message, from, to - 1, data); |
| 1061 | memset(from, data, to - from); |
| 1062 | } |
| 1063 | |
| 1064 | static int check_bytes_and_report(struct kmem_cache *s, struct slab *slab, |
| 1065 | u8 *object, char *what, |
| 1066 | u8 *start, unsigned int value, unsigned int bytes) |
| 1067 | { |
| 1068 | u8 *fault; |
| 1069 | u8 *end; |
| 1070 | u8 *addr = slab_address(slab); |
| 1071 | |
| 1072 | metadata_access_enable(); |
| 1073 | fault = memchr_inv(kasan_reset_tag(start), value, bytes); |
| 1074 | metadata_access_disable(); |
| 1075 | if (!fault) |
| 1076 | return 1; |
| 1077 | |
| 1078 | end = start + bytes; |
| 1079 | while (end > fault && end[-1] == value) |
| 1080 | end--; |
| 1081 | |
| 1082 | if (slab_add_kunit_errors()) |
| 1083 | goto skip_bug_print; |
| 1084 | |
| 1085 | slab_bug(s, "%s overwritten", what); |
| 1086 | pr_err("0x%p-0x%p @offset=%tu. First byte 0x%x instead of 0x%x\n", |
| 1087 | fault, end - 1, fault - addr, |
| 1088 | fault[0], value); |
| 1089 | print_trailer(s, slab, object); |
| 1090 | add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE); |
| 1091 | |
| 1092 | skip_bug_print: |
| 1093 | restore_bytes(s, what, value, fault, end); |
| 1094 | return 0; |
| 1095 | } |
| 1096 | |
| 1097 | /* |
| 1098 | * Object layout: |
| 1099 | * |
| 1100 | * object address |
| 1101 | * Bytes of the object to be managed. |
| 1102 | * If the freepointer may overlay the object then the free |
| 1103 | * pointer is at the middle of the object. |
| 1104 | * |
| 1105 | * Poisoning uses 0x6b (POISON_FREE) and the last byte is |
| 1106 | * 0xa5 (POISON_END) |
| 1107 | * |
| 1108 | * object + s->object_size |
| 1109 | * Padding to reach word boundary. This is also used for Redzoning. |
| 1110 | * Padding is extended by another word if Redzoning is enabled and |
| 1111 | * object_size == inuse. |
| 1112 | * |
| 1113 | * We fill with 0xbb (RED_INACTIVE) for inactive objects and with |
| 1114 | * 0xcc (RED_ACTIVE) for objects in use. |
| 1115 | * |
| 1116 | * object + s->inuse |
| 1117 | * Meta data starts here. |
| 1118 | * |
| 1119 | * A. Free pointer (if we cannot overwrite object on free) |
| 1120 | * B. Tracking data for SLAB_STORE_USER |
| 1121 | * C. Original request size for kmalloc object (SLAB_STORE_USER enabled) |
| 1122 | * D. Padding to reach required alignment boundary or at minimum |
| 1123 | * one word if debugging is on to be able to detect writes |
| 1124 | * before the word boundary. |
| 1125 | * |
| 1126 | * Padding is done using 0x5a (POISON_INUSE) |
| 1127 | * |
| 1128 | * object + s->size |
| 1129 | * Nothing is used beyond s->size. |
| 1130 | * |
| 1131 | * If slabcaches are merged then the object_size and inuse boundaries are mostly |
| 1132 | * ignored. And therefore no slab options that rely on these boundaries |
| 1133 | * may be used with merged slabcaches. |
| 1134 | */ |
| 1135 | |
| 1136 | static int check_pad_bytes(struct kmem_cache *s, struct slab *slab, u8 *p) |
| 1137 | { |
| 1138 | unsigned long off = get_info_end(s); /* The end of info */ |
| 1139 | |
| 1140 | if (s->flags & SLAB_STORE_USER) { |
| 1141 | /* We also have user information there */ |
| 1142 | off += 2 * sizeof(struct track); |
| 1143 | |
| 1144 | if (s->flags & SLAB_KMALLOC) |
| 1145 | off += sizeof(unsigned int); |
| 1146 | } |
| 1147 | |
| 1148 | off += kasan_metadata_size(s, false); |
| 1149 | |
| 1150 | if (size_from_object(s) == off) |
| 1151 | return 1; |
| 1152 | |
| 1153 | return check_bytes_and_report(s, slab, p, "Object padding", |
| 1154 | p + off, POISON_INUSE, size_from_object(s) - off); |
| 1155 | } |
| 1156 | |
| 1157 | /* Check the pad bytes at the end of a slab page */ |
| 1158 | static void slab_pad_check(struct kmem_cache *s, struct slab *slab) |
| 1159 | { |
| 1160 | u8 *start; |
| 1161 | u8 *fault; |
| 1162 | u8 *end; |
| 1163 | u8 *pad; |
| 1164 | int length; |
| 1165 | int remainder; |
| 1166 | |
| 1167 | if (!(s->flags & SLAB_POISON)) |
| 1168 | return; |
| 1169 | |
| 1170 | start = slab_address(slab); |
| 1171 | length = slab_size(slab); |
| 1172 | end = start + length; |
| 1173 | remainder = length % s->size; |
| 1174 | if (!remainder) |
| 1175 | return; |
| 1176 | |
| 1177 | pad = end - remainder; |
| 1178 | metadata_access_enable(); |
| 1179 | fault = memchr_inv(kasan_reset_tag(pad), POISON_INUSE, remainder); |
| 1180 | metadata_access_disable(); |
| 1181 | if (!fault) |
| 1182 | return; |
| 1183 | while (end > fault && end[-1] == POISON_INUSE) |
| 1184 | end--; |
| 1185 | |
| 1186 | slab_err(s, slab, "Padding overwritten. 0x%p-0x%p @offset=%tu", |
| 1187 | fault, end - 1, fault - start); |
| 1188 | print_section(KERN_ERR, "Padding ", pad, remainder); |
| 1189 | |
| 1190 | restore_bytes(s, "slab padding", POISON_INUSE, fault, end); |
| 1191 | } |
| 1192 | |
| 1193 | static int check_object(struct kmem_cache *s, struct slab *slab, |
| 1194 | void *object, u8 val) |
| 1195 | { |
| 1196 | u8 *p = object; |
| 1197 | u8 *endobject = object + s->object_size; |
| 1198 | unsigned int orig_size; |
| 1199 | |
| 1200 | if (s->flags & SLAB_RED_ZONE) { |
| 1201 | if (!check_bytes_and_report(s, slab, object, "Left Redzone", |
| 1202 | object - s->red_left_pad, val, s->red_left_pad)) |
| 1203 | return 0; |
| 1204 | |
| 1205 | if (!check_bytes_and_report(s, slab, object, "Right Redzone", |
| 1206 | endobject, val, s->inuse - s->object_size)) |
| 1207 | return 0; |
| 1208 | |
| 1209 | if (slub_debug_orig_size(s) && val == SLUB_RED_ACTIVE) { |
| 1210 | orig_size = get_orig_size(s, object); |
| 1211 | |
| 1212 | if (s->object_size > orig_size && |
| 1213 | !check_bytes_and_report(s, slab, object, |
| 1214 | "kmalloc Redzone", p + orig_size, |
| 1215 | val, s->object_size - orig_size)) { |
| 1216 | return 0; |
| 1217 | } |
| 1218 | } |
| 1219 | } else { |
| 1220 | if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) { |
| 1221 | check_bytes_and_report(s, slab, p, "Alignment padding", |
| 1222 | endobject, POISON_INUSE, |
| 1223 | s->inuse - s->object_size); |
| 1224 | } |
| 1225 | } |
| 1226 | |
| 1227 | if (s->flags & SLAB_POISON) { |
| 1228 | if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) && |
| 1229 | (!check_bytes_and_report(s, slab, p, "Poison", p, |
| 1230 | POISON_FREE, s->object_size - 1) || |
| 1231 | !check_bytes_and_report(s, slab, p, "End Poison", |
| 1232 | p + s->object_size - 1, POISON_END, 1))) |
| 1233 | return 0; |
| 1234 | /* |
| 1235 | * check_pad_bytes cleans up on its own. |
| 1236 | */ |
| 1237 | check_pad_bytes(s, slab, p); |
| 1238 | } |
| 1239 | |
| 1240 | if (!freeptr_outside_object(s) && val == SLUB_RED_ACTIVE) |
| 1241 | /* |
| 1242 | * Object and freepointer overlap. Cannot check |
| 1243 | * freepointer while object is allocated. |
| 1244 | */ |
| 1245 | return 1; |
| 1246 | |
| 1247 | /* Check free pointer validity */ |
| 1248 | if (!check_valid_pointer(s, slab, get_freepointer(s, p))) { |
| 1249 | object_err(s, slab, p, "Freepointer corrupt"); |
| 1250 | /* |
| 1251 | * No choice but to zap it and thus lose the remainder |
| 1252 | * of the free objects in this slab. May cause |
| 1253 | * another error because the object count is now wrong. |
| 1254 | */ |
| 1255 | set_freepointer(s, p, NULL); |
| 1256 | return 0; |
| 1257 | } |
| 1258 | return 1; |
| 1259 | } |
| 1260 | |
| 1261 | static int check_slab(struct kmem_cache *s, struct slab *slab) |
| 1262 | { |
| 1263 | int maxobj; |
| 1264 | |
| 1265 | if (!folio_test_slab(slab_folio(slab))) { |
| 1266 | slab_err(s, slab, "Not a valid slab page"); |
| 1267 | return 0; |
| 1268 | } |
| 1269 | |
| 1270 | maxobj = order_objects(slab_order(slab), s->size); |
| 1271 | if (slab->objects > maxobj) { |
| 1272 | slab_err(s, slab, "objects %u > max %u", |
| 1273 | slab->objects, maxobj); |
| 1274 | return 0; |
| 1275 | } |
| 1276 | if (slab->inuse > slab->objects) { |
| 1277 | slab_err(s, slab, "inuse %u > max %u", |
| 1278 | slab->inuse, slab->objects); |
| 1279 | return 0; |
| 1280 | } |
| 1281 | /* Slab_pad_check fixes things up after itself */ |
| 1282 | slab_pad_check(s, slab); |
| 1283 | return 1; |
| 1284 | } |
| 1285 | |
| 1286 | /* |
| 1287 | * Determine if a certain object in a slab is on the freelist. Must hold the |
| 1288 | * slab lock to guarantee that the chains are in a consistent state. |
| 1289 | */ |
| 1290 | static int on_freelist(struct kmem_cache *s, struct slab *slab, void *search) |
| 1291 | { |
| 1292 | int nr = 0; |
| 1293 | void *fp; |
| 1294 | void *object = NULL; |
| 1295 | int max_objects; |
| 1296 | |
| 1297 | fp = slab->freelist; |
| 1298 | while (fp && nr <= slab->objects) { |
| 1299 | if (fp == search) |
| 1300 | return 1; |
| 1301 | if (!check_valid_pointer(s, slab, fp)) { |
| 1302 | if (object) { |
| 1303 | object_err(s, slab, object, |
| 1304 | "Freechain corrupt"); |
| 1305 | set_freepointer(s, object, NULL); |
| 1306 | } else { |
| 1307 | slab_err(s, slab, "Freepointer corrupt"); |
| 1308 | slab->freelist = NULL; |
| 1309 | slab->inuse = slab->objects; |
| 1310 | slab_fix(s, "Freelist cleared"); |
| 1311 | return 0; |
| 1312 | } |
| 1313 | break; |
| 1314 | } |
| 1315 | object = fp; |
| 1316 | fp = get_freepointer(s, object); |
| 1317 | nr++; |
| 1318 | } |
| 1319 | |
| 1320 | max_objects = order_objects(slab_order(slab), s->size); |
| 1321 | if (max_objects > MAX_OBJS_PER_PAGE) |
| 1322 | max_objects = MAX_OBJS_PER_PAGE; |
| 1323 | |
| 1324 | if (slab->objects != max_objects) { |
| 1325 | slab_err(s, slab, "Wrong number of objects. Found %d but should be %d", |
| 1326 | slab->objects, max_objects); |
| 1327 | slab->objects = max_objects; |
| 1328 | slab_fix(s, "Number of objects adjusted"); |
| 1329 | } |
| 1330 | if (slab->inuse != slab->objects - nr) { |
| 1331 | slab_err(s, slab, "Wrong object count. Counter is %d but counted were %d", |
| 1332 | slab->inuse, slab->objects - nr); |
| 1333 | slab->inuse = slab->objects - nr; |
| 1334 | slab_fix(s, "Object count adjusted"); |
| 1335 | } |
| 1336 | return search == NULL; |
| 1337 | } |
| 1338 | |
| 1339 | static void trace(struct kmem_cache *s, struct slab *slab, void *object, |
| 1340 | int alloc) |
| 1341 | { |
| 1342 | if (s->flags & SLAB_TRACE) { |
| 1343 | pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n", |
| 1344 | s->name, |
| 1345 | alloc ? "alloc" : "free", |
| 1346 | object, slab->inuse, |
| 1347 | slab->freelist); |
| 1348 | |
| 1349 | if (!alloc) |
| 1350 | print_section(KERN_INFO, "Object ", (void *)object, |
| 1351 | s->object_size); |
| 1352 | |
| 1353 | dump_stack(); |
| 1354 | } |
| 1355 | } |
| 1356 | |
| 1357 | /* |
| 1358 | * Tracking of fully allocated slabs for debugging purposes. |
| 1359 | */ |
| 1360 | static void add_full(struct kmem_cache *s, |
| 1361 | struct kmem_cache_node *n, struct slab *slab) |
| 1362 | { |
| 1363 | if (!(s->flags & SLAB_STORE_USER)) |
| 1364 | return; |
| 1365 | |
| 1366 | lockdep_assert_held(&n->list_lock); |
| 1367 | list_add(&slab->slab_list, &n->full); |
| 1368 | } |
| 1369 | |
| 1370 | static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct slab *slab) |
| 1371 | { |
| 1372 | if (!(s->flags & SLAB_STORE_USER)) |
| 1373 | return; |
| 1374 | |
| 1375 | lockdep_assert_held(&n->list_lock); |
| 1376 | list_del(&slab->slab_list); |
| 1377 | } |
| 1378 | |
| 1379 | static inline unsigned long node_nr_slabs(struct kmem_cache_node *n) |
| 1380 | { |
| 1381 | return atomic_long_read(&n->nr_slabs); |
| 1382 | } |
| 1383 | |
| 1384 | static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects) |
| 1385 | { |
| 1386 | struct kmem_cache_node *n = get_node(s, node); |
| 1387 | |
| 1388 | /* |
| 1389 | * May be called early in order to allocate a slab for the |
| 1390 | * kmem_cache_node structure. Solve the chicken-egg |
| 1391 | * dilemma by deferring the increment of the count during |
| 1392 | * bootstrap (see early_kmem_cache_node_alloc). |
| 1393 | */ |
| 1394 | if (likely(n)) { |
| 1395 | atomic_long_inc(&n->nr_slabs); |
| 1396 | atomic_long_add(objects, &n->total_objects); |
| 1397 | } |
| 1398 | } |
| 1399 | static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects) |
| 1400 | { |
| 1401 | struct kmem_cache_node *n = get_node(s, node); |
| 1402 | |
| 1403 | atomic_long_dec(&n->nr_slabs); |
| 1404 | atomic_long_sub(objects, &n->total_objects); |
| 1405 | } |
| 1406 | |
| 1407 | /* Object debug checks for alloc/free paths */ |
| 1408 | static void setup_object_debug(struct kmem_cache *s, void *object) |
| 1409 | { |
| 1410 | if (!kmem_cache_debug_flags(s, SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)) |
| 1411 | return; |
| 1412 | |
| 1413 | init_object(s, object, SLUB_RED_INACTIVE); |
| 1414 | init_tracking(s, object); |
| 1415 | } |
| 1416 | |
| 1417 | static |
| 1418 | void setup_slab_debug(struct kmem_cache *s, struct slab *slab, void *addr) |
| 1419 | { |
| 1420 | if (!kmem_cache_debug_flags(s, SLAB_POISON)) |
| 1421 | return; |
| 1422 | |
| 1423 | metadata_access_enable(); |
| 1424 | memset(kasan_reset_tag(addr), POISON_INUSE, slab_size(slab)); |
| 1425 | metadata_access_disable(); |
| 1426 | } |
| 1427 | |
| 1428 | static inline int alloc_consistency_checks(struct kmem_cache *s, |
| 1429 | struct slab *slab, void *object) |
| 1430 | { |
| 1431 | if (!check_slab(s, slab)) |
| 1432 | return 0; |
| 1433 | |
| 1434 | if (!check_valid_pointer(s, slab, object)) { |
| 1435 | object_err(s, slab, object, "Freelist Pointer check fails"); |
| 1436 | return 0; |
| 1437 | } |
| 1438 | |
| 1439 | if (!check_object(s, slab, object, SLUB_RED_INACTIVE)) |
| 1440 | return 0; |
| 1441 | |
| 1442 | return 1; |
| 1443 | } |
| 1444 | |
| 1445 | static noinline bool alloc_debug_processing(struct kmem_cache *s, |
| 1446 | struct slab *slab, void *object, int orig_size) |
| 1447 | { |
| 1448 | if (s->flags & SLAB_CONSISTENCY_CHECKS) { |
| 1449 | if (!alloc_consistency_checks(s, slab, object)) |
| 1450 | goto bad; |
| 1451 | } |
| 1452 | |
| 1453 | /* Success. Perform special debug activities for allocs */ |
| 1454 | trace(s, slab, object, 1); |
| 1455 | set_orig_size(s, object, orig_size); |
| 1456 | init_object(s, object, SLUB_RED_ACTIVE); |
| 1457 | return true; |
| 1458 | |
| 1459 | bad: |
| 1460 | if (folio_test_slab(slab_folio(slab))) { |
| 1461 | /* |
| 1462 | * If this is a slab page then lets do the best we can |
| 1463 | * to avoid issues in the future. Marking all objects |
| 1464 | * as used avoids touching the remaining objects. |
| 1465 | */ |
| 1466 | slab_fix(s, "Marking all objects used"); |
| 1467 | slab->inuse = slab->objects; |
| 1468 | slab->freelist = NULL; |
| 1469 | } |
| 1470 | return false; |
| 1471 | } |
| 1472 | |
| 1473 | static inline int free_consistency_checks(struct kmem_cache *s, |
| 1474 | struct slab *slab, void *object, unsigned long addr) |
| 1475 | { |
| 1476 | if (!check_valid_pointer(s, slab, object)) { |
| 1477 | slab_err(s, slab, "Invalid object pointer 0x%p", object); |
| 1478 | return 0; |
| 1479 | } |
| 1480 | |
| 1481 | if (on_freelist(s, slab, object)) { |
| 1482 | object_err(s, slab, object, "Object already free"); |
| 1483 | return 0; |
| 1484 | } |
| 1485 | |
| 1486 | if (!check_object(s, slab, object, SLUB_RED_ACTIVE)) |
| 1487 | return 0; |
| 1488 | |
| 1489 | if (unlikely(s != slab->slab_cache)) { |
| 1490 | if (!folio_test_slab(slab_folio(slab))) { |
| 1491 | slab_err(s, slab, "Attempt to free object(0x%p) outside of slab", |
| 1492 | object); |
| 1493 | } else if (!slab->slab_cache) { |
| 1494 | pr_err("SLUB <none>: no slab for object 0x%p.\n", |
| 1495 | object); |
| 1496 | dump_stack(); |
| 1497 | } else |
| 1498 | object_err(s, slab, object, |
| 1499 | "page slab pointer corrupt."); |
| 1500 | return 0; |
| 1501 | } |
| 1502 | return 1; |
| 1503 | } |
| 1504 | |
| 1505 | /* |
| 1506 | * Parse a block of slub_debug options. Blocks are delimited by ';' |
| 1507 | * |
| 1508 | * @str: start of block |
| 1509 | * @flags: returns parsed flags, or DEBUG_DEFAULT_FLAGS if none specified |
| 1510 | * @slabs: return start of list of slabs, or NULL when there's no list |
| 1511 | * @init: assume this is initial parsing and not per-kmem-create parsing |
| 1512 | * |
| 1513 | * returns the start of next block if there's any, or NULL |
| 1514 | */ |
| 1515 | static char * |
| 1516 | parse_slub_debug_flags(char *str, slab_flags_t *flags, char **slabs, bool init) |
| 1517 | { |
| 1518 | bool higher_order_disable = false; |
| 1519 | |
| 1520 | /* Skip any completely empty blocks */ |
| 1521 | while (*str && *str == ';') |
| 1522 | str++; |
| 1523 | |
| 1524 | if (*str == ',') { |
| 1525 | /* |
| 1526 | * No options but restriction on slabs. This means full |
| 1527 | * debugging for slabs matching a pattern. |
| 1528 | */ |
| 1529 | *flags = DEBUG_DEFAULT_FLAGS; |
| 1530 | goto check_slabs; |
| 1531 | } |
| 1532 | *flags = 0; |
| 1533 | |
| 1534 | /* Determine which debug features should be switched on */ |
| 1535 | for (; *str && *str != ',' && *str != ';'; str++) { |
| 1536 | switch (tolower(*str)) { |
| 1537 | case '-': |
| 1538 | *flags = 0; |
| 1539 | break; |
| 1540 | case 'f': |
| 1541 | *flags |= SLAB_CONSISTENCY_CHECKS; |
| 1542 | break; |
| 1543 | case 'z': |
| 1544 | *flags |= SLAB_RED_ZONE; |
| 1545 | break; |
| 1546 | case 'p': |
| 1547 | *flags |= SLAB_POISON; |
| 1548 | break; |
| 1549 | case 'u': |
| 1550 | *flags |= SLAB_STORE_USER; |
| 1551 | break; |
| 1552 | case 't': |
| 1553 | *flags |= SLAB_TRACE; |
| 1554 | break; |
| 1555 | case 'a': |
| 1556 | *flags |= SLAB_FAILSLAB; |
| 1557 | break; |
| 1558 | case 'o': |
| 1559 | /* |
| 1560 | * Avoid enabling debugging on caches if its minimum |
| 1561 | * order would increase as a result. |
| 1562 | */ |
| 1563 | higher_order_disable = true; |
| 1564 | break; |
| 1565 | default: |
| 1566 | if (init) |
| 1567 | pr_err("slub_debug option '%c' unknown. skipped\n", *str); |
| 1568 | } |
| 1569 | } |
| 1570 | check_slabs: |
| 1571 | if (*str == ',') |
| 1572 | *slabs = ++str; |
| 1573 | else |
| 1574 | *slabs = NULL; |
| 1575 | |
| 1576 | /* Skip over the slab list */ |
| 1577 | while (*str && *str != ';') |
| 1578 | str++; |
| 1579 | |
| 1580 | /* Skip any completely empty blocks */ |
| 1581 | while (*str && *str == ';') |
| 1582 | str++; |
| 1583 | |
| 1584 | if (init && higher_order_disable) |
| 1585 | disable_higher_order_debug = 1; |
| 1586 | |
| 1587 | if (*str) |
| 1588 | return str; |
| 1589 | else |
| 1590 | return NULL; |
| 1591 | } |
| 1592 | |
| 1593 | static int __init setup_slub_debug(char *str) |
| 1594 | { |
| 1595 | slab_flags_t flags; |
| 1596 | slab_flags_t global_flags; |
| 1597 | char *saved_str; |
| 1598 | char *slab_list; |
| 1599 | bool global_slub_debug_changed = false; |
| 1600 | bool slab_list_specified = false; |
| 1601 | |
| 1602 | global_flags = DEBUG_DEFAULT_FLAGS; |
| 1603 | if (*str++ != '=' || !*str) |
| 1604 | /* |
| 1605 | * No options specified. Switch on full debugging. |
| 1606 | */ |
| 1607 | goto out; |
| 1608 | |
| 1609 | saved_str = str; |
| 1610 | while (str) { |
| 1611 | str = parse_slub_debug_flags(str, &flags, &slab_list, true); |
| 1612 | |
| 1613 | if (!slab_list) { |
| 1614 | global_flags = flags; |
| 1615 | global_slub_debug_changed = true; |
| 1616 | } else { |
| 1617 | slab_list_specified = true; |
| 1618 | if (flags & SLAB_STORE_USER) |
| 1619 | stack_depot_request_early_init(); |
| 1620 | } |
| 1621 | } |
| 1622 | |
| 1623 | /* |
| 1624 | * For backwards compatibility, a single list of flags with list of |
| 1625 | * slabs means debugging is only changed for those slabs, so the global |
| 1626 | * slub_debug should be unchanged (0 or DEBUG_DEFAULT_FLAGS, depending |
| 1627 | * on CONFIG_SLUB_DEBUG_ON). We can extended that to multiple lists as |
| 1628 | * long as there is no option specifying flags without a slab list. |
| 1629 | */ |
| 1630 | if (slab_list_specified) { |
| 1631 | if (!global_slub_debug_changed) |
| 1632 | global_flags = slub_debug; |
| 1633 | slub_debug_string = saved_str; |
| 1634 | } |
| 1635 | out: |
| 1636 | slub_debug = global_flags; |
| 1637 | if (slub_debug & SLAB_STORE_USER) |
| 1638 | stack_depot_request_early_init(); |
| 1639 | if (slub_debug != 0 || slub_debug_string) |
| 1640 | static_branch_enable(&slub_debug_enabled); |
| 1641 | else |
| 1642 | static_branch_disable(&slub_debug_enabled); |
| 1643 | if ((static_branch_unlikely(&init_on_alloc) || |
| 1644 | static_branch_unlikely(&init_on_free)) && |
| 1645 | (slub_debug & SLAB_POISON)) |
| 1646 | pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n"); |
| 1647 | return 1; |
| 1648 | } |
| 1649 | |
| 1650 | __setup("slub_debug", setup_slub_debug); |
| 1651 | |
| 1652 | /* |
| 1653 | * kmem_cache_flags - apply debugging options to the cache |
| 1654 | * @object_size: the size of an object without meta data |
| 1655 | * @flags: flags to set |
| 1656 | * @name: name of the cache |
| 1657 | * |
| 1658 | * Debug option(s) are applied to @flags. In addition to the debug |
| 1659 | * option(s), if a slab name (or multiple) is specified i.e. |
| 1660 | * slub_debug=<Debug-Options>,<slab name1>,<slab name2> ... |
| 1661 | * then only the select slabs will receive the debug option(s). |
| 1662 | */ |
| 1663 | slab_flags_t kmem_cache_flags(unsigned int object_size, |
| 1664 | slab_flags_t flags, const char *name) |
| 1665 | { |
| 1666 | char *iter; |
| 1667 | size_t len; |
| 1668 | char *next_block; |
| 1669 | slab_flags_t block_flags; |
| 1670 | slab_flags_t slub_debug_local = slub_debug; |
| 1671 | |
| 1672 | if (flags & SLAB_NO_USER_FLAGS) |
| 1673 | return flags; |
| 1674 | |
| 1675 | /* |
| 1676 | * If the slab cache is for debugging (e.g. kmemleak) then |
| 1677 | * don't store user (stack trace) information by default, |
| 1678 | * but let the user enable it via the command line below. |
| 1679 | */ |
| 1680 | if (flags & SLAB_NOLEAKTRACE) |
| 1681 | slub_debug_local &= ~SLAB_STORE_USER; |
| 1682 | |
| 1683 | len = strlen(name); |
| 1684 | next_block = slub_debug_string; |
| 1685 | /* Go through all blocks of debug options, see if any matches our slab's name */ |
| 1686 | while (next_block) { |
| 1687 | next_block = parse_slub_debug_flags(next_block, &block_flags, &iter, false); |
| 1688 | if (!iter) |
| 1689 | continue; |
| 1690 | /* Found a block that has a slab list, search it */ |
| 1691 | while (*iter) { |
| 1692 | char *end, *glob; |
| 1693 | size_t cmplen; |
| 1694 | |
| 1695 | end = strchrnul(iter, ','); |
| 1696 | if (next_block && next_block < end) |
| 1697 | end = next_block - 1; |
| 1698 | |
| 1699 | glob = strnchr(iter, end - iter, '*'); |
| 1700 | if (glob) |
| 1701 | cmplen = glob - iter; |
| 1702 | else |
| 1703 | cmplen = max_t(size_t, len, (end - iter)); |
| 1704 | |
| 1705 | if (!strncmp(name, iter, cmplen)) { |
| 1706 | flags |= block_flags; |
| 1707 | return flags; |
| 1708 | } |
| 1709 | |
| 1710 | if (!*end || *end == ';') |
| 1711 | break; |
| 1712 | iter = end + 1; |
| 1713 | } |
| 1714 | } |
| 1715 | |
| 1716 | return flags | slub_debug_local; |
| 1717 | } |
| 1718 | #else /* !CONFIG_SLUB_DEBUG */ |
| 1719 | static inline void setup_object_debug(struct kmem_cache *s, void *object) {} |
| 1720 | static inline |
| 1721 | void setup_slab_debug(struct kmem_cache *s, struct slab *slab, void *addr) {} |
| 1722 | |
| 1723 | static inline bool alloc_debug_processing(struct kmem_cache *s, |
| 1724 | struct slab *slab, void *object, int orig_size) { return true; } |
| 1725 | |
| 1726 | static inline bool free_debug_processing(struct kmem_cache *s, |
| 1727 | struct slab *slab, void *head, void *tail, int *bulk_cnt, |
| 1728 | unsigned long addr, depot_stack_handle_t handle) { return true; } |
| 1729 | |
| 1730 | static inline void slab_pad_check(struct kmem_cache *s, struct slab *slab) {} |
| 1731 | static inline int check_object(struct kmem_cache *s, struct slab *slab, |
| 1732 | void *object, u8 val) { return 1; } |
| 1733 | static inline depot_stack_handle_t set_track_prepare(void) { return 0; } |
| 1734 | static inline void set_track(struct kmem_cache *s, void *object, |
| 1735 | enum track_item alloc, unsigned long addr) {} |
| 1736 | static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n, |
| 1737 | struct slab *slab) {} |
| 1738 | static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, |
| 1739 | struct slab *slab) {} |
| 1740 | slab_flags_t kmem_cache_flags(unsigned int object_size, |
| 1741 | slab_flags_t flags, const char *name) |
| 1742 | { |
| 1743 | return flags; |
| 1744 | } |
| 1745 | #define slub_debug 0 |
| 1746 | |
| 1747 | #define disable_higher_order_debug 0 |
| 1748 | |
| 1749 | static inline unsigned long node_nr_slabs(struct kmem_cache_node *n) |
| 1750 | { return 0; } |
| 1751 | static inline void inc_slabs_node(struct kmem_cache *s, int node, |
| 1752 | int objects) {} |
| 1753 | static inline void dec_slabs_node(struct kmem_cache *s, int node, |
| 1754 | int objects) {} |
| 1755 | |
| 1756 | #ifndef CONFIG_SLUB_TINY |
| 1757 | static bool freelist_corrupted(struct kmem_cache *s, struct slab *slab, |
| 1758 | void **freelist, void *nextfree) |
| 1759 | { |
| 1760 | return false; |
| 1761 | } |
| 1762 | #endif |
| 1763 | #endif /* CONFIG_SLUB_DEBUG */ |
| 1764 | |
| 1765 | /* |
| 1766 | * Hooks for other subsystems that check memory allocations. In a typical |
| 1767 | * production configuration these hooks all should produce no code at all. |
| 1768 | */ |
| 1769 | static __always_inline bool slab_free_hook(struct kmem_cache *s, |
| 1770 | void *x, bool init) |
| 1771 | { |
| 1772 | kmemleak_free_recursive(x, s->flags); |
| 1773 | kmsan_slab_free(s, x); |
| 1774 | |
| 1775 | debug_check_no_locks_freed(x, s->object_size); |
| 1776 | |
| 1777 | if (!(s->flags & SLAB_DEBUG_OBJECTS)) |
| 1778 | debug_check_no_obj_freed(x, s->object_size); |
| 1779 | |
| 1780 | /* Use KCSAN to help debug racy use-after-free. */ |
| 1781 | if (!(s->flags & SLAB_TYPESAFE_BY_RCU)) |
| 1782 | __kcsan_check_access(x, s->object_size, |
| 1783 | KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT); |
| 1784 | |
| 1785 | /* |
| 1786 | * As memory initialization might be integrated into KASAN, |
| 1787 | * kasan_slab_free and initialization memset's must be |
| 1788 | * kept together to avoid discrepancies in behavior. |
| 1789 | * |
| 1790 | * The initialization memset's clear the object and the metadata, |
| 1791 | * but don't touch the SLAB redzone. |
| 1792 | */ |
| 1793 | if (init) { |
| 1794 | int rsize; |
| 1795 | |
| 1796 | if (!kasan_has_integrated_init()) |
| 1797 | memset(kasan_reset_tag(x), 0, s->object_size); |
| 1798 | rsize = (s->flags & SLAB_RED_ZONE) ? s->red_left_pad : 0; |
| 1799 | memset((char *)kasan_reset_tag(x) + s->inuse, 0, |
| 1800 | s->size - s->inuse - rsize); |
| 1801 | } |
| 1802 | /* KASAN might put x into memory quarantine, delaying its reuse. */ |
| 1803 | return kasan_slab_free(s, x, init); |
| 1804 | } |
| 1805 | |
| 1806 | static inline bool slab_free_freelist_hook(struct kmem_cache *s, |
| 1807 | void **head, void **tail, |
| 1808 | int *cnt) |
| 1809 | { |
| 1810 | |
| 1811 | void *object; |
| 1812 | void *next = *head; |
| 1813 | void *old_tail = *tail ? *tail : *head; |
| 1814 | |
| 1815 | if (is_kfence_address(next)) { |
| 1816 | slab_free_hook(s, next, false); |
| 1817 | return true; |
| 1818 | } |
| 1819 | |
| 1820 | /* Head and tail of the reconstructed freelist */ |
| 1821 | *head = NULL; |
| 1822 | *tail = NULL; |
| 1823 | |
| 1824 | do { |
| 1825 | object = next; |
| 1826 | next = get_freepointer(s, object); |
| 1827 | |
| 1828 | /* If object's reuse doesn't have to be delayed */ |
| 1829 | if (!slab_free_hook(s, object, slab_want_init_on_free(s))) { |
| 1830 | /* Move object to the new freelist */ |
| 1831 | set_freepointer(s, object, *head); |
| 1832 | *head = object; |
| 1833 | if (!*tail) |
| 1834 | *tail = object; |
| 1835 | } else { |
| 1836 | /* |
| 1837 | * Adjust the reconstructed freelist depth |
| 1838 | * accordingly if object's reuse is delayed. |
| 1839 | */ |
| 1840 | --(*cnt); |
| 1841 | } |
| 1842 | } while (object != old_tail); |
| 1843 | |
| 1844 | if (*head == *tail) |
| 1845 | *tail = NULL; |
| 1846 | |
| 1847 | return *head != NULL; |
| 1848 | } |
| 1849 | |
| 1850 | static void *setup_object(struct kmem_cache *s, void *object) |
| 1851 | { |
| 1852 | setup_object_debug(s, object); |
| 1853 | object = kasan_init_slab_obj(s, object); |
| 1854 | if (unlikely(s->ctor)) { |
| 1855 | kasan_unpoison_object_data(s, object); |
| 1856 | s->ctor(object); |
| 1857 | kasan_poison_object_data(s, object); |
| 1858 | } |
| 1859 | return object; |
| 1860 | } |
| 1861 | |
| 1862 | /* |
| 1863 | * Slab allocation and freeing |
| 1864 | */ |
| 1865 | static inline struct slab *alloc_slab_page(gfp_t flags, int node, |
| 1866 | struct kmem_cache_order_objects oo) |
| 1867 | { |
| 1868 | struct folio *folio; |
| 1869 | struct slab *slab; |
| 1870 | unsigned int order = oo_order(oo); |
| 1871 | |
| 1872 | if (node == NUMA_NO_NODE) |
| 1873 | folio = (struct folio *)alloc_pages(flags, order); |
| 1874 | else |
| 1875 | folio = (struct folio *)__alloc_pages_node(node, flags, order); |
| 1876 | |
| 1877 | if (!folio) |
| 1878 | return NULL; |
| 1879 | |
| 1880 | slab = folio_slab(folio); |
| 1881 | __folio_set_slab(folio); |
| 1882 | /* Make the flag visible before any changes to folio->mapping */ |
| 1883 | smp_wmb(); |
| 1884 | if (folio_is_pfmemalloc(folio)) |
| 1885 | slab_set_pfmemalloc(slab); |
| 1886 | |
| 1887 | return slab; |
| 1888 | } |
| 1889 | |
| 1890 | #ifdef CONFIG_SLAB_FREELIST_RANDOM |
| 1891 | /* Pre-initialize the random sequence cache */ |
| 1892 | static int init_cache_random_seq(struct kmem_cache *s) |
| 1893 | { |
| 1894 | unsigned int count = oo_objects(s->oo); |
| 1895 | int err; |
| 1896 | |
| 1897 | /* Bailout if already initialised */ |
| 1898 | if (s->random_seq) |
| 1899 | return 0; |
| 1900 | |
| 1901 | err = cache_random_seq_create(s, count, GFP_KERNEL); |
| 1902 | if (err) { |
| 1903 | pr_err("SLUB: Unable to initialize free list for %s\n", |
| 1904 | s->name); |
| 1905 | return err; |
| 1906 | } |
| 1907 | |
| 1908 | /* Transform to an offset on the set of pages */ |
| 1909 | if (s->random_seq) { |
| 1910 | unsigned int i; |
| 1911 | |
| 1912 | for (i = 0; i < count; i++) |
| 1913 | s->random_seq[i] *= s->size; |
| 1914 | } |
| 1915 | return 0; |
| 1916 | } |
| 1917 | |
| 1918 | /* Initialize each random sequence freelist per cache */ |
| 1919 | static void __init init_freelist_randomization(void) |
| 1920 | { |
| 1921 | struct kmem_cache *s; |
| 1922 | |
| 1923 | mutex_lock(&slab_mutex); |
| 1924 | |
| 1925 | list_for_each_entry(s, &slab_caches, list) |
| 1926 | init_cache_random_seq(s); |
| 1927 | |
| 1928 | mutex_unlock(&slab_mutex); |
| 1929 | } |
| 1930 | |
| 1931 | /* Get the next entry on the pre-computed freelist randomized */ |
| 1932 | static void *next_freelist_entry(struct kmem_cache *s, struct slab *slab, |
| 1933 | unsigned long *pos, void *start, |
| 1934 | unsigned long page_limit, |
| 1935 | unsigned long freelist_count) |
| 1936 | { |
| 1937 | unsigned int idx; |
| 1938 | |
| 1939 | /* |
| 1940 | * If the target page allocation failed, the number of objects on the |
| 1941 | * page might be smaller than the usual size defined by the cache. |
| 1942 | */ |
| 1943 | do { |
| 1944 | idx = s->random_seq[*pos]; |
| 1945 | *pos += 1; |
| 1946 | if (*pos >= freelist_count) |
| 1947 | *pos = 0; |
| 1948 | } while (unlikely(idx >= page_limit)); |
| 1949 | |
| 1950 | return (char *)start + idx; |
| 1951 | } |
| 1952 | |
| 1953 | /* Shuffle the single linked freelist based on a random pre-computed sequence */ |
| 1954 | static bool shuffle_freelist(struct kmem_cache *s, struct slab *slab) |
| 1955 | { |
| 1956 | void *start; |
| 1957 | void *cur; |
| 1958 | void *next; |
| 1959 | unsigned long idx, pos, page_limit, freelist_count; |
| 1960 | |
| 1961 | if (slab->objects < 2 || !s->random_seq) |
| 1962 | return false; |
| 1963 | |
| 1964 | freelist_count = oo_objects(s->oo); |
| 1965 | pos = get_random_u32_below(freelist_count); |
| 1966 | |
| 1967 | page_limit = slab->objects * s->size; |
| 1968 | start = fixup_red_left(s, slab_address(slab)); |
| 1969 | |
| 1970 | /* First entry is used as the base of the freelist */ |
| 1971 | cur = next_freelist_entry(s, slab, &pos, start, page_limit, |
| 1972 | freelist_count); |
| 1973 | cur = setup_object(s, cur); |
| 1974 | slab->freelist = cur; |
| 1975 | |
| 1976 | for (idx = 1; idx < slab->objects; idx++) { |
| 1977 | next = next_freelist_entry(s, slab, &pos, start, page_limit, |
| 1978 | freelist_count); |
| 1979 | next = setup_object(s, next); |
| 1980 | set_freepointer(s, cur, next); |
| 1981 | cur = next; |
| 1982 | } |
| 1983 | set_freepointer(s, cur, NULL); |
| 1984 | |
| 1985 | return true; |
| 1986 | } |
| 1987 | #else |
| 1988 | static inline int init_cache_random_seq(struct kmem_cache *s) |
| 1989 | { |
| 1990 | return 0; |
| 1991 | } |
| 1992 | static inline void init_freelist_randomization(void) { } |
| 1993 | static inline bool shuffle_freelist(struct kmem_cache *s, struct slab *slab) |
| 1994 | { |
| 1995 | return false; |
| 1996 | } |
| 1997 | #endif /* CONFIG_SLAB_FREELIST_RANDOM */ |
| 1998 | |
| 1999 | static struct slab *allocate_slab(struct kmem_cache *s, gfp_t flags, int node) |
| 2000 | { |
| 2001 | struct slab *slab; |
| 2002 | struct kmem_cache_order_objects oo = s->oo; |
| 2003 | gfp_t alloc_gfp; |
| 2004 | void *start, *p, *next; |
| 2005 | int idx; |
| 2006 | bool shuffle; |
| 2007 | |
| 2008 | flags &= gfp_allowed_mask; |
| 2009 | |
| 2010 | flags |= s->allocflags; |
| 2011 | |
| 2012 | /* |
| 2013 | * Let the initial higher-order allocation fail under memory pressure |
| 2014 | * so we fall-back to the minimum order allocation. |
| 2015 | */ |
| 2016 | alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL; |
| 2017 | if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min)) |
| 2018 | alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~__GFP_RECLAIM; |
| 2019 | |
| 2020 | slab = alloc_slab_page(alloc_gfp, node, oo); |
| 2021 | if (unlikely(!slab)) { |
| 2022 | oo = s->min; |
| 2023 | alloc_gfp = flags; |
| 2024 | /* |
| 2025 | * Allocation may have failed due to fragmentation. |
| 2026 | * Try a lower order alloc if possible |
| 2027 | */ |
| 2028 | slab = alloc_slab_page(alloc_gfp, node, oo); |
| 2029 | if (unlikely(!slab)) |
| 2030 | return NULL; |
| 2031 | stat(s, ORDER_FALLBACK); |
| 2032 | } |
| 2033 | |
| 2034 | slab->objects = oo_objects(oo); |
| 2035 | slab->inuse = 0; |
| 2036 | slab->frozen = 0; |
| 2037 | |
| 2038 | account_slab(slab, oo_order(oo), s, flags); |
| 2039 | |
| 2040 | slab->slab_cache = s; |
| 2041 | |
| 2042 | kasan_poison_slab(slab); |
| 2043 | |
| 2044 | start = slab_address(slab); |
| 2045 | |
| 2046 | setup_slab_debug(s, slab, start); |
| 2047 | |
| 2048 | shuffle = shuffle_freelist(s, slab); |
| 2049 | |
| 2050 | if (!shuffle) { |
| 2051 | start = fixup_red_left(s, start); |
| 2052 | start = setup_object(s, start); |
| 2053 | slab->freelist = start; |
| 2054 | for (idx = 0, p = start; idx < slab->objects - 1; idx++) { |
| 2055 | next = p + s->size; |
| 2056 | next = setup_object(s, next); |
| 2057 | set_freepointer(s, p, next); |
| 2058 | p = next; |
| 2059 | } |
| 2060 | set_freepointer(s, p, NULL); |
| 2061 | } |
| 2062 | |
| 2063 | return slab; |
| 2064 | } |
| 2065 | |
| 2066 | static struct slab *new_slab(struct kmem_cache *s, gfp_t flags, int node) |
| 2067 | { |
| 2068 | if (unlikely(flags & GFP_SLAB_BUG_MASK)) |
| 2069 | flags = kmalloc_fix_flags(flags); |
| 2070 | |
| 2071 | WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO)); |
| 2072 | |
| 2073 | return allocate_slab(s, |
| 2074 | flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node); |
| 2075 | } |
| 2076 | |
| 2077 | static void __free_slab(struct kmem_cache *s, struct slab *slab) |
| 2078 | { |
| 2079 | struct folio *folio = slab_folio(slab); |
| 2080 | int order = folio_order(folio); |
| 2081 | int pages = 1 << order; |
| 2082 | |
| 2083 | __slab_clear_pfmemalloc(slab); |
| 2084 | folio->mapping = NULL; |
| 2085 | /* Make the mapping reset visible before clearing the flag */ |
| 2086 | smp_wmb(); |
| 2087 | __folio_clear_slab(folio); |
| 2088 | mm_account_reclaimed_pages(pages); |
| 2089 | unaccount_slab(slab, order, s); |
| 2090 | __free_pages(&folio->page, order); |
| 2091 | } |
| 2092 | |
| 2093 | static void rcu_free_slab(struct rcu_head *h) |
| 2094 | { |
| 2095 | struct slab *slab = container_of(h, struct slab, rcu_head); |
| 2096 | |
| 2097 | __free_slab(slab->slab_cache, slab); |
| 2098 | } |
| 2099 | |
| 2100 | static void free_slab(struct kmem_cache *s, struct slab *slab) |
| 2101 | { |
| 2102 | if (kmem_cache_debug_flags(s, SLAB_CONSISTENCY_CHECKS)) { |
| 2103 | void *p; |
| 2104 | |
| 2105 | slab_pad_check(s, slab); |
| 2106 | for_each_object(p, s, slab_address(slab), slab->objects) |
| 2107 | check_object(s, slab, p, SLUB_RED_INACTIVE); |
| 2108 | } |
| 2109 | |
| 2110 | if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU)) |
| 2111 | call_rcu(&slab->rcu_head, rcu_free_slab); |
| 2112 | else |
| 2113 | __free_slab(s, slab); |
| 2114 | } |
| 2115 | |
| 2116 | static void discard_slab(struct kmem_cache *s, struct slab *slab) |
| 2117 | { |
| 2118 | dec_slabs_node(s, slab_nid(slab), slab->objects); |
| 2119 | free_slab(s, slab); |
| 2120 | } |
| 2121 | |
| 2122 | /* |
| 2123 | * Management of partially allocated slabs. |
| 2124 | */ |
| 2125 | static inline void |
| 2126 | __add_partial(struct kmem_cache_node *n, struct slab *slab, int tail) |
| 2127 | { |
| 2128 | n->nr_partial++; |
| 2129 | if (tail == DEACTIVATE_TO_TAIL) |
| 2130 | list_add_tail(&slab->slab_list, &n->partial); |
| 2131 | else |
| 2132 | list_add(&slab->slab_list, &n->partial); |
| 2133 | } |
| 2134 | |
| 2135 | static inline void add_partial(struct kmem_cache_node *n, |
| 2136 | struct slab *slab, int tail) |
| 2137 | { |
| 2138 | lockdep_assert_held(&n->list_lock); |
| 2139 | __add_partial(n, slab, tail); |
| 2140 | } |
| 2141 | |
| 2142 | static inline void remove_partial(struct kmem_cache_node *n, |
| 2143 | struct slab *slab) |
| 2144 | { |
| 2145 | lockdep_assert_held(&n->list_lock); |
| 2146 | list_del(&slab->slab_list); |
| 2147 | n->nr_partial--; |
| 2148 | } |
| 2149 | |
| 2150 | /* |
| 2151 | * Called only for kmem_cache_debug() caches instead of acquire_slab(), with a |
| 2152 | * slab from the n->partial list. Remove only a single object from the slab, do |
| 2153 | * the alloc_debug_processing() checks and leave the slab on the list, or move |
| 2154 | * it to full list if it was the last free object. |
| 2155 | */ |
| 2156 | static void *alloc_single_from_partial(struct kmem_cache *s, |
| 2157 | struct kmem_cache_node *n, struct slab *slab, int orig_size) |
| 2158 | { |
| 2159 | void *object; |
| 2160 | |
| 2161 | lockdep_assert_held(&n->list_lock); |
| 2162 | |
| 2163 | object = slab->freelist; |
| 2164 | slab->freelist = get_freepointer(s, object); |
| 2165 | slab->inuse++; |
| 2166 | |
| 2167 | if (!alloc_debug_processing(s, slab, object, orig_size)) { |
| 2168 | remove_partial(n, slab); |
| 2169 | return NULL; |
| 2170 | } |
| 2171 | |
| 2172 | if (slab->inuse == slab->objects) { |
| 2173 | remove_partial(n, slab); |
| 2174 | add_full(s, n, slab); |
| 2175 | } |
| 2176 | |
| 2177 | return object; |
| 2178 | } |
| 2179 | |
| 2180 | /* |
| 2181 | * Called only for kmem_cache_debug() caches to allocate from a freshly |
| 2182 | * allocated slab. Allocate a single object instead of whole freelist |
| 2183 | * and put the slab to the partial (or full) list. |
| 2184 | */ |
| 2185 | static void *alloc_single_from_new_slab(struct kmem_cache *s, |
| 2186 | struct slab *slab, int orig_size) |
| 2187 | { |
| 2188 | int nid = slab_nid(slab); |
| 2189 | struct kmem_cache_node *n = get_node(s, nid); |
| 2190 | unsigned long flags; |
| 2191 | void *object; |
| 2192 | |
| 2193 | |
| 2194 | object = slab->freelist; |
| 2195 | slab->freelist = get_freepointer(s, object); |
| 2196 | slab->inuse = 1; |
| 2197 | |
| 2198 | if (!alloc_debug_processing(s, slab, object, orig_size)) |
| 2199 | /* |
| 2200 | * It's not really expected that this would fail on a |
| 2201 | * freshly allocated slab, but a concurrent memory |
| 2202 | * corruption in theory could cause that. |
| 2203 | */ |
| 2204 | return NULL; |
| 2205 | |
| 2206 | spin_lock_irqsave(&n->list_lock, flags); |
| 2207 | |
| 2208 | if (slab->inuse == slab->objects) |
| 2209 | add_full(s, n, slab); |
| 2210 | else |
| 2211 | add_partial(n, slab, DEACTIVATE_TO_HEAD); |
| 2212 | |
| 2213 | inc_slabs_node(s, nid, slab->objects); |
| 2214 | spin_unlock_irqrestore(&n->list_lock, flags); |
| 2215 | |
| 2216 | return object; |
| 2217 | } |
| 2218 | |
| 2219 | /* |
| 2220 | * Remove slab from the partial list, freeze it and |
| 2221 | * return the pointer to the freelist. |
| 2222 | * |
| 2223 | * Returns a list of objects or NULL if it fails. |
| 2224 | */ |
| 2225 | static inline void *acquire_slab(struct kmem_cache *s, |
| 2226 | struct kmem_cache_node *n, struct slab *slab, |
| 2227 | int mode) |
| 2228 | { |
| 2229 | void *freelist; |
| 2230 | unsigned long counters; |
| 2231 | struct slab new; |
| 2232 | |
| 2233 | lockdep_assert_held(&n->list_lock); |
| 2234 | |
| 2235 | /* |
| 2236 | * Zap the freelist and set the frozen bit. |
| 2237 | * The old freelist is the list of objects for the |
| 2238 | * per cpu allocation list. |
| 2239 | */ |
| 2240 | freelist = slab->freelist; |
| 2241 | counters = slab->counters; |
| 2242 | new.counters = counters; |
| 2243 | if (mode) { |
| 2244 | new.inuse = slab->objects; |
| 2245 | new.freelist = NULL; |
| 2246 | } else { |
| 2247 | new.freelist = freelist; |
| 2248 | } |
| 2249 | |
| 2250 | VM_BUG_ON(new.frozen); |
| 2251 | new.frozen = 1; |
| 2252 | |
| 2253 | if (!__slab_update_freelist(s, slab, |
| 2254 | freelist, counters, |
| 2255 | new.freelist, new.counters, |
| 2256 | "acquire_slab")) |
| 2257 | return NULL; |
| 2258 | |
| 2259 | remove_partial(n, slab); |
| 2260 | WARN_ON(!freelist); |
| 2261 | return freelist; |
| 2262 | } |
| 2263 | |
| 2264 | #ifdef CONFIG_SLUB_CPU_PARTIAL |
| 2265 | static void put_cpu_partial(struct kmem_cache *s, struct slab *slab, int drain); |
| 2266 | #else |
| 2267 | static inline void put_cpu_partial(struct kmem_cache *s, struct slab *slab, |
| 2268 | int drain) { } |
| 2269 | #endif |
| 2270 | static inline bool pfmemalloc_match(struct slab *slab, gfp_t gfpflags); |
| 2271 | |
| 2272 | /* |
| 2273 | * Try to allocate a partial slab from a specific node. |
| 2274 | */ |
| 2275 | static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n, |
| 2276 | struct partial_context *pc) |
| 2277 | { |
| 2278 | struct slab *slab, *slab2; |
| 2279 | void *object = NULL; |
| 2280 | unsigned long flags; |
| 2281 | unsigned int partial_slabs = 0; |
| 2282 | |
| 2283 | /* |
| 2284 | * Racy check. If we mistakenly see no partial slabs then we |
| 2285 | * just allocate an empty slab. If we mistakenly try to get a |
| 2286 | * partial slab and there is none available then get_partial() |
| 2287 | * will return NULL. |
| 2288 | */ |
| 2289 | if (!n || !n->nr_partial) |
| 2290 | return NULL; |
| 2291 | |
| 2292 | spin_lock_irqsave(&n->list_lock, flags); |
| 2293 | list_for_each_entry_safe(slab, slab2, &n->partial, slab_list) { |
| 2294 | void *t; |
| 2295 | |
| 2296 | if (!pfmemalloc_match(slab, pc->flags)) |
| 2297 | continue; |
| 2298 | |
| 2299 | if (IS_ENABLED(CONFIG_SLUB_TINY) || kmem_cache_debug(s)) { |
| 2300 | object = alloc_single_from_partial(s, n, slab, |
| 2301 | pc->orig_size); |
| 2302 | if (object) |
| 2303 | break; |
| 2304 | continue; |
| 2305 | } |
| 2306 | |
| 2307 | t = acquire_slab(s, n, slab, object == NULL); |
| 2308 | if (!t) |
| 2309 | break; |
| 2310 | |
| 2311 | if (!object) { |
| 2312 | *pc->slab = slab; |
| 2313 | stat(s, ALLOC_FROM_PARTIAL); |
| 2314 | object = t; |
| 2315 | } else { |
| 2316 | put_cpu_partial(s, slab, 0); |
| 2317 | stat(s, CPU_PARTIAL_NODE); |
| 2318 | partial_slabs++; |
| 2319 | } |
| 2320 | #ifdef CONFIG_SLUB_CPU_PARTIAL |
| 2321 | if (!kmem_cache_has_cpu_partial(s) |
| 2322 | || partial_slabs > s->cpu_partial_slabs / 2) |
| 2323 | break; |
| 2324 | #else |
| 2325 | break; |
| 2326 | #endif |
| 2327 | |
| 2328 | } |
| 2329 | spin_unlock_irqrestore(&n->list_lock, flags); |
| 2330 | return object; |
| 2331 | } |
| 2332 | |
| 2333 | /* |
| 2334 | * Get a slab from somewhere. Search in increasing NUMA distances. |
| 2335 | */ |
| 2336 | static void *get_any_partial(struct kmem_cache *s, struct partial_context *pc) |
| 2337 | { |
| 2338 | #ifdef CONFIG_NUMA |
| 2339 | struct zonelist *zonelist; |
| 2340 | struct zoneref *z; |
| 2341 | struct zone *zone; |
| 2342 | enum zone_type highest_zoneidx = gfp_zone(pc->flags); |
| 2343 | void *object; |
| 2344 | unsigned int cpuset_mems_cookie; |
| 2345 | |
| 2346 | /* |
| 2347 | * The defrag ratio allows a configuration of the tradeoffs between |
| 2348 | * inter node defragmentation and node local allocations. A lower |
| 2349 | * defrag_ratio increases the tendency to do local allocations |
| 2350 | * instead of attempting to obtain partial slabs from other nodes. |
| 2351 | * |
| 2352 | * If the defrag_ratio is set to 0 then kmalloc() always |
| 2353 | * returns node local objects. If the ratio is higher then kmalloc() |
| 2354 | * may return off node objects because partial slabs are obtained |
| 2355 | * from other nodes and filled up. |
| 2356 | * |
| 2357 | * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100 |
| 2358 | * (which makes defrag_ratio = 1000) then every (well almost) |
| 2359 | * allocation will first attempt to defrag slab caches on other nodes. |
| 2360 | * This means scanning over all nodes to look for partial slabs which |
| 2361 | * may be expensive if we do it every time we are trying to find a slab |
| 2362 | * with available objects. |
| 2363 | */ |
| 2364 | if (!s->remote_node_defrag_ratio || |
| 2365 | get_cycles() % 1024 > s->remote_node_defrag_ratio) |
| 2366 | return NULL; |
| 2367 | |
| 2368 | do { |
| 2369 | cpuset_mems_cookie = read_mems_allowed_begin(); |
| 2370 | zonelist = node_zonelist(mempolicy_slab_node(), pc->flags); |
| 2371 | for_each_zone_zonelist(zone, z, zonelist, highest_zoneidx) { |
| 2372 | struct kmem_cache_node *n; |
| 2373 | |
| 2374 | n = get_node(s, zone_to_nid(zone)); |
| 2375 | |
| 2376 | if (n && cpuset_zone_allowed(zone, pc->flags) && |
| 2377 | n->nr_partial > s->min_partial) { |
| 2378 | object = get_partial_node(s, n, pc); |
| 2379 | if (object) { |
| 2380 | /* |
| 2381 | * Don't check read_mems_allowed_retry() |
| 2382 | * here - if mems_allowed was updated in |
| 2383 | * parallel, that was a harmless race |
| 2384 | * between allocation and the cpuset |
| 2385 | * update |
| 2386 | */ |
| 2387 | return object; |
| 2388 | } |
| 2389 | } |
| 2390 | } |
| 2391 | } while (read_mems_allowed_retry(cpuset_mems_cookie)); |
| 2392 | #endif /* CONFIG_NUMA */ |
| 2393 | return NULL; |
| 2394 | } |
| 2395 | |
| 2396 | /* |
| 2397 | * Get a partial slab, lock it and return it. |
| 2398 | */ |
| 2399 | static void *get_partial(struct kmem_cache *s, int node, struct partial_context *pc) |
| 2400 | { |
| 2401 | void *object; |
| 2402 | int searchnode = node; |
| 2403 | |
| 2404 | if (node == NUMA_NO_NODE) |
| 2405 | searchnode = numa_mem_id(); |
| 2406 | |
| 2407 | object = get_partial_node(s, get_node(s, searchnode), pc); |
| 2408 | if (object || node != NUMA_NO_NODE) |
| 2409 | return object; |
| 2410 | |
| 2411 | return get_any_partial(s, pc); |
| 2412 | } |
| 2413 | |
| 2414 | #ifndef CONFIG_SLUB_TINY |
| 2415 | |
| 2416 | #ifdef CONFIG_PREEMPTION |
| 2417 | /* |
| 2418 | * Calculate the next globally unique transaction for disambiguation |
| 2419 | * during cmpxchg. The transactions start with the cpu number and are then |
| 2420 | * incremented by CONFIG_NR_CPUS. |
| 2421 | */ |
| 2422 | #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS) |
| 2423 | #else |
| 2424 | /* |
| 2425 | * No preemption supported therefore also no need to check for |
| 2426 | * different cpus. |
| 2427 | */ |
| 2428 | #define TID_STEP 1 |
| 2429 | #endif /* CONFIG_PREEMPTION */ |
| 2430 | |
| 2431 | static inline unsigned long next_tid(unsigned long tid) |
| 2432 | { |
| 2433 | return tid + TID_STEP; |
| 2434 | } |
| 2435 | |
| 2436 | #ifdef SLUB_DEBUG_CMPXCHG |
| 2437 | static inline unsigned int tid_to_cpu(unsigned long tid) |
| 2438 | { |
| 2439 | return tid % TID_STEP; |
| 2440 | } |
| 2441 | |
| 2442 | static inline unsigned long tid_to_event(unsigned long tid) |
| 2443 | { |
| 2444 | return tid / TID_STEP; |
| 2445 | } |
| 2446 | #endif |
| 2447 | |
| 2448 | static inline unsigned int init_tid(int cpu) |
| 2449 | { |
| 2450 | return cpu; |
| 2451 | } |
| 2452 | |
| 2453 | static inline void note_cmpxchg_failure(const char *n, |
| 2454 | const struct kmem_cache *s, unsigned long tid) |
| 2455 | { |
| 2456 | #ifdef SLUB_DEBUG_CMPXCHG |
| 2457 | unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid); |
| 2458 | |
| 2459 | pr_info("%s %s: cmpxchg redo ", n, s->name); |
| 2460 | |
| 2461 | #ifdef CONFIG_PREEMPTION |
| 2462 | if (tid_to_cpu(tid) != tid_to_cpu(actual_tid)) |
| 2463 | pr_warn("due to cpu change %d -> %d\n", |
| 2464 | tid_to_cpu(tid), tid_to_cpu(actual_tid)); |
| 2465 | else |
| 2466 | #endif |
| 2467 | if (tid_to_event(tid) != tid_to_event(actual_tid)) |
| 2468 | pr_warn("due to cpu running other code. Event %ld->%ld\n", |
| 2469 | tid_to_event(tid), tid_to_event(actual_tid)); |
| 2470 | else |
| 2471 | pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n", |
| 2472 | actual_tid, tid, next_tid(tid)); |
| 2473 | #endif |
| 2474 | stat(s, CMPXCHG_DOUBLE_CPU_FAIL); |
| 2475 | } |
| 2476 | |
| 2477 | static void init_kmem_cache_cpus(struct kmem_cache *s) |
| 2478 | { |
| 2479 | int cpu; |
| 2480 | struct kmem_cache_cpu *c; |
| 2481 | |
| 2482 | for_each_possible_cpu(cpu) { |
| 2483 | c = per_cpu_ptr(s->cpu_slab, cpu); |
| 2484 | local_lock_init(&c->lock); |
| 2485 | c->tid = init_tid(cpu); |
| 2486 | } |
| 2487 | } |
| 2488 | |
| 2489 | /* |
| 2490 | * Finishes removing the cpu slab. Merges cpu's freelist with slab's freelist, |
| 2491 | * unfreezes the slabs and puts it on the proper list. |
| 2492 | * Assumes the slab has been already safely taken away from kmem_cache_cpu |
| 2493 | * by the caller. |
| 2494 | */ |
| 2495 | static void deactivate_slab(struct kmem_cache *s, struct slab *slab, |
| 2496 | void *freelist) |
| 2497 | { |
| 2498 | enum slab_modes { M_NONE, M_PARTIAL, M_FREE, M_FULL_NOLIST }; |
| 2499 | struct kmem_cache_node *n = get_node(s, slab_nid(slab)); |
| 2500 | int free_delta = 0; |
| 2501 | enum slab_modes mode = M_NONE; |
| 2502 | void *nextfree, *freelist_iter, *freelist_tail; |
| 2503 | int tail = DEACTIVATE_TO_HEAD; |
| 2504 | unsigned long flags = 0; |
| 2505 | struct slab new; |
| 2506 | struct slab old; |
| 2507 | |
| 2508 | if (slab->freelist) { |
| 2509 | stat(s, DEACTIVATE_REMOTE_FREES); |
| 2510 | tail = DEACTIVATE_TO_TAIL; |
| 2511 | } |
| 2512 | |
| 2513 | /* |
| 2514 | * Stage one: Count the objects on cpu's freelist as free_delta and |
| 2515 | * remember the last object in freelist_tail for later splicing. |
| 2516 | */ |
| 2517 | freelist_tail = NULL; |
| 2518 | freelist_iter = freelist; |
| 2519 | while (freelist_iter) { |
| 2520 | nextfree = get_freepointer(s, freelist_iter); |
| 2521 | |
| 2522 | /* |
| 2523 | * If 'nextfree' is invalid, it is possible that the object at |
| 2524 | * 'freelist_iter' is already corrupted. So isolate all objects |
| 2525 | * starting at 'freelist_iter' by skipping them. |
| 2526 | */ |
| 2527 | if (freelist_corrupted(s, slab, &freelist_iter, nextfree)) |
| 2528 | break; |
| 2529 | |
| 2530 | freelist_tail = freelist_iter; |
| 2531 | free_delta++; |
| 2532 | |
| 2533 | freelist_iter = nextfree; |
| 2534 | } |
| 2535 | |
| 2536 | /* |
| 2537 | * Stage two: Unfreeze the slab while splicing the per-cpu |
| 2538 | * freelist to the head of slab's freelist. |
| 2539 | * |
| 2540 | * Ensure that the slab is unfrozen while the list presence |
| 2541 | * reflects the actual number of objects during unfreeze. |
| 2542 | * |
| 2543 | * We first perform cmpxchg holding lock and insert to list |
| 2544 | * when it succeed. If there is mismatch then the slab is not |
| 2545 | * unfrozen and number of objects in the slab may have changed. |
| 2546 | * Then release lock and retry cmpxchg again. |
| 2547 | */ |
| 2548 | redo: |
| 2549 | |
| 2550 | old.freelist = READ_ONCE(slab->freelist); |
| 2551 | old.counters = READ_ONCE(slab->counters); |
| 2552 | VM_BUG_ON(!old.frozen); |
| 2553 | |
| 2554 | /* Determine target state of the slab */ |
| 2555 | new.counters = old.counters; |
| 2556 | if (freelist_tail) { |
| 2557 | new.inuse -= free_delta; |
| 2558 | set_freepointer(s, freelist_tail, old.freelist); |
| 2559 | new.freelist = freelist; |
| 2560 | } else |
| 2561 | new.freelist = old.freelist; |
| 2562 | |
| 2563 | new.frozen = 0; |
| 2564 | |
| 2565 | if (!new.inuse && n->nr_partial >= s->min_partial) { |
| 2566 | mode = M_FREE; |
| 2567 | } else if (new.freelist) { |
| 2568 | mode = M_PARTIAL; |
| 2569 | /* |
| 2570 | * Taking the spinlock removes the possibility that |
| 2571 | * acquire_slab() will see a slab that is frozen |
| 2572 | */ |
| 2573 | spin_lock_irqsave(&n->list_lock, flags); |
| 2574 | } else { |
| 2575 | mode = M_FULL_NOLIST; |
| 2576 | } |
| 2577 | |
| 2578 | |
| 2579 | if (!slab_update_freelist(s, slab, |
| 2580 | old.freelist, old.counters, |
| 2581 | new.freelist, new.counters, |
| 2582 | "unfreezing slab")) { |
| 2583 | if (mode == M_PARTIAL) |
| 2584 | spin_unlock_irqrestore(&n->list_lock, flags); |
| 2585 | goto redo; |
| 2586 | } |
| 2587 | |
| 2588 | |
| 2589 | if (mode == M_PARTIAL) { |
| 2590 | add_partial(n, slab, tail); |
| 2591 | spin_unlock_irqrestore(&n->list_lock, flags); |
| 2592 | stat(s, tail); |
| 2593 | } else if (mode == M_FREE) { |
| 2594 | stat(s, DEACTIVATE_EMPTY); |
| 2595 | discard_slab(s, slab); |
| 2596 | stat(s, FREE_SLAB); |
| 2597 | } else if (mode == M_FULL_NOLIST) { |
| 2598 | stat(s, DEACTIVATE_FULL); |
| 2599 | } |
| 2600 | } |
| 2601 | |
| 2602 | #ifdef CONFIG_SLUB_CPU_PARTIAL |
| 2603 | static void __unfreeze_partials(struct kmem_cache *s, struct slab *partial_slab) |
| 2604 | { |
| 2605 | struct kmem_cache_node *n = NULL, *n2 = NULL; |
| 2606 | struct slab *slab, *slab_to_discard = NULL; |
| 2607 | unsigned long flags = 0; |
| 2608 | |
| 2609 | while (partial_slab) { |
| 2610 | struct slab new; |
| 2611 | struct slab old; |
| 2612 | |
| 2613 | slab = partial_slab; |
| 2614 | partial_slab = slab->next; |
| 2615 | |
| 2616 | n2 = get_node(s, slab_nid(slab)); |
| 2617 | if (n != n2) { |
| 2618 | if (n) |
| 2619 | spin_unlock_irqrestore(&n->list_lock, flags); |
| 2620 | |
| 2621 | n = n2; |
| 2622 | spin_lock_irqsave(&n->list_lock, flags); |
| 2623 | } |
| 2624 | |
| 2625 | do { |
| 2626 | |
| 2627 | old.freelist = slab->freelist; |
| 2628 | old.counters = slab->counters; |
| 2629 | VM_BUG_ON(!old.frozen); |
| 2630 | |
| 2631 | new.counters = old.counters; |
| 2632 | new.freelist = old.freelist; |
| 2633 | |
| 2634 | new.frozen = 0; |
| 2635 | |
| 2636 | } while (!__slab_update_freelist(s, slab, |
| 2637 | old.freelist, old.counters, |
| 2638 | new.freelist, new.counters, |
| 2639 | "unfreezing slab")); |
| 2640 | |
| 2641 | if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) { |
| 2642 | slab->next = slab_to_discard; |
| 2643 | slab_to_discard = slab; |
| 2644 | } else { |
| 2645 | add_partial(n, slab, DEACTIVATE_TO_TAIL); |
| 2646 | stat(s, FREE_ADD_PARTIAL); |
| 2647 | } |
| 2648 | } |
| 2649 | |
| 2650 | if (n) |
| 2651 | spin_unlock_irqrestore(&n->list_lock, flags); |
| 2652 | |
| 2653 | while (slab_to_discard) { |
| 2654 | slab = slab_to_discard; |
| 2655 | slab_to_discard = slab_to_discard->next; |
| 2656 | |
| 2657 | stat(s, DEACTIVATE_EMPTY); |
| 2658 | discard_slab(s, slab); |
| 2659 | stat(s, FREE_SLAB); |
| 2660 | } |
| 2661 | } |
| 2662 | |
| 2663 | /* |
| 2664 | * Unfreeze all the cpu partial slabs. |
| 2665 | */ |
| 2666 | static void unfreeze_partials(struct kmem_cache *s) |
| 2667 | { |
| 2668 | struct slab *partial_slab; |
| 2669 | unsigned long flags; |
| 2670 | |
| 2671 | local_lock_irqsave(&s->cpu_slab->lock, flags); |
| 2672 | partial_slab = this_cpu_read(s->cpu_slab->partial); |
| 2673 | this_cpu_write(s->cpu_slab->partial, NULL); |
| 2674 | local_unlock_irqrestore(&s->cpu_slab->lock, flags); |
| 2675 | |
| 2676 | if (partial_slab) |
| 2677 | __unfreeze_partials(s, partial_slab); |
| 2678 | } |
| 2679 | |
| 2680 | static void unfreeze_partials_cpu(struct kmem_cache *s, |
| 2681 | struct kmem_cache_cpu *c) |
| 2682 | { |
| 2683 | struct slab *partial_slab; |
| 2684 | |
| 2685 | partial_slab = slub_percpu_partial(c); |
| 2686 | c->partial = NULL; |
| 2687 | |
| 2688 | if (partial_slab) |
| 2689 | __unfreeze_partials(s, partial_slab); |
| 2690 | } |
| 2691 | |
| 2692 | /* |
| 2693 | * Put a slab that was just frozen (in __slab_free|get_partial_node) into a |
| 2694 | * partial slab slot if available. |
| 2695 | * |
| 2696 | * If we did not find a slot then simply move all the partials to the |
| 2697 | * per node partial list. |
| 2698 | */ |
| 2699 | static void put_cpu_partial(struct kmem_cache *s, struct slab *slab, int drain) |
| 2700 | { |
| 2701 | struct slab *oldslab; |
| 2702 | struct slab *slab_to_unfreeze = NULL; |
| 2703 | unsigned long flags; |
| 2704 | int slabs = 0; |
| 2705 | |
| 2706 | local_lock_irqsave(&s->cpu_slab->lock, flags); |
| 2707 | |
| 2708 | oldslab = this_cpu_read(s->cpu_slab->partial); |
| 2709 | |
| 2710 | if (oldslab) { |
| 2711 | if (drain && oldslab->slabs >= s->cpu_partial_slabs) { |
| 2712 | /* |
| 2713 | * Partial array is full. Move the existing set to the |
| 2714 | * per node partial list. Postpone the actual unfreezing |
| 2715 | * outside of the critical section. |
| 2716 | */ |
| 2717 | slab_to_unfreeze = oldslab; |
| 2718 | oldslab = NULL; |
| 2719 | } else { |
| 2720 | slabs = oldslab->slabs; |
| 2721 | } |
| 2722 | } |
| 2723 | |
| 2724 | slabs++; |
| 2725 | |
| 2726 | slab->slabs = slabs; |
| 2727 | slab->next = oldslab; |
| 2728 | |
| 2729 | this_cpu_write(s->cpu_slab->partial, slab); |
| 2730 | |
| 2731 | local_unlock_irqrestore(&s->cpu_slab->lock, flags); |
| 2732 | |
| 2733 | if (slab_to_unfreeze) { |
| 2734 | __unfreeze_partials(s, slab_to_unfreeze); |
| 2735 | stat(s, CPU_PARTIAL_DRAIN); |
| 2736 | } |
| 2737 | } |
| 2738 | |
| 2739 | #else /* CONFIG_SLUB_CPU_PARTIAL */ |
| 2740 | |
| 2741 | static inline void unfreeze_partials(struct kmem_cache *s) { } |
| 2742 | static inline void unfreeze_partials_cpu(struct kmem_cache *s, |
| 2743 | struct kmem_cache_cpu *c) { } |
| 2744 | |
| 2745 | #endif /* CONFIG_SLUB_CPU_PARTIAL */ |
| 2746 | |
| 2747 | static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c) |
| 2748 | { |
| 2749 | unsigned long flags; |
| 2750 | struct slab *slab; |
| 2751 | void *freelist; |
| 2752 | |
| 2753 | local_lock_irqsave(&s->cpu_slab->lock, flags); |
| 2754 | |
| 2755 | slab = c->slab; |
| 2756 | freelist = c->freelist; |
| 2757 | |
| 2758 | c->slab = NULL; |
| 2759 | c->freelist = NULL; |
| 2760 | c->tid = next_tid(c->tid); |
| 2761 | |
| 2762 | local_unlock_irqrestore(&s->cpu_slab->lock, flags); |
| 2763 | |
| 2764 | if (slab) { |
| 2765 | deactivate_slab(s, slab, freelist); |
| 2766 | stat(s, CPUSLAB_FLUSH); |
| 2767 | } |
| 2768 | } |
| 2769 | |
| 2770 | static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu) |
| 2771 | { |
| 2772 | struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu); |
| 2773 | void *freelist = c->freelist; |
| 2774 | struct slab *slab = c->slab; |
| 2775 | |
| 2776 | c->slab = NULL; |
| 2777 | c->freelist = NULL; |
| 2778 | c->tid = next_tid(c->tid); |
| 2779 | |
| 2780 | if (slab) { |
| 2781 | deactivate_slab(s, slab, freelist); |
| 2782 | stat(s, CPUSLAB_FLUSH); |
| 2783 | } |
| 2784 | |
| 2785 | unfreeze_partials_cpu(s, c); |
| 2786 | } |
| 2787 | |
| 2788 | struct slub_flush_work { |
| 2789 | struct work_struct work; |
| 2790 | struct kmem_cache *s; |
| 2791 | bool skip; |
| 2792 | }; |
| 2793 | |
| 2794 | /* |
| 2795 | * Flush cpu slab. |
| 2796 | * |
| 2797 | * Called from CPU work handler with migration disabled. |
| 2798 | */ |
| 2799 | static void flush_cpu_slab(struct work_struct *w) |
| 2800 | { |
| 2801 | struct kmem_cache *s; |
| 2802 | struct kmem_cache_cpu *c; |
| 2803 | struct slub_flush_work *sfw; |
| 2804 | |
| 2805 | sfw = container_of(w, struct slub_flush_work, work); |
| 2806 | |
| 2807 | s = sfw->s; |
| 2808 | c = this_cpu_ptr(s->cpu_slab); |
| 2809 | |
| 2810 | if (c->slab) |
| 2811 | flush_slab(s, c); |
| 2812 | |
| 2813 | unfreeze_partials(s); |
| 2814 | } |
| 2815 | |
| 2816 | static bool has_cpu_slab(int cpu, struct kmem_cache *s) |
| 2817 | { |
| 2818 | struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu); |
| 2819 | |
| 2820 | return c->slab || slub_percpu_partial(c); |
| 2821 | } |
| 2822 | |
| 2823 | static DEFINE_MUTEX(flush_lock); |
| 2824 | static DEFINE_PER_CPU(struct slub_flush_work, slub_flush); |
| 2825 | |
| 2826 | static void flush_all_cpus_locked(struct kmem_cache *s) |
| 2827 | { |
| 2828 | struct slub_flush_work *sfw; |
| 2829 | unsigned int cpu; |
| 2830 | |
| 2831 | lockdep_assert_cpus_held(); |
| 2832 | mutex_lock(&flush_lock); |
| 2833 | |
| 2834 | for_each_online_cpu(cpu) { |
| 2835 | sfw = &per_cpu(slub_flush, cpu); |
| 2836 | if (!has_cpu_slab(cpu, s)) { |
| 2837 | sfw->skip = true; |
| 2838 | continue; |
| 2839 | } |
| 2840 | INIT_WORK(&sfw->work, flush_cpu_slab); |
| 2841 | sfw->skip = false; |
| 2842 | sfw->s = s; |
| 2843 | queue_work_on(cpu, flushwq, &sfw->work); |
| 2844 | } |
| 2845 | |
| 2846 | for_each_online_cpu(cpu) { |
| 2847 | sfw = &per_cpu(slub_flush, cpu); |
| 2848 | if (sfw->skip) |
| 2849 | continue; |
| 2850 | flush_work(&sfw->work); |
| 2851 | } |
| 2852 | |
| 2853 | mutex_unlock(&flush_lock); |
| 2854 | } |
| 2855 | |
| 2856 | static void flush_all(struct kmem_cache *s) |
| 2857 | { |
| 2858 | cpus_read_lock(); |
| 2859 | flush_all_cpus_locked(s); |
| 2860 | cpus_read_unlock(); |
| 2861 | } |
| 2862 | |
| 2863 | /* |
| 2864 | * Use the cpu notifier to insure that the cpu slabs are flushed when |
| 2865 | * necessary. |
| 2866 | */ |
| 2867 | static int slub_cpu_dead(unsigned int cpu) |
| 2868 | { |
| 2869 | struct kmem_cache *s; |
| 2870 | |
| 2871 | mutex_lock(&slab_mutex); |
| 2872 | list_for_each_entry(s, &slab_caches, list) |
| 2873 | __flush_cpu_slab(s, cpu); |
| 2874 | mutex_unlock(&slab_mutex); |
| 2875 | return 0; |
| 2876 | } |
| 2877 | |
| 2878 | #else /* CONFIG_SLUB_TINY */ |
| 2879 | static inline void flush_all_cpus_locked(struct kmem_cache *s) { } |
| 2880 | static inline void flush_all(struct kmem_cache *s) { } |
| 2881 | static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu) { } |
| 2882 | static inline int slub_cpu_dead(unsigned int cpu) { return 0; } |
| 2883 | #endif /* CONFIG_SLUB_TINY */ |
| 2884 | |
| 2885 | /* |
| 2886 | * Check if the objects in a per cpu structure fit numa |
| 2887 | * locality expectations. |
| 2888 | */ |
| 2889 | static inline int node_match(struct slab *slab, int node) |
| 2890 | { |
| 2891 | #ifdef CONFIG_NUMA |
| 2892 | if (node != NUMA_NO_NODE && slab_nid(slab) != node) |
| 2893 | return 0; |
| 2894 | #endif |
| 2895 | return 1; |
| 2896 | } |
| 2897 | |
| 2898 | #ifdef CONFIG_SLUB_DEBUG |
| 2899 | static int count_free(struct slab *slab) |
| 2900 | { |
| 2901 | return slab->objects - slab->inuse; |
| 2902 | } |
| 2903 | |
| 2904 | static inline unsigned long node_nr_objs(struct kmem_cache_node *n) |
| 2905 | { |
| 2906 | return atomic_long_read(&n->total_objects); |
| 2907 | } |
| 2908 | |
| 2909 | /* Supports checking bulk free of a constructed freelist */ |
| 2910 | static inline bool free_debug_processing(struct kmem_cache *s, |
| 2911 | struct slab *slab, void *head, void *tail, int *bulk_cnt, |
| 2912 | unsigned long addr, depot_stack_handle_t handle) |
| 2913 | { |
| 2914 | bool checks_ok = false; |
| 2915 | void *object = head; |
| 2916 | int cnt = 0; |
| 2917 | |
| 2918 | if (s->flags & SLAB_CONSISTENCY_CHECKS) { |
| 2919 | if (!check_slab(s, slab)) |
| 2920 | goto out; |
| 2921 | } |
| 2922 | |
| 2923 | if (slab->inuse < *bulk_cnt) { |
| 2924 | slab_err(s, slab, "Slab has %d allocated objects but %d are to be freed\n", |
| 2925 | slab->inuse, *bulk_cnt); |
| 2926 | goto out; |
| 2927 | } |
| 2928 | |
| 2929 | next_object: |
| 2930 | |
| 2931 | if (++cnt > *bulk_cnt) |
| 2932 | goto out_cnt; |
| 2933 | |
| 2934 | if (s->flags & SLAB_CONSISTENCY_CHECKS) { |
| 2935 | if (!free_consistency_checks(s, slab, object, addr)) |
| 2936 | goto out; |
| 2937 | } |
| 2938 | |
| 2939 | if (s->flags & SLAB_STORE_USER) |
| 2940 | set_track_update(s, object, TRACK_FREE, addr, handle); |
| 2941 | trace(s, slab, object, 0); |
| 2942 | /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */ |
| 2943 | init_object(s, object, SLUB_RED_INACTIVE); |
| 2944 | |
| 2945 | /* Reached end of constructed freelist yet? */ |
| 2946 | if (object != tail) { |
| 2947 | object = get_freepointer(s, object); |
| 2948 | goto next_object; |
| 2949 | } |
| 2950 | checks_ok = true; |
| 2951 | |
| 2952 | out_cnt: |
| 2953 | if (cnt != *bulk_cnt) { |
| 2954 | slab_err(s, slab, "Bulk free expected %d objects but found %d\n", |
| 2955 | *bulk_cnt, cnt); |
| 2956 | *bulk_cnt = cnt; |
| 2957 | } |
| 2958 | |
| 2959 | out: |
| 2960 | |
| 2961 | if (!checks_ok) |
| 2962 | slab_fix(s, "Object at 0x%p not freed", object); |
| 2963 | |
| 2964 | return checks_ok; |
| 2965 | } |
| 2966 | #endif /* CONFIG_SLUB_DEBUG */ |
| 2967 | |
| 2968 | #if defined(CONFIG_SLUB_DEBUG) || defined(SLAB_SUPPORTS_SYSFS) |
| 2969 | static unsigned long count_partial(struct kmem_cache_node *n, |
| 2970 | int (*get_count)(struct slab *)) |
| 2971 | { |
| 2972 | unsigned long flags; |
| 2973 | unsigned long x = 0; |
| 2974 | struct slab *slab; |
| 2975 | |
| 2976 | spin_lock_irqsave(&n->list_lock, flags); |
| 2977 | list_for_each_entry(slab, &n->partial, slab_list) |
| 2978 | x += get_count(slab); |
| 2979 | spin_unlock_irqrestore(&n->list_lock, flags); |
| 2980 | return x; |
| 2981 | } |
| 2982 | #endif /* CONFIG_SLUB_DEBUG || SLAB_SUPPORTS_SYSFS */ |
| 2983 | |
| 2984 | #ifdef CONFIG_SLUB_DEBUG |
| 2985 | static noinline void |
| 2986 | slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid) |
| 2987 | { |
| 2988 | static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL, |
| 2989 | DEFAULT_RATELIMIT_BURST); |
| 2990 | int node; |
| 2991 | struct kmem_cache_node *n; |
| 2992 | |
| 2993 | if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs)) |
| 2994 | return; |
| 2995 | |
| 2996 | pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n", |
| 2997 | nid, gfpflags, &gfpflags); |
| 2998 | pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n", |
| 2999 | s->name, s->object_size, s->size, oo_order(s->oo), |
| 3000 | oo_order(s->min)); |
| 3001 | |
| 3002 | if (oo_order(s->min) > get_order(s->object_size)) |
| 3003 | pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n", |
| 3004 | s->name); |
| 3005 | |
| 3006 | for_each_kmem_cache_node(s, node, n) { |
| 3007 | unsigned long nr_slabs; |
| 3008 | unsigned long nr_objs; |
| 3009 | unsigned long nr_free; |
| 3010 | |
| 3011 | nr_free = count_partial(n, count_free); |
| 3012 | nr_slabs = node_nr_slabs(n); |
| 3013 | nr_objs = node_nr_objs(n); |
| 3014 | |
| 3015 | pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n", |
| 3016 | node, nr_slabs, nr_objs, nr_free); |
| 3017 | } |
| 3018 | } |
| 3019 | #else /* CONFIG_SLUB_DEBUG */ |
| 3020 | static inline void |
| 3021 | slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid) { } |
| 3022 | #endif |
| 3023 | |
| 3024 | static inline bool pfmemalloc_match(struct slab *slab, gfp_t gfpflags) |
| 3025 | { |
| 3026 | if (unlikely(slab_test_pfmemalloc(slab))) |
| 3027 | return gfp_pfmemalloc_allowed(gfpflags); |
| 3028 | |
| 3029 | return true; |
| 3030 | } |
| 3031 | |
| 3032 | #ifndef CONFIG_SLUB_TINY |
| 3033 | static inline bool |
| 3034 | __update_cpu_freelist_fast(struct kmem_cache *s, |
| 3035 | void *freelist_old, void *freelist_new, |
| 3036 | unsigned long tid) |
| 3037 | { |
| 3038 | freelist_aba_t old = { .freelist = freelist_old, .counter = tid }; |
| 3039 | freelist_aba_t new = { .freelist = freelist_new, .counter = next_tid(tid) }; |
| 3040 | |
| 3041 | return this_cpu_try_cmpxchg_freelist(s->cpu_slab->freelist_tid.full, |
| 3042 | &old.full, new.full); |
| 3043 | } |
| 3044 | |
| 3045 | /* |
| 3046 | * Check the slab->freelist and either transfer the freelist to the |
| 3047 | * per cpu freelist or deactivate the slab. |
| 3048 | * |
| 3049 | * The slab is still frozen if the return value is not NULL. |
| 3050 | * |
| 3051 | * If this function returns NULL then the slab has been unfrozen. |
| 3052 | */ |
| 3053 | static inline void *get_freelist(struct kmem_cache *s, struct slab *slab) |
| 3054 | { |
| 3055 | struct slab new; |
| 3056 | unsigned long counters; |
| 3057 | void *freelist; |
| 3058 | |
| 3059 | lockdep_assert_held(this_cpu_ptr(&s->cpu_slab->lock)); |
| 3060 | |
| 3061 | do { |
| 3062 | freelist = slab->freelist; |
| 3063 | counters = slab->counters; |
| 3064 | |
| 3065 | new.counters = counters; |
| 3066 | VM_BUG_ON(!new.frozen); |
| 3067 | |
| 3068 | new.inuse = slab->objects; |
| 3069 | new.frozen = freelist != NULL; |
| 3070 | |
| 3071 | } while (!__slab_update_freelist(s, slab, |
| 3072 | freelist, counters, |
| 3073 | NULL, new.counters, |
| 3074 | "get_freelist")); |
| 3075 | |
| 3076 | return freelist; |
| 3077 | } |
| 3078 | |
| 3079 | /* |
| 3080 | * Slow path. The lockless freelist is empty or we need to perform |
| 3081 | * debugging duties. |
| 3082 | * |
| 3083 | * Processing is still very fast if new objects have been freed to the |
| 3084 | * regular freelist. In that case we simply take over the regular freelist |
| 3085 | * as the lockless freelist and zap the regular freelist. |
| 3086 | * |
| 3087 | * If that is not working then we fall back to the partial lists. We take the |
| 3088 | * first element of the freelist as the object to allocate now and move the |
| 3089 | * rest of the freelist to the lockless freelist. |
| 3090 | * |
| 3091 | * And if we were unable to get a new slab from the partial slab lists then |
| 3092 | * we need to allocate a new slab. This is the slowest path since it involves |
| 3093 | * a call to the page allocator and the setup of a new slab. |
| 3094 | * |
| 3095 | * Version of __slab_alloc to use when we know that preemption is |
| 3096 | * already disabled (which is the case for bulk allocation). |
| 3097 | */ |
| 3098 | static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node, |
| 3099 | unsigned long addr, struct kmem_cache_cpu *c, unsigned int orig_size) |
| 3100 | { |
| 3101 | void *freelist; |
| 3102 | struct slab *slab; |
| 3103 | unsigned long flags; |
| 3104 | struct partial_context pc; |
| 3105 | |
| 3106 | stat(s, ALLOC_SLOWPATH); |
| 3107 | |
| 3108 | reread_slab: |
| 3109 | |
| 3110 | slab = READ_ONCE(c->slab); |
| 3111 | if (!slab) { |
| 3112 | /* |
| 3113 | * if the node is not online or has no normal memory, just |
| 3114 | * ignore the node constraint |
| 3115 | */ |
| 3116 | if (unlikely(node != NUMA_NO_NODE && |
| 3117 | !node_isset(node, slab_nodes))) |
| 3118 | node = NUMA_NO_NODE; |
| 3119 | goto new_slab; |
| 3120 | } |
| 3121 | redo: |
| 3122 | |
| 3123 | if (unlikely(!node_match(slab, node))) { |
| 3124 | /* |
| 3125 | * same as above but node_match() being false already |
| 3126 | * implies node != NUMA_NO_NODE |
| 3127 | */ |
| 3128 | if (!node_isset(node, slab_nodes)) { |
| 3129 | node = NUMA_NO_NODE; |
| 3130 | } else { |
| 3131 | stat(s, ALLOC_NODE_MISMATCH); |
| 3132 | goto deactivate_slab; |
| 3133 | } |
| 3134 | } |
| 3135 | |
| 3136 | /* |
| 3137 | * By rights, we should be searching for a slab page that was |
| 3138 | * PFMEMALLOC but right now, we are losing the pfmemalloc |
| 3139 | * information when the page leaves the per-cpu allocator |
| 3140 | */ |
| 3141 | if (unlikely(!pfmemalloc_match(slab, gfpflags))) |
| 3142 | goto deactivate_slab; |
| 3143 | |
| 3144 | /* must check again c->slab in case we got preempted and it changed */ |
| 3145 | local_lock_irqsave(&s->cpu_slab->lock, flags); |
| 3146 | if (unlikely(slab != c->slab)) { |
| 3147 | local_unlock_irqrestore(&s->cpu_slab->lock, flags); |
| 3148 | goto reread_slab; |
| 3149 | } |
| 3150 | freelist = c->freelist; |
| 3151 | if (freelist) |
| 3152 | goto load_freelist; |
| 3153 | |
| 3154 | freelist = get_freelist(s, slab); |
| 3155 | |
| 3156 | if (!freelist) { |
| 3157 | c->slab = NULL; |
| 3158 | c->tid = next_tid(c->tid); |
| 3159 | local_unlock_irqrestore(&s->cpu_slab->lock, flags); |
| 3160 | stat(s, DEACTIVATE_BYPASS); |
| 3161 | goto new_slab; |
| 3162 | } |
| 3163 | |
| 3164 | stat(s, ALLOC_REFILL); |
| 3165 | |
| 3166 | load_freelist: |
| 3167 | |
| 3168 | lockdep_assert_held(this_cpu_ptr(&s->cpu_slab->lock)); |
| 3169 | |
| 3170 | /* |
| 3171 | * freelist is pointing to the list of objects to be used. |
| 3172 | * slab is pointing to the slab from which the objects are obtained. |
| 3173 | * That slab must be frozen for per cpu allocations to work. |
| 3174 | */ |
| 3175 | VM_BUG_ON(!c->slab->frozen); |
| 3176 | c->freelist = get_freepointer(s, freelist); |
| 3177 | c->tid = next_tid(c->tid); |
| 3178 | local_unlock_irqrestore(&s->cpu_slab->lock, flags); |
| 3179 | return freelist; |
| 3180 | |
| 3181 | deactivate_slab: |
| 3182 | |
| 3183 | local_lock_irqsave(&s->cpu_slab->lock, flags); |
| 3184 | if (slab != c->slab) { |
| 3185 | local_unlock_irqrestore(&s->cpu_slab->lock, flags); |
| 3186 | goto reread_slab; |
| 3187 | } |
| 3188 | freelist = c->freelist; |
| 3189 | c->slab = NULL; |
| 3190 | c->freelist = NULL; |
| 3191 | c->tid = next_tid(c->tid); |
| 3192 | local_unlock_irqrestore(&s->cpu_slab->lock, flags); |
| 3193 | deactivate_slab(s, slab, freelist); |
| 3194 | |
| 3195 | new_slab: |
| 3196 | |
| 3197 | if (slub_percpu_partial(c)) { |
| 3198 | local_lock_irqsave(&s->cpu_slab->lock, flags); |
| 3199 | if (unlikely(c->slab)) { |
| 3200 | local_unlock_irqrestore(&s->cpu_slab->lock, flags); |
| 3201 | goto reread_slab; |
| 3202 | } |
| 3203 | if (unlikely(!slub_percpu_partial(c))) { |
| 3204 | local_unlock_irqrestore(&s->cpu_slab->lock, flags); |
| 3205 | /* we were preempted and partial list got empty */ |
| 3206 | goto new_objects; |
| 3207 | } |
| 3208 | |
| 3209 | slab = c->slab = slub_percpu_partial(c); |
| 3210 | slub_set_percpu_partial(c, slab); |
| 3211 | local_unlock_irqrestore(&s->cpu_slab->lock, flags); |
| 3212 | stat(s, CPU_PARTIAL_ALLOC); |
| 3213 | goto redo; |
| 3214 | } |
| 3215 | |
| 3216 | new_objects: |
| 3217 | |
| 3218 | pc.flags = gfpflags; |
| 3219 | pc.slab = &slab; |
| 3220 | pc.orig_size = orig_size; |
| 3221 | freelist = get_partial(s, node, &pc); |
| 3222 | if (freelist) |
| 3223 | goto check_new_slab; |
| 3224 | |
| 3225 | slub_put_cpu_ptr(s->cpu_slab); |
| 3226 | slab = new_slab(s, gfpflags, node); |
| 3227 | c = slub_get_cpu_ptr(s->cpu_slab); |
| 3228 | |
| 3229 | if (unlikely(!slab)) { |
| 3230 | slab_out_of_memory(s, gfpflags, node); |
| 3231 | return NULL; |
| 3232 | } |
| 3233 | |
| 3234 | stat(s, ALLOC_SLAB); |
| 3235 | |
| 3236 | if (kmem_cache_debug(s)) { |
| 3237 | freelist = alloc_single_from_new_slab(s, slab, orig_size); |
| 3238 | |
| 3239 | if (unlikely(!freelist)) |
| 3240 | goto new_objects; |
| 3241 | |
| 3242 | if (s->flags & SLAB_STORE_USER) |
| 3243 | set_track(s, freelist, TRACK_ALLOC, addr); |
| 3244 | |
| 3245 | return freelist; |
| 3246 | } |
| 3247 | |
| 3248 | /* |
| 3249 | * No other reference to the slab yet so we can |
| 3250 | * muck around with it freely without cmpxchg |
| 3251 | */ |
| 3252 | freelist = slab->freelist; |
| 3253 | slab->freelist = NULL; |
| 3254 | slab->inuse = slab->objects; |
| 3255 | slab->frozen = 1; |
| 3256 | |
| 3257 | inc_slabs_node(s, slab_nid(slab), slab->objects); |
| 3258 | |
| 3259 | check_new_slab: |
| 3260 | |
| 3261 | if (kmem_cache_debug(s)) { |
| 3262 | /* |
| 3263 | * For debug caches here we had to go through |
| 3264 | * alloc_single_from_partial() so just store the tracking info |
| 3265 | * and return the object |
| 3266 | */ |
| 3267 | if (s->flags & SLAB_STORE_USER) |
| 3268 | set_track(s, freelist, TRACK_ALLOC, addr); |
| 3269 | |
| 3270 | return freelist; |
| 3271 | } |
| 3272 | |
| 3273 | if (unlikely(!pfmemalloc_match(slab, gfpflags))) { |
| 3274 | /* |
| 3275 | * For !pfmemalloc_match() case we don't load freelist so that |
| 3276 | * we don't make further mismatched allocations easier. |
| 3277 | */ |
| 3278 | deactivate_slab(s, slab, get_freepointer(s, freelist)); |
| 3279 | return freelist; |
| 3280 | } |
| 3281 | |
| 3282 | retry_load_slab: |
| 3283 | |
| 3284 | local_lock_irqsave(&s->cpu_slab->lock, flags); |
| 3285 | if (unlikely(c->slab)) { |
| 3286 | void *flush_freelist = c->freelist; |
| 3287 | struct slab *flush_slab = c->slab; |
| 3288 | |
| 3289 | c->slab = NULL; |
| 3290 | c->freelist = NULL; |
| 3291 | c->tid = next_tid(c->tid); |
| 3292 | |
| 3293 | local_unlock_irqrestore(&s->cpu_slab->lock, flags); |
| 3294 | |
| 3295 | deactivate_slab(s, flush_slab, flush_freelist); |
| 3296 | |
| 3297 | stat(s, CPUSLAB_FLUSH); |
| 3298 | |
| 3299 | goto retry_load_slab; |
| 3300 | } |
| 3301 | c->slab = slab; |
| 3302 | |
| 3303 | goto load_freelist; |
| 3304 | } |
| 3305 | |
| 3306 | /* |
| 3307 | * A wrapper for ___slab_alloc() for contexts where preemption is not yet |
| 3308 | * disabled. Compensates for possible cpu changes by refetching the per cpu area |
| 3309 | * pointer. |
| 3310 | */ |
| 3311 | static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node, |
| 3312 | unsigned long addr, struct kmem_cache_cpu *c, unsigned int orig_size) |
| 3313 | { |
| 3314 | void *p; |
| 3315 | |
| 3316 | #ifdef CONFIG_PREEMPT_COUNT |
| 3317 | /* |
| 3318 | * We may have been preempted and rescheduled on a different |
| 3319 | * cpu before disabling preemption. Need to reload cpu area |
| 3320 | * pointer. |
| 3321 | */ |
| 3322 | c = slub_get_cpu_ptr(s->cpu_slab); |
| 3323 | #endif |
| 3324 | |
| 3325 | p = ___slab_alloc(s, gfpflags, node, addr, c, orig_size); |
| 3326 | #ifdef CONFIG_PREEMPT_COUNT |
| 3327 | slub_put_cpu_ptr(s->cpu_slab); |
| 3328 | #endif |
| 3329 | return p; |
| 3330 | } |
| 3331 | |
| 3332 | static __always_inline void *__slab_alloc_node(struct kmem_cache *s, |
| 3333 | gfp_t gfpflags, int node, unsigned long addr, size_t orig_size) |
| 3334 | { |
| 3335 | struct kmem_cache_cpu *c; |
| 3336 | struct slab *slab; |
| 3337 | unsigned long tid; |
| 3338 | void *object; |
| 3339 | |
| 3340 | redo: |
| 3341 | /* |
| 3342 | * Must read kmem_cache cpu data via this cpu ptr. Preemption is |
| 3343 | * enabled. We may switch back and forth between cpus while |
| 3344 | * reading from one cpu area. That does not matter as long |
| 3345 | * as we end up on the original cpu again when doing the cmpxchg. |
| 3346 | * |
| 3347 | * We must guarantee that tid and kmem_cache_cpu are retrieved on the |
| 3348 | * same cpu. We read first the kmem_cache_cpu pointer and use it to read |
| 3349 | * the tid. If we are preempted and switched to another cpu between the |
| 3350 | * two reads, it's OK as the two are still associated with the same cpu |
| 3351 | * and cmpxchg later will validate the cpu. |
| 3352 | */ |
| 3353 | c = raw_cpu_ptr(s->cpu_slab); |
| 3354 | tid = READ_ONCE(c->tid); |
| 3355 | |
| 3356 | /* |
| 3357 | * Irqless object alloc/free algorithm used here depends on sequence |
| 3358 | * of fetching cpu_slab's data. tid should be fetched before anything |
| 3359 | * on c to guarantee that object and slab associated with previous tid |
| 3360 | * won't be used with current tid. If we fetch tid first, object and |
| 3361 | * slab could be one associated with next tid and our alloc/free |
| 3362 | * request will be failed. In this case, we will retry. So, no problem. |
| 3363 | */ |
| 3364 | barrier(); |
| 3365 | |
| 3366 | /* |
| 3367 | * The transaction ids are globally unique per cpu and per operation on |
| 3368 | * a per cpu queue. Thus they can be guarantee that the cmpxchg_double |
| 3369 | * occurs on the right processor and that there was no operation on the |
| 3370 | * linked list in between. |
| 3371 | */ |
| 3372 | |
| 3373 | object = c->freelist; |
| 3374 | slab = c->slab; |
| 3375 | |
| 3376 | if (!USE_LOCKLESS_FAST_PATH() || |
| 3377 | unlikely(!object || !slab || !node_match(slab, node))) { |
| 3378 | object = __slab_alloc(s, gfpflags, node, addr, c, orig_size); |
| 3379 | } else { |
| 3380 | void *next_object = get_freepointer_safe(s, object); |
| 3381 | |
| 3382 | /* |
| 3383 | * The cmpxchg will only match if there was no additional |
| 3384 | * operation and if we are on the right processor. |
| 3385 | * |
| 3386 | * The cmpxchg does the following atomically (without lock |
| 3387 | * semantics!) |
| 3388 | * 1. Relocate first pointer to the current per cpu area. |
| 3389 | * 2. Verify that tid and freelist have not been changed |
| 3390 | * 3. If they were not changed replace tid and freelist |
| 3391 | * |
| 3392 | * Since this is without lock semantics the protection is only |
| 3393 | * against code executing on this cpu *not* from access by |
| 3394 | * other cpus. |
| 3395 | */ |
| 3396 | if (unlikely(!__update_cpu_freelist_fast(s, object, next_object, tid))) { |
| 3397 | note_cmpxchg_failure("slab_alloc", s, tid); |
| 3398 | goto redo; |
| 3399 | } |
| 3400 | prefetch_freepointer(s, next_object); |
| 3401 | stat(s, ALLOC_FASTPATH); |
| 3402 | } |
| 3403 | |
| 3404 | return object; |
| 3405 | } |
| 3406 | #else /* CONFIG_SLUB_TINY */ |
| 3407 | static void *__slab_alloc_node(struct kmem_cache *s, |
| 3408 | gfp_t gfpflags, int node, unsigned long addr, size_t orig_size) |
| 3409 | { |
| 3410 | struct partial_context pc; |
| 3411 | struct slab *slab; |
| 3412 | void *object; |
| 3413 | |
| 3414 | pc.flags = gfpflags; |
| 3415 | pc.slab = &slab; |
| 3416 | pc.orig_size = orig_size; |
| 3417 | object = get_partial(s, node, &pc); |
| 3418 | |
| 3419 | if (object) |
| 3420 | return object; |
| 3421 | |
| 3422 | slab = new_slab(s, gfpflags, node); |
| 3423 | if (unlikely(!slab)) { |
| 3424 | slab_out_of_memory(s, gfpflags, node); |
| 3425 | return NULL; |
| 3426 | } |
| 3427 | |
| 3428 | object = alloc_single_from_new_slab(s, slab, orig_size); |
| 3429 | |
| 3430 | return object; |
| 3431 | } |
| 3432 | #endif /* CONFIG_SLUB_TINY */ |
| 3433 | |
| 3434 | /* |
| 3435 | * If the object has been wiped upon free, make sure it's fully initialized by |
| 3436 | * zeroing out freelist pointer. |
| 3437 | */ |
| 3438 | static __always_inline void maybe_wipe_obj_freeptr(struct kmem_cache *s, |
| 3439 | void *obj) |
| 3440 | { |
| 3441 | if (unlikely(slab_want_init_on_free(s)) && obj) |
| 3442 | memset((void *)((char *)kasan_reset_tag(obj) + s->offset), |
| 3443 | 0, sizeof(void *)); |
| 3444 | } |
| 3445 | |
| 3446 | /* |
| 3447 | * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc) |
| 3448 | * have the fastpath folded into their functions. So no function call |
| 3449 | * overhead for requests that can be satisfied on the fastpath. |
| 3450 | * |
| 3451 | * The fastpath works by first checking if the lockless freelist can be used. |
| 3452 | * If not then __slab_alloc is called for slow processing. |
| 3453 | * |
| 3454 | * Otherwise we can simply pick the next object from the lockless free list. |
| 3455 | */ |
| 3456 | static __fastpath_inline void *slab_alloc_node(struct kmem_cache *s, struct list_lru *lru, |
| 3457 | gfp_t gfpflags, int node, unsigned long addr, size_t orig_size) |
| 3458 | { |
| 3459 | void *object; |
| 3460 | struct obj_cgroup *objcg = NULL; |
| 3461 | bool init = false; |
| 3462 | |
| 3463 | s = slab_pre_alloc_hook(s, lru, &objcg, 1, gfpflags); |
| 3464 | if (!s) |
| 3465 | return NULL; |
| 3466 | |
| 3467 | object = kfence_alloc(s, orig_size, gfpflags); |
| 3468 | if (unlikely(object)) |
| 3469 | goto out; |
| 3470 | |
| 3471 | object = __slab_alloc_node(s, gfpflags, node, addr, orig_size); |
| 3472 | |
| 3473 | maybe_wipe_obj_freeptr(s, object); |
| 3474 | init = slab_want_init_on_alloc(gfpflags, s); |
| 3475 | |
| 3476 | out: |
| 3477 | /* |
| 3478 | * When init equals 'true', like for kzalloc() family, only |
| 3479 | * @orig_size bytes might be zeroed instead of s->object_size |
| 3480 | */ |
| 3481 | slab_post_alloc_hook(s, objcg, gfpflags, 1, &object, init, orig_size); |
| 3482 | |
| 3483 | return object; |
| 3484 | } |
| 3485 | |
| 3486 | static __fastpath_inline void *slab_alloc(struct kmem_cache *s, struct list_lru *lru, |
| 3487 | gfp_t gfpflags, unsigned long addr, size_t orig_size) |
| 3488 | { |
| 3489 | return slab_alloc_node(s, lru, gfpflags, NUMA_NO_NODE, addr, orig_size); |
| 3490 | } |
| 3491 | |
| 3492 | static __fastpath_inline |
| 3493 | void *__kmem_cache_alloc_lru(struct kmem_cache *s, struct list_lru *lru, |
| 3494 | gfp_t gfpflags) |
| 3495 | { |
| 3496 | void *ret = slab_alloc(s, lru, gfpflags, _RET_IP_, s->object_size); |
| 3497 | |
| 3498 | trace_kmem_cache_alloc(_RET_IP_, ret, s, gfpflags, NUMA_NO_NODE); |
| 3499 | |
| 3500 | return ret; |
| 3501 | } |
| 3502 | |
| 3503 | void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags) |
| 3504 | { |
| 3505 | return __kmem_cache_alloc_lru(s, NULL, gfpflags); |
| 3506 | } |
| 3507 | EXPORT_SYMBOL(kmem_cache_alloc); |
| 3508 | |
| 3509 | void *kmem_cache_alloc_lru(struct kmem_cache *s, struct list_lru *lru, |
| 3510 | gfp_t gfpflags) |
| 3511 | { |
| 3512 | return __kmem_cache_alloc_lru(s, lru, gfpflags); |
| 3513 | } |
| 3514 | EXPORT_SYMBOL(kmem_cache_alloc_lru); |
| 3515 | |
| 3516 | void *__kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, |
| 3517 | int node, size_t orig_size, |
| 3518 | unsigned long caller) |
| 3519 | { |
| 3520 | return slab_alloc_node(s, NULL, gfpflags, node, |
| 3521 | caller, orig_size); |
| 3522 | } |
| 3523 | |
| 3524 | void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node) |
| 3525 | { |
| 3526 | void *ret = slab_alloc_node(s, NULL, gfpflags, node, _RET_IP_, s->object_size); |
| 3527 | |
| 3528 | trace_kmem_cache_alloc(_RET_IP_, ret, s, gfpflags, node); |
| 3529 | |
| 3530 | return ret; |
| 3531 | } |
| 3532 | EXPORT_SYMBOL(kmem_cache_alloc_node); |
| 3533 | |
| 3534 | static noinline void free_to_partial_list( |
| 3535 | struct kmem_cache *s, struct slab *slab, |
| 3536 | void *head, void *tail, int bulk_cnt, |
| 3537 | unsigned long addr) |
| 3538 | { |
| 3539 | struct kmem_cache_node *n = get_node(s, slab_nid(slab)); |
| 3540 | struct slab *slab_free = NULL; |
| 3541 | int cnt = bulk_cnt; |
| 3542 | unsigned long flags; |
| 3543 | depot_stack_handle_t handle = 0; |
| 3544 | |
| 3545 | if (s->flags & SLAB_STORE_USER) |
| 3546 | handle = set_track_prepare(); |
| 3547 | |
| 3548 | spin_lock_irqsave(&n->list_lock, flags); |
| 3549 | |
| 3550 | if (free_debug_processing(s, slab, head, tail, &cnt, addr, handle)) { |
| 3551 | void *prior = slab->freelist; |
| 3552 | |
| 3553 | /* Perform the actual freeing while we still hold the locks */ |
| 3554 | slab->inuse -= cnt; |
| 3555 | set_freepointer(s, tail, prior); |
| 3556 | slab->freelist = head; |
| 3557 | |
| 3558 | /* |
| 3559 | * If the slab is empty, and node's partial list is full, |
| 3560 | * it should be discarded anyway no matter it's on full or |
| 3561 | * partial list. |
| 3562 | */ |
| 3563 | if (slab->inuse == 0 && n->nr_partial >= s->min_partial) |
| 3564 | slab_free = slab; |
| 3565 | |
| 3566 | if (!prior) { |
| 3567 | /* was on full list */ |
| 3568 | remove_full(s, n, slab); |
| 3569 | if (!slab_free) { |
| 3570 | add_partial(n, slab, DEACTIVATE_TO_TAIL); |
| 3571 | stat(s, FREE_ADD_PARTIAL); |
| 3572 | } |
| 3573 | } else if (slab_free) { |
| 3574 | remove_partial(n, slab); |
| 3575 | stat(s, FREE_REMOVE_PARTIAL); |
| 3576 | } |
| 3577 | } |
| 3578 | |
| 3579 | if (slab_free) { |
| 3580 | /* |
| 3581 | * Update the counters while still holding n->list_lock to |
| 3582 | * prevent spurious validation warnings |
| 3583 | */ |
| 3584 | dec_slabs_node(s, slab_nid(slab_free), slab_free->objects); |
| 3585 | } |
| 3586 | |
| 3587 | spin_unlock_irqrestore(&n->list_lock, flags); |
| 3588 | |
| 3589 | if (slab_free) { |
| 3590 | stat(s, FREE_SLAB); |
| 3591 | free_slab(s, slab_free); |
| 3592 | } |
| 3593 | } |
| 3594 | |
| 3595 | /* |
| 3596 | * Slow path handling. This may still be called frequently since objects |
| 3597 | * have a longer lifetime than the cpu slabs in most processing loads. |
| 3598 | * |
| 3599 | * So we still attempt to reduce cache line usage. Just take the slab |
| 3600 | * lock and free the item. If there is no additional partial slab |
| 3601 | * handling required then we can return immediately. |
| 3602 | */ |
| 3603 | static void __slab_free(struct kmem_cache *s, struct slab *slab, |
| 3604 | void *head, void *tail, int cnt, |
| 3605 | unsigned long addr) |
| 3606 | |
| 3607 | { |
| 3608 | void *prior; |
| 3609 | int was_frozen; |
| 3610 | struct slab new; |
| 3611 | unsigned long counters; |
| 3612 | struct kmem_cache_node *n = NULL; |
| 3613 | unsigned long flags; |
| 3614 | |
| 3615 | stat(s, FREE_SLOWPATH); |
| 3616 | |
| 3617 | if (kfence_free(head)) |
| 3618 | return; |
| 3619 | |
| 3620 | if (IS_ENABLED(CONFIG_SLUB_TINY) || kmem_cache_debug(s)) { |
| 3621 | free_to_partial_list(s, slab, head, tail, cnt, addr); |
| 3622 | return; |
| 3623 | } |
| 3624 | |
| 3625 | do { |
| 3626 | if (unlikely(n)) { |
| 3627 | spin_unlock_irqrestore(&n->list_lock, flags); |
| 3628 | n = NULL; |
| 3629 | } |
| 3630 | prior = slab->freelist; |
| 3631 | counters = slab->counters; |
| 3632 | set_freepointer(s, tail, prior); |
| 3633 | new.counters = counters; |
| 3634 | was_frozen = new.frozen; |
| 3635 | new.inuse -= cnt; |
| 3636 | if ((!new.inuse || !prior) && !was_frozen) { |
| 3637 | |
| 3638 | if (kmem_cache_has_cpu_partial(s) && !prior) { |
| 3639 | |
| 3640 | /* |
| 3641 | * Slab was on no list before and will be |
| 3642 | * partially empty |
| 3643 | * We can defer the list move and instead |
| 3644 | * freeze it. |
| 3645 | */ |
| 3646 | new.frozen = 1; |
| 3647 | |
| 3648 | } else { /* Needs to be taken off a list */ |
| 3649 | |
| 3650 | n = get_node(s, slab_nid(slab)); |
| 3651 | /* |
| 3652 | * Speculatively acquire the list_lock. |
| 3653 | * If the cmpxchg does not succeed then we may |
| 3654 | * drop the list_lock without any processing. |
| 3655 | * |
| 3656 | * Otherwise the list_lock will synchronize with |
| 3657 | * other processors updating the list of slabs. |
| 3658 | */ |
| 3659 | spin_lock_irqsave(&n->list_lock, flags); |
| 3660 | |
| 3661 | } |
| 3662 | } |
| 3663 | |
| 3664 | } while (!slab_update_freelist(s, slab, |
| 3665 | prior, counters, |
| 3666 | head, new.counters, |
| 3667 | "__slab_free")); |
| 3668 | |
| 3669 | if (likely(!n)) { |
| 3670 | |
| 3671 | if (likely(was_frozen)) { |
| 3672 | /* |
| 3673 | * The list lock was not taken therefore no list |
| 3674 | * activity can be necessary. |
| 3675 | */ |
| 3676 | stat(s, FREE_FROZEN); |
| 3677 | } else if (new.frozen) { |
| 3678 | /* |
| 3679 | * If we just froze the slab then put it onto the |
| 3680 | * per cpu partial list. |
| 3681 | */ |
| 3682 | put_cpu_partial(s, slab, 1); |
| 3683 | stat(s, CPU_PARTIAL_FREE); |
| 3684 | } |
| 3685 | |
| 3686 | return; |
| 3687 | } |
| 3688 | |
| 3689 | if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) |
| 3690 | goto slab_empty; |
| 3691 | |
| 3692 | /* |
| 3693 | * Objects left in the slab. If it was not on the partial list before |
| 3694 | * then add it. |
| 3695 | */ |
| 3696 | if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) { |
| 3697 | remove_full(s, n, slab); |
| 3698 | add_partial(n, slab, DEACTIVATE_TO_TAIL); |
| 3699 | stat(s, FREE_ADD_PARTIAL); |
| 3700 | } |
| 3701 | spin_unlock_irqrestore(&n->list_lock, flags); |
| 3702 | return; |
| 3703 | |
| 3704 | slab_empty: |
| 3705 | if (prior) { |
| 3706 | /* |
| 3707 | * Slab on the partial list. |
| 3708 | */ |
| 3709 | remove_partial(n, slab); |
| 3710 | stat(s, FREE_REMOVE_PARTIAL); |
| 3711 | } else { |
| 3712 | /* Slab must be on the full list */ |
| 3713 | remove_full(s, n, slab); |
| 3714 | } |
| 3715 | |
| 3716 | spin_unlock_irqrestore(&n->list_lock, flags); |
| 3717 | stat(s, FREE_SLAB); |
| 3718 | discard_slab(s, slab); |
| 3719 | } |
| 3720 | |
| 3721 | #ifndef CONFIG_SLUB_TINY |
| 3722 | /* |
| 3723 | * Fastpath with forced inlining to produce a kfree and kmem_cache_free that |
| 3724 | * can perform fastpath freeing without additional function calls. |
| 3725 | * |
| 3726 | * The fastpath is only possible if we are freeing to the current cpu slab |
| 3727 | * of this processor. This typically the case if we have just allocated |
| 3728 | * the item before. |
| 3729 | * |
| 3730 | * If fastpath is not possible then fall back to __slab_free where we deal |
| 3731 | * with all sorts of special processing. |
| 3732 | * |
| 3733 | * Bulk free of a freelist with several objects (all pointing to the |
| 3734 | * same slab) possible by specifying head and tail ptr, plus objects |
| 3735 | * count (cnt). Bulk free indicated by tail pointer being set. |
| 3736 | */ |
| 3737 | static __always_inline void do_slab_free(struct kmem_cache *s, |
| 3738 | struct slab *slab, void *head, void *tail, |
| 3739 | int cnt, unsigned long addr) |
| 3740 | { |
| 3741 | void *tail_obj = tail ? : head; |
| 3742 | struct kmem_cache_cpu *c; |
| 3743 | unsigned long tid; |
| 3744 | void **freelist; |
| 3745 | |
| 3746 | redo: |
| 3747 | /* |
| 3748 | * Determine the currently cpus per cpu slab. |
| 3749 | * The cpu may change afterward. However that does not matter since |
| 3750 | * data is retrieved via this pointer. If we are on the same cpu |
| 3751 | * during the cmpxchg then the free will succeed. |
| 3752 | */ |
| 3753 | c = raw_cpu_ptr(s->cpu_slab); |
| 3754 | tid = READ_ONCE(c->tid); |
| 3755 | |
| 3756 | /* Same with comment on barrier() in slab_alloc_node() */ |
| 3757 | barrier(); |
| 3758 | |
| 3759 | if (unlikely(slab != c->slab)) { |
| 3760 | __slab_free(s, slab, head, tail_obj, cnt, addr); |
| 3761 | return; |
| 3762 | } |
| 3763 | |
| 3764 | if (USE_LOCKLESS_FAST_PATH()) { |
| 3765 | freelist = READ_ONCE(c->freelist); |
| 3766 | |
| 3767 | set_freepointer(s, tail_obj, freelist); |
| 3768 | |
| 3769 | if (unlikely(!__update_cpu_freelist_fast(s, freelist, head, tid))) { |
| 3770 | note_cmpxchg_failure("slab_free", s, tid); |
| 3771 | goto redo; |
| 3772 | } |
| 3773 | } else { |
| 3774 | /* Update the free list under the local lock */ |
| 3775 | local_lock(&s->cpu_slab->lock); |
| 3776 | c = this_cpu_ptr(s->cpu_slab); |
| 3777 | if (unlikely(slab != c->slab)) { |
| 3778 | local_unlock(&s->cpu_slab->lock); |
| 3779 | goto redo; |
| 3780 | } |
| 3781 | tid = c->tid; |
| 3782 | freelist = c->freelist; |
| 3783 | |
| 3784 | set_freepointer(s, tail_obj, freelist); |
| 3785 | c->freelist = head; |
| 3786 | c->tid = next_tid(tid); |
| 3787 | |
| 3788 | local_unlock(&s->cpu_slab->lock); |
| 3789 | } |
| 3790 | stat(s, FREE_FASTPATH); |
| 3791 | } |
| 3792 | #else /* CONFIG_SLUB_TINY */ |
| 3793 | static void do_slab_free(struct kmem_cache *s, |
| 3794 | struct slab *slab, void *head, void *tail, |
| 3795 | int cnt, unsigned long addr) |
| 3796 | { |
| 3797 | void *tail_obj = tail ? : head; |
| 3798 | |
| 3799 | __slab_free(s, slab, head, tail_obj, cnt, addr); |
| 3800 | } |
| 3801 | #endif /* CONFIG_SLUB_TINY */ |
| 3802 | |
| 3803 | static __fastpath_inline void slab_free(struct kmem_cache *s, struct slab *slab, |
| 3804 | void *head, void *tail, void **p, int cnt, |
| 3805 | unsigned long addr) |
| 3806 | { |
| 3807 | memcg_slab_free_hook(s, slab, p, cnt); |
| 3808 | /* |
| 3809 | * With KASAN enabled slab_free_freelist_hook modifies the freelist |
| 3810 | * to remove objects, whose reuse must be delayed. |
| 3811 | */ |
| 3812 | if (slab_free_freelist_hook(s, &head, &tail, &cnt)) |
| 3813 | do_slab_free(s, slab, head, tail, cnt, addr); |
| 3814 | } |
| 3815 | |
| 3816 | #ifdef CONFIG_KASAN_GENERIC |
| 3817 | void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr) |
| 3818 | { |
| 3819 | do_slab_free(cache, virt_to_slab(x), x, NULL, 1, addr); |
| 3820 | } |
| 3821 | #endif |
| 3822 | |
| 3823 | void __kmem_cache_free(struct kmem_cache *s, void *x, unsigned long caller) |
| 3824 | { |
| 3825 | slab_free(s, virt_to_slab(x), x, NULL, &x, 1, caller); |
| 3826 | } |
| 3827 | |
| 3828 | void kmem_cache_free(struct kmem_cache *s, void *x) |
| 3829 | { |
| 3830 | s = cache_from_obj(s, x); |
| 3831 | if (!s) |
| 3832 | return; |
| 3833 | trace_kmem_cache_free(_RET_IP_, x, s); |
| 3834 | slab_free(s, virt_to_slab(x), x, NULL, &x, 1, _RET_IP_); |
| 3835 | } |
| 3836 | EXPORT_SYMBOL(kmem_cache_free); |
| 3837 | |
| 3838 | struct detached_freelist { |
| 3839 | struct slab *slab; |
| 3840 | void *tail; |
| 3841 | void *freelist; |
| 3842 | int cnt; |
| 3843 | struct kmem_cache *s; |
| 3844 | }; |
| 3845 | |
| 3846 | /* |
| 3847 | * This function progressively scans the array with free objects (with |
| 3848 | * a limited look ahead) and extract objects belonging to the same |
| 3849 | * slab. It builds a detached freelist directly within the given |
| 3850 | * slab/objects. This can happen without any need for |
| 3851 | * synchronization, because the objects are owned by running process. |
| 3852 | * The freelist is build up as a single linked list in the objects. |
| 3853 | * The idea is, that this detached freelist can then be bulk |
| 3854 | * transferred to the real freelist(s), but only requiring a single |
| 3855 | * synchronization primitive. Look ahead in the array is limited due |
| 3856 | * to performance reasons. |
| 3857 | */ |
| 3858 | static inline |
| 3859 | int build_detached_freelist(struct kmem_cache *s, size_t size, |
| 3860 | void **p, struct detached_freelist *df) |
| 3861 | { |
| 3862 | int lookahead = 3; |
| 3863 | void *object; |
| 3864 | struct folio *folio; |
| 3865 | size_t same; |
| 3866 | |
| 3867 | object = p[--size]; |
| 3868 | folio = virt_to_folio(object); |
| 3869 | if (!s) { |
| 3870 | /* Handle kalloc'ed objects */ |
| 3871 | if (unlikely(!folio_test_slab(folio))) { |
| 3872 | free_large_kmalloc(folio, object); |
| 3873 | df->slab = NULL; |
| 3874 | return size; |
| 3875 | } |
| 3876 | /* Derive kmem_cache from object */ |
| 3877 | df->slab = folio_slab(folio); |
| 3878 | df->s = df->slab->slab_cache; |
| 3879 | } else { |
| 3880 | df->slab = folio_slab(folio); |
| 3881 | df->s = cache_from_obj(s, object); /* Support for memcg */ |
| 3882 | } |
| 3883 | |
| 3884 | /* Start new detached freelist */ |
| 3885 | df->tail = object; |
| 3886 | df->freelist = object; |
| 3887 | df->cnt = 1; |
| 3888 | |
| 3889 | if (is_kfence_address(object)) |
| 3890 | return size; |
| 3891 | |
| 3892 | set_freepointer(df->s, object, NULL); |
| 3893 | |
| 3894 | same = size; |
| 3895 | while (size) { |
| 3896 | object = p[--size]; |
| 3897 | /* df->slab is always set at this point */ |
| 3898 | if (df->slab == virt_to_slab(object)) { |
| 3899 | /* Opportunity build freelist */ |
| 3900 | set_freepointer(df->s, object, df->freelist); |
| 3901 | df->freelist = object; |
| 3902 | df->cnt++; |
| 3903 | same--; |
| 3904 | if (size != same) |
| 3905 | swap(p[size], p[same]); |
| 3906 | continue; |
| 3907 | } |
| 3908 | |
| 3909 | /* Limit look ahead search */ |
| 3910 | if (!--lookahead) |
| 3911 | break; |
| 3912 | } |
| 3913 | |
| 3914 | return same; |
| 3915 | } |
| 3916 | |
| 3917 | /* Note that interrupts must be enabled when calling this function. */ |
| 3918 | void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p) |
| 3919 | { |
| 3920 | if (!size) |
| 3921 | return; |
| 3922 | |
| 3923 | do { |
| 3924 | struct detached_freelist df; |
| 3925 | |
| 3926 | size = build_detached_freelist(s, size, p, &df); |
| 3927 | if (!df.slab) |
| 3928 | continue; |
| 3929 | |
| 3930 | slab_free(df.s, df.slab, df.freelist, df.tail, &p[size], df.cnt, |
| 3931 | _RET_IP_); |
| 3932 | } while (likely(size)); |
| 3933 | } |
| 3934 | EXPORT_SYMBOL(kmem_cache_free_bulk); |
| 3935 | |
| 3936 | #ifndef CONFIG_SLUB_TINY |
| 3937 | static inline int __kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, |
| 3938 | size_t size, void **p, struct obj_cgroup *objcg) |
| 3939 | { |
| 3940 | struct kmem_cache_cpu *c; |
| 3941 | unsigned long irqflags; |
| 3942 | int i; |
| 3943 | |
| 3944 | /* |
| 3945 | * Drain objects in the per cpu slab, while disabling local |
| 3946 | * IRQs, which protects against PREEMPT and interrupts |
| 3947 | * handlers invoking normal fastpath. |
| 3948 | */ |
| 3949 | c = slub_get_cpu_ptr(s->cpu_slab); |
| 3950 | local_lock_irqsave(&s->cpu_slab->lock, irqflags); |
| 3951 | |
| 3952 | for (i = 0; i < size; i++) { |
| 3953 | void *object = kfence_alloc(s, s->object_size, flags); |
| 3954 | |
| 3955 | if (unlikely(object)) { |
| 3956 | p[i] = object; |
| 3957 | continue; |
| 3958 | } |
| 3959 | |
| 3960 | object = c->freelist; |
| 3961 | if (unlikely(!object)) { |
| 3962 | /* |
| 3963 | * We may have removed an object from c->freelist using |
| 3964 | * the fastpath in the previous iteration; in that case, |
| 3965 | * c->tid has not been bumped yet. |
| 3966 | * Since ___slab_alloc() may reenable interrupts while |
| 3967 | * allocating memory, we should bump c->tid now. |
| 3968 | */ |
| 3969 | c->tid = next_tid(c->tid); |
| 3970 | |
| 3971 | local_unlock_irqrestore(&s->cpu_slab->lock, irqflags); |
| 3972 | |
| 3973 | /* |
| 3974 | * Invoking slow path likely have side-effect |
| 3975 | * of re-populating per CPU c->freelist |
| 3976 | */ |
| 3977 | p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE, |
| 3978 | _RET_IP_, c, s->object_size); |
| 3979 | if (unlikely(!p[i])) |
| 3980 | goto error; |
| 3981 | |
| 3982 | c = this_cpu_ptr(s->cpu_slab); |
| 3983 | maybe_wipe_obj_freeptr(s, p[i]); |
| 3984 | |
| 3985 | local_lock_irqsave(&s->cpu_slab->lock, irqflags); |
| 3986 | |
| 3987 | continue; /* goto for-loop */ |
| 3988 | } |
| 3989 | c->freelist = get_freepointer(s, object); |
| 3990 | p[i] = object; |
| 3991 | maybe_wipe_obj_freeptr(s, p[i]); |
| 3992 | } |
| 3993 | c->tid = next_tid(c->tid); |
| 3994 | local_unlock_irqrestore(&s->cpu_slab->lock, irqflags); |
| 3995 | slub_put_cpu_ptr(s->cpu_slab); |
| 3996 | |
| 3997 | return i; |
| 3998 | |
| 3999 | error: |
| 4000 | slub_put_cpu_ptr(s->cpu_slab); |
| 4001 | slab_post_alloc_hook(s, objcg, flags, i, p, false, s->object_size); |
| 4002 | kmem_cache_free_bulk(s, i, p); |
| 4003 | return 0; |
| 4004 | |
| 4005 | } |
| 4006 | #else /* CONFIG_SLUB_TINY */ |
| 4007 | static int __kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, |
| 4008 | size_t size, void **p, struct obj_cgroup *objcg) |
| 4009 | { |
| 4010 | int i; |
| 4011 | |
| 4012 | for (i = 0; i < size; i++) { |
| 4013 | void *object = kfence_alloc(s, s->object_size, flags); |
| 4014 | |
| 4015 | if (unlikely(object)) { |
| 4016 | p[i] = object; |
| 4017 | continue; |
| 4018 | } |
| 4019 | |
| 4020 | p[i] = __slab_alloc_node(s, flags, NUMA_NO_NODE, |
| 4021 | _RET_IP_, s->object_size); |
| 4022 | if (unlikely(!p[i])) |
| 4023 | goto error; |
| 4024 | |
| 4025 | maybe_wipe_obj_freeptr(s, p[i]); |
| 4026 | } |
| 4027 | |
| 4028 | return i; |
| 4029 | |
| 4030 | error: |
| 4031 | slab_post_alloc_hook(s, objcg, flags, i, p, false, s->object_size); |
| 4032 | kmem_cache_free_bulk(s, i, p); |
| 4033 | return 0; |
| 4034 | } |
| 4035 | #endif /* CONFIG_SLUB_TINY */ |
| 4036 | |
| 4037 | /* Note that interrupts must be enabled when calling this function. */ |
| 4038 | int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size, |
| 4039 | void **p) |
| 4040 | { |
| 4041 | int i; |
| 4042 | struct obj_cgroup *objcg = NULL; |
| 4043 | |
| 4044 | if (!size) |
| 4045 | return 0; |
| 4046 | |
| 4047 | /* memcg and kmem_cache debug support */ |
| 4048 | s = slab_pre_alloc_hook(s, NULL, &objcg, size, flags); |
| 4049 | if (unlikely(!s)) |
| 4050 | return 0; |
| 4051 | |
| 4052 | i = __kmem_cache_alloc_bulk(s, flags, size, p, objcg); |
| 4053 | |
| 4054 | /* |
| 4055 | * memcg and kmem_cache debug support and memory initialization. |
| 4056 | * Done outside of the IRQ disabled fastpath loop. |
| 4057 | */ |
| 4058 | if (i != 0) |
| 4059 | slab_post_alloc_hook(s, objcg, flags, size, p, |
| 4060 | slab_want_init_on_alloc(flags, s), s->object_size); |
| 4061 | return i; |
| 4062 | } |
| 4063 | EXPORT_SYMBOL(kmem_cache_alloc_bulk); |
| 4064 | |
| 4065 | |
| 4066 | /* |
| 4067 | * Object placement in a slab is made very easy because we always start at |
| 4068 | * offset 0. If we tune the size of the object to the alignment then we can |
| 4069 | * get the required alignment by putting one properly sized object after |
| 4070 | * another. |
| 4071 | * |
| 4072 | * Notice that the allocation order determines the sizes of the per cpu |
| 4073 | * caches. Each processor has always one slab available for allocations. |
| 4074 | * Increasing the allocation order reduces the number of times that slabs |
| 4075 | * must be moved on and off the partial lists and is therefore a factor in |
| 4076 | * locking overhead. |
| 4077 | */ |
| 4078 | |
| 4079 | /* |
| 4080 | * Minimum / Maximum order of slab pages. This influences locking overhead |
| 4081 | * and slab fragmentation. A higher order reduces the number of partial slabs |
| 4082 | * and increases the number of allocations possible without having to |
| 4083 | * take the list_lock. |
| 4084 | */ |
| 4085 | static unsigned int slub_min_order; |
| 4086 | static unsigned int slub_max_order = |
| 4087 | IS_ENABLED(CONFIG_SLUB_TINY) ? 1 : PAGE_ALLOC_COSTLY_ORDER; |
| 4088 | static unsigned int slub_min_objects; |
| 4089 | |
| 4090 | /* |
| 4091 | * Calculate the order of allocation given an slab object size. |
| 4092 | * |
| 4093 | * The order of allocation has significant impact on performance and other |
| 4094 | * system components. Generally order 0 allocations should be preferred since |
| 4095 | * order 0 does not cause fragmentation in the page allocator. Larger objects |
| 4096 | * be problematic to put into order 0 slabs because there may be too much |
| 4097 | * unused space left. We go to a higher order if more than 1/16th of the slab |
| 4098 | * would be wasted. |
| 4099 | * |
| 4100 | * In order to reach satisfactory performance we must ensure that a minimum |
| 4101 | * number of objects is in one slab. Otherwise we may generate too much |
| 4102 | * activity on the partial lists which requires taking the list_lock. This is |
| 4103 | * less a concern for large slabs though which are rarely used. |
| 4104 | * |
| 4105 | * slub_max_order specifies the order where we begin to stop considering the |
| 4106 | * number of objects in a slab as critical. If we reach slub_max_order then |
| 4107 | * we try to keep the page order as low as possible. So we accept more waste |
| 4108 | * of space in favor of a small page order. |
| 4109 | * |
| 4110 | * Higher order allocations also allow the placement of more objects in a |
| 4111 | * slab and thereby reduce object handling overhead. If the user has |
| 4112 | * requested a higher minimum order then we start with that one instead of |
| 4113 | * the smallest order which will fit the object. |
| 4114 | */ |
| 4115 | static inline unsigned int calc_slab_order(unsigned int size, |
| 4116 | unsigned int min_objects, unsigned int max_order, |
| 4117 | unsigned int fract_leftover) |
| 4118 | { |
| 4119 | unsigned int min_order = slub_min_order; |
| 4120 | unsigned int order; |
| 4121 | |
| 4122 | if (order_objects(min_order, size) > MAX_OBJS_PER_PAGE) |
| 4123 | return get_order(size * MAX_OBJS_PER_PAGE) - 1; |
| 4124 | |
| 4125 | for (order = max(min_order, (unsigned int)get_order(min_objects * size)); |
| 4126 | order <= max_order; order++) { |
| 4127 | |
| 4128 | unsigned int slab_size = (unsigned int)PAGE_SIZE << order; |
| 4129 | unsigned int rem; |
| 4130 | |
| 4131 | rem = slab_size % size; |
| 4132 | |
| 4133 | if (rem <= slab_size / fract_leftover) |
| 4134 | break; |
| 4135 | } |
| 4136 | |
| 4137 | return order; |
| 4138 | } |
| 4139 | |
| 4140 | static inline int calculate_order(unsigned int size) |
| 4141 | { |
| 4142 | unsigned int order; |
| 4143 | unsigned int min_objects; |
| 4144 | unsigned int max_objects; |
| 4145 | unsigned int nr_cpus; |
| 4146 | |
| 4147 | /* |
| 4148 | * Attempt to find best configuration for a slab. This |
| 4149 | * works by first attempting to generate a layout with |
| 4150 | * the best configuration and backing off gradually. |
| 4151 | * |
| 4152 | * First we increase the acceptable waste in a slab. Then |
| 4153 | * we reduce the minimum objects required in a slab. |
| 4154 | */ |
| 4155 | min_objects = slub_min_objects; |
| 4156 | if (!min_objects) { |
| 4157 | /* |
| 4158 | * Some architectures will only update present cpus when |
| 4159 | * onlining them, so don't trust the number if it's just 1. But |
| 4160 | * we also don't want to use nr_cpu_ids always, as on some other |
| 4161 | * architectures, there can be many possible cpus, but never |
| 4162 | * onlined. Here we compromise between trying to avoid too high |
| 4163 | * order on systems that appear larger than they are, and too |
| 4164 | * low order on systems that appear smaller than they are. |
| 4165 | */ |
| 4166 | nr_cpus = num_present_cpus(); |
| 4167 | if (nr_cpus <= 1) |
| 4168 | nr_cpus = nr_cpu_ids; |
| 4169 | min_objects = 4 * (fls(nr_cpus) + 1); |
| 4170 | } |
| 4171 | max_objects = order_objects(slub_max_order, size); |
| 4172 | min_objects = min(min_objects, max_objects); |
| 4173 | |
| 4174 | while (min_objects > 1) { |
| 4175 | unsigned int fraction; |
| 4176 | |
| 4177 | fraction = 16; |
| 4178 | while (fraction >= 4) { |
| 4179 | order = calc_slab_order(size, min_objects, |
| 4180 | slub_max_order, fraction); |
| 4181 | if (order <= slub_max_order) |
| 4182 | return order; |
| 4183 | fraction /= 2; |
| 4184 | } |
| 4185 | min_objects--; |
| 4186 | } |
| 4187 | |
| 4188 | /* |
| 4189 | * We were unable to place multiple objects in a slab. Now |
| 4190 | * lets see if we can place a single object there. |
| 4191 | */ |
| 4192 | order = calc_slab_order(size, 1, slub_max_order, 1); |
| 4193 | if (order <= slub_max_order) |
| 4194 | return order; |
| 4195 | |
| 4196 | /* |
| 4197 | * Doh this slab cannot be placed using slub_max_order. |
| 4198 | */ |
| 4199 | order = calc_slab_order(size, 1, MAX_ORDER, 1); |
| 4200 | if (order <= MAX_ORDER) |
| 4201 | return order; |
| 4202 | return -ENOSYS; |
| 4203 | } |
| 4204 | |
| 4205 | static void |
| 4206 | init_kmem_cache_node(struct kmem_cache_node *n) |
| 4207 | { |
| 4208 | n->nr_partial = 0; |
| 4209 | spin_lock_init(&n->list_lock); |
| 4210 | INIT_LIST_HEAD(&n->partial); |
| 4211 | #ifdef CONFIG_SLUB_DEBUG |
| 4212 | atomic_long_set(&n->nr_slabs, 0); |
| 4213 | atomic_long_set(&n->total_objects, 0); |
| 4214 | INIT_LIST_HEAD(&n->full); |
| 4215 | #endif |
| 4216 | } |
| 4217 | |
| 4218 | #ifndef CONFIG_SLUB_TINY |
| 4219 | static inline int alloc_kmem_cache_cpus(struct kmem_cache *s) |
| 4220 | { |
| 4221 | BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE < |
| 4222 | NR_KMALLOC_TYPES * KMALLOC_SHIFT_HIGH * |
| 4223 | sizeof(struct kmem_cache_cpu)); |
| 4224 | |
| 4225 | /* |
| 4226 | * Must align to double word boundary for the double cmpxchg |
| 4227 | * instructions to work; see __pcpu_double_call_return_bool(). |
| 4228 | */ |
| 4229 | s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu), |
| 4230 | 2 * sizeof(void *)); |
| 4231 | |
| 4232 | if (!s->cpu_slab) |
| 4233 | return 0; |
| 4234 | |
| 4235 | init_kmem_cache_cpus(s); |
| 4236 | |
| 4237 | return 1; |
| 4238 | } |
| 4239 | #else |
| 4240 | static inline int alloc_kmem_cache_cpus(struct kmem_cache *s) |
| 4241 | { |
| 4242 | return 1; |
| 4243 | } |
| 4244 | #endif /* CONFIG_SLUB_TINY */ |
| 4245 | |
| 4246 | static struct kmem_cache *kmem_cache_node; |
| 4247 | |
| 4248 | /* |
| 4249 | * No kmalloc_node yet so do it by hand. We know that this is the first |
| 4250 | * slab on the node for this slabcache. There are no concurrent accesses |
| 4251 | * possible. |
| 4252 | * |
| 4253 | * Note that this function only works on the kmem_cache_node |
| 4254 | * when allocating for the kmem_cache_node. This is used for bootstrapping |
| 4255 | * memory on a fresh node that has no slab structures yet. |
| 4256 | */ |
| 4257 | static void early_kmem_cache_node_alloc(int node) |
| 4258 | { |
| 4259 | struct slab *slab; |
| 4260 | struct kmem_cache_node *n; |
| 4261 | |
| 4262 | BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node)); |
| 4263 | |
| 4264 | slab = new_slab(kmem_cache_node, GFP_NOWAIT, node); |
| 4265 | |
| 4266 | BUG_ON(!slab); |
| 4267 | inc_slabs_node(kmem_cache_node, slab_nid(slab), slab->objects); |
| 4268 | if (slab_nid(slab) != node) { |
| 4269 | pr_err("SLUB: Unable to allocate memory from node %d\n", node); |
| 4270 | pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n"); |
| 4271 | } |
| 4272 | |
| 4273 | n = slab->freelist; |
| 4274 | BUG_ON(!n); |
| 4275 | #ifdef CONFIG_SLUB_DEBUG |
| 4276 | init_object(kmem_cache_node, n, SLUB_RED_ACTIVE); |
| 4277 | init_tracking(kmem_cache_node, n); |
| 4278 | #endif |
| 4279 | n = kasan_slab_alloc(kmem_cache_node, n, GFP_KERNEL, false); |
| 4280 | slab->freelist = get_freepointer(kmem_cache_node, n); |
| 4281 | slab->inuse = 1; |
| 4282 | kmem_cache_node->node[node] = n; |
| 4283 | init_kmem_cache_node(n); |
| 4284 | inc_slabs_node(kmem_cache_node, node, slab->objects); |
| 4285 | |
| 4286 | /* |
| 4287 | * No locks need to be taken here as it has just been |
| 4288 | * initialized and there is no concurrent access. |
| 4289 | */ |
| 4290 | __add_partial(n, slab, DEACTIVATE_TO_HEAD); |
| 4291 | } |
| 4292 | |
| 4293 | static void free_kmem_cache_nodes(struct kmem_cache *s) |
| 4294 | { |
| 4295 | int node; |
| 4296 | struct kmem_cache_node *n; |
| 4297 | |
| 4298 | for_each_kmem_cache_node(s, node, n) { |
| 4299 | s->node[node] = NULL; |
| 4300 | kmem_cache_free(kmem_cache_node, n); |
| 4301 | } |
| 4302 | } |
| 4303 | |
| 4304 | void __kmem_cache_release(struct kmem_cache *s) |
| 4305 | { |
| 4306 | cache_random_seq_destroy(s); |
| 4307 | #ifndef CONFIG_SLUB_TINY |
| 4308 | free_percpu(s->cpu_slab); |
| 4309 | #endif |
| 4310 | free_kmem_cache_nodes(s); |
| 4311 | } |
| 4312 | |
| 4313 | static int init_kmem_cache_nodes(struct kmem_cache *s) |
| 4314 | { |
| 4315 | int node; |
| 4316 | |
| 4317 | for_each_node_mask(node, slab_nodes) { |
| 4318 | struct kmem_cache_node *n; |
| 4319 | |
| 4320 | if (slab_state == DOWN) { |
| 4321 | early_kmem_cache_node_alloc(node); |
| 4322 | continue; |
| 4323 | } |
| 4324 | n = kmem_cache_alloc_node(kmem_cache_node, |
| 4325 | GFP_KERNEL, node); |
| 4326 | |
| 4327 | if (!n) { |
| 4328 | free_kmem_cache_nodes(s); |
| 4329 | return 0; |
| 4330 | } |
| 4331 | |
| 4332 | init_kmem_cache_node(n); |
| 4333 | s->node[node] = n; |
| 4334 | } |
| 4335 | return 1; |
| 4336 | } |
| 4337 | |
| 4338 | static void set_cpu_partial(struct kmem_cache *s) |
| 4339 | { |
| 4340 | #ifdef CONFIG_SLUB_CPU_PARTIAL |
| 4341 | unsigned int nr_objects; |
| 4342 | |
| 4343 | /* |
| 4344 | * cpu_partial determined the maximum number of objects kept in the |
| 4345 | * per cpu partial lists of a processor. |
| 4346 | * |
| 4347 | * Per cpu partial lists mainly contain slabs that just have one |
| 4348 | * object freed. If they are used for allocation then they can be |
| 4349 | * filled up again with minimal effort. The slab will never hit the |
| 4350 | * per node partial lists and therefore no locking will be required. |
| 4351 | * |
| 4352 | * For backwards compatibility reasons, this is determined as number |
| 4353 | * of objects, even though we now limit maximum number of pages, see |
| 4354 | * slub_set_cpu_partial() |
| 4355 | */ |
| 4356 | if (!kmem_cache_has_cpu_partial(s)) |
| 4357 | nr_objects = 0; |
| 4358 | else if (s->size >= PAGE_SIZE) |
| 4359 | nr_objects = 6; |
| 4360 | else if (s->size >= 1024) |
| 4361 | nr_objects = 24; |
| 4362 | else if (s->size >= 256) |
| 4363 | nr_objects = 52; |
| 4364 | else |
| 4365 | nr_objects = 120; |
| 4366 | |
| 4367 | slub_set_cpu_partial(s, nr_objects); |
| 4368 | #endif |
| 4369 | } |
| 4370 | |
| 4371 | /* |
| 4372 | * calculate_sizes() determines the order and the distribution of data within |
| 4373 | * a slab object. |
| 4374 | */ |
| 4375 | static int calculate_sizes(struct kmem_cache *s) |
| 4376 | { |
| 4377 | slab_flags_t flags = s->flags; |
| 4378 | unsigned int size = s->object_size; |
| 4379 | unsigned int order; |
| 4380 | |
| 4381 | /* |
| 4382 | * Round up object size to the next word boundary. We can only |
| 4383 | * place the free pointer at word boundaries and this determines |
| 4384 | * the possible location of the free pointer. |
| 4385 | */ |
| 4386 | size = ALIGN(size, sizeof(void *)); |
| 4387 | |
| 4388 | #ifdef CONFIG_SLUB_DEBUG |
| 4389 | /* |
| 4390 | * Determine if we can poison the object itself. If the user of |
| 4391 | * the slab may touch the object after free or before allocation |
| 4392 | * then we should never poison the object itself. |
| 4393 | */ |
| 4394 | if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) && |
| 4395 | !s->ctor) |
| 4396 | s->flags |= __OBJECT_POISON; |
| 4397 | else |
| 4398 | s->flags &= ~__OBJECT_POISON; |
| 4399 | |
| 4400 | |
| 4401 | /* |
| 4402 | * If we are Redzoning then check if there is some space between the |
| 4403 | * end of the object and the free pointer. If not then add an |
| 4404 | * additional word to have some bytes to store Redzone information. |
| 4405 | */ |
| 4406 | if ((flags & SLAB_RED_ZONE) && size == s->object_size) |
| 4407 | size += sizeof(void *); |
| 4408 | #endif |
| 4409 | |
| 4410 | /* |
| 4411 | * With that we have determined the number of bytes in actual use |
| 4412 | * by the object and redzoning. |
| 4413 | */ |
| 4414 | s->inuse = size; |
| 4415 | |
| 4416 | if (slub_debug_orig_size(s) || |
| 4417 | (flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) || |
| 4418 | ((flags & SLAB_RED_ZONE) && s->object_size < sizeof(void *)) || |
| 4419 | s->ctor) { |
| 4420 | /* |
| 4421 | * Relocate free pointer after the object if it is not |
| 4422 | * permitted to overwrite the first word of the object on |
| 4423 | * kmem_cache_free. |
| 4424 | * |
| 4425 | * This is the case if we do RCU, have a constructor or |
| 4426 | * destructor, are poisoning the objects, or are |
| 4427 | * redzoning an object smaller than sizeof(void *). |
| 4428 | * |
| 4429 | * The assumption that s->offset >= s->inuse means free |
| 4430 | * pointer is outside of the object is used in the |
| 4431 | * freeptr_outside_object() function. If that is no |
| 4432 | * longer true, the function needs to be modified. |
| 4433 | */ |
| 4434 | s->offset = size; |
| 4435 | size += sizeof(void *); |
| 4436 | } else { |
| 4437 | /* |
| 4438 | * Store freelist pointer near middle of object to keep |
| 4439 | * it away from the edges of the object to avoid small |
| 4440 | * sized over/underflows from neighboring allocations. |
| 4441 | */ |
| 4442 | s->offset = ALIGN_DOWN(s->object_size / 2, sizeof(void *)); |
| 4443 | } |
| 4444 | |
| 4445 | #ifdef CONFIG_SLUB_DEBUG |
| 4446 | if (flags & SLAB_STORE_USER) { |
| 4447 | /* |
| 4448 | * Need to store information about allocs and frees after |
| 4449 | * the object. |
| 4450 | */ |
| 4451 | size += 2 * sizeof(struct track); |
| 4452 | |
| 4453 | /* Save the original kmalloc request size */ |
| 4454 | if (flags & SLAB_KMALLOC) |
| 4455 | size += sizeof(unsigned int); |
| 4456 | } |
| 4457 | #endif |
| 4458 | |
| 4459 | kasan_cache_create(s, &size, &s->flags); |
| 4460 | #ifdef CONFIG_SLUB_DEBUG |
| 4461 | if (flags & SLAB_RED_ZONE) { |
| 4462 | /* |
| 4463 | * Add some empty padding so that we can catch |
| 4464 | * overwrites from earlier objects rather than let |
| 4465 | * tracking information or the free pointer be |
| 4466 | * corrupted if a user writes before the start |
| 4467 | * of the object. |
| 4468 | */ |
| 4469 | size += sizeof(void *); |
| 4470 | |
| 4471 | s->red_left_pad = sizeof(void *); |
| 4472 | s->red_left_pad = ALIGN(s->red_left_pad, s->align); |
| 4473 | size += s->red_left_pad; |
| 4474 | } |
| 4475 | #endif |
| 4476 | |
| 4477 | /* |
| 4478 | * SLUB stores one object immediately after another beginning from |
| 4479 | * offset 0. In order to align the objects we have to simply size |
| 4480 | * each object to conform to the alignment. |
| 4481 | */ |
| 4482 | size = ALIGN(size, s->align); |
| 4483 | s->size = size; |
| 4484 | s->reciprocal_size = reciprocal_value(size); |
| 4485 | order = calculate_order(size); |
| 4486 | |
| 4487 | if ((int)order < 0) |
| 4488 | return 0; |
| 4489 | |
| 4490 | s->allocflags = 0; |
| 4491 | if (order) |
| 4492 | s->allocflags |= __GFP_COMP; |
| 4493 | |
| 4494 | if (s->flags & SLAB_CACHE_DMA) |
| 4495 | s->allocflags |= GFP_DMA; |
| 4496 | |
| 4497 | if (s->flags & SLAB_CACHE_DMA32) |
| 4498 | s->allocflags |= GFP_DMA32; |
| 4499 | |
| 4500 | if (s->flags & SLAB_RECLAIM_ACCOUNT) |
| 4501 | s->allocflags |= __GFP_RECLAIMABLE; |
| 4502 | |
| 4503 | /* |
| 4504 | * Determine the number of objects per slab |
| 4505 | */ |
| 4506 | s->oo = oo_make(order, size); |
| 4507 | s->min = oo_make(get_order(size), size); |
| 4508 | |
| 4509 | return !!oo_objects(s->oo); |
| 4510 | } |
| 4511 | |
| 4512 | static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags) |
| 4513 | { |
| 4514 | s->flags = kmem_cache_flags(s->size, flags, s->name); |
| 4515 | #ifdef CONFIG_SLAB_FREELIST_HARDENED |
| 4516 | s->random = get_random_long(); |
| 4517 | #endif |
| 4518 | |
| 4519 | if (!calculate_sizes(s)) |
| 4520 | goto error; |
| 4521 | if (disable_higher_order_debug) { |
| 4522 | /* |
| 4523 | * Disable debugging flags that store metadata if the min slab |
| 4524 | * order increased. |
| 4525 | */ |
| 4526 | if (get_order(s->size) > get_order(s->object_size)) { |
| 4527 | s->flags &= ~DEBUG_METADATA_FLAGS; |
| 4528 | s->offset = 0; |
| 4529 | if (!calculate_sizes(s)) |
| 4530 | goto error; |
| 4531 | } |
| 4532 | } |
| 4533 | |
| 4534 | #ifdef system_has_freelist_aba |
| 4535 | if (system_has_freelist_aba() && !(s->flags & SLAB_NO_CMPXCHG)) { |
| 4536 | /* Enable fast mode */ |
| 4537 | s->flags |= __CMPXCHG_DOUBLE; |
| 4538 | } |
| 4539 | #endif |
| 4540 | |
| 4541 | /* |
| 4542 | * The larger the object size is, the more slabs we want on the partial |
| 4543 | * list to avoid pounding the page allocator excessively. |
| 4544 | */ |
| 4545 | s->min_partial = min_t(unsigned long, MAX_PARTIAL, ilog2(s->size) / 2); |
| 4546 | s->min_partial = max_t(unsigned long, MIN_PARTIAL, s->min_partial); |
| 4547 | |
| 4548 | set_cpu_partial(s); |
| 4549 | |
| 4550 | #ifdef CONFIG_NUMA |
| 4551 | s->remote_node_defrag_ratio = 1000; |
| 4552 | #endif |
| 4553 | |
| 4554 | /* Initialize the pre-computed randomized freelist if slab is up */ |
| 4555 | if (slab_state >= UP) { |
| 4556 | if (init_cache_random_seq(s)) |
| 4557 | goto error; |
| 4558 | } |
| 4559 | |
| 4560 | if (!init_kmem_cache_nodes(s)) |
| 4561 | goto error; |
| 4562 | |
| 4563 | if (alloc_kmem_cache_cpus(s)) |
| 4564 | return 0; |
| 4565 | |
| 4566 | error: |
| 4567 | __kmem_cache_release(s); |
| 4568 | return -EINVAL; |
| 4569 | } |
| 4570 | |
| 4571 | static void list_slab_objects(struct kmem_cache *s, struct slab *slab, |
| 4572 | const char *text) |
| 4573 | { |
| 4574 | #ifdef CONFIG_SLUB_DEBUG |
| 4575 | void *addr = slab_address(slab); |
| 4576 | void *p; |
| 4577 | |
| 4578 | slab_err(s, slab, text, s->name); |
| 4579 | |
| 4580 | spin_lock(&object_map_lock); |
| 4581 | __fill_map(object_map, s, slab); |
| 4582 | |
| 4583 | for_each_object(p, s, addr, slab->objects) { |
| 4584 | |
| 4585 | if (!test_bit(__obj_to_index(s, addr, p), object_map)) { |
| 4586 | pr_err("Object 0x%p @offset=%tu\n", p, p - addr); |
| 4587 | print_tracking(s, p); |
| 4588 | } |
| 4589 | } |
| 4590 | spin_unlock(&object_map_lock); |
| 4591 | #endif |
| 4592 | } |
| 4593 | |
| 4594 | /* |
| 4595 | * Attempt to free all partial slabs on a node. |
| 4596 | * This is called from __kmem_cache_shutdown(). We must take list_lock |
| 4597 | * because sysfs file might still access partial list after the shutdowning. |
| 4598 | */ |
| 4599 | static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n) |
| 4600 | { |
| 4601 | LIST_HEAD(discard); |
| 4602 | struct slab *slab, *h; |
| 4603 | |
| 4604 | BUG_ON(irqs_disabled()); |
| 4605 | spin_lock_irq(&n->list_lock); |
| 4606 | list_for_each_entry_safe(slab, h, &n->partial, slab_list) { |
| 4607 | if (!slab->inuse) { |
| 4608 | remove_partial(n, slab); |
| 4609 | list_add(&slab->slab_list, &discard); |
| 4610 | } else { |
| 4611 | list_slab_objects(s, slab, |
| 4612 | "Objects remaining in %s on __kmem_cache_shutdown()"); |
| 4613 | } |
| 4614 | } |
| 4615 | spin_unlock_irq(&n->list_lock); |
| 4616 | |
| 4617 | list_for_each_entry_safe(slab, h, &discard, slab_list) |
| 4618 | discard_slab(s, slab); |
| 4619 | } |
| 4620 | |
| 4621 | bool __kmem_cache_empty(struct kmem_cache *s) |
| 4622 | { |
| 4623 | int node; |
| 4624 | struct kmem_cache_node *n; |
| 4625 | |
| 4626 | for_each_kmem_cache_node(s, node, n) |
| 4627 | if (n->nr_partial || node_nr_slabs(n)) |
| 4628 | return false; |
| 4629 | return true; |
| 4630 | } |
| 4631 | |
| 4632 | /* |
| 4633 | * Release all resources used by a slab cache. |
| 4634 | */ |
| 4635 | int __kmem_cache_shutdown(struct kmem_cache *s) |
| 4636 | { |
| 4637 | int node; |
| 4638 | struct kmem_cache_node *n; |
| 4639 | |
| 4640 | flush_all_cpus_locked(s); |
| 4641 | /* Attempt to free all objects */ |
| 4642 | for_each_kmem_cache_node(s, node, n) { |
| 4643 | free_partial(s, n); |
| 4644 | if (n->nr_partial || node_nr_slabs(n)) |
| 4645 | return 1; |
| 4646 | } |
| 4647 | return 0; |
| 4648 | } |
| 4649 | |
| 4650 | #ifdef CONFIG_PRINTK |
| 4651 | void __kmem_obj_info(struct kmem_obj_info *kpp, void *object, struct slab *slab) |
| 4652 | { |
| 4653 | void *base; |
| 4654 | int __maybe_unused i; |
| 4655 | unsigned int objnr; |
| 4656 | void *objp; |
| 4657 | void *objp0; |
| 4658 | struct kmem_cache *s = slab->slab_cache; |
| 4659 | struct track __maybe_unused *trackp; |
| 4660 | |
| 4661 | kpp->kp_ptr = object; |
| 4662 | kpp->kp_slab = slab; |
| 4663 | kpp->kp_slab_cache = s; |
| 4664 | base = slab_address(slab); |
| 4665 | objp0 = kasan_reset_tag(object); |
| 4666 | #ifdef CONFIG_SLUB_DEBUG |
| 4667 | objp = restore_red_left(s, objp0); |
| 4668 | #else |
| 4669 | objp = objp0; |
| 4670 | #endif |
| 4671 | objnr = obj_to_index(s, slab, objp); |
| 4672 | kpp->kp_data_offset = (unsigned long)((char *)objp0 - (char *)objp); |
| 4673 | objp = base + s->size * objnr; |
| 4674 | kpp->kp_objp = objp; |
| 4675 | if (WARN_ON_ONCE(objp < base || objp >= base + slab->objects * s->size |
| 4676 | || (objp - base) % s->size) || |
| 4677 | !(s->flags & SLAB_STORE_USER)) |
| 4678 | return; |
| 4679 | #ifdef CONFIG_SLUB_DEBUG |
| 4680 | objp = fixup_red_left(s, objp); |
| 4681 | trackp = get_track(s, objp, TRACK_ALLOC); |
| 4682 | kpp->kp_ret = (void *)trackp->addr; |
| 4683 | #ifdef CONFIG_STACKDEPOT |
| 4684 | { |
| 4685 | depot_stack_handle_t handle; |
| 4686 | unsigned long *entries; |
| 4687 | unsigned int nr_entries; |
| 4688 | |
| 4689 | handle = READ_ONCE(trackp->handle); |
| 4690 | if (handle) { |
| 4691 | nr_entries = stack_depot_fetch(handle, &entries); |
| 4692 | for (i = 0; i < KS_ADDRS_COUNT && i < nr_entries; i++) |
| 4693 | kpp->kp_stack[i] = (void *)entries[i]; |
| 4694 | } |
| 4695 | |
| 4696 | trackp = get_track(s, objp, TRACK_FREE); |
| 4697 | handle = READ_ONCE(trackp->handle); |
| 4698 | if (handle) { |
| 4699 | nr_entries = stack_depot_fetch(handle, &entries); |
| 4700 | for (i = 0; i < KS_ADDRS_COUNT && i < nr_entries; i++) |
| 4701 | kpp->kp_free_stack[i] = (void *)entries[i]; |
| 4702 | } |
| 4703 | } |
| 4704 | #endif |
| 4705 | #endif |
| 4706 | } |
| 4707 | #endif |
| 4708 | |
| 4709 | /******************************************************************** |
| 4710 | * Kmalloc subsystem |
| 4711 | *******************************************************************/ |
| 4712 | |
| 4713 | static int __init setup_slub_min_order(char *str) |
| 4714 | { |
| 4715 | get_option(&str, (int *)&slub_min_order); |
| 4716 | |
| 4717 | return 1; |
| 4718 | } |
| 4719 | |
| 4720 | __setup("slub_min_order=", setup_slub_min_order); |
| 4721 | |
| 4722 | static int __init setup_slub_max_order(char *str) |
| 4723 | { |
| 4724 | get_option(&str, (int *)&slub_max_order); |
| 4725 | slub_max_order = min_t(unsigned int, slub_max_order, MAX_ORDER); |
| 4726 | |
| 4727 | return 1; |
| 4728 | } |
| 4729 | |
| 4730 | __setup("slub_max_order=", setup_slub_max_order); |
| 4731 | |
| 4732 | static int __init setup_slub_min_objects(char *str) |
| 4733 | { |
| 4734 | get_option(&str, (int *)&slub_min_objects); |
| 4735 | |
| 4736 | return 1; |
| 4737 | } |
| 4738 | |
| 4739 | __setup("slub_min_objects=", setup_slub_min_objects); |
| 4740 | |
| 4741 | #ifdef CONFIG_HARDENED_USERCOPY |
| 4742 | /* |
| 4743 | * Rejects incorrectly sized objects and objects that are to be copied |
| 4744 | * to/from userspace but do not fall entirely within the containing slab |
| 4745 | * cache's usercopy region. |
| 4746 | * |
| 4747 | * Returns NULL if check passes, otherwise const char * to name of cache |
| 4748 | * to indicate an error. |
| 4749 | */ |
| 4750 | void __check_heap_object(const void *ptr, unsigned long n, |
| 4751 | const struct slab *slab, bool to_user) |
| 4752 | { |
| 4753 | struct kmem_cache *s; |
| 4754 | unsigned int offset; |
| 4755 | bool is_kfence = is_kfence_address(ptr); |
| 4756 | |
| 4757 | ptr = kasan_reset_tag(ptr); |
| 4758 | |
| 4759 | /* Find object and usable object size. */ |
| 4760 | s = slab->slab_cache; |
| 4761 | |
| 4762 | /* Reject impossible pointers. */ |
| 4763 | if (ptr < slab_address(slab)) |
| 4764 | usercopy_abort("SLUB object not in SLUB page?!", NULL, |
| 4765 | to_user, 0, n); |
| 4766 | |
| 4767 | /* Find offset within object. */ |
| 4768 | if (is_kfence) |
| 4769 | offset = ptr - kfence_object_start(ptr); |
| 4770 | else |
| 4771 | offset = (ptr - slab_address(slab)) % s->size; |
| 4772 | |
| 4773 | /* Adjust for redzone and reject if within the redzone. */ |
| 4774 | if (!is_kfence && kmem_cache_debug_flags(s, SLAB_RED_ZONE)) { |
| 4775 | if (offset < s->red_left_pad) |
| 4776 | usercopy_abort("SLUB object in left red zone", |
| 4777 | s->name, to_user, offset, n); |
| 4778 | offset -= s->red_left_pad; |
| 4779 | } |
| 4780 | |
| 4781 | /* Allow address range falling entirely within usercopy region. */ |
| 4782 | if (offset >= s->useroffset && |
| 4783 | offset - s->useroffset <= s->usersize && |
| 4784 | n <= s->useroffset - offset + s->usersize) |
| 4785 | return; |
| 4786 | |
| 4787 | usercopy_abort("SLUB object", s->name, to_user, offset, n); |
| 4788 | } |
| 4789 | #endif /* CONFIG_HARDENED_USERCOPY */ |
| 4790 | |
| 4791 | #define SHRINK_PROMOTE_MAX 32 |
| 4792 | |
| 4793 | /* |
| 4794 | * kmem_cache_shrink discards empty slabs and promotes the slabs filled |
| 4795 | * up most to the head of the partial lists. New allocations will then |
| 4796 | * fill those up and thus they can be removed from the partial lists. |
| 4797 | * |
| 4798 | * The slabs with the least items are placed last. This results in them |
| 4799 | * being allocated from last increasing the chance that the last objects |
| 4800 | * are freed in them. |
| 4801 | */ |
| 4802 | static int __kmem_cache_do_shrink(struct kmem_cache *s) |
| 4803 | { |
| 4804 | int node; |
| 4805 | int i; |
| 4806 | struct kmem_cache_node *n; |
| 4807 | struct slab *slab; |
| 4808 | struct slab *t; |
| 4809 | struct list_head discard; |
| 4810 | struct list_head promote[SHRINK_PROMOTE_MAX]; |
| 4811 | unsigned long flags; |
| 4812 | int ret = 0; |
| 4813 | |
| 4814 | for_each_kmem_cache_node(s, node, n) { |
| 4815 | INIT_LIST_HEAD(&discard); |
| 4816 | for (i = 0; i < SHRINK_PROMOTE_MAX; i++) |
| 4817 | INIT_LIST_HEAD(promote + i); |
| 4818 | |
| 4819 | spin_lock_irqsave(&n->list_lock, flags); |
| 4820 | |
| 4821 | /* |
| 4822 | * Build lists of slabs to discard or promote. |
| 4823 | * |
| 4824 | * Note that concurrent frees may occur while we hold the |
| 4825 | * list_lock. slab->inuse here is the upper limit. |
| 4826 | */ |
| 4827 | list_for_each_entry_safe(slab, t, &n->partial, slab_list) { |
| 4828 | int free = slab->objects - slab->inuse; |
| 4829 | |
| 4830 | /* Do not reread slab->inuse */ |
| 4831 | barrier(); |
| 4832 | |
| 4833 | /* We do not keep full slabs on the list */ |
| 4834 | BUG_ON(free <= 0); |
| 4835 | |
| 4836 | if (free == slab->objects) { |
| 4837 | list_move(&slab->slab_list, &discard); |
| 4838 | n->nr_partial--; |
| 4839 | dec_slabs_node(s, node, slab->objects); |
| 4840 | } else if (free <= SHRINK_PROMOTE_MAX) |
| 4841 | list_move(&slab->slab_list, promote + free - 1); |
| 4842 | } |
| 4843 | |
| 4844 | /* |
| 4845 | * Promote the slabs filled up most to the head of the |
| 4846 | * partial list. |
| 4847 | */ |
| 4848 | for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--) |
| 4849 | list_splice(promote + i, &n->partial); |
| 4850 | |
| 4851 | spin_unlock_irqrestore(&n->list_lock, flags); |
| 4852 | |
| 4853 | /* Release empty slabs */ |
| 4854 | list_for_each_entry_safe(slab, t, &discard, slab_list) |
| 4855 | free_slab(s, slab); |
| 4856 | |
| 4857 | if (node_nr_slabs(n)) |
| 4858 | ret = 1; |
| 4859 | } |
| 4860 | |
| 4861 | return ret; |
| 4862 | } |
| 4863 | |
| 4864 | int __kmem_cache_shrink(struct kmem_cache *s) |
| 4865 | { |
| 4866 | flush_all(s); |
| 4867 | return __kmem_cache_do_shrink(s); |
| 4868 | } |
| 4869 | |
| 4870 | static int slab_mem_going_offline_callback(void *arg) |
| 4871 | { |
| 4872 | struct kmem_cache *s; |
| 4873 | |
| 4874 | mutex_lock(&slab_mutex); |
| 4875 | list_for_each_entry(s, &slab_caches, list) { |
| 4876 | flush_all_cpus_locked(s); |
| 4877 | __kmem_cache_do_shrink(s); |
| 4878 | } |
| 4879 | mutex_unlock(&slab_mutex); |
| 4880 | |
| 4881 | return 0; |
| 4882 | } |
| 4883 | |
| 4884 | static void slab_mem_offline_callback(void *arg) |
| 4885 | { |
| 4886 | struct memory_notify *marg = arg; |
| 4887 | int offline_node; |
| 4888 | |
| 4889 | offline_node = marg->status_change_nid_normal; |
| 4890 | |
| 4891 | /* |
| 4892 | * If the node still has available memory. we need kmem_cache_node |
| 4893 | * for it yet. |
| 4894 | */ |
| 4895 | if (offline_node < 0) |
| 4896 | return; |
| 4897 | |
| 4898 | mutex_lock(&slab_mutex); |
| 4899 | node_clear(offline_node, slab_nodes); |
| 4900 | /* |
| 4901 | * We no longer free kmem_cache_node structures here, as it would be |
| 4902 | * racy with all get_node() users, and infeasible to protect them with |
| 4903 | * slab_mutex. |
| 4904 | */ |
| 4905 | mutex_unlock(&slab_mutex); |
| 4906 | } |
| 4907 | |
| 4908 | static int slab_mem_going_online_callback(void *arg) |
| 4909 | { |
| 4910 | struct kmem_cache_node *n; |
| 4911 | struct kmem_cache *s; |
| 4912 | struct memory_notify *marg = arg; |
| 4913 | int nid = marg->status_change_nid_normal; |
| 4914 | int ret = 0; |
| 4915 | |
| 4916 | /* |
| 4917 | * If the node's memory is already available, then kmem_cache_node is |
| 4918 | * already created. Nothing to do. |
| 4919 | */ |
| 4920 | if (nid < 0) |
| 4921 | return 0; |
| 4922 | |
| 4923 | /* |
| 4924 | * We are bringing a node online. No memory is available yet. We must |
| 4925 | * allocate a kmem_cache_node structure in order to bring the node |
| 4926 | * online. |
| 4927 | */ |
| 4928 | mutex_lock(&slab_mutex); |
| 4929 | list_for_each_entry(s, &slab_caches, list) { |
| 4930 | /* |
| 4931 | * The structure may already exist if the node was previously |
| 4932 | * onlined and offlined. |
| 4933 | */ |
| 4934 | if (get_node(s, nid)) |
| 4935 | continue; |
| 4936 | /* |
| 4937 | * XXX: kmem_cache_alloc_node will fallback to other nodes |
| 4938 | * since memory is not yet available from the node that |
| 4939 | * is brought up. |
| 4940 | */ |
| 4941 | n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL); |
| 4942 | if (!n) { |
| 4943 | ret = -ENOMEM; |
| 4944 | goto out; |
| 4945 | } |
| 4946 | init_kmem_cache_node(n); |
| 4947 | s->node[nid] = n; |
| 4948 | } |
| 4949 | /* |
| 4950 | * Any cache created after this point will also have kmem_cache_node |
| 4951 | * initialized for the new node. |
| 4952 | */ |
| 4953 | node_set(nid, slab_nodes); |
| 4954 | out: |
| 4955 | mutex_unlock(&slab_mutex); |
| 4956 | return ret; |
| 4957 | } |
| 4958 | |
| 4959 | static int slab_memory_callback(struct notifier_block *self, |
| 4960 | unsigned long action, void *arg) |
| 4961 | { |
| 4962 | int ret = 0; |
| 4963 | |
| 4964 | switch (action) { |
| 4965 | case MEM_GOING_ONLINE: |
| 4966 | ret = slab_mem_going_online_callback(arg); |
| 4967 | break; |
| 4968 | case MEM_GOING_OFFLINE: |
| 4969 | ret = slab_mem_going_offline_callback(arg); |
| 4970 | break; |
| 4971 | case MEM_OFFLINE: |
| 4972 | case MEM_CANCEL_ONLINE: |
| 4973 | slab_mem_offline_callback(arg); |
| 4974 | break; |
| 4975 | case MEM_ONLINE: |
| 4976 | case MEM_CANCEL_OFFLINE: |
| 4977 | break; |
| 4978 | } |
| 4979 | if (ret) |
| 4980 | ret = notifier_from_errno(ret); |
| 4981 | else |
| 4982 | ret = NOTIFY_OK; |
| 4983 | return ret; |
| 4984 | } |
| 4985 | |
| 4986 | /******************************************************************** |
| 4987 | * Basic setup of slabs |
| 4988 | *******************************************************************/ |
| 4989 | |
| 4990 | /* |
| 4991 | * Used for early kmem_cache structures that were allocated using |
| 4992 | * the page allocator. Allocate them properly then fix up the pointers |
| 4993 | * that may be pointing to the wrong kmem_cache structure. |
| 4994 | */ |
| 4995 | |
| 4996 | static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache) |
| 4997 | { |
| 4998 | int node; |
| 4999 | struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT); |
| 5000 | struct kmem_cache_node *n; |
| 5001 | |
| 5002 | memcpy(s, static_cache, kmem_cache->object_size); |
| 5003 | |
| 5004 | /* |
| 5005 | * This runs very early, and only the boot processor is supposed to be |
| 5006 | * up. Even if it weren't true, IRQs are not up so we couldn't fire |
| 5007 | * IPIs around. |
| 5008 | */ |
| 5009 | __flush_cpu_slab(s, smp_processor_id()); |
| 5010 | for_each_kmem_cache_node(s, node, n) { |
| 5011 | struct slab *p; |
| 5012 | |
| 5013 | list_for_each_entry(p, &n->partial, slab_list) |
| 5014 | p->slab_cache = s; |
| 5015 | |
| 5016 | #ifdef CONFIG_SLUB_DEBUG |
| 5017 | list_for_each_entry(p, &n->full, slab_list) |
| 5018 | p->slab_cache = s; |
| 5019 | #endif |
| 5020 | } |
| 5021 | list_add(&s->list, &slab_caches); |
| 5022 | return s; |
| 5023 | } |
| 5024 | |
| 5025 | void __init kmem_cache_init(void) |
| 5026 | { |
| 5027 | static __initdata struct kmem_cache boot_kmem_cache, |
| 5028 | boot_kmem_cache_node; |
| 5029 | int node; |
| 5030 | |
| 5031 | if (debug_guardpage_minorder()) |
| 5032 | slub_max_order = 0; |
| 5033 | |
| 5034 | /* Print slub debugging pointers without hashing */ |
| 5035 | if (__slub_debug_enabled()) |
| 5036 | no_hash_pointers_enable(NULL); |
| 5037 | |
| 5038 | kmem_cache_node = &boot_kmem_cache_node; |
| 5039 | kmem_cache = &boot_kmem_cache; |
| 5040 | |
| 5041 | /* |
| 5042 | * Initialize the nodemask for which we will allocate per node |
| 5043 | * structures. Here we don't need taking slab_mutex yet. |
| 5044 | */ |
| 5045 | for_each_node_state(node, N_NORMAL_MEMORY) |
| 5046 | node_set(node, slab_nodes); |
| 5047 | |
| 5048 | create_boot_cache(kmem_cache_node, "kmem_cache_node", |
| 5049 | sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN, 0, 0); |
| 5050 | |
| 5051 | hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI); |
| 5052 | |
| 5053 | /* Able to allocate the per node structures */ |
| 5054 | slab_state = PARTIAL; |
| 5055 | |
| 5056 | create_boot_cache(kmem_cache, "kmem_cache", |
| 5057 | offsetof(struct kmem_cache, node) + |
| 5058 | nr_node_ids * sizeof(struct kmem_cache_node *), |
| 5059 | SLAB_HWCACHE_ALIGN, 0, 0); |
| 5060 | |
| 5061 | kmem_cache = bootstrap(&boot_kmem_cache); |
| 5062 | kmem_cache_node = bootstrap(&boot_kmem_cache_node); |
| 5063 | |
| 5064 | /* Now we can use the kmem_cache to allocate kmalloc slabs */ |
| 5065 | setup_kmalloc_cache_index_table(); |
| 5066 | create_kmalloc_caches(0); |
| 5067 | |
| 5068 | /* Setup random freelists for each cache */ |
| 5069 | init_freelist_randomization(); |
| 5070 | |
| 5071 | cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL, |
| 5072 | slub_cpu_dead); |
| 5073 | |
| 5074 | pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n", |
| 5075 | cache_line_size(), |
| 5076 | slub_min_order, slub_max_order, slub_min_objects, |
| 5077 | nr_cpu_ids, nr_node_ids); |
| 5078 | } |
| 5079 | |
| 5080 | void __init kmem_cache_init_late(void) |
| 5081 | { |
| 5082 | #ifndef CONFIG_SLUB_TINY |
| 5083 | flushwq = alloc_workqueue("slub_flushwq", WQ_MEM_RECLAIM, 0); |
| 5084 | WARN_ON(!flushwq); |
| 5085 | #endif |
| 5086 | } |
| 5087 | |
| 5088 | struct kmem_cache * |
| 5089 | __kmem_cache_alias(const char *name, unsigned int size, unsigned int align, |
| 5090 | slab_flags_t flags, void (*ctor)(void *)) |
| 5091 | { |
| 5092 | struct kmem_cache *s; |
| 5093 | |
| 5094 | s = find_mergeable(size, align, flags, name, ctor); |
| 5095 | if (s) { |
| 5096 | if (sysfs_slab_alias(s, name)) |
| 5097 | return NULL; |
| 5098 | |
| 5099 | s->refcount++; |
| 5100 | |
| 5101 | /* |
| 5102 | * Adjust the object sizes so that we clear |
| 5103 | * the complete object on kzalloc. |
| 5104 | */ |
| 5105 | s->object_size = max(s->object_size, size); |
| 5106 | s->inuse = max(s->inuse, ALIGN(size, sizeof(void *))); |
| 5107 | } |
| 5108 | |
| 5109 | return s; |
| 5110 | } |
| 5111 | |
| 5112 | int __kmem_cache_create(struct kmem_cache *s, slab_flags_t flags) |
| 5113 | { |
| 5114 | int err; |
| 5115 | |
| 5116 | err = kmem_cache_open(s, flags); |
| 5117 | if (err) |
| 5118 | return err; |
| 5119 | |
| 5120 | /* Mutex is not taken during early boot */ |
| 5121 | if (slab_state <= UP) |
| 5122 | return 0; |
| 5123 | |
| 5124 | err = sysfs_slab_add(s); |
| 5125 | if (err) { |
| 5126 | __kmem_cache_release(s); |
| 5127 | return err; |
| 5128 | } |
| 5129 | |
| 5130 | if (s->flags & SLAB_STORE_USER) |
| 5131 | debugfs_slab_add(s); |
| 5132 | |
| 5133 | return 0; |
| 5134 | } |
| 5135 | |
| 5136 | #ifdef SLAB_SUPPORTS_SYSFS |
| 5137 | static int count_inuse(struct slab *slab) |
| 5138 | { |
| 5139 | return slab->inuse; |
| 5140 | } |
| 5141 | |
| 5142 | static int count_total(struct slab *slab) |
| 5143 | { |
| 5144 | return slab->objects; |
| 5145 | } |
| 5146 | #endif |
| 5147 | |
| 5148 | #ifdef CONFIG_SLUB_DEBUG |
| 5149 | static void validate_slab(struct kmem_cache *s, struct slab *slab, |
| 5150 | unsigned long *obj_map) |
| 5151 | { |
| 5152 | void *p; |
| 5153 | void *addr = slab_address(slab); |
| 5154 | |
| 5155 | if (!check_slab(s, slab) || !on_freelist(s, slab, NULL)) |
| 5156 | return; |
| 5157 | |
| 5158 | /* Now we know that a valid freelist exists */ |
| 5159 | __fill_map(obj_map, s, slab); |
| 5160 | for_each_object(p, s, addr, slab->objects) { |
| 5161 | u8 val = test_bit(__obj_to_index(s, addr, p), obj_map) ? |
| 5162 | SLUB_RED_INACTIVE : SLUB_RED_ACTIVE; |
| 5163 | |
| 5164 | if (!check_object(s, slab, p, val)) |
| 5165 | break; |
| 5166 | } |
| 5167 | } |
| 5168 | |
| 5169 | static int validate_slab_node(struct kmem_cache *s, |
| 5170 | struct kmem_cache_node *n, unsigned long *obj_map) |
| 5171 | { |
| 5172 | unsigned long count = 0; |
| 5173 | struct slab *slab; |
| 5174 | unsigned long flags; |
| 5175 | |
| 5176 | spin_lock_irqsave(&n->list_lock, flags); |
| 5177 | |
| 5178 | list_for_each_entry(slab, &n->partial, slab_list) { |
| 5179 | validate_slab(s, slab, obj_map); |
| 5180 | count++; |
| 5181 | } |
| 5182 | if (count != n->nr_partial) { |
| 5183 | pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n", |
| 5184 | s->name, count, n->nr_partial); |
| 5185 | slab_add_kunit_errors(); |
| 5186 | } |
| 5187 | |
| 5188 | if (!(s->flags & SLAB_STORE_USER)) |
| 5189 | goto out; |
| 5190 | |
| 5191 | list_for_each_entry(slab, &n->full, slab_list) { |
| 5192 | validate_slab(s, slab, obj_map); |
| 5193 | count++; |
| 5194 | } |
| 5195 | if (count != node_nr_slabs(n)) { |
| 5196 | pr_err("SLUB: %s %ld slabs counted but counter=%ld\n", |
| 5197 | s->name, count, node_nr_slabs(n)); |
| 5198 | slab_add_kunit_errors(); |
| 5199 | } |
| 5200 | |
| 5201 | out: |
| 5202 | spin_unlock_irqrestore(&n->list_lock, flags); |
| 5203 | return count; |
| 5204 | } |
| 5205 | |
| 5206 | long validate_slab_cache(struct kmem_cache *s) |
| 5207 | { |
| 5208 | int node; |
| 5209 | unsigned long count = 0; |
| 5210 | struct kmem_cache_node *n; |
| 5211 | unsigned long *obj_map; |
| 5212 | |
| 5213 | obj_map = bitmap_alloc(oo_objects(s->oo), GFP_KERNEL); |
| 5214 | if (!obj_map) |
| 5215 | return -ENOMEM; |
| 5216 | |
| 5217 | flush_all(s); |
| 5218 | for_each_kmem_cache_node(s, node, n) |
| 5219 | count += validate_slab_node(s, n, obj_map); |
| 5220 | |
| 5221 | bitmap_free(obj_map); |
| 5222 | |
| 5223 | return count; |
| 5224 | } |
| 5225 | EXPORT_SYMBOL(validate_slab_cache); |
| 5226 | |
| 5227 | #ifdef CONFIG_DEBUG_FS |
| 5228 | /* |
| 5229 | * Generate lists of code addresses where slabcache objects are allocated |
| 5230 | * and freed. |
| 5231 | */ |
| 5232 | |
| 5233 | struct location { |
| 5234 | depot_stack_handle_t handle; |
| 5235 | unsigned long count; |
| 5236 | unsigned long addr; |
| 5237 | unsigned long waste; |
| 5238 | long long sum_time; |
| 5239 | long min_time; |
| 5240 | long max_time; |
| 5241 | long min_pid; |
| 5242 | long max_pid; |
| 5243 | DECLARE_BITMAP(cpus, NR_CPUS); |
| 5244 | nodemask_t nodes; |
| 5245 | }; |
| 5246 | |
| 5247 | struct loc_track { |
| 5248 | unsigned long max; |
| 5249 | unsigned long count; |
| 5250 | struct location *loc; |
| 5251 | loff_t idx; |
| 5252 | }; |
| 5253 | |
| 5254 | static struct dentry *slab_debugfs_root; |
| 5255 | |
| 5256 | static void free_loc_track(struct loc_track *t) |
| 5257 | { |
| 5258 | if (t->max) |
| 5259 | free_pages((unsigned long)t->loc, |
| 5260 | get_order(sizeof(struct location) * t->max)); |
| 5261 | } |
| 5262 | |
| 5263 | static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags) |
| 5264 | { |
| 5265 | struct location *l; |
| 5266 | int order; |
| 5267 | |
| 5268 | order = get_order(sizeof(struct location) * max); |
| 5269 | |
| 5270 | l = (void *)__get_free_pages(flags, order); |
| 5271 | if (!l) |
| 5272 | return 0; |
| 5273 | |
| 5274 | if (t->count) { |
| 5275 | memcpy(l, t->loc, sizeof(struct location) * t->count); |
| 5276 | free_loc_track(t); |
| 5277 | } |
| 5278 | t->max = max; |
| 5279 | t->loc = l; |
| 5280 | return 1; |
| 5281 | } |
| 5282 | |
| 5283 | static int add_location(struct loc_track *t, struct kmem_cache *s, |
| 5284 | const struct track *track, |
| 5285 | unsigned int orig_size) |
| 5286 | { |
| 5287 | long start, end, pos; |
| 5288 | struct location *l; |
| 5289 | unsigned long caddr, chandle, cwaste; |
| 5290 | unsigned long age = jiffies - track->when; |
| 5291 | depot_stack_handle_t handle = 0; |
| 5292 | unsigned int waste = s->object_size - orig_size; |
| 5293 | |
| 5294 | #ifdef CONFIG_STACKDEPOT |
| 5295 | handle = READ_ONCE(track->handle); |
| 5296 | #endif |
| 5297 | start = -1; |
| 5298 | end = t->count; |
| 5299 | |
| 5300 | for ( ; ; ) { |
| 5301 | pos = start + (end - start + 1) / 2; |
| 5302 | |
| 5303 | /* |
| 5304 | * There is nothing at "end". If we end up there |
| 5305 | * we need to add something to before end. |
| 5306 | */ |
| 5307 | if (pos == end) |
| 5308 | break; |
| 5309 | |
| 5310 | l = &t->loc[pos]; |
| 5311 | caddr = l->addr; |
| 5312 | chandle = l->handle; |
| 5313 | cwaste = l->waste; |
| 5314 | if ((track->addr == caddr) && (handle == chandle) && |
| 5315 | (waste == cwaste)) { |
| 5316 | |
| 5317 | l->count++; |
| 5318 | if (track->when) { |
| 5319 | l->sum_time += age; |
| 5320 | if (age < l->min_time) |
| 5321 | l->min_time = age; |
| 5322 | if (age > l->max_time) |
| 5323 | l->max_time = age; |
| 5324 | |
| 5325 | if (track->pid < l->min_pid) |
| 5326 | l->min_pid = track->pid; |
| 5327 | if (track->pid > l->max_pid) |
| 5328 | l->max_pid = track->pid; |
| 5329 | |
| 5330 | cpumask_set_cpu(track->cpu, |
| 5331 | to_cpumask(l->cpus)); |
| 5332 | } |
| 5333 | node_set(page_to_nid(virt_to_page(track)), l->nodes); |
| 5334 | return 1; |
| 5335 | } |
| 5336 | |
| 5337 | if (track->addr < caddr) |
| 5338 | end = pos; |
| 5339 | else if (track->addr == caddr && handle < chandle) |
| 5340 | end = pos; |
| 5341 | else if (track->addr == caddr && handle == chandle && |
| 5342 | waste < cwaste) |
| 5343 | end = pos; |
| 5344 | else |
| 5345 | start = pos; |
| 5346 | } |
| 5347 | |
| 5348 | /* |
| 5349 | * Not found. Insert new tracking element. |
| 5350 | */ |
| 5351 | if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC)) |
| 5352 | return 0; |
| 5353 | |
| 5354 | l = t->loc + pos; |
| 5355 | if (pos < t->count) |
| 5356 | memmove(l + 1, l, |
| 5357 | (t->count - pos) * sizeof(struct location)); |
| 5358 | t->count++; |
| 5359 | l->count = 1; |
| 5360 | l->addr = track->addr; |
| 5361 | l->sum_time = age; |
| 5362 | l->min_time = age; |
| 5363 | l->max_time = age; |
| 5364 | l->min_pid = track->pid; |
| 5365 | l->max_pid = track->pid; |
| 5366 | l->handle = handle; |
| 5367 | l->waste = waste; |
| 5368 | cpumask_clear(to_cpumask(l->cpus)); |
| 5369 | cpumask_set_cpu(track->cpu, to_cpumask(l->cpus)); |
| 5370 | nodes_clear(l->nodes); |
| 5371 | node_set(page_to_nid(virt_to_page(track)), l->nodes); |
| 5372 | return 1; |
| 5373 | } |
| 5374 | |
| 5375 | static void process_slab(struct loc_track *t, struct kmem_cache *s, |
| 5376 | struct slab *slab, enum track_item alloc, |
| 5377 | unsigned long *obj_map) |
| 5378 | { |
| 5379 | void *addr = slab_address(slab); |
| 5380 | bool is_alloc = (alloc == TRACK_ALLOC); |
| 5381 | void *p; |
| 5382 | |
| 5383 | __fill_map(obj_map, s, slab); |
| 5384 | |
| 5385 | for_each_object(p, s, addr, slab->objects) |
| 5386 | if (!test_bit(__obj_to_index(s, addr, p), obj_map)) |
| 5387 | add_location(t, s, get_track(s, p, alloc), |
| 5388 | is_alloc ? get_orig_size(s, p) : |
| 5389 | s->object_size); |
| 5390 | } |
| 5391 | #endif /* CONFIG_DEBUG_FS */ |
| 5392 | #endif /* CONFIG_SLUB_DEBUG */ |
| 5393 | |
| 5394 | #ifdef SLAB_SUPPORTS_SYSFS |
| 5395 | enum slab_stat_type { |
| 5396 | SL_ALL, /* All slabs */ |
| 5397 | SL_PARTIAL, /* Only partially allocated slabs */ |
| 5398 | SL_CPU, /* Only slabs used for cpu caches */ |
| 5399 | SL_OBJECTS, /* Determine allocated objects not slabs */ |
| 5400 | SL_TOTAL /* Determine object capacity not slabs */ |
| 5401 | }; |
| 5402 | |
| 5403 | #define SO_ALL (1 << SL_ALL) |
| 5404 | #define SO_PARTIAL (1 << SL_PARTIAL) |
| 5405 | #define SO_CPU (1 << SL_CPU) |
| 5406 | #define SO_OBJECTS (1 << SL_OBJECTS) |
| 5407 | #define SO_TOTAL (1 << SL_TOTAL) |
| 5408 | |
| 5409 | static ssize_t show_slab_objects(struct kmem_cache *s, |
| 5410 | char *buf, unsigned long flags) |
| 5411 | { |
| 5412 | unsigned long total = 0; |
| 5413 | int node; |
| 5414 | int x; |
| 5415 | unsigned long *nodes; |
| 5416 | int len = 0; |
| 5417 | |
| 5418 | nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL); |
| 5419 | if (!nodes) |
| 5420 | return -ENOMEM; |
| 5421 | |
| 5422 | if (flags & SO_CPU) { |
| 5423 | int cpu; |
| 5424 | |
| 5425 | for_each_possible_cpu(cpu) { |
| 5426 | struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, |
| 5427 | cpu); |
| 5428 | int node; |
| 5429 | struct slab *slab; |
| 5430 | |
| 5431 | slab = READ_ONCE(c->slab); |
| 5432 | if (!slab) |
| 5433 | continue; |
| 5434 | |
| 5435 | node = slab_nid(slab); |
| 5436 | if (flags & SO_TOTAL) |
| 5437 | x = slab->objects; |
| 5438 | else if (flags & SO_OBJECTS) |
| 5439 | x = slab->inuse; |
| 5440 | else |
| 5441 | x = 1; |
| 5442 | |
| 5443 | total += x; |
| 5444 | nodes[node] += x; |
| 5445 | |
| 5446 | #ifdef CONFIG_SLUB_CPU_PARTIAL |
| 5447 | slab = slub_percpu_partial_read_once(c); |
| 5448 | if (slab) { |
| 5449 | node = slab_nid(slab); |
| 5450 | if (flags & SO_TOTAL) |
| 5451 | WARN_ON_ONCE(1); |
| 5452 | else if (flags & SO_OBJECTS) |
| 5453 | WARN_ON_ONCE(1); |
| 5454 | else |
| 5455 | x = slab->slabs; |
| 5456 | total += x; |
| 5457 | nodes[node] += x; |
| 5458 | } |
| 5459 | #endif |
| 5460 | } |
| 5461 | } |
| 5462 | |
| 5463 | /* |
| 5464 | * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex" |
| 5465 | * already held which will conflict with an existing lock order: |
| 5466 | * |
| 5467 | * mem_hotplug_lock->slab_mutex->kernfs_mutex |
| 5468 | * |
| 5469 | * We don't really need mem_hotplug_lock (to hold off |
| 5470 | * slab_mem_going_offline_callback) here because slab's memory hot |
| 5471 | * unplug code doesn't destroy the kmem_cache->node[] data. |
| 5472 | */ |
| 5473 | |
| 5474 | #ifdef CONFIG_SLUB_DEBUG |
| 5475 | if (flags & SO_ALL) { |
| 5476 | struct kmem_cache_node *n; |
| 5477 | |
| 5478 | for_each_kmem_cache_node(s, node, n) { |
| 5479 | |
| 5480 | if (flags & SO_TOTAL) |
| 5481 | x = node_nr_objs(n); |
| 5482 | else if (flags & SO_OBJECTS) |
| 5483 | x = node_nr_objs(n) - count_partial(n, count_free); |
| 5484 | else |
| 5485 | x = node_nr_slabs(n); |
| 5486 | total += x; |
| 5487 | nodes[node] += x; |
| 5488 | } |
| 5489 | |
| 5490 | } else |
| 5491 | #endif |
| 5492 | if (flags & SO_PARTIAL) { |
| 5493 | struct kmem_cache_node *n; |
| 5494 | |
| 5495 | for_each_kmem_cache_node(s, node, n) { |
| 5496 | if (flags & SO_TOTAL) |
| 5497 | x = count_partial(n, count_total); |
| 5498 | else if (flags & SO_OBJECTS) |
| 5499 | x = count_partial(n, count_inuse); |
| 5500 | else |
| 5501 | x = n->nr_partial; |
| 5502 | total += x; |
| 5503 | nodes[node] += x; |
| 5504 | } |
| 5505 | } |
| 5506 | |
| 5507 | len += sysfs_emit_at(buf, len, "%lu", total); |
| 5508 | #ifdef CONFIG_NUMA |
| 5509 | for (node = 0; node < nr_node_ids; node++) { |
| 5510 | if (nodes[node]) |
| 5511 | len += sysfs_emit_at(buf, len, " N%d=%lu", |
| 5512 | node, nodes[node]); |
| 5513 | } |
| 5514 | #endif |
| 5515 | len += sysfs_emit_at(buf, len, "\n"); |
| 5516 | kfree(nodes); |
| 5517 | |
| 5518 | return len; |
| 5519 | } |
| 5520 | |
| 5521 | #define to_slab_attr(n) container_of(n, struct slab_attribute, attr) |
| 5522 | #define to_slab(n) container_of(n, struct kmem_cache, kobj) |
| 5523 | |
| 5524 | struct slab_attribute { |
| 5525 | struct attribute attr; |
| 5526 | ssize_t (*show)(struct kmem_cache *s, char *buf); |
| 5527 | ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count); |
| 5528 | }; |
| 5529 | |
| 5530 | #define SLAB_ATTR_RO(_name) \ |
| 5531 | static struct slab_attribute _name##_attr = __ATTR_RO_MODE(_name, 0400) |
| 5532 | |
| 5533 | #define SLAB_ATTR(_name) \ |
| 5534 | static struct slab_attribute _name##_attr = __ATTR_RW_MODE(_name, 0600) |
| 5535 | |
| 5536 | static ssize_t slab_size_show(struct kmem_cache *s, char *buf) |
| 5537 | { |
| 5538 | return sysfs_emit(buf, "%u\n", s->size); |
| 5539 | } |
| 5540 | SLAB_ATTR_RO(slab_size); |
| 5541 | |
| 5542 | static ssize_t align_show(struct kmem_cache *s, char *buf) |
| 5543 | { |
| 5544 | return sysfs_emit(buf, "%u\n", s->align); |
| 5545 | } |
| 5546 | SLAB_ATTR_RO(align); |
| 5547 | |
| 5548 | static ssize_t object_size_show(struct kmem_cache *s, char *buf) |
| 5549 | { |
| 5550 | return sysfs_emit(buf, "%u\n", s->object_size); |
| 5551 | } |
| 5552 | SLAB_ATTR_RO(object_size); |
| 5553 | |
| 5554 | static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf) |
| 5555 | { |
| 5556 | return sysfs_emit(buf, "%u\n", oo_objects(s->oo)); |
| 5557 | } |
| 5558 | SLAB_ATTR_RO(objs_per_slab); |
| 5559 | |
| 5560 | static ssize_t order_show(struct kmem_cache *s, char *buf) |
| 5561 | { |
| 5562 | return sysfs_emit(buf, "%u\n", oo_order(s->oo)); |
| 5563 | } |
| 5564 | SLAB_ATTR_RO(order); |
| 5565 | |
| 5566 | static ssize_t min_partial_show(struct kmem_cache *s, char *buf) |
| 5567 | { |
| 5568 | return sysfs_emit(buf, "%lu\n", s->min_partial); |
| 5569 | } |
| 5570 | |
| 5571 | static ssize_t min_partial_store(struct kmem_cache *s, const char *buf, |
| 5572 | size_t length) |
| 5573 | { |
| 5574 | unsigned long min; |
| 5575 | int err; |
| 5576 | |
| 5577 | err = kstrtoul(buf, 10, &min); |
| 5578 | if (err) |
| 5579 | return err; |
| 5580 | |
| 5581 | s->min_partial = min; |
| 5582 | return length; |
| 5583 | } |
| 5584 | SLAB_ATTR(min_partial); |
| 5585 | |
| 5586 | static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf) |
| 5587 | { |
| 5588 | unsigned int nr_partial = 0; |
| 5589 | #ifdef CONFIG_SLUB_CPU_PARTIAL |
| 5590 | nr_partial = s->cpu_partial; |
| 5591 | #endif |
| 5592 | |
| 5593 | return sysfs_emit(buf, "%u\n", nr_partial); |
| 5594 | } |
| 5595 | |
| 5596 | static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf, |
| 5597 | size_t length) |
| 5598 | { |
| 5599 | unsigned int objects; |
| 5600 | int err; |
| 5601 | |
| 5602 | err = kstrtouint(buf, 10, &objects); |
| 5603 | if (err) |
| 5604 | return err; |
| 5605 | if (objects && !kmem_cache_has_cpu_partial(s)) |
| 5606 | return -EINVAL; |
| 5607 | |
| 5608 | slub_set_cpu_partial(s, objects); |
| 5609 | flush_all(s); |
| 5610 | return length; |
| 5611 | } |
| 5612 | SLAB_ATTR(cpu_partial); |
| 5613 | |
| 5614 | static ssize_t ctor_show(struct kmem_cache *s, char *buf) |
| 5615 | { |
| 5616 | if (!s->ctor) |
| 5617 | return 0; |
| 5618 | return sysfs_emit(buf, "%pS\n", s->ctor); |
| 5619 | } |
| 5620 | SLAB_ATTR_RO(ctor); |
| 5621 | |
| 5622 | static ssize_t aliases_show(struct kmem_cache *s, char *buf) |
| 5623 | { |
| 5624 | return sysfs_emit(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1); |
| 5625 | } |
| 5626 | SLAB_ATTR_RO(aliases); |
| 5627 | |
| 5628 | static ssize_t partial_show(struct kmem_cache *s, char *buf) |
| 5629 | { |
| 5630 | return show_slab_objects(s, buf, SO_PARTIAL); |
| 5631 | } |
| 5632 | SLAB_ATTR_RO(partial); |
| 5633 | |
| 5634 | static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf) |
| 5635 | { |
| 5636 | return show_slab_objects(s, buf, SO_CPU); |
| 5637 | } |
| 5638 | SLAB_ATTR_RO(cpu_slabs); |
| 5639 | |
| 5640 | static ssize_t objects_partial_show(struct kmem_cache *s, char *buf) |
| 5641 | { |
| 5642 | return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS); |
| 5643 | } |
| 5644 | SLAB_ATTR_RO(objects_partial); |
| 5645 | |
| 5646 | static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf) |
| 5647 | { |
| 5648 | int objects = 0; |
| 5649 | int slabs = 0; |
| 5650 | int cpu __maybe_unused; |
| 5651 | int len = 0; |
| 5652 | |
| 5653 | #ifdef CONFIG_SLUB_CPU_PARTIAL |
| 5654 | for_each_online_cpu(cpu) { |
| 5655 | struct slab *slab; |
| 5656 | |
| 5657 | slab = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu)); |
| 5658 | |
| 5659 | if (slab) |
| 5660 | slabs += slab->slabs; |
| 5661 | } |
| 5662 | #endif |
| 5663 | |
| 5664 | /* Approximate half-full slabs, see slub_set_cpu_partial() */ |
| 5665 | objects = (slabs * oo_objects(s->oo)) / 2; |
| 5666 | len += sysfs_emit_at(buf, len, "%d(%d)", objects, slabs); |
| 5667 | |
| 5668 | #ifdef CONFIG_SLUB_CPU_PARTIAL |
| 5669 | for_each_online_cpu(cpu) { |
| 5670 | struct slab *slab; |
| 5671 | |
| 5672 | slab = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu)); |
| 5673 | if (slab) { |
| 5674 | slabs = READ_ONCE(slab->slabs); |
| 5675 | objects = (slabs * oo_objects(s->oo)) / 2; |
| 5676 | len += sysfs_emit_at(buf, len, " C%d=%d(%d)", |
| 5677 | cpu, objects, slabs); |
| 5678 | } |
| 5679 | } |
| 5680 | #endif |
| 5681 | len += sysfs_emit_at(buf, len, "\n"); |
| 5682 | |
| 5683 | return len; |
| 5684 | } |
| 5685 | SLAB_ATTR_RO(slabs_cpu_partial); |
| 5686 | |
| 5687 | static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf) |
| 5688 | { |
| 5689 | return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT)); |
| 5690 | } |
| 5691 | SLAB_ATTR_RO(reclaim_account); |
| 5692 | |
| 5693 | static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf) |
| 5694 | { |
| 5695 | return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN)); |
| 5696 | } |
| 5697 | SLAB_ATTR_RO(hwcache_align); |
| 5698 | |
| 5699 | #ifdef CONFIG_ZONE_DMA |
| 5700 | static ssize_t cache_dma_show(struct kmem_cache *s, char *buf) |
| 5701 | { |
| 5702 | return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA)); |
| 5703 | } |
| 5704 | SLAB_ATTR_RO(cache_dma); |
| 5705 | #endif |
| 5706 | |
| 5707 | #ifdef CONFIG_HARDENED_USERCOPY |
| 5708 | static ssize_t usersize_show(struct kmem_cache *s, char *buf) |
| 5709 | { |
| 5710 | return sysfs_emit(buf, "%u\n", s->usersize); |
| 5711 | } |
| 5712 | SLAB_ATTR_RO(usersize); |
| 5713 | #endif |
| 5714 | |
| 5715 | static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf) |
| 5716 | { |
| 5717 | return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU)); |
| 5718 | } |
| 5719 | SLAB_ATTR_RO(destroy_by_rcu); |
| 5720 | |
| 5721 | #ifdef CONFIG_SLUB_DEBUG |
| 5722 | static ssize_t slabs_show(struct kmem_cache *s, char *buf) |
| 5723 | { |
| 5724 | return show_slab_objects(s, buf, SO_ALL); |
| 5725 | } |
| 5726 | SLAB_ATTR_RO(slabs); |
| 5727 | |
| 5728 | static ssize_t total_objects_show(struct kmem_cache *s, char *buf) |
| 5729 | { |
| 5730 | return show_slab_objects(s, buf, SO_ALL|SO_TOTAL); |
| 5731 | } |
| 5732 | SLAB_ATTR_RO(total_objects); |
| 5733 | |
| 5734 | static ssize_t objects_show(struct kmem_cache *s, char *buf) |
| 5735 | { |
| 5736 | return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS); |
| 5737 | } |
| 5738 | SLAB_ATTR_RO(objects); |
| 5739 | |
| 5740 | static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf) |
| 5741 | { |
| 5742 | return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS)); |
| 5743 | } |
| 5744 | SLAB_ATTR_RO(sanity_checks); |
| 5745 | |
| 5746 | static ssize_t trace_show(struct kmem_cache *s, char *buf) |
| 5747 | { |
| 5748 | return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TRACE)); |
| 5749 | } |
| 5750 | SLAB_ATTR_RO(trace); |
| 5751 | |
| 5752 | static ssize_t red_zone_show(struct kmem_cache *s, char *buf) |
| 5753 | { |
| 5754 | return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE)); |
| 5755 | } |
| 5756 | |
| 5757 | SLAB_ATTR_RO(red_zone); |
| 5758 | |
| 5759 | static ssize_t poison_show(struct kmem_cache *s, char *buf) |
| 5760 | { |
| 5761 | return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_POISON)); |
| 5762 | } |
| 5763 | |
| 5764 | SLAB_ATTR_RO(poison); |
| 5765 | |
| 5766 | static ssize_t store_user_show(struct kmem_cache *s, char *buf) |
| 5767 | { |
| 5768 | return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_STORE_USER)); |
| 5769 | } |
| 5770 | |
| 5771 | SLAB_ATTR_RO(store_user); |
| 5772 | |
| 5773 | static ssize_t validate_show(struct kmem_cache *s, char *buf) |
| 5774 | { |
| 5775 | return 0; |
| 5776 | } |
| 5777 | |
| 5778 | static ssize_t validate_store(struct kmem_cache *s, |
| 5779 | const char *buf, size_t length) |
| 5780 | { |
| 5781 | int ret = -EINVAL; |
| 5782 | |
| 5783 | if (buf[0] == '1' && kmem_cache_debug(s)) { |
| 5784 | ret = validate_slab_cache(s); |
| 5785 | if (ret >= 0) |
| 5786 | ret = length; |
| 5787 | } |
| 5788 | return ret; |
| 5789 | } |
| 5790 | SLAB_ATTR(validate); |
| 5791 | |
| 5792 | #endif /* CONFIG_SLUB_DEBUG */ |
| 5793 | |
| 5794 | #ifdef CONFIG_FAILSLAB |
| 5795 | static ssize_t failslab_show(struct kmem_cache *s, char *buf) |
| 5796 | { |
| 5797 | return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB)); |
| 5798 | } |
| 5799 | |
| 5800 | static ssize_t failslab_store(struct kmem_cache *s, const char *buf, |
| 5801 | size_t length) |
| 5802 | { |
| 5803 | if (s->refcount > 1) |
| 5804 | return -EINVAL; |
| 5805 | |
| 5806 | if (buf[0] == '1') |
| 5807 | WRITE_ONCE(s->flags, s->flags | SLAB_FAILSLAB); |
| 5808 | else |
| 5809 | WRITE_ONCE(s->flags, s->flags & ~SLAB_FAILSLAB); |
| 5810 | |
| 5811 | return length; |
| 5812 | } |
| 5813 | SLAB_ATTR(failslab); |
| 5814 | #endif |
| 5815 | |
| 5816 | static ssize_t shrink_show(struct kmem_cache *s, char *buf) |
| 5817 | { |
| 5818 | return 0; |
| 5819 | } |
| 5820 | |
| 5821 | static ssize_t shrink_store(struct kmem_cache *s, |
| 5822 | const char *buf, size_t length) |
| 5823 | { |
| 5824 | if (buf[0] == '1') |
| 5825 | kmem_cache_shrink(s); |
| 5826 | else |
| 5827 | return -EINVAL; |
| 5828 | return length; |
| 5829 | } |
| 5830 | SLAB_ATTR(shrink); |
| 5831 | |
| 5832 | #ifdef CONFIG_NUMA |
| 5833 | static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf) |
| 5834 | { |
| 5835 | return sysfs_emit(buf, "%u\n", s->remote_node_defrag_ratio / 10); |
| 5836 | } |
| 5837 | |
| 5838 | static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s, |
| 5839 | const char *buf, size_t length) |
| 5840 | { |
| 5841 | unsigned int ratio; |
| 5842 | int err; |
| 5843 | |
| 5844 | err = kstrtouint(buf, 10, &ratio); |
| 5845 | if (err) |
| 5846 | return err; |
| 5847 | if (ratio > 100) |
| 5848 | return -ERANGE; |
| 5849 | |
| 5850 | s->remote_node_defrag_ratio = ratio * 10; |
| 5851 | |
| 5852 | return length; |
| 5853 | } |
| 5854 | SLAB_ATTR(remote_node_defrag_ratio); |
| 5855 | #endif |
| 5856 | |
| 5857 | #ifdef CONFIG_SLUB_STATS |
| 5858 | static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si) |
| 5859 | { |
| 5860 | unsigned long sum = 0; |
| 5861 | int cpu; |
| 5862 | int len = 0; |
| 5863 | int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL); |
| 5864 | |
| 5865 | if (!data) |
| 5866 | return -ENOMEM; |
| 5867 | |
| 5868 | for_each_online_cpu(cpu) { |
| 5869 | unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si]; |
| 5870 | |
| 5871 | data[cpu] = x; |
| 5872 | sum += x; |
| 5873 | } |
| 5874 | |
| 5875 | len += sysfs_emit_at(buf, len, "%lu", sum); |
| 5876 | |
| 5877 | #ifdef CONFIG_SMP |
| 5878 | for_each_online_cpu(cpu) { |
| 5879 | if (data[cpu]) |
| 5880 | len += sysfs_emit_at(buf, len, " C%d=%u", |
| 5881 | cpu, data[cpu]); |
| 5882 | } |
| 5883 | #endif |
| 5884 | kfree(data); |
| 5885 | len += sysfs_emit_at(buf, len, "\n"); |
| 5886 | |
| 5887 | return len; |
| 5888 | } |
| 5889 | |
| 5890 | static void clear_stat(struct kmem_cache *s, enum stat_item si) |
| 5891 | { |
| 5892 | int cpu; |
| 5893 | |
| 5894 | for_each_online_cpu(cpu) |
| 5895 | per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0; |
| 5896 | } |
| 5897 | |
| 5898 | #define STAT_ATTR(si, text) \ |
| 5899 | static ssize_t text##_show(struct kmem_cache *s, char *buf) \ |
| 5900 | { \ |
| 5901 | return show_stat(s, buf, si); \ |
| 5902 | } \ |
| 5903 | static ssize_t text##_store(struct kmem_cache *s, \ |
| 5904 | const char *buf, size_t length) \ |
| 5905 | { \ |
| 5906 | if (buf[0] != '0') \ |
| 5907 | return -EINVAL; \ |
| 5908 | clear_stat(s, si); \ |
| 5909 | return length; \ |
| 5910 | } \ |
| 5911 | SLAB_ATTR(text); \ |
| 5912 | |
| 5913 | STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath); |
| 5914 | STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath); |
| 5915 | STAT_ATTR(FREE_FASTPATH, free_fastpath); |
| 5916 | STAT_ATTR(FREE_SLOWPATH, free_slowpath); |
| 5917 | STAT_ATTR(FREE_FROZEN, free_frozen); |
| 5918 | STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial); |
| 5919 | STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial); |
| 5920 | STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial); |
| 5921 | STAT_ATTR(ALLOC_SLAB, alloc_slab); |
| 5922 | STAT_ATTR(ALLOC_REFILL, alloc_refill); |
| 5923 | STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch); |
| 5924 | STAT_ATTR(FREE_SLAB, free_slab); |
| 5925 | STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush); |
| 5926 | STAT_ATTR(DEACTIVATE_FULL, deactivate_full); |
| 5927 | STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty); |
| 5928 | STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head); |
| 5929 | STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail); |
| 5930 | STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees); |
| 5931 | STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass); |
| 5932 | STAT_ATTR(ORDER_FALLBACK, order_fallback); |
| 5933 | STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail); |
| 5934 | STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail); |
| 5935 | STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc); |
| 5936 | STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free); |
| 5937 | STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node); |
| 5938 | STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain); |
| 5939 | #endif /* CONFIG_SLUB_STATS */ |
| 5940 | |
| 5941 | #ifdef CONFIG_KFENCE |
| 5942 | static ssize_t skip_kfence_show(struct kmem_cache *s, char *buf) |
| 5943 | { |
| 5944 | return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_SKIP_KFENCE)); |
| 5945 | } |
| 5946 | |
| 5947 | static ssize_t skip_kfence_store(struct kmem_cache *s, |
| 5948 | const char *buf, size_t length) |
| 5949 | { |
| 5950 | int ret = length; |
| 5951 | |
| 5952 | if (buf[0] == '0') |
| 5953 | s->flags &= ~SLAB_SKIP_KFENCE; |
| 5954 | else if (buf[0] == '1') |
| 5955 | s->flags |= SLAB_SKIP_KFENCE; |
| 5956 | else |
| 5957 | ret = -EINVAL; |
| 5958 | |
| 5959 | return ret; |
| 5960 | } |
| 5961 | SLAB_ATTR(skip_kfence); |
| 5962 | #endif |
| 5963 | |
| 5964 | static struct attribute *slab_attrs[] = { |
| 5965 | &slab_size_attr.attr, |
| 5966 | &object_size_attr.attr, |
| 5967 | &objs_per_slab_attr.attr, |
| 5968 | &order_attr.attr, |
| 5969 | &min_partial_attr.attr, |
| 5970 | &cpu_partial_attr.attr, |
| 5971 | &objects_partial_attr.attr, |
| 5972 | &partial_attr.attr, |
| 5973 | &cpu_slabs_attr.attr, |
| 5974 | &ctor_attr.attr, |
| 5975 | &aliases_attr.attr, |
| 5976 | &align_attr.attr, |
| 5977 | &hwcache_align_attr.attr, |
| 5978 | &reclaim_account_attr.attr, |
| 5979 | &destroy_by_rcu_attr.attr, |
| 5980 | &shrink_attr.attr, |
| 5981 | &slabs_cpu_partial_attr.attr, |
| 5982 | #ifdef CONFIG_SLUB_DEBUG |
| 5983 | &total_objects_attr.attr, |
| 5984 | &objects_attr.attr, |
| 5985 | &slabs_attr.attr, |
| 5986 | &sanity_checks_attr.attr, |
| 5987 | &trace_attr.attr, |
| 5988 | &red_zone_attr.attr, |
| 5989 | &poison_attr.attr, |
| 5990 | &store_user_attr.attr, |
| 5991 | &validate_attr.attr, |
| 5992 | #endif |
| 5993 | #ifdef CONFIG_ZONE_DMA |
| 5994 | &cache_dma_attr.attr, |
| 5995 | #endif |
| 5996 | #ifdef CONFIG_NUMA |
| 5997 | &remote_node_defrag_ratio_attr.attr, |
| 5998 | #endif |
| 5999 | #ifdef CONFIG_SLUB_STATS |
| 6000 | &alloc_fastpath_attr.attr, |
| 6001 | &alloc_slowpath_attr.attr, |
| 6002 | &free_fastpath_attr.attr, |
| 6003 | &free_slowpath_attr.attr, |
| 6004 | &free_frozen_attr.attr, |
| 6005 | &free_add_partial_attr.attr, |
| 6006 | &free_remove_partial_attr.attr, |
| 6007 | &alloc_from_partial_attr.attr, |
| 6008 | &alloc_slab_attr.attr, |
| 6009 | &alloc_refill_attr.attr, |
| 6010 | &alloc_node_mismatch_attr.attr, |
| 6011 | &free_slab_attr.attr, |
| 6012 | &cpuslab_flush_attr.attr, |
| 6013 | &deactivate_full_attr.attr, |
| 6014 | &deactivate_empty_attr.attr, |
| 6015 | &deactivate_to_head_attr.attr, |
| 6016 | &deactivate_to_tail_attr.attr, |
| 6017 | &deactivate_remote_frees_attr.attr, |
| 6018 | &deactivate_bypass_attr.attr, |
| 6019 | &order_fallback_attr.attr, |
| 6020 | &cmpxchg_double_fail_attr.attr, |
| 6021 | &cmpxchg_double_cpu_fail_attr.attr, |
| 6022 | &cpu_partial_alloc_attr.attr, |
| 6023 | &cpu_partial_free_attr.attr, |
| 6024 | &cpu_partial_node_attr.attr, |
| 6025 | &cpu_partial_drain_attr.attr, |
| 6026 | #endif |
| 6027 | #ifdef CONFIG_FAILSLAB |
| 6028 | &failslab_attr.attr, |
| 6029 | #endif |
| 6030 | #ifdef CONFIG_HARDENED_USERCOPY |
| 6031 | &usersize_attr.attr, |
| 6032 | #endif |
| 6033 | #ifdef CONFIG_KFENCE |
| 6034 | &skip_kfence_attr.attr, |
| 6035 | #endif |
| 6036 | |
| 6037 | NULL |
| 6038 | }; |
| 6039 | |
| 6040 | static const struct attribute_group slab_attr_group = { |
| 6041 | .attrs = slab_attrs, |
| 6042 | }; |
| 6043 | |
| 6044 | static ssize_t slab_attr_show(struct kobject *kobj, |
| 6045 | struct attribute *attr, |
| 6046 | char *buf) |
| 6047 | { |
| 6048 | struct slab_attribute *attribute; |
| 6049 | struct kmem_cache *s; |
| 6050 | |
| 6051 | attribute = to_slab_attr(attr); |
| 6052 | s = to_slab(kobj); |
| 6053 | |
| 6054 | if (!attribute->show) |
| 6055 | return -EIO; |
| 6056 | |
| 6057 | return attribute->show(s, buf); |
| 6058 | } |
| 6059 | |
| 6060 | static ssize_t slab_attr_store(struct kobject *kobj, |
| 6061 | struct attribute *attr, |
| 6062 | const char *buf, size_t len) |
| 6063 | { |
| 6064 | struct slab_attribute *attribute; |
| 6065 | struct kmem_cache *s; |
| 6066 | |
| 6067 | attribute = to_slab_attr(attr); |
| 6068 | s = to_slab(kobj); |
| 6069 | |
| 6070 | if (!attribute->store) |
| 6071 | return -EIO; |
| 6072 | |
| 6073 | return attribute->store(s, buf, len); |
| 6074 | } |
| 6075 | |
| 6076 | static void kmem_cache_release(struct kobject *k) |
| 6077 | { |
| 6078 | slab_kmem_cache_release(to_slab(k)); |
| 6079 | } |
| 6080 | |
| 6081 | static const struct sysfs_ops slab_sysfs_ops = { |
| 6082 | .show = slab_attr_show, |
| 6083 | .store = slab_attr_store, |
| 6084 | }; |
| 6085 | |
| 6086 | static const struct kobj_type slab_ktype = { |
| 6087 | .sysfs_ops = &slab_sysfs_ops, |
| 6088 | .release = kmem_cache_release, |
| 6089 | }; |
| 6090 | |
| 6091 | static struct kset *slab_kset; |
| 6092 | |
| 6093 | static inline struct kset *cache_kset(struct kmem_cache *s) |
| 6094 | { |
| 6095 | return slab_kset; |
| 6096 | } |
| 6097 | |
| 6098 | #define ID_STR_LENGTH 32 |
| 6099 | |
| 6100 | /* Create a unique string id for a slab cache: |
| 6101 | * |
| 6102 | * Format :[flags-]size |
| 6103 | */ |
| 6104 | static char *create_unique_id(struct kmem_cache *s) |
| 6105 | { |
| 6106 | char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL); |
| 6107 | char *p = name; |
| 6108 | |
| 6109 | if (!name) |
| 6110 | return ERR_PTR(-ENOMEM); |
| 6111 | |
| 6112 | *p++ = ':'; |
| 6113 | /* |
| 6114 | * First flags affecting slabcache operations. We will only |
| 6115 | * get here for aliasable slabs so we do not need to support |
| 6116 | * too many flags. The flags here must cover all flags that |
| 6117 | * are matched during merging to guarantee that the id is |
| 6118 | * unique. |
| 6119 | */ |
| 6120 | if (s->flags & SLAB_CACHE_DMA) |
| 6121 | *p++ = 'd'; |
| 6122 | if (s->flags & SLAB_CACHE_DMA32) |
| 6123 | *p++ = 'D'; |
| 6124 | if (s->flags & SLAB_RECLAIM_ACCOUNT) |
| 6125 | *p++ = 'a'; |
| 6126 | if (s->flags & SLAB_CONSISTENCY_CHECKS) |
| 6127 | *p++ = 'F'; |
| 6128 | if (s->flags & SLAB_ACCOUNT) |
| 6129 | *p++ = 'A'; |
| 6130 | if (p != name + 1) |
| 6131 | *p++ = '-'; |
| 6132 | p += snprintf(p, ID_STR_LENGTH - (p - name), "%07u", s->size); |
| 6133 | |
| 6134 | if (WARN_ON(p > name + ID_STR_LENGTH - 1)) { |
| 6135 | kfree(name); |
| 6136 | return ERR_PTR(-EINVAL); |
| 6137 | } |
| 6138 | kmsan_unpoison_memory(name, p - name); |
| 6139 | return name; |
| 6140 | } |
| 6141 | |
| 6142 | static int sysfs_slab_add(struct kmem_cache *s) |
| 6143 | { |
| 6144 | int err; |
| 6145 | const char *name; |
| 6146 | struct kset *kset = cache_kset(s); |
| 6147 | int unmergeable = slab_unmergeable(s); |
| 6148 | |
| 6149 | if (!unmergeable && disable_higher_order_debug && |
| 6150 | (slub_debug & DEBUG_METADATA_FLAGS)) |
| 6151 | unmergeable = 1; |
| 6152 | |
| 6153 | if (unmergeable) { |
| 6154 | /* |
| 6155 | * Slabcache can never be merged so we can use the name proper. |
| 6156 | * This is typically the case for debug situations. In that |
| 6157 | * case we can catch duplicate names easily. |
| 6158 | */ |
| 6159 | sysfs_remove_link(&slab_kset->kobj, s->name); |
| 6160 | name = s->name; |
| 6161 | } else { |
| 6162 | /* |
| 6163 | * Create a unique name for the slab as a target |
| 6164 | * for the symlinks. |
| 6165 | */ |
| 6166 | name = create_unique_id(s); |
| 6167 | if (IS_ERR(name)) |
| 6168 | return PTR_ERR(name); |
| 6169 | } |
| 6170 | |
| 6171 | s->kobj.kset = kset; |
| 6172 | err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name); |
| 6173 | if (err) |
| 6174 | goto out; |
| 6175 | |
| 6176 | err = sysfs_create_group(&s->kobj, &slab_attr_group); |
| 6177 | if (err) |
| 6178 | goto out_del_kobj; |
| 6179 | |
| 6180 | if (!unmergeable) { |
| 6181 | /* Setup first alias */ |
| 6182 | sysfs_slab_alias(s, s->name); |
| 6183 | } |
| 6184 | out: |
| 6185 | if (!unmergeable) |
| 6186 | kfree(name); |
| 6187 | return err; |
| 6188 | out_del_kobj: |
| 6189 | kobject_del(&s->kobj); |
| 6190 | goto out; |
| 6191 | } |
| 6192 | |
| 6193 | void sysfs_slab_unlink(struct kmem_cache *s) |
| 6194 | { |
| 6195 | if (slab_state >= FULL) |
| 6196 | kobject_del(&s->kobj); |
| 6197 | } |
| 6198 | |
| 6199 | void sysfs_slab_release(struct kmem_cache *s) |
| 6200 | { |
| 6201 | if (slab_state >= FULL) |
| 6202 | kobject_put(&s->kobj); |
| 6203 | } |
| 6204 | |
| 6205 | /* |
| 6206 | * Need to buffer aliases during bootup until sysfs becomes |
| 6207 | * available lest we lose that information. |
| 6208 | */ |
| 6209 | struct saved_alias { |
| 6210 | struct kmem_cache *s; |
| 6211 | const char *name; |
| 6212 | struct saved_alias *next; |
| 6213 | }; |
| 6214 | |
| 6215 | static struct saved_alias *alias_list; |
| 6216 | |
| 6217 | static int sysfs_slab_alias(struct kmem_cache *s, const char *name) |
| 6218 | { |
| 6219 | struct saved_alias *al; |
| 6220 | |
| 6221 | if (slab_state == FULL) { |
| 6222 | /* |
| 6223 | * If we have a leftover link then remove it. |
| 6224 | */ |
| 6225 | sysfs_remove_link(&slab_kset->kobj, name); |
| 6226 | return sysfs_create_link(&slab_kset->kobj, &s->kobj, name); |
| 6227 | } |
| 6228 | |
| 6229 | al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL); |
| 6230 | if (!al) |
| 6231 | return -ENOMEM; |
| 6232 | |
| 6233 | al->s = s; |
| 6234 | al->name = name; |
| 6235 | al->next = alias_list; |
| 6236 | alias_list = al; |
| 6237 | kmsan_unpoison_memory(al, sizeof(*al)); |
| 6238 | return 0; |
| 6239 | } |
| 6240 | |
| 6241 | static int __init slab_sysfs_init(void) |
| 6242 | { |
| 6243 | struct kmem_cache *s; |
| 6244 | int err; |
| 6245 | |
| 6246 | mutex_lock(&slab_mutex); |
| 6247 | |
| 6248 | slab_kset = kset_create_and_add("slab", NULL, kernel_kobj); |
| 6249 | if (!slab_kset) { |
| 6250 | mutex_unlock(&slab_mutex); |
| 6251 | pr_err("Cannot register slab subsystem.\n"); |
| 6252 | return -ENOMEM; |
| 6253 | } |
| 6254 | |
| 6255 | slab_state = FULL; |
| 6256 | |
| 6257 | list_for_each_entry(s, &slab_caches, list) { |
| 6258 | err = sysfs_slab_add(s); |
| 6259 | if (err) |
| 6260 | pr_err("SLUB: Unable to add boot slab %s to sysfs\n", |
| 6261 | s->name); |
| 6262 | } |
| 6263 | |
| 6264 | while (alias_list) { |
| 6265 | struct saved_alias *al = alias_list; |
| 6266 | |
| 6267 | alias_list = alias_list->next; |
| 6268 | err = sysfs_slab_alias(al->s, al->name); |
| 6269 | if (err) |
| 6270 | pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n", |
| 6271 | al->name); |
| 6272 | kfree(al); |
| 6273 | } |
| 6274 | |
| 6275 | mutex_unlock(&slab_mutex); |
| 6276 | return 0; |
| 6277 | } |
| 6278 | late_initcall(slab_sysfs_init); |
| 6279 | #endif /* SLAB_SUPPORTS_SYSFS */ |
| 6280 | |
| 6281 | #if defined(CONFIG_SLUB_DEBUG) && defined(CONFIG_DEBUG_FS) |
| 6282 | static int slab_debugfs_show(struct seq_file *seq, void *v) |
| 6283 | { |
| 6284 | struct loc_track *t = seq->private; |
| 6285 | struct location *l; |
| 6286 | unsigned long idx; |
| 6287 | |
| 6288 | idx = (unsigned long) t->idx; |
| 6289 | if (idx < t->count) { |
| 6290 | l = &t->loc[idx]; |
| 6291 | |
| 6292 | seq_printf(seq, "%7ld ", l->count); |
| 6293 | |
| 6294 | if (l->addr) |
| 6295 | seq_printf(seq, "%pS", (void *)l->addr); |
| 6296 | else |
| 6297 | seq_puts(seq, "<not-available>"); |
| 6298 | |
| 6299 | if (l->waste) |
| 6300 | seq_printf(seq, " waste=%lu/%lu", |
| 6301 | l->count * l->waste, l->waste); |
| 6302 | |
| 6303 | if (l->sum_time != l->min_time) { |
| 6304 | seq_printf(seq, " age=%ld/%llu/%ld", |
| 6305 | l->min_time, div_u64(l->sum_time, l->count), |
| 6306 | l->max_time); |
| 6307 | } else |
| 6308 | seq_printf(seq, " age=%ld", l->min_time); |
| 6309 | |
| 6310 | if (l->min_pid != l->max_pid) |
| 6311 | seq_printf(seq, " pid=%ld-%ld", l->min_pid, l->max_pid); |
| 6312 | else |
| 6313 | seq_printf(seq, " pid=%ld", |
| 6314 | l->min_pid); |
| 6315 | |
| 6316 | if (num_online_cpus() > 1 && !cpumask_empty(to_cpumask(l->cpus))) |
| 6317 | seq_printf(seq, " cpus=%*pbl", |
| 6318 | cpumask_pr_args(to_cpumask(l->cpus))); |
| 6319 | |
| 6320 | if (nr_online_nodes > 1 && !nodes_empty(l->nodes)) |
| 6321 | seq_printf(seq, " nodes=%*pbl", |
| 6322 | nodemask_pr_args(&l->nodes)); |
| 6323 | |
| 6324 | #ifdef CONFIG_STACKDEPOT |
| 6325 | { |
| 6326 | depot_stack_handle_t handle; |
| 6327 | unsigned long *entries; |
| 6328 | unsigned int nr_entries, j; |
| 6329 | |
| 6330 | handle = READ_ONCE(l->handle); |
| 6331 | if (handle) { |
| 6332 | nr_entries = stack_depot_fetch(handle, &entries); |
| 6333 | seq_puts(seq, "\n"); |
| 6334 | for (j = 0; j < nr_entries; j++) |
| 6335 | seq_printf(seq, " %pS\n", (void *)entries[j]); |
| 6336 | } |
| 6337 | } |
| 6338 | #endif |
| 6339 | seq_puts(seq, "\n"); |
| 6340 | } |
| 6341 | |
| 6342 | if (!idx && !t->count) |
| 6343 | seq_puts(seq, "No data\n"); |
| 6344 | |
| 6345 | return 0; |
| 6346 | } |
| 6347 | |
| 6348 | static void slab_debugfs_stop(struct seq_file *seq, void *v) |
| 6349 | { |
| 6350 | } |
| 6351 | |
| 6352 | static void *slab_debugfs_next(struct seq_file *seq, void *v, loff_t *ppos) |
| 6353 | { |
| 6354 | struct loc_track *t = seq->private; |
| 6355 | |
| 6356 | t->idx = ++(*ppos); |
| 6357 | if (*ppos <= t->count) |
| 6358 | return ppos; |
| 6359 | |
| 6360 | return NULL; |
| 6361 | } |
| 6362 | |
| 6363 | static int cmp_loc_by_count(const void *a, const void *b, const void *data) |
| 6364 | { |
| 6365 | struct location *loc1 = (struct location *)a; |
| 6366 | struct location *loc2 = (struct location *)b; |
| 6367 | |
| 6368 | if (loc1->count > loc2->count) |
| 6369 | return -1; |
| 6370 | else |
| 6371 | return 1; |
| 6372 | } |
| 6373 | |
| 6374 | static void *slab_debugfs_start(struct seq_file *seq, loff_t *ppos) |
| 6375 | { |
| 6376 | struct loc_track *t = seq->private; |
| 6377 | |
| 6378 | t->idx = *ppos; |
| 6379 | return ppos; |
| 6380 | } |
| 6381 | |
| 6382 | static const struct seq_operations slab_debugfs_sops = { |
| 6383 | .start = slab_debugfs_start, |
| 6384 | .next = slab_debugfs_next, |
| 6385 | .stop = slab_debugfs_stop, |
| 6386 | .show = slab_debugfs_show, |
| 6387 | }; |
| 6388 | |
| 6389 | static int slab_debug_trace_open(struct inode *inode, struct file *filep) |
| 6390 | { |
| 6391 | |
| 6392 | struct kmem_cache_node *n; |
| 6393 | enum track_item alloc; |
| 6394 | int node; |
| 6395 | struct loc_track *t = __seq_open_private(filep, &slab_debugfs_sops, |
| 6396 | sizeof(struct loc_track)); |
| 6397 | struct kmem_cache *s = file_inode(filep)->i_private; |
| 6398 | unsigned long *obj_map; |
| 6399 | |
| 6400 | if (!t) |
| 6401 | return -ENOMEM; |
| 6402 | |
| 6403 | obj_map = bitmap_alloc(oo_objects(s->oo), GFP_KERNEL); |
| 6404 | if (!obj_map) { |
| 6405 | seq_release_private(inode, filep); |
| 6406 | return -ENOMEM; |
| 6407 | } |
| 6408 | |
| 6409 | if (strcmp(filep->f_path.dentry->d_name.name, "alloc_traces") == 0) |
| 6410 | alloc = TRACK_ALLOC; |
| 6411 | else |
| 6412 | alloc = TRACK_FREE; |
| 6413 | |
| 6414 | if (!alloc_loc_track(t, PAGE_SIZE / sizeof(struct location), GFP_KERNEL)) { |
| 6415 | bitmap_free(obj_map); |
| 6416 | seq_release_private(inode, filep); |
| 6417 | return -ENOMEM; |
| 6418 | } |
| 6419 | |
| 6420 | for_each_kmem_cache_node(s, node, n) { |
| 6421 | unsigned long flags; |
| 6422 | struct slab *slab; |
| 6423 | |
| 6424 | if (!node_nr_slabs(n)) |
| 6425 | continue; |
| 6426 | |
| 6427 | spin_lock_irqsave(&n->list_lock, flags); |
| 6428 | list_for_each_entry(slab, &n->partial, slab_list) |
| 6429 | process_slab(t, s, slab, alloc, obj_map); |
| 6430 | list_for_each_entry(slab, &n->full, slab_list) |
| 6431 | process_slab(t, s, slab, alloc, obj_map); |
| 6432 | spin_unlock_irqrestore(&n->list_lock, flags); |
| 6433 | } |
| 6434 | |
| 6435 | /* Sort locations by count */ |
| 6436 | sort_r(t->loc, t->count, sizeof(struct location), |
| 6437 | cmp_loc_by_count, NULL, NULL); |
| 6438 | |
| 6439 | bitmap_free(obj_map); |
| 6440 | return 0; |
| 6441 | } |
| 6442 | |
| 6443 | static int slab_debug_trace_release(struct inode *inode, struct file *file) |
| 6444 | { |
| 6445 | struct seq_file *seq = file->private_data; |
| 6446 | struct loc_track *t = seq->private; |
| 6447 | |
| 6448 | free_loc_track(t); |
| 6449 | return seq_release_private(inode, file); |
| 6450 | } |
| 6451 | |
| 6452 | static const struct file_operations slab_debugfs_fops = { |
| 6453 | .open = slab_debug_trace_open, |
| 6454 | .read = seq_read, |
| 6455 | .llseek = seq_lseek, |
| 6456 | .release = slab_debug_trace_release, |
| 6457 | }; |
| 6458 | |
| 6459 | static void debugfs_slab_add(struct kmem_cache *s) |
| 6460 | { |
| 6461 | struct dentry *slab_cache_dir; |
| 6462 | |
| 6463 | if (unlikely(!slab_debugfs_root)) |
| 6464 | return; |
| 6465 | |
| 6466 | slab_cache_dir = debugfs_create_dir(s->name, slab_debugfs_root); |
| 6467 | |
| 6468 | debugfs_create_file("alloc_traces", 0400, |
| 6469 | slab_cache_dir, s, &slab_debugfs_fops); |
| 6470 | |
| 6471 | debugfs_create_file("free_traces", 0400, |
| 6472 | slab_cache_dir, s, &slab_debugfs_fops); |
| 6473 | } |
| 6474 | |
| 6475 | void debugfs_slab_release(struct kmem_cache *s) |
| 6476 | { |
| 6477 | debugfs_lookup_and_remove(s->name, slab_debugfs_root); |
| 6478 | } |
| 6479 | |
| 6480 | static int __init slab_debugfs_init(void) |
| 6481 | { |
| 6482 | struct kmem_cache *s; |
| 6483 | |
| 6484 | slab_debugfs_root = debugfs_create_dir("slab", NULL); |
| 6485 | |
| 6486 | list_for_each_entry(s, &slab_caches, list) |
| 6487 | if (s->flags & SLAB_STORE_USER) |
| 6488 | debugfs_slab_add(s); |
| 6489 | |
| 6490 | return 0; |
| 6491 | |
| 6492 | } |
| 6493 | __initcall(slab_debugfs_init); |
| 6494 | #endif |
| 6495 | /* |
| 6496 | * The /proc/slabinfo ABI |
| 6497 | */ |
| 6498 | #ifdef CONFIG_SLUB_DEBUG |
| 6499 | void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo) |
| 6500 | { |
| 6501 | unsigned long nr_slabs = 0; |
| 6502 | unsigned long nr_objs = 0; |
| 6503 | unsigned long nr_free = 0; |
| 6504 | int node; |
| 6505 | struct kmem_cache_node *n; |
| 6506 | |
| 6507 | for_each_kmem_cache_node(s, node, n) { |
| 6508 | nr_slabs += node_nr_slabs(n); |
| 6509 | nr_objs += node_nr_objs(n); |
| 6510 | nr_free += count_partial(n, count_free); |
| 6511 | } |
| 6512 | |
| 6513 | sinfo->active_objs = nr_objs - nr_free; |
| 6514 | sinfo->num_objs = nr_objs; |
| 6515 | sinfo->active_slabs = nr_slabs; |
| 6516 | sinfo->num_slabs = nr_slabs; |
| 6517 | sinfo->objects_per_slab = oo_objects(s->oo); |
| 6518 | sinfo->cache_order = oo_order(s->oo); |
| 6519 | } |
| 6520 | |
| 6521 | void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s) |
| 6522 | { |
| 6523 | } |
| 6524 | |
| 6525 | ssize_t slabinfo_write(struct file *file, const char __user *buffer, |
| 6526 | size_t count, loff_t *ppos) |
| 6527 | { |
| 6528 | return -EIO; |
| 6529 | } |
| 6530 | #endif /* CONFIG_SLUB_DEBUG */ |