1// SPDX-License-Identifier: GPL-2.0
   2/*
   3 * SLUB: A slab allocator that limits cache line use instead of queuing
   4 * objects in per cpu and per node lists.
   5 *
   6 * The allocator synchronizes using per slab locks or atomic operations
   7 * and only uses a centralized lock to manage a pool of partial slabs.
   8 *
   9 * (C) 2007 SGI, Christoph Lameter
  10 * (C) 2011 Linux Foundation, Christoph Lameter
  11 */
  12
  13#include <linux/mm.h>
  14#include <linux/swap.h> /* mm_account_reclaimed_pages() */
  15#include <linux/module.h>
  16#include <linux/bit_spinlock.h>
  17#include <linux/interrupt.h>
  18#include <linux/swab.h>
  19#include <linux/bitops.h>
  20#include <linux/slab.h>
  21#include "slab.h"
  22#include <linux/proc_fs.h>
  23#include <linux/seq_file.h>
  24#include <linux/kasan.h>
  25#include <linux/kmsan.h>
  26#include <linux/cpu.h>
  27#include <linux/cpuset.h>
  28#include <linux/mempolicy.h>
  29#include <linux/ctype.h>
  30#include <linux/stackdepot.h>
  31#include <linux/debugobjects.h>
  32#include <linux/kallsyms.h>
  33#include <linux/kfence.h>
  34#include <linux/memory.h>
  35#include <linux/math64.h>
  36#include <linux/fault-inject.h>
  37#include <linux/kmemleak.h>
  38#include <linux/stacktrace.h>
  39#include <linux/prefetch.h>
  40#include <linux/memcontrol.h>
  41#include <linux/random.h>
  42#include <kunit/test.h>
  43#include <kunit/test-bug.h>
  44#include <linux/sort.h>
  45
  46#include <linux/debugfs.h>
  47#include <trace/events/kmem.h>
  48
  49#include "internal.h"
  50
  51/*
  52 * Lock order:
  53 *   1. slab_mutex (Global Mutex)
  54 *   2. node->list_lock (Spinlock)
  55 *   3. kmem_cache->cpu_slab->lock (Local lock)
  56 *   4. slab_lock(slab) (Only on some arches)
  57 *   5. object_map_lock (Only for debugging)
  58 *
  59 *   slab_mutex
  60 *
  61 *   The role of the slab_mutex is to protect the list of all the slabs
  62 *   and to synchronize major metadata changes to slab cache structures.
  63 *   Also synchronizes memory hotplug callbacks.
  64 *
  65 *   slab_lock
  66 *
  67 *   The slab_lock is a wrapper around the page lock, thus it is a bit
  68 *   spinlock.
  69 *
  70 *   The slab_lock is only used on arches that do not have the ability
  71 *   to do a cmpxchg_double. It only protects:
  72 *
  73 *	A. slab->freelist	-> List of free objects in a slab
  74 *	B. slab->inuse		-> Number of objects in use
  75 *	C. slab->objects	-> Number of objects in slab
  76 *	D. slab->frozen		-> frozen state
  77 *
  78 *   Frozen slabs
  79 *
  80 *   If a slab is frozen then it is exempt from list management. It is
  81 *   the cpu slab which is actively allocated from by the processor that
  82 *   froze it and it is not on any list. The processor that froze the
  83 *   slab is the one who can perform list operations on the slab. Other
  84 *   processors may put objects onto the freelist but the processor that
  85 *   froze the slab is the only one that can retrieve the objects from the
  86 *   slab's freelist.
  87 *
  88 *   CPU partial slabs
  89 *
  90 *   The partially empty slabs cached on the CPU partial list are used
  91 *   for performance reasons, which speeds up the allocation process.
  92 *   These slabs are not frozen, but are also exempt from list management,
  93 *   by clearing the PG_workingset flag when moving out of the node
  94 *   partial list. Please see __slab_free() for more details.
  95 *
  96 *   To sum up, the current scheme is:
  97 *   - node partial slab: PG_Workingset && !frozen
  98 *   - cpu partial slab: !PG_Workingset && !frozen
  99 *   - cpu slab: !PG_Workingset && frozen
 100 *   - full slab: !PG_Workingset && !frozen
 101 *
 102 *   list_lock
 103 *
 104 *   The list_lock protects the partial and full list on each node and
 105 *   the partial slab counter. If taken then no new slabs may be added or
 106 *   removed from the lists nor make the number of partial slabs be modified.
 107 *   (Note that the total number of slabs is an atomic value that may be
 108 *   modified without taking the list lock).
 109 *
 110 *   The list_lock is a centralized lock and thus we avoid taking it as
 111 *   much as possible. As long as SLUB does not have to handle partial
 112 *   slabs, operations can continue without any centralized lock. F.e.
 113 *   allocating a long series of objects that fill up slabs does not require
 114 *   the list lock.
 115 *
 116 *   For debug caches, all allocations are forced to go through a list_lock
 117 *   protected region to serialize against concurrent validation.
 118 *
 119 *   cpu_slab->lock local lock
 120 *
 121 *   This locks protect slowpath manipulation of all kmem_cache_cpu fields
 122 *   except the stat counters. This is a percpu structure manipulated only by
 123 *   the local cpu, so the lock protects against being preempted or interrupted
 124 *   by an irq. Fast path operations rely on lockless operations instead.
 125 *
 126 *   On PREEMPT_RT, the local lock neither disables interrupts nor preemption
 127 *   which means the lockless fastpath cannot be used as it might interfere with
 128 *   an in-progress slow path operations. In this case the local lock is always
 129 *   taken but it still utilizes the freelist for the common operations.
 130 *
 131 *   lockless fastpaths
 132 *
 133 *   The fast path allocation (slab_alloc_node()) and freeing (do_slab_free())
 134 *   are fully lockless when satisfied from the percpu slab (and when
 135 *   cmpxchg_double is possible to use, otherwise slab_lock is taken).
 136 *   They also don't disable preemption or migration or irqs. They rely on
 137 *   the transaction id (tid) field to detect being preempted or moved to
 138 *   another cpu.
 139 *
 140 *   irq, preemption, migration considerations
 141 *
 142 *   Interrupts are disabled as part of list_lock or local_lock operations, or
 143 *   around the slab_lock operation, in order to make the slab allocator safe
 144 *   to use in the context of an irq.
 145 *
 146 *   In addition, preemption (or migration on PREEMPT_RT) is disabled in the
 147 *   allocation slowpath, bulk allocation, and put_cpu_partial(), so that the
 148 *   local cpu doesn't change in the process and e.g. the kmem_cache_cpu pointer
 149 *   doesn't have to be revalidated in each section protected by the local lock.
 150 *
 151 * SLUB assigns one slab for allocation to each processor.
 152 * Allocations only occur from these slabs called cpu slabs.
 153 *
 154 * Slabs with free elements are kept on a partial list and during regular
 155 * operations no list for full slabs is used. If an object in a full slab is
 156 * freed then the slab will show up again on the partial lists.
 157 * We track full slabs for debugging purposes though because otherwise we
 158 * cannot scan all objects.
 159 *
 160 * Slabs are freed when they become empty. Teardown and setup is
 161 * minimal so we rely on the page allocators per cpu caches for
 162 * fast frees and allocs.
 163 *
 164 * slab->frozen		The slab is frozen and exempt from list processing.
 165 * 			This means that the slab is dedicated to a purpose
 166 * 			such as satisfying allocations for a specific
 167 * 			processor. Objects may be freed in the slab while
 168 * 			it is frozen but slab_free will then skip the usual
 169 * 			list operations. It is up to the processor holding
 170 * 			the slab to integrate the slab into the slab lists
 171 * 			when the slab is no longer needed.
 172 *
 173 * 			One use of this flag is to mark slabs that are
 174 * 			used for allocations. Then such a slab becomes a cpu
 175 * 			slab. The cpu slab may be equipped with an additional
 176 * 			freelist that allows lockless access to
 177 * 			free objects in addition to the regular freelist
 178 * 			that requires the slab lock.
 179 *
 180 * SLAB_DEBUG_FLAGS	Slab requires special handling due to debug
 181 * 			options set. This moves	slab handling out of
 182 * 			the fast path and disables lockless freelists.
 183 */
 184
 185/*
 186 * We could simply use migrate_disable()/enable() but as long as it's a
 187 * function call even on !PREEMPT_RT, use inline preempt_disable() there.
 188 */
 189#ifndef CONFIG_PREEMPT_RT
 190#define slub_get_cpu_ptr(var)		get_cpu_ptr(var)
 191#define slub_put_cpu_ptr(var)		put_cpu_ptr(var)
 192#define USE_LOCKLESS_FAST_PATH()	(true)
 193#else
 194#define slub_get_cpu_ptr(var)		\
 195({					\
 196	migrate_disable();		\
 197	this_cpu_ptr(var);		\
 198})
 199#define slub_put_cpu_ptr(var)		\
 200do {					\
 201	(void)(var);			\
 202	migrate_enable();		\
 203} while (0)
 204#define USE_LOCKLESS_FAST_PATH()	(false)
 205#endif
 206
 207#ifndef CONFIG_SLUB_TINY
 208#define __fastpath_inline __always_inline
 209#else
 210#define __fastpath_inline
 211#endif
 212
 213#ifdef CONFIG_SLUB_DEBUG
 214#ifdef CONFIG_SLUB_DEBUG_ON
 215DEFINE_STATIC_KEY_TRUE(slub_debug_enabled);
 216#else
 217DEFINE_STATIC_KEY_FALSE(slub_debug_enabled);
 218#endif
 219#endif		/* CONFIG_SLUB_DEBUG */
 220
 221/* Structure holding parameters for get_partial() call chain */
 222struct partial_context {
 223	gfp_t flags;
 224	unsigned int orig_size;
 225	void *object;
 226};
 227
 228static inline bool kmem_cache_debug(struct kmem_cache *s)
 229{
 230	return kmem_cache_debug_flags(s, SLAB_DEBUG_FLAGS);
 231}
 232
 233static inline bool slub_debug_orig_size(struct kmem_cache *s)
 234{
 235	return (kmem_cache_debug_flags(s, SLAB_STORE_USER) &&
 236			(s->flags & SLAB_KMALLOC));
 237}
 238
 239void *fixup_red_left(struct kmem_cache *s, void *p)
 240{
 241	if (kmem_cache_debug_flags(s, SLAB_RED_ZONE))
 242		p += s->red_left_pad;
 243
 244	return p;
 245}
 246
 247static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
 248{
 249#ifdef CONFIG_SLUB_CPU_PARTIAL
 250	return !kmem_cache_debug(s);
 251#else
 252	return false;
 253#endif
 254}
 255
 256/*
 257 * Issues still to be resolved:
 258 *
 259 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
 260 *
 261 * - Variable sizing of the per node arrays
 262 */
 263
 264/* Enable to log cmpxchg failures */
 265#undef SLUB_DEBUG_CMPXCHG
 266
 267#ifndef CONFIG_SLUB_TINY
 268/*
 269 * Minimum number of partial slabs. These will be left on the partial
 270 * lists even if they are empty. kmem_cache_shrink may reclaim them.
 271 */
 272#define MIN_PARTIAL 5
 273
 274/*
 275 * Maximum number of desirable partial slabs.
 276 * The existence of more partial slabs makes kmem_cache_shrink
 277 * sort the partial list by the number of objects in use.
 278 */
 279#define MAX_PARTIAL 10
 280#else
 281#define MIN_PARTIAL 0
 282#define MAX_PARTIAL 0
 283#endif
 284
 285#define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
 286				SLAB_POISON | SLAB_STORE_USER)
 287
 288/*
 289 * These debug flags cannot use CMPXCHG because there might be consistency
 290 * issues when checking or reading debug information
 291 */
 292#define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
 293				SLAB_TRACE)
 294
 295
 296/*
 297 * Debugging flags that require metadata to be stored in the slab.  These get
 298 * disabled when slub_debug=O is used and a cache's min order increases with
 299 * metadata.
 300 */
 301#define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
 302
 303#define OO_SHIFT	16
 304#define OO_MASK		((1 << OO_SHIFT) - 1)
 305#define MAX_OBJS_PER_PAGE	32767 /* since slab.objects is u15 */
 306
 307/* Internal SLUB flags */
 308/* Poison object */
 309#define __OBJECT_POISON		((slab_flags_t __force)0x80000000U)
 310/* Use cmpxchg_double */
 311
 312#ifdef system_has_freelist_aba
 313#define __CMPXCHG_DOUBLE	((slab_flags_t __force)0x40000000U)
 314#else
 315#define __CMPXCHG_DOUBLE	((slab_flags_t __force)0U)
 316#endif
 317
 318/*
 319 * Tracking user of a slab.
 320 */
 321#define TRACK_ADDRS_COUNT 16
 322struct track {
 323	unsigned long addr;	/* Called from address */
 324#ifdef CONFIG_STACKDEPOT
 325	depot_stack_handle_t handle;
 326#endif
 327	int cpu;		/* Was running on cpu */
 328	int pid;		/* Pid context */
 329	unsigned long when;	/* When did the operation occur */
 330};
 331
 332enum track_item { TRACK_ALLOC, TRACK_FREE };
 333
 334#ifdef SLAB_SUPPORTS_SYSFS
 335static int sysfs_slab_add(struct kmem_cache *);
 336static int sysfs_slab_alias(struct kmem_cache *, const char *);
 337#else
 338static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
 339static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
 340							{ return 0; }
 341#endif
 342
 343#if defined(CONFIG_DEBUG_FS) && defined(CONFIG_SLUB_DEBUG)
 344static void debugfs_slab_add(struct kmem_cache *);
 345#else
 346static inline void debugfs_slab_add(struct kmem_cache *s) { }
 347#endif
 348
 349enum stat_item {
 350	ALLOC_FASTPATH,		/* Allocation from cpu slab */
 351	ALLOC_SLOWPATH,		/* Allocation by getting a new cpu slab */
 352	FREE_FASTPATH,		/* Free to cpu slab */
 353	FREE_SLOWPATH,		/* Freeing not to cpu slab */
 354	FREE_FROZEN,		/* Freeing to frozen slab */
 355	FREE_ADD_PARTIAL,	/* Freeing moves slab to partial list */
 356	FREE_REMOVE_PARTIAL,	/* Freeing removes last object */
 357	ALLOC_FROM_PARTIAL,	/* Cpu slab acquired from node partial list */
 358	ALLOC_SLAB,		/* Cpu slab acquired from page allocator */
 359	ALLOC_REFILL,		/* Refill cpu slab from slab freelist */
 360	ALLOC_NODE_MISMATCH,	/* Switching cpu slab */
 361	FREE_SLAB,		/* Slab freed to the page allocator */
 362	CPUSLAB_FLUSH,		/* Abandoning of the cpu slab */
 363	DEACTIVATE_FULL,	/* Cpu slab was full when deactivated */
 364	DEACTIVATE_EMPTY,	/* Cpu slab was empty when deactivated */
 365	DEACTIVATE_TO_HEAD,	/* Cpu slab was moved to the head of partials */
 366	DEACTIVATE_TO_TAIL,	/* Cpu slab was moved to the tail of partials */
 367	DEACTIVATE_REMOTE_FREES,/* Slab contained remotely freed objects */
 368	DEACTIVATE_BYPASS,	/* Implicit deactivation */
 369	ORDER_FALLBACK,		/* Number of times fallback was necessary */
 370	CMPXCHG_DOUBLE_CPU_FAIL,/* Failures of this_cpu_cmpxchg_double */
 371	CMPXCHG_DOUBLE_FAIL,	/* Failures of slab freelist update */
 372	CPU_PARTIAL_ALLOC,	/* Used cpu partial on alloc */
 373	CPU_PARTIAL_FREE,	/* Refill cpu partial on free */
 374	CPU_PARTIAL_NODE,	/* Refill cpu partial from node partial */
 375	CPU_PARTIAL_DRAIN,	/* Drain cpu partial to node partial */
 376	NR_SLUB_STAT_ITEMS
 377};
 378
 379#ifndef CONFIG_SLUB_TINY
 380/*
 381 * When changing the layout, make sure freelist and tid are still compatible
 382 * with this_cpu_cmpxchg_double() alignment requirements.
 383 */
 384struct kmem_cache_cpu {
 385	union {
 386		struct {
 387			void **freelist;	/* Pointer to next available object */
 388			unsigned long tid;	/* Globally unique transaction id */
 389		};
 390		freelist_aba_t freelist_tid;
 391	};
 392	struct slab *slab;	/* The slab from which we are allocating */
 393#ifdef CONFIG_SLUB_CPU_PARTIAL
 394	struct slab *partial;	/* Partially allocated frozen slabs */
 395#endif
 396	local_lock_t lock;	/* Protects the fields above */
 397#ifdef CONFIG_SLUB_STATS
 398	unsigned int stat[NR_SLUB_STAT_ITEMS];
 399#endif
 400};
 401#endif /* CONFIG_SLUB_TINY */
 402
 403static inline void stat(const struct kmem_cache *s, enum stat_item si)
 404{
 405#ifdef CONFIG_SLUB_STATS
 406	/*
 407	 * The rmw is racy on a preemptible kernel but this is acceptable, so
 408	 * avoid this_cpu_add()'s irq-disable overhead.
 409	 */
 410	raw_cpu_inc(s->cpu_slab->stat[si]);
 411#endif
 412}
 413
 414static inline
 415void stat_add(const struct kmem_cache *s, enum stat_item si, int v)
 416{
 417#ifdef CONFIG_SLUB_STATS
 418	raw_cpu_add(s->cpu_slab->stat[si], v);
 419#endif
 420}
 421
 422/*
 423 * The slab lists for all objects.
 424 */
 425struct kmem_cache_node {
 426	spinlock_t list_lock;
 427	unsigned long nr_partial;
 428	struct list_head partial;
 429#ifdef CONFIG_SLUB_DEBUG
 430	atomic_long_t nr_slabs;
 431	atomic_long_t total_objects;
 432	struct list_head full;
 433#endif
 434};
 435
 436static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
 437{
 438	return s->node[node];
 439}
 440
 441/*
 442 * Iterator over all nodes. The body will be executed for each node that has
 443 * a kmem_cache_node structure allocated (which is true for all online nodes)
 444 */
 445#define for_each_kmem_cache_node(__s, __node, __n) \
 446	for (__node = 0; __node < nr_node_ids; __node++) \
 447		 if ((__n = get_node(__s, __node)))
 448
 449/*
 450 * Tracks for which NUMA nodes we have kmem_cache_nodes allocated.
 451 * Corresponds to node_state[N_NORMAL_MEMORY], but can temporarily
 452 * differ during memory hotplug/hotremove operations.
 453 * Protected by slab_mutex.
 454 */
 455static nodemask_t slab_nodes;
 456
 457#ifndef CONFIG_SLUB_TINY
 458/*
 459 * Workqueue used for flush_cpu_slab().
 460 */
 461static struct workqueue_struct *flushwq;
 462#endif
 463
 464/********************************************************************
 465 * 			Core slab cache functions
 466 *******************************************************************/
 467
 468/*
 469 * freeptr_t represents a SLUB freelist pointer, which might be encoded
 470 * and not dereferenceable if CONFIG_SLAB_FREELIST_HARDENED is enabled.
 471 */
 472typedef struct { unsigned long v; } freeptr_t;
 473
 474/*
 475 * Returns freelist pointer (ptr). With hardening, this is obfuscated
 476 * with an XOR of the address where the pointer is held and a per-cache
 477 * random number.
 478 */
 479static inline freeptr_t freelist_ptr_encode(const struct kmem_cache *s,
 480					    void *ptr, unsigned long ptr_addr)
 481{
 482	unsigned long encoded;
 483
 484#ifdef CONFIG_SLAB_FREELIST_HARDENED
 485	encoded = (unsigned long)ptr ^ s->random ^ swab(ptr_addr);
 486#else
 487	encoded = (unsigned long)ptr;
 488#endif
 489	return (freeptr_t){.v = encoded};
 490}
 491
 492static inline void *freelist_ptr_decode(const struct kmem_cache *s,
 493					freeptr_t ptr, unsigned long ptr_addr)
 494{
 495	void *decoded;
 496
 497#ifdef CONFIG_SLAB_FREELIST_HARDENED
 498	decoded = (void *)(ptr.v ^ s->random ^ swab(ptr_addr));
 499#else
 500	decoded = (void *)ptr.v;
 501#endif
 502	return decoded;
 503}
 504
 505static inline void *get_freepointer(struct kmem_cache *s, void *object)
 506{
 507	unsigned long ptr_addr;
 508	freeptr_t p;
 509
 510	object = kasan_reset_tag(object);
 511	ptr_addr = (unsigned long)object + s->offset;
 512	p = *(freeptr_t *)(ptr_addr);
 513	return freelist_ptr_decode(s, p, ptr_addr);
 514}
 515
 516#ifndef CONFIG_SLUB_TINY
 517static void prefetch_freepointer(const struct kmem_cache *s, void *object)
 518{
 519	prefetchw(object + s->offset);
 520}
 521#endif
 522
 523/*
 524 * When running under KMSAN, get_freepointer_safe() may return an uninitialized
 525 * pointer value in the case the current thread loses the race for the next
 526 * memory chunk in the freelist. In that case this_cpu_cmpxchg_double() in
 527 * slab_alloc_node() will fail, so the uninitialized value won't be used, but
 528 * KMSAN will still check all arguments of cmpxchg because of imperfect
 529 * handling of inline assembly.
 530 * To work around this problem, we apply __no_kmsan_checks to ensure that
 531 * get_freepointer_safe() returns initialized memory.
 532 */
 533__no_kmsan_checks
 534static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
 535{
 536	unsigned long freepointer_addr;
 537	freeptr_t p;
 538
 539	if (!debug_pagealloc_enabled_static())
 540		return get_freepointer(s, object);
 541
 542	object = kasan_reset_tag(object);
 543	freepointer_addr = (unsigned long)object + s->offset;
 544	copy_from_kernel_nofault(&p, (freeptr_t *)freepointer_addr, sizeof(p));
 545	return freelist_ptr_decode(s, p, freepointer_addr);
 546}
 547
 548static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
 549{
 550	unsigned long freeptr_addr = (unsigned long)object + s->offset;
 551
 552#ifdef CONFIG_SLAB_FREELIST_HARDENED
 553	BUG_ON(object == fp); /* naive detection of double free or corruption */
 554#endif
 555
 556	freeptr_addr = (unsigned long)kasan_reset_tag((void *)freeptr_addr);
 557	*(freeptr_t *)freeptr_addr = freelist_ptr_encode(s, fp, freeptr_addr);
 558}
 559
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 560/* Loop over all objects in a slab */
 561#define for_each_object(__p, __s, __addr, __objects) \
 562	for (__p = fixup_red_left(__s, __addr); \
 563		__p < (__addr) + (__objects) * (__s)->size; \
 564		__p += (__s)->size)
 565
 566static inline unsigned int order_objects(unsigned int order, unsigned int size)
 567{
 568	return ((unsigned int)PAGE_SIZE << order) / size;
 569}
 570
 571static inline struct kmem_cache_order_objects oo_make(unsigned int order,
 572		unsigned int size)
 573{
 574	struct kmem_cache_order_objects x = {
 575		(order << OO_SHIFT) + order_objects(order, size)
 576	};
 577
 578	return x;
 579}
 580
 581static inline unsigned int oo_order(struct kmem_cache_order_objects x)
 582{
 583	return x.x >> OO_SHIFT;
 584}
 585
 586static inline unsigned int oo_objects(struct kmem_cache_order_objects x)
 587{
 588	return x.x & OO_MASK;
 589}
 590
 591#ifdef CONFIG_SLUB_CPU_PARTIAL
 592static void slub_set_cpu_partial(struct kmem_cache *s, unsigned int nr_objects)
 593{
 594	unsigned int nr_slabs;
 595
 596	s->cpu_partial = nr_objects;
 597
 598	/*
 599	 * We take the number of objects but actually limit the number of
 600	 * slabs on the per cpu partial list, in order to limit excessive
 601	 * growth of the list. For simplicity we assume that the slabs will
 602	 * be half-full.
 603	 */
 604	nr_slabs = DIV_ROUND_UP(nr_objects * 2, oo_objects(s->oo));
 605	s->cpu_partial_slabs = nr_slabs;
 606}
 607#else
 608static inline void
 609slub_set_cpu_partial(struct kmem_cache *s, unsigned int nr_objects)
 610{
 611}
 612#endif /* CONFIG_SLUB_CPU_PARTIAL */
 613
 614/*
 615 * Per slab locking using the pagelock
 616 */
 617static __always_inline void slab_lock(struct slab *slab)
 618{
 619	struct page *page = slab_page(slab);
 620
 621	VM_BUG_ON_PAGE(PageTail(page), page);
 622	bit_spin_lock(PG_locked, &page->flags);
 623}
 624
 625static __always_inline void slab_unlock(struct slab *slab)
 626{
 627	struct page *page = slab_page(slab);
 628
 629	VM_BUG_ON_PAGE(PageTail(page), page);
 630	bit_spin_unlock(PG_locked, &page->flags);
 631}
 632
 633static inline bool
 634__update_freelist_fast(struct slab *slab,
 635		      void *freelist_old, unsigned long counters_old,
 636		      void *freelist_new, unsigned long counters_new)
 637{
 638#ifdef system_has_freelist_aba
 639	freelist_aba_t old = { .freelist = freelist_old, .counter = counters_old };
 640	freelist_aba_t new = { .freelist = freelist_new, .counter = counters_new };
 641
 642	return try_cmpxchg_freelist(&slab->freelist_counter.full, &old.full, new.full);
 643#else
 644	return false;
 645#endif
 646}
 647
 648static inline bool
 649__update_freelist_slow(struct slab *slab,
 650		      void *freelist_old, unsigned long counters_old,
 651		      void *freelist_new, unsigned long counters_new)
 652{
 653	bool ret = false;
 654
 655	slab_lock(slab);
 656	if (slab->freelist == freelist_old &&
 657	    slab->counters == counters_old) {
 658		slab->freelist = freelist_new;
 659		slab->counters = counters_new;
 660		ret = true;
 661	}
 662	slab_unlock(slab);
 663
 664	return ret;
 665}
 666
 667/*
 668 * Interrupts must be disabled (for the fallback code to work right), typically
 669 * by an _irqsave() lock variant. On PREEMPT_RT the preempt_disable(), which is
 670 * part of bit_spin_lock(), is sufficient because the policy is not to allow any
 671 * allocation/ free operation in hardirq context. Therefore nothing can
 672 * interrupt the operation.
 673 */
 674static inline bool __slab_update_freelist(struct kmem_cache *s, struct slab *slab,
 675		void *freelist_old, unsigned long counters_old,
 676		void *freelist_new, unsigned long counters_new,
 677		const char *n)
 678{
 679	bool ret;
 680
 681	if (USE_LOCKLESS_FAST_PATH())
 682		lockdep_assert_irqs_disabled();
 683
 684	if (s->flags & __CMPXCHG_DOUBLE) {
 685		ret = __update_freelist_fast(slab, freelist_old, counters_old,
 686				            freelist_new, counters_new);
 687	} else {
 688		ret = __update_freelist_slow(slab, freelist_old, counters_old,
 689				            freelist_new, counters_new);
 690	}
 691	if (likely(ret))
 692		return true;
 693
 694	cpu_relax();
 695	stat(s, CMPXCHG_DOUBLE_FAIL);
 696
 697#ifdef SLUB_DEBUG_CMPXCHG
 698	pr_info("%s %s: cmpxchg double redo ", n, s->name);
 699#endif
 700
 701	return false;
 702}
 703
 704static inline bool slab_update_freelist(struct kmem_cache *s, struct slab *slab,
 705		void *freelist_old, unsigned long counters_old,
 706		void *freelist_new, unsigned long counters_new,
 707		const char *n)
 708{
 709	bool ret;
 710
 711	if (s->flags & __CMPXCHG_DOUBLE) {
 712		ret = __update_freelist_fast(slab, freelist_old, counters_old,
 713				            freelist_new, counters_new);
 714	} else {
 715		unsigned long flags;
 716
 717		local_irq_save(flags);
 718		ret = __update_freelist_slow(slab, freelist_old, counters_old,
 719				            freelist_new, counters_new);
 720		local_irq_restore(flags);
 721	}
 722	if (likely(ret))
 723		return true;
 724
 725	cpu_relax();
 726	stat(s, CMPXCHG_DOUBLE_FAIL);
 727
 728#ifdef SLUB_DEBUG_CMPXCHG
 729	pr_info("%s %s: cmpxchg double redo ", n, s->name);
 730#endif
 731
 732	return false;
 733}
 734
 735#ifdef CONFIG_SLUB_DEBUG
 736static unsigned long object_map[BITS_TO_LONGS(MAX_OBJS_PER_PAGE)];
 737static DEFINE_SPINLOCK(object_map_lock);
 738
 739static void __fill_map(unsigned long *obj_map, struct kmem_cache *s,
 740		       struct slab *slab)
 741{
 742	void *addr = slab_address(slab);
 743	void *p;
 744
 745	bitmap_zero(obj_map, slab->objects);
 746
 747	for (p = slab->freelist; p; p = get_freepointer(s, p))
 748		set_bit(__obj_to_index(s, addr, p), obj_map);
 749}
 750
 751#if IS_ENABLED(CONFIG_KUNIT)
 752static bool slab_add_kunit_errors(void)
 753{
 754	struct kunit_resource *resource;
 755
 756	if (!kunit_get_current_test())
 757		return false;
 758
 759	resource = kunit_find_named_resource(current->kunit_test, "slab_errors");
 760	if (!resource)
 761		return false;
 762
 763	(*(int *)resource->data)++;
 764	kunit_put_resource(resource);
 765	return true;
 766}
 767#else
 768static inline bool slab_add_kunit_errors(void) { return false; }
 769#endif
 770
 771static inline unsigned int size_from_object(struct kmem_cache *s)
 772{
 773	if (s->flags & SLAB_RED_ZONE)
 774		return s->size - s->red_left_pad;
 775
 776	return s->size;
 777}
 778
 779static inline void *restore_red_left(struct kmem_cache *s, void *p)
 780{
 781	if (s->flags & SLAB_RED_ZONE)
 782		p -= s->red_left_pad;
 783
 784	return p;
 785}
 786
 787/*
 788 * Debug settings:
 789 */
 790#if defined(CONFIG_SLUB_DEBUG_ON)
 791static slab_flags_t slub_debug = DEBUG_DEFAULT_FLAGS;
 792#else
 793static slab_flags_t slub_debug;
 794#endif
 795
 796static char *slub_debug_string;
 797static int disable_higher_order_debug;
 798
 799/*
 800 * slub is about to manipulate internal object metadata.  This memory lies
 801 * outside the range of the allocated object, so accessing it would normally
 802 * be reported by kasan as a bounds error.  metadata_access_enable() is used
 803 * to tell kasan that these accesses are OK.
 804 */
 805static inline void metadata_access_enable(void)
 806{
 807	kasan_disable_current();
 808}
 809
 810static inline void metadata_access_disable(void)
 811{
 812	kasan_enable_current();
 813}
 814
 815/*
 816 * Object debugging
 817 */
 818
 819/* Verify that a pointer has an address that is valid within a slab page */
 820static inline int check_valid_pointer(struct kmem_cache *s,
 821				struct slab *slab, void *object)
 822{
 823	void *base;
 824
 825	if (!object)
 826		return 1;
 827
 828	base = slab_address(slab);
 829	object = kasan_reset_tag(object);
 830	object = restore_red_left(s, object);
 831	if (object < base || object >= base + slab->objects * s->size ||
 832		(object - base) % s->size) {
 833		return 0;
 834	}
 835
 836	return 1;
 837}
 838
 839static void print_section(char *level, char *text, u8 *addr,
 840			  unsigned int length)
 841{
 842	metadata_access_enable();
 843	print_hex_dump(level, text, DUMP_PREFIX_ADDRESS,
 844			16, 1, kasan_reset_tag((void *)addr), length, 1);
 845	metadata_access_disable();
 846}
 847
 848/*
 849 * See comment in calculate_sizes().
 850 */
 851static inline bool freeptr_outside_object(struct kmem_cache *s)
 852{
 853	return s->offset >= s->inuse;
 854}
 855
 856/*
 857 * Return offset of the end of info block which is inuse + free pointer if
 858 * not overlapping with object.
 859 */
 860static inline unsigned int get_info_end(struct kmem_cache *s)
 861{
 862	if (freeptr_outside_object(s))
 863		return s->inuse + sizeof(void *);
 864	else
 865		return s->inuse;
 866}
 867
 868static struct track *get_track(struct kmem_cache *s, void *object,
 869	enum track_item alloc)
 870{
 871	struct track *p;
 872
 873	p = object + get_info_end(s);
 874
 875	return kasan_reset_tag(p + alloc);
 876}
 877
 878#ifdef CONFIG_STACKDEPOT
 879static noinline depot_stack_handle_t set_track_prepare(void)
 880{
 881	depot_stack_handle_t handle;
 882	unsigned long entries[TRACK_ADDRS_COUNT];
 883	unsigned int nr_entries;
 884
 885	nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 3);
 886	handle = stack_depot_save(entries, nr_entries, GFP_NOWAIT);
 887
 888	return handle;
 889}
 890#else
 891static inline depot_stack_handle_t set_track_prepare(void)
 892{
 893	return 0;
 894}
 895#endif
 896
 897static void set_track_update(struct kmem_cache *s, void *object,
 898			     enum track_item alloc, unsigned long addr,
 899			     depot_stack_handle_t handle)
 900{
 901	struct track *p = get_track(s, object, alloc);
 902
 903#ifdef CONFIG_STACKDEPOT
 904	p->handle = handle;
 905#endif
 906	p->addr = addr;
 907	p->cpu = smp_processor_id();
 908	p->pid = current->pid;
 909	p->when = jiffies;
 910}
 911
 912static __always_inline void set_track(struct kmem_cache *s, void *object,
 913				      enum track_item alloc, unsigned long addr)
 914{
 915	depot_stack_handle_t handle = set_track_prepare();
 916
 917	set_track_update(s, object, alloc, addr, handle);
 918}
 919
 920static void init_tracking(struct kmem_cache *s, void *object)
 921{
 922	struct track *p;
 923
 924	if (!(s->flags & SLAB_STORE_USER))
 925		return;
 926
 927	p = get_track(s, object, TRACK_ALLOC);
 928	memset(p, 0, 2*sizeof(struct track));
 929}
 930
 931static void print_track(const char *s, struct track *t, unsigned long pr_time)
 932{
 933	depot_stack_handle_t handle __maybe_unused;
 934
 935	if (!t->addr)
 936		return;
 937
 938	pr_err("%s in %pS age=%lu cpu=%u pid=%d\n",
 939	       s, (void *)t->addr, pr_time - t->when, t->cpu, t->pid);
 940#ifdef CONFIG_STACKDEPOT
 941	handle = READ_ONCE(t->handle);
 942	if (handle)
 943		stack_depot_print(handle);
 944	else
 945		pr_err("object allocation/free stack trace missing\n");
 946#endif
 947}
 948
 949void print_tracking(struct kmem_cache *s, void *object)
 950{
 951	unsigned long pr_time = jiffies;
 952	if (!(s->flags & SLAB_STORE_USER))
 953		return;
 954
 955	print_track("Allocated", get_track(s, object, TRACK_ALLOC), pr_time);
 956	print_track("Freed", get_track(s, object, TRACK_FREE), pr_time);
 957}
 958
 959static void print_slab_info(const struct slab *slab)
 960{
 961	struct folio *folio = (struct folio *)slab_folio(slab);
 962
 963	pr_err("Slab 0x%p objects=%u used=%u fp=0x%p flags=%pGp\n",
 964	       slab, slab->objects, slab->inuse, slab->freelist,
 965	       folio_flags(folio, 0));
 966}
 967
 968/*
 969 * kmalloc caches has fixed sizes (mostly power of 2), and kmalloc() API
 970 * family will round up the real request size to these fixed ones, so
 971 * there could be an extra area than what is requested. Save the original
 972 * request size in the meta data area, for better debug and sanity check.
 973 */
 974static inline void set_orig_size(struct kmem_cache *s,
 975				void *object, unsigned int orig_size)
 976{
 977	void *p = kasan_reset_tag(object);
 978	unsigned int kasan_meta_size;
 979
 980	if (!slub_debug_orig_size(s))
 981		return;
 982
 983	/*
 984	 * KASAN can save its free meta data inside of the object at offset 0.
 985	 * If this meta data size is larger than 'orig_size', it will overlap
 986	 * the data redzone in [orig_size+1, object_size]. Thus, we adjust
 987	 * 'orig_size' to be as at least as big as KASAN's meta data.
 988	 */
 989	kasan_meta_size = kasan_metadata_size(s, true);
 990	if (kasan_meta_size > orig_size)
 991		orig_size = kasan_meta_size;
 992
 993	p += get_info_end(s);
 994	p += sizeof(struct track) * 2;
 995
 996	*(unsigned int *)p = orig_size;
 997}
 998
 999static inline unsigned int get_orig_size(struct kmem_cache *s, void *object)
1000{
1001	void *p = kasan_reset_tag(object);
1002
1003	if (!slub_debug_orig_size(s))
1004		return s->object_size;
1005
1006	p += get_info_end(s);
1007	p += sizeof(struct track) * 2;
1008
1009	return *(unsigned int *)p;
1010}
1011
1012void skip_orig_size_check(struct kmem_cache *s, const void *object)
1013{
1014	set_orig_size(s, (void *)object, s->object_size);
1015}
1016
1017static void slab_bug(struct kmem_cache *s, char *fmt, ...)
1018{
1019	struct va_format vaf;
1020	va_list args;
1021
1022	va_start(args, fmt);
1023	vaf.fmt = fmt;
1024	vaf.va = &args;
1025	pr_err("=============================================================================\n");
1026	pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
1027	pr_err("-----------------------------------------------------------------------------\n\n");
1028	va_end(args);
1029}
1030
1031__printf(2, 3)
1032static void slab_fix(struct kmem_cache *s, char *fmt, ...)
1033{
1034	struct va_format vaf;
1035	va_list args;
1036
1037	if (slab_add_kunit_errors())
1038		return;
1039
1040	va_start(args, fmt);
1041	vaf.fmt = fmt;
1042	vaf.va = &args;
1043	pr_err("FIX %s: %pV\n", s->name, &vaf);
1044	va_end(args);
1045}
1046
1047static void print_trailer(struct kmem_cache *s, struct slab *slab, u8 *p)
1048{
1049	unsigned int off;	/* Offset of last byte */
1050	u8 *addr = slab_address(slab);
1051
1052	print_tracking(s, p);
1053
1054	print_slab_info(slab);
1055
1056	pr_err("Object 0x%p @offset=%tu fp=0x%p\n\n",
1057	       p, p - addr, get_freepointer(s, p));
1058
1059	if (s->flags & SLAB_RED_ZONE)
1060		print_section(KERN_ERR, "Redzone  ", p - s->red_left_pad,
1061			      s->red_left_pad);
1062	else if (p > addr + 16)
1063		print_section(KERN_ERR, "Bytes b4 ", p - 16, 16);
1064
1065	print_section(KERN_ERR,         "Object   ", p,
1066		      min_t(unsigned int, s->object_size, PAGE_SIZE));
1067	if (s->flags & SLAB_RED_ZONE)
1068		print_section(KERN_ERR, "Redzone  ", p + s->object_size,
1069			s->inuse - s->object_size);
1070
1071	off = get_info_end(s);
1072
1073	if (s->flags & SLAB_STORE_USER)
1074		off += 2 * sizeof(struct track);
1075
1076	if (slub_debug_orig_size(s))
1077		off += sizeof(unsigned int);
1078
1079	off += kasan_metadata_size(s, false);
1080
1081	if (off != size_from_object(s))
1082		/* Beginning of the filler is the free pointer */
1083		print_section(KERN_ERR, "Padding  ", p + off,
1084			      size_from_object(s) - off);
1085
1086	dump_stack();
1087}
1088
1089static void object_err(struct kmem_cache *s, struct slab *slab,
1090			u8 *object, char *reason)
1091{
1092	if (slab_add_kunit_errors())
1093		return;
1094
1095	slab_bug(s, "%s", reason);
1096	print_trailer(s, slab, object);
1097	add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
1098}
1099
1100static bool freelist_corrupted(struct kmem_cache *s, struct slab *slab,
1101			       void **freelist, void *nextfree)
1102{
1103	if ((s->flags & SLAB_CONSISTENCY_CHECKS) &&
1104	    !check_valid_pointer(s, slab, nextfree) && freelist) {
1105		object_err(s, slab, *freelist, "Freechain corrupt");
1106		*freelist = NULL;
1107		slab_fix(s, "Isolate corrupted freechain");
1108		return true;
1109	}
1110
1111	return false;
1112}
1113
1114static __printf(3, 4) void slab_err(struct kmem_cache *s, struct slab *slab,
1115			const char *fmt, ...)
1116{
1117	va_list args;
1118	char buf[100];
1119
1120	if (slab_add_kunit_errors())
1121		return;
1122
1123	va_start(args, fmt);
1124	vsnprintf(buf, sizeof(buf), fmt, args);
1125	va_end(args);
1126	slab_bug(s, "%s", buf);
1127	print_slab_info(slab);
1128	dump_stack();
1129	add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
1130}
1131
1132static void init_object(struct kmem_cache *s, void *object, u8 val)
1133{
1134	u8 *p = kasan_reset_tag(object);
1135	unsigned int poison_size = s->object_size;
1136
1137	if (s->flags & SLAB_RED_ZONE) {
1138		memset(p - s->red_left_pad, val, s->red_left_pad);
1139
1140		if (slub_debug_orig_size(s) && val == SLUB_RED_ACTIVE) {
1141			/*
1142			 * Redzone the extra allocated space by kmalloc than
1143			 * requested, and the poison size will be limited to
1144			 * the original request size accordingly.
1145			 */
1146			poison_size = get_orig_size(s, object);
1147		}
1148	}
1149
1150	if (s->flags & __OBJECT_POISON) {
1151		memset(p, POISON_FREE, poison_size - 1);
1152		p[poison_size - 1] = POISON_END;
1153	}
1154
1155	if (s->flags & SLAB_RED_ZONE)
1156		memset(p + poison_size, val, s->inuse - poison_size);
1157}
1158
1159static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
1160						void *from, void *to)
1161{
1162	slab_fix(s, "Restoring %s 0x%p-0x%p=0x%x", message, from, to - 1, data);
1163	memset(from, data, to - from);
1164}
1165
1166static int check_bytes_and_report(struct kmem_cache *s, struct slab *slab,
1167			u8 *object, char *what,
1168			u8 *start, unsigned int value, unsigned int bytes)
1169{
1170	u8 *fault;
1171	u8 *end;
1172	u8 *addr = slab_address(slab);
1173
1174	metadata_access_enable();
1175	fault = memchr_inv(kasan_reset_tag(start), value, bytes);
1176	metadata_access_disable();
1177	if (!fault)
1178		return 1;
1179
1180	end = start + bytes;
1181	while (end > fault && end[-1] == value)
1182		end--;
1183
1184	if (slab_add_kunit_errors())
1185		goto skip_bug_print;
1186
1187	slab_bug(s, "%s overwritten", what);
1188	pr_err("0x%p-0x%p @offset=%tu. First byte 0x%x instead of 0x%x\n",
1189					fault, end - 1, fault - addr,
1190					fault[0], value);
1191	print_trailer(s, slab, object);
1192	add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
1193
1194skip_bug_print:
1195	restore_bytes(s, what, value, fault, end);
1196	return 0;
1197}
1198
1199/*
1200 * Object layout:
1201 *
1202 * object address
1203 * 	Bytes of the object to be managed.
1204 * 	If the freepointer may overlay the object then the free
1205 *	pointer is at the middle of the object.
1206 *
1207 * 	Poisoning uses 0x6b (POISON_FREE) and the last byte is
1208 * 	0xa5 (POISON_END)
1209 *
1210 * object + s->object_size
1211 * 	Padding to reach word boundary. This is also used for Redzoning.
1212 * 	Padding is extended by another word if Redzoning is enabled and
1213 * 	object_size == inuse.
1214 *
1215 * 	We fill with 0xbb (RED_INACTIVE) for inactive objects and with
1216 * 	0xcc (RED_ACTIVE) for objects in use.
1217 *
1218 * object + s->inuse
1219 * 	Meta data starts here.
1220 *
1221 * 	A. Free pointer (if we cannot overwrite object on free)
1222 * 	B. Tracking data for SLAB_STORE_USER
1223 *	C. Original request size for kmalloc object (SLAB_STORE_USER enabled)
1224 *	D. Padding to reach required alignment boundary or at minimum
1225 * 		one word if debugging is on to be able to detect writes
1226 * 		before the word boundary.
1227 *
1228 *	Padding is done using 0x5a (POISON_INUSE)
1229 *
1230 * object + s->size
1231 * 	Nothing is used beyond s->size.
1232 *
1233 * If slabcaches are merged then the object_size and inuse boundaries are mostly
1234 * ignored. And therefore no slab options that rely on these boundaries
1235 * may be used with merged slabcaches.
1236 */
1237
1238static int check_pad_bytes(struct kmem_cache *s, struct slab *slab, u8 *p)
1239{
1240	unsigned long off = get_info_end(s);	/* The end of info */
1241
1242	if (s->flags & SLAB_STORE_USER) {
1243		/* We also have user information there */
1244		off += 2 * sizeof(struct track);
1245
1246		if (s->flags & SLAB_KMALLOC)
1247			off += sizeof(unsigned int);
1248	}
1249
1250	off += kasan_metadata_size(s, false);
1251
1252	if (size_from_object(s) == off)
1253		return 1;
1254
1255	return check_bytes_and_report(s, slab, p, "Object padding",
1256			p + off, POISON_INUSE, size_from_object(s) - off);
1257}
1258
1259/* Check the pad bytes at the end of a slab page */
1260static void slab_pad_check(struct kmem_cache *s, struct slab *slab)
1261{
1262	u8 *start;
1263	u8 *fault;
1264	u8 *end;
1265	u8 *pad;
1266	int length;
1267	int remainder;
1268
1269	if (!(s->flags & SLAB_POISON))
1270		return;
1271
1272	start = slab_address(slab);
1273	length = slab_size(slab);
1274	end = start + length;
1275	remainder = length % s->size;
1276	if (!remainder)
1277		return;
1278
1279	pad = end - remainder;
1280	metadata_access_enable();
1281	fault = memchr_inv(kasan_reset_tag(pad), POISON_INUSE, remainder);
1282	metadata_access_disable();
1283	if (!fault)
1284		return;
1285	while (end > fault && end[-1] == POISON_INUSE)
1286		end--;
1287
1288	slab_err(s, slab, "Padding overwritten. 0x%p-0x%p @offset=%tu",
1289			fault, end - 1, fault - start);
1290	print_section(KERN_ERR, "Padding ", pad, remainder);
1291
1292	restore_bytes(s, "slab padding", POISON_INUSE, fault, end);
1293}
1294
1295static int check_object(struct kmem_cache *s, struct slab *slab,
1296					void *object, u8 val)
1297{
1298	u8 *p = object;
1299	u8 *endobject = object + s->object_size;
1300	unsigned int orig_size, kasan_meta_size;
1301
1302	if (s->flags & SLAB_RED_ZONE) {
1303		if (!check_bytes_and_report(s, slab, object, "Left Redzone",
1304			object - s->red_left_pad, val, s->red_left_pad))
1305			return 0;
1306
1307		if (!check_bytes_and_report(s, slab, object, "Right Redzone",
1308			endobject, val, s->inuse - s->object_size))
1309			return 0;
1310
1311		if (slub_debug_orig_size(s) && val == SLUB_RED_ACTIVE) {
1312			orig_size = get_orig_size(s, object);
1313
1314			if (s->object_size > orig_size  &&
1315				!check_bytes_and_report(s, slab, object,
1316					"kmalloc Redzone", p + orig_size,
1317					val, s->object_size - orig_size)) {
1318				return 0;
1319			}
1320		}
1321	} else {
1322		if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
1323			check_bytes_and_report(s, slab, p, "Alignment padding",
1324				endobject, POISON_INUSE,
1325				s->inuse - s->object_size);
1326		}
1327	}
1328
1329	if (s->flags & SLAB_POISON) {
1330		if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON)) {
1331			/*
1332			 * KASAN can save its free meta data inside of the
1333			 * object at offset 0. Thus, skip checking the part of
1334			 * the redzone that overlaps with the meta data.
1335			 */
1336			kasan_meta_size = kasan_metadata_size(s, true);
1337			if (kasan_meta_size < s->object_size - 1 &&
1338			    !check_bytes_and_report(s, slab, p, "Poison",
1339					p + kasan_meta_size, POISON_FREE,
1340					s->object_size - kasan_meta_size - 1))
1341				return 0;
1342			if (kasan_meta_size < s->object_size &&
1343			    !check_bytes_and_report(s, slab, p, "End Poison",
1344					p + s->object_size - 1, POISON_END, 1))
1345				return 0;
1346		}
1347		/*
1348		 * check_pad_bytes cleans up on its own.
1349		 */
1350		check_pad_bytes(s, slab, p);
1351	}
1352
1353	if (!freeptr_outside_object(s) && val == SLUB_RED_ACTIVE)
1354		/*
1355		 * Object and freepointer overlap. Cannot check
1356		 * freepointer while object is allocated.
1357		 */
1358		return 1;
1359
1360	/* Check free pointer validity */
1361	if (!check_valid_pointer(s, slab, get_freepointer(s, p))) {
1362		object_err(s, slab, p, "Freepointer corrupt");
1363		/*
1364		 * No choice but to zap it and thus lose the remainder
1365		 * of the free objects in this slab. May cause
1366		 * another error because the object count is now wrong.
1367		 */
1368		set_freepointer(s, p, NULL);
1369		return 0;
1370	}
1371	return 1;
1372}
1373
1374static int check_slab(struct kmem_cache *s, struct slab *slab)
1375{
1376	int maxobj;
1377
1378	if (!folio_test_slab(slab_folio(slab))) {
1379		slab_err(s, slab, "Not a valid slab page");
1380		return 0;
1381	}
1382
1383	maxobj = order_objects(slab_order(slab), s->size);
1384	if (slab->objects > maxobj) {
1385		slab_err(s, slab, "objects %u > max %u",
1386			slab->objects, maxobj);
1387		return 0;
1388	}
1389	if (slab->inuse > slab->objects) {
1390		slab_err(s, slab, "inuse %u > max %u",
1391			slab->inuse, slab->objects);
1392		return 0;
1393	}
1394	/* Slab_pad_check fixes things up after itself */
1395	slab_pad_check(s, slab);
1396	return 1;
1397}
1398
1399/*
1400 * Determine if a certain object in a slab is on the freelist. Must hold the
1401 * slab lock to guarantee that the chains are in a consistent state.
1402 */
1403static int on_freelist(struct kmem_cache *s, struct slab *slab, void *search)
1404{
1405	int nr = 0;
1406	void *fp;
1407	void *object = NULL;
1408	int max_objects;
1409
1410	fp = slab->freelist;
1411	while (fp && nr <= slab->objects) {
1412		if (fp == search)
1413			return 1;
1414		if (!check_valid_pointer(s, slab, fp)) {
1415			if (object) {
1416				object_err(s, slab, object,
1417					"Freechain corrupt");
1418				set_freepointer(s, object, NULL);
1419			} else {
1420				slab_err(s, slab, "Freepointer corrupt");
1421				slab->freelist = NULL;
1422				slab->inuse = slab->objects;
1423				slab_fix(s, "Freelist cleared");
1424				return 0;
1425			}
1426			break;
1427		}
1428		object = fp;
1429		fp = get_freepointer(s, object);
1430		nr++;
1431	}
1432
1433	max_objects = order_objects(slab_order(slab), s->size);
1434	if (max_objects > MAX_OBJS_PER_PAGE)
1435		max_objects = MAX_OBJS_PER_PAGE;
1436
1437	if (slab->objects != max_objects) {
1438		slab_err(s, slab, "Wrong number of objects. Found %d but should be %d",
1439			 slab->objects, max_objects);
1440		slab->objects = max_objects;
1441		slab_fix(s, "Number of objects adjusted");
1442	}
1443	if (slab->inuse != slab->objects - nr) {
1444		slab_err(s, slab, "Wrong object count. Counter is %d but counted were %d",
1445			 slab->inuse, slab->objects - nr);
1446		slab->inuse = slab->objects - nr;
1447		slab_fix(s, "Object count adjusted");
1448	}
1449	return search == NULL;
1450}
1451
1452static void trace(struct kmem_cache *s, struct slab *slab, void *object,
1453								int alloc)
1454{
1455	if (s->flags & SLAB_TRACE) {
1456		pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
1457			s->name,
1458			alloc ? "alloc" : "free",
1459			object, slab->inuse,
1460			slab->freelist);
1461
1462		if (!alloc)
1463			print_section(KERN_INFO, "Object ", (void *)object,
1464					s->object_size);
1465
1466		dump_stack();
1467	}
1468}
1469
1470/*
1471 * Tracking of fully allocated slabs for debugging purposes.
1472 */
1473static void add_full(struct kmem_cache *s,
1474	struct kmem_cache_node *n, struct slab *slab)
1475{
1476	if (!(s->flags & SLAB_STORE_USER))
1477		return;
1478
1479	lockdep_assert_held(&n->list_lock);
1480	list_add(&slab->slab_list, &n->full);
1481}
1482
1483static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct slab *slab)
1484{
1485	if (!(s->flags & SLAB_STORE_USER))
1486		return;
1487
1488	lockdep_assert_held(&n->list_lock);
1489	list_del(&slab->slab_list);
1490}
1491
1492static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1493{
1494	return atomic_long_read(&n->nr_slabs);
1495}
1496
1497static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1498{
1499	struct kmem_cache_node *n = get_node(s, node);
1500
1501	/*
1502	 * May be called early in order to allocate a slab for the
1503	 * kmem_cache_node structure. Solve the chicken-egg
1504	 * dilemma by deferring the increment of the count during
1505	 * bootstrap (see early_kmem_cache_node_alloc).
1506	 */
1507	if (likely(n)) {
1508		atomic_long_inc(&n->nr_slabs);
1509		atomic_long_add(objects, &n->total_objects);
1510	}
1511}
1512static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1513{
1514	struct kmem_cache_node *n = get_node(s, node);
1515
1516	atomic_long_dec(&n->nr_slabs);
1517	atomic_long_sub(objects, &n->total_objects);
1518}
1519
1520/* Object debug checks for alloc/free paths */
1521static void setup_object_debug(struct kmem_cache *s, void *object)
1522{
1523	if (!kmem_cache_debug_flags(s, SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))
1524		return;
1525
1526	init_object(s, object, SLUB_RED_INACTIVE);
1527	init_tracking(s, object);
1528}
1529
1530static
1531void setup_slab_debug(struct kmem_cache *s, struct slab *slab, void *addr)
1532{
1533	if (!kmem_cache_debug_flags(s, SLAB_POISON))
1534		return;
1535
1536	metadata_access_enable();
1537	memset(kasan_reset_tag(addr), POISON_INUSE, slab_size(slab));
1538	metadata_access_disable();
1539}
1540
1541static inline int alloc_consistency_checks(struct kmem_cache *s,
1542					struct slab *slab, void *object)
1543{
1544	if (!check_slab(s, slab))
1545		return 0;
1546
1547	if (!check_valid_pointer(s, slab, object)) {
1548		object_err(s, slab, object, "Freelist Pointer check fails");
1549		return 0;
1550	}
1551
1552	if (!check_object(s, slab, object, SLUB_RED_INACTIVE))
1553		return 0;
1554
1555	return 1;
1556}
1557
1558static noinline bool alloc_debug_processing(struct kmem_cache *s,
1559			struct slab *slab, void *object, int orig_size)
1560{
1561	if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1562		if (!alloc_consistency_checks(s, slab, object))
1563			goto bad;
1564	}
1565
1566	/* Success. Perform special debug activities for allocs */
1567	trace(s, slab, object, 1);
1568	set_orig_size(s, object, orig_size);
1569	init_object(s, object, SLUB_RED_ACTIVE);
1570	return true;
1571
1572bad:
1573	if (folio_test_slab(slab_folio(slab))) {
1574		/*
1575		 * If this is a slab page then lets do the best we can
1576		 * to avoid issues in the future. Marking all objects
1577		 * as used avoids touching the remaining objects.
1578		 */
1579		slab_fix(s, "Marking all objects used");
1580		slab->inuse = slab->objects;
1581		slab->freelist = NULL;
1582	}
1583	return false;
1584}
1585
1586static inline int free_consistency_checks(struct kmem_cache *s,
1587		struct slab *slab, void *object, unsigned long addr)
1588{
1589	if (!check_valid_pointer(s, slab, object)) {
1590		slab_err(s, slab, "Invalid object pointer 0x%p", object);
1591		return 0;
1592	}
1593
1594	if (on_freelist(s, slab, object)) {
1595		object_err(s, slab, object, "Object already free");
1596		return 0;
1597	}
1598
1599	if (!check_object(s, slab, object, SLUB_RED_ACTIVE))
1600		return 0;
1601
1602	if (unlikely(s != slab->slab_cache)) {
1603		if (!folio_test_slab(slab_folio(slab))) {
1604			slab_err(s, slab, "Attempt to free object(0x%p) outside of slab",
1605				 object);
1606		} else if (!slab->slab_cache) {
1607			pr_err("SLUB <none>: no slab for object 0x%p.\n",
1608			       object);
1609			dump_stack();
1610		} else
1611			object_err(s, slab, object,
1612					"page slab pointer corrupt.");
1613		return 0;
1614	}
1615	return 1;
1616}
1617
1618/*
1619 * Parse a block of slub_debug options. Blocks are delimited by ';'
1620 *
1621 * @str:    start of block
1622 * @flags:  returns parsed flags, or DEBUG_DEFAULT_FLAGS if none specified
1623 * @slabs:  return start of list of slabs, or NULL when there's no list
1624 * @init:   assume this is initial parsing and not per-kmem-create parsing
1625 *
1626 * returns the start of next block if there's any, or NULL
1627 */
1628static char *
1629parse_slub_debug_flags(char *str, slab_flags_t *flags, char **slabs, bool init)
1630{
1631	bool higher_order_disable = false;
1632
1633	/* Skip any completely empty blocks */
1634	while (*str && *str == ';')
1635		str++;
1636
1637	if (*str == ',') {
1638		/*
1639		 * No options but restriction on slabs. This means full
1640		 * debugging for slabs matching a pattern.
1641		 */
1642		*flags = DEBUG_DEFAULT_FLAGS;
1643		goto check_slabs;
1644	}
1645	*flags = 0;
1646
1647	/* Determine which debug features should be switched on */
1648	for (; *str && *str != ',' && *str != ';'; str++) {
1649		switch (tolower(*str)) {
1650		case '-':
1651			*flags = 0;
1652			break;
1653		case 'f':
1654			*flags |= SLAB_CONSISTENCY_CHECKS;
1655			break;
1656		case 'z':
1657			*flags |= SLAB_RED_ZONE;
1658			break;
1659		case 'p':
1660			*flags |= SLAB_POISON;
1661			break;
1662		case 'u':
1663			*flags |= SLAB_STORE_USER;
1664			break;
1665		case 't':
1666			*flags |= SLAB_TRACE;
1667			break;
1668		case 'a':
1669			*flags |= SLAB_FAILSLAB;
1670			break;
1671		case 'o':
1672			/*
1673			 * Avoid enabling debugging on caches if its minimum
1674			 * order would increase as a result.
1675			 */
1676			higher_order_disable = true;
1677			break;
1678		default:
1679			if (init)
1680				pr_err("slub_debug option '%c' unknown. skipped\n", *str);
1681		}
1682	}
1683check_slabs:
1684	if (*str == ',')
1685		*slabs = ++str;
1686	else
1687		*slabs = NULL;
1688
1689	/* Skip over the slab list */
1690	while (*str && *str != ';')
1691		str++;
1692
1693	/* Skip any completely empty blocks */
1694	while (*str && *str == ';')
1695		str++;
1696
1697	if (init && higher_order_disable)
1698		disable_higher_order_debug = 1;
1699
1700	if (*str)
1701		return str;
1702	else
1703		return NULL;
1704}
1705
1706static int __init setup_slub_debug(char *str)
1707{
1708	slab_flags_t flags;
1709	slab_flags_t global_flags;
1710	char *saved_str;
1711	char *slab_list;
1712	bool global_slub_debug_changed = false;
1713	bool slab_list_specified = false;
1714
1715	global_flags = DEBUG_DEFAULT_FLAGS;
1716	if (*str++ != '=' || !*str)
1717		/*
1718		 * No options specified. Switch on full debugging.
1719		 */
1720		goto out;
1721
1722	saved_str = str;
1723	while (str) {
1724		str = parse_slub_debug_flags(str, &flags, &slab_list, true);
1725
1726		if (!slab_list) {
1727			global_flags = flags;
1728			global_slub_debug_changed = true;
1729		} else {
1730			slab_list_specified = true;
1731			if (flags & SLAB_STORE_USER)
1732				stack_depot_request_early_init();
1733		}
1734	}
1735
1736	/*
1737	 * For backwards compatibility, a single list of flags with list of
1738	 * slabs means debugging is only changed for those slabs, so the global
1739	 * slub_debug should be unchanged (0 or DEBUG_DEFAULT_FLAGS, depending
1740	 * on CONFIG_SLUB_DEBUG_ON). We can extended that to multiple lists as
1741	 * long as there is no option specifying flags without a slab list.
1742	 */
1743	if (slab_list_specified) {
1744		if (!global_slub_debug_changed)
1745			global_flags = slub_debug;
1746		slub_debug_string = saved_str;
1747	}
1748out:
1749	slub_debug = global_flags;
1750	if (slub_debug & SLAB_STORE_USER)
1751		stack_depot_request_early_init();
1752	if (slub_debug != 0 || slub_debug_string)
1753		static_branch_enable(&slub_debug_enabled);
1754	else
1755		static_branch_disable(&slub_debug_enabled);
1756	if ((static_branch_unlikely(&init_on_alloc) ||
1757	     static_branch_unlikely(&init_on_free)) &&
1758	    (slub_debug & SLAB_POISON))
1759		pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n");
1760	return 1;
1761}
1762
1763__setup("slub_debug", setup_slub_debug);
 
1764
1765/*
1766 * kmem_cache_flags - apply debugging options to the cache
1767 * @object_size:	the size of an object without meta data
1768 * @flags:		flags to set
1769 * @name:		name of the cache
1770 *
1771 * Debug option(s) are applied to @flags. In addition to the debug
1772 * option(s), if a slab name (or multiple) is specified i.e.
1773 * slub_debug=<Debug-Options>,<slab name1>,<slab name2> ...
1774 * then only the select slabs will receive the debug option(s).
1775 */
1776slab_flags_t kmem_cache_flags(unsigned int object_size,
1777	slab_flags_t flags, const char *name)
1778{
1779	char *iter;
1780	size_t len;
1781	char *next_block;
1782	slab_flags_t block_flags;
1783	slab_flags_t slub_debug_local = slub_debug;
1784
1785	if (flags & SLAB_NO_USER_FLAGS)
1786		return flags;
1787
1788	/*
1789	 * If the slab cache is for debugging (e.g. kmemleak) then
1790	 * don't store user (stack trace) information by default,
1791	 * but let the user enable it via the command line below.
1792	 */
1793	if (flags & SLAB_NOLEAKTRACE)
1794		slub_debug_local &= ~SLAB_STORE_USER;
1795
1796	len = strlen(name);
1797	next_block = slub_debug_string;
1798	/* Go through all blocks of debug options, see if any matches our slab's name */
1799	while (next_block) {
1800		next_block = parse_slub_debug_flags(next_block, &block_flags, &iter, false);
1801		if (!iter)
1802			continue;
1803		/* Found a block that has a slab list, search it */
1804		while (*iter) {
1805			char *end, *glob;
1806			size_t cmplen;
1807
1808			end = strchrnul(iter, ',');
1809			if (next_block && next_block < end)
1810				end = next_block - 1;
1811
1812			glob = strnchr(iter, end - iter, '*');
1813			if (glob)
1814				cmplen = glob - iter;
1815			else
1816				cmplen = max_t(size_t, len, (end - iter));
1817
1818			if (!strncmp(name, iter, cmplen)) {
1819				flags |= block_flags;
1820				return flags;
1821			}
1822
1823			if (!*end || *end == ';')
1824				break;
1825			iter = end + 1;
1826		}
1827	}
1828
1829	return flags | slub_debug_local;
1830}
1831#else /* !CONFIG_SLUB_DEBUG */
1832static inline void setup_object_debug(struct kmem_cache *s, void *object) {}
1833static inline
1834void setup_slab_debug(struct kmem_cache *s, struct slab *slab, void *addr) {}
1835
1836static inline bool alloc_debug_processing(struct kmem_cache *s,
1837	struct slab *slab, void *object, int orig_size) { return true; }
1838
1839static inline bool free_debug_processing(struct kmem_cache *s,
1840	struct slab *slab, void *head, void *tail, int *bulk_cnt,
1841	unsigned long addr, depot_stack_handle_t handle) { return true; }
1842
1843static inline void slab_pad_check(struct kmem_cache *s, struct slab *slab) {}
1844static inline int check_object(struct kmem_cache *s, struct slab *slab,
1845			void *object, u8 val) { return 1; }
1846static inline depot_stack_handle_t set_track_prepare(void) { return 0; }
1847static inline void set_track(struct kmem_cache *s, void *object,
1848			     enum track_item alloc, unsigned long addr) {}
1849static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1850					struct slab *slab) {}
1851static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1852					struct slab *slab) {}
1853slab_flags_t kmem_cache_flags(unsigned int object_size,
1854	slab_flags_t flags, const char *name)
1855{
1856	return flags;
1857}
1858#define slub_debug 0
1859
1860#define disable_higher_order_debug 0
1861
1862static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1863							{ return 0; }
1864static inline void inc_slabs_node(struct kmem_cache *s, int node,
1865							int objects) {}
1866static inline void dec_slabs_node(struct kmem_cache *s, int node,
1867							int objects) {}
1868
1869#ifndef CONFIG_SLUB_TINY
1870static bool freelist_corrupted(struct kmem_cache *s, struct slab *slab,
1871			       void **freelist, void *nextfree)
1872{
1873	return false;
1874}
1875#endif
1876#endif /* CONFIG_SLUB_DEBUG */
1877
1878static inline enum node_stat_item cache_vmstat_idx(struct kmem_cache *s)
1879{
1880	return (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1881		NR_SLAB_RECLAIMABLE_B : NR_SLAB_UNRECLAIMABLE_B;
1882}
1883
1884#ifdef CONFIG_MEMCG_KMEM
1885static inline void memcg_free_slab_cgroups(struct slab *slab)
1886{
1887	kfree(slab_objcgs(slab));
1888	slab->memcg_data = 0;
1889}
1890
1891static inline size_t obj_full_size(struct kmem_cache *s)
1892{
1893	/*
1894	 * For each accounted object there is an extra space which is used
1895	 * to store obj_cgroup membership. Charge it too.
1896	 */
1897	return s->size + sizeof(struct obj_cgroup *);
1898}
1899
1900/*
1901 * Returns false if the allocation should fail.
1902 */
1903static bool __memcg_slab_pre_alloc_hook(struct kmem_cache *s,
1904					struct list_lru *lru,
1905					struct obj_cgroup **objcgp,
1906					size_t objects, gfp_t flags)
1907{
1908	/*
1909	 * The obtained objcg pointer is safe to use within the current scope,
1910	 * defined by current task or set_active_memcg() pair.
1911	 * obj_cgroup_get() is used to get a permanent reference.
1912	 */
1913	struct obj_cgroup *objcg = current_obj_cgroup();
1914	if (!objcg)
1915		return true;
1916
1917	if (lru) {
1918		int ret;
1919		struct mem_cgroup *memcg;
1920
1921		memcg = get_mem_cgroup_from_objcg(objcg);
1922		ret = memcg_list_lru_alloc(memcg, lru, flags);
1923		css_put(&memcg->css);
1924
1925		if (ret)
1926			return false;
1927	}
1928
1929	if (obj_cgroup_charge(objcg, flags, objects * obj_full_size(s)))
1930		return false;
1931
1932	*objcgp = objcg;
1933	return true;
1934}
1935
1936/*
1937 * Returns false if the allocation should fail.
1938 */
1939static __fastpath_inline
1940bool memcg_slab_pre_alloc_hook(struct kmem_cache *s, struct list_lru *lru,
1941			       struct obj_cgroup **objcgp, size_t objects,
1942			       gfp_t flags)
1943{
1944	if (!memcg_kmem_online())
1945		return true;
1946
1947	if (likely(!(flags & __GFP_ACCOUNT) && !(s->flags & SLAB_ACCOUNT)))
1948		return true;
1949
1950	return likely(__memcg_slab_pre_alloc_hook(s, lru, objcgp, objects,
1951						  flags));
1952}
1953
1954static void __memcg_slab_post_alloc_hook(struct kmem_cache *s,
1955					 struct obj_cgroup *objcg,
1956					 gfp_t flags, size_t size,
1957					 void **p)
1958{
1959	struct slab *slab;
1960	unsigned long off;
1961	size_t i;
1962
1963	flags &= gfp_allowed_mask;
1964
1965	for (i = 0; i < size; i++) {
1966		if (likely(p[i])) {
1967			slab = virt_to_slab(p[i]);
1968
1969			if (!slab_objcgs(slab) &&
1970			    memcg_alloc_slab_cgroups(slab, s, flags, false)) {
1971				obj_cgroup_uncharge(objcg, obj_full_size(s));
1972				continue;
1973			}
1974
1975			off = obj_to_index(s, slab, p[i]);
1976			obj_cgroup_get(objcg);
1977			slab_objcgs(slab)[off] = objcg;
1978			mod_objcg_state(objcg, slab_pgdat(slab),
1979					cache_vmstat_idx(s), obj_full_size(s));
1980		} else {
1981			obj_cgroup_uncharge(objcg, obj_full_size(s));
1982		}
1983	}
1984}
1985
1986static __fastpath_inline
1987void memcg_slab_post_alloc_hook(struct kmem_cache *s, struct obj_cgroup *objcg,
1988				gfp_t flags, size_t size, void **p)
1989{
1990	if (likely(!memcg_kmem_online() || !objcg))
1991		return;
1992
1993	return __memcg_slab_post_alloc_hook(s, objcg, flags, size, p);
1994}
1995
1996static void __memcg_slab_free_hook(struct kmem_cache *s, struct slab *slab,
1997				   void **p, int objects,
1998				   struct obj_cgroup **objcgs)
1999{
2000	for (int i = 0; i < objects; i++) {
2001		struct obj_cgroup *objcg;
2002		unsigned int off;
2003
2004		off = obj_to_index(s, slab, p[i]);
2005		objcg = objcgs[off];
2006		if (!objcg)
2007			continue;
2008
2009		objcgs[off] = NULL;
2010		obj_cgroup_uncharge(objcg, obj_full_size(s));
2011		mod_objcg_state(objcg, slab_pgdat(slab), cache_vmstat_idx(s),
2012				-obj_full_size(s));
2013		obj_cgroup_put(objcg);
2014	}
2015}
2016
2017static __fastpath_inline
2018void memcg_slab_free_hook(struct kmem_cache *s, struct slab *slab, void **p,
2019			  int objects)
2020{
2021	struct obj_cgroup **objcgs;
2022
2023	if (!memcg_kmem_online())
2024		return;
2025
2026	objcgs = slab_objcgs(slab);
2027	if (likely(!objcgs))
2028		return;
2029
2030	__memcg_slab_free_hook(s, slab, p, objects, objcgs);
2031}
2032
2033static inline
2034void memcg_slab_alloc_error_hook(struct kmem_cache *s, int objects,
2035			   struct obj_cgroup *objcg)
2036{
2037	if (objcg)
2038		obj_cgroup_uncharge(objcg, objects * obj_full_size(s));
2039}
2040#else /* CONFIG_MEMCG_KMEM */
2041static inline struct mem_cgroup *memcg_from_slab_obj(void *ptr)
2042{
2043	return NULL;
2044}
2045
2046static inline void memcg_free_slab_cgroups(struct slab *slab)
2047{
2048}
2049
2050static inline bool memcg_slab_pre_alloc_hook(struct kmem_cache *s,
2051					     struct list_lru *lru,
2052					     struct obj_cgroup **objcgp,
2053					     size_t objects, gfp_t flags)
2054{
2055	return true;
2056}
2057
2058static inline void memcg_slab_post_alloc_hook(struct kmem_cache *s,
2059					      struct obj_cgroup *objcg,
2060					      gfp_t flags, size_t size,
2061					      void **p)
2062{
2063}
2064
2065static inline void memcg_slab_free_hook(struct kmem_cache *s, struct slab *slab,
2066					void **p, int objects)
2067{
2068}
2069
2070static inline
2071void memcg_slab_alloc_error_hook(struct kmem_cache *s, int objects,
2072				 struct obj_cgroup *objcg)
2073{
2074}
2075#endif /* CONFIG_MEMCG_KMEM */
2076
2077/*
2078 * Hooks for other subsystems that check memory allocations. In a typical
2079 * production configuration these hooks all should produce no code at all.
2080 *
2081 * Returns true if freeing of the object can proceed, false if its reuse
2082 * was delayed by KASAN quarantine, or it was returned to KFENCE.
2083 */
2084static __always_inline
2085bool slab_free_hook(struct kmem_cache *s, void *x, bool init)
2086{
2087	kmemleak_free_recursive(x, s->flags);
2088	kmsan_slab_free(s, x);
2089
2090	debug_check_no_locks_freed(x, s->object_size);
2091
2092	if (!(s->flags & SLAB_DEBUG_OBJECTS))
2093		debug_check_no_obj_freed(x, s->object_size);
2094
2095	/* Use KCSAN to help debug racy use-after-free. */
2096	if (!(s->flags & SLAB_TYPESAFE_BY_RCU))
2097		__kcsan_check_access(x, s->object_size,
2098				     KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT);
2099
2100	if (kfence_free(x))
2101		return false;
2102
2103	/*
2104	 * As memory initialization might be integrated into KASAN,
2105	 * kasan_slab_free and initialization memset's must be
2106	 * kept together to avoid discrepancies in behavior.
2107	 *
2108	 * The initialization memset's clear the object and the metadata,
2109	 * but don't touch the SLAB redzone.
 
 
 
2110	 */
2111	if (unlikely(init)) {
2112		int rsize;
 
2113
 
2114		if (!kasan_has_integrated_init())
2115			memset(kasan_reset_tag(x), 0, s->object_size);
2116		rsize = (s->flags & SLAB_RED_ZONE) ? s->red_left_pad : 0;
2117		memset((char *)kasan_reset_tag(x) + s->inuse, 0,
2118		       s->size - s->inuse - rsize);
2119	}
2120	/* KASAN might put x into memory quarantine, delaying its reuse. */
2121	return !kasan_slab_free(s, x, init);
2122}
2123
2124static inline bool slab_free_freelist_hook(struct kmem_cache *s,
2125					   void **head, void **tail,
2126					   int *cnt)
2127{
2128
2129	void *object;
2130	void *next = *head;
2131	void *old_tail = *tail;
2132	bool init;
2133
2134	if (is_kfence_address(next)) {
2135		slab_free_hook(s, next, false);
2136		return false;
2137	}
2138
2139	/* Head and tail of the reconstructed freelist */
2140	*head = NULL;
2141	*tail = NULL;
2142
2143	init = slab_want_init_on_free(s);
2144
2145	do {
2146		object = next;
2147		next = get_freepointer(s, object);
2148
2149		/* If object's reuse doesn't have to be delayed */
2150		if (likely(slab_free_hook(s, object, init))) {
2151			/* Move object to the new freelist */
2152			set_freepointer(s, object, *head);
2153			*head = object;
2154			if (!*tail)
2155				*tail = object;
2156		} else {
2157			/*
2158			 * Adjust the reconstructed freelist depth
2159			 * accordingly if object's reuse is delayed.
2160			 */
2161			--(*cnt);
2162		}
2163	} while (object != old_tail);
2164
2165	return *head != NULL;
2166}
2167
2168static void *setup_object(struct kmem_cache *s, void *object)
2169{
2170	setup_object_debug(s, object);
2171	object = kasan_init_slab_obj(s, object);
2172	if (unlikely(s->ctor)) {
2173		kasan_unpoison_new_object(s, object);
2174		s->ctor(object);
2175		kasan_poison_new_object(s, object);
2176	}
2177	return object;
2178}
2179
2180/*
2181 * Slab allocation and freeing
2182 */
2183static inline struct slab *alloc_slab_page(gfp_t flags, int node,
2184		struct kmem_cache_order_objects oo)
2185{
2186	struct folio *folio;
2187	struct slab *slab;
2188	unsigned int order = oo_order(oo);
2189
2190	folio = (struct folio *)alloc_pages_node(node, flags, order);
2191	if (!folio)
2192		return NULL;
2193
2194	slab = folio_slab(folio);
2195	__folio_set_slab(folio);
2196	/* Make the flag visible before any changes to folio->mapping */
2197	smp_wmb();
2198	if (folio_is_pfmemalloc(folio))
2199		slab_set_pfmemalloc(slab);
2200
2201	return slab;
2202}
2203
2204#ifdef CONFIG_SLAB_FREELIST_RANDOM
2205/* Pre-initialize the random sequence cache */
2206static int init_cache_random_seq(struct kmem_cache *s)
2207{
2208	unsigned int count = oo_objects(s->oo);
2209	int err;
2210
2211	/* Bailout if already initialised */
2212	if (s->random_seq)
2213		return 0;
2214
2215	err = cache_random_seq_create(s, count, GFP_KERNEL);
2216	if (err) {
2217		pr_err("SLUB: Unable to initialize free list for %s\n",
2218			s->name);
2219		return err;
2220	}
2221
2222	/* Transform to an offset on the set of pages */
2223	if (s->random_seq) {
2224		unsigned int i;
2225
2226		for (i = 0; i < count; i++)
2227			s->random_seq[i] *= s->size;
2228	}
2229	return 0;
2230}
2231
2232/* Initialize each random sequence freelist per cache */
2233static void __init init_freelist_randomization(void)
2234{
2235	struct kmem_cache *s;
2236
2237	mutex_lock(&slab_mutex);
2238
2239	list_for_each_entry(s, &slab_caches, list)
2240		init_cache_random_seq(s);
2241
2242	mutex_unlock(&slab_mutex);
2243}
2244
2245/* Get the next entry on the pre-computed freelist randomized */
2246static void *next_freelist_entry(struct kmem_cache *s, struct slab *slab,
2247				unsigned long *pos, void *start,
2248				unsigned long page_limit,
2249				unsigned long freelist_count)
2250{
2251	unsigned int idx;
2252
2253	/*
2254	 * If the target page allocation failed, the number of objects on the
2255	 * page might be smaller than the usual size defined by the cache.
2256	 */
2257	do {
2258		idx = s->random_seq[*pos];
2259		*pos += 1;
2260		if (*pos >= freelist_count)
2261			*pos = 0;
2262	} while (unlikely(idx >= page_limit));
2263
2264	return (char *)start + idx;
2265}
2266
2267/* Shuffle the single linked freelist based on a random pre-computed sequence */
2268static bool shuffle_freelist(struct kmem_cache *s, struct slab *slab)
2269{
2270	void *start;
2271	void *cur;
2272	void *next;
2273	unsigned long idx, pos, page_limit, freelist_count;
2274
2275	if (slab->objects < 2 || !s->random_seq)
2276		return false;
2277
2278	freelist_count = oo_objects(s->oo);
2279	pos = get_random_u32_below(freelist_count);
2280
2281	page_limit = slab->objects * s->size;
2282	start = fixup_red_left(s, slab_address(slab));
2283
2284	/* First entry is used as the base of the freelist */
2285	cur = next_freelist_entry(s, slab, &pos, start, page_limit,
2286				freelist_count);
2287	cur = setup_object(s, cur);
2288	slab->freelist = cur;
2289
2290	for (idx = 1; idx < slab->objects; idx++) {
2291		next = next_freelist_entry(s, slab, &pos, start, page_limit,
2292			freelist_count);
2293		next = setup_object(s, next);
2294		set_freepointer(s, cur, next);
2295		cur = next;
2296	}
2297	set_freepointer(s, cur, NULL);
2298
2299	return true;
2300}
2301#else
2302static inline int init_cache_random_seq(struct kmem_cache *s)
2303{
2304	return 0;
2305}
2306static inline void init_freelist_randomization(void) { }
2307static inline bool shuffle_freelist(struct kmem_cache *s, struct slab *slab)
2308{
2309	return false;
2310}
2311#endif /* CONFIG_SLAB_FREELIST_RANDOM */
2312
2313static __always_inline void account_slab(struct slab *slab, int order,
2314					 struct kmem_cache *s, gfp_t gfp)
2315{
2316	if (memcg_kmem_online() && (s->flags & SLAB_ACCOUNT))
2317		memcg_alloc_slab_cgroups(slab, s, gfp, true);
2318
2319	mod_node_page_state(slab_pgdat(slab), cache_vmstat_idx(s),
2320			    PAGE_SIZE << order);
2321}
2322
2323static __always_inline void unaccount_slab(struct slab *slab, int order,
2324					   struct kmem_cache *s)
2325{
2326	if (memcg_kmem_online())
2327		memcg_free_slab_cgroups(slab);
2328
2329	mod_node_page_state(slab_pgdat(slab), cache_vmstat_idx(s),
2330			    -(PAGE_SIZE << order));
2331}
2332
2333static struct slab *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
2334{
2335	struct slab *slab;
2336	struct kmem_cache_order_objects oo = s->oo;
2337	gfp_t alloc_gfp;
2338	void *start, *p, *next;
2339	int idx;
2340	bool shuffle;
2341
2342	flags &= gfp_allowed_mask;
2343
2344	flags |= s->allocflags;
2345
2346	/*
2347	 * Let the initial higher-order allocation fail under memory pressure
2348	 * so we fall-back to the minimum order allocation.
2349	 */
2350	alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
2351	if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
2352		alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~__GFP_RECLAIM;
2353
2354	slab = alloc_slab_page(alloc_gfp, node, oo);
2355	if (unlikely(!slab)) {
2356		oo = s->min;
2357		alloc_gfp = flags;
2358		/*
2359		 * Allocation may have failed due to fragmentation.
2360		 * Try a lower order alloc if possible
2361		 */
2362		slab = alloc_slab_page(alloc_gfp, node, oo);
2363		if (unlikely(!slab))
2364			return NULL;
2365		stat(s, ORDER_FALLBACK);
2366	}
2367
2368	slab->objects = oo_objects(oo);
2369	slab->inuse = 0;
2370	slab->frozen = 0;
2371
2372	account_slab(slab, oo_order(oo), s, flags);
2373
2374	slab->slab_cache = s;
2375
2376	kasan_poison_slab(slab);
2377
2378	start = slab_address(slab);
2379
2380	setup_slab_debug(s, slab, start);
2381
2382	shuffle = shuffle_freelist(s, slab);
2383
2384	if (!shuffle) {
2385		start = fixup_red_left(s, start);
2386		start = setup_object(s, start);
2387		slab->freelist = start;
2388		for (idx = 0, p = start; idx < slab->objects - 1; idx++) {
2389			next = p + s->size;
2390			next = setup_object(s, next);
2391			set_freepointer(s, p, next);
2392			p = next;
2393		}
2394		set_freepointer(s, p, NULL);
2395	}
2396
2397	return slab;
2398}
2399
2400static struct slab *new_slab(struct kmem_cache *s, gfp_t flags, int node)
2401{
2402	if (unlikely(flags & GFP_SLAB_BUG_MASK))
2403		flags = kmalloc_fix_flags(flags);
2404
2405	WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO));
2406
2407	return allocate_slab(s,
2408		flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
2409}
2410
2411static void __free_slab(struct kmem_cache *s, struct slab *slab)
2412{
2413	struct folio *folio = slab_folio(slab);
2414	int order = folio_order(folio);
2415	int pages = 1 << order;
2416
2417	__slab_clear_pfmemalloc(slab);
2418	folio->mapping = NULL;
2419	/* Make the mapping reset visible before clearing the flag */
2420	smp_wmb();
2421	__folio_clear_slab(folio);
2422	mm_account_reclaimed_pages(pages);
2423	unaccount_slab(slab, order, s);
2424	__free_pages(&folio->page, order);
2425}
2426
2427static void rcu_free_slab(struct rcu_head *h)
2428{
2429	struct slab *slab = container_of(h, struct slab, rcu_head);
2430
2431	__free_slab(slab->slab_cache, slab);
2432}
2433
2434static void free_slab(struct kmem_cache *s, struct slab *slab)
2435{
2436	if (kmem_cache_debug_flags(s, SLAB_CONSISTENCY_CHECKS)) {
2437		void *p;
2438
2439		slab_pad_check(s, slab);
2440		for_each_object(p, s, slab_address(slab), slab->objects)
2441			check_object(s, slab, p, SLUB_RED_INACTIVE);
2442	}
2443
2444	if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU))
2445		call_rcu(&slab->rcu_head, rcu_free_slab);
2446	else
2447		__free_slab(s, slab);
2448}
2449
2450static void discard_slab(struct kmem_cache *s, struct slab *slab)
2451{
2452	dec_slabs_node(s, slab_nid(slab), slab->objects);
2453	free_slab(s, slab);
2454}
2455
2456/*
2457 * SLUB reuses PG_workingset bit to keep track of whether it's on
2458 * the per-node partial list.
2459 */
2460static inline bool slab_test_node_partial(const struct slab *slab)
2461{
2462	return folio_test_workingset((struct folio *)slab_folio(slab));
2463}
2464
2465static inline void slab_set_node_partial(struct slab *slab)
2466{
2467	set_bit(PG_workingset, folio_flags(slab_folio(slab), 0));
2468}
2469
2470static inline void slab_clear_node_partial(struct slab *slab)
2471{
2472	clear_bit(PG_workingset, folio_flags(slab_folio(slab), 0));
2473}
2474
2475/*
2476 * Management of partially allocated slabs.
2477 */
2478static inline void
2479__add_partial(struct kmem_cache_node *n, struct slab *slab, int tail)
2480{
2481	n->nr_partial++;
2482	if (tail == DEACTIVATE_TO_TAIL)
2483		list_add_tail(&slab->slab_list, &n->partial);
2484	else
2485		list_add(&slab->slab_list, &n->partial);
2486	slab_set_node_partial(slab);
2487}
2488
2489static inline void add_partial(struct kmem_cache_node *n,
2490				struct slab *slab, int tail)
2491{
2492	lockdep_assert_held(&n->list_lock);
2493	__add_partial(n, slab, tail);
2494}
2495
2496static inline void remove_partial(struct kmem_cache_node *n,
2497					struct slab *slab)
2498{
2499	lockdep_assert_held(&n->list_lock);
2500	list_del(&slab->slab_list);
2501	slab_clear_node_partial(slab);
2502	n->nr_partial--;
2503}
2504
2505/*
2506 * Called only for kmem_cache_debug() caches instead of remove_partial(), with a
2507 * slab from the n->partial list. Remove only a single object from the slab, do
2508 * the alloc_debug_processing() checks and leave the slab on the list, or move
2509 * it to full list if it was the last free object.
2510 */
2511static void *alloc_single_from_partial(struct kmem_cache *s,
2512		struct kmem_cache_node *n, struct slab *slab, int orig_size)
2513{
2514	void *object;
2515
2516	lockdep_assert_held(&n->list_lock);
2517
2518	object = slab->freelist;
2519	slab->freelist = get_freepointer(s, object);
2520	slab->inuse++;
2521
2522	if (!alloc_debug_processing(s, slab, object, orig_size)) {
2523		remove_partial(n, slab);
2524		return NULL;
2525	}
2526
2527	if (slab->inuse == slab->objects) {
2528		remove_partial(n, slab);
2529		add_full(s, n, slab);
2530	}
2531
2532	return object;
2533}
2534
2535/*
2536 * Called only for kmem_cache_debug() caches to allocate from a freshly
2537 * allocated slab. Allocate a single object instead of whole freelist
2538 * and put the slab to the partial (or full) list.
2539 */
2540static void *alloc_single_from_new_slab(struct kmem_cache *s,
2541					struct slab *slab, int orig_size)
2542{
2543	int nid = slab_nid(slab);
2544	struct kmem_cache_node *n = get_node(s, nid);
2545	unsigned long flags;
2546	void *object;
2547
2548
2549	object = slab->freelist;
2550	slab->freelist = get_freepointer(s, object);
2551	slab->inuse = 1;
2552
2553	if (!alloc_debug_processing(s, slab, object, orig_size))
2554		/*
2555		 * It's not really expected that this would fail on a
2556		 * freshly allocated slab, but a concurrent memory
2557		 * corruption in theory could cause that.
2558		 */
2559		return NULL;
2560
2561	spin_lock_irqsave(&n->list_lock, flags);
2562
2563	if (slab->inuse == slab->objects)
2564		add_full(s, n, slab);
2565	else
2566		add_partial(n, slab, DEACTIVATE_TO_HEAD);
2567
2568	inc_slabs_node(s, nid, slab->objects);
2569	spin_unlock_irqrestore(&n->list_lock, flags);
2570
2571	return object;
2572}
2573
2574#ifdef CONFIG_SLUB_CPU_PARTIAL
2575static void put_cpu_partial(struct kmem_cache *s, struct slab *slab, int drain);
2576#else
2577static inline void put_cpu_partial(struct kmem_cache *s, struct slab *slab,
2578				   int drain) { }
2579#endif
2580static inline bool pfmemalloc_match(struct slab *slab, gfp_t gfpflags);
2581
2582/*
2583 * Try to allocate a partial slab from a specific node.
2584 */
2585static struct slab *get_partial_node(struct kmem_cache *s,
2586				     struct kmem_cache_node *n,
2587				     struct partial_context *pc)
2588{
2589	struct slab *slab, *slab2, *partial = NULL;
2590	unsigned long flags;
2591	unsigned int partial_slabs = 0;
2592
2593	/*
2594	 * Racy check. If we mistakenly see no partial slabs then we
2595	 * just allocate an empty slab. If we mistakenly try to get a
2596	 * partial slab and there is none available then get_partial()
2597	 * will return NULL.
2598	 */
2599	if (!n || !n->nr_partial)
2600		return NULL;
2601
2602	spin_lock_irqsave(&n->list_lock, flags);
2603	list_for_each_entry_safe(slab, slab2, &n->partial, slab_list) {
2604		if (!pfmemalloc_match(slab, pc->flags))
2605			continue;
2606
2607		if (IS_ENABLED(CONFIG_SLUB_TINY) || kmem_cache_debug(s)) {
2608			void *object = alloc_single_from_partial(s, n, slab,
2609							pc->orig_size);
2610			if (object) {
2611				partial = slab;
2612				pc->object = object;
2613				break;
2614			}
2615			continue;
2616		}
2617
2618		remove_partial(n, slab);
2619
2620		if (!partial) {
2621			partial = slab;
2622			stat(s, ALLOC_FROM_PARTIAL);
2623		} else {
2624			put_cpu_partial(s, slab, 0);
2625			stat(s, CPU_PARTIAL_NODE);
2626			partial_slabs++;
2627		}
2628#ifdef CONFIG_SLUB_CPU_PARTIAL
2629		if (!kmem_cache_has_cpu_partial(s)
2630			|| partial_slabs > s->cpu_partial_slabs / 2)
2631			break;
2632#else
2633		break;
2634#endif
2635
2636	}
2637	spin_unlock_irqrestore(&n->list_lock, flags);
2638	return partial;
2639}
2640
2641/*
2642 * Get a slab from somewhere. Search in increasing NUMA distances.
2643 */
2644static struct slab *get_any_partial(struct kmem_cache *s,
2645				    struct partial_context *pc)
2646{
2647#ifdef CONFIG_NUMA
2648	struct zonelist *zonelist;
2649	struct zoneref *z;
2650	struct zone *zone;
2651	enum zone_type highest_zoneidx = gfp_zone(pc->flags);
2652	struct slab *slab;
2653	unsigned int cpuset_mems_cookie;
2654
2655	/*
2656	 * The defrag ratio allows a configuration of the tradeoffs between
2657	 * inter node defragmentation and node local allocations. A lower
2658	 * defrag_ratio increases the tendency to do local allocations
2659	 * instead of attempting to obtain partial slabs from other nodes.
2660	 *
2661	 * If the defrag_ratio is set to 0 then kmalloc() always
2662	 * returns node local objects. If the ratio is higher then kmalloc()
2663	 * may return off node objects because partial slabs are obtained
2664	 * from other nodes and filled up.
2665	 *
2666	 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
2667	 * (which makes defrag_ratio = 1000) then every (well almost)
2668	 * allocation will first attempt to defrag slab caches on other nodes.
2669	 * This means scanning over all nodes to look for partial slabs which
2670	 * may be expensive if we do it every time we are trying to find a slab
2671	 * with available objects.
2672	 */
2673	if (!s->remote_node_defrag_ratio ||
2674			get_cycles() % 1024 > s->remote_node_defrag_ratio)
2675		return NULL;
2676
2677	do {
2678		cpuset_mems_cookie = read_mems_allowed_begin();
2679		zonelist = node_zonelist(mempolicy_slab_node(), pc->flags);
2680		for_each_zone_zonelist(zone, z, zonelist, highest_zoneidx) {
2681			struct kmem_cache_node *n;
2682
2683			n = get_node(s, zone_to_nid(zone));
2684
2685			if (n && cpuset_zone_allowed(zone, pc->flags) &&
2686					n->nr_partial > s->min_partial) {
2687				slab = get_partial_node(s, n, pc);
2688				if (slab) {
2689					/*
2690					 * Don't check read_mems_allowed_retry()
2691					 * here - if mems_allowed was updated in
2692					 * parallel, that was a harmless race
2693					 * between allocation and the cpuset
2694					 * update
2695					 */
2696					return slab;
2697				}
2698			}
2699		}
2700	} while (read_mems_allowed_retry(cpuset_mems_cookie));
2701#endif	/* CONFIG_NUMA */
2702	return NULL;
2703}
2704
2705/*
2706 * Get a partial slab, lock it and return it.
2707 */
2708static struct slab *get_partial(struct kmem_cache *s, int node,
2709				struct partial_context *pc)
2710{
2711	struct slab *slab;
2712	int searchnode = node;
2713
2714	if (node == NUMA_NO_NODE)
2715		searchnode = numa_mem_id();
2716
2717	slab = get_partial_node(s, get_node(s, searchnode), pc);
2718	if (slab || node != NUMA_NO_NODE)
2719		return slab;
2720
2721	return get_any_partial(s, pc);
2722}
2723
2724#ifndef CONFIG_SLUB_TINY
2725
2726#ifdef CONFIG_PREEMPTION
2727/*
2728 * Calculate the next globally unique transaction for disambiguation
2729 * during cmpxchg. The transactions start with the cpu number and are then
2730 * incremented by CONFIG_NR_CPUS.
2731 */
2732#define TID_STEP  roundup_pow_of_two(CONFIG_NR_CPUS)
2733#else
2734/*
2735 * No preemption supported therefore also no need to check for
2736 * different cpus.
2737 */
2738#define TID_STEP 1
2739#endif /* CONFIG_PREEMPTION */
2740
2741static inline unsigned long next_tid(unsigned long tid)
2742{
2743	return tid + TID_STEP;
2744}
2745
2746#ifdef SLUB_DEBUG_CMPXCHG
2747static inline unsigned int tid_to_cpu(unsigned long tid)
2748{
2749	return tid % TID_STEP;
2750}
2751
2752static inline unsigned long tid_to_event(unsigned long tid)
2753{
2754	return tid / TID_STEP;
2755}
2756#endif
2757
2758static inline unsigned int init_tid(int cpu)
2759{
2760	return cpu;
2761}
2762
2763static inline void note_cmpxchg_failure(const char *n,
2764		const struct kmem_cache *s, unsigned long tid)
2765{
2766#ifdef SLUB_DEBUG_CMPXCHG
2767	unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
2768
2769	pr_info("%s %s: cmpxchg redo ", n, s->name);
2770
2771#ifdef CONFIG_PREEMPTION
2772	if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
2773		pr_warn("due to cpu change %d -> %d\n",
2774			tid_to_cpu(tid), tid_to_cpu(actual_tid));
2775	else
2776#endif
2777	if (tid_to_event(tid) != tid_to_event(actual_tid))
2778		pr_warn("due to cpu running other code. Event %ld->%ld\n",
2779			tid_to_event(tid), tid_to_event(actual_tid));
2780	else
2781		pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2782			actual_tid, tid, next_tid(tid));
2783#endif
2784	stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
2785}
2786
2787static void init_kmem_cache_cpus(struct kmem_cache *s)
2788{
2789	int cpu;
2790	struct kmem_cache_cpu *c;
2791
2792	for_each_possible_cpu(cpu) {
2793		c = per_cpu_ptr(s->cpu_slab, cpu);
2794		local_lock_init(&c->lock);
2795		c->tid = init_tid(cpu);
2796	}
2797}
2798
2799/*
2800 * Finishes removing the cpu slab. Merges cpu's freelist with slab's freelist,
2801 * unfreezes the slabs and puts it on the proper list.
2802 * Assumes the slab has been already safely taken away from kmem_cache_cpu
2803 * by the caller.
2804 */
2805static void deactivate_slab(struct kmem_cache *s, struct slab *slab,
2806			    void *freelist)
2807{
2808	struct kmem_cache_node *n = get_node(s, slab_nid(slab));
2809	int free_delta = 0;
2810	void *nextfree, *freelist_iter, *freelist_tail;
2811	int tail = DEACTIVATE_TO_HEAD;
2812	unsigned long flags = 0;
2813	struct slab new;
2814	struct slab old;
2815
2816	if (slab->freelist) {
2817		stat(s, DEACTIVATE_REMOTE_FREES);
2818		tail = DEACTIVATE_TO_TAIL;
2819	}
2820
2821	/*
2822	 * Stage one: Count the objects on cpu's freelist as free_delta and
2823	 * remember the last object in freelist_tail for later splicing.
2824	 */
2825	freelist_tail = NULL;
2826	freelist_iter = freelist;
2827	while (freelist_iter) {
2828		nextfree = get_freepointer(s, freelist_iter);
2829
2830		/*
2831		 * If 'nextfree' is invalid, it is possible that the object at
2832		 * 'freelist_iter' is already corrupted.  So isolate all objects
2833		 * starting at 'freelist_iter' by skipping them.
2834		 */
2835		if (freelist_corrupted(s, slab, &freelist_iter, nextfree))
2836			break;
2837
2838		freelist_tail = freelist_iter;
2839		free_delta++;
2840
2841		freelist_iter = nextfree;
2842	}
2843
2844	/*
2845	 * Stage two: Unfreeze the slab while splicing the per-cpu
2846	 * freelist to the head of slab's freelist.
2847	 */
2848	do {
2849		old.freelist = READ_ONCE(slab->freelist);
2850		old.counters = READ_ONCE(slab->counters);
2851		VM_BUG_ON(!old.frozen);
2852
2853		/* Determine target state of the slab */
2854		new.counters = old.counters;
2855		new.frozen = 0;
2856		if (freelist_tail) {
2857			new.inuse -= free_delta;
2858			set_freepointer(s, freelist_tail, old.freelist);
2859			new.freelist = freelist;
2860		} else {
2861			new.freelist = old.freelist;
2862		}
2863	} while (!slab_update_freelist(s, slab,
2864		old.freelist, old.counters,
2865		new.freelist, new.counters,
2866		"unfreezing slab"));
2867
2868	/*
2869	 * Stage three: Manipulate the slab list based on the updated state.
2870	 */
2871	if (!new.inuse && n->nr_partial >= s->min_partial) {
2872		stat(s, DEACTIVATE_EMPTY);
2873		discard_slab(s, slab);
2874		stat(s, FREE_SLAB);
2875	} else if (new.freelist) {
2876		spin_lock_irqsave(&n->list_lock, flags);
2877		add_partial(n, slab, tail);
2878		spin_unlock_irqrestore(&n->list_lock, flags);
2879		stat(s, tail);
2880	} else {
2881		stat(s, DEACTIVATE_FULL);
2882	}
2883}
2884
2885#ifdef CONFIG_SLUB_CPU_PARTIAL
2886static void __put_partials(struct kmem_cache *s, struct slab *partial_slab)
2887{
2888	struct kmem_cache_node *n = NULL, *n2 = NULL;
2889	struct slab *slab, *slab_to_discard = NULL;
2890	unsigned long flags = 0;
2891
2892	while (partial_slab) {
2893		slab = partial_slab;
2894		partial_slab = slab->next;
2895
2896		n2 = get_node(s, slab_nid(slab));
2897		if (n != n2) {
2898			if (n)
2899				spin_unlock_irqrestore(&n->list_lock, flags);
2900
2901			n = n2;
2902			spin_lock_irqsave(&n->list_lock, flags);
2903		}
2904
2905		if (unlikely(!slab->inuse && n->nr_partial >= s->min_partial)) {
2906			slab->next = slab_to_discard;
2907			slab_to_discard = slab;
2908		} else {
2909			add_partial(n, slab, DEACTIVATE_TO_TAIL);
2910			stat(s, FREE_ADD_PARTIAL);
2911		}
2912	}
2913
2914	if (n)
2915		spin_unlock_irqrestore(&n->list_lock, flags);
2916
2917	while (slab_to_discard) {
2918		slab = slab_to_discard;
2919		slab_to_discard = slab_to_discard->next;
2920
2921		stat(s, DEACTIVATE_EMPTY);
2922		discard_slab(s, slab);
2923		stat(s, FREE_SLAB);
2924	}
2925}
2926
2927/*
2928 * Put all the cpu partial slabs to the node partial list.
2929 */
2930static void put_partials(struct kmem_cache *s)
2931{
2932	struct slab *partial_slab;
2933	unsigned long flags;
2934
2935	local_lock_irqsave(&s->cpu_slab->lock, flags);
2936	partial_slab = this_cpu_read(s->cpu_slab->partial);
2937	this_cpu_write(s->cpu_slab->partial, NULL);
2938	local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2939
2940	if (partial_slab)
2941		__put_partials(s, partial_slab);
2942}
2943
2944static void put_partials_cpu(struct kmem_cache *s,
2945			     struct kmem_cache_cpu *c)
2946{
2947	struct slab *partial_slab;
2948
2949	partial_slab = slub_percpu_partial(c);
2950	c->partial = NULL;
2951
2952	if (partial_slab)
2953		__put_partials(s, partial_slab);
2954}
2955
2956/*
2957 * Put a slab into a partial slab slot if available.
2958 *
2959 * If we did not find a slot then simply move all the partials to the
2960 * per node partial list.
2961 */
2962static void put_cpu_partial(struct kmem_cache *s, struct slab *slab, int drain)
2963{
2964	struct slab *oldslab;
2965	struct slab *slab_to_put = NULL;
2966	unsigned long flags;
2967	int slabs = 0;
2968
2969	local_lock_irqsave(&s->cpu_slab->lock, flags);
2970
2971	oldslab = this_cpu_read(s->cpu_slab->partial);
2972
2973	if (oldslab) {
2974		if (drain && oldslab->slabs >= s->cpu_partial_slabs) {
2975			/*
2976			 * Partial array is full. Move the existing set to the
2977			 * per node partial list. Postpone the actual unfreezing
2978			 * outside of the critical section.
2979			 */
2980			slab_to_put = oldslab;
2981			oldslab = NULL;
2982		} else {
2983			slabs = oldslab->slabs;
2984		}
2985	}
2986
2987	slabs++;
2988
2989	slab->slabs = slabs;
2990	slab->next = oldslab;
2991
2992	this_cpu_write(s->cpu_slab->partial, slab);
2993
2994	local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2995
2996	if (slab_to_put) {
2997		__put_partials(s, slab_to_put);
2998		stat(s, CPU_PARTIAL_DRAIN);
2999	}
3000}
3001
3002#else	/* CONFIG_SLUB_CPU_PARTIAL */
3003
3004static inline void put_partials(struct kmem_cache *s) { }
3005static inline void put_partials_cpu(struct kmem_cache *s,
3006				    struct kmem_cache_cpu *c) { }
3007
3008#endif	/* CONFIG_SLUB_CPU_PARTIAL */
3009
3010static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
3011{
3012	unsigned long flags;
3013	struct slab *slab;
3014	void *freelist;
3015
3016	local_lock_irqsave(&s->cpu_slab->lock, flags);
3017
3018	slab = c->slab;
3019	freelist = c->freelist;
3020
3021	c->slab = NULL;
3022	c->freelist = NULL;
3023	c->tid = next_tid(c->tid);
3024
3025	local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3026
3027	if (slab) {
3028		deactivate_slab(s, slab, freelist);
3029		stat(s, CPUSLAB_FLUSH);
3030	}
3031}
3032
3033static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
3034{
3035	struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
3036	void *freelist = c->freelist;
3037	struct slab *slab = c->slab;
3038
3039	c->slab = NULL;
3040	c->freelist = NULL;
3041	c->tid = next_tid(c->tid);
3042
3043	if (slab) {
3044		deactivate_slab(s, slab, freelist);
3045		stat(s, CPUSLAB_FLUSH);
3046	}
3047
3048	put_partials_cpu(s, c);
3049}
3050
3051struct slub_flush_work {
3052	struct work_struct work;
3053	struct kmem_cache *s;
3054	bool skip;
3055};
3056
3057/*
3058 * Flush cpu slab.
3059 *
3060 * Called from CPU work handler with migration disabled.
3061 */
3062static void flush_cpu_slab(struct work_struct *w)
3063{
3064	struct kmem_cache *s;
3065	struct kmem_cache_cpu *c;
3066	struct slub_flush_work *sfw;
3067
3068	sfw = container_of(w, struct slub_flush_work, work);
3069
3070	s = sfw->s;
3071	c = this_cpu_ptr(s->cpu_slab);
3072
3073	if (c->slab)
3074		flush_slab(s, c);
3075
3076	put_partials(s);
3077}
3078
3079static bool has_cpu_slab(int cpu, struct kmem_cache *s)
3080{
3081	struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
3082
3083	return c->slab || slub_percpu_partial(c);
3084}
3085
3086static DEFINE_MUTEX(flush_lock);
3087static DEFINE_PER_CPU(struct slub_flush_work, slub_flush);
3088
3089static void flush_all_cpus_locked(struct kmem_cache *s)
3090{
3091	struct slub_flush_work *sfw;
3092	unsigned int cpu;
3093
3094	lockdep_assert_cpus_held();
3095	mutex_lock(&flush_lock);
3096
3097	for_each_online_cpu(cpu) {
3098		sfw = &per_cpu(slub_flush, cpu);
3099		if (!has_cpu_slab(cpu, s)) {
3100			sfw->skip = true;
3101			continue;
3102		}
3103		INIT_WORK(&sfw->work, flush_cpu_slab);
3104		sfw->skip = false;
3105		sfw->s = s;
3106		queue_work_on(cpu, flushwq, &sfw->work);
3107	}
3108
3109	for_each_online_cpu(cpu) {
3110		sfw = &per_cpu(slub_flush, cpu);
3111		if (sfw->skip)
3112			continue;
3113		flush_work(&sfw->work);
3114	}
3115
3116	mutex_unlock(&flush_lock);
3117}
3118
3119static void flush_all(struct kmem_cache *s)
3120{
3121	cpus_read_lock();
3122	flush_all_cpus_locked(s);
3123	cpus_read_unlock();
3124}
3125
3126/*
3127 * Use the cpu notifier to insure that the cpu slabs are flushed when
3128 * necessary.
3129 */
3130static int slub_cpu_dead(unsigned int cpu)
3131{
3132	struct kmem_cache *s;
3133
3134	mutex_lock(&slab_mutex);
3135	list_for_each_entry(s, &slab_caches, list)
3136		__flush_cpu_slab(s, cpu);
3137	mutex_unlock(&slab_mutex);
3138	return 0;
3139}
3140
3141#else /* CONFIG_SLUB_TINY */
3142static inline void flush_all_cpus_locked(struct kmem_cache *s) { }
3143static inline void flush_all(struct kmem_cache *s) { }
3144static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu) { }
3145static inline int slub_cpu_dead(unsigned int cpu) { return 0; }
3146#endif /* CONFIG_SLUB_TINY */
3147
3148/*
3149 * Check if the objects in a per cpu structure fit numa
3150 * locality expectations.
3151 */
3152static inline int node_match(struct slab *slab, int node)
3153{
3154#ifdef CONFIG_NUMA
3155	if (node != NUMA_NO_NODE && slab_nid(slab) != node)
3156		return 0;
3157#endif
3158	return 1;
3159}
3160
3161#ifdef CONFIG_SLUB_DEBUG
3162static int count_free(struct slab *slab)
3163{
3164	return slab->objects - slab->inuse;
3165}
3166
3167static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
3168{
3169	return atomic_long_read(&n->total_objects);
3170}
3171
3172/* Supports checking bulk free of a constructed freelist */
3173static inline bool free_debug_processing(struct kmem_cache *s,
3174	struct slab *slab, void *head, void *tail, int *bulk_cnt,
3175	unsigned long addr, depot_stack_handle_t handle)
3176{
3177	bool checks_ok = false;
3178	void *object = head;
3179	int cnt = 0;
3180
3181	if (s->flags & SLAB_CONSISTENCY_CHECKS) {
3182		if (!check_slab(s, slab))
3183			goto out;
3184	}
3185
3186	if (slab->inuse < *bulk_cnt) {
3187		slab_err(s, slab, "Slab has %d allocated objects but %d are to be freed\n",
3188			 slab->inuse, *bulk_cnt);
3189		goto out;
3190	}
3191
3192next_object:
3193
3194	if (++cnt > *bulk_cnt)
3195		goto out_cnt;
3196
3197	if (s->flags & SLAB_CONSISTENCY_CHECKS) {
3198		if (!free_consistency_checks(s, slab, object, addr))
3199			goto out;
3200	}
3201
3202	if (s->flags & SLAB_STORE_USER)
3203		set_track_update(s, object, TRACK_FREE, addr, handle);
3204	trace(s, slab, object, 0);
3205	/* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
3206	init_object(s, object, SLUB_RED_INACTIVE);
3207
3208	/* Reached end of constructed freelist yet? */
3209	if (object != tail) {
3210		object = get_freepointer(s, object);
3211		goto next_object;
3212	}
3213	checks_ok = true;
3214
3215out_cnt:
3216	if (cnt != *bulk_cnt) {
3217		slab_err(s, slab, "Bulk free expected %d objects but found %d\n",
3218			 *bulk_cnt, cnt);
3219		*bulk_cnt = cnt;
3220	}
3221
3222out:
3223
3224	if (!checks_ok)
3225		slab_fix(s, "Object at 0x%p not freed", object);
3226
3227	return checks_ok;
3228}
3229#endif /* CONFIG_SLUB_DEBUG */
3230
3231#if defined(CONFIG_SLUB_DEBUG) || defined(SLAB_SUPPORTS_SYSFS)
3232static unsigned long count_partial(struct kmem_cache_node *n,
3233					int (*get_count)(struct slab *))
3234{
3235	unsigned long flags;
3236	unsigned long x = 0;
3237	struct slab *slab;
3238
3239	spin_lock_irqsave(&n->list_lock, flags);
3240	list_for_each_entry(slab, &n->partial, slab_list)
3241		x += get_count(slab);
3242	spin_unlock_irqrestore(&n->list_lock, flags);
3243	return x;
3244}
3245#endif /* CONFIG_SLUB_DEBUG || SLAB_SUPPORTS_SYSFS */
3246
3247#ifdef CONFIG_SLUB_DEBUG
3248static noinline void
3249slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
3250{
3251	static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
3252				      DEFAULT_RATELIMIT_BURST);
3253	int node;
3254	struct kmem_cache_node *n;
3255
3256	if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
3257		return;
3258
3259	pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
3260		nid, gfpflags, &gfpflags);
3261	pr_warn("  cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
3262		s->name, s->object_size, s->size, oo_order(s->oo),
3263		oo_order(s->min));
3264
3265	if (oo_order(s->min) > get_order(s->object_size))
3266		pr_warn("  %s debugging increased min order, use slub_debug=O to disable.\n",
3267			s->name);
3268
3269	for_each_kmem_cache_node(s, node, n) {
3270		unsigned long nr_slabs;
3271		unsigned long nr_objs;
3272		unsigned long nr_free;
3273
3274		nr_free  = count_partial(n, count_free);
3275		nr_slabs = node_nr_slabs(n);
3276		nr_objs  = node_nr_objs(n);
3277
3278		pr_warn("  node %d: slabs: %ld, objs: %ld, free: %ld\n",
3279			node, nr_slabs, nr_objs, nr_free);
3280	}
3281}
3282#else /* CONFIG_SLUB_DEBUG */
3283static inline void
3284slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid) { }
3285#endif
3286
3287static inline bool pfmemalloc_match(struct slab *slab, gfp_t gfpflags)
3288{
3289	if (unlikely(slab_test_pfmemalloc(slab)))
3290		return gfp_pfmemalloc_allowed(gfpflags);
3291
3292	return true;
3293}
3294
3295#ifndef CONFIG_SLUB_TINY
3296static inline bool
3297__update_cpu_freelist_fast(struct kmem_cache *s,
3298			   void *freelist_old, void *freelist_new,
3299			   unsigned long tid)
3300{
3301	freelist_aba_t old = { .freelist = freelist_old, .counter = tid };
3302	freelist_aba_t new = { .freelist = freelist_new, .counter = next_tid(tid) };
3303
3304	return this_cpu_try_cmpxchg_freelist(s->cpu_slab->freelist_tid.full,
3305					     &old.full, new.full);
3306}
3307
3308/*
3309 * Check the slab->freelist and either transfer the freelist to the
3310 * per cpu freelist or deactivate the slab.
3311 *
3312 * The slab is still frozen if the return value is not NULL.
3313 *
3314 * If this function returns NULL then the slab has been unfrozen.
3315 */
3316static inline void *get_freelist(struct kmem_cache *s, struct slab *slab)
3317{
3318	struct slab new;
3319	unsigned long counters;
3320	void *freelist;
3321
3322	lockdep_assert_held(this_cpu_ptr(&s->cpu_slab->lock));
3323
3324	do {
3325		freelist = slab->freelist;
3326		counters = slab->counters;
3327
3328		new.counters = counters;
3329		VM_BUG_ON(!new.frozen);
3330
3331		new.inuse = slab->objects;
3332		new.frozen = freelist != NULL;
3333
3334	} while (!__slab_update_freelist(s, slab,
3335		freelist, counters,
3336		NULL, new.counters,
3337		"get_freelist"));
3338
3339	return freelist;
3340}
3341
3342/*
3343 * Freeze the partial slab and return the pointer to the freelist.
3344 */
3345static inline void *freeze_slab(struct kmem_cache *s, struct slab *slab)
3346{
3347	struct slab new;
3348	unsigned long counters;
3349	void *freelist;
3350
3351	do {
3352		freelist = slab->freelist;
3353		counters = slab->counters;
3354
3355		new.counters = counters;
3356		VM_BUG_ON(new.frozen);
3357
3358		new.inuse = slab->objects;
3359		new.frozen = 1;
3360
3361	} while (!slab_update_freelist(s, slab,
3362		freelist, counters,
3363		NULL, new.counters,
3364		"freeze_slab"));
3365
3366	return freelist;
3367}
3368
3369/*
3370 * Slow path. The lockless freelist is empty or we need to perform
3371 * debugging duties.
3372 *
3373 * Processing is still very fast if new objects have been freed to the
3374 * regular freelist. In that case we simply take over the regular freelist
3375 * as the lockless freelist and zap the regular freelist.
3376 *
3377 * If that is not working then we fall back to the partial lists. We take the
3378 * first element of the freelist as the object to allocate now and move the
3379 * rest of the freelist to the lockless freelist.
3380 *
3381 * And if we were unable to get a new slab from the partial slab lists then
3382 * we need to allocate a new slab. This is the slowest path since it involves
3383 * a call to the page allocator and the setup of a new slab.
3384 *
3385 * Version of __slab_alloc to use when we know that preemption is
3386 * already disabled (which is the case for bulk allocation).
3387 */
3388static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
3389			  unsigned long addr, struct kmem_cache_cpu *c, unsigned int orig_size)
3390{
3391	void *freelist;
3392	struct slab *slab;
3393	unsigned long flags;
3394	struct partial_context pc;
3395
3396	stat(s, ALLOC_SLOWPATH);
3397
3398reread_slab:
3399
3400	slab = READ_ONCE(c->slab);
3401	if (!slab) {
3402		/*
3403		 * if the node is not online or has no normal memory, just
3404		 * ignore the node constraint
3405		 */
3406		if (unlikely(node != NUMA_NO_NODE &&
3407			     !node_isset(node, slab_nodes)))
3408			node = NUMA_NO_NODE;
3409		goto new_slab;
3410	}
3411
3412	if (unlikely(!node_match(slab, node))) {
3413		/*
3414		 * same as above but node_match() being false already
3415		 * implies node != NUMA_NO_NODE
3416		 */
3417		if (!node_isset(node, slab_nodes)) {
3418			node = NUMA_NO_NODE;
3419		} else {
3420			stat(s, ALLOC_NODE_MISMATCH);
3421			goto deactivate_slab;
3422		}
3423	}
3424
3425	/*
3426	 * By rights, we should be searching for a slab page that was
3427	 * PFMEMALLOC but right now, we are losing the pfmemalloc
3428	 * information when the page leaves the per-cpu allocator
3429	 */
3430	if (unlikely(!pfmemalloc_match(slab, gfpflags)))
3431		goto deactivate_slab;
3432
3433	/* must check again c->slab in case we got preempted and it changed */
3434	local_lock_irqsave(&s->cpu_slab->lock, flags);
3435	if (unlikely(slab != c->slab)) {
3436		local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3437		goto reread_slab;
3438	}
3439	freelist = c->freelist;
3440	if (freelist)
3441		goto load_freelist;
3442
3443	freelist = get_freelist(s, slab);
3444
3445	if (!freelist) {
3446		c->slab = NULL;
3447		c->tid = next_tid(c->tid);
3448		local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3449		stat(s, DEACTIVATE_BYPASS);
3450		goto new_slab;
3451	}
3452
3453	stat(s, ALLOC_REFILL);
3454
3455load_freelist:
3456
3457	lockdep_assert_held(this_cpu_ptr(&s->cpu_slab->lock));
3458
3459	/*
3460	 * freelist is pointing to the list of objects to be used.
3461	 * slab is pointing to the slab from which the objects are obtained.
3462	 * That slab must be frozen for per cpu allocations to work.
3463	 */
3464	VM_BUG_ON(!c->slab->frozen);
3465	c->freelist = get_freepointer(s, freelist);
3466	c->tid = next_tid(c->tid);
3467	local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3468	return freelist;
3469
3470deactivate_slab:
3471
3472	local_lock_irqsave(&s->cpu_slab->lock, flags);
3473	if (slab != c->slab) {
3474		local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3475		goto reread_slab;
3476	}
3477	freelist = c->freelist;
3478	c->slab = NULL;
3479	c->freelist = NULL;
3480	c->tid = next_tid(c->tid);
3481	local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3482	deactivate_slab(s, slab, freelist);
3483
3484new_slab:
3485
3486#ifdef CONFIG_SLUB_CPU_PARTIAL
3487	while (slub_percpu_partial(c)) {
3488		local_lock_irqsave(&s->cpu_slab->lock, flags);
3489		if (unlikely(c->slab)) {
3490			local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3491			goto reread_slab;
3492		}
3493		if (unlikely(!slub_percpu_partial(c))) {
3494			local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3495			/* we were preempted and partial list got empty */
3496			goto new_objects;
3497		}
3498
3499		slab = slub_percpu_partial(c);
3500		slub_set_percpu_partial(c, slab);
3501		local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3502		stat(s, CPU_PARTIAL_ALLOC);
3503
3504		if (unlikely(!node_match(slab, node) ||
3505			     !pfmemalloc_match(slab, gfpflags))) {
3506			slab->next = NULL;
3507			__put_partials(s, slab);
3508			continue;
 
 
3509		}
3510
3511		freelist = freeze_slab(s, slab);
3512		goto retry_load_slab;
 
 
3513	}
3514#endif
3515
3516new_objects:
3517
3518	pc.flags = gfpflags;
3519	pc.orig_size = orig_size;
3520	slab = get_partial(s, node, &pc);
3521	if (slab) {
3522		if (kmem_cache_debug(s)) {
3523			freelist = pc.object;
3524			/*
3525			 * For debug caches here we had to go through
3526			 * alloc_single_from_partial() so just store the
3527			 * tracking info and return the object.
3528			 */
3529			if (s->flags & SLAB_STORE_USER)
3530				set_track(s, freelist, TRACK_ALLOC, addr);
3531
3532			return freelist;
3533		}
3534
3535		freelist = freeze_slab(s, slab);
3536		goto retry_load_slab;
3537	}
3538
3539	slub_put_cpu_ptr(s->cpu_slab);
3540	slab = new_slab(s, gfpflags, node);
3541	c = slub_get_cpu_ptr(s->cpu_slab);
3542
3543	if (unlikely(!slab)) {
3544		slab_out_of_memory(s, gfpflags, node);
3545		return NULL;
3546	}
3547
3548	stat(s, ALLOC_SLAB);
3549
3550	if (kmem_cache_debug(s)) {
3551		freelist = alloc_single_from_new_slab(s, slab, orig_size);
3552
3553		if (unlikely(!freelist))
3554			goto new_objects;
3555
3556		if (s->flags & SLAB_STORE_USER)
3557			set_track(s, freelist, TRACK_ALLOC, addr);
3558
3559		return freelist;
3560	}
3561
3562	/*
3563	 * No other reference to the slab yet so we can
3564	 * muck around with it freely without cmpxchg
3565	 */
3566	freelist = slab->freelist;
3567	slab->freelist = NULL;
3568	slab->inuse = slab->objects;
3569	slab->frozen = 1;
3570
3571	inc_slabs_node(s, slab_nid(slab), slab->objects);
3572
3573	if (unlikely(!pfmemalloc_match(slab, gfpflags))) {
3574		/*
3575		 * For !pfmemalloc_match() case we don't load freelist so that
3576		 * we don't make further mismatched allocations easier.
3577		 */
3578		deactivate_slab(s, slab, get_freepointer(s, freelist));
3579		return freelist;
3580	}
3581
3582retry_load_slab:
3583
3584	local_lock_irqsave(&s->cpu_slab->lock, flags);
3585	if (unlikely(c->slab)) {
3586		void *flush_freelist = c->freelist;
3587		struct slab *flush_slab = c->slab;
3588
3589		c->slab = NULL;
3590		c->freelist = NULL;
3591		c->tid = next_tid(c->tid);
3592
3593		local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3594
3595		deactivate_slab(s, flush_slab, flush_freelist);
3596
3597		stat(s, CPUSLAB_FLUSH);
3598
3599		goto retry_load_slab;
3600	}
3601	c->slab = slab;
3602
3603	goto load_freelist;
3604}
3605
3606/*
3607 * A wrapper for ___slab_alloc() for contexts where preemption is not yet
3608 * disabled. Compensates for possible cpu changes by refetching the per cpu area
3609 * pointer.
3610 */
3611static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
3612			  unsigned long addr, struct kmem_cache_cpu *c, unsigned int orig_size)
3613{
3614	void *p;
3615
3616#ifdef CONFIG_PREEMPT_COUNT
3617	/*
3618	 * We may have been preempted and rescheduled on a different
3619	 * cpu before disabling preemption. Need to reload cpu area
3620	 * pointer.
3621	 */
3622	c = slub_get_cpu_ptr(s->cpu_slab);
3623#endif
3624
3625	p = ___slab_alloc(s, gfpflags, node, addr, c, orig_size);
3626#ifdef CONFIG_PREEMPT_COUNT
3627	slub_put_cpu_ptr(s->cpu_slab);
3628#endif
3629	return p;
3630}
3631
3632static __always_inline void *__slab_alloc_node(struct kmem_cache *s,
3633		gfp_t gfpflags, int node, unsigned long addr, size_t orig_size)
3634{
3635	struct kmem_cache_cpu *c;
3636	struct slab *slab;
3637	unsigned long tid;
3638	void *object;
3639
3640redo:
3641	/*
3642	 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
3643	 * enabled. We may switch back and forth between cpus while
3644	 * reading from one cpu area. That does not matter as long
3645	 * as we end up on the original cpu again when doing the cmpxchg.
3646	 *
3647	 * We must guarantee that tid and kmem_cache_cpu are retrieved on the
3648	 * same cpu. We read first the kmem_cache_cpu pointer and use it to read
3649	 * the tid. If we are preempted and switched to another cpu between the
3650	 * two reads, it's OK as the two are still associated with the same cpu
3651	 * and cmpxchg later will validate the cpu.
3652	 */
3653	c = raw_cpu_ptr(s->cpu_slab);
3654	tid = READ_ONCE(c->tid);
3655
3656	/*
3657	 * Irqless object alloc/free algorithm used here depends on sequence
3658	 * of fetching cpu_slab's data. tid should be fetched before anything
3659	 * on c to guarantee that object and slab associated with previous tid
3660	 * won't be used with current tid. If we fetch tid first, object and
3661	 * slab could be one associated with next tid and our alloc/free
3662	 * request will be failed. In this case, we will retry. So, no problem.
3663	 */
3664	barrier();
3665
3666	/*
3667	 * The transaction ids are globally unique per cpu and per operation on
3668	 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
3669	 * occurs on the right processor and that there was no operation on the
3670	 * linked list in between.
3671	 */
3672
3673	object = c->freelist;
3674	slab = c->slab;
3675
3676	if (!USE_LOCKLESS_FAST_PATH() ||
3677	    unlikely(!object || !slab || !node_match(slab, node))) {
3678		object = __slab_alloc(s, gfpflags, node, addr, c, orig_size);
3679	} else {
3680		void *next_object = get_freepointer_safe(s, object);
3681
3682		/*
3683		 * The cmpxchg will only match if there was no additional
3684		 * operation and if we are on the right processor.
3685		 *
3686		 * The cmpxchg does the following atomically (without lock
3687		 * semantics!)
3688		 * 1. Relocate first pointer to the current per cpu area.
3689		 * 2. Verify that tid and freelist have not been changed
3690		 * 3. If they were not changed replace tid and freelist
3691		 *
3692		 * Since this is without lock semantics the protection is only
3693		 * against code executing on this cpu *not* from access by
3694		 * other cpus.
3695		 */
3696		if (unlikely(!__update_cpu_freelist_fast(s, object, next_object, tid))) {
3697			note_cmpxchg_failure("slab_alloc", s, tid);
3698			goto redo;
3699		}
3700		prefetch_freepointer(s, next_object);
3701		stat(s, ALLOC_FASTPATH);
3702	}
3703
3704	return object;
3705}
3706#else /* CONFIG_SLUB_TINY */
3707static void *__slab_alloc_node(struct kmem_cache *s,
3708		gfp_t gfpflags, int node, unsigned long addr, size_t orig_size)
3709{
3710	struct partial_context pc;
3711	struct slab *slab;
3712	void *object;
3713
3714	pc.flags = gfpflags;
3715	pc.orig_size = orig_size;
3716	slab = get_partial(s, node, &pc);
3717
3718	if (slab)
3719		return pc.object;
3720
3721	slab = new_slab(s, gfpflags, node);
3722	if (unlikely(!slab)) {
3723		slab_out_of_memory(s, gfpflags, node);
3724		return NULL;
3725	}
3726
3727	object = alloc_single_from_new_slab(s, slab, orig_size);
3728
3729	return object;
3730}
3731#endif /* CONFIG_SLUB_TINY */
3732
3733/*
3734 * If the object has been wiped upon free, make sure it's fully initialized by
3735 * zeroing out freelist pointer.
3736 */
3737static __always_inline void maybe_wipe_obj_freeptr(struct kmem_cache *s,
3738						   void *obj)
3739{
3740	if (unlikely(slab_want_init_on_free(s)) && obj)
 
3741		memset((void *)((char *)kasan_reset_tag(obj) + s->offset),
3742			0, sizeof(void *));
3743}
3744
3745noinline int should_failslab(struct kmem_cache *s, gfp_t gfpflags)
3746{
3747	if (__should_failslab(s, gfpflags))
3748		return -ENOMEM;
3749	return 0;
3750}
3751ALLOW_ERROR_INJECTION(should_failslab, ERRNO);
3752
3753static __fastpath_inline
3754struct kmem_cache *slab_pre_alloc_hook(struct kmem_cache *s,
3755				       struct list_lru *lru,
3756				       struct obj_cgroup **objcgp,
3757				       size_t size, gfp_t flags)
3758{
3759	flags &= gfp_allowed_mask;
3760
3761	might_alloc(flags);
3762
3763	if (unlikely(should_failslab(s, flags)))
3764		return NULL;
3765
3766	if (unlikely(!memcg_slab_pre_alloc_hook(s, lru, objcgp, size, flags)))
3767		return NULL;
3768
3769	return s;
3770}
3771
3772static __fastpath_inline
3773void slab_post_alloc_hook(struct kmem_cache *s,	struct obj_cgroup *objcg,
3774			  gfp_t flags, size_t size, void **p, bool init,
3775			  unsigned int orig_size)
3776{
3777	unsigned int zero_size = s->object_size;
3778	bool kasan_init = init;
3779	size_t i;
3780	gfp_t init_flags = flags & gfp_allowed_mask;
3781
3782	/*
3783	 * For kmalloc object, the allocated memory size(object_size) is likely
3784	 * larger than the requested size(orig_size). If redzone check is
3785	 * enabled for the extra space, don't zero it, as it will be redzoned
3786	 * soon. The redzone operation for this extra space could be seen as a
3787	 * replacement of current poisoning under certain debug option, and
3788	 * won't break other sanity checks.
3789	 */
3790	if (kmem_cache_debug_flags(s, SLAB_STORE_USER | SLAB_RED_ZONE) &&
3791	    (s->flags & SLAB_KMALLOC))
3792		zero_size = orig_size;
3793
3794	/*
3795	 * When slub_debug is enabled, avoid memory initialization integrated
3796	 * into KASAN and instead zero out the memory via the memset below with
3797	 * the proper size. Otherwise, KASAN might overwrite SLUB redzones and
3798	 * cause false-positive reports. This does not lead to a performance
3799	 * penalty on production builds, as slub_debug is not intended to be
3800	 * enabled there.
3801	 */
3802	if (__slub_debug_enabled())
3803		kasan_init = false;
3804
3805	/*
3806	 * As memory initialization might be integrated into KASAN,
3807	 * kasan_slab_alloc and initialization memset must be
3808	 * kept together to avoid discrepancies in behavior.
3809	 *
3810	 * As p[i] might get tagged, memset and kmemleak hook come after KASAN.
3811	 */
3812	for (i = 0; i < size; i++) {
3813		p[i] = kasan_slab_alloc(s, p[i], init_flags, kasan_init);
3814		if (p[i] && init && (!kasan_init ||
3815				     !kasan_has_integrated_init()))
3816			memset(p[i], 0, zero_size);
3817		kmemleak_alloc_recursive(p[i], s->object_size, 1,
3818					 s->flags, init_flags);
3819		kmsan_slab_alloc(s, p[i], init_flags);
3820	}
3821
3822	memcg_slab_post_alloc_hook(s, objcg, flags, size, p);
3823}
3824
3825/*
3826 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
3827 * have the fastpath folded into their functions. So no function call
3828 * overhead for requests that can be satisfied on the fastpath.
3829 *
3830 * The fastpath works by first checking if the lockless freelist can be used.
3831 * If not then __slab_alloc is called for slow processing.
3832 *
3833 * Otherwise we can simply pick the next object from the lockless free list.
3834 */
3835static __fastpath_inline void *slab_alloc_node(struct kmem_cache *s, struct list_lru *lru,
3836		gfp_t gfpflags, int node, unsigned long addr, size_t orig_size)
3837{
3838	void *object;
3839	struct obj_cgroup *objcg = NULL;
3840	bool init = false;
3841
3842	s = slab_pre_alloc_hook(s, lru, &objcg, 1, gfpflags);
3843	if (unlikely(!s))
3844		return NULL;
3845
3846	object = kfence_alloc(s, orig_size, gfpflags);
3847	if (unlikely(object))
3848		goto out;
3849
3850	object = __slab_alloc_node(s, gfpflags, node, addr, orig_size);
3851
3852	maybe_wipe_obj_freeptr(s, object);
3853	init = slab_want_init_on_alloc(gfpflags, s);
3854
3855out:
3856	/*
3857	 * When init equals 'true', like for kzalloc() family, only
3858	 * @orig_size bytes might be zeroed instead of s->object_size
3859	 */
3860	slab_post_alloc_hook(s, objcg, gfpflags, 1, &object, init, orig_size);
3861
3862	return object;
3863}
3864
3865void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
3866{
3867	void *ret = slab_alloc_node(s, NULL, gfpflags, NUMA_NO_NODE, _RET_IP_,
3868				    s->object_size);
3869
3870	trace_kmem_cache_alloc(_RET_IP_, ret, s, gfpflags, NUMA_NO_NODE);
3871
3872	return ret;
3873}
3874EXPORT_SYMBOL(kmem_cache_alloc);
3875
3876void *kmem_cache_alloc_lru(struct kmem_cache *s, struct list_lru *lru,
3877			   gfp_t gfpflags)
3878{
3879	void *ret = slab_alloc_node(s, lru, gfpflags, NUMA_NO_NODE, _RET_IP_,
3880				    s->object_size);
3881
3882	trace_kmem_cache_alloc(_RET_IP_, ret, s, gfpflags, NUMA_NO_NODE);
3883
3884	return ret;
3885}
3886EXPORT_SYMBOL(kmem_cache_alloc_lru);
3887
3888/**
3889 * kmem_cache_alloc_node - Allocate an object on the specified node
3890 * @s: The cache to allocate from.
3891 * @gfpflags: See kmalloc().
3892 * @node: node number of the target node.
3893 *
3894 * Identical to kmem_cache_alloc but it will allocate memory on the given
3895 * node, which can improve the performance for cpu bound structures.
3896 *
3897 * Fallback to other node is possible if __GFP_THISNODE is not set.
3898 *
3899 * Return: pointer to the new object or %NULL in case of error
3900 */
3901void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
3902{
3903	void *ret = slab_alloc_node(s, NULL, gfpflags, node, _RET_IP_, s->object_size);
3904
3905	trace_kmem_cache_alloc(_RET_IP_, ret, s, gfpflags, node);
3906
3907	return ret;
3908}
3909EXPORT_SYMBOL(kmem_cache_alloc_node);
3910
3911/*
3912 * To avoid unnecessary overhead, we pass through large allocation requests
3913 * directly to the page allocator. We use __GFP_COMP, because we will need to
3914 * know the allocation order to free the pages properly in kfree.
3915 */
3916static void *__kmalloc_large_node(size_t size, gfp_t flags, int node)
3917{
3918	struct folio *folio;
3919	void *ptr = NULL;
3920	unsigned int order = get_order(size);
3921
3922	if (unlikely(flags & GFP_SLAB_BUG_MASK))
3923		flags = kmalloc_fix_flags(flags);
3924
3925	flags |= __GFP_COMP;
3926	folio = (struct folio *)alloc_pages_node(node, flags, order);
3927	if (folio) {
3928		ptr = folio_address(folio);
3929		lruvec_stat_mod_folio(folio, NR_SLAB_UNRECLAIMABLE_B,
3930				      PAGE_SIZE << order);
3931	}
3932
3933	ptr = kasan_kmalloc_large(ptr, size, flags);
3934	/* As ptr might get tagged, call kmemleak hook after KASAN. */
3935	kmemleak_alloc(ptr, size, 1, flags);
3936	kmsan_kmalloc_large(ptr, size, flags);
3937
3938	return ptr;
3939}
3940
3941void *kmalloc_large(size_t size, gfp_t flags)
3942{
3943	void *ret = __kmalloc_large_node(size, flags, NUMA_NO_NODE);
3944
3945	trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << get_order(size),
3946		      flags, NUMA_NO_NODE);
3947	return ret;
3948}
3949EXPORT_SYMBOL(kmalloc_large);
3950
3951void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3952{
3953	void *ret = __kmalloc_large_node(size, flags, node);
3954
3955	trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << get_order(size),
3956		      flags, node);
3957	return ret;
3958}
3959EXPORT_SYMBOL(kmalloc_large_node);
3960
3961static __always_inline
3962void *__do_kmalloc_node(size_t size, gfp_t flags, int node,
3963			unsigned long caller)
3964{
3965	struct kmem_cache *s;
3966	void *ret;
3967
3968	if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3969		ret = __kmalloc_large_node(size, flags, node);
3970		trace_kmalloc(caller, ret, size,
3971			      PAGE_SIZE << get_order(size), flags, node);
3972		return ret;
3973	}
3974
3975	if (unlikely(!size))
3976		return ZERO_SIZE_PTR;
3977
3978	s = kmalloc_slab(size, flags, caller);
3979
3980	ret = slab_alloc_node(s, NULL, flags, node, caller, size);
3981	ret = kasan_kmalloc(s, ret, size, flags);
3982	trace_kmalloc(caller, ret, size, s->size, flags, node);
3983	return ret;
3984}
3985
3986void *__kmalloc_node(size_t size, gfp_t flags, int node)
3987{
3988	return __do_kmalloc_node(size, flags, node, _RET_IP_);
3989}
3990EXPORT_SYMBOL(__kmalloc_node);
3991
3992void *__kmalloc(size_t size, gfp_t flags)
3993{
3994	return __do_kmalloc_node(size, flags, NUMA_NO_NODE, _RET_IP_);
3995}
3996EXPORT_SYMBOL(__kmalloc);
3997
3998void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3999				  int node, unsigned long caller)
4000{
4001	return __do_kmalloc_node(size, flags, node, caller);
4002}
4003EXPORT_SYMBOL(__kmalloc_node_track_caller);
4004
4005void *kmalloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
4006{
4007	void *ret = slab_alloc_node(s, NULL, gfpflags, NUMA_NO_NODE,
4008					    _RET_IP_, size);
4009
4010	trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags, NUMA_NO_NODE);
4011
4012	ret = kasan_kmalloc(s, ret, size, gfpflags);
4013	return ret;
4014}
4015EXPORT_SYMBOL(kmalloc_trace);
4016
4017void *kmalloc_node_trace(struct kmem_cache *s, gfp_t gfpflags,
4018			 int node, size_t size)
4019{
4020	void *ret = slab_alloc_node(s, NULL, gfpflags, node, _RET_IP_, size);
4021
4022	trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags, node);
4023
4024	ret = kasan_kmalloc(s, ret, size, gfpflags);
4025	return ret;
4026}
4027EXPORT_SYMBOL(kmalloc_node_trace);
4028
4029static noinline void free_to_partial_list(
4030	struct kmem_cache *s, struct slab *slab,
4031	void *head, void *tail, int bulk_cnt,
4032	unsigned long addr)
4033{
4034	struct kmem_cache_node *n = get_node(s, slab_nid(slab));
4035	struct slab *slab_free = NULL;
4036	int cnt = bulk_cnt;
4037	unsigned long flags;
4038	depot_stack_handle_t handle = 0;
4039
4040	if (s->flags & SLAB_STORE_USER)
4041		handle = set_track_prepare();
4042
4043	spin_lock_irqsave(&n->list_lock, flags);
4044
4045	if (free_debug_processing(s, slab, head, tail, &cnt, addr, handle)) {
4046		void *prior = slab->freelist;
4047
4048		/* Perform the actual freeing while we still hold the locks */
4049		slab->inuse -= cnt;
4050		set_freepointer(s, tail, prior);
4051		slab->freelist = head;
4052
4053		/*
4054		 * If the slab is empty, and node's partial list is full,
4055		 * it should be discarded anyway no matter it's on full or
4056		 * partial list.
4057		 */
4058		if (slab->inuse == 0 && n->nr_partial >= s->min_partial)
4059			slab_free = slab;
4060
4061		if (!prior) {
4062			/* was on full list */
4063			remove_full(s, n, slab);
4064			if (!slab_free) {
4065				add_partial(n, slab, DEACTIVATE_TO_TAIL);
4066				stat(s, FREE_ADD_PARTIAL);
4067			}
4068		} else if (slab_free) {
4069			remove_partial(n, slab);
4070			stat(s, FREE_REMOVE_PARTIAL);
4071		}
4072	}
4073
4074	if (slab_free) {
4075		/*
4076		 * Update the counters while still holding n->list_lock to
4077		 * prevent spurious validation warnings
4078		 */
4079		dec_slabs_node(s, slab_nid(slab_free), slab_free->objects);
4080	}
4081
4082	spin_unlock_irqrestore(&n->list_lock, flags);
4083
4084	if (slab_free) {
4085		stat(s, FREE_SLAB);
4086		free_slab(s, slab_free);
4087	}
4088}
4089
4090/*
4091 * Slow path handling. This may still be called frequently since objects
4092 * have a longer lifetime than the cpu slabs in most processing loads.
4093 *
4094 * So we still attempt to reduce cache line usage. Just take the slab
4095 * lock and free the item. If there is no additional partial slab
4096 * handling required then we can return immediately.
4097 */
4098static void __slab_free(struct kmem_cache *s, struct slab *slab,
4099			void *head, void *tail, int cnt,
4100			unsigned long addr)
4101
4102{
4103	void *prior;
4104	int was_frozen;
4105	struct slab new;
4106	unsigned long counters;
4107	struct kmem_cache_node *n = NULL;
4108	unsigned long flags;
4109	bool on_node_partial;
4110
4111	stat(s, FREE_SLOWPATH);
4112
4113	if (IS_ENABLED(CONFIG_SLUB_TINY) || kmem_cache_debug(s)) {
4114		free_to_partial_list(s, slab, head, tail, cnt, addr);
4115		return;
4116	}
4117
4118	do {
4119		if (unlikely(n)) {
4120			spin_unlock_irqrestore(&n->list_lock, flags);
4121			n = NULL;
4122		}
4123		prior = slab->freelist;
4124		counters = slab->counters;
4125		set_freepointer(s, tail, prior);
4126		new.counters = counters;
4127		was_frozen = new.frozen;
4128		new.inuse -= cnt;
4129		if ((!new.inuse || !prior) && !was_frozen) {
4130			/* Needs to be taken off a list */
4131			if (!kmem_cache_has_cpu_partial(s) || prior) {
4132
4133				n = get_node(s, slab_nid(slab));
4134				/*
4135				 * Speculatively acquire the list_lock.
4136				 * If the cmpxchg does not succeed then we may
4137				 * drop the list_lock without any processing.
4138				 *
4139				 * Otherwise the list_lock will synchronize with
4140				 * other processors updating the list of slabs.
4141				 */
4142				spin_lock_irqsave(&n->list_lock, flags);
4143
4144				on_node_partial = slab_test_node_partial(slab);
4145			}
4146		}
4147
4148	} while (!slab_update_freelist(s, slab,
4149		prior, counters,
4150		head, new.counters,
4151		"__slab_free"));
4152
4153	if (likely(!n)) {
4154
4155		if (likely(was_frozen)) {
4156			/*
4157			 * The list lock was not taken therefore no list
4158			 * activity can be necessary.
4159			 */
4160			stat(s, FREE_FROZEN);
4161		} else if (kmem_cache_has_cpu_partial(s) && !prior) {
4162			/*
4163			 * If we started with a full slab then put it onto the
4164			 * per cpu partial list.
4165			 */
4166			put_cpu_partial(s, slab, 1);
4167			stat(s, CPU_PARTIAL_FREE);
4168		}
4169
4170		return;
4171	}
4172
4173	/*
4174	 * This slab was partially empty but not on the per-node partial list,
4175	 * in which case we shouldn't manipulate its list, just return.
4176	 */
4177	if (prior && !on_node_partial) {
4178		spin_unlock_irqrestore(&n->list_lock, flags);
4179		return;
4180	}
4181
4182	if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
4183		goto slab_empty;
4184
4185	/*
4186	 * Objects left in the slab. If it was not on the partial list before
4187	 * then add it.
4188	 */
4189	if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
4190		remove_full(s, n, slab);
4191		add_partial(n, slab, DEACTIVATE_TO_TAIL);
4192		stat(s, FREE_ADD_PARTIAL);
4193	}
4194	spin_unlock_irqrestore(&n->list_lock, flags);
4195	return;
4196
4197slab_empty:
4198	if (prior) {
4199		/*
4200		 * Slab on the partial list.
4201		 */
4202		remove_partial(n, slab);
4203		stat(s, FREE_REMOVE_PARTIAL);
4204	} else {
4205		/* Slab must be on the full list */
4206		remove_full(s, n, slab);
4207	}
4208
4209	spin_unlock_irqrestore(&n->list_lock, flags);
4210	stat(s, FREE_SLAB);
4211	discard_slab(s, slab);
4212}
4213
4214#ifndef CONFIG_SLUB_TINY
4215/*
4216 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
4217 * can perform fastpath freeing without additional function calls.
4218 *
4219 * The fastpath is only possible if we are freeing to the current cpu slab
4220 * of this processor. This typically the case if we have just allocated
4221 * the item before.
4222 *
4223 * If fastpath is not possible then fall back to __slab_free where we deal
4224 * with all sorts of special processing.
4225 *
4226 * Bulk free of a freelist with several objects (all pointing to the
4227 * same slab) possible by specifying head and tail ptr, plus objects
4228 * count (cnt). Bulk free indicated by tail pointer being set.
4229 */
4230static __always_inline void do_slab_free(struct kmem_cache *s,
4231				struct slab *slab, void *head, void *tail,
4232				int cnt, unsigned long addr)
4233{
4234	struct kmem_cache_cpu *c;
4235	unsigned long tid;
4236	void **freelist;
4237
4238redo:
4239	/*
4240	 * Determine the currently cpus per cpu slab.
4241	 * The cpu may change afterward. However that does not matter since
4242	 * data is retrieved via this pointer. If we are on the same cpu
4243	 * during the cmpxchg then the free will succeed.
4244	 */
4245	c = raw_cpu_ptr(s->cpu_slab);
4246	tid = READ_ONCE(c->tid);
4247
4248	/* Same with comment on barrier() in slab_alloc_node() */
4249	barrier();
4250
4251	if (unlikely(slab != c->slab)) {
4252		__slab_free(s, slab, head, tail, cnt, addr);
4253		return;
4254	}
4255
4256	if (USE_LOCKLESS_FAST_PATH()) {
4257		freelist = READ_ONCE(c->freelist);
4258
4259		set_freepointer(s, tail, freelist);
4260
4261		if (unlikely(!__update_cpu_freelist_fast(s, freelist, head, tid))) {
4262			note_cmpxchg_failure("slab_free", s, tid);
4263			goto redo;
4264		}
4265	} else {
4266		/* Update the free list under the local lock */
4267		local_lock(&s->cpu_slab->lock);
4268		c = this_cpu_ptr(s->cpu_slab);
4269		if (unlikely(slab != c->slab)) {
4270			local_unlock(&s->cpu_slab->lock);
4271			goto redo;
4272		}
4273		tid = c->tid;
4274		freelist = c->freelist;
4275
4276		set_freepointer(s, tail, freelist);
4277		c->freelist = head;
4278		c->tid = next_tid(tid);
4279
4280		local_unlock(&s->cpu_slab->lock);
4281	}
4282	stat_add(s, FREE_FASTPATH, cnt);
4283}
4284#else /* CONFIG_SLUB_TINY */
4285static void do_slab_free(struct kmem_cache *s,
4286				struct slab *slab, void *head, void *tail,
4287				int cnt, unsigned long addr)
4288{
4289	__slab_free(s, slab, head, tail, cnt, addr);
4290}
4291#endif /* CONFIG_SLUB_TINY */
4292
4293static __fastpath_inline
4294void slab_free(struct kmem_cache *s, struct slab *slab, void *object,
4295	       unsigned long addr)
4296{
4297	memcg_slab_free_hook(s, slab, &object, 1);
4298
4299	if (likely(slab_free_hook(s, object, slab_want_init_on_free(s))))
4300		do_slab_free(s, slab, object, object, 1, addr);
4301}
4302
4303static __fastpath_inline
4304void slab_free_bulk(struct kmem_cache *s, struct slab *slab, void *head,
4305		    void *tail, void **p, int cnt, unsigned long addr)
4306{
4307	memcg_slab_free_hook(s, slab, p, cnt);
4308	/*
4309	 * With KASAN enabled slab_free_freelist_hook modifies the freelist
4310	 * to remove objects, whose reuse must be delayed.
4311	 */
4312	if (likely(slab_free_freelist_hook(s, &head, &tail, &cnt)))
4313		do_slab_free(s, slab, head, tail, cnt, addr);
4314}
4315
4316#ifdef CONFIG_KASAN_GENERIC
4317void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
4318{
4319	do_slab_free(cache, virt_to_slab(x), x, x, 1, addr);
4320}
4321#endif
4322
4323static inline struct kmem_cache *virt_to_cache(const void *obj)
4324{
4325	struct slab *slab;
4326
4327	slab = virt_to_slab(obj);
4328	if (WARN_ONCE(!slab, "%s: Object is not a Slab page!\n", __func__))
4329		return NULL;
4330	return slab->slab_cache;
4331}
4332
4333static inline struct kmem_cache *cache_from_obj(struct kmem_cache *s, void *x)
4334{
4335	struct kmem_cache *cachep;
4336
4337	if (!IS_ENABLED(CONFIG_SLAB_FREELIST_HARDENED) &&
4338	    !kmem_cache_debug_flags(s, SLAB_CONSISTENCY_CHECKS))
4339		return s;
4340
4341	cachep = virt_to_cache(x);
4342	if (WARN(cachep && cachep != s,
4343		 "%s: Wrong slab cache. %s but object is from %s\n",
4344		 __func__, s->name, cachep->name))
4345		print_tracking(cachep, x);
4346	return cachep;
4347}
4348
4349/**
4350 * kmem_cache_free - Deallocate an object
4351 * @s: The cache the allocation was from.
4352 * @x: The previously allocated object.
4353 *
4354 * Free an object which was previously allocated from this
4355 * cache.
4356 */
4357void kmem_cache_free(struct kmem_cache *s, void *x)
4358{
4359	s = cache_from_obj(s, x);
4360	if (!s)
4361		return;
4362	trace_kmem_cache_free(_RET_IP_, x, s);
4363	slab_free(s, virt_to_slab(x), x, _RET_IP_);
4364}
4365EXPORT_SYMBOL(kmem_cache_free);
4366
4367static void free_large_kmalloc(struct folio *folio, void *object)
4368{
4369	unsigned int order = folio_order(folio);
4370
4371	if (WARN_ON_ONCE(order == 0))
4372		pr_warn_once("object pointer: 0x%p\n", object);
4373
4374	kmemleak_free(object);
4375	kasan_kfree_large(object);
4376	kmsan_kfree_large(object);
4377
4378	lruvec_stat_mod_folio(folio, NR_SLAB_UNRECLAIMABLE_B,
4379			      -(PAGE_SIZE << order));
4380	folio_put(folio);
4381}
4382
4383/**
4384 * kfree - free previously allocated memory
4385 * @object: pointer returned by kmalloc() or kmem_cache_alloc()
4386 *
4387 * If @object is NULL, no operation is performed.
4388 */
4389void kfree(const void *object)
4390{
4391	struct folio *folio;
4392	struct slab *slab;
4393	struct kmem_cache *s;
4394	void *x = (void *)object;
4395
4396	trace_kfree(_RET_IP_, object);
4397
4398	if (unlikely(ZERO_OR_NULL_PTR(object)))
4399		return;
4400
4401	folio = virt_to_folio(object);
4402	if (unlikely(!folio_test_slab(folio))) {
4403		free_large_kmalloc(folio, (void *)object);
4404		return;
4405	}
4406
4407	slab = folio_slab(folio);
4408	s = slab->slab_cache;
4409	slab_free(s, slab, x, _RET_IP_);
4410}
4411EXPORT_SYMBOL(kfree);
4412
4413struct detached_freelist {
4414	struct slab *slab;
4415	void *tail;
4416	void *freelist;
4417	int cnt;
4418	struct kmem_cache *s;
4419};
4420
4421/*
4422 * This function progressively scans the array with free objects (with
4423 * a limited look ahead) and extract objects belonging to the same
4424 * slab.  It builds a detached freelist directly within the given
4425 * slab/objects.  This can happen without any need for
4426 * synchronization, because the objects are owned by running process.
4427 * The freelist is build up as a single linked list in the objects.
4428 * The idea is, that this detached freelist can then be bulk
4429 * transferred to the real freelist(s), but only requiring a single
4430 * synchronization primitive.  Look ahead in the array is limited due
4431 * to performance reasons.
4432 */
4433static inline
4434int build_detached_freelist(struct kmem_cache *s, size_t size,
4435			    void **p, struct detached_freelist *df)
4436{
4437	int lookahead = 3;
4438	void *object;
4439	struct folio *folio;
4440	size_t same;
4441
4442	object = p[--size];
4443	folio = virt_to_folio(object);
4444	if (!s) {
4445		/* Handle kalloc'ed objects */
4446		if (unlikely(!folio_test_slab(folio))) {
4447			free_large_kmalloc(folio, object);
4448			df->slab = NULL;
4449			return size;
4450		}
4451		/* Derive kmem_cache from object */
4452		df->slab = folio_slab(folio);
4453		df->s = df->slab->slab_cache;
4454	} else {
4455		df->slab = folio_slab(folio);
4456		df->s = cache_from_obj(s, object); /* Support for memcg */
4457	}
4458
4459	/* Start new detached freelist */
4460	df->tail = object;
4461	df->freelist = object;
4462	df->cnt = 1;
4463
4464	if (is_kfence_address(object))
4465		return size;
4466
4467	set_freepointer(df->s, object, NULL);
4468
4469	same = size;
4470	while (size) {
4471		object = p[--size];
4472		/* df->slab is always set at this point */
4473		if (df->slab == virt_to_slab(object)) {
4474			/* Opportunity build freelist */
4475			set_freepointer(df->s, object, df->freelist);
4476			df->freelist = object;
4477			df->cnt++;
4478			same--;
4479			if (size != same)
4480				swap(p[size], p[same]);
4481			continue;
4482		}
4483
4484		/* Limit look ahead search */
4485		if (!--lookahead)
4486			break;
4487	}
4488
4489	return same;
4490}
4491
4492/*
4493 * Internal bulk free of objects that were not initialised by the post alloc
4494 * hooks and thus should not be processed by the free hooks
4495 */
4496static void __kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
4497{
4498	if (!size)
4499		return;
4500
4501	do {
4502		struct detached_freelist df;
4503
4504		size = build_detached_freelist(s, size, p, &df);
4505		if (!df.slab)
4506			continue;
4507
4508		do_slab_free(df.s, df.slab, df.freelist, df.tail, df.cnt,
4509			     _RET_IP_);
4510	} while (likely(size));
4511}
4512
4513/* Note that interrupts must be enabled when calling this function. */
4514void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
4515{
4516	if (!size)
4517		return;
4518
4519	do {
4520		struct detached_freelist df;
4521
4522		size = build_detached_freelist(s, size, p, &df);
4523		if (!df.slab)
4524			continue;
4525
4526		slab_free_bulk(df.s, df.slab, df.freelist, df.tail, &p[size],
4527			       df.cnt, _RET_IP_);
4528	} while (likely(size));
4529}
4530EXPORT_SYMBOL(kmem_cache_free_bulk);
4531
4532#ifndef CONFIG_SLUB_TINY
4533static inline
4534int __kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
4535			    void **p)
4536{
4537	struct kmem_cache_cpu *c;
4538	unsigned long irqflags;
4539	int i;
4540
4541	/*
4542	 * Drain objects in the per cpu slab, while disabling local
4543	 * IRQs, which protects against PREEMPT and interrupts
4544	 * handlers invoking normal fastpath.
4545	 */
4546	c = slub_get_cpu_ptr(s->cpu_slab);
4547	local_lock_irqsave(&s->cpu_slab->lock, irqflags);
4548
4549	for (i = 0; i < size; i++) {
4550		void *object = kfence_alloc(s, s->object_size, flags);
4551
4552		if (unlikely(object)) {
4553			p[i] = object;
4554			continue;
4555		}
4556
4557		object = c->freelist;
4558		if (unlikely(!object)) {
4559			/*
4560			 * We may have removed an object from c->freelist using
4561			 * the fastpath in the previous iteration; in that case,
4562			 * c->tid has not been bumped yet.
4563			 * Since ___slab_alloc() may reenable interrupts while
4564			 * allocating memory, we should bump c->tid now.
4565			 */
4566			c->tid = next_tid(c->tid);
4567
4568			local_unlock_irqrestore(&s->cpu_slab->lock, irqflags);
4569
4570			/*
4571			 * Invoking slow path likely have side-effect
4572			 * of re-populating per CPU c->freelist
4573			 */
4574			p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
4575					    _RET_IP_, c, s->object_size);
4576			if (unlikely(!p[i]))
4577				goto error;
4578
4579			c = this_cpu_ptr(s->cpu_slab);
4580			maybe_wipe_obj_freeptr(s, p[i]);
4581
4582			local_lock_irqsave(&s->cpu_slab->lock, irqflags);
4583
4584			continue; /* goto for-loop */
4585		}
4586		c->freelist = get_freepointer(s, object);
4587		p[i] = object;
4588		maybe_wipe_obj_freeptr(s, p[i]);
4589		stat(s, ALLOC_FASTPATH);
4590	}
4591	c->tid = next_tid(c->tid);
4592	local_unlock_irqrestore(&s->cpu_slab->lock, irqflags);
4593	slub_put_cpu_ptr(s->cpu_slab);
4594
4595	return i;
4596
4597error:
4598	slub_put_cpu_ptr(s->cpu_slab);
4599	__kmem_cache_free_bulk(s, i, p);
4600	return 0;
4601
4602}
4603#else /* CONFIG_SLUB_TINY */
4604static int __kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags,
4605				   size_t size, void **p)
4606{
4607	int i;
4608
4609	for (i = 0; i < size; i++) {
4610		void *object = kfence_alloc(s, s->object_size, flags);
4611
4612		if (unlikely(object)) {
4613			p[i] = object;
4614			continue;
4615		}
4616
4617		p[i] = __slab_alloc_node(s, flags, NUMA_NO_NODE,
4618					 _RET_IP_, s->object_size);
4619		if (unlikely(!p[i]))
4620			goto error;
4621
4622		maybe_wipe_obj_freeptr(s, p[i]);
4623	}
4624
4625	return i;
4626
4627error:
4628	__kmem_cache_free_bulk(s, i, p);
4629	return 0;
4630}
4631#endif /* CONFIG_SLUB_TINY */
4632
4633/* Note that interrupts must be enabled when calling this function. */
4634int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
4635			  void **p)
4636{
4637	int i;
4638	struct obj_cgroup *objcg = NULL;
4639
4640	if (!size)
4641		return 0;
4642
4643	/* memcg and kmem_cache debug support */
4644	s = slab_pre_alloc_hook(s, NULL, &objcg, size, flags);
4645	if (unlikely(!s))
4646		return 0;
4647
4648	i = __kmem_cache_alloc_bulk(s, flags, size, p);
4649
4650	/*
4651	 * memcg and kmem_cache debug support and memory initialization.
4652	 * Done outside of the IRQ disabled fastpath loop.
4653	 */
4654	if (likely(i != 0)) {
4655		slab_post_alloc_hook(s, objcg, flags, size, p,
4656			slab_want_init_on_alloc(flags, s), s->object_size);
4657	} else {
4658		memcg_slab_alloc_error_hook(s, size, objcg);
4659	}
4660
4661	return i;
4662}
4663EXPORT_SYMBOL(kmem_cache_alloc_bulk);
4664
4665
4666/*
4667 * Object placement in a slab is made very easy because we always start at
4668 * offset 0. If we tune the size of the object to the alignment then we can
4669 * get the required alignment by putting one properly sized object after
4670 * another.
4671 *
4672 * Notice that the allocation order determines the sizes of the per cpu
4673 * caches. Each processor has always one slab available for allocations.
4674 * Increasing the allocation order reduces the number of times that slabs
4675 * must be moved on and off the partial lists and is therefore a factor in
4676 * locking overhead.
4677 */
4678
4679/*
4680 * Minimum / Maximum order of slab pages. This influences locking overhead
4681 * and slab fragmentation. A higher order reduces the number of partial slabs
4682 * and increases the number of allocations possible without having to
4683 * take the list_lock.
4684 */
4685static unsigned int slub_min_order;
4686static unsigned int slub_max_order =
4687	IS_ENABLED(CONFIG_SLUB_TINY) ? 1 : PAGE_ALLOC_COSTLY_ORDER;
4688static unsigned int slub_min_objects;
4689
4690/*
4691 * Calculate the order of allocation given an slab object size.
4692 *
4693 * The order of allocation has significant impact on performance and other
4694 * system components. Generally order 0 allocations should be preferred since
4695 * order 0 does not cause fragmentation in the page allocator. Larger objects
4696 * be problematic to put into order 0 slabs because there may be too much
4697 * unused space left. We go to a higher order if more than 1/16th of the slab
4698 * would be wasted.
4699 *
4700 * In order to reach satisfactory performance we must ensure that a minimum
4701 * number of objects is in one slab. Otherwise we may generate too much
4702 * activity on the partial lists which requires taking the list_lock. This is
4703 * less a concern for large slabs though which are rarely used.
4704 *
4705 * slub_max_order specifies the order where we begin to stop considering the
4706 * number of objects in a slab as critical. If we reach slub_max_order then
4707 * we try to keep the page order as low as possible. So we accept more waste
4708 * of space in favor of a small page order.
4709 *
4710 * Higher order allocations also allow the placement of more objects in a
4711 * slab and thereby reduce object handling overhead. If the user has
4712 * requested a higher minimum order then we start with that one instead of
4713 * the smallest order which will fit the object.
4714 */
4715static inline unsigned int calc_slab_order(unsigned int size,
4716		unsigned int min_order, unsigned int max_order,
4717		unsigned int fract_leftover)
4718{
4719	unsigned int order;
4720
4721	for (order = min_order; order <= max_order; order++) {
4722
4723		unsigned int slab_size = (unsigned int)PAGE_SIZE << order;
4724		unsigned int rem;
4725
4726		rem = slab_size % size;
4727
4728		if (rem <= slab_size / fract_leftover)
4729			break;
4730	}
4731
4732	return order;
4733}
4734
4735static inline int calculate_order(unsigned int size)
4736{
4737	unsigned int order;
4738	unsigned int min_objects;
4739	unsigned int max_objects;
4740	unsigned int min_order;
4741
4742	min_objects = slub_min_objects;
4743	if (!min_objects) {
4744		/*
4745		 * Some architectures will only update present cpus when
4746		 * onlining them, so don't trust the number if it's just 1. But
4747		 * we also don't want to use nr_cpu_ids always, as on some other
4748		 * architectures, there can be many possible cpus, but never
4749		 * onlined. Here we compromise between trying to avoid too high
4750		 * order on systems that appear larger than they are, and too
4751		 * low order on systems that appear smaller than they are.
4752		 */
4753		unsigned int nr_cpus = num_present_cpus();
4754		if (nr_cpus <= 1)
4755			nr_cpus = nr_cpu_ids;
4756		min_objects = 4 * (fls(nr_cpus) + 1);
4757	}
4758	/* min_objects can't be 0 because get_order(0) is undefined */
4759	max_objects = max(order_objects(slub_max_order, size), 1U);
4760	min_objects = min(min_objects, max_objects);
4761
4762	min_order = max_t(unsigned int, slub_min_order,
4763			  get_order(min_objects * size));
4764	if (order_objects(min_order, size) > MAX_OBJS_PER_PAGE)
4765		return get_order(size * MAX_OBJS_PER_PAGE) - 1;
4766
4767	/*
4768	 * Attempt to find best configuration for a slab. This works by first
4769	 * attempting to generate a layout with the best possible configuration
4770	 * and backing off gradually.
4771	 *
4772	 * We start with accepting at most 1/16 waste and try to find the
4773	 * smallest order from min_objects-derived/slub_min_order up to
4774	 * slub_max_order that will satisfy the constraint. Note that increasing
4775	 * the order can only result in same or less fractional waste, not more.
4776	 *
4777	 * If that fails, we increase the acceptable fraction of waste and try
4778	 * again. The last iteration with fraction of 1/2 would effectively
4779	 * accept any waste and give us the order determined by min_objects, as
4780	 * long as at least single object fits within slub_max_order.
4781	 */
4782	for (unsigned int fraction = 16; fraction > 1; fraction /= 2) {
4783		order = calc_slab_order(size, min_order, slub_max_order,
4784					fraction);
4785		if (order <= slub_max_order)
4786			return order;
4787	}
4788
4789	/*
4790	 * Doh this slab cannot be placed using slub_max_order.
4791	 */
4792	order = get_order(size);
4793	if (order <= MAX_PAGE_ORDER)
4794		return order;
4795	return -ENOSYS;
4796}
4797
4798static void
4799init_kmem_cache_node(struct kmem_cache_node *n)
4800{
4801	n->nr_partial = 0;
4802	spin_lock_init(&n->list_lock);
4803	INIT_LIST_HEAD(&n->partial);
4804#ifdef CONFIG_SLUB_DEBUG
4805	atomic_long_set(&n->nr_slabs, 0);
4806	atomic_long_set(&n->total_objects, 0);
4807	INIT_LIST_HEAD(&n->full);
4808#endif
4809}
4810
4811#ifndef CONFIG_SLUB_TINY
4812static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
4813{
4814	BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
4815			NR_KMALLOC_TYPES * KMALLOC_SHIFT_HIGH *
4816			sizeof(struct kmem_cache_cpu));
4817
4818	/*
4819	 * Must align to double word boundary for the double cmpxchg
4820	 * instructions to work; see __pcpu_double_call_return_bool().
4821	 */
4822	s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
4823				     2 * sizeof(void *));
4824
4825	if (!s->cpu_slab)
4826		return 0;
4827
4828	init_kmem_cache_cpus(s);
4829
4830	return 1;
4831}
4832#else
4833static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
4834{
4835	return 1;
4836}
4837#endif /* CONFIG_SLUB_TINY */
4838
4839static struct kmem_cache *kmem_cache_node;
4840
4841/*
4842 * No kmalloc_node yet so do it by hand. We know that this is the first
4843 * slab on the node for this slabcache. There are no concurrent accesses
4844 * possible.
4845 *
4846 * Note that this function only works on the kmem_cache_node
4847 * when allocating for the kmem_cache_node. This is used for bootstrapping
4848 * memory on a fresh node that has no slab structures yet.
4849 */
4850static void early_kmem_cache_node_alloc(int node)
4851{
4852	struct slab *slab;
4853	struct kmem_cache_node *n;
4854
4855	BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
4856
4857	slab = new_slab(kmem_cache_node, GFP_NOWAIT, node);
4858
4859	BUG_ON(!slab);
4860	inc_slabs_node(kmem_cache_node, slab_nid(slab), slab->objects);
4861	if (slab_nid(slab) != node) {
4862		pr_err("SLUB: Unable to allocate memory from node %d\n", node);
4863		pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
4864	}
4865
4866	n = slab->freelist;
4867	BUG_ON(!n);
4868#ifdef CONFIG_SLUB_DEBUG
4869	init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
4870	init_tracking(kmem_cache_node, n);
4871#endif
4872	n = kasan_slab_alloc(kmem_cache_node, n, GFP_KERNEL, false);
4873	slab->freelist = get_freepointer(kmem_cache_node, n);
4874	slab->inuse = 1;
4875	kmem_cache_node->node[node] = n;
4876	init_kmem_cache_node(n);
4877	inc_slabs_node(kmem_cache_node, node, slab->objects);
4878
4879	/*
4880	 * No locks need to be taken here as it has just been
4881	 * initialized and there is no concurrent access.
4882	 */
4883	__add_partial(n, slab, DEACTIVATE_TO_HEAD);
4884}
4885
4886static void free_kmem_cache_nodes(struct kmem_cache *s)
4887{
4888	int node;
4889	struct kmem_cache_node *n;
4890
4891	for_each_kmem_cache_node(s, node, n) {
4892		s->node[node] = NULL;
4893		kmem_cache_free(kmem_cache_node, n);
4894	}
4895}
4896
4897void __kmem_cache_release(struct kmem_cache *s)
4898{
4899	cache_random_seq_destroy(s);
4900#ifndef CONFIG_SLUB_TINY
4901	free_percpu(s->cpu_slab);
4902#endif
4903	free_kmem_cache_nodes(s);
4904}
4905
4906static int init_kmem_cache_nodes(struct kmem_cache *s)
4907{
4908	int node;
4909
4910	for_each_node_mask(node, slab_nodes) {
4911		struct kmem_cache_node *n;
4912
4913		if (slab_state == DOWN) {
4914			early_kmem_cache_node_alloc(node);
4915			continue;
4916		}
4917		n = kmem_cache_alloc_node(kmem_cache_node,
4918						GFP_KERNEL, node);
4919
4920		if (!n) {
4921			free_kmem_cache_nodes(s);
4922			return 0;
4923		}
4924
4925		init_kmem_cache_node(n);
4926		s->node[node] = n;
4927	}
4928	return 1;
4929}
4930
4931static void set_cpu_partial(struct kmem_cache *s)
4932{
4933#ifdef CONFIG_SLUB_CPU_PARTIAL
4934	unsigned int nr_objects;
4935
4936	/*
4937	 * cpu_partial determined the maximum number of objects kept in the
4938	 * per cpu partial lists of a processor.
4939	 *
4940	 * Per cpu partial lists mainly contain slabs that just have one
4941	 * object freed. If they are used for allocation then they can be
4942	 * filled up again with minimal effort. The slab will never hit the
4943	 * per node partial lists and therefore no locking will be required.
4944	 *
4945	 * For backwards compatibility reasons, this is determined as number
4946	 * of objects, even though we now limit maximum number of pages, see
4947	 * slub_set_cpu_partial()
4948	 */
4949	if (!kmem_cache_has_cpu_partial(s))
4950		nr_objects = 0;
4951	else if (s->size >= PAGE_SIZE)
4952		nr_objects = 6;
4953	else if (s->size >= 1024)
4954		nr_objects = 24;
4955	else if (s->size >= 256)
4956		nr_objects = 52;
4957	else
4958		nr_objects = 120;
4959
4960	slub_set_cpu_partial(s, nr_objects);
4961#endif
4962}
4963
4964/*
4965 * calculate_sizes() determines the order and the distribution of data within
4966 * a slab object.
4967 */
4968static int calculate_sizes(struct kmem_cache *s)
4969{
4970	slab_flags_t flags = s->flags;
4971	unsigned int size = s->object_size;
4972	unsigned int order;
4973
4974	/*
4975	 * Round up object size to the next word boundary. We can only
4976	 * place the free pointer at word boundaries and this determines
4977	 * the possible location of the free pointer.
4978	 */
4979	size = ALIGN(size, sizeof(void *));
4980
4981#ifdef CONFIG_SLUB_DEBUG
4982	/*
4983	 * Determine if we can poison the object itself. If the user of
4984	 * the slab may touch the object after free or before allocation
4985	 * then we should never poison the object itself.
4986	 */
4987	if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) &&
4988			!s->ctor)
4989		s->flags |= __OBJECT_POISON;
4990	else
4991		s->flags &= ~__OBJECT_POISON;
4992
4993
4994	/*
4995	 * If we are Redzoning then check if there is some space between the
4996	 * end of the object and the free pointer. If not then add an
4997	 * additional word to have some bytes to store Redzone information.
4998	 */
4999	if ((flags & SLAB_RED_ZONE) && size == s->object_size)
5000		size += sizeof(void *);
5001#endif
5002
5003	/*
5004	 * With that we have determined the number of bytes in actual use
5005	 * by the object and redzoning.
5006	 */
5007	s->inuse = size;
5008
5009	if (slub_debug_orig_size(s) ||
5010	    (flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) ||
5011	    ((flags & SLAB_RED_ZONE) && s->object_size < sizeof(void *)) ||
5012	    s->ctor) {
5013		/*
5014		 * Relocate free pointer after the object if it is not
5015		 * permitted to overwrite the first word of the object on
5016		 * kmem_cache_free.
5017		 *
5018		 * This is the case if we do RCU, have a constructor or
5019		 * destructor, are poisoning the objects, or are
5020		 * redzoning an object smaller than sizeof(void *).
5021		 *
5022		 * The assumption that s->offset >= s->inuse means free
5023		 * pointer is outside of the object is used in the
5024		 * freeptr_outside_object() function. If that is no
5025		 * longer true, the function needs to be modified.
5026		 */
5027		s->offset = size;
5028		size += sizeof(void *);
5029	} else {
5030		/*
5031		 * Store freelist pointer near middle of object to keep
5032		 * it away from the edges of the object to avoid small
5033		 * sized over/underflows from neighboring allocations.
5034		 */
5035		s->offset = ALIGN_DOWN(s->object_size / 2, sizeof(void *));
5036	}
5037
5038#ifdef CONFIG_SLUB_DEBUG
5039	if (flags & SLAB_STORE_USER) {
5040		/*
5041		 * Need to store information about allocs and frees after
5042		 * the object.
5043		 */
5044		size += 2 * sizeof(struct track);
5045
5046		/* Save the original kmalloc request size */
5047		if (flags & SLAB_KMALLOC)
5048			size += sizeof(unsigned int);
5049	}
5050#endif
5051
5052	kasan_cache_create(s, &size, &s->flags);
5053#ifdef CONFIG_SLUB_DEBUG
5054	if (flags & SLAB_RED_ZONE) {
5055		/*
5056		 * Add some empty padding so that we can catch
5057		 * overwrites from earlier objects rather than let
5058		 * tracking information or the free pointer be
5059		 * corrupted if a user writes before the start
5060		 * of the object.
5061		 */
5062		size += sizeof(void *);
5063
5064		s->red_left_pad = sizeof(void *);
5065		s->red_left_pad = ALIGN(s->red_left_pad, s->align);
5066		size += s->red_left_pad;
5067	}
5068#endif
5069
5070	/*
5071	 * SLUB stores one object immediately after another beginning from
5072	 * offset 0. In order to align the objects we have to simply size
5073	 * each object to conform to the alignment.
5074	 */
5075	size = ALIGN(size, s->align);
5076	s->size = size;
5077	s->reciprocal_size = reciprocal_value(size);
5078	order = calculate_order(size);
5079
5080	if ((int)order < 0)
5081		return 0;
5082
5083	s->allocflags = 0;
5084	if (order)
5085		s->allocflags |= __GFP_COMP;
5086
5087	if (s->flags & SLAB_CACHE_DMA)
5088		s->allocflags |= GFP_DMA;
5089
5090	if (s->flags & SLAB_CACHE_DMA32)
5091		s->allocflags |= GFP_DMA32;
5092
5093	if (s->flags & SLAB_RECLAIM_ACCOUNT)
5094		s->allocflags |= __GFP_RECLAIMABLE;
5095
5096	/*
5097	 * Determine the number of objects per slab
5098	 */
5099	s->oo = oo_make(order, size);
5100	s->min = oo_make(get_order(size), size);
5101
5102	return !!oo_objects(s->oo);
5103}
5104
5105static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags)
5106{
5107	s->flags = kmem_cache_flags(s->size, flags, s->name);
5108#ifdef CONFIG_SLAB_FREELIST_HARDENED
5109	s->random = get_random_long();
5110#endif
5111
5112	if (!calculate_sizes(s))
5113		goto error;
5114	if (disable_higher_order_debug) {
5115		/*
5116		 * Disable debugging flags that store metadata if the min slab
5117		 * order increased.
5118		 */
5119		if (get_order(s->size) > get_order(s->object_size)) {
5120			s->flags &= ~DEBUG_METADATA_FLAGS;
5121			s->offset = 0;
5122			if (!calculate_sizes(s))
5123				goto error;
5124		}
5125	}
5126
5127#ifdef system_has_freelist_aba
5128	if (system_has_freelist_aba() && !(s->flags & SLAB_NO_CMPXCHG)) {
5129		/* Enable fast mode */
5130		s->flags |= __CMPXCHG_DOUBLE;
5131	}
5132#endif
5133
5134	/*
5135	 * The larger the object size is, the more slabs we want on the partial
5136	 * list to avoid pounding the page allocator excessively.
5137	 */
5138	s->min_partial = min_t(unsigned long, MAX_PARTIAL, ilog2(s->size) / 2);
5139	s->min_partial = max_t(unsigned long, MIN_PARTIAL, s->min_partial);
5140
5141	set_cpu_partial(s);
5142
5143#ifdef CONFIG_NUMA
5144	s->remote_node_defrag_ratio = 1000;
5145#endif
5146
5147	/* Initialize the pre-computed randomized freelist if slab is up */
5148	if (slab_state >= UP) {
5149		if (init_cache_random_seq(s))
5150			goto error;
5151	}
5152
5153	if (!init_kmem_cache_nodes(s))
5154		goto error;
5155
5156	if (alloc_kmem_cache_cpus(s))
5157		return 0;
5158
5159error:
5160	__kmem_cache_release(s);
5161	return -EINVAL;
5162}
5163
5164static void list_slab_objects(struct kmem_cache *s, struct slab *slab,
5165			      const char *text)
5166{
5167#ifdef CONFIG_SLUB_DEBUG
5168	void *addr = slab_address(slab);
5169	void *p;
5170
5171	slab_err(s, slab, text, s->name);
5172
5173	spin_lock(&object_map_lock);
5174	__fill_map(object_map, s, slab);
5175
5176	for_each_object(p, s, addr, slab->objects) {
5177
5178		if (!test_bit(__obj_to_index(s, addr, p), object_map)) {
5179			pr_err("Object 0x%p @offset=%tu\n", p, p - addr);
5180			print_tracking(s, p);
5181		}
5182	}
5183	spin_unlock(&object_map_lock);
5184#endif
5185}
5186
5187/*
5188 * Attempt to free all partial slabs on a node.
5189 * This is called from __kmem_cache_shutdown(). We must take list_lock
5190 * because sysfs file might still access partial list after the shutdowning.
5191 */
5192static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
5193{
5194	LIST_HEAD(discard);
5195	struct slab *slab, *h;
5196
5197	BUG_ON(irqs_disabled());
5198	spin_lock_irq(&n->list_lock);
5199	list_for_each_entry_safe(slab, h, &n->partial, slab_list) {
5200		if (!slab->inuse) {
5201			remove_partial(n, slab);
5202			list_add(&slab->slab_list, &discard);
5203		} else {
5204			list_slab_objects(s, slab,
5205			  "Objects remaining in %s on __kmem_cache_shutdown()");
5206		}
5207	}
5208	spin_unlock_irq(&n->list_lock);
5209
5210	list_for_each_entry_safe(slab, h, &discard, slab_list)
5211		discard_slab(s, slab);
5212}
5213
5214bool __kmem_cache_empty(struct kmem_cache *s)
5215{
5216	int node;
5217	struct kmem_cache_node *n;
5218
5219	for_each_kmem_cache_node(s, node, n)
5220		if (n->nr_partial || node_nr_slabs(n))
5221			return false;
5222	return true;
5223}
5224
5225/*
5226 * Release all resources used by a slab cache.
5227 */
5228int __kmem_cache_shutdown(struct kmem_cache *s)
5229{
5230	int node;
5231	struct kmem_cache_node *n;
5232
5233	flush_all_cpus_locked(s);
5234	/* Attempt to free all objects */
5235	for_each_kmem_cache_node(s, node, n) {
5236		free_partial(s, n);
5237		if (n->nr_partial || node_nr_slabs(n))
5238			return 1;
5239	}
5240	return 0;
5241}
5242
5243#ifdef CONFIG_PRINTK
5244void __kmem_obj_info(struct kmem_obj_info *kpp, void *object, struct slab *slab)
5245{
5246	void *base;
5247	int __maybe_unused i;
5248	unsigned int objnr;
5249	void *objp;
5250	void *objp0;
5251	struct kmem_cache *s = slab->slab_cache;
5252	struct track __maybe_unused *trackp;
5253
5254	kpp->kp_ptr = object;
5255	kpp->kp_slab = slab;
5256	kpp->kp_slab_cache = s;
5257	base = slab_address(slab);
5258	objp0 = kasan_reset_tag(object);
5259#ifdef CONFIG_SLUB_DEBUG
5260	objp = restore_red_left(s, objp0);
5261#else
5262	objp = objp0;
5263#endif
5264	objnr = obj_to_index(s, slab, objp);
5265	kpp->kp_data_offset = (unsigned long)((char *)objp0 - (char *)objp);
5266	objp = base + s->size * objnr;
5267	kpp->kp_objp = objp;
5268	if (WARN_ON_ONCE(objp < base || objp >= base + slab->objects * s->size
5269			 || (objp - base) % s->size) ||
5270	    !(s->flags & SLAB_STORE_USER))
5271		return;
5272#ifdef CONFIG_SLUB_DEBUG
5273	objp = fixup_red_left(s, objp);
5274	trackp = get_track(s, objp, TRACK_ALLOC);
5275	kpp->kp_ret = (void *)trackp->addr;
5276#ifdef CONFIG_STACKDEPOT
5277	{
5278		depot_stack_handle_t handle;
5279		unsigned long *entries;
5280		unsigned int nr_entries;
5281
5282		handle = READ_ONCE(trackp->handle);
5283		if (handle) {
5284			nr_entries = stack_depot_fetch(handle, &entries);
5285			for (i = 0; i < KS_ADDRS_COUNT && i < nr_entries; i++)
5286				kpp->kp_stack[i] = (void *)entries[i];
5287		}
5288
5289		trackp = get_track(s, objp, TRACK_FREE);
5290		handle = READ_ONCE(trackp->handle);
5291		if (handle) {
5292			nr_entries = stack_depot_fetch(handle, &entries);
5293			for (i = 0; i < KS_ADDRS_COUNT && i < nr_entries; i++)
5294				kpp->kp_free_stack[i] = (void *)entries[i];
5295		}
5296	}
5297#endif
5298#endif
5299}
5300#endif
5301
5302/********************************************************************
5303 *		Kmalloc subsystem
5304 *******************************************************************/
5305
5306static int __init setup_slub_min_order(char *str)
5307{
5308	get_option(&str, (int *)&slub_min_order);
5309
5310	if (slub_min_order > slub_max_order)
5311		slub_max_order = slub_min_order;
5312
5313	return 1;
5314}
5315
5316__setup("slub_min_order=", setup_slub_min_order);
 
 
5317
5318static int __init setup_slub_max_order(char *str)
5319{
5320	get_option(&str, (int *)&slub_max_order);
5321	slub_max_order = min_t(unsigned int, slub_max_order, MAX_PAGE_ORDER);
5322
5323	if (slub_min_order > slub_max_order)
5324		slub_min_order = slub_max_order;
5325
5326	return 1;
5327}
5328
5329__setup("slub_max_order=", setup_slub_max_order);
 
5330
5331static int __init setup_slub_min_objects(char *str)
5332{
5333	get_option(&str, (int *)&slub_min_objects);
5334
5335	return 1;
5336}
5337
5338__setup("slub_min_objects=", setup_slub_min_objects);
 
5339
5340#ifdef CONFIG_HARDENED_USERCOPY
5341/*
5342 * Rejects incorrectly sized objects and objects that are to be copied
5343 * to/from userspace but do not fall entirely within the containing slab
5344 * cache's usercopy region.
5345 *
5346 * Returns NULL if check passes, otherwise const char * to name of cache
5347 * to indicate an error.
5348 */
5349void __check_heap_object(const void *ptr, unsigned long n,
5350			 const struct slab *slab, bool to_user)
5351{
5352	struct kmem_cache *s;
5353	unsigned int offset;
5354	bool is_kfence = is_kfence_address(ptr);
5355
5356	ptr = kasan_reset_tag(ptr);
5357
5358	/* Find object and usable object size. */
5359	s = slab->slab_cache;
5360
5361	/* Reject impossible pointers. */
5362	if (ptr < slab_address(slab))
5363		usercopy_abort("SLUB object not in SLUB page?!", NULL,
5364			       to_user, 0, n);
5365
5366	/* Find offset within object. */
5367	if (is_kfence)
5368		offset = ptr - kfence_object_start(ptr);
5369	else
5370		offset = (ptr - slab_address(slab)) % s->size;
5371
5372	/* Adjust for redzone and reject if within the redzone. */
5373	if (!is_kfence && kmem_cache_debug_flags(s, SLAB_RED_ZONE)) {
5374		if (offset < s->red_left_pad)
5375			usercopy_abort("SLUB object in left red zone",
5376				       s->name, to_user, offset, n);
5377		offset -= s->red_left_pad;
5378	}
5379
5380	/* Allow address range falling entirely within usercopy region. */
5381	if (offset >= s->useroffset &&
5382	    offset - s->useroffset <= s->usersize &&
5383	    n <= s->useroffset - offset + s->usersize)
5384		return;
5385
5386	usercopy_abort("SLUB object", s->name, to_user, offset, n);
5387}
5388#endif /* CONFIG_HARDENED_USERCOPY */
5389
5390#define SHRINK_PROMOTE_MAX 32
5391
5392/*
5393 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
5394 * up most to the head of the partial lists. New allocations will then
5395 * fill those up and thus they can be removed from the partial lists.
5396 *
5397 * The slabs with the least items are placed last. This results in them
5398 * being allocated from last increasing the chance that the last objects
5399 * are freed in them.
5400 */
5401static int __kmem_cache_do_shrink(struct kmem_cache *s)
5402{
5403	int node;
5404	int i;
5405	struct kmem_cache_node *n;
5406	struct slab *slab;
5407	struct slab *t;
5408	struct list_head discard;
5409	struct list_head promote[SHRINK_PROMOTE_MAX];
5410	unsigned long flags;
5411	int ret = 0;
5412
5413	for_each_kmem_cache_node(s, node, n) {
5414		INIT_LIST_HEAD(&discard);
5415		for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
5416			INIT_LIST_HEAD(promote + i);
5417
5418		spin_lock_irqsave(&n->list_lock, flags);
5419
5420		/*
5421		 * Build lists of slabs to discard or promote.
5422		 *
5423		 * Note that concurrent frees may occur while we hold the
5424		 * list_lock. slab->inuse here is the upper limit.
5425		 */
5426		list_for_each_entry_safe(slab, t, &n->partial, slab_list) {
5427			int free = slab->objects - slab->inuse;
5428
5429			/* Do not reread slab->inuse */
5430			barrier();
5431
5432			/* We do not keep full slabs on the list */
5433			BUG_ON(free <= 0);
5434
5435			if (free == slab->objects) {
5436				list_move(&slab->slab_list, &discard);
5437				slab_clear_node_partial(slab);
5438				n->nr_partial--;
5439				dec_slabs_node(s, node, slab->objects);
5440			} else if (free <= SHRINK_PROMOTE_MAX)
5441				list_move(&slab->slab_list, promote + free - 1);
5442		}
5443
5444		/*
5445		 * Promote the slabs filled up most to the head of the
5446		 * partial list.
5447		 */
5448		for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
5449			list_splice(promote + i, &n->partial);
5450
5451		spin_unlock_irqrestore(&n->list_lock, flags);
5452
5453		/* Release empty slabs */
5454		list_for_each_entry_safe(slab, t, &discard, slab_list)
5455			free_slab(s, slab);
5456
5457		if (node_nr_slabs(n))
5458			ret = 1;
5459	}
5460
5461	return ret;
5462}
5463
5464int __kmem_cache_shrink(struct kmem_cache *s)
5465{
5466	flush_all(s);
5467	return __kmem_cache_do_shrink(s);
5468}
5469
5470static int slab_mem_going_offline_callback(void *arg)
5471{
5472	struct kmem_cache *s;
5473
5474	mutex_lock(&slab_mutex);
5475	list_for_each_entry(s, &slab_caches, list) {
5476		flush_all_cpus_locked(s);
5477		__kmem_cache_do_shrink(s);
5478	}
5479	mutex_unlock(&slab_mutex);
5480
5481	return 0;
5482}
5483
5484static void slab_mem_offline_callback(void *arg)
5485{
5486	struct memory_notify *marg = arg;
5487	int offline_node;
5488
5489	offline_node = marg->status_change_nid_normal;
5490
5491	/*
5492	 * If the node still has available memory. we need kmem_cache_node
5493	 * for it yet.
5494	 */
5495	if (offline_node < 0)
5496		return;
5497
5498	mutex_lock(&slab_mutex);
5499	node_clear(offline_node, slab_nodes);
5500	/*
5501	 * We no longer free kmem_cache_node structures here, as it would be
5502	 * racy with all get_node() users, and infeasible to protect them with
5503	 * slab_mutex.
5504	 */
5505	mutex_unlock(&slab_mutex);
5506}
5507
5508static int slab_mem_going_online_callback(void *arg)
5509{
5510	struct kmem_cache_node *n;
5511	struct kmem_cache *s;
5512	struct memory_notify *marg = arg;
5513	int nid = marg->status_change_nid_normal;
5514	int ret = 0;
5515
5516	/*
5517	 * If the node's memory is already available, then kmem_cache_node is
5518	 * already created. Nothing to do.
5519	 */
5520	if (nid < 0)
5521		return 0;
5522
5523	/*
5524	 * We are bringing a node online. No memory is available yet. We must
5525	 * allocate a kmem_cache_node structure in order to bring the node
5526	 * online.
5527	 */
5528	mutex_lock(&slab_mutex);
5529	list_for_each_entry(s, &slab_caches, list) {
5530		/*
5531		 * The structure may already exist if the node was previously
5532		 * onlined and offlined.
5533		 */
5534		if (get_node(s, nid))
5535			continue;
5536		/*
5537		 * XXX: kmem_cache_alloc_node will fallback to other nodes
5538		 *      since memory is not yet available from the node that
5539		 *      is brought up.
5540		 */
5541		n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
5542		if (!n) {
5543			ret = -ENOMEM;
5544			goto out;
5545		}
5546		init_kmem_cache_node(n);
5547		s->node[nid] = n;
5548	}
5549	/*
5550	 * Any cache created after this point will also have kmem_cache_node
5551	 * initialized for the new node.
5552	 */
5553	node_set(nid, slab_nodes);
5554out:
5555	mutex_unlock(&slab_mutex);
5556	return ret;
5557}
5558
5559static int slab_memory_callback(struct notifier_block *self,
5560				unsigned long action, void *arg)
5561{
5562	int ret = 0;
5563
5564	switch (action) {
5565	case MEM_GOING_ONLINE:
5566		ret = slab_mem_going_online_callback(arg);
5567		break;
5568	case MEM_GOING_OFFLINE:
5569		ret = slab_mem_going_offline_callback(arg);
5570		break;
5571	case MEM_OFFLINE:
5572	case MEM_CANCEL_ONLINE:
5573		slab_mem_offline_callback(arg);
5574		break;
5575	case MEM_ONLINE:
5576	case MEM_CANCEL_OFFLINE:
5577		break;
5578	}
5579	if (ret)
5580		ret = notifier_from_errno(ret);
5581	else
5582		ret = NOTIFY_OK;
5583	return ret;
5584}
5585
5586/********************************************************************
5587 *			Basic setup of slabs
5588 *******************************************************************/
5589
5590/*
5591 * Used for early kmem_cache structures that were allocated using
5592 * the page allocator. Allocate them properly then fix up the pointers
5593 * that may be pointing to the wrong kmem_cache structure.
5594 */
5595
5596static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
5597{
5598	int node;
5599	struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
5600	struct kmem_cache_node *n;
5601
5602	memcpy(s, static_cache, kmem_cache->object_size);
5603
5604	/*
5605	 * This runs very early, and only the boot processor is supposed to be
5606	 * up.  Even if it weren't true, IRQs are not up so we couldn't fire
5607	 * IPIs around.
5608	 */
5609	__flush_cpu_slab(s, smp_processor_id());
5610	for_each_kmem_cache_node(s, node, n) {
5611		struct slab *p;
5612
5613		list_for_each_entry(p, &n->partial, slab_list)
5614			p->slab_cache = s;
5615
5616#ifdef CONFIG_SLUB_DEBUG
5617		list_for_each_entry(p, &n->full, slab_list)
5618			p->slab_cache = s;
5619#endif
5620	}
5621	list_add(&s->list, &slab_caches);
5622	return s;
5623}
5624
5625void __init kmem_cache_init(void)
5626{
5627	static __initdata struct kmem_cache boot_kmem_cache,
5628		boot_kmem_cache_node;
5629	int node;
5630
5631	if (debug_guardpage_minorder())
5632		slub_max_order = 0;
5633
5634	/* Print slub debugging pointers without hashing */
5635	if (__slub_debug_enabled())
5636		no_hash_pointers_enable(NULL);
5637
5638	kmem_cache_node = &boot_kmem_cache_node;
5639	kmem_cache = &boot_kmem_cache;
5640
5641	/*
5642	 * Initialize the nodemask for which we will allocate per node
5643	 * structures. Here we don't need taking slab_mutex yet.
5644	 */
5645	for_each_node_state(node, N_NORMAL_MEMORY)
5646		node_set(node, slab_nodes);
5647
5648	create_boot_cache(kmem_cache_node, "kmem_cache_node",
5649		sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN, 0, 0);
5650
5651	hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
5652
5653	/* Able to allocate the per node structures */
5654	slab_state = PARTIAL;
5655
5656	create_boot_cache(kmem_cache, "kmem_cache",
5657			offsetof(struct kmem_cache, node) +
5658				nr_node_ids * sizeof(struct kmem_cache_node *),
5659		       SLAB_HWCACHE_ALIGN, 0, 0);
5660
5661	kmem_cache = bootstrap(&boot_kmem_cache);
5662	kmem_cache_node = bootstrap(&boot_kmem_cache_node);
5663
5664	/* Now we can use the kmem_cache to allocate kmalloc slabs */
5665	setup_kmalloc_cache_index_table();
5666	create_kmalloc_caches(0);
5667
5668	/* Setup random freelists for each cache */
5669	init_freelist_randomization();
5670
5671	cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL,
5672				  slub_cpu_dead);
5673
5674	pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n",
5675		cache_line_size(),
5676		slub_min_order, slub_max_order, slub_min_objects,
5677		nr_cpu_ids, nr_node_ids);
5678}
5679
5680void __init kmem_cache_init_late(void)
5681{
5682#ifndef CONFIG_SLUB_TINY
5683	flushwq = alloc_workqueue("slub_flushwq", WQ_MEM_RECLAIM, 0);
5684	WARN_ON(!flushwq);
5685#endif
5686}
5687
5688struct kmem_cache *
5689__kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
5690		   slab_flags_t flags, void (*ctor)(void *))
5691{
5692	struct kmem_cache *s;
5693
5694	s = find_mergeable(size, align, flags, name, ctor);
5695	if (s) {
5696		if (sysfs_slab_alias(s, name))
5697			return NULL;
5698
5699		s->refcount++;
5700
5701		/*
5702		 * Adjust the object sizes so that we clear
5703		 * the complete object on kzalloc.
5704		 */
5705		s->object_size = max(s->object_size, size);
5706		s->inuse = max(s->inuse, ALIGN(size, sizeof(void *)));
5707	}
5708
5709	return s;
5710}
5711
5712int __kmem_cache_create(struct kmem_cache *s, slab_flags_t flags)
5713{
5714	int err;
5715
5716	err = kmem_cache_open(s, flags);
5717	if (err)
5718		return err;
5719
5720	/* Mutex is not taken during early boot */
5721	if (slab_state <= UP)
5722		return 0;
5723
5724	err = sysfs_slab_add(s);
5725	if (err) {
5726		__kmem_cache_release(s);
5727		return err;
5728	}
5729
5730	if (s->flags & SLAB_STORE_USER)
5731		debugfs_slab_add(s);
5732
5733	return 0;
5734}
5735
5736#ifdef SLAB_SUPPORTS_SYSFS
5737static int count_inuse(struct slab *slab)
5738{
5739	return slab->inuse;
5740}
5741
5742static int count_total(struct slab *slab)
5743{
5744	return slab->objects;
5745}
5746#endif
5747
5748#ifdef CONFIG_SLUB_DEBUG
5749static void validate_slab(struct kmem_cache *s, struct slab *slab,
5750			  unsigned long *obj_map)
5751{
5752	void *p;
5753	void *addr = slab_address(slab);
5754
5755	if (!check_slab(s, slab) || !on_freelist(s, slab, NULL))
5756		return;
5757
5758	/* Now we know that a valid freelist exists */
5759	__fill_map(obj_map, s, slab);
5760	for_each_object(p, s, addr, slab->objects) {
5761		u8 val = test_bit(__obj_to_index(s, addr, p), obj_map) ?
5762			 SLUB_RED_INACTIVE : SLUB_RED_ACTIVE;
5763
5764		if (!check_object(s, slab, p, val))
5765			break;
5766	}
5767}
5768
5769static int validate_slab_node(struct kmem_cache *s,
5770		struct kmem_cache_node *n, unsigned long *obj_map)
5771{
5772	unsigned long count = 0;
5773	struct slab *slab;
5774	unsigned long flags;
5775
5776	spin_lock_irqsave(&n->list_lock, flags);
5777
5778	list_for_each_entry(slab, &n->partial, slab_list) {
5779		validate_slab(s, slab, obj_map);
5780		count++;
5781	}
5782	if (count != n->nr_partial) {
5783		pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
5784		       s->name, count, n->nr_partial);
5785		slab_add_kunit_errors();
5786	}
5787
5788	if (!(s->flags & SLAB_STORE_USER))
5789		goto out;
5790
5791	list_for_each_entry(slab, &n->full, slab_list) {
5792		validate_slab(s, slab, obj_map);
5793		count++;
5794	}
5795	if (count != node_nr_slabs(n)) {
5796		pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
5797		       s->name, count, node_nr_slabs(n));
5798		slab_add_kunit_errors();
5799	}
5800
5801out:
5802	spin_unlock_irqrestore(&n->list_lock, flags);
5803	return count;
5804}
5805
5806long validate_slab_cache(struct kmem_cache *s)
5807{
5808	int node;
5809	unsigned long count = 0;
5810	struct kmem_cache_node *n;
5811	unsigned long *obj_map;
5812
5813	obj_map = bitmap_alloc(oo_objects(s->oo), GFP_KERNEL);
5814	if (!obj_map)
5815		return -ENOMEM;
5816
5817	flush_all(s);
5818	for_each_kmem_cache_node(s, node, n)
5819		count += validate_slab_node(s, n, obj_map);
5820
5821	bitmap_free(obj_map);
5822
5823	return count;
5824}
5825EXPORT_SYMBOL(validate_slab_cache);
5826
5827#ifdef CONFIG_DEBUG_FS
5828/*
5829 * Generate lists of code addresses where slabcache objects are allocated
5830 * and freed.
5831 */
5832
5833struct location {
5834	depot_stack_handle_t handle;
5835	unsigned long count;
5836	unsigned long addr;
5837	unsigned long waste;
5838	long long sum_time;
5839	long min_time;
5840	long max_time;
5841	long min_pid;
5842	long max_pid;
5843	DECLARE_BITMAP(cpus, NR_CPUS);
5844	nodemask_t nodes;
5845};
5846
5847struct loc_track {
5848	unsigned long max;
5849	unsigned long count;
5850	struct location *loc;
5851	loff_t idx;
5852};
5853
5854static struct dentry *slab_debugfs_root;
5855
5856static void free_loc_track(struct loc_track *t)
5857{
5858	if (t->max)
5859		free_pages((unsigned long)t->loc,
5860			get_order(sizeof(struct location) * t->max));
5861}
5862
5863static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
5864{
5865	struct location *l;
5866	int order;
5867
5868	order = get_order(sizeof(struct location) * max);
5869
5870	l = (void *)__get_free_pages(flags, order);
5871	if (!l)
5872		return 0;
5873
5874	if (t->count) {
5875		memcpy(l, t->loc, sizeof(struct location) * t->count);
5876		free_loc_track(t);
5877	}
5878	t->max = max;
5879	t->loc = l;
5880	return 1;
5881}
5882
5883static int add_location(struct loc_track *t, struct kmem_cache *s,
5884				const struct track *track,
5885				unsigned int orig_size)
5886{
5887	long start, end, pos;
5888	struct location *l;
5889	unsigned long caddr, chandle, cwaste;
5890	unsigned long age = jiffies - track->when;
5891	depot_stack_handle_t handle = 0;
5892	unsigned int waste = s->object_size - orig_size;
5893
5894#ifdef CONFIG_STACKDEPOT
5895	handle = READ_ONCE(track->handle);
5896#endif
5897	start = -1;
5898	end = t->count;
5899
5900	for ( ; ; ) {
5901		pos = start + (end - start + 1) / 2;
5902
5903		/*
5904		 * There is nothing at "end". If we end up there
5905		 * we need to add something to before end.
5906		 */
5907		if (pos == end)
5908			break;
5909
5910		l = &t->loc[pos];
5911		caddr = l->addr;
5912		chandle = l->handle;
5913		cwaste = l->waste;
5914		if ((track->addr == caddr) && (handle == chandle) &&
5915			(waste == cwaste)) {
5916
5917			l->count++;
5918			if (track->when) {
5919				l->sum_time += age;
5920				if (age < l->min_time)
5921					l->min_time = age;
5922				if (age > l->max_time)
5923					l->max_time = age;
5924
5925				if (track->pid < l->min_pid)
5926					l->min_pid = track->pid;
5927				if (track->pid > l->max_pid)
5928					l->max_pid = track->pid;
5929
5930				cpumask_set_cpu(track->cpu,
5931						to_cpumask(l->cpus));
5932			}
5933			node_set(page_to_nid(virt_to_page(track)), l->nodes);
5934			return 1;
5935		}
5936
5937		if (track->addr < caddr)
5938			end = pos;
5939		else if (track->addr == caddr && handle < chandle)
5940			end = pos;
5941		else if (track->addr == caddr && handle == chandle &&
5942				waste < cwaste)
5943			end = pos;
5944		else
5945			start = pos;
5946	}
5947
5948	/*
5949	 * Not found. Insert new tracking element.
5950	 */
5951	if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
5952		return 0;
5953
5954	l = t->loc + pos;
5955	if (pos < t->count)
5956		memmove(l + 1, l,
5957			(t->count - pos) * sizeof(struct location));
5958	t->count++;
5959	l->count = 1;
5960	l->addr = track->addr;
5961	l->sum_time = age;
5962	l->min_time = age;
5963	l->max_time = age;
5964	l->min_pid = track->pid;
5965	l->max_pid = track->pid;
5966	l->handle = handle;
5967	l->waste = waste;
5968	cpumask_clear(to_cpumask(l->cpus));
5969	cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
5970	nodes_clear(l->nodes);
5971	node_set(page_to_nid(virt_to_page(track)), l->nodes);
5972	return 1;
5973}
5974
5975static void process_slab(struct loc_track *t, struct kmem_cache *s,
5976		struct slab *slab, enum track_item alloc,
5977		unsigned long *obj_map)
5978{
5979	void *addr = slab_address(slab);
5980	bool is_alloc = (alloc == TRACK_ALLOC);
5981	void *p;
5982
5983	__fill_map(obj_map, s, slab);
5984
5985	for_each_object(p, s, addr, slab->objects)
5986		if (!test_bit(__obj_to_index(s, addr, p), obj_map))
5987			add_location(t, s, get_track(s, p, alloc),
5988				     is_alloc ? get_orig_size(s, p) :
5989						s->object_size);
5990}
5991#endif  /* CONFIG_DEBUG_FS   */
5992#endif	/* CONFIG_SLUB_DEBUG */
5993
5994#ifdef SLAB_SUPPORTS_SYSFS
5995enum slab_stat_type {
5996	SL_ALL,			/* All slabs */
5997	SL_PARTIAL,		/* Only partially allocated slabs */
5998	SL_CPU,			/* Only slabs used for cpu caches */
5999	SL_OBJECTS,		/* Determine allocated objects not slabs */
6000	SL_TOTAL		/* Determine object capacity not slabs */
6001};
6002
6003#define SO_ALL		(1 << SL_ALL)
6004#define SO_PARTIAL	(1 << SL_PARTIAL)
6005#define SO_CPU		(1 << SL_CPU)
6006#define SO_OBJECTS	(1 << SL_OBJECTS)
6007#define SO_TOTAL	(1 << SL_TOTAL)
6008
6009static ssize_t show_slab_objects(struct kmem_cache *s,
6010				 char *buf, unsigned long flags)
6011{
6012	unsigned long total = 0;
6013	int node;
6014	int x;
6015	unsigned long *nodes;
6016	int len = 0;
6017
6018	nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL);
6019	if (!nodes)
6020		return -ENOMEM;
6021
6022	if (flags & SO_CPU) {
6023		int cpu;
6024
6025		for_each_possible_cpu(cpu) {
6026			struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
6027							       cpu);
6028			int node;
6029			struct slab *slab;
6030
6031			slab = READ_ONCE(c->slab);
6032			if (!slab)
6033				continue;
6034
6035			node = slab_nid(slab);
6036			if (flags & SO_TOTAL)
6037				x = slab->objects;
6038			else if (flags & SO_OBJECTS)
6039				x = slab->inuse;
6040			else
6041				x = 1;
6042
6043			total += x;
6044			nodes[node] += x;
6045
6046#ifdef CONFIG_SLUB_CPU_PARTIAL
6047			slab = slub_percpu_partial_read_once(c);
6048			if (slab) {
6049				node = slab_nid(slab);
6050				if (flags & SO_TOTAL)
6051					WARN_ON_ONCE(1);
6052				else if (flags & SO_OBJECTS)
6053					WARN_ON_ONCE(1);
6054				else
6055					x = slab->slabs;
6056				total += x;
6057				nodes[node] += x;
6058			}
6059#endif
6060		}
6061	}
6062
6063	/*
6064	 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
6065	 * already held which will conflict with an existing lock order:
6066	 *
6067	 * mem_hotplug_lock->slab_mutex->kernfs_mutex
6068	 *
6069	 * We don't really need mem_hotplug_lock (to hold off
6070	 * slab_mem_going_offline_callback) here because slab's memory hot
6071	 * unplug code doesn't destroy the kmem_cache->node[] data.
6072	 */
6073
6074#ifdef CONFIG_SLUB_DEBUG
6075	if (flags & SO_ALL) {
6076		struct kmem_cache_node *n;
6077
6078		for_each_kmem_cache_node(s, node, n) {
6079
6080			if (flags & SO_TOTAL)
6081				x = node_nr_objs(n);
6082			else if (flags & SO_OBJECTS)
6083				x = node_nr_objs(n) - count_partial(n, count_free);
6084			else
6085				x = node_nr_slabs(n);
6086			total += x;
6087			nodes[node] += x;
6088		}
6089
6090	} else
6091#endif
6092	if (flags & SO_PARTIAL) {
6093		struct kmem_cache_node *n;
6094
6095		for_each_kmem_cache_node(s, node, n) {
6096			if (flags & SO_TOTAL)
6097				x = count_partial(n, count_total);
6098			else if (flags & SO_OBJECTS)
6099				x = count_partial(n, count_inuse);
6100			else
6101				x = n->nr_partial;
6102			total += x;
6103			nodes[node] += x;
6104		}
6105	}
6106
6107	len += sysfs_emit_at(buf, len, "%lu", total);
6108#ifdef CONFIG_NUMA
6109	for (node = 0; node < nr_node_ids; node++) {
6110		if (nodes[node])
6111			len += sysfs_emit_at(buf, len, " N%d=%lu",
6112					     node, nodes[node]);
6113	}
6114#endif
6115	len += sysfs_emit_at(buf, len, "\n");
6116	kfree(nodes);
6117
6118	return len;
6119}
6120
6121#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
6122#define to_slab(n) container_of(n, struct kmem_cache, kobj)
6123
6124struct slab_attribute {
6125	struct attribute attr;
6126	ssize_t (*show)(struct kmem_cache *s, char *buf);
6127	ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
6128};
6129
6130#define SLAB_ATTR_RO(_name) \
6131	static struct slab_attribute _name##_attr = __ATTR_RO_MODE(_name, 0400)
6132
6133#define SLAB_ATTR(_name) \
6134	static struct slab_attribute _name##_attr = __ATTR_RW_MODE(_name, 0600)
6135
6136static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
6137{
6138	return sysfs_emit(buf, "%u\n", s->size);
6139}
6140SLAB_ATTR_RO(slab_size);
6141
6142static ssize_t align_show(struct kmem_cache *s, char *buf)
6143{
6144	return sysfs_emit(buf, "%u\n", s->align);
6145}
6146SLAB_ATTR_RO(align);
6147
6148static ssize_t object_size_show(struct kmem_cache *s, char *buf)
6149{
6150	return sysfs_emit(buf, "%u\n", s->object_size);
6151}
6152SLAB_ATTR_RO(object_size);
6153
6154static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
6155{
6156	return sysfs_emit(buf, "%u\n", oo_objects(s->oo));
6157}
6158SLAB_ATTR_RO(objs_per_slab);
6159
6160static ssize_t order_show(struct kmem_cache *s, char *buf)
6161{
6162	return sysfs_emit(buf, "%u\n", oo_order(s->oo));
6163}
6164SLAB_ATTR_RO(order);
6165
6166static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
6167{
6168	return sysfs_emit(buf, "%lu\n", s->min_partial);
6169}
6170
6171static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
6172				 size_t length)
6173{
6174	unsigned long min;
6175	int err;
6176
6177	err = kstrtoul(buf, 10, &min);
6178	if (err)
6179		return err;
6180
6181	s->min_partial = min;
6182	return length;
6183}
6184SLAB_ATTR(min_partial);
6185
6186static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
6187{
6188	unsigned int nr_partial = 0;
6189#ifdef CONFIG_SLUB_CPU_PARTIAL
6190	nr_partial = s->cpu_partial;
6191#endif
6192
6193	return sysfs_emit(buf, "%u\n", nr_partial);
6194}
6195
6196static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
6197				 size_t length)
6198{
6199	unsigned int objects;
6200	int err;
6201
6202	err = kstrtouint(buf, 10, &objects);
6203	if (err)
6204		return err;
6205	if (objects && !kmem_cache_has_cpu_partial(s))
6206		return -EINVAL;
6207
6208	slub_set_cpu_partial(s, objects);
6209	flush_all(s);
6210	return length;
6211}
6212SLAB_ATTR(cpu_partial);
6213
6214static ssize_t ctor_show(struct kmem_cache *s, char *buf)
6215{
6216	if (!s->ctor)
6217		return 0;
6218	return sysfs_emit(buf, "%pS\n", s->ctor);
6219}
6220SLAB_ATTR_RO(ctor);
6221
6222static ssize_t aliases_show(struct kmem_cache *s, char *buf)
6223{
6224	return sysfs_emit(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
6225}
6226SLAB_ATTR_RO(aliases);
6227
6228static ssize_t partial_show(struct kmem_cache *s, char *buf)
6229{
6230	return show_slab_objects(s, buf, SO_PARTIAL);
6231}
6232SLAB_ATTR_RO(partial);
6233
6234static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
6235{
6236	return show_slab_objects(s, buf, SO_CPU);
6237}
6238SLAB_ATTR_RO(cpu_slabs);
6239
6240static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
6241{
6242	return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
6243}
6244SLAB_ATTR_RO(objects_partial);
6245
6246static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
6247{
6248	int objects = 0;
6249	int slabs = 0;
6250	int cpu __maybe_unused;
6251	int len = 0;
6252
6253#ifdef CONFIG_SLUB_CPU_PARTIAL
6254	for_each_online_cpu(cpu) {
6255		struct slab *slab;
6256
6257		slab = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
6258
6259		if (slab)
6260			slabs += slab->slabs;
6261	}
6262#endif
6263
6264	/* Approximate half-full slabs, see slub_set_cpu_partial() */
6265	objects = (slabs * oo_objects(s->oo)) / 2;
6266	len += sysfs_emit_at(buf, len, "%d(%d)", objects, slabs);
6267
6268#ifdef CONFIG_SLUB_CPU_PARTIAL
6269	for_each_online_cpu(cpu) {
6270		struct slab *slab;
6271
6272		slab = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
6273		if (slab) {
6274			slabs = READ_ONCE(slab->slabs);
6275			objects = (slabs * oo_objects(s->oo)) / 2;
6276			len += sysfs_emit_at(buf, len, " C%d=%d(%d)",
6277					     cpu, objects, slabs);
6278		}
6279	}
6280#endif
6281	len += sysfs_emit_at(buf, len, "\n");
6282
6283	return len;
6284}
6285SLAB_ATTR_RO(slabs_cpu_partial);
6286
6287static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
6288{
6289	return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
6290}
6291SLAB_ATTR_RO(reclaim_account);
6292
6293static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
6294{
6295	return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
6296}
6297SLAB_ATTR_RO(hwcache_align);
6298
6299#ifdef CONFIG_ZONE_DMA
6300static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
6301{
6302	return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
6303}
6304SLAB_ATTR_RO(cache_dma);
6305#endif
6306
6307#ifdef CONFIG_HARDENED_USERCOPY
6308static ssize_t usersize_show(struct kmem_cache *s, char *buf)
6309{
6310	return sysfs_emit(buf, "%u\n", s->usersize);
6311}
6312SLAB_ATTR_RO(usersize);
6313#endif
6314
6315static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
6316{
6317	return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU));
6318}
6319SLAB_ATTR_RO(destroy_by_rcu);
6320
6321#ifdef CONFIG_SLUB_DEBUG
6322static ssize_t slabs_show(struct kmem_cache *s, char *buf)
6323{
6324	return show_slab_objects(s, buf, SO_ALL);
6325}
6326SLAB_ATTR_RO(slabs);
6327
6328static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
6329{
6330	return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
6331}
6332SLAB_ATTR_RO(total_objects);
6333
6334static ssize_t objects_show(struct kmem_cache *s, char *buf)
6335{
6336	return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
6337}
6338SLAB_ATTR_RO(objects);
6339
6340static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
6341{
6342	return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
6343}
6344SLAB_ATTR_RO(sanity_checks);
6345
6346static ssize_t trace_show(struct kmem_cache *s, char *buf)
6347{
6348	return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TRACE));
6349}
6350SLAB_ATTR_RO(trace);
6351
6352static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
6353{
6354	return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
6355}
6356
6357SLAB_ATTR_RO(red_zone);
6358
6359static ssize_t poison_show(struct kmem_cache *s, char *buf)
6360{
6361	return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_POISON));
6362}
6363
6364SLAB_ATTR_RO(poison);
6365
6366static ssize_t store_user_show(struct kmem_cache *s, char *buf)
6367{
6368	return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
6369}
6370
6371SLAB_ATTR_RO(store_user);
6372
6373static ssize_t validate_show(struct kmem_cache *s, char *buf)
6374{
6375	return 0;
6376}
6377
6378static ssize_t validate_store(struct kmem_cache *s,
6379			const char *buf, size_t length)
6380{
6381	int ret = -EINVAL;
6382
6383	if (buf[0] == '1' && kmem_cache_debug(s)) {
6384		ret = validate_slab_cache(s);
6385		if (ret >= 0)
6386			ret = length;
6387	}
6388	return ret;
6389}
6390SLAB_ATTR(validate);
6391
6392#endif /* CONFIG_SLUB_DEBUG */
6393
6394#ifdef CONFIG_FAILSLAB
6395static ssize_t failslab_show(struct kmem_cache *s, char *buf)
6396{
6397	return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
6398}
6399
6400static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
6401				size_t length)
6402{
6403	if (s->refcount > 1)
6404		return -EINVAL;
6405
6406	if (buf[0] == '1')
6407		WRITE_ONCE(s->flags, s->flags | SLAB_FAILSLAB);
6408	else
6409		WRITE_ONCE(s->flags, s->flags & ~SLAB_FAILSLAB);
6410
6411	return length;
6412}
6413SLAB_ATTR(failslab);
6414#endif
6415
6416static ssize_t shrink_show(struct kmem_cache *s, char *buf)
6417{
6418	return 0;
6419}
6420
6421static ssize_t shrink_store(struct kmem_cache *s,
6422			const char *buf, size_t length)
6423{
6424	if (buf[0] == '1')
6425		kmem_cache_shrink(s);
6426	else
6427		return -EINVAL;
6428	return length;
6429}
6430SLAB_ATTR(shrink);
6431
6432#ifdef CONFIG_NUMA
6433static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
6434{
6435	return sysfs_emit(buf, "%u\n", s->remote_node_defrag_ratio / 10);
6436}
6437
6438static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
6439				const char *buf, size_t length)
6440{
6441	unsigned int ratio;
6442	int err;
6443
6444	err = kstrtouint(buf, 10, &ratio);
6445	if (err)
6446		return err;
6447	if (ratio > 100)
6448		return -ERANGE;
6449
6450	s->remote_node_defrag_ratio = ratio * 10;
6451
6452	return length;
6453}
6454SLAB_ATTR(remote_node_defrag_ratio);
6455#endif
6456
6457#ifdef CONFIG_SLUB_STATS
6458static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
6459{
6460	unsigned long sum  = 0;
6461	int cpu;
6462	int len = 0;
6463	int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL);
6464
6465	if (!data)
6466		return -ENOMEM;
6467
6468	for_each_online_cpu(cpu) {
6469		unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
6470
6471		data[cpu] = x;
6472		sum += x;
6473	}
6474
6475	len += sysfs_emit_at(buf, len, "%lu", sum);
6476
6477#ifdef CONFIG_SMP
6478	for_each_online_cpu(cpu) {
6479		if (data[cpu])
6480			len += sysfs_emit_at(buf, len, " C%d=%u",
6481					     cpu, data[cpu]);
6482	}
6483#endif
6484	kfree(data);
6485	len += sysfs_emit_at(buf, len, "\n");
6486
6487	return len;
6488}
6489
6490static void clear_stat(struct kmem_cache *s, enum stat_item si)
6491{
6492	int cpu;
6493
6494	for_each_online_cpu(cpu)
6495		per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
6496}
6497
6498#define STAT_ATTR(si, text) 					\
6499static ssize_t text##_show(struct kmem_cache *s, char *buf)	\
6500{								\
6501	return show_stat(s, buf, si);				\
6502}								\
6503static ssize_t text##_store(struct kmem_cache *s,		\
6504				const char *buf, size_t length)	\
6505{								\
6506	if (buf[0] != '0')					\
6507		return -EINVAL;					\
6508	clear_stat(s, si);					\
6509	return length;						\
6510}								\
6511SLAB_ATTR(text);						\
6512
6513STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
6514STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
6515STAT_ATTR(FREE_FASTPATH, free_fastpath);
6516STAT_ATTR(FREE_SLOWPATH, free_slowpath);
6517STAT_ATTR(FREE_FROZEN, free_frozen);
6518STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
6519STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
6520STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
6521STAT_ATTR(ALLOC_SLAB, alloc_slab);
6522STAT_ATTR(ALLOC_REFILL, alloc_refill);
6523STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
6524STAT_ATTR(FREE_SLAB, free_slab);
6525STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
6526STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
6527STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
6528STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
6529STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
6530STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
6531STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
6532STAT_ATTR(ORDER_FALLBACK, order_fallback);
6533STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
6534STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
6535STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
6536STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
6537STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
6538STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
6539#endif	/* CONFIG_SLUB_STATS */
6540
6541#ifdef CONFIG_KFENCE
6542static ssize_t skip_kfence_show(struct kmem_cache *s, char *buf)
6543{
6544	return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_SKIP_KFENCE));
6545}
6546
6547static ssize_t skip_kfence_store(struct kmem_cache *s,
6548			const char *buf, size_t length)
6549{
6550	int ret = length;
6551
6552	if (buf[0] == '0')
6553		s->flags &= ~SLAB_SKIP_KFENCE;
6554	else if (buf[0] == '1')
6555		s->flags |= SLAB_SKIP_KFENCE;
6556	else
6557		ret = -EINVAL;
6558
6559	return ret;
6560}
6561SLAB_ATTR(skip_kfence);
6562#endif
6563
6564static struct attribute *slab_attrs[] = {
6565	&slab_size_attr.attr,
6566	&object_size_attr.attr,
6567	&objs_per_slab_attr.attr,
6568	&order_attr.attr,
6569	&min_partial_attr.attr,
6570	&cpu_partial_attr.attr,
6571	&objects_partial_attr.attr,
6572	&partial_attr.attr,
6573	&cpu_slabs_attr.attr,
6574	&ctor_attr.attr,
6575	&aliases_attr.attr,
6576	&align_attr.attr,
6577	&hwcache_align_attr.attr,
6578	&reclaim_account_attr.attr,
6579	&destroy_by_rcu_attr.attr,
6580	&shrink_attr.attr,
6581	&slabs_cpu_partial_attr.attr,
6582#ifdef CONFIG_SLUB_DEBUG
6583	&total_objects_attr.attr,
6584	&objects_attr.attr,
6585	&slabs_attr.attr,
6586	&sanity_checks_attr.attr,
6587	&trace_attr.attr,
6588	&red_zone_attr.attr,
6589	&poison_attr.attr,
6590	&store_user_attr.attr,
6591	&validate_attr.attr,
6592#endif
6593#ifdef CONFIG_ZONE_DMA
6594	&cache_dma_attr.attr,
6595#endif
6596#ifdef CONFIG_NUMA
6597	&remote_node_defrag_ratio_attr.attr,
6598#endif
6599#ifdef CONFIG_SLUB_STATS
6600	&alloc_fastpath_attr.attr,
6601	&alloc_slowpath_attr.attr,
6602	&free_fastpath_attr.attr,
6603	&free_slowpath_attr.attr,
6604	&free_frozen_attr.attr,
6605	&free_add_partial_attr.attr,
6606	&free_remove_partial_attr.attr,
6607	&alloc_from_partial_attr.attr,
6608	&alloc_slab_attr.attr,
6609	&alloc_refill_attr.attr,
6610	&alloc_node_mismatch_attr.attr,
6611	&free_slab_attr.attr,
6612	&cpuslab_flush_attr.attr,
6613	&deactivate_full_attr.attr,
6614	&deactivate_empty_attr.attr,
6615	&deactivate_to_head_attr.attr,
6616	&deactivate_to_tail_attr.attr,
6617	&deactivate_remote_frees_attr.attr,
6618	&deactivate_bypass_attr.attr,
6619	&order_fallback_attr.attr,
6620	&cmpxchg_double_fail_attr.attr,
6621	&cmpxchg_double_cpu_fail_attr.attr,
6622	&cpu_partial_alloc_attr.attr,
6623	&cpu_partial_free_attr.attr,
6624	&cpu_partial_node_attr.attr,
6625	&cpu_partial_drain_attr.attr,
6626#endif
6627#ifdef CONFIG_FAILSLAB
6628	&failslab_attr.attr,
6629#endif
6630#ifdef CONFIG_HARDENED_USERCOPY
6631	&usersize_attr.attr,
6632#endif
6633#ifdef CONFIG_KFENCE
6634	&skip_kfence_attr.attr,
6635#endif
6636
6637	NULL
6638};
6639
6640static const struct attribute_group slab_attr_group = {
6641	.attrs = slab_attrs,
6642};
6643
6644static ssize_t slab_attr_show(struct kobject *kobj,
6645				struct attribute *attr,
6646				char *buf)
6647{
6648	struct slab_attribute *attribute;
6649	struct kmem_cache *s;
6650
6651	attribute = to_slab_attr(attr);
6652	s = to_slab(kobj);
6653
6654	if (!attribute->show)
6655		return -EIO;
6656
6657	return attribute->show(s, buf);
6658}
6659
6660static ssize_t slab_attr_store(struct kobject *kobj,
6661				struct attribute *attr,
6662				const char *buf, size_t len)
6663{
6664	struct slab_attribute *attribute;
6665	struct kmem_cache *s;
6666
6667	attribute = to_slab_attr(attr);
6668	s = to_slab(kobj);
6669
6670	if (!attribute->store)
6671		return -EIO;
6672
6673	return attribute->store(s, buf, len);
6674}
6675
6676static void kmem_cache_release(struct kobject *k)
6677{
6678	slab_kmem_cache_release(to_slab(k));
6679}
6680
6681static const struct sysfs_ops slab_sysfs_ops = {
6682	.show = slab_attr_show,
6683	.store = slab_attr_store,
6684};
6685
6686static const struct kobj_type slab_ktype = {
6687	.sysfs_ops = &slab_sysfs_ops,
6688	.release = kmem_cache_release,
6689};
6690
6691static struct kset *slab_kset;
6692
6693static inline struct kset *cache_kset(struct kmem_cache *s)
6694{
6695	return slab_kset;
6696}
6697
6698#define ID_STR_LENGTH 32
6699
6700/* Create a unique string id for a slab cache:
6701 *
6702 * Format	:[flags-]size
6703 */
6704static char *create_unique_id(struct kmem_cache *s)
6705{
6706	char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
6707	char *p = name;
6708
6709	if (!name)
6710		return ERR_PTR(-ENOMEM);
6711
6712	*p++ = ':';
6713	/*
6714	 * First flags affecting slabcache operations. We will only
6715	 * get here for aliasable slabs so we do not need to support
6716	 * too many flags. The flags here must cover all flags that
6717	 * are matched during merging to guarantee that the id is
6718	 * unique.
6719	 */
6720	if (s->flags & SLAB_CACHE_DMA)
6721		*p++ = 'd';
6722	if (s->flags & SLAB_CACHE_DMA32)
6723		*p++ = 'D';
6724	if (s->flags & SLAB_RECLAIM_ACCOUNT)
6725		*p++ = 'a';
6726	if (s->flags & SLAB_CONSISTENCY_CHECKS)
6727		*p++ = 'F';
6728	if (s->flags & SLAB_ACCOUNT)
6729		*p++ = 'A';
6730	if (p != name + 1)
6731		*p++ = '-';
6732	p += snprintf(p, ID_STR_LENGTH - (p - name), "%07u", s->size);
6733
6734	if (WARN_ON(p > name + ID_STR_LENGTH - 1)) {
6735		kfree(name);
6736		return ERR_PTR(-EINVAL);
6737	}
6738	kmsan_unpoison_memory(name, p - name);
6739	return name;
6740}
6741
6742static int sysfs_slab_add(struct kmem_cache *s)
6743{
6744	int err;
6745	const char *name;
6746	struct kset *kset = cache_kset(s);
6747	int unmergeable = slab_unmergeable(s);
6748
6749	if (!unmergeable && disable_higher_order_debug &&
6750			(slub_debug & DEBUG_METADATA_FLAGS))
6751		unmergeable = 1;
6752
6753	if (unmergeable) {
6754		/*
6755		 * Slabcache can never be merged so we can use the name proper.
6756		 * This is typically the case for debug situations. In that
6757		 * case we can catch duplicate names easily.
6758		 */
6759		sysfs_remove_link(&slab_kset->kobj, s->name);
6760		name = s->name;
6761	} else {
6762		/*
6763		 * Create a unique name for the slab as a target
6764		 * for the symlinks.
6765		 */
6766		name = create_unique_id(s);
6767		if (IS_ERR(name))
6768			return PTR_ERR(name);
6769	}
6770
6771	s->kobj.kset = kset;
6772	err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
6773	if (err)
6774		goto out;
6775
6776	err = sysfs_create_group(&s->kobj, &slab_attr_group);
6777	if (err)
6778		goto out_del_kobj;
6779
6780	if (!unmergeable) {
6781		/* Setup first alias */
6782		sysfs_slab_alias(s, s->name);
6783	}
6784out:
6785	if (!unmergeable)
6786		kfree(name);
6787	return err;
6788out_del_kobj:
6789	kobject_del(&s->kobj);
6790	goto out;
6791}
6792
6793void sysfs_slab_unlink(struct kmem_cache *s)
6794{
6795	if (slab_state >= FULL)
6796		kobject_del(&s->kobj);
6797}
6798
6799void sysfs_slab_release(struct kmem_cache *s)
6800{
6801	if (slab_state >= FULL)
6802		kobject_put(&s->kobj);
6803}
6804
6805/*
6806 * Need to buffer aliases during bootup until sysfs becomes
6807 * available lest we lose that information.
6808 */
6809struct saved_alias {
6810	struct kmem_cache *s;
6811	const char *name;
6812	struct saved_alias *next;
6813};
6814
6815static struct saved_alias *alias_list;
6816
6817static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
6818{
6819	struct saved_alias *al;
6820
6821	if (slab_state == FULL) {
6822		/*
6823		 * If we have a leftover link then remove it.
6824		 */
6825		sysfs_remove_link(&slab_kset->kobj, name);
6826		return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
6827	}
6828
6829	al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
6830	if (!al)
6831		return -ENOMEM;
6832
6833	al->s = s;
6834	al->name = name;
6835	al->next = alias_list;
6836	alias_list = al;
6837	kmsan_unpoison_memory(al, sizeof(*al));
6838	return 0;
6839}
6840
6841static int __init slab_sysfs_init(void)
6842{
6843	struct kmem_cache *s;
6844	int err;
6845
6846	mutex_lock(&slab_mutex);
6847
6848	slab_kset = kset_create_and_add("slab", NULL, kernel_kobj);
6849	if (!slab_kset) {
6850		mutex_unlock(&slab_mutex);
6851		pr_err("Cannot register slab subsystem.\n");
6852		return -ENOMEM;
6853	}
6854
6855	slab_state = FULL;
6856
6857	list_for_each_entry(s, &slab_caches, list) {
6858		err = sysfs_slab_add(s);
6859		if (err)
6860			pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
6861			       s->name);
6862	}
6863
6864	while (alias_list) {
6865		struct saved_alias *al = alias_list;
6866
6867		alias_list = alias_list->next;
6868		err = sysfs_slab_alias(al->s, al->name);
6869		if (err)
6870			pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
6871			       al->name);
6872		kfree(al);
6873	}
6874
6875	mutex_unlock(&slab_mutex);
6876	return 0;
6877}
6878late_initcall(slab_sysfs_init);
6879#endif /* SLAB_SUPPORTS_SYSFS */
6880
6881#if defined(CONFIG_SLUB_DEBUG) && defined(CONFIG_DEBUG_FS)
6882static int slab_debugfs_show(struct seq_file *seq, void *v)
6883{
6884	struct loc_track *t = seq->private;
6885	struct location *l;
6886	unsigned long idx;
6887
6888	idx = (unsigned long) t->idx;
6889	if (idx < t->count) {
6890		l = &t->loc[idx];
6891
6892		seq_printf(seq, "%7ld ", l->count);
6893
6894		if (l->addr)
6895			seq_printf(seq, "%pS", (void *)l->addr);
6896		else
6897			seq_puts(seq, "<not-available>");
6898
6899		if (l->waste)
6900			seq_printf(seq, " waste=%lu/%lu",
6901				l->count * l->waste, l->waste);
6902
6903		if (l->sum_time != l->min_time) {
6904			seq_printf(seq, " age=%ld/%llu/%ld",
6905				l->min_time, div_u64(l->sum_time, l->count),
6906				l->max_time);
6907		} else
6908			seq_printf(seq, " age=%ld", l->min_time);
6909
6910		if (l->min_pid != l->max_pid)
6911			seq_printf(seq, " pid=%ld-%ld", l->min_pid, l->max_pid);
6912		else
6913			seq_printf(seq, " pid=%ld",
6914				l->min_pid);
6915
6916		if (num_online_cpus() > 1 && !cpumask_empty(to_cpumask(l->cpus)))
6917			seq_printf(seq, " cpus=%*pbl",
6918				 cpumask_pr_args(to_cpumask(l->cpus)));
6919
6920		if (nr_online_nodes > 1 && !nodes_empty(l->nodes))
6921			seq_printf(seq, " nodes=%*pbl",
6922				 nodemask_pr_args(&l->nodes));
6923
6924#ifdef CONFIG_STACKDEPOT
6925		{
6926			depot_stack_handle_t handle;
6927			unsigned long *entries;
6928			unsigned int nr_entries, j;
6929
6930			handle = READ_ONCE(l->handle);
6931			if (handle) {
6932				nr_entries = stack_depot_fetch(handle, &entries);
6933				seq_puts(seq, "\n");
6934				for (j = 0; j < nr_entries; j++)
6935					seq_printf(seq, "        %pS\n", (void *)entries[j]);
6936			}
6937		}
6938#endif
6939		seq_puts(seq, "\n");
6940	}
6941
6942	if (!idx && !t->count)
6943		seq_puts(seq, "No data\n");
6944
6945	return 0;
6946}
6947
6948static void slab_debugfs_stop(struct seq_file *seq, void *v)
6949{
6950}
6951
6952static void *slab_debugfs_next(struct seq_file *seq, void *v, loff_t *ppos)
6953{
6954	struct loc_track *t = seq->private;
6955
6956	t->idx = ++(*ppos);
6957	if (*ppos <= t->count)
6958		return ppos;
6959
6960	return NULL;
6961}
6962
6963static int cmp_loc_by_count(const void *a, const void *b, const void *data)
6964{
6965	struct location *loc1 = (struct location *)a;
6966	struct location *loc2 = (struct location *)b;
6967
6968	if (loc1->count > loc2->count)
6969		return -1;
6970	else
6971		return 1;
6972}
6973
6974static void *slab_debugfs_start(struct seq_file *seq, loff_t *ppos)
6975{
6976	struct loc_track *t = seq->private;
6977
6978	t->idx = *ppos;
6979	return ppos;
6980}
6981
6982static const struct seq_operations slab_debugfs_sops = {
6983	.start  = slab_debugfs_start,
6984	.next   = slab_debugfs_next,
6985	.stop   = slab_debugfs_stop,
6986	.show   = slab_debugfs_show,
6987};
6988
6989static int slab_debug_trace_open(struct inode *inode, struct file *filep)
6990{
6991
6992	struct kmem_cache_node *n;
6993	enum track_item alloc;
6994	int node;
6995	struct loc_track *t = __seq_open_private(filep, &slab_debugfs_sops,
6996						sizeof(struct loc_track));
6997	struct kmem_cache *s = file_inode(filep)->i_private;
6998	unsigned long *obj_map;
6999
7000	if (!t)
7001		return -ENOMEM;
7002
7003	obj_map = bitmap_alloc(oo_objects(s->oo), GFP_KERNEL);
7004	if (!obj_map) {
7005		seq_release_private(inode, filep);
7006		return -ENOMEM;
7007	}
7008
7009	if (strcmp(filep->f_path.dentry->d_name.name, "alloc_traces") == 0)
7010		alloc = TRACK_ALLOC;
7011	else
7012		alloc = TRACK_FREE;
7013
7014	if (!alloc_loc_track(t, PAGE_SIZE / sizeof(struct location), GFP_KERNEL)) {
7015		bitmap_free(obj_map);
7016		seq_release_private(inode, filep);
7017		return -ENOMEM;
7018	}
7019
7020	for_each_kmem_cache_node(s, node, n) {
7021		unsigned long flags;
7022		struct slab *slab;
7023
7024		if (!node_nr_slabs(n))
7025			continue;
7026
7027		spin_lock_irqsave(&n->list_lock, flags);
7028		list_for_each_entry(slab, &n->partial, slab_list)
7029			process_slab(t, s, slab, alloc, obj_map);
7030		list_for_each_entry(slab, &n->full, slab_list)
7031			process_slab(t, s, slab, alloc, obj_map);
7032		spin_unlock_irqrestore(&n->list_lock, flags);
7033	}
7034
7035	/* Sort locations by count */
7036	sort_r(t->loc, t->count, sizeof(struct location),
7037		cmp_loc_by_count, NULL, NULL);
7038
7039	bitmap_free(obj_map);
7040	return 0;
7041}
7042
7043static int slab_debug_trace_release(struct inode *inode, struct file *file)
7044{
7045	struct seq_file *seq = file->private_data;
7046	struct loc_track *t = seq->private;
7047
7048	free_loc_track(t);
7049	return seq_release_private(inode, file);
7050}
7051
7052static const struct file_operations slab_debugfs_fops = {
7053	.open    = slab_debug_trace_open,
7054	.read    = seq_read,
7055	.llseek  = seq_lseek,
7056	.release = slab_debug_trace_release,
7057};
7058
7059static void debugfs_slab_add(struct kmem_cache *s)
7060{
7061	struct dentry *slab_cache_dir;
7062
7063	if (unlikely(!slab_debugfs_root))
7064		return;
7065
7066	slab_cache_dir = debugfs_create_dir(s->name, slab_debugfs_root);
7067
7068	debugfs_create_file("alloc_traces", 0400,
7069		slab_cache_dir, s, &slab_debugfs_fops);
7070
7071	debugfs_create_file("free_traces", 0400,
7072		slab_cache_dir, s, &slab_debugfs_fops);
7073}
7074
7075void debugfs_slab_release(struct kmem_cache *s)
7076{
7077	debugfs_lookup_and_remove(s->name, slab_debugfs_root);
7078}
7079
7080static int __init slab_debugfs_init(void)
7081{
7082	struct kmem_cache *s;
7083
7084	slab_debugfs_root = debugfs_create_dir("slab", NULL);
7085
7086	list_for_each_entry(s, &slab_caches, list)
7087		if (s->flags & SLAB_STORE_USER)
7088			debugfs_slab_add(s);
7089
7090	return 0;
7091
7092}
7093__initcall(slab_debugfs_init);
7094#endif
7095/*
7096 * The /proc/slabinfo ABI
7097 */
7098#ifdef CONFIG_SLUB_DEBUG
7099void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
7100{
7101	unsigned long nr_slabs = 0;
7102	unsigned long nr_objs = 0;
7103	unsigned long nr_free = 0;
7104	int node;
7105	struct kmem_cache_node *n;
7106
7107	for_each_kmem_cache_node(s, node, n) {
7108		nr_slabs += node_nr_slabs(n);
7109		nr_objs += node_nr_objs(n);
7110		nr_free += count_partial(n, count_free);
7111	}
7112
7113	sinfo->active_objs = nr_objs - nr_free;
7114	sinfo->num_objs = nr_objs;
7115	sinfo->active_slabs = nr_slabs;
7116	sinfo->num_slabs = nr_slabs;
7117	sinfo->objects_per_slab = oo_objects(s->oo);
7118	sinfo->cache_order = oo_order(s->oo);
7119}
7120
7121void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
7122{
7123}
7124
7125ssize_t slabinfo_write(struct file *file, const char __user *buffer,
7126		       size_t count, loff_t *ppos)
7127{
7128	return -EIO;
7129}
7130#endif /* CONFIG_SLUB_DEBUG */