~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/mm/slub.c

Version: ~ [ linux-5.11-rc3 ] ~ [ linux-5.10.7 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.89 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.167 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.215 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.251 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.251 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.19.8 ] ~ [ linux-3.18.140 ] ~ [ linux-3.17.8 ] ~ [ linux-3.16.85 ] ~ [ linux-3.15.10 ] ~ [ linux-3.14.79 ] ~ [ linux-3.13.11 ] ~ [ linux-3.12.74 ] ~ [ linux-3.11.10 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  1 /*
  2  * SLUB: A slab allocator that limits cache line use instead of queuing
  3  * objects in per cpu and per node lists.
  4  *
  5  * The allocator synchronizes using per slab locks or atomic operatios
  6  * and only uses a centralized lock to manage a pool of partial slabs.
  7  *
  8  * (C) 2007 SGI, Christoph Lameter
  9  * (C) 2011 Linux Foundation, Christoph Lameter
 10  */
 11 
 12 #include <linux/mm.h>
 13 #include <linux/swap.h> /* struct reclaim_state */
 14 #include <linux/module.h>
 15 #include <linux/bit_spinlock.h>
 16 #include <linux/interrupt.h>
 17 #include <linux/bitops.h>
 18 #include <linux/slab.h>
 19 #include "slab.h"
 20 #include <linux/proc_fs.h>
 21 #include <linux/notifier.h>
 22 #include <linux/seq_file.h>
 23 #include <linux/kmemcheck.h>
 24 #include <linux/cpu.h>
 25 #include <linux/cpuset.h>
 26 #include <linux/mempolicy.h>
 27 #include <linux/ctype.h>
 28 #include <linux/debugobjects.h>
 29 #include <linux/kallsyms.h>
 30 #include <linux/memory.h>
 31 #include <linux/math64.h>
 32 #include <linux/fault-inject.h>
 33 #include <linux/stacktrace.h>
 34 #include <linux/prefetch.h>
 35 #include <linux/memcontrol.h>
 36 
 37 #include <trace/events/kmem.h>
 38 
 39 #include "internal.h"
 40 
 41 /*
 42  * Lock order:
 43  *   1. slab_mutex (Global Mutex)
 44  *   2. node->list_lock
 45  *   3. slab_lock(page) (Only on some arches and for debugging)
 46  *
 47  *   slab_mutex
 48  *
 49  *   The role of the slab_mutex is to protect the list of all the slabs
 50  *   and to synchronize major metadata changes to slab cache structures.
 51  *
 52  *   The slab_lock is only used for debugging and on arches that do not
 53  *   have the ability to do a cmpxchg_double. It only protects the second
 54  *   double word in the page struct. Meaning
 55  *      A. page->freelist       -> List of object free in a page
 56  *      B. page->counters       -> Counters of objects
 57  *      C. page->frozen         -> frozen state
 58  *
 59  *   If a slab is frozen then it is exempt from list management. It is not
 60  *   on any list. The processor that froze the slab is the one who can
 61  *   perform list operations on the page. Other processors may put objects
 62  *   onto the freelist but the processor that froze the slab is the only
 63  *   one that can retrieve the objects from the page's freelist.
 64  *
 65  *   The list_lock protects the partial and full list on each node and
 66  *   the partial slab counter. If taken then no new slabs may be added or
 67  *   removed from the lists nor make the number of partial slabs be modified.
 68  *   (Note that the total number of slabs is an atomic value that may be
 69  *   modified without taking the list lock).
 70  *
 71  *   The list_lock is a centralized lock and thus we avoid taking it as
 72  *   much as possible. As long as SLUB does not have to handle partial
 73  *   slabs, operations can continue without any centralized lock. F.e.
 74  *   allocating a long series of objects that fill up slabs does not require
 75  *   the list lock.
 76  *   Interrupts are disabled during allocation and deallocation in order to
 77  *   make the slab allocator safe to use in the context of an irq. In addition
 78  *   interrupts are disabled to ensure that the processor does not change
 79  *   while handling per_cpu slabs, due to kernel preemption.
 80  *
 81  * SLUB assigns one slab for allocation to each processor.
 82  * Allocations only occur from these slabs called cpu slabs.
 83  *
 84  * Slabs with free elements are kept on a partial list and during regular
 85  * operations no list for full slabs is used. If an object in a full slab is
 86  * freed then the slab will show up again on the partial lists.
 87  * We track full slabs for debugging purposes though because otherwise we
 88  * cannot scan all objects.
 89  *
 90  * Slabs are freed when they become empty. Teardown and setup is
 91  * minimal so we rely on the page allocators per cpu caches for
 92  * fast frees and allocs.
 93  *
 94  * Overloading of page flags that are otherwise used for LRU management.
 95  *
 96  * PageActive           The slab is frozen and exempt from list processing.
 97  *                      This means that the slab is dedicated to a purpose
 98  *                      such as satisfying allocations for a specific
 99  *                      processor. Objects may be freed in the slab while
100  *                      it is frozen but slab_free will then skip the usual
101  *                      list operations. It is up to the processor holding
102  *                      the slab to integrate the slab into the slab lists
103  *                      when the slab is no longer needed.
104  *
105  *                      One use of this flag is to mark slabs that are
106  *                      used for allocations. Then such a slab becomes a cpu
107  *                      slab. The cpu slab may be equipped with an additional
108  *                      freelist that allows lockless access to
109  *                      free objects in addition to the regular freelist
110  *                      that requires the slab lock.
111  *
112  * PageError            Slab requires special handling due to debug
113  *                      options set. This moves slab handling out of
114  *                      the fast path and disables lockless freelists.
115  */
116 
117 static inline int kmem_cache_debug(struct kmem_cache *s)
118 {
119 #ifdef CONFIG_SLUB_DEBUG
120         return unlikely(s->flags & SLAB_DEBUG_FLAGS);
121 #else
122         return 0;
123 #endif
124 }
125 
126 static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
127 {
128 #ifdef CONFIG_SLUB_CPU_PARTIAL
129         return !kmem_cache_debug(s);
130 #else
131         return false;
132 #endif
133 }
134 
135 /*
136  * Issues still to be resolved:
137  *
138  * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
139  *
140  * - Variable sizing of the per node arrays
141  */
142 
143 /* Enable to test recovery from slab corruption on boot */
144 #undef SLUB_RESILIENCY_TEST
145 
146 /* Enable to log cmpxchg failures */
147 #undef SLUB_DEBUG_CMPXCHG
148 
149 /*
150  * Mininum number of partial slabs. These will be left on the partial
151  * lists even if they are empty. kmem_cache_shrink may reclaim them.
152  */
153 #define MIN_PARTIAL 5
154 
155 /*
156  * Maximum number of desirable partial slabs.
157  * The existence of more partial slabs makes kmem_cache_shrink
158  * sort the partial list by the number of objects in use.
159  */
160 #define MAX_PARTIAL 10
161 
162 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
163                                 SLAB_POISON | SLAB_STORE_USER)
164 
165 /*
166  * Debugging flags that require metadata to be stored in the slab.  These get
167  * disabled when slub_debug=O is used and a cache's min order increases with
168  * metadata.
169  */
170 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
171 
172 /*
173  * Set of flags that will prevent slab merging
174  */
175 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
176                 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
177                 SLAB_FAILSLAB)
178 
179 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
180                 SLAB_CACHE_DMA | SLAB_NOTRACK)
181 
182 #define OO_SHIFT        16
183 #define OO_MASK         ((1 << OO_SHIFT) - 1)
184 #define MAX_OBJS_PER_PAGE       32767 /* since page.objects is u15 */
185 
186 /* Internal SLUB flags */
187 #define __OBJECT_POISON         0x80000000UL /* Poison object */
188 #define __CMPXCHG_DOUBLE        0x40000000UL /* Use cmpxchg_double */
189 
190 #ifdef CONFIG_SMP
191 static struct notifier_block slab_notifier;
192 #endif
193 
194 /*
195  * Tracking user of a slab.
196  */
197 #define TRACK_ADDRS_COUNT 16
198 struct track {
199         unsigned long addr;     /* Called from address */
200 #ifdef CONFIG_STACKTRACE
201         unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
202 #endif
203         int cpu;                /* Was running on cpu */
204         int pid;                /* Pid context */
205         unsigned long when;     /* When did the operation occur */
206 };
207 
208 enum track_item { TRACK_ALLOC, TRACK_FREE };
209 
210 #ifdef CONFIG_SYSFS
211 static int sysfs_slab_add(struct kmem_cache *);
212 static int sysfs_slab_alias(struct kmem_cache *, const char *);
213 static void memcg_propagate_slab_attrs(struct kmem_cache *s);
214 #else
215 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
216 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
217                                                         { return 0; }
218 static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
219 #endif
220 
221 static inline void stat(const struct kmem_cache *s, enum stat_item si)
222 {
223 #ifdef CONFIG_SLUB_STATS
224         /*
225          * The rmw is racy on a preemptible kernel but this is acceptable, so
226          * avoid this_cpu_add()'s irq-disable overhead.
227          */
228         raw_cpu_inc(s->cpu_slab->stat[si]);
229 #endif
230 }
231 
232 /********************************************************************
233  *                      Core slab cache functions
234  *******************************************************************/
235 
236 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
237 {
238         return s->node[node];
239 }
240 
241 /* Verify that a pointer has an address that is valid within a slab page */
242 static inline int check_valid_pointer(struct kmem_cache *s,
243                                 struct page *page, const void *object)
244 {
245         void *base;
246 
247         if (!object)
248                 return 1;
249 
250         base = page_address(page);
251         if (object < base || object >= base + page->objects * s->size ||
252                 (object - base) % s->size) {
253                 return 0;
254         }
255 
256         return 1;
257 }
258 
259 static inline void *get_freepointer(struct kmem_cache *s, void *object)
260 {
261         return *(void **)(object + s->offset);
262 }
263 
264 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
265 {
266         prefetch(object + s->offset);
267 }
268 
269 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
270 {
271         void *p;
272 
273 #ifdef CONFIG_DEBUG_PAGEALLOC
274         probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
275 #else
276         p = get_freepointer(s, object);
277 #endif
278         return p;
279 }
280 
281 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
282 {
283         *(void **)(object + s->offset) = fp;
284 }
285 
286 /* Loop over all objects in a slab */
287 #define for_each_object(__p, __s, __addr, __objects) \
288         for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
289                         __p += (__s)->size)
290 
291 /* Determine object index from a given position */
292 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
293 {
294         return (p - addr) / s->size;
295 }
296 
297 static inline size_t slab_ksize(const struct kmem_cache *s)
298 {
299 #ifdef CONFIG_SLUB_DEBUG
300         /*
301          * Debugging requires use of the padding between object
302          * and whatever may come after it.
303          */
304         if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
305                 return s->object_size;
306 
307 #endif
308         /*
309          * If we have the need to store the freelist pointer
310          * back there or track user information then we can
311          * only use the space before that information.
312          */
313         if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
314                 return s->inuse;
315         /*
316          * Else we can use all the padding etc for the allocation
317          */
318         return s->size;
319 }
320 
321 static inline int order_objects(int order, unsigned long size, int reserved)
322 {
323         return ((PAGE_SIZE << order) - reserved) / size;
324 }
325 
326 static inline struct kmem_cache_order_objects oo_make(int order,
327                 unsigned long size, int reserved)
328 {
329         struct kmem_cache_order_objects x = {
330                 (order << OO_SHIFT) + order_objects(order, size, reserved)
331         };
332 
333         return x;
334 }
335 
336 static inline int oo_order(struct kmem_cache_order_objects x)
337 {
338         return x.x >> OO_SHIFT;
339 }
340 
341 static inline int oo_objects(struct kmem_cache_order_objects x)
342 {
343         return x.x & OO_MASK;
344 }
345 
346 /*
347  * Per slab locking using the pagelock
348  */
349 static __always_inline void slab_lock(struct page *page)
350 {
351         bit_spin_lock(PG_locked, &page->flags);
352 }
353 
354 static __always_inline void slab_unlock(struct page *page)
355 {
356         __bit_spin_unlock(PG_locked, &page->flags);
357 }
358 
359 static inline void set_page_slub_counters(struct page *page, unsigned long counters_new)
360 {
361         struct page tmp;
362         tmp.counters = counters_new;
363         /*
364          * page->counters can cover frozen/inuse/objects as well
365          * as page->_count.  If we assign to ->counters directly
366          * we run the risk of losing updates to page->_count, so
367          * be careful and only assign to the fields we need.
368          */
369         page->frozen  = tmp.frozen;
370         page->inuse   = tmp.inuse;
371         page->objects = tmp.objects;
372 }
373 
374 /* Interrupts must be disabled (for the fallback code to work right) */
375 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
376                 void *freelist_old, unsigned long counters_old,
377                 void *freelist_new, unsigned long counters_new,
378                 const char *n)
379 {
380         VM_BUG_ON(!irqs_disabled());
381 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
382     defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
383         if (s->flags & __CMPXCHG_DOUBLE) {
384                 if (cmpxchg_double(&page->freelist, &page->counters,
385                         freelist_old, counters_old,
386                         freelist_new, counters_new))
387                 return 1;
388         } else
389 #endif
390         {
391                 slab_lock(page);
392                 if (page->freelist == freelist_old &&
393                                         page->counters == counters_old) {
394                         page->freelist = freelist_new;
395                         set_page_slub_counters(page, counters_new);
396                         slab_unlock(page);
397                         return 1;
398                 }
399                 slab_unlock(page);
400         }
401 
402         cpu_relax();
403         stat(s, CMPXCHG_DOUBLE_FAIL);
404 
405 #ifdef SLUB_DEBUG_CMPXCHG
406         pr_info("%s %s: cmpxchg double redo ", n, s->name);
407 #endif
408 
409         return 0;
410 }
411 
412 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
413                 void *freelist_old, unsigned long counters_old,
414                 void *freelist_new, unsigned long counters_new,
415                 const char *n)
416 {
417 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
418     defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
419         if (s->flags & __CMPXCHG_DOUBLE) {
420                 if (cmpxchg_double(&page->freelist, &page->counters,
421                         freelist_old, counters_old,
422                         freelist_new, counters_new))
423                 return 1;
424         } else
425 #endif
426         {
427                 unsigned long flags;
428 
429                 local_irq_save(flags);
430                 slab_lock(page);
431                 if (page->freelist == freelist_old &&
432                                         page->counters == counters_old) {
433                         page->freelist = freelist_new;
434                         set_page_slub_counters(page, counters_new);
435                         slab_unlock(page);
436                         local_irq_restore(flags);
437                         return 1;
438                 }
439                 slab_unlock(page);
440                 local_irq_restore(flags);
441         }
442 
443         cpu_relax();
444         stat(s, CMPXCHG_DOUBLE_FAIL);
445 
446 #ifdef SLUB_DEBUG_CMPXCHG
447         pr_info("%s %s: cmpxchg double redo ", n, s->name);
448 #endif
449 
450         return 0;
451 }
452 
453 #ifdef CONFIG_SLUB_DEBUG
454 /*
455  * Determine a map of object in use on a page.
456  *
457  * Node listlock must be held to guarantee that the page does
458  * not vanish from under us.
459  */
460 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
461 {
462         void *p;
463         void *addr = page_address(page);
464 
465         for (p = page->freelist; p; p = get_freepointer(s, p))
466                 set_bit(slab_index(p, s, addr), map);
467 }
468 
469 /*
470  * Debug settings:
471  */
472 #ifdef CONFIG_SLUB_DEBUG_ON
473 static int slub_debug = DEBUG_DEFAULT_FLAGS;
474 #else
475 static int slub_debug;
476 #endif
477 
478 static char *slub_debug_slabs;
479 static int disable_higher_order_debug;
480 
481 /*
482  * Object debugging
483  */
484 static void print_section(char *text, u8 *addr, unsigned int length)
485 {
486         print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
487                         length, 1);
488 }
489 
490 static struct track *get_track(struct kmem_cache *s, void *object,
491         enum track_item alloc)
492 {
493         struct track *p;
494 
495         if (s->offset)
496                 p = object + s->offset + sizeof(void *);
497         else
498                 p = object + s->inuse;
499 
500         return p + alloc;
501 }
502 
503 static void set_track(struct kmem_cache *s, void *object,
504                         enum track_item alloc, unsigned long addr)
505 {
506         struct track *p = get_track(s, object, alloc);
507 
508         if (addr) {
509 #ifdef CONFIG_STACKTRACE
510                 struct stack_trace trace;
511                 int i;
512 
513                 trace.nr_entries = 0;
514                 trace.max_entries = TRACK_ADDRS_COUNT;
515                 trace.entries = p->addrs;
516                 trace.skip = 3;
517                 save_stack_trace(&trace);
518 
519                 /* See rant in lockdep.c */
520                 if (trace.nr_entries != 0 &&
521                     trace.entries[trace.nr_entries - 1] == ULONG_MAX)
522                         trace.nr_entries--;
523 
524                 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
525                         p->addrs[i] = 0;
526 #endif
527                 p->addr = addr;
528                 p->cpu = smp_processor_id();
529                 p->pid = current->pid;
530                 p->when = jiffies;
531         } else
532                 memset(p, 0, sizeof(struct track));
533 }
534 
535 static void init_tracking(struct kmem_cache *s, void *object)
536 {
537         if (!(s->flags & SLAB_STORE_USER))
538                 return;
539 
540         set_track(s, object, TRACK_FREE, 0UL);
541         set_track(s, object, TRACK_ALLOC, 0UL);
542 }
543 
544 static void print_track(const char *s, struct track *t)
545 {
546         if (!t->addr)
547                 return;
548 
549         pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
550                s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
551 #ifdef CONFIG_STACKTRACE
552         {
553                 int i;
554                 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
555                         if (t->addrs[i])
556                                 pr_err("\t%pS\n", (void *)t->addrs[i]);
557                         else
558                                 break;
559         }
560 #endif
561 }
562 
563 static void print_tracking(struct kmem_cache *s, void *object)
564 {
565         if (!(s->flags & SLAB_STORE_USER))
566                 return;
567 
568         print_track("Allocated", get_track(s, object, TRACK_ALLOC));
569         print_track("Freed", get_track(s, object, TRACK_FREE));
570 }
571 
572 static void print_page_info(struct page *page)
573 {
574         pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
575                page, page->objects, page->inuse, page->freelist, page->flags);
576 
577 }
578 
579 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
580 {
581         struct va_format vaf;
582         va_list args;
583 
584         va_start(args, fmt);
585         vaf.fmt = fmt;
586         vaf.va = &args;
587         pr_err("=============================================================================\n");
588         pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
589         pr_err("-----------------------------------------------------------------------------\n\n");
590 
591         add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
592         va_end(args);
593 }
594 
595 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
596 {
597         struct va_format vaf;
598         va_list args;
599 
600         va_start(args, fmt);
601         vaf.fmt = fmt;
602         vaf.va = &args;
603         pr_err("FIX %s: %pV\n", s->name, &vaf);
604         va_end(args);
605 }
606 
607 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
608 {
609         unsigned int off;       /* Offset of last byte */
610         u8 *addr = page_address(page);
611 
612         print_tracking(s, p);
613 
614         print_page_info(page);
615 
616         pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
617                p, p - addr, get_freepointer(s, p));
618 
619         if (p > addr + 16)
620                 print_section("Bytes b4 ", p - 16, 16);
621 
622         print_section("Object ", p, min_t(unsigned long, s->object_size,
623                                 PAGE_SIZE));
624         if (s->flags & SLAB_RED_ZONE)
625                 print_section("Redzone ", p + s->object_size,
626                         s->inuse - s->object_size);
627 
628         if (s->offset)
629                 off = s->offset + sizeof(void *);
630         else
631                 off = s->inuse;
632 
633         if (s->flags & SLAB_STORE_USER)
634                 off += 2 * sizeof(struct track);
635 
636         if (off != s->size)
637                 /* Beginning of the filler is the free pointer */
638                 print_section("Padding ", p + off, s->size - off);
639 
640         dump_stack();
641 }
642 
643 static void object_err(struct kmem_cache *s, struct page *page,
644                         u8 *object, char *reason)
645 {
646         slab_bug(s, "%s", reason);
647         print_trailer(s, page, object);
648 }
649 
650 static void slab_err(struct kmem_cache *s, struct page *page,
651                         const char *fmt, ...)
652 {
653         va_list args;
654         char buf[100];
655 
656         va_start(args, fmt);
657         vsnprintf(buf, sizeof(buf), fmt, args);
658         va_end(args);
659         slab_bug(s, "%s", buf);
660         print_page_info(page);
661         dump_stack();
662 }
663 
664 static void init_object(struct kmem_cache *s, void *object, u8 val)
665 {
666         u8 *p = object;
667 
668         if (s->flags & __OBJECT_POISON) {
669                 memset(p, POISON_FREE, s->object_size - 1);
670                 p[s->object_size - 1] = POISON_END;
671         }
672 
673         if (s->flags & SLAB_RED_ZONE)
674                 memset(p + s->object_size, val, s->inuse - s->object_size);
675 }
676 
677 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
678                                                 void *from, void *to)
679 {
680         slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
681         memset(from, data, to - from);
682 }
683 
684 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
685                         u8 *object, char *what,
686                         u8 *start, unsigned int value, unsigned int bytes)
687 {
688         u8 *fault;
689         u8 *end;
690 
691         fault = memchr_inv(start, value, bytes);
692         if (!fault)
693                 return 1;
694 
695         end = start + bytes;
696         while (end > fault && end[-1] == value)
697                 end--;
698 
699         slab_bug(s, "%s overwritten", what);
700         pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
701                                         fault, end - 1, fault[0], value);
702         print_trailer(s, page, object);
703 
704         restore_bytes(s, what, value, fault, end);
705         return 0;
706 }
707 
708 /*
709  * Object layout:
710  *
711  * object address
712  *      Bytes of the object to be managed.
713  *      If the freepointer may overlay the object then the free
714  *      pointer is the first word of the object.
715  *
716  *      Poisoning uses 0x6b (POISON_FREE) and the last byte is
717  *      0xa5 (POISON_END)
718  *
719  * object + s->object_size
720  *      Padding to reach word boundary. This is also used for Redzoning.
721  *      Padding is extended by another word if Redzoning is enabled and
722  *      object_size == inuse.
723  *
724  *      We fill with 0xbb (RED_INACTIVE) for inactive objects and with
725  *      0xcc (RED_ACTIVE) for objects in use.
726  *
727  * object + s->inuse
728  *      Meta data starts here.
729  *
730  *      A. Free pointer (if we cannot overwrite object on free)
731  *      B. Tracking data for SLAB_STORE_USER
732  *      C. Padding to reach required alignment boundary or at mininum
733  *              one word if debugging is on to be able to detect writes
734  *              before the word boundary.
735  *
736  *      Padding is done using 0x5a (POISON_INUSE)
737  *
738  * object + s->size
739  *      Nothing is used beyond s->size.
740  *
741  * If slabcaches are merged then the object_size and inuse boundaries are mostly
742  * ignored. And therefore no slab options that rely on these boundaries
743  * may be used with merged slabcaches.
744  */
745 
746 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
747 {
748         unsigned long off = s->inuse;   /* The end of info */
749 
750         if (s->offset)
751                 /* Freepointer is placed after the object. */
752                 off += sizeof(void *);
753 
754         if (s->flags & SLAB_STORE_USER)
755                 /* We also have user information there */
756                 off += 2 * sizeof(struct track);
757 
758         if (s->size == off)
759                 return 1;
760 
761         return check_bytes_and_report(s, page, p, "Object padding",
762                                 p + off, POISON_INUSE, s->size - off);
763 }
764 
765 /* Check the pad bytes at the end of a slab page */
766 static int slab_pad_check(struct kmem_cache *s, struct page *page)
767 {
768         u8 *start;
769         u8 *fault;
770         u8 *end;
771         int length;
772         int remainder;
773 
774         if (!(s->flags & SLAB_POISON))
775                 return 1;
776 
777         start = page_address(page);
778         length = (PAGE_SIZE << compound_order(page)) - s->reserved;
779         end = start + length;
780         remainder = length % s->size;
781         if (!remainder)
782                 return 1;
783 
784         fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
785         if (!fault)
786                 return 1;
787         while (end > fault && end[-1] == POISON_INUSE)
788                 end--;
789 
790         slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
791         print_section("Padding ", end - remainder, remainder);
792 
793         restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
794         return 0;
795 }
796 
797 static int check_object(struct kmem_cache *s, struct page *page,
798                                         void *object, u8 val)
799 {
800         u8 *p = object;
801         u8 *endobject = object + s->object_size;
802 
803         if (s->flags & SLAB_RED_ZONE) {
804                 if (!check_bytes_and_report(s, page, object, "Redzone",
805                         endobject, val, s->inuse - s->object_size))
806                         return 0;
807         } else {
808                 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
809                         check_bytes_and_report(s, page, p, "Alignment padding",
810                                 endobject, POISON_INUSE,
811                                 s->inuse - s->object_size);
812                 }
813         }
814 
815         if (s->flags & SLAB_POISON) {
816                 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
817                         (!check_bytes_and_report(s, page, p, "Poison", p,
818                                         POISON_FREE, s->object_size - 1) ||
819                          !check_bytes_and_report(s, page, p, "Poison",
820                                 p + s->object_size - 1, POISON_END, 1)))
821                         return 0;
822                 /*
823                  * check_pad_bytes cleans up on its own.
824                  */
825                 check_pad_bytes(s, page, p);
826         }
827 
828         if (!s->offset && val == SLUB_RED_ACTIVE)
829                 /*
830                  * Object and freepointer overlap. Cannot check
831                  * freepointer while object is allocated.
832                  */
833                 return 1;
834 
835         /* Check free pointer validity */
836         if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
837                 object_err(s, page, p, "Freepointer corrupt");
838                 /*
839                  * No choice but to zap it and thus lose the remainder
840                  * of the free objects in this slab. May cause
841                  * another error because the object count is now wrong.
842                  */
843                 set_freepointer(s, p, NULL);
844                 return 0;
845         }
846         return 1;
847 }
848 
849 static int check_slab(struct kmem_cache *s, struct page *page)
850 {
851         int maxobj;
852 
853         VM_BUG_ON(!irqs_disabled());
854 
855         if (!PageSlab(page)) {
856                 slab_err(s, page, "Not a valid slab page");
857                 return 0;
858         }
859 
860         maxobj = order_objects(compound_order(page), s->size, s->reserved);
861         if (page->objects > maxobj) {
862                 slab_err(s, page, "objects %u > max %u",
863                         s->name, page->objects, maxobj);
864                 return 0;
865         }
866         if (page->inuse > page->objects) {
867                 slab_err(s, page, "inuse %u > max %u",
868                         s->name, page->inuse, page->objects);
869                 return 0;
870         }
871         /* Slab_pad_check fixes things up after itself */
872         slab_pad_check(s, page);
873         return 1;
874 }
875 
876 /*
877  * Determine if a certain object on a page is on the freelist. Must hold the
878  * slab lock to guarantee that the chains are in a consistent state.
879  */
880 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
881 {
882         int nr = 0;
883         void *fp;
884         void *object = NULL;
885         unsigned long max_objects;
886 
887         fp = page->freelist;
888         while (fp && nr <= page->objects) {
889                 if (fp == search)
890                         return 1;
891                 if (!check_valid_pointer(s, page, fp)) {
892                         if (object) {
893                                 object_err(s, page, object,
894                                         "Freechain corrupt");
895                                 set_freepointer(s, object, NULL);
896                         } else {
897                                 slab_err(s, page, "Freepointer corrupt");
898                                 page->freelist = NULL;
899                                 page->inuse = page->objects;
900                                 slab_fix(s, "Freelist cleared");
901                                 return 0;
902                         }
903                         break;
904                 }
905                 object = fp;
906                 fp = get_freepointer(s, object);
907                 nr++;
908         }
909 
910         max_objects = order_objects(compound_order(page), s->size, s->reserved);
911         if (max_objects > MAX_OBJS_PER_PAGE)
912                 max_objects = MAX_OBJS_PER_PAGE;
913 
914         if (page->objects != max_objects) {
915                 slab_err(s, page, "Wrong number of objects. Found %d but "
916                         "should be %d", page->objects, max_objects);
917                 page->objects = max_objects;
918                 slab_fix(s, "Number of objects adjusted.");
919         }
920         if (page->inuse != page->objects - nr) {
921                 slab_err(s, page, "Wrong object count. Counter is %d but "
922                         "counted were %d", page->inuse, page->objects - nr);
923                 page->inuse = page->objects - nr;
924                 slab_fix(s, "Object count adjusted.");
925         }
926         return search == NULL;
927 }
928 
929 static void trace(struct kmem_cache *s, struct page *page, void *object,
930                                                                 int alloc)
931 {
932         if (s->flags & SLAB_TRACE) {
933                 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
934                         s->name,
935                         alloc ? "alloc" : "free",
936                         object, page->inuse,
937                         page->freelist);
938 
939                 if (!alloc)
940                         print_section("Object ", (void *)object,
941                                         s->object_size);
942 
943                 dump_stack();
944         }
945 }
946 
947 /*
948  * Hooks for other subsystems that check memory allocations. In a typical
949  * production configuration these hooks all should produce no code at all.
950  */
951 static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
952 {
953         kmemleak_alloc(ptr, size, 1, flags);
954 }
955 
956 static inline void kfree_hook(const void *x)
957 {
958         kmemleak_free(x);
959 }
960 
961 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
962 {
963         flags &= gfp_allowed_mask;
964         lockdep_trace_alloc(flags);
965         might_sleep_if(flags & __GFP_WAIT);
966 
967         return should_failslab(s->object_size, flags, s->flags);
968 }
969 
970 static inline void slab_post_alloc_hook(struct kmem_cache *s,
971                                         gfp_t flags, void *object)
972 {
973         flags &= gfp_allowed_mask;
974         kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
975         kmemleak_alloc_recursive(object, s->object_size, 1, s->flags, flags);
976 }
977 
978 static inline void slab_free_hook(struct kmem_cache *s, void *x)
979 {
980         kmemleak_free_recursive(x, s->flags);
981 
982         /*
983          * Trouble is that we may no longer disable interrupts in the fast path
984          * So in order to make the debug calls that expect irqs to be
985          * disabled we need to disable interrupts temporarily.
986          */
987 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
988         {
989                 unsigned long flags;
990 
991                 local_irq_save(flags);
992                 kmemcheck_slab_free(s, x, s->object_size);
993                 debug_check_no_locks_freed(x, s->object_size);
994                 local_irq_restore(flags);
995         }
996 #endif
997         if (!(s->flags & SLAB_DEBUG_OBJECTS))
998                 debug_check_no_obj_freed(x, s->object_size);
999 }
1000 
1001 /*
1002  * Tracking of fully allocated slabs for debugging purposes.
1003  */
1004 static void add_full(struct kmem_cache *s,
1005         struct kmem_cache_node *n, struct page *page)
1006 {
1007         if (!(s->flags & SLAB_STORE_USER))
1008                 return;
1009 
1010         lockdep_assert_held(&n->list_lock);
1011         list_add(&page->lru, &n->full);
1012 }
1013 
1014 static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
1015 {
1016         if (!(s->flags & SLAB_STORE_USER))
1017                 return;
1018 
1019         lockdep_assert_held(&n->list_lock);
1020         list_del(&page->lru);
1021 }
1022 
1023 /* Tracking of the number of slabs for debugging purposes */
1024 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1025 {
1026         struct kmem_cache_node *n = get_node(s, node);
1027 
1028         return atomic_long_read(&n->nr_slabs);
1029 }
1030 
1031 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1032 {
1033         return atomic_long_read(&n->nr_slabs);
1034 }
1035 
1036 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1037 {
1038         struct kmem_cache_node *n = get_node(s, node);
1039 
1040         /*
1041          * May be called early in order to allocate a slab for the
1042          * kmem_cache_node structure. Solve the chicken-egg
1043          * dilemma by deferring the increment of the count during
1044          * bootstrap (see early_kmem_cache_node_alloc).
1045          */
1046         if (likely(n)) {
1047                 atomic_long_inc(&n->nr_slabs);
1048                 atomic_long_add(objects, &n->total_objects);
1049         }
1050 }
1051 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1052 {
1053         struct kmem_cache_node *n = get_node(s, node);
1054 
1055         atomic_long_dec(&n->nr_slabs);
1056         atomic_long_sub(objects, &n->total_objects);
1057 }
1058 
1059 /* Object debug checks for alloc/free paths */
1060 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1061                                                                 void *object)
1062 {
1063         if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1064                 return;
1065 
1066         init_object(s, object, SLUB_RED_INACTIVE);
1067         init_tracking(s, object);
1068 }
1069 
1070 static noinline int alloc_debug_processing(struct kmem_cache *s,
1071                                         struct page *page,
1072                                         void *object, unsigned long addr)
1073 {
1074         if (!check_slab(s, page))
1075                 goto bad;
1076 
1077         if (!check_valid_pointer(s, page, object)) {
1078                 object_err(s, page, object, "Freelist Pointer check fails");
1079                 goto bad;
1080         }
1081 
1082         if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1083                 goto bad;
1084 
1085         /* Success perform special debug activities for allocs */
1086         if (s->flags & SLAB_STORE_USER)
1087                 set_track(s, object, TRACK_ALLOC, addr);
1088         trace(s, page, object, 1);
1089         init_object(s, object, SLUB_RED_ACTIVE);
1090         return 1;
1091 
1092 bad:
1093         if (PageSlab(page)) {
1094                 /*
1095                  * If this is a slab page then lets do the best we can
1096                  * to avoid issues in the future. Marking all objects
1097                  * as used avoids touching the remaining objects.
1098                  */
1099                 slab_fix(s, "Marking all objects used");
1100                 page->inuse = page->objects;
1101                 page->freelist = NULL;
1102         }
1103         return 0;
1104 }
1105 
1106 static noinline struct kmem_cache_node *free_debug_processing(
1107         struct kmem_cache *s, struct page *page, void *object,
1108         unsigned long addr, unsigned long *flags)
1109 {
1110         struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1111 
1112         spin_lock_irqsave(&n->list_lock, *flags);
1113         slab_lock(page);
1114 
1115         if (!check_slab(s, page))
1116                 goto fail;
1117 
1118         if (!check_valid_pointer(s, page, object)) {
1119                 slab_err(s, page, "Invalid object pointer 0x%p", object);
1120                 goto fail;
1121         }
1122 
1123         if (on_freelist(s, page, object)) {
1124                 object_err(s, page, object, "Object already free");
1125                 goto fail;
1126         }
1127 
1128         if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1129                 goto out;
1130 
1131         if (unlikely(s != page->slab_cache)) {
1132                 if (!PageSlab(page)) {
1133                         slab_err(s, page, "Attempt to free object(0x%p) "
1134                                 "outside of slab", object);
1135                 } else if (!page->slab_cache) {
1136                         pr_err("SLUB <none>: no slab for object 0x%p.\n",
1137                                object);
1138                         dump_stack();
1139                 } else
1140                         object_err(s, page, object,
1141                                         "page slab pointer corrupt.");
1142                 goto fail;
1143         }
1144 
1145         if (s->flags & SLAB_STORE_USER)
1146                 set_track(s, object, TRACK_FREE, addr);
1147         trace(s, page, object, 0);
1148         init_object(s, object, SLUB_RED_INACTIVE);
1149 out:
1150         slab_unlock(page);
1151         /*
1152          * Keep node_lock to preserve integrity
1153          * until the object is actually freed
1154          */
1155         return n;
1156 
1157 fail:
1158         slab_unlock(page);
1159         spin_unlock_irqrestore(&n->list_lock, *flags);
1160         slab_fix(s, "Object at 0x%p not freed", object);
1161         return NULL;
1162 }
1163 
1164 static int __init setup_slub_debug(char *str)
1165 {
1166         slub_debug = DEBUG_DEFAULT_FLAGS;
1167         if (*str++ != '=' || !*str)
1168                 /*
1169                  * No options specified. Switch on full debugging.
1170                  */
1171                 goto out;
1172 
1173         if (*str == ',')
1174                 /*
1175                  * No options but restriction on slabs. This means full
1176                  * debugging for slabs matching a pattern.
1177                  */
1178                 goto check_slabs;
1179 
1180         if (tolower(*str) == 'o') {
1181                 /*
1182                  * Avoid enabling debugging on caches if its minimum order
1183                  * would increase as a result.
1184                  */
1185                 disable_higher_order_debug = 1;
1186                 goto out;
1187         }
1188 
1189         slub_debug = 0;
1190         if (*str == '-')
1191                 /*
1192                  * Switch off all debugging measures.
1193                  */
1194                 goto out;
1195 
1196         /*
1197          * Determine which debug features should be switched on
1198          */
1199         for (; *str && *str != ','; str++) {
1200                 switch (tolower(*str)) {
1201                 case 'f':
1202                         slub_debug |= SLAB_DEBUG_FREE;
1203                         break;
1204                 case 'z':
1205                         slub_debug |= SLAB_RED_ZONE;
1206                         break;
1207                 case 'p':
1208                         slub_debug |= SLAB_POISON;
1209                         break;
1210                 case 'u':
1211                         slub_debug |= SLAB_STORE_USER;
1212                         break;
1213                 case 't':
1214                         slub_debug |= SLAB_TRACE;
1215                         break;
1216                 case 'a':
1217                         slub_debug |= SLAB_FAILSLAB;
1218                         break;
1219                 default:
1220                         pr_err("slub_debug option '%c' unknown. skipped\n",
1221                                *str);
1222                 }
1223         }
1224 
1225 check_slabs:
1226         if (*str == ',')
1227                 slub_debug_slabs = str + 1;
1228 out:
1229         return 1;
1230 }
1231 
1232 __setup("slub_debug", setup_slub_debug);
1233 
1234 static unsigned long kmem_cache_flags(unsigned long object_size,
1235         unsigned long flags, const char *name,
1236         void (*ctor)(void *))
1237 {
1238         /*
1239          * Enable debugging if selected on the kernel commandline.
1240          */
1241         if (slub_debug && (!slub_debug_slabs || (name &&
1242                 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))))
1243                 flags |= slub_debug;
1244 
1245         return flags;
1246 }
1247 #else
1248 static inline void setup_object_debug(struct kmem_cache *s,
1249                         struct page *page, void *object) {}
1250 
1251 static inline int alloc_debug_processing(struct kmem_cache *s,
1252         struct page *page, void *object, unsigned long addr) { return 0; }
1253 
1254 static inline struct kmem_cache_node *free_debug_processing(
1255         struct kmem_cache *s, struct page *page, void *object,
1256         unsigned long addr, unsigned long *flags) { return NULL; }
1257 
1258 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1259                         { return 1; }
1260 static inline int check_object(struct kmem_cache *s, struct page *page,
1261                         void *object, u8 val) { return 1; }
1262 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1263                                         struct page *page) {}
1264 static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1265                                         struct page *page) {}
1266 static inline unsigned long kmem_cache_flags(unsigned long object_size,
1267         unsigned long flags, const char *name,
1268         void (*ctor)(void *))
1269 {
1270         return flags;
1271 }
1272 #define slub_debug 0
1273 
1274 #define disable_higher_order_debug 0
1275 
1276 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1277                                                         { return 0; }
1278 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1279                                                         { return 0; }
1280 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1281                                                         int objects) {}
1282 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1283                                                         int objects) {}
1284 
1285 static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1286 {
1287         kmemleak_alloc(ptr, size, 1, flags);
1288 }
1289 
1290 static inline void kfree_hook(const void *x)
1291 {
1292         kmemleak_free(x);
1293 }
1294 
1295 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1296                                                         { return 0; }
1297 
1298 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1299                 void *object)
1300 {
1301         kmemleak_alloc_recursive(object, s->object_size, 1, s->flags,
1302                 flags & gfp_allowed_mask);
1303 }
1304 
1305 static inline void slab_free_hook(struct kmem_cache *s, void *x)
1306 {
1307         kmemleak_free_recursive(x, s->flags);
1308 }
1309 
1310 #endif /* CONFIG_SLUB_DEBUG */
1311 
1312 /*
1313  * Slab allocation and freeing
1314  */
1315 static inline struct page *alloc_slab_page(struct kmem_cache *s,
1316                 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1317 {
1318         struct page *page;
1319         int order = oo_order(oo);
1320 
1321         flags |= __GFP_NOTRACK;
1322 
1323         if (memcg_charge_slab(s, flags, order))
1324                 return NULL;
1325 
1326         if (node == NUMA_NO_NODE)
1327                 page = alloc_pages(flags, order);
1328         else
1329                 page = alloc_pages_exact_node(node, flags, order);
1330 
1331         if (!page)
1332                 memcg_uncharge_slab(s, order);
1333 
1334         return page;
1335 }
1336 
1337 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1338 {
1339         struct page *page;
1340         struct kmem_cache_order_objects oo = s->oo;
1341         gfp_t alloc_gfp;
1342 
1343         flags &= gfp_allowed_mask;
1344 
1345         if (flags & __GFP_WAIT)
1346                 local_irq_enable();
1347 
1348         flags |= s->allocflags;
1349 
1350         /*
1351          * Let the initial higher-order allocation fail under memory pressure
1352          * so we fall-back to the minimum order allocation.
1353          */
1354         alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1355 
1356         page = alloc_slab_page(s, alloc_gfp, node, oo);
1357         if (unlikely(!page)) {
1358                 oo = s->min;
1359                 alloc_gfp = flags;
1360                 /*
1361                  * Allocation may have failed due to fragmentation.
1362                  * Try a lower order alloc if possible
1363                  */
1364                 page = alloc_slab_page(s, alloc_gfp, node, oo);
1365 
1366                 if (page)
1367                         stat(s, ORDER_FALLBACK);
1368         }
1369 
1370         if (kmemcheck_enabled && page
1371                 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1372                 int pages = 1 << oo_order(oo);
1373 
1374                 kmemcheck_alloc_shadow(page, oo_order(oo), alloc_gfp, node);
1375 
1376                 /*
1377                  * Objects from caches that have a constructor don't get
1378                  * cleared when they're allocated, so we need to do it here.
1379                  */
1380                 if (s->ctor)
1381                         kmemcheck_mark_uninitialized_pages(page, pages);
1382                 else
1383                         kmemcheck_mark_unallocated_pages(page, pages);
1384         }
1385 
1386         if (flags & __GFP_WAIT)
1387                 local_irq_disable();
1388         if (!page)
1389                 return NULL;
1390 
1391         page->objects = oo_objects(oo);
1392         mod_zone_page_state(page_zone(page),
1393                 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1394                 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1395                 1 << oo_order(oo));
1396 
1397         return page;
1398 }
1399 
1400 static void setup_object(struct kmem_cache *s, struct page *page,
1401                                 void *object)
1402 {
1403         setup_object_debug(s, page, object);
1404         if (unlikely(s->ctor))
1405                 s->ctor(object);
1406 }
1407 
1408 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1409 {
1410         struct page *page;
1411         void *start;
1412         void *last;
1413         void *p;
1414         int order;
1415 
1416         BUG_ON(flags & GFP_SLAB_BUG_MASK);
1417 
1418         page = allocate_slab(s,
1419                 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1420         if (!page)
1421                 goto out;
1422 
1423         order = compound_order(page);
1424         inc_slabs_node(s, page_to_nid(page), page->objects);
1425         page->slab_cache = s;
1426         __SetPageSlab(page);
1427         if (page->pfmemalloc)
1428                 SetPageSlabPfmemalloc(page);
1429 
1430         start = page_address(page);
1431 
1432         if (unlikely(s->flags & SLAB_POISON))
1433                 memset(start, POISON_INUSE, PAGE_SIZE << order);
1434 
1435         last = start;
1436         for_each_object(p, s, start, page->objects) {
1437                 setup_object(s, page, last);
1438                 set_freepointer(s, last, p);
1439                 last = p;
1440         }
1441         setup_object(s, page, last);
1442         set_freepointer(s, last, NULL);
1443 
1444         page->freelist = start;
1445         page->inuse = page->objects;
1446         page->frozen = 1;
1447 out:
1448         return page;
1449 }
1450 
1451 static void __free_slab(struct kmem_cache *s, struct page *page)
1452 {
1453         int order = compound_order(page);
1454         int pages = 1 << order;
1455 
1456         if (kmem_cache_debug(s)) {
1457                 void *p;
1458 
1459                 slab_pad_check(s, page);
1460                 for_each_object(p, s, page_address(page),
1461                                                 page->objects)
1462                         check_object(s, page, p, SLUB_RED_INACTIVE);
1463         }
1464 
1465         kmemcheck_free_shadow(page, compound_order(page));
1466 
1467         mod_zone_page_state(page_zone(page),
1468                 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1469                 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1470                 -pages);
1471 
1472         __ClearPageSlabPfmemalloc(page);
1473         __ClearPageSlab(page);
1474 
1475         page_mapcount_reset(page);
1476         if (current->reclaim_state)
1477                 current->reclaim_state->reclaimed_slab += pages;
1478         __free_pages(page, order);
1479         memcg_uncharge_slab(s, order);
1480 }
1481 
1482 #define need_reserve_slab_rcu                                           \
1483         (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1484 
1485 static void rcu_free_slab(struct rcu_head *h)
1486 {
1487         struct page *page;
1488 
1489         if (need_reserve_slab_rcu)
1490                 page = virt_to_head_page(h);
1491         else
1492                 page = container_of((struct list_head *)h, struct page, lru);
1493 
1494         __free_slab(page->slab_cache, page);
1495 }
1496 
1497 static void free_slab(struct kmem_cache *s, struct page *page)
1498 {
1499         if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1500                 struct rcu_head *head;
1501 
1502                 if (need_reserve_slab_rcu) {
1503                         int order = compound_order(page);
1504                         int offset = (PAGE_SIZE << order) - s->reserved;
1505 
1506                         VM_BUG_ON(s->reserved != sizeof(*head));
1507                         head = page_address(page) + offset;
1508                 } else {
1509                         /*
1510                          * RCU free overloads the RCU head over the LRU
1511                          */
1512                         head = (void *)&page->lru;
1513                 }
1514 
1515                 call_rcu(head, rcu_free_slab);
1516         } else
1517                 __free_slab(s, page);
1518 }
1519 
1520 static void discard_slab(struct kmem_cache *s, struct page *page)
1521 {
1522         dec_slabs_node(s, page_to_nid(page), page->objects);
1523         free_slab(s, page);
1524 }
1525 
1526 /*
1527  * Management of partially allocated slabs.
1528  */
1529 static inline void
1530 __add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1531 {
1532         n->nr_partial++;
1533         if (tail == DEACTIVATE_TO_TAIL)
1534                 list_add_tail(&page->lru, &n->partial);
1535         else
1536                 list_add(&page->lru, &n->partial);
1537 }
1538 
1539 static inline void add_partial(struct kmem_cache_node *n,
1540                                 struct page *page, int tail)
1541 {
1542         lockdep_assert_held(&n->list_lock);
1543         __add_partial(n, page, tail);
1544 }
1545 
1546 static inline void
1547 __remove_partial(struct kmem_cache_node *n, struct page *page)
1548 {
1549         list_del(&page->lru);
1550         n->nr_partial--;
1551 }
1552 
1553 static inline void remove_partial(struct kmem_cache_node *n,
1554                                         struct page *page)
1555 {
1556         lockdep_assert_held(&n->list_lock);
1557         __remove_partial(n, page);
1558 }
1559 
1560 /*
1561  * Remove slab from the partial list, freeze it and
1562  * return the pointer to the freelist.
1563  *
1564  * Returns a list of objects or NULL if it fails.
1565  */
1566 static inline void *acquire_slab(struct kmem_cache *s,
1567                 struct kmem_cache_node *n, struct page *page,
1568                 int mode, int *objects)
1569 {
1570         void *freelist;
1571         unsigned long counters;
1572         struct page new;
1573 
1574         lockdep_assert_held(&n->list_lock);
1575 
1576         /*
1577          * Zap the freelist and set the frozen bit.
1578          * The old freelist is the list of objects for the
1579          * per cpu allocation list.
1580          */
1581         freelist = page->freelist;
1582         counters = page->counters;
1583         new.counters = counters;
1584         *objects = new.objects - new.inuse;
1585         if (mode) {
1586                 new.inuse = page->objects;
1587                 new.freelist = NULL;
1588         } else {
1589                 new.freelist = freelist;
1590         }
1591 
1592         VM_BUG_ON(new.frozen);
1593         new.frozen = 1;
1594 
1595         if (!__cmpxchg_double_slab(s, page,
1596                         freelist, counters,
1597                         new.freelist, new.counters,
1598                         "acquire_slab"))
1599                 return NULL;
1600 
1601         remove_partial(n, page);
1602         WARN_ON(!freelist);
1603         return freelist;
1604 }
1605 
1606 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1607 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1608 
1609 /*
1610  * Try to allocate a partial slab from a specific node.
1611  */
1612 static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1613                                 struct kmem_cache_cpu *c, gfp_t flags)
1614 {
1615         struct page *page, *page2;
1616         void *object = NULL;
1617         int available = 0;
1618         int objects;
1619 
1620         /*
1621          * Racy check. If we mistakenly see no partial slabs then we
1622          * just allocate an empty slab. If we mistakenly try to get a
1623          * partial slab and there is none available then get_partials()
1624          * will return NULL.
1625          */
1626         if (!n || !n->nr_partial)
1627                 return NULL;
1628 
1629         spin_lock(&n->list_lock);
1630         list_for_each_entry_safe(page, page2, &n->partial, lru) {
1631                 void *t;
1632 
1633                 if (!pfmemalloc_match(page, flags))
1634                         continue;
1635 
1636                 t = acquire_slab(s, n, page, object == NULL, &objects);
1637                 if (!t)
1638                         break;
1639 
1640                 available += objects;
1641                 if (!object) {
1642                         c->page = page;
1643                         stat(s, ALLOC_FROM_PARTIAL);
1644                         object = t;
1645                 } else {
1646                         put_cpu_partial(s, page, 0);
1647                         stat(s, CPU_PARTIAL_NODE);
1648                 }
1649                 if (!kmem_cache_has_cpu_partial(s)
1650                         || available > s->cpu_partial / 2)
1651                         break;
1652 
1653         }
1654         spin_unlock(&n->list_lock);
1655         return object;
1656 }
1657 
1658 /*
1659  * Get a page from somewhere. Search in increasing NUMA distances.
1660  */
1661 static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1662                 struct kmem_cache_cpu *c)
1663 {
1664 #ifdef CONFIG_NUMA
1665         struct zonelist *zonelist;
1666         struct zoneref *z;
1667         struct zone *zone;
1668         enum zone_type high_zoneidx = gfp_zone(flags);
1669         void *object;
1670         unsigned int cpuset_mems_cookie;
1671 
1672         /*
1673          * The defrag ratio allows a configuration of the tradeoffs between
1674          * inter node defragmentation and node local allocations. A lower
1675          * defrag_ratio increases the tendency to do local allocations
1676          * instead of attempting to obtain partial slabs from other nodes.
1677          *
1678          * If the defrag_ratio is set to 0 then kmalloc() always
1679          * returns node local objects. If the ratio is higher then kmalloc()
1680          * may return off node objects because partial slabs are obtained
1681          * from other nodes and filled up.
1682          *
1683          * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1684          * defrag_ratio = 1000) then every (well almost) allocation will
1685          * first attempt to defrag slab caches on other nodes. This means
1686          * scanning over all nodes to look for partial slabs which may be
1687          * expensive if we do it every time we are trying to find a slab
1688          * with available objects.
1689          */
1690         if (!s->remote_node_defrag_ratio ||
1691                         get_cycles() % 1024 > s->remote_node_defrag_ratio)
1692                 return NULL;
1693 
1694         do {
1695                 cpuset_mems_cookie = read_mems_allowed_begin();
1696                 zonelist = node_zonelist(mempolicy_slab_node(), flags);
1697                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1698                         struct kmem_cache_node *n;
1699 
1700                         n = get_node(s, zone_to_nid(zone));
1701 
1702                         if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1703                                         n->nr_partial > s->min_partial) {
1704                                 object = get_partial_node(s, n, c, flags);
1705                                 if (object) {
1706                                         /*
1707                                          * Don't check read_mems_allowed_retry()
1708                                          * here - if mems_allowed was updated in
1709                                          * parallel, that was a harmless race
1710                                          * between allocation and the cpuset
1711                                          * update
1712                                          */
1713                                         return object;
1714                                 }
1715                         }
1716                 }
1717         } while (read_mems_allowed_retry(cpuset_mems_cookie));
1718 #endif
1719         return NULL;
1720 }
1721 
1722 /*
1723  * Get a partial page, lock it and return it.
1724  */
1725 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1726                 struct kmem_cache_cpu *c)
1727 {
1728         void *object;
1729         int searchnode = (node == NUMA_NO_NODE) ? numa_mem_id() : node;
1730 
1731         object = get_partial_node(s, get_node(s, searchnode), c, flags);
1732         if (object || node != NUMA_NO_NODE)
1733                 return object;
1734 
1735         return get_any_partial(s, flags, c);
1736 }
1737 
1738 #ifdef CONFIG_PREEMPT
1739 /*
1740  * Calculate the next globally unique transaction for disambiguiation
1741  * during cmpxchg. The transactions start with the cpu number and are then
1742  * incremented by CONFIG_NR_CPUS.
1743  */
1744 #define TID_STEP  roundup_pow_of_two(CONFIG_NR_CPUS)
1745 #else
1746 /*
1747  * No preemption supported therefore also no need to check for
1748  * different cpus.
1749  */
1750 #define TID_STEP 1
1751 #endif
1752 
1753 static inline unsigned long next_tid(unsigned long tid)
1754 {
1755         return tid + TID_STEP;
1756 }
1757 
1758 static inline unsigned int tid_to_cpu(unsigned long tid)
1759 {
1760         return tid % TID_STEP;
1761 }
1762 
1763 static inline unsigned long tid_to_event(unsigned long tid)
1764 {
1765         return tid / TID_STEP;
1766 }
1767 
1768 static inline unsigned int init_tid(int cpu)
1769 {
1770         return cpu;
1771 }
1772 
1773 static inline void note_cmpxchg_failure(const char *n,
1774                 const struct kmem_cache *s, unsigned long tid)
1775 {
1776 #ifdef SLUB_DEBUG_CMPXCHG
1777         unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1778 
1779         pr_info("%s %s: cmpxchg redo ", n, s->name);
1780 
1781 #ifdef CONFIG_PREEMPT
1782         if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1783                 pr_warn("due to cpu change %d -> %d\n",
1784                         tid_to_cpu(tid), tid_to_cpu(actual_tid));
1785         else
1786 #endif
1787         if (tid_to_event(tid) != tid_to_event(actual_tid))
1788                 pr_warn("due to cpu running other code. Event %ld->%ld\n",
1789                         tid_to_event(tid), tid_to_event(actual_tid));
1790         else
1791                 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
1792                         actual_tid, tid, next_tid(tid));
1793 #endif
1794         stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1795 }
1796 
1797 static void init_kmem_cache_cpus(struct kmem_cache *s)
1798 {
1799         int cpu;
1800 
1801         for_each_possible_cpu(cpu)
1802                 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1803 }
1804 
1805 /*
1806  * Remove the cpu slab
1807  */
1808 static void deactivate_slab(struct kmem_cache *s, struct page *page,
1809                                 void *freelist)
1810 {
1811         enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1812         struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1813         int lock = 0;
1814         enum slab_modes l = M_NONE, m = M_NONE;
1815         void *nextfree;
1816         int tail = DEACTIVATE_TO_HEAD;
1817         struct page new;
1818         struct page old;
1819 
1820         if (page->freelist) {
1821                 stat(s, DEACTIVATE_REMOTE_FREES);
1822                 tail = DEACTIVATE_TO_TAIL;
1823         }
1824 
1825         /*
1826          * Stage one: Free all available per cpu objects back
1827          * to the page freelist while it is still frozen. Leave the
1828          * last one.
1829          *
1830          * There is no need to take the list->lock because the page
1831          * is still frozen.
1832          */
1833         while (freelist && (nextfree = get_freepointer(s, freelist))) {
1834                 void *prior;
1835                 unsigned long counters;
1836 
1837                 do {
1838                         prior = page->freelist;
1839                         counters = page->counters;
1840                         set_freepointer(s, freelist, prior);
1841                         new.counters = counters;
1842                         new.inuse--;
1843                         VM_BUG_ON(!new.frozen);
1844 
1845                 } while (!__cmpxchg_double_slab(s, page,
1846                         prior, counters,
1847                         freelist, new.counters,
1848                         "drain percpu freelist"));
1849 
1850                 freelist = nextfree;
1851         }
1852 
1853         /*
1854          * Stage two: Ensure that the page is unfrozen while the
1855          * list presence reflects the actual number of objects
1856          * during unfreeze.
1857          *
1858          * We setup the list membership and then perform a cmpxchg
1859          * with the count. If there is a mismatch then the page
1860          * is not unfrozen but the page is on the wrong list.
1861          *
1862          * Then we restart the process which may have to remove
1863          * the page from the list that we just put it on again
1864          * because the number of objects in the slab may have
1865          * changed.
1866          */
1867 redo:
1868 
1869         old.freelist = page->freelist;
1870         old.counters = page->counters;
1871         VM_BUG_ON(!old.frozen);
1872 
1873         /* Determine target state of the slab */
1874         new.counters = old.counters;
1875         if (freelist) {
1876                 new.inuse--;
1877                 set_freepointer(s, freelist, old.freelist);
1878                 new.freelist = freelist;
1879         } else
1880                 new.freelist = old.freelist;
1881 
1882         new.frozen = 0;
1883 
1884         if (!new.inuse && n->nr_partial >= s->min_partial)
1885                 m = M_FREE;
1886         else if (new.freelist) {
1887                 m = M_PARTIAL;
1888                 if (!lock) {
1889                         lock = 1;
1890                         /*
1891                          * Taking the spinlock removes the possiblity
1892                          * that acquire_slab() will see a slab page that
1893                          * is frozen
1894                          */
1895                         spin_lock(&n->list_lock);
1896                 }
1897         } else {
1898                 m = M_FULL;
1899                 if (kmem_cache_debug(s) && !lock) {
1900                         lock = 1;
1901                         /*
1902                          * This also ensures that the scanning of full
1903                          * slabs from diagnostic functions will not see
1904                          * any frozen slabs.
1905                          */
1906                         spin_lock(&n->list_lock);
1907                 }
1908         }
1909 
1910         if (l != m) {
1911 
1912                 if (l == M_PARTIAL)
1913 
1914                         remove_partial(n, page);
1915 
1916                 else if (l == M_FULL)
1917 
1918                         remove_full(s, n, page);
1919 
1920                 if (m == M_PARTIAL) {
1921 
1922                         add_partial(n, page, tail);
1923                         stat(s, tail);
1924 
1925                 } else if (m == M_FULL) {
1926 
1927                         stat(s, DEACTIVATE_FULL);
1928                         add_full(s, n, page);
1929 
1930                 }
1931         }
1932 
1933         l = m;
1934         if (!__cmpxchg_double_slab(s, page,
1935                                 old.freelist, old.counters,
1936                                 new.freelist, new.counters,
1937                                 "unfreezing slab"))
1938                 goto redo;
1939 
1940         if (lock)
1941                 spin_unlock(&n->list_lock);
1942 
1943         if (m == M_FREE) {
1944                 stat(s, DEACTIVATE_EMPTY);
1945                 discard_slab(s, page);
1946                 stat(s, FREE_SLAB);
1947         }
1948 }
1949 
1950 /*
1951  * Unfreeze all the cpu partial slabs.
1952  *
1953  * This function must be called with interrupts disabled
1954  * for the cpu using c (or some other guarantee must be there
1955  * to guarantee no concurrent accesses).
1956  */
1957 static void unfreeze_partials(struct kmem_cache *s,
1958                 struct kmem_cache_cpu *c)
1959 {
1960 #ifdef CONFIG_SLUB_CPU_PARTIAL
1961         struct kmem_cache_node *n = NULL, *n2 = NULL;
1962         struct page *page, *discard_page = NULL;
1963 
1964         while ((page = c->partial)) {
1965                 struct page new;
1966                 struct page old;
1967 
1968                 c->partial = page->next;
1969 
1970                 n2 = get_node(s, page_to_nid(page));
1971                 if (n != n2) {
1972                         if (n)
1973                                 spin_unlock(&n->list_lock);
1974 
1975                         n = n2;
1976                         spin_lock(&n->list_lock);
1977                 }
1978 
1979                 do {
1980 
1981                         old.freelist = page->freelist;
1982                         old.counters = page->counters;
1983                         VM_BUG_ON(!old.frozen);
1984 
1985                         new.counters = old.counters;
1986                         new.freelist = old.freelist;
1987 
1988                         new.frozen = 0;
1989 
1990                 } while (!__cmpxchg_double_slab(s, page,
1991                                 old.freelist, old.counters,
1992                                 new.freelist, new.counters,
1993                                 "unfreezing slab"));
1994 
1995                 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
1996                         page->next = discard_page;
1997                         discard_page = page;
1998                 } else {
1999                         add_partial(n, page, DEACTIVATE_TO_TAIL);
2000                         stat(s, FREE_ADD_PARTIAL);
2001                 }
2002         }
2003 
2004         if (n)
2005                 spin_unlock(&n->list_lock);
2006 
2007         while (discard_page) {
2008                 page = discard_page;
2009                 discard_page = discard_page->next;
2010 
2011                 stat(s, DEACTIVATE_EMPTY);
2012                 discard_slab(s, page);
2013                 stat(s, FREE_SLAB);
2014         }
2015 #endif
2016 }
2017 
2018 /*
2019  * Put a page that was just frozen (in __slab_free) into a partial page
2020  * slot if available. This is done without interrupts disabled and without
2021  * preemption disabled. The cmpxchg is racy and may put the partial page
2022  * onto a random cpus partial slot.
2023  *
2024  * If we did not find a slot then simply move all the partials to the
2025  * per node partial list.
2026  */
2027 static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2028 {
2029 #ifdef CONFIG_SLUB_CPU_PARTIAL
2030         struct page *oldpage;
2031         int pages;
2032         int pobjects;
2033 
2034         do {
2035                 pages = 0;
2036                 pobjects = 0;
2037                 oldpage = this_cpu_read(s->cpu_slab->partial);
2038 
2039                 if (oldpage) {
2040                         pobjects = oldpage->pobjects;
2041                         pages = oldpage->pages;
2042                         if (drain && pobjects > s->cpu_partial) {
2043                                 unsigned long flags;
2044                                 /*
2045                                  * partial array is full. Move the existing
2046                                  * set to the per node partial list.
2047                                  */
2048                                 local_irq_save(flags);
2049                                 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2050                                 local_irq_restore(flags);
2051                                 oldpage = NULL;
2052                                 pobjects = 0;
2053                                 pages = 0;
2054                                 stat(s, CPU_PARTIAL_DRAIN);
2055                         }
2056                 }
2057 
2058                 pages++;
2059                 pobjects += page->objects - page->inuse;
2060 
2061                 page->pages = pages;
2062                 page->pobjects = pobjects;
2063                 page->next = oldpage;
2064 
2065         } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2066                                                                 != oldpage);
2067 #endif
2068 }
2069 
2070 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2071 {
2072         stat(s, CPUSLAB_FLUSH);
2073         deactivate_slab(s, c->page, c->freelist);
2074 
2075         c->tid = next_tid(c->tid);
2076         c->page = NULL;
2077         c->freelist = NULL;
2078 }
2079 
2080 /*
2081  * Flush cpu slab.
2082  *
2083  * Called from IPI handler with interrupts disabled.
2084  */
2085 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2086 {
2087         struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2088 
2089         if (likely(c)) {
2090                 if (c->page)
2091                         flush_slab(s, c);
2092 
2093                 unfreeze_partials(s, c);
2094         }
2095 }
2096 
2097 static void flush_cpu_slab(void *d)
2098 {
2099         struct kmem_cache *s = d;
2100 
2101         __flush_cpu_slab(s, smp_processor_id());
2102 }
2103 
2104 static bool has_cpu_slab(int cpu, void *info)
2105 {
2106         struct kmem_cache *s = info;
2107         struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2108 
2109         return c->page || c->partial;
2110 }
2111 
2112 static void flush_all(struct kmem_cache *s)
2113 {
2114         on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2115 }
2116 
2117 /*
2118  * Check if the objects in a per cpu structure fit numa
2119  * locality expectations.
2120  */
2121 static inline int node_match(struct page *page, int node)
2122 {
2123 #ifdef CONFIG_NUMA
2124         if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2125                 return 0;
2126 #endif
2127         return 1;
2128 }
2129 
2130 #ifdef CONFIG_SLUB_DEBUG
2131 static int count_free(struct page *page)
2132 {
2133         return page->objects - page->inuse;
2134 }
2135 
2136 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2137 {
2138         return atomic_long_read(&n->total_objects);
2139 }
2140 #endif /* CONFIG_SLUB_DEBUG */
2141 
2142 #if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2143 static unsigned long count_partial(struct kmem_cache_node *n,
2144                                         int (*get_count)(struct page *))
2145 {
2146         unsigned long flags;
2147         unsigned long x = 0;
2148         struct page *page;
2149 
2150         spin_lock_irqsave(&n->list_lock, flags);
2151         list_for_each_entry(page, &n->partial, lru)
2152                 x += get_count(page);
2153         spin_unlock_irqrestore(&n->list_lock, flags);
2154         return x;
2155 }
2156 #endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2157 
2158 static noinline void
2159 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2160 {
2161 #ifdef CONFIG_SLUB_DEBUG
2162         static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2163                                       DEFAULT_RATELIMIT_BURST);
2164         int node;
2165 
2166         if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2167                 return;
2168 
2169         pr_warn("SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2170                 nid, gfpflags);
2171         pr_warn("  cache: %s, object size: %d, buffer size: %d, default order: %d, min order: %d\n",
2172                 s->name, s->object_size, s->size, oo_order(s->oo),
2173                 oo_order(s->min));
2174 
2175         if (oo_order(s->min) > get_order(s->object_size))
2176                 pr_warn("  %s debugging increased min order, use slub_debug=O to disable.\n",
2177                         s->name);
2178 
2179         for_each_online_node(node) {
2180                 struct kmem_cache_node *n = get_node(s, node);
2181                 unsigned long nr_slabs;
2182                 unsigned long nr_objs;
2183                 unsigned long nr_free;
2184 
2185                 if (!n)
2186                         continue;
2187 
2188                 nr_free  = count_partial(n, count_free);
2189                 nr_slabs = node_nr_slabs(n);
2190                 nr_objs  = node_nr_objs(n);
2191 
2192                 pr_warn("  node %d: slabs: %ld, objs: %ld, free: %ld\n",
2193                         node, nr_slabs, nr_objs, nr_free);
2194         }
2195 #endif
2196 }
2197 
2198 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2199                         int node, struct kmem_cache_cpu **pc)
2200 {
2201         void *freelist;
2202         struct kmem_cache_cpu *c = *pc;
2203         struct page *page;
2204 
2205         freelist = get_partial(s, flags, node, c);
2206 
2207         if (freelist)
2208                 return freelist;
2209 
2210         page = new_slab(s, flags, node);
2211         if (page) {
2212                 c = raw_cpu_ptr(s->cpu_slab);
2213                 if (c->page)
2214                         flush_slab(s, c);
2215 
2216                 /*
2217                  * No other reference to the page yet so we can
2218                  * muck around with it freely without cmpxchg
2219                  */
2220                 freelist = page->freelist;
2221                 page->freelist = NULL;
2222 
2223                 stat(s, ALLOC_SLAB);
2224                 c->page = page;
2225                 *pc = c;
2226         } else
2227                 freelist = NULL;
2228 
2229         return freelist;
2230 }
2231 
2232 static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2233 {
2234         if (unlikely(PageSlabPfmemalloc(page)))
2235                 return gfp_pfmemalloc_allowed(gfpflags);
2236 
2237         return true;
2238 }
2239 
2240 /*
2241  * Check the page->freelist of a page and either transfer the freelist to the
2242  * per cpu freelist or deactivate the page.
2243  *
2244  * The page is still frozen if the return value is not NULL.
2245  *
2246  * If this function returns NULL then the page has been unfrozen.
2247  *
2248  * This function must be called with interrupt disabled.
2249  */
2250 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2251 {
2252         struct page new;
2253         unsigned long counters;
2254         void *freelist;
2255 
2256         do {
2257                 freelist = page->freelist;
2258                 counters = page->counters;
2259 
2260                 new.counters = counters;
2261                 VM_BUG_ON(!new.frozen);
2262 
2263                 new.inuse = page->objects;
2264                 new.frozen = freelist != NULL;
2265 
2266         } while (!__cmpxchg_double_slab(s, page,
2267                 freelist, counters,
2268                 NULL, new.counters,
2269                 "get_freelist"));
2270 
2271         return freelist;
2272 }
2273 
2274 /*
2275  * Slow path. The lockless freelist is empty or we need to perform
2276  * debugging duties.
2277  *
2278  * Processing is still very fast if new objects have been freed to the
2279  * regular freelist. In that case we simply take over the regular freelist
2280  * as the lockless freelist and zap the regular freelist.
2281  *
2282  * If that is not working then we fall back to the partial lists. We take the
2283  * first element of the freelist as the object to allocate now and move the
2284  * rest of the freelist to the lockless freelist.
2285  *
2286  * And if we were unable to get a new slab from the partial slab lists then
2287  * we need to allocate a new slab. This is the slowest path since it involves
2288  * a call to the page allocator and the setup of a new slab.
2289  */
2290 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2291                           unsigned long addr, struct kmem_cache_cpu *c)
2292 {
2293         void *freelist;
2294         struct page *page;
2295         unsigned long flags;
2296 
2297         local_irq_save(flags);
2298 #ifdef CONFIG_PREEMPT
2299         /*
2300          * We may have been preempted and rescheduled on a different
2301          * cpu before disabling interrupts. Need to reload cpu area
2302          * pointer.
2303          */
2304         c = this_cpu_ptr(s->cpu_slab);
2305 #endif
2306 
2307         page = c->page;
2308         if (!page)
2309                 goto new_slab;
2310 redo:
2311 
2312         if (unlikely(!node_match(page, node))) {
2313                 stat(s, ALLOC_NODE_MISMATCH);
2314                 deactivate_slab(s, page, c->freelist);
2315                 c->page = NULL;
2316                 c->freelist = NULL;
2317                 goto new_slab;
2318         }
2319 
2320         /*
2321          * By rights, we should be searching for a slab page that was
2322          * PFMEMALLOC but right now, we are losing the pfmemalloc
2323          * information when the page leaves the per-cpu allocator
2324          */
2325         if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2326                 deactivate_slab(s, page, c->freelist);
2327                 c->page = NULL;
2328                 c->freelist = NULL;
2329                 goto new_slab;
2330         }
2331 
2332         /* must check again c->freelist in case of cpu migration or IRQ */
2333         freelist = c->freelist;
2334         if (freelist)
2335                 goto load_freelist;
2336 
2337         freelist = get_freelist(s, page);
2338 
2339         if (!freelist) {
2340                 c->page = NULL;
2341                 stat(s, DEACTIVATE_BYPASS);
2342                 goto new_slab;
2343         }
2344 
2345         stat(s, ALLOC_REFILL);
2346 
2347 load_freelist:
2348         /*
2349          * freelist is pointing to the list of objects to be used.
2350          * page is pointing to the page from which the objects are obtained.
2351          * That page must be frozen for per cpu allocations to work.
2352          */
2353         VM_BUG_ON(!c->page->frozen);
2354         c->freelist = get_freepointer(s, freelist);
2355         c->tid = next_tid(c->tid);
2356         local_irq_restore(flags);
2357         return freelist;
2358 
2359 new_slab:
2360 
2361         if (c->partial) {
2362                 page = c->page = c->partial;
2363                 c->partial = page->next;
2364                 stat(s, CPU_PARTIAL_ALLOC);
2365                 c->freelist = NULL;
2366                 goto redo;
2367         }
2368 
2369         freelist = new_slab_objects(s, gfpflags, node, &c);
2370 
2371         if (unlikely(!freelist)) {
2372                 slab_out_of_memory(s, gfpflags, node);
2373                 local_irq_restore(flags);
2374                 return NULL;
2375         }
2376 
2377         page = c->page;
2378         if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2379                 goto load_freelist;
2380 
2381         /* Only entered in the debug case */
2382         if (kmem_cache_debug(s) &&
2383                         !alloc_debug_processing(s, page, freelist, addr))
2384                 goto new_slab;  /* Slab failed checks. Next slab needed */
2385 
2386         deactivate_slab(s, page, get_freepointer(s, freelist));
2387         c->page = NULL;
2388         c->freelist = NULL;
2389         local_irq_restore(flags);
2390         return freelist;
2391 }
2392 
2393 /*
2394  * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2395  * have the fastpath folded into their functions. So no function call
2396  * overhead for requests that can be satisfied on the fastpath.
2397  *
2398  * The fastpath works by first checking if the lockless freelist can be used.
2399  * If not then __slab_alloc is called for slow processing.
2400  *
2401  * Otherwise we can simply pick the next object from the lockless free list.
2402  */
2403 static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2404                 gfp_t gfpflags, int node, unsigned long addr)
2405 {
2406         void **object;
2407         struct kmem_cache_cpu *c;
2408         struct page *page;
2409         unsigned long tid;
2410 
2411         if (slab_pre_alloc_hook(s, gfpflags))
2412                 return NULL;
2413 
2414         s = memcg_kmem_get_cache(s, gfpflags);
2415 redo:
2416         /*
2417          * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2418          * enabled. We may switch back and forth between cpus while
2419          * reading from one cpu area. That does not matter as long
2420          * as we end up on the original cpu again when doing the cmpxchg.
2421          *
2422          * Preemption is disabled for the retrieval of the tid because that
2423          * must occur from the current processor. We cannot allow rescheduling
2424          * on a different processor between the determination of the pointer
2425          * and the retrieval of the tid.
2426          */
2427         preempt_disable();
2428         c = this_cpu_ptr(s->cpu_slab);
2429 
2430         /*
2431          * The transaction ids are globally unique per cpu and per operation on
2432          * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2433          * occurs on the right processor and that there was no operation on the
2434          * linked list in between.
2435          */
2436         tid = c->tid;
2437         preempt_enable();
2438 
2439         object = c->freelist;
2440         page = c->page;
2441         if (unlikely(!object || !node_match(page, node))) {
2442                 object = __slab_alloc(s, gfpflags, node, addr, c);
2443                 stat(s, ALLOC_SLOWPATH);
2444         } else {
2445                 void *next_object = get_freepointer_safe(s, object);
2446 
2447                 /*
2448                  * The cmpxchg will only match if there was no additional
2449                  * operation and if we are on the right processor.
2450                  *
2451                  * The cmpxchg does the following atomically (without lock
2452                  * semantics!)
2453                  * 1. Relocate first pointer to the current per cpu area.
2454                  * 2. Verify that tid and freelist have not been changed
2455                  * 3. If they were not changed replace tid and freelist
2456                  *
2457                  * Since this is without lock semantics the protection is only
2458                  * against code executing on this cpu *not* from access by
2459                  * other cpus.
2460                  */
2461                 if (unlikely(!this_cpu_cmpxchg_double(
2462                                 s->cpu_slab->freelist, s->cpu_slab->tid,
2463                                 object, tid,
2464                                 next_object, next_tid(tid)))) {
2465 
2466                         note_cmpxchg_failure("slab_alloc", s, tid);
2467                         goto redo;
2468                 }
2469                 prefetch_freepointer(s, next_object);
2470                 stat(s, ALLOC_FASTPATH);
2471         }
2472 
2473         if (unlikely(gfpflags & __GFP_ZERO) && object)
2474                 memset(object, 0, s->object_size);
2475 
2476         slab_post_alloc_hook(s, gfpflags, object);
2477 
2478         return object;
2479 }
2480 
2481 static __always_inline void *slab_alloc(struct kmem_cache *s,
2482                 gfp_t gfpflags, unsigned long addr)
2483 {
2484         return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2485 }
2486 
2487 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2488 {
2489         void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2490 
2491         trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2492                                 s->size, gfpflags);
2493 
2494         return ret;
2495 }
2496 EXPORT_SYMBOL(kmem_cache_alloc);
2497 
2498 #ifdef CONFIG_TRACING
2499 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2500 {
2501         void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2502         trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2503         return ret;
2504 }
2505 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2506 #endif
2507 
2508 #ifdef CONFIG_NUMA
2509 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2510 {
2511         void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2512 
2513         trace_kmem_cache_alloc_node(_RET_IP_, ret,
2514                                     s->object_size, s->size, gfpflags, node);
2515 
2516         return ret;
2517 }
2518 EXPORT_SYMBOL(kmem_cache_alloc_node);
2519 
2520 #ifdef CONFIG_TRACING
2521 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2522                                     gfp_t gfpflags,
2523                                     int node, size_t size)
2524 {
2525         void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2526 
2527         trace_kmalloc_node(_RET_IP_, ret,
2528                            size, s->size, gfpflags, node);
2529         return ret;
2530 }
2531 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2532 #endif
2533 #endif
2534 
2535 /*
2536  * Slow patch handling. This may still be called frequently since objects
2537  * have a longer lifetime than the cpu slabs in most processing loads.
2538  *
2539  * So we still attempt to reduce cache line usage. Just take the slab
2540  * lock and free the item. If there is no additional partial page
2541  * handling required then we can return immediately.
2542  */
2543 static void __slab_free(struct kmem_cache *s, struct page *page,
2544                         void *x, unsigned long addr)
2545 {
2546         void *prior;
2547         void **object = (void *)x;
2548         int was_frozen;
2549         struct page new;
2550         unsigned long counters;
2551         struct kmem_cache_node *n = NULL;
2552         unsigned long uninitialized_var(flags);
2553 
2554         stat(s, FREE_SLOWPATH);
2555 
2556         if (kmem_cache_debug(s) &&
2557                 !(n = free_debug_processing(s, page, x, addr, &flags)))
2558                 return;
2559 
2560         do {
2561                 if (unlikely(n)) {
2562                         spin_unlock_irqrestore(&n->list_lock, flags);
2563                         n = NULL;
2564                 }
2565                 prior = page->freelist;
2566                 counters = page->counters;
2567                 set_freepointer(s, object, prior);
2568                 new.counters = counters;
2569                 was_frozen = new.frozen;
2570                 new.inuse--;
2571                 if ((!new.inuse || !prior) && !was_frozen) {
2572 
2573                         if (kmem_cache_has_cpu_partial(s) && !prior) {
2574 
2575                                 /*
2576                                  * Slab was on no list before and will be
2577                                  * partially empty
2578                                  * We can defer the list move and instead
2579                                  * freeze it.
2580                                  */
2581                                 new.frozen = 1;
2582 
2583                         } else { /* Needs to be taken off a list */
2584 
2585                                 n = get_node(s, page_to_nid(page));
2586                                 /*
2587                                  * Speculatively acquire the list_lock.
2588                                  * If the cmpxchg does not succeed then we may
2589                                  * drop the list_lock without any processing.
2590                                  *
2591                                  * Otherwise the list_lock will synchronize with
2592                                  * other processors updating the list of slabs.
2593                                  */
2594                                 spin_lock_irqsave(&n->list_lock, flags);
2595 
2596                         }
2597                 }
2598 
2599         } while (!cmpxchg_double_slab(s, page,
2600                 prior, counters,
2601                 object, new.counters,
2602                 "__slab_free"));
2603 
2604         if (likely(!n)) {
2605 
2606                 /*
2607                  * If we just froze the page then put it onto the
2608                  * per cpu partial list.
2609                  */
2610                 if (new.frozen && !was_frozen) {
2611                         put_cpu_partial(s, page, 1);
2612                         stat(s, CPU_PARTIAL_FREE);
2613                 }
2614                 /*
2615                  * The list lock was not taken therefore no list
2616                  * activity can be necessary.
2617                  */
2618                 if (was_frozen)
2619                         stat(s, FREE_FROZEN);
2620                 return;
2621         }
2622 
2623         if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
2624                 goto slab_empty;
2625 
2626         /*
2627          * Objects left in the slab. If it was not on the partial list before
2628          * then add it.
2629          */
2630         if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2631                 if (kmem_cache_debug(s))
2632                         remove_full(s, n, page);
2633                 add_partial(n, page, DEACTIVATE_TO_TAIL);
2634                 stat(s, FREE_ADD_PARTIAL);
2635         }
2636         spin_unlock_irqrestore(&n->list_lock, flags);
2637         return;
2638 
2639 slab_empty:
2640         if (prior) {
2641                 /*
2642                  * Slab on the partial list.
2643                  */
2644                 remove_partial(n, page);
2645                 stat(s, FREE_REMOVE_PARTIAL);
2646         } else {
2647                 /* Slab must be on the full list */
2648                 remove_full(s, n, page);
2649         }
2650 
2651         spin_unlock_irqrestore(&n->list_lock, flags);
2652         stat(s, FREE_SLAB);
2653         discard_slab(s, page);
2654 }
2655 
2656 /*
2657  * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2658  * can perform fastpath freeing without additional function calls.
2659  *
2660  * The fastpath is only possible if we are freeing to the current cpu slab
2661  * of this processor. This typically the case if we have just allocated
2662  * the item before.
2663  *
2664  * If fastpath is not possible then fall back to __slab_free where we deal
2665  * with all sorts of special processing.
2666  */
2667 static __always_inline void slab_free(struct kmem_cache *s,
2668                         struct page *page, void *x, unsigned long addr)
2669 {
2670         void **object = (void *)x;
2671         struct kmem_cache_cpu *c;
2672         unsigned long tid;
2673 
2674         slab_free_hook(s, x);
2675 
2676 redo:
2677         /*
2678          * Determine the currently cpus per cpu slab.
2679          * The cpu may change afterward. However that does not matter since
2680          * data is retrieved via this pointer. If we are on the same cpu
2681          * during the cmpxchg then the free will succedd.
2682          */
2683         preempt_disable();
2684         c = this_cpu_ptr(s->cpu_slab);
2685 
2686         tid = c->tid;
2687         preempt_enable();
2688 
2689         if (likely(page == c->page)) {
2690                 set_freepointer(s, object, c->freelist);
2691 
2692                 if (unlikely(!this_cpu_cmpxchg_double(
2693                                 s->cpu_slab->freelist, s->cpu_slab->tid,
2694                                 c->freelist, tid,
2695                                 object, next_tid(tid)))) {
2696 
2697                         note_cmpxchg_failure("slab_free", s, tid);
2698                         goto redo;
2699                 }
2700                 stat(s, FREE_FASTPATH);
2701         } else
2702                 __slab_free(s, page, x, addr);
2703 
2704 }
2705 
2706 void kmem_cache_free(struct kmem_cache *s, void *x)
2707 {
2708         s = cache_from_obj(s, x);
2709         if (!s)
2710                 return;
2711         slab_free(s, virt_to_head_page(x), x, _RET_IP_);
2712         trace_kmem_cache_free(_RET_IP_, x);
2713 }
2714 EXPORT_SYMBOL(kmem_cache_free);
2715 
2716 /*
2717  * Object placement in a slab is made very easy because we always start at
2718  * offset 0. If we tune the size of the object to the alignment then we can
2719  * get the required alignment by putting one properly sized object after
2720  * another.
2721  *
2722  * Notice that the allocation order determines the sizes of the per cpu
2723  * caches. Each processor has always one slab available for allocations.
2724  * Increasing the allocation order reduces the number of times that slabs
2725  * must be moved on and off the partial lists and is therefore a factor in
2726  * locking overhead.
2727  */
2728 
2729 /*
2730  * Mininum / Maximum order of slab pages. This influences locking overhead
2731  * and slab fragmentation. A higher order reduces the number of partial slabs
2732  * and increases the number of allocations possible without having to
2733  * take the list_lock.
2734  */
2735 static int slub_min_order;
2736 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2737 static int slub_min_objects;
2738 
2739 /*
2740  * Merge control. If this is set then no merging of slab caches will occur.
2741  * (Could be removed. This was introduced to pacify the merge skeptics.)
2742  */
2743 static int slub_nomerge;
2744 
2745 /*
2746  * Calculate the order of allocation given an slab object size.
2747  *
2748  * The order of allocation has significant impact on performance and other
2749  * system components. Generally order 0 allocations should be preferred since
2750  * order 0 does not cause fragmentation in the page allocator. Larger objects
2751  * be problematic to put into order 0 slabs because there may be too much
2752  * unused space left. We go to a higher order if more than 1/16th of the slab
2753  * would be wasted.
2754  *
2755  * In order to reach satisfactory performance we must ensure that a minimum
2756  * number of objects is in one slab. Otherwise we may generate too much
2757  * activity on the partial lists which requires taking the list_lock. This is
2758  * less a concern for large slabs though which are rarely used.
2759  *
2760  * slub_max_order specifies the order where we begin to stop considering the
2761  * number of objects in a slab as critical. If we reach slub_max_order then
2762  * we try to keep the page order as low as possible. So we accept more waste
2763  * of space in favor of a small page order.
2764  *
2765  * Higher order allocations also allow the placement of more objects in a
2766  * slab and thereby reduce object handling overhead. If the user has
2767  * requested a higher mininum order then we start with that one instead of
2768  * the smallest order which will fit the object.
2769  */
2770 static inline int slab_order(int size, int min_objects,
2771                                 int max_order, int fract_leftover, int reserved)
2772 {
2773         int order;
2774         int rem;
2775         int min_order = slub_min_order;
2776 
2777         if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2778                 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2779 
2780         for (order = max(min_order,
2781                                 fls(min_objects * size - 1) - PAGE_SHIFT);
2782                         order <= max_order; order++) {
2783 
2784                 unsigned long slab_size = PAGE_SIZE << order;
2785 
2786                 if (slab_size < min_objects * size + reserved)
2787                         continue;
2788 
2789                 rem = (slab_size - reserved) % size;
2790 
2791                 if (rem <= slab_size / fract_leftover)
2792                         break;
2793 
2794         }
2795 
2796         return order;
2797 }
2798 
2799 static inline int calculate_order(int size, int reserved)
2800 {
2801         int order;
2802         int min_objects;
2803         int fraction;
2804         int max_objects;
2805 
2806         /*
2807          * Attempt to find best configuration for a slab. This
2808          * works by first attempting to generate a layout with
2809          * the best configuration and backing off gradually.
2810          *
2811          * First we reduce the acceptable waste in a slab. Then
2812          * we reduce the minimum objects required in a slab.
2813          */
2814         min_objects = slub_min_objects;
2815         if (!min_objects)
2816                 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2817         max_objects = order_objects(slub_max_order, size, reserved);
2818         min_objects = min(min_objects, max_objects);
2819 
2820         while (min_objects > 1) {
2821                 fraction = 16;
2822                 while (fraction >= 4) {
2823                         order = slab_order(size, min_objects,
2824                                         slub_max_order, fraction, reserved);
2825                         if (order <= slub_max_order)
2826                                 return order;
2827                         fraction /= 2;
2828                 }
2829                 min_objects--;
2830         }
2831 
2832         /*
2833          * We were unable to place multiple objects in a slab. Now
2834          * lets see if we can place a single object there.
2835          */
2836         order = slab_order(size, 1, slub_max_order, 1, reserved);
2837         if (order <= slub_max_order)
2838                 return order;
2839 
2840         /*
2841          * Doh this slab cannot be placed using slub_max_order.
2842          */
2843         order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2844         if (order < MAX_ORDER)
2845                 return order;
2846         return -ENOSYS;
2847 }
2848 
2849 static void
2850 init_kmem_cache_node(struct kmem_cache_node *n)
2851 {
2852         n->nr_partial = 0;
2853         spin_lock_init(&n->list_lock);
2854         INIT_LIST_HEAD(&n->partial);
2855 #ifdef CONFIG_SLUB_DEBUG
2856         atomic_long_set(&n->nr_slabs, 0);
2857         atomic_long_set(&n->total_objects, 0);
2858         INIT_LIST_HEAD(&n->full);
2859 #endif
2860 }
2861 
2862 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2863 {
2864         BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2865                         KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
2866 
2867         /*
2868          * Must align to double word boundary for the double cmpxchg
2869          * instructions to work; see __pcpu_double_call_return_bool().
2870          */
2871         s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2872                                      2 * sizeof(void *));
2873 
2874         if (!s->cpu_slab)
2875                 return 0;
2876 
2877         init_kmem_cache_cpus(s);
2878 
2879         return 1;
2880 }
2881 
2882 static struct kmem_cache *kmem_cache_node;
2883 
2884 /*
2885  * No kmalloc_node yet so do it by hand. We know that this is the first
2886  * slab on the node for this slabcache. There are no concurrent accesses
2887  * possible.
2888  *
2889  * Note that this function only works on the kmem_cache_node
2890  * when allocating for the kmem_cache_node. This is used for bootstrapping
2891  * memory on a fresh node that has no slab structures yet.
2892  */
2893 static void early_kmem_cache_node_alloc(int node)
2894 {
2895         struct page *page;
2896         struct kmem_cache_node *n;
2897 
2898         BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2899 
2900         page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2901 
2902         BUG_ON(!page);
2903         if (page_to_nid(page) != node) {
2904                 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
2905                 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
2906         }
2907 
2908         n = page->freelist;
2909         BUG_ON(!n);
2910         page->freelist = get_freepointer(kmem_cache_node, n);
2911         page->inuse = 1;
2912         page->frozen = 0;
2913         kmem_cache_node->node[node] = n;
2914 #ifdef CONFIG_SLUB_DEBUG
2915         init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2916         init_tracking(kmem_cache_node, n);
2917 #endif
2918         init_kmem_cache_node(n);
2919         inc_slabs_node(kmem_cache_node, node, page->objects);
2920 
2921         /*
2922          * No locks need to be taken here as it has just been
2923          * initialized and there is no concurrent access.
2924          */
2925         __add_partial(n, page, DEACTIVATE_TO_HEAD);
2926 }
2927 
2928 static void free_kmem_cache_nodes(struct kmem_cache *s)
2929 {
2930         int node;
2931 
2932         for_each_node_state(node, N_NORMAL_MEMORY) {
2933                 struct kmem_cache_node *n = s->node[node];
2934 
2935                 if (n)
2936                         kmem_cache_free(kmem_cache_node, n);
2937 
2938                 s->node[node] = NULL;
2939         }
2940 }
2941 
2942 static int init_kmem_cache_nodes(struct kmem_cache *s)
2943 {
2944         int node;
2945 
2946         for_each_node_state(node, N_NORMAL_MEMORY) {
2947                 struct kmem_cache_node *n;
2948 
2949                 if (slab_state == DOWN) {
2950                         early_kmem_cache_node_alloc(node);
2951                         continue;
2952                 }
2953                 n = kmem_cache_alloc_node(kmem_cache_node,
2954                                                 GFP_KERNEL, node);
2955 
2956                 if (!n) {
2957                         free_kmem_cache_nodes(s);
2958                         return 0;
2959                 }
2960 
2961                 s->node[node] = n;
2962                 init_kmem_cache_node(n);
2963         }
2964         return 1;
2965 }
2966 
2967 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2968 {
2969         if (min < MIN_PARTIAL)
2970                 min = MIN_PARTIAL;
2971         else if (min > MAX_PARTIAL)
2972                 min = MAX_PARTIAL;
2973         s->min_partial = min;
2974 }
2975 
2976 /*
2977  * calculate_sizes() determines the order and the distribution of data within
2978  * a slab object.
2979  */
2980 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2981 {
2982         unsigned long flags = s->flags;
2983         unsigned long size = s->object_size;
2984         int order;
2985 
2986         /*
2987          * Round up object size to the next word boundary. We can only
2988          * place the free pointer at word boundaries and this determines
2989          * the possible location of the free pointer.
2990          */
2991         size = ALIGN(size, sizeof(void *));
2992 
2993 #ifdef CONFIG_SLUB_DEBUG
2994         /*
2995          * Determine if we can poison the object itself. If the user of
2996          * the slab may touch the object after free or before allocation
2997          * then we should never poison the object itself.
2998          */
2999         if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
3000                         !s->ctor)
3001                 s->flags |= __OBJECT_POISON;
3002         else
3003                 s->flags &= ~__OBJECT_POISON;
3004 
3005 
3006         /*
3007          * If we are Redzoning then check if there is some space between the
3008          * end of the object and the free pointer. If not then add an
3009          * additional word to have some bytes to store Redzone information.
3010          */
3011         if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3012                 size += sizeof(void *);
3013 #endif
3014 
3015         /*
3016          * With that we have determined the number of bytes in actual use
3017          * by the object. This is the potential offset to the free pointer.
3018          */
3019         s->inuse = size;
3020 
3021         if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
3022                 s->ctor)) {
3023                 /*
3024                  * Relocate free pointer after the object if it is not
3025                  * permitted to overwrite the first word of the object on
3026                  * kmem_cache_free.
3027                  *
3028                  * This is the case if we do RCU, have a constructor or
3029                  * destructor or are poisoning the objects.
3030                  */
3031                 s->offset = size;
3032                 size += sizeof(void *);
3033         }
3034 
3035 #ifdef CONFIG_SLUB_DEBUG
3036         if (flags & SLAB_STORE_USER)
3037                 /*
3038                  * Need to store information about allocs and frees after
3039                  * the object.
3040                  */
3041                 size += 2 * sizeof(struct track);
3042 
3043         if (flags & SLAB_RED_ZONE)
3044                 /*
3045                  * Add some empty padding so that we can catch
3046                  * overwrites from earlier objects rather than let
3047                  * tracking information or the free pointer be
3048                  * corrupted if a user writes before the start
3049                  * of the object.
3050                  */
3051                 size += sizeof(void *);
3052 #endif
3053 
3054         /*
3055          * SLUB stores one object immediately after another beginning from
3056          * offset 0. In order to align the objects we have to simply size
3057          * each object to conform to the alignment.
3058          */
3059         size = ALIGN(size, s->align);
3060         s->size = size;
3061         if (forced_order >= 0)
3062                 order = forced_order;
3063         else
3064                 order = calculate_order(size, s->reserved);
3065 
3066         if (order < 0)
3067                 return 0;
3068 
3069         s->allocflags = 0;
3070         if (order)
3071                 s->allocflags |= __GFP_COMP;
3072 
3073         if (s->flags & SLAB_CACHE_DMA)
3074                 s->allocflags |= GFP_DMA;
3075 
3076         if (s->flags & SLAB_RECLAIM_ACCOUNT)
3077                 s->allocflags |= __GFP_RECLAIMABLE;
3078 
3079         /*
3080          * Determine the number of objects per slab
3081          */
3082         s->oo = oo_make(order, size, s->reserved);
3083         s->min = oo_make(get_order(size), size, s->reserved);
3084         if (oo_objects(s->oo) > oo_objects(s->max))
3085                 s->max = s->oo;
3086 
3087         return !!oo_objects(s->oo);
3088 }
3089 
3090 static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3091 {
3092         s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3093         s->reserved = 0;
3094 
3095         if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3096                 s->reserved = sizeof(struct rcu_head);
3097 
3098         if (!calculate_sizes(s, -1))
3099                 goto error;
3100         if (disable_higher_order_debug) {
3101                 /*
3102                  * Disable debugging flags that store metadata if the min slab
3103                  * order increased.
3104                  */
3105                 if (get_order(s->size) > get_order(s->object_size)) {
3106                         s->flags &= ~DEBUG_METADATA_FLAGS;
3107                         s->offset = 0;
3108                         if (!calculate_sizes(s, -1))
3109                                 goto error;
3110                 }
3111         }
3112 
3113 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3114     defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3115         if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3116                 /* Enable fast mode */
3117                 s->flags |= __CMPXCHG_DOUBLE;
3118 #endif
3119 
3120         /*
3121          * The larger the object size is, the more pages we want on the partial
3122          * list to avoid pounding the page allocator excessively.
3123          */
3124         set_min_partial(s, ilog2(s->size) / 2);
3125 
3126         /*
3127          * cpu_partial determined the maximum number of objects kept in the
3128          * per cpu partial lists of a processor.
3129          *
3130          * Per cpu partial lists mainly contain slabs that just have one
3131          * object freed. If they are used for allocation then they can be
3132          * filled up again with minimal effort. The slab will never hit the
3133          * per node partial lists and therefore no locking will be required.
3134          *
3135          * This setting also determines
3136          *
3137          * A) The number of objects from per cpu partial slabs dumped to the
3138          *    per node list when we reach the limit.
3139          * B) The number of objects in cpu partial slabs to extract from the
3140          *    per node list when we run out of per cpu objects. We only fetch
3141          *    50% to keep some capacity around for frees.
3142          */
3143         if (!kmem_cache_has_cpu_partial(s))
3144                 s->cpu_partial = 0;
3145         else if (s->size >= PAGE_SIZE)
3146                 s->cpu_partial = 2;
3147         else if (s->size >= 1024)
3148                 s->cpu_partial = 6;
3149         else if (s->size >= 256)
3150                 s->cpu_partial = 13;
3151         else
3152                 s->cpu_partial = 30;
3153 
3154 #ifdef CONFIG_NUMA
3155         s->remote_node_defrag_ratio = 1000;
3156 #endif
3157         if (!init_kmem_cache_nodes(s))
3158                 goto error;
3159 
3160         if (alloc_kmem_cache_cpus(s))
3161                 return 0;
3162 
3163         free_kmem_cache_nodes(s);
3164 error:
3165         if (flags & SLAB_PANIC)
3166                 panic("Cannot create slab %s size=%lu realsize=%u "
3167                         "order=%u offset=%u flags=%lx\n",
3168                         s->name, (unsigned long)s->size, s->size,
3169                         oo_order(s->oo), s->offset, flags);
3170         return -EINVAL;
3171 }
3172 
3173 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3174                                                         const char *text)
3175 {
3176 #ifdef CONFIG_SLUB_DEBUG
3177         void *addr = page_address(page);
3178         void *p;
3179         unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3180                                      sizeof(long), GFP_ATOMIC);
3181         if (!map)
3182                 return;
3183         slab_err(s, page, text, s->name);
3184         slab_lock(page);
3185 
3186         get_map(s, page, map);
3187         for_each_object(p, s, addr, page->objects) {
3188 
3189                 if (!test_bit(slab_index(p, s, addr), map)) {
3190                         pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3191                         print_tracking(s, p);
3192                 }
3193         }
3194         slab_unlock(page);
3195         kfree(map);
3196 #endif
3197 }
3198 
3199 /*
3200  * Attempt to free all partial slabs on a node.
3201  * This is called from kmem_cache_close(). We must be the last thread
3202  * using the cache and therefore we do not need to lock anymore.
3203  */
3204 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3205 {
3206         struct page *page, *h;
3207 
3208         list_for_each_entry_safe(page, h, &n->partial, lru) {
3209                 if (!page->inuse) {
3210                         __remove_partial(n, page);
3211                         discard_slab(s, page);
3212                 } else {
3213                         list_slab_objects(s, page,
3214                         "Objects remaining in %s on kmem_cache_close()");
3215                 }
3216         }
3217 }
3218 
3219 /*
3220  * Release all resources used by a slab cache.
3221  */
3222 static inline int kmem_cache_close(struct kmem_cache *s)
3223 {
3224         int node;
3225 
3226         flush_all(s);
3227         /* Attempt to free all objects */
3228         for_each_node_state(node, N_NORMAL_MEMORY) {
3229                 struct kmem_cache_node *n = get_node(s, node);
3230 
3231                 free_partial(s, n);
3232                 if (n->nr_partial || slabs_node(s, node))
3233                         return 1;
3234         }
3235         free_percpu(s->cpu_slab);
3236         free_kmem_cache_nodes(s);
3237         return 0;
3238 }
3239 
3240 int __kmem_cache_shutdown(struct kmem_cache *s)
3241 {
3242         return kmem_cache_close(s);
3243 }
3244 
3245 /********************************************************************
3246  *              Kmalloc subsystem
3247  *******************************************************************/
3248 
3249 static int __init setup_slub_min_order(char *str)
3250 {
3251         get_option(&str, &slub_min_order);
3252 
3253         return 1;
3254 }
3255 
3256 __setup("slub_min_order=", setup_slub_min_order);
3257 
3258 static int __init setup_slub_max_order(char *str)
3259 {
3260         get_option(&str, &slub_max_order);
3261         slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3262 
3263         return 1;
3264 }
3265 
3266 __setup("slub_max_order=", setup_slub_max_order);
3267 
3268 static int __init setup_slub_min_objects(char *str)
3269 {
3270         get_option(&str, &slub_min_objects);
3271 
3272         return 1;
3273 }
3274 
3275 __setup("slub_min_objects=", setup_slub_min_objects);
3276 
3277 static int __init setup_slub_nomerge(char *str)
3278 {
3279         slub_nomerge = 1;
3280         return 1;
3281 }
3282 
3283 __setup("slub_nomerge", setup_slub_nomerge);
3284 
3285 void *__kmalloc(size_t size, gfp_t flags)
3286 {
3287         struct kmem_cache *s;
3288         void *ret;
3289 
3290         if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3291                 return kmalloc_large(size, flags);
3292 
3293         s = kmalloc_slab(size, flags);
3294 
3295         if (unlikely(ZERO_OR_NULL_PTR(s)))
3296                 return s;
3297 
3298         ret = slab_alloc(s, flags, _RET_IP_);
3299 
3300         trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3301 
3302         return ret;
3303 }
3304 EXPORT_SYMBOL(__kmalloc);
3305 
3306 #ifdef CONFIG_NUMA
3307 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3308 {
3309         struct page *page;
3310         void *ptr = NULL;
3311 
3312         flags |= __GFP_COMP | __GFP_NOTRACK;
3313         page = alloc_kmem_pages_node(node, flags, get_order(size));
3314         if (page)
3315                 ptr = page_address(page);
3316 
3317         kmalloc_large_node_hook(ptr, size, flags);
3318         return ptr;
3319 }
3320 
3321 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3322 {
3323         struct kmem_cache *s;
3324         void *ret;
3325 
3326         if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3327                 ret = kmalloc_large_node(size, flags, node);
3328 
3329                 trace_kmalloc_node(_RET_IP_, ret,
3330                                    size, PAGE_SIZE << get_order(size),
3331                                    flags, node);
3332 
3333                 return ret;
3334         }
3335 
3336         s = kmalloc_slab(size, flags);
3337 
3338         if (unlikely(ZERO_OR_NULL_PTR(s)))
3339                 return s;
3340 
3341         ret = slab_alloc_node(s, flags, node, _RET_IP_);
3342 
3343         trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3344 
3345         return ret;
3346 }
3347 EXPORT_SYMBOL(__kmalloc_node);
3348 #endif
3349 
3350 size_t ksize(const void *object)
3351 {
3352         struct page *page;
3353 
3354         if (unlikely(object == ZERO_SIZE_PTR))
3355                 return 0;
3356 
3357         page = virt_to_head_page(object);
3358 
3359         if (unlikely(!PageSlab(page))) {
3360                 WARN_ON(!PageCompound(page));
3361                 return PAGE_SIZE << compound_order(page);
3362         }
3363 
3364         return slab_ksize(page->slab_cache);
3365 }
3366 EXPORT_SYMBOL(ksize);
3367 
3368 void kfree(const void *x)
3369 {
3370         struct page *page;
3371         void *object = (void *)x;
3372 
3373         trace_kfree(_RET_IP_, x);
3374 
3375         if (unlikely(ZERO_OR_NULL_PTR(x)))
3376                 return;
3377 
3378         page = virt_to_head_page(x);
3379         if (unlikely(!PageSlab(page))) {
3380                 BUG_ON(!PageCompound(page));
3381                 kfree_hook(x);
3382                 __free_kmem_pages(page, compound_order(page));
3383                 return;
3384         }
3385         slab_free(page->slab_cache, page, object, _RET_IP_);
3386 }
3387 EXPORT_SYMBOL(kfree);
3388 
3389 /*
3390  * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3391  * the remaining slabs by the number of items in use. The slabs with the
3392  * most items in use come first. New allocations will then fill those up
3393  * and thus they can be removed from the partial lists.
3394  *
3395  * The slabs with the least items are placed last. This results in them
3396  * being allocated from last increasing the chance that the last objects
3397  * are freed in them.
3398  */
3399 int __kmem_cache_shrink(struct kmem_cache *s)
3400 {
3401         int node;
3402         int i;
3403         struct kmem_cache_node *n;
3404         struct page *page;
3405         struct page *t;
3406         int objects = oo_objects(s->max);
3407         struct list_head *slabs_by_inuse =
3408                 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3409         unsigned long flags;
3410 
3411         if (!slabs_by_inuse)
3412                 return -ENOMEM;
3413 
3414         flush_all(s);
3415         for_each_node_state(node, N_NORMAL_MEMORY) {
3416                 n = get_node(s, node);
3417 
3418                 if (!n->nr_partial)
3419                         continue;
3420 
3421                 for (i = 0; i < objects; i++)
3422                         INIT_LIST_HEAD(slabs_by_inuse + i);
3423 
3424                 spin_lock_irqsave(&n->list_lock, flags);
3425 
3426                 /*
3427                  * Build lists indexed by the items in use in each slab.
3428                  *
3429                  * Note that concurrent frees may occur while we hold the
3430                  * list_lock. page->inuse here is the upper limit.
3431                  */
3432                 list_for_each_entry_safe(page, t, &n->partial, lru) {
3433                         list_move(&page->lru, slabs_by_inuse + page->inuse);
3434                         if (!page->inuse)
3435                                 n->nr_partial--;
3436                 }
3437 
3438                 /*
3439                  * Rebuild the partial list with the slabs filled up most
3440                  * first and the least used slabs at the end.
3441                  */
3442                 for (i = objects - 1; i > 0; i--)
3443                         list_splice(slabs_by_inuse + i, n->partial.prev);
3444 
3445                 spin_unlock_irqrestore(&n->list_lock, flags);
3446 
3447                 /* Release empty slabs */
3448                 list_for_each_entry_safe(page, t, slabs_by_inuse, lru)
3449                         discard_slab(s, page);
3450         }
3451 
3452         kfree(slabs_by_inuse);
3453         return 0;
3454 }
3455 
3456 static int slab_mem_going_offline_callback(void *arg)
3457 {
3458         struct kmem_cache *s;
3459 
3460         mutex_lock(&slab_mutex);
3461         list_for_each_entry(s, &slab_caches, list)
3462                 __kmem_cache_shrink(s);
3463         mutex_unlock(&slab_mutex);
3464 
3465         return 0;
3466 }
3467 
3468 static void slab_mem_offline_callback(void *arg)
3469 {
3470         struct kmem_cache_node *n;
3471         struct kmem_cache *s;
3472         struct memory_notify *marg = arg;
3473         int offline_node;
3474 
3475         offline_node = marg->status_change_nid_normal;
3476 
3477         /*
3478          * If the node still has available memory. we need kmem_cache_node
3479          * for it yet.
3480          */
3481         if (offline_node < 0)
3482                 return;
3483 
3484         mutex_lock(&slab_mutex);
3485         list_for_each_entry(s, &slab_caches, list) {
3486                 n = get_node(s, offline_node);
3487                 if (n) {
3488                         /*
3489                          * if n->nr_slabs > 0, slabs still exist on the node
3490                          * that is going down. We were unable to free them,
3491                          * and offline_pages() function shouldn't call this
3492                          * callback. So, we must fail.
3493                          */
3494                         BUG_ON(slabs_node(s, offline_node));
3495 
3496                         s->node[offline_node] = NULL;
3497                         kmem_cache_free(kmem_cache_node, n);
3498                 }
3499         }
3500         mutex_unlock(&slab_mutex);
3501 }
3502 
3503 static int slab_mem_going_online_callback(void *arg)
3504 {
3505         struct kmem_cache_node *n;
3506         struct kmem_cache *s;
3507         struct memory_notify *marg = arg;
3508         int nid = marg->status_change_nid_normal;
3509         int ret = 0;
3510 
3511         /*
3512          * If the node's memory is already available, then kmem_cache_node is
3513          * already created. Nothing to do.
3514          */
3515         if (nid < 0)
3516                 return 0;
3517 
3518         /*
3519          * We are bringing a node online. No memory is available yet. We must
3520          * allocate a kmem_cache_node structure in order to bring the node
3521          * online.
3522          */
3523         mutex_lock(&slab_mutex);
3524         list_for_each_entry(s, &slab_caches, list) {
3525                 /*
3526                  * XXX: kmem_cache_alloc_node will fallback to other nodes
3527                  *      since memory is not yet available from the node that
3528                  *      is brought up.
3529                  */
3530                 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3531                 if (!n) {
3532                         ret = -ENOMEM;
3533                         goto out;
3534                 }
3535                 init_kmem_cache_node(n);
3536                 s->node[nid] = n;
3537         }
3538 out:
3539         mutex_unlock(&slab_mutex);
3540         return ret;
3541 }
3542 
3543 static int slab_memory_callback(struct notifier_block *self,
3544                                 unsigned long action, void *arg)
3545 {
3546         int ret = 0;
3547 
3548         switch (action) {
3549         case MEM_GOING_ONLINE:
3550                 ret = slab_mem_going_online_callback(arg);
3551                 break;
3552         case MEM_GOING_OFFLINE:
3553                 ret = slab_mem_going_offline_callback(arg);
3554                 break;
3555         case MEM_OFFLINE:
3556         case MEM_CANCEL_ONLINE:
3557                 slab_mem_offline_callback(arg);
3558                 break;
3559         case MEM_ONLINE:
3560         case MEM_CANCEL_OFFLINE:
3561                 break;
3562         }
3563         if (ret)
3564                 ret = notifier_from_errno(ret);
3565         else
3566                 ret = NOTIFY_OK;
3567         return ret;
3568 }
3569 
3570 static struct notifier_block slab_memory_callback_nb = {
3571         .notifier_call = slab_memory_callback,
3572         .priority = SLAB_CALLBACK_PRI,
3573 };
3574 
3575 /********************************************************************
3576  *                      Basic setup of slabs
3577  *******************************************************************/
3578 
3579 /*
3580  * Used for early kmem_cache structures that were allocated using
3581  * the page allocator. Allocate them properly then fix up the pointers
3582  * that may be pointing to the wrong kmem_cache structure.
3583  */
3584 
3585 static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
3586 {
3587         int node;
3588         struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
3589 
3590         memcpy(s, static_cache, kmem_cache->object_size);
3591 
3592         /*
3593          * This runs very early, and only the boot processor is supposed to be
3594          * up.  Even if it weren't true, IRQs are not up so we couldn't fire
3595          * IPIs around.
3596          */
3597         __flush_cpu_slab(s, smp_processor_id());
3598         for_each_node_state(node, N_NORMAL_MEMORY) {
3599                 struct kmem_cache_node *n = get_node(s, node);
3600                 struct page *p;
3601 
3602                 if (n) {
3603                         list_for_each_entry(p, &n->partial, lru)
3604                                 p->slab_cache = s;
3605 
3606 #ifdef CONFIG_SLUB_DEBUG
3607                         list_for_each_entry(p, &n->full, lru)
3608                                 p->slab_cache = s;
3609 #endif
3610                 }
3611         }
3612         list_add(&s->list, &slab_caches);
3613         return s;
3614 }
3615 
3616 void __init kmem_cache_init(void)
3617 {
3618         static __initdata struct kmem_cache boot_kmem_cache,
3619                 boot_kmem_cache_node;
3620 
3621         if (debug_guardpage_minorder())
3622                 slub_max_order = 0;
3623 
3624         kmem_cache_node = &boot_kmem_cache_node;
3625         kmem_cache = &boot_kmem_cache;
3626 
3627         create_boot_cache(kmem_cache_node, "kmem_cache_node",
3628                 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
3629 
3630         register_hotmemory_notifier(&slab_memory_callback_nb);
3631 
3632         /* Able to allocate the per node structures */
3633         slab_state = PARTIAL;
3634 
3635         create_boot_cache(kmem_cache, "kmem_cache",
3636                         offsetof(struct kmem_cache, node) +
3637                                 nr_node_ids * sizeof(struct kmem_cache_node *),
3638                        SLAB_HWCACHE_ALIGN);
3639 
3640         kmem_cache = bootstrap(&boot_kmem_cache);
3641 
3642         /*
3643          * Allocate kmem_cache_node properly from the kmem_cache slab.
3644          * kmem_cache_node is separately allocated so no need to
3645          * update any list pointers.
3646          */
3647         kmem_cache_node = bootstrap(&boot_kmem_cache_node);
3648 
3649         /* Now we can use the kmem_cache to allocate kmalloc slabs */
3650         create_kmalloc_caches(0);
3651 
3652 #ifdef CONFIG_SMP
3653         register_cpu_notifier(&slab_notifier);
3654 #endif
3655 
3656         pr_info("SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d, CPUs=%d, Nodes=%d\n",
3657                 cache_line_size(),
3658                 slub_min_order, slub_max_order, slub_min_objects,
3659                 nr_cpu_ids, nr_node_ids);
3660 }
3661 
3662 void __init kmem_cache_init_late(void)
3663 {
3664 }
3665 
3666 /*
3667  * Find a mergeable slab cache
3668  */
3669 static int slab_unmergeable(struct kmem_cache *s)
3670 {
3671         if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3672                 return 1;
3673 
3674         if (!is_root_cache(s))
3675                 return 1;
3676 
3677         if (s->ctor)
3678                 return 1;
3679 
3680         /*
3681          * We may have set a slab to be unmergeable during bootstrap.
3682          */
3683         if (s->refcount < 0)
3684                 return 1;
3685 
3686         return 0;
3687 }
3688 
3689 static struct kmem_cache *find_mergeable(size_t size, size_t align,
3690                 unsigned long flags, const char *name, void (*ctor)(void *))
3691 {
3692         struct kmem_cache *s;
3693 
3694         if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3695                 return NULL;
3696 
3697         if (ctor)
3698                 return NULL;
3699 
3700         size = ALIGN(size, sizeof(void *));
3701         align = calculate_alignment(flags, align, size);
3702         size = ALIGN(size, align);
3703         flags = kmem_cache_flags(size, flags, name, NULL);
3704 
3705         list_for_each_entry(s, &slab_caches, list) {
3706                 if (slab_unmergeable(s))
3707                         continue;
3708 
3709                 if (size > s->size)
3710                         continue;
3711 
3712                 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3713                         continue;
3714                 /*
3715                  * Check if alignment is compatible.
3716                  * Courtesy of Adrian Drzewiecki
3717                  */
3718                 if ((s->size & ~(align - 1)) != s->size)
3719                         continue;
3720 
3721                 if (s->size - size >= sizeof(void *))
3722                         continue;
3723 
3724                 return s;
3725         }
3726         return NULL;
3727 }
3728 
3729 struct kmem_cache *
3730 __kmem_cache_alias(const char *name, size_t size, size_t align,
3731                    unsigned long flags, void (*ctor)(void *))
3732 {
3733         struct kmem_cache *s;
3734 
3735         s = find_mergeable(size, align, flags, name, ctor);
3736         if (s) {
3737                 int i;
3738                 struct kmem_cache *c;
3739 
3740                 s->refcount++;
3741 
3742                 /*
3743                  * Adjust the object sizes so that we clear
3744                  * the complete object on kzalloc.
3745                  */
3746                 s->object_size = max(s->object_size, (int)size);
3747                 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3748 
3749                 for_each_memcg_cache_index(i) {
3750                         c = cache_from_memcg_idx(s, i);
3751                         if (!c)
3752                                 continue;
3753                         c->object_size = s->object_size;
3754                         c->inuse = max_t(int, c->inuse,
3755                                          ALIGN(size, sizeof(void *)));
3756                 }
3757 
3758                 if (sysfs_slab_alias(s, name)) {
3759                         s->refcount--;
3760                         s = NULL;
3761                 }
3762         }
3763 
3764         return s;
3765 }
3766 
3767 int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
3768 {
3769         int err;
3770 
3771         err = kmem_cache_open(s, flags);
3772         if (err)
3773                 return err;
3774 
3775         /* Mutex is not taken during early boot */
3776         if (slab_state <= UP)
3777                 return 0;
3778 
3779         memcg_propagate_slab_attrs(s);
3780         err = sysfs_slab_add(s);
3781         if (err)
3782                 kmem_cache_close(s);
3783 
3784         return err;
3785 }
3786 
3787 #ifdef CONFIG_SMP
3788 /*
3789  * Use the cpu notifier to insure that the cpu slabs are flushed when
3790  * necessary.
3791  */
3792 static int slab_cpuup_callback(struct notifier_block *nfb,
3793                 unsigned long action, void *hcpu)
3794 {
3795         long cpu = (long)hcpu;
3796         struct kmem_cache *s;
3797         unsigned long flags;
3798 
3799         switch (action) {
3800         case CPU_UP_CANCELED:
3801         case CPU_UP_CANCELED_FROZEN:
3802         case CPU_DEAD:
3803         case CPU_DEAD_FROZEN:
3804                 mutex_lock(&slab_mutex);
3805                 list_for_each_entry(s, &slab_caches, list) {
3806                         local_irq_save(flags);
3807                         __flush_cpu_slab(s, cpu);
3808                         local_irq_restore(flags);
3809                 }
3810                 mutex_unlock(&slab_mutex);
3811                 break;
3812         default:
3813                 break;
3814         }
3815         return NOTIFY_OK;
3816 }
3817 
3818 static struct notifier_block slab_notifier = {
3819         .notifier_call = slab_cpuup_callback
3820 };
3821 
3822 #endif
3823 
3824 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3825 {
3826         struct kmem_cache *s;
3827         void *ret;
3828 
3829         if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3830                 return kmalloc_large(size, gfpflags);
3831 
3832         s = kmalloc_slab(size, gfpflags);
3833 
3834         if (unlikely(ZERO_OR_NULL_PTR(s)))
3835                 return s;
3836 
3837         ret = slab_alloc(s, gfpflags, caller);
3838 
3839         /* Honor the call site pointer we received. */
3840         trace_kmalloc(caller, ret, size, s->size, gfpflags);
3841 
3842         return ret;
3843 }
3844 
3845 #ifdef CONFIG_NUMA
3846 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3847                                         int node, unsigned long caller)
3848 {
3849         struct kmem_cache *s;
3850         void *ret;
3851 
3852         if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3853                 ret = kmalloc_large_node(size, gfpflags, node);
3854 
3855                 trace_kmalloc_node(caller, ret,
3856                                    size, PAGE_SIZE << get_order(size),
3857                                    gfpflags, node);
3858 
3859                 return ret;
3860         }
3861 
3862         s = kmalloc_slab(size, gfpflags);
3863 
3864         if (unlikely(ZERO_OR_NULL_PTR(s)))
3865                 return s;
3866 
3867         ret = slab_alloc_node(s, gfpflags, node, caller);
3868 
3869         /* Honor the call site pointer we received. */
3870         trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3871 
3872         return ret;
3873 }
3874 #endif
3875 
3876 #ifdef CONFIG_SYSFS
3877 static int count_inuse(struct page *page)
3878 {
3879         return page->inuse;
3880 }
3881 
3882 static int count_total(struct page *page)
3883 {
3884         return page->objects;
3885 }
3886 #endif
3887 
3888 #ifdef CONFIG_SLUB_DEBUG
3889 static int validate_slab(struct kmem_cache *s, struct page *page,
3890                                                 unsigned long *map)
3891 {
3892         void *p;
3893         void *addr = page_address(page);
3894 
3895         if (!check_slab(s, page) ||
3896                         !on_freelist(s, page, NULL))
3897                 return 0;
3898 
3899         /* Now we know that a valid freelist exists */
3900         bitmap_zero(map, page->objects);
3901 
3902         get_map(s, page, map);
3903         for_each_object(p, s, addr, page->objects) {
3904                 if (test_bit(slab_index(p, s, addr), map))
3905                         if (!check_object(s, page, p, SLUB_RED_INACTIVE))
3906                                 return 0;
3907         }
3908 
3909         for_each_object(p, s, addr, page->objects)
3910                 if (!test_bit(slab_index(p, s, addr), map))
3911                         if (!check_object(s, page, p, SLUB_RED_ACTIVE))
3912                                 return 0;
3913         return 1;
3914 }
3915 
3916 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3917                                                 unsigned long *map)
3918 {
3919         slab_lock(page);
3920         validate_slab(s, page, map);
3921         slab_unlock(page);
3922 }
3923 
3924 static int validate_slab_node(struct kmem_cache *s,
3925                 struct kmem_cache_node *n, unsigned long *map)
3926 {
3927         unsigned long count = 0;
3928         struct page *page;
3929         unsigned long flags;
3930 
3931         spin_lock_irqsave(&n->list_lock, flags);
3932 
3933         list_for_each_entry(page, &n->partial, lru) {
3934                 validate_slab_slab(s, page, map);
3935                 count++;
3936         }
3937         if (count != n->nr_partial)
3938                 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
3939                        s->name, count, n->nr_partial);
3940 
3941         if (!(s->flags & SLAB_STORE_USER))
3942                 goto out;
3943 
3944         list_for_each_entry(page, &n->full, lru) {
3945                 validate_slab_slab(s, page, map);
3946                 count++;
3947         }
3948         if (count != atomic_long_read(&n->nr_slabs))
3949                 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
3950                        s->name, count, atomic_long_read(&n->nr_slabs));
3951 
3952 out:
3953         spin_unlock_irqrestore(&n->list_lock, flags);
3954         return count;
3955 }
3956 
3957 static long validate_slab_cache(struct kmem_cache *s)
3958 {
3959         int node;
3960         unsigned long count = 0;
3961         unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3962                                 sizeof(unsigned long), GFP_KERNEL);
3963 
3964         if (!map)
3965                 return -ENOMEM;
3966 
3967         flush_all(s);
3968         for_each_node_state(node, N_NORMAL_MEMORY) {
3969                 struct kmem_cache_node *n = get_node(s, node);
3970 
3971                 count += validate_slab_node(s, n, map);
3972         }
3973         kfree(map);
3974         return count;
3975 }
3976 /*
3977  * Generate lists of code addresses where slabcache objects are allocated
3978  * and freed.
3979  */
3980 
3981 struct location {
3982         unsigned long count;
3983         unsigned long addr;
3984         long long sum_time;
3985         long min_time;
3986         long max_time;
3987         long min_pid;
3988         long max_pid;
3989         DECLARE_BITMAP(cpus, NR_CPUS);
3990         nodemask_t nodes;
3991 };
3992 
3993 struct loc_track {
3994         unsigned long max;
3995         unsigned long count;
3996         struct location *loc;
3997 };
3998 
3999 static void free_loc_track(struct loc_track *t)
4000 {
4001         if (t->max)
4002                 free_pages((unsigned long)t->loc,
4003                         get_order(sizeof(struct location) * t->max));
4004 }
4005 
4006 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4007 {
4008         struct location *l;
4009         int order;
4010 
4011         order = get_order(sizeof(struct location) * max);
4012 
4013         l = (void *)__get_free_pages(flags, order);
4014         if (!l)
4015                 return 0;
4016 
4017         if (t->count) {
4018                 memcpy(l, t->loc, sizeof(struct location) * t->count);
4019                 free_loc_track(t);
4020         }
4021         t->max = max;
4022         t->loc = l;
4023         return 1;
4024 }
4025 
4026 static int add_location(struct loc_track *t, struct kmem_cache *s,
4027                                 const struct track *track)
4028 {
4029         long start, end, pos;
4030         struct location *l;
4031         unsigned long caddr;
4032         unsigned long age = jiffies - track->when;
4033 
4034         start = -1;
4035         end = t->count;
4036 
4037         for ( ; ; ) {
4038                 pos = start + (end - start + 1) / 2;
4039 
4040                 /*
4041                  * There is nothing at "end". If we end up there
4042                  * we need to add something to before end.
4043                  */
4044                 if (pos == end)
4045                         break;
4046 
4047                 caddr = t->loc[pos].addr;
4048                 if (track->addr == caddr) {
4049 
4050                         l = &t->loc[pos];
4051                         l->count++;
4052                         if (track->when) {
4053                                 l->sum_time += age;
4054                                 if (age < l->min_time)
4055                                         l->min_time = age;
4056                                 if (age > l->max_time)
4057                                         l->max_time = age;
4058 
4059                                 if (track->pid < l->min_pid)
4060                                         l->min_pid = track->pid;
4061                                 if (track->pid > l->max_pid)
4062                                         l->max_pid = track->pid;
4063 
4064                                 cpumask_set_cpu(track->cpu,
4065                                                 to_cpumask(l->cpus));
4066                         }
4067                         node_set(page_to_nid(virt_to_page(track)), l->nodes);
4068                         return 1;
4069                 }
4070 
4071                 if (track->addr < caddr)
4072                         end = pos;
4073                 else
4074                         start = pos;
4075         }
4076 
4077         /*
4078          * Not found. Insert new tracking element.
4079          */
4080         if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4081                 return 0;
4082 
4083         l = t->loc + pos;
4084         if (pos < t->count)
4085                 memmove(l + 1, l,
4086                         (t->count - pos) * sizeof(struct location));
4087         t->count++;
4088         l->count = 1;
4089         l->addr = track->addr;
4090         l->sum_time = age;
4091         l->min_time = age;
4092         l->max_time = age;
4093         l->min_pid = track->pid;
4094         l->max_pid = track->pid;
4095         cpumask_clear(to_cpumask(l->cpus));
4096         cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4097         nodes_clear(l->nodes);
4098         node_set(page_to_nid(virt_to_page(track)), l->nodes);
4099         return 1;
4100 }
4101 
4102 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4103                 struct page *page, enum track_item alloc,
4104                 unsigned long *map)
4105 {
4106         void *addr = page_address(page);
4107         void *p;
4108 
4109         bitmap_zero(map, page->objects);
4110         get_map(s, page, map);
4111 
4112         for_each_object(p, s, addr, page->objects)
4113                 if (!test_bit(slab_index(p, s, addr), map))
4114                         add_location(t, s, get_track(s, p, alloc));
4115 }
4116 
4117 static int list_locations(struct kmem_cache *s, char *buf,
4118                                         enum track_item alloc)
4119 {
4120         int len = 0;
4121         unsigned long i;
4122         struct loc_track t = { 0, 0, NULL };
4123         int node;
4124         unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4125                                      sizeof(unsigned long), GFP_KERNEL);
4126 
4127         if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4128                                      GFP_TEMPORARY)) {
4129                 kfree(map);
4130                 return sprintf(buf, "Out of memory\n");
4131         }
4132         /* Push back cpu slabs */
4133         flush_all(s);
4134 
4135         for_each_node_state(node, N_NORMAL_MEMORY) {
4136                 struct kmem_cache_node *n = get_node(s, node);
4137                 unsigned long flags;
4138                 struct page *page;
4139 
4140                 if (!atomic_long_read(&n->nr_slabs))
4141                         continue;
4142 
4143                 spin_lock_irqsave(&n->list_lock, flags);
4144                 list_for_each_entry(page, &n->partial, lru)
4145                         process_slab(&t, s, page, alloc, map);
4146                 list_for_each_entry(page, &n->full, lru)
4147                         process_slab(&t, s, page, alloc, map);
4148                 spin_unlock_irqrestore(&n->list_lock, flags);
4149         }
4150 
4151         for (i = 0; i < t.count; i++) {
4152                 struct location *l = &t.loc[i];
4153 
4154                 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4155                         break;
4156                 len += sprintf(buf + len, "%7ld ", l->count);
4157 
4158                 if (l->addr)
4159                         len += sprintf(buf + len, "%pS", (void *)l->addr);
4160                 else
4161                         len += sprintf(buf + len, "<not-available>");
4162 
4163                 if (l->sum_time != l->min_time) {
4164                         len += sprintf(buf + len, " age=%ld/%ld/%ld",
4165                                 l->min_time,
4166                                 (long)div_u64(l->sum_time, l->count),
4167                                 l->max_time);
4168                 } else
4169                         len += sprintf(buf + len, " age=%ld",
4170                                 l->min_time);
4171 
4172                 if (l->min_pid != l->max_pid)
4173                         len += sprintf(buf + len, " pid=%ld-%ld",
4174                                 l->min_pid, l->max_pid);
4175                 else
4176                         len += sprintf(buf + len, " pid=%ld",
4177                                 l->min_pid);
4178 
4179                 if (num_online_cpus() > 1 &&
4180                                 !cpumask_empty(to_cpumask(l->cpus)) &&
4181                                 len < PAGE_SIZE - 60) {
4182                         len += sprintf(buf + len, " cpus=");
4183                         len += cpulist_scnprintf(buf + len,
4184                                                  PAGE_SIZE - len - 50,
4185                                                  to_cpumask(l->cpus));
4186                 }
4187 
4188                 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4189                                 len < PAGE_SIZE - 60) {
4190                         len += sprintf(buf + len, " nodes=");
4191                         len += nodelist_scnprintf(buf + len,
4192                                                   PAGE_SIZE - len - 50,
4193                                                   l->nodes);
4194                 }
4195 
4196                 len += sprintf(buf + len, "\n");
4197         }
4198 
4199         free_loc_track(&t);
4200         kfree(map);
4201         if (!t.count)
4202                 len += sprintf(buf, "No data\n");
4203         return len;
4204 }
4205 #endif
4206 
4207 #ifdef SLUB_RESILIENCY_TEST
4208 static void resiliency_test(void)
4209 {
4210         u8 *p;
4211 
4212         BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4213 
4214         pr_err("SLUB resiliency testing\n");
4215         pr_err("-----------------------\n");
4216         pr_err("A. Corruption after allocation\n");
4217 
4218         p = kzalloc(16, GFP_KERNEL);
4219         p[16] = 0x12;
4220         pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4221                p + 16);
4222 
4223         validate_slab_cache(kmalloc_caches[4]);
4224 
4225         /* Hmmm... The next two are dangerous */
4226         p = kzalloc(32, GFP_KERNEL);
4227         p[32 + sizeof(void *)] = 0x34;
4228         pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4229                p);
4230         pr_err("If allocated object is overwritten then not detectable\n\n");
4231 
4232         validate_slab_cache(kmalloc_caches[5]);
4233         p = kzalloc(64, GFP_KERNEL);
4234         p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4235         *p = 0x56;
4236         pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4237                p);
4238         pr_err("If allocated object is overwritten then not detectable\n\n");
4239         validate_slab_cache(kmalloc_caches[6]);
4240 
4241         pr_err("\nB. Corruption after free\n");
4242         p = kzalloc(128, GFP_KERNEL);
4243         kfree(p);
4244         *p = 0x78;
4245         pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4246         validate_slab_cache(kmalloc_caches[7]);
4247 
4248         p = kzalloc(256, GFP_KERNEL);
4249         kfree(p);
4250         p[50] = 0x9a;
4251         pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4252         validate_slab_cache(kmalloc_caches[8]);
4253 
4254         p = kzalloc(512, GFP_KERNEL);
4255         kfree(p);
4256         p[512] = 0xab;
4257         pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4258         validate_slab_cache(kmalloc_caches[9]);
4259 }
4260 #else
4261 #ifdef CONFIG_SYSFS
4262 static void resiliency_test(void) {};
4263 #endif
4264 #endif
4265 
4266 #ifdef CONFIG_SYSFS
4267 enum slab_stat_type {
4268         SL_ALL,                 /* All slabs */
4269         SL_PARTIAL,             /* Only partially allocated slabs */
4270         SL_CPU,                 /* Only slabs used for cpu caches */
4271         SL_OBJECTS,             /* Determine allocated objects not slabs */
4272         SL_TOTAL                /* Determine object capacity not slabs */
4273 };
4274 
4275 #define SO_ALL          (1 << SL_ALL)
4276 #define SO_PARTIAL      (1 << SL_PARTIAL)
4277 #define SO_CPU          (1 << SL_CPU)
4278 #define SO_OBJECTS      (1 << SL_OBJECTS)
4279 #define SO_TOTAL        (1 << SL_TOTAL)
4280 
4281 static ssize_t show_slab_objects(struct kmem_cache *s,
4282                             char *buf, unsigned long flags)
4283 {
4284         unsigned long total = 0;
4285         int node;
4286         int x;
4287         unsigned long *nodes;
4288 
4289         nodes = kzalloc(sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4290         if (!nodes)
4291                 return -ENOMEM;
4292 
4293         if (flags & SO_CPU) {
4294                 int cpu;
4295 
4296                 for_each_possible_cpu(cpu) {
4297                         struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4298                                                                cpu);
4299                         int node;
4300                         struct page *page;
4301 
4302                         page = ACCESS_ONCE(c->page);
4303                         if (!page)
4304                                 continue;
4305 
4306                         node = page_to_nid(page);
4307                         if (flags & SO_TOTAL)
4308                                 x = page->objects;
4309                         else if (flags & SO_OBJECTS)
4310                                 x = page->inuse;
4311                         else
4312                                 x = 1;
4313 
4314                         total += x;
4315                         nodes[node] += x;
4316 
4317                         page = ACCESS_ONCE(c->partial);
4318                         if (page) {
4319                                 node = page_to_nid(page);
4320                                 if (flags & SO_TOTAL)
4321                                         WARN_ON_ONCE(1);
4322                                 else if (flags & SO_OBJECTS)
4323                                         WARN_ON_ONCE(1);
4324                                 else
4325                                         x = page->pages;
4326                                 total += x;
4327                                 nodes[node] += x;
4328                         }
4329                 }
4330         }
4331 
4332         /*
4333          * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
4334          * already held which will conflict with an existing lock order:
4335          *
4336          * mem_hotplug_lock->slab_mutex->kernfs_mutex
4337          *
4338          * We don't really need mem_hotplug_lock (to hold off
4339          * slab_mem_going_offline_callback) here because slab's memory hot
4340          * unplug code doesn't destroy the kmem_cache->node[] data.
4341          */
4342 
4343 #ifdef CONFIG_SLUB_DEBUG
4344         if (flags & SO_ALL) {
4345                 for_each_node_state(node, N_NORMAL_MEMORY) {
4346                         struct kmem_cache_node *n = get_node(s, node);
4347 
4348                         if (flags & SO_TOTAL)
4349                                 x = atomic_long_read(&n->total_objects);
4350                         else if (flags & SO_OBJECTS)
4351                                 x = atomic_long_read(&n->total_objects) -
4352                                         count_partial(n, count_free);
4353                         else
4354                                 x = atomic_long_read(&n->nr_slabs);
4355                         total += x;
4356                         nodes[node] += x;
4357                 }
4358 
4359         } else
4360 #endif
4361         if (flags & SO_PARTIAL) {
4362                 for_each_node_state(node, N_NORMAL_MEMORY) {
4363                         struct kmem_cache_node *n = get_node(s, node);
4364 
4365                         if (flags & SO_TOTAL)
4366                                 x = count_partial(n, count_total);
4367                         else if (flags & SO_OBJECTS)
4368                                 x = count_partial(n, count_inuse);
4369                         else
4370                                 x = n->nr_partial;
4371                         total += x;
4372                         nodes[node] += x;
4373                 }
4374         }
4375         x = sprintf(buf, "%lu", total);
4376 #ifdef CONFIG_NUMA
4377         for_each_node_state(node, N_NORMAL_MEMORY)
4378                 if (nodes[node])
4379                         x += sprintf(buf + x, " N%d=%lu",
4380                                         node, nodes[node]);
4381 #endif
4382         kfree(nodes);
4383         return x + sprintf(buf + x, "\n");
4384 }
4385 
4386 #ifdef CONFIG_SLUB_DEBUG
4387 static int any_slab_objects(struct kmem_cache *s)
4388 {
4389         int node;
4390 
4391         for_each_online_node(node) {
4392                 struct kmem_cache_node *n = get_node(s, node);
4393 
4394                 if (!n)
4395                         continue;
4396 
4397                 if (atomic_long_read(&n->total_objects))
4398                         return 1;
4399         }
4400         return 0;
4401 }
4402 #endif
4403 
4404 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4405 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4406 
4407 struct slab_attribute {
4408         struct attribute attr;
4409         ssize_t (*show)(struct kmem_cache *s, char *buf);
4410         ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4411 };
4412 
4413 #define SLAB_ATTR_RO(_name) \
4414         static struct slab_attribute _name##_attr = \
4415         __ATTR(_name, 0400, _name##_show, NULL)
4416 
4417 #define SLAB_ATTR(_name) \
4418         static struct slab_attribute _name##_attr =  \
4419         __ATTR(_name, 0600, _name##_show, _name##_store)
4420 
4421 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4422 {
4423         return sprintf(buf, "%d\n", s->size);
4424 }
4425 SLAB_ATTR_RO(slab_size);
4426 
4427 static ssize_t align_show(struct kmem_cache *s, char *buf)
4428 {
4429         return sprintf(buf, "%d\n", s->align);
4430 }
4431 SLAB_ATTR_RO(align);
4432 
4433 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4434 {
4435         return sprintf(buf, "%d\n", s->object_size);
4436 }
4437 SLAB_ATTR_RO(object_size);
4438 
4439 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4440 {
4441         return sprintf(buf, "%d\n", oo_objects(s->oo));
4442 }
4443 SLAB_ATTR_RO(objs_per_slab);
4444 
4445 static ssize_t order_store(struct kmem_cache *s,
4446                                 const char *buf, size_t length)
4447 {
4448         unsigned long order;
4449         int err;
4450 
4451         err = kstrtoul(buf, 10, &order);
4452         if (err)
4453                 return err;
4454 
4455         if (order > slub_max_order || order < slub_min_order)
4456                 return -EINVAL;
4457 
4458         calculate_sizes(s, order);
4459         return length;
4460 }
4461 
4462 static ssize_t order_show(struct kmem_cache *s, char *buf)
4463 {
4464         return sprintf(buf, "%d\n", oo_order(s->oo));
4465 }
4466 SLAB_ATTR(order);
4467 
4468 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4469 {
4470         return sprintf(buf, "%lu\n", s->min_partial);
4471 }
4472 
4473 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4474                                  size_t length)
4475 {
4476         unsigned long min;
4477         int err;
4478 
4479         err = kstrtoul(buf, 10, &min);
4480         if (err)
4481                 return err;
4482 
4483         set_min_partial(s, min);
4484         return length;
4485 }
4486 SLAB_ATTR(min_partial);
4487 
4488 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4489 {
4490         return sprintf(buf, "%u\n", s->cpu_partial);
4491 }
4492 
4493 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4494                                  size_t length)
4495 {
4496         unsigned long objects;
4497         int err;
4498 
4499         err = kstrtoul(buf, 10, &objects);
4500         if (err)
4501                 return err;
4502         if (objects && !kmem_cache_has_cpu_partial(s))
4503                 return -EINVAL;
4504 
4505         s->cpu_partial = objects;
4506         flush_all(s);
4507         return length;
4508 }
4509 SLAB_ATTR(cpu_partial);
4510 
4511 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4512 {
4513         if (!s->ctor)
4514                 return 0;
4515         return sprintf(buf, "%pS\n", s->ctor);
4516 }
4517 SLAB_ATTR_RO(ctor);
4518 
4519 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4520 {
4521         return sprintf(buf, "%d\n", s->refcount - 1);
4522 }
4523 SLAB_ATTR_RO(aliases);
4524 
4525 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4526 {
4527         return show_slab_objects(s, buf, SO_PARTIAL);
4528 }
4529 SLAB_ATTR_RO(partial);
4530 
4531 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4532 {
4533         return show_slab_objects(s, buf, SO_CPU);
4534 }
4535 SLAB_ATTR_RO(cpu_slabs);
4536 
4537 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4538 {
4539         return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4540 }
4541 SLAB_ATTR_RO(objects);
4542 
4543 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4544 {
4545         return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4546 }
4547 SLAB_ATTR_RO(objects_partial);
4548 
4549 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4550 {
4551         int objects = 0;
4552         int pages = 0;
4553         int cpu;
4554         int len;
4555 
4556         for_each_online_cpu(cpu) {
4557                 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4558 
4559                 if (page) {
4560                         pages += page->pages;
4561                         objects += page->pobjects;
4562                 }
4563         }
4564 
4565         len = sprintf(buf, "%d(%d)", objects, pages);
4566 
4567 #ifdef CONFIG_SMP
4568         for_each_online_cpu(cpu) {
4569                 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4570 
4571                 if (page && len < PAGE_SIZE - 20)
4572                         len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4573                                 page->pobjects, page->pages);
4574         }
4575 #endif
4576         return len + sprintf(buf + len, "\n");
4577 }
4578 SLAB_ATTR_RO(slabs_cpu_partial);
4579 
4580 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4581 {
4582         return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4583 }
4584 
4585 static ssize_t reclaim_account_store(struct kmem_cache *s,
4586                                 const char *buf, size_t length)
4587 {
4588         s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4589         if (buf[0] == '1')
4590                 s->flags |= SLAB_RECLAIM_ACCOUNT;
4591         return length;
4592 }
4593 SLAB_ATTR(reclaim_account);
4594 
4595 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4596 {
4597         return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4598 }
4599 SLAB_ATTR_RO(hwcache_align);
4600 
4601 #ifdef CONFIG_ZONE_DMA
4602 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4603 {
4604         return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4605 }
4606 SLAB_ATTR_RO(cache_dma);
4607 #endif
4608 
4609 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4610 {
4611         return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4612 }
4613 SLAB_ATTR_RO(destroy_by_rcu);
4614 
4615 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4616 {
4617         return sprintf(buf, "%d\n", s->reserved);
4618 }
4619 SLAB_ATTR_RO(reserved);
4620 
4621 #ifdef CONFIG_SLUB_DEBUG
4622 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4623 {
4624         return show_slab_objects(s, buf, SO_ALL);
4625 }
4626 SLAB_ATTR_RO(slabs);
4627 
4628 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4629 {
4630         return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4631 }
4632 SLAB_ATTR_RO(total_objects);
4633 
4634 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4635 {
4636         return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4637 }
4638 
4639 static ssize_t sanity_checks_store(struct kmem_cache *s,
4640                                 const char *buf, size_t length)
4641 {
4642         s->flags &= ~SLAB_DEBUG_FREE;
4643         if (buf[0] == '1') {
4644                 s->flags &= ~__CMPXCHG_DOUBLE;
4645                 s->flags |= SLAB_DEBUG_FREE;
4646         }
4647         return length;
4648 }
4649 SLAB_ATTR(sanity_checks);
4650 
4651 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4652 {
4653         return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4654 }
4655 
4656 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4657                                                         size_t length)
4658 {
4659         s->flags &= ~SLAB_TRACE;
4660         if (buf[0] == '1') {
4661                 s->flags &= ~__CMPXCHG_DOUBLE;
4662                 s->flags |= SLAB_TRACE;
4663         }
4664         return length;
4665 }
4666 SLAB_ATTR(trace);
4667 
4668 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4669 {
4670         return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4671 }
4672 
4673 static ssize_t red_zone_store(struct kmem_cache *s,
4674                                 const char *buf, size_t length)
4675 {
4676         if (any_slab_objects(s))
4677                 return -EBUSY;
4678 
4679         s->flags &= ~SLAB_RED_ZONE;
4680         if (buf[0] == '1') {
4681                 s->flags &= ~__CMPXCHG_DOUBLE;
4682                 s->flags |= SLAB_RED_ZONE;
4683         }
4684         calculate_sizes(s, -1);
4685         return length;
4686 }
4687 SLAB_ATTR(red_zone);
4688 
4689 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4690 {
4691         return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4692 }
4693 
4694 static ssize_t poison_store(struct kmem_cache *s,
4695                                 const char *buf, size_t length)
4696 {
4697         if (any_slab_objects(s))
4698                 return -EBUSY;
4699 
4700         s->flags &= ~SLAB_POISON;
4701         if (buf[0] == '1') {
4702                 s->flags &= ~__CMPXCHG_DOUBLE;
4703                 s->flags |= SLAB_POISON;
4704         }
4705         calculate_sizes(s, -1);
4706         return length;
4707 }
4708 SLAB_ATTR(poison);
4709 
4710 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4711 {
4712         return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4713 }
4714 
4715 static ssize_t store_user_store(struct kmem_cache *s,
4716                                 const char *buf, size_t length)
4717 {
4718         if (any_slab_objects(s))
4719                 return -EBUSY;
4720 
4721         s->flags &= ~SLAB_STORE_USER;
4722         if (buf[0] == '1') {
4723                 s->flags &= ~__CMPXCHG_DOUBLE;
4724                 s->flags |= SLAB_STORE_USER;
4725         }
4726         calculate_sizes(s, -1);
4727         return length;
4728 }
4729 SLAB_ATTR(store_user);
4730 
4731 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4732 {
4733         return 0;
4734 }
4735 
4736 static ssize_t validate_store(struct kmem_cache *s,
4737                         const char *buf, size_t length)
4738 {
4739         int ret = -EINVAL;
4740 
4741         if (buf[0] == '1') {
4742                 ret = validate_slab_cache(s);
4743                 if (ret >= 0)
4744                         ret = length;
4745         }
4746         return ret;
4747 }
4748 SLAB_ATTR(validate);
4749 
4750 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4751 {
4752         if (!(s->flags & SLAB_STORE_USER))
4753                 return -ENOSYS;
4754         return list_locations(s, buf, TRACK_ALLOC);
4755 }
4756 SLAB_ATTR_RO(alloc_calls);
4757 
4758 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4759 {
4760         if (!(s->flags & SLAB_STORE_USER))
4761                 return -ENOSYS;
4762         return list_locations(s, buf, TRACK_FREE);
4763 }
4764 SLAB_ATTR_RO(free_calls);
4765 #endif /* CONFIG_SLUB_DEBUG */
4766 
4767 #ifdef CONFIG_FAILSLAB
4768 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4769 {
4770         return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4771 }
4772 
4773 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4774                                                         size_t length)
4775 {
4776         s->flags &= ~SLAB_FAILSLAB;
4777         if (buf[0] == '1')
4778                 s->flags |= SLAB_FAILSLAB;
4779         return length;
4780 }
4781 SLAB_ATTR(failslab);
4782 #endif
4783 
4784 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4785 {
4786         return 0;
4787 }
4788 
4789 static ssize_t shrink_store(struct kmem_cache *s,
4790                         const char *buf, size_t length)
4791 {
4792         if (buf[0] == '1') {
4793                 int rc = kmem_cache_shrink(s);
4794 
4795                 if (rc)
4796                         return rc;
4797         } else
4798                 return -EINVAL;
4799         return length;
4800 }
4801 SLAB_ATTR(shrink);
4802 
4803 #ifdef CONFIG_NUMA
4804 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4805 {
4806         return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4807 }
4808 
4809 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4810                                 const char *buf, size_t length)
4811 {
4812         unsigned long ratio;
4813         int err;
4814 
4815         err = kstrtoul(buf, 10, &ratio);
4816         if (err)
4817                 return err;
4818 
4819         if (ratio <= 100)
4820                 s->remote_node_defrag_ratio = ratio * 10;
4821 
4822         return length;
4823 }
4824 SLAB_ATTR(remote_node_defrag_ratio);
4825 #endif
4826 
4827 #ifdef CONFIG_SLUB_STATS
4828 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4829 {
4830         unsigned long sum  = 0;
4831         int cpu;
4832         int len;
4833         int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4834 
4835         if (!data)
4836                 return -ENOMEM;
4837 
4838         for_each_online_cpu(cpu) {
4839                 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4840 
4841                 data[cpu] = x;
4842                 sum += x;
4843         }
4844 
4845         len = sprintf(buf, "%lu", sum);
4846 
4847 #ifdef CONFIG_SMP
4848         for_each_online_cpu(cpu) {
4849                 if (data[cpu] && len < PAGE_SIZE - 20)
4850                         len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4851         }
4852 #endif
4853         kfree(data);
4854         return len + sprintf(buf + len, "\n");
4855 }
4856 
4857 static void clear_stat(struct kmem_cache *s, enum stat_item si)
4858 {
4859         int cpu;
4860 
4861         for_each_online_cpu(cpu)
4862                 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4863 }
4864 
4865 #define STAT_ATTR(si, text)                                     \
4866 static ssize_t text##_show(struct kmem_cache *s, char *buf)     \
4867 {                                                               \
4868         return show_stat(s, buf, si);                           \
4869 }                                                               \
4870 static ssize_t text##_store(struct kmem_cache *s,               \
4871                                 const char *buf, size_t length) \
4872 {                                                               \
4873         if (buf[0] != '')                                      \
4874                 return -EINVAL;                                 \
4875         clear_stat(s, si);                                      \
4876         return length;                                          \
4877 }                                                               \
4878 SLAB_ATTR(text);                                                \
4879 
4880 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4881 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4882 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4883 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4884 STAT_ATTR(FREE_FROZEN, free_frozen);
4885 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4886 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4887 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4888 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4889 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4890 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
4891 STAT_ATTR(FREE_SLAB, free_slab);
4892 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4893 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4894 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4895 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4896 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4897 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4898 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
4899 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4900 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
4901 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
4902 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
4903 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
4904 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
4905 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
4906 #endif
4907 
4908 static struct attribute *slab_attrs[] = {
4909         &slab_size_attr.attr,
4910         &object_size_attr.attr,
4911         &objs_per_slab_attr.attr,
4912         &order_attr.attr,
4913         &min_partial_attr.attr,
4914         &cpu_partial_attr.attr,
4915         &objects_attr.attr,
4916         &objects_partial_attr.attr,
4917         &partial_attr.attr,
4918         &cpu_slabs_attr.attr,
4919         &ctor_attr.attr,
4920         &aliases_attr.attr,
4921         &align_attr.attr,
4922         &hwcache_align_attr.attr,
4923         &reclaim_account_attr.attr,
4924         &destroy_by_rcu_attr.attr,
4925         &shrink_attr.attr,
4926         &reserved_attr.attr,
4927         &slabs_cpu_partial_attr.attr,
4928 #ifdef CONFIG_SLUB_DEBUG
4929         &total_objects_attr.attr,
4930         &slabs_attr.attr,
4931         &sanity_checks_attr.attr,
4932         &trace_attr.attr,
4933         &red_zone_attr.attr,
4934         &poison_attr.attr,
4935         &store_user_attr.attr,
4936         &validate_attr.attr,
4937         &alloc_calls_attr.attr,
4938         &free_calls_attr.attr,
4939 #endif
4940 #ifdef CONFIG_ZONE_DMA
4941         &cache_dma_attr.attr,
4942 #endif
4943 #ifdef CONFIG_NUMA
4944         &remote_node_defrag_ratio_attr.attr,
4945 #endif
4946 #ifdef CONFIG_SLUB_STATS
4947         &alloc_fastpath_attr.attr,
4948         &alloc_slowpath_attr.attr,
4949         &free_fastpath_attr.attr,
4950         &free_slowpath_attr.attr,
4951         &free_frozen_attr.attr,
4952         &free_add_partial_attr.attr,
4953         &free_remove_partial_attr.attr,
4954         &alloc_from_partial_attr.attr,
4955         &alloc_slab_attr.attr,
4956         &alloc_refill_attr.attr,
4957         &alloc_node_mismatch_attr.attr,
4958         &free_slab_attr.attr,
4959         &cpuslab_flush_attr.attr,
4960         &deactivate_full_attr.attr,
4961         &deactivate_empty_attr.attr,
4962         &deactivate_to_head_attr.attr,
4963         &deactivate_to_tail_attr.attr,
4964         &deactivate_remote_frees_attr.attr,
4965         &deactivate_bypass_attr.attr,
4966         &order_fallback_attr.attr,
4967         &cmpxchg_double_fail_attr.attr,
4968         &cmpxchg_double_cpu_fail_attr.attr,
4969         &cpu_partial_alloc_attr.attr,
4970         &cpu_partial_free_attr.attr,
4971         &cpu_partial_node_attr.attr,
4972         &cpu_partial_drain_attr.attr,
4973 #endif
4974 #ifdef CONFIG_FAILSLAB
4975         &failslab_attr.attr,
4976 #endif
4977 
4978         NULL
4979 };
4980 
4981 static struct attribute_group slab_attr_group = {
4982         .attrs = slab_attrs,
4983 };
4984 
4985 static ssize_t slab_attr_show(struct kobject *kobj,
4986                                 struct attribute *attr,
4987                                 char *buf)
4988 {
4989         struct slab_attribute *attribute;
4990         struct kmem_cache *s;
4991         int err;
4992 
4993         attribute = to_slab_attr(attr);
4994         s = to_slab(kobj);
4995 
4996         if (!attribute->show)
4997                 return -EIO;
4998 
4999         err = attribute->show(s, buf);
5000 
5001         return err;
5002 }
5003 
5004 static ssize_t slab_attr_store(struct kobject *kobj,
5005                                 struct attribute *attr,
5006                                 const char *buf, size_t len)
5007 {
5008         struct slab_attribute *attribute;
5009         struct kmem_cache *s;
5010         int err;
5011 
5012         attribute = to_slab_attr(attr);
5013         s = to_slab(kobj);
5014 
5015         if (!attribute->store)
5016                 return -EIO;
5017 
5018         err = attribute->store(s, buf, len);
5019 #ifdef CONFIG_MEMCG_KMEM
5020         if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5021                 int i;
5022 
5023                 mutex_lock(&slab_mutex);
5024                 if (s->max_attr_size < len)
5025                         s->max_attr_size = len;
5026 
5027                 /*
5028                  * This is a best effort propagation, so this function's return
5029                  * value will be determined by the parent cache only. This is
5030                  * basically because not all attributes will have a well
5031                  * defined semantics for rollbacks - most of the actions will
5032                  * have permanent effects.
5033                  *
5034                  * Returning the error value of any of the children that fail
5035                  * is not 100 % defined, in the sense that users seeing the
5036                  * error code won't be able to know anything about the state of
5037                  * the cache.
5038                  *
5039                  * Only returning the error code for the parent cache at least
5040                  * has well defined semantics. The cache being written to
5041                  * directly either failed or succeeded, in which case we loop
5042                  * through the descendants with best-effort propagation.
5043                  */
5044                 for_each_memcg_cache_index(i) {
5045                         struct kmem_cache *c = cache_from_memcg_idx(s, i);
5046                         if (c)
5047                                 attribute->store(c, buf, len);
5048                 }
5049                 mutex_unlock(&slab_mutex);
5050         }
5051 #endif
5052         return err;
5053 }
5054 
5055 static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5056 {
5057 #ifdef CONFIG_MEMCG_KMEM
5058         int i;
5059         char *buffer = NULL;
5060         struct kmem_cache *root_cache;
5061 
5062         if (is_root_cache(s))
5063                 return;
5064 
5065         root_cache = s->memcg_params->root_cache;
5066 
5067         /*
5068          * This mean this cache had no attribute written. Therefore, no point
5069          * in copying default values around
5070          */
5071         if (!root_cache->max_attr_size)
5072                 return;
5073 
5074         for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5075                 char mbuf[64];
5076                 char *buf;
5077                 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5078                 ssize_t len;
5079 
5080                 if (!attr || !attr->store || !attr->show)
5081                         continue;
5082 
5083                 /*
5084                  * It is really bad that we have to allocate here, so we will
5085                  * do it only as a fallback. If we actually allocate, though,
5086                  * we can just use the allocated buffer until the end.
5087                  *
5088                  * Most of the slub attributes will tend to be very small in
5089                  * size, but sysfs allows buffers up to a page, so they can
5090                  * theoretically happen.
5091                  */
5092                 if (buffer)
5093                         buf = buffer;
5094                 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf))
5095                         buf = mbuf;
5096                 else {
5097                         buffer = (char *) get_zeroed_page(GFP_KERNEL);
5098                         if (WARN_ON(!buffer))
5099                                 continue;
5100                         buf = buffer;
5101                 }
5102 
5103                 len = attr->show(root_cache, buf);
5104                 if (len > 0)
5105                         attr->store(s, buf, len);
5106         }
5107 
5108         if (buffer)
5109                 free_page((unsigned long)buffer);
5110 #endif
5111 }
5112 
5113 static void kmem_cache_release(struct kobject *k)
5114 {
5115         slab_kmem_cache_release(to_slab(k));
5116 }
5117 
5118 static const struct sysfs_ops slab_sysfs_ops = {
5119         .show = slab_attr_show,
5120         .store = slab_attr_store,
5121 };
5122 
5123 static struct kobj_type slab_ktype = {
5124         .sysfs_ops = &slab_sysfs_ops,
5125         .release = kmem_cache_release,
5126 };
5127 
5128 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5129 {
5130         struct kobj_type *ktype = get_ktype(kobj);
5131 
5132         if (ktype == &slab_ktype)
5133                 return 1;
5134         return 0;
5135 }
5136 
5137 static const struct kset_uevent_ops slab_uevent_ops = {
5138         .filter = uevent_filter,
5139 };
5140 
5141 static struct kset *slab_kset;
5142 
5143 static inline struct kset *cache_kset(struct kmem_cache *s)
5144 {
5145 #ifdef CONFIG_MEMCG_KMEM
5146         if (!is_root_cache(s))
5147                 return s->memcg_params->root_cache->memcg_kset;
5148 #endif
5149         return slab_kset;
5150 }
5151 
5152 #define ID_STR_LENGTH 64
5153 
5154 /* Create a unique string id for a slab cache:
5155  *
5156  * Format       :[flags-]size
5157  */
5158 static char *create_unique_id(struct kmem_cache *s)
5159 {
5160         char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5161         char *p = name;
5162 
5163         BUG_ON(!name);
5164 
5165         *p++ = ':';
5166         /*
5167          * First flags affecting slabcache operations. We will only
5168          * get here for aliasable slabs so we do not need to support
5169          * too many flags. The flags here must cover all flags that
5170          * are matched during merging to guarantee that the id is
5171          * unique.
5172          */
5173         if (s->flags & SLAB_CACHE_DMA)
5174                 *p++ = 'd';
5175         if (s->flags & SLAB_RECLAIM_ACCOUNT)
5176                 *p++ = 'a';
5177         if (s->flags & SLAB_DEBUG_FREE)
5178                 *p++ = 'F';
5179         if (!(s->flags & SLAB_NOTRACK))
5180                 *p++ = 't';
5181         if (p != name + 1)
5182                 *p++ = '-';
5183         p += sprintf(p, "%07d", s->size);
5184 
5185 #ifdef CONFIG_MEMCG_KMEM
5186         if (!is_root_cache(s))
5187                 p += sprintf(p, "-%08d",
5188                                 memcg_cache_id(s->memcg_params->memcg));
5189 #endif
5190 
5191         BUG_ON(p > name + ID_STR_LENGTH - 1);
5192         return name;
5193 }
5194 
5195 static int sysfs_slab_add(struct kmem_cache *s)
5196 {
5197         int err;
5198         const char *name;
5199         int unmergeable = slab_unmergeable(s);
5200 
5201         if (unmergeable) {
5202                 /*
5203                  * Slabcache can never be merged so we can use the name proper.
5204                  * This is typically the case for debug situations. In that
5205                  * case we can catch duplicate names easily.
5206                  */
5207                 sysfs_remove_link(&slab_kset->kobj, s->name);
5208                 name = s->name;
5209         } else {
5210                 /*
5211                  * Create a unique name for the slab as a target
5212                  * for the symlinks.
5213                  */
5214                 name = create_unique_id(s);
5215         }
5216 
5217         s->kobj.kset = cache_kset(s);
5218         err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5219         if (err)
5220                 goto out_put_kobj;
5221 
5222         err = sysfs_create_group(&s->kobj, &slab_attr_group);
5223         if (err)
5224                 goto out_del_kobj;
5225 
5226 #ifdef CONFIG_MEMCG_KMEM
5227         if (is_root_cache(s)) {
5228                 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5229                 if (!s->memcg_kset) {
5230                         err = -ENOMEM;
5231                         goto out_del_kobj;
5232                 }
5233         }
5234 #endif
5235 
5236         kobject_uevent(&s->kobj, KOBJ_ADD);
5237         if (!unmergeable) {
5238                 /* Setup first alias */
5239                 sysfs_slab_alias(s, s->name);
5240         }
5241 out:
5242         if (!unmergeable)
5243                 kfree(name);
5244         return err;
5245 out_del_kobj:
5246         kobject_del(&s->kobj);
5247 out_put_kobj:
5248         kobject_put(&s->kobj);
5249         goto out;
5250 }
5251 
5252 void sysfs_slab_remove(struct kmem_cache *s)
5253 {
5254         if (slab_state < FULL)
5255                 /*
5256                  * Sysfs has not been setup yet so no need to remove the
5257                  * cache from sysfs.
5258                  */
5259                 return;
5260 
5261 #ifdef CONFIG_MEMCG_KMEM
5262         kset_unregister(s->memcg_kset);
5263 #endif
5264         kobject_uevent(&s->kobj, KOBJ_REMOVE);
5265         kobject_del(&s->kobj);
5266         kobject_put(&s->kobj);
5267 }
5268 
5269 /*
5270  * Need to buffer aliases during bootup until sysfs becomes
5271  * available lest we lose that information.
5272  */
5273 struct saved_alias {
5274         struct kmem_cache *s;
5275         const char *name;
5276         struct saved_alias *next;
5277 };
5278 
5279 static struct saved_alias *alias_list;
5280 
5281 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5282 {
5283         struct saved_alias *al;
5284 
5285         if (slab_state == FULL) {
5286                 /*
5287                  * If we have a leftover link then remove it.
5288                  */
5289                 sysfs_remove_link(&slab_kset->kobj, name);
5290                 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5291         }
5292 
5293         al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5294         if (!al)
5295                 return -ENOMEM;
5296 
5297         al->s = s;
5298         al->name = name;
5299         al->next = alias_list;
5300         alias_list = al;
5301         return 0;
5302 }
5303 
5304 static int __init slab_sysfs_init(void)
5305 {
5306         struct kmem_cache *s;
5307         int err;
5308 
5309         mutex_lock(&slab_mutex);
5310 
5311         slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5312         if (!slab_kset) {
5313                 mutex_unlock(&slab_mutex);
5314                 pr_err("Cannot register slab subsystem.\n");
5315                 return -ENOSYS;
5316         }
5317 
5318         slab_state = FULL;
5319 
5320         list_for_each_entry(s, &slab_caches, list) {
5321                 err = sysfs_slab_add(s);
5322                 if (err)
5323                         pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5324                                s->name);
5325         }
5326 
5327         while (alias_list) {
5328                 struct saved_alias *al = alias_list;
5329 
5330                 alias_list = alias_list->next;
5331                 err = sysfs_slab_alias(al->s, al->name);
5332                 if (err)
5333                         pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5334                                al->name);
5335                 kfree(al);
5336         }
5337 
5338         mutex_unlock(&slab_mutex);
5339         resiliency_test();
5340         return 0;
5341 }
5342 
5343 __initcall(slab_sysfs_init);
5344 #endif /* CONFIG_SYSFS */
5345 
5346 /*
5347  * The /proc/slabinfo ABI
5348  */
5349 #ifdef CONFIG_SLABINFO
5350 void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5351 {
5352         unsigned long nr_slabs = 0;
5353         unsigned long nr_objs = 0;
5354         unsigned long nr_free = 0;
5355         int node;
5356 
5357         for_each_online_node(node) {
5358                 struct kmem_cache_node *n = get_node(s, node);
5359 
5360                 if (!n)
5361                         continue;
5362 
5363                 nr_slabs += node_nr_slabs(n);
5364                 nr_objs += node_nr_objs(n);
5365                 nr_free += count_partial(n, count_free);
5366         }
5367 
5368         sinfo->active_objs = nr_objs - nr_free;
5369         sinfo->num_objs = nr_objs;
5370         sinfo->active_slabs = nr_slabs;
5371         sinfo->num_slabs = nr_slabs;
5372         sinfo->objects_per_slab = oo_objects(s->oo);
5373         sinfo->cache_order = oo_order(s->oo);
5374 }
5375 
5376 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5377 {
5378 }
5379 
5380 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5381                        size_t count, loff_t *ppos)
5382 {
5383         return -EIO;
5384 }
5385 #endif /* CONFIG_SLABINFO */
5386 

~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

kernel.org | git.kernel.org | LWN.net | Project Home | Wiki (Japanese) | Wiki (English) | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

osdn.jp