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

TOMOYO Linux Cross Reference
Linux/mm/slub.c

Version: ~ [ linux-5.6 ] ~ [ linux-5.5.13 ] ~ [ linux-5.4.28 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.113 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.174 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.217 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.217 ] ~ [ 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-3.16.82 ] ~ [ 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-3.9.11 ] ~ [ linux-3.8.13 ] ~ [ linux-3.7.10 ] ~ [ linux-3.6.11 ] ~ [ linux-3.5.7 ] ~ [ linux-3.4.113 ] ~ [ linux-3.3.8 ] ~ [ linux-3.2.102 ] ~ [ linux-3.1.10 ] ~ [ linux-3.0.101 ] ~ [ 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 <linux/proc_fs.h>
 20 #include <linux/seq_file.h>
 21 #include <linux/kmemcheck.h>
 22 #include <linux/cpu.h>
 23 #include <linux/cpuset.h>
 24 #include <linux/mempolicy.h>
 25 #include <linux/ctype.h>
 26 #include <linux/debugobjects.h>
 27 #include <linux/kallsyms.h>
 28 #include <linux/memory.h>
 29 #include <linux/math64.h>
 30 #include <linux/fault-inject.h>
 31 #include <linux/stacktrace.h>
 32 #include <linux/prefetch.h>
 33 
 34 #include <trace/events/kmem.h>
 35 
 36 /*
 37  * Lock order:
 38  *   1. slub_lock (Global Semaphore)
 39  *   2. node->list_lock
 40  *   3. slab_lock(page) (Only on some arches and for debugging)
 41  *
 42  *   slub_lock
 43  *
 44  *   The role of the slub_lock is to protect the list of all the slabs
 45  *   and to synchronize major metadata changes to slab cache structures.
 46  *
 47  *   The slab_lock is only used for debugging and on arches that do not
 48  *   have the ability to do a cmpxchg_double. It only protects the second
 49  *   double word in the page struct. Meaning
 50  *      A. page->freelist       -> List of object free in a page
 51  *      B. page->counters       -> Counters of objects
 52  *      C. page->frozen         -> frozen state
 53  *
 54  *   If a slab is frozen then it is exempt from list management. It is not
 55  *   on any list. The processor that froze the slab is the one who can
 56  *   perform list operations on the page. Other processors may put objects
 57  *   onto the freelist but the processor that froze the slab is the only
 58  *   one that can retrieve the objects from the page's freelist.
 59  *
 60  *   The list_lock protects the partial and full list on each node and
 61  *   the partial slab counter. If taken then no new slabs may be added or
 62  *   removed from the lists nor make the number of partial slabs be modified.
 63  *   (Note that the total number of slabs is an atomic value that may be
 64  *   modified without taking the list lock).
 65  *
 66  *   The list_lock is a centralized lock and thus we avoid taking it as
 67  *   much as possible. As long as SLUB does not have to handle partial
 68  *   slabs, operations can continue without any centralized lock. F.e.
 69  *   allocating a long series of objects that fill up slabs does not require
 70  *   the list lock.
 71  *   Interrupts are disabled during allocation and deallocation in order to
 72  *   make the slab allocator safe to use in the context of an irq. In addition
 73  *   interrupts are disabled to ensure that the processor does not change
 74  *   while handling per_cpu slabs, due to kernel preemption.
 75  *
 76  * SLUB assigns one slab for allocation to each processor.
 77  * Allocations only occur from these slabs called cpu slabs.
 78  *
 79  * Slabs with free elements are kept on a partial list and during regular
 80  * operations no list for full slabs is used. If an object in a full slab is
 81  * freed then the slab will show up again on the partial lists.
 82  * We track full slabs for debugging purposes though because otherwise we
 83  * cannot scan all objects.
 84  *
 85  * Slabs are freed when they become empty. Teardown and setup is
 86  * minimal so we rely on the page allocators per cpu caches for
 87  * fast frees and allocs.
 88  *
 89  * Overloading of page flags that are otherwise used for LRU management.
 90  *
 91  * PageActive           The slab is frozen and exempt from list processing.
 92  *                      This means that the slab is dedicated to a purpose
 93  *                      such as satisfying allocations for a specific
 94  *                      processor. Objects may be freed in the slab while
 95  *                      it is frozen but slab_free will then skip the usual
 96  *                      list operations. It is up to the processor holding
 97  *                      the slab to integrate the slab into the slab lists
 98  *                      when the slab is no longer needed.
 99  *
100  *                      One use of this flag is to mark slabs that are
101  *                      used for allocations. Then such a slab becomes a cpu
102  *                      slab. The cpu slab may be equipped with an additional
103  *                      freelist that allows lockless access to
104  *                      free objects in addition to the regular freelist
105  *                      that requires the slab lock.
106  *
107  * PageError            Slab requires special handling due to debug
108  *                      options set. This moves slab handling out of
109  *                      the fast path and disables lockless freelists.
110  */
111 
112 #define SLAB_DEBUG_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
113                 SLAB_TRACE | SLAB_DEBUG_FREE)
114 
115 static inline int kmem_cache_debug(struct kmem_cache *s)
116 {
117 #ifdef CONFIG_SLUB_DEBUG
118         return unlikely(s->flags & SLAB_DEBUG_FLAGS);
119 #else
120         return 0;
121 #endif
122 }
123 
124 /*
125  * Issues still to be resolved:
126  *
127  * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
128  *
129  * - Variable sizing of the per node arrays
130  */
131 
132 /* Enable to test recovery from slab corruption on boot */
133 #undef SLUB_RESILIENCY_TEST
134 
135 /* Enable to log cmpxchg failures */
136 #undef SLUB_DEBUG_CMPXCHG
137 
138 /*
139  * Mininum number of partial slabs. These will be left on the partial
140  * lists even if they are empty. kmem_cache_shrink may reclaim them.
141  */
142 #define MIN_PARTIAL 5
143 
144 /*
145  * Maximum number of desirable partial slabs.
146  * The existence of more partial slabs makes kmem_cache_shrink
147  * sort the partial list by the number of objects in the.
148  */
149 #define MAX_PARTIAL 10
150 
151 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
152                                 SLAB_POISON | SLAB_STORE_USER)
153 
154 /*
155  * Debugging flags that require metadata to be stored in the slab.  These get
156  * disabled when slub_debug=O is used and a cache's min order increases with
157  * metadata.
158  */
159 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
160 
161 /*
162  * Set of flags that will prevent slab merging
163  */
164 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
165                 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
166                 SLAB_FAILSLAB)
167 
168 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
169                 SLAB_CACHE_DMA | SLAB_NOTRACK)
170 
171 #define OO_SHIFT        16
172 #define OO_MASK         ((1 << OO_SHIFT) - 1)
173 #define MAX_OBJS_PER_PAGE       32767 /* since page.objects is u15 */
174 
175 /* Internal SLUB flags */
176 #define __OBJECT_POISON         0x80000000UL /* Poison object */
177 #define __CMPXCHG_DOUBLE        0x40000000UL /* Use cmpxchg_double */
178 
179 static int kmem_size = sizeof(struct kmem_cache);
180 
181 #ifdef CONFIG_SMP
182 static struct notifier_block slab_notifier;
183 #endif
184 
185 static enum {
186         DOWN,           /* No slab functionality available */
187         PARTIAL,        /* Kmem_cache_node works */
188         UP,             /* Everything works but does not show up in sysfs */
189         SYSFS           /* Sysfs up */
190 } slab_state = DOWN;
191 
192 /* A list of all slab caches on the system */
193 static DECLARE_RWSEM(slub_lock);
194 static LIST_HEAD(slab_caches);
195 
196 /*
197  * Tracking user of a slab.
198  */
199 #define TRACK_ADDRS_COUNT 16
200 struct track {
201         unsigned long addr;     /* Called from address */
202 #ifdef CONFIG_STACKTRACE
203         unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
204 #endif
205         int cpu;                /* Was running on cpu */
206         int pid;                /* Pid context */
207         unsigned long when;     /* When did the operation occur */
208 };
209 
210 enum track_item { TRACK_ALLOC, TRACK_FREE };
211 
212 #ifdef CONFIG_SYSFS
213 static int sysfs_slab_add(struct kmem_cache *);
214 static int sysfs_slab_alias(struct kmem_cache *, const char *);
215 static void sysfs_slab_remove(struct kmem_cache *);
216 
217 #else
218 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
219 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
220                                                         { return 0; }
221 static inline void sysfs_slab_remove(struct kmem_cache *s)
222 {
223         kfree(s->name);
224         kfree(s);
225 }
226 
227 #endif
228 
229 static inline void stat(const struct kmem_cache *s, enum stat_item si)
230 {
231 #ifdef CONFIG_SLUB_STATS
232         __this_cpu_inc(s->cpu_slab->stat[si]);
233 #endif
234 }
235 
236 /********************************************************************
237  *                      Core slab cache functions
238  *******************************************************************/
239 
240 int slab_is_available(void)
241 {
242         return slab_state >= UP;
243 }
244 
245 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
246 {
247         return s->node[node];
248 }
249 
250 /* Verify that a pointer has an address that is valid within a slab page */
251 static inline int check_valid_pointer(struct kmem_cache *s,
252                                 struct page *page, const void *object)
253 {
254         void *base;
255 
256         if (!object)
257                 return 1;
258 
259         base = page_address(page);
260         if (object < base || object >= base + page->objects * s->size ||
261                 (object - base) % s->size) {
262                 return 0;
263         }
264 
265         return 1;
266 }
267 
268 static inline void *get_freepointer(struct kmem_cache *s, void *object)
269 {
270         return *(void **)(object + s->offset);
271 }
272 
273 static void prefetch_freepointer(const struct kmem_cache *s, void *object)
274 {
275         prefetch(object + s->offset);
276 }
277 
278 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
279 {
280         void *p;
281 
282 #ifdef CONFIG_DEBUG_PAGEALLOC
283         probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
284 #else
285         p = get_freepointer(s, object);
286 #endif
287         return p;
288 }
289 
290 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
291 {
292         *(void **)(object + s->offset) = fp;
293 }
294 
295 /* Loop over all objects in a slab */
296 #define for_each_object(__p, __s, __addr, __objects) \
297         for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
298                         __p += (__s)->size)
299 
300 /* Determine object index from a given position */
301 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
302 {
303         return (p - addr) / s->size;
304 }
305 
306 static inline size_t slab_ksize(const struct kmem_cache *s)
307 {
308 #ifdef CONFIG_SLUB_DEBUG
309         /*
310          * Debugging requires use of the padding between object
311          * and whatever may come after it.
312          */
313         if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
314                 return s->objsize;
315 
316 #endif
317         /*
318          * If we have the need to store the freelist pointer
319          * back there or track user information then we can
320          * only use the space before that information.
321          */
322         if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
323                 return s->inuse;
324         /*
325          * Else we can use all the padding etc for the allocation
326          */
327         return s->size;
328 }
329 
330 static inline int order_objects(int order, unsigned long size, int reserved)
331 {
332         return ((PAGE_SIZE << order) - reserved) / size;
333 }
334 
335 static inline struct kmem_cache_order_objects oo_make(int order,
336                 unsigned long size, int reserved)
337 {
338         struct kmem_cache_order_objects x = {
339                 (order << OO_SHIFT) + order_objects(order, size, reserved)
340         };
341 
342         return x;
343 }
344 
345 static inline int oo_order(struct kmem_cache_order_objects x)
346 {
347         return x.x >> OO_SHIFT;
348 }
349 
350 static inline int oo_objects(struct kmem_cache_order_objects x)
351 {
352         return x.x & OO_MASK;
353 }
354 
355 /*
356  * Per slab locking using the pagelock
357  */
358 static __always_inline void slab_lock(struct page *page)
359 {
360         bit_spin_lock(PG_locked, &page->flags);
361 }
362 
363 static __always_inline void slab_unlock(struct page *page)
364 {
365         __bit_spin_unlock(PG_locked, &page->flags);
366 }
367 
368 /* Interrupts must be disabled (for the fallback code to work right) */
369 static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
370                 void *freelist_old, unsigned long counters_old,
371                 void *freelist_new, unsigned long counters_new,
372                 const char *n)
373 {
374         VM_BUG_ON(!irqs_disabled());
375 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
376     defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
377         if (s->flags & __CMPXCHG_DOUBLE) {
378                 if (cmpxchg_double(&page->freelist, &page->counters,
379                         freelist_old, counters_old,
380                         freelist_new, counters_new))
381                 return 1;
382         } else
383 #endif
384         {
385                 slab_lock(page);
386                 if (page->freelist == freelist_old && page->counters == counters_old) {
387                         page->freelist = freelist_new;
388                         page->counters = counters_new;
389                         slab_unlock(page);
390                         return 1;
391                 }
392                 slab_unlock(page);
393         }
394 
395         cpu_relax();
396         stat(s, CMPXCHG_DOUBLE_FAIL);
397 
398 #ifdef SLUB_DEBUG_CMPXCHG
399         printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
400 #endif
401 
402         return 0;
403 }
404 
405 static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
406                 void *freelist_old, unsigned long counters_old,
407                 void *freelist_new, unsigned long counters_new,
408                 const char *n)
409 {
410 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
411     defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
412         if (s->flags & __CMPXCHG_DOUBLE) {
413                 if (cmpxchg_double(&page->freelist, &page->counters,
414                         freelist_old, counters_old,
415                         freelist_new, counters_new))
416                 return 1;
417         } else
418 #endif
419         {
420                 unsigned long flags;
421 
422                 local_irq_save(flags);
423                 slab_lock(page);
424                 if (page->freelist == freelist_old && page->counters == counters_old) {
425                         page->freelist = freelist_new;
426                         page->counters = counters_new;
427                         slab_unlock(page);
428                         local_irq_restore(flags);
429                         return 1;
430                 }
431                 slab_unlock(page);
432                 local_irq_restore(flags);
433         }
434 
435         cpu_relax();
436         stat(s, CMPXCHG_DOUBLE_FAIL);
437 
438 #ifdef SLUB_DEBUG_CMPXCHG
439         printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
440 #endif
441 
442         return 0;
443 }
444 
445 #ifdef CONFIG_SLUB_DEBUG
446 /*
447  * Determine a map of object in use on a page.
448  *
449  * Node listlock must be held to guarantee that the page does
450  * not vanish from under us.
451  */
452 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
453 {
454         void *p;
455         void *addr = page_address(page);
456 
457         for (p = page->freelist; p; p = get_freepointer(s, p))
458                 set_bit(slab_index(p, s, addr), map);
459 }
460 
461 /*
462  * Debug settings:
463  */
464 #ifdef CONFIG_SLUB_DEBUG_ON
465 static int slub_debug = DEBUG_DEFAULT_FLAGS;
466 #else
467 static int slub_debug;
468 #endif
469 
470 static char *slub_debug_slabs;
471 static int disable_higher_order_debug;
472 
473 /*
474  * Object debugging
475  */
476 static void print_section(char *text, u8 *addr, unsigned int length)
477 {
478         print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
479                         length, 1);
480 }
481 
482 static struct track *get_track(struct kmem_cache *s, void *object,
483         enum track_item alloc)
484 {
485         struct track *p;
486 
487         if (s->offset)
488                 p = object + s->offset + sizeof(void *);
489         else
490                 p = object + s->inuse;
491 
492         return p + alloc;
493 }
494 
495 static void set_track(struct kmem_cache *s, void *object,
496                         enum track_item alloc, unsigned long addr)
497 {
498         struct track *p = get_track(s, object, alloc);
499 
500         if (addr) {
501 #ifdef CONFIG_STACKTRACE
502                 struct stack_trace trace;
503                 int i;
504 
505                 trace.nr_entries = 0;
506                 trace.max_entries = TRACK_ADDRS_COUNT;
507                 trace.entries = p->addrs;
508                 trace.skip = 3;
509                 save_stack_trace(&trace);
510 
511                 /* See rant in lockdep.c */
512                 if (trace.nr_entries != 0 &&
513                     trace.entries[trace.nr_entries - 1] == ULONG_MAX)
514                         trace.nr_entries--;
515 
516                 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
517                         p->addrs[i] = 0;
518 #endif
519                 p->addr = addr;
520                 p->cpu = smp_processor_id();
521                 p->pid = current->pid;
522                 p->when = jiffies;
523         } else
524                 memset(p, 0, sizeof(struct track));
525 }
526 
527 static void init_tracking(struct kmem_cache *s, void *object)
528 {
529         if (!(s->flags & SLAB_STORE_USER))
530                 return;
531 
532         set_track(s, object, TRACK_FREE, 0UL);
533         set_track(s, object, TRACK_ALLOC, 0UL);
534 }
535 
536 static void print_track(const char *s, struct track *t)
537 {
538         if (!t->addr)
539                 return;
540 
541         printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
542                 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
543 #ifdef CONFIG_STACKTRACE
544         {
545                 int i;
546                 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
547                         if (t->addrs[i])
548                                 printk(KERN_ERR "\t%pS\n", (void *)t->addrs[i]);
549                         else
550                                 break;
551         }
552 #endif
553 }
554 
555 static void print_tracking(struct kmem_cache *s, void *object)
556 {
557         if (!(s->flags & SLAB_STORE_USER))
558                 return;
559 
560         print_track("Allocated", get_track(s, object, TRACK_ALLOC));
561         print_track("Freed", get_track(s, object, TRACK_FREE));
562 }
563 
564 static void print_page_info(struct page *page)
565 {
566         printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
567                 page, page->objects, page->inuse, page->freelist, page->flags);
568 
569 }
570 
571 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
572 {
573         va_list args;
574         char buf[100];
575 
576         va_start(args, fmt);
577         vsnprintf(buf, sizeof(buf), fmt, args);
578         va_end(args);
579         printk(KERN_ERR "========================================"
580                         "=====================================\n");
581         printk(KERN_ERR "BUG %s (%s): %s\n", s->name, print_tainted(), buf);
582         printk(KERN_ERR "----------------------------------------"
583                         "-------------------------------------\n\n");
584 }
585 
586 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
587 {
588         va_list args;
589         char buf[100];
590 
591         va_start(args, fmt);
592         vsnprintf(buf, sizeof(buf), fmt, args);
593         va_end(args);
594         printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
595 }
596 
597 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
598 {
599         unsigned int off;       /* Offset of last byte */
600         u8 *addr = page_address(page);
601 
602         print_tracking(s, p);
603 
604         print_page_info(page);
605 
606         printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
607                         p, p - addr, get_freepointer(s, p));
608 
609         if (p > addr + 16)
610                 print_section("Bytes b4 ", p - 16, 16);
611 
612         print_section("Object ", p, min_t(unsigned long, s->objsize,
613                                 PAGE_SIZE));
614         if (s->flags & SLAB_RED_ZONE)
615                 print_section("Redzone ", p + s->objsize,
616                         s->inuse - s->objsize);
617 
618         if (s->offset)
619                 off = s->offset + sizeof(void *);
620         else
621                 off = s->inuse;
622 
623         if (s->flags & SLAB_STORE_USER)
624                 off += 2 * sizeof(struct track);
625 
626         if (off != s->size)
627                 /* Beginning of the filler is the free pointer */
628                 print_section("Padding ", p + off, s->size - off);
629 
630         dump_stack();
631 }
632 
633 static void object_err(struct kmem_cache *s, struct page *page,
634                         u8 *object, char *reason)
635 {
636         slab_bug(s, "%s", reason);
637         print_trailer(s, page, object);
638 }
639 
640 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
641 {
642         va_list args;
643         char buf[100];
644 
645         va_start(args, fmt);
646         vsnprintf(buf, sizeof(buf), fmt, args);
647         va_end(args);
648         slab_bug(s, "%s", buf);
649         print_page_info(page);
650         dump_stack();
651 }
652 
653 static void init_object(struct kmem_cache *s, void *object, u8 val)
654 {
655         u8 *p = object;
656 
657         if (s->flags & __OBJECT_POISON) {
658                 memset(p, POISON_FREE, s->objsize - 1);
659                 p[s->objsize - 1] = POISON_END;
660         }
661 
662         if (s->flags & SLAB_RED_ZONE)
663                 memset(p + s->objsize, val, s->inuse - s->objsize);
664 }
665 
666 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
667                                                 void *from, void *to)
668 {
669         slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
670         memset(from, data, to - from);
671 }
672 
673 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
674                         u8 *object, char *what,
675                         u8 *start, unsigned int value, unsigned int bytes)
676 {
677         u8 *fault;
678         u8 *end;
679 
680         fault = memchr_inv(start, value, bytes);
681         if (!fault)
682                 return 1;
683 
684         end = start + bytes;
685         while (end > fault && end[-1] == value)
686                 end--;
687 
688         slab_bug(s, "%s overwritten", what);
689         printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
690                                         fault, end - 1, fault[0], value);
691         print_trailer(s, page, object);
692 
693         restore_bytes(s, what, value, fault, end);
694         return 0;
695 }
696 
697 /*
698  * Object layout:
699  *
700  * object address
701  *      Bytes of the object to be managed.
702  *      If the freepointer may overlay the object then the free
703  *      pointer is the first word of the object.
704  *
705  *      Poisoning uses 0x6b (POISON_FREE) and the last byte is
706  *      0xa5 (POISON_END)
707  *
708  * object + s->objsize
709  *      Padding to reach word boundary. This is also used for Redzoning.
710  *      Padding is extended by another word if Redzoning is enabled and
711  *      objsize == inuse.
712  *
713  *      We fill with 0xbb (RED_INACTIVE) for inactive objects and with
714  *      0xcc (RED_ACTIVE) for objects in use.
715  *
716  * object + s->inuse
717  *      Meta data starts here.
718  *
719  *      A. Free pointer (if we cannot overwrite object on free)
720  *      B. Tracking data for SLAB_STORE_USER
721  *      C. Padding to reach required alignment boundary or at mininum
722  *              one word if debugging is on to be able to detect writes
723  *              before the word boundary.
724  *
725  *      Padding is done using 0x5a (POISON_INUSE)
726  *
727  * object + s->size
728  *      Nothing is used beyond s->size.
729  *
730  * If slabcaches are merged then the objsize and inuse boundaries are mostly
731  * ignored. And therefore no slab options that rely on these boundaries
732  * may be used with merged slabcaches.
733  */
734 
735 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
736 {
737         unsigned long off = s->inuse;   /* The end of info */
738 
739         if (s->offset)
740                 /* Freepointer is placed after the object. */
741                 off += sizeof(void *);
742 
743         if (s->flags & SLAB_STORE_USER)
744                 /* We also have user information there */
745                 off += 2 * sizeof(struct track);
746 
747         if (s->size == off)
748                 return 1;
749 
750         return check_bytes_and_report(s, page, p, "Object padding",
751                                 p + off, POISON_INUSE, s->size - off);
752 }
753 
754 /* Check the pad bytes at the end of a slab page */
755 static int slab_pad_check(struct kmem_cache *s, struct page *page)
756 {
757         u8 *start;
758         u8 *fault;
759         u8 *end;
760         int length;
761         int remainder;
762 
763         if (!(s->flags & SLAB_POISON))
764                 return 1;
765 
766         start = page_address(page);
767         length = (PAGE_SIZE << compound_order(page)) - s->reserved;
768         end = start + length;
769         remainder = length % s->size;
770         if (!remainder)
771                 return 1;
772 
773         fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
774         if (!fault)
775                 return 1;
776         while (end > fault && end[-1] == POISON_INUSE)
777                 end--;
778 
779         slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
780         print_section("Padding ", end - remainder, remainder);
781 
782         restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
783         return 0;
784 }
785 
786 static int check_object(struct kmem_cache *s, struct page *page,
787                                         void *object, u8 val)
788 {
789         u8 *p = object;
790         u8 *endobject = object + s->objsize;
791 
792         if (s->flags & SLAB_RED_ZONE) {
793                 if (!check_bytes_and_report(s, page, object, "Redzone",
794                         endobject, val, s->inuse - s->objsize))
795                         return 0;
796         } else {
797                 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
798                         check_bytes_and_report(s, page, p, "Alignment padding",
799                                 endobject, POISON_INUSE, s->inuse - s->objsize);
800                 }
801         }
802 
803         if (s->flags & SLAB_POISON) {
804                 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
805                         (!check_bytes_and_report(s, page, p, "Poison", p,
806                                         POISON_FREE, s->objsize - 1) ||
807                          !check_bytes_and_report(s, page, p, "Poison",
808                                 p + s->objsize - 1, POISON_END, 1)))
809                         return 0;
810                 /*
811                  * check_pad_bytes cleans up on its own.
812                  */
813                 check_pad_bytes(s, page, p);
814         }
815 
816         if (!s->offset && val == SLUB_RED_ACTIVE)
817                 /*
818                  * Object and freepointer overlap. Cannot check
819                  * freepointer while object is allocated.
820                  */
821                 return 1;
822 
823         /* Check free pointer validity */
824         if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
825                 object_err(s, page, p, "Freepointer corrupt");
826                 /*
827                  * No choice but to zap it and thus lose the remainder
828                  * of the free objects in this slab. May cause
829                  * another error because the object count is now wrong.
830                  */
831                 set_freepointer(s, p, NULL);
832                 return 0;
833         }
834         return 1;
835 }
836 
837 static int check_slab(struct kmem_cache *s, struct page *page)
838 {
839         int maxobj;
840 
841         VM_BUG_ON(!irqs_disabled());
842 
843         if (!PageSlab(page)) {
844                 slab_err(s, page, "Not a valid slab page");
845                 return 0;
846         }
847 
848         maxobj = order_objects(compound_order(page), s->size, s->reserved);
849         if (page->objects > maxobj) {
850                 slab_err(s, page, "objects %u > max %u",
851                         s->name, page->objects, maxobj);
852                 return 0;
853         }
854         if (page->inuse > page->objects) {
855                 slab_err(s, page, "inuse %u > max %u",
856                         s->name, page->inuse, page->objects);
857                 return 0;
858         }
859         /* Slab_pad_check fixes things up after itself */
860         slab_pad_check(s, page);
861         return 1;
862 }
863 
864 /*
865  * Determine if a certain object on a page is on the freelist. Must hold the
866  * slab lock to guarantee that the chains are in a consistent state.
867  */
868 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
869 {
870         int nr = 0;
871         void *fp;
872         void *object = NULL;
873         unsigned long max_objects;
874 
875         fp = page->freelist;
876         while (fp && nr <= page->objects) {
877                 if (fp == search)
878                         return 1;
879                 if (!check_valid_pointer(s, page, fp)) {
880                         if (object) {
881                                 object_err(s, page, object,
882                                         "Freechain corrupt");
883                                 set_freepointer(s, object, NULL);
884                                 break;
885                         } else {
886                                 slab_err(s, page, "Freepointer corrupt");
887                                 page->freelist = NULL;
888                                 page->inuse = page->objects;
889                                 slab_fix(s, "Freelist cleared");
890                                 return 0;
891                         }
892                         break;
893                 }
894                 object = fp;
895                 fp = get_freepointer(s, object);
896                 nr++;
897         }
898 
899         max_objects = order_objects(compound_order(page), s->size, s->reserved);
900         if (max_objects > MAX_OBJS_PER_PAGE)
901                 max_objects = MAX_OBJS_PER_PAGE;
902 
903         if (page->objects != max_objects) {
904                 slab_err(s, page, "Wrong number of objects. Found %d but "
905                         "should be %d", page->objects, max_objects);
906                 page->objects = max_objects;
907                 slab_fix(s, "Number of objects adjusted.");
908         }
909         if (page->inuse != page->objects - nr) {
910                 slab_err(s, page, "Wrong object count. Counter is %d but "
911                         "counted were %d", page->inuse, page->objects - nr);
912                 page->inuse = page->objects - nr;
913                 slab_fix(s, "Object count adjusted.");
914         }
915         return search == NULL;
916 }
917 
918 static void trace(struct kmem_cache *s, struct page *page, void *object,
919                                                                 int alloc)
920 {
921         if (s->flags & SLAB_TRACE) {
922                 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
923                         s->name,
924                         alloc ? "alloc" : "free",
925                         object, page->inuse,
926                         page->freelist);
927 
928                 if (!alloc)
929                         print_section("Object ", (void *)object, s->objsize);
930 
931                 dump_stack();
932         }
933 }
934 
935 /*
936  * Hooks for other subsystems that check memory allocations. In a typical
937  * production configuration these hooks all should produce no code at all.
938  */
939 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
940 {
941         flags &= gfp_allowed_mask;
942         lockdep_trace_alloc(flags);
943         might_sleep_if(flags & __GFP_WAIT);
944 
945         return should_failslab(s->objsize, flags, s->flags);
946 }
947 
948 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object)
949 {
950         flags &= gfp_allowed_mask;
951         kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
952         kmemleak_alloc_recursive(object, s->objsize, 1, s->flags, flags);
953 }
954 
955 static inline void slab_free_hook(struct kmem_cache *s, void *x)
956 {
957         kmemleak_free_recursive(x, s->flags);
958 
959         /*
960          * Trouble is that we may no longer disable interupts in the fast path
961          * So in order to make the debug calls that expect irqs to be
962          * disabled we need to disable interrupts temporarily.
963          */
964 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
965         {
966                 unsigned long flags;
967 
968                 local_irq_save(flags);
969                 kmemcheck_slab_free(s, x, s->objsize);
970                 debug_check_no_locks_freed(x, s->objsize);
971                 local_irq_restore(flags);
972         }
973 #endif
974         if (!(s->flags & SLAB_DEBUG_OBJECTS))
975                 debug_check_no_obj_freed(x, s->objsize);
976 }
977 
978 /*
979  * Tracking of fully allocated slabs for debugging purposes.
980  *
981  * list_lock must be held.
982  */
983 static void add_full(struct kmem_cache *s,
984         struct kmem_cache_node *n, struct page *page)
985 {
986         if (!(s->flags & SLAB_STORE_USER))
987                 return;
988 
989         list_add(&page->lru, &n->full);
990 }
991 
992 /*
993  * list_lock must be held.
994  */
995 static void remove_full(struct kmem_cache *s, struct page *page)
996 {
997         if (!(s->flags & SLAB_STORE_USER))
998                 return;
999 
1000         list_del(&page->lru);
1001 }
1002 
1003 /* Tracking of the number of slabs for debugging purposes */
1004 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1005 {
1006         struct kmem_cache_node *n = get_node(s, node);
1007 
1008         return atomic_long_read(&n->nr_slabs);
1009 }
1010 
1011 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1012 {
1013         return atomic_long_read(&n->nr_slabs);
1014 }
1015 
1016 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1017 {
1018         struct kmem_cache_node *n = get_node(s, node);
1019 
1020         /*
1021          * May be called early in order to allocate a slab for the
1022          * kmem_cache_node structure. Solve the chicken-egg
1023          * dilemma by deferring the increment of the count during
1024          * bootstrap (see early_kmem_cache_node_alloc).
1025          */
1026         if (n) {
1027                 atomic_long_inc(&n->nr_slabs);
1028                 atomic_long_add(objects, &n->total_objects);
1029         }
1030 }
1031 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1032 {
1033         struct kmem_cache_node *n = get_node(s, node);
1034 
1035         atomic_long_dec(&n->nr_slabs);
1036         atomic_long_sub(objects, &n->total_objects);
1037 }
1038 
1039 /* Object debug checks for alloc/free paths */
1040 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1041                                                                 void *object)
1042 {
1043         if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1044                 return;
1045 
1046         init_object(s, object, SLUB_RED_INACTIVE);
1047         init_tracking(s, object);
1048 }
1049 
1050 static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page,
1051                                         void *object, unsigned long addr)
1052 {
1053         if (!check_slab(s, page))
1054                 goto bad;
1055 
1056         if (!check_valid_pointer(s, page, object)) {
1057                 object_err(s, page, object, "Freelist Pointer check fails");
1058                 goto bad;
1059         }
1060 
1061         if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1062                 goto bad;
1063 
1064         /* Success perform special debug activities for allocs */
1065         if (s->flags & SLAB_STORE_USER)
1066                 set_track(s, object, TRACK_ALLOC, addr);
1067         trace(s, page, object, 1);
1068         init_object(s, object, SLUB_RED_ACTIVE);
1069         return 1;
1070 
1071 bad:
1072         if (PageSlab(page)) {
1073                 /*
1074                  * If this is a slab page then lets do the best we can
1075                  * to avoid issues in the future. Marking all objects
1076                  * as used avoids touching the remaining objects.
1077                  */
1078                 slab_fix(s, "Marking all objects used");
1079                 page->inuse = page->objects;
1080                 page->freelist = NULL;
1081         }
1082         return 0;
1083 }
1084 
1085 static noinline int free_debug_processing(struct kmem_cache *s,
1086                  struct page *page, void *object, unsigned long addr)
1087 {
1088         unsigned long flags;
1089         int rc = 0;
1090 
1091         local_irq_save(flags);
1092         slab_lock(page);
1093 
1094         if (!check_slab(s, page))
1095                 goto fail;
1096 
1097         if (!check_valid_pointer(s, page, object)) {
1098                 slab_err(s, page, "Invalid object pointer 0x%p", object);
1099                 goto fail;
1100         }
1101 
1102         if (on_freelist(s, page, object)) {
1103                 object_err(s, page, object, "Object already free");
1104                 goto fail;
1105         }
1106 
1107         if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1108                 goto out;
1109 
1110         if (unlikely(s != page->slab)) {
1111                 if (!PageSlab(page)) {
1112                         slab_err(s, page, "Attempt to free object(0x%p) "
1113                                 "outside of slab", object);
1114                 } else if (!page->slab) {
1115                         printk(KERN_ERR
1116                                 "SLUB <none>: no slab for object 0x%p.\n",
1117                                                 object);
1118                         dump_stack();
1119                 } else
1120                         object_err(s, page, object,
1121                                         "page slab pointer corrupt.");
1122                 goto fail;
1123         }
1124 
1125         if (s->flags & SLAB_STORE_USER)
1126                 set_track(s, object, TRACK_FREE, addr);
1127         trace(s, page, object, 0);
1128         init_object(s, object, SLUB_RED_INACTIVE);
1129         rc = 1;
1130 out:
1131         slab_unlock(page);
1132         local_irq_restore(flags);
1133         return rc;
1134 
1135 fail:
1136         slab_fix(s, "Object at 0x%p not freed", object);
1137         goto out;
1138 }
1139 
1140 static int __init setup_slub_debug(char *str)
1141 {
1142         slub_debug = DEBUG_DEFAULT_FLAGS;
1143         if (*str++ != '=' || !*str)
1144                 /*
1145                  * No options specified. Switch on full debugging.
1146                  */
1147                 goto out;
1148 
1149         if (*str == ',')
1150                 /*
1151                  * No options but restriction on slabs. This means full
1152                  * debugging for slabs matching a pattern.
1153                  */
1154                 goto check_slabs;
1155 
1156         if (tolower(*str) == 'o') {
1157                 /*
1158                  * Avoid enabling debugging on caches if its minimum order
1159                  * would increase as a result.
1160                  */
1161                 disable_higher_order_debug = 1;
1162                 goto out;
1163         }
1164 
1165         slub_debug = 0;
1166         if (*str == '-')
1167                 /*
1168                  * Switch off all debugging measures.
1169                  */
1170                 goto out;
1171 
1172         /*
1173          * Determine which debug features should be switched on
1174          */
1175         for (; *str && *str != ','; str++) {
1176                 switch (tolower(*str)) {
1177                 case 'f':
1178                         slub_debug |= SLAB_DEBUG_FREE;
1179                         break;
1180                 case 'z':
1181                         slub_debug |= SLAB_RED_ZONE;
1182                         break;
1183                 case 'p':
1184                         slub_debug |= SLAB_POISON;
1185                         break;
1186                 case 'u':
1187                         slub_debug |= SLAB_STORE_USER;
1188                         break;
1189                 case 't':
1190                         slub_debug |= SLAB_TRACE;
1191                         break;
1192                 case 'a':
1193                         slub_debug |= SLAB_FAILSLAB;
1194                         break;
1195                 default:
1196                         printk(KERN_ERR "slub_debug option '%c' "
1197                                 "unknown. skipped\n", *str);
1198                 }
1199         }
1200 
1201 check_slabs:
1202         if (*str == ',')
1203                 slub_debug_slabs = str + 1;
1204 out:
1205         return 1;
1206 }
1207 
1208 __setup("slub_debug", setup_slub_debug);
1209 
1210 static unsigned long kmem_cache_flags(unsigned long objsize,
1211         unsigned long flags, const char *name,
1212         void (*ctor)(void *))
1213 {
1214         /*
1215          * Enable debugging if selected on the kernel commandline.
1216          */
1217         if (slub_debug && (!slub_debug_slabs ||
1218                 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1219                 flags |= slub_debug;
1220 
1221         return flags;
1222 }
1223 #else
1224 static inline void setup_object_debug(struct kmem_cache *s,
1225                         struct page *page, void *object) {}
1226 
1227 static inline int alloc_debug_processing(struct kmem_cache *s,
1228         struct page *page, void *object, unsigned long addr) { return 0; }
1229 
1230 static inline int free_debug_processing(struct kmem_cache *s,
1231         struct page *page, void *object, unsigned long addr) { return 0; }
1232 
1233 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1234                         { return 1; }
1235 static inline int check_object(struct kmem_cache *s, struct page *page,
1236                         void *object, u8 val) { return 1; }
1237 static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1238                                         struct page *page) {}
1239 static inline void remove_full(struct kmem_cache *s, struct page *page) {}
1240 static inline unsigned long kmem_cache_flags(unsigned long objsize,
1241         unsigned long flags, const char *name,
1242         void (*ctor)(void *))
1243 {
1244         return flags;
1245 }
1246 #define slub_debug 0
1247 
1248 #define disable_higher_order_debug 0
1249 
1250 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1251                                                         { return 0; }
1252 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1253                                                         { return 0; }
1254 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1255                                                         int objects) {}
1256 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1257                                                         int objects) {}
1258 
1259 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1260                                                         { return 0; }
1261 
1262 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1263                 void *object) {}
1264 
1265 static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
1266 
1267 #endif /* CONFIG_SLUB_DEBUG */
1268 
1269 /*
1270  * Slab allocation and freeing
1271  */
1272 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1273                                         struct kmem_cache_order_objects oo)
1274 {
1275         int order = oo_order(oo);
1276 
1277         flags |= __GFP_NOTRACK;
1278 
1279         if (node == NUMA_NO_NODE)
1280                 return alloc_pages(flags, order);
1281         else
1282                 return alloc_pages_exact_node(node, flags, order);
1283 }
1284 
1285 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1286 {
1287         struct page *page;
1288         struct kmem_cache_order_objects oo = s->oo;
1289         gfp_t alloc_gfp;
1290 
1291         flags &= gfp_allowed_mask;
1292 
1293         if (flags & __GFP_WAIT)
1294                 local_irq_enable();
1295 
1296         flags |= s->allocflags;
1297 
1298         /*
1299          * Let the initial higher-order allocation fail under memory pressure
1300          * so we fall-back to the minimum order allocation.
1301          */
1302         alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1303 
1304         page = alloc_slab_page(alloc_gfp, node, oo);
1305         if (unlikely(!page)) {
1306                 oo = s->min;
1307                 /*
1308                  * Allocation may have failed due to fragmentation.
1309                  * Try a lower order alloc if possible
1310                  */
1311                 page = alloc_slab_page(flags, node, oo);
1312 
1313                 if (page)
1314                         stat(s, ORDER_FALLBACK);
1315         }
1316 
1317         if (flags & __GFP_WAIT)
1318                 local_irq_disable();
1319 
1320         if (!page)
1321                 return NULL;
1322 
1323         if (kmemcheck_enabled
1324                 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1325                 int pages = 1 << oo_order(oo);
1326 
1327                 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1328 
1329                 /*
1330                  * Objects from caches that have a constructor don't get
1331                  * cleared when they're allocated, so we need to do it here.
1332                  */
1333                 if (s->ctor)
1334                         kmemcheck_mark_uninitialized_pages(page, pages);
1335                 else
1336                         kmemcheck_mark_unallocated_pages(page, pages);
1337         }
1338 
1339         page->objects = oo_objects(oo);
1340         mod_zone_page_state(page_zone(page),
1341                 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1342                 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1343                 1 << oo_order(oo));
1344 
1345         return page;
1346 }
1347 
1348 static void setup_object(struct kmem_cache *s, struct page *page,
1349                                 void *object)
1350 {
1351         setup_object_debug(s, page, object);
1352         if (unlikely(s->ctor))
1353                 s->ctor(object);
1354 }
1355 
1356 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1357 {
1358         struct page *page;
1359         void *start;
1360         void *last;
1361         void *p;
1362 
1363         BUG_ON(flags & GFP_SLAB_BUG_MASK);
1364 
1365         page = allocate_slab(s,
1366                 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1367         if (!page)
1368                 goto out;
1369 
1370         inc_slabs_node(s, page_to_nid(page), page->objects);
1371         page->slab = s;
1372         page->flags |= 1 << PG_slab;
1373 
1374         start = page_address(page);
1375 
1376         if (unlikely(s->flags & SLAB_POISON))
1377                 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1378 
1379         last = start;
1380         for_each_object(p, s, start, page->objects) {
1381                 setup_object(s, page, last);
1382                 set_freepointer(s, last, p);
1383                 last = p;
1384         }
1385         setup_object(s, page, last);
1386         set_freepointer(s, last, NULL);
1387 
1388         page->freelist = start;
1389         page->inuse = page->objects;
1390         page->frozen = 1;
1391 out:
1392         return page;
1393 }
1394 
1395 static void __free_slab(struct kmem_cache *s, struct page *page)
1396 {
1397         int order = compound_order(page);
1398         int pages = 1 << order;
1399 
1400         if (kmem_cache_debug(s)) {
1401                 void *p;
1402 
1403                 slab_pad_check(s, page);
1404                 for_each_object(p, s, page_address(page),
1405                                                 page->objects)
1406                         check_object(s, page, p, SLUB_RED_INACTIVE);
1407         }
1408 
1409         kmemcheck_free_shadow(page, compound_order(page));
1410 
1411         mod_zone_page_state(page_zone(page),
1412                 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1413                 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1414                 -pages);
1415 
1416         __ClearPageSlab(page);
1417         reset_page_mapcount(page);
1418         if (current->reclaim_state)
1419                 current->reclaim_state->reclaimed_slab += pages;
1420         __free_pages(page, order);
1421 }
1422 
1423 #define need_reserve_slab_rcu                                           \
1424         (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1425 
1426 static void rcu_free_slab(struct rcu_head *h)
1427 {
1428         struct page *page;
1429 
1430         if (need_reserve_slab_rcu)
1431                 page = virt_to_head_page(h);
1432         else
1433                 page = container_of((struct list_head *)h, struct page, lru);
1434 
1435         __free_slab(page->slab, page);
1436 }
1437 
1438 static void free_slab(struct kmem_cache *s, struct page *page)
1439 {
1440         if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1441                 struct rcu_head *head;
1442 
1443                 if (need_reserve_slab_rcu) {
1444                         int order = compound_order(page);
1445                         int offset = (PAGE_SIZE << order) - s->reserved;
1446 
1447                         VM_BUG_ON(s->reserved != sizeof(*head));
1448                         head = page_address(page) + offset;
1449                 } else {
1450                         /*
1451                          * RCU free overloads the RCU head over the LRU
1452                          */
1453                         head = (void *)&page->lru;
1454                 }
1455 
1456                 call_rcu(head, rcu_free_slab);
1457         } else
1458                 __free_slab(s, page);
1459 }
1460 
1461 static void discard_slab(struct kmem_cache *s, struct page *page)
1462 {
1463         dec_slabs_node(s, page_to_nid(page), page->objects);
1464         free_slab(s, page);
1465 }
1466 
1467 /*
1468  * Management of partially allocated slabs.
1469  *
1470  * list_lock must be held.
1471  */
1472 static inline void add_partial(struct kmem_cache_node *n,
1473                                 struct page *page, int tail)
1474 {
1475         n->nr_partial++;
1476         if (tail == DEACTIVATE_TO_TAIL)
1477                 list_add_tail(&page->lru, &n->partial);
1478         else
1479                 list_add(&page->lru, &n->partial);
1480 }
1481 
1482 /*
1483  * list_lock must be held.
1484  */
1485 static inline void remove_partial(struct kmem_cache_node *n,
1486                                         struct page *page)
1487 {
1488         list_del(&page->lru);
1489         n->nr_partial--;
1490 }
1491 
1492 /*
1493  * Lock slab, remove from the partial list and put the object into the
1494  * per cpu freelist.
1495  *
1496  * Returns a list of objects or NULL if it fails.
1497  *
1498  * Must hold list_lock.
1499  */
1500 static inline void *acquire_slab(struct kmem_cache *s,
1501                 struct kmem_cache_node *n, struct page *page,
1502                 int mode)
1503 {
1504         void *freelist;
1505         unsigned long counters;
1506         struct page new;
1507 
1508         /*
1509          * Zap the freelist and set the frozen bit.
1510          * The old freelist is the list of objects for the
1511          * per cpu allocation list.
1512          */
1513         do {
1514                 freelist = page->freelist;
1515                 counters = page->counters;
1516                 new.counters = counters;
1517                 if (mode) {
1518                         new.inuse = page->objects;
1519                         new.freelist = NULL;
1520                 } else {
1521                         new.freelist = freelist;
1522                 }
1523 
1524                 VM_BUG_ON(new.frozen);
1525                 new.frozen = 1;
1526 
1527         } while (!__cmpxchg_double_slab(s, page,
1528                         freelist, counters,
1529                         new.freelist, new.counters,
1530                         "lock and freeze"));
1531 
1532         remove_partial(n, page);
1533         return freelist;
1534 }
1535 
1536 static int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1537 
1538 /*
1539  * Try to allocate a partial slab from a specific node.
1540  */
1541 static void *get_partial_node(struct kmem_cache *s,
1542                 struct kmem_cache_node *n, struct kmem_cache_cpu *c)
1543 {
1544         struct page *page, *page2;
1545         void *object = NULL;
1546 
1547         /*
1548          * Racy check. If we mistakenly see no partial slabs then we
1549          * just allocate an empty slab. If we mistakenly try to get a
1550          * partial slab and there is none available then get_partials()
1551          * will return NULL.
1552          */
1553         if (!n || !n->nr_partial)
1554                 return NULL;
1555 
1556         spin_lock(&n->list_lock);
1557         list_for_each_entry_safe(page, page2, &n->partial, lru) {
1558                 void *t = acquire_slab(s, n, page, object == NULL);
1559                 int available;
1560 
1561                 if (!t)
1562                         break;
1563 
1564                 if (!object) {
1565                         c->page = page;
1566                         c->node = page_to_nid(page);
1567                         stat(s, ALLOC_FROM_PARTIAL);
1568                         object = t;
1569                         available =  page->objects - page->inuse;
1570                 } else {
1571                         available = put_cpu_partial(s, page, 0);
1572                         stat(s, CPU_PARTIAL_NODE);
1573                 }
1574                 if (kmem_cache_debug(s) || available > s->cpu_partial / 2)
1575                         break;
1576 
1577         }
1578         spin_unlock(&n->list_lock);
1579         return object;
1580 }
1581 
1582 /*
1583  * Get a page from somewhere. Search in increasing NUMA distances.
1584  */
1585 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags,
1586                 struct kmem_cache_cpu *c)
1587 {
1588 #ifdef CONFIG_NUMA
1589         struct zonelist *zonelist;
1590         struct zoneref *z;
1591         struct zone *zone;
1592         enum zone_type high_zoneidx = gfp_zone(flags);
1593         void *object;
1594         unsigned int cpuset_mems_cookie;
1595 
1596         /*
1597          * The defrag ratio allows a configuration of the tradeoffs between
1598          * inter node defragmentation and node local allocations. A lower
1599          * defrag_ratio increases the tendency to do local allocations
1600          * instead of attempting to obtain partial slabs from other nodes.
1601          *
1602          * If the defrag_ratio is set to 0 then kmalloc() always
1603          * returns node local objects. If the ratio is higher then kmalloc()
1604          * may return off node objects because partial slabs are obtained
1605          * from other nodes and filled up.
1606          *
1607          * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1608          * defrag_ratio = 1000) then every (well almost) allocation will
1609          * first attempt to defrag slab caches on other nodes. This means
1610          * scanning over all nodes to look for partial slabs which may be
1611          * expensive if we do it every time we are trying to find a slab
1612          * with available objects.
1613          */
1614         if (!s->remote_node_defrag_ratio ||
1615                         get_cycles() % 1024 > s->remote_node_defrag_ratio)
1616                 return NULL;
1617 
1618         do {
1619                 cpuset_mems_cookie = get_mems_allowed();
1620                 zonelist = node_zonelist(slab_node(), flags);
1621                 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1622                         struct kmem_cache_node *n;
1623 
1624                         n = get_node(s, zone_to_nid(zone));
1625 
1626                         if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1627                                         n->nr_partial > s->min_partial) {
1628                                 object = get_partial_node(s, n, c);
1629                                 if (object) {
1630                                         /*
1631                                          * Return the object even if
1632                                          * put_mems_allowed indicated that
1633                                          * the cpuset mems_allowed was
1634                                          * updated in parallel. It's a
1635                                          * harmless race between the alloc
1636                                          * and the cpuset update.
1637                                          */
1638                                         put_mems_allowed(cpuset_mems_cookie);
1639                                         return object;
1640                                 }
1641                         }
1642                 }
1643         } while (!put_mems_allowed(cpuset_mems_cookie));
1644 #endif
1645         return NULL;
1646 }
1647 
1648 /*
1649  * Get a partial page, lock it and return it.
1650  */
1651 static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1652                 struct kmem_cache_cpu *c)
1653 {
1654         void *object;
1655         int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1656 
1657         object = get_partial_node(s, get_node(s, searchnode), c);
1658         if (object || node != NUMA_NO_NODE)
1659                 return object;
1660 
1661         return get_any_partial(s, flags, c);
1662 }
1663 
1664 #ifdef CONFIG_PREEMPT
1665 /*
1666  * Calculate the next globally unique transaction for disambiguiation
1667  * during cmpxchg. The transactions start with the cpu number and are then
1668  * incremented by CONFIG_NR_CPUS.
1669  */
1670 #define TID_STEP  roundup_pow_of_two(CONFIG_NR_CPUS)
1671 #else
1672 /*
1673  * No preemption supported therefore also no need to check for
1674  * different cpus.
1675  */
1676 #define TID_STEP 1
1677 #endif
1678 
1679 static inline unsigned long next_tid(unsigned long tid)
1680 {
1681         return tid + TID_STEP;
1682 }
1683 
1684 static inline unsigned int tid_to_cpu(unsigned long tid)
1685 {
1686         return tid % TID_STEP;
1687 }
1688 
1689 static inline unsigned long tid_to_event(unsigned long tid)
1690 {
1691         return tid / TID_STEP;
1692 }
1693 
1694 static inline unsigned int init_tid(int cpu)
1695 {
1696         return cpu;
1697 }
1698 
1699 static inline void note_cmpxchg_failure(const char *n,
1700                 const struct kmem_cache *s, unsigned long tid)
1701 {
1702 #ifdef SLUB_DEBUG_CMPXCHG
1703         unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1704 
1705         printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1706 
1707 #ifdef CONFIG_PREEMPT
1708         if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1709                 printk("due to cpu change %d -> %d\n",
1710                         tid_to_cpu(tid), tid_to_cpu(actual_tid));
1711         else
1712 #endif
1713         if (tid_to_event(tid) != tid_to_event(actual_tid))
1714                 printk("due to cpu running other code. Event %ld->%ld\n",
1715                         tid_to_event(tid), tid_to_event(actual_tid));
1716         else
1717                 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1718                         actual_tid, tid, next_tid(tid));
1719 #endif
1720         stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1721 }
1722 
1723 void init_kmem_cache_cpus(struct kmem_cache *s)
1724 {
1725         int cpu;
1726 
1727         for_each_possible_cpu(cpu)
1728                 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1729 }
1730 
1731 /*
1732  * Remove the cpu slab
1733  */
1734 static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1735 {
1736         enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1737         struct page *page = c->page;
1738         struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1739         int lock = 0;
1740         enum slab_modes l = M_NONE, m = M_NONE;
1741         void *freelist;
1742         void *nextfree;
1743         int tail = DEACTIVATE_TO_HEAD;
1744         struct page new;
1745         struct page old;
1746 
1747         if (page->freelist) {
1748                 stat(s, DEACTIVATE_REMOTE_FREES);
1749                 tail = DEACTIVATE_TO_TAIL;
1750         }
1751 
1752         c->tid = next_tid(c->tid);
1753         c->page = NULL;
1754         freelist = c->freelist;
1755         c->freelist = NULL;
1756 
1757         /*
1758          * Stage one: Free all available per cpu objects back
1759          * to the page freelist while it is still frozen. Leave the
1760          * last one.
1761          *
1762          * There is no need to take the list->lock because the page
1763          * is still frozen.
1764          */
1765         while (freelist && (nextfree = get_freepointer(s, freelist))) {
1766                 void *prior;
1767                 unsigned long counters;
1768 
1769                 do {
1770                         prior = page->freelist;
1771                         counters = page->counters;
1772                         set_freepointer(s, freelist, prior);
1773                         new.counters = counters;
1774                         new.inuse--;
1775                         VM_BUG_ON(!new.frozen);
1776 
1777                 } while (!__cmpxchg_double_slab(s, page,
1778                         prior, counters,
1779                         freelist, new.counters,
1780                         "drain percpu freelist"));
1781 
1782                 freelist = nextfree;
1783         }
1784 
1785         /*
1786          * Stage two: Ensure that the page is unfrozen while the
1787          * list presence reflects the actual number of objects
1788          * during unfreeze.
1789          *
1790          * We setup the list membership and then perform a cmpxchg
1791          * with the count. If there is a mismatch then the page
1792          * is not unfrozen but the page is on the wrong list.
1793          *
1794          * Then we restart the process which may have to remove
1795          * the page from the list that we just put it on again
1796          * because the number of objects in the slab may have
1797          * changed.
1798          */
1799 redo:
1800 
1801         old.freelist = page->freelist;
1802         old.counters = page->counters;
1803         VM_BUG_ON(!old.frozen);
1804 
1805         /* Determine target state of the slab */
1806         new.counters = old.counters;
1807         if (freelist) {
1808                 new.inuse--;
1809                 set_freepointer(s, freelist, old.freelist);
1810                 new.freelist = freelist;
1811         } else
1812                 new.freelist = old.freelist;
1813 
1814         new.frozen = 0;
1815 
1816         if (!new.inuse && n->nr_partial > s->min_partial)
1817                 m = M_FREE;
1818         else if (new.freelist) {
1819                 m = M_PARTIAL;
1820                 if (!lock) {
1821                         lock = 1;
1822                         /*
1823                          * Taking the spinlock removes the possiblity
1824                          * that acquire_slab() will see a slab page that
1825                          * is frozen
1826                          */
1827                         spin_lock(&n->list_lock);
1828                 }
1829         } else {
1830                 m = M_FULL;
1831                 if (kmem_cache_debug(s) && !lock) {
1832                         lock = 1;
1833                         /*
1834                          * This also ensures that the scanning of full
1835                          * slabs from diagnostic functions will not see
1836                          * any frozen slabs.
1837                          */
1838                         spin_lock(&n->list_lock);
1839                 }
1840         }
1841 
1842         if (l != m) {
1843 
1844                 if (l == M_PARTIAL)
1845 
1846                         remove_partial(n, page);
1847 
1848                 else if (l == M_FULL)
1849 
1850                         remove_full(s, page);
1851 
1852                 if (m == M_PARTIAL) {
1853 
1854                         add_partial(n, page, tail);
1855                         stat(s, tail);
1856 
1857                 } else if (m == M_FULL) {
1858 
1859                         stat(s, DEACTIVATE_FULL);
1860                         add_full(s, n, page);
1861 
1862                 }
1863         }
1864 
1865         l = m;
1866         if (!__cmpxchg_double_slab(s, page,
1867                                 old.freelist, old.counters,
1868                                 new.freelist, new.counters,
1869                                 "unfreezing slab"))
1870                 goto redo;
1871 
1872         if (lock)
1873                 spin_unlock(&n->list_lock);
1874 
1875         if (m == M_FREE) {
1876                 stat(s, DEACTIVATE_EMPTY);
1877                 discard_slab(s, page);
1878                 stat(s, FREE_SLAB);
1879         }
1880 }
1881 
1882 /* Unfreeze all the cpu partial slabs */
1883 static void unfreeze_partials(struct kmem_cache *s)
1884 {
1885         struct kmem_cache_node *n = NULL, *n2 = NULL;
1886         struct kmem_cache_cpu *c = this_cpu_ptr(s->cpu_slab);
1887         struct page *page, *discard_page = NULL;
1888 
1889         while ((page = c->partial)) {
1890                 struct page new;
1891                 struct page old;
1892 
1893                 c->partial = page->next;
1894 
1895                 n2 = get_node(s, page_to_nid(page));
1896                 if (n != n2) {
1897                         if (n)
1898                                 spin_unlock(&n->list_lock);
1899 
1900                         n = n2;
1901                         spin_lock(&n->list_lock);
1902                 }
1903 
1904                 do {
1905 
1906                         old.freelist = page->freelist;
1907                         old.counters = page->counters;
1908                         VM_BUG_ON(!old.frozen);
1909 
1910                         new.counters = old.counters;
1911                         new.freelist = old.freelist;
1912 
1913                         new.frozen = 0;
1914 
1915                 } while (!cmpxchg_double_slab(s, page,
1916                                 old.freelist, old.counters,
1917                                 new.freelist, new.counters,
1918                                 "unfreezing slab"));
1919 
1920                 if (unlikely(!new.inuse && n->nr_partial > s->min_partial)) {
1921                         page->next = discard_page;
1922                         discard_page = page;
1923                 } else {
1924                         add_partial(n, page, DEACTIVATE_TO_TAIL);
1925                         stat(s, FREE_ADD_PARTIAL);
1926                 }
1927         }
1928 
1929         if (n)
1930                 spin_unlock(&n->list_lock);
1931 
1932         while (discard_page) {
1933                 page = discard_page;
1934                 discard_page = discard_page->next;
1935 
1936                 stat(s, DEACTIVATE_EMPTY);
1937                 discard_slab(s, page);
1938                 stat(s, FREE_SLAB);
1939         }
1940 }
1941 
1942 /*
1943  * Put a page that was just frozen (in __slab_free) into a partial page
1944  * slot if available. This is done without interrupts disabled and without
1945  * preemption disabled. The cmpxchg is racy and may put the partial page
1946  * onto a random cpus partial slot.
1947  *
1948  * If we did not find a slot then simply move all the partials to the
1949  * per node partial list.
1950  */
1951 int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
1952 {
1953         struct page *oldpage;
1954         int pages;
1955         int pobjects;
1956 
1957         do {
1958                 pages = 0;
1959                 pobjects = 0;
1960                 oldpage = this_cpu_read(s->cpu_slab->partial);
1961 
1962                 if (oldpage) {
1963                         pobjects = oldpage->pobjects;
1964                         pages = oldpage->pages;
1965                         if (drain && pobjects > s->cpu_partial) {
1966                                 unsigned long flags;
1967                                 /*
1968                                  * partial array is full. Move the existing
1969                                  * set to the per node partial list.
1970                                  */
1971                                 local_irq_save(flags);
1972                                 unfreeze_partials(s);
1973                                 local_irq_restore(flags);
1974                                 pobjects = 0;
1975                                 pages = 0;
1976                                 stat(s, CPU_PARTIAL_DRAIN);
1977                         }
1978                 }
1979 
1980                 pages++;
1981                 pobjects += page->objects - page->inuse;
1982 
1983                 page->pages = pages;
1984                 page->pobjects = pobjects;
1985                 page->next = oldpage;
1986 
1987         } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page) != oldpage);
1988         return pobjects;
1989 }
1990 
1991 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1992 {
1993         stat(s, CPUSLAB_FLUSH);
1994         deactivate_slab(s, c);
1995 }
1996 
1997 /*
1998  * Flush cpu slab.
1999  *
2000  * Called from IPI handler with interrupts disabled.
2001  */
2002 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2003 {
2004         struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2005 
2006         if (likely(c)) {
2007                 if (c->page)
2008                         flush_slab(s, c);
2009 
2010                 unfreeze_partials(s);
2011         }
2012 }
2013 
2014 static void flush_cpu_slab(void *d)
2015 {
2016         struct kmem_cache *s = d;
2017 
2018         __flush_cpu_slab(s, smp_processor_id());
2019 }
2020 
2021 static bool has_cpu_slab(int cpu, void *info)
2022 {
2023         struct kmem_cache *s = info;
2024         struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2025 
2026         return c->page || c->partial;
2027 }
2028 
2029 static void flush_all(struct kmem_cache *s)
2030 {
2031         on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2032 }
2033 
2034 /*
2035  * Check if the objects in a per cpu structure fit numa
2036  * locality expectations.
2037  */
2038 static inline int node_match(struct kmem_cache_cpu *c, int node)
2039 {
2040 #ifdef CONFIG_NUMA
2041         if (node != NUMA_NO_NODE && c->node != node)
2042                 return 0;
2043 #endif
2044         return 1;
2045 }
2046 
2047 static int count_free(struct page *page)
2048 {
2049         return page->objects - page->inuse;
2050 }
2051 
2052 static unsigned long count_partial(struct kmem_cache_node *n,
2053                                         int (*get_count)(struct page *))
2054 {
2055         unsigned long flags;
2056         unsigned long x = 0;
2057         struct page *page;
2058 
2059         spin_lock_irqsave(&n->list_lock, flags);
2060         list_for_each_entry(page, &n->partial, lru)
2061                 x += get_count(page);
2062         spin_unlock_irqrestore(&n->list_lock, flags);
2063         return x;
2064 }
2065 
2066 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2067 {
2068 #ifdef CONFIG_SLUB_DEBUG
2069         return atomic_long_read(&n->total_objects);
2070 #else
2071         return 0;
2072 #endif
2073 }
2074 
2075 static noinline void
2076 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2077 {
2078         int node;
2079 
2080         printk(KERN_WARNING
2081                 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2082                 nid, gfpflags);
2083         printk(KERN_WARNING "  cache: %s, object size: %d, buffer size: %d, "
2084                 "default order: %d, min order: %d\n", s->name, s->objsize,
2085                 s->size, oo_order(s->oo), oo_order(s->min));
2086 
2087         if (oo_order(s->min) > get_order(s->objsize))
2088                 printk(KERN_WARNING "  %s debugging increased min order, use "
2089                        "slub_debug=O to disable.\n", s->name);
2090 
2091         for_each_online_node(node) {
2092                 struct kmem_cache_node *n = get_node(s, node);
2093                 unsigned long nr_slabs;
2094                 unsigned long nr_objs;
2095                 unsigned long nr_free;
2096 
2097                 if (!n)
2098                         continue;
2099 
2100                 nr_free  = count_partial(n, count_free);
2101                 nr_slabs = node_nr_slabs(n);
2102                 nr_objs  = node_nr_objs(n);
2103 
2104                 printk(KERN_WARNING
2105                         "  node %d: slabs: %ld, objs: %ld, free: %ld\n",
2106                         node, nr_slabs, nr_objs, nr_free);
2107         }
2108 }
2109 
2110 static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2111                         int node, struct kmem_cache_cpu **pc)
2112 {
2113         void *object;
2114         struct kmem_cache_cpu *c;
2115         struct page *page = new_slab(s, flags, node);
2116 
2117         if (page) {
2118                 c = __this_cpu_ptr(s->cpu_slab);
2119                 if (c->page)
2120                         flush_slab(s, c);
2121 
2122                 /*
2123                  * No other reference to the page yet so we can
2124                  * muck around with it freely without cmpxchg
2125                  */
2126                 object = page->freelist;
2127                 page->freelist = NULL;
2128 
2129                 stat(s, ALLOC_SLAB);
2130                 c->node = page_to_nid(page);
2131                 c->page = page;
2132                 *pc = c;
2133         } else
2134                 object = NULL;
2135 
2136         return object;
2137 }
2138 
2139 /*
2140  * Check the page->freelist of a page and either transfer the freelist to the per cpu freelist
2141  * or deactivate the page.
2142  *
2143  * The page is still frozen if the return value is not NULL.
2144  *
2145  * If this function returns NULL then the page has been unfrozen.
2146  */
2147 static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2148 {
2149         struct page new;
2150         unsigned long counters;
2151         void *freelist;
2152 
2153         do {
2154                 freelist = page->freelist;
2155                 counters = page->counters;
2156                 new.counters = counters;
2157                 VM_BUG_ON(!new.frozen);
2158 
2159                 new.inuse = page->objects;
2160                 new.frozen = freelist != NULL;
2161 
2162         } while (!cmpxchg_double_slab(s, page,
2163                 freelist, counters,
2164                 NULL, new.counters,
2165                 "get_freelist"));
2166 
2167         return freelist;
2168 }
2169 
2170 /*
2171  * Slow path. The lockless freelist is empty or we need to perform
2172  * debugging duties.
2173  *
2174  * Processing is still very fast if new objects have been freed to the
2175  * regular freelist. In that case we simply take over the regular freelist
2176  * as the lockless freelist and zap the regular freelist.
2177  *
2178  * If that is not working then we fall back to the partial lists. We take the
2179  * first element of the freelist as the object to allocate now and move the
2180  * rest of the freelist to the lockless freelist.
2181  *
2182  * And if we were unable to get a new slab from the partial slab lists then
2183  * we need to allocate a new slab. This is the slowest path since it involves
2184  * a call to the page allocator and the setup of a new slab.
2185  */
2186 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2187                           unsigned long addr, struct kmem_cache_cpu *c)
2188 {
2189         void **object;
2190         unsigned long flags;
2191 
2192         local_irq_save(flags);
2193 #ifdef CONFIG_PREEMPT
2194         /*
2195          * We may have been preempted and rescheduled on a different
2196          * cpu before disabling interrupts. Need to reload cpu area
2197          * pointer.
2198          */
2199         c = this_cpu_ptr(s->cpu_slab);
2200 #endif
2201 
2202         if (!c->page)
2203                 goto new_slab;
2204 redo:
2205         if (unlikely(!node_match(c, node))) {
2206                 stat(s, ALLOC_NODE_MISMATCH);
2207                 deactivate_slab(s, c);
2208                 goto new_slab;
2209         }
2210 
2211         /* must check again c->freelist in case of cpu migration or IRQ */
2212         object = c->freelist;
2213         if (object)
2214                 goto load_freelist;
2215 
2216         stat(s, ALLOC_SLOWPATH);
2217 
2218         object = get_freelist(s, c->page);
2219 
2220         if (!object) {
2221                 c->page = NULL;
2222                 stat(s, DEACTIVATE_BYPASS);
2223                 goto new_slab;
2224         }
2225 
2226         stat(s, ALLOC_REFILL);
2227 
2228 load_freelist:
2229         c->freelist = get_freepointer(s, object);
2230         c->tid = next_tid(c->tid);
2231         local_irq_restore(flags);
2232         return object;
2233 
2234 new_slab:
2235 
2236         if (c->partial) {
2237                 c->page = c->partial;
2238                 c->partial = c->page->next;
2239                 c->node = page_to_nid(c->page);
2240                 stat(s, CPU_PARTIAL_ALLOC);
2241                 c->freelist = NULL;
2242                 goto redo;
2243         }
2244 
2245         /* Then do expensive stuff like retrieving pages from the partial lists */
2246         object = get_partial(s, gfpflags, node, c);
2247 
2248         if (unlikely(!object)) {
2249 
2250                 object = new_slab_objects(s, gfpflags, node, &c);
2251 
2252                 if (unlikely(!object)) {
2253                         if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
2254                                 slab_out_of_memory(s, gfpflags, node);
2255 
2256                         local_irq_restore(flags);
2257                         return NULL;
2258                 }
2259         }
2260 
2261         if (likely(!kmem_cache_debug(s)))
2262                 goto load_freelist;
2263 
2264         /* Only entered in the debug case */
2265         if (!alloc_debug_processing(s, c->page, object, addr))
2266                 goto new_slab;  /* Slab failed checks. Next slab needed */
2267 
2268         c->freelist = get_freepointer(s, object);
2269         deactivate_slab(s, c);
2270         c->node = NUMA_NO_NODE;
2271         local_irq_restore(flags);
2272         return object;
2273 }
2274 
2275 /*
2276  * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2277  * have the fastpath folded into their functions. So no function call
2278  * overhead for requests that can be satisfied on the fastpath.
2279  *
2280  * The fastpath works by first checking if the lockless freelist can be used.
2281  * If not then __slab_alloc is called for slow processing.
2282  *
2283  * Otherwise we can simply pick the next object from the lockless free list.
2284  */
2285 static __always_inline void *slab_alloc(struct kmem_cache *s,
2286                 gfp_t gfpflags, int node, unsigned long addr)
2287 {
2288         void **object;
2289         struct kmem_cache_cpu *c;
2290         unsigned long tid;
2291 
2292         if (slab_pre_alloc_hook(s, gfpflags))
2293                 return NULL;
2294 
2295 redo:
2296 
2297         /*
2298          * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2299          * enabled. We may switch back and forth between cpus while
2300          * reading from one cpu area. That does not matter as long
2301          * as we end up on the original cpu again when doing the cmpxchg.
2302          */
2303         c = __this_cpu_ptr(s->cpu_slab);
2304 
2305         /*
2306          * The transaction ids are globally unique per cpu and per operation on
2307          * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2308          * occurs on the right processor and that there was no operation on the
2309          * linked list in between.
2310          */
2311         tid = c->tid;
2312         barrier();
2313 
2314         object = c->freelist;
2315         if (unlikely(!object || !node_match(c, node)))
2316 
2317                 object = __slab_alloc(s, gfpflags, node, addr, c);
2318 
2319         else {
2320                 void *next_object = get_freepointer_safe(s, object);
2321 
2322                 /*
2323                  * The cmpxchg will only match if there was no additional
2324                  * operation and if we are on the right processor.
2325                  *
2326                  * The cmpxchg does the following atomically (without lock semantics!)
2327                  * 1. Relocate first pointer to the current per cpu area.
2328                  * 2. Verify that tid and freelist have not been changed
2329                  * 3. If they were not changed replace tid and freelist
2330                  *
2331                  * Since this is without lock semantics the protection is only against
2332                  * code executing on this cpu *not* from access by other cpus.
2333                  */
2334                 if (unlikely(!this_cpu_cmpxchg_double(
2335                                 s->cpu_slab->freelist, s->cpu_slab->tid,
2336                                 object, tid,
2337                                 next_object, next_tid(tid)))) {
2338 
2339                         note_cmpxchg_failure("slab_alloc", s, tid);
2340                         goto redo;
2341                 }
2342                 prefetch_freepointer(s, next_object);
2343                 stat(s, ALLOC_FASTPATH);
2344         }
2345 
2346         if (unlikely(gfpflags & __GFP_ZERO) && object)
2347                 memset(object, 0, s->objsize);
2348 
2349         slab_post_alloc_hook(s, gfpflags, object);
2350 
2351         return object;
2352 }
2353 
2354 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2355 {
2356         void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2357 
2358         trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags);
2359 
2360         return ret;
2361 }
2362 EXPORT_SYMBOL(kmem_cache_alloc);
2363 
2364 #ifdef CONFIG_TRACING
2365 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2366 {
2367         void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2368         trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2369         return ret;
2370 }
2371 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2372 
2373 void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
2374 {
2375         void *ret = kmalloc_order(size, flags, order);
2376         trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
2377         return ret;
2378 }
2379 EXPORT_SYMBOL(kmalloc_order_trace);
2380 #endif
2381 
2382 #ifdef CONFIG_NUMA
2383 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2384 {
2385         void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2386 
2387         trace_kmem_cache_alloc_node(_RET_IP_, ret,
2388                                     s->objsize, s->size, gfpflags, node);
2389 
2390         return ret;
2391 }
2392 EXPORT_SYMBOL(kmem_cache_alloc_node);
2393 
2394 #ifdef CONFIG_TRACING
2395 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2396                                     gfp_t gfpflags,
2397                                     int node, size_t size)
2398 {
2399         void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2400 
2401         trace_kmalloc_node(_RET_IP_, ret,
2402                            size, s->size, gfpflags, node);
2403         return ret;
2404 }
2405 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2406 #endif
2407 #endif
2408 
2409 /*
2410  * Slow patch handling. This may still be called frequently since objects
2411  * have a longer lifetime than the cpu slabs in most processing loads.
2412  *
2413  * So we still attempt to reduce cache line usage. Just take the slab
2414  * lock and free the item. If there is no additional partial page
2415  * handling required then we can return immediately.
2416  */
2417 static void __slab_free(struct kmem_cache *s, struct page *page,
2418                         void *x, unsigned long addr)
2419 {
2420         void *prior;
2421         void **object = (void *)x;
2422         int was_frozen;
2423         int inuse;
2424         struct page new;
2425         unsigned long counters;
2426         struct kmem_cache_node *n = NULL;
2427         unsigned long uninitialized_var(flags);
2428 
2429         stat(s, FREE_SLOWPATH);
2430 
2431         if (kmem_cache_debug(s) && !free_debug_processing(s, page, x, addr))
2432                 return;
2433 
2434         do {
2435                 prior = page->freelist;
2436                 counters = page->counters;
2437                 set_freepointer(s, object, prior);
2438                 new.counters = counters;
2439                 was_frozen = new.frozen;
2440                 new.inuse--;
2441                 if ((!new.inuse || !prior) && !was_frozen && !n) {
2442 
2443                         if (!kmem_cache_debug(s) && !prior)
2444 
2445                                 /*
2446                                  * Slab was on no list before and will be partially empty
2447                                  * We can defer the list move and instead freeze it.
2448                                  */
2449                                 new.frozen = 1;
2450 
2451                         else { /* Needs to be taken off a list */
2452 
2453                                 n = get_node(s, page_to_nid(page));
2454                                 /*
2455                                  * Speculatively acquire the list_lock.
2456                                  * If the cmpxchg does not succeed then we may
2457                                  * drop the list_lock without any processing.
2458                                  *
2459                                  * Otherwise the list_lock will synchronize with
2460                                  * other processors updating the list of slabs.
2461                                  */
2462                                 spin_lock_irqsave(&n->list_lock, flags);
2463 
2464                         }
2465                 }
2466                 inuse = new.inuse;
2467 
2468         } while (!cmpxchg_double_slab(s, page,
2469                 prior, counters,
2470                 object, new.counters,
2471                 "__slab_free"));
2472 
2473         if (likely(!n)) {
2474 
2475                 /*
2476                  * If we just froze the page then put it onto the
2477                  * per cpu partial list.
2478                  */
2479                 if (new.frozen && !was_frozen) {
2480                         put_cpu_partial(s, page, 1);
2481                         stat(s, CPU_PARTIAL_FREE);
2482                 }
2483                 /*
2484                  * The list lock was not taken therefore no list
2485                  * activity can be necessary.
2486                  */
2487                 if (was_frozen)
2488                         stat(s, FREE_FROZEN);
2489                 return;
2490         }
2491 
2492         /*
2493          * was_frozen may have been set after we acquired the list_lock in
2494          * an earlier loop. So we need to check it here again.
2495          */
2496         if (was_frozen)
2497                 stat(s, FREE_FROZEN);
2498         else {
2499                 if (unlikely(!inuse && n->nr_partial > s->min_partial))
2500                         goto slab_empty;
2501 
2502                 /*
2503                  * Objects left in the slab. If it was not on the partial list before
2504                  * then add it.
2505                  */
2506                 if (unlikely(!prior)) {
2507                         remove_full(s, page);
2508                         add_partial(n, page, DEACTIVATE_TO_TAIL);
2509                         stat(s, FREE_ADD_PARTIAL);
2510                 }
2511         }
2512         spin_unlock_irqrestore(&n->list_lock, flags);
2513         return;
2514 
2515 slab_empty:
2516         if (prior) {
2517                 /*
2518                  * Slab on the partial list.
2519                  */
2520                 remove_partial(n, page);
2521                 stat(s, FREE_REMOVE_PARTIAL);
2522         } else
2523                 /* Slab must be on the full list */
2524                 remove_full(s, page);
2525 
2526         spin_unlock_irqrestore(&n->list_lock, flags);
2527         stat(s, FREE_SLAB);
2528         discard_slab(s, page);
2529 }
2530 
2531 /*
2532  * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2533  * can perform fastpath freeing without additional function calls.
2534  *
2535  * The fastpath is only possible if we are freeing to the current cpu slab
2536  * of this processor. This typically the case if we have just allocated
2537  * the item before.
2538  *
2539  * If fastpath is not possible then fall back to __slab_free where we deal
2540  * with all sorts of special processing.
2541  */
2542 static __always_inline void slab_free(struct kmem_cache *s,
2543                         struct page *page, void *x, unsigned long addr)
2544 {
2545         void **object = (void *)x;
2546         struct kmem_cache_cpu *c;
2547         unsigned long tid;
2548 
2549         slab_free_hook(s, x);
2550 
2551 redo:
2552         /*
2553          * Determine the currently cpus per cpu slab.
2554          * The cpu may change afterward. However that does not matter since
2555          * data is retrieved via this pointer. If we are on the same cpu
2556          * during the cmpxchg then the free will succedd.
2557          */
2558         c = __this_cpu_ptr(s->cpu_slab);
2559 
2560         tid = c->tid;
2561         barrier();
2562 
2563         if (likely(page == c->page)) {
2564                 set_freepointer(s, object, c->freelist);
2565 
2566                 if (unlikely(!this_cpu_cmpxchg_double(
2567                                 s->cpu_slab->freelist, s->cpu_slab->tid,
2568                                 c->freelist, tid,
2569                                 object, next_tid(tid)))) {
2570 
2571                         note_cmpxchg_failure("slab_free", s, tid);
2572                         goto redo;
2573                 }
2574                 stat(s, FREE_FASTPATH);
2575         } else
2576                 __slab_free(s, page, x, addr);
2577 
2578 }
2579 
2580 void kmem_cache_free(struct kmem_cache *s, void *x)
2581 {
2582         struct page *page;
2583 
2584         page = virt_to_head_page(x);
2585 
2586         slab_free(s, page, x, _RET_IP_);
2587 
2588         trace_kmem_cache_free(_RET_IP_, x);
2589 }
2590 EXPORT_SYMBOL(kmem_cache_free);
2591 
2592 /*
2593  * Object placement in a slab is made very easy because we always start at
2594  * offset 0. If we tune the size of the object to the alignment then we can
2595  * get the required alignment by putting one properly sized object after
2596  * another.
2597  *
2598  * Notice that the allocation order determines the sizes of the per cpu
2599  * caches. Each processor has always one slab available for allocations.
2600  * Increasing the allocation order reduces the number of times that slabs
2601  * must be moved on and off the partial lists and is therefore a factor in
2602  * locking overhead.
2603  */
2604 
2605 /*
2606  * Mininum / Maximum order of slab pages. This influences locking overhead
2607  * and slab fragmentation. A higher order reduces the number of partial slabs
2608  * and increases the number of allocations possible without having to
2609  * take the list_lock.
2610  */
2611 static int slub_min_order;
2612 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2613 static int slub_min_objects;
2614 
2615 /*
2616  * Merge control. If this is set then no merging of slab caches will occur.
2617  * (Could be removed. This was introduced to pacify the merge skeptics.)
2618  */
2619 static int slub_nomerge;
2620 
2621 /*
2622  * Calculate the order of allocation given an slab object size.
2623  *
2624  * The order of allocation has significant impact on performance and other
2625  * system components. Generally order 0 allocations should be preferred since
2626  * order 0 does not cause fragmentation in the page allocator. Larger objects
2627  * be problematic to put into order 0 slabs because there may be too much
2628  * unused space left. We go to a higher order if more than 1/16th of the slab
2629  * would be wasted.
2630  *
2631  * In order to reach satisfactory performance we must ensure that a minimum
2632  * number of objects is in one slab. Otherwise we may generate too much
2633  * activity on the partial lists which requires taking the list_lock. This is
2634  * less a concern for large slabs though which are rarely used.
2635  *
2636  * slub_max_order specifies the order where we begin to stop considering the
2637  * number of objects in a slab as critical. If we reach slub_max_order then
2638  * we try to keep the page order as low as possible. So we accept more waste
2639  * of space in favor of a small page order.
2640  *
2641  * Higher order allocations also allow the placement of more objects in a
2642  * slab and thereby reduce object handling overhead. If the user has
2643  * requested a higher mininum order then we start with that one instead of
2644  * the smallest order which will fit the object.
2645  */
2646 static inline int slab_order(int size, int min_objects,
2647                                 int max_order, int fract_leftover, int reserved)
2648 {
2649         int order;
2650         int rem;
2651         int min_order = slub_min_order;
2652 
2653         if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2654                 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2655 
2656         for (order = max(min_order,
2657                                 fls(min_objects * size - 1) - PAGE_SHIFT);
2658                         order <= max_order; order++) {
2659 
2660                 unsigned long slab_size = PAGE_SIZE << order;
2661 
2662                 if (slab_size < min_objects * size + reserved)
2663                         continue;
2664 
2665                 rem = (slab_size - reserved) % size;
2666 
2667                 if (rem <= slab_size / fract_leftover)
2668                         break;
2669 
2670         }
2671 
2672         return order;
2673 }
2674 
2675 static inline int calculate_order(int size, int reserved)
2676 {
2677         int order;
2678         int min_objects;
2679         int fraction;
2680         int max_objects;
2681 
2682         /*
2683          * Attempt to find best configuration for a slab. This
2684          * works by first attempting to generate a layout with
2685          * the best configuration and backing off gradually.
2686          *
2687          * First we reduce the acceptable waste in a slab. Then
2688          * we reduce the minimum objects required in a slab.
2689          */
2690         min_objects = slub_min_objects;
2691         if (!min_objects)
2692                 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2693         max_objects = order_objects(slub_max_order, size, reserved);
2694         min_objects = min(min_objects, max_objects);
2695 
2696         while (min_objects > 1) {
2697                 fraction = 16;
2698                 while (fraction >= 4) {
2699                         order = slab_order(size, min_objects,
2700                                         slub_max_order, fraction, reserved);
2701                         if (order <= slub_max_order)
2702                                 return order;
2703                         fraction /= 2;
2704                 }
2705                 min_objects--;
2706         }
2707 
2708         /*
2709          * We were unable to place multiple objects in a slab. Now
2710          * lets see if we can place a single object there.
2711          */
2712         order = slab_order(size, 1, slub_max_order, 1, reserved);
2713         if (order <= slub_max_order)
2714                 return order;
2715 
2716         /*
2717          * Doh this slab cannot be placed using slub_max_order.
2718          */
2719         order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2720         if (order < MAX_ORDER)
2721                 return order;
2722         return -ENOSYS;
2723 }
2724 
2725 /*
2726  * Figure out what the alignment of the objects will be.
2727  */
2728 static unsigned long calculate_alignment(unsigned long flags,
2729                 unsigned long align, unsigned long size)
2730 {
2731         /*
2732          * If the user wants hardware cache aligned objects then follow that
2733          * suggestion if the object is sufficiently large.
2734          *
2735          * The hardware cache alignment cannot override the specified
2736          * alignment though. If that is greater then use it.
2737          */
2738         if (flags & SLAB_HWCACHE_ALIGN) {
2739                 unsigned long ralign = cache_line_size();
2740                 while (size <= ralign / 2)
2741                         ralign /= 2;
2742                 align = max(align, ralign);
2743         }
2744 
2745         if (align < ARCH_SLAB_MINALIGN)
2746                 align = ARCH_SLAB_MINALIGN;
2747 
2748         return ALIGN(align, sizeof(void *));
2749 }
2750 
2751 static void
2752 init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
2753 {
2754         n->nr_partial = 0;
2755         spin_lock_init(&n->list_lock);
2756         INIT_LIST_HEAD(&n->partial);
2757 #ifdef CONFIG_SLUB_DEBUG
2758         atomic_long_set(&n->nr_slabs, 0);
2759         atomic_long_set(&n->total_objects, 0);
2760         INIT_LIST_HEAD(&n->full);
2761 #endif
2762 }
2763 
2764 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2765 {
2766         BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2767                         SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu));
2768 
2769         /*
2770          * Must align to double word boundary for the double cmpxchg
2771          * instructions to work; see __pcpu_double_call_return_bool().
2772          */
2773         s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2774                                      2 * sizeof(void *));
2775 
2776         if (!s->cpu_slab)
2777                 return 0;
2778 
2779         init_kmem_cache_cpus(s);
2780 
2781         return 1;
2782 }
2783 
2784 static struct kmem_cache *kmem_cache_node;
2785 
2786 /*
2787  * No kmalloc_node yet so do it by hand. We know that this is the first
2788  * slab on the node for this slabcache. There are no concurrent accesses
2789  * possible.
2790  *
2791  * Note that this function only works on the kmalloc_node_cache
2792  * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2793  * memory on a fresh node that has no slab structures yet.
2794  */
2795 static void early_kmem_cache_node_alloc(int node)
2796 {
2797         struct page *page;
2798         struct kmem_cache_node *n;
2799 
2800         BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2801 
2802         page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2803 
2804         BUG_ON(!page);
2805         if (page_to_nid(page) != node) {
2806                 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2807                                 "node %d\n", node);
2808                 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2809                                 "in order to be able to continue\n");
2810         }
2811 
2812         n = page->freelist;
2813         BUG_ON(!n);
2814         page->freelist = get_freepointer(kmem_cache_node, n);
2815         page->inuse = 1;
2816         page->frozen = 0;
2817         kmem_cache_node->node[node] = n;
2818 #ifdef CONFIG_SLUB_DEBUG
2819         init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2820         init_tracking(kmem_cache_node, n);
2821 #endif
2822         init_kmem_cache_node(n, kmem_cache_node);
2823         inc_slabs_node(kmem_cache_node, node, page->objects);
2824 
2825         add_partial(n, page, DEACTIVATE_TO_HEAD);
2826 }
2827 
2828 static void free_kmem_cache_nodes(struct kmem_cache *s)
2829 {
2830         int node;
2831 
2832         for_each_node_state(node, N_NORMAL_MEMORY) {
2833                 struct kmem_cache_node *n = s->node[node];
2834 
2835                 if (n)
2836                         kmem_cache_free(kmem_cache_node, n);
2837 
2838                 s->node[node] = NULL;
2839         }
2840 }
2841 
2842 static int init_kmem_cache_nodes(struct kmem_cache *s)
2843 {
2844         int node;
2845 
2846         for_each_node_state(node, N_NORMAL_MEMORY) {
2847                 struct kmem_cache_node *n;
2848 
2849                 if (slab_state == DOWN) {
2850                         early_kmem_cache_node_alloc(node);
2851                         continue;
2852                 }
2853                 n = kmem_cache_alloc_node(kmem_cache_node,
2854                                                 GFP_KERNEL, node);
2855 
2856                 if (!n) {
2857                         free_kmem_cache_nodes(s);
2858                         return 0;
2859                 }
2860 
2861                 s->node[node] = n;
2862                 init_kmem_cache_node(n, s);
2863         }
2864         return 1;
2865 }
2866 
2867 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2868 {
2869         if (min < MIN_PARTIAL)
2870                 min = MIN_PARTIAL;
2871         else if (min > MAX_PARTIAL)
2872                 min = MAX_PARTIAL;
2873         s->min_partial = min;
2874 }
2875 
2876 /*
2877  * calculate_sizes() determines the order and the distribution of data within
2878  * a slab object.
2879  */
2880 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2881 {
2882         unsigned long flags = s->flags;
2883         unsigned long size = s->objsize;
2884         unsigned long align = s->align;
2885         int order;
2886 
2887         /*
2888          * Round up object size to the next word boundary. We can only
2889          * place the free pointer at word boundaries and this determines
2890          * the possible location of the free pointer.
2891          */
2892         size = ALIGN(size, sizeof(void *));
2893 
2894 #ifdef CONFIG_SLUB_DEBUG
2895         /*
2896          * Determine if we can poison the object itself. If the user of
2897          * the slab may touch the object after free or before allocation
2898          * then we should never poison the object itself.
2899          */
2900         if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2901                         !s->ctor)
2902                 s->flags |= __OBJECT_POISON;
2903         else
2904                 s->flags &= ~__OBJECT_POISON;
2905 
2906 
2907         /*
2908          * If we are Redzoning then check if there is some space between the
2909          * end of the object and the free pointer. If not then add an
2910          * additional word to have some bytes to store Redzone information.
2911          */
2912         if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2913                 size += sizeof(void *);
2914 #endif
2915 
2916         /*
2917          * With that we have determined the number of bytes in actual use
2918          * by the object. This is the potential offset to the free pointer.
2919          */
2920         s->inuse = size;
2921 
2922         if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2923                 s->ctor)) {
2924                 /*
2925                  * Relocate free pointer after the object if it is not
2926                  * permitted to overwrite the first word of the object on
2927                  * kmem_cache_free.
2928                  *
2929                  * This is the case if we do RCU, have a constructor or
2930                  * destructor or are poisoning the objects.
2931                  */
2932                 s->offset = size;
2933                 size += sizeof(void *);
2934         }
2935 
2936 #ifdef CONFIG_SLUB_DEBUG
2937         if (flags & SLAB_STORE_USER)
2938                 /*
2939                  * Need to store information about allocs and frees after
2940                  * the object.
2941                  */
2942                 size += 2 * sizeof(struct track);
2943 
2944         if (flags & SLAB_RED_ZONE)
2945                 /*
2946                  * Add some empty padding so that we can catch
2947                  * overwrites from earlier objects rather than let
2948                  * tracking information or the free pointer be
2949                  * corrupted if a user writes before the start
2950                  * of the object.
2951                  */
2952                 size += sizeof(void *);
2953 #endif
2954 
2955         /*
2956          * Determine the alignment based on various parameters that the
2957          * user specified and the dynamic determination of cache line size
2958          * on bootup.
2959          */
2960         align = calculate_alignment(flags, align, s->objsize);
2961         s->align = align;
2962 
2963         /*
2964          * SLUB stores one object immediately after another beginning from
2965          * offset 0. In order to align the objects we have to simply size
2966          * each object to conform to the alignment.
2967          */
2968         size = ALIGN(size, align);
2969         s->size = size;
2970         if (forced_order >= 0)
2971                 order = forced_order;
2972         else
2973                 order = calculate_order(size, s->reserved);
2974 
2975         if (order < 0)
2976                 return 0;
2977 
2978         s->allocflags = 0;
2979         if (order)
2980                 s->allocflags |= __GFP_COMP;
2981 
2982         if (s->flags & SLAB_CACHE_DMA)
2983                 s->allocflags |= SLUB_DMA;
2984 
2985         if (s->flags & SLAB_RECLAIM_ACCOUNT)
2986                 s->allocflags |= __GFP_RECLAIMABLE;
2987 
2988         /*
2989          * Determine the number of objects per slab
2990          */
2991         s->oo = oo_make(order, size, s->reserved);
2992         s->min = oo_make(get_order(size), size, s->reserved);
2993         if (oo_objects(s->oo) > oo_objects(s->max))
2994                 s->max = s->oo;
2995 
2996         return !!oo_objects(s->oo);
2997 
2998 }
2999 
3000 static int kmem_cache_open(struct kmem_cache *s,
3001                 const char *name, size_t size,
3002                 size_t align, unsigned long flags,
3003                 void (*ctor)(void *))
3004 {
3005         memset(s, 0, kmem_size);
3006         s->name = name;
3007         s->ctor = ctor;
3008         s->objsize = size;
3009         s->align = align;
3010         s->flags = kmem_cache_flags(size, flags, name, ctor);
3011         s->reserved = 0;
3012 
3013         if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3014                 s->reserved = sizeof(struct rcu_head);
3015 
3016         if (!calculate_sizes(s, -1))
3017                 goto error;
3018         if (disable_higher_order_debug) {
3019                 /*
3020                  * Disable debugging flags that store metadata if the min slab
3021                  * order increased.
3022                  */
3023                 if (get_order(s->size) > get_order(s->objsize)) {
3024                         s->flags &= ~DEBUG_METADATA_FLAGS;
3025                         s->offset = 0;
3026                         if (!calculate_sizes(s, -1))
3027                                 goto error;
3028                 }
3029         }
3030 
3031 #if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3032     defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3033         if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3034                 /* Enable fast mode */
3035                 s->flags |= __CMPXCHG_DOUBLE;
3036 #endif
3037 
3038         /*
3039          * The larger the object size is, the more pages we want on the partial
3040          * list to avoid pounding the page allocator excessively.
3041          */
3042         set_min_partial(s, ilog2(s->size) / 2);
3043 
3044         /*
3045          * cpu_partial determined the maximum number of objects kept in the
3046          * per cpu partial lists of a processor.
3047          *
3048          * Per cpu partial lists mainly contain slabs that just have one
3049          * object freed. If they are used for allocation then they can be
3050          * filled up again with minimal effort. The slab will never hit the
3051          * per node partial lists and therefore no locking will be required.
3052          *
3053          * This setting also determines
3054          *
3055          * A) The number of objects from per cpu partial slabs dumped to the
3056          *    per node list when we reach the limit.
3057          * B) The number of objects in cpu partial slabs to extract from the
3058          *    per node list when we run out of per cpu objects. We only fetch 50%
3059          *    to keep some capacity around for frees.
3060          */
3061         if (kmem_cache_debug(s))
3062                 s->cpu_partial = 0;
3063         else if (s->size >= PAGE_SIZE)
3064                 s->cpu_partial = 2;
3065         else if (s->size >= 1024)
3066                 s->cpu_partial = 6;
3067         else if (s->size >= 256)
3068                 s->cpu_partial = 13;
3069         else
3070                 s->cpu_partial = 30;
3071 
3072         s->refcount = 1;
3073 #ifdef CONFIG_NUMA
3074         s->remote_node_defrag_ratio = 1000;
3075 #endif
3076         if (!init_kmem_cache_nodes(s))
3077                 goto error;
3078 
3079         if (alloc_kmem_cache_cpus(s))
3080                 return 1;
3081 
3082         free_kmem_cache_nodes(s);
3083 error:
3084         if (flags & SLAB_PANIC)
3085                 panic("Cannot create slab %s size=%lu realsize=%u "
3086                         "order=%u offset=%u flags=%lx\n",
3087                         s->name, (unsigned long)size, s->size, oo_order(s->oo),
3088                         s->offset, flags);
3089         return 0;
3090 }
3091 
3092 /*
3093  * Determine the size of a slab object
3094  */
3095 unsigned int kmem_cache_size(struct kmem_cache *s)
3096 {
3097         return s->objsize;
3098 }
3099 EXPORT_SYMBOL(kmem_cache_size);
3100 
3101 static void list_slab_objects(struct kmem_cache *s, struct page *page,
3102                                                         const char *text)
3103 {
3104 #ifdef CONFIG_SLUB_DEBUG
3105         void *addr = page_address(page);
3106         void *p;
3107         unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3108                                      sizeof(long), GFP_ATOMIC);
3109         if (!map)
3110                 return;
3111         slab_err(s, page, "%s", text);
3112         slab_lock(page);
3113 
3114         get_map(s, page, map);
3115         for_each_object(p, s, addr, page->objects) {
3116 
3117                 if (!test_bit(slab_index(p, s, addr), map)) {
3118                         printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
3119                                                         p, p - addr);
3120                         print_tracking(s, p);
3121                 }
3122         }
3123         slab_unlock(page);
3124         kfree(map);
3125 #endif
3126 }
3127 
3128 /*
3129  * Attempt to free all partial slabs on a node.
3130  * This is called from kmem_cache_close(). We must be the last thread
3131  * using the cache and therefore we do not need to lock anymore.
3132  */
3133 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3134 {
3135         struct page *page, *h;
3136 
3137         list_for_each_entry_safe(page, h, &n->partial, lru) {
3138                 if (!page->inuse) {
3139                         remove_partial(n, page);
3140                         discard_slab(s, page);
3141                 } else {
3142                         list_slab_objects(s, page,
3143                                 "Objects remaining on kmem_cache_close()");
3144                 }
3145         }
3146 }
3147 
3148 /*
3149  * Release all resources used by a slab cache.
3150  */
3151 static inline int kmem_cache_close(struct kmem_cache *s)
3152 {
3153         int node;
3154 
3155         flush_all(s);
3156         free_percpu(s->cpu_slab);
3157         /* Attempt to free all objects */
3158         for_each_node_state(node, N_NORMAL_MEMORY) {
3159                 struct kmem_cache_node *n = get_node(s, node);
3160 
3161                 free_partial(s, n);
3162                 if (n->nr_partial || slabs_node(s, node))
3163                         return 1;
3164         }
3165         free_kmem_cache_nodes(s);
3166         return 0;
3167 }
3168 
3169 /*
3170  * Close a cache and release the kmem_cache structure
3171  * (must be used for caches created using kmem_cache_create)
3172  */
3173 void kmem_cache_destroy(struct kmem_cache *s)
3174 {
3175         down_write(&slub_lock);
3176         s->refcount--;
3177         if (!s->refcount) {
3178                 list_del(&s->list);
3179                 up_write(&slub_lock);
3180                 if (kmem_cache_close(s)) {
3181                         printk(KERN_ERR "SLUB %s: %s called for cache that "
3182                                 "still has objects.\n", s->name, __func__);
3183                         dump_stack();
3184                 }
3185                 if (s->flags & SLAB_DESTROY_BY_RCU)
3186                         rcu_barrier();
3187                 sysfs_slab_remove(s);
3188         } else
3189                 up_write(&slub_lock);
3190 }
3191 EXPORT_SYMBOL(kmem_cache_destroy);
3192 
3193 /********************************************************************
3194  *              Kmalloc subsystem
3195  *******************************************************************/
3196 
3197 struct kmem_cache *kmalloc_caches[SLUB_PAGE_SHIFT];
3198 EXPORT_SYMBOL(kmalloc_caches);
3199 
3200 static struct kmem_cache *kmem_cache;
3201 
3202 #ifdef CONFIG_ZONE_DMA
3203 static struct kmem_cache *kmalloc_dma_caches[SLUB_PAGE_SHIFT];
3204 #endif
3205 
3206 static int __init setup_slub_min_order(char *str)
3207 {
3208         get_option(&str, &slub_min_order);
3209 
3210         return 1;
3211 }
3212 
3213 __setup("slub_min_order=", setup_slub_min_order);
3214 
3215 static int __init setup_slub_max_order(char *str)
3216 {
3217         get_option(&str, &slub_max_order);
3218         slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3219 
3220         return 1;
3221 }
3222 
3223 __setup("slub_max_order=", setup_slub_max_order);
3224 
3225 static int __init setup_slub_min_objects(char *str)
3226 {
3227         get_option(&str, &slub_min_objects);
3228 
3229         return 1;
3230 }
3231 
3232 __setup("slub_min_objects=", setup_slub_min_objects);
3233 
3234 static int __init setup_slub_nomerge(char *str)
3235 {
3236         slub_nomerge = 1;
3237         return 1;
3238 }
3239 
3240 __setup("slub_nomerge", setup_slub_nomerge);
3241 
3242 static struct kmem_cache *__init create_kmalloc_cache(const char *name,
3243                                                 int size, unsigned int flags)
3244 {
3245         struct kmem_cache *s;
3246 
3247         s = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3248 
3249         /*
3250          * This function is called with IRQs disabled during early-boot on
3251          * single CPU so there's no need to take slub_lock here.
3252          */
3253         if (!kmem_cache_open(s, name, size, ARCH_KMALLOC_MINALIGN,
3254                                                                 flags, NULL))
3255                 goto panic;
3256 
3257         list_add(&s->list, &slab_caches);
3258         return s;
3259 
3260 panic:
3261         panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
3262         return NULL;
3263 }
3264 
3265 /*
3266  * Conversion table for small slabs sizes / 8 to the index in the
3267  * kmalloc array. This is necessary for slabs < 192 since we have non power
3268  * of two cache sizes there. The size of larger slabs can be determined using
3269  * fls.
3270  */
3271 static s8 size_index[24] = {
3272         3,      /* 8 */
3273         4,      /* 16 */
3274         5,      /* 24 */
3275         5,      /* 32 */
3276         6,      /* 40 */
3277         6,      /* 48 */
3278         6,      /* 56 */
3279         6,      /* 64 */
3280         1,      /* 72 */
3281         1,      /* 80 */
3282         1,      /* 88 */
3283         1,      /* 96 */
3284         7,      /* 104 */
3285         7,      /* 112 */
3286         7,      /* 120 */
3287         7,      /* 128 */
3288         2,      /* 136 */
3289         2,      /* 144 */
3290         2,      /* 152 */
3291         2,      /* 160 */
3292         2,      /* 168 */
3293         2,      /* 176 */
3294         2,      /* 184 */
3295         2       /* 192 */
3296 };
3297 
3298 static inline int size_index_elem(size_t bytes)
3299 {
3300         return (bytes - 1) / 8;
3301 }
3302 
3303 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
3304 {
3305         int index;
3306 
3307         if (size <= 192) {
3308                 if (!size)
3309                         return ZERO_SIZE_PTR;
3310 
3311                 index = size_index[size_index_elem(size)];
3312         } else
3313                 index = fls(size - 1);
3314 
3315 #ifdef CONFIG_ZONE_DMA
3316         if (unlikely((flags & SLUB_DMA)))
3317                 return kmalloc_dma_caches[index];
3318 
3319 #endif
3320         return kmalloc_caches[index];
3321 }
3322 
3323 void *__kmalloc(size_t size, gfp_t flags)
3324 {
3325         struct kmem_cache *s;
3326         void *ret;
3327 
3328         if (unlikely(size > SLUB_MAX_SIZE))
3329                 return kmalloc_large(size, flags);
3330 
3331         s = get_slab(size, flags);
3332 
3333         if (unlikely(ZERO_OR_NULL_PTR(s)))
3334                 return s;
3335 
3336         ret = slab_alloc(s, flags, NUMA_NO_NODE, _RET_IP_);
3337 
3338         trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3339 
3340         return ret;
3341 }
3342 EXPORT_SYMBOL(__kmalloc);
3343 
3344 #ifdef CONFIG_NUMA
3345 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3346 {
3347         struct page *page;
3348         void *ptr = NULL;
3349 
3350         flags |= __GFP_COMP | __GFP_NOTRACK;
3351         page = alloc_pages_node(node, flags, get_order(size));
3352         if (page)
3353                 ptr = page_address(page);
3354 
3355         kmemleak_alloc(ptr, size, 1, flags);
3356         return ptr;
3357 }
3358 
3359 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3360 {
3361         struct kmem_cache *s;
3362         void *ret;
3363 
3364         if (unlikely(size > SLUB_MAX_SIZE)) {
3365                 ret = kmalloc_large_node(size, flags, node);
3366 
3367                 trace_kmalloc_node(_RET_IP_, ret,
3368                                    size, PAGE_SIZE << get_order(size),
3369                                    flags, node);
3370 
3371                 return ret;
3372         }
3373 
3374         s = get_slab(size, flags);
3375 
3376         if (unlikely(ZERO_OR_NULL_PTR(s)))
3377                 return s;
3378 
3379         ret = slab_alloc(s, flags, node, _RET_IP_);
3380 
3381         trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3382 
3383         return ret;
3384 }
3385 EXPORT_SYMBOL(__kmalloc_node);
3386 #endif
3387 
3388 size_t ksize(const void *object)
3389 {
3390         struct page *page;
3391 
3392         if (unlikely(object == ZERO_SIZE_PTR))
3393                 return 0;
3394 
3395         page = virt_to_head_page(object);
3396 
3397         if (unlikely(!PageSlab(page))) {
3398                 WARN_ON(!PageCompound(page));
3399                 return PAGE_SIZE << compound_order(page);
3400         }
3401 
3402         return slab_ksize(page->slab);
3403 }
3404 EXPORT_SYMBOL(ksize);
3405 
3406 #ifdef CONFIG_SLUB_DEBUG
3407 bool verify_mem_not_deleted(const void *x)
3408 {
3409         struct page *page;
3410         void *object = (void *)x;
3411         unsigned long flags;
3412         bool rv;
3413 
3414         if (unlikely(ZERO_OR_NULL_PTR(x)))
3415                 return false;
3416 
3417         local_irq_save(flags);
3418 
3419         page = virt_to_head_page(x);
3420         if (unlikely(!PageSlab(page))) {
3421                 /* maybe it was from stack? */
3422                 rv = true;
3423                 goto out_unlock;
3424         }
3425 
3426         slab_lock(page);
3427         if (on_freelist(page->slab, page, object)) {
3428                 object_err(page->slab, page, object, "Object is on free-list");
3429                 rv = false;
3430         } else {
3431                 rv = true;
3432         }
3433         slab_unlock(page);
3434 
3435 out_unlock:
3436         local_irq_restore(flags);
3437         return rv;
3438 }
3439 EXPORT_SYMBOL(verify_mem_not_deleted);
3440 #endif
3441 
3442 void kfree(const void *x)
3443 {
3444         struct page *page;
3445         void *object = (void *)x;
3446 
3447         trace_kfree(_RET_IP_, x);
3448 
3449         if (unlikely(ZERO_OR_NULL_PTR(x)))
3450                 return;
3451 
3452         page = virt_to_head_page(x);
3453         if (unlikely(!PageSlab(page))) {
3454                 BUG_ON(!PageCompound(page));
3455                 kmemleak_free(x);
3456                 put_page(page);
3457                 return;
3458         }
3459         slab_free(page->slab, page, object, _RET_IP_);
3460 }
3461 EXPORT_SYMBOL(kfree);
3462 
3463 /*
3464  * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3465  * the remaining slabs by the number of items in use. The slabs with the
3466  * most items in use come first. New allocations will then fill those up
3467  * and thus they can be removed from the partial lists.
3468  *
3469  * The slabs with the least items are placed last. This results in them
3470  * being allocated from last increasing the chance that the last objects
3471  * are freed in them.
3472  */
3473 int kmem_cache_shrink(struct kmem_cache *s)
3474 {
3475         int node;
3476         int i;
3477         struct kmem_cache_node *n;
3478         struct page *page;
3479         struct page *t;
3480         int objects = oo_objects(s->max);
3481         struct list_head *slabs_by_inuse =
3482                 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3483         unsigned long flags;
3484 
3485         if (!slabs_by_inuse)
3486                 return -ENOMEM;
3487 
3488         flush_all(s);
3489         for_each_node_state(node, N_NORMAL_MEMORY) {
3490                 n = get_node(s, node);
3491 
3492                 if (!n->nr_partial)
3493                         continue;
3494 
3495                 for (i = 0; i < objects; i++)
3496                         INIT_LIST_HEAD(slabs_by_inuse + i);
3497 
3498                 spin_lock_irqsave(&n->list_lock, flags);
3499 
3500                 /*
3501                  * Build lists indexed by the items in use in each slab.
3502                  *
3503                  * Note that concurrent frees may occur while we hold the
3504                  * list_lock. page->inuse here is the upper limit.
3505                  */
3506                 list_for_each_entry_safe(page, t, &n->partial, lru) {
3507                         list_move(&page->lru, slabs_by_inuse + page->inuse);
3508                         if (!page->inuse)
3509                                 n->nr_partial--;
3510                 }
3511 
3512                 /*
3513                  * Rebuild the partial list with the slabs filled up most
3514                  * first and the least used slabs at the end.
3515                  */
3516                 for (i = objects - 1; i > 0; i--)
3517                         list_splice(slabs_by_inuse + i, n->partial.prev);
3518 
3519                 spin_unlock_irqrestore(&n->list_lock, flags);
3520 
3521                 /* Release empty slabs */
3522                 list_for_each_entry_safe(page, t, slabs_by_inuse, lru)
3523                         discard_slab(s, page);
3524         }
3525 
3526         kfree(slabs_by_inuse);
3527         return 0;
3528 }
3529 EXPORT_SYMBOL(kmem_cache_shrink);
3530 
3531 #if defined(CONFIG_MEMORY_HOTPLUG)
3532 static int slab_mem_going_offline_callback(void *arg)
3533 {
3534         struct kmem_cache *s;
3535 
3536         down_read(&slub_lock);
3537         list_for_each_entry(s, &slab_caches, list)
3538                 kmem_cache_shrink(s);
3539         up_read(&slub_lock);
3540 
3541         return 0;
3542 }
3543 
3544 static void slab_mem_offline_callback(void *arg)
3545 {
3546         struct kmem_cache_node *n;
3547         struct kmem_cache *s;
3548         struct memory_notify *marg = arg;
3549         int offline_node;
3550 
3551         offline_node = marg->status_change_nid;
3552 
3553         /*
3554          * If the node still has available memory. we need kmem_cache_node
3555          * for it yet.
3556          */
3557         if (offline_node < 0)
3558                 return;
3559 
3560         down_read(&slub_lock);
3561         list_for_each_entry(s, &slab_caches, list) {
3562                 n = get_node(s, offline_node);
3563                 if (n) {
3564                         /*
3565                          * if n->nr_slabs > 0, slabs still exist on the node
3566                          * that is going down. We were unable to free them,
3567                          * and offline_pages() function shouldn't call this
3568                          * callback. So, we must fail.
3569                          */
3570                         BUG_ON(slabs_node(s, offline_node));
3571 
3572                         s->node[offline_node] = NULL;
3573                         kmem_cache_free(kmem_cache_node, n);
3574                 }
3575         }
3576         up_read(&slub_lock);
3577 }
3578 
3579 static int slab_mem_going_online_callback(void *arg)
3580 {
3581         struct kmem_cache_node *n;
3582         struct kmem_cache *s;
3583         struct memory_notify *marg = arg;
3584         int nid = marg->status_change_nid;
3585         int ret = 0;
3586 
3587         /*
3588          * If the node's memory is already available, then kmem_cache_node is
3589          * already created. Nothing to do.
3590          */
3591         if (nid < 0)
3592                 return 0;
3593 
3594         /*
3595          * We are bringing a node online. No memory is available yet. We must
3596          * allocate a kmem_cache_node structure in order to bring the node
3597          * online.
3598          */
3599         down_read(&slub_lock);
3600         list_for_each_entry(s, &slab_caches, list) {
3601                 /*
3602                  * XXX: kmem_cache_alloc_node will fallback to other nodes
3603                  *      since memory is not yet available from the node that
3604                  *      is brought up.
3605                  */
3606                 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3607                 if (!n) {
3608                         ret = -ENOMEM;
3609                         goto out;
3610                 }
3611                 init_kmem_cache_node(n, s);
3612                 s->node[nid] = n;
3613         }
3614 out:
3615         up_read(&slub_lock);
3616         return ret;
3617 }
3618 
3619 static int slab_memory_callback(struct notifier_block *self,
3620                                 unsigned long action, void *arg)
3621 {
3622         int ret = 0;
3623 
3624         switch (action) {
3625         case MEM_GOING_ONLINE:
3626                 ret = slab_mem_going_online_callback(arg);
3627                 break;
3628         case MEM_GOING_OFFLINE:
3629                 ret = slab_mem_going_offline_callback(arg);
3630                 break;
3631         case MEM_OFFLINE:
3632         case MEM_CANCEL_ONLINE:
3633                 slab_mem_offline_callback(arg);
3634                 break;
3635         case MEM_ONLINE:
3636         case MEM_CANCEL_OFFLINE:
3637                 break;
3638         }
3639         if (ret)
3640                 ret = notifier_from_errno(ret);
3641         else
3642                 ret = NOTIFY_OK;
3643         return ret;
3644 }
3645 
3646 #endif /* CONFIG_MEMORY_HOTPLUG */
3647 
3648 /********************************************************************
3649  *                      Basic setup of slabs
3650  *******************************************************************/
3651 
3652 /*
3653  * Used for early kmem_cache structures that were allocated using
3654  * the page allocator
3655  */
3656 
3657 static void __init kmem_cache_bootstrap_fixup(struct kmem_cache *s)
3658 {
3659         int node;
3660 
3661         list_add(&s->list, &slab_caches);
3662         s->refcount = -1;
3663 
3664         for_each_node_state(node, N_NORMAL_MEMORY) {
3665                 struct kmem_cache_node *n = get_node(s, node);
3666                 struct page *p;
3667 
3668                 if (n) {
3669                         list_for_each_entry(p, &n->partial, lru)
3670                                 p->slab = s;
3671 
3672 #ifdef CONFIG_SLUB_DEBUG
3673                         list_for_each_entry(p, &n->full, lru)
3674                                 p->slab = s;
3675 #endif
3676                 }
3677         }
3678 }
3679 
3680 void __init kmem_cache_init(void)
3681 {
3682         int i;
3683         int caches = 0;
3684         struct kmem_cache *temp_kmem_cache;
3685         int order;
3686         struct kmem_cache *temp_kmem_cache_node;
3687         unsigned long kmalloc_size;
3688 
3689         if (debug_guardpage_minorder())
3690                 slub_max_order = 0;
3691 
3692         kmem_size = offsetof(struct kmem_cache, node) +
3693                                 nr_node_ids * sizeof(struct kmem_cache_node *);
3694 
3695         /* Allocate two kmem_caches from the page allocator */
3696         kmalloc_size = ALIGN(kmem_size, cache_line_size());
3697         order = get_order(2 * kmalloc_size);
3698         kmem_cache = (void *)__get_free_pages(GFP_NOWAIT, order);
3699 
3700         /*
3701          * Must first have the slab cache available for the allocations of the
3702          * struct kmem_cache_node's. There is special bootstrap code in
3703          * kmem_cache_open for slab_state == DOWN.
3704          */
3705         kmem_cache_node = (void *)kmem_cache + kmalloc_size;
3706 
3707         kmem_cache_open(kmem_cache_node, "kmem_cache_node",
3708                 sizeof(struct kmem_cache_node),
3709                 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3710 
3711         hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3712 
3713         /* Able to allocate the per node structures */
3714         slab_state = PARTIAL;
3715 
3716         temp_kmem_cache = kmem_cache;
3717         kmem_cache_open(kmem_cache, "kmem_cache", kmem_size,
3718                 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3719         kmem_cache = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3720         memcpy(kmem_cache, temp_kmem_cache, kmem_size);
3721 
3722         /*
3723          * Allocate kmem_cache_node properly from the kmem_cache slab.
3724          * kmem_cache_node is separately allocated so no need to
3725          * update any list pointers.
3726          */
3727         temp_kmem_cache_node = kmem_cache_node;
3728 
3729         kmem_cache_node = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3730         memcpy(kmem_cache_node, temp_kmem_cache_node, kmem_size);
3731 
3732         kmem_cache_bootstrap_fixup(kmem_cache_node);
3733 
3734         caches++;
3735         kmem_cache_bootstrap_fixup(kmem_cache);
3736         caches++;
3737         /* Free temporary boot structure */
3738         free_pages((unsigned long)temp_kmem_cache, order);
3739 
3740         /* Now we can use the kmem_cache to allocate kmalloc slabs */
3741 
3742         /*
3743          * Patch up the size_index table if we have strange large alignment
3744          * requirements for the kmalloc array. This is only the case for
3745          * MIPS it seems. The standard arches will not generate any code here.
3746          *
3747          * Largest permitted alignment is 256 bytes due to the way we
3748          * handle the index determination for the smaller caches.
3749          *
3750          * Make sure that nothing crazy happens if someone starts tinkering
3751          * around with ARCH_KMALLOC_MINALIGN
3752          */
3753         BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3754                 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3755 
3756         for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
3757                 int elem = size_index_elem(i);
3758                 if (elem >= ARRAY_SIZE(size_index))
3759                         break;
3760                 size_index[elem] = KMALLOC_SHIFT_LOW;
3761         }
3762 
3763         if (KMALLOC_MIN_SIZE == 64) {
3764                 /*
3765                  * The 96 byte size cache is not used if the alignment
3766                  * is 64 byte.
3767                  */
3768                 for (i = 64 + 8; i <= 96; i += 8)
3769                         size_index[size_index_elem(i)] = 7;
3770         } else if (KMALLOC_MIN_SIZE == 128) {
3771                 /*
3772                  * The 192 byte sized cache is not used if the alignment
3773                  * is 128 byte. Redirect kmalloc to use the 256 byte cache
3774                  * instead.
3775                  */
3776                 for (i = 128 + 8; i <= 192; i += 8)
3777                         size_index[size_index_elem(i)] = 8;
3778         }
3779 
3780         /* Caches that are not of the two-to-the-power-of size */
3781         if (KMALLOC_MIN_SIZE <= 32) {
3782                 kmalloc_caches[1] = create_kmalloc_cache("kmalloc-96", 96, 0);
3783                 caches++;
3784         }
3785 
3786         if (KMALLOC_MIN_SIZE <= 64) {
3787                 kmalloc_caches[2] = create_kmalloc_cache("kmalloc-192", 192, 0);
3788                 caches++;
3789         }
3790 
3791         for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3792                 kmalloc_caches[i] = create_kmalloc_cache("kmalloc", 1 << i, 0);
3793                 caches++;
3794         }
3795 
3796         slab_state = UP;
3797 
3798         /* Provide the correct kmalloc names now that the caches are up */
3799         if (KMALLOC_MIN_SIZE <= 32) {
3800                 kmalloc_caches[1]->name = kstrdup(kmalloc_caches[1]->name, GFP_NOWAIT);
3801                 BUG_ON(!kmalloc_caches[1]->name);
3802         }
3803 
3804         if (KMALLOC_MIN_SIZE <= 64) {
3805                 kmalloc_caches[2]->name = kstrdup(kmalloc_caches[2]->name, GFP_NOWAIT);
3806                 BUG_ON(!kmalloc_caches[2]->name);
3807         }
3808 
3809         for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3810                 char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
3811 
3812                 BUG_ON(!s);
3813                 kmalloc_caches[i]->name = s;
3814         }
3815 
3816 #ifdef CONFIG_SMP
3817         register_cpu_notifier(&slab_notifier);
3818 #endif
3819 
3820 #ifdef CONFIG_ZONE_DMA
3821         for (i = 0; i < SLUB_PAGE_SHIFT; i++) {
3822                 struct kmem_cache *s = kmalloc_caches[i];
3823 
3824                 if (s && s->size) {
3825                         char *name = kasprintf(GFP_NOWAIT,
3826                                  "dma-kmalloc-%d", s->objsize);
3827 
3828                         BUG_ON(!name);
3829                         kmalloc_dma_caches[i] = create_kmalloc_cache(name,
3830                                 s->objsize, SLAB_CACHE_DMA);
3831                 }
3832         }
3833 #endif
3834         printk(KERN_INFO
3835                 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3836                 " CPUs=%d, Nodes=%d\n",
3837                 caches, cache_line_size(),
3838                 slub_min_order, slub_max_order, slub_min_objects,
3839                 nr_cpu_ids, nr_node_ids);
3840 }
3841 
3842 void __init kmem_cache_init_late(void)
3843 {
3844 }
3845 
3846 /*
3847  * Find a mergeable slab cache
3848  */
3849 static int slab_unmergeable(struct kmem_cache *s)
3850 {
3851         if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3852                 return 1;
3853 
3854         if (s->ctor)
3855                 return 1;
3856 
3857         /*
3858          * We may have set a slab to be unmergeable during bootstrap.
3859          */
3860         if (s->refcount < 0)
3861                 return 1;
3862 
3863         return 0;
3864 }
3865 
3866 static struct kmem_cache *find_mergeable(size_t size,
3867                 size_t align, unsigned long flags, const char *name,
3868                 void (*ctor)(void *))
3869 {
3870         struct kmem_cache *s;
3871 
3872         if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3873                 return NULL;
3874 
3875         if (ctor)
3876                 return NULL;
3877 
3878         size = ALIGN(size, sizeof(void *));
3879         align = calculate_alignment(flags, align, size);
3880         size = ALIGN(size, align);
3881         flags = kmem_cache_flags(size, flags, name, NULL);
3882 
3883         list_for_each_entry(s, &slab_caches, list) {
3884                 if (slab_unmergeable(s))
3885                         continue;
3886 
3887                 if (size > s->size)
3888                         continue;
3889 
3890                 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3891                                 continue;
3892                 /*
3893                  * Check if alignment is compatible.
3894                  * Courtesy of Adrian Drzewiecki
3895                  */
3896                 if ((s->size & ~(align - 1)) != s->size)
3897                         continue;
3898 
3899                 if (s->size - size >= sizeof(void *))
3900                         continue;
3901 
3902                 return s;
3903         }
3904         return NULL;
3905 }
3906 
3907 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3908                 size_t align, unsigned long flags, void (*ctor)(void *))
3909 {
3910         struct kmem_cache *s;
3911         char *n;
3912 
3913         if (WARN_ON(!name))
3914                 return NULL;
3915 
3916         down_write(&slub_lock);
3917         s = find_mergeable(size, align, flags, name, ctor);
3918         if (s) {
3919                 s->refcount++;
3920                 /*
3921                  * Adjust the object sizes so that we clear
3922                  * the complete object on kzalloc.
3923                  */
3924                 s->objsize = max(s->objsize, (int)size);
3925                 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3926 
3927                 if (sysfs_slab_alias(s, name)) {
3928                         s->refcount--;
3929                         goto err;
3930                 }
3931                 up_write(&slub_lock);
3932                 return s;
3933         }
3934 
3935         n = kstrdup(name, GFP_KERNEL);
3936         if (!n)
3937                 goto err;
3938 
3939         s = kmalloc(kmem_size, GFP_KERNEL);
3940         if (s) {
3941                 if (kmem_cache_open(s, n,
3942                                 size, align, flags, ctor)) {
3943                         list_add(&s->list, &slab_caches);
3944                         up_write(&slub_lock);
3945                         if (sysfs_slab_add(s)) {
3946                                 down_write(&slub_lock);
3947                                 list_del(&s->list);
3948                                 kfree(n);
3949                                 kfree(s);
3950                                 goto err;
3951                         }
3952                         return s;
3953                 }
3954                 kfree(n);
3955                 kfree(s);
3956         }
3957 err:
3958         up_write(&slub_lock);
3959 
3960         if (flags & SLAB_PANIC)
3961                 panic("Cannot create slabcache %s\n", name);
3962         else
3963                 s = NULL;
3964         return s;
3965 }
3966 EXPORT_SYMBOL(kmem_cache_create);
3967 
3968 #ifdef CONFIG_SMP
3969 /*
3970  * Use the cpu notifier to insure that the cpu slabs are flushed when
3971  * necessary.
3972  */
3973 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3974                 unsigned long action, void *hcpu)
3975 {
3976         long cpu = (long)hcpu;
3977         struct kmem_cache *s;
3978         unsigned long flags;
3979 
3980         switch (action) {
3981         case CPU_UP_CANCELED:
3982         case CPU_UP_CANCELED_FROZEN:
3983         case CPU_DEAD:
3984         case CPU_DEAD_FROZEN:
3985                 down_read(&slub_lock);
3986                 list_for_each_entry(s, &slab_caches, list) {
3987                         local_irq_save(flags);
3988                         __flush_cpu_slab(s, cpu);
3989                         local_irq_restore(flags);
3990                 }
3991                 up_read(&slub_lock);
3992                 break;
3993         default:
3994                 break;
3995         }
3996         return NOTIFY_OK;
3997 }
3998 
3999 static struct notifier_block __cpuinitdata slab_notifier = {
4000         .notifier_call = slab_cpuup_callback
4001 };
4002 
4003 #endif
4004 
4005 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4006 {
4007         struct kmem_cache *s;
4008         void *ret;
4009 
4010         if (unlikely(size > SLUB_MAX_SIZE))
4011                 return kmalloc_large(size, gfpflags);
4012 
4013         s = get_slab(size, gfpflags);
4014 
4015         if (unlikely(ZERO_OR_NULL_PTR(s)))
4016                 return s;
4017 
4018         ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, caller);
4019 
4020         /* Honor the call site pointer we received. */
4021         trace_kmalloc(caller, ret, size, s->size, gfpflags);
4022 
4023         return ret;
4024 }
4025 
4026 #ifdef CONFIG_NUMA
4027 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4028                                         int node, unsigned long caller)
4029 {
4030         struct kmem_cache *s;
4031         void *ret;
4032 
4033         if (unlikely(size > SLUB_MAX_SIZE)) {
4034                 ret = kmalloc_large_node(size, gfpflags, node);
4035 
4036                 trace_kmalloc_node(caller, ret,
4037                                    size, PAGE_SIZE << get_order(size),
4038                                    gfpflags, node);
4039 
4040                 return ret;
4041         }
4042 
4043         s = get_slab(size, gfpflags);
4044 
4045         if (unlikely(ZERO_OR_NULL_PTR(s)))
4046                 return s;
4047 
4048         ret = slab_alloc(s, gfpflags, node, caller);
4049 
4050         /* Honor the call site pointer we received. */
4051         trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4052 
4053         return ret;
4054 }
4055 #endif
4056 
4057 #ifdef CONFIG_SYSFS
4058 static int count_inuse(struct page *page)
4059 {
4060         return page->inuse;
4061 }
4062 
4063 static int count_total(struct page *page)
4064 {
4065         return page->objects;
4066 }
4067 #endif
4068 
4069 #ifdef CONFIG_SLUB_DEBUG
4070 static int validate_slab(struct kmem_cache *s, struct page *page,
4071                                                 unsigned long *map)
4072 {
4073         void *p;
4074         void *addr = page_address(page);
4075 
4076         if (!check_slab(s, page) ||
4077                         !on_freelist(s, page, NULL))
4078                 return 0;
4079 
4080         /* Now we know that a valid freelist exists */
4081         bitmap_zero(map, page->objects);
4082 
4083         get_map(s, page, map);
4084         for_each_object(p, s, addr, page->objects) {
4085                 if (test_bit(slab_index(p, s, addr), map))
4086                         if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4087                                 return 0;
4088         }
4089 
4090         for_each_object(p, s, addr, page->objects)
4091                 if (!test_bit(slab_index(p, s, addr), map))
4092                         if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4093                                 return 0;
4094         return 1;
4095 }
4096 
4097 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4098                                                 unsigned long *map)
4099 {
4100         slab_lock(page);
4101         validate_slab(s, page, map);
4102         slab_unlock(page);
4103 }
4104 
4105 static int validate_slab_node(struct kmem_cache *s,
4106                 struct kmem_cache_node *n, unsigned long *map)
4107 {
4108         unsigned long count = 0;
4109         struct page *page;
4110         unsigned long flags;
4111 
4112         spin_lock_irqsave(&n->list_lock, flags);
4113 
4114         list_for_each_entry(page, &n->partial, lru) {
4115                 validate_slab_slab(s, page, map);
4116                 count++;
4117         }
4118         if (count != n->nr_partial)
4119                 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
4120                         "counter=%ld\n", s->name, count, n->nr_partial);
4121 
4122         if (!(s->flags & SLAB_STORE_USER))
4123                 goto out;
4124 
4125         list_for_each_entry(page, &n->full, lru) {
4126                 validate_slab_slab(s, page, map);
4127                 count++;
4128         }
4129         if (count != atomic_long_read(&n->nr_slabs))
4130                 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
4131                         "counter=%ld\n", s->name, count,
4132                         atomic_long_read(&n->nr_slabs));
4133 
4134 out:
4135         spin_unlock_irqrestore(&n->list_lock, flags);
4136         return count;
4137 }
4138 
4139 static long validate_slab_cache(struct kmem_cache *s)
4140 {
4141         int node;
4142         unsigned long count = 0;
4143         unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4144                                 sizeof(unsigned long), GFP_KERNEL);
4145 
4146         if (!map)
4147                 return -ENOMEM;
4148 
4149         flush_all(s);
4150         for_each_node_state(node, N_NORMAL_MEMORY) {
4151                 struct kmem_cache_node *n = get_node(s, node);
4152 
4153                 count += validate_slab_node(s, n, map);
4154         }
4155         kfree(map);
4156         return count;
4157 }
4158 /*
4159  * Generate lists of code addresses where slabcache objects are allocated
4160  * and freed.
4161  */
4162 
4163 struct location {
4164         unsigned long count;
4165         unsigned long addr;
4166         long long sum_time;
4167         long min_time;
4168         long max_time;
4169         long min_pid;
4170         long max_pid;
4171         DECLARE_BITMAP(cpus, NR_CPUS);
4172         nodemask_t nodes;
4173 };
4174 
4175 struct loc_track {
4176         unsigned long max;
4177         unsigned long count;
4178         struct location *loc;
4179 };
4180 
4181 static void free_loc_track(struct loc_track *t)
4182 {
4183         if (t->max)
4184                 free_pages((unsigned long)t->loc,
4185                         get_order(sizeof(struct location) * t->max));
4186 }
4187 
4188 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4189 {
4190         struct location *l;
4191         int order;
4192 
4193         order = get_order(sizeof(struct location) * max);
4194 
4195         l = (void *)__get_free_pages(flags, order);
4196         if (!l)
4197                 return 0;
4198 
4199         if (t->count) {
4200                 memcpy(l, t->loc, sizeof(struct location) * t->count);
4201                 free_loc_track(t);
4202         }
4203         t->max = max;
4204         t->loc = l;
4205         return 1;
4206 }
4207 
4208 static int add_location(struct loc_track *t, struct kmem_cache *s,
4209                                 const struct track *track)
4210 {
4211         long start, end, pos;
4212         struct location *l;
4213         unsigned long caddr;
4214         unsigned long age = jiffies - track->when;
4215 
4216         start = -1;
4217         end = t->count;
4218 
4219         for ( ; ; ) {
4220                 pos = start + (end - start + 1) / 2;
4221 
4222                 /*
4223                  * There is nothing at "end". If we end up there
4224                  * we need to add something to before end.
4225                  */
4226                 if (pos == end)
4227                         break;
4228 
4229                 caddr = t->loc[pos].addr;
4230                 if (track->addr == caddr) {
4231 
4232                         l = &t->loc[pos];
4233                         l->count++;
4234                         if (track->when) {
4235                                 l->sum_time += age;
4236                                 if (age < l->min_time)
4237                                         l->min_time = age;
4238                                 if (age > l->max_time)
4239                                         l->max_time = age;
4240 
4241                                 if (track->pid < l->min_pid)
4242                                         l->min_pid = track->pid;
4243                                 if (track->pid > l->max_pid)
4244                                         l->max_pid = track->pid;
4245 
4246                                 cpumask_set_cpu(track->cpu,
4247                                                 to_cpumask(l->cpus));
4248                         }
4249                         node_set(page_to_nid(virt_to_page(track)), l->nodes);
4250                         return 1;
4251                 }
4252 
4253                 if (track->addr < caddr)
4254                         end = pos;
4255                 else
4256                         start = pos;
4257         }
4258 
4259         /*
4260          * Not found. Insert new tracking element.
4261          */
4262         if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4263                 return 0;
4264 
4265         l = t->loc + pos;
4266         if (pos < t->count)
4267                 memmove(l + 1, l,
4268                         (t->count - pos) * sizeof(struct location));
4269         t->count++;
4270         l->count = 1;
4271         l->addr = track->addr;
4272         l->sum_time = age;
4273         l->min_time = age;
4274         l->max_time = age;
4275         l->min_pid = track->pid;
4276         l->max_pid = track->pid;
4277         cpumask_clear(to_cpumask(l->cpus));
4278         cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4279         nodes_clear(l->nodes);
4280         node_set(page_to_nid(virt_to_page(track)), l->nodes);
4281         return 1;
4282 }
4283 
4284 static void process_slab(struct loc_track *t, struct kmem_cache *s,
4285                 struct page *page, enum track_item alloc,
4286                 unsigned long *map)
4287 {
4288         void *addr = page_address(page);
4289         void *p;
4290 
4291         bitmap_zero(map, page->objects);
4292         get_map(s, page, map);
4293 
4294         for_each_object(p, s, addr, page->objects)
4295                 if (!test_bit(slab_index(p, s, addr), map))
4296                         add_location(t, s, get_track(s, p, alloc));
4297 }
4298 
4299 static int list_locations(struct kmem_cache *s, char *buf,
4300                                         enum track_item alloc)
4301 {
4302         int len = 0;
4303         unsigned long i;
4304         struct loc_track t = { 0, 0, NULL };
4305         int node;
4306         unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4307                                      sizeof(unsigned long), GFP_KERNEL);
4308 
4309         if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4310                                      GFP_TEMPORARY)) {
4311                 kfree(map);
4312                 return sprintf(buf, "Out of memory\n");
4313         }
4314         /* Push back cpu slabs */
4315         flush_all(s);
4316 
4317         for_each_node_state(node, N_NORMAL_MEMORY) {
4318                 struct kmem_cache_node *n = get_node(s, node);
4319                 unsigned long flags;
4320                 struct page *page;
4321 
4322                 if (!atomic_long_read(&n->nr_slabs))
4323                         continue;
4324 
4325                 spin_lock_irqsave(&n->list_lock, flags);
4326                 list_for_each_entry(page, &n->partial, lru)
4327                         process_slab(&t, s, page, alloc, map);
4328                 list_for_each_entry(page, &n->full, lru)
4329                         process_slab(&t, s, page, alloc, map);
4330                 spin_unlock_irqrestore(&n->list_lock, flags);
4331         }
4332 
4333         for (i = 0; i < t.count; i++) {
4334                 struct location *l = &t.loc[i];
4335 
4336                 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4337                         break;
4338                 len += sprintf(buf + len, "%7ld ", l->count);
4339 
4340                 if (l->addr)
4341                         len += sprintf(buf + len, "%pS", (void *)l->addr);
4342                 else
4343                         len += sprintf(buf + len, "<not-available>");
4344 
4345                 if (l->sum_time != l->min_time) {
4346                         len += sprintf(buf + len, " age=%ld/%ld/%ld",
4347                                 l->min_time,
4348                                 (long)div_u64(l->sum_time, l->count),
4349                                 l->max_time);
4350                 } else
4351                         len += sprintf(buf + len, " age=%ld",
4352                                 l->min_time);
4353 
4354                 if (l->min_pid != l->max_pid)
4355                         len += sprintf(buf + len, " pid=%ld-%ld",
4356                                 l->min_pid, l->max_pid);
4357                 else
4358                         len += sprintf(buf + len, " pid=%ld",
4359                                 l->min_pid);
4360 
4361                 if (num_online_cpus() > 1 &&
4362                                 !cpumask_empty(to_cpumask(l->cpus)) &&
4363                                 len < PAGE_SIZE - 60) {
4364                         len += sprintf(buf + len, " cpus=");
4365                         len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4366                                                  to_cpumask(l->cpus));
4367                 }
4368 
4369                 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4370                                 len < PAGE_SIZE - 60) {
4371                         len += sprintf(buf + len, " nodes=");
4372                         len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4373                                         l->nodes);
4374                 }
4375 
4376                 len += sprintf(buf + len, "\n");
4377         }
4378 
4379         free_loc_track(&t);
4380         kfree(map);
4381         if (!t.count)
4382                 len += sprintf(buf, "No data\n");
4383         return len;
4384 }
4385 #endif
4386 
4387 #ifdef SLUB_RESILIENCY_TEST
4388 static void resiliency_test(void)
4389 {
4390         u8 *p;
4391 
4392         BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || SLUB_PAGE_SHIFT < 10);
4393 
4394         printk(KERN_ERR "SLUB resiliency testing\n");
4395         printk(KERN_ERR "-----------------------\n");
4396         printk(KERN_ERR "A. Corruption after allocation\n");
4397 
4398         p = kzalloc(16, GFP_KERNEL);
4399         p[16] = 0x12;
4400         printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
4401                         " 0x12->0x%p\n\n", p + 16);
4402 
4403         validate_slab_cache(kmalloc_caches[4]);
4404 
4405         /* Hmmm... The next two are dangerous */
4406         p = kzalloc(32, GFP_KERNEL);
4407         p[32 + sizeof(void *)] = 0x34;
4408         printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
4409                         " 0x34 -> -0x%p\n", p);
4410         printk(KERN_ERR
4411                 "If allocated object is overwritten then not detectable\n\n");
4412 
4413         validate_slab_cache(kmalloc_caches[5]);
4414         p = kzalloc(64, GFP_KERNEL);
4415         p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4416         *p = 0x56;
4417         printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4418                                                                         p);
4419         printk(KERN_ERR
4420                 "If allocated object is overwritten then not detectable\n\n");
4421         validate_slab_cache(kmalloc_caches[6]);
4422 
4423         printk(KERN_ERR "\nB. Corruption after free\n");
4424         p = kzalloc(128, GFP_KERNEL);
4425         kfree(p);
4426         *p = 0x78;
4427         printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4428         validate_slab_cache(kmalloc_caches[7]);
4429 
4430         p = kzalloc(256, GFP_KERNEL);
4431         kfree(p);
4432         p[50] = 0x9a;
4433         printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
4434                         p);
4435         validate_slab_cache(kmalloc_caches[8]);
4436 
4437         p = kzalloc(512, GFP_KERNEL);
4438         kfree(p);
4439         p[512] = 0xab;
4440         printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4441         validate_slab_cache(kmalloc_caches[9]);
4442 }
4443 #else
4444 #ifdef CONFIG_SYSFS
4445 static void resiliency_test(void) {};
4446 #endif
4447 #endif
4448 
4449 #ifdef CONFIG_SYSFS
4450 enum slab_stat_type {
4451         SL_ALL,                 /* All slabs */
4452         SL_PARTIAL,             /* Only partially allocated slabs */
4453         SL_CPU,                 /* Only slabs used for cpu caches */
4454         SL_OBJECTS,             /* Determine allocated objects not slabs */
4455         SL_TOTAL                /* Determine object capacity not slabs */
4456 };
4457 
4458 #define SO_ALL          (1 << SL_ALL)
4459 #define SO_PARTIAL      (1 << SL_PARTIAL)
4460 #define SO_CPU          (1 << SL_CPU)
4461 #define SO_OBJECTS      (1 << SL_OBJECTS)
4462 #define SO_TOTAL        (1 << SL_TOTAL)
4463 
4464 static ssize_t show_slab_objects(struct kmem_cache *s,
4465                             char *buf, unsigned long flags)
4466 {
4467         unsigned long total = 0;
4468         int node;
4469         int x;
4470         unsigned long *nodes;
4471         unsigned long *per_cpu;
4472 
4473         nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4474         if (!nodes)
4475                 return -ENOMEM;
4476         per_cpu = nodes + nr_node_ids;
4477 
4478         if (flags & SO_CPU) {
4479                 int cpu;
4480 
4481                 for_each_possible_cpu(cpu) {
4482                         struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
4483                         int node = ACCESS_ONCE(c->node);
4484                         struct page *page;
4485 
4486                         if (node < 0)
4487                                 continue;
4488                         page = ACCESS_ONCE(c->page);
4489                         if (page) {
4490                                 if (flags & SO_TOTAL)
4491                                         x = page->objects;
4492                                 else if (flags & SO_OBJECTS)
4493                                         x = page->inuse;
4494                                 else
4495                                         x = 1;
4496 
4497                                 total += x;
4498                                 nodes[node] += x;
4499                         }
4500                         page = c->partial;
4501 
4502                         if (page) {
4503                                 node = page_to_nid(page);
4504                                 if (flags & SO_TOTAL)
4505                                         WARN_ON_ONCE(1);
4506                                 else if (flags & SO_OBJECTS)
4507                                         WARN_ON_ONCE(1);
4508                                 else
4509                                         x = page->pages;
4510                                 total += x;
4511                                 nodes[node] += x;
4512                         }
4513                         per_cpu[node]++;
4514                 }
4515         }
4516 
4517         lock_memory_hotplug();
4518 #ifdef CONFIG_SLUB_DEBUG
4519         if (flags & SO_ALL) {
4520                 for_each_node_state(node, N_NORMAL_MEMORY) {
4521                         struct kmem_cache_node *n = get_node(s, node);
4522 
4523                 if (flags & SO_TOTAL)
4524                         x = atomic_long_read(&n->total_objects);
4525                 else if (flags & SO_OBJECTS)
4526                         x = atomic_long_read(&n->total_objects) -
4527                                 count_partial(n, count_free);
4528 
4529                         else
4530                                 x = atomic_long_read(&n->nr_slabs);
4531                         total += x;
4532                         nodes[node] += x;
4533                 }
4534 
4535         } else
4536 #endif
4537         if (flags & SO_PARTIAL) {
4538                 for_each_node_state(node, N_NORMAL_MEMORY) {
4539                         struct kmem_cache_node *n = get_node(s, node);
4540 
4541                         if (flags & SO_TOTAL)
4542                                 x = count_partial(n, count_total);
4543                         else if (flags & SO_OBJECTS)
4544                                 x = count_partial(n, count_inuse);
4545                         else
4546                                 x = n->nr_partial;
4547                         total += x;
4548                         nodes[node] += x;
4549                 }
4550         }
4551         x = sprintf(buf, "%lu", total);
4552 #ifdef CONFIG_NUMA
4553         for_each_node_state(node, N_NORMAL_MEMORY)
4554                 if (nodes[node])
4555                         x += sprintf(buf + x, " N%d=%lu",
4556                                         node, nodes[node]);
4557 #endif
4558         unlock_memory_hotplug();
4559         kfree(nodes);
4560         return x + sprintf(buf + x, "\n");
4561 }
4562 
4563 #ifdef CONFIG_SLUB_DEBUG
4564 static int any_slab_objects(struct kmem_cache *s)
4565 {
4566         int node;
4567 
4568         for_each_online_node(node) {
4569                 struct kmem_cache_node *n = get_node(s, node);
4570 
4571                 if (!n)
4572                         continue;
4573 
4574                 if (atomic_long_read(&n->total_objects))
4575                         return 1;
4576         }
4577         return 0;
4578 }
4579 #endif
4580 
4581 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4582 #define to_slab(n) container_of(n, struct kmem_cache, kobj)
4583 
4584 struct slab_attribute {
4585         struct attribute attr;
4586         ssize_t (*show)(struct kmem_cache *s, char *buf);
4587         ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4588 };
4589 
4590 #define SLAB_ATTR_RO(_name) \
4591         static struct slab_attribute _name##_attr = \
4592         __ATTR(_name, 0400, _name##_show, NULL)
4593 
4594 #define SLAB_ATTR(_name) \
4595         static struct slab_attribute _name##_attr =  \
4596         __ATTR(_name, 0600, _name##_show, _name##_store)
4597 
4598 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4599 {
4600         return sprintf(buf, "%d\n", s->size);
4601 }
4602 SLAB_ATTR_RO(slab_size);
4603 
4604 static ssize_t align_show(struct kmem_cache *s, char *buf)
4605 {
4606         return sprintf(buf, "%d\n", s->align);
4607 }
4608 SLAB_ATTR_RO(align);
4609 
4610 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4611 {
4612         return sprintf(buf, "%d\n", s->objsize);
4613 }
4614 SLAB_ATTR_RO(object_size);
4615 
4616 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4617 {
4618         return sprintf(buf, "%d\n", oo_objects(s->oo));
4619 }
4620 SLAB_ATTR_RO(objs_per_slab);
4621 
4622 static ssize_t order_store(struct kmem_cache *s,
4623                                 const char *buf, size_t length)
4624 {
4625         unsigned long order;
4626         int err;
4627 
4628         err = strict_strtoul(buf, 10, &order);
4629         if (err)
4630                 return err;
4631 
4632         if (order > slub_max_order || order < slub_min_order)
4633                 return -EINVAL;
4634 
4635         calculate_sizes(s, order);
4636         return length;
4637 }
4638 
4639 static ssize_t order_show(struct kmem_cache *s, char *buf)
4640 {
4641         return sprintf(buf, "%d\n", oo_order(s->oo));
4642 }
4643 SLAB_ATTR(order);
4644 
4645 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4646 {
4647         return sprintf(buf, "%lu\n", s->min_partial);
4648 }
4649 
4650 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4651                                  size_t length)
4652 {
4653         unsigned long min;
4654         int err;
4655 
4656         err = strict_strtoul(buf, 10, &min);
4657         if (err)
4658                 return err;
4659 
4660         set_min_partial(s, min);
4661         return length;
4662 }
4663 SLAB_ATTR(min_partial);
4664 
4665 static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4666 {
4667         return sprintf(buf, "%u\n", s->cpu_partial);
4668 }
4669 
4670 static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4671                                  size_t length)
4672 {
4673         unsigned long objects;
4674         int err;
4675 
4676         err = strict_strtoul(buf, 10, &objects);
4677         if (err)
4678                 return err;
4679         if (objects && kmem_cache_debug(s))
4680                 return -EINVAL;
4681 
4682         s->cpu_partial = objects;
4683         flush_all(s);
4684         return length;
4685 }
4686 SLAB_ATTR(cpu_partial);
4687 
4688 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4689 {
4690         if (!s->ctor)
4691                 return 0;
4692         return sprintf(buf, "%pS\n", s->ctor);
4693 }
4694 SLAB_ATTR_RO(ctor);
4695 
4696 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4697 {
4698         return sprintf(buf, "%d\n", s->refcount - 1);
4699 }
4700 SLAB_ATTR_RO(aliases);
4701 
4702 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4703 {
4704         return show_slab_objects(s, buf, SO_PARTIAL);
4705 }
4706 SLAB_ATTR_RO(partial);
4707 
4708 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4709 {
4710         return show_slab_objects(s, buf, SO_CPU);
4711 }
4712 SLAB_ATTR_RO(cpu_slabs);
4713 
4714 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4715 {
4716         return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4717 }
4718 SLAB_ATTR_RO(objects);
4719 
4720 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4721 {
4722         return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4723 }
4724 SLAB_ATTR_RO(objects_partial);
4725 
4726 static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4727 {
4728         int objects = 0;
4729         int pages = 0;
4730         int cpu;
4731         int len;
4732 
4733         for_each_online_cpu(cpu) {
4734                 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4735 
4736                 if (page) {
4737                         pages += page->pages;
4738                         objects += page->pobjects;
4739                 }
4740         }
4741 
4742         len = sprintf(buf, "%d(%d)", objects, pages);
4743 
4744 #ifdef CONFIG_SMP
4745         for_each_online_cpu(cpu) {
4746                 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4747 
4748                 if (page && len < PAGE_SIZE - 20)
4749                         len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4750                                 page->pobjects, page->pages);
4751         }
4752 #endif
4753         return len + sprintf(buf + len, "\n");
4754 }
4755 SLAB_ATTR_RO(slabs_cpu_partial);
4756 
4757 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4758 {
4759         return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4760 }
4761 
4762 static ssize_t reclaim_account_store(struct kmem_cache *s,
4763                                 const char *buf, size_t length)
4764 {
4765         s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4766         if (buf[0] == '1')
4767                 s->flags |= SLAB_RECLAIM_ACCOUNT;
4768         return length;
4769 }
4770 SLAB_ATTR(reclaim_account);
4771 
4772 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4773 {
4774         return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4775 }
4776 SLAB_ATTR_RO(hwcache_align);
4777 
4778 #ifdef CONFIG_ZONE_DMA
4779 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4780 {
4781         return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4782 }
4783 SLAB_ATTR_RO(cache_dma);
4784 #endif
4785 
4786 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4787 {
4788         return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4789 }
4790 SLAB_ATTR_RO(destroy_by_rcu);
4791 
4792 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4793 {
4794         return sprintf(buf, "%d\n", s->reserved);
4795 }
4796 SLAB_ATTR_RO(reserved);
4797 
4798 #ifdef CONFIG_SLUB_DEBUG
4799 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4800 {
4801         return show_slab_objects(s, buf, SO_ALL);
4802 }
4803 SLAB_ATTR_RO(slabs);
4804 
4805 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4806 {
4807         return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4808 }
4809 SLAB_ATTR_RO(total_objects);
4810 
4811 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4812 {
4813         return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4814 }
4815 
4816 static ssize_t sanity_checks_store(struct kmem_cache *s,
4817                                 const char *buf, size_t length)
4818 {
4819         s->flags &= ~SLAB_DEBUG_FREE;
4820         if (buf[0] == '1') {
4821                 s->flags &= ~__CMPXCHG_DOUBLE;
4822                 s->flags |= SLAB_DEBUG_FREE;
4823         }
4824         return length;
4825 }
4826 SLAB_ATTR(sanity_checks);
4827 
4828 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4829 {
4830         return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4831 }
4832 
4833 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4834                                                         size_t length)
4835 {
4836         s->flags &= ~SLAB_TRACE;
4837         if (buf[0] == '1') {
4838                 s->flags &= ~__CMPXCHG_DOUBLE;
4839                 s->flags |= SLAB_TRACE;
4840         }
4841         return length;
4842 }
4843 SLAB_ATTR(trace);
4844 
4845 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4846 {
4847         return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4848 }
4849 
4850 static ssize_t red_zone_store(struct kmem_cache *s,
4851                                 const char *buf, size_t length)
4852 {
4853         if (any_slab_objects(s))
4854                 return -EBUSY;
4855 
4856         s->flags &= ~SLAB_RED_ZONE;
4857         if (buf[0] == '1') {
4858                 s->flags &= ~__CMPXCHG_DOUBLE;
4859                 s->flags |= SLAB_RED_ZONE;
4860         }
4861         calculate_sizes(s, -1);
4862         return length;
4863 }
4864 SLAB_ATTR(red_zone);
4865 
4866 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4867 {
4868         return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4869 }
4870 
4871 static ssize_t poison_store(struct kmem_cache *s,
4872                                 const char *buf, size_t length)
4873 {
4874         if (any_slab_objects(s))
4875                 return -EBUSY;
4876 
4877         s->flags &= ~SLAB_POISON;
4878         if (buf[0] == '1') {
4879                 s->flags &= ~__CMPXCHG_DOUBLE;
4880                 s->flags |= SLAB_POISON;
4881         }
4882         calculate_sizes(s, -1);
4883         return length;
4884 }
4885 SLAB_ATTR(poison);
4886 
4887 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4888 {
4889         return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4890 }
4891 
4892 static ssize_t store_user_store(struct kmem_cache *s,
4893                                 const char *buf, size_t length)
4894 {
4895         if (any_slab_objects(s))
4896                 return -EBUSY;
4897 
4898         s->flags &= ~SLAB_STORE_USER;
4899         if (buf[0] == '1') {
4900                 s->flags &= ~__CMPXCHG_DOUBLE;
4901                 s->flags |= SLAB_STORE_USER;
4902         }
4903         calculate_sizes(s, -1);
4904         return length;
4905 }
4906 SLAB_ATTR(store_user);
4907 
4908 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4909 {
4910         return 0;
4911 }
4912 
4913 static ssize_t validate_store(struct kmem_cache *s,
4914                         const char *buf, size_t length)
4915 {
4916         int ret = -EINVAL;
4917 
4918         if (buf[0] == '1') {
4919                 ret = validate_slab_cache(s);
4920                 if (ret >= 0)
4921                         ret = length;
4922         }
4923         return ret;
4924 }
4925 SLAB_ATTR(validate);
4926 
4927 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4928 {
4929         if (!(s->flags & SLAB_STORE_USER))
4930                 return -ENOSYS;
4931         return list_locations(s, buf, TRACK_ALLOC);
4932 }
4933 SLAB_ATTR_RO(alloc_calls);
4934 
4935 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4936 {
4937         if (!(s->flags & SLAB_STORE_USER))
4938                 return -ENOSYS;
4939         return list_locations(s, buf, TRACK_FREE);
4940 }
4941 SLAB_ATTR_RO(free_calls);
4942 #endif /* CONFIG_SLUB_DEBUG */
4943 
4944 #ifdef CONFIG_FAILSLAB
4945 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4946 {
4947         return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4948 }
4949 
4950 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4951                                                         size_t length)
4952 {
4953         s->flags &= ~SLAB_FAILSLAB;
4954         if (buf[0] == '1')
4955                 s->flags |= SLAB_FAILSLAB;
4956         return length;
4957 }
4958 SLAB_ATTR(failslab);
4959 #endif
4960 
4961 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4962 {
4963         return 0;
4964 }
4965 
4966 static ssize_t shrink_store(struct kmem_cache *s,
4967                         const char *buf, size_t length)
4968 {
4969         if (buf[0] == '1') {
4970                 int rc = kmem_cache_shrink(s);
4971 
4972                 if (rc)
4973                         return rc;
4974         } else
4975                 return -EINVAL;
4976         return length;
4977 }
4978 SLAB_ATTR(shrink);
4979 
4980 #ifdef CONFIG_NUMA
4981 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4982 {
4983         return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4984 }
4985 
4986 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4987                                 const char *buf, size_t length)
4988 {
4989         unsigned long ratio;
4990         int err;
4991 
4992         err = strict_strtoul(buf, 10, &ratio);
4993         if (err)
4994                 return err;
4995 
4996         if (ratio <= 100)
4997                 s->remote_node_defrag_ratio = ratio * 10;
4998 
4999         return length;
5000 }
5001 SLAB_ATTR(remote_node_defrag_ratio);
5002 #endif
5003 
5004 #ifdef CONFIG_SLUB_STATS
5005 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5006 {
5007         unsigned long sum  = 0;
5008         int cpu;
5009         int len;
5010         int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
5011 
5012         if (!data)
5013                 return -ENOMEM;
5014 
5015         for_each_online_cpu(cpu) {
5016                 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5017 
5018                 data[cpu] = x;
5019                 sum += x;
5020         }
5021 
5022         len = sprintf(buf, "%lu", sum);
5023 
5024 #ifdef CONFIG_SMP
5025         for_each_online_cpu(cpu) {
5026                 if (data[cpu] && len < PAGE_SIZE - 20)
5027                         len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5028         }
5029 #endif
5030         kfree(data);
5031         return len + sprintf(buf + len, "\n");
5032 }
5033 
5034 static void clear_stat(struct kmem_cache *s, enum stat_item si)
5035 {
5036         int cpu;
5037 
5038         for_each_online_cpu(cpu)
5039                 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5040 }
5041 
5042 #define STAT_ATTR(si, text)                                     \
5043 static ssize_t text##_show(struct kmem_cache *s, char *buf)     \
5044 {                                                               \
5045         return show_stat(s, buf, si);                           \
5046 }                                                               \
5047 static ssize_t text##_store(struct kmem_cache *s,               \
5048                                 const char *buf, size_t length) \
5049 {                                                               \
5050         if (buf[0] != '')                                      \
5051                 return -EINVAL;                                 \
5052         clear_stat(s, si);                                      \
5053         return length;                                          \
5054 }                                                               \
5055 SLAB_ATTR(text);                                                \
5056 
5057 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5058 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5059 STAT_ATTR(FREE_FASTPATH, free_fastpath);
5060 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5061 STAT_ATTR(FREE_FROZEN, free_frozen);
5062 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5063 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5064 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5065 STAT_ATTR(ALLOC_SLAB, alloc_slab);
5066 STAT_ATTR(ALLOC_REFILL, alloc_refill);
5067 STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5068 STAT_ATTR(FREE_SLAB, free_slab);
5069 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5070 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5071 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5072 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5073 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5074 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5075 STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5076 STAT_ATTR(ORDER_FALLBACK, order_fallback);
5077 STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5078 STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5079 STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5080 STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5081 STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5082 STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5083 #endif
5084 
5085 static struct attribute *slab_attrs[] = {
5086         &slab_size_attr.attr,
5087         &object_size_attr.attr,
5088         &objs_per_slab_attr.attr,
5089         &order_attr.attr,
5090         &min_partial_attr.attr,
5091         &cpu_partial_attr.attr,
5092         &objects_attr.attr,
5093         &objects_partial_attr.attr,
5094         &partial_attr.attr,
5095         &cpu_slabs_attr.attr,
5096         &ctor_attr.attr,
5097         &aliases_attr.attr,
5098         &align_attr.attr,
5099         &hwcache_align_attr.attr,
5100         &reclaim_account_attr.attr,
5101         &destroy_by_rcu_attr.attr,
5102         &shrink_attr.attr,
5103         &reserved_attr.attr,
5104         &slabs_cpu_partial_attr.attr,
5105 #ifdef CONFIG_SLUB_DEBUG
5106         &total_objects_attr.attr,
5107         &slabs_attr.attr,
5108         &sanity_checks_attr.attr,
5109         &trace_attr.attr,
5110         &red_zone_attr.attr,
5111         &poison_attr.attr,
5112         &store_user_attr.attr,
5113         &validate_attr.attr,
5114         &alloc_calls_attr.attr,
5115         &free_calls_attr.attr,
5116 #endif
5117 #ifdef CONFIG_ZONE_DMA
5118         &cache_dma_attr.attr,
5119 #endif
5120 #ifdef CONFIG_NUMA
5121         &remote_node_defrag_ratio_attr.attr,
5122 #endif
5123 #ifdef CONFIG_SLUB_STATS
5124         &alloc_fastpath_attr.attr,
5125         &alloc_slowpath_attr.attr,
5126         &free_fastpath_attr.attr,
5127         &free_slowpath_attr.attr,
5128         &free_frozen_attr.attr,
5129         &free_add_partial_attr.attr,
5130         &free_remove_partial_attr.attr,
5131         &alloc_from_partial_attr.attr,
5132         &alloc_slab_attr.attr,
5133         &alloc_refill_attr.attr,
5134         &alloc_node_mismatch_attr.attr,
5135         &free_slab_attr.attr,
5136         &cpuslab_flush_attr.attr,
5137         &deactivate_full_attr.attr,
5138         &deactivate_empty_attr.attr,
5139         &deactivate_to_head_attr.attr,
5140         &deactivate_to_tail_attr.attr,
5141         &deactivate_remote_frees_attr.attr,
5142         &deactivate_bypass_attr.attr,
5143         &order_fallback_attr.attr,
5144         &cmpxchg_double_fail_attr.attr,
5145         &cmpxchg_double_cpu_fail_attr.attr,
5146         &cpu_partial_alloc_attr.attr,
5147         &cpu_partial_free_attr.attr,
5148         &cpu_partial_node_attr.attr,
5149         &cpu_partial_drain_attr.attr,
5150 #endif
5151 #ifdef CONFIG_FAILSLAB
5152         &failslab_attr.attr,
5153 #endif
5154 
5155         NULL
5156 };
5157 
5158 static struct attribute_group slab_attr_group = {
5159         .attrs = slab_attrs,
5160 };
5161 
5162 static ssize_t slab_attr_show(struct kobject *kobj,
5163                                 struct attribute *attr,
5164                                 char *buf)
5165 {
5166         struct slab_attribute *attribute;
5167         struct kmem_cache *s;
5168         int err;
5169 
5170         attribute = to_slab_attr(attr);
5171         s = to_slab(kobj);
5172 
5173         if (!attribute->show)
5174                 return -EIO;
5175 
5176         err = attribute->show(s, buf);
5177 
5178         return err;
5179 }
5180 
5181 static ssize_t slab_attr_store(struct kobject *kobj,
5182                                 struct attribute *attr,
5183                                 const char *buf, size_t len)
5184 {
5185         struct slab_attribute *attribute;
5186         struct kmem_cache *s;
5187         int err;
5188 
5189         attribute = to_slab_attr(attr);
5190         s = to_slab(kobj);
5191 
5192         if (!attribute->store)
5193                 return -EIO;
5194 
5195         err = attribute->store(s, buf, len);
5196 
5197         return err;
5198 }
5199 
5200 static void kmem_cache_release(struct kobject *kobj)
5201 {
5202         struct kmem_cache *s = to_slab(kobj);
5203 
5204         kfree(s->name);
5205         kfree(s);
5206 }
5207 
5208 static const struct sysfs_ops slab_sysfs_ops = {
5209         .show = slab_attr_show,
5210         .store = slab_attr_store,
5211 };
5212 
5213 static struct kobj_type slab_ktype = {
5214         .sysfs_ops = &slab_sysfs_ops,
5215         .release = kmem_cache_release
5216 };
5217 
5218 static int uevent_filter(struct kset *kset, struct kobject *kobj)
5219 {
5220         struct kobj_type *ktype = get_ktype(kobj);
5221 
5222         if (ktype == &slab_ktype)
5223                 return 1;
5224         return 0;
5225 }
5226 
5227 static const struct kset_uevent_ops slab_uevent_ops = {
5228         .filter = uevent_filter,
5229 };
5230 
5231 static struct kset *slab_kset;
5232 
5233 #define ID_STR_LENGTH 64
5234 
5235 /* Create a unique string id for a slab cache:
5236  *
5237  * Format       :[flags-]size
5238  */
5239 static char *create_unique_id(struct kmem_cache *s)
5240 {
5241         char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5242         char *p = name;
5243 
5244         BUG_ON(!name);
5245 
5246         *p++ = ':';
5247         /*
5248          * First flags affecting slabcache operations. We will only
5249          * get here for aliasable slabs so we do not need to support
5250          * too many flags. The flags here must cover all flags that
5251          * are matched during merging to guarantee that the id is
5252          * unique.
5253          */
5254         if (s->flags & SLAB_CACHE_DMA)
5255                 *p++ = 'd';
5256         if (s->flags & SLAB_RECLAIM_ACCOUNT)
5257                 *p++ = 'a';
5258         if (s->flags & SLAB_DEBUG_FREE)
5259                 *p++ = 'F';
5260         if (!(s->flags & SLAB_NOTRACK))
5261                 *p++ = 't';
5262         if (p != name + 1)
5263                 *p++ = '-';
5264         p += sprintf(p, "%07d", s->size);
5265         BUG_ON(p > name + ID_STR_LENGTH - 1);
5266         return name;
5267 }
5268 
5269 static int sysfs_slab_add(struct kmem_cache *s)
5270 {
5271         int err;
5272         const char *name;
5273         int unmergeable;
5274 
5275         if (slab_state < SYSFS)
5276                 /* Defer until later */
5277                 return 0;
5278 
5279         unmergeable = slab_unmergeable(s);
5280         if (unmergeable) {
5281                 /*
5282                  * Slabcache can never be merged so we can use the name proper.
5283                  * This is typically the case for debug situations. In that
5284                  * case we can catch duplicate names easily.
5285                  */
5286                 sysfs_remove_link(&slab_kset->kobj, s->name);
5287                 name = s->name;
5288         } else {
5289                 /*
5290                  * Create a unique name for the slab as a target
5291                  * for the symlinks.
5292                  */
5293                 name = create_unique_id(s);
5294         }
5295 
5296         s->kobj.kset = slab_kset;
5297         err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
5298         if (err) {
5299                 kobject_put(&s->kobj);
5300                 return err;
5301         }
5302 
5303         err = sysfs_create_group(&s->kobj, &slab_attr_group);
5304         if (err) {
5305                 kobject_del(&s->kobj);
5306                 kobject_put(&s->kobj);
5307                 return err;
5308         }
5309         kobject_uevent(&s->kobj, KOBJ_ADD);
5310         if (!unmergeable) {
5311                 /* Setup first alias */
5312                 sysfs_slab_alias(s, s->name);
5313                 kfree(name);
5314         }
5315         return 0;
5316 }
5317 
5318 static void sysfs_slab_remove(struct kmem_cache *s)
5319 {
5320         if (slab_state < SYSFS)
5321                 /*
5322                  * Sysfs has not been setup yet so no need to remove the
5323                  * cache from sysfs.
5324                  */
5325                 return;
5326 
5327         kobject_uevent(&s->kobj, KOBJ_REMOVE);
5328         kobject_del(&s->kobj);
5329         kobject_put(&s->kobj);
5330 }
5331 
5332 /*
5333  * Need to buffer aliases during bootup until sysfs becomes
5334  * available lest we lose that information.
5335  */
5336 struct saved_alias {
5337         struct kmem_cache *s;
5338         const char *name;
5339         struct saved_alias *next;
5340 };
5341 
5342 static struct saved_alias *alias_list;
5343 
5344 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5345 {
5346         struct saved_alias *al;
5347 
5348         if (slab_state == SYSFS) {
5349                 /*
5350                  * If we have a leftover link then remove it.
5351                  */
5352                 sysfs_remove_link(&slab_kset->kobj, name);
5353                 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5354         }
5355 
5356         al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5357         if (!al)
5358                 return -ENOMEM;
5359 
5360         al->s = s;
5361         al->name = name;
5362         al->next = alias_list;
5363         alias_list = al;
5364         return 0;
5365 }
5366 
5367 static int __init slab_sysfs_init(void)
5368 {
5369         struct kmem_cache *s;
5370         int err;
5371 
5372         down_write(&slub_lock);
5373 
5374         slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5375         if (!slab_kset) {
5376                 up_write(&slub_lock);
5377                 printk(KERN_ERR "Cannot register slab subsystem.\n");
5378                 return -ENOSYS;
5379         }
5380 
5381         slab_state = SYSFS;
5382 
5383         list_for_each_entry(s, &slab_caches, list) {
5384                 err = sysfs_slab_add(s);
5385                 if (err)
5386                         printk(KERN_ERR "SLUB: Unable to add boot slab %s"
5387                                                 " to sysfs\n", s->name);
5388         }
5389 
5390         while (alias_list) {
5391                 struct saved_alias *al = alias_list;
5392 
5393                 alias_list = alias_list->next;
5394                 err = sysfs_slab_alias(al->s, al->name);
5395                 if (err)
5396                         printk(KERN_ERR "SLUB: Unable to add boot slab alias"
5397                                         " %s to sysfs\n", s->name);
5398                 kfree(al);
5399         }
5400 
5401         up_write(&slub_lock);
5402         resiliency_test();
5403         return 0;
5404 }
5405 
5406 __initcall(slab_sysfs_init);
5407 #endif /* CONFIG_SYSFS */
5408 
5409 /*
5410  * The /proc/slabinfo ABI
5411  */
5412 #ifdef CONFIG_SLABINFO
5413 static void print_slabinfo_header(struct seq_file *m)
5414 {
5415         seq_puts(m, "slabinfo - version: 2.1\n");
5416         seq_puts(m, "# name            <active_objs> <num_objs> <objsize> "
5417                  "<objperslab> <pagesperslab>");
5418         seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
5419         seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
5420         seq_putc(m, '\n');
5421 }
5422 
5423 static void *s_start(struct seq_file *m, loff_t *pos)
5424 {
5425         loff_t n = *pos;
5426 
5427         down_read(&slub_lock);
5428         if (!n)
5429                 print_slabinfo_header(m);
5430 
5431         return seq_list_start(&slab_caches, *pos);
5432 }
5433 
5434 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
5435 {
5436         return seq_list_next(p, &slab_caches, pos);
5437 }
5438 
5439 static void s_stop(struct seq_file *m, void *p)
5440 {
5441         up_read(&slub_lock);
5442 }
5443 
5444 static int s_show(struct seq_file *m, void *p)
5445 {
5446         unsigned long nr_partials = 0;
5447         unsigned long nr_slabs = 0;
5448         unsigned long nr_inuse = 0;
5449         unsigned long nr_objs = 0;
5450         unsigned long nr_free = 0;
5451         struct kmem_cache *s;
5452         int node;
5453 
5454         s = list_entry(p, struct kmem_cache, list);
5455 
5456         for_each_online_node(node) {
5457                 struct kmem_cache_node *n = get_node(s, node);
5458 
5459                 if (!n)
5460                         continue;
5461 
5462                 nr_partials += n->nr_partial;
5463                 nr_slabs += atomic_long_read(&n->nr_slabs);
5464                 nr_objs += atomic_long_read(&n->total_objects);
5465                 nr_free += count_partial(n, count_free);
5466         }
5467 
5468         nr_inuse = nr_objs - nr_free;
5469 
5470         seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
5471                    nr_objs, s->size, oo_objects(s->oo),
5472                    (1 << oo_order(s->oo)));
5473         seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
5474         seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
5475                    0UL);
5476         seq_putc(m, '\n');
5477         return 0;
5478 }
5479 
5480 static const struct seq_operations slabinfo_op = {
5481         .start = s_start,
5482         .next = s_next,
5483         .stop = s_stop,
5484         .show = s_show,
5485 };
5486 
5487 static int slabinfo_open(struct inode *inode, struct file *file)
5488 {
5489         return seq_open(file, &slabinfo_op);
5490 }
5491 
5492 static const struct file_operations proc_slabinfo_operations = {
5493         .open           = slabinfo_open,
5494         .read           = seq_read,
5495         .llseek         = seq_lseek,
5496         .release        = seq_release,
5497 };
5498 
5499 static int __init slab_proc_init(void)
5500 {
5501         proc_create("slabinfo", S_IRUSR, NULL, &proc_slabinfo_operations);
5502         return 0;
5503 }
5504 module_init(slab_proc_init);
5505 #endif /* CONFIG_SLABINFO */
5506 

~ [ 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