mm/slob.c

   1 /*
   2  * SLOB Allocator: Simple List Of Blocks
   3  *
   4  * Matt Mackall <mpm@selenic.com> 12/30/03
   5  *
   6  * NUMA support by Paul Mundt, 2007.
   7  *
   8  * How SLOB works:
   9  *
  10  * The core of SLOB is a traditional K&R style heap allocator, with
  11  * support for returning aligned objects. The granularity of this
  12  * allocator is as little as 2 bytes, however typically most architectures
  13  * will require 4 bytes on 32-bit and 8 bytes on 64-bit.
  14  *
  15  * The slob heap is a set of linked list of pages from alloc_pages(),
  16  * and within each page, there is a singly-linked list of free blocks
  17  * (slob_t). The heap is grown on demand. To reduce fragmentation,
  18  * heap pages are segregated into three lists, with objects less than
  19  * 256 bytes, objects less than 1024 bytes, and all other objects.
  20  *
  21  * Allocation from heap involves first searching for a page with
  22  * sufficient free blocks (using a next-fit-like approach) followed by
  23  * a first-fit scan of the page. Deallocation inserts objects back
  24  * into the free list in address order, so this is effectively an
  25  * address-ordered first fit.
  26  *
  27  * Above this is an implementation of kmalloc/kfree. Blocks returned
  28  * from kmalloc are prepended with a 4-byte header with the kmalloc size.
  29  * If kmalloc is asked for objects of PAGE_SIZE or larger, it calls
  30  * alloc_pages() directly, allocating compound pages so the page order
  31  * does not have to be separately tracked, and also stores the exact
  32  * allocation size in page->private so that it can be used to accurately
  33  * provide ksize(). These objects are detected in kfree() because slob_page()
  34  * is false for them.
  35  *
  36  * SLAB is emulated on top of SLOB by simply calling constructors and
  37  * destructors for every SLAB allocation. Objects are returned with the
  38  * 4-byte alignment unless the SLAB_HWCACHE_ALIGN flag is set, in which
  39  * case the low-level allocator will fragment blocks to create the proper
  40  * alignment. Again, objects of page-size or greater are allocated by
  41  * calling alloc_pages(). As SLAB objects know their size, no separate
  42  * size bookkeeping is necessary and there is essentially no allocation
  43  * space overhead, and compound pages aren't needed for multi-page
  44  * allocations.
  45  *
  46  * NUMA support in SLOB is fairly simplistic, pushing most of the real
  47  * logic down to the page allocator, and simply doing the node accounting
  48  * on the upper levels. In the event that a node id is explicitly
  49  * provided, alloc_pages_exact_node() with the specified node id is used
  50  * instead. The common case (or when the node id isn't explicitly provided)
  51  * will default to the current node, as per numa_node_id().
  52  *
  53  * Node aware pages are still inserted in to the global freelist, and
  54  * these are scanned for by matching against the node id encoded in the
  55  * page flags. As a result, block allocations that can be satisfied from
  56  * the freelist will only be done so on pages residing on the same node,
  57  * in order to prevent random node placement.
  58  */
  59
  60 #include <linux/kernel.h>
  61 #include <linux/slab.h>
  62 #include <linux/mm.h>
  63 #include <linux/swap.h> /* struct reclaim_state */
  64 #include <linux/cache.h>
  65 #include <linux/init.h>
  66 #include <linux/export.h>
  67 #include <linux/rcupdate.h>
  68 #include <linux/list.h>
  69 #include <linux/kmemleak.h>
  70
  71 #include <trace/events/kmem.h>
  72
  73 #include <linux/atomic.h>
  74
  75 /*
  76  * slob_block has a field 'units', which indicates size of block if +ve,
  77  * or offset of next block if -ve (in SLOB_UNITs).
  78  *
  79  * Free blocks of size 1 unit simply contain the offset of the next block.
  80  * Those with larger size contain their size in the first SLOB_UNIT of
  81  * memory, and the offset of the next free block in the second SLOB_UNIT.
  82  */
  83 #if PAGE_SIZE <= (32767 * 2)
  84 typedef s16 slobidx_t;
  85 #else
  86 typedef s32 slobidx_t;
  87 #endif
  88
  89 struct slob_block {
  90         slobidx_t units;
  91 };
  92 typedef struct slob_block slob_t;
  93
  94 /*
  95  * free_slob_page: call before a slob_page is returned to the page allocator.
  96  */
  97 static inline void free_slob_page(struct page *sp)
  98 {
  99         reset_page_mapcount(sp);
 100         sp->mapping = NULL;
 101 }
 102
 103 /*
 104  * All partially free slob pages go on these lists.
 105  */
 106 #define SLOB_BREAK1 256
 107 #define SLOB_BREAK2 1024
 108 static LIST_HEAD(free_slob_small);
 109 static LIST_HEAD(free_slob_medium);
 110 static LIST_HEAD(free_slob_large);
 111
 112 /*
 113  * is_slob_page: True for all slob pages (false for bigblock pages)
 114  */
 115 static inline int is_slob_page(struct page *sp)
 116 {
 117         return PageSlab(sp);
 118 }
 119
 120 static inline void set_slob_page(struct page *sp)
 121 {
 122         __SetPageSlab(sp);
 123 }
 124
 125 static inline void clear_slob_page(struct page *sp)
 126 {
 127         __ClearPageSlab(sp);
 128 }
 129
 130 static inline struct page *slob_page(const void *addr)
 131 {
 132         return virt_to_page(addr);
 133 }
 134
 135 /*
 136  * slob_page_free: true for pages on free_slob_pages list.
 137  */
 138 static inline int slob_page_free(struct page *sp)
 139 {
 140         return PageSlobFree(sp);
 141 }
 142
 143 static void set_slob_page_free(struct page *sp, struct list_head *list)
 144 {
 145         list_add(&sp->list, list);
 146         __SetPageSlobFree(sp);
 147 }
 148
 149 static inline void clear_slob_page_free(struct page *sp)
 150 {
 151         list_del(&sp->list);
 152         __ClearPageSlobFree(sp);
 153 }
 154
 155 #define SLOB_UNIT sizeof(slob_t)
 156 #define SLOB_UNITS(size) (((size) + SLOB_UNIT - 1)/SLOB_UNIT)
 157 #define SLOB_ALIGN L1_CACHE_BYTES
 158
 159 /*
 160  * struct slob_rcu is inserted at the tail of allocated slob blocks, which
 161  * were created with a SLAB_DESTROY_BY_RCU slab. slob_rcu is used to free
 162  * the block using call_rcu.
 163  */
 164 struct slob_rcu {
 165         struct rcu_head head;
 166         int size;
 167 };
 168
 169 /*
 170  * slob_lock protects all slob allocator structures.
 171  */
 172 static DEFINE_SPINLOCK(slob_lock);
 173
 174 /*
 175  * Encode the given size and next info into a free slob block s.
 176  */
 177 static void set_slob(slob_t *s, slobidx_t size, slob_t *next)
 178 {
 179         slob_t *base = (slob_t *)((unsigned long)s & PAGE_MASK);
 180         slobidx_t offset = next - base;
 181
 182         if (size > 1) {
 183                 s[0].units = size;
 184                 s[1].units = offset;
 185         } else
 186                 s[0].units = -offset;
 187 }
 188
 189 /*
 190  * Return the size of a slob block.
 191  */
 192 static slobidx_t slob_units(slob_t *s)
 193 {
 194         if (s->units > 0)
 195                 return s->units;
 196         return 1;
 197 }
 198
 199 /*
 200  * Return the next free slob block pointer after this one.
 201  */
 202 static slob_t *slob_next(slob_t *s)
 203 {
 204         slob_t *base = (slob_t *)((unsigned long)s & PAGE_MASK);
 205         slobidx_t next;
 206
 207         if (s[0].units < 0)
 208                 next = -s[0].units;
 209         else
 210                 next = s[1].units;
 211         return base+next;
 212 }
 213
 214 /*
 215  * Returns true if s is the last free block in its page.
 216  */
 217 static int slob_last(slob_t *s)
 218 {
 219         return !((unsigned long)slob_next(s) & ~PAGE_MASK);
 220 }
 221
 222 static void *slob_new_pages(gfp_t gfp, int order, int node)
 223 {
 224         void *page;
 225
 226 #ifdef CONFIG_NUMA
 227         if (node != -1)
 228                 page = alloc_pages_exact_node(node, gfp, order);
 229         else
 230 #endif
 231                 page = alloc_pages(gfp, order);
 232
 233         if (!page)
 234                 return NULL;
 235
 236         return page_address(page);
 237 }
 238
 239 static void slob_free_pages(void *b, int order)
 240 {
 241         if (current->reclaim_state)
 242                 current->reclaim_state->reclaimed_slab += 1 << order;
 243         free_pages((unsigned long)b, order);
 244 }
 245
 246 /*
 247  * Allocate a slob block within a given slob_page sp.
 248  */
 249 static void *slob_page_alloc(struct page *sp, size_t size, int align)
 250 {
 251         slob_t *prev, *cur, *aligned = NULL;
 252         int delta = 0, units = SLOB_UNITS(size);
 253
 254         for (prev = NULL, cur = sp->freelist; ; prev = cur, cur = slob_next(cur)) {
 255                 slobidx_t avail = slob_units(cur);
 256
 257                 if (align) {
 258                         aligned = (slob_t *)ALIGN((unsigned long)cur, align);
 259                         delta = aligned - cur;
 260                 }
 261                 if (avail >= units + delta) { /* room enough? */
 262                         slob_t *next;
 263
 264                         if (delta) { /* need to fragment head to align? */
 265                                 next = slob_next(cur);
 266                                 set_slob(aligned, avail - delta, next);
 267                                 set_slob(cur, delta, aligned);
 268                                 prev = cur;
 269                                 cur = aligned;
 270                                 avail = slob_units(cur);
 271                         }
 272
 273                         next = slob_next(cur);
 274                         if (avail == units) { /* exact fit? unlink. */
 275                                 if (prev)
 276                                         set_slob(prev, slob_units(prev), next);
 277                                 else
 278                                         sp->freelist = next;
 279                         } else { /* fragment */
 280                                 if (prev)
 281                                         set_slob(prev, slob_units(prev), cur + units);
 282                                 else
 283                                         sp->freelist = cur + units;
 284                                 set_slob(cur + units, avail - units, next);
 285                         }
 286
 287                         sp->units -= units;
 288                         if (!sp->units)
 289                                 clear_slob_page_free(sp);
 290                         return cur;
 291                 }
 292                 if (slob_last(cur))
 293                         return NULL;
 294         }
 295 }
 296
 297 /*
 298  * slob_alloc: entry point into the slob allocator.
 299  */
 300 static void *slob_alloc(size_t size, gfp_t gfp, int align, int node)
 301 {
 302         struct page *sp;
 303         struct list_head *prev;
 304         struct list_head *slob_list;
 305         slob_t *b = NULL;
 306         unsigned long flags;
 307
 308         if (size < SLOB_BREAK1)
 309                 slob_list = &free_slob_small;
 310         else if (size < SLOB_BREAK2)
 311                 slob_list = &free_slob_medium;
 312         else
 313                 slob_list = &free_slob_large;
 314
 315         spin_lock_irqsave(&slob_lock, flags);
 316         /* Iterate through each partially free page, try to find room */
 317         list_for_each_entry(sp, slob_list, list) {
 318 #ifdef CONFIG_NUMA
 319                 /*
 320                  * If there's a node specification, search for a partial
 321                  * page with a matching node id in the freelist.
 322                  */
 323                 if (node != -1 && page_to_nid(sp) != node)
 324                         continue;
 325 #endif
 326                 /* Enough room on this page? */
 327                 if (sp->units < SLOB_UNITS(size))
 328                         continue;
 329
 330                 /* Attempt to alloc */
 331                 prev = sp->list.prev;
 332                 b = slob_page_alloc(sp, size, align);
 333                 if (!b)
 334                         continue;
 335
 336                 /* Improve fragment distribution and reduce our average
 337                  * search time by starting our next search here. (see
 338                  * Knuth vol 1, sec 2.5, pg 449) */
 339                 if (prev != slob_list->prev &&
 340                                 slob_list->next != prev->next)
 341                         list_move_tail(slob_list, prev->next);
 342                 break;
 343         }
 344         spin_unlock_irqrestore(&slob_lock, flags);
 345
 346         /* Not enough space: must allocate a new page */
 347         if (!b) {
 348                 b = slob_new_pages(gfp & ~__GFP_ZERO, 0, node);
 349                 if (!b)
 350                         return NULL;
 351                 sp = slob_page(b);
 352                 set_slob_page(sp);
 353
 354                 spin_lock_irqsave(&slob_lock, flags);
 355                 sp->units = SLOB_UNITS(PAGE_SIZE);
 356                 sp->freelist = b;
 357                 INIT_LIST_HEAD(&sp->list);
 358                 set_slob(b, SLOB_UNITS(PAGE_SIZE), b + SLOB_UNITS(PAGE_SIZE));
 359                 set_slob_page_free(sp, slob_list);
 360                 b = slob_page_alloc(sp, size, align);
 361                 BUG_ON(!b);
 362                 spin_unlock_irqrestore(&slob_lock, flags);
 363         }
 364         if (unlikely((gfp & __GFP_ZERO) && b))
 365                 memset(b, 0, size);
 366         return b;
 367 }
 368
 369 /*
 370  * slob_free: entry point into the slob allocator.
 371  */
 372 static void slob_free(void *block, int size)
 373 {
 374         struct page *sp;
 375         slob_t *prev, *next, *b = (slob_t *)block;
 376         slobidx_t units;
 377         unsigned long flags;
 378         struct list_head *slob_list;
 379
 380         if (unlikely(ZERO_OR_NULL_PTR(block)))
 381                 return;
 382         BUG_ON(!size);
 383
 384         sp = slob_page(block);
 385         units = SLOB_UNITS(size);
 386
 387         spin_lock_irqsave(&slob_lock, flags);
 388
 389         if (sp->units + units == SLOB_UNITS(PAGE_SIZE)) {
 390                 /* Go directly to page allocator. Do not pass slob allocator */
 391                 if (slob_page_free(sp))
 392                         clear_slob_page_free(sp);
 393                 spin_unlock_irqrestore(&slob_lock, flags);
 394                 clear_slob_page(sp);
 395                 free_slob_page(sp);
 396                 slob_free_pages(b, 0);
 397                 return;
 398         }
 399
 400         if (!slob_page_free(sp)) {
 401                 /* This slob page is about to become partially free. Easy! */
 402                 sp->units = units;
 403                 sp->freelist = b;
 404                 set_slob(b, units,
 405                         (void *)((unsigned long)(b +
 406                                         SLOB_UNITS(PAGE_SIZE)) & PAGE_MASK));
 407                 if (size < SLOB_BREAK1)
 408                         slob_list = &free_slob_small;
 409                 else if (size < SLOB_BREAK2)
 410                         slob_list = &free_slob_medium;
 411                 else
 412                         slob_list = &free_slob_large;
 413                 set_slob_page_free(sp, slob_list);
 414                 goto out;
 415         }
 416
 417         /*
 418          * Otherwise the page is already partially free, so find reinsertion
 419          * point.
 420          */
 421         sp->units += units;
 422
 423         if (b < (slob_t *)sp->freelist) {
 424                 if (b + units == sp->freelist) {
 425                         units += slob_units(sp->freelist);
 426                         sp->freelist = slob_next(sp->freelist);
 427                 }
 428                 set_slob(b, units, sp->freelist);
 429                 sp->freelist = b;
 430         } else {
 431                 prev = sp->freelist;
 432                 next = slob_next(prev);
 433                 while (b > next) {
 434                         prev = next;
 435                         next = slob_next(prev);
 436                 }
 437
 438                 if (!slob_last(prev) && b + units == next) {
 439                         units += slob_units(next);
 440                         set_slob(b, units, slob_next(next));
 441                 } else
 442                         set_slob(b, units, next);
 443
 444                 if (prev + slob_units(prev) == b) {
 445                         units = slob_units(b) + slob_units(prev);
 446                         set_slob(prev, units, slob_next(b));
 447                 } else
 448                         set_slob(prev, slob_units(prev), b);
 449         }
 450 out:
 451         spin_unlock_irqrestore(&slob_lock, flags);
 452 }
 453
 454 /*
 455  * End of slob allocator proper. Begin kmem_cache_alloc and kmalloc frontend.
 456  */
 457
 458 void *__kmalloc_node(size_t size, gfp_t gfp, int node)
 459 {
 460         unsigned int *m;
 461         int align = max(ARCH_KMALLOC_MINALIGN, ARCH_SLAB_MINALIGN);
 462         void *ret;
 463
 464         gfp &= gfp_allowed_mask;
 465
 466         lockdep_trace_alloc(gfp);
 467
 468         if (size < PAGE_SIZE - align) {
 469                 if (!size)
 470                         return ZERO_SIZE_PTR;
 471
 472                 m = slob_alloc(size + align, gfp, align, node);
 473
 474                 if (!m)
 475                         return NULL;
 476                 *m = size;
 477                 ret = (void *)m + align;
 478
 479                 trace_kmalloc_node(_RET_IP_, ret,
 480                                    size, size + align, gfp, node);
 481         } else {
 482                 unsigned int order = get_order(size);
 483
 484                 if (likely(order))
 485                         gfp |= __GFP_COMP;
 486                 ret = slob_new_pages(gfp, order, node);
 487                 if (ret) {
 488                         struct page *page;
 489                         page = virt_to_page(ret);
 490                         page->private = size;
 491                 }
 492
 493                 trace_kmalloc_node(_RET_IP_, ret,
 494                                    size, PAGE_SIZE << order, gfp, node);
 495         }
 496
 497         kmemleak_alloc(ret, size, 1, gfp);
 498         return ret;
 499 }
 500 EXPORT_SYMBOL(__kmalloc_node);
 501
 502 void kfree(const void *block)
 503 {
 504         struct page *sp;
 505
 506         trace_kfree(_RET_IP_, block);
 507
 508         if (unlikely(ZERO_OR_NULL_PTR(block)))
 509                 return;
 510         kmemleak_free(block);
 511
 512         sp = slob_page(block);
 513         if (is_slob_page(sp)) {
 514                 int align = max(ARCH_KMALLOC_MINALIGN, ARCH_SLAB_MINALIGN);
 515                 unsigned int *m = (unsigned int *)(block - align);
 516                 slob_free(m, *m + align);
 517         } else
 518                 put_page(sp);
 519 }
 520 EXPORT_SYMBOL(kfree);
 521
 522 /* can't use ksize for kmem_cache_alloc memory, only kmalloc */
 523 size_t ksize(const void *block)
 524 {
 525         struct page *sp;
 526
 527         BUG_ON(!block);
 528         if (unlikely(block == ZERO_SIZE_PTR))
 529                 return 0;
 530
 531         sp = slob_page(block);
 532         if (is_slob_page(sp)) {
 533                 int align = max(ARCH_KMALLOC_MINALIGN, ARCH_SLAB_MINALIGN);
 534                 unsigned int *m = (unsigned int *)(block - align);
 535                 return SLOB_UNITS(*m) * SLOB_UNIT;
 536         } else
 537                 return sp->private;
 538 }
 539 EXPORT_SYMBOL(ksize);
 540
 541 struct kmem_cache {
 542         unsigned int size, align;
 543         unsigned long flags;
 544         const char *name;
 545         void (*ctor)(void *);
 546 };
 547
 548 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
 549         size_t align, unsigned long flags, void (*ctor)(void *))
 550 {
 551         struct kmem_cache *c;
 552
 553         c = slob_alloc(sizeof(struct kmem_cache),
 554                 GFP_KERNEL, ARCH_KMALLOC_MINALIGN, -1);
 555
 556         if (c) {
 557                 c->name = name;
 558                 c->size = size;
 559                 if (flags & SLAB_DESTROY_BY_RCU) {
 560                         /* leave room for rcu footer at the end of object */
 561                         c->size += sizeof(struct slob_rcu);
 562                 }
 563                 c->flags = flags;
 564                 c->ctor = ctor;
 565                 /* ignore alignment unless it's forced */
 566                 c->align = (flags & SLAB_HWCACHE_ALIGN) ? SLOB_ALIGN : 0;
 567                 if (c->align < ARCH_SLAB_MINALIGN)
 568                         c->align = ARCH_SLAB_MINALIGN;
 569                 if (c->align < align)
 570                         c->align = align;
 571         } else if (flags & SLAB_PANIC)
 572                 panic("Cannot create slab cache %s\n", name);
 573
 574         kmemleak_alloc(c, sizeof(struct kmem_cache), 1, GFP_KERNEL);
 575         return c;
 576 }
 577 EXPORT_SYMBOL(kmem_cache_create);
 578
 579 void kmem_cache_destroy(struct kmem_cache *c)
 580 {
 581         kmemleak_free(c);
 582         if (c->flags & SLAB_DESTROY_BY_RCU)
 583                 rcu_barrier();
 584         slob_free(c, sizeof(struct kmem_cache));
 585 }
 586 EXPORT_SYMBOL(kmem_cache_destroy);
 587
 588 void *kmem_cache_alloc_node(struct kmem_cache *c, gfp_t flags, int node)
 589 {
 590         void *b;
 591
 592         flags &= gfp_allowed_mask;
 593
 594         lockdep_trace_alloc(flags);
 595
 596         if (c->size < PAGE_SIZE) {
 597                 b = slob_alloc(c->size, flags, c->align, node);
 598                 trace_kmem_cache_alloc_node(_RET_IP_, b, c->size,
 599                                             SLOB_UNITS(c->size) * SLOB_UNIT,
 600                                             flags, node);
 601         } else {
 602                 b = slob_new_pages(flags, get_order(c->size), node);
 603                 trace_kmem_cache_alloc_node(_RET_IP_, b, c->size,
 604                                             PAGE_SIZE << get_order(c->size),
 605                                             flags, node);
 606         }
 607
 608         if (c->ctor)
 609                 c->ctor(b);
 610
 611         kmemleak_alloc_recursive(b, c->size, 1, c->flags, flags);
 612         return b;
 613 }
 614 EXPORT_SYMBOL(kmem_cache_alloc_node);
 615
 616 static void __kmem_cache_free(void *b, int size)
 617 {
 618         if (size < PAGE_SIZE)
 619                 slob_free(b, size);
 620         else
 621                 slob_free_pages(b, get_order(size));
 622 }
 623
 624 static void kmem_rcu_free(struct rcu_head *head)
 625 {
 626         struct slob_rcu *slob_rcu = (struct slob_rcu *)head;
 627         void *b = (void *)slob_rcu - (slob_rcu->size - sizeof(struct slob_rcu));
 628
 629         __kmem_cache_free(b, slob_rcu->size);
 630 }
 631
 632 void kmem_cache_free(struct kmem_cache *c, void *b)
 633 {
 634         kmemleak_free_recursive(b, c->flags);
 635         if (unlikely(c->flags & SLAB_DESTROY_BY_RCU)) {
 636                 struct slob_rcu *slob_rcu;
 637                 slob_rcu = b + (c->size - sizeof(struct slob_rcu));
 638                 slob_rcu->size = c->size;
 639                 call_rcu(&slob_rcu->head, kmem_rcu_free);
 640         } else {
 641                 __kmem_cache_free(b, c->size);
 642         }
 643
 644         trace_kmem_cache_free(_RET_IP_, b);
 645 }
 646 EXPORT_SYMBOL(kmem_cache_free);
 647
 648 unsigned int kmem_cache_size(struct kmem_cache *c)
 649 {
 650         return c->size;
 651 }
 652 EXPORT_SYMBOL(kmem_cache_size);
 653
 654 int kmem_cache_shrink(struct kmem_cache *d)
 655 {
 656         return 0;
 657 }
 658 EXPORT_SYMBOL(kmem_cache_shrink);
 659
 660 static unsigned int slob_ready __read_mostly;
 661
 662 int slab_is_available(void)
 663 {
 664         return slob_ready;
 665 }
 666
 667 void __init kmem_cache_init(void)
 668 {
 669         slob_ready = 1;
 670 }
 671
 672 void __init kmem_cache_init_late(void)
 673 {
 674         /* Nothing to do */
 675 }