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  * All partially free slob pages go on these lists.
  96  */
  97 #define SLOB_BREAK1 256
  98 #define SLOB_BREAK2 1024
  99 static LIST_HEAD(free_slob_small);
 100 static LIST_HEAD(free_slob_medium);
 101 static LIST_HEAD(free_slob_large);
 102
 103 /*
 104  * slob_page_free: true for pages on free_slob_pages list.
 105  */
 106 static inline int slob_page_free(struct page *sp)
 107 {
 108         return PageSlobFree(sp);
 109 }
 110
 111 static void set_slob_page_free(struct page *sp, struct list_head *list)
 112 {
 113         list_add(&sp->list, list);
 114         __SetPageSlobFree(sp);
 115 }
 116
 117 static inline void clear_slob_page_free(struct page *sp)
 118 {
 119         list_del(&sp->list);
 120         __ClearPageSlobFree(sp);
 121 }
 122
 123 #define SLOB_UNIT sizeof(slob_t)
 124 #define SLOB_UNITS(size) (((size) + SLOB_UNIT - 1)/SLOB_UNIT)
 125 #define SLOB_ALIGN L1_CACHE_BYTES
 126
 127 /*
 128  * struct slob_rcu is inserted at the tail of allocated slob blocks, which
 129  * were created with a SLAB_DESTROY_BY_RCU slab. slob_rcu is used to free
 130  * the block using call_rcu.
 131  */
 132 struct slob_rcu {
 133         struct rcu_head head;
 134         int size;
 135 };
 136
 137 /*
 138  * slob_lock protects all slob allocator structures.
 139  */
 140 static DEFINE_SPINLOCK(slob_lock);
 141
 142 /*
 143  * Encode the given size and next info into a free slob block s.
 144  */
 145 static void set_slob(slob_t *s, slobidx_t size, slob_t *next)
 146 {
 147         slob_t *base = (slob_t *)((unsigned long)s & PAGE_MASK);
 148         slobidx_t offset = next - base;
 149
 150         if (size > 1) {
 151                 s[0].units = size;
 152                 s[1].units = offset;
 153         } else
 154                 s[0].units = -offset;
 155 }
 156
 157 /*
 158  * Return the size of a slob block.
 159  */
 160 static slobidx_t slob_units(slob_t *s)
 161 {
 162         if (s->units > 0)
 163                 return s->units;
 164         return 1;
 165 }
 166
 167 /*
 168  * Return the next free slob block pointer after this one.
 169  */
 170 static slob_t *slob_next(slob_t *s)
 171 {
 172         slob_t *base = (slob_t *)((unsigned long)s & PAGE_MASK);
 173         slobidx_t next;
 174
 175         if (s[0].units < 0)
 176                 next = -s[0].units;
 177         else
 178                 next = s[1].units;
 179         return base+next;
 180 }
 181
 182 /*
 183  * Returns true if s is the last free block in its page.
 184  */
 185 static int slob_last(slob_t *s)
 186 {
 187         return !((unsigned long)slob_next(s) & ~PAGE_MASK);
 188 }
 189
 190 static void *slob_new_pages(gfp_t gfp, int order, int node)
 191 {
 192         void *page;
 193
 194 #ifdef CONFIG_NUMA
 195         if (node != -1)
 196                 page = alloc_pages_exact_node(node, gfp, order);
 197         else
 198 #endif
 199                 page = alloc_pages(gfp, order);
 200
 201         if (!page)
 202                 return NULL;
 203
 204         return page_address(page);
 205 }
 206
 207 static void slob_free_pages(void *b, int order)
 208 {
 209         if (current->reclaim_state)
 210                 current->reclaim_state->reclaimed_slab += 1 << order;
 211         free_pages((unsigned long)b, order);
 212 }
 213
 214 /*
 215  * Allocate a slob block within a given slob_page sp.
 216  */
 217 static void *slob_page_alloc(struct page *sp, size_t size, int align)
 218 {
 219         slob_t *prev, *cur, *aligned = NULL;
 220         int delta = 0, units = SLOB_UNITS(size);
 221
 222         for (prev = NULL, cur = sp->freelist; ; prev = cur, cur = slob_next(cur)) {
 223                 slobidx_t avail = slob_units(cur);
 224
 225                 if (align) {
 226                         aligned = (slob_t *)ALIGN((unsigned long)cur, align);
 227                         delta = aligned - cur;
 228                 }
 229                 if (avail >= units + delta) { /* room enough? */
 230                         slob_t *next;
 231
 232                         if (delta) { /* need to fragment head to align? */
 233                                 next = slob_next(cur);
 234                                 set_slob(aligned, avail - delta, next);
 235                                 set_slob(cur, delta, aligned);
 236                                 prev = cur;
 237                                 cur = aligned;
 238                                 avail = slob_units(cur);
 239                         }
 240
 241                         next = slob_next(cur);
 242                         if (avail == units) { /* exact fit? unlink. */
 243                                 if (prev)
 244                                         set_slob(prev, slob_units(prev), next);
 245                                 else
 246                                         sp->freelist = next;
 247                         } else { /* fragment */
 248                                 if (prev)
 249                                         set_slob(prev, slob_units(prev), cur + units);
 250                                 else
 251                                         sp->freelist = cur + units;
 252                                 set_slob(cur + units, avail - units, next);
 253                         }
 254
 255                         sp->units -= units;
 256                         if (!sp->units)
 257                                 clear_slob_page_free(sp);
 258                         return cur;
 259                 }
 260                 if (slob_last(cur))
 261                         return NULL;
 262         }
 263 }
 264
 265 /*
 266  * slob_alloc: entry point into the slob allocator.
 267  */
 268 static void *slob_alloc(size_t size, gfp_t gfp, int align, int node)
 269 {
 270         struct page *sp;
 271         struct list_head *prev;
 272         struct list_head *slob_list;
 273         slob_t *b = NULL;
 274         unsigned long flags;
 275
 276         if (size < SLOB_BREAK1)
 277                 slob_list = &free_slob_small;
 278         else if (size < SLOB_BREAK2)
 279                 slob_list = &free_slob_medium;
 280         else
 281                 slob_list = &free_slob_large;
 282
 283         spin_lock_irqsave(&slob_lock, flags);
 284         /* Iterate through each partially free page, try to find room */
 285         list_for_each_entry(sp, slob_list, list) {
 286 #ifdef CONFIG_NUMA
 287                 /*
 288                  * If there's a node specification, search for a partial
 289                  * page with a matching node id in the freelist.
 290                  */
 291                 if (node != -1 && page_to_nid(sp) != node)
 292                         continue;
 293 #endif
 294                 /* Enough room on this page? */
 295                 if (sp->units < SLOB_UNITS(size))
 296                         continue;
 297
 298                 /* Attempt to alloc */
 299                 prev = sp->list.prev;
 300                 b = slob_page_alloc(sp, size, align);
 301                 if (!b)
 302                         continue;
 303
 304                 /* Improve fragment distribution and reduce our average
 305                  * search time by starting our next search here. (see
 306                  * Knuth vol 1, sec 2.5, pg 449) */
 307                 if (prev != slob_list->prev &&
 308                                 slob_list->next != prev->next)
 309                         list_move_tail(slob_list, prev->next);
 310                 break;
 311         }
 312         spin_unlock_irqrestore(&slob_lock, flags);
 313
 314         /* Not enough space: must allocate a new page */
 315         if (!b) {
 316                 b = slob_new_pages(gfp & ~__GFP_ZERO, 0, node);
 317                 if (!b)
 318                         return NULL;
 319                 sp = virt_to_page(b);
 320                 __SetPageSlab(sp);
 321
 322                 spin_lock_irqsave(&slob_lock, flags);
 323                 sp->units = SLOB_UNITS(PAGE_SIZE);
 324                 sp->freelist = b;
 325                 INIT_LIST_HEAD(&sp->list);
 326                 set_slob(b, SLOB_UNITS(PAGE_SIZE), b + SLOB_UNITS(PAGE_SIZE));
 327                 set_slob_page_free(sp, slob_list);
 328                 b = slob_page_alloc(sp, size, align);
 329                 BUG_ON(!b);
 330                 spin_unlock_irqrestore(&slob_lock, flags);
 331         }
 332         if (unlikely((gfp & __GFP_ZERO) && b))
 333                 memset(b, 0, size);
 334         return b;
 335 }
 336
 337 /*
 338  * slob_free: entry point into the slob allocator.
 339  */
 340 static void slob_free(void *block, int size)
 341 {
 342         struct page *sp;
 343         slob_t *prev, *next, *b = (slob_t *)block;
 344         slobidx_t units;
 345         unsigned long flags;
 346         struct list_head *slob_list;
 347
 348         if (unlikely(ZERO_OR_NULL_PTR(block)))
 349                 return;
 350         BUG_ON(!size);
 351
 352         sp = virt_to_page(block);
 353         units = SLOB_UNITS(size);
 354
 355         spin_lock_irqsave(&slob_lock, flags);
 356
 357         if (sp->units + units == SLOB_UNITS(PAGE_SIZE)) {
 358                 /* Go directly to page allocator. Do not pass slob allocator */
 359                 if (slob_page_free(sp))
 360                         clear_slob_page_free(sp);
 361                 spin_unlock_irqrestore(&slob_lock, flags);
 362                 __ClearPageSlab(sp);
 363                 reset_page_mapcount(sp);
 364                 slob_free_pages(b, 0);
 365                 return;
 366         }
 367
 368         if (!slob_page_free(sp)) {
 369                 /* This slob page is about to become partially free. Easy! */
 370                 sp->units = units;
 371                 sp->freelist = b;
 372                 set_slob(b, units,
 373                         (void *)((unsigned long)(b +
 374                                         SLOB_UNITS(PAGE_SIZE)) & PAGE_MASK));
 375                 if (size < SLOB_BREAK1)
 376                         slob_list = &free_slob_small;
 377                 else if (size < SLOB_BREAK2)
 378                         slob_list = &free_slob_medium;
 379                 else
 380                         slob_list = &free_slob_large;
 381                 set_slob_page_free(sp, slob_list);
 382                 goto out;
 383         }
 384
 385         /*
 386          * Otherwise the page is already partially free, so find reinsertion
 387          * point.
 388          */
 389         sp->units += units;
 390
 391         if (b < (slob_t *)sp->freelist) {
 392                 if (b + units == sp->freelist) {
 393                         units += slob_units(sp->freelist);
 394                         sp->freelist = slob_next(sp->freelist);
 395                 }
 396                 set_slob(b, units, sp->freelist);
 397                 sp->freelist = b;
 398         } else {
 399                 prev = sp->freelist;
 400                 next = slob_next(prev);
 401                 while (b > next) {
 402                         prev = next;
 403                         next = slob_next(prev);
 404                 }
 405
 406                 if (!slob_last(prev) && b + units == next) {
 407                         units += slob_units(next);
 408                         set_slob(b, units, slob_next(next));
 409                 } else
 410                         set_slob(b, units, next);
 411
 412                 if (prev + slob_units(prev) == b) {
 413                         units = slob_units(b) + slob_units(prev);
 414                         set_slob(prev, units, slob_next(b));
 415                 } else
 416                         set_slob(prev, slob_units(prev), b);
 417         }
 418 out:
 419         spin_unlock_irqrestore(&slob_lock, flags);
 420 }
 421
 422 /*
 423  * End of slob allocator proper. Begin kmem_cache_alloc and kmalloc frontend.
 424  */
 425
 426 void *__kmalloc_node(size_t size, gfp_t gfp, int node)
 427 {
 428         unsigned int *m;
 429         int align = max(ARCH_KMALLOC_MINALIGN, ARCH_SLAB_MINALIGN);
 430         void *ret;
 431
 432         gfp &= gfp_allowed_mask;
 433
 434         lockdep_trace_alloc(gfp);
 435
 436         if (size < PAGE_SIZE - align) {
 437                 if (!size)
 438                         return ZERO_SIZE_PTR;
 439
 440                 m = slob_alloc(size + align, gfp, align, node);
 441
 442                 if (!m)
 443                         return NULL;
 444                 *m = size;
 445                 ret = (void *)m + align;
 446
 447                 trace_kmalloc_node(_RET_IP_, ret,
 448                                    size, size + align, gfp, node);
 449         } else {
 450                 unsigned int order = get_order(size);
 451
 452                 if (likely(order))
 453                         gfp |= __GFP_COMP;
 454                 ret = slob_new_pages(gfp, order, node);
 455                 if (ret) {
 456                         struct page *page;
 457                         page = virt_to_page(ret);
 458                         page->private = size;
 459                 }
 460
 461                 trace_kmalloc_node(_RET_IP_, ret,
 462                                    size, PAGE_SIZE << order, gfp, node);
 463         }
 464
 465         kmemleak_alloc(ret, size, 1, gfp);
 466         return ret;
 467 }
 468 EXPORT_SYMBOL(__kmalloc_node);
 469
 470 void kfree(const void *block)
 471 {
 472         struct page *sp;
 473
 474         trace_kfree(_RET_IP_, block);
 475
 476         if (unlikely(ZERO_OR_NULL_PTR(block)))
 477                 return;
 478         kmemleak_free(block);
 479
 480         sp = virt_to_page(block);
 481         if (PageSlab(sp)) {
 482                 int align = max(ARCH_KMALLOC_MINALIGN, ARCH_SLAB_MINALIGN);
 483                 unsigned int *m = (unsigned int *)(block - align);
 484                 slob_free(m, *m + align);
 485         } else
 486                 put_page(sp);
 487 }
 488 EXPORT_SYMBOL(kfree);
 489
 490 /* can't use ksize for kmem_cache_alloc memory, only kmalloc */
 491 size_t ksize(const void *block)
 492 {
 493         struct page *sp;
 494
 495         BUG_ON(!block);
 496         if (unlikely(block == ZERO_SIZE_PTR))
 497                 return 0;
 498
 499         sp = virt_to_page(block);
 500         if (PageSlab(sp)) {
 501                 int align = max(ARCH_KMALLOC_MINALIGN, ARCH_SLAB_MINALIGN);
 502                 unsigned int *m = (unsigned int *)(block - align);
 503                 return SLOB_UNITS(*m) * SLOB_UNIT;
 504         } else
 505                 return sp->private;
 506 }
 507 EXPORT_SYMBOL(ksize);
 508
 509 struct kmem_cache {
 510         unsigned int size, align;
 511         unsigned long flags;
 512         const char *name;
 513         void (*ctor)(void *);
 514 };
 515
 516 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
 517         size_t align, unsigned long flags, void (*ctor)(void *))
 518 {
 519         struct kmem_cache *c;
 520
 521         c = slob_alloc(sizeof(struct kmem_cache),
 522                 GFP_KERNEL, ARCH_KMALLOC_MINALIGN, -1);
 523
 524         if (c) {
 525                 c->name = name;
 526                 c->size = size;
 527                 if (flags & SLAB_DESTROY_BY_RCU) {
 528                         /* leave room for rcu footer at the end of object */
 529                         c->size += sizeof(struct slob_rcu);
 530                 }
 531                 c->flags = flags;
 532                 c->ctor = ctor;
 533                 /* ignore alignment unless it's forced */
 534                 c->align = (flags & SLAB_HWCACHE_ALIGN) ? SLOB_ALIGN : 0;
 535                 if (c->align < ARCH_SLAB_MINALIGN)
 536                         c->align = ARCH_SLAB_MINALIGN;
 537                 if (c->align < align)
 538                         c->align = align;
 539         } else if (flags & SLAB_PANIC)
 540                 panic("Cannot create slab cache %s\n", name);
 541
 542         kmemleak_alloc(c, sizeof(struct kmem_cache), 1, GFP_KERNEL);
 543         return c;
 544 }
 545 EXPORT_SYMBOL(kmem_cache_create);
 546
 547 void kmem_cache_destroy(struct kmem_cache *c)
 548 {
 549         kmemleak_free(c);
 550         if (c->flags & SLAB_DESTROY_BY_RCU)
 551                 rcu_barrier();
 552         slob_free(c, sizeof(struct kmem_cache));
 553 }
 554 EXPORT_SYMBOL(kmem_cache_destroy);
 555
 556 void *kmem_cache_alloc_node(struct kmem_cache *c, gfp_t flags, int node)
 557 {
 558         void *b;
 559
 560         flags &= gfp_allowed_mask;
 561
 562         lockdep_trace_alloc(flags);
 563
 564         if (c->size < PAGE_SIZE) {
 565                 b = slob_alloc(c->size, flags, c->align, node);
 566                 trace_kmem_cache_alloc_node(_RET_IP_, b, c->size,
 567                                             SLOB_UNITS(c->size) * SLOB_UNIT,
 568                                             flags, node);
 569         } else {
 570                 b = slob_new_pages(flags, get_order(c->size), node);
 571                 trace_kmem_cache_alloc_node(_RET_IP_, b, c->size,
 572                                             PAGE_SIZE << get_order(c->size),
 573                                             flags, node);
 574         }
 575
 576         if (c->ctor)
 577                 c->ctor(b);
 578
 579         kmemleak_alloc_recursive(b, c->size, 1, c->flags, flags);
 580         return b;
 581 }
 582 EXPORT_SYMBOL(kmem_cache_alloc_node);
 583
 584 static void __kmem_cache_free(void *b, int size)
 585 {
 586         if (size < PAGE_SIZE)
 587                 slob_free(b, size);
 588         else
 589                 slob_free_pages(b, get_order(size));
 590 }
 591
 592 static void kmem_rcu_free(struct rcu_head *head)
 593 {
 594         struct slob_rcu *slob_rcu = (struct slob_rcu *)head;
 595         void *b = (void *)slob_rcu - (slob_rcu->size - sizeof(struct slob_rcu));
 596
 597         __kmem_cache_free(b, slob_rcu->size);
 598 }
 599
 600 void kmem_cache_free(struct kmem_cache *c, void *b)
 601 {
 602         kmemleak_free_recursive(b, c->flags);
 603         if (unlikely(c->flags & SLAB_DESTROY_BY_RCU)) {
 604                 struct slob_rcu *slob_rcu;
 605                 slob_rcu = b + (c->size - sizeof(struct slob_rcu));
 606                 slob_rcu->size = c->size;
 607                 call_rcu(&slob_rcu->head, kmem_rcu_free);
 608         } else {
 609                 __kmem_cache_free(b, c->size);
 610         }
 611
 612         trace_kmem_cache_free(_RET_IP_, b);
 613 }
 614 EXPORT_SYMBOL(kmem_cache_free);
 615
 616 unsigned int kmem_cache_size(struct kmem_cache *c)
 617 {
 618         return c->size;
 619 }
 620 EXPORT_SYMBOL(kmem_cache_size);
 621
 622 int kmem_cache_shrink(struct kmem_cache *d)
 623 {
 624         return 0;
 625 }
 626 EXPORT_SYMBOL(kmem_cache_shrink);
 627
 628 static unsigned int slob_ready __read_mostly;
 629
 630 int slab_is_available(void)
 631 {
 632         return slob_ready;
 633 }
 634
 635 void __init kmem_cache_init(void)
 636 {
 637         slob_ready = 1;
 638 }
 639
 640 void __init kmem_cache_init_late(void)
 641 {
 642         /* Nothing to do */
 643 }