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 linked list of pages from alloc_pages(), and
  16  * within each page, there is a singly-linked list of free blocks (slob_t).
  17  * The heap is grown on demand and allocation from the heap is currently
  18  * first-fit.
  19  *
  20  * Above this is an implementation of kmalloc/kfree. Blocks returned
  21  * from kmalloc are prepended with a 4-byte header with the kmalloc size.
  22  * If kmalloc is asked for objects of PAGE_SIZE or larger, it calls
  23  * alloc_pages() directly, allocating compound pages so the page order
  24  * does not have to be separately tracked, and also stores the exact
  25  * allocation size in page->private so that it can be used to accurately
  26  * provide ksize(). These objects are detected in kfree() because slob_page()
  27  * is false for them.
  28  *
  29  * SLAB is emulated on top of SLOB by simply calling constructors and
  30  * destructors for every SLAB allocation. Objects are returned with the
  31  * 4-byte alignment unless the SLAB_HWCACHE_ALIGN flag is set, in which
  32  * case the low-level allocator will fragment blocks to create the proper
  33  * alignment. Again, objects of page-size or greater are allocated by
  34  * calling alloc_pages(). As SLAB objects know their size, no separate
  35  * size bookkeeping is necessary and there is essentially no allocation
  36  * space overhead, and compound pages aren't needed for multi-page
  37  * allocations.
  38  *
  39  * NUMA support in SLOB is fairly simplistic, pushing most of the real
  40  * logic down to the page allocator, and simply doing the node accounting
  41  * on the upper levels. In the event that a node id is explicitly
  42  * provided, alloc_pages_node() with the specified node id is used
  43  * instead. The common case (or when the node id isn't explicitly provided)
  44  * will default to the current node, as per numa_node_id().
  45  *
  46  * Node aware pages are still inserted in to the global freelist, and
  47  * these are scanned for by matching against the node id encoded in the
  48  * page flags. As a result, block allocations that can be satisfied from
  49  * the freelist will only be done so on pages residing on the same node,
  50  * in order to prevent random node placement.
  51  */
  52
  53 #include <linux/kernel.h>
  54 #include <linux/slab.h>
  55 #include <linux/mm.h>
  56 #include <linux/cache.h>
  57 #include <linux/init.h>
  58 #include <linux/module.h>
  59 #include <linux/rcupdate.h>
  60 #include <linux/list.h>
  61 #include <asm/atomic.h>
  62
  63 /*
  64  * slob_block has a field 'units', which indicates size of block if +ve,
  65  * or offset of next block if -ve (in SLOB_UNITs).
  66  *
  67  * Free blocks of size 1 unit simply contain the offset of the next block.
  68  * Those with larger size contain their size in the first SLOB_UNIT of
  69  * memory, and the offset of the next free block in the second SLOB_UNIT.
  70  */
  71 #if PAGE_SIZE <= (32767 * 2)
  72 typedef s16 slobidx_t;
  73 #else
  74 typedef s32 slobidx_t;
  75 #endif
  76
  77 struct slob_block {
  78         slobidx_t units;
  79 };
  80 typedef struct slob_block slob_t;
  81
  82 /*
  83  * We use struct page fields to manage some slob allocation aspects,
  84  * however to avoid the horrible mess in include/linux/mm_types.h, we'll
  85  * just define our own struct page type variant here.
  86  */
  87 struct slob_page {
  88         union {
  89                 struct {
  90                         unsigned long flags;    /* mandatory */
  91                         atomic_t _count;        /* mandatory */
  92                         slobidx_t units;        /* free units left in page */
  93                         unsigned long pad[2];
  94                         slob_t *free;           /* first free slob_t in page */
  95                         struct list_head list;  /* linked list of free pages */
  96                 };
  97                 struct page page;
  98         };
  99 };
 100 static inline void struct_slob_page_wrong_size(void)
 101 { BUILD_BUG_ON(sizeof(struct slob_page) != sizeof(struct page)); }
 102
 103 /*
 104  * free_slob_page: call before a slob_page is returned to the page allocator.
 105  */
 106 static inline void free_slob_page(struct slob_page *sp)
 107 {
 108         reset_page_mapcount(&sp->page);
 109         sp->page.mapping = NULL;
 110 }
 111
 112 /*
 113  * All (partially) free slob pages go on this list.
 114  */
 115 static LIST_HEAD(free_slob_pages);
 116
 117 /*
 118  * slob_page: True for all slob pages (false for bigblock pages)
 119  */
 120 static inline int slob_page(struct slob_page *sp)
 121 {
 122         return test_bit(PG_active, &sp->flags);
 123 }
 124
 125 static inline void set_slob_page(struct slob_page *sp)
 126 {
 127         __set_bit(PG_active, &sp->flags);
 128 }
 129
 130 static inline void clear_slob_page(struct slob_page *sp)
 131 {
 132         __clear_bit(PG_active, &sp->flags);
 133 }
 134
 135 /*
 136  * slob_page_free: true for pages on free_slob_pages list.
 137  */
 138 static inline int slob_page_free(struct slob_page *sp)
 139 {
 140         return test_bit(PG_private, &sp->flags);
 141 }
 142
 143 static inline void set_slob_page_free(struct slob_page *sp)
 144 {
 145         list_add(&sp->list, &free_slob_pages);
 146         __set_bit(PG_private, &sp->flags);
 147 }
 148
 149 static inline void clear_slob_page_free(struct slob_page *sp)
 150 {
 151         list_del(&sp->list);
 152         __clear_bit(PG_private, &sp->flags);
 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_page(gfp_t gfp, int order, int node)
 223 {
 224         void *page;
 225
 226 #ifdef CONFIG_NUMA
 227         if (node != -1)
 228                 page = alloc_pages_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 /*
 240  * Allocate a slob block within a given slob_page sp.
 241  */
 242 static void *slob_page_alloc(struct slob_page *sp, size_t size, int align)
 243 {
 244         slob_t *prev, *cur, *aligned = 0;
 245         int delta = 0, units = SLOB_UNITS(size);
 246
 247         for (prev = NULL, cur = sp->free; ; prev = cur, cur = slob_next(cur)) {
 248                 slobidx_t avail = slob_units(cur);
 249
 250                 if (align) {
 251                         aligned = (slob_t *)ALIGN((unsigned long)cur, align);
 252                         delta = aligned - cur;
 253                 }
 254                 if (avail >= units + delta) { /* room enough? */
 255                         slob_t *next;
 256
 257                         if (delta) { /* need to fragment head to align? */
 258                                 next = slob_next(cur);
 259                                 set_slob(aligned, avail - delta, next);
 260                                 set_slob(cur, delta, aligned);
 261                                 prev = cur;
 262                                 cur = aligned;
 263                                 avail = slob_units(cur);
 264                         }
 265
 266                         next = slob_next(cur);
 267                         if (avail == units) { /* exact fit? unlink. */
 268                                 if (prev)
 269                                         set_slob(prev, slob_units(prev), next);
 270                                 else
 271                                         sp->free = next;
 272                         } else { /* fragment */
 273                                 if (prev)
 274                                         set_slob(prev, slob_units(prev), cur + units);
 275                                 else
 276                                         sp->free = cur + units;
 277                                 set_slob(cur + units, avail - units, next);
 278                         }
 279
 280                         sp->units -= units;
 281                         if (!sp->units)
 282                                 clear_slob_page_free(sp);
 283                         return cur;
 284                 }
 285                 if (slob_last(cur))
 286                         return NULL;
 287         }
 288 }
 289
 290 /*
 291  * slob_alloc: entry point into the slob allocator.
 292  */
 293 static void *slob_alloc(size_t size, gfp_t gfp, int align, int node)
 294 {
 295         struct slob_page *sp;
 296         struct list_head *prev;
 297         slob_t *b = NULL;
 298         unsigned long flags;
 299
 300         spin_lock_irqsave(&slob_lock, flags);
 301         /* Iterate through each partially free page, try to find room */
 302         list_for_each_entry(sp, &free_slob_pages, list) {
 303 #ifdef CONFIG_NUMA
 304                 /*
 305                  * If there's a node specification, search for a partial
 306                  * page with a matching node id in the freelist.
 307                  */
 308                 if (node != -1 && page_to_nid(&sp->page) != node)
 309                         continue;
 310 #endif
 311                 /* Enough room on this page? */
 312                 if (sp->units < SLOB_UNITS(size))
 313                         continue;
 314
 315                 /* Attempt to alloc */
 316                 prev = sp->list.prev;
 317                 b = slob_page_alloc(sp, size, align);
 318                 if (!b)
 319                         continue;
 320
 321                 /* Improve fragment distribution and reduce our average
 322                  * search time by starting our next search here. (see
 323                  * Knuth vol 1, sec 2.5, pg 449) */
 324                 if (prev != free_slob_pages.prev &&
 325                                 free_slob_pages.next != prev->next)
 326                         list_move_tail(&free_slob_pages, prev->next);
 327                 break;
 328         }
 329         spin_unlock_irqrestore(&slob_lock, flags);
 330
 331         /* Not enough space: must allocate a new page */
 332         if (!b) {
 333                 b = slob_new_page(gfp & ~__GFP_ZERO, 0, node);
 334                 if (!b)
 335                         return 0;
 336                 sp = (struct slob_page *)virt_to_page(b);
 337                 set_slob_page(sp);
 338
 339                 spin_lock_irqsave(&slob_lock, flags);
 340                 sp->units = SLOB_UNITS(PAGE_SIZE);
 341                 sp->free = b;
 342                 INIT_LIST_HEAD(&sp->list);
 343                 set_slob(b, SLOB_UNITS(PAGE_SIZE), b + SLOB_UNITS(PAGE_SIZE));
 344                 set_slob_page_free(sp);
 345                 b = slob_page_alloc(sp, size, align);
 346                 BUG_ON(!b);
 347                 spin_unlock_irqrestore(&slob_lock, flags);
 348         }
 349         if (unlikely((gfp & __GFP_ZERO) && b))
 350                 memset(b, 0, size);
 351         return b;
 352 }
 353
 354 /*
 355  * slob_free: entry point into the slob allocator.
 356  */
 357 static void slob_free(void *block, int size)
 358 {
 359         struct slob_page *sp;
 360         slob_t *prev, *next, *b = (slob_t *)block;
 361         slobidx_t units;
 362         unsigned long flags;
 363
 364         if (unlikely(ZERO_OR_NULL_PTR(block)))
 365                 return;
 366         BUG_ON(!size);
 367
 368         sp = (struct slob_page *)virt_to_page(block);
 369         units = SLOB_UNITS(size);
 370
 371         spin_lock_irqsave(&slob_lock, flags);
 372
 373         if (sp->units + units == SLOB_UNITS(PAGE_SIZE)) {
 374                 /* Go directly to page allocator. Do not pass slob allocator */
 375                 if (slob_page_free(sp))
 376                         clear_slob_page_free(sp);
 377                 clear_slob_page(sp);
 378                 free_slob_page(sp);
 379                 free_page((unsigned long)b);
 380                 goto out;
 381         }
 382
 383         if (!slob_page_free(sp)) {
 384                 /* This slob page is about to become partially free. Easy! */
 385                 sp->units = units;
 386                 sp->free = b;
 387                 set_slob(b, units,
 388                         (void *)((unsigned long)(b +
 389                                         SLOB_UNITS(PAGE_SIZE)) & PAGE_MASK));
 390                 set_slob_page_free(sp);
 391                 goto out;
 392         }
 393
 394         /*
 395          * Otherwise the page is already partially free, so find reinsertion
 396          * point.
 397          */
 398         sp->units += units;
 399
 400         if (b < sp->free) {
 401                 set_slob(b, units, sp->free);
 402                 sp->free = b;
 403         } else {
 404                 prev = sp->free;
 405                 next = slob_next(prev);
 406                 while (b > next) {
 407                         prev = next;
 408                         next = slob_next(prev);
 409                 }
 410
 411                 if (!slob_last(prev) && b + units == next) {
 412                         units += slob_units(next);
 413                         set_slob(b, units, slob_next(next));
 414                 } else
 415                         set_slob(b, units, next);
 416
 417                 if (prev + slob_units(prev) == b) {
 418                         units = slob_units(b) + slob_units(prev);
 419                         set_slob(prev, units, slob_next(b));
 420                 } else
 421                         set_slob(prev, slob_units(prev), b);
 422         }
 423 out:
 424         spin_unlock_irqrestore(&slob_lock, flags);
 425 }
 426
 427 /*
 428  * End of slob allocator proper. Begin kmem_cache_alloc and kmalloc frontend.
 429  */
 430
 431 #ifndef ARCH_KMALLOC_MINALIGN
 432 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long)
 433 #endif
 434
 435 #ifndef ARCH_SLAB_MINALIGN
 436 #define ARCH_SLAB_MINALIGN __alignof__(unsigned long)
 437 #endif
 438
 439 void *__kmalloc_node(size_t size, gfp_t gfp, int node)
 440 {
 441         unsigned int *m;
 442         int align = max(ARCH_KMALLOC_MINALIGN, ARCH_SLAB_MINALIGN);
 443
 444         if (size < PAGE_SIZE - align) {
 445                 if (!size)
 446                         return ZERO_SIZE_PTR;
 447
 448                 m = slob_alloc(size + align, gfp, align, node);
 449                 if (m)
 450                         *m = size;
 451                 return (void *)m + align;
 452         } else {
 453                 void *ret;
 454
 455                 ret = slob_new_page(gfp | __GFP_COMP, get_order(size), node);
 456                 if (ret) {
 457                         struct page *page;
 458                         page = virt_to_page(ret);
 459                         page->private = size;
 460                 }
 461                 return ret;
 462         }
 463 }
 464 EXPORT_SYMBOL(__kmalloc_node);
 465
 466 void kfree(const void *block)
 467 {
 468         struct slob_page *sp;
 469
 470         if (unlikely(ZERO_OR_NULL_PTR(block)))
 471                 return;
 472
 473         sp = (struct slob_page *)virt_to_page(block);
 474         if (slob_page(sp)) {
 475                 int align = max(ARCH_KMALLOC_MINALIGN, ARCH_SLAB_MINALIGN);
 476                 unsigned int *m = (unsigned int *)(block - align);
 477                 slob_free(m, *m + align);
 478         } else
 479                 put_page(&sp->page);
 480 }
 481 EXPORT_SYMBOL(kfree);
 482
 483 /* can't use ksize for kmem_cache_alloc memory, only kmalloc */
 484 size_t ksize(const void *block)
 485 {
 486         struct slob_page *sp;
 487
 488         BUG_ON(!block);
 489         if (unlikely(block == ZERO_SIZE_PTR))
 490                 return 0;
 491
 492         sp = (struct slob_page *)virt_to_page(block);
 493         if (slob_page(sp))
 494                 return ((slob_t *)block - 1)->units + SLOB_UNIT;
 495         else
 496                 return sp->page.private;
 497 }
 498 EXPORT_SYMBOL(ksize);
 499
 500 struct kmem_cache {
 501         unsigned int size, align;
 502         unsigned long flags;
 503         const char *name;
 504         void (*ctor)(struct kmem_cache *, void *);
 505 };
 506
 507 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
 508         size_t align, unsigned long flags,
 509         void (*ctor)(struct kmem_cache *, void *))
 510 {
 511         struct kmem_cache *c;
 512
 513         c = slob_alloc(sizeof(struct kmem_cache), flags, 0, -1);
 514
 515         if (c) {
 516                 c->name = name;
 517                 c->size = size;
 518                 if (flags & SLAB_DESTROY_BY_RCU) {
 519                         /* leave room for rcu footer at the end of object */
 520                         c->size += sizeof(struct slob_rcu);
 521                 }
 522                 c->flags = flags;
 523                 c->ctor = ctor;
 524                 /* ignore alignment unless it's forced */
 525                 c->align = (flags & SLAB_HWCACHE_ALIGN) ? SLOB_ALIGN : 0;
 526                 if (c->align < ARCH_SLAB_MINALIGN)
 527                         c->align = ARCH_SLAB_MINALIGN;
 528                 if (c->align < align)
 529                         c->align = align;
 530         } else if (flags & SLAB_PANIC)
 531                 panic("Cannot create slab cache %s\n", name);
 532
 533         return c;
 534 }
 535 EXPORT_SYMBOL(kmem_cache_create);
 536
 537 void kmem_cache_destroy(struct kmem_cache *c)
 538 {
 539         slob_free(c, sizeof(struct kmem_cache));
 540 }
 541 EXPORT_SYMBOL(kmem_cache_destroy);
 542
 543 void *kmem_cache_alloc_node(struct kmem_cache *c, gfp_t flags, int node)
 544 {
 545         void *b;
 546
 547         if (c->size < PAGE_SIZE)
 548                 b = slob_alloc(c->size, flags, c->align, node);
 549         else
 550                 b = slob_new_page(flags, get_order(c->size), node);
 551
 552         if (c->ctor)
 553                 c->ctor(c, b);
 554
 555         return b;
 556 }
 557 EXPORT_SYMBOL(kmem_cache_alloc_node);
 558
 559 static void __kmem_cache_free(void *b, int size)
 560 {
 561         if (size < PAGE_SIZE)
 562                 slob_free(b, size);
 563         else
 564                 free_pages((unsigned long)b, get_order(size));
 565 }
 566
 567 static void kmem_rcu_free(struct rcu_head *head)
 568 {
 569         struct slob_rcu *slob_rcu = (struct slob_rcu *)head;
 570         void *b = (void *)slob_rcu - (slob_rcu->size - sizeof(struct slob_rcu));
 571
 572         __kmem_cache_free(b, slob_rcu->size);
 573 }
 574
 575 void kmem_cache_free(struct kmem_cache *c, void *b)
 576 {
 577         if (unlikely(c->flags & SLAB_DESTROY_BY_RCU)) {
 578                 struct slob_rcu *slob_rcu;
 579                 slob_rcu = b + (c->size - sizeof(struct slob_rcu));
 580                 INIT_RCU_HEAD(&slob_rcu->head);
 581                 slob_rcu->size = c->size;
 582                 call_rcu(&slob_rcu->head, kmem_rcu_free);
 583         } else {
 584                 __kmem_cache_free(b, c->size);
 585         }
 586 }
 587 EXPORT_SYMBOL(kmem_cache_free);
 588
 589 unsigned int kmem_cache_size(struct kmem_cache *c)
 590 {
 591         return c->size;
 592 }
 593 EXPORT_SYMBOL(kmem_cache_size);
 594
 595 const char *kmem_cache_name(struct kmem_cache *c)
 596 {
 597         return c->name;
 598 }
 599 EXPORT_SYMBOL(kmem_cache_name);
 600
 601 int kmem_cache_shrink(struct kmem_cache *d)
 602 {
 603         return 0;
 604 }
 605 EXPORT_SYMBOL(kmem_cache_shrink);
 606
 607 int kmem_ptr_validate(struct kmem_cache *a, const void *b)
 608 {
 609         return 0;
 610 }
 611
 612 static unsigned int slob_ready __read_mostly;
 613
 614 int slab_is_available(void)
 615 {
 616         return slob_ready;
 617 }
 618
 619 void __init kmem_cache_init(void)
 620 {
 621         slob_ready = 1;
 622 }