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_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/cache.h>
  64 #include <linux/init.h>
  65 #include <linux/module.h>
  66 #include <linux/rcupdate.h>
  67 #include <linux/list.h>
  68 #include <asm/atomic.h>
  69
  70 /*
  71  * slob_block has a field 'units', which indicates size of block if +ve,
  72  * or offset of next block if -ve (in SLOB_UNITs).
  73  *
  74  * Free blocks of size 1 unit simply contain the offset of the next block.
  75  * Those with larger size contain their size in the first SLOB_UNIT of
  76  * memory, and the offset of the next free block in the second SLOB_UNIT.
  77  */
  78 #if PAGE_SIZE <= (32767 * 2)
  79 typedef s16 slobidx_t;
  80 #else
  81 typedef s32 slobidx_t;
  82 #endif
  83
  84 struct slob_block {
  85         slobidx_t units;
  86 };
  87 typedef struct slob_block slob_t;
  88
  89 /*
  90  * We use struct page fields to manage some slob allocation aspects,
  91  * however to avoid the horrible mess in include/linux/mm_types.h, we'll
  92  * just define our own struct page type variant here.
  93  */
  94 struct slob_page {
  95         union {
  96                 struct {
  97                         unsigned long flags;    /* mandatory */
  98                         atomic_t _count;        /* mandatory */
  99                         slobidx_t units;        /* free units left in page */
 100                         unsigned long pad[2];
 101                         slob_t *free;           /* first free slob_t in page */
 102                         struct list_head list;  /* linked list of free pages */
 103                 };
 104                 struct page page;
 105         };
 106 };
 107 static inline void struct_slob_page_wrong_size(void)
 108 { BUILD_BUG_ON(sizeof(struct slob_page) != sizeof(struct page)); }
 109
 110 /*
 111  * free_slob_page: call before a slob_page is returned to the page allocator.
 112  */
 113 static inline void free_slob_page(struct slob_page *sp)
 114 {
 115         reset_page_mapcount(&sp->page);
 116         sp->page.mapping = NULL;
 117 }
 118
 119 /*
 120  * All partially free slob pages go on these lists.
 121  */
 122 #define SLOB_BREAK1 256
 123 #define SLOB_BREAK2 1024
 124 static LIST_HEAD(free_slob_small);
 125 static LIST_HEAD(free_slob_medium);
 126 static LIST_HEAD(free_slob_large);
 127
 128 /*
 129  * slob_page: True for all slob pages (false for bigblock pages)
 130  */
 131 static inline int slob_page(struct slob_page *sp)
 132 {
 133         return test_bit(PG_active, &sp->flags);
 134 }
 135
 136 static inline void set_slob_page(struct slob_page *sp)
 137 {
 138         __set_bit(PG_active, &sp->flags);
 139 }
 140
 141 static inline void clear_slob_page(struct slob_page *sp)
 142 {
 143         __clear_bit(PG_active, &sp->flags);
 144 }
 145
 146 /*
 147  * slob_page_free: true for pages on free_slob_pages list.
 148  */
 149 static inline int slob_page_free(struct slob_page *sp)
 150 {
 151         return test_bit(PG_private, &sp->flags);
 152 }
 153
 154 static void set_slob_page_free(struct slob_page *sp, struct list_head *list)
 155 {
 156         list_add(&sp->list, list);
 157         __set_bit(PG_private, &sp->flags);
 158 }
 159
 160 static inline void clear_slob_page_free(struct slob_page *sp)
 161 {
 162         list_del(&sp->list);
 163         __clear_bit(PG_private, &sp->flags);
 164 }
 165
 166 #define SLOB_UNIT sizeof(slob_t)
 167 #define SLOB_UNITS(size) (((size) + SLOB_UNIT - 1)/SLOB_UNIT)
 168 #define SLOB_ALIGN L1_CACHE_BYTES
 169
 170 /*
 171  * struct slob_rcu is inserted at the tail of allocated slob blocks, which
 172  * were created with a SLAB_DESTROY_BY_RCU slab. slob_rcu is used to free
 173  * the block using call_rcu.
 174  */
 175 struct slob_rcu {
 176         struct rcu_head head;
 177         int size;
 178 };
 179
 180 /*
 181  * slob_lock protects all slob allocator structures.
 182  */
 183 static DEFINE_SPINLOCK(slob_lock);
 184
 185 /*
 186  * Encode the given size and next info into a free slob block s.
 187  */
 188 static void set_slob(slob_t *s, slobidx_t size, slob_t *next)
 189 {
 190         slob_t *base = (slob_t *)((unsigned long)s & PAGE_MASK);
 191         slobidx_t offset = next - base;
 192
 193         if (size > 1) {
 194                 s[0].units = size;
 195                 s[1].units = offset;
 196         } else
 197                 s[0].units = -offset;
 198 }
 199
 200 /*
 201  * Return the size of a slob block.
 202  */
 203 static slobidx_t slob_units(slob_t *s)
 204 {
 205         if (s->units > 0)
 206                 return s->units;
 207         return 1;
 208 }
 209
 210 /*
 211  * Return the next free slob block pointer after this one.
 212  */
 213 static slob_t *slob_next(slob_t *s)
 214 {
 215         slob_t *base = (slob_t *)((unsigned long)s & PAGE_MASK);
 216         slobidx_t next;
 217
 218         if (s[0].units < 0)
 219                 next = -s[0].units;
 220         else
 221                 next = s[1].units;
 222         return base+next;
 223 }
 224
 225 /*
 226  * Returns true if s is the last free block in its page.
 227  */
 228 static int slob_last(slob_t *s)
 229 {
 230         return !((unsigned long)slob_next(s) & ~PAGE_MASK);
 231 }
 232
 233 static void *slob_new_page(gfp_t gfp, int order, int node)
 234 {
 235         void *page;
 236
 237 #ifdef CONFIG_NUMA
 238         if (node != -1)
 239                 page = alloc_pages_node(node, gfp, order);
 240         else
 241 #endif
 242                 page = alloc_pages(gfp, order);
 243
 244         if (!page)
 245                 return NULL;
 246
 247         return page_address(page);
 248 }
 249
 250 /*
 251  * Allocate a slob block within a given slob_page sp.
 252  */
 253 static void *slob_page_alloc(struct slob_page *sp, size_t size, int align)
 254 {
 255         slob_t *prev, *cur, *aligned = 0;
 256         int delta = 0, units = SLOB_UNITS(size);
 257
 258         for (prev = NULL, cur = sp->free; ; prev = cur, cur = slob_next(cur)) {
 259                 slobidx_t avail = slob_units(cur);
 260
 261                 if (align) {
 262                         aligned = (slob_t *)ALIGN((unsigned long)cur, align);
 263                         delta = aligned - cur;
 264                 }
 265                 if (avail >= units + delta) { /* room enough? */
 266                         slob_t *next;
 267
 268                         if (delta) { /* need to fragment head to align? */
 269                                 next = slob_next(cur);
 270                                 set_slob(aligned, avail - delta, next);
 271                                 set_slob(cur, delta, aligned);
 272                                 prev = cur;
 273                                 cur = aligned;
 274                                 avail = slob_units(cur);
 275                         }
 276
 277                         next = slob_next(cur);
 278                         if (avail == units) { /* exact fit? unlink. */
 279                                 if (prev)
 280                                         set_slob(prev, slob_units(prev), next);
 281                                 else
 282                                         sp->free = next;
 283                         } else { /* fragment */
 284                                 if (prev)
 285                                         set_slob(prev, slob_units(prev), cur + units);
 286                                 else
 287                                         sp->free = cur + units;
 288                                 set_slob(cur + units, avail - units, next);
 289                         }
 290
 291                         sp->units -= units;
 292                         if (!sp->units)
 293                                 clear_slob_page_free(sp);
 294                         return cur;
 295                 }
 296                 if (slob_last(cur))
 297                         return NULL;
 298         }
 299 }
 300
 301 /*
 302  * slob_alloc: entry point into the slob allocator.
 303  */
 304 static void *slob_alloc(size_t size, gfp_t gfp, int align, int node)
 305 {
 306         struct slob_page *sp;
 307         struct list_head *prev;
 308         struct list_head *slob_list;
 309         slob_t *b = NULL;
 310         unsigned long flags;
 311
 312         if (size < SLOB_BREAK1)
 313                 slob_list = &free_slob_small;
 314         else if (size < SLOB_BREAK2)
 315                 slob_list = &free_slob_medium;
 316         else
 317                 slob_list = &free_slob_large;
 318
 319         spin_lock_irqsave(&slob_lock, flags);
 320         /* Iterate through each partially free page, try to find room */
 321         list_for_each_entry(sp, slob_list, list) {
 322 #ifdef CONFIG_NUMA
 323                 /*
 324                  * If there's a node specification, search for a partial
 325                  * page with a matching node id in the freelist.
 326                  */
 327                 if (node != -1 && page_to_nid(&sp->page) != node)
 328                         continue;
 329 #endif
 330                 /* Enough room on this page? */
 331                 if (sp->units < SLOB_UNITS(size))
 332                         continue;
 333
 334                 /* Attempt to alloc */
 335                 prev = sp->list.prev;
 336                 b = slob_page_alloc(sp, size, align);
 337                 if (!b)
 338                         continue;
 339
 340                 /* Improve fragment distribution and reduce our average
 341                  * search time by starting our next search here. (see
 342                  * Knuth vol 1, sec 2.5, pg 449) */
 343                 if (prev != slob_list->prev &&
 344                                 slob_list->next != prev->next)
 345                         list_move_tail(slob_list, prev->next);
 346                 break;
 347         }
 348         spin_unlock_irqrestore(&slob_lock, flags);
 349
 350         /* Not enough space: must allocate a new page */
 351         if (!b) {
 352                 b = slob_new_page(gfp & ~__GFP_ZERO, 0, node);
 353                 if (!b)
 354                         return 0;
 355                 sp = (struct slob_page *)virt_to_page(b);
 356                 set_slob_page(sp);
 357
 358                 spin_lock_irqsave(&slob_lock, flags);
 359                 sp->units = SLOB_UNITS(PAGE_SIZE);
 360                 sp->free = b;
 361                 INIT_LIST_HEAD(&sp->list);
 362                 set_slob(b, SLOB_UNITS(PAGE_SIZE), b + SLOB_UNITS(PAGE_SIZE));
 363                 set_slob_page_free(sp, slob_list);
 364                 b = slob_page_alloc(sp, size, align);
 365                 BUG_ON(!b);
 366                 spin_unlock_irqrestore(&slob_lock, flags);
 367         }
 368         if (unlikely((gfp & __GFP_ZERO) && b))
 369                 memset(b, 0, size);
 370         return b;
 371 }
 372
 373 /*
 374  * slob_free: entry point into the slob allocator.
 375  */
 376 static void slob_free(void *block, int size)
 377 {
 378         struct slob_page *sp;
 379         slob_t *prev, *next, *b = (slob_t *)block;
 380         slobidx_t units;
 381         unsigned long flags;
 382
 383         if (unlikely(ZERO_OR_NULL_PTR(block)))
 384                 return;
 385         BUG_ON(!size);
 386
 387         sp = (struct slob_page *)virt_to_page(block);
 388         units = SLOB_UNITS(size);
 389
 390         spin_lock_irqsave(&slob_lock, flags);
 391
 392         if (sp->units + units == SLOB_UNITS(PAGE_SIZE)) {
 393                 /* Go directly to page allocator. Do not pass slob allocator */
 394                 if (slob_page_free(sp))
 395                         clear_slob_page_free(sp);
 396                 clear_slob_page(sp);
 397                 free_slob_page(sp);
 398                 free_page((unsigned long)b);
 399                 goto out;
 400         }
 401
 402         if (!slob_page_free(sp)) {
 403                 /* This slob page is about to become partially free. Easy! */
 404                 sp->units = units;
 405                 sp->free = b;
 406                 set_slob(b, units,
 407                         (void *)((unsigned long)(b +
 408                                         SLOB_UNITS(PAGE_SIZE)) & PAGE_MASK));
 409                 set_slob_page_free(sp, &free_slob_small);
 410                 goto out;
 411         }
 412
 413         /*
 414          * Otherwise the page is already partially free, so find reinsertion
 415          * point.
 416          */
 417         sp->units += units;
 418
 419         if (b < sp->free) {
 420                 if (b + units == sp->free) {
 421                         units += slob_units(sp->free);
 422                         sp->free = slob_next(sp->free);
 423                 }
 424                 set_slob(b, units, sp->free);
 425                 sp->free = b;
 426         } else {
 427                 prev = sp->free;
 428                 next = slob_next(prev);
 429                 while (b > next) {
 430                         prev = next;
 431                         next = slob_next(prev);
 432                 }
 433
 434                 if (!slob_last(prev) && b + units == next) {
 435                         units += slob_units(next);
 436                         set_slob(b, units, slob_next(next));
 437                 } else
 438                         set_slob(b, units, next);
 439
 440                 if (prev + slob_units(prev) == b) {
 441                         units = slob_units(b) + slob_units(prev);
 442                         set_slob(prev, units, slob_next(b));
 443                 } else
 444                         set_slob(prev, slob_units(prev), b);
 445         }
 446 out:
 447         spin_unlock_irqrestore(&slob_lock, flags);
 448 }
 449
 450 /*
 451  * End of slob allocator proper. Begin kmem_cache_alloc and kmalloc frontend.
 452  */
 453
 454 #ifndef ARCH_KMALLOC_MINALIGN
 455 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long)
 456 #endif
 457
 458 #ifndef ARCH_SLAB_MINALIGN
 459 #define ARCH_SLAB_MINALIGN __alignof__(unsigned long)
 460 #endif
 461
 462 void *__kmalloc_node(size_t size, gfp_t gfp, int node)
 463 {
 464         unsigned int *m;
 465         int align = max(ARCH_KMALLOC_MINALIGN, ARCH_SLAB_MINALIGN);
 466
 467         if (size < PAGE_SIZE - align) {
 468                 if (!size)
 469                         return ZERO_SIZE_PTR;
 470
 471                 m = slob_alloc(size + align, gfp, align, node);
 472                 if (m)
 473                         *m = size;
 474                 return (void *)m + align;
 475         } else {
 476                 void *ret;
 477
 478                 ret = slob_new_page(gfp | __GFP_COMP, get_order(size), node);
 479                 if (ret) {
 480                         struct page *page;
 481                         page = virt_to_page(ret);
 482                         page->private = size;
 483                 }
 484                 return ret;
 485         }
 486 }
 487 EXPORT_SYMBOL(__kmalloc_node);
 488
 489 void kfree(const void *block)
 490 {
 491         struct slob_page *sp;
 492
 493         if (unlikely(ZERO_OR_NULL_PTR(block)))
 494                 return;
 495
 496         sp = (struct slob_page *)virt_to_page(block);
 497         if (slob_page(sp)) {
 498                 int align = max(ARCH_KMALLOC_MINALIGN, ARCH_SLAB_MINALIGN);
 499                 unsigned int *m = (unsigned int *)(block - align);
 500                 slob_free(m, *m + align);
 501         } else
 502                 put_page(&sp->page);
 503 }
 504 EXPORT_SYMBOL(kfree);
 505
 506 /* can't use ksize for kmem_cache_alloc memory, only kmalloc */
 507 size_t ksize(const void *block)
 508 {
 509         struct slob_page *sp;
 510
 511         BUG_ON(!block);
 512         if (unlikely(block == ZERO_SIZE_PTR))
 513                 return 0;
 514
 515         sp = (struct slob_page *)virt_to_page(block);
 516         if (slob_page(sp))
 517                 return ((slob_t *)block - 1)->units + SLOB_UNIT;
 518         else
 519                 return sp->page.private;
 520 }
 521 EXPORT_SYMBOL(ksize);
 522
 523 struct kmem_cache {
 524         unsigned int size, align;
 525         unsigned long flags;
 526         const char *name;
 527         void (*ctor)(struct kmem_cache *, void *);
 528 };
 529
 530 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
 531         size_t align, unsigned long flags,
 532         void (*ctor)(struct kmem_cache *, void *))
 533 {
 534         struct kmem_cache *c;
 535
 536         c = slob_alloc(sizeof(struct kmem_cache), flags, 0, -1);
 537
 538         if (c) {
 539                 c->name = name;
 540                 c->size = size;
 541                 if (flags & SLAB_DESTROY_BY_RCU) {
 542                         /* leave room for rcu footer at the end of object */
 543                         c->size += sizeof(struct slob_rcu);
 544                 }
 545                 c->flags = flags;
 546                 c->ctor = ctor;
 547                 /* ignore alignment unless it's forced */
 548                 c->align = (flags & SLAB_HWCACHE_ALIGN) ? SLOB_ALIGN : 0;
 549                 if (c->align < ARCH_SLAB_MINALIGN)
 550                         c->align = ARCH_SLAB_MINALIGN;
 551                 if (c->align < align)
 552                         c->align = align;
 553         } else if (flags & SLAB_PANIC)
 554                 panic("Cannot create slab cache %s\n", name);
 555
 556         return c;
 557 }
 558 EXPORT_SYMBOL(kmem_cache_create);
 559
 560 void kmem_cache_destroy(struct kmem_cache *c)
 561 {
 562         slob_free(c, sizeof(struct kmem_cache));
 563 }
 564 EXPORT_SYMBOL(kmem_cache_destroy);
 565
 566 void *kmem_cache_alloc_node(struct kmem_cache *c, gfp_t flags, int node)
 567 {
 568         void *b;
 569
 570         if (c->size < PAGE_SIZE)
 571                 b = slob_alloc(c->size, flags, c->align, node);
 572         else
 573                 b = slob_new_page(flags, get_order(c->size), node);
 574
 575         if (c->ctor)
 576                 c->ctor(c, b);
 577
 578         return b;
 579 }
 580 EXPORT_SYMBOL(kmem_cache_alloc_node);
 581
 582 static void __kmem_cache_free(void *b, int size)
 583 {
 584         if (size < PAGE_SIZE)
 585                 slob_free(b, size);
 586         else
 587                 free_pages((unsigned long)b, get_order(size));
 588 }
 589
 590 static void kmem_rcu_free(struct rcu_head *head)
 591 {
 592         struct slob_rcu *slob_rcu = (struct slob_rcu *)head;
 593         void *b = (void *)slob_rcu - (slob_rcu->size - sizeof(struct slob_rcu));
 594
 595         __kmem_cache_free(b, slob_rcu->size);
 596 }
 597
 598 void kmem_cache_free(struct kmem_cache *c, void *b)
 599 {
 600         if (unlikely(c->flags & SLAB_DESTROY_BY_RCU)) {
 601                 struct slob_rcu *slob_rcu;
 602                 slob_rcu = b + (c->size - sizeof(struct slob_rcu));
 603                 INIT_RCU_HEAD(&slob_rcu->head);
 604                 slob_rcu->size = c->size;
 605                 call_rcu(&slob_rcu->head, kmem_rcu_free);
 606         } else {
 607                 __kmem_cache_free(b, c->size);
 608         }
 609 }
 610 EXPORT_SYMBOL(kmem_cache_free);
 611
 612 unsigned int kmem_cache_size(struct kmem_cache *c)
 613 {
 614         return c->size;
 615 }
 616 EXPORT_SYMBOL(kmem_cache_size);
 617
 618 const char *kmem_cache_name(struct kmem_cache *c)
 619 {
 620         return c->name;
 621 }
 622 EXPORT_SYMBOL(kmem_cache_name);
 623
 624 int kmem_cache_shrink(struct kmem_cache *d)
 625 {
 626         return 0;
 627 }
 628 EXPORT_SYMBOL(kmem_cache_shrink);
 629
 630 int kmem_ptr_validate(struct kmem_cache *a, const void *b)
 631 {
 632         return 0;
 633 }
 634
 635 static unsigned int slob_ready __read_mostly;
 636
 637 int slab_is_available(void)
 638 {
 639         return slob_ready;
 640 }
 641
 642 void __init kmem_cache_init(void)
 643 {
 644         slob_ready = 1;
 645 }