2 * SLOB Allocator: Simple List Of Blocks
4 * Matt Mackall <mpm@selenic.com> 12/30/03
6 * NUMA support by Paul Mundt, 2007.
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.
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
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()
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
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().
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.
53 #include <linux/kernel.h>
54 #include <linux/slab.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>
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).
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.
71 #if PAGE_SIZE <= (32767 * 2)
72 typedef s16 slobidx_t;
74 typedef s32 slobidx_t;
80 typedef struct slob_block slob_t;
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.
90 unsigned long flags; /* mandatory */
91 atomic_t _count; /* mandatory */
92 slobidx_t units; /* free units left in page */
94 slob_t *free; /* first free slob_t in page */
95 struct list_head list; /* linked list of free pages */
100 static inline void struct_slob_page_wrong_size(void)
101 { BUILD_BUG_ON(sizeof(struct slob_page) != sizeof(struct page)); }
104 * free_slob_page: call before a slob_page is returned to the page allocator.
106 static inline void free_slob_page(struct slob_page *sp)
108 reset_page_mapcount(&sp->page);
109 sp->page.mapping = NULL;
113 * All (partially) free slob pages go on this list.
115 static LIST_HEAD(free_slob_pages);
118 * slob_page: True for all slob pages (false for bigblock pages)
120 static inline int slob_page(struct slob_page *sp)
122 return test_bit(PG_active, &sp->flags);
125 static inline void set_slob_page(struct slob_page *sp)
127 __set_bit(PG_active, &sp->flags);
130 static inline void clear_slob_page(struct slob_page *sp)
132 __clear_bit(PG_active, &sp->flags);
136 * slob_page_free: true for pages on free_slob_pages list.
138 static inline int slob_page_free(struct slob_page *sp)
140 return test_bit(PG_private, &sp->flags);
143 static inline void set_slob_page_free(struct slob_page *sp)
145 list_add(&sp->list, &free_slob_pages);
146 __set_bit(PG_private, &sp->flags);
149 static inline void clear_slob_page_free(struct slob_page *sp)
152 __clear_bit(PG_private, &sp->flags);
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
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.
165 struct rcu_head head;
170 * slob_lock protects all slob allocator structures.
172 static DEFINE_SPINLOCK(slob_lock);
175 * Encode the given size and next info into a free slob block s.
177 static void set_slob(slob_t *s, slobidx_t size, slob_t *next)
179 slob_t *base = (slob_t *)((unsigned long)s & PAGE_MASK);
180 slobidx_t offset = next - base;
186 s[0].units = -offset;
190 * Return the size of a slob block.
192 static slobidx_t slob_units(slob_t *s)
200 * Return the next free slob block pointer after this one.
202 static slob_t *slob_next(slob_t *s)
204 slob_t *base = (slob_t *)((unsigned long)s & PAGE_MASK);
215 * Returns true if s is the last free block in its page.
217 static int slob_last(slob_t *s)
219 return !((unsigned long)slob_next(s) & ~PAGE_MASK);
222 static void *slob_new_page(gfp_t gfp, int order, int node)
228 page = alloc_pages_node(node, gfp, order);
231 page = alloc_pages(gfp, order);
236 return page_address(page);
240 * Allocate a slob block within a given slob_page sp.
242 static void *slob_page_alloc(struct slob_page *sp, size_t size, int align)
244 slob_t *prev, *cur, *aligned = 0;
245 int delta = 0, units = SLOB_UNITS(size);
247 for (prev = NULL, cur = sp->free; ; prev = cur, cur = slob_next(cur)) {
248 slobidx_t avail = slob_units(cur);
251 aligned = (slob_t *)ALIGN((unsigned long)cur, align);
252 delta = aligned - cur;
254 if (avail >= units + delta) { /* room enough? */
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);
263 avail = slob_units(cur);
266 next = slob_next(cur);
267 if (avail == units) { /* exact fit? unlink. */
269 set_slob(prev, slob_units(prev), next);
272 } else { /* fragment */
274 set_slob(prev, slob_units(prev), cur + units);
276 sp->free = cur + units;
277 set_slob(cur + units, avail - units, next);
282 clear_slob_page_free(sp);
291 * slob_alloc: entry point into the slob allocator.
293 static void *slob_alloc(size_t size, gfp_t gfp, int align, int node)
295 struct slob_page *sp;
296 struct list_head *prev;
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) {
305 * If there's a node specification, search for a partial
306 * page with a matching node id in the freelist.
308 if (node != -1 && page_to_nid(&sp->page) != node)
311 /* Enough room on this page? */
312 if (sp->units < SLOB_UNITS(size))
315 /* Attempt to alloc */
316 prev = sp->list.prev;
317 b = slob_page_alloc(sp, size, align);
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 (free_slob_pages.next != prev->next)
325 list_move_tail(&free_slob_pages, prev->next);
328 spin_unlock_irqrestore(&slob_lock, flags);
330 /* Not enough space: must allocate a new page */
332 b = slob_new_page(gfp, 0, node);
335 sp = (struct slob_page *)virt_to_page(b);
338 spin_lock_irqsave(&slob_lock, flags);
339 sp->units = SLOB_UNITS(PAGE_SIZE);
341 INIT_LIST_HEAD(&sp->list);
342 set_slob(b, SLOB_UNITS(PAGE_SIZE), b + SLOB_UNITS(PAGE_SIZE));
343 set_slob_page_free(sp);
344 b = slob_page_alloc(sp, size, align);
346 spin_unlock_irqrestore(&slob_lock, flags);
348 if (unlikely((gfp & __GFP_ZERO) && b))
354 * slob_free: entry point into the slob allocator.
356 static void slob_free(void *block, int size)
358 struct slob_page *sp;
359 slob_t *prev, *next, *b = (slob_t *)block;
363 if (unlikely(ZERO_OR_NULL_PTR(block)))
367 sp = (struct slob_page *)virt_to_page(block);
368 units = SLOB_UNITS(size);
370 spin_lock_irqsave(&slob_lock, flags);
372 if (sp->units + units == SLOB_UNITS(PAGE_SIZE)) {
373 /* Go directly to page allocator. Do not pass slob allocator */
374 if (slob_page_free(sp))
375 clear_slob_page_free(sp);
378 free_page((unsigned long)b);
382 if (!slob_page_free(sp)) {
383 /* This slob page is about to become partially free. Easy! */
387 (void *)((unsigned long)(b +
388 SLOB_UNITS(PAGE_SIZE)) & PAGE_MASK));
389 set_slob_page_free(sp);
394 * Otherwise the page is already partially free, so find reinsertion
400 set_slob(b, units, sp->free);
404 next = slob_next(prev);
407 next = slob_next(prev);
410 if (!slob_last(prev) && b + units == next) {
411 units += slob_units(next);
412 set_slob(b, units, slob_next(next));
414 set_slob(b, units, next);
416 if (prev + slob_units(prev) == b) {
417 units = slob_units(b) + slob_units(prev);
418 set_slob(prev, units, slob_next(b));
420 set_slob(prev, slob_units(prev), b);
423 spin_unlock_irqrestore(&slob_lock, flags);
427 * End of slob allocator proper. Begin kmem_cache_alloc and kmalloc frontend.
430 #ifndef ARCH_KMALLOC_MINALIGN
431 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long)
434 #ifndef ARCH_SLAB_MINALIGN
435 #define ARCH_SLAB_MINALIGN __alignof__(unsigned long)
438 void *__kmalloc_node(size_t size, gfp_t gfp, int node)
441 int align = max(ARCH_KMALLOC_MINALIGN, ARCH_SLAB_MINALIGN);
443 if (size < PAGE_SIZE - align) {
445 return ZERO_SIZE_PTR;
447 m = slob_alloc(size + align, gfp, align, node);
450 return (void *)m + align;
454 ret = slob_new_page(gfp | __GFP_COMP, get_order(size), node);
457 page = virt_to_page(ret);
458 page->private = size;
463 EXPORT_SYMBOL(__kmalloc_node);
465 void kfree(const void *block)
467 struct slob_page *sp;
469 if (unlikely(ZERO_OR_NULL_PTR(block)))
472 sp = (struct slob_page *)virt_to_page(block);
474 int align = max(ARCH_KMALLOC_MINALIGN, ARCH_SLAB_MINALIGN);
475 unsigned int *m = (unsigned int *)(block - align);
476 slob_free(m, *m + align);
480 EXPORT_SYMBOL(kfree);
482 /* can't use ksize for kmem_cache_alloc memory, only kmalloc */
483 size_t ksize(const void *block)
485 struct slob_page *sp;
488 if (unlikely(block == ZERO_SIZE_PTR))
491 sp = (struct slob_page *)virt_to_page(block);
493 return ((slob_t *)block - 1)->units + SLOB_UNIT;
495 return sp->page.private;
499 unsigned int size, align;
502 void (*ctor)(struct kmem_cache *, void *);
505 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
506 size_t align, unsigned long flags,
507 void (*ctor)(struct kmem_cache *, void *))
509 struct kmem_cache *c;
511 c = slob_alloc(sizeof(struct kmem_cache), flags, 0, -1);
516 if (flags & SLAB_DESTROY_BY_RCU) {
517 /* leave room for rcu footer at the end of object */
518 c->size += sizeof(struct slob_rcu);
522 /* ignore alignment unless it's forced */
523 c->align = (flags & SLAB_HWCACHE_ALIGN) ? SLOB_ALIGN : 0;
524 if (c->align < ARCH_SLAB_MINALIGN)
525 c->align = ARCH_SLAB_MINALIGN;
526 if (c->align < align)
528 } else if (flags & SLAB_PANIC)
529 panic("Cannot create slab cache %s\n", name);
533 EXPORT_SYMBOL(kmem_cache_create);
535 void kmem_cache_destroy(struct kmem_cache *c)
537 slob_free(c, sizeof(struct kmem_cache));
539 EXPORT_SYMBOL(kmem_cache_destroy);
541 void *kmem_cache_alloc_node(struct kmem_cache *c, gfp_t flags, int node)
545 if (c->size < PAGE_SIZE)
546 b = slob_alloc(c->size, flags, c->align, node);
548 b = slob_new_page(flags, get_order(c->size), node);
555 EXPORT_SYMBOL(kmem_cache_alloc_node);
557 static void __kmem_cache_free(void *b, int size)
559 if (size < PAGE_SIZE)
562 free_pages((unsigned long)b, get_order(size));
565 static void kmem_rcu_free(struct rcu_head *head)
567 struct slob_rcu *slob_rcu = (struct slob_rcu *)head;
568 void *b = (void *)slob_rcu - (slob_rcu->size - sizeof(struct slob_rcu));
570 __kmem_cache_free(b, slob_rcu->size);
573 void kmem_cache_free(struct kmem_cache *c, void *b)
575 if (unlikely(c->flags & SLAB_DESTROY_BY_RCU)) {
576 struct slob_rcu *slob_rcu;
577 slob_rcu = b + (c->size - sizeof(struct slob_rcu));
578 INIT_RCU_HEAD(&slob_rcu->head);
579 slob_rcu->size = c->size;
580 call_rcu(&slob_rcu->head, kmem_rcu_free);
582 __kmem_cache_free(b, c->size);
585 EXPORT_SYMBOL(kmem_cache_free);
587 unsigned int kmem_cache_size(struct kmem_cache *c)
591 EXPORT_SYMBOL(kmem_cache_size);
593 const char *kmem_cache_name(struct kmem_cache *c)
597 EXPORT_SYMBOL(kmem_cache_name);
599 int kmem_cache_shrink(struct kmem_cache *d)
603 EXPORT_SYMBOL(kmem_cache_shrink);
605 int kmem_ptr_validate(struct kmem_cache *a, const void *b)
610 static unsigned int slob_ready __read_mostly;
612 int slab_is_available(void)
617 void __init kmem_cache_init(void)