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 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.
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.
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()
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
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().
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.
60 #include <linux/kernel.h>
61 #include <linux/slab.h>
63 #include <linux/swap.h> /* struct reclaim_state */
64 #include <linux/cache.h>
65 #include <linux/init.h>
66 #include <linux/module.h>
67 #include <linux/rcupdate.h>
68 #include <linux/list.h>
69 #include <trace/kmemtrace.h>
70 #include <asm/atomic.h>
73 * slob_block has a field 'units', which indicates size of block if +ve,
74 * or offset of next block if -ve (in SLOB_UNITs).
76 * Free blocks of size 1 unit simply contain the offset of the next block.
77 * Those with larger size contain their size in the first SLOB_UNIT of
78 * memory, and the offset of the next free block in the second SLOB_UNIT.
80 #if PAGE_SIZE <= (32767 * 2)
81 typedef s16 slobidx_t;
83 typedef s32 slobidx_t;
89 typedef struct slob_block slob_t;
92 * We use struct page fields to manage some slob allocation aspects,
93 * however to avoid the horrible mess in include/linux/mm_types.h, we'll
94 * just define our own struct page type variant here.
99 unsigned long flags; /* mandatory */
100 atomic_t _count; /* mandatory */
101 slobidx_t units; /* free units left in page */
102 unsigned long pad[2];
103 slob_t *free; /* first free slob_t in page */
104 struct list_head list; /* linked list of free pages */
109 static inline void struct_slob_page_wrong_size(void)
110 { BUILD_BUG_ON(sizeof(struct slob_page) != sizeof(struct page)); }
113 * free_slob_page: call before a slob_page is returned to the page allocator.
115 static inline void free_slob_page(struct slob_page *sp)
117 reset_page_mapcount(&sp->page);
118 sp->page.mapping = NULL;
122 * All partially free slob pages go on these lists.
124 #define SLOB_BREAK1 256
125 #define SLOB_BREAK2 1024
126 static LIST_HEAD(free_slob_small);
127 static LIST_HEAD(free_slob_medium);
128 static LIST_HEAD(free_slob_large);
131 * is_slob_page: True for all slob pages (false for bigblock pages)
133 static inline int is_slob_page(struct slob_page *sp)
135 return PageSlobPage((struct page *)sp);
138 static inline void set_slob_page(struct slob_page *sp)
140 __SetPageSlobPage((struct page *)sp);
143 static inline void clear_slob_page(struct slob_page *sp)
145 __ClearPageSlobPage((struct page *)sp);
148 static inline struct slob_page *slob_page(const void *addr)
150 return (struct slob_page *)virt_to_page(addr);
154 * slob_page_free: true for pages on free_slob_pages list.
156 static inline int slob_page_free(struct slob_page *sp)
158 return PageSlobFree((struct page *)sp);
161 static void set_slob_page_free(struct slob_page *sp, struct list_head *list)
163 list_add(&sp->list, list);
164 __SetPageSlobFree((struct page *)sp);
167 static inline void clear_slob_page_free(struct slob_page *sp)
170 __ClearPageSlobFree((struct page *)sp);
173 #define SLOB_UNIT sizeof(slob_t)
174 #define SLOB_UNITS(size) (((size) + SLOB_UNIT - 1)/SLOB_UNIT)
175 #define SLOB_ALIGN L1_CACHE_BYTES
178 * struct slob_rcu is inserted at the tail of allocated slob blocks, which
179 * were created with a SLAB_DESTROY_BY_RCU slab. slob_rcu is used to free
180 * the block using call_rcu.
183 struct rcu_head head;
188 * slob_lock protects all slob allocator structures.
190 static DEFINE_SPINLOCK(slob_lock);
193 * Encode the given size and next info into a free slob block s.
195 static void set_slob(slob_t *s, slobidx_t size, slob_t *next)
197 slob_t *base = (slob_t *)((unsigned long)s & PAGE_MASK);
198 slobidx_t offset = next - base;
204 s[0].units = -offset;
208 * Return the size of a slob block.
210 static slobidx_t slob_units(slob_t *s)
218 * Return the next free slob block pointer after this one.
220 static slob_t *slob_next(slob_t *s)
222 slob_t *base = (slob_t *)((unsigned long)s & PAGE_MASK);
233 * Returns true if s is the last free block in its page.
235 static int slob_last(slob_t *s)
237 return !((unsigned long)slob_next(s) & ~PAGE_MASK);
240 static void *slob_new_pages(gfp_t gfp, int order, int node)
246 page = alloc_pages_node(node, gfp, order);
249 page = alloc_pages(gfp, order);
254 return page_address(page);
257 static void slob_free_pages(void *b, int order)
259 if (current->reclaim_state)
260 current->reclaim_state->reclaimed_slab += 1 << order;
261 free_pages((unsigned long)b, order);
265 * Allocate a slob block within a given slob_page sp.
267 static void *slob_page_alloc(struct slob_page *sp, size_t size, int align)
269 slob_t *prev, *cur, *aligned = NULL;
270 int delta = 0, units = SLOB_UNITS(size);
272 for (prev = NULL, cur = sp->free; ; prev = cur, cur = slob_next(cur)) {
273 slobidx_t avail = slob_units(cur);
276 aligned = (slob_t *)ALIGN((unsigned long)cur, align);
277 delta = aligned - cur;
279 if (avail >= units + delta) { /* room enough? */
282 if (delta) { /* need to fragment head to align? */
283 next = slob_next(cur);
284 set_slob(aligned, avail - delta, next);
285 set_slob(cur, delta, aligned);
288 avail = slob_units(cur);
291 next = slob_next(cur);
292 if (avail == units) { /* exact fit? unlink. */
294 set_slob(prev, slob_units(prev), next);
297 } else { /* fragment */
299 set_slob(prev, slob_units(prev), cur + units);
301 sp->free = cur + units;
302 set_slob(cur + units, avail - units, next);
307 clear_slob_page_free(sp);
316 * slob_alloc: entry point into the slob allocator.
318 static void *slob_alloc(size_t size, gfp_t gfp, int align, int node)
320 struct slob_page *sp;
321 struct list_head *prev;
322 struct list_head *slob_list;
326 if (size < SLOB_BREAK1)
327 slob_list = &free_slob_small;
328 else if (size < SLOB_BREAK2)
329 slob_list = &free_slob_medium;
331 slob_list = &free_slob_large;
333 spin_lock_irqsave(&slob_lock, flags);
334 /* Iterate through each partially free page, try to find room */
335 list_for_each_entry(sp, slob_list, list) {
338 * If there's a node specification, search for a partial
339 * page with a matching node id in the freelist.
341 if (node != -1 && page_to_nid(&sp->page) != node)
344 /* Enough room on this page? */
345 if (sp->units < SLOB_UNITS(size))
348 /* Attempt to alloc */
349 prev = sp->list.prev;
350 b = slob_page_alloc(sp, size, align);
354 /* Improve fragment distribution and reduce our average
355 * search time by starting our next search here. (see
356 * Knuth vol 1, sec 2.5, pg 449) */
357 if (prev != slob_list->prev &&
358 slob_list->next != prev->next)
359 list_move_tail(slob_list, prev->next);
362 spin_unlock_irqrestore(&slob_lock, flags);
364 /* Not enough space: must allocate a new page */
366 b = slob_new_pages(gfp & ~__GFP_ZERO, 0, node);
372 spin_lock_irqsave(&slob_lock, flags);
373 sp->units = SLOB_UNITS(PAGE_SIZE);
375 INIT_LIST_HEAD(&sp->list);
376 set_slob(b, SLOB_UNITS(PAGE_SIZE), b + SLOB_UNITS(PAGE_SIZE));
377 set_slob_page_free(sp, slob_list);
378 b = slob_page_alloc(sp, size, align);
380 spin_unlock_irqrestore(&slob_lock, flags);
382 if (unlikely((gfp & __GFP_ZERO) && b))
388 * slob_free: entry point into the slob allocator.
390 static void slob_free(void *block, int size)
392 struct slob_page *sp;
393 slob_t *prev, *next, *b = (slob_t *)block;
397 if (unlikely(ZERO_OR_NULL_PTR(block)))
401 sp = slob_page(block);
402 units = SLOB_UNITS(size);
404 spin_lock_irqsave(&slob_lock, flags);
406 if (sp->units + units == SLOB_UNITS(PAGE_SIZE)) {
407 /* Go directly to page allocator. Do not pass slob allocator */
408 if (slob_page_free(sp))
409 clear_slob_page_free(sp);
410 spin_unlock_irqrestore(&slob_lock, flags);
413 slob_free_pages(b, 0);
417 if (!slob_page_free(sp)) {
418 /* This slob page is about to become partially free. Easy! */
422 (void *)((unsigned long)(b +
423 SLOB_UNITS(PAGE_SIZE)) & PAGE_MASK));
424 set_slob_page_free(sp, &free_slob_small);
429 * Otherwise the page is already partially free, so find reinsertion
435 if (b + units == sp->free) {
436 units += slob_units(sp->free);
437 sp->free = slob_next(sp->free);
439 set_slob(b, units, sp->free);
443 next = slob_next(prev);
446 next = slob_next(prev);
449 if (!slob_last(prev) && b + units == next) {
450 units += slob_units(next);
451 set_slob(b, units, slob_next(next));
453 set_slob(b, units, next);
455 if (prev + slob_units(prev) == b) {
456 units = slob_units(b) + slob_units(prev);
457 set_slob(prev, units, slob_next(b));
459 set_slob(prev, slob_units(prev), b);
462 spin_unlock_irqrestore(&slob_lock, flags);
466 * End of slob allocator proper. Begin kmem_cache_alloc and kmalloc frontend.
469 #ifndef ARCH_KMALLOC_MINALIGN
470 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long)
473 #ifndef ARCH_SLAB_MINALIGN
474 #define ARCH_SLAB_MINALIGN __alignof__(unsigned long)
477 void *__kmalloc_node(size_t size, gfp_t gfp, int node)
480 int align = max(ARCH_KMALLOC_MINALIGN, ARCH_SLAB_MINALIGN);
483 lockdep_trace_alloc(gfp);
485 if (size < PAGE_SIZE - align) {
487 return ZERO_SIZE_PTR;
489 m = slob_alloc(size + align, gfp, align, node);
494 ret = (void *)m + align;
496 trace_kmalloc_node(_RET_IP_, ret,
497 size, size + align, gfp, node);
499 unsigned int order = get_order(size);
501 ret = slob_new_pages(gfp | __GFP_COMP, get_order(size), node);
504 page = virt_to_page(ret);
505 page->private = size;
508 trace_kmalloc_node(_RET_IP_, ret,
509 size, PAGE_SIZE << order, gfp, node);
514 EXPORT_SYMBOL(__kmalloc_node);
516 void kfree(const void *block)
518 struct slob_page *sp;
520 trace_kfree(_RET_IP_, block);
522 if (unlikely(ZERO_OR_NULL_PTR(block)))
525 sp = slob_page(block);
526 if (is_slob_page(sp)) {
527 int align = max(ARCH_KMALLOC_MINALIGN, ARCH_SLAB_MINALIGN);
528 unsigned int *m = (unsigned int *)(block - align);
529 slob_free(m, *m + align);
533 EXPORT_SYMBOL(kfree);
535 /* can't use ksize for kmem_cache_alloc memory, only kmalloc */
536 size_t ksize(const void *block)
538 struct slob_page *sp;
541 if (unlikely(block == ZERO_SIZE_PTR))
544 sp = slob_page(block);
545 if (is_slob_page(sp)) {
546 int align = max(ARCH_KMALLOC_MINALIGN, ARCH_SLAB_MINALIGN);
547 unsigned int *m = (unsigned int *)(block - align);
548 return SLOB_UNITS(*m) * SLOB_UNIT;
550 return sp->page.private;
552 EXPORT_SYMBOL(ksize);
555 unsigned int size, align;
558 void (*ctor)(void *);
561 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
562 size_t align, unsigned long flags, void (*ctor)(void *))
564 struct kmem_cache *c;
566 c = slob_alloc(sizeof(struct kmem_cache),
567 GFP_KERNEL, ARCH_KMALLOC_MINALIGN, -1);
572 if (flags & SLAB_DESTROY_BY_RCU) {
573 /* leave room for rcu footer at the end of object */
574 c->size += sizeof(struct slob_rcu);
578 /* ignore alignment unless it's forced */
579 c->align = (flags & SLAB_HWCACHE_ALIGN) ? SLOB_ALIGN : 0;
580 if (c->align < ARCH_SLAB_MINALIGN)
581 c->align = ARCH_SLAB_MINALIGN;
582 if (c->align < align)
584 } else if (flags & SLAB_PANIC)
585 panic("Cannot create slab cache %s\n", name);
589 EXPORT_SYMBOL(kmem_cache_create);
591 void kmem_cache_destroy(struct kmem_cache *c)
593 slob_free(c, sizeof(struct kmem_cache));
595 EXPORT_SYMBOL(kmem_cache_destroy);
597 void *kmem_cache_alloc_node(struct kmem_cache *c, gfp_t flags, int node)
601 if (c->size < PAGE_SIZE) {
602 b = slob_alloc(c->size, flags, c->align, node);
603 trace_kmem_cache_alloc_node(_RET_IP_, b, c->size,
604 SLOB_UNITS(c->size) * SLOB_UNIT,
607 b = slob_new_pages(flags, get_order(c->size), node);
608 trace_kmem_cache_alloc_node(_RET_IP_, b, c->size,
609 PAGE_SIZE << get_order(c->size),
618 EXPORT_SYMBOL(kmem_cache_alloc_node);
620 static void __kmem_cache_free(void *b, int size)
622 if (size < PAGE_SIZE)
625 slob_free_pages(b, get_order(size));
628 static void kmem_rcu_free(struct rcu_head *head)
630 struct slob_rcu *slob_rcu = (struct slob_rcu *)head;
631 void *b = (void *)slob_rcu - (slob_rcu->size - sizeof(struct slob_rcu));
633 __kmem_cache_free(b, slob_rcu->size);
636 void kmem_cache_free(struct kmem_cache *c, void *b)
638 if (unlikely(c->flags & SLAB_DESTROY_BY_RCU)) {
639 struct slob_rcu *slob_rcu;
640 slob_rcu = b + (c->size - sizeof(struct slob_rcu));
641 INIT_RCU_HEAD(&slob_rcu->head);
642 slob_rcu->size = c->size;
643 call_rcu(&slob_rcu->head, kmem_rcu_free);
645 __kmem_cache_free(b, c->size);
648 trace_kmem_cache_free(_RET_IP_, b);
650 EXPORT_SYMBOL(kmem_cache_free);
652 unsigned int kmem_cache_size(struct kmem_cache *c)
656 EXPORT_SYMBOL(kmem_cache_size);
658 const char *kmem_cache_name(struct kmem_cache *c)
662 EXPORT_SYMBOL(kmem_cache_name);
664 int kmem_cache_shrink(struct kmem_cache *d)
668 EXPORT_SYMBOL(kmem_cache_shrink);
670 int kmem_ptr_validate(struct kmem_cache *a, const void *b)
675 static unsigned int slob_ready __read_mostly;
677 int slab_is_available(void)
682 void __init kmem_cache_init(void)