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_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().
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>
65 #include <linux/swap.h> /* struct reclaim_state */
66 #include <linux/cache.h>
67 #include <linux/init.h>
68 #include <linux/export.h>
69 #include <linux/rcupdate.h>
70 #include <linux/list.h>
71 #include <linux/kmemleak.h>
73 #include <trace/events/kmem.h>
75 #include <linux/atomic.h>
78 * slob_block has a field 'units', which indicates size of block if +ve,
79 * or offset of next block if -ve (in SLOB_UNITs).
81 * Free blocks of size 1 unit simply contain the offset of the next block.
82 * Those with larger size contain their size in the first SLOB_UNIT of
83 * memory, and the offset of the next free block in the second SLOB_UNIT.
85 #if PAGE_SIZE <= (32767 * 2)
86 typedef s16 slobidx_t;
88 typedef s32 slobidx_t;
94 typedef struct slob_block slob_t;
97 * All partially free slob pages go on these lists.
99 #define SLOB_BREAK1 256
100 #define SLOB_BREAK2 1024
101 static LIST_HEAD(free_slob_small);
102 static LIST_HEAD(free_slob_medium);
103 static LIST_HEAD(free_slob_large);
106 * slob_page_free: true for pages on free_slob_pages list.
108 static inline int slob_page_free(struct page *sp)
110 return PageSlobFree(sp);
113 static void set_slob_page_free(struct page *sp, struct list_head *list)
115 list_add(&sp->list, list);
116 __SetPageSlobFree(sp);
119 static inline void clear_slob_page_free(struct page *sp)
122 __ClearPageSlobFree(sp);
125 #define SLOB_UNIT sizeof(slob_t)
126 #define SLOB_UNITS(size) (((size) + SLOB_UNIT - 1)/SLOB_UNIT)
127 #define SLOB_ALIGN L1_CACHE_BYTES
130 * struct slob_rcu is inserted at the tail of allocated slob blocks, which
131 * were created with a SLAB_DESTROY_BY_RCU slab. slob_rcu is used to free
132 * the block using call_rcu.
135 struct rcu_head head;
140 * slob_lock protects all slob allocator structures.
142 static DEFINE_SPINLOCK(slob_lock);
145 * Encode the given size and next info into a free slob block s.
147 static void set_slob(slob_t *s, slobidx_t size, slob_t *next)
149 slob_t *base = (slob_t *)((unsigned long)s & PAGE_MASK);
150 slobidx_t offset = next - base;
156 s[0].units = -offset;
160 * Return the size of a slob block.
162 static slobidx_t slob_units(slob_t *s)
170 * Return the next free slob block pointer after this one.
172 static slob_t *slob_next(slob_t *s)
174 slob_t *base = (slob_t *)((unsigned long)s & PAGE_MASK);
185 * Returns true if s is the last free block in its page.
187 static int slob_last(slob_t *s)
189 return !((unsigned long)slob_next(s) & ~PAGE_MASK);
192 static void *slob_new_pages(gfp_t gfp, int order, int node)
197 if (node != NUMA_NO_NODE)
198 page = alloc_pages_exact_node(node, gfp, order);
201 page = alloc_pages(gfp, order);
206 return page_address(page);
209 static void slob_free_pages(void *b, int order)
211 if (current->reclaim_state)
212 current->reclaim_state->reclaimed_slab += 1 << order;
213 free_pages((unsigned long)b, order);
217 * Allocate a slob block within a given slob_page sp.
219 static void *slob_page_alloc(struct page *sp, size_t size, int align)
221 slob_t *prev, *cur, *aligned = NULL;
222 int delta = 0, units = SLOB_UNITS(size);
224 for (prev = NULL, cur = sp->freelist; ; prev = cur, cur = slob_next(cur)) {
225 slobidx_t avail = slob_units(cur);
228 aligned = (slob_t *)ALIGN((unsigned long)cur, align);
229 delta = aligned - cur;
231 if (avail >= units + delta) { /* room enough? */
234 if (delta) { /* need to fragment head to align? */
235 next = slob_next(cur);
236 set_slob(aligned, avail - delta, next);
237 set_slob(cur, delta, aligned);
240 avail = slob_units(cur);
243 next = slob_next(cur);
244 if (avail == units) { /* exact fit? unlink. */
246 set_slob(prev, slob_units(prev), next);
249 } else { /* fragment */
251 set_slob(prev, slob_units(prev), cur + units);
253 sp->freelist = cur + units;
254 set_slob(cur + units, avail - units, next);
259 clear_slob_page_free(sp);
268 * slob_alloc: entry point into the slob allocator.
270 static void *slob_alloc(size_t size, gfp_t gfp, int align, int node)
273 struct list_head *prev;
274 struct list_head *slob_list;
278 if (size < SLOB_BREAK1)
279 slob_list = &free_slob_small;
280 else if (size < SLOB_BREAK2)
281 slob_list = &free_slob_medium;
283 slob_list = &free_slob_large;
285 spin_lock_irqsave(&slob_lock, flags);
286 /* Iterate through each partially free page, try to find room */
287 list_for_each_entry(sp, slob_list, list) {
290 * If there's a node specification, search for a partial
291 * page with a matching node id in the freelist.
293 if (node != NUMA_NO_NODE && page_to_nid(sp) != node)
296 /* Enough room on this page? */
297 if (sp->units < SLOB_UNITS(size))
300 /* Attempt to alloc */
301 prev = sp->list.prev;
302 b = slob_page_alloc(sp, size, align);
306 /* Improve fragment distribution and reduce our average
307 * search time by starting our next search here. (see
308 * Knuth vol 1, sec 2.5, pg 449) */
309 if (prev != slob_list->prev &&
310 slob_list->next != prev->next)
311 list_move_tail(slob_list, prev->next);
314 spin_unlock_irqrestore(&slob_lock, flags);
316 /* Not enough space: must allocate a new page */
318 b = slob_new_pages(gfp & ~__GFP_ZERO, 0, node);
321 sp = virt_to_page(b);
324 spin_lock_irqsave(&slob_lock, flags);
325 sp->units = SLOB_UNITS(PAGE_SIZE);
327 INIT_LIST_HEAD(&sp->list);
328 set_slob(b, SLOB_UNITS(PAGE_SIZE), b + SLOB_UNITS(PAGE_SIZE));
329 set_slob_page_free(sp, slob_list);
330 b = slob_page_alloc(sp, size, align);
332 spin_unlock_irqrestore(&slob_lock, flags);
334 if (unlikely((gfp & __GFP_ZERO) && b))
340 * slob_free: entry point into the slob allocator.
342 static void slob_free(void *block, int size)
345 slob_t *prev, *next, *b = (slob_t *)block;
348 struct list_head *slob_list;
350 if (unlikely(ZERO_OR_NULL_PTR(block)))
354 sp = virt_to_page(block);
355 units = SLOB_UNITS(size);
357 spin_lock_irqsave(&slob_lock, flags);
359 if (sp->units + units == SLOB_UNITS(PAGE_SIZE)) {
360 /* Go directly to page allocator. Do not pass slob allocator */
361 if (slob_page_free(sp))
362 clear_slob_page_free(sp);
363 spin_unlock_irqrestore(&slob_lock, flags);
365 reset_page_mapcount(sp);
366 slob_free_pages(b, 0);
370 if (!slob_page_free(sp)) {
371 /* This slob page is about to become partially free. Easy! */
375 (void *)((unsigned long)(b +
376 SLOB_UNITS(PAGE_SIZE)) & PAGE_MASK));
377 if (size < SLOB_BREAK1)
378 slob_list = &free_slob_small;
379 else if (size < SLOB_BREAK2)
380 slob_list = &free_slob_medium;
382 slob_list = &free_slob_large;
383 set_slob_page_free(sp, slob_list);
388 * Otherwise the page is already partially free, so find reinsertion
393 if (b < (slob_t *)sp->freelist) {
394 if (b + units == sp->freelist) {
395 units += slob_units(sp->freelist);
396 sp->freelist = slob_next(sp->freelist);
398 set_slob(b, units, sp->freelist);
402 next = slob_next(prev);
405 next = slob_next(prev);
408 if (!slob_last(prev) && b + units == next) {
409 units += slob_units(next);
410 set_slob(b, units, slob_next(next));
412 set_slob(b, units, next);
414 if (prev + slob_units(prev) == b) {
415 units = slob_units(b) + slob_units(prev);
416 set_slob(prev, units, slob_next(b));
418 set_slob(prev, slob_units(prev), b);
421 spin_unlock_irqrestore(&slob_lock, flags);
425 * End of slob allocator proper. Begin kmem_cache_alloc and kmalloc frontend.
428 static __always_inline void *
429 __do_kmalloc_node(size_t size, gfp_t gfp, int node, unsigned long caller)
432 int align = max_t(size_t, ARCH_KMALLOC_MINALIGN, ARCH_SLAB_MINALIGN);
435 gfp &= gfp_allowed_mask;
437 lockdep_trace_alloc(gfp);
439 if (size < PAGE_SIZE - align) {
441 return ZERO_SIZE_PTR;
443 m = slob_alloc(size + align, gfp, align, node);
448 ret = (void *)m + align;
450 trace_kmalloc_node(caller, ret,
451 size, size + align, gfp, node);
453 unsigned int order = get_order(size);
457 ret = slob_new_pages(gfp, order, node);
460 page = virt_to_page(ret);
461 page->private = size;
464 trace_kmalloc_node(caller, ret,
465 size, PAGE_SIZE << order, gfp, node);
468 kmemleak_alloc(ret, size, 1, gfp);
472 void *__kmalloc_node(size_t size, gfp_t gfp, int node)
474 return __do_kmalloc_node(size, gfp, node, _RET_IP_);
476 EXPORT_SYMBOL(__kmalloc_node);
478 #ifdef CONFIG_TRACING
479 void *__kmalloc_track_caller(size_t size, gfp_t gfp, unsigned long caller)
481 return __do_kmalloc_node(size, gfp, NUMA_NO_NODE, caller);
485 void *__kmalloc_node_track_caller(size_t size, gfp_t gfp,
486 int node, unsigned long caller)
488 return __do_kmalloc_node(size, gfp, node, caller);
493 void kfree(const void *block)
497 trace_kfree(_RET_IP_, block);
499 if (unlikely(ZERO_OR_NULL_PTR(block)))
501 kmemleak_free(block);
503 sp = virt_to_page(block);
505 int align = max_t(size_t, ARCH_KMALLOC_MINALIGN, ARCH_SLAB_MINALIGN);
506 unsigned int *m = (unsigned int *)(block - align);
507 slob_free(m, *m + align);
511 EXPORT_SYMBOL(kfree);
513 /* can't use ksize for kmem_cache_alloc memory, only kmalloc */
514 size_t ksize(const void *block)
519 if (unlikely(block == ZERO_SIZE_PTR))
522 sp = virt_to_page(block);
524 int align = max_t(size_t, ARCH_KMALLOC_MINALIGN, ARCH_SLAB_MINALIGN);
525 unsigned int *m = (unsigned int *)(block - align);
526 return SLOB_UNITS(*m) * SLOB_UNIT;
530 EXPORT_SYMBOL(ksize);
532 int __kmem_cache_create(struct kmem_cache *c, unsigned long flags)
534 size_t align = c->size;
536 if (flags & SLAB_DESTROY_BY_RCU) {
537 /* leave room for rcu footer at the end of object */
538 c->size += sizeof(struct slob_rcu);
541 /* ignore alignment unless it's forced */
542 c->align = (flags & SLAB_HWCACHE_ALIGN) ? SLOB_ALIGN : 0;
543 if (c->align < ARCH_SLAB_MINALIGN)
544 c->align = ARCH_SLAB_MINALIGN;
545 if (c->align < align)
551 void *kmem_cache_alloc_node(struct kmem_cache *c, gfp_t flags, int node)
555 flags &= gfp_allowed_mask;
557 lockdep_trace_alloc(flags);
559 if (c->size < PAGE_SIZE) {
560 b = slob_alloc(c->size, flags, c->align, node);
561 trace_kmem_cache_alloc_node(_RET_IP_, b, c->size,
562 SLOB_UNITS(c->size) * SLOB_UNIT,
565 b = slob_new_pages(flags, get_order(c->size), node);
566 trace_kmem_cache_alloc_node(_RET_IP_, b, c->size,
567 PAGE_SIZE << get_order(c->size),
574 kmemleak_alloc_recursive(b, c->size, 1, c->flags, flags);
577 EXPORT_SYMBOL(kmem_cache_alloc_node);
579 static void __kmem_cache_free(void *b, int size)
581 if (size < PAGE_SIZE)
584 slob_free_pages(b, get_order(size));
587 static void kmem_rcu_free(struct rcu_head *head)
589 struct slob_rcu *slob_rcu = (struct slob_rcu *)head;
590 void *b = (void *)slob_rcu - (slob_rcu->size - sizeof(struct slob_rcu));
592 __kmem_cache_free(b, slob_rcu->size);
595 void kmem_cache_free(struct kmem_cache *c, void *b)
597 kmemleak_free_recursive(b, c->flags);
598 if (unlikely(c->flags & SLAB_DESTROY_BY_RCU)) {
599 struct slob_rcu *slob_rcu;
600 slob_rcu = b + (c->size - sizeof(struct slob_rcu));
601 slob_rcu->size = c->size;
602 call_rcu(&slob_rcu->head, kmem_rcu_free);
604 __kmem_cache_free(b, c->size);
607 trace_kmem_cache_free(_RET_IP_, b);
609 EXPORT_SYMBOL(kmem_cache_free);
611 unsigned int kmem_cache_size(struct kmem_cache *c)
615 EXPORT_SYMBOL(kmem_cache_size);
617 int __kmem_cache_shutdown(struct kmem_cache *c)
619 /* No way to check for remaining objects */
623 int kmem_cache_shrink(struct kmem_cache *d)
627 EXPORT_SYMBOL(kmem_cache_shrink);
629 struct kmem_cache kmem_cache_boot = {
630 .name = "kmem_cache",
631 .size = sizeof(struct kmem_cache),
633 .align = ARCH_KMALLOC_MINALIGN,
636 void __init kmem_cache_init(void)
638 kmem_cache = &kmem_cache_boot;
642 void __init kmem_cache_init_late(void)