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.
32 * These objects are detected in kfree() because PageSlab()
35 * SLAB is emulated on top of SLOB by simply calling constructors and
36 * destructors for every SLAB allocation. Objects are returned with the
37 * 4-byte alignment unless the SLAB_HWCACHE_ALIGN flag is set, in which
38 * case the low-level allocator will fragment blocks to create the proper
39 * alignment. Again, objects of page-size or greater are allocated by
40 * calling alloc_pages(). As SLAB objects know their size, no separate
41 * size bookkeeping is necessary and there is essentially no allocation
42 * space overhead, and compound pages aren't needed for multi-page
45 * NUMA support in SLOB is fairly simplistic, pushing most of the real
46 * logic down to the page allocator, and simply doing the node accounting
47 * on the upper levels. In the event that a node id is explicitly
48 * provided, alloc_pages_exact_node() with the specified node id is used
49 * instead. The common case (or when the node id isn't explicitly provided)
50 * will default to the current node, as per numa_node_id().
52 * Node aware pages are still inserted in to the global freelist, and
53 * these are scanned for by matching against the node id encoded in the
54 * page flags. As a result, block allocations that can be satisfied from
55 * the freelist will only be done so on pages residing on the same node,
56 * in order to prevent random node placement.
59 #include <linux/kernel.h>
60 #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/export.h>
67 #include <linux/rcupdate.h>
68 #include <linux/list.h>
69 #include <linux/kmemleak.h>
71 #include <trace/events/kmem.h>
73 #include <linux/atomic.h>
77 * slob_block has a field 'units', which indicates size of block if +ve,
78 * or offset of next block if -ve (in SLOB_UNITs).
80 * Free blocks of size 1 unit simply contain the offset of the next block.
81 * Those with larger size contain their size in the first SLOB_UNIT of
82 * memory, and the offset of the next free block in the second SLOB_UNIT.
84 #if PAGE_SIZE <= (32767 * 2)
85 typedef s16 slobidx_t;
87 typedef s32 slobidx_t;
93 typedef struct slob_block slob_t;
96 * All partially free slob pages go on these lists.
98 #define SLOB_BREAK1 256
99 #define SLOB_BREAK2 1024
100 static LIST_HEAD(free_slob_small);
101 static LIST_HEAD(free_slob_medium);
102 static LIST_HEAD(free_slob_large);
105 * slob_page_free: true for pages on free_slob_pages list.
107 static inline int slob_page_free(struct page *sp)
109 return PageSlobFree(sp);
112 static void set_slob_page_free(struct page *sp, struct list_head *list)
114 list_add(&sp->lru, list);
115 __SetPageSlobFree(sp);
118 static inline void clear_slob_page_free(struct page *sp)
121 __ClearPageSlobFree(sp);
124 #define SLOB_UNIT sizeof(slob_t)
125 #define SLOB_UNITS(size) DIV_ROUND_UP(size, SLOB_UNIT)
128 * struct slob_rcu is inserted at the tail of allocated slob blocks, which
129 * were created with a SLAB_DESTROY_BY_RCU slab. slob_rcu is used to free
130 * the block using call_rcu.
133 struct rcu_head head;
138 * slob_lock protects all slob allocator structures.
140 static DEFINE_SPINLOCK(slob_lock);
143 * Encode the given size and next info into a free slob block s.
145 static void set_slob(slob_t *s, slobidx_t size, slob_t *next)
147 slob_t *base = (slob_t *)((unsigned long)s & PAGE_MASK);
148 slobidx_t offset = next - base;
154 s[0].units = -offset;
158 * Return the size of a slob block.
160 static slobidx_t slob_units(slob_t *s)
168 * Return the next free slob block pointer after this one.
170 static slob_t *slob_next(slob_t *s)
172 slob_t *base = (slob_t *)((unsigned long)s & PAGE_MASK);
183 * Returns true if s is the last free block in its page.
185 static int slob_last(slob_t *s)
187 return !((unsigned long)slob_next(s) & ~PAGE_MASK);
190 static void *slob_new_pages(gfp_t gfp, int order, int node)
195 if (node != NUMA_NO_NODE)
196 page = alloc_pages_exact_node(node, gfp, order);
199 page = alloc_pages(gfp, order);
204 return page_address(page);
207 static void slob_free_pages(void *b, int order)
209 if (current->reclaim_state)
210 current->reclaim_state->reclaimed_slab += 1 << order;
211 free_pages((unsigned long)b, order);
215 * Allocate a slob block within a given slob_page sp.
217 static void *slob_page_alloc(struct page *sp, size_t size, int align)
219 slob_t *prev, *cur, *aligned = NULL;
220 int delta = 0, units = SLOB_UNITS(size);
222 for (prev = NULL, cur = sp->freelist; ; prev = cur, cur = slob_next(cur)) {
223 slobidx_t avail = slob_units(cur);
226 aligned = (slob_t *)ALIGN((unsigned long)cur, align);
227 delta = aligned - cur;
229 if (avail >= units + delta) { /* room enough? */
232 if (delta) { /* need to fragment head to align? */
233 next = slob_next(cur);
234 set_slob(aligned, avail - delta, next);
235 set_slob(cur, delta, aligned);
238 avail = slob_units(cur);
241 next = slob_next(cur);
242 if (avail == units) { /* exact fit? unlink. */
244 set_slob(prev, slob_units(prev), next);
247 } else { /* fragment */
249 set_slob(prev, slob_units(prev), cur + units);
251 sp->freelist = cur + units;
252 set_slob(cur + units, avail - units, next);
257 clear_slob_page_free(sp);
266 * slob_alloc: entry point into the slob allocator.
268 static void *slob_alloc(size_t size, gfp_t gfp, int align, int node)
271 struct list_head *prev;
272 struct list_head *slob_list;
276 if (size < SLOB_BREAK1)
277 slob_list = &free_slob_small;
278 else if (size < SLOB_BREAK2)
279 slob_list = &free_slob_medium;
281 slob_list = &free_slob_large;
283 spin_lock_irqsave(&slob_lock, flags);
284 /* Iterate through each partially free page, try to find room */
285 list_for_each_entry(sp, slob_list, lru) {
288 * If there's a node specification, search for a partial
289 * page with a matching node id in the freelist.
291 if (node != NUMA_NO_NODE && page_to_nid(sp) != node)
294 /* Enough room on this page? */
295 if (sp->units < SLOB_UNITS(size))
298 /* Attempt to alloc */
300 b = slob_page_alloc(sp, size, align);
304 /* Improve fragment distribution and reduce our average
305 * search time by starting our next search here. (see
306 * Knuth vol 1, sec 2.5, pg 449) */
307 if (prev != slob_list->prev &&
308 slob_list->next != prev->next)
309 list_move_tail(slob_list, prev->next);
312 spin_unlock_irqrestore(&slob_lock, flags);
314 /* Not enough space: must allocate a new page */
316 b = slob_new_pages(gfp & ~__GFP_ZERO, 0, node);
319 sp = virt_to_page(b);
322 spin_lock_irqsave(&slob_lock, flags);
323 sp->units = SLOB_UNITS(PAGE_SIZE);
325 INIT_LIST_HEAD(&sp->lru);
326 set_slob(b, SLOB_UNITS(PAGE_SIZE), b + SLOB_UNITS(PAGE_SIZE));
327 set_slob_page_free(sp, slob_list);
328 b = slob_page_alloc(sp, size, align);
330 spin_unlock_irqrestore(&slob_lock, flags);
332 if (unlikely((gfp & __GFP_ZERO) && b))
338 * slob_free: entry point into the slob allocator.
340 static void slob_free(void *block, int size)
343 slob_t *prev, *next, *b = (slob_t *)block;
346 struct list_head *slob_list;
348 if (unlikely(ZERO_OR_NULL_PTR(block)))
352 sp = virt_to_page(block);
353 units = SLOB_UNITS(size);
355 spin_lock_irqsave(&slob_lock, flags);
357 if (sp->units + units == SLOB_UNITS(PAGE_SIZE)) {
358 /* Go directly to page allocator. Do not pass slob allocator */
359 if (slob_page_free(sp))
360 clear_slob_page_free(sp);
361 spin_unlock_irqrestore(&slob_lock, flags);
363 page_mapcount_reset(sp);
364 slob_free_pages(b, 0);
368 if (!slob_page_free(sp)) {
369 /* This slob page is about to become partially free. Easy! */
373 (void *)((unsigned long)(b +
374 SLOB_UNITS(PAGE_SIZE)) & PAGE_MASK));
375 if (size < SLOB_BREAK1)
376 slob_list = &free_slob_small;
377 else if (size < SLOB_BREAK2)
378 slob_list = &free_slob_medium;
380 slob_list = &free_slob_large;
381 set_slob_page_free(sp, slob_list);
386 * Otherwise the page is already partially free, so find reinsertion
391 if (b < (slob_t *)sp->freelist) {
392 if (b + units == sp->freelist) {
393 units += slob_units(sp->freelist);
394 sp->freelist = slob_next(sp->freelist);
396 set_slob(b, units, sp->freelist);
400 next = slob_next(prev);
403 next = slob_next(prev);
406 if (!slob_last(prev) && b + units == next) {
407 units += slob_units(next);
408 set_slob(b, units, slob_next(next));
410 set_slob(b, units, next);
412 if (prev + slob_units(prev) == b) {
413 units = slob_units(b) + slob_units(prev);
414 set_slob(prev, units, slob_next(b));
416 set_slob(prev, slob_units(prev), b);
419 spin_unlock_irqrestore(&slob_lock, flags);
423 * End of slob allocator proper. Begin kmem_cache_alloc and kmalloc frontend.
426 static __always_inline void *
427 __do_kmalloc_node(size_t size, gfp_t gfp, int node, unsigned long caller)
430 int align = max_t(size_t, ARCH_KMALLOC_MINALIGN, ARCH_SLAB_MINALIGN);
433 gfp &= gfp_allowed_mask;
435 lockdep_trace_alloc(gfp);
437 if (size < PAGE_SIZE - align) {
439 return ZERO_SIZE_PTR;
441 m = slob_alloc(size + align, gfp, align, node);
446 ret = (void *)m + align;
448 trace_kmalloc_node(caller, ret,
449 size, size + align, gfp, node);
451 unsigned int order = get_order(size);
455 ret = slob_new_pages(gfp, order, node);
457 trace_kmalloc_node(caller, ret,
458 size, PAGE_SIZE << order, gfp, node);
461 kmemleak_alloc(ret, size, 1, gfp);
465 void *__kmalloc(size_t size, gfp_t gfp)
467 return __do_kmalloc_node(size, gfp, NUMA_NO_NODE, _RET_IP_);
469 EXPORT_SYMBOL(__kmalloc);
471 void *__kmalloc_track_caller(size_t size, gfp_t gfp, unsigned long caller)
473 return __do_kmalloc_node(size, gfp, NUMA_NO_NODE, caller);
477 void *__kmalloc_node_track_caller(size_t size, gfp_t gfp,
478 int node, unsigned long caller)
480 return __do_kmalloc_node(size, gfp, node, caller);
484 void kfree(const void *block)
488 trace_kfree(_RET_IP_, block);
490 if (unlikely(ZERO_OR_NULL_PTR(block)))
492 kmemleak_free(block);
494 sp = virt_to_page(block);
496 int align = max_t(size_t, ARCH_KMALLOC_MINALIGN, ARCH_SLAB_MINALIGN);
497 unsigned int *m = (unsigned int *)(block - align);
498 slob_free(m, *m + align);
500 __free_pages(sp, compound_order(sp));
502 EXPORT_SYMBOL(kfree);
504 /* can't use ksize for kmem_cache_alloc memory, only kmalloc */
505 size_t ksize(const void *block)
512 if (unlikely(block == ZERO_SIZE_PTR))
515 sp = virt_to_page(block);
516 if (unlikely(!PageSlab(sp)))
517 return PAGE_SIZE << compound_order(sp);
519 align = max_t(size_t, ARCH_KMALLOC_MINALIGN, ARCH_SLAB_MINALIGN);
520 m = (unsigned int *)(block - align);
521 return SLOB_UNITS(*m) * SLOB_UNIT;
523 EXPORT_SYMBOL(ksize);
525 int __kmem_cache_create(struct kmem_cache *c, unsigned long flags)
527 if (flags & SLAB_DESTROY_BY_RCU) {
528 /* leave room for rcu footer at the end of object */
529 c->size += sizeof(struct slob_rcu);
535 void *slob_alloc_node(struct kmem_cache *c, gfp_t flags, int node)
539 flags &= gfp_allowed_mask;
541 lockdep_trace_alloc(flags);
543 if (c->size < PAGE_SIZE) {
544 b = slob_alloc(c->size, flags, c->align, node);
545 trace_kmem_cache_alloc_node(_RET_IP_, b, c->object_size,
546 SLOB_UNITS(c->size) * SLOB_UNIT,
549 b = slob_new_pages(flags, get_order(c->size), node);
550 trace_kmem_cache_alloc_node(_RET_IP_, b, c->object_size,
551 PAGE_SIZE << get_order(c->size),
558 kmemleak_alloc_recursive(b, c->size, 1, c->flags, flags);
561 EXPORT_SYMBOL(slob_alloc_node);
563 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
565 return slob_alloc_node(cachep, flags, NUMA_NO_NODE);
567 EXPORT_SYMBOL(kmem_cache_alloc);
570 void *__kmalloc_node(size_t size, gfp_t gfp, int node)
572 return __do_kmalloc_node(size, gfp, node, _RET_IP_);
574 EXPORT_SYMBOL(__kmalloc_node);
576 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t gfp, int node)
578 return slob_alloc_node(cachep, gfp, node);
580 EXPORT_SYMBOL(kmem_cache_alloc_node);
583 static void __kmem_cache_free(void *b, int size)
585 if (size < PAGE_SIZE)
588 slob_free_pages(b, get_order(size));
591 static void kmem_rcu_free(struct rcu_head *head)
593 struct slob_rcu *slob_rcu = (struct slob_rcu *)head;
594 void *b = (void *)slob_rcu - (slob_rcu->size - sizeof(struct slob_rcu));
596 __kmem_cache_free(b, slob_rcu->size);
599 void kmem_cache_free(struct kmem_cache *c, void *b)
601 kmemleak_free_recursive(b, c->flags);
602 if (unlikely(c->flags & SLAB_DESTROY_BY_RCU)) {
603 struct slob_rcu *slob_rcu;
604 slob_rcu = b + (c->size - sizeof(struct slob_rcu));
605 slob_rcu->size = c->size;
606 call_rcu(&slob_rcu->head, kmem_rcu_free);
608 __kmem_cache_free(b, c->size);
611 trace_kmem_cache_free(_RET_IP_, b);
613 EXPORT_SYMBOL(kmem_cache_free);
615 int __kmem_cache_shutdown(struct kmem_cache *c)
617 /* No way to check for remaining objects */
621 int __kmem_cache_shrink(struct kmem_cache *d)
626 struct kmem_cache kmem_cache_boot = {
627 .name = "kmem_cache",
628 .size = sizeof(struct kmem_cache),
630 .align = ARCH_KMALLOC_MINALIGN,
633 void __init kmem_cache_init(void)
635 kmem_cache = &kmem_cache_boot;
639 void __init kmem_cache_init_late(void)