Loading...
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_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().
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/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 <linux/kmemleak.h>
70
71#include <trace/events/kmem.h>
72
73#include <linux/atomic.h>
74
75/*
76 * slob_block has a field 'units', which indicates size of block if +ve,
77 * or offset of next block if -ve (in SLOB_UNITs).
78 *
79 * Free blocks of size 1 unit simply contain the offset of the next block.
80 * Those with larger size contain their size in the first SLOB_UNIT of
81 * memory, and the offset of the next free block in the second SLOB_UNIT.
82 */
83#if PAGE_SIZE <= (32767 * 2)
84typedef s16 slobidx_t;
85#else
86typedef s32 slobidx_t;
87#endif
88
89struct slob_block {
90 slobidx_t units;
91};
92typedef struct slob_block slob_t;
93
94/*
95 * We use struct page fields to manage some slob allocation aspects,
96 * however to avoid the horrible mess in include/linux/mm_types.h, we'll
97 * just define our own struct page type variant here.
98 */
99struct slob_page {
100 union {
101 struct {
102 unsigned long flags; /* mandatory */
103 atomic_t _count; /* mandatory */
104 slobidx_t units; /* free units left in page */
105 unsigned long pad[2];
106 slob_t *free; /* first free slob_t in page */
107 struct list_head list; /* linked list of free pages */
108 };
109 struct page page;
110 };
111};
112static inline void struct_slob_page_wrong_size(void)
113{ BUILD_BUG_ON(sizeof(struct slob_page) != sizeof(struct page)); }
114
115/*
116 * free_slob_page: call before a slob_page is returned to the page allocator.
117 */
118static inline void free_slob_page(struct slob_page *sp)
119{
120 reset_page_mapcount(&sp->page);
121 sp->page.mapping = NULL;
122}
123
124/*
125 * All partially free slob pages go on these lists.
126 */
127#define SLOB_BREAK1 256
128#define SLOB_BREAK2 1024
129static LIST_HEAD(free_slob_small);
130static LIST_HEAD(free_slob_medium);
131static LIST_HEAD(free_slob_large);
132
133/*
134 * is_slob_page: True for all slob pages (false for bigblock pages)
135 */
136static inline int is_slob_page(struct slob_page *sp)
137{
138 return PageSlab((struct page *)sp);
139}
140
141static inline void set_slob_page(struct slob_page *sp)
142{
143 __SetPageSlab((struct page *)sp);
144}
145
146static inline void clear_slob_page(struct slob_page *sp)
147{
148 __ClearPageSlab((struct page *)sp);
149}
150
151static inline struct slob_page *slob_page(const void *addr)
152{
153 return (struct slob_page *)virt_to_page(addr);
154}
155
156/*
157 * slob_page_free: true for pages on free_slob_pages list.
158 */
159static inline int slob_page_free(struct slob_page *sp)
160{
161 return PageSlobFree((struct page *)sp);
162}
163
164static void set_slob_page_free(struct slob_page *sp, struct list_head *list)
165{
166 list_add(&sp->list, list);
167 __SetPageSlobFree((struct page *)sp);
168}
169
170static inline void clear_slob_page_free(struct slob_page *sp)
171{
172 list_del(&sp->list);
173 __ClearPageSlobFree((struct page *)sp);
174}
175
176#define SLOB_UNIT sizeof(slob_t)
177#define SLOB_UNITS(size) (((size) + SLOB_UNIT - 1)/SLOB_UNIT)
178#define SLOB_ALIGN L1_CACHE_BYTES
179
180/*
181 * struct slob_rcu is inserted at the tail of allocated slob blocks, which
182 * were created with a SLAB_DESTROY_BY_RCU slab. slob_rcu is used to free
183 * the block using call_rcu.
184 */
185struct slob_rcu {
186 struct rcu_head head;
187 int size;
188};
189
190/*
191 * slob_lock protects all slob allocator structures.
192 */
193static DEFINE_SPINLOCK(slob_lock);
194
195/*
196 * Encode the given size and next info into a free slob block s.
197 */
198static void set_slob(slob_t *s, slobidx_t size, slob_t *next)
199{
200 slob_t *base = (slob_t *)((unsigned long)s & PAGE_MASK);
201 slobidx_t offset = next - base;
202
203 if (size > 1) {
204 s[0].units = size;
205 s[1].units = offset;
206 } else
207 s[0].units = -offset;
208}
209
210/*
211 * Return the size of a slob block.
212 */
213static slobidx_t slob_units(slob_t *s)
214{
215 if (s->units > 0)
216 return s->units;
217 return 1;
218}
219
220/*
221 * Return the next free slob block pointer after this one.
222 */
223static slob_t *slob_next(slob_t *s)
224{
225 slob_t *base = (slob_t *)((unsigned long)s & PAGE_MASK);
226 slobidx_t next;
227
228 if (s[0].units < 0)
229 next = -s[0].units;
230 else
231 next = s[1].units;
232 return base+next;
233}
234
235/*
236 * Returns true if s is the last free block in its page.
237 */
238static int slob_last(slob_t *s)
239{
240 return !((unsigned long)slob_next(s) & ~PAGE_MASK);
241}
242
243static void *slob_new_pages(gfp_t gfp, int order, int node)
244{
245 void *page;
246
247#ifdef CONFIG_NUMA
248 if (node != -1)
249 page = alloc_pages_exact_node(node, gfp, order);
250 else
251#endif
252 page = alloc_pages(gfp, order);
253
254 if (!page)
255 return NULL;
256
257 return page_address(page);
258}
259
260static void slob_free_pages(void *b, int order)
261{
262 if (current->reclaim_state)
263 current->reclaim_state->reclaimed_slab += 1 << order;
264 free_pages((unsigned long)b, order);
265}
266
267/*
268 * Allocate a slob block within a given slob_page sp.
269 */
270static void *slob_page_alloc(struct slob_page *sp, size_t size, int align)
271{
272 slob_t *prev, *cur, *aligned = NULL;
273 int delta = 0, units = SLOB_UNITS(size);
274
275 for (prev = NULL, cur = sp->free; ; prev = cur, cur = slob_next(cur)) {
276 slobidx_t avail = slob_units(cur);
277
278 if (align) {
279 aligned = (slob_t *)ALIGN((unsigned long)cur, align);
280 delta = aligned - cur;
281 }
282 if (avail >= units + delta) { /* room enough? */
283 slob_t *next;
284
285 if (delta) { /* need to fragment head to align? */
286 next = slob_next(cur);
287 set_slob(aligned, avail - delta, next);
288 set_slob(cur, delta, aligned);
289 prev = cur;
290 cur = aligned;
291 avail = slob_units(cur);
292 }
293
294 next = slob_next(cur);
295 if (avail == units) { /* exact fit? unlink. */
296 if (prev)
297 set_slob(prev, slob_units(prev), next);
298 else
299 sp->free = next;
300 } else { /* fragment */
301 if (prev)
302 set_slob(prev, slob_units(prev), cur + units);
303 else
304 sp->free = cur + units;
305 set_slob(cur + units, avail - units, next);
306 }
307
308 sp->units -= units;
309 if (!sp->units)
310 clear_slob_page_free(sp);
311 return cur;
312 }
313 if (slob_last(cur))
314 return NULL;
315 }
316}
317
318/*
319 * slob_alloc: entry point into the slob allocator.
320 */
321static void *slob_alloc(size_t size, gfp_t gfp, int align, int node)
322{
323 struct slob_page *sp;
324 struct list_head *prev;
325 struct list_head *slob_list;
326 slob_t *b = NULL;
327 unsigned long flags;
328
329 if (size < SLOB_BREAK1)
330 slob_list = &free_slob_small;
331 else if (size < SLOB_BREAK2)
332 slob_list = &free_slob_medium;
333 else
334 slob_list = &free_slob_large;
335
336 spin_lock_irqsave(&slob_lock, flags);
337 /* Iterate through each partially free page, try to find room */
338 list_for_each_entry(sp, slob_list, list) {
339#ifdef CONFIG_NUMA
340 /*
341 * If there's a node specification, search for a partial
342 * page with a matching node id in the freelist.
343 */
344 if (node != -1 && page_to_nid(&sp->page) != node)
345 continue;
346#endif
347 /* Enough room on this page? */
348 if (sp->units < SLOB_UNITS(size))
349 continue;
350
351 /* Attempt to alloc */
352 prev = sp->list.prev;
353 b = slob_page_alloc(sp, size, align);
354 if (!b)
355 continue;
356
357 /* Improve fragment distribution and reduce our average
358 * search time by starting our next search here. (see
359 * Knuth vol 1, sec 2.5, pg 449) */
360 if (prev != slob_list->prev &&
361 slob_list->next != prev->next)
362 list_move_tail(slob_list, prev->next);
363 break;
364 }
365 spin_unlock_irqrestore(&slob_lock, flags);
366
367 /* Not enough space: must allocate a new page */
368 if (!b) {
369 b = slob_new_pages(gfp & ~__GFP_ZERO, 0, node);
370 if (!b)
371 return NULL;
372 sp = slob_page(b);
373 set_slob_page(sp);
374
375 spin_lock_irqsave(&slob_lock, flags);
376 sp->units = SLOB_UNITS(PAGE_SIZE);
377 sp->free = b;
378 INIT_LIST_HEAD(&sp->list);
379 set_slob(b, SLOB_UNITS(PAGE_SIZE), b + SLOB_UNITS(PAGE_SIZE));
380 set_slob_page_free(sp, slob_list);
381 b = slob_page_alloc(sp, size, align);
382 BUG_ON(!b);
383 spin_unlock_irqrestore(&slob_lock, flags);
384 }
385 if (unlikely((gfp & __GFP_ZERO) && b))
386 memset(b, 0, size);
387 return b;
388}
389
390/*
391 * slob_free: entry point into the slob allocator.
392 */
393static void slob_free(void *block, int size)
394{
395 struct slob_page *sp;
396 slob_t *prev, *next, *b = (slob_t *)block;
397 slobidx_t units;
398 unsigned long flags;
399 struct list_head *slob_list;
400
401 if (unlikely(ZERO_OR_NULL_PTR(block)))
402 return;
403 BUG_ON(!size);
404
405 sp = slob_page(block);
406 units = SLOB_UNITS(size);
407
408 spin_lock_irqsave(&slob_lock, flags);
409
410 if (sp->units + units == SLOB_UNITS(PAGE_SIZE)) {
411 /* Go directly to page allocator. Do not pass slob allocator */
412 if (slob_page_free(sp))
413 clear_slob_page_free(sp);
414 spin_unlock_irqrestore(&slob_lock, flags);
415 clear_slob_page(sp);
416 free_slob_page(sp);
417 slob_free_pages(b, 0);
418 return;
419 }
420
421 if (!slob_page_free(sp)) {
422 /* This slob page is about to become partially free. Easy! */
423 sp->units = units;
424 sp->free = b;
425 set_slob(b, units,
426 (void *)((unsigned long)(b +
427 SLOB_UNITS(PAGE_SIZE)) & PAGE_MASK));
428 if (size < SLOB_BREAK1)
429 slob_list = &free_slob_small;
430 else if (size < SLOB_BREAK2)
431 slob_list = &free_slob_medium;
432 else
433 slob_list = &free_slob_large;
434 set_slob_page_free(sp, slob_list);
435 goto out;
436 }
437
438 /*
439 * Otherwise the page is already partially free, so find reinsertion
440 * point.
441 */
442 sp->units += units;
443
444 if (b < sp->free) {
445 if (b + units == sp->free) {
446 units += slob_units(sp->free);
447 sp->free = slob_next(sp->free);
448 }
449 set_slob(b, units, sp->free);
450 sp->free = b;
451 } else {
452 prev = sp->free;
453 next = slob_next(prev);
454 while (b > next) {
455 prev = next;
456 next = slob_next(prev);
457 }
458
459 if (!slob_last(prev) && b + units == next) {
460 units += slob_units(next);
461 set_slob(b, units, slob_next(next));
462 } else
463 set_slob(b, units, next);
464
465 if (prev + slob_units(prev) == b) {
466 units = slob_units(b) + slob_units(prev);
467 set_slob(prev, units, slob_next(b));
468 } else
469 set_slob(prev, slob_units(prev), b);
470 }
471out:
472 spin_unlock_irqrestore(&slob_lock, flags);
473}
474
475/*
476 * End of slob allocator proper. Begin kmem_cache_alloc and kmalloc frontend.
477 */
478
479void *__kmalloc_node(size_t size, gfp_t gfp, int node)
480{
481 unsigned int *m;
482 int align = max(ARCH_KMALLOC_MINALIGN, ARCH_SLAB_MINALIGN);
483 void *ret;
484
485 gfp &= gfp_allowed_mask;
486
487 lockdep_trace_alloc(gfp);
488
489 if (size < PAGE_SIZE - align) {
490 if (!size)
491 return ZERO_SIZE_PTR;
492
493 m = slob_alloc(size + align, gfp, align, node);
494
495 if (!m)
496 return NULL;
497 *m = size;
498 ret = (void *)m + align;
499
500 trace_kmalloc_node(_RET_IP_, ret,
501 size, size + align, gfp, node);
502 } else {
503 unsigned int order = get_order(size);
504
505 if (likely(order))
506 gfp |= __GFP_COMP;
507 ret = slob_new_pages(gfp, order, node);
508 if (ret) {
509 struct page *page;
510 page = virt_to_page(ret);
511 page->private = size;
512 }
513
514 trace_kmalloc_node(_RET_IP_, ret,
515 size, PAGE_SIZE << order, gfp, node);
516 }
517
518 kmemleak_alloc(ret, size, 1, gfp);
519 return ret;
520}
521EXPORT_SYMBOL(__kmalloc_node);
522
523void kfree(const void *block)
524{
525 struct slob_page *sp;
526
527 trace_kfree(_RET_IP_, block);
528
529 if (unlikely(ZERO_OR_NULL_PTR(block)))
530 return;
531 kmemleak_free(block);
532
533 sp = slob_page(block);
534 if (is_slob_page(sp)) {
535 int align = max(ARCH_KMALLOC_MINALIGN, ARCH_SLAB_MINALIGN);
536 unsigned int *m = (unsigned int *)(block - align);
537 slob_free(m, *m + align);
538 } else
539 put_page(&sp->page);
540}
541EXPORT_SYMBOL(kfree);
542
543/* can't use ksize for kmem_cache_alloc memory, only kmalloc */
544size_t ksize(const void *block)
545{
546 struct slob_page *sp;
547
548 BUG_ON(!block);
549 if (unlikely(block == ZERO_SIZE_PTR))
550 return 0;
551
552 sp = slob_page(block);
553 if (is_slob_page(sp)) {
554 int align = max(ARCH_KMALLOC_MINALIGN, ARCH_SLAB_MINALIGN);
555 unsigned int *m = (unsigned int *)(block - align);
556 return SLOB_UNITS(*m) * SLOB_UNIT;
557 } else
558 return sp->page.private;
559}
560EXPORT_SYMBOL(ksize);
561
562struct kmem_cache {
563 unsigned int size, align;
564 unsigned long flags;
565 const char *name;
566 void (*ctor)(void *);
567};
568
569struct kmem_cache *kmem_cache_create(const char *name, size_t size,
570 size_t align, unsigned long flags, void (*ctor)(void *))
571{
572 struct kmem_cache *c;
573
574 c = slob_alloc(sizeof(struct kmem_cache),
575 GFP_KERNEL, ARCH_KMALLOC_MINALIGN, -1);
576
577 if (c) {
578 c->name = name;
579 c->size = size;
580 if (flags & SLAB_DESTROY_BY_RCU) {
581 /* leave room for rcu footer at the end of object */
582 c->size += sizeof(struct slob_rcu);
583 }
584 c->flags = flags;
585 c->ctor = ctor;
586 /* ignore alignment unless it's forced */
587 c->align = (flags & SLAB_HWCACHE_ALIGN) ? SLOB_ALIGN : 0;
588 if (c->align < ARCH_SLAB_MINALIGN)
589 c->align = ARCH_SLAB_MINALIGN;
590 if (c->align < align)
591 c->align = align;
592 } else if (flags & SLAB_PANIC)
593 panic("Cannot create slab cache %s\n", name);
594
595 kmemleak_alloc(c, sizeof(struct kmem_cache), 1, GFP_KERNEL);
596 return c;
597}
598EXPORT_SYMBOL(kmem_cache_create);
599
600void kmem_cache_destroy(struct kmem_cache *c)
601{
602 kmemleak_free(c);
603 if (c->flags & SLAB_DESTROY_BY_RCU)
604 rcu_barrier();
605 slob_free(c, sizeof(struct kmem_cache));
606}
607EXPORT_SYMBOL(kmem_cache_destroy);
608
609void *kmem_cache_alloc_node(struct kmem_cache *c, gfp_t flags, int node)
610{
611 void *b;
612
613 flags &= gfp_allowed_mask;
614
615 lockdep_trace_alloc(flags);
616
617 if (c->size < PAGE_SIZE) {
618 b = slob_alloc(c->size, flags, c->align, node);
619 trace_kmem_cache_alloc_node(_RET_IP_, b, c->size,
620 SLOB_UNITS(c->size) * SLOB_UNIT,
621 flags, node);
622 } else {
623 b = slob_new_pages(flags, get_order(c->size), node);
624 trace_kmem_cache_alloc_node(_RET_IP_, b, c->size,
625 PAGE_SIZE << get_order(c->size),
626 flags, node);
627 }
628
629 if (c->ctor)
630 c->ctor(b);
631
632 kmemleak_alloc_recursive(b, c->size, 1, c->flags, flags);
633 return b;
634}
635EXPORT_SYMBOL(kmem_cache_alloc_node);
636
637static void __kmem_cache_free(void *b, int size)
638{
639 if (size < PAGE_SIZE)
640 slob_free(b, size);
641 else
642 slob_free_pages(b, get_order(size));
643}
644
645static void kmem_rcu_free(struct rcu_head *head)
646{
647 struct slob_rcu *slob_rcu = (struct slob_rcu *)head;
648 void *b = (void *)slob_rcu - (slob_rcu->size - sizeof(struct slob_rcu));
649
650 __kmem_cache_free(b, slob_rcu->size);
651}
652
653void kmem_cache_free(struct kmem_cache *c, void *b)
654{
655 kmemleak_free_recursive(b, c->flags);
656 if (unlikely(c->flags & SLAB_DESTROY_BY_RCU)) {
657 struct slob_rcu *slob_rcu;
658 slob_rcu = b + (c->size - sizeof(struct slob_rcu));
659 slob_rcu->size = c->size;
660 call_rcu(&slob_rcu->head, kmem_rcu_free);
661 } else {
662 __kmem_cache_free(b, c->size);
663 }
664
665 trace_kmem_cache_free(_RET_IP_, b);
666}
667EXPORT_SYMBOL(kmem_cache_free);
668
669unsigned int kmem_cache_size(struct kmem_cache *c)
670{
671 return c->size;
672}
673EXPORT_SYMBOL(kmem_cache_size);
674
675int kmem_cache_shrink(struct kmem_cache *d)
676{
677 return 0;
678}
679EXPORT_SYMBOL(kmem_cache_shrink);
680
681static unsigned int slob_ready __read_mostly;
682
683int slab_is_available(void)
684{
685 return slob_ready;
686}
687
688void __init kmem_cache_init(void)
689{
690 slob_ready = 1;
691}
692
693void __init kmem_cache_init_late(void)
694{
695 /* Nothing to do */
696}
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.
32 * These objects are detected in kfree() because PageSlab()
33 * is false for them.
34 *
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
43 * allocations.
44 *
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_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().
51 *
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.
57 */
58
59#include <linux/kernel.h>
60#include <linux/slab.h>
61
62#include <linux/mm.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>
70
71#include <trace/events/kmem.h>
72
73#include <linux/atomic.h>
74
75#include "slab.h"
76/*
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).
79 *
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.
83 */
84#if PAGE_SIZE <= (32767 * 2)
85typedef s16 slobidx_t;
86#else
87typedef s32 slobidx_t;
88#endif
89
90struct slob_block {
91 slobidx_t units;
92};
93typedef struct slob_block slob_t;
94
95/*
96 * All partially free slob pages go on these lists.
97 */
98#define SLOB_BREAK1 256
99#define SLOB_BREAK2 1024
100static LIST_HEAD(free_slob_small);
101static LIST_HEAD(free_slob_medium);
102static LIST_HEAD(free_slob_large);
103
104/*
105 * slob_page_free: true for pages on free_slob_pages list.
106 */
107static inline int slob_page_free(struct page *sp)
108{
109 return PageSlobFree(sp);
110}
111
112static void set_slob_page_free(struct page *sp, struct list_head *list)
113{
114 list_add(&sp->lru, list);
115 __SetPageSlobFree(sp);
116}
117
118static inline void clear_slob_page_free(struct page *sp)
119{
120 list_del(&sp->lru);
121 __ClearPageSlobFree(sp);
122}
123
124#define SLOB_UNIT sizeof(slob_t)
125#define SLOB_UNITS(size) DIV_ROUND_UP(size, SLOB_UNIT)
126
127/*
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.
131 */
132struct slob_rcu {
133 struct rcu_head head;
134 int size;
135};
136
137/*
138 * slob_lock protects all slob allocator structures.
139 */
140static DEFINE_SPINLOCK(slob_lock);
141
142/*
143 * Encode the given size and next info into a free slob block s.
144 */
145static void set_slob(slob_t *s, slobidx_t size, slob_t *next)
146{
147 slob_t *base = (slob_t *)((unsigned long)s & PAGE_MASK);
148 slobidx_t offset = next - base;
149
150 if (size > 1) {
151 s[0].units = size;
152 s[1].units = offset;
153 } else
154 s[0].units = -offset;
155}
156
157/*
158 * Return the size of a slob block.
159 */
160static slobidx_t slob_units(slob_t *s)
161{
162 if (s->units > 0)
163 return s->units;
164 return 1;
165}
166
167/*
168 * Return the next free slob block pointer after this one.
169 */
170static slob_t *slob_next(slob_t *s)
171{
172 slob_t *base = (slob_t *)((unsigned long)s & PAGE_MASK);
173 slobidx_t next;
174
175 if (s[0].units < 0)
176 next = -s[0].units;
177 else
178 next = s[1].units;
179 return base+next;
180}
181
182/*
183 * Returns true if s is the last free block in its page.
184 */
185static int slob_last(slob_t *s)
186{
187 return !((unsigned long)slob_next(s) & ~PAGE_MASK);
188}
189
190static void *slob_new_pages(gfp_t gfp, int order, int node)
191{
192 void *page;
193
194#ifdef CONFIG_NUMA
195 if (node != NUMA_NO_NODE)
196 page = __alloc_pages_node(node, gfp, order);
197 else
198#endif
199 page = alloc_pages(gfp, order);
200
201 if (!page)
202 return NULL;
203
204 return page_address(page);
205}
206
207static void slob_free_pages(void *b, int order)
208{
209 if (current->reclaim_state)
210 current->reclaim_state->reclaimed_slab += 1 << order;
211 free_pages((unsigned long)b, order);
212}
213
214/*
215 * Allocate a slob block within a given slob_page sp.
216 */
217static void *slob_page_alloc(struct page *sp, size_t size, int align)
218{
219 slob_t *prev, *cur, *aligned = NULL;
220 int delta = 0, units = SLOB_UNITS(size);
221
222 for (prev = NULL, cur = sp->freelist; ; prev = cur, cur = slob_next(cur)) {
223 slobidx_t avail = slob_units(cur);
224
225 if (align) {
226 aligned = (slob_t *)ALIGN((unsigned long)cur, align);
227 delta = aligned - cur;
228 }
229 if (avail >= units + delta) { /* room enough? */
230 slob_t *next;
231
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);
236 prev = cur;
237 cur = aligned;
238 avail = slob_units(cur);
239 }
240
241 next = slob_next(cur);
242 if (avail == units) { /* exact fit? unlink. */
243 if (prev)
244 set_slob(prev, slob_units(prev), next);
245 else
246 sp->freelist = next;
247 } else { /* fragment */
248 if (prev)
249 set_slob(prev, slob_units(prev), cur + units);
250 else
251 sp->freelist = cur + units;
252 set_slob(cur + units, avail - units, next);
253 }
254
255 sp->units -= units;
256 if (!sp->units)
257 clear_slob_page_free(sp);
258 return cur;
259 }
260 if (slob_last(cur))
261 return NULL;
262 }
263}
264
265/*
266 * slob_alloc: entry point into the slob allocator.
267 */
268static void *slob_alloc(size_t size, gfp_t gfp, int align, int node)
269{
270 struct page *sp;
271 struct list_head *prev;
272 struct list_head *slob_list;
273 slob_t *b = NULL;
274 unsigned long flags;
275
276 if (size < SLOB_BREAK1)
277 slob_list = &free_slob_small;
278 else if (size < SLOB_BREAK2)
279 slob_list = &free_slob_medium;
280 else
281 slob_list = &free_slob_large;
282
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) {
286#ifdef CONFIG_NUMA
287 /*
288 * If there's a node specification, search for a partial
289 * page with a matching node id in the freelist.
290 */
291 if (node != NUMA_NO_NODE && page_to_nid(sp) != node)
292 continue;
293#endif
294 /* Enough room on this page? */
295 if (sp->units < SLOB_UNITS(size))
296 continue;
297
298 /* Attempt to alloc */
299 prev = sp->lru.prev;
300 b = slob_page_alloc(sp, size, align);
301 if (!b)
302 continue;
303
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);
310 break;
311 }
312 spin_unlock_irqrestore(&slob_lock, flags);
313
314 /* Not enough space: must allocate a new page */
315 if (!b) {
316 b = slob_new_pages(gfp & ~__GFP_ZERO, 0, node);
317 if (!b)
318 return NULL;
319 sp = virt_to_page(b);
320 __SetPageSlab(sp);
321
322 spin_lock_irqsave(&slob_lock, flags);
323 sp->units = SLOB_UNITS(PAGE_SIZE);
324 sp->freelist = b;
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);
329 BUG_ON(!b);
330 spin_unlock_irqrestore(&slob_lock, flags);
331 }
332 if (unlikely((gfp & __GFP_ZERO) && b))
333 memset(b, 0, size);
334 return b;
335}
336
337/*
338 * slob_free: entry point into the slob allocator.
339 */
340static void slob_free(void *block, int size)
341{
342 struct page *sp;
343 slob_t *prev, *next, *b = (slob_t *)block;
344 slobidx_t units;
345 unsigned long flags;
346 struct list_head *slob_list;
347
348 if (unlikely(ZERO_OR_NULL_PTR(block)))
349 return;
350 BUG_ON(!size);
351
352 sp = virt_to_page(block);
353 units = SLOB_UNITS(size);
354
355 spin_lock_irqsave(&slob_lock, flags);
356
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);
362 __ClearPageSlab(sp);
363 page_mapcount_reset(sp);
364 slob_free_pages(b, 0);
365 return;
366 }
367
368 if (!slob_page_free(sp)) {
369 /* This slob page is about to become partially free. Easy! */
370 sp->units = units;
371 sp->freelist = b;
372 set_slob(b, units,
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;
379 else
380 slob_list = &free_slob_large;
381 set_slob_page_free(sp, slob_list);
382 goto out;
383 }
384
385 /*
386 * Otherwise the page is already partially free, so find reinsertion
387 * point.
388 */
389 sp->units += units;
390
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);
395 }
396 set_slob(b, units, sp->freelist);
397 sp->freelist = b;
398 } else {
399 prev = sp->freelist;
400 next = slob_next(prev);
401 while (b > next) {
402 prev = next;
403 next = slob_next(prev);
404 }
405
406 if (!slob_last(prev) && b + units == next) {
407 units += slob_units(next);
408 set_slob(b, units, slob_next(next));
409 } else
410 set_slob(b, units, next);
411
412 if (prev + slob_units(prev) == b) {
413 units = slob_units(b) + slob_units(prev);
414 set_slob(prev, units, slob_next(b));
415 } else
416 set_slob(prev, slob_units(prev), b);
417 }
418out:
419 spin_unlock_irqrestore(&slob_lock, flags);
420}
421
422/*
423 * End of slob allocator proper. Begin kmem_cache_alloc and kmalloc frontend.
424 */
425
426static __always_inline void *
427__do_kmalloc_node(size_t size, gfp_t gfp, int node, unsigned long caller)
428{
429 unsigned int *m;
430 int align = max_t(size_t, ARCH_KMALLOC_MINALIGN, ARCH_SLAB_MINALIGN);
431 void *ret;
432
433 gfp &= gfp_allowed_mask;
434
435 lockdep_trace_alloc(gfp);
436
437 if (size < PAGE_SIZE - align) {
438 if (!size)
439 return ZERO_SIZE_PTR;
440
441 m = slob_alloc(size + align, gfp, align, node);
442
443 if (!m)
444 return NULL;
445 *m = size;
446 ret = (void *)m + align;
447
448 trace_kmalloc_node(caller, ret,
449 size, size + align, gfp, node);
450 } else {
451 unsigned int order = get_order(size);
452
453 if (likely(order))
454 gfp |= __GFP_COMP;
455 ret = slob_new_pages(gfp, order, node);
456
457 trace_kmalloc_node(caller, ret,
458 size, PAGE_SIZE << order, gfp, node);
459 }
460
461 kmemleak_alloc(ret, size, 1, gfp);
462 return ret;
463}
464
465void *__kmalloc(size_t size, gfp_t gfp)
466{
467 return __do_kmalloc_node(size, gfp, NUMA_NO_NODE, _RET_IP_);
468}
469EXPORT_SYMBOL(__kmalloc);
470
471void *__kmalloc_track_caller(size_t size, gfp_t gfp, unsigned long caller)
472{
473 return __do_kmalloc_node(size, gfp, NUMA_NO_NODE, caller);
474}
475
476#ifdef CONFIG_NUMA
477void *__kmalloc_node_track_caller(size_t size, gfp_t gfp,
478 int node, unsigned long caller)
479{
480 return __do_kmalloc_node(size, gfp, node, caller);
481}
482#endif
483
484void kfree(const void *block)
485{
486 struct page *sp;
487
488 trace_kfree(_RET_IP_, block);
489
490 if (unlikely(ZERO_OR_NULL_PTR(block)))
491 return;
492 kmemleak_free(block);
493
494 sp = virt_to_page(block);
495 if (PageSlab(sp)) {
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);
499 } else
500 __free_pages(sp, compound_order(sp));
501}
502EXPORT_SYMBOL(kfree);
503
504/* can't use ksize for kmem_cache_alloc memory, only kmalloc */
505size_t ksize(const void *block)
506{
507 struct page *sp;
508 int align;
509 unsigned int *m;
510
511 BUG_ON(!block);
512 if (unlikely(block == ZERO_SIZE_PTR))
513 return 0;
514
515 sp = virt_to_page(block);
516 if (unlikely(!PageSlab(sp)))
517 return PAGE_SIZE << compound_order(sp);
518
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;
522}
523EXPORT_SYMBOL(ksize);
524
525int __kmem_cache_create(struct kmem_cache *c, unsigned long flags)
526{
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);
530 }
531 c->flags = flags;
532 return 0;
533}
534
535static void *slob_alloc_node(struct kmem_cache *c, gfp_t flags, int node)
536{
537 void *b;
538
539 flags &= gfp_allowed_mask;
540
541 lockdep_trace_alloc(flags);
542
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,
547 flags, node);
548 } else {
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),
552 flags, node);
553 }
554
555 if (b && c->ctor)
556 c->ctor(b);
557
558 kmemleak_alloc_recursive(b, c->size, 1, c->flags, flags);
559 return b;
560}
561
562void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
563{
564 return slob_alloc_node(cachep, flags, NUMA_NO_NODE);
565}
566EXPORT_SYMBOL(kmem_cache_alloc);
567
568#ifdef CONFIG_NUMA
569void *__kmalloc_node(size_t size, gfp_t gfp, int node)
570{
571 return __do_kmalloc_node(size, gfp, node, _RET_IP_);
572}
573EXPORT_SYMBOL(__kmalloc_node);
574
575void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t gfp, int node)
576{
577 return slob_alloc_node(cachep, gfp, node);
578}
579EXPORT_SYMBOL(kmem_cache_alloc_node);
580#endif
581
582static void __kmem_cache_free(void *b, int size)
583{
584 if (size < PAGE_SIZE)
585 slob_free(b, size);
586 else
587 slob_free_pages(b, get_order(size));
588}
589
590static 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
598void kmem_cache_free(struct kmem_cache *c, void *b)
599{
600 kmemleak_free_recursive(b, c->flags);
601 if (unlikely(c->flags & SLAB_DESTROY_BY_RCU)) {
602 struct slob_rcu *slob_rcu;
603 slob_rcu = b + (c->size - sizeof(struct slob_rcu));
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 trace_kmem_cache_free(_RET_IP_, b);
611}
612EXPORT_SYMBOL(kmem_cache_free);
613
614void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
615{
616 __kmem_cache_free_bulk(s, size, p);
617}
618EXPORT_SYMBOL(kmem_cache_free_bulk);
619
620int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
621 void **p)
622{
623 return __kmem_cache_alloc_bulk(s, flags, size, p);
624}
625EXPORT_SYMBOL(kmem_cache_alloc_bulk);
626
627int __kmem_cache_shutdown(struct kmem_cache *c)
628{
629 /* No way to check for remaining objects */
630 return 0;
631}
632
633void __kmem_cache_release(struct kmem_cache *c)
634{
635}
636
637int __kmem_cache_shrink(struct kmem_cache *d, bool deactivate)
638{
639 return 0;
640}
641
642struct kmem_cache kmem_cache_boot = {
643 .name = "kmem_cache",
644 .size = sizeof(struct kmem_cache),
645 .flags = SLAB_PANIC,
646 .align = ARCH_KMALLOC_MINALIGN,
647};
648
649void __init kmem_cache_init(void)
650{
651 kmem_cache = &kmem_cache_boot;
652 slab_state = UP;
653}
654
655void __init kmem_cache_init_late(void)
656{
657 slab_state = FULL;
658}