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// SPDX-License-Identifier: GPL-2.0
2/*
3 * SLOB Allocator: Simple List Of Blocks
4 *
5 * Matt Mackall <mpm@selenic.com> 12/30/03
6 *
7 * NUMA support by Paul Mundt, 2007.
8 *
9 * How SLOB works:
10 *
11 * The core of SLOB is a traditional K&R style heap allocator, with
12 * support for returning aligned objects. The granularity of this
13 * allocator is as little as 2 bytes, however typically most architectures
14 * will require 4 bytes on 32-bit and 8 bytes on 64-bit.
15 *
16 * The slob heap is a set of linked list of pages from alloc_pages(),
17 * and within each page, there is a singly-linked list of free blocks
18 * (slob_t). The heap is grown on demand. To reduce fragmentation,
19 * heap pages are segregated into three lists, with objects less than
20 * 256 bytes, objects less than 1024 bytes, and all other objects.
21 *
22 * Allocation from heap involves first searching for a page with
23 * sufficient free blocks (using a next-fit-like approach) followed by
24 * a first-fit scan of the page. Deallocation inserts objects back
25 * into the free list in address order, so this is effectively an
26 * address-ordered first fit.
27 *
28 * Above this is an implementation of kmalloc/kfree. Blocks returned
29 * from kmalloc are prepended with a 4-byte header with the kmalloc size.
30 * If kmalloc is asked for objects of PAGE_SIZE or larger, it calls
31 * alloc_pages() directly, allocating compound pages so the page order
32 * does not have to be separately tracked.
33 * These objects are detected in kfree() because PageSlab()
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_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
63#include <linux/mm.h>
64#include <linux/swap.h> /* struct reclaim_state */
65#include <linux/cache.h>
66#include <linux/init.h>
67#include <linux/export.h>
68#include <linux/rcupdate.h>
69#include <linux/list.h>
70#include <linux/kmemleak.h>
71
72#include <trace/events/kmem.h>
73
74#include <linux/atomic.h>
75
76#include "slab.h"
77/*
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).
80 *
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.
84 */
85#if PAGE_SIZE <= (32767 * 2)
86typedef s16 slobidx_t;
87#else
88typedef s32 slobidx_t;
89#endif
90
91struct slob_block {
92 slobidx_t units;
93};
94typedef struct slob_block slob_t;
95
96/*
97 * All partially free slob pages go on these lists.
98 */
99#define SLOB_BREAK1 256
100#define SLOB_BREAK2 1024
101static LIST_HEAD(free_slob_small);
102static LIST_HEAD(free_slob_medium);
103static LIST_HEAD(free_slob_large);
104
105/*
106 * slob_page_free: true for pages on free_slob_pages list.
107 */
108static inline int slob_page_free(struct page *sp)
109{
110 return PageSlobFree(sp);
111}
112
113static void set_slob_page_free(struct page *sp, struct list_head *list)
114{
115 list_add(&sp->slab_list, list);
116 __SetPageSlobFree(sp);
117}
118
119static inline void clear_slob_page_free(struct page *sp)
120{
121 list_del(&sp->slab_list);
122 __ClearPageSlobFree(sp);
123}
124
125#define SLOB_UNIT sizeof(slob_t)
126#define SLOB_UNITS(size) DIV_ROUND_UP(size, SLOB_UNIT)
127
128/*
129 * struct slob_rcu is inserted at the tail of allocated slob blocks, which
130 * were created with a SLAB_TYPESAFE_BY_RCU slab. slob_rcu is used to free
131 * the block using call_rcu.
132 */
133struct slob_rcu {
134 struct rcu_head head;
135 int size;
136};
137
138/*
139 * slob_lock protects all slob allocator structures.
140 */
141static DEFINE_SPINLOCK(slob_lock);
142
143/*
144 * Encode the given size and next info into a free slob block s.
145 */
146static void set_slob(slob_t *s, slobidx_t size, slob_t *next)
147{
148 slob_t *base = (slob_t *)((unsigned long)s & PAGE_MASK);
149 slobidx_t offset = next - base;
150
151 if (size > 1) {
152 s[0].units = size;
153 s[1].units = offset;
154 } else
155 s[0].units = -offset;
156}
157
158/*
159 * Return the size of a slob block.
160 */
161static slobidx_t slob_units(slob_t *s)
162{
163 if (s->units > 0)
164 return s->units;
165 return 1;
166}
167
168/*
169 * Return the next free slob block pointer after this one.
170 */
171static slob_t *slob_next(slob_t *s)
172{
173 slob_t *base = (slob_t *)((unsigned long)s & PAGE_MASK);
174 slobidx_t next;
175
176 if (s[0].units < 0)
177 next = -s[0].units;
178 else
179 next = s[1].units;
180 return base+next;
181}
182
183/*
184 * Returns true if s is the last free block in its page.
185 */
186static int slob_last(slob_t *s)
187{
188 return !((unsigned long)slob_next(s) & ~PAGE_MASK);
189}
190
191static void *slob_new_pages(gfp_t gfp, int order, int node)
192{
193 struct page *page;
194
195#ifdef CONFIG_NUMA
196 if (node != NUMA_NO_NODE)
197 page = __alloc_pages_node(node, gfp, order);
198 else
199#endif
200 page = alloc_pages(gfp, order);
201
202 if (!page)
203 return NULL;
204
205 mod_node_page_state(page_pgdat(page), NR_SLAB_UNRECLAIMABLE_B,
206 PAGE_SIZE << order);
207 return page_address(page);
208}
209
210static void slob_free_pages(void *b, int order)
211{
212 struct page *sp = virt_to_page(b);
213
214 if (current->reclaim_state)
215 current->reclaim_state->reclaimed_slab += 1 << order;
216
217 mod_node_page_state(page_pgdat(sp), NR_SLAB_UNRECLAIMABLE_B,
218 -(PAGE_SIZE << order));
219 __free_pages(sp, order);
220}
221
222/*
223 * slob_page_alloc() - Allocate a slob block within a given slob_page sp.
224 * @sp: Page to look in.
225 * @size: Size of the allocation.
226 * @align: Allocation alignment.
227 * @align_offset: Offset in the allocated block that will be aligned.
228 * @page_removed_from_list: Return parameter.
229 *
230 * Tries to find a chunk of memory at least @size bytes big within @page.
231 *
232 * Return: Pointer to memory if allocated, %NULL otherwise. If the
233 * allocation fills up @page then the page is removed from the
234 * freelist, in this case @page_removed_from_list will be set to
235 * true (set to false otherwise).
236 */
237static void *slob_page_alloc(struct page *sp, size_t size, int align,
238 int align_offset, bool *page_removed_from_list)
239{
240 slob_t *prev, *cur, *aligned = NULL;
241 int delta = 0, units = SLOB_UNITS(size);
242
243 *page_removed_from_list = false;
244 for (prev = NULL, cur = sp->freelist; ; prev = cur, cur = slob_next(cur)) {
245 slobidx_t avail = slob_units(cur);
246
247 /*
248 * 'aligned' will hold the address of the slob block so that the
249 * address 'aligned'+'align_offset' is aligned according to the
250 * 'align' parameter. This is for kmalloc() which prepends the
251 * allocated block with its size, so that the block itself is
252 * aligned when needed.
253 */
254 if (align) {
255 aligned = (slob_t *)
256 (ALIGN((unsigned long)cur + align_offset, align)
257 - align_offset);
258 delta = aligned - cur;
259 }
260 if (avail >= units + delta) { /* room enough? */
261 slob_t *next;
262
263 if (delta) { /* need to fragment head to align? */
264 next = slob_next(cur);
265 set_slob(aligned, avail - delta, next);
266 set_slob(cur, delta, aligned);
267 prev = cur;
268 cur = aligned;
269 avail = slob_units(cur);
270 }
271
272 next = slob_next(cur);
273 if (avail == units) { /* exact fit? unlink. */
274 if (prev)
275 set_slob(prev, slob_units(prev), next);
276 else
277 sp->freelist = next;
278 } else { /* fragment */
279 if (prev)
280 set_slob(prev, slob_units(prev), cur + units);
281 else
282 sp->freelist = cur + units;
283 set_slob(cur + units, avail - units, next);
284 }
285
286 sp->units -= units;
287 if (!sp->units) {
288 clear_slob_page_free(sp);
289 *page_removed_from_list = true;
290 }
291 return cur;
292 }
293 if (slob_last(cur))
294 return NULL;
295 }
296}
297
298/*
299 * slob_alloc: entry point into the slob allocator.
300 */
301static void *slob_alloc(size_t size, gfp_t gfp, int align, int node,
302 int align_offset)
303{
304 struct page *sp;
305 struct list_head *slob_list;
306 slob_t *b = NULL;
307 unsigned long flags;
308 bool _unused;
309
310 if (size < SLOB_BREAK1)
311 slob_list = &free_slob_small;
312 else if (size < SLOB_BREAK2)
313 slob_list = &free_slob_medium;
314 else
315 slob_list = &free_slob_large;
316
317 spin_lock_irqsave(&slob_lock, flags);
318 /* Iterate through each partially free page, try to find room */
319 list_for_each_entry(sp, slob_list, slab_list) {
320 bool page_removed_from_list = false;
321#ifdef CONFIG_NUMA
322 /*
323 * If there's a node specification, search for a partial
324 * page with a matching node id in the freelist.
325 */
326 if (node != NUMA_NO_NODE && page_to_nid(sp) != node)
327 continue;
328#endif
329 /* Enough room on this page? */
330 if (sp->units < SLOB_UNITS(size))
331 continue;
332
333 b = slob_page_alloc(sp, size, align, align_offset, &page_removed_from_list);
334 if (!b)
335 continue;
336
337 /*
338 * If slob_page_alloc() removed sp from the list then we
339 * cannot call list functions on sp. If so allocation
340 * did not fragment the page anyway so optimisation is
341 * unnecessary.
342 */
343 if (!page_removed_from_list) {
344 /*
345 * Improve fragment distribution and reduce our average
346 * search time by starting our next search here. (see
347 * Knuth vol 1, sec 2.5, pg 449)
348 */
349 if (!list_is_first(&sp->slab_list, slob_list))
350 list_rotate_to_front(&sp->slab_list, slob_list);
351 }
352 break;
353 }
354 spin_unlock_irqrestore(&slob_lock, flags);
355
356 /* Not enough space: must allocate a new page */
357 if (!b) {
358 b = slob_new_pages(gfp & ~__GFP_ZERO, 0, node);
359 if (!b)
360 return NULL;
361 sp = virt_to_page(b);
362 __SetPageSlab(sp);
363
364 spin_lock_irqsave(&slob_lock, flags);
365 sp->units = SLOB_UNITS(PAGE_SIZE);
366 sp->freelist = b;
367 INIT_LIST_HEAD(&sp->slab_list);
368 set_slob(b, SLOB_UNITS(PAGE_SIZE), b + SLOB_UNITS(PAGE_SIZE));
369 set_slob_page_free(sp, slob_list);
370 b = slob_page_alloc(sp, size, align, align_offset, &_unused);
371 BUG_ON(!b);
372 spin_unlock_irqrestore(&slob_lock, flags);
373 }
374 if (unlikely(gfp & __GFP_ZERO))
375 memset(b, 0, size);
376 return b;
377}
378
379/*
380 * slob_free: entry point into the slob allocator.
381 */
382static void slob_free(void *block, int size)
383{
384 struct page *sp;
385 slob_t *prev, *next, *b = (slob_t *)block;
386 slobidx_t units;
387 unsigned long flags;
388 struct list_head *slob_list;
389
390 if (unlikely(ZERO_OR_NULL_PTR(block)))
391 return;
392 BUG_ON(!size);
393
394 sp = virt_to_page(block);
395 units = SLOB_UNITS(size);
396
397 spin_lock_irqsave(&slob_lock, flags);
398
399 if (sp->units + units == SLOB_UNITS(PAGE_SIZE)) {
400 /* Go directly to page allocator. Do not pass slob allocator */
401 if (slob_page_free(sp))
402 clear_slob_page_free(sp);
403 spin_unlock_irqrestore(&slob_lock, flags);
404 __ClearPageSlab(sp);
405 page_mapcount_reset(sp);
406 slob_free_pages(b, 0);
407 return;
408 }
409
410 if (!slob_page_free(sp)) {
411 /* This slob page is about to become partially free. Easy! */
412 sp->units = units;
413 sp->freelist = b;
414 set_slob(b, units,
415 (void *)((unsigned long)(b +
416 SLOB_UNITS(PAGE_SIZE)) & PAGE_MASK));
417 if (size < SLOB_BREAK1)
418 slob_list = &free_slob_small;
419 else if (size < SLOB_BREAK2)
420 slob_list = &free_slob_medium;
421 else
422 slob_list = &free_slob_large;
423 set_slob_page_free(sp, slob_list);
424 goto out;
425 }
426
427 /*
428 * Otherwise the page is already partially free, so find reinsertion
429 * point.
430 */
431 sp->units += units;
432
433 if (b < (slob_t *)sp->freelist) {
434 if (b + units == sp->freelist) {
435 units += slob_units(sp->freelist);
436 sp->freelist = slob_next(sp->freelist);
437 }
438 set_slob(b, units, sp->freelist);
439 sp->freelist = b;
440 } else {
441 prev = sp->freelist;
442 next = slob_next(prev);
443 while (b > next) {
444 prev = next;
445 next = slob_next(prev);
446 }
447
448 if (!slob_last(prev) && b + units == next) {
449 units += slob_units(next);
450 set_slob(b, units, slob_next(next));
451 } else
452 set_slob(b, units, next);
453
454 if (prev + slob_units(prev) == b) {
455 units = slob_units(b) + slob_units(prev);
456 set_slob(prev, units, slob_next(b));
457 } else
458 set_slob(prev, slob_units(prev), b);
459 }
460out:
461 spin_unlock_irqrestore(&slob_lock, flags);
462}
463
464/*
465 * End of slob allocator proper. Begin kmem_cache_alloc and kmalloc frontend.
466 */
467
468static __always_inline void *
469__do_kmalloc_node(size_t size, gfp_t gfp, int node, unsigned long caller)
470{
471 unsigned int *m;
472 int minalign = max_t(size_t, ARCH_KMALLOC_MINALIGN, ARCH_SLAB_MINALIGN);
473 void *ret;
474
475 gfp &= gfp_allowed_mask;
476
477 fs_reclaim_acquire(gfp);
478 fs_reclaim_release(gfp);
479
480 if (size < PAGE_SIZE - minalign) {
481 int align = minalign;
482
483 /*
484 * For power of two sizes, guarantee natural alignment for
485 * kmalloc()'d objects.
486 */
487 if (is_power_of_2(size))
488 align = max(minalign, (int) size);
489
490 if (!size)
491 return ZERO_SIZE_PTR;
492
493 m = slob_alloc(size + minalign, gfp, align, node, minalign);
494
495 if (!m)
496 return NULL;
497 *m = size;
498 ret = (void *)m + minalign;
499
500 trace_kmalloc_node(caller, ret,
501 size, size + minalign, 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
509 trace_kmalloc_node(caller, ret,
510 size, PAGE_SIZE << order, gfp, node);
511 }
512
513 kmemleak_alloc(ret, size, 1, gfp);
514 return ret;
515}
516
517void *__kmalloc(size_t size, gfp_t gfp)
518{
519 return __do_kmalloc_node(size, gfp, NUMA_NO_NODE, _RET_IP_);
520}
521EXPORT_SYMBOL(__kmalloc);
522
523void *__kmalloc_track_caller(size_t size, gfp_t gfp, unsigned long caller)
524{
525 return __do_kmalloc_node(size, gfp, NUMA_NO_NODE, caller);
526}
527EXPORT_SYMBOL(__kmalloc_track_caller);
528
529#ifdef CONFIG_NUMA
530void *__kmalloc_node_track_caller(size_t size, gfp_t gfp,
531 int node, unsigned long caller)
532{
533 return __do_kmalloc_node(size, gfp, node, caller);
534}
535EXPORT_SYMBOL(__kmalloc_node_track_caller);
536#endif
537
538void kfree(const void *block)
539{
540 struct page *sp;
541
542 trace_kfree(_RET_IP_, block);
543
544 if (unlikely(ZERO_OR_NULL_PTR(block)))
545 return;
546 kmemleak_free(block);
547
548 sp = virt_to_page(block);
549 if (PageSlab(sp)) {
550 int align = max_t(size_t, ARCH_KMALLOC_MINALIGN, ARCH_SLAB_MINALIGN);
551 unsigned int *m = (unsigned int *)(block - align);
552 slob_free(m, *m + align);
553 } else {
554 unsigned int order = compound_order(sp);
555 mod_node_page_state(page_pgdat(sp), NR_SLAB_UNRECLAIMABLE_B,
556 -(PAGE_SIZE << order));
557 __free_pages(sp, order);
558
559 }
560}
561EXPORT_SYMBOL(kfree);
562
563/* can't use ksize for kmem_cache_alloc memory, only kmalloc */
564size_t __ksize(const void *block)
565{
566 struct page *sp;
567 int align;
568 unsigned int *m;
569
570 BUG_ON(!block);
571 if (unlikely(block == ZERO_SIZE_PTR))
572 return 0;
573
574 sp = virt_to_page(block);
575 if (unlikely(!PageSlab(sp)))
576 return page_size(sp);
577
578 align = max_t(size_t, ARCH_KMALLOC_MINALIGN, ARCH_SLAB_MINALIGN);
579 m = (unsigned int *)(block - align);
580 return SLOB_UNITS(*m) * SLOB_UNIT;
581}
582EXPORT_SYMBOL(__ksize);
583
584int __kmem_cache_create(struct kmem_cache *c, slab_flags_t flags)
585{
586 if (flags & SLAB_TYPESAFE_BY_RCU) {
587 /* leave room for rcu footer at the end of object */
588 c->size += sizeof(struct slob_rcu);
589 }
590 c->flags = flags;
591 return 0;
592}
593
594static void *slob_alloc_node(struct kmem_cache *c, gfp_t flags, int node)
595{
596 void *b;
597
598 flags &= gfp_allowed_mask;
599
600 fs_reclaim_acquire(flags);
601 fs_reclaim_release(flags);
602
603 if (c->size < PAGE_SIZE) {
604 b = slob_alloc(c->size, flags, c->align, node, 0);
605 trace_kmem_cache_alloc_node(_RET_IP_, b, c->object_size,
606 SLOB_UNITS(c->size) * SLOB_UNIT,
607 flags, node);
608 } else {
609 b = slob_new_pages(flags, get_order(c->size), node);
610 trace_kmem_cache_alloc_node(_RET_IP_, b, c->object_size,
611 PAGE_SIZE << get_order(c->size),
612 flags, node);
613 }
614
615 if (b && c->ctor) {
616 WARN_ON_ONCE(flags & __GFP_ZERO);
617 c->ctor(b);
618 }
619
620 kmemleak_alloc_recursive(b, c->size, 1, c->flags, flags);
621 return b;
622}
623
624void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
625{
626 return slob_alloc_node(cachep, flags, NUMA_NO_NODE);
627}
628EXPORT_SYMBOL(kmem_cache_alloc);
629
630#ifdef CONFIG_NUMA
631void *__kmalloc_node(size_t size, gfp_t gfp, int node)
632{
633 return __do_kmalloc_node(size, gfp, node, _RET_IP_);
634}
635EXPORT_SYMBOL(__kmalloc_node);
636
637void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t gfp, int node)
638{
639 return slob_alloc_node(cachep, gfp, node);
640}
641EXPORT_SYMBOL(kmem_cache_alloc_node);
642#endif
643
644static void __kmem_cache_free(void *b, int size)
645{
646 if (size < PAGE_SIZE)
647 slob_free(b, size);
648 else
649 slob_free_pages(b, get_order(size));
650}
651
652static void kmem_rcu_free(struct rcu_head *head)
653{
654 struct slob_rcu *slob_rcu = (struct slob_rcu *)head;
655 void *b = (void *)slob_rcu - (slob_rcu->size - sizeof(struct slob_rcu));
656
657 __kmem_cache_free(b, slob_rcu->size);
658}
659
660void kmem_cache_free(struct kmem_cache *c, void *b)
661{
662 kmemleak_free_recursive(b, c->flags);
663 if (unlikely(c->flags & SLAB_TYPESAFE_BY_RCU)) {
664 struct slob_rcu *slob_rcu;
665 slob_rcu = b + (c->size - sizeof(struct slob_rcu));
666 slob_rcu->size = c->size;
667 call_rcu(&slob_rcu->head, kmem_rcu_free);
668 } else {
669 __kmem_cache_free(b, c->size);
670 }
671
672 trace_kmem_cache_free(_RET_IP_, b);
673}
674EXPORT_SYMBOL(kmem_cache_free);
675
676void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
677{
678 __kmem_cache_free_bulk(s, size, p);
679}
680EXPORT_SYMBOL(kmem_cache_free_bulk);
681
682int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
683 void **p)
684{
685 return __kmem_cache_alloc_bulk(s, flags, size, p);
686}
687EXPORT_SYMBOL(kmem_cache_alloc_bulk);
688
689int __kmem_cache_shutdown(struct kmem_cache *c)
690{
691 /* No way to check for remaining objects */
692 return 0;
693}
694
695void __kmem_cache_release(struct kmem_cache *c)
696{
697}
698
699int __kmem_cache_shrink(struct kmem_cache *d)
700{
701 return 0;
702}
703
704struct kmem_cache kmem_cache_boot = {
705 .name = "kmem_cache",
706 .size = sizeof(struct kmem_cache),
707 .flags = SLAB_PANIC,
708 .align = ARCH_KMALLOC_MINALIGN,
709};
710
711void __init kmem_cache_init(void)
712{
713 kmem_cache = &kmem_cache_boot;
714 slab_state = UP;
715}
716
717void __init kmem_cache_init_late(void)
718{
719 slab_state = FULL;
720}