• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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_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().
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)
85 typedef s16 slobidx_t;
86 #else
87 typedef s32 slobidx_t;
88 #endif
89 
90 struct slob_block {
91 	slobidx_t units;
92 };
93 typedef 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
100 static LIST_HEAD(free_slob_small);
101 static LIST_HEAD(free_slob_medium);
102 static LIST_HEAD(free_slob_large);
103 
104 /*
105  * slob_page_free: true for pages on free_slob_pages list.
106  */
slob_page_free(struct page * sp)107 static inline int slob_page_free(struct page *sp)
108 {
109 	return PageSlobFree(sp);
110 }
111 
set_slob_page_free(struct page * sp,struct list_head * list)112 static void set_slob_page_free(struct page *sp, struct list_head *list)
113 {
114 	list_add(&sp->lru, list);
115 	__SetPageSlobFree(sp);
116 }
117 
clear_slob_page_free(struct page * sp)118 static 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  */
132 struct slob_rcu {
133 	struct rcu_head head;
134 	int size;
135 };
136 
137 /*
138  * slob_lock protects all slob allocator structures.
139  */
140 static DEFINE_SPINLOCK(slob_lock);
141 
142 /*
143  * Encode the given size and next info into a free slob block s.
144  */
set_slob(slob_t * s,slobidx_t size,slob_t * next)145 static 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  */
slob_units(slob_t * s)160 static 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  */
slob_next(slob_t * s)170 static 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  */
slob_last(slob_t * s)185 static int slob_last(slob_t *s)
186 {
187 	return !((unsigned long)slob_next(s) & ~PAGE_MASK);
188 }
189 
slob_new_pages(gfp_t gfp,int order,int node)190 static 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_exact_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 
slob_free_pages(void * b,int order)207 static 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  */
slob_page_alloc(struct page * sp,size_t size,int align)217 static 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  */
slob_alloc(size_t size,gfp_t gfp,int align,int node)268 static 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  */
slob_free(void * block,int size)340 static 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 	}
418 out:
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 
426 static __always_inline void *
__do_kmalloc_node(size_t size,gfp_t gfp,int node,unsigned long caller)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 
__kmalloc(size_t size,gfp_t gfp)465 void *__kmalloc(size_t size, gfp_t gfp)
466 {
467 	return __do_kmalloc_node(size, gfp, NUMA_NO_NODE, _RET_IP_);
468 }
469 EXPORT_SYMBOL(__kmalloc);
470 
__kmalloc_track_caller(size_t size,gfp_t gfp,unsigned long caller)471 void *__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
__kmalloc_node_track_caller(size_t size,gfp_t gfp,int node,unsigned long caller)477 void *__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 
kfree(const void * block)484 void 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 }
502 EXPORT_SYMBOL(kfree);
503 
504 /* can't use ksize for kmem_cache_alloc memory, only kmalloc */
ksize(const void * block)505 size_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 }
523 EXPORT_SYMBOL(ksize);
524 
__kmem_cache_create(struct kmem_cache * c,unsigned long flags)525 int __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 
slob_alloc_node(struct kmem_cache * c,gfp_t flags,int node)535 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 EXPORT_SYMBOL(slob_alloc_node);
562 
kmem_cache_alloc(struct kmem_cache * cachep,gfp_t flags)563 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
564 {
565 	return slob_alloc_node(cachep, flags, NUMA_NO_NODE);
566 }
567 EXPORT_SYMBOL(kmem_cache_alloc);
568 
569 #ifdef CONFIG_NUMA
__kmalloc_node(size_t size,gfp_t gfp,int node)570 void *__kmalloc_node(size_t size, gfp_t gfp, int node)
571 {
572 	return __do_kmalloc_node(size, gfp, node, _RET_IP_);
573 }
574 EXPORT_SYMBOL(__kmalloc_node);
575 
kmem_cache_alloc_node(struct kmem_cache * cachep,gfp_t gfp,int node)576 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t gfp, int node)
577 {
578 	return slob_alloc_node(cachep, gfp, node);
579 }
580 EXPORT_SYMBOL(kmem_cache_alloc_node);
581 #endif
582 
__kmem_cache_free(void * b,int size)583 static void __kmem_cache_free(void *b, int size)
584 {
585 	if (size < PAGE_SIZE)
586 		slob_free(b, size);
587 	else
588 		slob_free_pages(b, get_order(size));
589 }
590 
kmem_rcu_free(struct rcu_head * head)591 static void kmem_rcu_free(struct rcu_head *head)
592 {
593 	struct slob_rcu *slob_rcu = (struct slob_rcu *)head;
594 	void *b = (void *)slob_rcu - (slob_rcu->size - sizeof(struct slob_rcu));
595 
596 	__kmem_cache_free(b, slob_rcu->size);
597 }
598 
kmem_cache_free(struct kmem_cache * c,void * b)599 void kmem_cache_free(struct kmem_cache *c, void *b)
600 {
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);
607 	} else {
608 		__kmem_cache_free(b, c->size);
609 	}
610 
611 	trace_kmem_cache_free(_RET_IP_, b);
612 }
613 EXPORT_SYMBOL(kmem_cache_free);
614 
__kmem_cache_shutdown(struct kmem_cache * c)615 int __kmem_cache_shutdown(struct kmem_cache *c)
616 {
617 	/* No way to check for remaining objects */
618 	return 0;
619 }
620 
__kmem_cache_shrink(struct kmem_cache * d)621 int __kmem_cache_shrink(struct kmem_cache *d)
622 {
623 	return 0;
624 }
625 
626 struct kmem_cache kmem_cache_boot = {
627 	.name = "kmem_cache",
628 	.size = sizeof(struct kmem_cache),
629 	.flags = SLAB_PANIC,
630 	.align = ARCH_KMALLOC_MINALIGN,
631 };
632 
kmem_cache_init(void)633 void __init kmem_cache_init(void)
634 {
635 	kmem_cache = &kmem_cache_boot;
636 	slab_state = UP;
637 }
638 
kmem_cache_init_late(void)639 void __init kmem_cache_init_late(void)
640 {
641 	slab_state = FULL;
642 }
643