10 #include <ccan/build_assert/build_assert.h>
11 #include <ccan/likely/likely.h>
12 #include <ccan/alignof/alignof.h>
13 #include <ccan/short_types/short_types.h>
14 #include <ccan/compiler/compiler.h>
18 Inspired by (and parts taken from) Andrew Tridgell's alloc_mmap:
19 http://samba.org/~tridge/junkcode/alloc_mmap/
21 Copyright (C) Andrew Tridgell 2007
23 This library is free software; you can redistribute it and/or
24 modify it under the terms of the GNU Lesser General Public
25 License as published by the Free Software Foundation; either
26 version 2 of the License, or (at your option) any later version.
28 This library is distributed in the hope that it will be useful,
29 but WITHOUT ANY WARRANTY; without even the implied warranty of
30 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
31 Lesser General Public License for more details.
33 You should have received a copy of the GNU Lesser General Public
34 License along with this library; if not, write to the Free Software
35 Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
38 /* We divide the pool into this many large pages (nearest power of 2) */
39 #define MAX_LARGE_PAGES (256UL)
41 /* 32 small pages == 1 large page. */
42 #define BITS_FROM_SMALL_TO_LARGE_PAGE 5
44 #define MAX_SMALL_PAGES (MAX_LARGE_PAGES << BITS_FROM_SMALL_TO_LARGE_PAGE)
46 /* Smallest pool size for this scheme: 128-byte small pages. That's
47 * 9/13% overhead for 32/64 bit. */
48 #define MIN_USEFUL_SIZE (MAX_SMALL_PAGES * 128)
50 /* Every 4 buckets, we jump up a power of 2. ...8 10 12 14 16 20 24 28 32... */
51 #define INTER_BUCKET_SPACE 4
53 #define SMALL_PAGES_PER_LARGE_PAGE (1 << BITS_FROM_SMALL_TO_LARGE_PAGE)
55 /* FIXME: Figure this out properly. */
56 #define MAX_SIZE (1 << 30)
58 /* How few object to fit in a page before using a larger one? (8) */
59 #define MAX_PAGE_OBJECT_ORDER 3
61 #define BITS_PER_LONG (sizeof(long) * CHAR_BIT)
64 u32 elements_per_page;
70 /* Bitmap of which pages are large. */
71 unsigned long pagesize[MAX_LARGE_PAGES / BITS_PER_LONG];
73 /* List of unused small/large pages. */
77 /* List of huge allocs. */
80 /* This is less defined: we have two buckets for each power of 2 */
81 struct bucket_state bs[1];
85 unsigned long next, prev;
86 unsigned long off, len;
91 /* FIXME: We can just count all-0 and all-1 used[] elements. */
92 unsigned elements_used : 25;
94 unsigned long used[1]; /* One bit per element. */
98 * Every 4 buckets, the size doubles.
99 * Between buckets, sizes increase linearly.
101 * eg. bucket 40 = 2^10 = 1024
102 * bucket 41 = 2^10 + 2^10*4 = 1024 + 256
103 * bucket 42 = 2^10 + 2^10*4 = 1024 + 512
104 * bucket 43 = 2^10 + 2^10*4 = 1024 + 768
105 * bucket 45 = 2^11 = 2048
107 * Care is taken to handle low numbered buckets, at cost of overflow.
109 static unsigned long bucket_to_size(unsigned int bucket)
111 unsigned long base = 1UL << (bucket / INTER_BUCKET_SPACE);
112 return base + ((bucket % INTER_BUCKET_SPACE)
113 << (bucket / INTER_BUCKET_SPACE))
114 / INTER_BUCKET_SPACE;
119 * fls(size/2) == 3. 1 << 3 == 8, so we're 2 too large, out of a possible
120 * 8 too large. That's 1/4 of the way to the next power of 2 == 1 bucket.
122 * We make sure we round up. Note that this fails on 32 bit at size
123 * 1879048193 (around bucket 120).
125 static unsigned int size_to_bucket(unsigned long size)
127 unsigned int base = afls(size/2);
128 unsigned long overshoot;
130 overshoot = size - (1UL << base);
131 return base * INTER_BUCKET_SPACE
132 + ((overshoot * INTER_BUCKET_SPACE + (1UL << base)-1) >> base);
135 static unsigned int small_page_bits(unsigned long poolsize)
137 return afls(poolsize / MAX_SMALL_PAGES - 1);
140 static struct page_header *from_pgnum(struct header *head,
144 return (struct page_header *)((char *)head + (pgnum << sp_bits));
147 static u16 to_pgnum(struct header *head, void *p, unsigned sp_bits)
149 return ((char *)p - (char *)head) >> sp_bits;
152 static size_t used_size(unsigned int num_elements)
154 return align_up(num_elements, BITS_PER_LONG) / CHAR_BIT;
158 * We always align the first entry to the lower power of 2.
159 * eg. the 12-byte bucket gets 8-byte aligned. The 4096-byte bucket
160 * gets 4096-byte aligned.
162 static unsigned long page_header_size(unsigned int align_bits,
163 unsigned long num_elements)
167 size = sizeof(struct page_header)
168 - sizeof(((struct page_header *)0)->used)
169 + used_size(num_elements);
170 return align_up(size, 1UL << align_bits);
173 static void add_to_list(struct header *head,
174 u16 *list, struct page_header *ph, unsigned sp_bits)
176 unsigned long h = *list, offset = to_pgnum(head, ph, sp_bits);
180 struct page_header *prev = from_pgnum(head, h, sp_bits);
181 assert(prev->prev == 0);
188 static void del_from_list(struct header *head,
189 u16 *list, struct page_header *ph, unsigned sp_bits)
195 struct page_header *prev = from_pgnum(head, ph->prev, sp_bits);
196 prev->next = ph->next;
199 struct page_header *next = from_pgnum(head, ph->next, sp_bits);
200 next->prev = ph->prev;
204 static u16 pop_from_list(struct header *head,
206 unsigned int sp_bits)
209 struct page_header *ph = from_pgnum(head, h, sp_bits);
214 from_pgnum(head, *list, sp_bits)->prev = 0;
219 static void add_to_huge_list(struct header *head, struct huge_alloc *ha)
221 unsigned long h = head->huge;
222 unsigned long offset = (char *)ha - (char *)head;
226 struct huge_alloc *prev = (void *)((char *)head + h);
227 assert(prev->prev == 0);
234 static void del_from_huge(struct header *head, struct huge_alloc *ha)
238 head->huge = ha->next;
240 struct huge_alloc *prev = (void *)((char *)head + ha->prev);
241 prev->next = ha->next;
244 struct huge_alloc *next = (void *)((char *)head + ha->next);
245 next->prev = ha->prev;
249 static void add_small_page_to_freelist(struct header *head,
250 struct page_header *ph,
251 unsigned int sp_bits)
253 add_to_list(head, &head->small_free_list, ph, sp_bits);
256 static void add_large_page_to_freelist(struct header *head,
257 struct page_header *ph,
258 unsigned int sp_bits)
260 add_to_list(head, &head->large_free_list, ph, sp_bits);
263 static void add_to_bucket_list(struct header *head,
264 struct bucket_state *bs,
265 struct page_header *ph,
266 unsigned int sp_bits)
268 add_to_list(head, &bs->page_list, ph, sp_bits);
271 static void del_from_bucket_list(struct header *head,
272 struct bucket_state *bs,
273 struct page_header *ph,
274 unsigned int sp_bits)
276 del_from_list(head, &bs->page_list, ph, sp_bits);
279 static void del_from_bucket_full_list(struct header *head,
280 struct bucket_state *bs,
281 struct page_header *ph,
282 unsigned int sp_bits)
284 del_from_list(head, &bs->full_list, ph, sp_bits);
287 static void add_to_bucket_full_list(struct header *head,
288 struct bucket_state *bs,
289 struct page_header *ph,
290 unsigned int sp_bits)
292 add_to_list(head, &bs->full_list, ph, sp_bits);
295 static void clear_bit(unsigned long bitmap[], unsigned int off)
297 bitmap[off / BITS_PER_LONG] &= ~(1UL << (off % BITS_PER_LONG));
300 static bool test_bit(const unsigned long bitmap[], unsigned int off)
302 return bitmap[off / BITS_PER_LONG] & (1UL << (off % BITS_PER_LONG));
305 static void set_bit(unsigned long bitmap[], unsigned int off)
307 bitmap[off / BITS_PER_LONG] |= (1UL << (off % BITS_PER_LONG));
310 /* There must be a bit to be found. */
311 static unsigned int find_free_bit(const unsigned long bitmap[])
315 for (i = 0; bitmap[i] == -1UL; i++);
316 return (i*BITS_PER_LONG) + affsl(~bitmap[i]) - 1;
319 /* How many elements can we fit in a page? */
320 static unsigned long elements_per_page(unsigned long align_bits,
324 unsigned long num, overhead;
326 /* First approximation: no extra room for bitmap. */
327 overhead = align_up(sizeof(struct page_header), 1UL << align_bits);
328 num = (psize - overhead) / esize;
330 while (page_header_size(align_bits, num) + esize * num > psize)
335 static bool large_page_bucket(unsigned int bucket, unsigned int sp_bits)
337 unsigned long max_smallsize;
339 /* Note: this doesn't take into account page header. */
340 max_smallsize = (1UL << sp_bits) >> MAX_PAGE_OBJECT_ORDER;
342 return bucket_to_size(bucket) > max_smallsize;
345 static unsigned int max_bucket(unsigned int lp_bits)
347 return (lp_bits - MAX_PAGE_OBJECT_ORDER) * INTER_BUCKET_SPACE;
350 void alloc_init(void *pool, unsigned long poolsize)
352 struct header *head = pool;
353 struct page_header *ph;
354 unsigned int lp_bits, sp_bits, num_buckets;
355 unsigned long header_size, i;
357 if (poolsize < MIN_USEFUL_SIZE) {
358 tiny_alloc_init(pool, poolsize);
362 /* We rely on page numbers fitting in 16 bit. */
363 BUILD_ASSERT(MAX_SMALL_PAGES < 65536);
365 sp_bits = small_page_bits(poolsize);
366 lp_bits = sp_bits + BITS_FROM_SMALL_TO_LARGE_PAGE;
368 num_buckets = max_bucket(lp_bits);
371 header_size = sizeof(*head) + sizeof(head->bs) * (num_buckets-1);
373 memset(head, 0, header_size);
374 for (i = 0; i < num_buckets; i++) {
375 unsigned long pagesize;
377 if (large_page_bucket(i, sp_bits))
378 pagesize = 1UL << lp_bits;
380 pagesize = 1UL << sp_bits;
382 head->bs[i].elements_per_page
383 = elements_per_page(i / INTER_BUCKET_SPACE,
388 /* They start as all large pages. */
389 memset(head->pagesize, 0xFF, sizeof(head->pagesize));
390 /* FIXME: small pages for last bit? */
392 /* Split first page into small pages. */
393 assert(header_size < (1UL << lp_bits));
394 clear_bit(head->pagesize, 0);
396 /* Skip over page(s) used by header, add rest to free list */
397 for (i = align_up(header_size, (1UL << sp_bits)) >> sp_bits;
398 i < SMALL_PAGES_PER_LARGE_PAGE;
400 ph = from_pgnum(head, i, sp_bits);
401 ph->elements_used = 0;
402 add_small_page_to_freelist(head, ph, sp_bits);
405 /* Add the rest of the pages as large pages. */
406 i = SMALL_PAGES_PER_LARGE_PAGE;
407 while ((i << sp_bits) + (1UL << lp_bits) <= poolsize) {
408 assert(i < MAX_SMALL_PAGES);
409 ph = from_pgnum(head, i, sp_bits);
410 ph->elements_used = 0;
411 add_large_page_to_freelist(head, ph, sp_bits);
412 i += SMALL_PAGES_PER_LARGE_PAGE;
416 /* A large page worth of small pages are free: delete them from free list. */
417 static void del_large_from_small_free_list(struct header *head,
418 struct page_header *ph,
419 unsigned int sp_bits)
423 for (i = 0; i < SMALL_PAGES_PER_LARGE_PAGE; i++) {
424 del_from_list(head, &head->small_free_list,
425 (struct page_header *)((char *)ph
431 static bool all_empty(struct header *head,
437 for (i = 0; i < SMALL_PAGES_PER_LARGE_PAGE; i++) {
438 struct page_header *ph = from_pgnum(head, pgnum + i, sp_bits);
439 if (ph->elements_used)
445 static void recombine_small_pages(struct header *head, unsigned long poolsize,
446 unsigned int sp_bits)
449 unsigned int lp_bits = sp_bits + BITS_FROM_SMALL_TO_LARGE_PAGE;
451 /* Look for small pages to coalesce, after first large page. */
452 for (i = SMALL_PAGES_PER_LARGE_PAGE;
453 i < (poolsize >> lp_bits) << BITS_FROM_SMALL_TO_LARGE_PAGE;
454 i += SMALL_PAGES_PER_LARGE_PAGE) {
455 /* Already a large page? */
456 if (test_bit(head->pagesize, i / SMALL_PAGES_PER_LARGE_PAGE))
458 if (all_empty(head, i, sp_bits)) {
459 struct page_header *ph = from_pgnum(head, i, sp_bits);
460 set_bit(head->pagesize,
461 i / SMALL_PAGES_PER_LARGE_PAGE);
462 del_large_from_small_free_list(head, ph, sp_bits);
463 add_large_page_to_freelist(head, ph, sp_bits);
468 static u16 get_large_page(struct header *head, unsigned long poolsize,
469 unsigned int sp_bits)
473 page = pop_from_list(head, &head->large_free_list, sp_bits);
477 recombine_small_pages(head, poolsize, sp_bits);
479 return pop_from_list(head, &head->large_free_list, sp_bits);
482 /* Returns small page. */
483 static unsigned long break_up_large_page(struct header *head,
484 unsigned int sp_bits,
489 clear_bit(head->pagesize, lpage >> BITS_FROM_SMALL_TO_LARGE_PAGE);
491 for (i = 1; i < SMALL_PAGES_PER_LARGE_PAGE; i++) {
492 struct page_header *ph = from_pgnum(head, lpage + i, sp_bits);
493 /* Initialize this: huge_alloc reads it. */
494 ph->elements_used = 0;
495 add_small_page_to_freelist(head, ph, sp_bits);
501 static u16 get_small_page(struct header *head, unsigned long poolsize,
502 unsigned int sp_bits)
506 ret = pop_from_list(head, &head->small_free_list, sp_bits);
509 ret = get_large_page(head, poolsize, sp_bits);
511 ret = break_up_large_page(head, sp_bits, ret);
515 static bool huge_allocated(struct header *head, unsigned long offset)
518 struct huge_alloc *ha;
520 for (i = head->huge; i; i = ha->next) {
521 ha = (void *)((char *)head + i);
522 if (ha->off <= offset && ha->off + ha->len > offset)
528 /* They want something really big. Aim for contiguous pages (slow). */
529 static COLD void *huge_alloc(void *pool, unsigned long poolsize,
530 unsigned long size, unsigned long align)
532 struct header *head = pool;
533 struct huge_alloc *ha;
534 unsigned long i, sp_bits, lp_bits, num, header_size;
536 sp_bits = small_page_bits(poolsize);
537 lp_bits = sp_bits + BITS_FROM_SMALL_TO_LARGE_PAGE;
539 /* Allocate tracking structure optimistically. */
540 ha = alloc_get(pool, poolsize, sizeof(*ha), ALIGNOF(*ha));
544 /* First search for contiguous small pages... */
545 header_size = sizeof(*head) + sizeof(head->bs) * (max_bucket(lp_bits)-1);
548 for (i = (header_size + (1UL << sp_bits) - 1) >> sp_bits;
549 i << sp_bits < poolsize;
551 struct page_header *pg;
552 unsigned long off = (i << sp_bits);
554 /* Skip over large pages. */
555 if (test_bit(head->pagesize, i >> BITS_FROM_SMALL_TO_LARGE_PAGE)) {
556 i += (1UL << BITS_FROM_SMALL_TO_LARGE_PAGE)-1;
560 /* Does this page meet alignment requirements? */
561 if (!num && off % align != 0)
564 /* FIXME: This makes us O(n^2). */
565 if (huge_allocated(head, off)) {
570 pg = (struct page_header *)((char *)head + off);
571 if (pg->elements_used) {
577 if (num << sp_bits >= size) {
580 /* Remove from free list. */
581 for (pgnum = i; pgnum > i - num; pgnum--) {
582 pg = from_pgnum(head, pgnum, sp_bits);
584 &head->small_free_list,
587 ha->off = (i - num + 1) << sp_bits;
588 ha->len = num << sp_bits;
593 /* Now search for large pages... */
594 recombine_small_pages(head, poolsize, sp_bits);
597 for (i = (header_size + (1UL << lp_bits) - 1) >> lp_bits;
598 (i << lp_bits) < poolsize; i++) {
599 struct page_header *pg;
600 unsigned long off = (i << lp_bits);
602 /* Ignore small pages. */
603 if (!test_bit(head->pagesize, i))
606 /* Does this page meet alignment requirements? */
607 if (!num && off % align != 0)
610 /* FIXME: This makes us O(n^2). */
611 if (huge_allocated(head, off)) {
616 pg = (struct page_header *)((char *)head + off);
617 if (pg->elements_used) {
623 if (num << lp_bits >= size) {
626 /* Remove from free list. */
627 for (pgnum = i; pgnum > i - num; pgnum--) {
628 pg = from_pgnum(head, pgnum, lp_bits);
630 &head->large_free_list,
633 ha->off = (i - num + 1) << lp_bits;
634 ha->len = num << lp_bits;
639 /* Unable to satisfy: free huge alloc structure. */
640 alloc_free(pool, poolsize, ha);
644 add_to_huge_list(pool, ha);
645 return (char *)pool + ha->off;
649 huge_free(struct header *head, unsigned long poolsize, void *free)
651 unsigned long i, off, pgnum, free_off = (char *)free - (char *)head;
652 unsigned int sp_bits, lp_bits;
653 struct huge_alloc *ha;
655 for (i = head->huge; i; i = ha->next) {
656 ha = (void *)((char *)head + i);
657 if (free_off == ha->off)
662 /* Free up all the pages, delete and free ha */
663 sp_bits = small_page_bits(poolsize);
664 lp_bits = sp_bits + BITS_FROM_SMALL_TO_LARGE_PAGE;
665 pgnum = free_off >> sp_bits;
667 if (test_bit(head->pagesize, pgnum >> BITS_FROM_SMALL_TO_LARGE_PAGE)) {
669 off < ha->off + ha->len;
670 off += 1UL << lp_bits) {
671 add_large_page_to_freelist(head,
672 (void *)((char *)head + off),
677 off < ha->off + ha->len;
678 off += 1UL << sp_bits) {
679 add_small_page_to_freelist(head,
680 (void *)((char *)head + off),
684 del_from_huge(head, ha);
685 alloc_free(head, poolsize, ha);
688 static COLD unsigned long huge_size(struct header *head, void *p)
690 unsigned long i, off = (char *)p - (char *)head;
691 struct huge_alloc *ha;
693 for (i = head->huge; i; i = ha->next) {
694 ha = (void *)((char *)head + i);
695 if (off == ha->off) {
702 void *alloc_get(void *pool, unsigned long poolsize,
703 unsigned long size, unsigned long align)
705 struct header *head = pool;
708 struct bucket_state *bs;
709 struct page_header *ph;
710 unsigned int sp_bits;
712 if (poolsize < MIN_USEFUL_SIZE) {
713 return tiny_alloc_get(pool, poolsize, size, align);
716 size = align_up(size, align);
719 bucket = size_to_bucket(size);
721 sp_bits = small_page_bits(poolsize);
723 if (bucket >= max_bucket(sp_bits + BITS_FROM_SMALL_TO_LARGE_PAGE)) {
724 return huge_alloc(pool, poolsize, size, align);
727 bs = &head->bs[bucket];
729 if (!bs->page_list) {
730 struct page_header *ph;
732 if (large_page_bucket(bucket, sp_bits))
733 bs->page_list = get_large_page(head, poolsize,
736 bs->page_list = get_small_page(head, poolsize,
738 /* FIXME: Try large-aligned alloc? Header stuffing? */
739 if (unlikely(!bs->page_list))
741 ph = from_pgnum(head, bs->page_list, sp_bits);
743 ph->elements_used = 0;
745 memset(ph->used, 0, used_size(bs->elements_per_page));
748 ph = from_pgnum(head, bs->page_list, sp_bits);
750 i = find_free_bit(ph->used);
751 set_bit(ph->used, i);
754 /* check if this page is now full */
755 if (unlikely(ph->elements_used == bs->elements_per_page)) {
756 del_from_bucket_list(head, bs, ph, sp_bits);
757 add_to_bucket_full_list(head, bs, ph, sp_bits);
760 return (char *)ph + page_header_size(ph->bucket / INTER_BUCKET_SPACE,
761 bs->elements_per_page)
762 + i * bucket_to_size(bucket);
765 void alloc_free(void *pool, unsigned long poolsize, void *free)
767 struct header *head = pool;
768 struct bucket_state *bs;
769 unsigned int sp_bits;
770 unsigned long i, pgnum, pgoffset, offset = (char *)free - (char *)pool;
772 struct page_header *ph;
774 if (poolsize < MIN_USEFUL_SIZE) {
775 return tiny_alloc_free(pool, poolsize, free);
778 /* Get page header. */
779 sp_bits = small_page_bits(poolsize);
780 pgnum = offset >> sp_bits;
782 /* Big page? Round down further. */
783 if (test_bit(head->pagesize, pgnum >> BITS_FROM_SMALL_TO_LARGE_PAGE)) {
785 pgnum &= ~(SMALL_PAGES_PER_LARGE_PAGE - 1);
789 /* Step back to page header. */
790 ph = from_pgnum(head, pgnum, sp_bits);
791 if ((void *)ph == free) {
792 huge_free(head, poolsize, free);
796 bs = &head->bs[ph->bucket];
797 pgoffset = offset - (pgnum << sp_bits)
798 - page_header_size(ph->bucket / INTER_BUCKET_SPACE,
799 bs->elements_per_page);
801 if (unlikely(ph->elements_used == bs->elements_per_page)) {
802 del_from_bucket_full_list(head, bs, ph, sp_bits);
803 add_to_bucket_list(head, bs, ph, sp_bits);
806 /* Which element are we? */
807 i = pgoffset / bucket_to_size(ph->bucket);
808 clear_bit(ph->used, i);
811 if (unlikely(ph->elements_used == 0)) {
812 bs = &head->bs[ph->bucket];
813 del_from_bucket_list(head, bs, ph, sp_bits);
815 add_small_page_to_freelist(head, ph, sp_bits);
817 add_large_page_to_freelist(head, ph, sp_bits);
821 unsigned long alloc_size(void *pool, unsigned long poolsize, void *p)
823 struct header *head = pool;
824 unsigned int pgnum, sp_bits;
825 unsigned long offset = (char *)p - (char *)pool;
826 struct page_header *ph;
828 if (poolsize < MIN_USEFUL_SIZE)
829 return tiny_alloc_size(pool, poolsize, p);
831 /* Get page header. */
832 sp_bits = small_page_bits(poolsize);
833 pgnum = offset >> sp_bits;
835 /* Big page? Round down further. */
836 if (test_bit(head->pagesize, pgnum >> BITS_FROM_SMALL_TO_LARGE_PAGE))
837 pgnum &= ~(SMALL_PAGES_PER_LARGE_PAGE - 1);
839 /* Step back to page header. */
840 ph = from_pgnum(head, pgnum, sp_bits);
842 return huge_size(head, p);
844 return bucket_to_size(ph->bucket);
847 /* Useful for gdb breakpoints. */
848 static bool check_fail(void)
853 static unsigned long count_bits(const unsigned long bitmap[],
856 unsigned long i, count = 0;
858 while (limit >= BITS_PER_LONG) {
859 count += popcount(bitmap[0]);
861 limit -= BITS_PER_LONG;
864 for (i = 0; i < limit; i++)
865 if (test_bit(bitmap, i))
870 static bool out_of_bounds(unsigned long pgnum,
871 unsigned int sp_bits,
872 unsigned long pagesize,
873 unsigned long poolsize)
875 if (((pgnum << sp_bits) >> sp_bits) != pgnum)
878 if ((pgnum << sp_bits) > poolsize)
881 return ((pgnum << sp_bits) + pagesize > poolsize);
884 static bool check_bucket(struct header *head,
885 unsigned long poolsize,
886 unsigned long pages[],
887 struct bucket_state *bs,
891 struct page_header *ph;
892 unsigned long taken, i, prev, pagesize, sp_bits, lp_bits;
894 sp_bits = small_page_bits(poolsize);
895 lp_bits = sp_bits + BITS_FROM_SMALL_TO_LARGE_PAGE;
897 lp_bucket = large_page_bucket(bindex, sp_bits);
899 pagesize = 1UL << (lp_bucket ? lp_bits : sp_bits);
901 /* This many elements fit? */
902 taken = page_header_size(bindex / INTER_BUCKET_SPACE,
903 bs->elements_per_page);
904 taken += bucket_to_size(bindex) * bs->elements_per_page;
905 if (taken > pagesize)
908 /* One more wouldn't fit? */
909 taken = page_header_size(bindex / INTER_BUCKET_SPACE,
910 bs->elements_per_page + 1);
911 taken += bucket_to_size(bindex) * (bs->elements_per_page + 1);
912 if (taken <= pagesize)
915 /* Walk used list. */
917 for (i = bs->page_list; i; i = ph->next) {
919 if (out_of_bounds(i, sp_bits, pagesize, poolsize))
921 /* Wrong size page? */
922 if (!!test_bit(head->pagesize, i >> BITS_FROM_SMALL_TO_LARGE_PAGE)
925 /* Large page not on boundary? */
926 if (lp_bucket && (i % SMALL_PAGES_PER_LARGE_PAGE) != 0)
928 ph = from_pgnum(head, i, sp_bits);
929 /* Linked list corrupt? */
930 if (ph->prev != prev)
932 /* Already seen this page? */
933 if (test_bit(pages, i))
937 if (ph->elements_used == 0)
939 if (ph->elements_used >= bs->elements_per_page)
941 /* Used bits don't agree? */
942 if (ph->elements_used != count_bits(ph->used,
943 bs->elements_per_page))
946 if (ph->bucket != bindex)
951 /* Walk full list. */
953 for (i = bs->full_list; i; i = ph->next) {
955 if (out_of_bounds(i, sp_bits, pagesize, poolsize))
957 /* Wrong size page? */
958 if (!!test_bit(head->pagesize, i >> BITS_FROM_SMALL_TO_LARGE_PAGE)
960 /* Large page not on boundary? */
961 if (lp_bucket && (i % SMALL_PAGES_PER_LARGE_PAGE) != 0)
963 ph = from_pgnum(head, i, sp_bits);
964 /* Linked list corrupt? */
965 if (ph->prev != prev)
967 /* Already seen this page? */
968 if (test_bit(pages, i))
972 if (ph->elements_used != bs->elements_per_page)
974 /* Used bits don't agree? */
975 if (ph->elements_used != count_bits(ph->used,
976 bs->elements_per_page))
979 if (ph->bucket != bindex)
986 bool alloc_check(void *pool, unsigned long poolsize)
988 struct header *head = pool;
989 unsigned long prev, i, lp_bits, sp_bits, header_size, num_buckets;
990 struct page_header *ph;
991 struct huge_alloc *ha;
992 unsigned long pages[MAX_SMALL_PAGES / BITS_PER_LONG] = { 0 };
994 if (poolsize < MIN_USEFUL_SIZE)
995 return tiny_alloc_check(pool, poolsize);
997 sp_bits = small_page_bits(poolsize);
998 lp_bits = sp_bits + BITS_FROM_SMALL_TO_LARGE_PAGE;
1000 num_buckets = max_bucket(lp_bits);
1002 header_size = sizeof(*head) + sizeof(head->bs) * (num_buckets-1);
1004 /* First, set all bits taken by header. */
1005 for (i = 0; i < header_size; i += (1UL << sp_bits))
1006 set_bit(pages, i >> sp_bits);
1008 /* Check small page free list. */
1010 for (i = head->small_free_list; i; i = ph->next) {
1012 if (out_of_bounds(i, sp_bits, 1UL << sp_bits, poolsize))
1013 return check_fail();
1015 if (test_bit(head->pagesize, i >> BITS_FROM_SMALL_TO_LARGE_PAGE))
1016 return check_fail();
1017 ph = from_pgnum(head, i, sp_bits);
1018 /* Linked list corrupt? */
1019 if (ph->prev != prev)
1020 return check_fail();
1021 /* Already seen this page? */
1022 if (test_bit(pages, i))
1023 return check_fail();
1028 /* Check large page free list. */
1030 for (i = head->large_free_list; i; i = ph->next) {
1032 if (out_of_bounds(i, sp_bits, 1UL << lp_bits, poolsize))
1033 return check_fail();
1034 /* Not large page? */
1035 if (!test_bit(head->pagesize, i >> BITS_FROM_SMALL_TO_LARGE_PAGE))
1036 return check_fail();
1037 /* Not page boundary? */
1038 if ((i % SMALL_PAGES_PER_LARGE_PAGE) != 0)
1039 return check_fail();
1040 ph = from_pgnum(head, i, sp_bits);
1041 /* Linked list corrupt? */
1042 if (ph->prev != prev)
1043 return check_fail();
1044 /* Already seen this page? */
1045 if (test_bit(pages, i))
1046 return check_fail();
1051 /* Check the buckets. */
1052 for (i = 0; i < max_bucket(lp_bits); i++) {
1053 struct bucket_state *bs = &head->bs[i];
1055 if (!check_bucket(head, poolsize, pages, bs, i))
1059 /* Check the huge alloc list. */
1061 for (i = head->huge; i; i = ha->next) {
1062 unsigned long pgbits, j;
1065 if (i >= poolsize || i + sizeof(*ha) > poolsize)
1066 return check_fail();
1067 ha = (void *)((char *)head + i);
1069 /* Check contents of ha. */
1070 if (ha->off > poolsize || ha->off + ha->len > poolsize)
1071 return check_fail();
1073 /* Large or small page? */
1074 pgbits = test_bit(head->pagesize, ha->off >> lp_bits)
1075 ? lp_bits : sp_bits;
1077 /* Not page boundary? */
1078 if ((ha->off % (1UL << pgbits)) != 0)
1079 return check_fail();
1081 /* Not page length? */
1082 if ((ha->len % (1UL << pgbits)) != 0)
1083 return check_fail();
1085 /* Linked list corrupt? */
1086 if (ha->prev != prev)
1087 return check_fail();
1089 for (j = ha->off; j < ha->off + ha->len; j += (1UL<<sp_bits)) {
1090 /* Already seen this page? */
1091 if (test_bit(pages, j >> sp_bits))
1092 return check_fail();
1093 set_bit(pages, j >> sp_bits);
1099 /* Make sure every page accounted for. */
1100 for (i = 0; i < poolsize >> sp_bits; i++) {
1101 if (!test_bit(pages, i))
1102 return check_fail();
1103 if (test_bit(head->pagesize,
1104 i >> BITS_FROM_SMALL_TO_LARGE_PAGE)) {
1105 /* Large page, skip rest. */
1106 i += SMALL_PAGES_PER_LARGE_PAGE - 1;
1113 static unsigned long print_overhead(FILE *out, const char *desc,
1114 unsigned long bytes,
1115 unsigned long poolsize)
1117 fprintf(out, "Overhead (%s): %lu bytes (%.3g%%)\n",
1118 desc, bytes, 100.0 * bytes / poolsize);
1122 static unsigned long count_list(struct header *head,
1124 unsigned int sp_bits,
1125 unsigned long *total_elems)
1127 struct page_header *p;
1128 unsigned long ret = 0;
1131 p = from_pgnum(head, pgnum, sp_bits);
1133 (*total_elems) += p->elements_used;
1140 static unsigned long visualize_bucket(FILE *out, struct header *head,
1141 unsigned int bucket,
1142 unsigned long poolsize,
1143 unsigned int sp_bits)
1145 unsigned long num_full, num_partial, num_pages, page_size,
1146 elems, hdr_min, hdr_size, elems_per_page, overhead = 0;
1148 elems_per_page = head->bs[bucket].elements_per_page;
1150 /* If we used byte-based bitmaps, we could get pg hdr to: */
1151 hdr_min = sizeof(struct page_header)
1152 - sizeof(((struct page_header *)0)->used)
1153 + align_up(elems_per_page, CHAR_BIT) / CHAR_BIT;
1154 hdr_size = page_header_size(bucket / INTER_BUCKET_SPACE,
1158 num_full = count_list(head, head->bs[bucket].full_list, sp_bits,
1160 num_partial = count_list(head, head->bs[bucket].page_list, sp_bits,
1162 num_pages = num_full + num_partial;
1166 fprintf(out, "Bucket %u (%lu bytes):"
1167 " %lu full, %lu partial = %lu elements\n",
1168 bucket, bucket_to_size(bucket), num_full, num_partial, elems);
1169 /* Strict requirement of page header size. */
1170 overhead += print_overhead(out, "page headers",
1171 hdr_min * num_pages, poolsize);
1172 /* Gap between minimal page header and actual start. */
1173 overhead += print_overhead(out, "page post-header alignments",
1174 (hdr_size - hdr_min) * num_pages, poolsize);
1175 /* Between last element and end of page. */
1176 page_size = (1UL << sp_bits);
1177 if (large_page_bucket(bucket, sp_bits))
1178 page_size <<= BITS_FROM_SMALL_TO_LARGE_PAGE;
1180 overhead += print_overhead(out, "page tails",
1181 (page_size - (hdr_size
1183 * bucket_to_size(bucket))))
1184 * num_pages, poolsize);
1188 void alloc_visualize(FILE *out, void *pool, unsigned long poolsize)
1190 struct header *head = pool;
1191 unsigned long i, lp_bits, sp_bits, header_size, num_buckets, count,
1194 fprintf(out, "Pool %p size %lu: (%s allocator)\n", pool, poolsize,
1195 poolsize < MIN_USEFUL_SIZE ? "tiny" : "standard");
1197 if (poolsize < MIN_USEFUL_SIZE) {
1198 tiny_alloc_visualize(out, pool, poolsize);
1202 sp_bits = small_page_bits(poolsize);
1203 lp_bits = sp_bits + BITS_FROM_SMALL_TO_LARGE_PAGE;
1205 num_buckets = max_bucket(lp_bits);
1206 header_size = sizeof(*head) + sizeof(head->bs) * (num_buckets-1);
1208 fprintf(out, "Large page size %lu, small page size %lu.\n",
1209 1UL << lp_bits, 1UL << sp_bits);
1210 overhead += print_overhead(out, "unused pool tail",
1211 poolsize % (1UL << lp_bits), poolsize);
1212 fprintf(out, "Main header %lu bytes (%lu small pages).\n",
1213 header_size, align_up(header_size, 1UL << sp_bits) >> sp_bits);
1214 overhead += print_overhead(out, "partial header page",
1215 align_up(header_size, 1UL << sp_bits)
1216 - header_size, poolsize);
1217 /* Total large pages. */
1218 i = count_bits(head->pagesize, poolsize >> lp_bits);
1220 count = i - count_list(head, head->large_free_list, sp_bits, NULL);
1221 fprintf(out, "%lu/%lu large pages used (%.3g%%)\n",
1222 count, i, count ? 100.0 * count / i : 0.0);
1224 /* Total small pages. */
1225 i = ((poolsize >> lp_bits) - i) << BITS_FROM_SMALL_TO_LARGE_PAGE;
1227 count = i - count_list(head, head->small_free_list, sp_bits, NULL);
1228 fprintf(out, "%lu/%lu small pages used (%.3g%%)\n",
1229 count, i, count ? 100.0 * count / i : 0.0);
1231 /* Summary of each bucket. */
1232 fprintf(out, "%lu buckets:\n", num_buckets);
1233 for (i = 0; i < num_buckets; i++)
1234 overhead += visualize_bucket(out, head, i, poolsize, sp_bits);
1236 print_overhead(out, "total", overhead, poolsize);