2 * Copyright (c) 1998 Matthew Dillon. All Rights Reserved.
3 * Redistribution and use in source and binary forms, with or without
4 * modification, are permitted provided that the following conditions
6 * 1. Redistributions of source code must retain the above copyright
7 * notice, this list of conditions and the following disclaimer.
8 * 2. Redistributions in binary form must reproduce the above copyright
9 * notice, this list of conditions and the following disclaimer in the
10 * documentation and/or other materials provided with the distribution.
11 * 4. Neither the name of the University nor the names of its contributors
12 * may be used to endorse or promote products derived from this software
13 * without specific prior written permission.
15 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS
16 * OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
17 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY
19 * DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE
21 * GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
22 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
23 * WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
24 * NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
25 * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
28 * BLIST.C - Bitmap allocator/deallocator, using a radix tree with hinting
30 * This module implements a general bitmap allocator/deallocator. The
31 * allocator eats around 2 bits per 'block'. The module does not
32 * try to interpret the meaning of a 'block' other than to return
33 * SWAPBLK_NONE on an allocation failure.
35 * A radix tree is used to maintain the bitmap. Two radix constants are
36 * involved: One for the bitmaps contained in the leaf nodes (typically
37 * 32), and one for the meta nodes (typically 16). Both meta and leaf
38 * nodes have a hint field. This field gives us a hint as to the largest
39 * free contiguous range of blocks under the node. It may contain a
40 * value that is too high, but will never contain a value that is too
41 * low. When the radix tree is searched, allocation failures in subtrees
44 * The radix tree also implements two collapsed states for meta nodes:
45 * the ALL-ALLOCATED state and the ALL-FREE state. If a meta node is
46 * in either of these two states, all information contained underneath
47 * the node is considered stale. These states are used to optimize
48 * allocation and freeing operations.
50 * The hinting greatly increases code efficiency for allocations while
51 * the general radix structure optimizes both allocations and frees. The
52 * radix tree should be able to operate well no matter how much
53 * fragmentation there is and no matter how large a bitmap is used.
55 * The blist code wires all necessary memory at creation time. Neither
56 * allocations nor frees require interaction with the memory subsystem.
57 * The non-blocking features of the blist code are used in the swap code
60 * LAYOUT: The radix tree is laid out recursively using a
61 * linear array. Each meta node is immediately followed (laid out
62 * sequentially in memory) by BLIST_META_RADIX lower level nodes. This
63 * is a recursive structure but one that can be easily scanned through
64 * a very simple 'skip' calculation. In order to support large radixes,
65 * portions of the tree may reside outside our memory allocation. We
66 * handle this with an early-termination optimization (when bighint is
67 * set to -1) on the scan. The memory allocation is only large enough
68 * to cover the number of blocks requested at creation time even if it
69 * must be encompassed in larger root-node radix.
71 * NOTE: the allocator cannot currently allocate more than
72 * BLIST_BMAP_RADIX blocks per call. It will panic with 'allocation too
73 * large' if you try. This is an area that could use improvement. The
74 * radix is large enough that this restriction does not effect the swap
75 * system, though. Currently only the allocation code is effected by
76 * this algorithmic unfeature. The freeing code can handle arbitrary
79 * This code can be compiled stand-alone for debugging.
82 #include <sys/cdefs.h>
83 __FBSDID("$FreeBSD$");
87 #include <sys/param.h>
88 #include <sys/systm.h>
90 #include <sys/kernel.h>
91 #include <sys/blist.h>
92 #include <sys/malloc.h>
94 #include <sys/mutex.h>
98 #ifndef BLIST_NO_DEBUG
102 #include <sys/types.h>
103 #include <sys/malloc.h>
109 #define malloc(a,b,c) calloc(a, 1)
110 #define free(a,b) free(a)
112 #include <sys/blist.h>
114 void panic(const char *ctl, ...);
119 * static support functions
122 static daddr_t blst_leaf_alloc(blmeta_t *scan, daddr_t blk, int count);
123 static daddr_t blst_meta_alloc(blmeta_t *scan, daddr_t blk,
124 daddr_t count, daddr_t radix, int skip);
125 static void blst_leaf_free(blmeta_t *scan, daddr_t relblk, int count);
126 static void blst_meta_free(blmeta_t *scan, daddr_t freeBlk, daddr_t count,
127 daddr_t radix, int skip, daddr_t blk);
128 static void blst_copy(blmeta_t *scan, daddr_t blk, daddr_t radix,
129 daddr_t skip, blist_t dest, daddr_t count);
130 static int blst_leaf_fill(blmeta_t *scan, daddr_t blk, int count);
131 static int blst_meta_fill(blmeta_t *scan, daddr_t allocBlk, daddr_t count,
132 daddr_t radix, int skip, daddr_t blk);
133 static daddr_t blst_radix_init(blmeta_t *scan, daddr_t radix,
134 int skip, daddr_t count);
136 static void blst_radix_print(blmeta_t *scan, daddr_t blk,
137 daddr_t radix, int skip, int tab);
141 static MALLOC_DEFINE(M_SWAP, "SWAP", "Swap space");
145 * blist_create() - create a blist capable of handling up to the specified
148 * blocks - must be greater than 0
149 * flags - malloc flags
151 * The smallest blist consists of a single leaf node capable of
152 * managing BLIST_BMAP_RADIX blocks.
156 blist_create(daddr_t blocks, int flags)
163 * Calculate radix and skip field used for scanning.
165 radix = BLIST_BMAP_RADIX;
167 while (radix < blocks) {
168 radix *= BLIST_META_RADIX;
169 skip = (skip + 1) * BLIST_META_RADIX;
172 bl = malloc(sizeof(struct blist), M_SWAP, flags | M_ZERO);
174 bl->bl_blocks = blocks;
175 bl->bl_radix = radix;
177 bl->bl_rootblks = 1 +
178 blst_radix_init(NULL, bl->bl_radix, bl->bl_skip, blocks);
179 bl->bl_root = malloc(sizeof(blmeta_t) * bl->bl_rootblks, M_SWAP, flags);
181 #if defined(BLIST_DEBUG)
183 "BLIST representing %lld blocks (%lld MB of swap)"
184 ", requiring %lldK of ram\n",
185 (long long)bl->bl_blocks,
186 (long long)bl->bl_blocks * 4 / 1024,
187 (long long)(bl->bl_rootblks * sizeof(blmeta_t) + 1023) / 1024
189 printf("BLIST raw radix tree contains %lld records\n",
190 (long long)bl->bl_rootblks);
192 blst_radix_init(bl->bl_root, bl->bl_radix, bl->bl_skip, blocks);
198 blist_destroy(blist_t bl)
200 free(bl->bl_root, M_SWAP);
205 * blist_alloc() - reserve space in the block bitmap. Return the base
206 * of a contiguous region or SWAPBLK_NONE if space could
211 blist_alloc(blist_t bl, daddr_t count)
213 daddr_t blk = SWAPBLK_NONE;
216 if (bl->bl_radix == BLIST_BMAP_RADIX)
217 blk = blst_leaf_alloc(bl->bl_root, 0, count);
219 blk = blst_meta_alloc(bl->bl_root, 0, count, bl->bl_radix, bl->bl_skip);
220 if (blk != SWAPBLK_NONE)
221 bl->bl_free -= count;
227 * blist_free() - free up space in the block bitmap. Return the base
228 * of a contiguous region. Panic if an inconsistancy is
233 blist_free(blist_t bl, daddr_t blkno, daddr_t count)
236 if (bl->bl_radix == BLIST_BMAP_RADIX)
237 blst_leaf_free(bl->bl_root, blkno, count);
239 blst_meta_free(bl->bl_root, blkno, count, bl->bl_radix, bl->bl_skip, 0);
240 bl->bl_free += count;
245 * blist_fill() - mark a region in the block bitmap as off-limits
246 * to the allocator (i.e. allocate it), ignoring any
247 * existing allocations. Return the number of blocks
248 * actually filled that were free before the call.
252 blist_fill(blist_t bl, daddr_t blkno, daddr_t count)
257 if (bl->bl_radix == BLIST_BMAP_RADIX)
258 filled = blst_leaf_fill(bl->bl_root, blkno, count);
260 filled = blst_meta_fill(bl->bl_root, blkno, count,
261 bl->bl_radix, bl->bl_skip, 0);
262 bl->bl_free -= filled;
269 * blist_resize() - resize an existing radix tree to handle the
270 * specified number of blocks. This will reallocate
271 * the tree and transfer the previous bitmap to the new
272 * one. When extending the tree you can specify whether
273 * the new blocks are to left allocated or freed.
277 blist_resize(blist_t *pbl, daddr_t count, int freenew, int flags)
279 blist_t newbl = blist_create(count, flags);
283 if (count > save->bl_blocks)
284 count = save->bl_blocks;
285 blst_copy(save->bl_root, 0, save->bl_radix, save->bl_skip, newbl, count);
288 * If resizing upwards, should we free the new space or not?
290 if (freenew && count < newbl->bl_blocks) {
291 blist_free(newbl, count, newbl->bl_blocks - count);
299 * blist_print() - dump radix tree
303 blist_print(blist_t bl)
306 blst_radix_print(bl->bl_root, 0, bl->bl_radix, bl->bl_skip, 4);
312 /************************************************************************
313 * ALLOCATION SUPPORT FUNCTIONS *
314 ************************************************************************
316 * These support functions do all the actual work. They may seem
317 * rather longish, but that's because I've commented them up. The
318 * actual code is straight forward.
323 * blist_leaf_alloc() - allocate at a leaf in the radix tree (a bitmap).
325 * This is the core of the allocator and is optimized for the 1 block
326 * and the BLIST_BMAP_RADIX block allocation cases. Other cases are
327 * somewhat slower. The 1 block allocation case is log2 and extremely
337 u_daddr_t orig = scan->u.bmu_bitmap;
341 * Optimize bitmap all-allocated case. Also, count = 1
342 * case assumes at least 1 bit is free in the bitmap, so
343 * we have to take care of this case here.
345 scan->bm_bighint = 0;
346 return(SWAPBLK_NONE);
350 * Optimized code to allocate one bit out of the bitmap
353 int j = BLIST_BMAP_RADIX/2;
356 mask = (u_daddr_t)-1 >> (BLIST_BMAP_RADIX/2);
359 if ((orig & mask) == 0) {
366 scan->u.bmu_bitmap &= ~((u_daddr_t)1 << r);
369 if (count <= BLIST_BMAP_RADIX) {
371 * non-optimized code to allocate N bits out of the bitmap.
372 * The more bits, the faster the code runs. It will run
373 * the slowest allocating 2 bits, but since there aren't any
374 * memory ops in the core loop (or shouldn't be, anyway),
375 * you probably won't notice the difference.
378 int n = BLIST_BMAP_RADIX - count;
381 mask = (u_daddr_t)-1 >> n;
383 for (j = 0; j <= n; ++j) {
384 if ((orig & mask) == mask) {
385 scan->u.bmu_bitmap &= ~mask;
392 * We couldn't allocate count in this subtree, update bighint.
394 scan->bm_bighint = count - 1;
395 return(SWAPBLK_NONE);
399 * blist_meta_alloc() - allocate at a meta in the radix tree.
401 * Attempt to allocate at a meta node. If we can't, we update
402 * bighint and return a failure. Updating bighint optimize future
403 * calls that hit this node. We have to check for our collapse cases
404 * and we have a few optimizations strewn in as well.
416 int next_skip = ((u_int)skip / BLIST_META_RADIX);
418 if (scan->u.bmu_avail == 0) {
420 * ALL-ALLOCATED special case
422 scan->bm_bighint = count;
423 return(SWAPBLK_NONE);
426 if (scan->u.bmu_avail == radix) {
427 radix /= BLIST_META_RADIX;
430 * ALL-FREE special case, initialize uninitialize
433 for (i = 1; i <= skip; i += next_skip) {
434 if (scan[i].bm_bighint == (daddr_t)-1)
436 if (next_skip == 1) {
437 scan[i].u.bmu_bitmap = (u_daddr_t)-1;
438 scan[i].bm_bighint = BLIST_BMAP_RADIX;
440 scan[i].bm_bighint = radix;
441 scan[i].u.bmu_avail = radix;
445 radix /= BLIST_META_RADIX;
448 for (i = 1; i <= skip; i += next_skip) {
449 if (count <= scan[i].bm_bighint) {
451 * count fits in object
454 if (next_skip == 1) {
455 r = blst_leaf_alloc(&scan[i], blk, count);
457 r = blst_meta_alloc(&scan[i], blk, count, radix, next_skip - 1);
459 if (r != SWAPBLK_NONE) {
460 scan->u.bmu_avail -= count;
461 if (scan->bm_bighint > scan->u.bmu_avail)
462 scan->bm_bighint = scan->u.bmu_avail;
465 } else if (scan[i].bm_bighint == (daddr_t)-1) {
470 } else if (count > radix) {
472 * count does not fit in object even if it were
475 panic("blist_meta_alloc: allocation too large");
481 * We couldn't allocate count in this subtree, update bighint.
483 if (scan->bm_bighint >= count)
484 scan->bm_bighint = count - 1;
485 return(SWAPBLK_NONE);
489 * BLST_LEAF_FREE() - free allocated block from leaf bitmap
500 * free some data in this bitmap
503 * 0000111111111110000
507 int n = blk & (BLIST_BMAP_RADIX - 1);
510 mask = ((u_daddr_t)-1 << n) &
511 ((u_daddr_t)-1 >> (BLIST_BMAP_RADIX - count - n));
513 if (scan->u.bmu_bitmap & mask)
514 panic("blst_radix_free: freeing free block");
515 scan->u.bmu_bitmap |= mask;
518 * We could probably do a better job here. We are required to make
519 * bighint at least as large as the biggest contiguous block of
520 * data. If we just shoehorn it, a little extra overhead will
521 * be incured on the next allocation (but only that one typically).
523 scan->bm_bighint = BLIST_BMAP_RADIX;
527 * BLST_META_FREE() - free allocated blocks from radix tree meta info
529 * This support routine frees a range of blocks from the bitmap.
530 * The range must be entirely enclosed by this radix node. If a
531 * meta node, we break the range down recursively to free blocks
532 * in subnodes (which means that this code can free an arbitrary
533 * range whereas the allocation code cannot allocate an arbitrary
547 int next_skip = ((u_int)skip / BLIST_META_RADIX);
550 printf("free (%llx,%lld) FROM (%llx,%lld)\n",
551 (long long)freeBlk, (long long)count,
552 (long long)blk, (long long)radix
556 if (scan->u.bmu_avail == 0) {
558 * ALL-ALLOCATED special case, with possible
559 * shortcut to ALL-FREE special case.
561 scan->u.bmu_avail = count;
562 scan->bm_bighint = count;
564 if (count != radix) {
565 for (i = 1; i <= skip; i += next_skip) {
566 if (scan[i].bm_bighint == (daddr_t)-1)
568 scan[i].bm_bighint = 0;
569 if (next_skip == 1) {
570 scan[i].u.bmu_bitmap = 0;
572 scan[i].u.bmu_avail = 0;
578 scan->u.bmu_avail += count;
579 /* scan->bm_bighint = radix; */
583 * ALL-FREE special case.
586 if (scan->u.bmu_avail == radix)
588 if (scan->u.bmu_avail > radix)
589 panic("blst_meta_free: freeing already free blocks (%lld) %lld/%lld",
590 (long long)count, (long long)scan->u.bmu_avail,
594 * Break the free down into its components
597 radix /= BLIST_META_RADIX;
599 i = (freeBlk - blk) / radix;
601 i = i * next_skip + 1;
603 while (i <= skip && blk < freeBlk + count) {
606 v = blk + radix - freeBlk;
610 if (scan->bm_bighint == (daddr_t)-1)
611 panic("blst_meta_free: freeing unexpected range");
613 if (next_skip == 1) {
614 blst_leaf_free(&scan[i], freeBlk, v);
616 blst_meta_free(&scan[i], freeBlk, v, radix, next_skip - 1, blk);
618 if (scan->bm_bighint < scan[i].bm_bighint)
619 scan->bm_bighint = scan[i].bm_bighint;
628 * BLIST_RADIX_COPY() - copy one radix tree to another
630 * Locates free space in the source tree and frees it in the destination
631 * tree. The space may not already be free in the destination.
634 static void blst_copy(
649 if (radix == BLIST_BMAP_RADIX) {
650 u_daddr_t v = scan->u.bmu_bitmap;
652 if (v == (u_daddr_t)-1) {
653 blist_free(dest, blk, count);
657 for (i = 0; i < BLIST_BMAP_RADIX && i < count; ++i) {
658 if (v & ((u_daddr_t)1 << i))
659 blist_free(dest, blk + i, 1);
669 if (scan->u.bmu_avail == 0) {
671 * Source all allocated, leave dest allocated
675 if (scan->u.bmu_avail == radix) {
677 * Source all free, free entire dest
680 blist_free(dest, blk, count);
682 blist_free(dest, blk, radix);
687 radix /= BLIST_META_RADIX;
688 next_skip = ((u_int)skip / BLIST_META_RADIX);
690 for (i = 1; count && i <= skip; i += next_skip) {
691 if (scan[i].bm_bighint == (daddr_t)-1)
694 if (count >= radix) {
722 * BLST_LEAF_FILL() - allocate specific blocks in leaf bitmap
724 * This routine allocates all blocks in the specified range
725 * regardless of any existing allocations in that range. Returns
726 * the number of blocks allocated by the call.
730 blst_leaf_fill(blmeta_t *scan, daddr_t blk, int count)
732 int n = blk & (BLIST_BMAP_RADIX - 1);
734 u_daddr_t mask, bitmap;
736 mask = ((u_daddr_t)-1 << n) &
737 ((u_daddr_t)-1 >> (BLIST_BMAP_RADIX - count - n));
739 /* Count the number of blocks we're about to allocate */
740 bitmap = scan->u.bmu_bitmap & mask;
741 for (nblks = 0; bitmap != 0; nblks++)
742 bitmap &= bitmap - 1;
744 scan->u.bmu_bitmap &= ~mask;
749 * BLIST_META_FILL() - allocate specific blocks at a meta node
751 * This routine allocates the specified range of blocks,
752 * regardless of any existing allocations in the range. The
753 * range must be within the extent of this node. Returns the
754 * number of blocks allocated by the call.
766 int next_skip = ((u_int)skip / BLIST_META_RADIX);
770 panic("blist_meta_fill: allocation too large");
771 if (count == radix || scan->u.bmu_avail == 0) {
773 * ALL-ALLOCATED special case
775 nblks = scan->u.bmu_avail;
776 scan->u.bmu_avail = 0;
777 scan->bm_bighint = count;
781 if (scan->u.bmu_avail == radix) {
782 radix /= BLIST_META_RADIX;
785 * ALL-FREE special case, initialize sublevel
787 for (i = 1; i <= skip; i += next_skip) {
788 if (scan[i].bm_bighint == (daddr_t)-1)
790 if (next_skip == 1) {
791 scan[i].u.bmu_bitmap = (u_daddr_t)-1;
792 scan[i].bm_bighint = BLIST_BMAP_RADIX;
794 scan[i].bm_bighint = radix;
795 scan[i].u.bmu_avail = radix;
799 radix /= BLIST_META_RADIX;
802 i = (allocBlk - blk) / radix;
804 i = i * next_skip + 1;
806 while (i <= skip && blk < allocBlk + count) {
809 v = blk + radix - allocBlk;
813 if (scan->bm_bighint == (daddr_t)-1)
814 panic("blst_meta_fill: filling unexpected range");
816 if (next_skip == 1) {
817 nblks += blst_leaf_fill(&scan[i], allocBlk, v);
819 nblks += blst_meta_fill(&scan[i], allocBlk, v,
820 radix, next_skip - 1, blk);
827 scan->u.bmu_avail -= nblks;
832 * BLST_RADIX_INIT() - initialize radix tree
834 * Initialize our meta structures and bitmaps and calculate the exact
835 * amount of space required to manage 'count' blocks - this space may
836 * be considerably less than the calculated radix due to the large
837 * RADIX values we use.
841 blst_radix_init(blmeta_t *scan, daddr_t radix, int skip, daddr_t count)
845 daddr_t memindex = 0;
851 if (radix == BLIST_BMAP_RADIX) {
853 scan->bm_bighint = 0;
854 scan->u.bmu_bitmap = 0;
860 * Meta node. If allocating the entire object we can special
861 * case it. However, we need to figure out how much memory
862 * is required to manage 'count' blocks, so we continue on anyway.
866 scan->bm_bighint = 0;
867 scan->u.bmu_avail = 0;
870 radix /= BLIST_META_RADIX;
871 next_skip = ((u_int)skip / BLIST_META_RADIX);
873 for (i = 1; i <= skip; i += next_skip) {
874 if (count >= radix) {
876 * Allocate the entire object
878 memindex = i + blst_radix_init(
879 ((scan) ? &scan[i] : NULL),
885 } else if (count > 0) {
887 * Allocate a partial object
889 memindex = i + blst_radix_init(
890 ((scan) ? &scan[i] : NULL),
898 * Add terminator and break out
901 scan[i].bm_bighint = (daddr_t)-1;
913 blst_radix_print(blmeta_t *scan, daddr_t blk, daddr_t radix, int skip, int tab)
919 if (radix == BLIST_BMAP_RADIX) {
921 "%*.*s(%08llx,%lld): bitmap %016llx big=%lld\n",
923 (long long)blk, (long long)radix,
924 (long long)scan->u.bmu_bitmap,
925 (long long)scan->bm_bighint
930 if (scan->u.bmu_avail == 0) {
932 "%*.*s(%08llx,%lld) ALL ALLOCATED\n",
939 if (scan->u.bmu_avail == radix) {
941 "%*.*s(%08llx,%lld) ALL FREE\n",
950 "%*.*s(%08llx,%lld): subtree (%lld/%lld) big=%lld {\n",
952 (long long)blk, (long long)radix,
953 (long long)scan->u.bmu_avail,
955 (long long)scan->bm_bighint
958 radix /= BLIST_META_RADIX;
959 next_skip = ((u_int)skip / BLIST_META_RADIX);
962 for (i = 1; i <= skip; i += next_skip) {
963 if (scan[i].bm_bighint == (daddr_t)-1) {
965 "%*.*s(%08llx,%lld): Terminator\n",
967 (long long)blk, (long long)radix
994 main(int ac, char **av)
1000 for (i = 1; i < ac; ++i) {
1001 const char *ptr = av[i];
1003 size = strtol(ptr, NULL, 0);
1007 fprintf(stderr, "Bad option: %s\n", ptr - 2);
1010 bl = blist_create(size, M_WAITOK);
1011 blist_free(bl, 0, size);
1016 long long count = 0;
1018 printf("%lld/%lld/%lld> ", (long long)bl->bl_free,
1019 (long long)size, (long long)bl->bl_radix);
1021 if (fgets(buf, sizeof(buf), stdin) == NULL)
1025 if (sscanf(buf + 1, "%lld", &count) == 1) {
1026 blist_resize(&bl, count, 1, M_WAITOK);
1034 if (sscanf(buf + 1, "%lld", &count) == 1) {
1035 daddr_t blk = blist_alloc(bl, count);
1036 printf(" R=%08llx\n", (long long)blk);
1042 if (sscanf(buf + 1, "%llx %lld", &da, &count) == 2) {
1043 blist_free(bl, da, count);
1049 if (sscanf(buf + 1, "%llx %lld", &da, &count) == 2) {
1051 blist_fill(bl, da, count));
1076 panic(const char *ctl, ...)
1081 vfprintf(stderr, ctl, va);
1082 fprintf(stderr, "\n");