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 daddr_t blst_leaf_fill(blmeta_t *scan, daddr_t blk, int count);
131 static daddr_t 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)
159 daddr_t nodes, radix;
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);
176 bl->bl_blocks = blocks;
177 bl->bl_radix = radix;
179 nodes = 1 + blst_radix_init(NULL, radix, bl->bl_skip, blocks);
180 bl->bl_root = malloc(nodes * sizeof(blmeta_t), M_SWAP, flags);
181 if (bl->bl_root == NULL) {
185 blst_radix_init(bl->bl_root, radix, bl->bl_skip, blocks);
187 #if defined(BLIST_DEBUG)
189 "BLIST representing %lld blocks (%lld MB of swap)"
190 ", requiring %lldK of ram\n",
191 (long long)bl->bl_blocks,
192 (long long)bl->bl_blocks * 4 / 1024,
193 (long long)(nodes * sizeof(blmeta_t) + 1023) / 1024
195 printf("BLIST raw radix tree contains %lld records\n",
203 blist_destroy(blist_t bl)
205 free(bl->bl_root, M_SWAP);
210 * blist_alloc() - reserve space in the block bitmap. Return the base
211 * of a contiguous region or SWAPBLK_NONE if space could
216 blist_alloc(blist_t bl, daddr_t count)
218 daddr_t blk = SWAPBLK_NONE;
221 if (bl->bl_radix == BLIST_BMAP_RADIX)
222 blk = blst_leaf_alloc(bl->bl_root, 0, count);
224 blk = blst_meta_alloc(bl->bl_root, 0, count, bl->bl_radix, bl->bl_skip);
225 if (blk != SWAPBLK_NONE)
226 bl->bl_free -= count;
232 * blist_free() - free up space in the block bitmap. Return the base
233 * of a contiguous region. Panic if an inconsistancy is
238 blist_free(blist_t bl, daddr_t blkno, daddr_t count)
241 if (bl->bl_radix == BLIST_BMAP_RADIX)
242 blst_leaf_free(bl->bl_root, blkno, count);
244 blst_meta_free(bl->bl_root, blkno, count, bl->bl_radix, bl->bl_skip, 0);
245 bl->bl_free += count;
250 * blist_fill() - mark a region in the block bitmap as off-limits
251 * to the allocator (i.e. allocate it), ignoring any
252 * existing allocations. Return the number of blocks
253 * actually filled that were free before the call.
257 blist_fill(blist_t bl, daddr_t blkno, daddr_t count)
262 if (bl->bl_radix == BLIST_BMAP_RADIX)
263 filled = blst_leaf_fill(bl->bl_root, blkno, count);
265 filled = blst_meta_fill(bl->bl_root, blkno, count,
266 bl->bl_radix, bl->bl_skip, 0);
267 bl->bl_free -= filled;
274 * blist_resize() - resize an existing radix tree to handle the
275 * specified number of blocks. This will reallocate
276 * the tree and transfer the previous bitmap to the new
277 * one. When extending the tree you can specify whether
278 * the new blocks are to left allocated or freed.
282 blist_resize(blist_t *pbl, daddr_t count, int freenew, int flags)
284 blist_t newbl = blist_create(count, flags);
288 if (count > save->bl_blocks)
289 count = save->bl_blocks;
290 blst_copy(save->bl_root, 0, save->bl_radix, save->bl_skip, newbl, count);
293 * If resizing upwards, should we free the new space or not?
295 if (freenew && count < newbl->bl_blocks) {
296 blist_free(newbl, count, newbl->bl_blocks - count);
304 * blist_print() - dump radix tree
308 blist_print(blist_t bl)
311 blst_radix_print(bl->bl_root, 0, bl->bl_radix, bl->bl_skip, 4);
317 /************************************************************************
318 * ALLOCATION SUPPORT FUNCTIONS *
319 ************************************************************************
321 * These support functions do all the actual work. They may seem
322 * rather longish, but that's because I've commented them up. The
323 * actual code is straight forward.
328 * blist_leaf_alloc() - allocate at a leaf in the radix tree (a bitmap).
330 * This is the core of the allocator and is optimized for the 1 block
331 * and the BLIST_BMAP_RADIX block allocation cases. Other cases are
332 * somewhat slower. The 1 block allocation case is log2 and extremely
342 u_daddr_t orig = scan->u.bmu_bitmap;
346 * Optimize bitmap all-allocated case. Also, count = 1
347 * case assumes at least 1 bit is free in the bitmap, so
348 * we have to take care of this case here.
350 scan->bm_bighint = 0;
351 return(SWAPBLK_NONE);
355 * Optimized code to allocate one bit out of the bitmap
358 int j = BLIST_BMAP_RADIX/2;
361 mask = (u_daddr_t)-1 >> (BLIST_BMAP_RADIX/2);
364 if ((orig & mask) == 0) {
371 scan->u.bmu_bitmap &= ~((u_daddr_t)1 << r);
374 if (count <= BLIST_BMAP_RADIX) {
376 * non-optimized code to allocate N bits out of the bitmap.
377 * The more bits, the faster the code runs. It will run
378 * the slowest allocating 2 bits, but since there aren't any
379 * memory ops in the core loop (or shouldn't be, anyway),
380 * you probably won't notice the difference.
383 int n = BLIST_BMAP_RADIX - count;
386 mask = (u_daddr_t)-1 >> n;
388 for (j = 0; j <= n; ++j) {
389 if ((orig & mask) == mask) {
390 scan->u.bmu_bitmap &= ~mask;
397 * We couldn't allocate count in this subtree, update bighint.
399 scan->bm_bighint = count - 1;
400 return(SWAPBLK_NONE);
404 * blist_meta_alloc() - allocate at a meta in the radix tree.
406 * Attempt to allocate at a meta node. If we can't, we update
407 * bighint and return a failure. Updating bighint optimize future
408 * calls that hit this node. We have to check for our collapse cases
409 * and we have a few optimizations strewn in as well.
421 int next_skip = ((u_int)skip / BLIST_META_RADIX);
423 if (scan->u.bmu_avail == 0) {
425 * ALL-ALLOCATED special case
427 scan->bm_bighint = 0;
428 return(SWAPBLK_NONE);
431 if (scan->u.bmu_avail == radix) {
432 radix /= BLIST_META_RADIX;
435 * ALL-FREE special case, initialize uninitialize
438 for (i = 1; i <= skip; i += next_skip) {
439 if (scan[i].bm_bighint == (daddr_t)-1)
441 if (next_skip == 1) {
442 scan[i].u.bmu_bitmap = (u_daddr_t)-1;
443 scan[i].bm_bighint = BLIST_BMAP_RADIX;
445 scan[i].bm_bighint = radix;
446 scan[i].u.bmu_avail = radix;
450 radix /= BLIST_META_RADIX;
453 for (i = 1; i <= skip; i += next_skip) {
454 if (count <= scan[i].bm_bighint) {
456 * count fits in object
459 if (next_skip == 1) {
460 r = blst_leaf_alloc(&scan[i], blk, count);
462 r = blst_meta_alloc(&scan[i], blk, count, radix, next_skip - 1);
464 if (r != SWAPBLK_NONE) {
465 scan->u.bmu_avail -= count;
466 if (scan->bm_bighint > scan->u.bmu_avail)
467 scan->bm_bighint = scan->u.bmu_avail;
470 } else if (scan[i].bm_bighint == (daddr_t)-1) {
475 } else if (count > radix) {
477 * count does not fit in object even if it were
480 panic("blist_meta_alloc: allocation too large");
486 * We couldn't allocate count in this subtree, update bighint.
488 if (scan->bm_bighint >= count)
489 scan->bm_bighint = count - 1;
490 return(SWAPBLK_NONE);
494 * BLST_LEAF_FREE() - free allocated block from leaf bitmap
505 * free some data in this bitmap
508 * 0000111111111110000
512 int n = blk & (BLIST_BMAP_RADIX - 1);
515 mask = ((u_daddr_t)-1 << n) &
516 ((u_daddr_t)-1 >> (BLIST_BMAP_RADIX - count - n));
518 if (scan->u.bmu_bitmap & mask)
519 panic("blst_radix_free: freeing free block");
520 scan->u.bmu_bitmap |= mask;
523 * We could probably do a better job here. We are required to make
524 * bighint at least as large as the biggest contiguous block of
525 * data. If we just shoehorn it, a little extra overhead will
526 * be incured on the next allocation (but only that one typically).
528 scan->bm_bighint = BLIST_BMAP_RADIX;
532 * BLST_META_FREE() - free allocated blocks from radix tree meta info
534 * This support routine frees a range of blocks from the bitmap.
535 * The range must be entirely enclosed by this radix node. If a
536 * meta node, we break the range down recursively to free blocks
537 * in subnodes (which means that this code can free an arbitrary
538 * range whereas the allocation code cannot allocate an arbitrary
552 int next_skip = ((u_int)skip / BLIST_META_RADIX);
555 printf("free (%llx,%lld) FROM (%llx,%lld)\n",
556 (long long)freeBlk, (long long)count,
557 (long long)blk, (long long)radix
561 if (scan->u.bmu_avail == 0) {
563 * ALL-ALLOCATED special case, with possible
564 * shortcut to ALL-FREE special case.
566 scan->u.bmu_avail = count;
567 scan->bm_bighint = count;
569 if (count != radix) {
570 for (i = 1; i <= skip; i += next_skip) {
571 if (scan[i].bm_bighint == (daddr_t)-1)
573 scan[i].bm_bighint = 0;
574 if (next_skip == 1) {
575 scan[i].u.bmu_bitmap = 0;
577 scan[i].u.bmu_avail = 0;
583 scan->u.bmu_avail += count;
584 /* scan->bm_bighint = radix; */
588 * ALL-FREE special case.
591 if (scan->u.bmu_avail == radix)
593 if (scan->u.bmu_avail > radix)
594 panic("blst_meta_free: freeing already free blocks (%lld) %lld/%lld",
595 (long long)count, (long long)scan->u.bmu_avail,
599 * Break the free down into its components
602 radix /= BLIST_META_RADIX;
604 i = (freeBlk - blk) / radix;
606 i = i * next_skip + 1;
608 while (i <= skip && blk < freeBlk + count) {
611 v = blk + radix - freeBlk;
615 if (scan->bm_bighint == (daddr_t)-1)
616 panic("blst_meta_free: freeing unexpected range");
618 if (next_skip == 1) {
619 blst_leaf_free(&scan[i], freeBlk, v);
621 blst_meta_free(&scan[i], freeBlk, v, radix, next_skip - 1, blk);
623 if (scan->bm_bighint < scan[i].bm_bighint)
624 scan->bm_bighint = scan[i].bm_bighint;
633 * BLIST_RADIX_COPY() - copy one radix tree to another
635 * Locates free space in the source tree and frees it in the destination
636 * tree. The space may not already be free in the destination.
639 static void blst_copy(
654 if (radix == BLIST_BMAP_RADIX) {
655 u_daddr_t v = scan->u.bmu_bitmap;
657 if (v == (u_daddr_t)-1) {
658 blist_free(dest, blk, count);
662 for (i = 0; i < BLIST_BMAP_RADIX && i < count; ++i) {
663 if (v & ((u_daddr_t)1 << i))
664 blist_free(dest, blk + i, 1);
674 if (scan->u.bmu_avail == 0) {
676 * Source all allocated, leave dest allocated
680 if (scan->u.bmu_avail == radix) {
682 * Source all free, free entire dest
685 blist_free(dest, blk, count);
687 blist_free(dest, blk, radix);
692 radix /= BLIST_META_RADIX;
693 next_skip = ((u_int)skip / BLIST_META_RADIX);
695 for (i = 1; count && i <= skip; i += next_skip) {
696 if (scan[i].bm_bighint == (daddr_t)-1)
699 if (count >= radix) {
727 * BLST_LEAF_FILL() - allocate specific blocks in leaf bitmap
729 * This routine allocates all blocks in the specified range
730 * regardless of any existing allocations in that range. Returns
731 * the number of blocks allocated by the call.
735 blst_leaf_fill(blmeta_t *scan, daddr_t blk, int count)
737 int n = blk & (BLIST_BMAP_RADIX - 1);
739 u_daddr_t mask, bitmap;
741 mask = ((u_daddr_t)-1 << n) &
742 ((u_daddr_t)-1 >> (BLIST_BMAP_RADIX - count - n));
744 /* Count the number of blocks we're about to allocate */
745 bitmap = scan->u.bmu_bitmap & mask;
746 for (nblks = 0; bitmap != 0; nblks++)
747 bitmap &= bitmap - 1;
749 scan->u.bmu_bitmap &= ~mask;
754 * BLIST_META_FILL() - allocate specific blocks at a meta node
756 * This routine allocates the specified range of blocks,
757 * regardless of any existing allocations in the range. The
758 * range must be within the extent of this node. Returns the
759 * number of blocks allocated by the call.
771 int next_skip = ((u_int)skip / BLIST_META_RADIX);
775 panic("blist_meta_fill: allocation too large");
776 if (count == radix || scan->u.bmu_avail == 0) {
778 * ALL-ALLOCATED special case
780 nblks = scan->u.bmu_avail;
781 scan->u.bmu_avail = 0;
782 scan->bm_bighint = 0;
786 if (scan->u.bmu_avail == radix) {
787 radix /= BLIST_META_RADIX;
790 * ALL-FREE special case, initialize sublevel
792 for (i = 1; i <= skip; i += next_skip) {
793 if (scan[i].bm_bighint == (daddr_t)-1)
795 if (next_skip == 1) {
796 scan[i].u.bmu_bitmap = (u_daddr_t)-1;
797 scan[i].bm_bighint = BLIST_BMAP_RADIX;
799 scan[i].bm_bighint = radix;
800 scan[i].u.bmu_avail = radix;
804 radix /= BLIST_META_RADIX;
807 i = (allocBlk - blk) / radix;
809 i = i * next_skip + 1;
811 while (i <= skip && blk < allocBlk + count) {
814 v = blk + radix - allocBlk;
818 if (scan->bm_bighint == (daddr_t)-1)
819 panic("blst_meta_fill: filling unexpected range");
821 if (next_skip == 1) {
822 nblks += blst_leaf_fill(&scan[i], allocBlk, v);
824 nblks += blst_meta_fill(&scan[i], allocBlk, v,
825 radix, next_skip - 1, blk);
832 scan->u.bmu_avail -= nblks;
837 * BLST_RADIX_INIT() - initialize radix tree
839 * Initialize our meta structures and bitmaps and calculate the exact
840 * amount of space required to manage 'count' blocks - this space may
841 * be considerably less than the calculated radix due to the large
842 * RADIX values we use.
846 blst_radix_init(blmeta_t *scan, daddr_t radix, int skip, daddr_t count)
850 daddr_t memindex = 0;
856 if (radix == BLIST_BMAP_RADIX) {
858 scan->bm_bighint = 0;
859 scan->u.bmu_bitmap = 0;
865 * Meta node. If allocating the entire object we can special
866 * case it. However, we need to figure out how much memory
867 * is required to manage 'count' blocks, so we continue on anyway.
871 scan->bm_bighint = 0;
872 scan->u.bmu_avail = 0;
875 radix /= BLIST_META_RADIX;
876 next_skip = ((u_int)skip / BLIST_META_RADIX);
878 for (i = 1; i <= skip; i += next_skip) {
879 if (count >= radix) {
881 * Allocate the entire object
883 memindex = i + blst_radix_init(
884 ((scan) ? &scan[i] : NULL),
890 } else if (count > 0) {
892 * Allocate a partial object
894 memindex = i + blst_radix_init(
895 ((scan) ? &scan[i] : NULL),
903 * Add terminator and break out
906 scan[i].bm_bighint = (daddr_t)-1;
918 blst_radix_print(blmeta_t *scan, daddr_t blk, daddr_t radix, int skip, int tab)
924 if (radix == BLIST_BMAP_RADIX) {
926 "%*.*s(%08llx,%lld): bitmap %016llx big=%lld\n",
928 (long long)blk, (long long)radix,
929 (long long)scan->u.bmu_bitmap,
930 (long long)scan->bm_bighint
935 if (scan->u.bmu_avail == 0) {
937 "%*.*s(%08llx,%lld) ALL ALLOCATED\n",
944 if (scan->u.bmu_avail == radix) {
946 "%*.*s(%08llx,%lld) ALL FREE\n",
955 "%*.*s(%08llx,%lld): subtree (%lld/%lld) big=%lld {\n",
957 (long long)blk, (long long)radix,
958 (long long)scan->u.bmu_avail,
960 (long long)scan->bm_bighint
963 radix /= BLIST_META_RADIX;
964 next_skip = ((u_int)skip / BLIST_META_RADIX);
967 for (i = 1; i <= skip; i += next_skip) {
968 if (scan[i].bm_bighint == (daddr_t)-1) {
970 "%*.*s(%08llx,%lld): Terminator\n",
972 (long long)blk, (long long)radix
999 main(int ac, char **av)
1005 for (i = 1; i < ac; ++i) {
1006 const char *ptr = av[i];
1008 size = strtol(ptr, NULL, 0);
1012 fprintf(stderr, "Bad option: %s\n", ptr - 2);
1015 bl = blist_create(size, M_WAITOK);
1016 blist_free(bl, 0, size);
1021 long long count = 0;
1023 printf("%lld/%lld/%lld> ", (long long)bl->bl_free,
1024 (long long)size, (long long)bl->bl_radix);
1026 if (fgets(buf, sizeof(buf), stdin) == NULL)
1030 if (sscanf(buf + 1, "%lld", &count) == 1) {
1031 blist_resize(&bl, count, 1, M_WAITOK);
1039 if (sscanf(buf + 1, "%lld", &count) == 1) {
1040 daddr_t blk = blist_alloc(bl, count);
1041 printf(" R=%08llx\n", (long long)blk);
1047 if (sscanf(buf + 1, "%llx %lld", &da, &count) == 2) {
1048 blist_free(bl, da, count);
1054 if (sscanf(buf + 1, "%llx %lld", &da, &count) == 2) {
1056 (intmax_t)blist_fill(bl, da, count));
1081 panic(const char *ctl, ...)
1086 vfprintf(stderr, ctl, va);
1087 fprintf(stderr, "\n");