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 layed out recursively using a
61 * linear array. Each meta node is immediately followed (layed 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 #define SWAPBLK_NONE ((daddr_t)-1)
104 #include <sys/types.h>
110 #define malloc(a,b,c) calloc(a, 1)
111 #define free(a,b) free(a)
113 typedef unsigned int u_daddr_t;
115 #include <sys/blist.h>
117 void panic(const char *ctl, ...);
122 * static support functions
125 static daddr_t blst_leaf_alloc(blmeta_t *scan, daddr_t blk, int count);
126 static daddr_t blst_meta_alloc(blmeta_t *scan, daddr_t blk,
127 daddr_t count, daddr_t radix, int skip);
128 static void blst_leaf_free(blmeta_t *scan, daddr_t relblk, int count);
129 static void blst_meta_free(blmeta_t *scan, daddr_t freeBlk, daddr_t count,
130 daddr_t radix, int skip, daddr_t blk);
131 static void blst_copy(blmeta_t *scan, daddr_t blk, daddr_t radix,
132 daddr_t skip, blist_t dest, daddr_t count);
133 static int blst_leaf_fill(blmeta_t *scan, daddr_t blk, int count);
134 static int blst_meta_fill(blmeta_t *scan, daddr_t allocBlk, daddr_t count,
135 daddr_t radix, int skip, daddr_t blk);
136 static daddr_t blst_radix_init(blmeta_t *scan, daddr_t radix,
137 int skip, daddr_t count);
139 static void blst_radix_print(blmeta_t *scan, daddr_t blk,
140 daddr_t radix, int skip, int tab);
144 static MALLOC_DEFINE(M_SWAP, "SWAP", "Swap space");
148 * blist_create() - create a blist capable of handling up to the specified
151 * blocks - must be greater than 0
152 * flags - malloc flags
154 * The smallest blist consists of a single leaf node capable of
155 * managing BLIST_BMAP_RADIX blocks.
159 blist_create(daddr_t blocks, int flags)
166 * Calculate radix and skip field used for scanning.
168 radix = BLIST_BMAP_RADIX;
170 while (radix < blocks) {
171 radix *= BLIST_META_RADIX;
172 skip = (skip + 1) * BLIST_META_RADIX;
175 bl = malloc(sizeof(struct blist), M_SWAP, flags | M_ZERO);
177 bl->bl_blocks = blocks;
178 bl->bl_radix = radix;
180 bl->bl_rootblks = 1 +
181 blst_radix_init(NULL, bl->bl_radix, bl->bl_skip, blocks);
182 bl->bl_root = malloc(sizeof(blmeta_t) * bl->bl_rootblks, M_SWAP, flags);
184 #if defined(BLIST_DEBUG)
186 "BLIST representing %lld blocks (%lld MB of swap)"
187 ", requiring %lldK of ram\n",
188 (long long)bl->bl_blocks,
189 (long long)bl->bl_blocks * 4 / 1024,
190 (long long)(bl->bl_rootblks * sizeof(blmeta_t) + 1023) / 1024
192 printf("BLIST raw radix tree contains %lld records\n",
193 (long long)bl->bl_rootblks);
195 blst_radix_init(bl->bl_root, bl->bl_radix, bl->bl_skip, blocks);
201 blist_destroy(blist_t bl)
203 free(bl->bl_root, M_SWAP);
208 * blist_alloc() - reserve space in the block bitmap. Return the base
209 * of a contiguous region or SWAPBLK_NONE if space could
214 blist_alloc(blist_t bl, daddr_t count)
216 daddr_t blk = SWAPBLK_NONE;
219 if (bl->bl_radix == BLIST_BMAP_RADIX)
220 blk = blst_leaf_alloc(bl->bl_root, 0, count);
222 blk = blst_meta_alloc(bl->bl_root, 0, count, bl->bl_radix, bl->bl_skip);
223 if (blk != SWAPBLK_NONE)
224 bl->bl_free -= count;
230 * blist_free() - free up space in the block bitmap. Return the base
231 * of a contiguous region. Panic if an inconsistancy is
236 blist_free(blist_t bl, daddr_t blkno, daddr_t count)
239 if (bl->bl_radix == BLIST_BMAP_RADIX)
240 blst_leaf_free(bl->bl_root, blkno, count);
242 blst_meta_free(bl->bl_root, blkno, count, bl->bl_radix, bl->bl_skip, 0);
243 bl->bl_free += count;
248 * blist_fill() - mark a region in the block bitmap as off-limits
249 * to the allocator (i.e. allocate it), ignoring any
250 * existing allocations. Return the number of blocks
251 * actually filled that were free before the call.
255 blist_fill(blist_t bl, daddr_t blkno, daddr_t count)
260 if (bl->bl_radix == BLIST_BMAP_RADIX)
261 filled = blst_leaf_fill(bl->bl_root, blkno, count);
263 filled = blst_meta_fill(bl->bl_root, blkno, count,
264 bl->bl_radix, bl->bl_skip, 0);
265 bl->bl_free -= filled;
272 * blist_resize() - resize an existing radix tree to handle the
273 * specified number of blocks. This will reallocate
274 * the tree and transfer the previous bitmap to the new
275 * one. When extending the tree you can specify whether
276 * the new blocks are to left allocated or freed.
280 blist_resize(blist_t *pbl, daddr_t count, int freenew, int flags)
282 blist_t newbl = blist_create(count, flags);
286 if (count > save->bl_blocks)
287 count = save->bl_blocks;
288 blst_copy(save->bl_root, 0, save->bl_radix, save->bl_skip, newbl, count);
291 * If resizing upwards, should we free the new space or not?
293 if (freenew && count < newbl->bl_blocks) {
294 blist_free(newbl, count, newbl->bl_blocks - count);
302 * blist_print() - dump radix tree
306 blist_print(blist_t bl)
309 blst_radix_print(bl->bl_root, 0, bl->bl_radix, bl->bl_skip, 4);
315 /************************************************************************
316 * ALLOCATION SUPPORT FUNCTIONS *
317 ************************************************************************
319 * These support functions do all the actual work. They may seem
320 * rather longish, but that's because I've commented them up. The
321 * actual code is straight forward.
326 * blist_leaf_alloc() - allocate at a leaf in the radix tree (a bitmap).
328 * This is the core of the allocator and is optimized for the 1 block
329 * and the BLIST_BMAP_RADIX block allocation cases. Other cases are
330 * somewhat slower. The 1 block allocation case is log2 and extremely
340 u_daddr_t orig = scan->u.bmu_bitmap;
344 * Optimize bitmap all-allocated case. Also, count = 1
345 * case assumes at least 1 bit is free in the bitmap, so
346 * we have to take care of this case here.
348 scan->bm_bighint = 0;
349 return(SWAPBLK_NONE);
353 * Optimized code to allocate one bit out of the bitmap
356 int j = BLIST_BMAP_RADIX/2;
359 mask = (u_daddr_t)-1 >> (BLIST_BMAP_RADIX/2);
362 if ((orig & mask) == 0) {
369 scan->u.bmu_bitmap &= ~(1 << r);
372 if (count <= BLIST_BMAP_RADIX) {
374 * non-optimized code to allocate N bits out of the bitmap.
375 * The more bits, the faster the code runs. It will run
376 * the slowest allocating 2 bits, but since there aren't any
377 * memory ops in the core loop (or shouldn't be, anyway),
378 * you probably won't notice the difference.
381 int n = BLIST_BMAP_RADIX - count;
384 mask = (u_daddr_t)-1 >> n;
386 for (j = 0; j <= n; ++j) {
387 if ((orig & mask) == mask) {
388 scan->u.bmu_bitmap &= ~mask;
395 * We couldn't allocate count in this subtree, update bighint.
397 scan->bm_bighint = count - 1;
398 return(SWAPBLK_NONE);
402 * blist_meta_alloc() - allocate at a meta in the radix tree.
404 * Attempt to allocate at a meta node. If we can't, we update
405 * bighint and return a failure. Updating bighint optimize future
406 * calls that hit this node. We have to check for our collapse cases
407 * and we have a few optimizations strewn in as well.
419 int next_skip = ((u_int)skip / BLIST_META_RADIX);
421 if (scan->u.bmu_avail == 0) {
423 * ALL-ALLOCATED special case
425 scan->bm_bighint = count;
426 return(SWAPBLK_NONE);
429 if (scan->u.bmu_avail == radix) {
430 radix /= BLIST_META_RADIX;
433 * ALL-FREE special case, initialize uninitialize
436 for (i = 1; i <= skip; i += next_skip) {
437 if (scan[i].bm_bighint == (daddr_t)-1)
439 if (next_skip == 1) {
440 scan[i].u.bmu_bitmap = (u_daddr_t)-1;
441 scan[i].bm_bighint = BLIST_BMAP_RADIX;
443 scan[i].bm_bighint = radix;
444 scan[i].u.bmu_avail = radix;
448 radix /= BLIST_META_RADIX;
451 for (i = 1; i <= skip; i += next_skip) {
452 if (count <= scan[i].bm_bighint) {
454 * count fits in object
457 if (next_skip == 1) {
458 r = blst_leaf_alloc(&scan[i], blk, count);
460 r = blst_meta_alloc(&scan[i], blk, count, radix, next_skip - 1);
462 if (r != SWAPBLK_NONE) {
463 scan->u.bmu_avail -= count;
464 if (scan->bm_bighint > scan->u.bmu_avail)
465 scan->bm_bighint = scan->u.bmu_avail;
468 } else if (scan[i].bm_bighint == (daddr_t)-1) {
473 } else if (count > radix) {
475 * count does not fit in object even if it were
478 panic("blist_meta_alloc: allocation too large");
484 * We couldn't allocate count in this subtree, update bighint.
486 if (scan->bm_bighint >= count)
487 scan->bm_bighint = count - 1;
488 return(SWAPBLK_NONE);
492 * BLST_LEAF_FREE() - free allocated block from leaf bitmap
503 * free some data in this bitmap
506 * 0000111111111110000
510 int n = blk & (BLIST_BMAP_RADIX - 1);
513 mask = ((u_daddr_t)-1 << n) &
514 ((u_daddr_t)-1 >> (BLIST_BMAP_RADIX - count - n));
516 if (scan->u.bmu_bitmap & mask)
517 panic("blst_radix_free: freeing free block");
518 scan->u.bmu_bitmap |= mask;
521 * We could probably do a better job here. We are required to make
522 * bighint at least as large as the biggest contiguous block of
523 * data. If we just shoehorn it, a little extra overhead will
524 * be incured on the next allocation (but only that one typically).
526 scan->bm_bighint = BLIST_BMAP_RADIX;
530 * BLST_META_FREE() - free allocated blocks from radix tree meta info
532 * This support routine frees a range of blocks from the bitmap.
533 * The range must be entirely enclosed by this radix node. If a
534 * meta node, we break the range down recursively to free blocks
535 * in subnodes (which means that this code can free an arbitrary
536 * range whereas the allocation code cannot allocate an arbitrary
550 int next_skip = ((u_int)skip / BLIST_META_RADIX);
553 printf("free (%llx,%lld) FROM (%llx,%lld)\n",
554 (long long)freeBlk, (long long)count,
555 (long long)blk, (long long)radix
559 if (scan->u.bmu_avail == 0) {
561 * ALL-ALLOCATED special case, with possible
562 * shortcut to ALL-FREE special case.
564 scan->u.bmu_avail = count;
565 scan->bm_bighint = count;
567 if (count != radix) {
568 for (i = 1; i <= skip; i += next_skip) {
569 if (scan[i].bm_bighint == (daddr_t)-1)
571 scan[i].bm_bighint = 0;
572 if (next_skip == 1) {
573 scan[i].u.bmu_bitmap = 0;
575 scan[i].u.bmu_avail = 0;
581 scan->u.bmu_avail += count;
582 /* scan->bm_bighint = radix; */
586 * ALL-FREE special case.
589 if (scan->u.bmu_avail == radix)
591 if (scan->u.bmu_avail > radix)
592 panic("blst_meta_free: freeing already free blocks (%lld) %lld/%lld",
593 (long long)count, (long long)scan->u.bmu_avail,
597 * Break the free down into its components
600 radix /= BLIST_META_RADIX;
602 i = (freeBlk - blk) / radix;
604 i = i * next_skip + 1;
606 while (i <= skip && blk < freeBlk + count) {
609 v = blk + radix - freeBlk;
613 if (scan->bm_bighint == (daddr_t)-1)
614 panic("blst_meta_free: freeing unexpected range");
616 if (next_skip == 1) {
617 blst_leaf_free(&scan[i], freeBlk, v);
619 blst_meta_free(&scan[i], freeBlk, v, radix, next_skip - 1, blk);
621 if (scan->bm_bighint < scan[i].bm_bighint)
622 scan->bm_bighint = scan[i].bm_bighint;
631 * BLIST_RADIX_COPY() - copy one radix tree to another
633 * Locates free space in the source tree and frees it in the destination
634 * tree. The space may not already be free in the destination.
637 static void blst_copy(
652 if (radix == BLIST_BMAP_RADIX) {
653 u_daddr_t v = scan->u.bmu_bitmap;
655 if (v == (u_daddr_t)-1) {
656 blist_free(dest, blk, count);
660 for (i = 0; i < BLIST_BMAP_RADIX && i < count; ++i) {
662 blist_free(dest, blk + i, 1);
672 if (scan->u.bmu_avail == 0) {
674 * Source all allocated, leave dest allocated
678 if (scan->u.bmu_avail == radix) {
680 * Source all free, free entire dest
683 blist_free(dest, blk, count);
685 blist_free(dest, blk, radix);
690 radix /= BLIST_META_RADIX;
691 next_skip = ((u_int)skip / BLIST_META_RADIX);
693 for (i = 1; count && i <= skip; i += next_skip) {
694 if (scan[i].bm_bighint == (daddr_t)-1)
697 if (count >= radix) {
725 * BLST_LEAF_FILL() - allocate specific blocks in leaf bitmap
727 * This routine allocates all blocks in the specified range
728 * regardless of any existing allocations in that range. Returns
729 * the number of blocks allocated by the call.
733 blst_leaf_fill(blmeta_t *scan, daddr_t blk, int count)
735 int n = blk & (BLIST_BMAP_RADIX - 1);
737 u_daddr_t mask, bitmap;
739 mask = ((u_daddr_t)-1 << n) &
740 ((u_daddr_t)-1 >> (BLIST_BMAP_RADIX - count - n));
742 /* Count the number of blocks we're about to allocate */
743 bitmap = scan->u.bmu_bitmap & mask;
744 for (nblks = 0; bitmap != 0; nblks++)
745 bitmap &= bitmap - 1;
747 scan->u.bmu_bitmap &= ~mask;
752 * BLIST_META_FILL() - allocate specific blocks at a meta node
754 * This routine allocates the specified range of blocks,
755 * regardless of any existing allocations in the range. The
756 * range must be within the extent of this node. Returns the
757 * number of blocks allocated by the call.
769 int next_skip = ((u_int)skip / BLIST_META_RADIX);
772 if (count == radix || scan->u.bmu_avail == 0) {
774 * ALL-ALLOCATED special case
776 nblks = scan->u.bmu_avail;
777 scan->u.bmu_avail = 0;
778 scan->bm_bighint = count;
782 if (scan->u.bmu_avail == radix) {
783 radix /= BLIST_META_RADIX;
786 * ALL-FREE special case, initialize sublevel
788 for (i = 1; i <= skip; i += next_skip) {
789 if (scan[i].bm_bighint == (daddr_t)-1)
791 if (next_skip == 1) {
792 scan[i].u.bmu_bitmap = (u_daddr_t)-1;
793 scan[i].bm_bighint = BLIST_BMAP_RADIX;
795 scan[i].bm_bighint = radix;
796 scan[i].u.bmu_avail = radix;
800 radix /= BLIST_META_RADIX;
804 panic("blist_meta_fill: allocation too large");
806 i = (allocBlk - blk) / radix;
808 i = i * next_skip + 1;
810 while (i <= skip && blk < allocBlk + count) {
813 v = blk + radix - allocBlk;
817 if (scan->bm_bighint == (daddr_t)-1)
818 panic("blst_meta_fill: filling unexpected range");
820 if (next_skip == 1) {
821 nblks += blst_leaf_fill(&scan[i], allocBlk, v);
823 nblks += blst_meta_fill(&scan[i], allocBlk, v,
824 radix, next_skip - 1, blk);
831 scan->u.bmu_avail -= nblks;
836 * BLST_RADIX_INIT() - initialize radix tree
838 * Initialize our meta structures and bitmaps and calculate the exact
839 * amount of space required to manage 'count' blocks - this space may
840 * be considerably less than the calculated radix due to the large
841 * RADIX values we use.
845 blst_radix_init(blmeta_t *scan, daddr_t radix, int skip, daddr_t count)
849 daddr_t memindex = 0;
855 if (radix == BLIST_BMAP_RADIX) {
857 scan->bm_bighint = 0;
858 scan->u.bmu_bitmap = 0;
864 * Meta node. If allocating the entire object we can special
865 * case it. However, we need to figure out how much memory
866 * is required to manage 'count' blocks, so we continue on anyway.
870 scan->bm_bighint = 0;
871 scan->u.bmu_avail = 0;
874 radix /= BLIST_META_RADIX;
875 next_skip = ((u_int)skip / BLIST_META_RADIX);
877 for (i = 1; i <= skip; i += next_skip) {
878 if (count >= radix) {
880 * Allocate the entire object
882 memindex = i + blst_radix_init(
883 ((scan) ? &scan[i] : NULL),
889 } else if (count > 0) {
891 * Allocate a partial object
893 memindex = i + blst_radix_init(
894 ((scan) ? &scan[i] : NULL),
902 * Add terminator and break out
905 scan[i].bm_bighint = (daddr_t)-1;
917 blst_radix_print(blmeta_t *scan, daddr_t blk, daddr_t radix, int skip, int tab)
923 if (radix == BLIST_BMAP_RADIX) {
925 "%*.*s(%08llx,%lld): bitmap %08llx big=%lld\n",
927 (long long)blk, (long long)radix,
928 (long long)scan->u.bmu_bitmap,
929 (long long)scan->bm_bighint
934 if (scan->u.bmu_avail == 0) {
936 "%*.*s(%08llx,%lld) ALL ALLOCATED\n",
943 if (scan->u.bmu_avail == radix) {
945 "%*.*s(%08llx,%lld) ALL FREE\n",
954 "%*.*s(%08llx,%lld): subtree (%lld/%lld) big=%lld {\n",
956 (long long)blk, (long long)radix,
957 (long long)scan->u.bmu_avail,
959 (long long)scan->bm_bighint
962 radix /= BLIST_META_RADIX;
963 next_skip = ((u_int)skip / BLIST_META_RADIX);
966 for (i = 1; i <= skip; i += next_skip) {
967 if (scan[i].bm_bighint == (daddr_t)-1) {
969 "%*.*s(%08llx,%lld): Terminator\n",
971 (long long)blk, (long long)radix
998 main(int ac, char **av)
1004 for (i = 1; i < ac; ++i) {
1005 const char *ptr = av[i];
1007 size = strtol(ptr, NULL, 0);
1011 fprintf(stderr, "Bad option: %s\n", ptr - 2);
1014 bl = blist_create(size, M_WAITOK);
1015 blist_free(bl, 0, size);
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);
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",
1048 (long long *)&da, (long long *)&count) == 2) {
1049 blist_free(bl, da, count);
1055 if (sscanf(buf + 1, "%llx %lld",
1056 (long long *)&da, (long long *)&count) == 2) {
1058 blist_fill(bl, da, count));
1083 panic(const char *ctl, ...)
1088 vfprintf(stderr, ctl, va);
1089 fprintf(stderr, "\n");