2 * BLIST.C - Bitmap allocator/deallocator, using a radix tree with hinting
4 * Copyright (c) 1998,2004 The DragonFly Project. All rights reserved.
6 * This code is derived from software contributed to The DragonFly Project
7 * by Matthew Dillon <dillon@backplane.com>
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
13 * 1. Redistributions of source code must retain the above copyright
14 * notice, this list of conditions and the following disclaimer.
15 * 2. Redistributions in binary form must reproduce the above copyright
16 * notice, this list of conditions and the following disclaimer in
17 * the documentation and/or other materials provided with the
19 * 3. Neither the name of The DragonFly Project nor the names of its
20 * contributors may be used to endorse or promote products derived
21 * from this software without specific, prior written permission.
23 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
24 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
25 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
26 * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
27 * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
28 * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
29 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
30 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
31 * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
32 * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
33 * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
37 * This module implements a general bitmap allocator/deallocator. The
38 * allocator eats around 2 bits per 'block'. The module does not
39 * try to interpret the meaning of a 'block' other then to return
40 * SWAPBLK_NONE on an allocation failure.
42 * A radix tree is used to maintain the bitmap. Two radix constants are
43 * involved: One for the bitmaps contained in the leaf nodes (typically
44 * 32), and one for the meta nodes (typically 16). Both meta and leaf
45 * nodes have a hint field. This field gives us a hint as to the largest
46 * free contiguous range of blocks under the node. It may contain a
47 * value that is too high, but will never contain a value that is too
48 * low. When the radix tree is searched, allocation failures in subtrees
51 * The radix tree also implements two collapsed states for meta nodes:
52 * the ALL-ALLOCATED state and the ALL-FREE state. If a meta node is
53 * in either of these two states, all information contained underneath
54 * the node is considered stale. These states are used to optimize
55 * allocation and freeing operations.
57 * The hinting greatly increases code efficiency for allocations while
58 * the general radix structure optimizes both allocations and frees. The
59 * radix tree should be able to operate well no matter how much
60 * fragmentation there is and no matter how large a bitmap is used.
62 * Unlike the rlist code, the blist code wires all necessary memory at
63 * creation time. Neither allocations nor frees require interaction with
64 * the memory subsystem. In contrast, the rlist code may allocate memory
65 * on an rlist_free() call. The non-blocking features of the blist code
66 * are used to great advantage in the swap code (vm/nswap_pager.c). The
67 * rlist code uses a little less overall memory then the blist code (but
68 * due to swap interleaving not all that much less), but the blist code
69 * scales much, much better.
71 * LAYOUT: The radix tree is layed out recursively using a
72 * linear array. Each meta node is immediately followed (layed out
73 * sequentially in memory) by BLIST_META_RADIX lower level nodes. This
74 * is a recursive structure but one that can be easily scanned through
75 * a very simple 'skip' calculation. In order to support large radixes,
76 * portions of the tree may reside outside our memory allocation. We
77 * handle this with an early-termination optimization (when bighint is
78 * set to -1) on the scan. The memory allocation is only large enough
79 * to cover the number of blocks requested at creation time even if it
80 * must be encompassed in larger root-node radix.
82 * NOTE: the allocator cannot currently allocate more then
83 * BLIST_BMAP_RADIX blocks per call. It will panic with 'allocation too
84 * large' if you try. This is an area that could use improvement. The
85 * radix is large enough that this restriction does not effect the swap
86 * system, though. Currently only the allocation code is effected by
87 * this algorithmic unfeature. The freeing code can handle arbitrary
90 * This code can be compiled stand-alone for debugging.
92 * $FreeBSD: src/sys/kern/subr_blist.c,v 1.5.2.2 2003/01/12 09:23:12 dillon Exp $
93 * $DragonFly: src/sys/kern/subr_blist.c,v 1.6 2006/09/05 00:55:45 dillon Exp $
98 #include <sys/param.h>
99 #include <sys/systm.h>
100 #include <sys/lock.h>
101 #include <sys/kernel.h>
102 #include <sys/blist.h>
103 #include <sys/malloc.h>
105 #include <vm/vm_object.h>
106 #include <vm/vm_kern.h>
107 #include <vm/vm_extern.h>
108 #include <vm/vm_page.h>
112 #ifndef BLIST_NO_DEBUG
116 #define SWAPBLK_NONE ((daddr_t)-1)
118 #include <sys/types.h>
124 #define kmalloc(a,b,c) malloc(a)
125 #define kfree(a,b) free(a)
127 typedef unsigned int u_daddr_t;
129 #include <sys/blist.h>
131 void panic(const char *ctl, ...);
136 * static support functions
139 static daddr_t blst_leaf_alloc(blmeta_t *scan, daddr_t blk, int count);
140 static daddr_t blst_meta_alloc(blmeta_t *scan, daddr_t blk,
141 daddr_t count, daddr_t radix, int skip);
142 static void blst_leaf_free(blmeta_t *scan, daddr_t relblk, int count);
143 static void blst_meta_free(blmeta_t *scan, daddr_t freeBlk, daddr_t count,
144 daddr_t radix, int skip, daddr_t blk);
145 static void blst_copy(blmeta_t *scan, daddr_t blk, daddr_t radix,
146 daddr_t skip, blist_t dest, daddr_t count);
147 static daddr_t blst_radix_init(blmeta_t *scan, daddr_t radix,
148 int skip, daddr_t count);
150 static void blst_radix_print(blmeta_t *scan, daddr_t blk,
151 daddr_t radix, int skip, int tab);
155 static MALLOC_DEFINE(M_SWAP, "SWAP", "Swap space");
159 * blist_create() - create a blist capable of handling up to the specified
162 * blocks must be greater then 0
164 * The smallest blist consists of a single leaf node capable of
165 * managing BLIST_BMAP_RADIX blocks.
169 blist_create(daddr_t blocks)
176 * Calculate radix and skip field used for scanning.
178 radix = BLIST_BMAP_RADIX;
180 while (radix < blocks) {
181 radix *= BLIST_META_RADIX;
182 skip = (skip + 1) * BLIST_META_RADIX;
185 bl = kmalloc(sizeof(struct blist), M_SWAP, M_WAITOK);
187 bzero(bl, sizeof(*bl));
189 bl->bl_blocks = blocks;
190 bl->bl_radix = radix;
192 bl->bl_rootblks = 1 +
193 blst_radix_init(NULL, bl->bl_radix, bl->bl_skip, blocks);
194 bl->bl_root = kmalloc(sizeof(blmeta_t) * bl->bl_rootblks, M_SWAP, M_WAITOK);
196 #if defined(BLIST_DEBUG)
198 "BLIST representing %d blocks (%d MB of swap)"
199 ", requiring %dK of ram\n",
201 bl->bl_blocks * 4 / 1024,
202 (bl->bl_rootblks * sizeof(blmeta_t) + 1023) / 1024
204 printf("BLIST raw radix tree contains %d records\n", bl->bl_rootblks);
206 blst_radix_init(bl->bl_root, bl->bl_radix, bl->bl_skip, blocks);
212 blist_destroy(blist_t bl)
214 kfree(bl->bl_root, M_SWAP);
219 * blist_alloc() - reserve space in the block bitmap. Return the base
220 * of a contiguous region or SWAPBLK_NONE if space could
225 blist_alloc(blist_t bl, daddr_t count)
227 daddr_t blk = SWAPBLK_NONE;
230 if (bl->bl_radix == BLIST_BMAP_RADIX)
231 blk = blst_leaf_alloc(bl->bl_root, 0, count);
233 blk = blst_meta_alloc(bl->bl_root, 0, count, bl->bl_radix, bl->bl_skip);
234 if (blk != SWAPBLK_NONE)
235 bl->bl_free -= count;
241 * blist_free() - free up space in the block bitmap. Return the base
242 * of a contiguous region. Panic if an inconsistancy is
247 blist_free(blist_t bl, daddr_t blkno, daddr_t count)
250 if (bl->bl_radix == BLIST_BMAP_RADIX)
251 blst_leaf_free(bl->bl_root, blkno, count);
253 blst_meta_free(bl->bl_root, blkno, count, bl->bl_radix, bl->bl_skip, 0);
254 bl->bl_free += count;
259 * blist_resize() - resize an existing radix tree to handle the
260 * specified number of blocks. This will reallocate
261 * the tree and transfer the previous bitmap to the new
262 * one. When extending the tree you can specify whether
263 * the new blocks are to left allocated or freed.
267 blist_resize(blist_t *pbl, daddr_t count, int freenew)
269 blist_t newbl = blist_create(count);
273 if (count > save->bl_blocks)
274 count = save->bl_blocks;
275 blst_copy(save->bl_root, 0, save->bl_radix, save->bl_skip, newbl, count);
278 * If resizing upwards, should we free the new space or not?
280 if (freenew && count < newbl->bl_blocks) {
281 blist_free(newbl, count, newbl->bl_blocks - count);
289 * blist_print() - dump radix tree
293 blist_print(blist_t bl)
296 blst_radix_print(bl->bl_root, 0, bl->bl_radix, bl->bl_skip, 4);
302 /************************************************************************
303 * ALLOCATION SUPPORT FUNCTIONS *
304 ************************************************************************
306 * These support functions do all the actual work. They may seem
307 * rather longish, but that's because I've commented them up. The
308 * actual code is straight forward.
313 * blist_leaf_alloc() - allocate at a leaf in the radix tree (a bitmap).
315 * This is the core of the allocator and is optimized for the 1 block
316 * and the BLIST_BMAP_RADIX block allocation cases. Other cases are
317 * somewhat slower. The 1 block allocation case is log2 and extremely
327 u_daddr_t orig = scan->u.bmu_bitmap;
331 * Optimize bitmap all-allocated case. Also, count = 1
332 * case assumes at least 1 bit is free in the bitmap, so
333 * we have to take care of this case here.
335 scan->bm_bighint = 0;
336 return(SWAPBLK_NONE);
340 * Optimized code to allocate one bit out of the bitmap
343 int j = BLIST_BMAP_RADIX/2;
346 mask = (u_daddr_t)-1 >> (BLIST_BMAP_RADIX/2);
349 if ((orig & mask) == 0) {
356 scan->u.bmu_bitmap &= ~(1 << r);
359 if (count <= BLIST_BMAP_RADIX) {
361 * non-optimized code to allocate N bits out of the bitmap.
362 * The more bits, the faster the code runs. It will run
363 * the slowest allocating 2 bits, but since there aren't any
364 * memory ops in the core loop (or shouldn't be, anyway),
365 * you probably won't notice the difference.
368 int n = BLIST_BMAP_RADIX - count;
371 mask = (u_daddr_t)-1 >> n;
373 for (j = 0; j <= n; ++j) {
374 if ((orig & mask) == mask) {
375 scan->u.bmu_bitmap &= ~mask;
382 * We couldn't allocate count in this subtree, update bighint.
384 scan->bm_bighint = count - 1;
385 return(SWAPBLK_NONE);
389 * blist_meta_alloc() - allocate at a meta in the radix tree.
391 * Attempt to allocate at a meta node. If we can't, we update
392 * bighint and return a failure. Updating bighint optimize future
393 * calls that hit this node. We have to check for our collapse cases
394 * and we have a few optimizations strewn in as well.
406 int next_skip = ((u_int)skip / BLIST_META_RADIX);
408 if (scan->u.bmu_avail == 0) {
410 * ALL-ALLOCATED special case
412 scan->bm_bighint = count;
413 return(SWAPBLK_NONE);
416 if (scan->u.bmu_avail == radix) {
417 radix /= BLIST_META_RADIX;
420 * ALL-FREE special case, initialize uninitialize
423 for (i = 1; i <= skip; i += next_skip) {
424 if (scan[i].bm_bighint == (daddr_t)-1)
426 if (next_skip == 1) {
427 scan[i].u.bmu_bitmap = (u_daddr_t)-1;
428 scan[i].bm_bighint = BLIST_BMAP_RADIX;
430 scan[i].bm_bighint = radix;
431 scan[i].u.bmu_avail = radix;
435 radix /= BLIST_META_RADIX;
438 for (i = 1; i <= skip; i += next_skip) {
439 if (count <= scan[i].bm_bighint) {
441 * count fits in object
444 if (next_skip == 1) {
445 r = blst_leaf_alloc(&scan[i], blk, count);
447 r = blst_meta_alloc(&scan[i], blk, count, radix, next_skip - 1);
449 if (r != SWAPBLK_NONE) {
450 scan->u.bmu_avail -= count;
451 if (scan->bm_bighint > scan->u.bmu_avail)
452 scan->bm_bighint = scan->u.bmu_avail;
455 } else if (scan[i].bm_bighint == (daddr_t)-1) {
460 } else if (count > radix) {
462 * count does not fit in object even if it were
465 panic("blist_meta_alloc: allocation too large");
471 * We couldn't allocate count in this subtree, update bighint.
473 if (scan->bm_bighint >= count)
474 scan->bm_bighint = count - 1;
475 return(SWAPBLK_NONE);
479 * BLST_LEAF_FREE() - free allocated block from leaf bitmap
490 * free some data in this bitmap
493 * 0000111111111110000
497 int n = blk & (BLIST_BMAP_RADIX - 1);
500 mask = ((u_daddr_t)-1 << n) &
501 ((u_daddr_t)-1 >> (BLIST_BMAP_RADIX - count - n));
503 if (scan->u.bmu_bitmap & mask)
504 panic("blst_radix_free: freeing free block");
505 scan->u.bmu_bitmap |= mask;
508 * We could probably do a better job here. We are required to make
509 * bighint at least as large as the biggest contiguous block of
510 * data. If we just shoehorn it, a little extra overhead will
511 * be incured on the next allocation (but only that one typically).
513 scan->bm_bighint = BLIST_BMAP_RADIX;
517 * BLST_META_FREE() - free allocated blocks from radix tree meta info
519 * This support routine frees a range of blocks from the bitmap.
520 * The range must be entirely enclosed by this radix node. If a
521 * meta node, we break the range down recursively to free blocks
522 * in subnodes (which means that this code can free an arbitrary
523 * range whereas the allocation code cannot allocate an arbitrary
537 int next_skip = ((u_int)skip / BLIST_META_RADIX);
540 printf("FREE (%x,%d) FROM (%x,%d)\n",
546 if (scan->u.bmu_avail == 0) {
548 * ALL-ALLOCATED special case, with possible
549 * shortcut to ALL-FREE special case.
551 scan->u.bmu_avail = count;
552 scan->bm_bighint = count;
554 if (count != radix) {
555 for (i = 1; i <= skip; i += next_skip) {
556 if (scan[i].bm_bighint == (daddr_t)-1)
558 scan[i].bm_bighint = 0;
559 if (next_skip == 1) {
560 scan[i].u.bmu_bitmap = 0;
562 scan[i].u.bmu_avail = 0;
568 scan->u.bmu_avail += count;
569 /* scan->bm_bighint = radix; */
573 * ALL-FREE special case.
576 if (scan->u.bmu_avail == radix)
578 if (scan->u.bmu_avail > radix)
579 panic("blst_meta_free: freeing already free blocks (%d) %d/%d", count, scan->u.bmu_avail, radix);
582 * Break the free down into its components
585 radix /= BLIST_META_RADIX;
587 i = (freeBlk - blk) / radix;
589 i = i * next_skip + 1;
591 while (i <= skip && blk < freeBlk + count) {
594 v = blk + radix - freeBlk;
598 if (scan->bm_bighint == (daddr_t)-1)
599 panic("blst_meta_free: freeing unexpected range");
601 if (next_skip == 1) {
602 blst_leaf_free(&scan[i], freeBlk, v);
604 blst_meta_free(&scan[i], freeBlk, v, radix, next_skip - 1, blk);
606 if (scan->bm_bighint < scan[i].bm_bighint)
607 scan->bm_bighint = scan[i].bm_bighint;
616 * BLIST_RADIX_COPY() - copy one radix tree to another
618 * Locates free space in the source tree and frees it in the destination
619 * tree. The space may not already be free in the destination.
622 static void blst_copy(
637 if (radix == BLIST_BMAP_RADIX) {
638 u_daddr_t v = scan->u.bmu_bitmap;
640 if (v == (u_daddr_t)-1) {
641 blist_free(dest, blk, count);
645 for (i = 0; i < BLIST_BMAP_RADIX && i < count; ++i) {
647 blist_free(dest, blk + i, 1);
657 if (scan->u.bmu_avail == 0) {
659 * Source all allocated, leave dest allocated
663 if (scan->u.bmu_avail == radix) {
665 * Source all free, free entire dest
668 blist_free(dest, blk, count);
670 blist_free(dest, blk, radix);
675 radix /= BLIST_META_RADIX;
676 next_skip = ((u_int)skip / BLIST_META_RADIX);
678 for (i = 1; count && i <= skip; i += next_skip) {
679 if (scan[i].bm_bighint == (daddr_t)-1)
682 if (count >= radix) {
710 * BLST_RADIX_INIT() - initialize radix tree
712 * Initialize our meta structures and bitmaps and calculate the exact
713 * amount of space required to manage 'count' blocks - this space may
714 * be considerably less then the calculated radix due to the large
715 * RADIX values we use.
719 blst_radix_init(blmeta_t *scan, daddr_t radix, int skip, daddr_t count)
723 daddr_t memindex = 0;
729 if (radix == BLIST_BMAP_RADIX) {
731 scan->bm_bighint = 0;
732 scan->u.bmu_bitmap = 0;
738 * Meta node. If allocating the entire object we can special
739 * case it. However, we need to figure out how much memory
740 * is required to manage 'count' blocks, so we continue on anyway.
744 scan->bm_bighint = 0;
745 scan->u.bmu_avail = 0;
748 radix /= BLIST_META_RADIX;
749 next_skip = ((u_int)skip / BLIST_META_RADIX);
751 for (i = 1; i <= skip; i += next_skip) {
752 if (count >= radix) {
754 * Allocate the entire object
756 memindex = i + blst_radix_init(
757 ((scan) ? &scan[i] : NULL),
763 } else if (count > 0) {
765 * Allocate a partial object
767 memindex = i + blst_radix_init(
768 ((scan) ? &scan[i] : NULL),
776 * Add terminator and break out
779 scan[i].bm_bighint = (daddr_t)-1;
791 blst_radix_print(blmeta_t *scan, daddr_t blk, daddr_t radix, int skip, int tab)
797 if (radix == BLIST_BMAP_RADIX) {
799 "%*.*s(%04x,%d): bitmap %08x big=%d\n",
808 if (scan->u.bmu_avail == 0) {
810 "%*.*s(%04x,%d) ALL ALLOCATED\n",
817 if (scan->u.bmu_avail == radix) {
819 "%*.*s(%04x,%d) ALL FREE\n",
828 "%*.*s(%04x,%d): subtree (%d/%d) big=%d {\n",
836 radix /= BLIST_META_RADIX;
837 next_skip = ((u_int)skip / BLIST_META_RADIX);
840 for (i = 1; i <= skip; i += next_skip) {
841 if (scan[i].bm_bighint == (daddr_t)-1) {
843 "%*.*s(%04x,%d): Terminator\n",
872 main(int ac, char **av)
878 for (i = 1; i < ac; ++i) {
879 const char *ptr = av[i];
881 size = strtol(ptr, NULL, 0);
885 fprintf(stderr, "Bad option: %s\n", ptr - 2);
888 bl = blist_create(size);
889 blist_free(bl, 0, size);
897 printf("%d/%d/%d> ", bl->bl_free, size, bl->bl_radix);
899 if (fgets(buf, sizeof(buf), stdin) == NULL)
903 if (sscanf(buf + 1, "%d", &count) == 1) {
904 blist_resize(&bl, count, 1);
912 if (sscanf(buf + 1, "%d", &count) == 1) {
913 daddr_t blk = blist_alloc(bl, count);
914 printf(" R=%04x\n", blk);
920 if (sscanf(buf + 1, "%x %d", &da, &count) == 2) {
921 blist_free(bl, da, count);
945 panic(const char *ctl, ...)
950 vfprintf(stderr, ctl, va);
951 fprintf(stderr, "\n");