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 1.8 2008/08/10 22:09:50 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 ((swblk_t)-1)
118 #include <sys/types.h>
124 #define kmalloc(a,b,c) malloc(a)
125 #define kfree(a,b) free(a)
127 #include <sys/blist.h>
129 void panic(const char *ctl
, ...);
134 * static support functions
137 static swblk_t
blst_leaf_alloc(blmeta_t
*scan
, swblk_t blk
, int count
);
138 static swblk_t
blst_meta_alloc(blmeta_t
*scan
, swblk_t blk
,
139 swblk_t count
, swblk_t radix
, int skip
);
140 static void blst_leaf_free(blmeta_t
*scan
, swblk_t relblk
, int count
);
141 static void blst_meta_free(blmeta_t
*scan
, swblk_t freeBlk
, swblk_t count
,
142 swblk_t radix
, int skip
, swblk_t blk
);
143 static void blst_copy(blmeta_t
*scan
, swblk_t blk
, swblk_t radix
,
144 swblk_t skip
, blist_t dest
, swblk_t count
);
145 static swblk_t
blst_radix_init(blmeta_t
*scan
, swblk_t radix
,
146 int skip
, swblk_t count
);
148 static void blst_radix_print(blmeta_t
*scan
, swblk_t blk
,
149 swblk_t radix
, int skip
, int tab
);
153 static MALLOC_DEFINE(M_SWAP
, "SWAP", "Swap space");
157 * blist_create() - create a blist capable of handling up to the specified
160 * blocks must be greater then 0
162 * The smallest blist consists of a single leaf node capable of
163 * managing BLIST_BMAP_RADIX blocks.
167 blist_create(swblk_t blocks
)
174 * Calculate radix and skip field used for scanning.
176 radix
= BLIST_BMAP_RADIX
;
178 while (radix
< blocks
) {
179 radix
*= BLIST_META_RADIX
;
180 skip
= (skip
+ 1) * BLIST_META_RADIX
;
183 bl
= kmalloc(sizeof(struct blist
), M_SWAP
, M_WAITOK
);
185 bzero(bl
, sizeof(*bl
));
187 bl
->bl_blocks
= blocks
;
188 bl
->bl_radix
= radix
;
190 bl
->bl_rootblks
= 1 +
191 blst_radix_init(NULL
, bl
->bl_radix
, bl
->bl_skip
, blocks
);
192 bl
->bl_root
= kmalloc(sizeof(blmeta_t
) * bl
->bl_rootblks
, M_SWAP
, M_WAITOK
);
194 #if defined(BLIST_DEBUG)
196 "BLIST representing %d blocks (%d MB of swap)"
197 ", requiring %dK of ram\n",
199 bl
->bl_blocks
* 4 / 1024,
200 (bl
->bl_rootblks
* sizeof(blmeta_t
) + 1023) / 1024
202 kprintf("BLIST raw radix tree contains %d records\n", bl
->bl_rootblks
);
204 blst_radix_init(bl
->bl_root
, bl
->bl_radix
, bl
->bl_skip
, blocks
);
210 blist_destroy(blist_t bl
)
212 kfree(bl
->bl_root
, M_SWAP
);
217 * blist_alloc() - reserve space in the block bitmap. Return the base
218 * of a contiguous region or SWAPBLK_NONE if space could
223 blist_alloc(blist_t bl
, swblk_t count
)
225 swblk_t blk
= SWAPBLK_NONE
;
228 if (bl
->bl_radix
== BLIST_BMAP_RADIX
)
229 blk
= blst_leaf_alloc(bl
->bl_root
, 0, count
);
231 blk
= blst_meta_alloc(bl
->bl_root
, 0, count
, bl
->bl_radix
, bl
->bl_skip
);
232 if (blk
!= SWAPBLK_NONE
)
233 bl
->bl_free
-= count
;
239 * blist_free() - free up space in the block bitmap. Return the base
240 * of a contiguous region. Panic if an inconsistancy is
245 blist_free(blist_t bl
, swblk_t blkno
, swblk_t count
)
248 if (bl
->bl_radix
== BLIST_BMAP_RADIX
)
249 blst_leaf_free(bl
->bl_root
, blkno
, count
);
251 blst_meta_free(bl
->bl_root
, blkno
, count
, bl
->bl_radix
, bl
->bl_skip
, 0);
252 bl
->bl_free
+= count
;
257 * blist_resize() - resize an existing radix tree to handle the
258 * specified number of blocks. This will reallocate
259 * the tree and transfer the previous bitmap to the new
260 * one. When extending the tree you can specify whether
261 * the new blocks are to left allocated or freed.
265 blist_resize(blist_t
*pbl
, swblk_t count
, int freenew
)
267 blist_t newbl
= blist_create(count
);
271 if (count
> save
->bl_blocks
)
272 count
= save
->bl_blocks
;
273 blst_copy(save
->bl_root
, 0, save
->bl_radix
, save
->bl_skip
, newbl
, count
);
276 * If resizing upwards, should we free the new space or not?
278 if (freenew
&& count
< newbl
->bl_blocks
) {
279 blist_free(newbl
, count
, newbl
->bl_blocks
- count
);
287 * blist_print() - dump radix tree
291 blist_print(blist_t bl
)
293 kprintf("BLIST {\n");
294 blst_radix_print(bl
->bl_root
, 0, bl
->bl_radix
, bl
->bl_skip
, 4);
300 /************************************************************************
301 * ALLOCATION SUPPORT FUNCTIONS *
302 ************************************************************************
304 * These support functions do all the actual work. They may seem
305 * rather longish, but that's because I've commented them up. The
306 * actual code is straight forward.
311 * blist_leaf_alloc() - allocate at a leaf in the radix tree (a bitmap).
313 * This is the core of the allocator and is optimized for the 1 block
314 * and the BLIST_BMAP_RADIX block allocation cases. Other cases are
315 * somewhat slower. The 1 block allocation case is log2 and extremely
320 blst_leaf_alloc(blmeta_t
*scan
, swblk_t blk
, int count
)
322 u_swblk_t orig
= scan
->u
.bmu_bitmap
;
326 * Optimize bitmap all-allocated case. Also, count = 1
327 * case assumes at least 1 bit is free in the bitmap, so
328 * we have to take care of this case here.
330 scan
->bm_bighint
= 0;
331 return(SWAPBLK_NONE
);
335 * Optimized code to allocate one bit out of the bitmap
338 int j
= BLIST_BMAP_RADIX
/2;
341 mask
= (u_swblk_t
)-1 >> (BLIST_BMAP_RADIX
/2);
344 if ((orig
& mask
) == 0) {
351 scan
->u
.bmu_bitmap
&= ~(1 << r
);
354 if (count
<= BLIST_BMAP_RADIX
) {
356 * non-optimized code to allocate N bits out of the bitmap.
357 * The more bits, the faster the code runs. It will run
358 * the slowest allocating 2 bits, but since there aren't any
359 * memory ops in the core loop (or shouldn't be, anyway),
360 * you probably won't notice the difference.
363 int n
= BLIST_BMAP_RADIX
- count
;
366 mask
= (u_swblk_t
)-1 >> n
;
368 for (j
= 0; j
<= n
; ++j
) {
369 if ((orig
& mask
) == mask
) {
370 scan
->u
.bmu_bitmap
&= ~mask
;
377 * We couldn't allocate count in this subtree, update bighint.
379 scan
->bm_bighint
= count
- 1;
380 return(SWAPBLK_NONE
);
384 * blist_meta_alloc() - allocate at a meta in the radix tree.
386 * Attempt to allocate at a meta node. If we can't, we update
387 * bighint and return a failure. Updating bighint optimize future
388 * calls that hit this node. We have to check for our collapse cases
389 * and we have a few optimizations strewn in as well.
393 blst_meta_alloc(blmeta_t
*scan
, swblk_t blk
, swblk_t count
,
394 swblk_t radix
, int skip
)
397 int next_skip
= ((u_int
)skip
/ BLIST_META_RADIX
);
399 if (scan
->u
.bmu_avail
== 0) {
401 * ALL-ALLOCATED special case
403 scan
->bm_bighint
= count
;
404 return(SWAPBLK_NONE
);
407 if (scan
->u
.bmu_avail
== radix
) {
408 radix
/= BLIST_META_RADIX
;
411 * ALL-FREE special case, initialize uninitialize
414 for (i
= 1; i
<= skip
; i
+= next_skip
) {
415 if (scan
[i
].bm_bighint
== (swblk_t
)-1)
417 if (next_skip
== 1) {
418 scan
[i
].u
.bmu_bitmap
= (u_swblk_t
)-1;
419 scan
[i
].bm_bighint
= BLIST_BMAP_RADIX
;
421 scan
[i
].bm_bighint
= radix
;
422 scan
[i
].u
.bmu_avail
= radix
;
426 radix
/= BLIST_META_RADIX
;
429 for (i
= 1; i
<= skip
; i
+= next_skip
) {
430 if (count
<= scan
[i
].bm_bighint
) {
432 * count fits in object
435 if (next_skip
== 1) {
436 r
= blst_leaf_alloc(&scan
[i
], blk
, count
);
438 r
= blst_meta_alloc(&scan
[i
], blk
, count
, radix
, next_skip
- 1);
440 if (r
!= SWAPBLK_NONE
) {
441 scan
->u
.bmu_avail
-= count
;
442 if (scan
->bm_bighint
> scan
->u
.bmu_avail
)
443 scan
->bm_bighint
= scan
->u
.bmu_avail
;
446 } else if (scan
[i
].bm_bighint
== (swblk_t
)-1) {
451 } else if (count
> radix
) {
453 * count does not fit in object even if it were
456 panic("blist_meta_alloc: allocation too large");
462 * We couldn't allocate count in this subtree, update bighint.
464 if (scan
->bm_bighint
>= count
)
465 scan
->bm_bighint
= count
- 1;
466 return(SWAPBLK_NONE
);
470 * BLST_LEAF_FREE() - free allocated block from leaf bitmap
475 blst_leaf_free(blmeta_t
*scan
, swblk_t blk
, int count
)
478 * free some data in this bitmap
481 * 0000111111111110000
485 int n
= blk
& (BLIST_BMAP_RADIX
- 1);
488 mask
= ((u_swblk_t
)-1 << n
) &
489 ((u_swblk_t
)-1 >> (BLIST_BMAP_RADIX
- count
- n
));
491 if (scan
->u
.bmu_bitmap
& mask
)
492 panic("blst_radix_free: freeing free block");
493 scan
->u
.bmu_bitmap
|= mask
;
496 * We could probably do a better job here. We are required to make
497 * bighint at least as large as the biggest contiguous block of
498 * data. If we just shoehorn it, a little extra overhead will
499 * be incured on the next allocation (but only that one typically).
501 scan
->bm_bighint
= BLIST_BMAP_RADIX
;
505 * BLST_META_FREE() - free allocated blocks from radix tree meta info
507 * This support routine frees a range of blocks from the bitmap.
508 * The range must be entirely enclosed by this radix node. If a
509 * meta node, we break the range down recursively to free blocks
510 * in subnodes (which means that this code can free an arbitrary
511 * range whereas the allocation code cannot allocate an arbitrary
516 blst_meta_free(blmeta_t
*scan
, swblk_t freeBlk
, swblk_t count
,
517 swblk_t radix
, int skip
, swblk_t blk
)
520 int next_skip
= ((u_int
)skip
/ BLIST_META_RADIX
);
523 kprintf("FREE (%x,%d) FROM (%x,%d)\n",
529 if (scan
->u
.bmu_avail
== 0) {
531 * ALL-ALLOCATED special case, with possible
532 * shortcut to ALL-FREE special case.
534 scan
->u
.bmu_avail
= count
;
535 scan
->bm_bighint
= count
;
537 if (count
!= radix
) {
538 for (i
= 1; i
<= skip
; i
+= next_skip
) {
539 if (scan
[i
].bm_bighint
== (swblk_t
)-1)
541 scan
[i
].bm_bighint
= 0;
542 if (next_skip
== 1) {
543 scan
[i
].u
.bmu_bitmap
= 0;
545 scan
[i
].u
.bmu_avail
= 0;
551 scan
->u
.bmu_avail
+= count
;
552 /* scan->bm_bighint = radix; */
556 * ALL-FREE special case.
559 if (scan
->u
.bmu_avail
== radix
)
561 if (scan
->u
.bmu_avail
> radix
)
562 panic("blst_meta_free: freeing already free blocks (%d) %d/%d", count
, scan
->u
.bmu_avail
, radix
);
565 * Break the free down into its components
568 radix
/= BLIST_META_RADIX
;
570 i
= (freeBlk
- blk
) / radix
;
572 i
= i
* next_skip
+ 1;
574 while (i
<= skip
&& blk
< freeBlk
+ count
) {
577 v
= blk
+ radix
- freeBlk
;
581 if (scan
->bm_bighint
== (swblk_t
)-1)
582 panic("blst_meta_free: freeing unexpected range");
584 if (next_skip
== 1) {
585 blst_leaf_free(&scan
[i
], freeBlk
, v
);
587 blst_meta_free(&scan
[i
], freeBlk
, v
, radix
, next_skip
- 1, blk
);
589 if (scan
->bm_bighint
< scan
[i
].bm_bighint
)
590 scan
->bm_bighint
= scan
[i
].bm_bighint
;
599 * BLIST_RADIX_COPY() - copy one radix tree to another
601 * Locates free space in the source tree and frees it in the destination
602 * tree. The space may not already be free in the destination.
606 blst_copy(blmeta_t
*scan
, swblk_t blk
, swblk_t radix
,
607 swblk_t skip
, blist_t dest
, swblk_t count
)
616 if (radix
== BLIST_BMAP_RADIX
) {
617 u_swblk_t v
= scan
->u
.bmu_bitmap
;
619 if (v
== (u_swblk_t
)-1) {
620 blist_free(dest
, blk
, count
);
624 for (i
= 0; i
< BLIST_BMAP_RADIX
&& i
< count
; ++i
) {
626 blist_free(dest
, blk
+ i
, 1);
636 if (scan
->u
.bmu_avail
== 0) {
638 * Source all allocated, leave dest allocated
642 if (scan
->u
.bmu_avail
== radix
) {
644 * Source all free, free entire dest
647 blist_free(dest
, blk
, count
);
649 blist_free(dest
, blk
, radix
);
654 radix
/= BLIST_META_RADIX
;
655 next_skip
= ((u_int
)skip
/ BLIST_META_RADIX
);
657 for (i
= 1; count
&& i
<= skip
; i
+= next_skip
) {
658 if (scan
[i
].bm_bighint
== (swblk_t
)-1)
661 if (count
>= radix
) {
689 * BLST_RADIX_INIT() - initialize radix tree
691 * Initialize our meta structures and bitmaps and calculate the exact
692 * amount of space required to manage 'count' blocks - this space may
693 * be considerably less then the calculated radix due to the large
694 * RADIX values we use.
698 blst_radix_init(blmeta_t
*scan
, swblk_t radix
, int skip
, swblk_t count
)
702 swblk_t memindex
= 0;
708 if (radix
== BLIST_BMAP_RADIX
) {
710 scan
->bm_bighint
= 0;
711 scan
->u
.bmu_bitmap
= 0;
717 * Meta node. If allocating the entire object we can special
718 * case it. However, we need to figure out how much memory
719 * is required to manage 'count' blocks, so we continue on anyway.
723 scan
->bm_bighint
= 0;
724 scan
->u
.bmu_avail
= 0;
727 radix
/= BLIST_META_RADIX
;
728 next_skip
= ((u_int
)skip
/ BLIST_META_RADIX
);
730 for (i
= 1; i
<= skip
; i
+= next_skip
) {
731 if (count
>= radix
) {
733 * Allocate the entire object
735 memindex
= i
+ blst_radix_init(
736 ((scan
) ? &scan
[i
] : NULL
),
742 } else if (count
> 0) {
744 * Allocate a partial object
746 memindex
= i
+ blst_radix_init(
747 ((scan
) ? &scan
[i
] : NULL
),
755 * Add terminator and break out
758 scan
[i
].bm_bighint
= (swblk_t
)-1;
770 blst_radix_print(blmeta_t
*scan
, swblk_t blk
, swblk_t radix
, int skip
, int tab
)
776 if (radix
== BLIST_BMAP_RADIX
) {
778 "%*.*s(%04x,%d): bitmap %08x big=%d\n",
787 if (scan
->u
.bmu_avail
== 0) {
789 "%*.*s(%04x,%d) ALL ALLOCATED\n",
796 if (scan
->u
.bmu_avail
== radix
) {
798 "%*.*s(%04x,%d) ALL FREE\n",
807 "%*.*s(%04x,%d): subtree (%d/%d) big=%d {\n",
815 radix
/= BLIST_META_RADIX
;
816 next_skip
= ((u_int
)skip
/ BLIST_META_RADIX
);
819 for (i
= 1; i
<= skip
; i
+= next_skip
) {
820 if (scan
[i
].bm_bighint
== (swblk_t
)-1) {
822 "%*.*s(%04x,%d): Terminator\n",
851 main(int ac
, char **av
)
857 for (i
= 1; i
< ac
; ++i
) {
858 const char *ptr
= av
[i
];
860 size
= strtol(ptr
, NULL
, 0);
864 fprintf(stderr
, "Bad option: %s\n", ptr
- 2);
867 bl
= blist_create(size
);
868 blist_free(bl
, 0, size
);
876 kprintf("%d/%d/%d> ", bl
->bl_free
, size
, bl
->bl_radix
);
878 if (fgets(buf
, sizeof(buf
), stdin
) == NULL
)
882 if (sscanf(buf
+ 1, "%d", &count
) == 1) {
883 blist_resize(&bl
, count
, 1);
891 if (sscanf(buf
+ 1, "%d", &count
) == 1) {
892 swblk_t blk
= blist_alloc(bl
, count
);
893 kprintf(" R=%04x\n", blk
);
899 if (sscanf(buf
+ 1, "%x %d", &da
, &count
) == 2) {
900 blist_free(bl
, da
, count
);
924 panic(const char *ctl
, ...)
929 vfprintf(stderr
, ctl
, va
);
930 fprintf(stderr
, "\n");