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 * NOTE: The radix may exceed 32 bits in order to support up to 2^31
91 * blocks. The first divison will drop the radix down and fit
92 * it within a signed 32 bit integer.
94 * This code can be compiled stand-alone for debugging.
96 * $FreeBSD: src/sys/kern/subr_blist.c,v 1.5.2.2 2003/01/12 09:23:12 dillon Exp $
97 * $DragonFly: src/sys/kern/subr_blist.c,v 1.8 2008/08/10 22:09:50 dillon Exp $
102 #include <sys/param.h>
103 #include <sys/systm.h>
104 #include <sys/lock.h>
105 #include <sys/kernel.h>
106 #include <sys/blist.h>
107 #include <sys/malloc.h>
109 #include <vm/vm_object.h>
110 #include <vm/vm_kern.h>
111 #include <vm/vm_extern.h>
112 #include <vm/vm_page.h>
116 #ifndef BLIST_NO_DEBUG
120 #define SWAPBLK_NONE ((swblk_t)-1)
122 #include <sys/types.h>
128 #define kmalloc(a,b,c) malloc(a)
129 #define kfree(a,b) free(a)
130 #define kprintf printf
131 #define KKASSERT(exp)
133 #include <sys/blist.h>
135 void panic(const char *ctl
, ...);
140 * static support functions
143 static swblk_t
blst_leaf_alloc(blmeta_t
*scan
, swblk_t blk
, int count
);
144 static swblk_t
blst_meta_alloc(blmeta_t
*scan
, swblk_t blk
,
145 swblk_t count
, int64_t radix
, int skip
);
146 static void blst_leaf_free(blmeta_t
*scan
, swblk_t relblk
, int count
);
147 static void blst_meta_free(blmeta_t
*scan
, swblk_t freeBlk
, swblk_t count
,
148 int64_t radix
, int skip
, swblk_t blk
);
149 static void blst_copy(blmeta_t
*scan
, swblk_t blk
, int64_t radix
,
150 swblk_t skip
, blist_t dest
, swblk_t count
);
151 static swblk_t
blst_radix_init(blmeta_t
*scan
, int64_t radix
,
152 int skip
, swblk_t count
);
154 static void blst_radix_print(blmeta_t
*scan
, swblk_t blk
,
155 int64_t radix
, int skip
, int tab
);
159 static MALLOC_DEFINE(M_SWAP
, "SWAP", "Swap space");
163 * blist_create() - create a blist capable of handling up to the specified
166 * blocks must be greater then 0
168 * The smallest blist consists of a single leaf node capable of
169 * managing BLIST_BMAP_RADIX blocks.
173 blist_create(swblk_t blocks
)
180 * Calculate radix and skip field used for scanning.
182 * Radix can exceed 32 bits even if swblk_t is limited to 32 bits.
184 radix
= BLIST_BMAP_RADIX
;
186 while (radix
< blocks
) {
187 radix
*= BLIST_META_RADIX
;
188 skip
= (skip
+ 1) * BLIST_META_RADIX
;
192 bl
= kmalloc(sizeof(struct blist
), M_SWAP
, M_WAITOK
);
194 bzero(bl
, sizeof(*bl
));
196 bl
->bl_blocks
= blocks
;
197 bl
->bl_radix
= radix
;
199 bl
->bl_rootblks
= 1 +
200 blst_radix_init(NULL
, bl
->bl_radix
, bl
->bl_skip
, blocks
);
201 bl
->bl_root
= kmalloc(sizeof(blmeta_t
) * bl
->bl_rootblks
, M_SWAP
, M_WAITOK
);
203 #if defined(BLIST_DEBUG)
205 "BLIST representing %d blocks (%d MB of swap)"
206 ", requiring %dK of ram\n",
208 bl
->bl_blocks
* 4 / 1024,
209 (bl
->bl_rootblks
* sizeof(blmeta_t
) + 1023) / 1024
211 kprintf("BLIST raw radix tree contains %d records\n", bl
->bl_rootblks
);
213 blst_radix_init(bl
->bl_root
, bl
->bl_radix
, bl
->bl_skip
, blocks
);
219 blist_destroy(blist_t bl
)
221 kfree(bl
->bl_root
, M_SWAP
);
226 * blist_alloc() - reserve space in the block bitmap. Return the base
227 * of a contiguous region or SWAPBLK_NONE if space could
232 blist_alloc(blist_t bl
, swblk_t count
)
234 swblk_t blk
= SWAPBLK_NONE
;
237 if (bl
->bl_radix
== BLIST_BMAP_RADIX
)
238 blk
= blst_leaf_alloc(bl
->bl_root
, 0, count
);
240 blk
= blst_meta_alloc(bl
->bl_root
, 0, count
, bl
->bl_radix
, bl
->bl_skip
);
241 if (blk
!= SWAPBLK_NONE
)
242 bl
->bl_free
-= count
;
248 * blist_free() - free up space in the block bitmap. Return the base
249 * of a contiguous region. Panic if an inconsistancy is
254 blist_free(blist_t bl
, swblk_t blkno
, swblk_t count
)
257 if (bl
->bl_radix
== BLIST_BMAP_RADIX
)
258 blst_leaf_free(bl
->bl_root
, blkno
, count
);
260 blst_meta_free(bl
->bl_root
, blkno
, count
, bl
->bl_radix
, bl
->bl_skip
, 0);
261 bl
->bl_free
+= count
;
266 * blist_resize() - resize an existing radix tree to handle the
267 * specified number of blocks. This will reallocate
268 * the tree and transfer the previous bitmap to the new
269 * one. When extending the tree you can specify whether
270 * the new blocks are to left allocated or freed.
274 blist_resize(blist_t
*pbl
, swblk_t count
, int freenew
)
276 blist_t newbl
= blist_create(count
);
280 if (count
> save
->bl_blocks
)
281 count
= save
->bl_blocks
;
282 blst_copy(save
->bl_root
, 0, save
->bl_radix
, save
->bl_skip
, newbl
, count
);
285 * If resizing upwards, should we free the new space or not?
287 if (freenew
&& count
< newbl
->bl_blocks
) {
288 blist_free(newbl
, count
, newbl
->bl_blocks
- count
);
296 * blist_print() - dump radix tree
300 blist_print(blist_t bl
)
302 kprintf("BLIST {\n");
303 blst_radix_print(bl
->bl_root
, 0, bl
->bl_radix
, bl
->bl_skip
, 4);
309 /************************************************************************
310 * ALLOCATION SUPPORT FUNCTIONS *
311 ************************************************************************
313 * These support functions do all the actual work. They may seem
314 * rather longish, but that's because I've commented them up. The
315 * actual code is straight forward.
320 * blist_leaf_alloc() - allocate at a leaf in the radix tree (a bitmap).
322 * This is the core of the allocator and is optimized for the 1 block
323 * and the BLIST_BMAP_RADIX block allocation cases. Other cases are
324 * somewhat slower. The 1 block allocation case is log2 and extremely
329 blst_leaf_alloc(blmeta_t
*scan
, swblk_t blk
, int count
)
331 u_swblk_t orig
= scan
->u
.bmu_bitmap
;
335 * Optimize bitmap all-allocated case. Also, count = 1
336 * case assumes at least 1 bit is free in the bitmap, so
337 * we have to take care of this case here.
339 scan
->bm_bighint
= 0;
340 return(SWAPBLK_NONE
);
344 * Optimized code to allocate one bit out of the bitmap
347 int j
= BLIST_BMAP_RADIX
/2;
350 mask
= (u_swblk_t
)-1 >> (BLIST_BMAP_RADIX
/2);
353 if ((orig
& mask
) == 0) {
360 scan
->u
.bmu_bitmap
&= ~(1 << r
);
363 if (count
<= BLIST_BMAP_RADIX
) {
365 * non-optimized code to allocate N bits out of the bitmap.
366 * The more bits, the faster the code runs. It will run
367 * the slowest allocating 2 bits, but since there aren't any
368 * memory ops in the core loop (or shouldn't be, anyway),
369 * you probably won't notice the difference.
372 int n
= BLIST_BMAP_RADIX
- count
;
375 mask
= (u_swblk_t
)-1 >> n
;
377 for (j
= 0; j
<= n
; ++j
) {
378 if ((orig
& mask
) == mask
) {
379 scan
->u
.bmu_bitmap
&= ~mask
;
386 * We couldn't allocate count in this subtree, update bighint.
388 scan
->bm_bighint
= count
- 1;
389 return(SWAPBLK_NONE
);
393 * blist_meta_alloc() - allocate at a meta in the radix tree.
395 * Attempt to allocate at a meta node. If we can't, we update
396 * bighint and return a failure. Updating bighint optimize future
397 * calls that hit this node. We have to check for our collapse cases
398 * and we have a few optimizations strewn in as well.
402 blst_meta_alloc(blmeta_t
*scan
, swblk_t blk
, swblk_t count
,
403 int64_t radix
, int skip
)
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
);
417 * note: radix may exceed 32 bits until first division.
419 if (scan
->u
.bmu_avail
== radix
) {
420 radix
/= BLIST_META_RADIX
;
423 * ALL-FREE special case, initialize uninitialize
426 for (i
= 1; i
<= skip
; i
+= next_skip
) {
427 if (scan
[i
].bm_bighint
== (swblk_t
)-1)
429 if (next_skip
== 1) {
430 scan
[i
].u
.bmu_bitmap
= (u_swblk_t
)-1;
431 scan
[i
].bm_bighint
= BLIST_BMAP_RADIX
;
433 scan
[i
].bm_bighint
= (swblk_t
)radix
;
434 scan
[i
].u
.bmu_avail
= (swblk_t
)radix
;
438 radix
/= BLIST_META_RADIX
;
441 for (i
= 1; i
<= skip
; i
+= next_skip
) {
442 if (count
<= scan
[i
].bm_bighint
) {
444 * count fits in object
447 if (next_skip
== 1) {
448 r
= blst_leaf_alloc(&scan
[i
], blk
, count
);
450 r
= blst_meta_alloc(&scan
[i
], blk
, count
, radix
, next_skip
- 1);
452 if (r
!= SWAPBLK_NONE
) {
453 scan
->u
.bmu_avail
-= count
;
454 if (scan
->bm_bighint
> scan
->u
.bmu_avail
)
455 scan
->bm_bighint
= scan
->u
.bmu_avail
;
458 } else if (scan
[i
].bm_bighint
== (swblk_t
)-1) {
463 } else if (count
> (swblk_t
)radix
) {
465 * count does not fit in object even if it were
468 panic("blist_meta_alloc: allocation too large");
470 blk
+= (swblk_t
)radix
;
474 * We couldn't allocate count in this subtree, update bighint.
476 if (scan
->bm_bighint
>= count
)
477 scan
->bm_bighint
= count
- 1;
478 return(SWAPBLK_NONE
);
482 * BLST_LEAF_FREE() - free allocated block from leaf bitmap
487 blst_leaf_free(blmeta_t
*scan
, swblk_t blk
, int count
)
490 * free some data in this bitmap
493 * 0000111111111110000
497 int n
= blk
& (BLIST_BMAP_RADIX
- 1);
500 mask
= ((u_swblk_t
)-1 << n
) &
501 ((u_swblk_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
528 blst_meta_free(blmeta_t
*scan
, swblk_t freeBlk
, swblk_t count
,
529 int64_t radix
, int skip
, swblk_t blk
)
532 int next_skip
= ((u_int
)skip
/ BLIST_META_RADIX
);
535 kprintf("FREE (%x,%d) FROM (%x,%lld)\n",
537 blk
, (long long)radix
542 * NOTE: radix may exceed 32 bits until first division.
544 if (scan
->u
.bmu_avail
== 0) {
546 * ALL-ALLOCATED special case, with possible
547 * shortcut to ALL-FREE special case.
549 scan
->u
.bmu_avail
= count
;
550 scan
->bm_bighint
= count
;
552 if (count
!= radix
) {
553 for (i
= 1; i
<= skip
; i
+= next_skip
) {
554 if (scan
[i
].bm_bighint
== (swblk_t
)-1)
556 scan
[i
].bm_bighint
= 0;
557 if (next_skip
== 1) {
558 scan
[i
].u
.bmu_bitmap
= 0;
560 scan
[i
].u
.bmu_avail
= 0;
566 scan
->u
.bmu_avail
+= count
;
567 /* scan->bm_bighint = radix; */
571 * ALL-FREE special case.
574 if (scan
->u
.bmu_avail
== radix
)
576 if (scan
->u
.bmu_avail
> radix
)
577 panic("blst_meta_free: freeing already free blocks (%d) %d/%lld", count
, scan
->u
.bmu_avail
, (long long)radix
);
580 * Break the free down into its components
583 radix
/= BLIST_META_RADIX
;
585 i
= (freeBlk
- blk
) / (swblk_t
)radix
;
586 blk
+= i
* (swblk_t
)radix
;
587 i
= i
* next_skip
+ 1;
589 while (i
<= skip
&& blk
< freeBlk
+ count
) {
592 v
= blk
+ (swblk_t
)radix
- freeBlk
;
596 if (scan
->bm_bighint
== (swblk_t
)-1)
597 panic("blst_meta_free: freeing unexpected range");
599 if (next_skip
== 1) {
600 blst_leaf_free(&scan
[i
], freeBlk
, v
);
602 blst_meta_free(&scan
[i
], freeBlk
, v
, radix
, next_skip
- 1, blk
);
604 if (scan
->bm_bighint
< scan
[i
].bm_bighint
)
605 scan
->bm_bighint
= scan
[i
].bm_bighint
;
608 blk
+= (swblk_t
)radix
;
614 * BLIST_RADIX_COPY() - copy one radix tree to another
616 * Locates free space in the source tree and frees it in the destination
617 * tree. The space may not already be free in the destination.
621 blst_copy(blmeta_t
*scan
, swblk_t blk
, int64_t radix
,
622 swblk_t skip
, blist_t dest
, swblk_t count
)
631 if (radix
== BLIST_BMAP_RADIX
) {
632 u_swblk_t v
= scan
->u
.bmu_bitmap
;
634 if (v
== (u_swblk_t
)-1) {
635 blist_free(dest
, blk
, count
);
639 for (i
= 0; i
< BLIST_BMAP_RADIX
&& i
< count
; ++i
) {
641 blist_free(dest
, blk
+ i
, 1);
651 if (scan
->u
.bmu_avail
== 0) {
653 * Source all allocated, leave dest allocated
657 if (scan
->u
.bmu_avail
== radix
) {
659 * Source all free, free entire dest
662 blist_free(dest
, blk
, count
);
664 blist_free(dest
, blk
, (swblk_t
)radix
);
669 radix
/= BLIST_META_RADIX
;
670 next_skip
= ((u_int
)skip
/ BLIST_META_RADIX
);
672 for (i
= 1; count
&& i
<= skip
; i
+= next_skip
) {
673 if (scan
[i
].bm_bighint
== (swblk_t
)-1)
676 if (count
>= (swblk_t
)radix
) {
685 count
-= (swblk_t
)radix
;
699 blk
+= (swblk_t
)radix
;
704 * BLST_RADIX_INIT() - initialize radix tree
706 * Initialize our meta structures and bitmaps and calculate the exact
707 * amount of space required to manage 'count' blocks - this space may
708 * be considerably less then the calculated radix due to the large
709 * RADIX values we use.
713 blst_radix_init(blmeta_t
*scan
, int64_t radix
, int skip
, swblk_t count
)
717 swblk_t memindex
= 0;
723 if (radix
== BLIST_BMAP_RADIX
) {
725 scan
->bm_bighint
= 0;
726 scan
->u
.bmu_bitmap
= 0;
732 * Meta node. If allocating the entire object we can special
733 * case it. However, we need to figure out how much memory
734 * is required to manage 'count' blocks, so we continue on anyway.
738 scan
->bm_bighint
= 0;
739 scan
->u
.bmu_avail
= 0;
742 radix
/= BLIST_META_RADIX
;
743 next_skip
= ((u_int
)skip
/ BLIST_META_RADIX
);
745 for (i
= 1; i
<= skip
; i
+= next_skip
) {
746 if (count
>= (swblk_t
)radix
) {
748 * Allocate the entire object
750 memindex
= i
+ blst_radix_init(
751 ((scan
) ? &scan
[i
] : NULL
),
756 count
-= (swblk_t
)radix
;
757 } else if (count
> 0) {
759 * Allocate a partial object
761 memindex
= i
+ blst_radix_init(
762 ((scan
) ? &scan
[i
] : NULL
),
770 * Add terminator and break out
773 scan
[i
].bm_bighint
= (swblk_t
)-1;
785 blst_radix_print(blmeta_t
*scan
, swblk_t blk
, int64_t radix
, int skip
, int tab
)
791 if (radix
== BLIST_BMAP_RADIX
) {
793 "%*.*s(%04x,%lld): bitmap %08x big=%d\n",
795 blk
, (long long)radix
,
802 if (scan
->u
.bmu_avail
== 0) {
804 "%*.*s(%04x,%lld) ALL ALLOCATED\n",
811 if (scan
->u
.bmu_avail
== radix
) {
813 "%*.*s(%04x,%lld) ALL FREE\n",
822 "%*.*s(%04x,%lld): subtree (%d/%lld) big=%d {\n",
824 blk
, (long long)radix
,
830 radix
/= BLIST_META_RADIX
;
831 next_skip
= ((u_int
)skip
/ BLIST_META_RADIX
);
834 for (i
= 1; i
<= skip
; i
+= next_skip
) {
835 if (scan
[i
].bm_bighint
== (swblk_t
)-1) {
837 "%*.*s(%04x,%lld): Terminator\n",
839 blk
, (long long)radix
851 blk
+= (swblk_t
)radix
;
866 main(int ac
, char **av
)
872 for (i
= 1; i
< ac
; ++i
) {
873 const char *ptr
= av
[i
];
875 size
= strtol(ptr
, NULL
, 0);
879 fprintf(stderr
, "Bad option: %s\n", ptr
- 2);
882 bl
= blist_create(size
);
883 blist_free(bl
, 0, size
);
891 kprintf("%d/%d/%lld> ",
892 bl
->bl_free
, size
, (long long)bl
->bl_radix
);
894 if (fgets(buf
, sizeof(buf
), stdin
) == NULL
)
898 if (sscanf(buf
+ 1, "%d", &count
) == 1) {
899 blist_resize(&bl
, count
, 1);
907 if (sscanf(buf
+ 1, "%d", &count
) == 1) {
908 swblk_t blk
= blist_alloc(bl
, count
);
909 kprintf(" R=%04x\n", blk
);
915 if (sscanf(buf
+ 1, "%x %d", &da
, &count
) == 2) {
916 blist_free(bl
, da
, count
);
940 panic(const char *ctl
, ...)
945 vfprintf(stderr
, ctl
, va
);
946 fprintf(stderr
, "\n");