FreeBSD/Linux Kernel Cross Reference
sys/kern/subr_blist.c
1 /*-
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
5 * are met:
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.
14 *
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.
26 */
27 /*
28 * BLIST.C - Bitmap allocator/deallocator, using a radix tree with hinting
29 *
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.
34 *
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
42 * update the hint.
43 *
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.
49 *
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.
54 *
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
58 * (vm/swap_pager.c).
59 *
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.
70 *
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
77 * ranges.
78 *
79 * This code can be compiled stand-alone for debugging.
80 */
81
82 #include <sys/cdefs.h>
83 __FBSDID("$FreeBSD$");
84
85 #ifdef _KERNEL
86
87 #include <sys/param.h>
88 #include <sys/systm.h>
89 #include <sys/lock.h>
90 #include <sys/kernel.h>
91 #include <sys/blist.h>
92 #include <sys/malloc.h>
93 #include <sys/proc.h>
94 #include <sys/mutex.h>
95
96 #else
97
98 #ifndef BLIST_NO_DEBUG
99 #define BLIST_DEBUG
100 #endif
101
102 #include <sys/types.h>
103 #include <sys/malloc.h>
104 #include <stdio.h>
105 #include <string.h>
106 #include <stdlib.h>
107 #include <stdarg.h>
108 #include <stdbool.h>
109
110 #define bitcount64(x) __bitcount64((uint64_t)(x))
111 #define malloc(a,b,c) calloc(a, 1)
112 #define free(a,b) free(a)
113
114 #include <sys/blist.h>
115
116 void panic(const char *ctl, ...);
117
118 #endif
119
120 /*
121 * static support functions
122 */
123
124 static daddr_t blst_leaf_alloc(blmeta_t *scan, daddr_t blk, int count,
125 daddr_t cursor);
126 static daddr_t blst_meta_alloc(blmeta_t *scan, daddr_t blk, daddr_t count,
127 daddr_t radix, daddr_t skip, daddr_t cursor);
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, daddr_t 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 daddr_t blst_leaf_fill(blmeta_t *scan, daddr_t blk, int count);
134 static daddr_t blst_meta_fill(blmeta_t *scan, daddr_t allocBlk, daddr_t count,
135 daddr_t radix, daddr_t skip, daddr_t blk);
136 static daddr_t blst_radix_init(blmeta_t *scan, daddr_t radix, daddr_t skip,
137 daddr_t count);
138 #ifndef _KERNEL
139 static void blst_radix_print(blmeta_t *scan, daddr_t blk, daddr_t radix,
140 daddr_t skip, int tab);
141 #endif
142
143 #ifdef _KERNEL
144 static MALLOC_DEFINE(M_SWAP, "SWAP", "Swap space");
145 #endif
146
147 /*
148 * blist_create() - create a blist capable of handling up to the specified
149 * number of blocks
150 *
151 * blocks - must be greater than 0
152 * flags - malloc flags
153 *
154 * The smallest blist consists of a single leaf node capable of
155 * managing BLIST_BMAP_RADIX blocks.
156 */
157
158 blist_t
159 blist_create(daddr_t blocks, int flags)
160 {
161 blist_t bl;
162 daddr_t nodes, radix, skip;
163
164 /*
165 * Calculate radix and skip field used for scanning.
166 */
167 radix = BLIST_BMAP_RADIX;
168 skip = 0;
169 while (radix < blocks) {
170 radix *= BLIST_META_RADIX;
171 skip = (skip + 1) * BLIST_META_RADIX;
172 }
173 nodes = 1 + blst_radix_init(NULL, radix, skip, blocks);
174
175 bl = malloc(sizeof(struct blist), M_SWAP, flags);
176 if (bl == NULL)
177 return (NULL);
178
179 bl->bl_blocks = blocks;
180 bl->bl_radix = radix;
181 bl->bl_skip = skip;
182 bl->bl_cursor = 0;
183 bl->bl_root = malloc(nodes * sizeof(blmeta_t), M_SWAP, flags);
184 if (bl->bl_root == NULL) {
185 free(bl, M_SWAP);
186 return (NULL);
187 }
188 blst_radix_init(bl->bl_root, radix, skip, blocks);
189
190 #if defined(BLIST_DEBUG)
191 printf(
192 "BLIST representing %lld blocks (%lld MB of swap)"
193 ", requiring %lldK of ram\n",
194 (long long)bl->bl_blocks,
195 (long long)bl->bl_blocks * 4 / 1024,
196 (long long)(nodes * sizeof(blmeta_t) + 1023) / 1024
197 );
198 printf("BLIST raw radix tree contains %lld records\n",
199 (long long)nodes);
200 #endif
201
202 return (bl);
203 }
204
205 void
206 blist_destroy(blist_t bl)
207 {
208 free(bl->bl_root, M_SWAP);
209 free(bl, M_SWAP);
210 }
211
212 /*
213 * blist_alloc() - reserve space in the block bitmap. Return the base
214 * of a contiguous region or SWAPBLK_NONE if space could
215 * not be allocated.
216 */
217
218 daddr_t
219 blist_alloc(blist_t bl, daddr_t count)
220 {
221 daddr_t blk;
222
223 /*
224 * This loop iterates at most twice. An allocation failure in the
225 * first iteration leads to a second iteration only if the cursor was
226 * non-zero. When the cursor is zero, an allocation failure will
227 * reduce the hint, stopping further iterations.
228 */
229 while (count <= bl->bl_root->bm_bighint) {
230 if (bl->bl_radix == BLIST_BMAP_RADIX)
231 blk = blst_leaf_alloc(bl->bl_root, 0, count,
232 bl->bl_cursor);
233 else
234 blk = blst_meta_alloc(bl->bl_root, 0, count,
235 bl->bl_radix, bl->bl_skip, bl->bl_cursor);
236 if (blk != SWAPBLK_NONE) {
237 bl->bl_cursor = blk + count;
238 return (blk);
239 } else if (bl->bl_cursor != 0)
240 bl->bl_cursor = 0;
241 }
242 return (SWAPBLK_NONE);
243 }
244
245 /*
246 * blist_avail() - return the number of free blocks.
247 */
248
249 daddr_t
250 blist_avail(blist_t bl)
251 {
252
253 if (bl->bl_radix == BLIST_BMAP_RADIX)
254 return (bitcount64(bl->bl_root->u.bmu_bitmap));
255 else
256 return (bl->bl_root->u.bmu_avail);
257 }
258
259 /*
260 * blist_free() - free up space in the block bitmap. Return the base
261 * of a contiguous region. Panic if an inconsistancy is
262 * found.
263 */
264
265 void
266 blist_free(blist_t bl, daddr_t blkno, daddr_t count)
267 {
268 if (bl) {
269 if (bl->bl_radix == BLIST_BMAP_RADIX)
270 blst_leaf_free(bl->bl_root, blkno, count);
271 else
272 blst_meta_free(bl->bl_root, blkno, count,
273 bl->bl_radix, bl->bl_skip, 0);
274 }
275 }
276
277 /*
278 * blist_fill() - mark a region in the block bitmap as off-limits
279 * to the allocator (i.e. allocate it), ignoring any
280 * existing allocations. Return the number of blocks
281 * actually filled that were free before the call.
282 */
283
284 daddr_t
285 blist_fill(blist_t bl, daddr_t blkno, daddr_t count)
286 {
287 daddr_t filled;
288
289 if (bl) {
290 if (bl->bl_radix == BLIST_BMAP_RADIX)
291 filled = blst_leaf_fill(bl->bl_root, blkno, count);
292 else
293 filled = blst_meta_fill(bl->bl_root, blkno, count,
294 bl->bl_radix, bl->bl_skip, 0);
295 return (filled);
296 }
297 return (0);
298 }
299
300 /*
301 * blist_resize() - resize an existing radix tree to handle the
302 * specified number of blocks. This will reallocate
303 * the tree and transfer the previous bitmap to the new
304 * one. When extending the tree you can specify whether
305 * the new blocks are to left allocated or freed.
306 */
307
308 void
309 blist_resize(blist_t *pbl, daddr_t count, int freenew, int flags)
310 {
311 blist_t newbl = blist_create(count, flags);
312 blist_t save = *pbl;
313
314 *pbl = newbl;
315 if (count > save->bl_blocks)
316 count = save->bl_blocks;
317 blst_copy(save->bl_root, 0, save->bl_radix, save->bl_skip, newbl, count);
318
319 /*
320 * If resizing upwards, should we free the new space or not?
321 */
322 if (freenew && count < newbl->bl_blocks) {
323 blist_free(newbl, count, newbl->bl_blocks - count);
324 }
325 blist_destroy(save);
326 }
327
328 #ifdef BLIST_DEBUG
329
330 /*
331 * blist_print() - dump radix tree
332 */
333
334 void
335 blist_print(blist_t bl)
336 {
337 printf("BLIST {\n");
338 blst_radix_print(bl->bl_root, 0, bl->bl_radix, bl->bl_skip, 4);
339 printf("}\n");
340 }
341
342 #endif
343
344 /************************************************************************
345 * ALLOCATION SUPPORT FUNCTIONS *
346 ************************************************************************
347 *
348 * These support functions do all the actual work. They may seem
349 * rather longish, but that's because I've commented them up. The
350 * actual code is straight forward.
351 *
352 */
353
354 /*
355 * blist_leaf_alloc() - allocate at a leaf in the radix tree (a bitmap).
356 *
357 * This is the core of the allocator and is optimized for the
358 * BLIST_BMAP_RADIX block allocation case. Otherwise, execution
359 * time is proportional to log2(count) + log2(BLIST_BMAP_RADIX).
360 */
361
362 static daddr_t
363 blst_leaf_alloc(blmeta_t *scan, daddr_t blk, int count, daddr_t cursor)
364 {
365 u_daddr_t mask;
366 int count1, hi, lo, mid, num_shifts, range1, range_ext;
367
368 if (count == BLIST_BMAP_RADIX) {
369 /*
370 * Optimize allocation of BLIST_BMAP_RADIX bits. If this wasn't
371 * a special case, then forming the final value of 'mask' below
372 * would require special handling to avoid an invalid left shift
373 * when count equals the number of bits in mask.
374 */
375 if (~scan->u.bmu_bitmap != 0) {
376 scan->bm_bighint = BLIST_BMAP_RADIX - 1;
377 return (SWAPBLK_NONE);
378 }
379 if (cursor != blk)
380 return (SWAPBLK_NONE);
381 scan->u.bmu_bitmap = 0;
382 scan->bm_bighint = 0;
383 return (blk);
384 }
385 range1 = 0;
386 count1 = count - 1;
387 num_shifts = fls(count1);
388 mask = scan->u.bmu_bitmap;
389 while (mask != 0 && num_shifts > 0) {
390 /*
391 * If bit i is set in mask, then bits in [i, i+range1] are set
392 * in scan->u.bmu_bitmap. The value of range1 is equal to
393 * count1 >> num_shifts. Grow range and reduce num_shifts to 0,
394 * while preserving these invariants. The updates to mask leave
395 * fewer bits set, but each bit that remains set represents a
396 * longer string of consecutive bits set in scan->u.bmu_bitmap.
397 */
398 num_shifts--;
399 range_ext = range1 + ((count1 >> num_shifts) & 1);
400 mask &= mask >> range_ext;
401 range1 += range_ext;
402 }
403 if (mask == 0) {
404 /*
405 * Update bighint. There is no allocation bigger than range1
406 * available in this leaf.
407 */
408 scan->bm_bighint = range1;
409 return (SWAPBLK_NONE);
410 }
411
412 /*
413 * Discard any candidates that appear before the cursor.
414 */
415 lo = cursor - blk;
416 mask &= ~(u_daddr_t)0 << lo;
417
418 if (mask == 0)
419 return (SWAPBLK_NONE);
420
421 /*
422 * The least significant set bit in mask marks the start of the first
423 * available range of sufficient size. Clear all the bits but that one,
424 * and then perform a binary search to find its position.
425 */
426 mask &= -mask;
427 hi = BLIST_BMAP_RADIX - count1;
428 while (lo + 1 < hi) {
429 mid = (lo + hi) >> 1;
430 if ((mask >> mid) != 0)
431 lo = mid;
432 else
433 hi = mid;
434 }
435
436 /*
437 * Set in mask exactly the bits being allocated, and clear them from
438 * the set of available bits.
439 */
440 mask = (mask << count) - mask;
441 scan->u.bmu_bitmap &= ~mask;
442 return (blk + lo);
443 }
444
445 /*
446 * blist_meta_alloc() - allocate at a meta in the radix tree.
447 *
448 * Attempt to allocate at a meta node. If we can't, we update
449 * bighint and return a failure. Updating bighint optimize future
450 * calls that hit this node. We have to check for our collapse cases
451 * and we have a few optimizations strewn in as well.
452 */
453
454 static daddr_t
455 blst_meta_alloc(blmeta_t *scan, daddr_t blk, daddr_t count, daddr_t radix,
456 daddr_t skip, daddr_t cursor)
457 {
458 daddr_t i, next_skip, r;
459 int child;
460 bool scan_from_start;
461
462 if (scan->u.bmu_avail < count) {
463 /*
464 * The meta node's hint must be too large if the allocation
465 * exceeds the number of free blocks. Reduce the hint, and
466 * return failure.
467 */
468 scan->bm_bighint = scan->u.bmu_avail;
469 return (SWAPBLK_NONE);
470 }
471 next_skip = skip / BLIST_META_RADIX;
472
473 /*
474 * An ALL-FREE meta node requires special handling before allocating
475 * any of its blocks.
476 */
477 if (scan->u.bmu_avail == radix) {
478 radix /= BLIST_META_RADIX;
479
480 /*
481 * Reinitialize each of the meta node's children. An ALL-FREE
482 * meta node cannot have a terminator in any subtree.
483 */
484 for (i = 1; i <= skip; i += next_skip) {
485 if (next_skip == 1)
486 scan[i].u.bmu_bitmap = (u_daddr_t)-1;
487 else
488 scan[i].u.bmu_avail = radix;
489 scan[i].bm_bighint = radix;
490 }
491 } else {
492 radix /= BLIST_META_RADIX;
493 }
494
495 if (count > radix) {
496 /*
497 * The allocation exceeds the number of blocks that are
498 * managed by a subtree of this meta node.
499 */
500 panic("allocation too large");
501 }
502 scan_from_start = cursor == blk;
503 child = (cursor - blk) / radix;
504 blk += child * radix;
505 for (i = 1 + child * next_skip; i <= skip; i += next_skip) {
506 if (count <= scan[i].bm_bighint) {
507 /*
508 * The allocation might fit in the i'th subtree.
509 */
510 if (next_skip == 1) {
511 r = blst_leaf_alloc(&scan[i], blk, count,
512 cursor > blk ? cursor : blk);
513 } else {
514 r = blst_meta_alloc(&scan[i], blk, count,
515 radix, next_skip - 1, cursor > blk ?
516 cursor : blk);
517 }
518 if (r != SWAPBLK_NONE) {
519 scan->u.bmu_avail -= count;
520 return (r);
521 }
522 } else if (scan[i].bm_bighint == (daddr_t)-1) {
523 /*
524 * Terminator
525 */
526 break;
527 }
528 blk += radix;
529 }
530
531 /*
532 * We couldn't allocate count in this subtree, update bighint.
533 */
534 if (scan_from_start && scan->bm_bighint >= count)
535 scan->bm_bighint = count - 1;
536
537 return (SWAPBLK_NONE);
538 }
539
540 /*
541 * BLST_LEAF_FREE() - free allocated block from leaf bitmap
542 *
543 */
544
545 static void
546 blst_leaf_free(
547 blmeta_t *scan,
548 daddr_t blk,
549 int count
550 ) {
551 /*
552 * free some data in this bitmap
553 *
554 * e.g.
555 * 0000111111111110000
556 * \_________/\__/
557 * v n
558 */
559 int n = blk & (BLIST_BMAP_RADIX - 1);
560 u_daddr_t mask;
561
562 mask = ((u_daddr_t)-1 << n) &
563 ((u_daddr_t)-1 >> (BLIST_BMAP_RADIX - count - n));
564
565 if (scan->u.bmu_bitmap & mask)
566 panic("blst_radix_free: freeing free block");
567 scan->u.bmu_bitmap |= mask;
568
569 /*
570 * We could probably do a better job here. We are required to make
571 * bighint at least as large as the biggest contiguous block of
572 * data. If we just shoehorn it, a little extra overhead will
573 * be incured on the next allocation (but only that one typically).
574 */
575 scan->bm_bighint = BLIST_BMAP_RADIX;
576 }
577
578 /*
579 * BLST_META_FREE() - free allocated blocks from radix tree meta info
580 *
581 * This support routine frees a range of blocks from the bitmap.
582 * The range must be entirely enclosed by this radix node. If a
583 * meta node, we break the range down recursively to free blocks
584 * in subnodes (which means that this code can free an arbitrary
585 * range whereas the allocation code cannot allocate an arbitrary
586 * range).
587 */
588
589 static void
590 blst_meta_free(blmeta_t *scan, daddr_t freeBlk, daddr_t count, daddr_t radix,
591 daddr_t skip, daddr_t blk)
592 {
593 daddr_t i, next_skip, v;
594 int child;
595
596 #if 0
597 printf("free (%llx,%lld) FROM (%llx,%lld)\n",
598 (long long)freeBlk, (long long)count,
599 (long long)blk, (long long)radix
600 );
601 #endif
602 next_skip = skip / BLIST_META_RADIX;
603
604 if (scan->u.bmu_avail == 0) {
605 /*
606 * ALL-ALLOCATED special case, with possible
607 * shortcut to ALL-FREE special case.
608 */
609 scan->u.bmu_avail = count;
610 scan->bm_bighint = count;
611
612 if (count != radix) {
613 for (i = 1; i <= skip; i += next_skip) {
614 if (scan[i].bm_bighint == (daddr_t)-1)
615 break;
616 scan[i].bm_bighint = 0;
617 if (next_skip == 1) {
618 scan[i].u.bmu_bitmap = 0;
619 } else {
620 scan[i].u.bmu_avail = 0;
621 }
622 }
623 /* fall through */
624 }
625 } else {
626 scan->u.bmu_avail += count;
627 /* scan->bm_bighint = radix; */
628 }
629
630 /*
631 * ALL-FREE special case.
632 */
633
634 if (scan->u.bmu_avail == radix)
635 return;
636 if (scan->u.bmu_avail > radix)
637 panic("blst_meta_free: freeing already free blocks (%lld) %lld/%lld",
638 (long long)count, (long long)scan->u.bmu_avail,
639 (long long)radix);
640
641 /*
642 * Break the free down into its components
643 */
644
645 radix /= BLIST_META_RADIX;
646
647 child = (freeBlk - blk) / radix;
648 blk += child * radix;
649 i = 1 + child * next_skip;
650 while (i <= skip && blk < freeBlk + count) {
651 v = blk + radix - freeBlk;
652 if (v > count)
653 v = count;
654
655 if (scan->bm_bighint == (daddr_t)-1)
656 panic("blst_meta_free: freeing unexpected range");
657
658 if (next_skip == 1) {
659 blst_leaf_free(&scan[i], freeBlk, v);
660 } else {
661 blst_meta_free(&scan[i], freeBlk, v, radix, next_skip - 1, blk);
662 }
663 if (scan->bm_bighint < scan[i].bm_bighint)
664 scan->bm_bighint = scan[i].bm_bighint;
665 count -= v;
666 freeBlk += v;
667 blk += radix;
668 i += next_skip;
669 }
670 }
671
672 /*
673 * BLIST_RADIX_COPY() - copy one radix tree to another
674 *
675 * Locates free space in the source tree and frees it in the destination
676 * tree. The space may not already be free in the destination.
677 */
678
679 static void blst_copy(
680 blmeta_t *scan,
681 daddr_t blk,
682 daddr_t radix,
683 daddr_t skip,
684 blist_t dest,
685 daddr_t count
686 ) {
687 daddr_t i, next_skip;
688
689 /*
690 * Leaf node
691 */
692
693 if (radix == BLIST_BMAP_RADIX) {
694 u_daddr_t v = scan->u.bmu_bitmap;
695
696 if (v == (u_daddr_t)-1) {
697 blist_free(dest, blk, count);
698 } else if (v != 0) {
699 int i;
700
701 for (i = 0; i < BLIST_BMAP_RADIX && i < count; ++i) {
702 if (v & ((u_daddr_t)1 << i))
703 blist_free(dest, blk + i, 1);
704 }
705 }
706 return;
707 }
708
709 /*
710 * Meta node
711 */
712
713 if (scan->u.bmu_avail == 0) {
714 /*
715 * Source all allocated, leave dest allocated
716 */
717 return;
718 }
719 if (scan->u.bmu_avail == radix) {
720 /*
721 * Source all free, free entire dest
722 */
723 if (count < radix)
724 blist_free(dest, blk, count);
725 else
726 blist_free(dest, blk, radix);
727 return;
728 }
729
730
731 radix /= BLIST_META_RADIX;
732 next_skip = skip / BLIST_META_RADIX;
733
734 for (i = 1; count && i <= skip; i += next_skip) {
735 if (scan[i].bm_bighint == (daddr_t)-1)
736 break;
737
738 if (count >= radix) {
739 blst_copy(
740 &scan[i],
741 blk,
742 radix,
743 next_skip - 1,
744 dest,
745 radix
746 );
747 count -= radix;
748 } else {
749 if (count) {
750 blst_copy(
751 &scan[i],
752 blk,
753 radix,
754 next_skip - 1,
755 dest,
756 count
757 );
758 }
759 count = 0;
760 }
761 blk += radix;
762 }
763 }
764
765 /*
766 * BLST_LEAF_FILL() - allocate specific blocks in leaf bitmap
767 *
768 * This routine allocates all blocks in the specified range
769 * regardless of any existing allocations in that range. Returns
770 * the number of blocks allocated by the call.
771 */
772
773 static daddr_t
774 blst_leaf_fill(blmeta_t *scan, daddr_t blk, int count)
775 {
776 int n = blk & (BLIST_BMAP_RADIX - 1);
777 daddr_t nblks;
778 u_daddr_t mask;
779
780 mask = ((u_daddr_t)-1 << n) &
781 ((u_daddr_t)-1 >> (BLIST_BMAP_RADIX - count - n));
782
783 /* Count the number of blocks that we are allocating. */
784 nblks = bitcount64(scan->u.bmu_bitmap & mask);
785
786 scan->u.bmu_bitmap &= ~mask;
787 return (nblks);
788 }
789
790 /*
791 * BLIST_META_FILL() - allocate specific blocks at a meta node
792 *
793 * This routine allocates the specified range of blocks,
794 * regardless of any existing allocations in the range. The
795 * range must be within the extent of this node. Returns the
796 * number of blocks allocated by the call.
797 */
798 static daddr_t
799 blst_meta_fill(blmeta_t *scan, daddr_t allocBlk, daddr_t count, daddr_t radix,
800 daddr_t skip, daddr_t blk)
801 {
802 daddr_t i, nblks, next_skip, v;
803 int child;
804
805 if (count > radix) {
806 /*
807 * The allocation exceeds the number of blocks that are
808 * managed by this meta node.
809 */
810 panic("allocation too large");
811 }
812 if (count == radix || scan->u.bmu_avail == 0) {
813 /*
814 * ALL-ALLOCATED special case
815 */
816 nblks = scan->u.bmu_avail;
817 scan->u.bmu_avail = 0;
818 scan->bm_bighint = 0;
819 return nblks;
820 }
821 next_skip = skip / BLIST_META_RADIX;
822
823 /*
824 * An ALL-FREE meta node requires special handling before allocating
825 * any of its blocks.
826 */
827 if (scan->u.bmu_avail == radix) {
828 radix /= BLIST_META_RADIX;
829
830 /*
831 * Reinitialize each of the meta node's children. An ALL-FREE
832 * meta node cannot have a terminator in any subtree.
833 */
834 for (i = 1; i <= skip; i += next_skip) {
835 if (next_skip == 1) {
836 scan[i].u.bmu_bitmap = (u_daddr_t)-1;
837 scan[i].bm_bighint = BLIST_BMAP_RADIX;
838 } else {
839 scan[i].bm_bighint = radix;
840 scan[i].u.bmu_avail = radix;
841 }
842 }
843 } else {
844 radix /= BLIST_META_RADIX;
845 }
846
847 nblks = 0;
848 child = (allocBlk - blk) / radix;
849 blk += child * radix;
850 i = 1 + child * next_skip;
851 while (i <= skip && blk < allocBlk + count) {
852 v = blk + radix - allocBlk;
853 if (v > count)
854 v = count;
855
856 if (scan->bm_bighint == (daddr_t)-1)
857 panic("blst_meta_fill: filling unexpected range");
858
859 if (next_skip == 1) {
860 nblks += blst_leaf_fill(&scan[i], allocBlk, v);
861 } else {
862 nblks += blst_meta_fill(&scan[i], allocBlk, v,
863 radix, next_skip - 1, blk);
864 }
865 count -= v;
866 allocBlk += v;
867 blk += radix;
868 i += next_skip;
869 }
870 scan->u.bmu_avail -= nblks;
871 return nblks;
872 }
873
874 /*
875 * BLST_RADIX_INIT() - initialize radix tree
876 *
877 * Initialize our meta structures and bitmaps and calculate the exact
878 * amount of space required to manage 'count' blocks - this space may
879 * be considerably less than the calculated radix due to the large
880 * RADIX values we use.
881 */
882
883 static daddr_t
884 blst_radix_init(blmeta_t *scan, daddr_t radix, daddr_t skip, daddr_t count)
885 {
886 daddr_t i, memindex, next_skip;
887
888 memindex = 0;
889
890 /*
891 * Leaf node
892 */
893
894 if (radix == BLIST_BMAP_RADIX) {
895 if (scan) {
896 scan->bm_bighint = 0;
897 scan->u.bmu_bitmap = 0;
898 }
899 return(memindex);
900 }
901
902 /*
903 * Meta node. If allocating the entire object we can special
904 * case it. However, we need to figure out how much memory
905 * is required to manage 'count' blocks, so we continue on anyway.
906 */
907
908 if (scan) {
909 scan->bm_bighint = 0;
910 scan->u.bmu_avail = 0;
911 }
912
913 radix /= BLIST_META_RADIX;
914 next_skip = skip / BLIST_META_RADIX;
915
916 for (i = 1; i <= skip; i += next_skip) {
917 if (count >= radix) {
918 /*
919 * Allocate the entire object
920 */
921 memindex = i + blst_radix_init(
922 ((scan) ? &scan[i] : NULL),
923 radix,
924 next_skip - 1,
925 radix
926 );
927 count -= radix;
928 } else if (count > 0) {
929 /*
930 * Allocate a partial object
931 */
932 memindex = i + blst_radix_init(
933 ((scan) ? &scan[i] : NULL),
934 radix,
935 next_skip - 1,
936 count
937 );
938 count = 0;
939 } else {
940 /*
941 * Add terminator and break out
942 */
943 if (scan)
944 scan[i].bm_bighint = (daddr_t)-1;
945 break;
946 }
947 }
948 if (memindex < i)
949 memindex = i;
950 return(memindex);
951 }
952
953 #ifdef BLIST_DEBUG
954
955 static void
956 blst_radix_print(blmeta_t *scan, daddr_t blk, daddr_t radix, daddr_t skip,
957 int tab)
958 {
959 daddr_t i, next_skip;
960
961 if (radix == BLIST_BMAP_RADIX) {
962 printf(
963 "%*.*s(%08llx,%lld): bitmap %016llx big=%lld\n",
964 tab, tab, "",
965 (long long)blk, (long long)radix,
966 (long long)scan->u.bmu_bitmap,
967 (long long)scan->bm_bighint
968 );
969 return;
970 }
971
972 if (scan->u.bmu_avail == 0) {
973 printf(
974 "%*.*s(%08llx,%lld) ALL ALLOCATED\n",
975 tab, tab, "",
976 (long long)blk,
977 (long long)radix
978 );
979 return;
980 }
981 if (scan->u.bmu_avail == radix) {
982 printf(
983 "%*.*s(%08llx,%lld) ALL FREE\n",
984 tab, tab, "",
985 (long long)blk,
986 (long long)radix
987 );
988 return;
989 }
990
991 printf(
992 "%*.*s(%08llx,%lld): subtree (%lld/%lld) big=%lld {\n",
993 tab, tab, "",
994 (long long)blk, (long long)radix,
995 (long long)scan->u.bmu_avail,
996 (long long)radix,
997 (long long)scan->bm_bighint
998 );
999
1000 radix /= BLIST_META_RADIX;
1001 next_skip = skip / BLIST_META_RADIX;
1002 tab += 4;
1003
1004 for (i = 1; i <= skip; i += next_skip) {
1005 if (scan[i].bm_bighint == (daddr_t)-1) {
1006 printf(
1007 "%*.*s(%08llx,%lld): Terminator\n",
1008 tab, tab, "",
1009 (long long)blk, (long long)radix
1010 );
1011 break;
1012 }
1013 blst_radix_print(
1014 &scan[i],
1015 blk,
1016 radix,
1017 next_skip - 1,
1018 tab
1019 );
1020 blk += radix;
1021 }
1022 tab -= 4;
1023
1024 printf(
1025 "%*.*s}\n",
1026 tab, tab, ""
1027 );
1028 }
1029
1030 #endif
1031
1032 #ifdef BLIST_DEBUG
1033
1034 int
1035 main(int ac, char **av)
1036 {
1037 int size = 1024;
1038 int i;
1039 blist_t bl;
1040
1041 for (i = 1; i < ac; ++i) {
1042 const char *ptr = av[i];
1043 if (*ptr != '-') {
1044 size = strtol(ptr, NULL, 0);
1045 continue;
1046 }
1047 ptr += 2;
1048 fprintf(stderr, "Bad option: %s\n", ptr - 2);
1049 exit(1);
1050 }
1051 bl = blist_create(size, M_WAITOK);
1052 blist_free(bl, 0, size);
1053
1054 for (;;) {
1055 char buf[1024];
1056 long long da = 0;
1057 long long count = 0;
1058
1059 printf("%lld/%lld/%lld> ", (long long)blist_avail(bl),
1060 (long long)size, (long long)bl->bl_radix);
1061 fflush(stdout);
1062 if (fgets(buf, sizeof(buf), stdin) == NULL)
1063 break;
1064 switch(buf[0]) {
1065 case 'r':
1066 if (sscanf(buf + 1, "%lld", &count) == 1) {
1067 blist_resize(&bl, count, 1, M_WAITOK);
1068 } else {
1069 printf("?\n");
1070 }
1071 case 'p':
1072 blist_print(bl);
1073 break;
1074 case 'a':
1075 if (sscanf(buf + 1, "%lld", &count) == 1) {
1076 daddr_t blk = blist_alloc(bl, count);
1077 printf(" R=%08llx\n", (long long)blk);
1078 } else {
1079 printf("?\n");
1080 }
1081 break;
1082 case 'f':
1083 if (sscanf(buf + 1, "%llx %lld", &da, &count) == 2) {
1084 blist_free(bl, da, count);
1085 } else {
1086 printf("?\n");
1087 }
1088 break;
1089 case 'l':
1090 if (sscanf(buf + 1, "%llx %lld", &da, &count) == 2) {
1091 printf(" n=%jd\n",
1092 (intmax_t)blist_fill(bl, da, count));
1093 } else {
1094 printf("?\n");
1095 }
1096 break;
1097 case '?':
1098 case 'h':
1099 puts(
1100 "p -print\n"
1101 "a %d -allocate\n"
1102 "f %x %d -free\n"
1103 "l %x %d -fill\n"
1104 "r %d -resize\n"
1105 "h/? -help"
1106 );
1107 break;
1108 default:
1109 printf("?\n");
1110 break;
1111 }
1112 }
1113 return(0);
1114 }
1115
1116 void
1117 panic(const char *ctl, ...)
1118 {
1119 va_list va;
1120
1121 va_start(va, ctl);
1122 vfprintf(stderr, ctl, va);
1123 fprintf(stderr, "\n");
1124 va_end(va);
1125 exit(1);
1126 }
1127
1128 #endif
1129
Cache object: 5112e3fbecbb9ac608b59a46b7edb3ad
|