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 * 3. 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 * 64), 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 affected 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 static daddr_t blst_leaf_alloc(blmeta_t *scan, daddr_t blk, int count, 124 daddr_t cursor); 125 static daddr_t blst_meta_alloc(blmeta_t *scan, daddr_t blk, daddr_t count, 126 daddr_t radix, daddr_t skip, daddr_t cursor); 127 static void blst_leaf_free(blmeta_t *scan, daddr_t relblk, int count); 128 static void blst_meta_free(blmeta_t *scan, daddr_t freeBlk, daddr_t count, 129 daddr_t radix, daddr_t skip, daddr_t blk); 130 static void blst_copy(blmeta_t *scan, daddr_t blk, daddr_t radix, 131 daddr_t skip, blist_t dest, daddr_t count); 132 static daddr_t blst_leaf_fill(blmeta_t *scan, daddr_t blk, int count); 133 static daddr_t blst_meta_fill(blmeta_t *scan, daddr_t allocBlk, daddr_t count, 134 daddr_t radix, daddr_t skip, daddr_t blk); 135 static daddr_t blst_radix_init(blmeta_t *scan, daddr_t radix, daddr_t skip, 136 daddr_t count); 137 #ifndef _KERNEL 138 static void blst_radix_print(blmeta_t *scan, daddr_t blk, daddr_t radix, 139 daddr_t skip, int tab); 140 #endif 141 142 #ifdef _KERNEL 143 static MALLOC_DEFINE(M_SWAP, "SWAP", "Swap space"); 144 #endif 145 146 /* 147 * blist_create() - create a blist capable of handling up to the specified 148 * number of blocks 149 * 150 * blocks - must be greater than 0 151 * flags - malloc flags 152 * 153 * The smallest blist consists of a single leaf node capable of 154 * managing BLIST_BMAP_RADIX blocks. 155 */ 156 blist_t 157 blist_create(daddr_t blocks, int flags) 158 { 159 blist_t bl; 160 daddr_t nodes, radix, skip; 161 162 /* 163 * Calculate radix and skip field used for scanning. 164 */ 165 radix = BLIST_BMAP_RADIX; 166 skip = 0; 167 while (radix < blocks) { 168 radix *= BLIST_META_RADIX; 169 skip = (skip + 1) * BLIST_META_RADIX; 170 } 171 nodes = 1 + blst_radix_init(NULL, radix, skip, blocks); 172 173 bl = malloc(sizeof(struct blist), M_SWAP, flags); 174 if (bl == NULL) 175 return (NULL); 176 177 bl->bl_blocks = blocks; 178 bl->bl_radix = radix; 179 bl->bl_skip = skip; 180 bl->bl_cursor = 0; 181 bl->bl_root = malloc(nodes * sizeof(blmeta_t), M_SWAP, flags); 182 if (bl->bl_root == NULL) { 183 free(bl, M_SWAP); 184 return (NULL); 185 } 186 blst_radix_init(bl->bl_root, radix, skip, blocks); 187 188 #if defined(BLIST_DEBUG) 189 printf( 190 "BLIST representing %lld blocks (%lld MB of swap)" 191 ", requiring %lldK of ram\n", 192 (long long)bl->bl_blocks, 193 (long long)bl->bl_blocks * 4 / 1024, 194 (long long)(nodes * sizeof(blmeta_t) + 1023) / 1024 195 ); 196 printf("BLIST raw radix tree contains %lld records\n", 197 (long long)nodes); 198 #endif 199 200 return (bl); 201 } 202 203 void 204 blist_destroy(blist_t bl) 205 { 206 free(bl->bl_root, M_SWAP); 207 free(bl, M_SWAP); 208 } 209 210 /* 211 * blist_alloc() - reserve space in the block bitmap. Return the base 212 * of a contiguous region or SWAPBLK_NONE if space could 213 * not be allocated. 214 */ 215 daddr_t 216 blist_alloc(blist_t bl, daddr_t count) 217 { 218 daddr_t blk; 219 220 /* 221 * This loop iterates at most twice. An allocation failure in the 222 * first iteration leads to a second iteration only if the cursor was 223 * non-zero. When the cursor is zero, an allocation failure will 224 * reduce the hint, stopping further iterations. 225 */ 226 while (count <= bl->bl_root->bm_bighint) { 227 if (bl->bl_radix == BLIST_BMAP_RADIX) 228 blk = blst_leaf_alloc(bl->bl_root, 0, count, 229 bl->bl_cursor); 230 else 231 blk = blst_meta_alloc(bl->bl_root, 0, count, 232 bl->bl_radix, bl->bl_skip, bl->bl_cursor); 233 if (blk != SWAPBLK_NONE) { 234 bl->bl_cursor = blk + count; 235 return (blk); 236 } else if (bl->bl_cursor != 0) 237 bl->bl_cursor = 0; 238 } 239 return (SWAPBLK_NONE); 240 } 241 242 /* 243 * blist_avail() - return the number of free blocks. 244 */ 245 daddr_t 246 blist_avail(blist_t bl) 247 { 248 249 if (bl->bl_radix == BLIST_BMAP_RADIX) 250 return (bitcount64(bl->bl_root->u.bmu_bitmap)); 251 else 252 return (bl->bl_root->u.bmu_avail); 253 } 254 255 /* 256 * blist_free() - free up space in the block bitmap. Return the base 257 * of a contiguous region. Panic if an inconsistancy is 258 * found. 259 */ 260 void 261 blist_free(blist_t bl, daddr_t blkno, daddr_t count) 262 { 263 if (bl) { 264 if (bl->bl_radix == BLIST_BMAP_RADIX) 265 blst_leaf_free(bl->bl_root, blkno, count); 266 else 267 blst_meta_free(bl->bl_root, blkno, count, 268 bl->bl_radix, bl->bl_skip, 0); 269 } 270 } 271 272 /* 273 * blist_fill() - mark a region in the block bitmap as off-limits 274 * to the allocator (i.e. allocate it), ignoring any 275 * existing allocations. Return the number of blocks 276 * actually filled that were free before the call. 277 */ 278 daddr_t 279 blist_fill(blist_t bl, daddr_t blkno, daddr_t count) 280 { 281 daddr_t filled; 282 283 if (bl) { 284 if (bl->bl_radix == BLIST_BMAP_RADIX) 285 filled = blst_leaf_fill(bl->bl_root, blkno, count); 286 else 287 filled = blst_meta_fill(bl->bl_root, blkno, count, 288 bl->bl_radix, bl->bl_skip, 0); 289 return (filled); 290 } 291 return (0); 292 } 293 294 /* 295 * blist_resize() - resize an existing radix tree to handle the 296 * specified number of blocks. This will reallocate 297 * the tree and transfer the previous bitmap to the new 298 * one. When extending the tree you can specify whether 299 * the new blocks are to left allocated or freed. 300 */ 301 void 302 blist_resize(blist_t *pbl, daddr_t count, int freenew, int flags) 303 { 304 blist_t newbl = blist_create(count, flags); 305 blist_t save = *pbl; 306 307 *pbl = newbl; 308 if (count > save->bl_blocks) 309 count = save->bl_blocks; 310 blst_copy(save->bl_root, 0, save->bl_radix, save->bl_skip, newbl, count); 311 312 /* 313 * If resizing upwards, should we free the new space or not? 314 */ 315 if (freenew && count < newbl->bl_blocks) { 316 blist_free(newbl, count, newbl->bl_blocks - count); 317 } 318 blist_destroy(save); 319 } 320 321 #ifdef BLIST_DEBUG 322 323 /* 324 * blist_print() - dump radix tree 325 */ 326 void 327 blist_print(blist_t bl) 328 { 329 printf("BLIST {\n"); 330 blst_radix_print(bl->bl_root, 0, bl->bl_radix, bl->bl_skip, 4); 331 printf("}\n"); 332 } 333 334 #endif 335 336 /************************************************************************ 337 * ALLOCATION SUPPORT FUNCTIONS * 338 ************************************************************************ 339 * 340 * These support functions do all the actual work. They may seem 341 * rather longish, but that's because I've commented them up. The 342 * actual code is straight forward. 343 * 344 */ 345 346 /* 347 * blist_leaf_alloc() - allocate at a leaf in the radix tree (a bitmap). 348 * 349 * This is the core of the allocator and is optimized for the 350 * BLIST_BMAP_RADIX block allocation case. Otherwise, execution 351 * time is proportional to log2(count) + log2(BLIST_BMAP_RADIX). 352 */ 353 static daddr_t 354 blst_leaf_alloc(blmeta_t *scan, daddr_t blk, int count, daddr_t cursor) 355 { 356 u_daddr_t mask; 357 int count1, hi, lo, mid, num_shifts, range1, range_ext; 358 359 if (count == BLIST_BMAP_RADIX) { 360 /* 361 * Optimize allocation of BLIST_BMAP_RADIX bits. If this wasn't 362 * a special case, then forming the final value of 'mask' below 363 * would require special handling to avoid an invalid left shift 364 * when count equals the number of bits in mask. 365 */ 366 if (~scan->u.bmu_bitmap != 0) { 367 scan->bm_bighint = BLIST_BMAP_RADIX - 1; 368 return (SWAPBLK_NONE); 369 } 370 if (cursor != blk) 371 return (SWAPBLK_NONE); 372 scan->u.bmu_bitmap = 0; 373 scan->bm_bighint = 0; 374 return (blk); 375 } 376 range1 = 0; 377 count1 = count - 1; 378 num_shifts = fls(count1); 379 mask = scan->u.bmu_bitmap; 380 while (mask != 0 && num_shifts > 0) { 381 /* 382 * If bit i is set in mask, then bits in [i, i+range1] are set 383 * in scan->u.bmu_bitmap. The value of range1 is equal to 384 * count1 >> num_shifts. Grow range and reduce num_shifts to 0, 385 * while preserving these invariants. The updates to mask leave 386 * fewer bits set, but each bit that remains set represents a 387 * longer string of consecutive bits set in scan->u.bmu_bitmap. 388 */ 389 num_shifts--; 390 range_ext = range1 + ((count1 >> num_shifts) & 1); 391 mask &= mask >> range_ext; 392 range1 += range_ext; 393 } 394 if (mask == 0) { 395 /* 396 * Update bighint. There is no allocation bigger than range1 397 * available in this leaf. 398 */ 399 scan->bm_bighint = range1; 400 return (SWAPBLK_NONE); 401 } 402 403 /* 404 * Discard any candidates that appear before the cursor. 405 */ 406 lo = cursor - blk; 407 mask &= ~(u_daddr_t)0 << lo; 408 409 if (mask == 0) 410 return (SWAPBLK_NONE); 411 412 /* 413 * The least significant set bit in mask marks the start of the first 414 * available range of sufficient size. Clear all the bits but that one, 415 * and then perform a binary search to find its position. 416 */ 417 mask &= -mask; 418 hi = BLIST_BMAP_RADIX - count1; 419 while (lo + 1 < hi) { 420 mid = (lo + hi) >> 1; 421 if ((mask >> mid) != 0) 422 lo = mid; 423 else 424 hi = mid; 425 } 426 427 /* 428 * Set in mask exactly the bits being allocated, and clear them from 429 * the set of available bits. 430 */ 431 mask = (mask << count) - mask; 432 scan->u.bmu_bitmap &= ~mask; 433 return (blk + lo); 434 } 435 436 /* 437 * blist_meta_alloc() - allocate at a meta in the radix tree. 438 * 439 * Attempt to allocate at a meta node. If we can't, we update 440 * bighint and return a failure. Updating bighint optimize future 441 * calls that hit this node. We have to check for our collapse cases 442 * and we have a few optimizations strewn in as well. 443 */ 444 static daddr_t 445 blst_meta_alloc(blmeta_t *scan, daddr_t blk, daddr_t count, daddr_t radix, 446 daddr_t skip, daddr_t cursor) 447 { 448 daddr_t i, next_skip, r; 449 int child; 450 bool scan_from_start; 451 452 if (scan->u.bmu_avail < count) { 453 /* 454 * The meta node's hint must be too large if the allocation 455 * exceeds the number of free blocks. Reduce the hint, and 456 * return failure. 457 */ 458 scan->bm_bighint = scan->u.bmu_avail; 459 return (SWAPBLK_NONE); 460 } 461 next_skip = skip / BLIST_META_RADIX; 462 463 /* 464 * An ALL-FREE meta node requires special handling before allocating 465 * any of its blocks. 466 */ 467 if (scan->u.bmu_avail == radix) { 468 radix /= BLIST_META_RADIX; 469 470 /* 471 * Reinitialize each of the meta node's children. An ALL-FREE 472 * meta node cannot have a terminator in any subtree. 473 */ 474 for (i = 1; i <= skip; i += next_skip) { 475 if (next_skip == 1) 476 scan[i].u.bmu_bitmap = (u_daddr_t)-1; 477 else 478 scan[i].u.bmu_avail = radix; 479 scan[i].bm_bighint = radix; 480 } 481 } else { 482 radix /= BLIST_META_RADIX; 483 } 484 485 if (count > radix) { 486 /* 487 * The allocation exceeds the number of blocks that are 488 * managed by a subtree of this meta node. 489 */ 490 panic("allocation too large"); 491 } 492 scan_from_start = cursor == blk; 493 child = (cursor - blk) / radix; 494 blk += child * radix; 495 for (i = 1 + child * next_skip; i <= skip; i += next_skip) { 496 if (count <= scan[i].bm_bighint) { 497 /* 498 * The allocation might fit in the i'th subtree. 499 */ 500 if (next_skip == 1) { 501 r = blst_leaf_alloc(&scan[i], blk, count, 502 cursor > blk ? cursor : blk); 503 } else { 504 r = blst_meta_alloc(&scan[i], blk, count, 505 radix, next_skip - 1, cursor > blk ? 506 cursor : blk); 507 } 508 if (r != SWAPBLK_NONE) { 509 scan->u.bmu_avail -= count; 510 return (r); 511 } 512 } else if (scan[i].bm_bighint == (daddr_t)-1) { 513 /* 514 * Terminator 515 */ 516 break; 517 } 518 blk += radix; 519 } 520 521 /* 522 * We couldn't allocate count in this subtree, update bighint. 523 */ 524 if (scan_from_start && scan->bm_bighint >= count) 525 scan->bm_bighint = count - 1; 526 527 return (SWAPBLK_NONE); 528 } 529 530 /* 531 * BLST_LEAF_FREE() - free allocated block from leaf bitmap 532 * 533 */ 534 static void 535 blst_leaf_free(blmeta_t *scan, daddr_t blk, int count) 536 { 537 /* 538 * free some data in this bitmap 539 * 540 * e.g. 541 * 0000111111111110000 542 * \_________/\__/ 543 * v n 544 */ 545 int n = blk & (BLIST_BMAP_RADIX - 1); 546 u_daddr_t mask; 547 548 mask = ((u_daddr_t)-1 << n) & 549 ((u_daddr_t)-1 >> (BLIST_BMAP_RADIX - count - n)); 550 551 if (scan->u.bmu_bitmap & mask) 552 panic("blst_radix_free: freeing free block"); 553 scan->u.bmu_bitmap |= mask; 554 555 /* 556 * We could probably do a better job here. We are required to make 557 * bighint at least as large as the biggest contiguous block of 558 * data. If we just shoehorn it, a little extra overhead will 559 * be incured on the next allocation (but only that one typically). 560 */ 561 scan->bm_bighint = BLIST_BMAP_RADIX; 562 } 563 564 /* 565 * BLST_META_FREE() - free allocated blocks from radix tree meta info 566 * 567 * This support routine frees a range of blocks from the bitmap. 568 * The range must be entirely enclosed by this radix node. If a 569 * meta node, we break the range down recursively to free blocks 570 * in subnodes (which means that this code can free an arbitrary 571 * range whereas the allocation code cannot allocate an arbitrary 572 * range). 573 */ 574 static void 575 blst_meta_free(blmeta_t *scan, daddr_t freeBlk, daddr_t count, daddr_t radix, 576 daddr_t skip, daddr_t blk) 577 { 578 daddr_t i, next_skip, v; 579 int child; 580 581 next_skip = skip / BLIST_META_RADIX; 582 583 if (scan->u.bmu_avail == 0) { 584 /* 585 * ALL-ALLOCATED special case, with possible 586 * shortcut to ALL-FREE special case. 587 */ 588 scan->u.bmu_avail = count; 589 scan->bm_bighint = count; 590 591 if (count != radix) { 592 for (i = 1; i <= skip; i += next_skip) { 593 if (scan[i].bm_bighint == (daddr_t)-1) 594 break; 595 scan[i].bm_bighint = 0; 596 if (next_skip == 1) { 597 scan[i].u.bmu_bitmap = 0; 598 } else { 599 scan[i].u.bmu_avail = 0; 600 } 601 } 602 /* fall through */ 603 } 604 } else { 605 scan->u.bmu_avail += count; 606 /* scan->bm_bighint = radix; */ 607 } 608 609 /* 610 * ALL-FREE special case. 611 */ 612 613 if (scan->u.bmu_avail == radix) 614 return; 615 if (scan->u.bmu_avail > radix) 616 panic("blst_meta_free: freeing already free blocks (%lld) %lld/%lld", 617 (long long)count, (long long)scan->u.bmu_avail, 618 (long long)radix); 619 620 /* 621 * Break the free down into its components 622 */ 623 624 radix /= BLIST_META_RADIX; 625 626 child = (freeBlk - blk) / radix; 627 blk += child * radix; 628 i = 1 + child * next_skip; 629 while (i <= skip && blk < freeBlk + count) { 630 v = blk + radix - freeBlk; 631 if (v > count) 632 v = count; 633 634 if (scan->bm_bighint == (daddr_t)-1) 635 panic("blst_meta_free: freeing unexpected range"); 636 637 if (next_skip == 1) { 638 blst_leaf_free(&scan[i], freeBlk, v); 639 } else { 640 blst_meta_free(&scan[i], freeBlk, v, radix, next_skip - 1, blk); 641 } 642 if (scan->bm_bighint < scan[i].bm_bighint) 643 scan->bm_bighint = scan[i].bm_bighint; 644 count -= v; 645 freeBlk += v; 646 blk += radix; 647 i += next_skip; 648 } 649 } 650 651 /* 652 * BLIST_RADIX_COPY() - copy one radix tree to another 653 * 654 * Locates free space in the source tree and frees it in the destination 655 * tree. The space may not already be free in the destination. 656 */ 657 static void 658 blst_copy(blmeta_t *scan, daddr_t blk, daddr_t radix, daddr_t skip, 659 blist_t dest, daddr_t count) 660 { 661 daddr_t i, next_skip; 662 663 /* 664 * Leaf node 665 */ 666 667 if (radix == BLIST_BMAP_RADIX) { 668 u_daddr_t v = scan->u.bmu_bitmap; 669 670 if (v == (u_daddr_t)-1) { 671 blist_free(dest, blk, count); 672 } else if (v != 0) { 673 int i; 674 675 for (i = 0; i < BLIST_BMAP_RADIX && i < count; ++i) { 676 if (v & ((u_daddr_t)1 << i)) 677 blist_free(dest, blk + i, 1); 678 } 679 } 680 return; 681 } 682 683 /* 684 * Meta node 685 */ 686 687 if (scan->u.bmu_avail == 0) { 688 /* 689 * Source all allocated, leave dest allocated 690 */ 691 return; 692 } 693 if (scan->u.bmu_avail == radix) { 694 /* 695 * Source all free, free entire dest 696 */ 697 if (count < radix) 698 blist_free(dest, blk, count); 699 else 700 blist_free(dest, blk, radix); 701 return; 702 } 703 704 705 radix /= BLIST_META_RADIX; 706 next_skip = skip / BLIST_META_RADIX; 707 708 for (i = 1; count && i <= skip; i += next_skip) { 709 if (scan[i].bm_bighint == (daddr_t)-1) 710 break; 711 712 if (count >= radix) { 713 blst_copy(&scan[i], blk, radix, next_skip - 1, dest, 714 radix); 715 count -= radix; 716 } else { 717 if (count) { 718 blst_copy(&scan[i], blk, radix, next_skip - 1, 719 dest, count); 720 } 721 count = 0; 722 } 723 blk += radix; 724 } 725 } 726 727 /* 728 * BLST_LEAF_FILL() - allocate specific blocks in leaf bitmap 729 * 730 * This routine allocates all blocks in the specified range 731 * regardless of any existing allocations in that range. Returns 732 * the number of blocks allocated by the call. 733 */ 734 static daddr_t 735 blst_leaf_fill(blmeta_t *scan, daddr_t blk, int count) 736 { 737 int n = blk & (BLIST_BMAP_RADIX - 1); 738 daddr_t nblks; 739 u_daddr_t mask; 740 741 mask = ((u_daddr_t)-1 << n) & 742 ((u_daddr_t)-1 >> (BLIST_BMAP_RADIX - count - n)); 743 744 /* Count the number of blocks that we are allocating. */ 745 nblks = bitcount64(scan->u.bmu_bitmap & mask); 746 747 scan->u.bmu_bitmap &= ~mask; 748 return (nblks); 749 } 750 751 /* 752 * BLIST_META_FILL() - allocate specific blocks at a meta node 753 * 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. 758 */ 759 static daddr_t 760 blst_meta_fill(blmeta_t *scan, daddr_t allocBlk, daddr_t count, daddr_t radix, 761 daddr_t skip, daddr_t blk) 762 { 763 daddr_t i, nblks, next_skip, v; 764 int child; 765 766 if (count > radix) { 767 /* 768 * The allocation exceeds the number of blocks that are 769 * managed by this meta node. 770 */ 771 panic("allocation too large"); 772 } 773 if (count == radix || scan->u.bmu_avail == 0) { 774 /* 775 * ALL-ALLOCATED special case 776 */ 777 nblks = scan->u.bmu_avail; 778 scan->u.bmu_avail = 0; 779 scan->bm_bighint = 0; 780 return nblks; 781 } 782 next_skip = skip / BLIST_META_RADIX; 783 784 /* 785 * An ALL-FREE meta node requires special handling before allocating 786 * any of its blocks. 787 */ 788 if (scan->u.bmu_avail == radix) { 789 radix /= BLIST_META_RADIX; 790 791 /* 792 * Reinitialize each of the meta node's children. An ALL-FREE 793 * meta node cannot have a terminator in any subtree. 794 */ 795 for (i = 1; i <= skip; i += next_skip) { 796 if (next_skip == 1) { 797 scan[i].u.bmu_bitmap = (u_daddr_t)-1; 798 scan[i].bm_bighint = BLIST_BMAP_RADIX; 799 } else { 800 scan[i].bm_bighint = radix; 801 scan[i].u.bmu_avail = radix; 802 } 803 } 804 } else { 805 radix /= BLIST_META_RADIX; 806 } 807 808 nblks = 0; 809 child = (allocBlk - blk) / radix; 810 blk += child * radix; 811 i = 1 + child * next_skip; 812 while (i <= skip && blk < allocBlk + count) { 813 v = blk + radix - allocBlk; 814 if (v > count) 815 v = count; 816 817 if (scan->bm_bighint == (daddr_t)-1) 818 panic("blst_meta_fill: filling unexpected range"); 819 820 if (next_skip == 1) { 821 nblks += blst_leaf_fill(&scan[i], allocBlk, v); 822 } else { 823 nblks += blst_meta_fill(&scan[i], allocBlk, v, 824 radix, next_skip - 1, blk); 825 } 826 count -= v; 827 allocBlk += v; 828 blk += radix; 829 i += next_skip; 830 } 831 scan->u.bmu_avail -= nblks; 832 return nblks; 833 } 834 835 /* 836 * BLST_RADIX_INIT() - initialize radix tree 837 * 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. 842 */ 843 static daddr_t 844 blst_radix_init(blmeta_t *scan, daddr_t radix, daddr_t skip, daddr_t count) 845 { 846 daddr_t i, memindex, next_skip; 847 848 memindex = 0; 849 850 /* 851 * Leaf node 852 */ 853 854 if (radix == BLIST_BMAP_RADIX) { 855 if (scan) { 856 scan->bm_bighint = 0; 857 scan->u.bmu_bitmap = 0; 858 } 859 return (memindex); 860 } 861 862 /* 863 * Meta node. If allocating the entire object we can special 864 * case it. However, we need to figure out how much memory 865 * is required to manage 'count' blocks, so we continue on anyway. 866 */ 867 868 if (scan) { 869 scan->bm_bighint = 0; 870 scan->u.bmu_avail = 0; 871 } 872 873 radix /= BLIST_META_RADIX; 874 next_skip = skip / BLIST_META_RADIX; 875 876 for (i = 1; i <= skip; i += next_skip) { 877 if (count >= radix) { 878 /* 879 * Allocate the entire object 880 */ 881 memindex = i + 882 blst_radix_init(((scan) ? &scan[i] : NULL), radix, 883 next_skip - 1, radix); 884 count -= radix; 885 } else if (count > 0) { 886 /* 887 * Allocate a partial object 888 */ 889 memindex = i + 890 blst_radix_init(((scan) ? &scan[i] : NULL), radix, 891 next_skip - 1, count); 892 count = 0; 893 } else { 894 /* 895 * Add terminator and break out 896 */ 897 if (scan) 898 scan[i].bm_bighint = (daddr_t)-1; 899 break; 900 } 901 } 902 if (memindex < i) 903 memindex = i; 904 return (memindex); 905 } 906 907 #ifdef BLIST_DEBUG 908 909 static void 910 blst_radix_print(blmeta_t *scan, daddr_t blk, daddr_t radix, daddr_t skip, 911 int tab) 912 { 913 daddr_t i, next_skip; 914 915 if (radix == BLIST_BMAP_RADIX) { 916 printf( 917 "%*.*s(%08llx,%lld): bitmap %016llx big=%lld\n", 918 tab, tab, "", 919 (long long)blk, (long long)radix, 920 (long long)scan->u.bmu_bitmap, 921 (long long)scan->bm_bighint 922 ); 923 return; 924 } 925 926 if (scan->u.bmu_avail == 0) { 927 printf( 928 "%*.*s(%08llx,%lld) ALL ALLOCATED\n", 929 tab, tab, "", 930 (long long)blk, 931 (long long)radix 932 ); 933 return; 934 } 935 if (scan->u.bmu_avail == radix) { 936 printf( 937 "%*.*s(%08llx,%lld) ALL FREE\n", 938 tab, tab, "", 939 (long long)blk, 940 (long long)radix 941 ); 942 return; 943 } 944 945 printf( 946 "%*.*s(%08llx,%lld): subtree (%lld/%lld) big=%lld {\n", 947 tab, tab, "", 948 (long long)blk, (long long)radix, 949 (long long)scan->u.bmu_avail, 950 (long long)radix, 951 (long long)scan->bm_bighint 952 ); 953 954 radix /= BLIST_META_RADIX; 955 next_skip = skip / BLIST_META_RADIX; 956 tab += 4; 957 958 for (i = 1; i <= skip; i += next_skip) { 959 if (scan[i].bm_bighint == (daddr_t)-1) { 960 printf( 961 "%*.*s(%08llx,%lld): Terminator\n", 962 tab, tab, "", 963 (long long)blk, (long long)radix 964 ); 965 break; 966 } 967 blst_radix_print(&scan[i], blk, radix, next_skip - 1, tab); 968 blk += radix; 969 } 970 tab -= 4; 971 972 printf( 973 "%*.*s}\n", 974 tab, tab, "" 975 ); 976 } 977 978 #endif 979 980 #ifdef BLIST_DEBUG 981 982 int 983 main(int ac, char **av) 984 { 985 int size = 1024; 986 int i; 987 blist_t bl; 988 989 for (i = 1; i < ac; ++i) { 990 const char *ptr = av[i]; 991 if (*ptr != '-') { 992 size = strtol(ptr, NULL, 0); 993 continue; 994 } 995 ptr += 2; 996 fprintf(stderr, "Bad option: %s\n", ptr - 2); 997 exit(1); 998 } 999 bl = blist_create(size, M_WAITOK); 1000 blist_free(bl, 0, size); 1001 1002 for (;;) { 1003 char buf[1024]; 1004 long long da = 0; 1005 long long count = 0; 1006 1007 printf("%lld/%lld/%lld> ", (long long)blist_avail(bl), 1008 (long long)size, (long long)bl->bl_radix); 1009 fflush(stdout); 1010 if (fgets(buf, sizeof(buf), stdin) == NULL) 1011 break; 1012 switch(buf[0]) { 1013 case 'r': 1014 if (sscanf(buf + 1, "%lld", &count) == 1) { 1015 blist_resize(&bl, count, 1, M_WAITOK); 1016 } else { 1017 printf("?\n"); 1018 } 1019 case 'p': 1020 blist_print(bl); 1021 break; 1022 case 'a': 1023 if (sscanf(buf + 1, "%lld", &count) == 1) { 1024 daddr_t blk = blist_alloc(bl, count); 1025 printf(" R=%08llx\n", (long long)blk); 1026 } else { 1027 printf("?\n"); 1028 } 1029 break; 1030 case 'f': 1031 if (sscanf(buf + 1, "%llx %lld", &da, &count) == 2) { 1032 blist_free(bl, da, count); 1033 } else { 1034 printf("?\n"); 1035 } 1036 break; 1037 case 'l': 1038 if (sscanf(buf + 1, "%llx %lld", &da, &count) == 2) { 1039 printf(" n=%jd\n", 1040 (intmax_t)blist_fill(bl, da, count)); 1041 } else { 1042 printf("?\n"); 1043 } 1044 break; 1045 case '?': 1046 case 'h': 1047 puts( 1048 "p -print\n" 1049 "a %d -allocate\n" 1050 "f %x %d -free\n" 1051 "l %x %d -fill\n" 1052 "r %d -resize\n" 1053 "h/? -help" 1054 ); 1055 break; 1056 default: 1057 printf("?\n"); 1058 break; 1059 } 1060 } 1061 return(0); 1062 } 1063 1064 void 1065 panic(const char *ctl, ...) 1066 { 1067 va_list va; 1068 1069 va_start(va, ctl); 1070 vfprintf(stderr, ctl, va); 1071 fprintf(stderr, "\n"); 1072 va_end(va); 1073 exit(1); 1074 } 1075 1076 #endif 1077