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 #define CTASSERT(expr) 114 115 #include <sys/blist.h> 116 117 void panic(const char *ctl, ...); 118 119 #endif 120 121 /* 122 * static support functions 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 cursor, daddr_t count, 127 u_daddr_t radix); 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 u_daddr_t radix); 131 static void blst_copy(blmeta_t *scan, daddr_t blk, daddr_t radix, 132 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 u_daddr_t radix); 136 static daddr_t blst_radix_init(blmeta_t *scan, daddr_t radix, daddr_t count); 137 #ifndef _KERNEL 138 static void blst_radix_print(blmeta_t *scan, daddr_t blk, daddr_t radix, 139 int tab); 140 #endif 141 142 #ifdef _KERNEL 143 static MALLOC_DEFINE(M_SWAP, "SWAP", "Swap space"); 144 #endif 145 146 CTASSERT(BLIST_BMAP_RADIX % BLIST_META_RADIX == 0); 147 148 /* 149 * For a subtree that can represent the state of up to 'radix' blocks, the 150 * number of leaf nodes of the subtree is L=radix/BLIST_BMAP_RADIX. If 'm' 151 * is short for BLIST_META_RADIX, then for a tree of height h with L=m**h 152 * leaf nodes, the total number of tree nodes is 1 + m + m**2 + ... + m**h, 153 * or, equivalently, (m**(h+1)-1)/(m-1). This quantity is called 'skip' 154 * in the 'meta' functions that process subtrees. Since integer division 155 * discards remainders, we can express this computation as 156 * skip = (m * m**h) / (m - 1) 157 * skip = (m * (radix / BLIST_BMAP_RADIX)) / (m - 1) 158 * and since m divides BLIST_BMAP_RADIX, we can simplify further to 159 * skip = (radix / (BLIST_BMAP_RADIX / m)) / (m - 1) 160 * skip = radix / ((BLIST_BMAP_RADIX / m) * (m - 1)) 161 * so that simple integer division by a constant can safely be used for the 162 * calculation. 163 */ 164 static inline daddr_t 165 radix_to_skip(daddr_t radix) 166 { 167 168 return (radix / 169 ((BLIST_BMAP_RADIX / BLIST_META_RADIX) * (BLIST_META_RADIX - 1))); 170 } 171 172 /* 173 * blist_create() - create a blist capable of handling up to the specified 174 * number of blocks 175 * 176 * blocks - must be greater than 0 177 * flags - malloc flags 178 * 179 * The smallest blist consists of a single leaf node capable of 180 * managing BLIST_BMAP_RADIX blocks. 181 */ 182 blist_t 183 blist_create(daddr_t blocks, int flags) 184 { 185 blist_t bl; 186 daddr_t nodes, radix; 187 188 /* 189 * Calculate the radix field used for scanning. 190 */ 191 radix = BLIST_BMAP_RADIX; 192 while (radix < blocks) { 193 radix *= BLIST_META_RADIX; 194 } 195 nodes = 1 + blst_radix_init(NULL, radix, blocks); 196 197 bl = malloc(sizeof(struct blist), M_SWAP, flags); 198 if (bl == NULL) 199 return (NULL); 200 201 bl->bl_blocks = blocks; 202 bl->bl_radix = radix; 203 bl->bl_cursor = 0; 204 bl->bl_root = malloc(nodes * sizeof(blmeta_t), M_SWAP, flags); 205 if (bl->bl_root == NULL) { 206 free(bl, M_SWAP); 207 return (NULL); 208 } 209 blst_radix_init(bl->bl_root, radix, blocks); 210 211 #if defined(BLIST_DEBUG) 212 printf( 213 "BLIST representing %lld blocks (%lld MB of swap)" 214 ", requiring %lldK of ram\n", 215 (long long)bl->bl_blocks, 216 (long long)bl->bl_blocks * 4 / 1024, 217 (long long)(nodes * sizeof(blmeta_t) + 1023) / 1024 218 ); 219 printf("BLIST raw radix tree contains %lld records\n", 220 (long long)nodes); 221 #endif 222 223 return (bl); 224 } 225 226 void 227 blist_destroy(blist_t bl) 228 { 229 free(bl->bl_root, M_SWAP); 230 free(bl, M_SWAP); 231 } 232 233 /* 234 * blist_alloc() - reserve space in the block bitmap. Return the base 235 * of a contiguous region or SWAPBLK_NONE if space could 236 * not be allocated. 237 */ 238 daddr_t 239 blist_alloc(blist_t bl, daddr_t count) 240 { 241 daddr_t blk; 242 243 /* 244 * This loop iterates at most twice. An allocation failure in the 245 * first iteration leads to a second iteration only if the cursor was 246 * non-zero. When the cursor is zero, an allocation failure will 247 * reduce the hint, stopping further iterations. 248 */ 249 while (count <= bl->bl_root->bm_bighint) { 250 blk = blst_meta_alloc(bl->bl_root, bl->bl_cursor, count, 251 bl->bl_radix); 252 if (blk != SWAPBLK_NONE) { 253 bl->bl_cursor = blk + count; 254 return (blk); 255 } else if (bl->bl_cursor != 0) 256 bl->bl_cursor = 0; 257 } 258 return (SWAPBLK_NONE); 259 } 260 261 /* 262 * blist_avail() - return the number of free blocks. 263 */ 264 daddr_t 265 blist_avail(blist_t bl) 266 { 267 268 if (bl->bl_radix == BLIST_BMAP_RADIX) 269 return (bitcount64(bl->bl_root->u.bmu_bitmap)); 270 else 271 return (bl->bl_root->u.bmu_avail); 272 } 273 274 /* 275 * blist_free() - free up space in the block bitmap. Return the base 276 * of a contiguous region. Panic if an inconsistancy is 277 * found. 278 */ 279 void 280 blist_free(blist_t bl, daddr_t blkno, daddr_t count) 281 { 282 283 blst_meta_free(bl->bl_root, blkno, count, bl->bl_radix); 284 } 285 286 /* 287 * blist_fill() - mark a region in the block bitmap as off-limits 288 * to the allocator (i.e. allocate it), ignoring any 289 * existing allocations. Return the number of blocks 290 * actually filled that were free before the call. 291 */ 292 daddr_t 293 blist_fill(blist_t bl, daddr_t blkno, daddr_t count) 294 { 295 296 return (blst_meta_fill(bl->bl_root, blkno, count, bl->bl_radix)); 297 } 298 299 /* 300 * blist_resize() - resize an existing radix tree to handle the 301 * specified number of blocks. This will reallocate 302 * the tree and transfer the previous bitmap to the new 303 * one. When extending the tree you can specify whether 304 * the new blocks are to left allocated or freed. 305 */ 306 void 307 blist_resize(blist_t *pbl, daddr_t count, int freenew, int flags) 308 { 309 blist_t newbl = blist_create(count, flags); 310 blist_t save = *pbl; 311 312 *pbl = newbl; 313 if (count > save->bl_blocks) 314 count = save->bl_blocks; 315 blst_copy(save->bl_root, 0, save->bl_radix, newbl, count); 316 317 /* 318 * If resizing upwards, should we free the new space or not? 319 */ 320 if (freenew && count < newbl->bl_blocks) { 321 blist_free(newbl, count, newbl->bl_blocks - count); 322 } 323 blist_destroy(save); 324 } 325 326 #ifdef BLIST_DEBUG 327 328 /* 329 * blist_print() - dump radix tree 330 */ 331 void 332 blist_print(blist_t bl) 333 { 334 printf("BLIST cursor = %08jx {\n", (uintmax_t)bl->bl_cursor); 335 blst_radix_print(bl->bl_root, 0, bl->bl_radix, 4); 336 printf("}\n"); 337 } 338 339 #endif 340 341 /************************************************************************ 342 * ALLOCATION SUPPORT FUNCTIONS * 343 ************************************************************************ 344 * 345 * These support functions do all the actual work. They may seem 346 * rather longish, but that's because I've commented them up. The 347 * actual code is straight forward. 348 * 349 */ 350 351 /* 352 * blist_leaf_alloc() - allocate at a leaf in the radix tree (a bitmap). 353 * 354 * This is the core of the allocator and is optimized for the 355 * BLIST_BMAP_RADIX block allocation case. Otherwise, execution 356 * time is proportional to log2(count) + log2(BLIST_BMAP_RADIX). 357 */ 358 static daddr_t 359 blst_leaf_alloc(blmeta_t *scan, daddr_t blk, int count, daddr_t cursor) 360 { 361 u_daddr_t mask; 362 int count1, hi, lo, mid, num_shifts, range1, range_ext; 363 364 if (count == BLIST_BMAP_RADIX) { 365 /* 366 * Optimize allocation of BLIST_BMAP_RADIX bits. If this wasn't 367 * a special case, then forming the final value of 'mask' below 368 * would require special handling to avoid an invalid left shift 369 * when count equals the number of bits in mask. 370 */ 371 if (~scan->u.bmu_bitmap != 0) { 372 scan->bm_bighint = BLIST_BMAP_RADIX - 1; 373 return (SWAPBLK_NONE); 374 } 375 if (cursor != blk) 376 return (SWAPBLK_NONE); 377 scan->u.bmu_bitmap = 0; 378 scan->bm_bighint = 0; 379 return (blk); 380 } 381 range1 = 0; 382 count1 = count - 1; 383 num_shifts = fls(count1); 384 mask = scan->u.bmu_bitmap; 385 while (mask != 0 && num_shifts > 0) { 386 /* 387 * If bit i is set in mask, then bits in [i, i+range1] are set 388 * in scan->u.bmu_bitmap. The value of range1 is equal to 389 * count1 >> num_shifts. Grow range and reduce num_shifts to 0, 390 * while preserving these invariants. The updates to mask leave 391 * fewer bits set, but each bit that remains set represents a 392 * longer string of consecutive bits set in scan->u.bmu_bitmap. 393 */ 394 num_shifts--; 395 range_ext = range1 + ((count1 >> num_shifts) & 1); 396 mask &= mask >> range_ext; 397 range1 += range_ext; 398 } 399 if (mask == 0) { 400 /* 401 * Update bighint. There is no allocation bigger than range1 402 * available in this leaf. 403 */ 404 scan->bm_bighint = range1; 405 return (SWAPBLK_NONE); 406 } 407 408 /* 409 * Discard any candidates that appear before the cursor. 410 */ 411 lo = cursor - blk; 412 mask &= ~(u_daddr_t)0 << lo; 413 414 if (mask == 0) 415 return (SWAPBLK_NONE); 416 417 /* 418 * The least significant set bit in mask marks the start of the first 419 * available range of sufficient size. Clear all the bits but that one, 420 * and then perform a binary search to find its position. 421 */ 422 mask &= -mask; 423 hi = BLIST_BMAP_RADIX - count1; 424 while (lo + 1 < hi) { 425 mid = (lo + hi) >> 1; 426 if ((mask >> mid) != 0) 427 lo = mid; 428 else 429 hi = mid; 430 } 431 432 /* 433 * Set in mask exactly the bits being allocated, and clear them from 434 * the set of available bits. 435 */ 436 mask = (mask << count) - mask; 437 scan->u.bmu_bitmap &= ~mask; 438 return (blk + lo); 439 } 440 441 /* 442 * blist_meta_alloc() - allocate at a meta in the radix tree. 443 * 444 * Attempt to allocate at a meta node. If we can't, we update 445 * bighint and return a failure. Updating bighint optimize future 446 * calls that hit this node. We have to check for our collapse cases 447 * and we have a few optimizations strewn in as well. 448 */ 449 static daddr_t 450 blst_meta_alloc(blmeta_t *scan, daddr_t cursor, daddr_t count, u_daddr_t radix) 451 { 452 daddr_t blk, i, next_skip, r, skip; 453 int child; 454 bool scan_from_start; 455 456 blk = cursor & -radix; 457 if (radix == BLIST_BMAP_RADIX) 458 return (blst_leaf_alloc(scan, blk, count, cursor)); 459 if (scan->u.bmu_avail < count) { 460 /* 461 * The meta node's hint must be too large if the allocation 462 * exceeds the number of free blocks. Reduce the hint, and 463 * return failure. 464 */ 465 scan->bm_bighint = scan->u.bmu_avail; 466 return (SWAPBLK_NONE); 467 } 468 skip = radix_to_skip(radix); 469 next_skip = skip / BLIST_META_RADIX; 470 471 /* 472 * An ALL-FREE meta node requires special handling before allocating 473 * any of its blocks. 474 */ 475 if (scan->u.bmu_avail == radix) { 476 radix /= BLIST_META_RADIX; 477 478 /* 479 * Reinitialize each of the meta node's children. An ALL-FREE 480 * meta node cannot have a terminator in any subtree. 481 */ 482 for (i = 1; i < skip; i += next_skip) { 483 if (next_skip == 1) 484 scan[i].u.bmu_bitmap = (u_daddr_t)-1; 485 else 486 scan[i].u.bmu_avail = radix; 487 scan[i].bm_bighint = radix; 488 } 489 } else { 490 radix /= BLIST_META_RADIX; 491 } 492 493 if (count > radix) { 494 /* 495 * The allocation exceeds the number of blocks that are 496 * managed by a subtree of this meta node. 497 */ 498 panic("allocation too large"); 499 } 500 scan_from_start = cursor == blk; 501 child = (cursor - blk) / radix; 502 blk += child * radix; 503 for (i = 1 + child * next_skip; i < skip; i += next_skip) { 504 if (count <= scan[i].bm_bighint) { 505 /* 506 * The allocation might fit in the i'th subtree. 507 */ 508 r = blst_meta_alloc(&scan[i], 509 cursor > blk ? cursor : blk, count, radix); 510 if (r != SWAPBLK_NONE) { 511 scan->u.bmu_avail -= count; 512 return (r); 513 } 514 } else if (scan[i].bm_bighint == (daddr_t)-1) { 515 /* 516 * Terminator 517 */ 518 break; 519 } 520 blk += radix; 521 } 522 523 /* 524 * We couldn't allocate count in this subtree, update bighint. 525 */ 526 if (scan_from_start && scan->bm_bighint >= count) 527 scan->bm_bighint = count - 1; 528 529 return (SWAPBLK_NONE); 530 } 531 532 /* 533 * BLST_LEAF_FREE() - free allocated block from leaf bitmap 534 * 535 */ 536 static void 537 blst_leaf_free(blmeta_t *scan, daddr_t blk, int count) 538 { 539 /* 540 * free some data in this bitmap 541 * 542 * e.g. 543 * 0000111111111110000 544 * \_________/\__/ 545 * v n 546 */ 547 int n = blk & (BLIST_BMAP_RADIX - 1); 548 u_daddr_t mask; 549 550 mask = ((u_daddr_t)-1 << n) & 551 ((u_daddr_t)-1 >> (BLIST_BMAP_RADIX - count - n)); 552 553 if (scan->u.bmu_bitmap & mask) 554 panic("blst_radix_free: freeing free block"); 555 scan->u.bmu_bitmap |= mask; 556 557 /* 558 * We could probably do a better job here. We are required to make 559 * bighint at least as large as the biggest contiguous block of 560 * data. If we just shoehorn it, a little extra overhead will 561 * be incured on the next allocation (but only that one typically). 562 */ 563 scan->bm_bighint = BLIST_BMAP_RADIX; 564 } 565 566 /* 567 * BLST_META_FREE() - free allocated blocks from radix tree meta info 568 * 569 * This support routine frees a range of blocks from the bitmap. 570 * The range must be entirely enclosed by this radix node. If a 571 * meta node, we break the range down recursively to free blocks 572 * in subnodes (which means that this code can free an arbitrary 573 * range whereas the allocation code cannot allocate an arbitrary 574 * range). 575 */ 576 static void 577 blst_meta_free(blmeta_t *scan, daddr_t freeBlk, daddr_t count, u_daddr_t radix) 578 { 579 daddr_t blk, i, next_skip, skip, v; 580 int child; 581 582 if (scan->bm_bighint == (daddr_t)-1) 583 panic("freeing invalid range"); 584 if (radix == BLIST_BMAP_RADIX) 585 return (blst_leaf_free(scan, freeBlk, count)); 586 skip = radix_to_skip(radix); 587 next_skip = skip / BLIST_META_RADIX; 588 589 if (scan->u.bmu_avail == 0) { 590 /* 591 * ALL-ALLOCATED special case, with possible 592 * shortcut to ALL-FREE special case. 593 */ 594 scan->u.bmu_avail = count; 595 scan->bm_bighint = count; 596 597 if (count != radix) { 598 for (i = 1; i < skip; i += next_skip) { 599 if (scan[i].bm_bighint == (daddr_t)-1) 600 break; 601 scan[i].bm_bighint = 0; 602 if (next_skip == 1) { 603 scan[i].u.bmu_bitmap = 0; 604 } else { 605 scan[i].u.bmu_avail = 0; 606 } 607 } 608 /* fall through */ 609 } 610 } else { 611 scan->u.bmu_avail += count; 612 /* scan->bm_bighint = radix; */ 613 } 614 615 /* 616 * ALL-FREE special case. 617 */ 618 619 if (scan->u.bmu_avail == radix) 620 return; 621 if (scan->u.bmu_avail > radix) 622 panic("blst_meta_free: freeing already free blocks (%lld) %lld/%lld", 623 (long long)count, (long long)scan->u.bmu_avail, 624 (long long)radix); 625 626 /* 627 * Break the free down into its components 628 */ 629 630 blk = freeBlk & -radix; 631 radix /= BLIST_META_RADIX; 632 633 child = (freeBlk - blk) / radix; 634 blk += child * radix; 635 i = 1 + child * next_skip; 636 while (i < skip && blk < freeBlk + count) { 637 v = blk + radix - freeBlk; 638 if (v > count) 639 v = count; 640 blst_meta_free(&scan[i], freeBlk, v, radix); 641 if (scan->bm_bighint < scan[i].bm_bighint) 642 scan->bm_bighint = scan[i].bm_bighint; 643 count -= v; 644 freeBlk += v; 645 blk += radix; 646 i += next_skip; 647 } 648 } 649 650 /* 651 * BLIST_RADIX_COPY() - copy one radix tree to another 652 * 653 * Locates free space in the source tree and frees it in the destination 654 * tree. The space may not already be free in the destination. 655 */ 656 static void 657 blst_copy(blmeta_t *scan, daddr_t blk, daddr_t radix, blist_t dest, 658 daddr_t count) 659 { 660 daddr_t i, next_skip, skip; 661 662 /* 663 * Leaf node 664 */ 665 666 if (radix == BLIST_BMAP_RADIX) { 667 u_daddr_t v = scan->u.bmu_bitmap; 668 669 if (v == (u_daddr_t)-1) { 670 blist_free(dest, blk, count); 671 } else if (v != 0) { 672 int i; 673 674 for (i = 0; i < BLIST_BMAP_RADIX && i < count; ++i) { 675 if (v & ((u_daddr_t)1 << i)) 676 blist_free(dest, blk + i, 1); 677 } 678 } 679 return; 680 } 681 682 /* 683 * Meta node 684 */ 685 686 if (scan->u.bmu_avail == 0) { 687 /* 688 * Source all allocated, leave dest allocated 689 */ 690 return; 691 } 692 if (scan->u.bmu_avail == radix) { 693 /* 694 * Source all free, free entire dest 695 */ 696 if (count < radix) 697 blist_free(dest, blk, count); 698 else 699 blist_free(dest, blk, radix); 700 return; 701 } 702 703 704 skip = radix_to_skip(radix); 705 next_skip = skip / BLIST_META_RADIX; 706 radix /= 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, dest, radix); 714 count -= radix; 715 } else { 716 if (count) { 717 blst_copy(&scan[i], blk, radix, dest, count); 718 } 719 count = 0; 720 } 721 blk += radix; 722 } 723 } 724 725 /* 726 * BLST_LEAF_FILL() - allocate specific blocks in leaf bitmap 727 * 728 * This routine allocates all blocks in the specified range 729 * regardless of any existing allocations in that range. Returns 730 * the number of blocks allocated by the call. 731 */ 732 static daddr_t 733 blst_leaf_fill(blmeta_t *scan, daddr_t blk, int count) 734 { 735 int n = blk & (BLIST_BMAP_RADIX - 1); 736 daddr_t nblks; 737 u_daddr_t mask; 738 739 mask = ((u_daddr_t)-1 << n) & 740 ((u_daddr_t)-1 >> (BLIST_BMAP_RADIX - count - n)); 741 742 /* Count the number of blocks that we are allocating. */ 743 nblks = bitcount64(scan->u.bmu_bitmap & mask); 744 745 scan->u.bmu_bitmap &= ~mask; 746 return (nblks); 747 } 748 749 /* 750 * BLIST_META_FILL() - allocate specific blocks at a meta node 751 * 752 * This routine allocates the specified range of blocks, 753 * regardless of any existing allocations in the range. The 754 * range must be within the extent of this node. Returns the 755 * number of blocks allocated by the call. 756 */ 757 static daddr_t 758 blst_meta_fill(blmeta_t *scan, daddr_t allocBlk, daddr_t count, u_daddr_t radix) 759 { 760 daddr_t blk, i, nblks, next_skip, skip, v; 761 int child; 762 763 if (scan->bm_bighint == (daddr_t)-1) 764 panic("filling invalid range"); 765 if (count > radix) { 766 /* 767 * The allocation exceeds the number of blocks that are 768 * managed by this node. 769 */ 770 panic("fill too large"); 771 } 772 if (radix == BLIST_BMAP_RADIX) 773 return (blst_leaf_fill(scan, allocBlk, count)); 774 if (count == radix || scan->u.bmu_avail == 0) { 775 /* 776 * ALL-ALLOCATED special case 777 */ 778 nblks = scan->u.bmu_avail; 779 scan->u.bmu_avail = 0; 780 scan->bm_bighint = 0; 781 return (nblks); 782 } 783 skip = radix_to_skip(radix); 784 next_skip = skip / BLIST_META_RADIX; 785 blk = allocBlk & -radix; 786 787 /* 788 * An ALL-FREE meta node requires special handling before allocating 789 * any of its blocks. 790 */ 791 if (scan->u.bmu_avail == radix) { 792 radix /= BLIST_META_RADIX; 793 794 /* 795 * Reinitialize each of the meta node's children. An ALL-FREE 796 * meta node cannot have a terminator in any subtree. 797 */ 798 for (i = 1; i < skip; i += next_skip) { 799 if (next_skip == 1) 800 scan[i].u.bmu_bitmap = (u_daddr_t)-1; 801 else 802 scan[i].u.bmu_avail = radix; 803 scan[i].bm_bighint = radix; 804 } 805 } else { 806 radix /= BLIST_META_RADIX; 807 } 808 809 nblks = 0; 810 child = (allocBlk - blk) / radix; 811 blk += child * radix; 812 i = 1 + child * next_skip; 813 while (i < skip && blk < allocBlk + count) { 814 v = blk + radix - allocBlk; 815 if (v > count) 816 v = count; 817 nblks += blst_meta_fill(&scan[i], allocBlk, v, radix); 818 count -= v; 819 allocBlk += v; 820 blk += radix; 821 i += next_skip; 822 } 823 scan->u.bmu_avail -= nblks; 824 return (nblks); 825 } 826 827 /* 828 * BLST_RADIX_INIT() - initialize radix tree 829 * 830 * Initialize our meta structures and bitmaps and calculate the exact 831 * amount of space required to manage 'count' blocks - this space may 832 * be considerably less than the calculated radix due to the large 833 * RADIX values we use. 834 */ 835 static daddr_t 836 blst_radix_init(blmeta_t *scan, daddr_t radix, daddr_t count) 837 { 838 daddr_t i, memindex, next_skip, skip; 839 840 memindex = 0; 841 842 /* 843 * Leaf node 844 */ 845 846 if (radix == BLIST_BMAP_RADIX) { 847 if (scan) { 848 scan->bm_bighint = 0; 849 scan->u.bmu_bitmap = 0; 850 } 851 return (memindex); 852 } 853 854 /* 855 * Meta node. If allocating the entire object we can special 856 * case it. However, we need to figure out how much memory 857 * is required to manage 'count' blocks, so we continue on anyway. 858 */ 859 860 if (scan) { 861 scan->bm_bighint = 0; 862 scan->u.bmu_avail = 0; 863 } 864 865 skip = radix_to_skip(radix); 866 next_skip = skip / BLIST_META_RADIX; 867 radix /= BLIST_META_RADIX; 868 869 for (i = 1; i < skip; i += next_skip) { 870 if (count >= radix) { 871 /* 872 * Allocate the entire object 873 */ 874 memindex = i + 875 blst_radix_init(((scan) ? &scan[i] : NULL), radix, 876 radix); 877 count -= radix; 878 } else if (count > 0) { 879 /* 880 * Allocate a partial object 881 */ 882 memindex = i + 883 blst_radix_init(((scan) ? &scan[i] : NULL), radix, 884 count); 885 count = 0; 886 } else { 887 /* 888 * Add terminator and break out 889 */ 890 if (scan) 891 scan[i].bm_bighint = (daddr_t)-1; 892 break; 893 } 894 } 895 if (memindex < i) 896 memindex = i; 897 return (memindex); 898 } 899 900 #ifdef BLIST_DEBUG 901 902 static void 903 blst_radix_print(blmeta_t *scan, daddr_t blk, daddr_t radix, int tab) 904 { 905 daddr_t i, next_skip, skip; 906 907 if (radix == BLIST_BMAP_RADIX) { 908 printf( 909 "%*.*s(%08llx,%lld): bitmap %016llx big=%lld\n", 910 tab, tab, "", 911 (long long)blk, (long long)radix, 912 (long long)scan->u.bmu_bitmap, 913 (long long)scan->bm_bighint 914 ); 915 return; 916 } 917 918 if (scan->u.bmu_avail == 0) { 919 printf( 920 "%*.*s(%08llx,%lld) ALL ALLOCATED\n", 921 tab, tab, "", 922 (long long)blk, 923 (long long)radix 924 ); 925 return; 926 } 927 if (scan->u.bmu_avail == radix) { 928 printf( 929 "%*.*s(%08llx,%lld) ALL FREE\n", 930 tab, tab, "", 931 (long long)blk, 932 (long long)radix 933 ); 934 return; 935 } 936 937 printf( 938 "%*.*s(%08llx,%lld): subtree (%lld/%lld) big=%lld {\n", 939 tab, tab, "", 940 (long long)blk, (long long)radix, 941 (long long)scan->u.bmu_avail, 942 (long long)radix, 943 (long long)scan->bm_bighint 944 ); 945 946 skip = radix_to_skip(radix); 947 next_skip = skip / BLIST_META_RADIX; 948 radix /= BLIST_META_RADIX; 949 tab += 4; 950 951 for (i = 1; i < skip; i += next_skip) { 952 if (scan[i].bm_bighint == (daddr_t)-1) { 953 printf( 954 "%*.*s(%08llx,%lld): Terminator\n", 955 tab, tab, "", 956 (long long)blk, (long long)radix 957 ); 958 break; 959 } 960 blst_radix_print(&scan[i], blk, radix, tab); 961 blk += radix; 962 } 963 tab -= 4; 964 965 printf( 966 "%*.*s}\n", 967 tab, tab, "" 968 ); 969 } 970 971 #endif 972 973 #ifdef BLIST_DEBUG 974 975 int 976 main(int ac, char **av) 977 { 978 int size = 1024; 979 int i; 980 blist_t bl; 981 982 for (i = 1; i < ac; ++i) { 983 const char *ptr = av[i]; 984 if (*ptr != '-') { 985 size = strtol(ptr, NULL, 0); 986 continue; 987 } 988 ptr += 2; 989 fprintf(stderr, "Bad option: %s\n", ptr - 2); 990 exit(1); 991 } 992 bl = blist_create(size, M_WAITOK); 993 blist_free(bl, 0, size); 994 995 for (;;) { 996 char buf[1024]; 997 long long da = 0; 998 long long count = 0; 999 1000 printf("%lld/%lld/%lld> ", (long long)blist_avail(bl), 1001 (long long)size, (long long)bl->bl_radix); 1002 fflush(stdout); 1003 if (fgets(buf, sizeof(buf), stdin) == NULL) 1004 break; 1005 switch(buf[0]) { 1006 case 'r': 1007 if (sscanf(buf + 1, "%lld", &count) == 1) { 1008 blist_resize(&bl, count, 1, M_WAITOK); 1009 } else { 1010 printf("?\n"); 1011 } 1012 case 'p': 1013 blist_print(bl); 1014 break; 1015 case 'a': 1016 if (sscanf(buf + 1, "%lld", &count) == 1) { 1017 daddr_t blk = blist_alloc(bl, count); 1018 printf(" R=%08llx\n", (long long)blk); 1019 } else { 1020 printf("?\n"); 1021 } 1022 break; 1023 case 'f': 1024 if (sscanf(buf + 1, "%llx %lld", &da, &count) == 2) { 1025 blist_free(bl, da, count); 1026 } else { 1027 printf("?\n"); 1028 } 1029 break; 1030 case 'l': 1031 if (sscanf(buf + 1, "%llx %lld", &da, &count) == 2) { 1032 printf(" n=%jd\n", 1033 (intmax_t)blist_fill(bl, da, count)); 1034 } else { 1035 printf("?\n"); 1036 } 1037 break; 1038 case '?': 1039 case 'h': 1040 puts( 1041 "p -print\n" 1042 "a %d -allocate\n" 1043 "f %x %d -free\n" 1044 "l %x %d -fill\n" 1045 "r %d -resize\n" 1046 "h/? -help" 1047 ); 1048 break; 1049 default: 1050 printf("?\n"); 1051 break; 1052 } 1053 } 1054 return(0); 1055 } 1056 1057 void 1058 panic(const char *ctl, ...) 1059 { 1060 va_list va; 1061 1062 va_start(va, ctl); 1063 vfprintf(stderr, ctl, va); 1064 fprintf(stderr, "\n"); 1065 va_end(va); 1066 exit(1); 1067 } 1068 1069 #endif 1070