| 1 | /* |
| 2 | * BLIST.C - Bitmap allocator/deallocator, using a radix tree with hinting |
| 3 | * |
| 4 | * (c)Copyright 1998, Matthew Dillon. Terms for use and redistribution |
| 5 | * are covered by the BSD Copyright as found in /usr/src/COPYRIGHT. |
| 6 | * |
| 7 | * This module implements a general bitmap allocator/deallocator. The |
| 8 | * allocator eats around 2 bits per 'block'. The module does not |
| 9 | * try to interpret the meaning of a 'block' other then to return |
| 10 | * SWAPBLK_NONE on an allocation failure. |
| 11 | * |
| 12 | * A radix tree is used to maintain the bitmap. Two radix constants are |
| 13 | * involved: One for the bitmaps contained in the leaf nodes (typically |
| 14 | * 32), and one for the meta nodes (typically 16). Both meta and leaf |
| 15 | * nodes have a hint field. This field gives us a hint as to the largest |
| 16 | * free contiguous range of blocks under the node. It may contain a |
| 17 | * value that is too high, but will never contain a value that is too |
| 18 | * low. When the radix tree is searched, allocation failures in subtrees |
| 19 | * update the hint. |
| 20 | * |
| 21 | * The radix tree also implements two collapsed states for meta nodes: |
| 22 | * the ALL-ALLOCATED state and the ALL-FREE state. If a meta node is |
| 23 | * in either of these two states, all information contained underneath |
| 24 | * the node is considered stale. These states are used to optimize |
| 25 | * allocation and freeing operations. |
| 26 | * |
| 27 | * The hinting greatly increases code efficiency for allocations while |
| 28 | * the general radix structure optimizes both allocations and frees. The |
| 29 | * radix tree should be able to operate well no matter how much |
| 30 | * fragmentation there is and no matter how large a bitmap is used. |
| 31 | * |
| 32 | * Unlike the rlist code, the blist code wires all necessary memory at |
| 33 | * creation time. Neither allocations nor frees require interaction with |
| 34 | * the memory subsystem. In contrast, the rlist code may allocate memory |
| 35 | * on an rlist_free() call. The non-blocking features of the blist code |
| 36 | * are used to great advantage in the swap code (vm/nswap_pager.c). The |
| 37 | * rlist code uses a little less overall memory then the blist code (but |
| 38 | * due to swap interleaving not all that much less), but the blist code |
| 39 | * scales much, much better. |
| 40 | * |
| 41 | * LAYOUT: The radix tree is layed out recursively using a |
| 42 | * linear array. Each meta node is immediately followed (layed out |
| 43 | * sequentially in memory) by BLIST_META_RADIX lower level nodes. This |
| 44 | * is a recursive structure but one that can be easily scanned through |
| 45 | * a very simple 'skip' calculation. In order to support large radixes, |
| 46 | * portions of the tree may reside outside our memory allocation. We |
| 47 | * handle this with an early-termination optimization (when bighint is |
| 48 | * set to -1) on the scan. The memory allocation is only large enough |
| 49 | * to cover the number of blocks requested at creation time even if it |
| 50 | * must be encompassed in larger root-node radix. |
| 51 | * |
| 52 | * NOTE: the allocator cannot currently allocate more then |
| 53 | * BLIST_BMAP_RADIX blocks per call. It will panic with 'allocation too |
| 54 | * large' if you try. This is an area that could use improvement. The |
| 55 | * radix is large enough that this restriction does not effect the swap |
| 56 | * system, though. Currently only the allocation code is effected by |
| 57 | * this algorithmic unfeature. The freeing code can handle arbitrary |
| 58 | * ranges. |
| 59 | * |
| 60 | * This code can be compiled stand-alone for debugging. |
| 61 | * |
| 62 | * $FreeBSD: src/sys/kern/subr_blist.c,v 1.5.2.1 2000/03/17 10:47:29 ps Exp $ |
| 63 | */ |
| 64 | |
| 65 | #if !defined(__APPLE__) |
| 66 | #ifdef _KERNEL |
| 67 | |
| 68 | #include <sys/param.h> |
| 69 | #include <sys/systm.h> |
| 70 | #include <sys/lock.h> |
| 71 | #include <sys/kernel.h> |
| 72 | #include <sys/blist.h> |
| 73 | #include <sys/malloc.h> |
| 74 | #include <vm/vm.h> |
| 75 | #include <vm/vm_object.h> |
| 76 | #include <vm/vm_kern.h> |
| 77 | #include <vm/vm_extern.h> |
| 78 | #include <vm/vm_page.h> |
| 79 | |
| 80 | #else |
| 81 | |
| 82 | #ifndef BLIST_NO_DEBUG |
| 83 | #define BLIST_DEBUG |
| 84 | #endif |
| 85 | |
| 86 | #define SWAPBLK_NONE ((daddr_t)-1) |
| 87 | |
| 88 | #include <sys/types.h> |
| 89 | #include <stdio.h> |
| 90 | #include <string.h> |
| 91 | #include <stdlib.h> |
| 92 | #include <stdarg.h> |
| 93 | |
| 94 | #define malloc(a,b,c) malloc(a) |
| 95 | #define free(a,b) free(a) |
| 96 | |
| 97 | typedef unsigned int u_daddr_t; |
| 98 | |
| 99 | #include <sys/blist.h> |
| 100 | |
| 101 | void panic(const char *ctl, ...); |
| 102 | |
| 103 | #endif |
| 104 | #else /* is MacOS X */ |
| 105 | #ifdef KERNEL |
| 106 | #ifndef _KERNEL |
| 107 | #define _KERNEL /* Solaris vs. Darwin */ |
| 108 | #endif |
| 109 | #endif |
| 110 | |
| 111 | typedef unsigned int u_daddr_t; |
| 112 | |
| 113 | #include <sys/param.h> |
| 114 | #include <sys/systm.h> |
| 115 | #include <sys/lock.h> |
| 116 | #include <sys/kernel.h> |
| 117 | /* #include <sys/blist.h> */ |
| 118 | #include "blist.h" |
| 119 | #include <sys/malloc.h> |
| 120 | |
| 121 | #define SWAPBLK_NONE ((daddr_t)-1) |
| 122 | #define malloc _MALLOC |
| 123 | #define free _FREE |
| 124 | #define M_SWAP M_TEMP |
| 125 | |
| 126 | #endif /* __APPLE__ */ |
| 127 | |
| 128 | /* |
| 129 | * static support functions |
| 130 | */ |
| 131 | |
| 132 | static daddr_t blst_leaf_alloc(blmeta_t *scan, daddr_t blk, int count); |
| 133 | static daddr_t blst_meta_alloc(blmeta_t *scan, daddr_t blk, |
| 134 | daddr_t count, daddr_t radix, int skip); |
| 135 | static void blst_leaf_free(blmeta_t *scan, daddr_t relblk, int count); |
| 136 | static void blst_meta_free(blmeta_t *scan, daddr_t freeBlk, daddr_t count, |
| 137 | daddr_t radix, int skip, daddr_t blk); |
| 138 | static void blst_copy(blmeta_t *scan, daddr_t blk, daddr_t radix, |
| 139 | daddr_t skip, blist_t dest, daddr_t count); |
| 140 | static daddr_t blst_radix_init(blmeta_t *scan, daddr_t radix, |
| 141 | int skip, daddr_t count); |
| 142 | #ifndef _KERNEL |
| 143 | static void blst_radix_print(blmeta_t *scan, daddr_t blk, |
| 144 | daddr_t radix, int skip, int tab); |
| 145 | #endif |
| 146 | |
| 147 | #if !defined(__APPLE__) |
| 148 | #ifdef _KERNEL |
| 149 | static MALLOC_DEFINE(M_SWAP, "SWAP" , "Swap space" ); |
| 150 | #endif |
| 151 | #endif /* __APPLE__ */ |
| 152 | |
| 153 | /* |
| 154 | * blist_create() - create a blist capable of handling up to the specified |
| 155 | * number of blocks |
| 156 | * |
| 157 | * blocks must be greater then 0 |
| 158 | * |
| 159 | * The smallest blist consists of a single leaf node capable of |
| 160 | * managing BLIST_BMAP_RADIX blocks. |
| 161 | */ |
| 162 | |
| 163 | blist_t |
| 164 | blist_create(daddr_t blocks) |
| 165 | { |
| 166 | blist_t bl; |
| 167 | int radix; |
| 168 | int skip = 0; |
| 169 | |
| 170 | /* |
| 171 | * Calculate radix and skip field used for scanning. |
| 172 | */ |
| 173 | radix = BLIST_BMAP_RADIX; |
| 174 | |
| 175 | while (radix < blocks) { |
| 176 | radix <<= BLIST_META_RADIX_SHIFT; |
| 177 | skip = (skip + 1) << BLIST_META_RADIX_SHIFT; |
| 178 | } |
| 179 | |
| 180 | bl = malloc(sizeof(struct blist), M_SWAP, M_WAITOK); |
| 181 | |
| 182 | bzero(bl, sizeof(*bl)); |
| 183 | |
| 184 | bl->bl_blocks = blocks; |
| 185 | bl->bl_radix = radix; |
| 186 | bl->bl_skip = skip; |
| 187 | bl->bl_rootblks = 1 + |
| 188 | blst_radix_init(NULL, bl->bl_radix, bl->bl_skip, blocks); |
| 189 | bl->bl_root = malloc(sizeof(blmeta_t) * bl->bl_rootblks, M_SWAP, M_WAITOK); |
| 190 | |
| 191 | #if defined(BLIST_DEBUG) |
| 192 | printf( |
| 193 | "BLIST representing %d blocks (%d MB of swap)" |
| 194 | ", requiring %dK of ram\n" , |
| 195 | bl->bl_blocks, |
| 196 | bl->bl_blocks * 4 / 1024, |
| 197 | (bl->bl_rootblks * sizeof(blmeta_t) + 1023) / 1024 |
| 198 | ); |
| 199 | printf("BLIST raw radix tree contains %d records\n" , bl->bl_rootblks); |
| 200 | #endif |
| 201 | blst_radix_init(bl->bl_root, bl->bl_radix, bl->bl_skip, blocks); |
| 202 | |
| 203 | return(bl); |
| 204 | } |
| 205 | |
| 206 | void |
| 207 | blist_destroy(blist_t bl) |
| 208 | { |
| 209 | free(bl->bl_root, M_SWAP); |
| 210 | free(bl, M_SWAP); |
| 211 | } |
| 212 | |
| 213 | /* |
| 214 | * blist_alloc() - reserve space in the block bitmap. Return the base |
| 215 | * of a contiguous region or SWAPBLK_NONE if space could |
| 216 | * not be allocated. |
| 217 | */ |
| 218 | |
| 219 | daddr_t |
| 220 | blist_alloc(blist_t bl, daddr_t count) |
| 221 | { |
| 222 | daddr_t blk = SWAPBLK_NONE; |
| 223 | |
| 224 | if (bl) { |
| 225 | if (bl->bl_radix == BLIST_BMAP_RADIX) |
| 226 | blk = blst_leaf_alloc(bl->bl_root, 0, count); |
| 227 | else |
| 228 | blk = blst_meta_alloc(bl->bl_root, 0, count, |
| 229 | bl->bl_radix, bl->bl_skip); |
| 230 | if (blk != SWAPBLK_NONE) |
| 231 | bl->bl_free -= count; |
| 232 | } |
| 233 | return(blk); |
| 234 | } |
| 235 | |
| 236 | /* |
| 237 | * blist_free() - free up space in the block bitmap. Return the base |
| 238 | * of a contiguous region. Panic if an inconsistancy is |
| 239 | * found. |
| 240 | */ |
| 241 | |
| 242 | void |
| 243 | blist_free(blist_t bl, daddr_t blkno, daddr_t count) |
| 244 | { |
| 245 | if (bl) { |
| 246 | if (bl->bl_radix == BLIST_BMAP_RADIX) |
| 247 | blst_leaf_free(bl->bl_root, blkno, count); |
| 248 | else |
| 249 | blst_meta_free(bl->bl_root, blkno, count, |
| 250 | bl->bl_radix, bl->bl_skip, 0); |
| 251 | bl->bl_free += count; |
| 252 | } |
| 253 | } |
| 254 | |
| 255 | /* |
| 256 | * blist_resize() - resize an existing radix tree to handle the |
| 257 | * specified number of blocks. This will reallocate |
| 258 | * the tree and transfer the previous bitmap to the new |
| 259 | * one. When extending the tree you can specify whether |
| 260 | * the new blocks are to left allocated or freed. |
| 261 | */ |
| 262 | |
| 263 | void |
| 264 | blist_resize(blist_t *pbl, daddr_t count, int freenew) |
| 265 | { |
| 266 | blist_t newbl = blist_create(count); |
| 267 | blist_t save = *pbl; |
| 268 | |
| 269 | *pbl = newbl; |
| 270 | if (count > save->bl_blocks) |
| 271 | count = save->bl_blocks; |
| 272 | blst_copy(save->bl_root, 0, save->bl_radix, save->bl_skip, newbl, count); |
| 273 | |
| 274 | /* |
| 275 | * If resizing upwards, should we free the new space or not? |
| 276 | */ |
| 277 | if (freenew && count < newbl->bl_blocks) |
| 278 | blist_free(newbl, count, newbl->bl_blocks - count); |
| 279 | blist_destroy(save); |
| 280 | } |
| 281 | |
| 282 | #ifdef BLIST_DEBUG |
| 283 | |
| 284 | /* |
| 285 | * blist_print() - dump radix tree |
| 286 | */ |
| 287 | |
| 288 | void |
| 289 | blist_print(blist_t bl) |
| 290 | { |
| 291 | printf("BLIST {\n" ); |
| 292 | blst_radix_print(bl->bl_root, 0, bl->bl_radix, bl->bl_skip, 4); |
| 293 | printf("}\n" ); |
| 294 | } |
| 295 | |
| 296 | #endif |
| 297 | |
| 298 | /************************************************************************ |
| 299 | * ALLOCATION SUPPORT FUNCTIONS * |
| 300 | ************************************************************************ |
| 301 | * |
| 302 | * These support functions do all the actual work. They may seem |
| 303 | * rather longish, but that's because I've commented them up. The |
| 304 | * actual code is straight forward. |
| 305 | * |
| 306 | */ |
| 307 | |
| 308 | /* |
| 309 | * blist_leaf_alloc() - allocate at a leaf in the radix tree (a bitmap). |
| 310 | * |
| 311 | * This is the core of the allocator and is optimized for the 1 block |
| 312 | * and the BLIST_BMAP_RADIX block allocation cases. Other cases are |
| 313 | * somewhat slower. The 1 block allocation case is log2 and extremely |
| 314 | * quick. |
| 315 | */ |
| 316 | |
| 317 | static daddr_t |
| 318 | blst_leaf_alloc(blmeta_t *scan, daddr_t blk, int count) |
| 319 | { |
| 320 | u_daddr_t orig = scan->u.bmu_bitmap; |
| 321 | |
| 322 | if (orig == 0) { |
| 323 | /* |
| 324 | * Optimize bitmap all-allocated case. Also, count = 1 |
| 325 | * case assumes at least 1 bit is free in the bitmap, so |
| 326 | * we have to take care of this case here. |
| 327 | */ |
| 328 | scan->bm_bighint = 0; |
| 329 | return(SWAPBLK_NONE); |
| 330 | } |
| 331 | if (count == 1) { |
| 332 | /* |
| 333 | * Optimized code to allocate one bit out of the bitmap |
| 334 | */ |
| 335 | u_daddr_t mask; |
| 336 | int j = BLIST_BMAP_RADIX/2; |
| 337 | int r = 0; |
| 338 | |
| 339 | mask = (u_daddr_t)-1 >> (BLIST_BMAP_RADIX/2); |
| 340 | |
| 341 | while (j) { |
| 342 | if ((orig & mask) == 0) { |
| 343 | r += j; |
| 344 | orig >>= j; |
| 345 | } |
| 346 | j >>= 1; |
| 347 | mask >>= j; |
| 348 | } |
| 349 | scan->u.bmu_bitmap &= ~(1 << r); |
| 350 | return(blk + r); |
| 351 | } |
| 352 | #if !defined(__APPLE__) |
| 353 | if (count <= BLIST_BMAP_RADIX) { |
| 354 | #else |
| 355 | if (count <= (int)BLIST_BMAP_RADIX) { |
| 356 | #endif /* __APPLE__ */ |
| 357 | /* |
| 358 | * non-optimized code to allocate N bits out of the bitmap. |
| 359 | * The more bits, the faster the code runs. It will run |
| 360 | * the slowest allocating 2 bits, but since there aren't any |
| 361 | * memory ops in the core loop (or shouldn't be, anyway), |
| 362 | * you probably won't notice the difference. |
| 363 | */ |
| 364 | int j; |
| 365 | int n = BLIST_BMAP_RADIX - count; |
| 366 | u_daddr_t mask; |
| 367 | |
| 368 | mask = (u_daddr_t)-1 >> n; |
| 369 | |
| 370 | for (j = 0; j <= n; ++j) { |
| 371 | if ((orig & mask) == mask) { |
| 372 | scan->u.bmu_bitmap &= ~mask; |
| 373 | return(blk + j); |
| 374 | } |
| 375 | mask = (mask << 1); |
| 376 | } |
| 377 | } |
| 378 | /* |
| 379 | * We couldn't allocate count in this subtree, update bighint. |
| 380 | */ |
| 381 | scan->bm_bighint = count - 1; |
| 382 | return(SWAPBLK_NONE); |
| 383 | } |
| 384 | |
| 385 | /* |
| 386 | * blist_meta_alloc() - allocate at a meta in the radix tree. |
| 387 | * |
| 388 | * Attempt to allocate at a meta node. If we can't, we update |
| 389 | * bighint and return a failure. Updating bighint optimize future |
| 390 | * calls that hit this node. We have to check for our collapse cases |
| 391 | * and we have a few optimizations strewn in as well. |
| 392 | */ |
| 393 | |
| 394 | static daddr_t |
| 395 | blst_meta_alloc(blmeta_t *scan, daddr_t blk, daddr_t count, daddr_t radix, |
| 396 | int skip) |
| 397 | { |
| 398 | int i; |
| 399 | int next_skip = (skip >> BLIST_META_RADIX_SHIFT); |
| 400 | |
| 401 | if (scan->u.bmu_avail == 0) { |
| 402 | /* |
| 403 | * ALL-ALLOCATED special case |
| 404 | */ |
| 405 | scan->bm_bighint = count; |
| 406 | return(SWAPBLK_NONE); |
| 407 | } |
| 408 | |
| 409 | if (scan->u.bmu_avail == radix) { |
| 410 | radix >>= BLIST_META_RADIX_SHIFT; |
| 411 | |
| 412 | /* |
| 413 | * ALL-FREE special case, initialize uninitialize |
| 414 | * sublevel. |
| 415 | */ |
| 416 | for (i = 1; i <= skip; i += next_skip) { |
| 417 | if (scan[i].bm_bighint == (daddr_t)-1) |
| 418 | break; |
| 419 | if (next_skip == 1) { |
| 420 | scan[i].u.bmu_bitmap = (u_daddr_t)-1; |
| 421 | scan[i].bm_bighint = BLIST_BMAP_RADIX; |
| 422 | } else { |
| 423 | scan[i].bm_bighint = radix; |
| 424 | scan[i].u.bmu_avail = radix; |
| 425 | } |
| 426 | } |
| 427 | } else { |
| 428 | radix >>= BLIST_META_RADIX_SHIFT; |
| 429 | } |
| 430 | |
| 431 | for (i = 1; i <= skip; i += next_skip) { |
| 432 | if (count <= scan[i].bm_bighint) { |
| 433 | /* |
| 434 | * count fits in object |
| 435 | */ |
| 436 | daddr_t r; |
| 437 | if (next_skip == 1) |
| 438 | r = blst_leaf_alloc(&scan[i], blk, count); |
| 439 | else |
| 440 | r = blst_meta_alloc(&scan[i], blk, count, |
| 441 | radix, next_skip - 1); |
| 442 | if (r != SWAPBLK_NONE) { |
| 443 | scan->u.bmu_avail -= count; |
| 444 | if (scan->bm_bighint > scan->u.bmu_avail) |
| 445 | scan->bm_bighint = scan->u.bmu_avail; |
| 446 | return r; |
| 447 | } |
| 448 | } else if (scan[i].bm_bighint == (daddr_t)-1) { |
| 449 | /* |
| 450 | * Terminator |
| 451 | */ |
| 452 | break; |
| 453 | } else if (count > radix) { |
| 454 | /* |
| 455 | * count does not fit in object even if it were |
| 456 | * complete free. |
| 457 | */ |
| 458 | panic("blist_meta_alloc: allocation too large" ); |
| 459 | } |
| 460 | blk += radix; |
| 461 | } |
| 462 | |
| 463 | /* |
| 464 | * We couldn't allocate count in this subtree, update bighint. |
| 465 | */ |
| 466 | if (scan->bm_bighint >= count) |
| 467 | scan->bm_bighint = count - 1; |
| 468 | return(SWAPBLK_NONE); |
| 469 | } |
| 470 | |
| 471 | /* |
| 472 | * BLST_LEAF_FREE() - free allocated block from leaf bitmap |
| 473 | * |
| 474 | */ |
| 475 | |
| 476 | static void |
| 477 | blst_leaf_free(blmeta_t *scan, daddr_t blk, int count) |
| 478 | { |
| 479 | /* |
| 480 | * free some data in this bitmap |
| 481 | * |
| 482 | * e.g. |
| 483 | * 0000111111111110000 |
| 484 | * \_________/\__/ |
| 485 | * v n |
| 486 | */ |
| 487 | int n = blk & (BLIST_BMAP_RADIX - 1); |
| 488 | u_daddr_t mask; |
| 489 | |
| 490 | mask = ((u_daddr_t)-1 << n) & |
| 491 | ((u_daddr_t)-1 >> (BLIST_BMAP_RADIX - count - n)); |
| 492 | |
| 493 | if (scan->u.bmu_bitmap & mask) |
| 494 | panic("blst_radix_free: freeing free block" ); |
| 495 | scan->u.bmu_bitmap |= mask; |
| 496 | |
| 497 | /* |
| 498 | * We could probably do a better job here. We are required to make |
| 499 | * bighint at least as large as the biggest contiguous block of |
| 500 | * data. If we just shoehorn it, a little extra overhead will |
| 501 | * be incured on the next allocation (but only that one typically). |
| 502 | */ |
| 503 | scan->bm_bighint = BLIST_BMAP_RADIX; |
| 504 | } |
| 505 | |
| 506 | /* |
| 507 | * BLST_META_FREE() - free allocated blocks from radix tree meta info |
| 508 | * |
| 509 | * This support routine frees a range of blocks from the bitmap. |
| 510 | * The range must be entirely enclosed by this radix node. If a |
| 511 | * meta node, we break the range down recursively to free blocks |
| 512 | * in subnodes (which means that this code can free an arbitrary |
| 513 | * range whereas the allocation code cannot allocate an arbitrary |
| 514 | * range). |
| 515 | */ |
| 516 | |
| 517 | static void |
| 518 | blst_meta_free(blmeta_t *scan, daddr_t freeBlk, daddr_t count, daddr_t radix, |
| 519 | int skip, daddr_t blk) |
| 520 | { |
| 521 | int i; |
| 522 | int next_skip = (skip >> BLIST_META_RADIX_SHIFT); |
| 523 | |
| 524 | #if 0 |
| 525 | printf("FREE (%x,%d) FROM (%x,%d)\n" , |
| 526 | freeBlk, count, |
| 527 | blk, radix |
| 528 | ); |
| 529 | #endif |
| 530 | |
| 531 | if (scan->u.bmu_avail == 0) { |
| 532 | /* |
| 533 | * ALL-ALLOCATED special case, with possible |
| 534 | * shortcut to ALL-FREE special case. |
| 535 | */ |
| 536 | scan->u.bmu_avail = count; |
| 537 | scan->bm_bighint = count; |
| 538 | |
| 539 | if (count != radix) { |
| 540 | for (i = 1; i <= skip; i += next_skip) { |
| 541 | if (scan[i].bm_bighint == (daddr_t)-1) |
| 542 | break; |
| 543 | scan[i].bm_bighint = 0; |
| 544 | if (next_skip == 1) |
| 545 | scan[i].u.bmu_bitmap = 0; |
| 546 | else |
| 547 | scan[i].u.bmu_avail = 0; |
| 548 | } |
| 549 | /* fall through */ |
| 550 | } |
| 551 | } else { |
| 552 | scan->u.bmu_avail += count; |
| 553 | /* scan->bm_bighint = radix; */ |
| 554 | } |
| 555 | |
| 556 | /* |
| 557 | * ALL-FREE special case. |
| 558 | */ |
| 559 | |
| 560 | if (scan->u.bmu_avail == radix) |
| 561 | return; |
| 562 | if (scan->u.bmu_avail > radix) |
| 563 | panic("blst_meta_free: freeing already free blocks (%d) %d/%d" , count, scan->u.bmu_avail, radix); |
| 564 | |
| 565 | /* |
| 566 | * Break the free down into its components |
| 567 | */ |
| 568 | |
| 569 | radix >>= BLIST_META_RADIX_SHIFT; |
| 570 | |
| 571 | i = (freeBlk - blk) / radix; |
| 572 | blk += i * radix; |
| 573 | i = i * next_skip + 1; |
| 574 | |
| 575 | while (i <= skip && blk < freeBlk + count) { |
| 576 | daddr_t v; |
| 577 | |
| 578 | v = blk + radix - freeBlk; |
| 579 | if (v > count) |
| 580 | v = count; |
| 581 | |
| 582 | if (scan->bm_bighint == (daddr_t)-1) |
| 583 | panic("blst_meta_free: freeing unexpected range" ); |
| 584 | |
| 585 | if (next_skip == 1) |
| 586 | blst_leaf_free(&scan[i], freeBlk, v); |
| 587 | else |
| 588 | blst_meta_free(&scan[i], freeBlk, v, radix, |
| 589 | next_skip - 1, blk); |
| 590 | if (scan->bm_bighint < scan[i].bm_bighint) |
| 591 | scan->bm_bighint = scan[i].bm_bighint; |
| 592 | count -= v; |
| 593 | freeBlk += v; |
| 594 | blk += radix; |
| 595 | i += next_skip; |
| 596 | } |
| 597 | } |
| 598 | |
| 599 | /* |
| 600 | * BLIST_RADIX_COPY() - copy one radix tree to another |
| 601 | * |
| 602 | * Locates free space in the source tree and frees it in the destination |
| 603 | * tree. The space may not already be free in the destination. |
| 604 | */ |
| 605 | |
| 606 | static void blst_copy(blmeta_t *scan, daddr_t blk, daddr_t radix, |
| 607 | daddr_t skip, blist_t dest, daddr_t count) |
| 608 | { |
| 609 | int next_skip; |
| 610 | int i; |
| 611 | |
| 612 | /* |
| 613 | * Leaf node |
| 614 | */ |
| 615 | |
| 616 | if (radix == BLIST_BMAP_RADIX) { |
| 617 | u_daddr_t v = scan->u.bmu_bitmap; |
| 618 | |
| 619 | if (v == (u_daddr_t)-1) { |
| 620 | blist_free(dest, blk, count); |
| 621 | } else if (v != 0) { |
| 622 | #if !defined(__APPLE__) |
| 623 | int i; |
| 624 | |
| 625 | for (i = 0; i < BLIST_BMAP_RADIX && i < count; ++i) |
| 626 | if (v & (1 << i)) |
| 627 | blist_free(dest, blk + i, 1); |
| 628 | #else |
| 629 | int j; /* Avoid shadow warnings */ |
| 630 | |
| 631 | for (j = 0; j < (int)BLIST_BMAP_RADIX && j < count; ++j) |
| 632 | if (v & (1 << j)) |
| 633 | blist_free(dest, blk + j, 1); |
| 634 | #endif /* __APPLE__ */ |
| 635 | } |
| 636 | return; |
| 637 | } |
| 638 | |
| 639 | /* |
| 640 | * Meta node |
| 641 | */ |
| 642 | |
| 643 | /* |
| 644 | * Source all allocated, leave dest allocated |
| 645 | */ |
| 646 | if (scan->u.bmu_avail == 0) |
| 647 | return; |
| 648 | if (scan->u.bmu_avail == radix) { |
| 649 | /* |
| 650 | * Source all free, free entire dest |
| 651 | */ |
| 652 | if (count < radix) |
| 653 | blist_free(dest, blk, count); |
| 654 | else |
| 655 | blist_free(dest, blk, radix); |
| 656 | return; |
| 657 | } |
| 658 | |
| 659 | radix >>= BLIST_META_RADIX_SHIFT; |
| 660 | next_skip = (skip >> BLIST_META_RADIX_SHIFT); |
| 661 | |
| 662 | for (i = 1; count && i <= skip; i += next_skip) { |
| 663 | if (scan[i].bm_bighint == (daddr_t)-1) |
| 664 | break; |
| 665 | |
| 666 | if (count >= radix) { |
| 667 | blst_copy( |
| 668 | &scan[i], |
| 669 | blk, |
| 670 | radix, |
| 671 | next_skip - 1, |
| 672 | dest, |
| 673 | radix |
| 674 | ); |
| 675 | count -= radix; |
| 676 | } else { |
| 677 | if (count) { |
| 678 | blst_copy( |
| 679 | &scan[i], |
| 680 | blk, |
| 681 | radix, |
| 682 | next_skip - 1, |
| 683 | dest, |
| 684 | count |
| 685 | ); |
| 686 | } |
| 687 | count = 0; |
| 688 | } |
| 689 | blk += radix; |
| 690 | } |
| 691 | } |
| 692 | |
| 693 | /* |
| 694 | * BLST_RADIX_INIT() - initialize radix tree |
| 695 | * |
| 696 | * Initialize our meta structures and bitmaps and calculate the exact |
| 697 | * amount of space required to manage 'count' blocks - this space may |
| 698 | * be considerably less then the calculated radix due to the large |
| 699 | * RADIX values we use. |
| 700 | */ |
| 701 | |
| 702 | static daddr_t |
| 703 | blst_radix_init(blmeta_t *scan, daddr_t radix, int skip, daddr_t count) |
| 704 | { |
| 705 | int i; |
| 706 | int next_skip; |
| 707 | daddr_t memindex = 0; |
| 708 | |
| 709 | /* |
| 710 | * Leaf node |
| 711 | */ |
| 712 | |
| 713 | if (radix == BLIST_BMAP_RADIX) { |
| 714 | if (scan) { |
| 715 | scan->bm_bighint = 0; |
| 716 | scan->u.bmu_bitmap = 0; |
| 717 | } |
| 718 | return(memindex); |
| 719 | } |
| 720 | |
| 721 | /* |
| 722 | * Meta node. If allocating the entire object we can special |
| 723 | * case it. However, we need to figure out how much memory |
| 724 | * is required to manage 'count' blocks, so we continue on anyway. |
| 725 | */ |
| 726 | |
| 727 | if (scan) { |
| 728 | scan->bm_bighint = 0; |
| 729 | scan->u.bmu_avail = 0; |
| 730 | } |
| 731 | |
| 732 | radix >>= BLIST_META_RADIX_SHIFT; |
| 733 | next_skip = (skip >> BLIST_META_RADIX_SHIFT); |
| 734 | |
| 735 | for (i = 1; i <= skip; i += next_skip) { |
| 736 | if (count >= radix) { |
| 737 | /* |
| 738 | * Allocate the entire object |
| 739 | */ |
| 740 | memindex = i + blst_radix_init( |
| 741 | ((scan) ? &scan[i] : NULL), |
| 742 | radix, |
| 743 | next_skip - 1, |
| 744 | radix |
| 745 | ); |
| 746 | count -= radix; |
| 747 | } else if (count > 0) { |
| 748 | /* |
| 749 | * Allocate a partial object |
| 750 | */ |
| 751 | memindex = i + blst_radix_init( |
| 752 | ((scan) ? &scan[i] : NULL), |
| 753 | radix, |
| 754 | next_skip - 1, |
| 755 | count |
| 756 | ); |
| 757 | count = 0; |
| 758 | } else { |
| 759 | /* |
| 760 | * Add terminator and break out |
| 761 | */ |
| 762 | if (scan) |
| 763 | scan[i].bm_bighint = (daddr_t)-1; |
| 764 | break; |
| 765 | } |
| 766 | } |
| 767 | if (memindex < i) |
| 768 | memindex = i; |
| 769 | return(memindex); |
| 770 | } |
| 771 | |
| 772 | #ifdef BLIST_DEBUG |
| 773 | |
| 774 | static void |
| 775 | blst_radix_print(blmeta_t *scan, daddr_t blk, daddr_t radix, int skip, int tab) |
| 776 | { |
| 777 | int i; |
| 778 | int next_skip; |
| 779 | int lastState = 0; |
| 780 | |
| 781 | if (radix == BLIST_BMAP_RADIX) { |
| 782 | printf( |
| 783 | "%*.*s(%04x,%d): bitmap %08x big=%d\n" , |
| 784 | tab, tab, "" , |
| 785 | blk, radix, |
| 786 | scan->u.bmu_bitmap, |
| 787 | scan->bm_bighint |
| 788 | ); |
| 789 | return; |
| 790 | } |
| 791 | |
| 792 | if (scan->u.bmu_avail == 0) { |
| 793 | printf( |
| 794 | "%*.*s(%04x,%d) ALL ALLOCATED\n" , |
| 795 | tab, tab, "" , |
| 796 | blk, |
| 797 | radix |
| 798 | ); |
| 799 | return; |
| 800 | } |
| 801 | if (scan->u.bmu_avail == radix) { |
| 802 | printf( |
| 803 | "%*.*s(%04x,%d) ALL FREE\n" , |
| 804 | tab, tab, "" , |
| 805 | blk, |
| 806 | radix |
| 807 | ); |
| 808 | return; |
| 809 | } |
| 810 | |
| 811 | printf( |
| 812 | "%*.*s(%04x,%d): subtree (%d/%d) big=%d {\n" , |
| 813 | tab, tab, "" , |
| 814 | blk, radix, |
| 815 | scan->u.bmu_avail, |
| 816 | radix, |
| 817 | scan->bm_bighint |
| 818 | ); |
| 819 | |
| 820 | radix >>= BLIST_META_RADIX_SHIFT; |
| 821 | next_skip = (skip >> BLIST_META_RADIX_SHIFT); |
| 822 | tab += 4; |
| 823 | |
| 824 | for (i = 1; i <= skip; i += next_skip) { |
| 825 | if (scan[i].bm_bighint == (daddr_t)-1) { |
| 826 | printf( |
| 827 | "%*.*s(%04x,%d): Terminator\n" , |
| 828 | tab, tab, "" , |
| 829 | blk, radix |
| 830 | ); |
| 831 | lastState = 0; |
| 832 | break; |
| 833 | } |
| 834 | blst_radix_print( |
| 835 | &scan[i], |
| 836 | blk, |
| 837 | radix, |
| 838 | next_skip - 1, |
| 839 | tab |
| 840 | ); |
| 841 | blk += radix; |
| 842 | } |
| 843 | tab -= 4; |
| 844 | |
| 845 | printf( |
| 846 | "%*.*s}\n" , |
| 847 | tab, tab, "" |
| 848 | ); |
| 849 | } |
| 850 | |
| 851 | #endif |
| 852 | |
| 853 | #ifdef BLIST_DEBUG |
| 854 | |
| 855 | int |
| 856 | main(int ac, char **av) |
| 857 | { |
| 858 | int size = 1024; |
| 859 | int i; |
| 860 | blist_t bl; |
| 861 | |
| 862 | for (i = 1; i < ac; ++i) { |
| 863 | const char *ptr = av[i]; |
| 864 | if (*ptr != '-') { |
| 865 | size = strtol(ptr, NULL, 0); |
| 866 | continue; |
| 867 | } |
| 868 | ptr += 2; |
| 869 | fprintf(stderr, "Bad option: %s\n" , ptr - 2); |
| 870 | exit(1); |
| 871 | } |
| 872 | bl = blist_create(size); |
| 873 | blist_free(bl, 0, size); |
| 874 | |
| 875 | for (;;) { |
| 876 | char buf[1024]; |
| 877 | daddr_t da = 0; |
| 878 | daddr_t count = 0; |
| 879 | |
| 880 | |
| 881 | printf("%d/%d/%d> " , bl->bl_free, size, bl->bl_radix); |
| 882 | fflush(stdout); |
| 883 | if (fgets(buf, sizeof(buf), stdin) == NULL) |
| 884 | break; |
| 885 | switch(buf[0]) { |
| 886 | case 'r': |
| 887 | if (sscanf(buf + 1, "%d" , &count) == 1) { |
| 888 | blist_resize(&bl, count, 1); |
| 889 | } else { |
| 890 | printf("?\n" ); |
| 891 | } |
| 892 | case 'p': |
| 893 | blist_print(bl); |
| 894 | break; |
| 895 | case 'a': |
| 896 | if (sscanf(buf + 1, "%d" , &count) == 1) { |
| 897 | daddr_t blk = blist_alloc(bl, count); |
| 898 | printf(" R=%04x\n" , blk); |
| 899 | } else { |
| 900 | printf("?\n" ); |
| 901 | } |
| 902 | break; |
| 903 | case 'f': |
| 904 | if (sscanf(buf + 1, "%x %d" , &da, &count) == 2) { |
| 905 | blist_free(bl, da, count); |
| 906 | } else { |
| 907 | printf("?\n" ); |
| 908 | } |
| 909 | break; |
| 910 | case '?': |
| 911 | case 'h': |
| 912 | puts( |
| 913 | "p -print\n" |
| 914 | "a %d -allocate\n" |
| 915 | "f %x %d -free\n" |
| 916 | "r %d -resize\n" |
| 917 | "h/? -help" |
| 918 | ); |
| 919 | break; |
| 920 | default: |
| 921 | printf("?\n" ); |
| 922 | break; |
| 923 | } |
| 924 | } |
| 925 | return(0); |
| 926 | } |
| 927 | |
| 928 | void |
| 929 | panic(const char *ctl, ...) |
| 930 | { |
| 931 | va_list va; |
| 932 | |
| 933 | va_start(va, ctl); |
| 934 | vfprintf(stderr, ctl, va); |
| 935 | fprintf(stderr, "\n" ); |
| 936 | va_end(va); |
| 937 | exit(1); |
| 938 | } |
| 939 | |
| 940 | #endif |
| 941 | |