1 /* 2 * Copyright (C) International Business Machines Corp., 2000-2004 3 * Portions Copyright (C) Tino Reichardt, 2012 4 * 5 * This program is free software; you can redistribute it and/or modify 6 * it under the terms of the GNU General Public License as published by 7 * the Free Software Foundation; either version 2 of the License, or 8 * (at your option) any later version. 9 * 10 * This program is distributed in the hope that it will be useful, 11 * but WITHOUT ANY WARRANTY; without even the implied warranty of 12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See 13 * the GNU General Public License for more details. 14 * 15 * You should have received a copy of the GNU General Public License 16 * along with this program; if not, write to the Free Software 17 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA 18 */ 19 20 #include <linux/fs.h> 21 #include <linux/slab.h> 22 #include "jfs_incore.h" 23 #include "jfs_superblock.h" 24 #include "jfs_dmap.h" 25 #include "jfs_imap.h" 26 #include "jfs_lock.h" 27 #include "jfs_metapage.h" 28 #include "jfs_debug.h" 29 #include "jfs_discard.h" 30 31 /* 32 * SERIALIZATION of the Block Allocation Map. 33 * 34 * the working state of the block allocation map is accessed in 35 * two directions: 36 * 37 * 1) allocation and free requests that start at the dmap 38 * level and move up through the dmap control pages (i.e. 39 * the vast majority of requests). 40 * 41 * 2) allocation requests that start at dmap control page 42 * level and work down towards the dmaps. 43 * 44 * the serialization scheme used here is as follows. 45 * 46 * requests which start at the bottom are serialized against each 47 * other through buffers and each requests holds onto its buffers 48 * as it works it way up from a single dmap to the required level 49 * of dmap control page. 50 * requests that start at the top are serialized against each other 51 * and request that start from the bottom by the multiple read/single 52 * write inode lock of the bmap inode. requests starting at the top 53 * take this lock in write mode while request starting at the bottom 54 * take the lock in read mode. a single top-down request may proceed 55 * exclusively while multiple bottoms-up requests may proceed 56 * simultaneously (under the protection of busy buffers). 57 * 58 * in addition to information found in dmaps and dmap control pages, 59 * the working state of the block allocation map also includes read/ 60 * write information maintained in the bmap descriptor (i.e. total 61 * free block count, allocation group level free block counts). 62 * a single exclusive lock (BMAP_LOCK) is used to guard this information 63 * in the face of multiple-bottoms up requests. 64 * (lock ordering: IREAD_LOCK, BMAP_LOCK); 65 * 66 * accesses to the persistent state of the block allocation map (limited 67 * to the persistent bitmaps in dmaps) is guarded by (busy) buffers. 68 */ 69 70 #define BMAP_LOCK_INIT(bmp) mutex_init(&bmp->db_bmaplock) 71 #define BMAP_LOCK(bmp) mutex_lock(&bmp->db_bmaplock) 72 #define BMAP_UNLOCK(bmp) mutex_unlock(&bmp->db_bmaplock) 73 74 /* 75 * forward references 76 */ 77 static void dbAllocBits(struct bmap * bmp, struct dmap * dp, s64 blkno, 78 int nblocks); 79 static void dbSplit(dmtree_t * tp, int leafno, int splitsz, int newval); 80 static int dbBackSplit(dmtree_t * tp, int leafno); 81 static int dbJoin(dmtree_t * tp, int leafno, int newval); 82 static void dbAdjTree(dmtree_t * tp, int leafno, int newval); 83 static int dbAdjCtl(struct bmap * bmp, s64 blkno, int newval, int alloc, 84 int level); 85 static int dbAllocAny(struct bmap * bmp, s64 nblocks, int l2nb, s64 * results); 86 static int dbAllocNext(struct bmap * bmp, struct dmap * dp, s64 blkno, 87 int nblocks); 88 static int dbAllocNear(struct bmap * bmp, struct dmap * dp, s64 blkno, 89 int nblocks, 90 int l2nb, s64 * results); 91 static int dbAllocDmap(struct bmap * bmp, struct dmap * dp, s64 blkno, 92 int nblocks); 93 static int dbAllocDmapLev(struct bmap * bmp, struct dmap * dp, int nblocks, 94 int l2nb, 95 s64 * results); 96 static int dbAllocAG(struct bmap * bmp, int agno, s64 nblocks, int l2nb, 97 s64 * results); 98 static int dbAllocCtl(struct bmap * bmp, s64 nblocks, int l2nb, s64 blkno, 99 s64 * results); 100 static int dbExtend(struct inode *ip, s64 blkno, s64 nblocks, s64 addnblocks); 101 static int dbFindBits(u32 word, int l2nb); 102 static int dbFindCtl(struct bmap * bmp, int l2nb, int level, s64 * blkno); 103 static int dbFindLeaf(dmtree_t * tp, int l2nb, int *leafidx); 104 static int dbFreeBits(struct bmap * bmp, struct dmap * dp, s64 blkno, 105 int nblocks); 106 static int dbFreeDmap(struct bmap * bmp, struct dmap * dp, s64 blkno, 107 int nblocks); 108 static int dbMaxBud(u8 * cp); 109 static int blkstol2(s64 nb); 110 111 static int cntlz(u32 value); 112 static int cnttz(u32 word); 113 114 static int dbAllocDmapBU(struct bmap * bmp, struct dmap * dp, s64 blkno, 115 int nblocks); 116 static int dbInitDmap(struct dmap * dp, s64 blkno, int nblocks); 117 static int dbInitDmapTree(struct dmap * dp); 118 static int dbInitTree(struct dmaptree * dtp); 119 static int dbInitDmapCtl(struct dmapctl * dcp, int level, int i); 120 static int dbGetL2AGSize(s64 nblocks); 121 122 /* 123 * buddy table 124 * 125 * table used for determining buddy sizes within characters of 126 * dmap bitmap words. the characters themselves serve as indexes 127 * into the table, with the table elements yielding the maximum 128 * binary buddy of free bits within the character. 129 */ 130 static const s8 budtab[256] = { 131 3, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 132 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 133 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 134 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 135 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 136 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 137 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 138 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 139 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 140 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 141 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 142 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 143 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 144 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 145 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 146 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, -1 147 }; 148 149 /* 150 * NAME: dbMount() 151 * 152 * FUNCTION: initializate the block allocation map. 153 * 154 * memory is allocated for the in-core bmap descriptor and 155 * the in-core descriptor is initialized from disk. 156 * 157 * PARAMETERS: 158 * ipbmap - pointer to in-core inode for the block map. 159 * 160 * RETURN VALUES: 161 * 0 - success 162 * -ENOMEM - insufficient memory 163 * -EIO - i/o error 164 */ 165 int dbMount(struct inode *ipbmap) 166 { 167 struct bmap *bmp; 168 struct dbmap_disk *dbmp_le; 169 struct metapage *mp; 170 int i; 171 172 /* 173 * allocate/initialize the in-memory bmap descriptor 174 */ 175 /* allocate memory for the in-memory bmap descriptor */ 176 bmp = kmalloc(sizeof(struct bmap), GFP_KERNEL); 177 if (bmp == NULL) 178 return -ENOMEM; 179 180 /* read the on-disk bmap descriptor. */ 181 mp = read_metapage(ipbmap, 182 BMAPBLKNO << JFS_SBI(ipbmap->i_sb)->l2nbperpage, 183 PSIZE, 0); 184 if (mp == NULL) { 185 kfree(bmp); 186 return -EIO; 187 } 188 189 /* copy the on-disk bmap descriptor to its in-memory version. */ 190 dbmp_le = (struct dbmap_disk *) mp->data; 191 bmp->db_mapsize = le64_to_cpu(dbmp_le->dn_mapsize); 192 bmp->db_nfree = le64_to_cpu(dbmp_le->dn_nfree); 193 bmp->db_l2nbperpage = le32_to_cpu(dbmp_le->dn_l2nbperpage); 194 bmp->db_numag = le32_to_cpu(dbmp_le->dn_numag); 195 bmp->db_maxlevel = le32_to_cpu(dbmp_le->dn_maxlevel); 196 bmp->db_maxag = le32_to_cpu(dbmp_le->dn_maxag); 197 bmp->db_agpref = le32_to_cpu(dbmp_le->dn_agpref); 198 bmp->db_aglevel = le32_to_cpu(dbmp_le->dn_aglevel); 199 bmp->db_agheight = le32_to_cpu(dbmp_le->dn_agheight); 200 bmp->db_agwidth = le32_to_cpu(dbmp_le->dn_agwidth); 201 bmp->db_agstart = le32_to_cpu(dbmp_le->dn_agstart); 202 bmp->db_agl2size = le32_to_cpu(dbmp_le->dn_agl2size); 203 for (i = 0; i < MAXAG; i++) 204 bmp->db_agfree[i] = le64_to_cpu(dbmp_le->dn_agfree[i]); 205 bmp->db_agsize = le64_to_cpu(dbmp_le->dn_agsize); 206 bmp->db_maxfreebud = dbmp_le->dn_maxfreebud; 207 208 /* release the buffer. */ 209 release_metapage(mp); 210 211 /* bind the bmap inode and the bmap descriptor to each other. */ 212 bmp->db_ipbmap = ipbmap; 213 JFS_SBI(ipbmap->i_sb)->bmap = bmp; 214 215 memset(bmp->db_active, 0, sizeof(bmp->db_active)); 216 217 /* 218 * allocate/initialize the bmap lock 219 */ 220 BMAP_LOCK_INIT(bmp); 221 222 return (0); 223 } 224 225 226 /* 227 * NAME: dbUnmount() 228 * 229 * FUNCTION: terminate the block allocation map in preparation for 230 * file system unmount. 231 * 232 * the in-core bmap descriptor is written to disk and 233 * the memory for this descriptor is freed. 234 * 235 * PARAMETERS: 236 * ipbmap - pointer to in-core inode for the block map. 237 * 238 * RETURN VALUES: 239 * 0 - success 240 * -EIO - i/o error 241 */ 242 int dbUnmount(struct inode *ipbmap, int mounterror) 243 { 244 struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap; 245 246 if (!(mounterror || isReadOnly(ipbmap))) 247 dbSync(ipbmap); 248 249 /* 250 * Invalidate the page cache buffers 251 */ 252 truncate_inode_pages(ipbmap->i_mapping, 0); 253 254 /* free the memory for the in-memory bmap. */ 255 kfree(bmp); 256 257 return (0); 258 } 259 260 /* 261 * dbSync() 262 */ 263 int dbSync(struct inode *ipbmap) 264 { 265 struct dbmap_disk *dbmp_le; 266 struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap; 267 struct metapage *mp; 268 int i; 269 270 /* 271 * write bmap global control page 272 */ 273 /* get the buffer for the on-disk bmap descriptor. */ 274 mp = read_metapage(ipbmap, 275 BMAPBLKNO << JFS_SBI(ipbmap->i_sb)->l2nbperpage, 276 PSIZE, 0); 277 if (mp == NULL) { 278 jfs_err("dbSync: read_metapage failed!"); 279 return -EIO; 280 } 281 /* copy the in-memory version of the bmap to the on-disk version */ 282 dbmp_le = (struct dbmap_disk *) mp->data; 283 dbmp_le->dn_mapsize = cpu_to_le64(bmp->db_mapsize); 284 dbmp_le->dn_nfree = cpu_to_le64(bmp->db_nfree); 285 dbmp_le->dn_l2nbperpage = cpu_to_le32(bmp->db_l2nbperpage); 286 dbmp_le->dn_numag = cpu_to_le32(bmp->db_numag); 287 dbmp_le->dn_maxlevel = cpu_to_le32(bmp->db_maxlevel); 288 dbmp_le->dn_maxag = cpu_to_le32(bmp->db_maxag); 289 dbmp_le->dn_agpref = cpu_to_le32(bmp->db_agpref); 290 dbmp_le->dn_aglevel = cpu_to_le32(bmp->db_aglevel); 291 dbmp_le->dn_agheight = cpu_to_le32(bmp->db_agheight); 292 dbmp_le->dn_agwidth = cpu_to_le32(bmp->db_agwidth); 293 dbmp_le->dn_agstart = cpu_to_le32(bmp->db_agstart); 294 dbmp_le->dn_agl2size = cpu_to_le32(bmp->db_agl2size); 295 for (i = 0; i < MAXAG; i++) 296 dbmp_le->dn_agfree[i] = cpu_to_le64(bmp->db_agfree[i]); 297 dbmp_le->dn_agsize = cpu_to_le64(bmp->db_agsize); 298 dbmp_le->dn_maxfreebud = bmp->db_maxfreebud; 299 300 /* write the buffer */ 301 write_metapage(mp); 302 303 /* 304 * write out dirty pages of bmap 305 */ 306 filemap_write_and_wait(ipbmap->i_mapping); 307 308 diWriteSpecial(ipbmap, 0); 309 310 return (0); 311 } 312 313 /* 314 * NAME: dbFree() 315 * 316 * FUNCTION: free the specified block range from the working block 317 * allocation map. 318 * 319 * the blocks will be free from the working map one dmap 320 * at a time. 321 * 322 * PARAMETERS: 323 * ip - pointer to in-core inode; 324 * blkno - starting block number to be freed. 325 * nblocks - number of blocks to be freed. 326 * 327 * RETURN VALUES: 328 * 0 - success 329 * -EIO - i/o error 330 */ 331 int dbFree(struct inode *ip, s64 blkno, s64 nblocks) 332 { 333 struct metapage *mp; 334 struct dmap *dp; 335 int nb, rc; 336 s64 lblkno, rem; 337 struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap; 338 struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap; 339 struct super_block *sb = ipbmap->i_sb; 340 341 IREAD_LOCK(ipbmap, RDWRLOCK_DMAP); 342 343 /* block to be freed better be within the mapsize. */ 344 if (unlikely((blkno == 0) || (blkno + nblocks > bmp->db_mapsize))) { 345 IREAD_UNLOCK(ipbmap); 346 printk(KERN_ERR "blkno = %Lx, nblocks = %Lx\n", 347 (unsigned long long) blkno, 348 (unsigned long long) nblocks); 349 jfs_error(ip->i_sb, 350 "dbFree: block to be freed is outside the map"); 351 return -EIO; 352 } 353 354 /** 355 * TRIM the blocks, when mounted with discard option 356 */ 357 if (JFS_SBI(sb)->flag & JFS_DISCARD) 358 if (JFS_SBI(sb)->minblks_trim <= nblocks) 359 jfs_issue_discard(ipbmap, blkno, nblocks); 360 361 /* 362 * free the blocks a dmap at a time. 363 */ 364 mp = NULL; 365 for (rem = nblocks; rem > 0; rem -= nb, blkno += nb) { 366 /* release previous dmap if any */ 367 if (mp) { 368 write_metapage(mp); 369 } 370 371 /* get the buffer for the current dmap. */ 372 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage); 373 mp = read_metapage(ipbmap, lblkno, PSIZE, 0); 374 if (mp == NULL) { 375 IREAD_UNLOCK(ipbmap); 376 return -EIO; 377 } 378 dp = (struct dmap *) mp->data; 379 380 /* determine the number of blocks to be freed from 381 * this dmap. 382 */ 383 nb = min(rem, BPERDMAP - (blkno & (BPERDMAP - 1))); 384 385 /* free the blocks. */ 386 if ((rc = dbFreeDmap(bmp, dp, blkno, nb))) { 387 jfs_error(ip->i_sb, "dbFree: error in block map\n"); 388 release_metapage(mp); 389 IREAD_UNLOCK(ipbmap); 390 return (rc); 391 } 392 } 393 394 /* write the last buffer. */ 395 write_metapage(mp); 396 397 IREAD_UNLOCK(ipbmap); 398 399 return (0); 400 } 401 402 403 /* 404 * NAME: dbUpdatePMap() 405 * 406 * FUNCTION: update the allocation state (free or allocate) of the 407 * specified block range in the persistent block allocation map. 408 * 409 * the blocks will be updated in the persistent map one 410 * dmap at a time. 411 * 412 * PARAMETERS: 413 * ipbmap - pointer to in-core inode for the block map. 414 * free - 'true' if block range is to be freed from the persistent 415 * map; 'false' if it is to be allocated. 416 * blkno - starting block number of the range. 417 * nblocks - number of contiguous blocks in the range. 418 * tblk - transaction block; 419 * 420 * RETURN VALUES: 421 * 0 - success 422 * -EIO - i/o error 423 */ 424 int 425 dbUpdatePMap(struct inode *ipbmap, 426 int free, s64 blkno, s64 nblocks, struct tblock * tblk) 427 { 428 int nblks, dbitno, wbitno, rbits; 429 int word, nbits, nwords; 430 struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap; 431 s64 lblkno, rem, lastlblkno; 432 u32 mask; 433 struct dmap *dp; 434 struct metapage *mp; 435 struct jfs_log *log; 436 int lsn, difft, diffp; 437 unsigned long flags; 438 439 /* the blocks better be within the mapsize. */ 440 if (blkno + nblocks > bmp->db_mapsize) { 441 printk(KERN_ERR "blkno = %Lx, nblocks = %Lx\n", 442 (unsigned long long) blkno, 443 (unsigned long long) nblocks); 444 jfs_error(ipbmap->i_sb, 445 "dbUpdatePMap: blocks are outside the map"); 446 return -EIO; 447 } 448 449 /* compute delta of transaction lsn from log syncpt */ 450 lsn = tblk->lsn; 451 log = (struct jfs_log *) JFS_SBI(tblk->sb)->log; 452 logdiff(difft, lsn, log); 453 454 /* 455 * update the block state a dmap at a time. 456 */ 457 mp = NULL; 458 lastlblkno = 0; 459 for (rem = nblocks; rem > 0; rem -= nblks, blkno += nblks) { 460 /* get the buffer for the current dmap. */ 461 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage); 462 if (lblkno != lastlblkno) { 463 if (mp) { 464 write_metapage(mp); 465 } 466 467 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 468 0); 469 if (mp == NULL) 470 return -EIO; 471 metapage_wait_for_io(mp); 472 } 473 dp = (struct dmap *) mp->data; 474 475 /* determine the bit number and word within the dmap of 476 * the starting block. also determine how many blocks 477 * are to be updated within this dmap. 478 */ 479 dbitno = blkno & (BPERDMAP - 1); 480 word = dbitno >> L2DBWORD; 481 nblks = min(rem, (s64)BPERDMAP - dbitno); 482 483 /* update the bits of the dmap words. the first and last 484 * words may only have a subset of their bits updated. if 485 * this is the case, we'll work against that word (i.e. 486 * partial first and/or last) only in a single pass. a 487 * single pass will also be used to update all words that 488 * are to have all their bits updated. 489 */ 490 for (rbits = nblks; rbits > 0; 491 rbits -= nbits, dbitno += nbits) { 492 /* determine the bit number within the word and 493 * the number of bits within the word. 494 */ 495 wbitno = dbitno & (DBWORD - 1); 496 nbits = min(rbits, DBWORD - wbitno); 497 498 /* check if only part of the word is to be updated. */ 499 if (nbits < DBWORD) { 500 /* update (free or allocate) the bits 501 * in this word. 502 */ 503 mask = 504 (ONES << (DBWORD - nbits) >> wbitno); 505 if (free) 506 dp->pmap[word] &= 507 cpu_to_le32(~mask); 508 else 509 dp->pmap[word] |= 510 cpu_to_le32(mask); 511 512 word += 1; 513 } else { 514 /* one or more words are to have all 515 * their bits updated. determine how 516 * many words and how many bits. 517 */ 518 nwords = rbits >> L2DBWORD; 519 nbits = nwords << L2DBWORD; 520 521 /* update (free or allocate) the bits 522 * in these words. 523 */ 524 if (free) 525 memset(&dp->pmap[word], 0, 526 nwords * 4); 527 else 528 memset(&dp->pmap[word], (int) ONES, 529 nwords * 4); 530 531 word += nwords; 532 } 533 } 534 535 /* 536 * update dmap lsn 537 */ 538 if (lblkno == lastlblkno) 539 continue; 540 541 lastlblkno = lblkno; 542 543 LOGSYNC_LOCK(log, flags); 544 if (mp->lsn != 0) { 545 /* inherit older/smaller lsn */ 546 logdiff(diffp, mp->lsn, log); 547 if (difft < diffp) { 548 mp->lsn = lsn; 549 550 /* move bp after tblock in logsync list */ 551 list_move(&mp->synclist, &tblk->synclist); 552 } 553 554 /* inherit younger/larger clsn */ 555 logdiff(difft, tblk->clsn, log); 556 logdiff(diffp, mp->clsn, log); 557 if (difft > diffp) 558 mp->clsn = tblk->clsn; 559 } else { 560 mp->log = log; 561 mp->lsn = lsn; 562 563 /* insert bp after tblock in logsync list */ 564 log->count++; 565 list_add(&mp->synclist, &tblk->synclist); 566 567 mp->clsn = tblk->clsn; 568 } 569 LOGSYNC_UNLOCK(log, flags); 570 } 571 572 /* write the last buffer. */ 573 if (mp) { 574 write_metapage(mp); 575 } 576 577 return (0); 578 } 579 580 581 /* 582 * NAME: dbNextAG() 583 * 584 * FUNCTION: find the preferred allocation group for new allocations. 585 * 586 * Within the allocation groups, we maintain a preferred 587 * allocation group which consists of a group with at least 588 * average free space. It is the preferred group that we target 589 * new inode allocation towards. The tie-in between inode 590 * allocation and block allocation occurs as we allocate the 591 * first (data) block of an inode and specify the inode (block) 592 * as the allocation hint for this block. 593 * 594 * We try to avoid having more than one open file growing in 595 * an allocation group, as this will lead to fragmentation. 596 * This differs from the old OS/2 method of trying to keep 597 * empty ags around for large allocations. 598 * 599 * PARAMETERS: 600 * ipbmap - pointer to in-core inode for the block map. 601 * 602 * RETURN VALUES: 603 * the preferred allocation group number. 604 */ 605 int dbNextAG(struct inode *ipbmap) 606 { 607 s64 avgfree; 608 int agpref; 609 s64 hwm = 0; 610 int i; 611 int next_best = -1; 612 struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap; 613 614 BMAP_LOCK(bmp); 615 616 /* determine the average number of free blocks within the ags. */ 617 avgfree = (u32)bmp->db_nfree / bmp->db_numag; 618 619 /* 620 * if the current preferred ag does not have an active allocator 621 * and has at least average freespace, return it 622 */ 623 agpref = bmp->db_agpref; 624 if ((atomic_read(&bmp->db_active[agpref]) == 0) && 625 (bmp->db_agfree[agpref] >= avgfree)) 626 goto unlock; 627 628 /* From the last preferred ag, find the next one with at least 629 * average free space. 630 */ 631 for (i = 0 ; i < bmp->db_numag; i++, agpref++) { 632 if (agpref == bmp->db_numag) 633 agpref = 0; 634 635 if (atomic_read(&bmp->db_active[agpref])) 636 /* open file is currently growing in this ag */ 637 continue; 638 if (bmp->db_agfree[agpref] >= avgfree) { 639 /* Return this one */ 640 bmp->db_agpref = agpref; 641 goto unlock; 642 } else if (bmp->db_agfree[agpref] > hwm) { 643 /* Less than avg. freespace, but best so far */ 644 hwm = bmp->db_agfree[agpref]; 645 next_best = agpref; 646 } 647 } 648 649 /* 650 * If no inactive ag was found with average freespace, use the 651 * next best 652 */ 653 if (next_best != -1) 654 bmp->db_agpref = next_best; 655 /* else leave db_agpref unchanged */ 656 unlock: 657 BMAP_UNLOCK(bmp); 658 659 /* return the preferred group. 660 */ 661 return (bmp->db_agpref); 662 } 663 664 /* 665 * NAME: dbAlloc() 666 * 667 * FUNCTION: attempt to allocate a specified number of contiguous free 668 * blocks from the working allocation block map. 669 * 670 * the block allocation policy uses hints and a multi-step 671 * approach. 672 * 673 * for allocation requests smaller than the number of blocks 674 * per dmap, we first try to allocate the new blocks 675 * immediately following the hint. if these blocks are not 676 * available, we try to allocate blocks near the hint. if 677 * no blocks near the hint are available, we next try to 678 * allocate within the same dmap as contains the hint. 679 * 680 * if no blocks are available in the dmap or the allocation 681 * request is larger than the dmap size, we try to allocate 682 * within the same allocation group as contains the hint. if 683 * this does not succeed, we finally try to allocate anywhere 684 * within the aggregate. 685 * 686 * we also try to allocate anywhere within the aggregate for 687 * for allocation requests larger than the allocation group 688 * size or requests that specify no hint value. 689 * 690 * PARAMETERS: 691 * ip - pointer to in-core inode; 692 * hint - allocation hint. 693 * nblocks - number of contiguous blocks in the range. 694 * results - on successful return, set to the starting block number 695 * of the newly allocated contiguous range. 696 * 697 * RETURN VALUES: 698 * 0 - success 699 * -ENOSPC - insufficient disk resources 700 * -EIO - i/o error 701 */ 702 int dbAlloc(struct inode *ip, s64 hint, s64 nblocks, s64 * results) 703 { 704 int rc, agno; 705 struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap; 706 struct bmap *bmp; 707 struct metapage *mp; 708 s64 lblkno, blkno; 709 struct dmap *dp; 710 int l2nb; 711 s64 mapSize; 712 int writers; 713 714 /* assert that nblocks is valid */ 715 assert(nblocks > 0); 716 717 /* get the log2 number of blocks to be allocated. 718 * if the number of blocks is not a log2 multiple, 719 * it will be rounded up to the next log2 multiple. 720 */ 721 l2nb = BLKSTOL2(nblocks); 722 723 bmp = JFS_SBI(ip->i_sb)->bmap; 724 725 mapSize = bmp->db_mapsize; 726 727 /* the hint should be within the map */ 728 if (hint >= mapSize) { 729 jfs_error(ip->i_sb, "dbAlloc: the hint is outside the map"); 730 return -EIO; 731 } 732 733 /* if the number of blocks to be allocated is greater than the 734 * allocation group size, try to allocate anywhere. 735 */ 736 if (l2nb > bmp->db_agl2size) { 737 IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP); 738 739 rc = dbAllocAny(bmp, nblocks, l2nb, results); 740 741 goto write_unlock; 742 } 743 744 /* 745 * If no hint, let dbNextAG recommend an allocation group 746 */ 747 if (hint == 0) 748 goto pref_ag; 749 750 /* we would like to allocate close to the hint. adjust the 751 * hint to the block following the hint since the allocators 752 * will start looking for free space starting at this point. 753 */ 754 blkno = hint + 1; 755 756 if (blkno >= bmp->db_mapsize) 757 goto pref_ag; 758 759 agno = blkno >> bmp->db_agl2size; 760 761 /* check if blkno crosses over into a new allocation group. 762 * if so, check if we should allow allocations within this 763 * allocation group. 764 */ 765 if ((blkno & (bmp->db_agsize - 1)) == 0) 766 /* check if the AG is currently being written to. 767 * if so, call dbNextAG() to find a non-busy 768 * AG with sufficient free space. 769 */ 770 if (atomic_read(&bmp->db_active[agno])) 771 goto pref_ag; 772 773 /* check if the allocation request size can be satisfied from a 774 * single dmap. if so, try to allocate from the dmap containing 775 * the hint using a tiered strategy. 776 */ 777 if (nblocks <= BPERDMAP) { 778 IREAD_LOCK(ipbmap, RDWRLOCK_DMAP); 779 780 /* get the buffer for the dmap containing the hint. 781 */ 782 rc = -EIO; 783 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage); 784 mp = read_metapage(ipbmap, lblkno, PSIZE, 0); 785 if (mp == NULL) 786 goto read_unlock; 787 788 dp = (struct dmap *) mp->data; 789 790 /* first, try to satisfy the allocation request with the 791 * blocks beginning at the hint. 792 */ 793 if ((rc = dbAllocNext(bmp, dp, blkno, (int) nblocks)) 794 != -ENOSPC) { 795 if (rc == 0) { 796 *results = blkno; 797 mark_metapage_dirty(mp); 798 } 799 800 release_metapage(mp); 801 goto read_unlock; 802 } 803 804 writers = atomic_read(&bmp->db_active[agno]); 805 if ((writers > 1) || 806 ((writers == 1) && (JFS_IP(ip)->active_ag != agno))) { 807 /* 808 * Someone else is writing in this allocation 809 * group. To avoid fragmenting, try another ag 810 */ 811 release_metapage(mp); 812 IREAD_UNLOCK(ipbmap); 813 goto pref_ag; 814 } 815 816 /* next, try to satisfy the allocation request with blocks 817 * near the hint. 818 */ 819 if ((rc = 820 dbAllocNear(bmp, dp, blkno, (int) nblocks, l2nb, results)) 821 != -ENOSPC) { 822 if (rc == 0) 823 mark_metapage_dirty(mp); 824 825 release_metapage(mp); 826 goto read_unlock; 827 } 828 829 /* try to satisfy the allocation request with blocks within 830 * the same dmap as the hint. 831 */ 832 if ((rc = dbAllocDmapLev(bmp, dp, (int) nblocks, l2nb, results)) 833 != -ENOSPC) { 834 if (rc == 0) 835 mark_metapage_dirty(mp); 836 837 release_metapage(mp); 838 goto read_unlock; 839 } 840 841 release_metapage(mp); 842 IREAD_UNLOCK(ipbmap); 843 } 844 845 /* try to satisfy the allocation request with blocks within 846 * the same allocation group as the hint. 847 */ 848 IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP); 849 if ((rc = dbAllocAG(bmp, agno, nblocks, l2nb, results)) != -ENOSPC) 850 goto write_unlock; 851 852 IWRITE_UNLOCK(ipbmap); 853 854 855 pref_ag: 856 /* 857 * Let dbNextAG recommend a preferred allocation group 858 */ 859 agno = dbNextAG(ipbmap); 860 IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP); 861 862 /* Try to allocate within this allocation group. if that fails, try to 863 * allocate anywhere in the map. 864 */ 865 if ((rc = dbAllocAG(bmp, agno, nblocks, l2nb, results)) == -ENOSPC) 866 rc = dbAllocAny(bmp, nblocks, l2nb, results); 867 868 write_unlock: 869 IWRITE_UNLOCK(ipbmap); 870 871 return (rc); 872 873 read_unlock: 874 IREAD_UNLOCK(ipbmap); 875 876 return (rc); 877 } 878 879 #ifdef _NOTYET 880 /* 881 * NAME: dbAllocExact() 882 * 883 * FUNCTION: try to allocate the requested extent; 884 * 885 * PARAMETERS: 886 * ip - pointer to in-core inode; 887 * blkno - extent address; 888 * nblocks - extent length; 889 * 890 * RETURN VALUES: 891 * 0 - success 892 * -ENOSPC - insufficient disk resources 893 * -EIO - i/o error 894 */ 895 int dbAllocExact(struct inode *ip, s64 blkno, int nblocks) 896 { 897 int rc; 898 struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap; 899 struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap; 900 struct dmap *dp; 901 s64 lblkno; 902 struct metapage *mp; 903 904 IREAD_LOCK(ipbmap, RDWRLOCK_DMAP); 905 906 /* 907 * validate extent request: 908 * 909 * note: defragfs policy: 910 * max 64 blocks will be moved. 911 * allocation request size must be satisfied from a single dmap. 912 */ 913 if (nblocks <= 0 || nblocks > BPERDMAP || blkno >= bmp->db_mapsize) { 914 IREAD_UNLOCK(ipbmap); 915 return -EINVAL; 916 } 917 918 if (nblocks > ((s64) 1 << bmp->db_maxfreebud)) { 919 /* the free space is no longer available */ 920 IREAD_UNLOCK(ipbmap); 921 return -ENOSPC; 922 } 923 924 /* read in the dmap covering the extent */ 925 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage); 926 mp = read_metapage(ipbmap, lblkno, PSIZE, 0); 927 if (mp == NULL) { 928 IREAD_UNLOCK(ipbmap); 929 return -EIO; 930 } 931 dp = (struct dmap *) mp->data; 932 933 /* try to allocate the requested extent */ 934 rc = dbAllocNext(bmp, dp, blkno, nblocks); 935 936 IREAD_UNLOCK(ipbmap); 937 938 if (rc == 0) 939 mark_metapage_dirty(mp); 940 941 release_metapage(mp); 942 943 return (rc); 944 } 945 #endif /* _NOTYET */ 946 947 /* 948 * NAME: dbReAlloc() 949 * 950 * FUNCTION: attempt to extend a current allocation by a specified 951 * number of blocks. 952 * 953 * this routine attempts to satisfy the allocation request 954 * by first trying to extend the existing allocation in 955 * place by allocating the additional blocks as the blocks 956 * immediately following the current allocation. if these 957 * blocks are not available, this routine will attempt to 958 * allocate a new set of contiguous blocks large enough 959 * to cover the existing allocation plus the additional 960 * number of blocks required. 961 * 962 * PARAMETERS: 963 * ip - pointer to in-core inode requiring allocation. 964 * blkno - starting block of the current allocation. 965 * nblocks - number of contiguous blocks within the current 966 * allocation. 967 * addnblocks - number of blocks to add to the allocation. 968 * results - on successful return, set to the starting block number 969 * of the existing allocation if the existing allocation 970 * was extended in place or to a newly allocated contiguous 971 * range if the existing allocation could not be extended 972 * in place. 973 * 974 * RETURN VALUES: 975 * 0 - success 976 * -ENOSPC - insufficient disk resources 977 * -EIO - i/o error 978 */ 979 int 980 dbReAlloc(struct inode *ip, 981 s64 blkno, s64 nblocks, s64 addnblocks, s64 * results) 982 { 983 int rc; 984 985 /* try to extend the allocation in place. 986 */ 987 if ((rc = dbExtend(ip, blkno, nblocks, addnblocks)) == 0) { 988 *results = blkno; 989 return (0); 990 } else { 991 if (rc != -ENOSPC) 992 return (rc); 993 } 994 995 /* could not extend the allocation in place, so allocate a 996 * new set of blocks for the entire request (i.e. try to get 997 * a range of contiguous blocks large enough to cover the 998 * existing allocation plus the additional blocks.) 999 */ 1000 return (dbAlloc 1001 (ip, blkno + nblocks - 1, addnblocks + nblocks, results)); 1002 } 1003 1004 1005 /* 1006 * NAME: dbExtend() 1007 * 1008 * FUNCTION: attempt to extend a current allocation by a specified 1009 * number of blocks. 1010 * 1011 * this routine attempts to satisfy the allocation request 1012 * by first trying to extend the existing allocation in 1013 * place by allocating the additional blocks as the blocks 1014 * immediately following the current allocation. 1015 * 1016 * PARAMETERS: 1017 * ip - pointer to in-core inode requiring allocation. 1018 * blkno - starting block of the current allocation. 1019 * nblocks - number of contiguous blocks within the current 1020 * allocation. 1021 * addnblocks - number of blocks to add to the allocation. 1022 * 1023 * RETURN VALUES: 1024 * 0 - success 1025 * -ENOSPC - insufficient disk resources 1026 * -EIO - i/o error 1027 */ 1028 static int dbExtend(struct inode *ip, s64 blkno, s64 nblocks, s64 addnblocks) 1029 { 1030 struct jfs_sb_info *sbi = JFS_SBI(ip->i_sb); 1031 s64 lblkno, lastblkno, extblkno; 1032 uint rel_block; 1033 struct metapage *mp; 1034 struct dmap *dp; 1035 int rc; 1036 struct inode *ipbmap = sbi->ipbmap; 1037 struct bmap *bmp; 1038 1039 /* 1040 * We don't want a non-aligned extent to cross a page boundary 1041 */ 1042 if (((rel_block = blkno & (sbi->nbperpage - 1))) && 1043 (rel_block + nblocks + addnblocks > sbi->nbperpage)) 1044 return -ENOSPC; 1045 1046 /* get the last block of the current allocation */ 1047 lastblkno = blkno + nblocks - 1; 1048 1049 /* determine the block number of the block following 1050 * the existing allocation. 1051 */ 1052 extblkno = lastblkno + 1; 1053 1054 IREAD_LOCK(ipbmap, RDWRLOCK_DMAP); 1055 1056 /* better be within the file system */ 1057 bmp = sbi->bmap; 1058 if (lastblkno < 0 || lastblkno >= bmp->db_mapsize) { 1059 IREAD_UNLOCK(ipbmap); 1060 jfs_error(ip->i_sb, 1061 "dbExtend: the block is outside the filesystem"); 1062 return -EIO; 1063 } 1064 1065 /* we'll attempt to extend the current allocation in place by 1066 * allocating the additional blocks as the blocks immediately 1067 * following the current allocation. we only try to extend the 1068 * current allocation in place if the number of additional blocks 1069 * can fit into a dmap, the last block of the current allocation 1070 * is not the last block of the file system, and the start of the 1071 * inplace extension is not on an allocation group boundary. 1072 */ 1073 if (addnblocks > BPERDMAP || extblkno >= bmp->db_mapsize || 1074 (extblkno & (bmp->db_agsize - 1)) == 0) { 1075 IREAD_UNLOCK(ipbmap); 1076 return -ENOSPC; 1077 } 1078 1079 /* get the buffer for the dmap containing the first block 1080 * of the extension. 1081 */ 1082 lblkno = BLKTODMAP(extblkno, bmp->db_l2nbperpage); 1083 mp = read_metapage(ipbmap, lblkno, PSIZE, 0); 1084 if (mp == NULL) { 1085 IREAD_UNLOCK(ipbmap); 1086 return -EIO; 1087 } 1088 1089 dp = (struct dmap *) mp->data; 1090 1091 /* try to allocate the blocks immediately following the 1092 * current allocation. 1093 */ 1094 rc = dbAllocNext(bmp, dp, extblkno, (int) addnblocks); 1095 1096 IREAD_UNLOCK(ipbmap); 1097 1098 /* were we successful ? */ 1099 if (rc == 0) 1100 write_metapage(mp); 1101 else 1102 /* we were not successful */ 1103 release_metapage(mp); 1104 1105 return (rc); 1106 } 1107 1108 1109 /* 1110 * NAME: dbAllocNext() 1111 * 1112 * FUNCTION: attempt to allocate the blocks of the specified block 1113 * range within a dmap. 1114 * 1115 * PARAMETERS: 1116 * bmp - pointer to bmap descriptor 1117 * dp - pointer to dmap. 1118 * blkno - starting block number of the range. 1119 * nblocks - number of contiguous free blocks of the range. 1120 * 1121 * RETURN VALUES: 1122 * 0 - success 1123 * -ENOSPC - insufficient disk resources 1124 * -EIO - i/o error 1125 * 1126 * serialization: IREAD_LOCK(ipbmap) held on entry/exit; 1127 */ 1128 static int dbAllocNext(struct bmap * bmp, struct dmap * dp, s64 blkno, 1129 int nblocks) 1130 { 1131 int dbitno, word, rembits, nb, nwords, wbitno, nw; 1132 int l2size; 1133 s8 *leaf; 1134 u32 mask; 1135 1136 if (dp->tree.leafidx != cpu_to_le32(LEAFIND)) { 1137 jfs_error(bmp->db_ipbmap->i_sb, 1138 "dbAllocNext: Corrupt dmap page"); 1139 return -EIO; 1140 } 1141 1142 /* pick up a pointer to the leaves of the dmap tree. 1143 */ 1144 leaf = dp->tree.stree + le32_to_cpu(dp->tree.leafidx); 1145 1146 /* determine the bit number and word within the dmap of the 1147 * starting block. 1148 */ 1149 dbitno = blkno & (BPERDMAP - 1); 1150 word = dbitno >> L2DBWORD; 1151 1152 /* check if the specified block range is contained within 1153 * this dmap. 1154 */ 1155 if (dbitno + nblocks > BPERDMAP) 1156 return -ENOSPC; 1157 1158 /* check if the starting leaf indicates that anything 1159 * is free. 1160 */ 1161 if (leaf[word] == NOFREE) 1162 return -ENOSPC; 1163 1164 /* check the dmaps words corresponding to block range to see 1165 * if the block range is free. not all bits of the first and 1166 * last words may be contained within the block range. if this 1167 * is the case, we'll work against those words (i.e. partial first 1168 * and/or last) on an individual basis (a single pass) and examine 1169 * the actual bits to determine if they are free. a single pass 1170 * will be used for all dmap words fully contained within the 1171 * specified range. within this pass, the leaves of the dmap 1172 * tree will be examined to determine if the blocks are free. a 1173 * single leaf may describe the free space of multiple dmap 1174 * words, so we may visit only a subset of the actual leaves 1175 * corresponding to the dmap words of the block range. 1176 */ 1177 for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) { 1178 /* determine the bit number within the word and 1179 * the number of bits within the word. 1180 */ 1181 wbitno = dbitno & (DBWORD - 1); 1182 nb = min(rembits, DBWORD - wbitno); 1183 1184 /* check if only part of the word is to be examined. 1185 */ 1186 if (nb < DBWORD) { 1187 /* check if the bits are free. 1188 */ 1189 mask = (ONES << (DBWORD - nb) >> wbitno); 1190 if ((mask & ~le32_to_cpu(dp->wmap[word])) != mask) 1191 return -ENOSPC; 1192 1193 word += 1; 1194 } else { 1195 /* one or more dmap words are fully contained 1196 * within the block range. determine how many 1197 * words and how many bits. 1198 */ 1199 nwords = rembits >> L2DBWORD; 1200 nb = nwords << L2DBWORD; 1201 1202 /* now examine the appropriate leaves to determine 1203 * if the blocks are free. 1204 */ 1205 while (nwords > 0) { 1206 /* does the leaf describe any free space ? 1207 */ 1208 if (leaf[word] < BUDMIN) 1209 return -ENOSPC; 1210 1211 /* determine the l2 number of bits provided 1212 * by this leaf. 1213 */ 1214 l2size = 1215 min((int)leaf[word], NLSTOL2BSZ(nwords)); 1216 1217 /* determine how many words were handled. 1218 */ 1219 nw = BUDSIZE(l2size, BUDMIN); 1220 1221 nwords -= nw; 1222 word += nw; 1223 } 1224 } 1225 } 1226 1227 /* allocate the blocks. 1228 */ 1229 return (dbAllocDmap(bmp, dp, blkno, nblocks)); 1230 } 1231 1232 1233 /* 1234 * NAME: dbAllocNear() 1235 * 1236 * FUNCTION: attempt to allocate a number of contiguous free blocks near 1237 * a specified block (hint) within a dmap. 1238 * 1239 * starting with the dmap leaf that covers the hint, we'll 1240 * check the next four contiguous leaves for sufficient free 1241 * space. if sufficient free space is found, we'll allocate 1242 * the desired free space. 1243 * 1244 * PARAMETERS: 1245 * bmp - pointer to bmap descriptor 1246 * dp - pointer to dmap. 1247 * blkno - block number to allocate near. 1248 * nblocks - actual number of contiguous free blocks desired. 1249 * l2nb - log2 number of contiguous free blocks desired. 1250 * results - on successful return, set to the starting block number 1251 * of the newly allocated range. 1252 * 1253 * RETURN VALUES: 1254 * 0 - success 1255 * -ENOSPC - insufficient disk resources 1256 * -EIO - i/o error 1257 * 1258 * serialization: IREAD_LOCK(ipbmap) held on entry/exit; 1259 */ 1260 static int 1261 dbAllocNear(struct bmap * bmp, 1262 struct dmap * dp, s64 blkno, int nblocks, int l2nb, s64 * results) 1263 { 1264 int word, lword, rc; 1265 s8 *leaf; 1266 1267 if (dp->tree.leafidx != cpu_to_le32(LEAFIND)) { 1268 jfs_error(bmp->db_ipbmap->i_sb, 1269 "dbAllocNear: Corrupt dmap page"); 1270 return -EIO; 1271 } 1272 1273 leaf = dp->tree.stree + le32_to_cpu(dp->tree.leafidx); 1274 1275 /* determine the word within the dmap that holds the hint 1276 * (i.e. blkno). also, determine the last word in the dmap 1277 * that we'll include in our examination. 1278 */ 1279 word = (blkno & (BPERDMAP - 1)) >> L2DBWORD; 1280 lword = min(word + 4, LPERDMAP); 1281 1282 /* examine the leaves for sufficient free space. 1283 */ 1284 for (; word < lword; word++) { 1285 /* does the leaf describe sufficient free space ? 1286 */ 1287 if (leaf[word] < l2nb) 1288 continue; 1289 1290 /* determine the block number within the file system 1291 * of the first block described by this dmap word. 1292 */ 1293 blkno = le64_to_cpu(dp->start) + (word << L2DBWORD); 1294 1295 /* if not all bits of the dmap word are free, get the 1296 * starting bit number within the dmap word of the required 1297 * string of free bits and adjust the block number with the 1298 * value. 1299 */ 1300 if (leaf[word] < BUDMIN) 1301 blkno += 1302 dbFindBits(le32_to_cpu(dp->wmap[word]), l2nb); 1303 1304 /* allocate the blocks. 1305 */ 1306 if ((rc = dbAllocDmap(bmp, dp, blkno, nblocks)) == 0) 1307 *results = blkno; 1308 1309 return (rc); 1310 } 1311 1312 return -ENOSPC; 1313 } 1314 1315 1316 /* 1317 * NAME: dbAllocAG() 1318 * 1319 * FUNCTION: attempt to allocate the specified number of contiguous 1320 * free blocks within the specified allocation group. 1321 * 1322 * unless the allocation group size is equal to the number 1323 * of blocks per dmap, the dmap control pages will be used to 1324 * find the required free space, if available. we start the 1325 * search at the highest dmap control page level which 1326 * distinctly describes the allocation group's free space 1327 * (i.e. the highest level at which the allocation group's 1328 * free space is not mixed in with that of any other group). 1329 * in addition, we start the search within this level at a 1330 * height of the dmapctl dmtree at which the nodes distinctly 1331 * describe the allocation group's free space. at this height, 1332 * the allocation group's free space may be represented by 1 1333 * or two sub-trees, depending on the allocation group size. 1334 * we search the top nodes of these subtrees left to right for 1335 * sufficient free space. if sufficient free space is found, 1336 * the subtree is searched to find the leftmost leaf that 1337 * has free space. once we have made it to the leaf, we 1338 * move the search to the next lower level dmap control page 1339 * corresponding to this leaf. we continue down the dmap control 1340 * pages until we find the dmap that contains or starts the 1341 * sufficient free space and we allocate at this dmap. 1342 * 1343 * if the allocation group size is equal to the dmap size, 1344 * we'll start at the dmap corresponding to the allocation 1345 * group and attempt the allocation at this level. 1346 * 1347 * the dmap control page search is also not performed if the 1348 * allocation group is completely free and we go to the first 1349 * dmap of the allocation group to do the allocation. this is 1350 * done because the allocation group may be part (not the first 1351 * part) of a larger binary buddy system, causing the dmap 1352 * control pages to indicate no free space (NOFREE) within 1353 * the allocation group. 1354 * 1355 * PARAMETERS: 1356 * bmp - pointer to bmap descriptor 1357 * agno - allocation group number. 1358 * nblocks - actual number of contiguous free blocks desired. 1359 * l2nb - log2 number of contiguous free blocks desired. 1360 * results - on successful return, set to the starting block number 1361 * of the newly allocated range. 1362 * 1363 * RETURN VALUES: 1364 * 0 - success 1365 * -ENOSPC - insufficient disk resources 1366 * -EIO - i/o error 1367 * 1368 * note: IWRITE_LOCK(ipmap) held on entry/exit; 1369 */ 1370 static int 1371 dbAllocAG(struct bmap * bmp, int agno, s64 nblocks, int l2nb, s64 * results) 1372 { 1373 struct metapage *mp; 1374 struct dmapctl *dcp; 1375 int rc, ti, i, k, m, n, agperlev; 1376 s64 blkno, lblkno; 1377 int budmin; 1378 1379 /* allocation request should not be for more than the 1380 * allocation group size. 1381 */ 1382 if (l2nb > bmp->db_agl2size) { 1383 jfs_error(bmp->db_ipbmap->i_sb, 1384 "dbAllocAG: allocation request is larger than the " 1385 "allocation group size"); 1386 return -EIO; 1387 } 1388 1389 /* determine the starting block number of the allocation 1390 * group. 1391 */ 1392 blkno = (s64) agno << bmp->db_agl2size; 1393 1394 /* check if the allocation group size is the minimum allocation 1395 * group size or if the allocation group is completely free. if 1396 * the allocation group size is the minimum size of BPERDMAP (i.e. 1397 * 1 dmap), there is no need to search the dmap control page (below) 1398 * that fully describes the allocation group since the allocation 1399 * group is already fully described by a dmap. in this case, we 1400 * just call dbAllocCtl() to search the dmap tree and allocate the 1401 * required space if available. 1402 * 1403 * if the allocation group is completely free, dbAllocCtl() is 1404 * also called to allocate the required space. this is done for 1405 * two reasons. first, it makes no sense searching the dmap control 1406 * pages for free space when we know that free space exists. second, 1407 * the dmap control pages may indicate that the allocation group 1408 * has no free space if the allocation group is part (not the first 1409 * part) of a larger binary buddy system. 1410 */ 1411 if (bmp->db_agsize == BPERDMAP 1412 || bmp->db_agfree[agno] == bmp->db_agsize) { 1413 rc = dbAllocCtl(bmp, nblocks, l2nb, blkno, results); 1414 if ((rc == -ENOSPC) && 1415 (bmp->db_agfree[agno] == bmp->db_agsize)) { 1416 printk(KERN_ERR "blkno = %Lx, blocks = %Lx\n", 1417 (unsigned long long) blkno, 1418 (unsigned long long) nblocks); 1419 jfs_error(bmp->db_ipbmap->i_sb, 1420 "dbAllocAG: dbAllocCtl failed in free AG"); 1421 } 1422 return (rc); 1423 } 1424 1425 /* the buffer for the dmap control page that fully describes the 1426 * allocation group. 1427 */ 1428 lblkno = BLKTOCTL(blkno, bmp->db_l2nbperpage, bmp->db_aglevel); 1429 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0); 1430 if (mp == NULL) 1431 return -EIO; 1432 dcp = (struct dmapctl *) mp->data; 1433 budmin = dcp->budmin; 1434 1435 if (dcp->leafidx != cpu_to_le32(CTLLEAFIND)) { 1436 jfs_error(bmp->db_ipbmap->i_sb, 1437 "dbAllocAG: Corrupt dmapctl page"); 1438 release_metapage(mp); 1439 return -EIO; 1440 } 1441 1442 /* search the subtree(s) of the dmap control page that describes 1443 * the allocation group, looking for sufficient free space. to begin, 1444 * determine how many allocation groups are represented in a dmap 1445 * control page at the control page level (i.e. L0, L1, L2) that 1446 * fully describes an allocation group. next, determine the starting 1447 * tree index of this allocation group within the control page. 1448 */ 1449 agperlev = 1450 (1 << (L2LPERCTL - (bmp->db_agheight << 1))) / bmp->db_agwidth; 1451 ti = bmp->db_agstart + bmp->db_agwidth * (agno & (agperlev - 1)); 1452 1453 /* dmap control page trees fan-out by 4 and a single allocation 1454 * group may be described by 1 or 2 subtrees within the ag level 1455 * dmap control page, depending upon the ag size. examine the ag's 1456 * subtrees for sufficient free space, starting with the leftmost 1457 * subtree. 1458 */ 1459 for (i = 0; i < bmp->db_agwidth; i++, ti++) { 1460 /* is there sufficient free space ? 1461 */ 1462 if (l2nb > dcp->stree[ti]) 1463 continue; 1464 1465 /* sufficient free space found in a subtree. now search down 1466 * the subtree to find the leftmost leaf that describes this 1467 * free space. 1468 */ 1469 for (k = bmp->db_agheight; k > 0; k--) { 1470 for (n = 0, m = (ti << 2) + 1; n < 4; n++) { 1471 if (l2nb <= dcp->stree[m + n]) { 1472 ti = m + n; 1473 break; 1474 } 1475 } 1476 if (n == 4) { 1477 jfs_error(bmp->db_ipbmap->i_sb, 1478 "dbAllocAG: failed descending stree"); 1479 release_metapage(mp); 1480 return -EIO; 1481 } 1482 } 1483 1484 /* determine the block number within the file system 1485 * that corresponds to this leaf. 1486 */ 1487 if (bmp->db_aglevel == 2) 1488 blkno = 0; 1489 else if (bmp->db_aglevel == 1) 1490 blkno &= ~(MAXL1SIZE - 1); 1491 else /* bmp->db_aglevel == 0 */ 1492 blkno &= ~(MAXL0SIZE - 1); 1493 1494 blkno += 1495 ((s64) (ti - le32_to_cpu(dcp->leafidx))) << budmin; 1496 1497 /* release the buffer in preparation for going down 1498 * the next level of dmap control pages. 1499 */ 1500 release_metapage(mp); 1501 1502 /* check if we need to continue to search down the lower 1503 * level dmap control pages. we need to if the number of 1504 * blocks required is less than maximum number of blocks 1505 * described at the next lower level. 1506 */ 1507 if (l2nb < budmin) { 1508 1509 /* search the lower level dmap control pages to get 1510 * the starting block number of the dmap that 1511 * contains or starts off the free space. 1512 */ 1513 if ((rc = 1514 dbFindCtl(bmp, l2nb, bmp->db_aglevel - 1, 1515 &blkno))) { 1516 if (rc == -ENOSPC) { 1517 jfs_error(bmp->db_ipbmap->i_sb, 1518 "dbAllocAG: control page " 1519 "inconsistent"); 1520 return -EIO; 1521 } 1522 return (rc); 1523 } 1524 } 1525 1526 /* allocate the blocks. 1527 */ 1528 rc = dbAllocCtl(bmp, nblocks, l2nb, blkno, results); 1529 if (rc == -ENOSPC) { 1530 jfs_error(bmp->db_ipbmap->i_sb, 1531 "dbAllocAG: unable to allocate blocks"); 1532 rc = -EIO; 1533 } 1534 return (rc); 1535 } 1536 1537 /* no space in the allocation group. release the buffer and 1538 * return -ENOSPC. 1539 */ 1540 release_metapage(mp); 1541 1542 return -ENOSPC; 1543 } 1544 1545 1546 /* 1547 * NAME: dbAllocAny() 1548 * 1549 * FUNCTION: attempt to allocate the specified number of contiguous 1550 * free blocks anywhere in the file system. 1551 * 1552 * dbAllocAny() attempts to find the sufficient free space by 1553 * searching down the dmap control pages, starting with the 1554 * highest level (i.e. L0, L1, L2) control page. if free space 1555 * large enough to satisfy the desired free space is found, the 1556 * desired free space is allocated. 1557 * 1558 * PARAMETERS: 1559 * bmp - pointer to bmap descriptor 1560 * nblocks - actual number of contiguous free blocks desired. 1561 * l2nb - log2 number of contiguous free blocks desired. 1562 * results - on successful return, set to the starting block number 1563 * of the newly allocated range. 1564 * 1565 * RETURN VALUES: 1566 * 0 - success 1567 * -ENOSPC - insufficient disk resources 1568 * -EIO - i/o error 1569 * 1570 * serialization: IWRITE_LOCK(ipbmap) held on entry/exit; 1571 */ 1572 static int dbAllocAny(struct bmap * bmp, s64 nblocks, int l2nb, s64 * results) 1573 { 1574 int rc; 1575 s64 blkno = 0; 1576 1577 /* starting with the top level dmap control page, search 1578 * down the dmap control levels for sufficient free space. 1579 * if free space is found, dbFindCtl() returns the starting 1580 * block number of the dmap that contains or starts off the 1581 * range of free space. 1582 */ 1583 if ((rc = dbFindCtl(bmp, l2nb, bmp->db_maxlevel, &blkno))) 1584 return (rc); 1585 1586 /* allocate the blocks. 1587 */ 1588 rc = dbAllocCtl(bmp, nblocks, l2nb, blkno, results); 1589 if (rc == -ENOSPC) { 1590 jfs_error(bmp->db_ipbmap->i_sb, 1591 "dbAllocAny: unable to allocate blocks"); 1592 return -EIO; 1593 } 1594 return (rc); 1595 } 1596 1597 1598 /* 1599 * NAME: dbDiscardAG() 1600 * 1601 * FUNCTION: attempt to discard (TRIM) all free blocks of specific AG 1602 * 1603 * algorithm: 1604 * 1) allocate blocks, as large as possible and save them 1605 * while holding IWRITE_LOCK on ipbmap 1606 * 2) trim all these saved block/length values 1607 * 3) mark the blocks free again 1608 * 1609 * benefit: 1610 * - we work only on one ag at some time, minimizing how long we 1611 * need to lock ipbmap 1612 * - reading / writing the fs is possible most time, even on 1613 * trimming 1614 * 1615 * downside: 1616 * - we write two times to the dmapctl and dmap pages 1617 * - but for me, this seems the best way, better ideas? 1618 * /TR 2012 1619 * 1620 * PARAMETERS: 1621 * ip - pointer to in-core inode 1622 * agno - ag to trim 1623 * minlen - minimum value of contiguous blocks 1624 * 1625 * RETURN VALUES: 1626 * s64 - actual number of blocks trimmed 1627 */ 1628 s64 dbDiscardAG(struct inode *ip, int agno, s64 minlen) 1629 { 1630 struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap; 1631 struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap; 1632 s64 nblocks, blkno; 1633 u64 trimmed = 0; 1634 int rc, l2nb; 1635 struct super_block *sb = ipbmap->i_sb; 1636 1637 struct range2trim { 1638 u64 blkno; 1639 u64 nblocks; 1640 } *totrim, *tt; 1641 1642 /* max blkno / nblocks pairs to trim */ 1643 int count = 0, range_cnt; 1644 u64 max_ranges; 1645 1646 /* prevent others from writing new stuff here, while trimming */ 1647 IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP); 1648 1649 nblocks = bmp->db_agfree[agno]; 1650 max_ranges = nblocks; 1651 do_div(max_ranges, minlen); 1652 range_cnt = min_t(u64, max_ranges + 1, 32 * 1024); 1653 totrim = kmalloc(sizeof(struct range2trim) * range_cnt, GFP_NOFS); 1654 if (totrim == NULL) { 1655 jfs_error(bmp->db_ipbmap->i_sb, 1656 "dbDiscardAG: no memory for trim array"); 1657 IWRITE_UNLOCK(ipbmap); 1658 return 0; 1659 } 1660 1661 tt = totrim; 1662 while (nblocks >= minlen) { 1663 l2nb = BLKSTOL2(nblocks); 1664 1665 /* 0 = okay, -EIO = fatal, -ENOSPC -> try smaller block */ 1666 rc = dbAllocAG(bmp, agno, nblocks, l2nb, &blkno); 1667 if (rc == 0) { 1668 tt->blkno = blkno; 1669 tt->nblocks = nblocks; 1670 tt++; count++; 1671 1672 /* the whole ag is free, trim now */ 1673 if (bmp->db_agfree[agno] == 0) 1674 break; 1675 1676 /* give a hint for the next while */ 1677 nblocks = bmp->db_agfree[agno]; 1678 continue; 1679 } else if (rc == -ENOSPC) { 1680 /* search for next smaller log2 block */ 1681 l2nb = BLKSTOL2(nblocks) - 1; 1682 nblocks = 1 << l2nb; 1683 } else { 1684 /* Trim any already allocated blocks */ 1685 jfs_error(bmp->db_ipbmap->i_sb, 1686 "dbDiscardAG: -EIO"); 1687 break; 1688 } 1689 1690 /* check, if our trim array is full */ 1691 if (unlikely(count >= range_cnt - 1)) 1692 break; 1693 } 1694 IWRITE_UNLOCK(ipbmap); 1695 1696 tt->nblocks = 0; /* mark the current end */ 1697 for (tt = totrim; tt->nblocks != 0; tt++) { 1698 /* when mounted with online discard, dbFree() will 1699 * call jfs_issue_discard() itself */ 1700 if (!(JFS_SBI(sb)->flag & JFS_DISCARD)) 1701 jfs_issue_discard(ip, tt->blkno, tt->nblocks); 1702 dbFree(ip, tt->blkno, tt->nblocks); 1703 trimmed += tt->nblocks; 1704 } 1705 kfree(totrim); 1706 1707 return trimmed; 1708 } 1709 1710 /* 1711 * NAME: dbFindCtl() 1712 * 1713 * FUNCTION: starting at a specified dmap control page level and block 1714 * number, search down the dmap control levels for a range of 1715 * contiguous free blocks large enough to satisfy an allocation 1716 * request for the specified number of free blocks. 1717 * 1718 * if sufficient contiguous free blocks are found, this routine 1719 * returns the starting block number within a dmap page that 1720 * contains or starts a range of contiqious free blocks that 1721 * is sufficient in size. 1722 * 1723 * PARAMETERS: 1724 * bmp - pointer to bmap descriptor 1725 * level - starting dmap control page level. 1726 * l2nb - log2 number of contiguous free blocks desired. 1727 * *blkno - on entry, starting block number for conducting the search. 1728 * on successful return, the first block within a dmap page 1729 * that contains or starts a range of contiguous free blocks. 1730 * 1731 * RETURN VALUES: 1732 * 0 - success 1733 * -ENOSPC - insufficient disk resources 1734 * -EIO - i/o error 1735 * 1736 * serialization: IWRITE_LOCK(ipbmap) held on entry/exit; 1737 */ 1738 static int dbFindCtl(struct bmap * bmp, int l2nb, int level, s64 * blkno) 1739 { 1740 int rc, leafidx, lev; 1741 s64 b, lblkno; 1742 struct dmapctl *dcp; 1743 int budmin; 1744 struct metapage *mp; 1745 1746 /* starting at the specified dmap control page level and block 1747 * number, search down the dmap control levels for the starting 1748 * block number of a dmap page that contains or starts off 1749 * sufficient free blocks. 1750 */ 1751 for (lev = level, b = *blkno; lev >= 0; lev--) { 1752 /* get the buffer of the dmap control page for the block 1753 * number and level (i.e. L0, L1, L2). 1754 */ 1755 lblkno = BLKTOCTL(b, bmp->db_l2nbperpage, lev); 1756 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0); 1757 if (mp == NULL) 1758 return -EIO; 1759 dcp = (struct dmapctl *) mp->data; 1760 budmin = dcp->budmin; 1761 1762 if (dcp->leafidx != cpu_to_le32(CTLLEAFIND)) { 1763 jfs_error(bmp->db_ipbmap->i_sb, 1764 "dbFindCtl: Corrupt dmapctl page"); 1765 release_metapage(mp); 1766 return -EIO; 1767 } 1768 1769 /* search the tree within the dmap control page for 1770 * sufficient free space. if sufficient free space is found, 1771 * dbFindLeaf() returns the index of the leaf at which 1772 * free space was found. 1773 */ 1774 rc = dbFindLeaf((dmtree_t *) dcp, l2nb, &leafidx); 1775 1776 /* release the buffer. 1777 */ 1778 release_metapage(mp); 1779 1780 /* space found ? 1781 */ 1782 if (rc) { 1783 if (lev != level) { 1784 jfs_error(bmp->db_ipbmap->i_sb, 1785 "dbFindCtl: dmap inconsistent"); 1786 return -EIO; 1787 } 1788 return -ENOSPC; 1789 } 1790 1791 /* adjust the block number to reflect the location within 1792 * the dmap control page (i.e. the leaf) at which free 1793 * space was found. 1794 */ 1795 b += (((s64) leafidx) << budmin); 1796 1797 /* we stop the search at this dmap control page level if 1798 * the number of blocks required is greater than or equal 1799 * to the maximum number of blocks described at the next 1800 * (lower) level. 1801 */ 1802 if (l2nb >= budmin) 1803 break; 1804 } 1805 1806 *blkno = b; 1807 return (0); 1808 } 1809 1810 1811 /* 1812 * NAME: dbAllocCtl() 1813 * 1814 * FUNCTION: attempt to allocate a specified number of contiguous 1815 * blocks starting within a specific dmap. 1816 * 1817 * this routine is called by higher level routines that search 1818 * the dmap control pages above the actual dmaps for contiguous 1819 * free space. the result of successful searches by these 1820 * routines are the starting block numbers within dmaps, with 1821 * the dmaps themselves containing the desired contiguous free 1822 * space or starting a contiguous free space of desired size 1823 * that is made up of the blocks of one or more dmaps. these 1824 * calls should not fail due to insufficent resources. 1825 * 1826 * this routine is called in some cases where it is not known 1827 * whether it will fail due to insufficient resources. more 1828 * specifically, this occurs when allocating from an allocation 1829 * group whose size is equal to the number of blocks per dmap. 1830 * in this case, the dmap control pages are not examined prior 1831 * to calling this routine (to save pathlength) and the call 1832 * might fail. 1833 * 1834 * for a request size that fits within a dmap, this routine relies 1835 * upon the dmap's dmtree to find the requested contiguous free 1836 * space. for request sizes that are larger than a dmap, the 1837 * requested free space will start at the first block of the 1838 * first dmap (i.e. blkno). 1839 * 1840 * PARAMETERS: 1841 * bmp - pointer to bmap descriptor 1842 * nblocks - actual number of contiguous free blocks to allocate. 1843 * l2nb - log2 number of contiguous free blocks to allocate. 1844 * blkno - starting block number of the dmap to start the allocation 1845 * from. 1846 * results - on successful return, set to the starting block number 1847 * of the newly allocated range. 1848 * 1849 * RETURN VALUES: 1850 * 0 - success 1851 * -ENOSPC - insufficient disk resources 1852 * -EIO - i/o error 1853 * 1854 * serialization: IWRITE_LOCK(ipbmap) held on entry/exit; 1855 */ 1856 static int 1857 dbAllocCtl(struct bmap * bmp, s64 nblocks, int l2nb, s64 blkno, s64 * results) 1858 { 1859 int rc, nb; 1860 s64 b, lblkno, n; 1861 struct metapage *mp; 1862 struct dmap *dp; 1863 1864 /* check if the allocation request is confined to a single dmap. 1865 */ 1866 if (l2nb <= L2BPERDMAP) { 1867 /* get the buffer for the dmap. 1868 */ 1869 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage); 1870 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0); 1871 if (mp == NULL) 1872 return -EIO; 1873 dp = (struct dmap *) mp->data; 1874 1875 /* try to allocate the blocks. 1876 */ 1877 rc = dbAllocDmapLev(bmp, dp, (int) nblocks, l2nb, results); 1878 if (rc == 0) 1879 mark_metapage_dirty(mp); 1880 1881 release_metapage(mp); 1882 1883 return (rc); 1884 } 1885 1886 /* allocation request involving multiple dmaps. it must start on 1887 * a dmap boundary. 1888 */ 1889 assert((blkno & (BPERDMAP - 1)) == 0); 1890 1891 /* allocate the blocks dmap by dmap. 1892 */ 1893 for (n = nblocks, b = blkno; n > 0; n -= nb, b += nb) { 1894 /* get the buffer for the dmap. 1895 */ 1896 lblkno = BLKTODMAP(b, bmp->db_l2nbperpage); 1897 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0); 1898 if (mp == NULL) { 1899 rc = -EIO; 1900 goto backout; 1901 } 1902 dp = (struct dmap *) mp->data; 1903 1904 /* the dmap better be all free. 1905 */ 1906 if (dp->tree.stree[ROOT] != L2BPERDMAP) { 1907 release_metapage(mp); 1908 jfs_error(bmp->db_ipbmap->i_sb, 1909 "dbAllocCtl: the dmap is not all free"); 1910 rc = -EIO; 1911 goto backout; 1912 } 1913 1914 /* determine how many blocks to allocate from this dmap. 1915 */ 1916 nb = min(n, (s64)BPERDMAP); 1917 1918 /* allocate the blocks from the dmap. 1919 */ 1920 if ((rc = dbAllocDmap(bmp, dp, b, nb))) { 1921 release_metapage(mp); 1922 goto backout; 1923 } 1924 1925 /* write the buffer. 1926 */ 1927 write_metapage(mp); 1928 } 1929 1930 /* set the results (starting block number) and return. 1931 */ 1932 *results = blkno; 1933 return (0); 1934 1935 /* something failed in handling an allocation request involving 1936 * multiple dmaps. we'll try to clean up by backing out any 1937 * allocation that has already happened for this request. if 1938 * we fail in backing out the allocation, we'll mark the file 1939 * system to indicate that blocks have been leaked. 1940 */ 1941 backout: 1942 1943 /* try to backout the allocations dmap by dmap. 1944 */ 1945 for (n = nblocks - n, b = blkno; n > 0; 1946 n -= BPERDMAP, b += BPERDMAP) { 1947 /* get the buffer for this dmap. 1948 */ 1949 lblkno = BLKTODMAP(b, bmp->db_l2nbperpage); 1950 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0); 1951 if (mp == NULL) { 1952 /* could not back out. mark the file system 1953 * to indicate that we have leaked blocks. 1954 */ 1955 jfs_error(bmp->db_ipbmap->i_sb, 1956 "dbAllocCtl: I/O Error: Block Leakage."); 1957 continue; 1958 } 1959 dp = (struct dmap *) mp->data; 1960 1961 /* free the blocks is this dmap. 1962 */ 1963 if (dbFreeDmap(bmp, dp, b, BPERDMAP)) { 1964 /* could not back out. mark the file system 1965 * to indicate that we have leaked blocks. 1966 */ 1967 release_metapage(mp); 1968 jfs_error(bmp->db_ipbmap->i_sb, 1969 "dbAllocCtl: Block Leakage."); 1970 continue; 1971 } 1972 1973 /* write the buffer. 1974 */ 1975 write_metapage(mp); 1976 } 1977 1978 return (rc); 1979 } 1980 1981 1982 /* 1983 * NAME: dbAllocDmapLev() 1984 * 1985 * FUNCTION: attempt to allocate a specified number of contiguous blocks 1986 * from a specified dmap. 1987 * 1988 * this routine checks if the contiguous blocks are available. 1989 * if so, nblocks of blocks are allocated; otherwise, ENOSPC is 1990 * returned. 1991 * 1992 * PARAMETERS: 1993 * mp - pointer to bmap descriptor 1994 * dp - pointer to dmap to attempt to allocate blocks from. 1995 * l2nb - log2 number of contiguous block desired. 1996 * nblocks - actual number of contiguous block desired. 1997 * results - on successful return, set to the starting block number 1998 * of the newly allocated range. 1999 * 2000 * RETURN VALUES: 2001 * 0 - success 2002 * -ENOSPC - insufficient disk resources 2003 * -EIO - i/o error 2004 * 2005 * serialization: IREAD_LOCK(ipbmap), e.g., from dbAlloc(), or 2006 * IWRITE_LOCK(ipbmap), e.g., dbAllocCtl(), held on entry/exit; 2007 */ 2008 static int 2009 dbAllocDmapLev(struct bmap * bmp, 2010 struct dmap * dp, int nblocks, int l2nb, s64 * results) 2011 { 2012 s64 blkno; 2013 int leafidx, rc; 2014 2015 /* can't be more than a dmaps worth of blocks */ 2016 assert(l2nb <= L2BPERDMAP); 2017 2018 /* search the tree within the dmap page for sufficient 2019 * free space. if sufficient free space is found, dbFindLeaf() 2020 * returns the index of the leaf at which free space was found. 2021 */ 2022 if (dbFindLeaf((dmtree_t *) & dp->tree, l2nb, &leafidx)) 2023 return -ENOSPC; 2024 2025 /* determine the block number within the file system corresponding 2026 * to the leaf at which free space was found. 2027 */ 2028 blkno = le64_to_cpu(dp->start) + (leafidx << L2DBWORD); 2029 2030 /* if not all bits of the dmap word are free, get the starting 2031 * bit number within the dmap word of the required string of free 2032 * bits and adjust the block number with this value. 2033 */ 2034 if (dp->tree.stree[leafidx + LEAFIND] < BUDMIN) 2035 blkno += dbFindBits(le32_to_cpu(dp->wmap[leafidx]), l2nb); 2036 2037 /* allocate the blocks */ 2038 if ((rc = dbAllocDmap(bmp, dp, blkno, nblocks)) == 0) 2039 *results = blkno; 2040 2041 return (rc); 2042 } 2043 2044 2045 /* 2046 * NAME: dbAllocDmap() 2047 * 2048 * FUNCTION: adjust the disk allocation map to reflect the allocation 2049 * of a specified block range within a dmap. 2050 * 2051 * this routine allocates the specified blocks from the dmap 2052 * through a call to dbAllocBits(). if the allocation of the 2053 * block range causes the maximum string of free blocks within 2054 * the dmap to change (i.e. the value of the root of the dmap's 2055 * dmtree), this routine will cause this change to be reflected 2056 * up through the appropriate levels of the dmap control pages 2057 * by a call to dbAdjCtl() for the L0 dmap control page that 2058 * covers this dmap. 2059 * 2060 * PARAMETERS: 2061 * bmp - pointer to bmap descriptor 2062 * dp - pointer to dmap to allocate the block range from. 2063 * blkno - starting block number of the block to be allocated. 2064 * nblocks - number of blocks to be allocated. 2065 * 2066 * RETURN VALUES: 2067 * 0 - success 2068 * -EIO - i/o error 2069 * 2070 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit; 2071 */ 2072 static int dbAllocDmap(struct bmap * bmp, struct dmap * dp, s64 blkno, 2073 int nblocks) 2074 { 2075 s8 oldroot; 2076 int rc; 2077 2078 /* save the current value of the root (i.e. maximum free string) 2079 * of the dmap tree. 2080 */ 2081 oldroot = dp->tree.stree[ROOT]; 2082 2083 /* allocate the specified (blocks) bits */ 2084 dbAllocBits(bmp, dp, blkno, nblocks); 2085 2086 /* if the root has not changed, done. */ 2087 if (dp->tree.stree[ROOT] == oldroot) 2088 return (0); 2089 2090 /* root changed. bubble the change up to the dmap control pages. 2091 * if the adjustment of the upper level control pages fails, 2092 * backout the bit allocation (thus making everything consistent). 2093 */ 2094 if ((rc = dbAdjCtl(bmp, blkno, dp->tree.stree[ROOT], 1, 0))) 2095 dbFreeBits(bmp, dp, blkno, nblocks); 2096 2097 return (rc); 2098 } 2099 2100 2101 /* 2102 * NAME: dbFreeDmap() 2103 * 2104 * FUNCTION: adjust the disk allocation map to reflect the allocation 2105 * of a specified block range within a dmap. 2106 * 2107 * this routine frees the specified blocks from the dmap through 2108 * a call to dbFreeBits(). if the deallocation of the block range 2109 * causes the maximum string of free blocks within the dmap to 2110 * change (i.e. the value of the root of the dmap's dmtree), this 2111 * routine will cause this change to be reflected up through the 2112 * appropriate levels of the dmap control pages by a call to 2113 * dbAdjCtl() for the L0 dmap control page that covers this dmap. 2114 * 2115 * PARAMETERS: 2116 * bmp - pointer to bmap descriptor 2117 * dp - pointer to dmap to free the block range from. 2118 * blkno - starting block number of the block to be freed. 2119 * nblocks - number of blocks to be freed. 2120 * 2121 * RETURN VALUES: 2122 * 0 - success 2123 * -EIO - i/o error 2124 * 2125 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit; 2126 */ 2127 static int dbFreeDmap(struct bmap * bmp, struct dmap * dp, s64 blkno, 2128 int nblocks) 2129 { 2130 s8 oldroot; 2131 int rc = 0, word; 2132 2133 /* save the current value of the root (i.e. maximum free string) 2134 * of the dmap tree. 2135 */ 2136 oldroot = dp->tree.stree[ROOT]; 2137 2138 /* free the specified (blocks) bits */ 2139 rc = dbFreeBits(bmp, dp, blkno, nblocks); 2140 2141 /* if error or the root has not changed, done. */ 2142 if (rc || (dp->tree.stree[ROOT] == oldroot)) 2143 return (rc); 2144 2145 /* root changed. bubble the change up to the dmap control pages. 2146 * if the adjustment of the upper level control pages fails, 2147 * backout the deallocation. 2148 */ 2149 if ((rc = dbAdjCtl(bmp, blkno, dp->tree.stree[ROOT], 0, 0))) { 2150 word = (blkno & (BPERDMAP - 1)) >> L2DBWORD; 2151 2152 /* as part of backing out the deallocation, we will have 2153 * to back split the dmap tree if the deallocation caused 2154 * the freed blocks to become part of a larger binary buddy 2155 * system. 2156 */ 2157 if (dp->tree.stree[word] == NOFREE) 2158 dbBackSplit((dmtree_t *) & dp->tree, word); 2159 2160 dbAllocBits(bmp, dp, blkno, nblocks); 2161 } 2162 2163 return (rc); 2164 } 2165 2166 2167 /* 2168 * NAME: dbAllocBits() 2169 * 2170 * FUNCTION: allocate a specified block range from a dmap. 2171 * 2172 * this routine updates the dmap to reflect the working 2173 * state allocation of the specified block range. it directly 2174 * updates the bits of the working map and causes the adjustment 2175 * of the binary buddy system described by the dmap's dmtree 2176 * leaves to reflect the bits allocated. it also causes the 2177 * dmap's dmtree, as a whole, to reflect the allocated range. 2178 * 2179 * PARAMETERS: 2180 * bmp - pointer to bmap descriptor 2181 * dp - pointer to dmap to allocate bits from. 2182 * blkno - starting block number of the bits to be allocated. 2183 * nblocks - number of bits to be allocated. 2184 * 2185 * RETURN VALUES: none 2186 * 2187 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit; 2188 */ 2189 static void dbAllocBits(struct bmap * bmp, struct dmap * dp, s64 blkno, 2190 int nblocks) 2191 { 2192 int dbitno, word, rembits, nb, nwords, wbitno, nw, agno; 2193 dmtree_t *tp = (dmtree_t *) & dp->tree; 2194 int size; 2195 s8 *leaf; 2196 2197 /* pick up a pointer to the leaves of the dmap tree */ 2198 leaf = dp->tree.stree + LEAFIND; 2199 2200 /* determine the bit number and word within the dmap of the 2201 * starting block. 2202 */ 2203 dbitno = blkno & (BPERDMAP - 1); 2204 word = dbitno >> L2DBWORD; 2205 2206 /* block range better be within the dmap */ 2207 assert(dbitno + nblocks <= BPERDMAP); 2208 2209 /* allocate the bits of the dmap's words corresponding to the block 2210 * range. not all bits of the first and last words may be contained 2211 * within the block range. if this is the case, we'll work against 2212 * those words (i.e. partial first and/or last) on an individual basis 2213 * (a single pass), allocating the bits of interest by hand and 2214 * updating the leaf corresponding to the dmap word. a single pass 2215 * will be used for all dmap words fully contained within the 2216 * specified range. within this pass, the bits of all fully contained 2217 * dmap words will be marked as free in a single shot and the leaves 2218 * will be updated. a single leaf may describe the free space of 2219 * multiple dmap words, so we may update only a subset of the actual 2220 * leaves corresponding to the dmap words of the block range. 2221 */ 2222 for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) { 2223 /* determine the bit number within the word and 2224 * the number of bits within the word. 2225 */ 2226 wbitno = dbitno & (DBWORD - 1); 2227 nb = min(rembits, DBWORD - wbitno); 2228 2229 /* check if only part of a word is to be allocated. 2230 */ 2231 if (nb < DBWORD) { 2232 /* allocate (set to 1) the appropriate bits within 2233 * this dmap word. 2234 */ 2235 dp->wmap[word] |= cpu_to_le32(ONES << (DBWORD - nb) 2236 >> wbitno); 2237 2238 /* update the leaf for this dmap word. in addition 2239 * to setting the leaf value to the binary buddy max 2240 * of the updated dmap word, dbSplit() will split 2241 * the binary system of the leaves if need be. 2242 */ 2243 dbSplit(tp, word, BUDMIN, 2244 dbMaxBud((u8 *) & dp->wmap[word])); 2245 2246 word += 1; 2247 } else { 2248 /* one or more dmap words are fully contained 2249 * within the block range. determine how many 2250 * words and allocate (set to 1) the bits of these 2251 * words. 2252 */ 2253 nwords = rembits >> L2DBWORD; 2254 memset(&dp->wmap[word], (int) ONES, nwords * 4); 2255 2256 /* determine how many bits. 2257 */ 2258 nb = nwords << L2DBWORD; 2259 2260 /* now update the appropriate leaves to reflect 2261 * the allocated words. 2262 */ 2263 for (; nwords > 0; nwords -= nw) { 2264 if (leaf[word] < BUDMIN) { 2265 jfs_error(bmp->db_ipbmap->i_sb, 2266 "dbAllocBits: leaf page " 2267 "corrupt"); 2268 break; 2269 } 2270 2271 /* determine what the leaf value should be 2272 * updated to as the minimum of the l2 number 2273 * of bits being allocated and the l2 number 2274 * of bits currently described by this leaf. 2275 */ 2276 size = min((int)leaf[word], NLSTOL2BSZ(nwords)); 2277 2278 /* update the leaf to reflect the allocation. 2279 * in addition to setting the leaf value to 2280 * NOFREE, dbSplit() will split the binary 2281 * system of the leaves to reflect the current 2282 * allocation (size). 2283 */ 2284 dbSplit(tp, word, size, NOFREE); 2285 2286 /* get the number of dmap words handled */ 2287 nw = BUDSIZE(size, BUDMIN); 2288 word += nw; 2289 } 2290 } 2291 } 2292 2293 /* update the free count for this dmap */ 2294 le32_add_cpu(&dp->nfree, -nblocks); 2295 2296 BMAP_LOCK(bmp); 2297 2298 /* if this allocation group is completely free, 2299 * update the maximum allocation group number if this allocation 2300 * group is the new max. 2301 */ 2302 agno = blkno >> bmp->db_agl2size; 2303 if (agno > bmp->db_maxag) 2304 bmp->db_maxag = agno; 2305 2306 /* update the free count for the allocation group and map */ 2307 bmp->db_agfree[agno] -= nblocks; 2308 bmp->db_nfree -= nblocks; 2309 2310 BMAP_UNLOCK(bmp); 2311 } 2312 2313 2314 /* 2315 * NAME: dbFreeBits() 2316 * 2317 * FUNCTION: free a specified block range from a dmap. 2318 * 2319 * this routine updates the dmap to reflect the working 2320 * state allocation of the specified block range. it directly 2321 * updates the bits of the working map and causes the adjustment 2322 * of the binary buddy system described by the dmap's dmtree 2323 * leaves to reflect the bits freed. it also causes the dmap's 2324 * dmtree, as a whole, to reflect the deallocated range. 2325 * 2326 * PARAMETERS: 2327 * bmp - pointer to bmap descriptor 2328 * dp - pointer to dmap to free bits from. 2329 * blkno - starting block number of the bits to be freed. 2330 * nblocks - number of bits to be freed. 2331 * 2332 * RETURN VALUES: 0 for success 2333 * 2334 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit; 2335 */ 2336 static int dbFreeBits(struct bmap * bmp, struct dmap * dp, s64 blkno, 2337 int nblocks) 2338 { 2339 int dbitno, word, rembits, nb, nwords, wbitno, nw, agno; 2340 dmtree_t *tp = (dmtree_t *) & dp->tree; 2341 int rc = 0; 2342 int size; 2343 2344 /* determine the bit number and word within the dmap of the 2345 * starting block. 2346 */ 2347 dbitno = blkno & (BPERDMAP - 1); 2348 word = dbitno >> L2DBWORD; 2349 2350 /* block range better be within the dmap. 2351 */ 2352 assert(dbitno + nblocks <= BPERDMAP); 2353 2354 /* free the bits of the dmaps words corresponding to the block range. 2355 * not all bits of the first and last words may be contained within 2356 * the block range. if this is the case, we'll work against those 2357 * words (i.e. partial first and/or last) on an individual basis 2358 * (a single pass), freeing the bits of interest by hand and updating 2359 * the leaf corresponding to the dmap word. a single pass will be used 2360 * for all dmap words fully contained within the specified range. 2361 * within this pass, the bits of all fully contained dmap words will 2362 * be marked as free in a single shot and the leaves will be updated. a 2363 * single leaf may describe the free space of multiple dmap words, 2364 * so we may update only a subset of the actual leaves corresponding 2365 * to the dmap words of the block range. 2366 * 2367 * dbJoin() is used to update leaf values and will join the binary 2368 * buddy system of the leaves if the new leaf values indicate this 2369 * should be done. 2370 */ 2371 for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) { 2372 /* determine the bit number within the word and 2373 * the number of bits within the word. 2374 */ 2375 wbitno = dbitno & (DBWORD - 1); 2376 nb = min(rembits, DBWORD - wbitno); 2377 2378 /* check if only part of a word is to be freed. 2379 */ 2380 if (nb < DBWORD) { 2381 /* free (zero) the appropriate bits within this 2382 * dmap word. 2383 */ 2384 dp->wmap[word] &= 2385 cpu_to_le32(~(ONES << (DBWORD - nb) 2386 >> wbitno)); 2387 2388 /* update the leaf for this dmap word. 2389 */ 2390 rc = dbJoin(tp, word, 2391 dbMaxBud((u8 *) & dp->wmap[word])); 2392 if (rc) 2393 return rc; 2394 2395 word += 1; 2396 } else { 2397 /* one or more dmap words are fully contained 2398 * within the block range. determine how many 2399 * words and free (zero) the bits of these words. 2400 */ 2401 nwords = rembits >> L2DBWORD; 2402 memset(&dp->wmap[word], 0, nwords * 4); 2403 2404 /* determine how many bits. 2405 */ 2406 nb = nwords << L2DBWORD; 2407 2408 /* now update the appropriate leaves to reflect 2409 * the freed words. 2410 */ 2411 for (; nwords > 0; nwords -= nw) { 2412 /* determine what the leaf value should be 2413 * updated to as the minimum of the l2 number 2414 * of bits being freed and the l2 (max) number 2415 * of bits that can be described by this leaf. 2416 */ 2417 size = 2418 min(LITOL2BSZ 2419 (word, L2LPERDMAP, BUDMIN), 2420 NLSTOL2BSZ(nwords)); 2421 2422 /* update the leaf. 2423 */ 2424 rc = dbJoin(tp, word, size); 2425 if (rc) 2426 return rc; 2427 2428 /* get the number of dmap words handled. 2429 */ 2430 nw = BUDSIZE(size, BUDMIN); 2431 word += nw; 2432 } 2433 } 2434 } 2435 2436 /* update the free count for this dmap. 2437 */ 2438 le32_add_cpu(&dp->nfree, nblocks); 2439 2440 BMAP_LOCK(bmp); 2441 2442 /* update the free count for the allocation group and 2443 * map. 2444 */ 2445 agno = blkno >> bmp->db_agl2size; 2446 bmp->db_nfree += nblocks; 2447 bmp->db_agfree[agno] += nblocks; 2448 2449 /* check if this allocation group is not completely free and 2450 * if it is currently the maximum (rightmost) allocation group. 2451 * if so, establish the new maximum allocation group number by 2452 * searching left for the first allocation group with allocation. 2453 */ 2454 if ((bmp->db_agfree[agno] == bmp->db_agsize && agno == bmp->db_maxag) || 2455 (agno == bmp->db_numag - 1 && 2456 bmp->db_agfree[agno] == (bmp-> db_mapsize & (BPERDMAP - 1)))) { 2457 while (bmp->db_maxag > 0) { 2458 bmp->db_maxag -= 1; 2459 if (bmp->db_agfree[bmp->db_maxag] != 2460 bmp->db_agsize) 2461 break; 2462 } 2463 2464 /* re-establish the allocation group preference if the 2465 * current preference is right of the maximum allocation 2466 * group. 2467 */ 2468 if (bmp->db_agpref > bmp->db_maxag) 2469 bmp->db_agpref = bmp->db_maxag; 2470 } 2471 2472 BMAP_UNLOCK(bmp); 2473 2474 return 0; 2475 } 2476 2477 2478 /* 2479 * NAME: dbAdjCtl() 2480 * 2481 * FUNCTION: adjust a dmap control page at a specified level to reflect 2482 * the change in a lower level dmap or dmap control page's 2483 * maximum string of free blocks (i.e. a change in the root 2484 * of the lower level object's dmtree) due to the allocation 2485 * or deallocation of a range of blocks with a single dmap. 2486 * 2487 * on entry, this routine is provided with the new value of 2488 * the lower level dmap or dmap control page root and the 2489 * starting block number of the block range whose allocation 2490 * or deallocation resulted in the root change. this range 2491 * is respresented by a single leaf of the current dmapctl 2492 * and the leaf will be updated with this value, possibly 2493 * causing a binary buddy system within the leaves to be 2494 * split or joined. the update may also cause the dmapctl's 2495 * dmtree to be updated. 2496 * 2497 * if the adjustment of the dmap control page, itself, causes its 2498 * root to change, this change will be bubbled up to the next dmap 2499 * control level by a recursive call to this routine, specifying 2500 * the new root value and the next dmap control page level to 2501 * be adjusted. 2502 * PARAMETERS: 2503 * bmp - pointer to bmap descriptor 2504 * blkno - the first block of a block range within a dmap. it is 2505 * the allocation or deallocation of this block range that 2506 * requires the dmap control page to be adjusted. 2507 * newval - the new value of the lower level dmap or dmap control 2508 * page root. 2509 * alloc - 'true' if adjustment is due to an allocation. 2510 * level - current level of dmap control page (i.e. L0, L1, L2) to 2511 * be adjusted. 2512 * 2513 * RETURN VALUES: 2514 * 0 - success 2515 * -EIO - i/o error 2516 * 2517 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit; 2518 */ 2519 static int 2520 dbAdjCtl(struct bmap * bmp, s64 blkno, int newval, int alloc, int level) 2521 { 2522 struct metapage *mp; 2523 s8 oldroot; 2524 int oldval; 2525 s64 lblkno; 2526 struct dmapctl *dcp; 2527 int rc, leafno, ti; 2528 2529 /* get the buffer for the dmap control page for the specified 2530 * block number and control page level. 2531 */ 2532 lblkno = BLKTOCTL(blkno, bmp->db_l2nbperpage, level); 2533 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0); 2534 if (mp == NULL) 2535 return -EIO; 2536 dcp = (struct dmapctl *) mp->data; 2537 2538 if (dcp->leafidx != cpu_to_le32(CTLLEAFIND)) { 2539 jfs_error(bmp->db_ipbmap->i_sb, 2540 "dbAdjCtl: Corrupt dmapctl page"); 2541 release_metapage(mp); 2542 return -EIO; 2543 } 2544 2545 /* determine the leaf number corresponding to the block and 2546 * the index within the dmap control tree. 2547 */ 2548 leafno = BLKTOCTLLEAF(blkno, dcp->budmin); 2549 ti = leafno + le32_to_cpu(dcp->leafidx); 2550 2551 /* save the current leaf value and the current root level (i.e. 2552 * maximum l2 free string described by this dmapctl). 2553 */ 2554 oldval = dcp->stree[ti]; 2555 oldroot = dcp->stree[ROOT]; 2556 2557 /* check if this is a control page update for an allocation. 2558 * if so, update the leaf to reflect the new leaf value using 2559 * dbSplit(); otherwise (deallocation), use dbJoin() to update 2560 * the leaf with the new value. in addition to updating the 2561 * leaf, dbSplit() will also split the binary buddy system of 2562 * the leaves, if required, and bubble new values within the 2563 * dmapctl tree, if required. similarly, dbJoin() will join 2564 * the binary buddy system of leaves and bubble new values up 2565 * the dmapctl tree as required by the new leaf value. 2566 */ 2567 if (alloc) { 2568 /* check if we are in the middle of a binary buddy 2569 * system. this happens when we are performing the 2570 * first allocation out of an allocation group that 2571 * is part (not the first part) of a larger binary 2572 * buddy system. if we are in the middle, back split 2573 * the system prior to calling dbSplit() which assumes 2574 * that it is at the front of a binary buddy system. 2575 */ 2576 if (oldval == NOFREE) { 2577 rc = dbBackSplit((dmtree_t *) dcp, leafno); 2578 if (rc) 2579 return rc; 2580 oldval = dcp->stree[ti]; 2581 } 2582 dbSplit((dmtree_t *) dcp, leafno, dcp->budmin, newval); 2583 } else { 2584 rc = dbJoin((dmtree_t *) dcp, leafno, newval); 2585 if (rc) 2586 return rc; 2587 } 2588 2589 /* check if the root of the current dmap control page changed due 2590 * to the update and if the current dmap control page is not at 2591 * the current top level (i.e. L0, L1, L2) of the map. if so (i.e. 2592 * root changed and this is not the top level), call this routine 2593 * again (recursion) for the next higher level of the mapping to 2594 * reflect the change in root for the current dmap control page. 2595 */ 2596 if (dcp->stree[ROOT] != oldroot) { 2597 /* are we below the top level of the map. if so, 2598 * bubble the root up to the next higher level. 2599 */ 2600 if (level < bmp->db_maxlevel) { 2601 /* bubble up the new root of this dmap control page to 2602 * the next level. 2603 */ 2604 if ((rc = 2605 dbAdjCtl(bmp, blkno, dcp->stree[ROOT], alloc, 2606 level + 1))) { 2607 /* something went wrong in bubbling up the new 2608 * root value, so backout the changes to the 2609 * current dmap control page. 2610 */ 2611 if (alloc) { 2612 dbJoin((dmtree_t *) dcp, leafno, 2613 oldval); 2614 } else { 2615 /* the dbJoin() above might have 2616 * caused a larger binary buddy system 2617 * to form and we may now be in the 2618 * middle of it. if this is the case, 2619 * back split the buddies. 2620 */ 2621 if (dcp->stree[ti] == NOFREE) 2622 dbBackSplit((dmtree_t *) 2623 dcp, leafno); 2624 dbSplit((dmtree_t *) dcp, leafno, 2625 dcp->budmin, oldval); 2626 } 2627 2628 /* release the buffer and return the error. 2629 */ 2630 release_metapage(mp); 2631 return (rc); 2632 } 2633 } else { 2634 /* we're at the top level of the map. update 2635 * the bmap control page to reflect the size 2636 * of the maximum free buddy system. 2637 */ 2638 assert(level == bmp->db_maxlevel); 2639 if (bmp->db_maxfreebud != oldroot) { 2640 jfs_error(bmp->db_ipbmap->i_sb, 2641 "dbAdjCtl: the maximum free buddy is " 2642 "not the old root"); 2643 } 2644 bmp->db_maxfreebud = dcp->stree[ROOT]; 2645 } 2646 } 2647 2648 /* write the buffer. 2649 */ 2650 write_metapage(mp); 2651 2652 return (0); 2653 } 2654 2655 2656 /* 2657 * NAME: dbSplit() 2658 * 2659 * FUNCTION: update the leaf of a dmtree with a new value, splitting 2660 * the leaf from the binary buddy system of the dmtree's 2661 * leaves, as required. 2662 * 2663 * PARAMETERS: 2664 * tp - pointer to the tree containing the leaf. 2665 * leafno - the number of the leaf to be updated. 2666 * splitsz - the size the binary buddy system starting at the leaf 2667 * must be split to, specified as the log2 number of blocks. 2668 * newval - the new value for the leaf. 2669 * 2670 * RETURN VALUES: none 2671 * 2672 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit; 2673 */ 2674 static void dbSplit(dmtree_t * tp, int leafno, int splitsz, int newval) 2675 { 2676 int budsz; 2677 int cursz; 2678 s8 *leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx); 2679 2680 /* check if the leaf needs to be split. 2681 */ 2682 if (leaf[leafno] > tp->dmt_budmin) { 2683 /* the split occurs by cutting the buddy system in half 2684 * at the specified leaf until we reach the specified 2685 * size. pick up the starting split size (current size 2686 * - 1 in l2) and the corresponding buddy size. 2687 */ 2688 cursz = leaf[leafno] - 1; 2689 budsz = BUDSIZE(cursz, tp->dmt_budmin); 2690 2691 /* split until we reach the specified size. 2692 */ 2693 while (cursz >= splitsz) { 2694 /* update the buddy's leaf with its new value. 2695 */ 2696 dbAdjTree(tp, leafno ^ budsz, cursz); 2697 2698 /* on to the next size and buddy. 2699 */ 2700 cursz -= 1; 2701 budsz >>= 1; 2702 } 2703 } 2704 2705 /* adjust the dmap tree to reflect the specified leaf's new 2706 * value. 2707 */ 2708 dbAdjTree(tp, leafno, newval); 2709 } 2710 2711 2712 /* 2713 * NAME: dbBackSplit() 2714 * 2715 * FUNCTION: back split the binary buddy system of dmtree leaves 2716 * that hold a specified leaf until the specified leaf 2717 * starts its own binary buddy system. 2718 * 2719 * the allocators typically perform allocations at the start 2720 * of binary buddy systems and dbSplit() is used to accomplish 2721 * any required splits. in some cases, however, allocation 2722 * may occur in the middle of a binary system and requires a 2723 * back split, with the split proceeding out from the middle of 2724 * the system (less efficient) rather than the start of the 2725 * system (more efficient). the cases in which a back split 2726 * is required are rare and are limited to the first allocation 2727 * within an allocation group which is a part (not first part) 2728 * of a larger binary buddy system and a few exception cases 2729 * in which a previous join operation must be backed out. 2730 * 2731 * PARAMETERS: 2732 * tp - pointer to the tree containing the leaf. 2733 * leafno - the number of the leaf to be updated. 2734 * 2735 * RETURN VALUES: none 2736 * 2737 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit; 2738 */ 2739 static int dbBackSplit(dmtree_t * tp, int leafno) 2740 { 2741 int budsz, bud, w, bsz, size; 2742 int cursz; 2743 s8 *leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx); 2744 2745 /* leaf should be part (not first part) of a binary 2746 * buddy system. 2747 */ 2748 assert(leaf[leafno] == NOFREE); 2749 2750 /* the back split is accomplished by iteratively finding the leaf 2751 * that starts the buddy system that contains the specified leaf and 2752 * splitting that system in two. this iteration continues until 2753 * the specified leaf becomes the start of a buddy system. 2754 * 2755 * determine maximum possible l2 size for the specified leaf. 2756 */ 2757 size = 2758 LITOL2BSZ(leafno, le32_to_cpu(tp->dmt_l2nleafs), 2759 tp->dmt_budmin); 2760 2761 /* determine the number of leaves covered by this size. this 2762 * is the buddy size that we will start with as we search for 2763 * the buddy system that contains the specified leaf. 2764 */ 2765 budsz = BUDSIZE(size, tp->dmt_budmin); 2766 2767 /* back split. 2768 */ 2769 while (leaf[leafno] == NOFREE) { 2770 /* find the leftmost buddy leaf. 2771 */ 2772 for (w = leafno, bsz = budsz;; bsz <<= 1, 2773 w = (w < bud) ? w : bud) { 2774 if (bsz >= le32_to_cpu(tp->dmt_nleafs)) { 2775 jfs_err("JFS: block map error in dbBackSplit"); 2776 return -EIO; 2777 } 2778 2779 /* determine the buddy. 2780 */ 2781 bud = w ^ bsz; 2782 2783 /* check if this buddy is the start of the system. 2784 */ 2785 if (leaf[bud] != NOFREE) { 2786 /* split the leaf at the start of the 2787 * system in two. 2788 */ 2789 cursz = leaf[bud] - 1; 2790 dbSplit(tp, bud, cursz, cursz); 2791 break; 2792 } 2793 } 2794 } 2795 2796 if (leaf[leafno] != size) { 2797 jfs_err("JFS: wrong leaf value in dbBackSplit"); 2798 return -EIO; 2799 } 2800 return 0; 2801 } 2802 2803 2804 /* 2805 * NAME: dbJoin() 2806 * 2807 * FUNCTION: update the leaf of a dmtree with a new value, joining 2808 * the leaf with other leaves of the dmtree into a multi-leaf 2809 * binary buddy system, as required. 2810 * 2811 * PARAMETERS: 2812 * tp - pointer to the tree containing the leaf. 2813 * leafno - the number of the leaf to be updated. 2814 * newval - the new value for the leaf. 2815 * 2816 * RETURN VALUES: none 2817 */ 2818 static int dbJoin(dmtree_t * tp, int leafno, int newval) 2819 { 2820 int budsz, buddy; 2821 s8 *leaf; 2822 2823 /* can the new leaf value require a join with other leaves ? 2824 */ 2825 if (newval >= tp->dmt_budmin) { 2826 /* pickup a pointer to the leaves of the tree. 2827 */ 2828 leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx); 2829 2830 /* try to join the specified leaf into a large binary 2831 * buddy system. the join proceeds by attempting to join 2832 * the specified leafno with its buddy (leaf) at new value. 2833 * if the join occurs, we attempt to join the left leaf 2834 * of the joined buddies with its buddy at new value + 1. 2835 * we continue to join until we find a buddy that cannot be 2836 * joined (does not have a value equal to the size of the 2837 * last join) or until all leaves have been joined into a 2838 * single system. 2839 * 2840 * get the buddy size (number of words covered) of 2841 * the new value. 2842 */ 2843 budsz = BUDSIZE(newval, tp->dmt_budmin); 2844 2845 /* try to join. 2846 */ 2847 while (budsz < le32_to_cpu(tp->dmt_nleafs)) { 2848 /* get the buddy leaf. 2849 */ 2850 buddy = leafno ^ budsz; 2851 2852 /* if the leaf's new value is greater than its 2853 * buddy's value, we join no more. 2854 */ 2855 if (newval > leaf[buddy]) 2856 break; 2857 2858 /* It shouldn't be less */ 2859 if (newval < leaf[buddy]) 2860 return -EIO; 2861 2862 /* check which (leafno or buddy) is the left buddy. 2863 * the left buddy gets to claim the blocks resulting 2864 * from the join while the right gets to claim none. 2865 * the left buddy is also eligible to participate in 2866 * a join at the next higher level while the right 2867 * is not. 2868 * 2869 */ 2870 if (leafno < buddy) { 2871 /* leafno is the left buddy. 2872 */ 2873 dbAdjTree(tp, buddy, NOFREE); 2874 } else { 2875 /* buddy is the left buddy and becomes 2876 * leafno. 2877 */ 2878 dbAdjTree(tp, leafno, NOFREE); 2879 leafno = buddy; 2880 } 2881 2882 /* on to try the next join. 2883 */ 2884 newval += 1; 2885 budsz <<= 1; 2886 } 2887 } 2888 2889 /* update the leaf value. 2890 */ 2891 dbAdjTree(tp, leafno, newval); 2892 2893 return 0; 2894 } 2895 2896 2897 /* 2898 * NAME: dbAdjTree() 2899 * 2900 * FUNCTION: update a leaf of a dmtree with a new value, adjusting 2901 * the dmtree, as required, to reflect the new leaf value. 2902 * the combination of any buddies must already be done before 2903 * this is called. 2904 * 2905 * PARAMETERS: 2906 * tp - pointer to the tree to be adjusted. 2907 * leafno - the number of the leaf to be updated. 2908 * newval - the new value for the leaf. 2909 * 2910 * RETURN VALUES: none 2911 */ 2912 static void dbAdjTree(dmtree_t * tp, int leafno, int newval) 2913 { 2914 int lp, pp, k; 2915 int max; 2916 2917 /* pick up the index of the leaf for this leafno. 2918 */ 2919 lp = leafno + le32_to_cpu(tp->dmt_leafidx); 2920 2921 /* is the current value the same as the old value ? if so, 2922 * there is nothing to do. 2923 */ 2924 if (tp->dmt_stree[lp] == newval) 2925 return; 2926 2927 /* set the new value. 2928 */ 2929 tp->dmt_stree[lp] = newval; 2930 2931 /* bubble the new value up the tree as required. 2932 */ 2933 for (k = 0; k < le32_to_cpu(tp->dmt_height); k++) { 2934 /* get the index of the first leaf of the 4 leaf 2935 * group containing the specified leaf (leafno). 2936 */ 2937 lp = ((lp - 1) & ~0x03) + 1; 2938 2939 /* get the index of the parent of this 4 leaf group. 2940 */ 2941 pp = (lp - 1) >> 2; 2942 2943 /* determine the maximum of the 4 leaves. 2944 */ 2945 max = TREEMAX(&tp->dmt_stree[lp]); 2946 2947 /* if the maximum of the 4 is the same as the 2948 * parent's value, we're done. 2949 */ 2950 if (tp->dmt_stree[pp] == max) 2951 break; 2952 2953 /* parent gets new value. 2954 */ 2955 tp->dmt_stree[pp] = max; 2956 2957 /* parent becomes leaf for next go-round. 2958 */ 2959 lp = pp; 2960 } 2961 } 2962 2963 2964 /* 2965 * NAME: dbFindLeaf() 2966 * 2967 * FUNCTION: search a dmtree_t for sufficient free blocks, returning 2968 * the index of a leaf describing the free blocks if 2969 * sufficient free blocks are found. 2970 * 2971 * the search starts at the top of the dmtree_t tree and 2972 * proceeds down the tree to the leftmost leaf with sufficient 2973 * free space. 2974 * 2975 * PARAMETERS: 2976 * tp - pointer to the tree to be searched. 2977 * l2nb - log2 number of free blocks to search for. 2978 * leafidx - return pointer to be set to the index of the leaf 2979 * describing at least l2nb free blocks if sufficient 2980 * free blocks are found. 2981 * 2982 * RETURN VALUES: 2983 * 0 - success 2984 * -ENOSPC - insufficient free blocks. 2985 */ 2986 static int dbFindLeaf(dmtree_t * tp, int l2nb, int *leafidx) 2987 { 2988 int ti, n = 0, k, x = 0; 2989 2990 /* first check the root of the tree to see if there is 2991 * sufficient free space. 2992 */ 2993 if (l2nb > tp->dmt_stree[ROOT]) 2994 return -ENOSPC; 2995 2996 /* sufficient free space available. now search down the tree 2997 * starting at the next level for the leftmost leaf that 2998 * describes sufficient free space. 2999 */ 3000 for (k = le32_to_cpu(tp->dmt_height), ti = 1; 3001 k > 0; k--, ti = ((ti + n) << 2) + 1) { 3002 /* search the four nodes at this level, starting from 3003 * the left. 3004 */ 3005 for (x = ti, n = 0; n < 4; n++) { 3006 /* sufficient free space found. move to the next 3007 * level (or quit if this is the last level). 3008 */ 3009 if (l2nb <= tp->dmt_stree[x + n]) 3010 break; 3011 } 3012 3013 /* better have found something since the higher 3014 * levels of the tree said it was here. 3015 */ 3016 assert(n < 4); 3017 } 3018 3019 /* set the return to the leftmost leaf describing sufficient 3020 * free space. 3021 */ 3022 *leafidx = x + n - le32_to_cpu(tp->dmt_leafidx); 3023 3024 return (0); 3025 } 3026 3027 3028 /* 3029 * NAME: dbFindBits() 3030 * 3031 * FUNCTION: find a specified number of binary buddy free bits within a 3032 * dmap bitmap word value. 3033 * 3034 * this routine searches the bitmap value for (1 << l2nb) free 3035 * bits at (1 << l2nb) alignments within the value. 3036 * 3037 * PARAMETERS: 3038 * word - dmap bitmap word value. 3039 * l2nb - number of free bits specified as a log2 number. 3040 * 3041 * RETURN VALUES: 3042 * starting bit number of free bits. 3043 */ 3044 static int dbFindBits(u32 word, int l2nb) 3045 { 3046 int bitno, nb; 3047 u32 mask; 3048 3049 /* get the number of bits. 3050 */ 3051 nb = 1 << l2nb; 3052 assert(nb <= DBWORD); 3053 3054 /* complement the word so we can use a mask (i.e. 0s represent 3055 * free bits) and compute the mask. 3056 */ 3057 word = ~word; 3058 mask = ONES << (DBWORD - nb); 3059 3060 /* scan the word for nb free bits at nb alignments. 3061 */ 3062 for (bitno = 0; mask != 0; bitno += nb, mask >>= nb) { 3063 if ((mask & word) == mask) 3064 break; 3065 } 3066 3067 ASSERT(bitno < 32); 3068 3069 /* return the bit number. 3070 */ 3071 return (bitno); 3072 } 3073 3074 3075 /* 3076 * NAME: dbMaxBud(u8 *cp) 3077 * 3078 * FUNCTION: determine the largest binary buddy string of free 3079 * bits within 32-bits of the map. 3080 * 3081 * PARAMETERS: 3082 * cp - pointer to the 32-bit value. 3083 * 3084 * RETURN VALUES: 3085 * largest binary buddy of free bits within a dmap word. 3086 */ 3087 static int dbMaxBud(u8 * cp) 3088 { 3089 signed char tmp1, tmp2; 3090 3091 /* check if the wmap word is all free. if so, the 3092 * free buddy size is BUDMIN. 3093 */ 3094 if (*((uint *) cp) == 0) 3095 return (BUDMIN); 3096 3097 /* check if the wmap word is half free. if so, the 3098 * free buddy size is BUDMIN-1. 3099 */ 3100 if (*((u16 *) cp) == 0 || *((u16 *) cp + 1) == 0) 3101 return (BUDMIN - 1); 3102 3103 /* not all free or half free. determine the free buddy 3104 * size thru table lookup using quarters of the wmap word. 3105 */ 3106 tmp1 = max(budtab[cp[2]], budtab[cp[3]]); 3107 tmp2 = max(budtab[cp[0]], budtab[cp[1]]); 3108 return (max(tmp1, tmp2)); 3109 } 3110 3111 3112 /* 3113 * NAME: cnttz(uint word) 3114 * 3115 * FUNCTION: determine the number of trailing zeros within a 32-bit 3116 * value. 3117 * 3118 * PARAMETERS: 3119 * value - 32-bit value to be examined. 3120 * 3121 * RETURN VALUES: 3122 * count of trailing zeros 3123 */ 3124 static int cnttz(u32 word) 3125 { 3126 int n; 3127 3128 for (n = 0; n < 32; n++, word >>= 1) { 3129 if (word & 0x01) 3130 break; 3131 } 3132 3133 return (n); 3134 } 3135 3136 3137 /* 3138 * NAME: cntlz(u32 value) 3139 * 3140 * FUNCTION: determine the number of leading zeros within a 32-bit 3141 * value. 3142 * 3143 * PARAMETERS: 3144 * value - 32-bit value to be examined. 3145 * 3146 * RETURN VALUES: 3147 * count of leading zeros 3148 */ 3149 static int cntlz(u32 value) 3150 { 3151 int n; 3152 3153 for (n = 0; n < 32; n++, value <<= 1) { 3154 if (value & HIGHORDER) 3155 break; 3156 } 3157 return (n); 3158 } 3159 3160 3161 /* 3162 * NAME: blkstol2(s64 nb) 3163 * 3164 * FUNCTION: convert a block count to its log2 value. if the block 3165 * count is not a l2 multiple, it is rounded up to the next 3166 * larger l2 multiple. 3167 * 3168 * PARAMETERS: 3169 * nb - number of blocks 3170 * 3171 * RETURN VALUES: 3172 * log2 number of blocks 3173 */ 3174 static int blkstol2(s64 nb) 3175 { 3176 int l2nb; 3177 s64 mask; /* meant to be signed */ 3178 3179 mask = (s64) 1 << (64 - 1); 3180 3181 /* count the leading bits. 3182 */ 3183 for (l2nb = 0; l2nb < 64; l2nb++, mask >>= 1) { 3184 /* leading bit found. 3185 */ 3186 if (nb & mask) { 3187 /* determine the l2 value. 3188 */ 3189 l2nb = (64 - 1) - l2nb; 3190 3191 /* check if we need to round up. 3192 */ 3193 if (~mask & nb) 3194 l2nb++; 3195 3196 return (l2nb); 3197 } 3198 } 3199 assert(0); 3200 return 0; /* fix compiler warning */ 3201 } 3202 3203 3204 /* 3205 * NAME: dbAllocBottomUp() 3206 * 3207 * FUNCTION: alloc the specified block range from the working block 3208 * allocation map. 3209 * 3210 * the blocks will be alloc from the working map one dmap 3211 * at a time. 3212 * 3213 * PARAMETERS: 3214 * ip - pointer to in-core inode; 3215 * blkno - starting block number to be freed. 3216 * nblocks - number of blocks to be freed. 3217 * 3218 * RETURN VALUES: 3219 * 0 - success 3220 * -EIO - i/o error 3221 */ 3222 int dbAllocBottomUp(struct inode *ip, s64 blkno, s64 nblocks) 3223 { 3224 struct metapage *mp; 3225 struct dmap *dp; 3226 int nb, rc; 3227 s64 lblkno, rem; 3228 struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap; 3229 struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap; 3230 3231 IREAD_LOCK(ipbmap, RDWRLOCK_DMAP); 3232 3233 /* block to be allocated better be within the mapsize. */ 3234 ASSERT(nblocks <= bmp->db_mapsize - blkno); 3235 3236 /* 3237 * allocate the blocks a dmap at a time. 3238 */ 3239 mp = NULL; 3240 for (rem = nblocks; rem > 0; rem -= nb, blkno += nb) { 3241 /* release previous dmap if any */ 3242 if (mp) { 3243 write_metapage(mp); 3244 } 3245 3246 /* get the buffer for the current dmap. */ 3247 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage); 3248 mp = read_metapage(ipbmap, lblkno, PSIZE, 0); 3249 if (mp == NULL) { 3250 IREAD_UNLOCK(ipbmap); 3251 return -EIO; 3252 } 3253 dp = (struct dmap *) mp->data; 3254 3255 /* determine the number of blocks to be allocated from 3256 * this dmap. 3257 */ 3258 nb = min(rem, BPERDMAP - (blkno & (BPERDMAP - 1))); 3259 3260 /* allocate the blocks. */ 3261 if ((rc = dbAllocDmapBU(bmp, dp, blkno, nb))) { 3262 release_metapage(mp); 3263 IREAD_UNLOCK(ipbmap); 3264 return (rc); 3265 } 3266 } 3267 3268 /* write the last buffer. */ 3269 write_metapage(mp); 3270 3271 IREAD_UNLOCK(ipbmap); 3272 3273 return (0); 3274 } 3275 3276 3277 static int dbAllocDmapBU(struct bmap * bmp, struct dmap * dp, s64 blkno, 3278 int nblocks) 3279 { 3280 int rc; 3281 int dbitno, word, rembits, nb, nwords, wbitno, agno; 3282 s8 oldroot; 3283 struct dmaptree *tp = (struct dmaptree *) & dp->tree; 3284 3285 /* save the current value of the root (i.e. maximum free string) 3286 * of the dmap tree. 3287 */ 3288 oldroot = tp->stree[ROOT]; 3289 3290 /* determine the bit number and word within the dmap of the 3291 * starting block. 3292 */ 3293 dbitno = blkno & (BPERDMAP - 1); 3294 word = dbitno >> L2DBWORD; 3295 3296 /* block range better be within the dmap */ 3297 assert(dbitno + nblocks <= BPERDMAP); 3298 3299 /* allocate the bits of the dmap's words corresponding to the block 3300 * range. not all bits of the first and last words may be contained 3301 * within the block range. if this is the case, we'll work against 3302 * those words (i.e. partial first and/or last) on an individual basis 3303 * (a single pass), allocating the bits of interest by hand and 3304 * updating the leaf corresponding to the dmap word. a single pass 3305 * will be used for all dmap words fully contained within the 3306 * specified range. within this pass, the bits of all fully contained 3307 * dmap words will be marked as free in a single shot and the leaves 3308 * will be updated. a single leaf may describe the free space of 3309 * multiple dmap words, so we may update only a subset of the actual 3310 * leaves corresponding to the dmap words of the block range. 3311 */ 3312 for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) { 3313 /* determine the bit number within the word and 3314 * the number of bits within the word. 3315 */ 3316 wbitno = dbitno & (DBWORD - 1); 3317 nb = min(rembits, DBWORD - wbitno); 3318 3319 /* check if only part of a word is to be allocated. 3320 */ 3321 if (nb < DBWORD) { 3322 /* allocate (set to 1) the appropriate bits within 3323 * this dmap word. 3324 */ 3325 dp->wmap[word] |= cpu_to_le32(ONES << (DBWORD - nb) 3326 >> wbitno); 3327 3328 word++; 3329 } else { 3330 /* one or more dmap words are fully contained 3331 * within the block range. determine how many 3332 * words and allocate (set to 1) the bits of these 3333 * words. 3334 */ 3335 nwords = rembits >> L2DBWORD; 3336 memset(&dp->wmap[word], (int) ONES, nwords * 4); 3337 3338 /* determine how many bits */ 3339 nb = nwords << L2DBWORD; 3340 word += nwords; 3341 } 3342 } 3343 3344 /* update the free count for this dmap */ 3345 le32_add_cpu(&dp->nfree, -nblocks); 3346 3347 /* reconstruct summary tree */ 3348 dbInitDmapTree(dp); 3349 3350 BMAP_LOCK(bmp); 3351 3352 /* if this allocation group is completely free, 3353 * update the highest active allocation group number 3354 * if this allocation group is the new max. 3355 */ 3356 agno = blkno >> bmp->db_agl2size; 3357 if (agno > bmp->db_maxag) 3358 bmp->db_maxag = agno; 3359 3360 /* update the free count for the allocation group and map */ 3361 bmp->db_agfree[agno] -= nblocks; 3362 bmp->db_nfree -= nblocks; 3363 3364 BMAP_UNLOCK(bmp); 3365 3366 /* if the root has not changed, done. */ 3367 if (tp->stree[ROOT] == oldroot) 3368 return (0); 3369 3370 /* root changed. bubble the change up to the dmap control pages. 3371 * if the adjustment of the upper level control pages fails, 3372 * backout the bit allocation (thus making everything consistent). 3373 */ 3374 if ((rc = dbAdjCtl(bmp, blkno, tp->stree[ROOT], 1, 0))) 3375 dbFreeBits(bmp, dp, blkno, nblocks); 3376 3377 return (rc); 3378 } 3379 3380 3381 /* 3382 * NAME: dbExtendFS() 3383 * 3384 * FUNCTION: extend bmap from blkno for nblocks; 3385 * dbExtendFS() updates bmap ready for dbAllocBottomUp(); 3386 * 3387 * L2 3388 * | 3389 * L1---------------------------------L1 3390 * | | 3391 * L0---------L0---------L0 L0---------L0---------L0 3392 * | | | | | | 3393 * d0,...,dn d0,...,dn d0,...,dn d0,...,dn d0,...,dn d0,.,dm; 3394 * L2L1L0d0,...,dnL0d0,...,dnL0d0,...,dnL1L0d0,...,dnL0d0,...,dnL0d0,..dm 3395 * 3396 * <---old---><----------------------------extend-----------------------> 3397 */ 3398 int dbExtendFS(struct inode *ipbmap, s64 blkno, s64 nblocks) 3399 { 3400 struct jfs_sb_info *sbi = JFS_SBI(ipbmap->i_sb); 3401 int nbperpage = sbi->nbperpage; 3402 int i, i0 = true, j, j0 = true, k, n; 3403 s64 newsize; 3404 s64 p; 3405 struct metapage *mp, *l2mp, *l1mp = NULL, *l0mp = NULL; 3406 struct dmapctl *l2dcp, *l1dcp, *l0dcp; 3407 struct dmap *dp; 3408 s8 *l0leaf, *l1leaf, *l2leaf; 3409 struct bmap *bmp = sbi->bmap; 3410 int agno, l2agsize, oldl2agsize; 3411 s64 ag_rem; 3412 3413 newsize = blkno + nblocks; 3414 3415 jfs_info("dbExtendFS: blkno:%Ld nblocks:%Ld newsize:%Ld", 3416 (long long) blkno, (long long) nblocks, (long long) newsize); 3417 3418 /* 3419 * initialize bmap control page. 3420 * 3421 * all the data in bmap control page should exclude 3422 * the mkfs hidden dmap page. 3423 */ 3424 3425 /* update mapsize */ 3426 bmp->db_mapsize = newsize; 3427 bmp->db_maxlevel = BMAPSZTOLEV(bmp->db_mapsize); 3428 3429 /* compute new AG size */ 3430 l2agsize = dbGetL2AGSize(newsize); 3431 oldl2agsize = bmp->db_agl2size; 3432 3433 bmp->db_agl2size = l2agsize; 3434 bmp->db_agsize = 1 << l2agsize; 3435 3436 /* compute new number of AG */ 3437 agno = bmp->db_numag; 3438 bmp->db_numag = newsize >> l2agsize; 3439 bmp->db_numag += ((u32) newsize % (u32) bmp->db_agsize) ? 1 : 0; 3440 3441 /* 3442 * reconfigure db_agfree[] 3443 * from old AG configuration to new AG configuration; 3444 * 3445 * coalesce contiguous k (newAGSize/oldAGSize) AGs; 3446 * i.e., (AGi, ..., AGj) where i = k*n and j = k*(n+1) - 1 to AGn; 3447 * note: new AG size = old AG size * (2**x). 3448 */ 3449 if (l2agsize == oldl2agsize) 3450 goto extend; 3451 k = 1 << (l2agsize - oldl2agsize); 3452 ag_rem = bmp->db_agfree[0]; /* save agfree[0] */ 3453 for (i = 0, n = 0; i < agno; n++) { 3454 bmp->db_agfree[n] = 0; /* init collection point */ 3455 3456 /* coalesce contiguous k AGs; */ 3457 for (j = 0; j < k && i < agno; j++, i++) { 3458 /* merge AGi to AGn */ 3459 bmp->db_agfree[n] += bmp->db_agfree[i]; 3460 } 3461 } 3462 bmp->db_agfree[0] += ag_rem; /* restore agfree[0] */ 3463 3464 for (; n < MAXAG; n++) 3465 bmp->db_agfree[n] = 0; 3466 3467 /* 3468 * update highest active ag number 3469 */ 3470 3471 bmp->db_maxag = bmp->db_maxag / k; 3472 3473 /* 3474 * extend bmap 3475 * 3476 * update bit maps and corresponding level control pages; 3477 * global control page db_nfree, db_agfree[agno], db_maxfreebud; 3478 */ 3479 extend: 3480 /* get L2 page */ 3481 p = BMAPBLKNO + nbperpage; /* L2 page */ 3482 l2mp = read_metapage(ipbmap, p, PSIZE, 0); 3483 if (!l2mp) { 3484 jfs_error(ipbmap->i_sb, "dbExtendFS: L2 page could not be read"); 3485 return -EIO; 3486 } 3487 l2dcp = (struct dmapctl *) l2mp->data; 3488 3489 /* compute start L1 */ 3490 k = blkno >> L2MAXL1SIZE; 3491 l2leaf = l2dcp->stree + CTLLEAFIND + k; 3492 p = BLKTOL1(blkno, sbi->l2nbperpage); /* L1 page */ 3493 3494 /* 3495 * extend each L1 in L2 3496 */ 3497 for (; k < LPERCTL; k++, p += nbperpage) { 3498 /* get L1 page */ 3499 if (j0) { 3500 /* read in L1 page: (blkno & (MAXL1SIZE - 1)) */ 3501 l1mp = read_metapage(ipbmap, p, PSIZE, 0); 3502 if (l1mp == NULL) 3503 goto errout; 3504 l1dcp = (struct dmapctl *) l1mp->data; 3505 3506 /* compute start L0 */ 3507 j = (blkno & (MAXL1SIZE - 1)) >> L2MAXL0SIZE; 3508 l1leaf = l1dcp->stree + CTLLEAFIND + j; 3509 p = BLKTOL0(blkno, sbi->l2nbperpage); 3510 j0 = false; 3511 } else { 3512 /* assign/init L1 page */ 3513 l1mp = get_metapage(ipbmap, p, PSIZE, 0); 3514 if (l1mp == NULL) 3515 goto errout; 3516 3517 l1dcp = (struct dmapctl *) l1mp->data; 3518 3519 /* compute start L0 */ 3520 j = 0; 3521 l1leaf = l1dcp->stree + CTLLEAFIND; 3522 p += nbperpage; /* 1st L0 of L1.k */ 3523 } 3524 3525 /* 3526 * extend each L0 in L1 3527 */ 3528 for (; j < LPERCTL; j++) { 3529 /* get L0 page */ 3530 if (i0) { 3531 /* read in L0 page: (blkno & (MAXL0SIZE - 1)) */ 3532 3533 l0mp = read_metapage(ipbmap, p, PSIZE, 0); 3534 if (l0mp == NULL) 3535 goto errout; 3536 l0dcp = (struct dmapctl *) l0mp->data; 3537 3538 /* compute start dmap */ 3539 i = (blkno & (MAXL0SIZE - 1)) >> 3540 L2BPERDMAP; 3541 l0leaf = l0dcp->stree + CTLLEAFIND + i; 3542 p = BLKTODMAP(blkno, 3543 sbi->l2nbperpage); 3544 i0 = false; 3545 } else { 3546 /* assign/init L0 page */ 3547 l0mp = get_metapage(ipbmap, p, PSIZE, 0); 3548 if (l0mp == NULL) 3549 goto errout; 3550 3551 l0dcp = (struct dmapctl *) l0mp->data; 3552 3553 /* compute start dmap */ 3554 i = 0; 3555 l0leaf = l0dcp->stree + CTLLEAFIND; 3556 p += nbperpage; /* 1st dmap of L0.j */ 3557 } 3558 3559 /* 3560 * extend each dmap in L0 3561 */ 3562 for (; i < LPERCTL; i++) { 3563 /* 3564 * reconstruct the dmap page, and 3565 * initialize corresponding parent L0 leaf 3566 */ 3567 if ((n = blkno & (BPERDMAP - 1))) { 3568 /* read in dmap page: */ 3569 mp = read_metapage(ipbmap, p, 3570 PSIZE, 0); 3571 if (mp == NULL) 3572 goto errout; 3573 n = min(nblocks, (s64)BPERDMAP - n); 3574 } else { 3575 /* assign/init dmap page */ 3576 mp = read_metapage(ipbmap, p, 3577 PSIZE, 0); 3578 if (mp == NULL) 3579 goto errout; 3580 3581 n = min(nblocks, (s64)BPERDMAP); 3582 } 3583 3584 dp = (struct dmap *) mp->data; 3585 *l0leaf = dbInitDmap(dp, blkno, n); 3586 3587 bmp->db_nfree += n; 3588 agno = le64_to_cpu(dp->start) >> l2agsize; 3589 bmp->db_agfree[agno] += n; 3590 3591 write_metapage(mp); 3592 3593 l0leaf++; 3594 p += nbperpage; 3595 3596 blkno += n; 3597 nblocks -= n; 3598 if (nblocks == 0) 3599 break; 3600 } /* for each dmap in a L0 */ 3601 3602 /* 3603 * build current L0 page from its leaves, and 3604 * initialize corresponding parent L1 leaf 3605 */ 3606 *l1leaf = dbInitDmapCtl(l0dcp, 0, ++i); 3607 write_metapage(l0mp); 3608 l0mp = NULL; 3609 3610 if (nblocks) 3611 l1leaf++; /* continue for next L0 */ 3612 else { 3613 /* more than 1 L0 ? */ 3614 if (j > 0) 3615 break; /* build L1 page */ 3616 else { 3617 /* summarize in global bmap page */ 3618 bmp->db_maxfreebud = *l1leaf; 3619 release_metapage(l1mp); 3620 release_metapage(l2mp); 3621 goto finalize; 3622 } 3623 } 3624 } /* for each L0 in a L1 */ 3625 3626 /* 3627 * build current L1 page from its leaves, and 3628 * initialize corresponding parent L2 leaf 3629 */ 3630 *l2leaf = dbInitDmapCtl(l1dcp, 1, ++j); 3631 write_metapage(l1mp); 3632 l1mp = NULL; 3633 3634 if (nblocks) 3635 l2leaf++; /* continue for next L1 */ 3636 else { 3637 /* more than 1 L1 ? */ 3638 if (k > 0) 3639 break; /* build L2 page */ 3640 else { 3641 /* summarize in global bmap page */ 3642 bmp->db_maxfreebud = *l2leaf; 3643 release_metapage(l2mp); 3644 goto finalize; 3645 } 3646 } 3647 } /* for each L1 in a L2 */ 3648 3649 jfs_error(ipbmap->i_sb, 3650 "dbExtendFS: function has not returned as expected"); 3651 errout: 3652 if (l0mp) 3653 release_metapage(l0mp); 3654 if (l1mp) 3655 release_metapage(l1mp); 3656 release_metapage(l2mp); 3657 return -EIO; 3658 3659 /* 3660 * finalize bmap control page 3661 */ 3662 finalize: 3663 3664 return 0; 3665 } 3666 3667 3668 /* 3669 * dbFinalizeBmap() 3670 */ 3671 void dbFinalizeBmap(struct inode *ipbmap) 3672 { 3673 struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap; 3674 int actags, inactags, l2nl; 3675 s64 ag_rem, actfree, inactfree, avgfree; 3676 int i, n; 3677 3678 /* 3679 * finalize bmap control page 3680 */ 3681 //finalize: 3682 /* 3683 * compute db_agpref: preferred ag to allocate from 3684 * (the leftmost ag with average free space in it); 3685 */ 3686 //agpref: 3687 /* get the number of active ags and inacitve ags */ 3688 actags = bmp->db_maxag + 1; 3689 inactags = bmp->db_numag - actags; 3690 ag_rem = bmp->db_mapsize & (bmp->db_agsize - 1); /* ??? */ 3691 3692 /* determine how many blocks are in the inactive allocation 3693 * groups. in doing this, we must account for the fact that 3694 * the rightmost group might be a partial group (i.e. file 3695 * system size is not a multiple of the group size). 3696 */ 3697 inactfree = (inactags && ag_rem) ? 3698 ((inactags - 1) << bmp->db_agl2size) + ag_rem 3699 : inactags << bmp->db_agl2size; 3700 3701 /* determine how many free blocks are in the active 3702 * allocation groups plus the average number of free blocks 3703 * within the active ags. 3704 */ 3705 actfree = bmp->db_nfree - inactfree; 3706 avgfree = (u32) actfree / (u32) actags; 3707 3708 /* if the preferred allocation group has not average free space. 3709 * re-establish the preferred group as the leftmost 3710 * group with average free space. 3711 */ 3712 if (bmp->db_agfree[bmp->db_agpref] < avgfree) { 3713 for (bmp->db_agpref = 0; bmp->db_agpref < actags; 3714 bmp->db_agpref++) { 3715 if (bmp->db_agfree[bmp->db_agpref] >= avgfree) 3716 break; 3717 } 3718 if (bmp->db_agpref >= bmp->db_numag) { 3719 jfs_error(ipbmap->i_sb, 3720 "cannot find ag with average freespace"); 3721 } 3722 } 3723 3724 /* 3725 * compute db_aglevel, db_agheight, db_width, db_agstart: 3726 * an ag is covered in aglevel dmapctl summary tree, 3727 * at agheight level height (from leaf) with agwidth number of nodes 3728 * each, which starts at agstart index node of the smmary tree node 3729 * array; 3730 */ 3731 bmp->db_aglevel = BMAPSZTOLEV(bmp->db_agsize); 3732 l2nl = 3733 bmp->db_agl2size - (L2BPERDMAP + bmp->db_aglevel * L2LPERCTL); 3734 bmp->db_agheight = l2nl >> 1; 3735 bmp->db_agwidth = 1 << (l2nl - (bmp->db_agheight << 1)); 3736 for (i = 5 - bmp->db_agheight, bmp->db_agstart = 0, n = 1; i > 0; 3737 i--) { 3738 bmp->db_agstart += n; 3739 n <<= 2; 3740 } 3741 3742 } 3743 3744 3745 /* 3746 * NAME: dbInitDmap()/ujfs_idmap_page() 3747 * 3748 * FUNCTION: initialize working/persistent bitmap of the dmap page 3749 * for the specified number of blocks: 3750 * 3751 * at entry, the bitmaps had been initialized as free (ZEROS); 3752 * The number of blocks will only account for the actually 3753 * existing blocks. Blocks which don't actually exist in 3754 * the aggregate will be marked as allocated (ONES); 3755 * 3756 * PARAMETERS: 3757 * dp - pointer to page of map 3758 * nblocks - number of blocks this page 3759 * 3760 * RETURNS: NONE 3761 */ 3762 static int dbInitDmap(struct dmap * dp, s64 Blkno, int nblocks) 3763 { 3764 int blkno, w, b, r, nw, nb, i; 3765 3766 /* starting block number within the dmap */ 3767 blkno = Blkno & (BPERDMAP - 1); 3768 3769 if (blkno == 0) { 3770 dp->nblocks = dp->nfree = cpu_to_le32(nblocks); 3771 dp->start = cpu_to_le64(Blkno); 3772 3773 if (nblocks == BPERDMAP) { 3774 memset(&dp->wmap[0], 0, LPERDMAP * 4); 3775 memset(&dp->pmap[0], 0, LPERDMAP * 4); 3776 goto initTree; 3777 } 3778 } else { 3779 le32_add_cpu(&dp->nblocks, nblocks); 3780 le32_add_cpu(&dp->nfree, nblocks); 3781 } 3782 3783 /* word number containing start block number */ 3784 w = blkno >> L2DBWORD; 3785 3786 /* 3787 * free the bits corresponding to the block range (ZEROS): 3788 * note: not all bits of the first and last words may be contained 3789 * within the block range. 3790 */ 3791 for (r = nblocks; r > 0; r -= nb, blkno += nb) { 3792 /* number of bits preceding range to be freed in the word */ 3793 b = blkno & (DBWORD - 1); 3794 /* number of bits to free in the word */ 3795 nb = min(r, DBWORD - b); 3796 3797 /* is partial word to be freed ? */ 3798 if (nb < DBWORD) { 3799 /* free (set to 0) from the bitmap word */ 3800 dp->wmap[w] &= cpu_to_le32(~(ONES << (DBWORD - nb) 3801 >> b)); 3802 dp->pmap[w] &= cpu_to_le32(~(ONES << (DBWORD - nb) 3803 >> b)); 3804 3805 /* skip the word freed */ 3806 w++; 3807 } else { 3808 /* free (set to 0) contiguous bitmap words */ 3809 nw = r >> L2DBWORD; 3810 memset(&dp->wmap[w], 0, nw * 4); 3811 memset(&dp->pmap[w], 0, nw * 4); 3812 3813 /* skip the words freed */ 3814 nb = nw << L2DBWORD; 3815 w += nw; 3816 } 3817 } 3818 3819 /* 3820 * mark bits following the range to be freed (non-existing 3821 * blocks) as allocated (ONES) 3822 */ 3823 3824 if (blkno == BPERDMAP) 3825 goto initTree; 3826 3827 /* the first word beyond the end of existing blocks */ 3828 w = blkno >> L2DBWORD; 3829 3830 /* does nblocks fall on a 32-bit boundary ? */ 3831 b = blkno & (DBWORD - 1); 3832 if (b) { 3833 /* mark a partial word allocated */ 3834 dp->wmap[w] = dp->pmap[w] = cpu_to_le32(ONES >> b); 3835 w++; 3836 } 3837 3838 /* set the rest of the words in the page to allocated (ONES) */ 3839 for (i = w; i < LPERDMAP; i++) 3840 dp->pmap[i] = dp->wmap[i] = cpu_to_le32(ONES); 3841 3842 /* 3843 * init tree 3844 */ 3845 initTree: 3846 return (dbInitDmapTree(dp)); 3847 } 3848 3849 3850 /* 3851 * NAME: dbInitDmapTree()/ujfs_complete_dmap() 3852 * 3853 * FUNCTION: initialize summary tree of the specified dmap: 3854 * 3855 * at entry, bitmap of the dmap has been initialized; 3856 * 3857 * PARAMETERS: 3858 * dp - dmap to complete 3859 * blkno - starting block number for this dmap 3860 * treemax - will be filled in with max free for this dmap 3861 * 3862 * RETURNS: max free string at the root of the tree 3863 */ 3864 static int dbInitDmapTree(struct dmap * dp) 3865 { 3866 struct dmaptree *tp; 3867 s8 *cp; 3868 int i; 3869 3870 /* init fixed info of tree */ 3871 tp = &dp->tree; 3872 tp->nleafs = cpu_to_le32(LPERDMAP); 3873 tp->l2nleafs = cpu_to_le32(L2LPERDMAP); 3874 tp->leafidx = cpu_to_le32(LEAFIND); 3875 tp->height = cpu_to_le32(4); 3876 tp->budmin = BUDMIN; 3877 3878 /* init each leaf from corresponding wmap word: 3879 * note: leaf is set to NOFREE(-1) if all blocks of corresponding 3880 * bitmap word are allocated. 3881 */ 3882 cp = tp->stree + le32_to_cpu(tp->leafidx); 3883 for (i = 0; i < LPERDMAP; i++) 3884 *cp++ = dbMaxBud((u8 *) & dp->wmap[i]); 3885 3886 /* build the dmap's binary buddy summary tree */ 3887 return (dbInitTree(tp)); 3888 } 3889 3890 3891 /* 3892 * NAME: dbInitTree()/ujfs_adjtree() 3893 * 3894 * FUNCTION: initialize binary buddy summary tree of a dmap or dmapctl. 3895 * 3896 * at entry, the leaves of the tree has been initialized 3897 * from corresponding bitmap word or root of summary tree 3898 * of the child control page; 3899 * configure binary buddy system at the leaf level, then 3900 * bubble up the values of the leaf nodes up the tree. 3901 * 3902 * PARAMETERS: 3903 * cp - Pointer to the root of the tree 3904 * l2leaves- Number of leaf nodes as a power of 2 3905 * l2min - Number of blocks that can be covered by a leaf 3906 * as a power of 2 3907 * 3908 * RETURNS: max free string at the root of the tree 3909 */ 3910 static int dbInitTree(struct dmaptree * dtp) 3911 { 3912 int l2max, l2free, bsize, nextb, i; 3913 int child, parent, nparent; 3914 s8 *tp, *cp, *cp1; 3915 3916 tp = dtp->stree; 3917 3918 /* Determine the maximum free string possible for the leaves */ 3919 l2max = le32_to_cpu(dtp->l2nleafs) + dtp->budmin; 3920 3921 /* 3922 * configure the leaf levevl into binary buddy system 3923 * 3924 * Try to combine buddies starting with a buddy size of 1 3925 * (i.e. two leaves). At a buddy size of 1 two buddy leaves 3926 * can be combined if both buddies have a maximum free of l2min; 3927 * the combination will result in the left-most buddy leaf having 3928 * a maximum free of l2min+1. 3929 * After processing all buddies for a given size, process buddies 3930 * at the next higher buddy size (i.e. current size * 2) and 3931 * the next maximum free (current free + 1). 3932 * This continues until the maximum possible buddy combination 3933 * yields maximum free. 3934 */ 3935 for (l2free = dtp->budmin, bsize = 1; l2free < l2max; 3936 l2free++, bsize = nextb) { 3937 /* get next buddy size == current buddy pair size */ 3938 nextb = bsize << 1; 3939 3940 /* scan each adjacent buddy pair at current buddy size */ 3941 for (i = 0, cp = tp + le32_to_cpu(dtp->leafidx); 3942 i < le32_to_cpu(dtp->nleafs); 3943 i += nextb, cp += nextb) { 3944 /* coalesce if both adjacent buddies are max free */ 3945 if (*cp == l2free && *(cp + bsize) == l2free) { 3946 *cp = l2free + 1; /* left take right */ 3947 *(cp + bsize) = -1; /* right give left */ 3948 } 3949 } 3950 } 3951 3952 /* 3953 * bubble summary information of leaves up the tree. 3954 * 3955 * Starting at the leaf node level, the four nodes described by 3956 * the higher level parent node are compared for a maximum free and 3957 * this maximum becomes the value of the parent node. 3958 * when all lower level nodes are processed in this fashion then 3959 * move up to the next level (parent becomes a lower level node) and 3960 * continue the process for that level. 3961 */ 3962 for (child = le32_to_cpu(dtp->leafidx), 3963 nparent = le32_to_cpu(dtp->nleafs) >> 2; 3964 nparent > 0; nparent >>= 2, child = parent) { 3965 /* get index of 1st node of parent level */ 3966 parent = (child - 1) >> 2; 3967 3968 /* set the value of the parent node as the maximum 3969 * of the four nodes of the current level. 3970 */ 3971 for (i = 0, cp = tp + child, cp1 = tp + parent; 3972 i < nparent; i++, cp += 4, cp1++) 3973 *cp1 = TREEMAX(cp); 3974 } 3975 3976 return (*tp); 3977 } 3978 3979 3980 /* 3981 * dbInitDmapCtl() 3982 * 3983 * function: initialize dmapctl page 3984 */ 3985 static int dbInitDmapCtl(struct dmapctl * dcp, int level, int i) 3986 { /* start leaf index not covered by range */ 3987 s8 *cp; 3988 3989 dcp->nleafs = cpu_to_le32(LPERCTL); 3990 dcp->l2nleafs = cpu_to_le32(L2LPERCTL); 3991 dcp->leafidx = cpu_to_le32(CTLLEAFIND); 3992 dcp->height = cpu_to_le32(5); 3993 dcp->budmin = L2BPERDMAP + L2LPERCTL * level; 3994 3995 /* 3996 * initialize the leaves of current level that were not covered 3997 * by the specified input block range (i.e. the leaves have no 3998 * low level dmapctl or dmap). 3999 */ 4000 cp = &dcp->stree[CTLLEAFIND + i]; 4001 for (; i < LPERCTL; i++) 4002 *cp++ = NOFREE; 4003 4004 /* build the dmap's binary buddy summary tree */ 4005 return (dbInitTree((struct dmaptree *) dcp)); 4006 } 4007 4008 4009 /* 4010 * NAME: dbGetL2AGSize()/ujfs_getagl2size() 4011 * 4012 * FUNCTION: Determine log2(allocation group size) from aggregate size 4013 * 4014 * PARAMETERS: 4015 * nblocks - Number of blocks in aggregate 4016 * 4017 * RETURNS: log2(allocation group size) in aggregate blocks 4018 */ 4019 static int dbGetL2AGSize(s64 nblocks) 4020 { 4021 s64 sz; 4022 s64 m; 4023 int l2sz; 4024 4025 if (nblocks < BPERDMAP * MAXAG) 4026 return (L2BPERDMAP); 4027 4028 /* round up aggregate size to power of 2 */ 4029 m = ((u64) 1 << (64 - 1)); 4030 for (l2sz = 64; l2sz >= 0; l2sz--, m >>= 1) { 4031 if (m & nblocks) 4032 break; 4033 } 4034 4035 sz = (s64) 1 << l2sz; 4036 if (sz < nblocks) 4037 l2sz += 1; 4038 4039 /* agsize = roundupSize/max_number_of_ag */ 4040 return (l2sz - L2MAXAG); 4041 } 4042 4043 4044 /* 4045 * NAME: dbMapFileSizeToMapSize() 4046 * 4047 * FUNCTION: compute number of blocks the block allocation map file 4048 * can cover from the map file size; 4049 * 4050 * RETURNS: Number of blocks which can be covered by this block map file; 4051 */ 4052 4053 /* 4054 * maximum number of map pages at each level including control pages 4055 */ 4056 #define MAXL0PAGES (1 + LPERCTL) 4057 #define MAXL1PAGES (1 + LPERCTL * MAXL0PAGES) 4058 #define MAXL2PAGES (1 + LPERCTL * MAXL1PAGES) 4059 4060 /* 4061 * convert number of map pages to the zero origin top dmapctl level 4062 */ 4063 #define BMAPPGTOLEV(npages) \ 4064 (((npages) <= 3 + MAXL0PAGES) ? 0 : \ 4065 ((npages) <= 2 + MAXL1PAGES) ? 1 : 2) 4066 4067 s64 dbMapFileSizeToMapSize(struct inode * ipbmap) 4068 { 4069 struct super_block *sb = ipbmap->i_sb; 4070 s64 nblocks; 4071 s64 npages, ndmaps; 4072 int level, i; 4073 int complete, factor; 4074 4075 nblocks = ipbmap->i_size >> JFS_SBI(sb)->l2bsize; 4076 npages = nblocks >> JFS_SBI(sb)->l2nbperpage; 4077 level = BMAPPGTOLEV(npages); 4078 4079 /* At each level, accumulate the number of dmap pages covered by 4080 * the number of full child levels below it; 4081 * repeat for the last incomplete child level. 4082 */ 4083 ndmaps = 0; 4084 npages--; /* skip the first global control page */ 4085 /* skip higher level control pages above top level covered by map */ 4086 npages -= (2 - level); 4087 npages--; /* skip top level's control page */ 4088 for (i = level; i >= 0; i--) { 4089 factor = 4090 (i == 2) ? MAXL1PAGES : ((i == 1) ? MAXL0PAGES : 1); 4091 complete = (u32) npages / factor; 4092 ndmaps += complete * ((i == 2) ? LPERCTL * LPERCTL : 4093 ((i == 1) ? LPERCTL : 1)); 4094 4095 /* pages in last/incomplete child */ 4096 npages = (u32) npages % factor; 4097 /* skip incomplete child's level control page */ 4098 npages--; 4099 } 4100 4101 /* convert the number of dmaps into the number of blocks 4102 * which can be covered by the dmaps; 4103 */ 4104 nblocks = ndmaps << L2BPERDMAP; 4105 4106 return (nblocks); 4107 } 4108