1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 * Copyright (C) International Business Machines Corp., 2000-2004 4 * Portions Copyright (C) Tino Reichardt, 2012 5 */ 6 7 #include <linux/fs.h> 8 #include <linux/slab.h> 9 #include "jfs_incore.h" 10 #include "jfs_superblock.h" 11 #include "jfs_dmap.h" 12 #include "jfs_imap.h" 13 #include "jfs_lock.h" 14 #include "jfs_metapage.h" 15 #include "jfs_debug.h" 16 #include "jfs_discard.h" 17 18 /* 19 * SERIALIZATION of the Block Allocation Map. 20 * 21 * the working state of the block allocation map is accessed in 22 * two directions: 23 * 24 * 1) allocation and free requests that start at the dmap 25 * level and move up through the dmap control pages (i.e. 26 * the vast majority of requests). 27 * 28 * 2) allocation requests that start at dmap control page 29 * level and work down towards the dmaps. 30 * 31 * the serialization scheme used here is as follows. 32 * 33 * requests which start at the bottom are serialized against each 34 * other through buffers and each requests holds onto its buffers 35 * as it works it way up from a single dmap to the required level 36 * of dmap control page. 37 * requests that start at the top are serialized against each other 38 * and request that start from the bottom by the multiple read/single 39 * write inode lock of the bmap inode. requests starting at the top 40 * take this lock in write mode while request starting at the bottom 41 * take the lock in read mode. a single top-down request may proceed 42 * exclusively while multiple bottoms-up requests may proceed 43 * simultaneously (under the protection of busy buffers). 44 * 45 * in addition to information found in dmaps and dmap control pages, 46 * the working state of the block allocation map also includes read/ 47 * write information maintained in the bmap descriptor (i.e. total 48 * free block count, allocation group level free block counts). 49 * a single exclusive lock (BMAP_LOCK) is used to guard this information 50 * in the face of multiple-bottoms up requests. 51 * (lock ordering: IREAD_LOCK, BMAP_LOCK); 52 * 53 * accesses to the persistent state of the block allocation map (limited 54 * to the persistent bitmaps in dmaps) is guarded by (busy) buffers. 55 */ 56 57 #define BMAP_LOCK_INIT(bmp) mutex_init(&bmp->db_bmaplock) 58 #define BMAP_LOCK(bmp) mutex_lock(&bmp->db_bmaplock) 59 #define BMAP_UNLOCK(bmp) mutex_unlock(&bmp->db_bmaplock) 60 61 /* 62 * forward references 63 */ 64 static void dbAllocBits(struct bmap * bmp, struct dmap * dp, s64 blkno, 65 int nblocks); 66 static void dbSplit(dmtree_t * tp, int leafno, int splitsz, int newval); 67 static int dbBackSplit(dmtree_t * tp, int leafno); 68 static int dbJoin(dmtree_t * tp, int leafno, int newval); 69 static void dbAdjTree(dmtree_t * tp, int leafno, int newval); 70 static int dbAdjCtl(struct bmap * bmp, s64 blkno, int newval, int alloc, 71 int level); 72 static int dbAllocAny(struct bmap * bmp, s64 nblocks, int l2nb, s64 * results); 73 static int dbAllocNext(struct bmap * bmp, struct dmap * dp, s64 blkno, 74 int nblocks); 75 static int dbAllocNear(struct bmap * bmp, struct dmap * dp, s64 blkno, 76 int nblocks, 77 int l2nb, s64 * results); 78 static int dbAllocDmap(struct bmap * bmp, struct dmap * dp, s64 blkno, 79 int nblocks); 80 static int dbAllocDmapLev(struct bmap * bmp, struct dmap * dp, int nblocks, 81 int l2nb, 82 s64 * results); 83 static int dbAllocAG(struct bmap * bmp, int agno, s64 nblocks, int l2nb, 84 s64 * results); 85 static int dbAllocCtl(struct bmap * bmp, s64 nblocks, int l2nb, s64 blkno, 86 s64 * results); 87 static int dbExtend(struct inode *ip, s64 blkno, s64 nblocks, s64 addnblocks); 88 static int dbFindBits(u32 word, int l2nb); 89 static int dbFindCtl(struct bmap * bmp, int l2nb, int level, s64 * blkno); 90 static int dbFindLeaf(dmtree_t * tp, int l2nb, int *leafidx); 91 static int dbFreeBits(struct bmap * bmp, struct dmap * dp, s64 blkno, 92 int nblocks); 93 static int dbFreeDmap(struct bmap * bmp, struct dmap * dp, s64 blkno, 94 int nblocks); 95 static int dbMaxBud(u8 * cp); 96 static int blkstol2(s64 nb); 97 98 static int cntlz(u32 value); 99 static int cnttz(u32 word); 100 101 static int dbAllocDmapBU(struct bmap * bmp, struct dmap * dp, s64 blkno, 102 int nblocks); 103 static int dbInitDmap(struct dmap * dp, s64 blkno, int nblocks); 104 static int dbInitDmapTree(struct dmap * dp); 105 static int dbInitTree(struct dmaptree * dtp); 106 static int dbInitDmapCtl(struct dmapctl * dcp, int level, int i); 107 static int dbGetL2AGSize(s64 nblocks); 108 109 /* 110 * buddy table 111 * 112 * table used for determining buddy sizes within characters of 113 * dmap bitmap words. the characters themselves serve as indexes 114 * into the table, with the table elements yielding the maximum 115 * binary buddy of free bits within the character. 116 */ 117 static const s8 budtab[256] = { 118 3, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 119 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 120 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 121 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 122 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 123 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 124 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 125 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 126 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 127 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 128 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 129 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 130 2, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 131 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 132 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 133 2, 1, 1, 1, 1, 0, 0, 0, 1, 0, 0, 0, 1, 0, 0, -1 134 }; 135 136 /* 137 * NAME: dbMount() 138 * 139 * FUNCTION: initializate the block allocation map. 140 * 141 * memory is allocated for the in-core bmap descriptor and 142 * the in-core descriptor is initialized from disk. 143 * 144 * PARAMETERS: 145 * ipbmap - pointer to in-core inode for the block map. 146 * 147 * RETURN VALUES: 148 * 0 - success 149 * -ENOMEM - insufficient memory 150 * -EIO - i/o error 151 * -EINVAL - wrong bmap data 152 */ 153 int dbMount(struct inode *ipbmap) 154 { 155 struct bmap *bmp; 156 struct dbmap_disk *dbmp_le; 157 struct metapage *mp; 158 int i, err; 159 160 /* 161 * allocate/initialize the in-memory bmap descriptor 162 */ 163 /* allocate memory for the in-memory bmap descriptor */ 164 bmp = kmalloc(sizeof(struct bmap), GFP_KERNEL); 165 if (bmp == NULL) 166 return -ENOMEM; 167 168 /* read the on-disk bmap descriptor. */ 169 mp = read_metapage(ipbmap, 170 BMAPBLKNO << JFS_SBI(ipbmap->i_sb)->l2nbperpage, 171 PSIZE, 0); 172 if (mp == NULL) { 173 err = -EIO; 174 goto err_kfree_bmp; 175 } 176 177 /* copy the on-disk bmap descriptor to its in-memory version. */ 178 dbmp_le = (struct dbmap_disk *) mp->data; 179 bmp->db_mapsize = le64_to_cpu(dbmp_le->dn_mapsize); 180 bmp->db_nfree = le64_to_cpu(dbmp_le->dn_nfree); 181 bmp->db_l2nbperpage = le32_to_cpu(dbmp_le->dn_l2nbperpage); 182 bmp->db_numag = le32_to_cpu(dbmp_le->dn_numag); 183 if (!bmp->db_numag) { 184 err = -EINVAL; 185 goto err_release_metapage; 186 } 187 188 bmp->db_maxlevel = le32_to_cpu(dbmp_le->dn_maxlevel); 189 bmp->db_maxag = le32_to_cpu(dbmp_le->dn_maxag); 190 bmp->db_agpref = le32_to_cpu(dbmp_le->dn_agpref); 191 bmp->db_aglevel = le32_to_cpu(dbmp_le->dn_aglevel); 192 bmp->db_agheight = le32_to_cpu(dbmp_le->dn_agheight); 193 bmp->db_agwidth = le32_to_cpu(dbmp_le->dn_agwidth); 194 bmp->db_agstart = le32_to_cpu(dbmp_le->dn_agstart); 195 bmp->db_agl2size = le32_to_cpu(dbmp_le->dn_agl2size); 196 if (bmp->db_agl2size > L2MAXL2SIZE - L2MAXAG) { 197 err = -EINVAL; 198 goto err_release_metapage; 199 } 200 201 if (((bmp->db_mapsize - 1) >> bmp->db_agl2size) > MAXAG) { 202 err = -EINVAL; 203 goto err_release_metapage; 204 } 205 206 for (i = 0; i < MAXAG; i++) 207 bmp->db_agfree[i] = le64_to_cpu(dbmp_le->dn_agfree[i]); 208 bmp->db_agsize = le64_to_cpu(dbmp_le->dn_agsize); 209 bmp->db_maxfreebud = dbmp_le->dn_maxfreebud; 210 211 /* release the buffer. */ 212 release_metapage(mp); 213 214 /* bind the bmap inode and the bmap descriptor to each other. */ 215 bmp->db_ipbmap = ipbmap; 216 JFS_SBI(ipbmap->i_sb)->bmap = bmp; 217 218 memset(bmp->db_active, 0, sizeof(bmp->db_active)); 219 220 /* 221 * allocate/initialize the bmap lock 222 */ 223 BMAP_LOCK_INIT(bmp); 224 225 return (0); 226 227 err_release_metapage: 228 release_metapage(mp); 229 err_kfree_bmp: 230 kfree(bmp); 231 return err; 232 } 233 234 235 /* 236 * NAME: dbUnmount() 237 * 238 * FUNCTION: terminate the block allocation map in preparation for 239 * file system unmount. 240 * 241 * the in-core bmap descriptor is written to disk and 242 * the memory for this descriptor is freed. 243 * 244 * PARAMETERS: 245 * ipbmap - pointer to in-core inode for the block map. 246 * 247 * RETURN VALUES: 248 * 0 - success 249 * -EIO - i/o error 250 */ 251 int dbUnmount(struct inode *ipbmap, int mounterror) 252 { 253 struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap; 254 255 if (!(mounterror || isReadOnly(ipbmap))) 256 dbSync(ipbmap); 257 258 /* 259 * Invalidate the page cache buffers 260 */ 261 truncate_inode_pages(ipbmap->i_mapping, 0); 262 263 /* free the memory for the in-memory bmap. */ 264 kfree(bmp); 265 266 return (0); 267 } 268 269 /* 270 * dbSync() 271 */ 272 int dbSync(struct inode *ipbmap) 273 { 274 struct dbmap_disk *dbmp_le; 275 struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap; 276 struct metapage *mp; 277 int i; 278 279 /* 280 * write bmap global control page 281 */ 282 /* get the buffer for the on-disk bmap descriptor. */ 283 mp = read_metapage(ipbmap, 284 BMAPBLKNO << JFS_SBI(ipbmap->i_sb)->l2nbperpage, 285 PSIZE, 0); 286 if (mp == NULL) { 287 jfs_err("dbSync: read_metapage failed!"); 288 return -EIO; 289 } 290 /* copy the in-memory version of the bmap to the on-disk version */ 291 dbmp_le = (struct dbmap_disk *) mp->data; 292 dbmp_le->dn_mapsize = cpu_to_le64(bmp->db_mapsize); 293 dbmp_le->dn_nfree = cpu_to_le64(bmp->db_nfree); 294 dbmp_le->dn_l2nbperpage = cpu_to_le32(bmp->db_l2nbperpage); 295 dbmp_le->dn_numag = cpu_to_le32(bmp->db_numag); 296 dbmp_le->dn_maxlevel = cpu_to_le32(bmp->db_maxlevel); 297 dbmp_le->dn_maxag = cpu_to_le32(bmp->db_maxag); 298 dbmp_le->dn_agpref = cpu_to_le32(bmp->db_agpref); 299 dbmp_le->dn_aglevel = cpu_to_le32(bmp->db_aglevel); 300 dbmp_le->dn_agheight = cpu_to_le32(bmp->db_agheight); 301 dbmp_le->dn_agwidth = cpu_to_le32(bmp->db_agwidth); 302 dbmp_le->dn_agstart = cpu_to_le32(bmp->db_agstart); 303 dbmp_le->dn_agl2size = cpu_to_le32(bmp->db_agl2size); 304 for (i = 0; i < MAXAG; i++) 305 dbmp_le->dn_agfree[i] = cpu_to_le64(bmp->db_agfree[i]); 306 dbmp_le->dn_agsize = cpu_to_le64(bmp->db_agsize); 307 dbmp_le->dn_maxfreebud = bmp->db_maxfreebud; 308 309 /* write the buffer */ 310 write_metapage(mp); 311 312 /* 313 * write out dirty pages of bmap 314 */ 315 filemap_write_and_wait(ipbmap->i_mapping); 316 317 diWriteSpecial(ipbmap, 0); 318 319 return (0); 320 } 321 322 /* 323 * NAME: dbFree() 324 * 325 * FUNCTION: free the specified block range from the working block 326 * allocation map. 327 * 328 * the blocks will be free from the working map one dmap 329 * at a time. 330 * 331 * PARAMETERS: 332 * ip - pointer to in-core inode; 333 * blkno - starting block number to be freed. 334 * nblocks - number of blocks to be freed. 335 * 336 * RETURN VALUES: 337 * 0 - success 338 * -EIO - i/o error 339 */ 340 int dbFree(struct inode *ip, s64 blkno, s64 nblocks) 341 { 342 struct metapage *mp; 343 struct dmap *dp; 344 int nb, rc; 345 s64 lblkno, rem; 346 struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap; 347 struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap; 348 struct super_block *sb = ipbmap->i_sb; 349 350 IREAD_LOCK(ipbmap, RDWRLOCK_DMAP); 351 352 /* block to be freed better be within the mapsize. */ 353 if (unlikely((blkno == 0) || (blkno + nblocks > bmp->db_mapsize))) { 354 IREAD_UNLOCK(ipbmap); 355 printk(KERN_ERR "blkno = %Lx, nblocks = %Lx\n", 356 (unsigned long long) blkno, 357 (unsigned long long) nblocks); 358 jfs_error(ip->i_sb, "block to be freed is outside the map\n"); 359 return -EIO; 360 } 361 362 /** 363 * TRIM the blocks, when mounted with discard option 364 */ 365 if (JFS_SBI(sb)->flag & JFS_DISCARD) 366 if (JFS_SBI(sb)->minblks_trim <= nblocks) 367 jfs_issue_discard(ipbmap, blkno, nblocks); 368 369 /* 370 * free the blocks a dmap at a time. 371 */ 372 mp = NULL; 373 for (rem = nblocks; rem > 0; rem -= nb, blkno += nb) { 374 /* release previous dmap if any */ 375 if (mp) { 376 write_metapage(mp); 377 } 378 379 /* get the buffer for the current dmap. */ 380 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage); 381 mp = read_metapage(ipbmap, lblkno, PSIZE, 0); 382 if (mp == NULL) { 383 IREAD_UNLOCK(ipbmap); 384 return -EIO; 385 } 386 dp = (struct dmap *) mp->data; 387 388 /* determine the number of blocks to be freed from 389 * this dmap. 390 */ 391 nb = min(rem, BPERDMAP - (blkno & (BPERDMAP - 1))); 392 393 /* free the blocks. */ 394 if ((rc = dbFreeDmap(bmp, dp, blkno, nb))) { 395 jfs_error(ip->i_sb, "error in block map\n"); 396 release_metapage(mp); 397 IREAD_UNLOCK(ipbmap); 398 return (rc); 399 } 400 } 401 402 /* write the last buffer. */ 403 if (mp) 404 write_metapage(mp); 405 406 IREAD_UNLOCK(ipbmap); 407 408 return (0); 409 } 410 411 412 /* 413 * NAME: dbUpdatePMap() 414 * 415 * FUNCTION: update the allocation state (free or allocate) of the 416 * specified block range in the persistent block allocation map. 417 * 418 * the blocks will be updated in the persistent map one 419 * dmap at a time. 420 * 421 * PARAMETERS: 422 * ipbmap - pointer to in-core inode for the block map. 423 * free - 'true' if block range is to be freed from the persistent 424 * map; 'false' if it is to be allocated. 425 * blkno - starting block number of the range. 426 * nblocks - number of contiguous blocks in the range. 427 * tblk - transaction block; 428 * 429 * RETURN VALUES: 430 * 0 - success 431 * -EIO - i/o error 432 */ 433 int 434 dbUpdatePMap(struct inode *ipbmap, 435 int free, s64 blkno, s64 nblocks, struct tblock * tblk) 436 { 437 int nblks, dbitno, wbitno, rbits; 438 int word, nbits, nwords; 439 struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap; 440 s64 lblkno, rem, lastlblkno; 441 u32 mask; 442 struct dmap *dp; 443 struct metapage *mp; 444 struct jfs_log *log; 445 int lsn, difft, diffp; 446 unsigned long flags; 447 448 /* the blocks better be within the mapsize. */ 449 if (blkno + nblocks > bmp->db_mapsize) { 450 printk(KERN_ERR "blkno = %Lx, nblocks = %Lx\n", 451 (unsigned long long) blkno, 452 (unsigned long long) nblocks); 453 jfs_error(ipbmap->i_sb, "blocks are outside the map\n"); 454 return -EIO; 455 } 456 457 /* compute delta of transaction lsn from log syncpt */ 458 lsn = tblk->lsn; 459 log = (struct jfs_log *) JFS_SBI(tblk->sb)->log; 460 logdiff(difft, lsn, log); 461 462 /* 463 * update the block state a dmap at a time. 464 */ 465 mp = NULL; 466 lastlblkno = 0; 467 for (rem = nblocks; rem > 0; rem -= nblks, blkno += nblks) { 468 /* get the buffer for the current dmap. */ 469 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage); 470 if (lblkno != lastlblkno) { 471 if (mp) { 472 write_metapage(mp); 473 } 474 475 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 476 0); 477 if (mp == NULL) 478 return -EIO; 479 metapage_wait_for_io(mp); 480 } 481 dp = (struct dmap *) mp->data; 482 483 /* determine the bit number and word within the dmap of 484 * the starting block. also determine how many blocks 485 * are to be updated within this dmap. 486 */ 487 dbitno = blkno & (BPERDMAP - 1); 488 word = dbitno >> L2DBWORD; 489 nblks = min(rem, (s64)BPERDMAP - dbitno); 490 491 /* update the bits of the dmap words. the first and last 492 * words may only have a subset of their bits updated. if 493 * this is the case, we'll work against that word (i.e. 494 * partial first and/or last) only in a single pass. a 495 * single pass will also be used to update all words that 496 * are to have all their bits updated. 497 */ 498 for (rbits = nblks; rbits > 0; 499 rbits -= nbits, dbitno += nbits) { 500 /* determine the bit number within the word and 501 * the number of bits within the word. 502 */ 503 wbitno = dbitno & (DBWORD - 1); 504 nbits = min(rbits, DBWORD - wbitno); 505 506 /* check if only part of the word is to be updated. */ 507 if (nbits < DBWORD) { 508 /* update (free or allocate) the bits 509 * in this word. 510 */ 511 mask = 512 (ONES << (DBWORD - nbits) >> wbitno); 513 if (free) 514 dp->pmap[word] &= 515 cpu_to_le32(~mask); 516 else 517 dp->pmap[word] |= 518 cpu_to_le32(mask); 519 520 word += 1; 521 } else { 522 /* one or more words are to have all 523 * their bits updated. determine how 524 * many words and how many bits. 525 */ 526 nwords = rbits >> L2DBWORD; 527 nbits = nwords << L2DBWORD; 528 529 /* update (free or allocate) the bits 530 * in these words. 531 */ 532 if (free) 533 memset(&dp->pmap[word], 0, 534 nwords * 4); 535 else 536 memset(&dp->pmap[word], (int) ONES, 537 nwords * 4); 538 539 word += nwords; 540 } 541 } 542 543 /* 544 * update dmap lsn 545 */ 546 if (lblkno == lastlblkno) 547 continue; 548 549 lastlblkno = lblkno; 550 551 LOGSYNC_LOCK(log, flags); 552 if (mp->lsn != 0) { 553 /* inherit older/smaller lsn */ 554 logdiff(diffp, mp->lsn, log); 555 if (difft < diffp) { 556 mp->lsn = lsn; 557 558 /* move bp after tblock in logsync list */ 559 list_move(&mp->synclist, &tblk->synclist); 560 } 561 562 /* inherit younger/larger clsn */ 563 logdiff(difft, tblk->clsn, log); 564 logdiff(diffp, mp->clsn, log); 565 if (difft > diffp) 566 mp->clsn = tblk->clsn; 567 } else { 568 mp->log = log; 569 mp->lsn = lsn; 570 571 /* insert bp after tblock in logsync list */ 572 log->count++; 573 list_add(&mp->synclist, &tblk->synclist); 574 575 mp->clsn = tblk->clsn; 576 } 577 LOGSYNC_UNLOCK(log, flags); 578 } 579 580 /* write the last buffer. */ 581 if (mp) { 582 write_metapage(mp); 583 } 584 585 return (0); 586 } 587 588 589 /* 590 * NAME: dbNextAG() 591 * 592 * FUNCTION: find the preferred allocation group for new allocations. 593 * 594 * Within the allocation groups, we maintain a preferred 595 * allocation group which consists of a group with at least 596 * average free space. It is the preferred group that we target 597 * new inode allocation towards. The tie-in between inode 598 * allocation and block allocation occurs as we allocate the 599 * first (data) block of an inode and specify the inode (block) 600 * as the allocation hint for this block. 601 * 602 * We try to avoid having more than one open file growing in 603 * an allocation group, as this will lead to fragmentation. 604 * This differs from the old OS/2 method of trying to keep 605 * empty ags around for large allocations. 606 * 607 * PARAMETERS: 608 * ipbmap - pointer to in-core inode for the block map. 609 * 610 * RETURN VALUES: 611 * the preferred allocation group number. 612 */ 613 int dbNextAG(struct inode *ipbmap) 614 { 615 s64 avgfree; 616 int agpref; 617 s64 hwm = 0; 618 int i; 619 int next_best = -1; 620 struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap; 621 622 BMAP_LOCK(bmp); 623 624 /* determine the average number of free blocks within the ags. */ 625 avgfree = (u32)bmp->db_nfree / bmp->db_numag; 626 627 /* 628 * if the current preferred ag does not have an active allocator 629 * and has at least average freespace, return it 630 */ 631 agpref = bmp->db_agpref; 632 if ((atomic_read(&bmp->db_active[agpref]) == 0) && 633 (bmp->db_agfree[agpref] >= avgfree)) 634 goto unlock; 635 636 /* From the last preferred ag, find the next one with at least 637 * average free space. 638 */ 639 for (i = 0 ; i < bmp->db_numag; i++, agpref++) { 640 if (agpref == bmp->db_numag) 641 agpref = 0; 642 643 if (atomic_read(&bmp->db_active[agpref])) 644 /* open file is currently growing in this ag */ 645 continue; 646 if (bmp->db_agfree[agpref] >= avgfree) { 647 /* Return this one */ 648 bmp->db_agpref = agpref; 649 goto unlock; 650 } else if (bmp->db_agfree[agpref] > hwm) { 651 /* Less than avg. freespace, but best so far */ 652 hwm = bmp->db_agfree[agpref]; 653 next_best = agpref; 654 } 655 } 656 657 /* 658 * If no inactive ag was found with average freespace, use the 659 * next best 660 */ 661 if (next_best != -1) 662 bmp->db_agpref = next_best; 663 /* else leave db_agpref unchanged */ 664 unlock: 665 BMAP_UNLOCK(bmp); 666 667 /* return the preferred group. 668 */ 669 return (bmp->db_agpref); 670 } 671 672 /* 673 * NAME: dbAlloc() 674 * 675 * FUNCTION: attempt to allocate a specified number of contiguous free 676 * blocks from the working allocation block map. 677 * 678 * the block allocation policy uses hints and a multi-step 679 * approach. 680 * 681 * for allocation requests smaller than the number of blocks 682 * per dmap, we first try to allocate the new blocks 683 * immediately following the hint. if these blocks are not 684 * available, we try to allocate blocks near the hint. if 685 * no blocks near the hint are available, we next try to 686 * allocate within the same dmap as contains the hint. 687 * 688 * if no blocks are available in the dmap or the allocation 689 * request is larger than the dmap size, we try to allocate 690 * within the same allocation group as contains the hint. if 691 * this does not succeed, we finally try to allocate anywhere 692 * within the aggregate. 693 * 694 * we also try to allocate anywhere within the aggregate 695 * for allocation requests larger than the allocation group 696 * size or requests that specify no hint value. 697 * 698 * PARAMETERS: 699 * ip - pointer to in-core inode; 700 * hint - allocation hint. 701 * nblocks - number of contiguous blocks in the range. 702 * results - on successful return, set to the starting block number 703 * of the newly allocated contiguous range. 704 * 705 * RETURN VALUES: 706 * 0 - success 707 * -ENOSPC - insufficient disk resources 708 * -EIO - i/o error 709 */ 710 int dbAlloc(struct inode *ip, s64 hint, s64 nblocks, s64 * results) 711 { 712 int rc, agno; 713 struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap; 714 struct bmap *bmp; 715 struct metapage *mp; 716 s64 lblkno, blkno; 717 struct dmap *dp; 718 int l2nb; 719 s64 mapSize; 720 int writers; 721 722 /* assert that nblocks is valid */ 723 assert(nblocks > 0); 724 725 /* get the log2 number of blocks to be allocated. 726 * if the number of blocks is not a log2 multiple, 727 * it will be rounded up to the next log2 multiple. 728 */ 729 l2nb = BLKSTOL2(nblocks); 730 731 bmp = JFS_SBI(ip->i_sb)->bmap; 732 733 mapSize = bmp->db_mapsize; 734 735 /* the hint should be within the map */ 736 if (hint >= mapSize) { 737 jfs_error(ip->i_sb, "the hint is outside the map\n"); 738 return -EIO; 739 } 740 741 /* if the number of blocks to be allocated is greater than the 742 * allocation group size, try to allocate anywhere. 743 */ 744 if (l2nb > bmp->db_agl2size) { 745 IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP); 746 747 rc = dbAllocAny(bmp, nblocks, l2nb, results); 748 749 goto write_unlock; 750 } 751 752 /* 753 * If no hint, let dbNextAG recommend an allocation group 754 */ 755 if (hint == 0) 756 goto pref_ag; 757 758 /* we would like to allocate close to the hint. adjust the 759 * hint to the block following the hint since the allocators 760 * will start looking for free space starting at this point. 761 */ 762 blkno = hint + 1; 763 764 if (blkno >= bmp->db_mapsize) 765 goto pref_ag; 766 767 agno = blkno >> bmp->db_agl2size; 768 769 /* check if blkno crosses over into a new allocation group. 770 * if so, check if we should allow allocations within this 771 * allocation group. 772 */ 773 if ((blkno & (bmp->db_agsize - 1)) == 0) 774 /* check if the AG is currently being written to. 775 * if so, call dbNextAG() to find a non-busy 776 * AG with sufficient free space. 777 */ 778 if (atomic_read(&bmp->db_active[agno])) 779 goto pref_ag; 780 781 /* check if the allocation request size can be satisfied from a 782 * single dmap. if so, try to allocate from the dmap containing 783 * the hint using a tiered strategy. 784 */ 785 if (nblocks <= BPERDMAP) { 786 IREAD_LOCK(ipbmap, RDWRLOCK_DMAP); 787 788 /* get the buffer for the dmap containing the hint. 789 */ 790 rc = -EIO; 791 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage); 792 mp = read_metapage(ipbmap, lblkno, PSIZE, 0); 793 if (mp == NULL) 794 goto read_unlock; 795 796 dp = (struct dmap *) mp->data; 797 798 /* first, try to satisfy the allocation request with the 799 * blocks beginning at the hint. 800 */ 801 if ((rc = dbAllocNext(bmp, dp, blkno, (int) nblocks)) 802 != -ENOSPC) { 803 if (rc == 0) { 804 *results = blkno; 805 mark_metapage_dirty(mp); 806 } 807 808 release_metapage(mp); 809 goto read_unlock; 810 } 811 812 writers = atomic_read(&bmp->db_active[agno]); 813 if ((writers > 1) || 814 ((writers == 1) && (JFS_IP(ip)->active_ag != agno))) { 815 /* 816 * Someone else is writing in this allocation 817 * group. To avoid fragmenting, try another ag 818 */ 819 release_metapage(mp); 820 IREAD_UNLOCK(ipbmap); 821 goto pref_ag; 822 } 823 824 /* next, try to satisfy the allocation request with blocks 825 * near the hint. 826 */ 827 if ((rc = 828 dbAllocNear(bmp, dp, blkno, (int) nblocks, l2nb, results)) 829 != -ENOSPC) { 830 if (rc == 0) 831 mark_metapage_dirty(mp); 832 833 release_metapage(mp); 834 goto read_unlock; 835 } 836 837 /* try to satisfy the allocation request with blocks within 838 * the same dmap as the hint. 839 */ 840 if ((rc = dbAllocDmapLev(bmp, dp, (int) nblocks, l2nb, results)) 841 != -ENOSPC) { 842 if (rc == 0) 843 mark_metapage_dirty(mp); 844 845 release_metapage(mp); 846 goto read_unlock; 847 } 848 849 release_metapage(mp); 850 IREAD_UNLOCK(ipbmap); 851 } 852 853 /* try to satisfy the allocation request with blocks within 854 * the same allocation group as the hint. 855 */ 856 IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP); 857 if ((rc = dbAllocAG(bmp, agno, nblocks, l2nb, results)) != -ENOSPC) 858 goto write_unlock; 859 860 IWRITE_UNLOCK(ipbmap); 861 862 863 pref_ag: 864 /* 865 * Let dbNextAG recommend a preferred allocation group 866 */ 867 agno = dbNextAG(ipbmap); 868 IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP); 869 870 /* Try to allocate within this allocation group. if that fails, try to 871 * allocate anywhere in the map. 872 */ 873 if ((rc = dbAllocAG(bmp, agno, nblocks, l2nb, results)) == -ENOSPC) 874 rc = dbAllocAny(bmp, nblocks, l2nb, results); 875 876 write_unlock: 877 IWRITE_UNLOCK(ipbmap); 878 879 return (rc); 880 881 read_unlock: 882 IREAD_UNLOCK(ipbmap); 883 884 return (rc); 885 } 886 887 /* 888 * NAME: dbReAlloc() 889 * 890 * FUNCTION: attempt to extend a current allocation by a specified 891 * number of blocks. 892 * 893 * this routine attempts to satisfy the allocation request 894 * by first trying to extend the existing allocation in 895 * place by allocating the additional blocks as the blocks 896 * immediately following the current allocation. if these 897 * blocks are not available, this routine will attempt to 898 * allocate a new set of contiguous blocks large enough 899 * to cover the existing allocation plus the additional 900 * number of blocks required. 901 * 902 * PARAMETERS: 903 * ip - pointer to in-core inode requiring allocation. 904 * blkno - starting block of the current allocation. 905 * nblocks - number of contiguous blocks within the current 906 * allocation. 907 * addnblocks - number of blocks to add to the allocation. 908 * results - on successful return, set to the starting block number 909 * of the existing allocation if the existing allocation 910 * was extended in place or to a newly allocated contiguous 911 * range if the existing allocation could not be extended 912 * in place. 913 * 914 * RETURN VALUES: 915 * 0 - success 916 * -ENOSPC - insufficient disk resources 917 * -EIO - i/o error 918 */ 919 int 920 dbReAlloc(struct inode *ip, 921 s64 blkno, s64 nblocks, s64 addnblocks, s64 * results) 922 { 923 int rc; 924 925 /* try to extend the allocation in place. 926 */ 927 if ((rc = dbExtend(ip, blkno, nblocks, addnblocks)) == 0) { 928 *results = blkno; 929 return (0); 930 } else { 931 if (rc != -ENOSPC) 932 return (rc); 933 } 934 935 /* could not extend the allocation in place, so allocate a 936 * new set of blocks for the entire request (i.e. try to get 937 * a range of contiguous blocks large enough to cover the 938 * existing allocation plus the additional blocks.) 939 */ 940 return (dbAlloc 941 (ip, blkno + nblocks - 1, addnblocks + nblocks, results)); 942 } 943 944 945 /* 946 * NAME: dbExtend() 947 * 948 * FUNCTION: attempt to extend a current allocation by a specified 949 * number of blocks. 950 * 951 * this routine attempts to satisfy the allocation request 952 * by first trying to extend the existing allocation in 953 * place by allocating the additional blocks as the blocks 954 * immediately following the current allocation. 955 * 956 * PARAMETERS: 957 * ip - pointer to in-core inode requiring allocation. 958 * blkno - starting block of the current allocation. 959 * nblocks - number of contiguous blocks within the current 960 * allocation. 961 * addnblocks - number of blocks to add to the allocation. 962 * 963 * RETURN VALUES: 964 * 0 - success 965 * -ENOSPC - insufficient disk resources 966 * -EIO - i/o error 967 */ 968 static int dbExtend(struct inode *ip, s64 blkno, s64 nblocks, s64 addnblocks) 969 { 970 struct jfs_sb_info *sbi = JFS_SBI(ip->i_sb); 971 s64 lblkno, lastblkno, extblkno; 972 uint rel_block; 973 struct metapage *mp; 974 struct dmap *dp; 975 int rc; 976 struct inode *ipbmap = sbi->ipbmap; 977 struct bmap *bmp; 978 979 /* 980 * We don't want a non-aligned extent to cross a page boundary 981 */ 982 if (((rel_block = blkno & (sbi->nbperpage - 1))) && 983 (rel_block + nblocks + addnblocks > sbi->nbperpage)) 984 return -ENOSPC; 985 986 /* get the last block of the current allocation */ 987 lastblkno = blkno + nblocks - 1; 988 989 /* determine the block number of the block following 990 * the existing allocation. 991 */ 992 extblkno = lastblkno + 1; 993 994 IREAD_LOCK(ipbmap, RDWRLOCK_DMAP); 995 996 /* better be within the file system */ 997 bmp = sbi->bmap; 998 if (lastblkno < 0 || lastblkno >= bmp->db_mapsize) { 999 IREAD_UNLOCK(ipbmap); 1000 jfs_error(ip->i_sb, "the block is outside the filesystem\n"); 1001 return -EIO; 1002 } 1003 1004 /* we'll attempt to extend the current allocation in place by 1005 * allocating the additional blocks as the blocks immediately 1006 * following the current allocation. we only try to extend the 1007 * current allocation in place if the number of additional blocks 1008 * can fit into a dmap, the last block of the current allocation 1009 * is not the last block of the file system, and the start of the 1010 * inplace extension is not on an allocation group boundary. 1011 */ 1012 if (addnblocks > BPERDMAP || extblkno >= bmp->db_mapsize || 1013 (extblkno & (bmp->db_agsize - 1)) == 0) { 1014 IREAD_UNLOCK(ipbmap); 1015 return -ENOSPC; 1016 } 1017 1018 /* get the buffer for the dmap containing the first block 1019 * of the extension. 1020 */ 1021 lblkno = BLKTODMAP(extblkno, bmp->db_l2nbperpage); 1022 mp = read_metapage(ipbmap, lblkno, PSIZE, 0); 1023 if (mp == NULL) { 1024 IREAD_UNLOCK(ipbmap); 1025 return -EIO; 1026 } 1027 1028 dp = (struct dmap *) mp->data; 1029 1030 /* try to allocate the blocks immediately following the 1031 * current allocation. 1032 */ 1033 rc = dbAllocNext(bmp, dp, extblkno, (int) addnblocks); 1034 1035 IREAD_UNLOCK(ipbmap); 1036 1037 /* were we successful ? */ 1038 if (rc == 0) 1039 write_metapage(mp); 1040 else 1041 /* we were not successful */ 1042 release_metapage(mp); 1043 1044 return (rc); 1045 } 1046 1047 1048 /* 1049 * NAME: dbAllocNext() 1050 * 1051 * FUNCTION: attempt to allocate the blocks of the specified block 1052 * range within a dmap. 1053 * 1054 * PARAMETERS: 1055 * bmp - pointer to bmap descriptor 1056 * dp - pointer to dmap. 1057 * blkno - starting block number of the range. 1058 * nblocks - number of contiguous free blocks of the range. 1059 * 1060 * RETURN VALUES: 1061 * 0 - success 1062 * -ENOSPC - insufficient disk resources 1063 * -EIO - i/o error 1064 * 1065 * serialization: IREAD_LOCK(ipbmap) held on entry/exit; 1066 */ 1067 static int dbAllocNext(struct bmap * bmp, struct dmap * dp, s64 blkno, 1068 int nblocks) 1069 { 1070 int dbitno, word, rembits, nb, nwords, wbitno, nw; 1071 int l2size; 1072 s8 *leaf; 1073 u32 mask; 1074 1075 if (dp->tree.leafidx != cpu_to_le32(LEAFIND)) { 1076 jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmap page\n"); 1077 return -EIO; 1078 } 1079 1080 /* pick up a pointer to the leaves of the dmap tree. 1081 */ 1082 leaf = dp->tree.stree + le32_to_cpu(dp->tree.leafidx); 1083 1084 /* determine the bit number and word within the dmap of the 1085 * starting block. 1086 */ 1087 dbitno = blkno & (BPERDMAP - 1); 1088 word = dbitno >> L2DBWORD; 1089 1090 /* check if the specified block range is contained within 1091 * this dmap. 1092 */ 1093 if (dbitno + nblocks > BPERDMAP) 1094 return -ENOSPC; 1095 1096 /* check if the starting leaf indicates that anything 1097 * is free. 1098 */ 1099 if (leaf[word] == NOFREE) 1100 return -ENOSPC; 1101 1102 /* check the dmaps words corresponding to block range to see 1103 * if the block range is free. not all bits of the first and 1104 * last words may be contained within the block range. if this 1105 * is the case, we'll work against those words (i.e. partial first 1106 * and/or last) on an individual basis (a single pass) and examine 1107 * the actual bits to determine if they are free. a single pass 1108 * will be used for all dmap words fully contained within the 1109 * specified range. within this pass, the leaves of the dmap 1110 * tree will be examined to determine if the blocks are free. a 1111 * single leaf may describe the free space of multiple dmap 1112 * words, so we may visit only a subset of the actual leaves 1113 * corresponding to the dmap words of the block range. 1114 */ 1115 for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) { 1116 /* determine the bit number within the word and 1117 * the number of bits within the word. 1118 */ 1119 wbitno = dbitno & (DBWORD - 1); 1120 nb = min(rembits, DBWORD - wbitno); 1121 1122 /* check if only part of the word is to be examined. 1123 */ 1124 if (nb < DBWORD) { 1125 /* check if the bits are free. 1126 */ 1127 mask = (ONES << (DBWORD - nb) >> wbitno); 1128 if ((mask & ~le32_to_cpu(dp->wmap[word])) != mask) 1129 return -ENOSPC; 1130 1131 word += 1; 1132 } else { 1133 /* one or more dmap words are fully contained 1134 * within the block range. determine how many 1135 * words and how many bits. 1136 */ 1137 nwords = rembits >> L2DBWORD; 1138 nb = nwords << L2DBWORD; 1139 1140 /* now examine the appropriate leaves to determine 1141 * if the blocks are free. 1142 */ 1143 while (nwords > 0) { 1144 /* does the leaf describe any free space ? 1145 */ 1146 if (leaf[word] < BUDMIN) 1147 return -ENOSPC; 1148 1149 /* determine the l2 number of bits provided 1150 * by this leaf. 1151 */ 1152 l2size = 1153 min_t(int, leaf[word], NLSTOL2BSZ(nwords)); 1154 1155 /* determine how many words were handled. 1156 */ 1157 nw = BUDSIZE(l2size, BUDMIN); 1158 1159 nwords -= nw; 1160 word += nw; 1161 } 1162 } 1163 } 1164 1165 /* allocate the blocks. 1166 */ 1167 return (dbAllocDmap(bmp, dp, blkno, nblocks)); 1168 } 1169 1170 1171 /* 1172 * NAME: dbAllocNear() 1173 * 1174 * FUNCTION: attempt to allocate a number of contiguous free blocks near 1175 * a specified block (hint) within a dmap. 1176 * 1177 * starting with the dmap leaf that covers the hint, we'll 1178 * check the next four contiguous leaves for sufficient free 1179 * space. if sufficient free space is found, we'll allocate 1180 * the desired free space. 1181 * 1182 * PARAMETERS: 1183 * bmp - pointer to bmap descriptor 1184 * dp - pointer to dmap. 1185 * blkno - block number to allocate near. 1186 * nblocks - actual number of contiguous free blocks desired. 1187 * l2nb - log2 number of contiguous free blocks desired. 1188 * results - on successful return, set to the starting block number 1189 * of the newly allocated range. 1190 * 1191 * RETURN VALUES: 1192 * 0 - success 1193 * -ENOSPC - insufficient disk resources 1194 * -EIO - i/o error 1195 * 1196 * serialization: IREAD_LOCK(ipbmap) held on entry/exit; 1197 */ 1198 static int 1199 dbAllocNear(struct bmap * bmp, 1200 struct dmap * dp, s64 blkno, int nblocks, int l2nb, s64 * results) 1201 { 1202 int word, lword, rc; 1203 s8 *leaf; 1204 1205 if (dp->tree.leafidx != cpu_to_le32(LEAFIND)) { 1206 jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmap page\n"); 1207 return -EIO; 1208 } 1209 1210 leaf = dp->tree.stree + le32_to_cpu(dp->tree.leafidx); 1211 1212 /* determine the word within the dmap that holds the hint 1213 * (i.e. blkno). also, determine the last word in the dmap 1214 * that we'll include in our examination. 1215 */ 1216 word = (blkno & (BPERDMAP - 1)) >> L2DBWORD; 1217 lword = min(word + 4, LPERDMAP); 1218 1219 /* examine the leaves for sufficient free space. 1220 */ 1221 for (; word < lword; word++) { 1222 /* does the leaf describe sufficient free space ? 1223 */ 1224 if (leaf[word] < l2nb) 1225 continue; 1226 1227 /* determine the block number within the file system 1228 * of the first block described by this dmap word. 1229 */ 1230 blkno = le64_to_cpu(dp->start) + (word << L2DBWORD); 1231 1232 /* if not all bits of the dmap word are free, get the 1233 * starting bit number within the dmap word of the required 1234 * string of free bits and adjust the block number with the 1235 * value. 1236 */ 1237 if (leaf[word] < BUDMIN) 1238 blkno += 1239 dbFindBits(le32_to_cpu(dp->wmap[word]), l2nb); 1240 1241 /* allocate the blocks. 1242 */ 1243 if ((rc = dbAllocDmap(bmp, dp, blkno, nblocks)) == 0) 1244 *results = blkno; 1245 1246 return (rc); 1247 } 1248 1249 return -ENOSPC; 1250 } 1251 1252 1253 /* 1254 * NAME: dbAllocAG() 1255 * 1256 * FUNCTION: attempt to allocate the specified number of contiguous 1257 * free blocks within the specified allocation group. 1258 * 1259 * unless the allocation group size is equal to the number 1260 * of blocks per dmap, the dmap control pages will be used to 1261 * find the required free space, if available. we start the 1262 * search at the highest dmap control page level which 1263 * distinctly describes the allocation group's free space 1264 * (i.e. the highest level at which the allocation group's 1265 * free space is not mixed in with that of any other group). 1266 * in addition, we start the search within this level at a 1267 * height of the dmapctl dmtree at which the nodes distinctly 1268 * describe the allocation group's free space. at this height, 1269 * the allocation group's free space may be represented by 1 1270 * or two sub-trees, depending on the allocation group size. 1271 * we search the top nodes of these subtrees left to right for 1272 * sufficient free space. if sufficient free space is found, 1273 * the subtree is searched to find the leftmost leaf that 1274 * has free space. once we have made it to the leaf, we 1275 * move the search to the next lower level dmap control page 1276 * corresponding to this leaf. we continue down the dmap control 1277 * pages until we find the dmap that contains or starts the 1278 * sufficient free space and we allocate at this dmap. 1279 * 1280 * if the allocation group size is equal to the dmap size, 1281 * we'll start at the dmap corresponding to the allocation 1282 * group and attempt the allocation at this level. 1283 * 1284 * the dmap control page search is also not performed if the 1285 * allocation group is completely free and we go to the first 1286 * dmap of the allocation group to do the allocation. this is 1287 * done because the allocation group may be part (not the first 1288 * part) of a larger binary buddy system, causing the dmap 1289 * control pages to indicate no free space (NOFREE) within 1290 * the allocation group. 1291 * 1292 * PARAMETERS: 1293 * bmp - pointer to bmap descriptor 1294 * agno - allocation group number. 1295 * nblocks - actual number of contiguous free blocks desired. 1296 * l2nb - log2 number of contiguous free blocks desired. 1297 * results - on successful return, set to the starting block number 1298 * of the newly allocated range. 1299 * 1300 * RETURN VALUES: 1301 * 0 - success 1302 * -ENOSPC - insufficient disk resources 1303 * -EIO - i/o error 1304 * 1305 * note: IWRITE_LOCK(ipmap) held on entry/exit; 1306 */ 1307 static int 1308 dbAllocAG(struct bmap * bmp, int agno, s64 nblocks, int l2nb, s64 * results) 1309 { 1310 struct metapage *mp; 1311 struct dmapctl *dcp; 1312 int rc, ti, i, k, m, n, agperlev; 1313 s64 blkno, lblkno; 1314 int budmin; 1315 1316 /* allocation request should not be for more than the 1317 * allocation group size. 1318 */ 1319 if (l2nb > bmp->db_agl2size) { 1320 jfs_error(bmp->db_ipbmap->i_sb, 1321 "allocation request is larger than the allocation group size\n"); 1322 return -EIO; 1323 } 1324 1325 /* determine the starting block number of the allocation 1326 * group. 1327 */ 1328 blkno = (s64) agno << bmp->db_agl2size; 1329 1330 /* check if the allocation group size is the minimum allocation 1331 * group size or if the allocation group is completely free. if 1332 * the allocation group size is the minimum size of BPERDMAP (i.e. 1333 * 1 dmap), there is no need to search the dmap control page (below) 1334 * that fully describes the allocation group since the allocation 1335 * group is already fully described by a dmap. in this case, we 1336 * just call dbAllocCtl() to search the dmap tree and allocate the 1337 * required space if available. 1338 * 1339 * if the allocation group is completely free, dbAllocCtl() is 1340 * also called to allocate the required space. this is done for 1341 * two reasons. first, it makes no sense searching the dmap control 1342 * pages for free space when we know that free space exists. second, 1343 * the dmap control pages may indicate that the allocation group 1344 * has no free space if the allocation group is part (not the first 1345 * part) of a larger binary buddy system. 1346 */ 1347 if (bmp->db_agsize == BPERDMAP 1348 || bmp->db_agfree[agno] == bmp->db_agsize) { 1349 rc = dbAllocCtl(bmp, nblocks, l2nb, blkno, results); 1350 if ((rc == -ENOSPC) && 1351 (bmp->db_agfree[agno] == bmp->db_agsize)) { 1352 printk(KERN_ERR "blkno = %Lx, blocks = %Lx\n", 1353 (unsigned long long) blkno, 1354 (unsigned long long) nblocks); 1355 jfs_error(bmp->db_ipbmap->i_sb, 1356 "dbAllocCtl failed in free AG\n"); 1357 } 1358 return (rc); 1359 } 1360 1361 /* the buffer for the dmap control page that fully describes the 1362 * allocation group. 1363 */ 1364 lblkno = BLKTOCTL(blkno, bmp->db_l2nbperpage, bmp->db_aglevel); 1365 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0); 1366 if (mp == NULL) 1367 return -EIO; 1368 dcp = (struct dmapctl *) mp->data; 1369 budmin = dcp->budmin; 1370 1371 if (dcp->leafidx != cpu_to_le32(CTLLEAFIND)) { 1372 jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmapctl page\n"); 1373 release_metapage(mp); 1374 return -EIO; 1375 } 1376 1377 /* search the subtree(s) of the dmap control page that describes 1378 * the allocation group, looking for sufficient free space. to begin, 1379 * determine how many allocation groups are represented in a dmap 1380 * control page at the control page level (i.e. L0, L1, L2) that 1381 * fully describes an allocation group. next, determine the starting 1382 * tree index of this allocation group within the control page. 1383 */ 1384 agperlev = 1385 (1 << (L2LPERCTL - (bmp->db_agheight << 1))) / bmp->db_agwidth; 1386 ti = bmp->db_agstart + bmp->db_agwidth * (agno & (agperlev - 1)); 1387 1388 /* dmap control page trees fan-out by 4 and a single allocation 1389 * group may be described by 1 or 2 subtrees within the ag level 1390 * dmap control page, depending upon the ag size. examine the ag's 1391 * subtrees for sufficient free space, starting with the leftmost 1392 * subtree. 1393 */ 1394 for (i = 0; i < bmp->db_agwidth; i++, ti++) { 1395 /* is there sufficient free space ? 1396 */ 1397 if (l2nb > dcp->stree[ti]) 1398 continue; 1399 1400 /* sufficient free space found in a subtree. now search down 1401 * the subtree to find the leftmost leaf that describes this 1402 * free space. 1403 */ 1404 for (k = bmp->db_agheight; k > 0; k--) { 1405 for (n = 0, m = (ti << 2) + 1; n < 4; n++) { 1406 if (l2nb <= dcp->stree[m + n]) { 1407 ti = m + n; 1408 break; 1409 } 1410 } 1411 if (n == 4) { 1412 jfs_error(bmp->db_ipbmap->i_sb, 1413 "failed descending stree\n"); 1414 release_metapage(mp); 1415 return -EIO; 1416 } 1417 } 1418 1419 /* determine the block number within the file system 1420 * that corresponds to this leaf. 1421 */ 1422 if (bmp->db_aglevel == 2) 1423 blkno = 0; 1424 else if (bmp->db_aglevel == 1) 1425 blkno &= ~(MAXL1SIZE - 1); 1426 else /* bmp->db_aglevel == 0 */ 1427 blkno &= ~(MAXL0SIZE - 1); 1428 1429 blkno += 1430 ((s64) (ti - le32_to_cpu(dcp->leafidx))) << budmin; 1431 1432 /* release the buffer in preparation for going down 1433 * the next level of dmap control pages. 1434 */ 1435 release_metapage(mp); 1436 1437 /* check if we need to continue to search down the lower 1438 * level dmap control pages. we need to if the number of 1439 * blocks required is less than maximum number of blocks 1440 * described at the next lower level. 1441 */ 1442 if (l2nb < budmin) { 1443 1444 /* search the lower level dmap control pages to get 1445 * the starting block number of the dmap that 1446 * contains or starts off the free space. 1447 */ 1448 if ((rc = 1449 dbFindCtl(bmp, l2nb, bmp->db_aglevel - 1, 1450 &blkno))) { 1451 if (rc == -ENOSPC) { 1452 jfs_error(bmp->db_ipbmap->i_sb, 1453 "control page inconsistent\n"); 1454 return -EIO; 1455 } 1456 return (rc); 1457 } 1458 } 1459 1460 /* allocate the blocks. 1461 */ 1462 rc = dbAllocCtl(bmp, nblocks, l2nb, blkno, results); 1463 if (rc == -ENOSPC) { 1464 jfs_error(bmp->db_ipbmap->i_sb, 1465 "unable to allocate blocks\n"); 1466 rc = -EIO; 1467 } 1468 return (rc); 1469 } 1470 1471 /* no space in the allocation group. release the buffer and 1472 * return -ENOSPC. 1473 */ 1474 release_metapage(mp); 1475 1476 return -ENOSPC; 1477 } 1478 1479 1480 /* 1481 * NAME: dbAllocAny() 1482 * 1483 * FUNCTION: attempt to allocate the specified number of contiguous 1484 * free blocks anywhere in the file system. 1485 * 1486 * dbAllocAny() attempts to find the sufficient free space by 1487 * searching down the dmap control pages, starting with the 1488 * highest level (i.e. L0, L1, L2) control page. if free space 1489 * large enough to satisfy the desired free space is found, the 1490 * desired free space is allocated. 1491 * 1492 * PARAMETERS: 1493 * bmp - pointer to bmap descriptor 1494 * nblocks - actual number of contiguous free blocks desired. 1495 * l2nb - log2 number of contiguous free blocks desired. 1496 * results - on successful return, set to the starting block number 1497 * of the newly allocated range. 1498 * 1499 * RETURN VALUES: 1500 * 0 - success 1501 * -ENOSPC - insufficient disk resources 1502 * -EIO - i/o error 1503 * 1504 * serialization: IWRITE_LOCK(ipbmap) held on entry/exit; 1505 */ 1506 static int dbAllocAny(struct bmap * bmp, s64 nblocks, int l2nb, s64 * results) 1507 { 1508 int rc; 1509 s64 blkno = 0; 1510 1511 /* starting with the top level dmap control page, search 1512 * down the dmap control levels for sufficient free space. 1513 * if free space is found, dbFindCtl() returns the starting 1514 * block number of the dmap that contains or starts off the 1515 * range of free space. 1516 */ 1517 if ((rc = dbFindCtl(bmp, l2nb, bmp->db_maxlevel, &blkno))) 1518 return (rc); 1519 1520 /* allocate the blocks. 1521 */ 1522 rc = dbAllocCtl(bmp, nblocks, l2nb, blkno, results); 1523 if (rc == -ENOSPC) { 1524 jfs_error(bmp->db_ipbmap->i_sb, "unable to allocate blocks\n"); 1525 return -EIO; 1526 } 1527 return (rc); 1528 } 1529 1530 1531 /* 1532 * NAME: dbDiscardAG() 1533 * 1534 * FUNCTION: attempt to discard (TRIM) all free blocks of specific AG 1535 * 1536 * algorithm: 1537 * 1) allocate blocks, as large as possible and save them 1538 * while holding IWRITE_LOCK on ipbmap 1539 * 2) trim all these saved block/length values 1540 * 3) mark the blocks free again 1541 * 1542 * benefit: 1543 * - we work only on one ag at some time, minimizing how long we 1544 * need to lock ipbmap 1545 * - reading / writing the fs is possible most time, even on 1546 * trimming 1547 * 1548 * downside: 1549 * - we write two times to the dmapctl and dmap pages 1550 * - but for me, this seems the best way, better ideas? 1551 * /TR 2012 1552 * 1553 * PARAMETERS: 1554 * ip - pointer to in-core inode 1555 * agno - ag to trim 1556 * minlen - minimum value of contiguous blocks 1557 * 1558 * RETURN VALUES: 1559 * s64 - actual number of blocks trimmed 1560 */ 1561 s64 dbDiscardAG(struct inode *ip, int agno, s64 minlen) 1562 { 1563 struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap; 1564 struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap; 1565 s64 nblocks, blkno; 1566 u64 trimmed = 0; 1567 int rc, l2nb; 1568 struct super_block *sb = ipbmap->i_sb; 1569 1570 struct range2trim { 1571 u64 blkno; 1572 u64 nblocks; 1573 } *totrim, *tt; 1574 1575 /* max blkno / nblocks pairs to trim */ 1576 int count = 0, range_cnt; 1577 u64 max_ranges; 1578 1579 /* prevent others from writing new stuff here, while trimming */ 1580 IWRITE_LOCK(ipbmap, RDWRLOCK_DMAP); 1581 1582 nblocks = bmp->db_agfree[agno]; 1583 max_ranges = nblocks; 1584 do_div(max_ranges, minlen); 1585 range_cnt = min_t(u64, max_ranges + 1, 32 * 1024); 1586 totrim = kmalloc_array(range_cnt, sizeof(struct range2trim), GFP_NOFS); 1587 if (totrim == NULL) { 1588 jfs_error(bmp->db_ipbmap->i_sb, "no memory for trim array\n"); 1589 IWRITE_UNLOCK(ipbmap); 1590 return 0; 1591 } 1592 1593 tt = totrim; 1594 while (nblocks >= minlen) { 1595 l2nb = BLKSTOL2(nblocks); 1596 1597 /* 0 = okay, -EIO = fatal, -ENOSPC -> try smaller block */ 1598 rc = dbAllocAG(bmp, agno, nblocks, l2nb, &blkno); 1599 if (rc == 0) { 1600 tt->blkno = blkno; 1601 tt->nblocks = nblocks; 1602 tt++; count++; 1603 1604 /* the whole ag is free, trim now */ 1605 if (bmp->db_agfree[agno] == 0) 1606 break; 1607 1608 /* give a hint for the next while */ 1609 nblocks = bmp->db_agfree[agno]; 1610 continue; 1611 } else if (rc == -ENOSPC) { 1612 /* search for next smaller log2 block */ 1613 l2nb = BLKSTOL2(nblocks) - 1; 1614 nblocks = 1LL << l2nb; 1615 } else { 1616 /* Trim any already allocated blocks */ 1617 jfs_error(bmp->db_ipbmap->i_sb, "-EIO\n"); 1618 break; 1619 } 1620 1621 /* check, if our trim array is full */ 1622 if (unlikely(count >= range_cnt - 1)) 1623 break; 1624 } 1625 IWRITE_UNLOCK(ipbmap); 1626 1627 tt->nblocks = 0; /* mark the current end */ 1628 for (tt = totrim; tt->nblocks != 0; tt++) { 1629 /* when mounted with online discard, dbFree() will 1630 * call jfs_issue_discard() itself */ 1631 if (!(JFS_SBI(sb)->flag & JFS_DISCARD)) 1632 jfs_issue_discard(ip, tt->blkno, tt->nblocks); 1633 dbFree(ip, tt->blkno, tt->nblocks); 1634 trimmed += tt->nblocks; 1635 } 1636 kfree(totrim); 1637 1638 return trimmed; 1639 } 1640 1641 /* 1642 * NAME: dbFindCtl() 1643 * 1644 * FUNCTION: starting at a specified dmap control page level and block 1645 * number, search down the dmap control levels for a range of 1646 * contiguous free blocks large enough to satisfy an allocation 1647 * request for the specified number of free blocks. 1648 * 1649 * if sufficient contiguous free blocks are found, this routine 1650 * returns the starting block number within a dmap page that 1651 * contains or starts a range of contiqious free blocks that 1652 * is sufficient in size. 1653 * 1654 * PARAMETERS: 1655 * bmp - pointer to bmap descriptor 1656 * level - starting dmap control page level. 1657 * l2nb - log2 number of contiguous free blocks desired. 1658 * *blkno - on entry, starting block number for conducting the search. 1659 * on successful return, the first block within a dmap page 1660 * that contains or starts a range of contiguous free blocks. 1661 * 1662 * RETURN VALUES: 1663 * 0 - success 1664 * -ENOSPC - insufficient disk resources 1665 * -EIO - i/o error 1666 * 1667 * serialization: IWRITE_LOCK(ipbmap) held on entry/exit; 1668 */ 1669 static int dbFindCtl(struct bmap * bmp, int l2nb, int level, s64 * blkno) 1670 { 1671 int rc, leafidx, lev; 1672 s64 b, lblkno; 1673 struct dmapctl *dcp; 1674 int budmin; 1675 struct metapage *mp; 1676 1677 /* starting at the specified dmap control page level and block 1678 * number, search down the dmap control levels for the starting 1679 * block number of a dmap page that contains or starts off 1680 * sufficient free blocks. 1681 */ 1682 for (lev = level, b = *blkno; lev >= 0; lev--) { 1683 /* get the buffer of the dmap control page for the block 1684 * number and level (i.e. L0, L1, L2). 1685 */ 1686 lblkno = BLKTOCTL(b, bmp->db_l2nbperpage, lev); 1687 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0); 1688 if (mp == NULL) 1689 return -EIO; 1690 dcp = (struct dmapctl *) mp->data; 1691 budmin = dcp->budmin; 1692 1693 if (dcp->leafidx != cpu_to_le32(CTLLEAFIND)) { 1694 jfs_error(bmp->db_ipbmap->i_sb, 1695 "Corrupt dmapctl page\n"); 1696 release_metapage(mp); 1697 return -EIO; 1698 } 1699 1700 /* search the tree within the dmap control page for 1701 * sufficient free space. if sufficient free space is found, 1702 * dbFindLeaf() returns the index of the leaf at which 1703 * free space was found. 1704 */ 1705 rc = dbFindLeaf((dmtree_t *) dcp, l2nb, &leafidx); 1706 1707 /* release the buffer. 1708 */ 1709 release_metapage(mp); 1710 1711 /* space found ? 1712 */ 1713 if (rc) { 1714 if (lev != level) { 1715 jfs_error(bmp->db_ipbmap->i_sb, 1716 "dmap inconsistent\n"); 1717 return -EIO; 1718 } 1719 return -ENOSPC; 1720 } 1721 1722 /* adjust the block number to reflect the location within 1723 * the dmap control page (i.e. the leaf) at which free 1724 * space was found. 1725 */ 1726 b += (((s64) leafidx) << budmin); 1727 1728 /* we stop the search at this dmap control page level if 1729 * the number of blocks required is greater than or equal 1730 * to the maximum number of blocks described at the next 1731 * (lower) level. 1732 */ 1733 if (l2nb >= budmin) 1734 break; 1735 } 1736 1737 *blkno = b; 1738 return (0); 1739 } 1740 1741 1742 /* 1743 * NAME: dbAllocCtl() 1744 * 1745 * FUNCTION: attempt to allocate a specified number of contiguous 1746 * blocks starting within a specific dmap. 1747 * 1748 * this routine is called by higher level routines that search 1749 * the dmap control pages above the actual dmaps for contiguous 1750 * free space. the result of successful searches by these 1751 * routines are the starting block numbers within dmaps, with 1752 * the dmaps themselves containing the desired contiguous free 1753 * space or starting a contiguous free space of desired size 1754 * that is made up of the blocks of one or more dmaps. these 1755 * calls should not fail due to insufficent resources. 1756 * 1757 * this routine is called in some cases where it is not known 1758 * whether it will fail due to insufficient resources. more 1759 * specifically, this occurs when allocating from an allocation 1760 * group whose size is equal to the number of blocks per dmap. 1761 * in this case, the dmap control pages are not examined prior 1762 * to calling this routine (to save pathlength) and the call 1763 * might fail. 1764 * 1765 * for a request size that fits within a dmap, this routine relies 1766 * upon the dmap's dmtree to find the requested contiguous free 1767 * space. for request sizes that are larger than a dmap, the 1768 * requested free space will start at the first block of the 1769 * first dmap (i.e. blkno). 1770 * 1771 * PARAMETERS: 1772 * bmp - pointer to bmap descriptor 1773 * nblocks - actual number of contiguous free blocks to allocate. 1774 * l2nb - log2 number of contiguous free blocks to allocate. 1775 * blkno - starting block number of the dmap to start the allocation 1776 * from. 1777 * results - on successful return, set to the starting block number 1778 * of the newly allocated range. 1779 * 1780 * RETURN VALUES: 1781 * 0 - success 1782 * -ENOSPC - insufficient disk resources 1783 * -EIO - i/o error 1784 * 1785 * serialization: IWRITE_LOCK(ipbmap) held on entry/exit; 1786 */ 1787 static int 1788 dbAllocCtl(struct bmap * bmp, s64 nblocks, int l2nb, s64 blkno, s64 * results) 1789 { 1790 int rc, nb; 1791 s64 b, lblkno, n; 1792 struct metapage *mp; 1793 struct dmap *dp; 1794 1795 /* check if the allocation request is confined to a single dmap. 1796 */ 1797 if (l2nb <= L2BPERDMAP) { 1798 /* get the buffer for the dmap. 1799 */ 1800 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage); 1801 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0); 1802 if (mp == NULL) 1803 return -EIO; 1804 dp = (struct dmap *) mp->data; 1805 1806 /* try to allocate the blocks. 1807 */ 1808 rc = dbAllocDmapLev(bmp, dp, (int) nblocks, l2nb, results); 1809 if (rc == 0) 1810 mark_metapage_dirty(mp); 1811 1812 release_metapage(mp); 1813 1814 return (rc); 1815 } 1816 1817 /* allocation request involving multiple dmaps. it must start on 1818 * a dmap boundary. 1819 */ 1820 assert((blkno & (BPERDMAP - 1)) == 0); 1821 1822 /* allocate the blocks dmap by dmap. 1823 */ 1824 for (n = nblocks, b = blkno; n > 0; n -= nb, b += nb) { 1825 /* get the buffer for the dmap. 1826 */ 1827 lblkno = BLKTODMAP(b, bmp->db_l2nbperpage); 1828 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0); 1829 if (mp == NULL) { 1830 rc = -EIO; 1831 goto backout; 1832 } 1833 dp = (struct dmap *) mp->data; 1834 1835 /* the dmap better be all free. 1836 */ 1837 if (dp->tree.stree[ROOT] != L2BPERDMAP) { 1838 release_metapage(mp); 1839 jfs_error(bmp->db_ipbmap->i_sb, 1840 "the dmap is not all free\n"); 1841 rc = -EIO; 1842 goto backout; 1843 } 1844 1845 /* determine how many blocks to allocate from this dmap. 1846 */ 1847 nb = min_t(s64, n, BPERDMAP); 1848 1849 /* allocate the blocks from the dmap. 1850 */ 1851 if ((rc = dbAllocDmap(bmp, dp, b, nb))) { 1852 release_metapage(mp); 1853 goto backout; 1854 } 1855 1856 /* write the buffer. 1857 */ 1858 write_metapage(mp); 1859 } 1860 1861 /* set the results (starting block number) and return. 1862 */ 1863 *results = blkno; 1864 return (0); 1865 1866 /* something failed in handling an allocation request involving 1867 * multiple dmaps. we'll try to clean up by backing out any 1868 * allocation that has already happened for this request. if 1869 * we fail in backing out the allocation, we'll mark the file 1870 * system to indicate that blocks have been leaked. 1871 */ 1872 backout: 1873 1874 /* try to backout the allocations dmap by dmap. 1875 */ 1876 for (n = nblocks - n, b = blkno; n > 0; 1877 n -= BPERDMAP, b += BPERDMAP) { 1878 /* get the buffer for this dmap. 1879 */ 1880 lblkno = BLKTODMAP(b, bmp->db_l2nbperpage); 1881 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0); 1882 if (mp == NULL) { 1883 /* could not back out. mark the file system 1884 * to indicate that we have leaked blocks. 1885 */ 1886 jfs_error(bmp->db_ipbmap->i_sb, 1887 "I/O Error: Block Leakage\n"); 1888 continue; 1889 } 1890 dp = (struct dmap *) mp->data; 1891 1892 /* free the blocks is this dmap. 1893 */ 1894 if (dbFreeDmap(bmp, dp, b, BPERDMAP)) { 1895 /* could not back out. mark the file system 1896 * to indicate that we have leaked blocks. 1897 */ 1898 release_metapage(mp); 1899 jfs_error(bmp->db_ipbmap->i_sb, "Block Leakage\n"); 1900 continue; 1901 } 1902 1903 /* write the buffer. 1904 */ 1905 write_metapage(mp); 1906 } 1907 1908 return (rc); 1909 } 1910 1911 1912 /* 1913 * NAME: dbAllocDmapLev() 1914 * 1915 * FUNCTION: attempt to allocate a specified number of contiguous blocks 1916 * from a specified dmap. 1917 * 1918 * this routine checks if the contiguous blocks are available. 1919 * if so, nblocks of blocks are allocated; otherwise, ENOSPC is 1920 * returned. 1921 * 1922 * PARAMETERS: 1923 * mp - pointer to bmap descriptor 1924 * dp - pointer to dmap to attempt to allocate blocks from. 1925 * l2nb - log2 number of contiguous block desired. 1926 * nblocks - actual number of contiguous block desired. 1927 * results - on successful return, set to the starting block number 1928 * of the newly allocated range. 1929 * 1930 * RETURN VALUES: 1931 * 0 - success 1932 * -ENOSPC - insufficient disk resources 1933 * -EIO - i/o error 1934 * 1935 * serialization: IREAD_LOCK(ipbmap), e.g., from dbAlloc(), or 1936 * IWRITE_LOCK(ipbmap), e.g., dbAllocCtl(), held on entry/exit; 1937 */ 1938 static int 1939 dbAllocDmapLev(struct bmap * bmp, 1940 struct dmap * dp, int nblocks, int l2nb, s64 * results) 1941 { 1942 s64 blkno; 1943 int leafidx, rc; 1944 1945 /* can't be more than a dmaps worth of blocks */ 1946 assert(l2nb <= L2BPERDMAP); 1947 1948 /* search the tree within the dmap page for sufficient 1949 * free space. if sufficient free space is found, dbFindLeaf() 1950 * returns the index of the leaf at which free space was found. 1951 */ 1952 if (dbFindLeaf((dmtree_t *) & dp->tree, l2nb, &leafidx)) 1953 return -ENOSPC; 1954 1955 /* determine the block number within the file system corresponding 1956 * to the leaf at which free space was found. 1957 */ 1958 blkno = le64_to_cpu(dp->start) + (leafidx << L2DBWORD); 1959 1960 /* if not all bits of the dmap word are free, get the starting 1961 * bit number within the dmap word of the required string of free 1962 * bits and adjust the block number with this value. 1963 */ 1964 if (dp->tree.stree[leafidx + LEAFIND] < BUDMIN) 1965 blkno += dbFindBits(le32_to_cpu(dp->wmap[leafidx]), l2nb); 1966 1967 /* allocate the blocks */ 1968 if ((rc = dbAllocDmap(bmp, dp, blkno, nblocks)) == 0) 1969 *results = blkno; 1970 1971 return (rc); 1972 } 1973 1974 1975 /* 1976 * NAME: dbAllocDmap() 1977 * 1978 * FUNCTION: adjust the disk allocation map to reflect the allocation 1979 * of a specified block range within a dmap. 1980 * 1981 * this routine allocates the specified blocks from the dmap 1982 * through a call to dbAllocBits(). if the allocation of the 1983 * block range causes the maximum string of free blocks within 1984 * the dmap to change (i.e. the value of the root of the dmap's 1985 * dmtree), this routine will cause this change to be reflected 1986 * up through the appropriate levels of the dmap control pages 1987 * by a call to dbAdjCtl() for the L0 dmap control page that 1988 * covers this dmap. 1989 * 1990 * PARAMETERS: 1991 * bmp - pointer to bmap descriptor 1992 * dp - pointer to dmap to allocate the block range from. 1993 * blkno - starting block number of the block to be allocated. 1994 * nblocks - number of blocks to be allocated. 1995 * 1996 * RETURN VALUES: 1997 * 0 - success 1998 * -EIO - i/o error 1999 * 2000 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit; 2001 */ 2002 static int dbAllocDmap(struct bmap * bmp, struct dmap * dp, s64 blkno, 2003 int nblocks) 2004 { 2005 s8 oldroot; 2006 int rc; 2007 2008 /* save the current value of the root (i.e. maximum free string) 2009 * of the dmap tree. 2010 */ 2011 oldroot = dp->tree.stree[ROOT]; 2012 2013 /* allocate the specified (blocks) bits */ 2014 dbAllocBits(bmp, dp, blkno, nblocks); 2015 2016 /* if the root has not changed, done. */ 2017 if (dp->tree.stree[ROOT] == oldroot) 2018 return (0); 2019 2020 /* root changed. bubble the change up to the dmap control pages. 2021 * if the adjustment of the upper level control pages fails, 2022 * backout the bit allocation (thus making everything consistent). 2023 */ 2024 if ((rc = dbAdjCtl(bmp, blkno, dp->tree.stree[ROOT], 1, 0))) 2025 dbFreeBits(bmp, dp, blkno, nblocks); 2026 2027 return (rc); 2028 } 2029 2030 2031 /* 2032 * NAME: dbFreeDmap() 2033 * 2034 * FUNCTION: adjust the disk allocation map to reflect the allocation 2035 * of a specified block range within a dmap. 2036 * 2037 * this routine frees the specified blocks from the dmap through 2038 * a call to dbFreeBits(). if the deallocation of the block range 2039 * causes the maximum string of free blocks within the dmap to 2040 * change (i.e. the value of the root of the dmap's dmtree), this 2041 * routine will cause this change to be reflected up through the 2042 * appropriate levels of the dmap control pages by a call to 2043 * dbAdjCtl() for the L0 dmap control page that covers this dmap. 2044 * 2045 * PARAMETERS: 2046 * bmp - pointer to bmap descriptor 2047 * dp - pointer to dmap to free the block range from. 2048 * blkno - starting block number of the block to be freed. 2049 * nblocks - number of blocks to be freed. 2050 * 2051 * RETURN VALUES: 2052 * 0 - success 2053 * -EIO - i/o error 2054 * 2055 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit; 2056 */ 2057 static int dbFreeDmap(struct bmap * bmp, struct dmap * dp, s64 blkno, 2058 int nblocks) 2059 { 2060 s8 oldroot; 2061 int rc = 0, word; 2062 2063 /* save the current value of the root (i.e. maximum free string) 2064 * of the dmap tree. 2065 */ 2066 oldroot = dp->tree.stree[ROOT]; 2067 2068 /* free the specified (blocks) bits */ 2069 rc = dbFreeBits(bmp, dp, blkno, nblocks); 2070 2071 /* if error or the root has not changed, done. */ 2072 if (rc || (dp->tree.stree[ROOT] == oldroot)) 2073 return (rc); 2074 2075 /* root changed. bubble the change up to the dmap control pages. 2076 * if the adjustment of the upper level control pages fails, 2077 * backout the deallocation. 2078 */ 2079 if ((rc = dbAdjCtl(bmp, blkno, dp->tree.stree[ROOT], 0, 0))) { 2080 word = (blkno & (BPERDMAP - 1)) >> L2DBWORD; 2081 2082 /* as part of backing out the deallocation, we will have 2083 * to back split the dmap tree if the deallocation caused 2084 * the freed blocks to become part of a larger binary buddy 2085 * system. 2086 */ 2087 if (dp->tree.stree[word] == NOFREE) 2088 dbBackSplit((dmtree_t *) & dp->tree, word); 2089 2090 dbAllocBits(bmp, dp, blkno, nblocks); 2091 } 2092 2093 return (rc); 2094 } 2095 2096 2097 /* 2098 * NAME: dbAllocBits() 2099 * 2100 * FUNCTION: allocate a specified block range from a dmap. 2101 * 2102 * this routine updates the dmap to reflect the working 2103 * state allocation of the specified block range. it directly 2104 * updates the bits of the working map and causes the adjustment 2105 * of the binary buddy system described by the dmap's dmtree 2106 * leaves to reflect the bits allocated. it also causes the 2107 * dmap's dmtree, as a whole, to reflect the allocated range. 2108 * 2109 * PARAMETERS: 2110 * bmp - pointer to bmap descriptor 2111 * dp - pointer to dmap to allocate bits from. 2112 * blkno - starting block number of the bits to be allocated. 2113 * nblocks - number of bits to be allocated. 2114 * 2115 * RETURN VALUES: none 2116 * 2117 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit; 2118 */ 2119 static void dbAllocBits(struct bmap * bmp, struct dmap * dp, s64 blkno, 2120 int nblocks) 2121 { 2122 int dbitno, word, rembits, nb, nwords, wbitno, nw, agno; 2123 dmtree_t *tp = (dmtree_t *) & dp->tree; 2124 int size; 2125 s8 *leaf; 2126 2127 /* pick up a pointer to the leaves of the dmap tree */ 2128 leaf = dp->tree.stree + LEAFIND; 2129 2130 /* determine the bit number and word within the dmap of the 2131 * starting block. 2132 */ 2133 dbitno = blkno & (BPERDMAP - 1); 2134 word = dbitno >> L2DBWORD; 2135 2136 /* block range better be within the dmap */ 2137 assert(dbitno + nblocks <= BPERDMAP); 2138 2139 /* allocate the bits of the dmap's words corresponding to the block 2140 * range. not all bits of the first and last words may be contained 2141 * within the block range. if this is the case, we'll work against 2142 * those words (i.e. partial first and/or last) on an individual basis 2143 * (a single pass), allocating the bits of interest by hand and 2144 * updating the leaf corresponding to the dmap word. a single pass 2145 * will be used for all dmap words fully contained within the 2146 * specified range. within this pass, the bits of all fully contained 2147 * dmap words will be marked as free in a single shot and the leaves 2148 * will be updated. a single leaf may describe the free space of 2149 * multiple dmap words, so we may update only a subset of the actual 2150 * leaves corresponding to the dmap words of the block range. 2151 */ 2152 for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) { 2153 /* determine the bit number within the word and 2154 * the number of bits within the word. 2155 */ 2156 wbitno = dbitno & (DBWORD - 1); 2157 nb = min(rembits, DBWORD - wbitno); 2158 2159 /* check if only part of a word is to be allocated. 2160 */ 2161 if (nb < DBWORD) { 2162 /* allocate (set to 1) the appropriate bits within 2163 * this dmap word. 2164 */ 2165 dp->wmap[word] |= cpu_to_le32(ONES << (DBWORD - nb) 2166 >> wbitno); 2167 2168 /* update the leaf for this dmap word. in addition 2169 * to setting the leaf value to the binary buddy max 2170 * of the updated dmap word, dbSplit() will split 2171 * the binary system of the leaves if need be. 2172 */ 2173 dbSplit(tp, word, BUDMIN, 2174 dbMaxBud((u8 *) & dp->wmap[word])); 2175 2176 word += 1; 2177 } else { 2178 /* one or more dmap words are fully contained 2179 * within the block range. determine how many 2180 * words and allocate (set to 1) the bits of these 2181 * words. 2182 */ 2183 nwords = rembits >> L2DBWORD; 2184 memset(&dp->wmap[word], (int) ONES, nwords * 4); 2185 2186 /* determine how many bits. 2187 */ 2188 nb = nwords << L2DBWORD; 2189 2190 /* now update the appropriate leaves to reflect 2191 * the allocated words. 2192 */ 2193 for (; nwords > 0; nwords -= nw) { 2194 if (leaf[word] < BUDMIN) { 2195 jfs_error(bmp->db_ipbmap->i_sb, 2196 "leaf page corrupt\n"); 2197 break; 2198 } 2199 2200 /* determine what the leaf value should be 2201 * updated to as the minimum of the l2 number 2202 * of bits being allocated and the l2 number 2203 * of bits currently described by this leaf. 2204 */ 2205 size = min_t(int, leaf[word], 2206 NLSTOL2BSZ(nwords)); 2207 2208 /* update the leaf to reflect the allocation. 2209 * in addition to setting the leaf value to 2210 * NOFREE, dbSplit() will split the binary 2211 * system of the leaves to reflect the current 2212 * allocation (size). 2213 */ 2214 dbSplit(tp, word, size, NOFREE); 2215 2216 /* get the number of dmap words handled */ 2217 nw = BUDSIZE(size, BUDMIN); 2218 word += nw; 2219 } 2220 } 2221 } 2222 2223 /* update the free count for this dmap */ 2224 le32_add_cpu(&dp->nfree, -nblocks); 2225 2226 BMAP_LOCK(bmp); 2227 2228 /* if this allocation group is completely free, 2229 * update the maximum allocation group number if this allocation 2230 * group is the new max. 2231 */ 2232 agno = blkno >> bmp->db_agl2size; 2233 if (agno > bmp->db_maxag) 2234 bmp->db_maxag = agno; 2235 2236 /* update the free count for the allocation group and map */ 2237 bmp->db_agfree[agno] -= nblocks; 2238 bmp->db_nfree -= nblocks; 2239 2240 BMAP_UNLOCK(bmp); 2241 } 2242 2243 2244 /* 2245 * NAME: dbFreeBits() 2246 * 2247 * FUNCTION: free a specified block range from a dmap. 2248 * 2249 * this routine updates the dmap to reflect the working 2250 * state allocation of the specified block range. it directly 2251 * updates the bits of the working map and causes the adjustment 2252 * of the binary buddy system described by the dmap's dmtree 2253 * leaves to reflect the bits freed. it also causes the dmap's 2254 * dmtree, as a whole, to reflect the deallocated range. 2255 * 2256 * PARAMETERS: 2257 * bmp - pointer to bmap descriptor 2258 * dp - pointer to dmap to free bits from. 2259 * blkno - starting block number of the bits to be freed. 2260 * nblocks - number of bits to be freed. 2261 * 2262 * RETURN VALUES: 0 for success 2263 * 2264 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit; 2265 */ 2266 static int dbFreeBits(struct bmap * bmp, struct dmap * dp, s64 blkno, 2267 int nblocks) 2268 { 2269 int dbitno, word, rembits, nb, nwords, wbitno, nw, agno; 2270 dmtree_t *tp = (dmtree_t *) & dp->tree; 2271 int rc = 0; 2272 int size; 2273 2274 /* determine the bit number and word within the dmap of the 2275 * starting block. 2276 */ 2277 dbitno = blkno & (BPERDMAP - 1); 2278 word = dbitno >> L2DBWORD; 2279 2280 /* block range better be within the dmap. 2281 */ 2282 assert(dbitno + nblocks <= BPERDMAP); 2283 2284 /* free the bits of the dmaps words corresponding to the block range. 2285 * not all bits of the first and last words may be contained within 2286 * the block range. if this is the case, we'll work against those 2287 * words (i.e. partial first and/or last) on an individual basis 2288 * (a single pass), freeing the bits of interest by hand and updating 2289 * the leaf corresponding to the dmap word. a single pass will be used 2290 * for all dmap words fully contained within the specified range. 2291 * within this pass, the bits of all fully contained dmap words will 2292 * be marked as free in a single shot and the leaves will be updated. a 2293 * single leaf may describe the free space of multiple dmap words, 2294 * so we may update only a subset of the actual leaves corresponding 2295 * to the dmap words of the block range. 2296 * 2297 * dbJoin() is used to update leaf values and will join the binary 2298 * buddy system of the leaves if the new leaf values indicate this 2299 * should be done. 2300 */ 2301 for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) { 2302 /* determine the bit number within the word and 2303 * the number of bits within the word. 2304 */ 2305 wbitno = dbitno & (DBWORD - 1); 2306 nb = min(rembits, DBWORD - wbitno); 2307 2308 /* check if only part of a word is to be freed. 2309 */ 2310 if (nb < DBWORD) { 2311 /* free (zero) the appropriate bits within this 2312 * dmap word. 2313 */ 2314 dp->wmap[word] &= 2315 cpu_to_le32(~(ONES << (DBWORD - nb) 2316 >> wbitno)); 2317 2318 /* update the leaf for this dmap word. 2319 */ 2320 rc = dbJoin(tp, word, 2321 dbMaxBud((u8 *) & dp->wmap[word])); 2322 if (rc) 2323 return rc; 2324 2325 word += 1; 2326 } else { 2327 /* one or more dmap words are fully contained 2328 * within the block range. determine how many 2329 * words and free (zero) the bits of these words. 2330 */ 2331 nwords = rembits >> L2DBWORD; 2332 memset(&dp->wmap[word], 0, nwords * 4); 2333 2334 /* determine how many bits. 2335 */ 2336 nb = nwords << L2DBWORD; 2337 2338 /* now update the appropriate leaves to reflect 2339 * the freed words. 2340 */ 2341 for (; nwords > 0; nwords -= nw) { 2342 /* determine what the leaf value should be 2343 * updated to as the minimum of the l2 number 2344 * of bits being freed and the l2 (max) number 2345 * of bits that can be described by this leaf. 2346 */ 2347 size = 2348 min(LITOL2BSZ 2349 (word, L2LPERDMAP, BUDMIN), 2350 NLSTOL2BSZ(nwords)); 2351 2352 /* update the leaf. 2353 */ 2354 rc = dbJoin(tp, word, size); 2355 if (rc) 2356 return rc; 2357 2358 /* get the number of dmap words handled. 2359 */ 2360 nw = BUDSIZE(size, BUDMIN); 2361 word += nw; 2362 } 2363 } 2364 } 2365 2366 /* update the free count for this dmap. 2367 */ 2368 le32_add_cpu(&dp->nfree, nblocks); 2369 2370 BMAP_LOCK(bmp); 2371 2372 /* update the free count for the allocation group and 2373 * map. 2374 */ 2375 agno = blkno >> bmp->db_agl2size; 2376 bmp->db_nfree += nblocks; 2377 bmp->db_agfree[agno] += nblocks; 2378 2379 /* check if this allocation group is not completely free and 2380 * if it is currently the maximum (rightmost) allocation group. 2381 * if so, establish the new maximum allocation group number by 2382 * searching left for the first allocation group with allocation. 2383 */ 2384 if ((bmp->db_agfree[agno] == bmp->db_agsize && agno == bmp->db_maxag) || 2385 (agno == bmp->db_numag - 1 && 2386 bmp->db_agfree[agno] == (bmp-> db_mapsize & (BPERDMAP - 1)))) { 2387 while (bmp->db_maxag > 0) { 2388 bmp->db_maxag -= 1; 2389 if (bmp->db_agfree[bmp->db_maxag] != 2390 bmp->db_agsize) 2391 break; 2392 } 2393 2394 /* re-establish the allocation group preference if the 2395 * current preference is right of the maximum allocation 2396 * group. 2397 */ 2398 if (bmp->db_agpref > bmp->db_maxag) 2399 bmp->db_agpref = bmp->db_maxag; 2400 } 2401 2402 BMAP_UNLOCK(bmp); 2403 2404 return 0; 2405 } 2406 2407 2408 /* 2409 * NAME: dbAdjCtl() 2410 * 2411 * FUNCTION: adjust a dmap control page at a specified level to reflect 2412 * the change in a lower level dmap or dmap control page's 2413 * maximum string of free blocks (i.e. a change in the root 2414 * of the lower level object's dmtree) due to the allocation 2415 * or deallocation of a range of blocks with a single dmap. 2416 * 2417 * on entry, this routine is provided with the new value of 2418 * the lower level dmap or dmap control page root and the 2419 * starting block number of the block range whose allocation 2420 * or deallocation resulted in the root change. this range 2421 * is respresented by a single leaf of the current dmapctl 2422 * and the leaf will be updated with this value, possibly 2423 * causing a binary buddy system within the leaves to be 2424 * split or joined. the update may also cause the dmapctl's 2425 * dmtree to be updated. 2426 * 2427 * if the adjustment of the dmap control page, itself, causes its 2428 * root to change, this change will be bubbled up to the next dmap 2429 * control level by a recursive call to this routine, specifying 2430 * the new root value and the next dmap control page level to 2431 * be adjusted. 2432 * PARAMETERS: 2433 * bmp - pointer to bmap descriptor 2434 * blkno - the first block of a block range within a dmap. it is 2435 * the allocation or deallocation of this block range that 2436 * requires the dmap control page to be adjusted. 2437 * newval - the new value of the lower level dmap or dmap control 2438 * page root. 2439 * alloc - 'true' if adjustment is due to an allocation. 2440 * level - current level of dmap control page (i.e. L0, L1, L2) to 2441 * be adjusted. 2442 * 2443 * RETURN VALUES: 2444 * 0 - success 2445 * -EIO - i/o error 2446 * 2447 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit; 2448 */ 2449 static int 2450 dbAdjCtl(struct bmap * bmp, s64 blkno, int newval, int alloc, int level) 2451 { 2452 struct metapage *mp; 2453 s8 oldroot; 2454 int oldval; 2455 s64 lblkno; 2456 struct dmapctl *dcp; 2457 int rc, leafno, ti; 2458 2459 /* get the buffer for the dmap control page for the specified 2460 * block number and control page level. 2461 */ 2462 lblkno = BLKTOCTL(blkno, bmp->db_l2nbperpage, level); 2463 mp = read_metapage(bmp->db_ipbmap, lblkno, PSIZE, 0); 2464 if (mp == NULL) 2465 return -EIO; 2466 dcp = (struct dmapctl *) mp->data; 2467 2468 if (dcp->leafidx != cpu_to_le32(CTLLEAFIND)) { 2469 jfs_error(bmp->db_ipbmap->i_sb, "Corrupt dmapctl page\n"); 2470 release_metapage(mp); 2471 return -EIO; 2472 } 2473 2474 /* determine the leaf number corresponding to the block and 2475 * the index within the dmap control tree. 2476 */ 2477 leafno = BLKTOCTLLEAF(blkno, dcp->budmin); 2478 ti = leafno + le32_to_cpu(dcp->leafidx); 2479 2480 /* save the current leaf value and the current root level (i.e. 2481 * maximum l2 free string described by this dmapctl). 2482 */ 2483 oldval = dcp->stree[ti]; 2484 oldroot = dcp->stree[ROOT]; 2485 2486 /* check if this is a control page update for an allocation. 2487 * if so, update the leaf to reflect the new leaf value using 2488 * dbSplit(); otherwise (deallocation), use dbJoin() to update 2489 * the leaf with the new value. in addition to updating the 2490 * leaf, dbSplit() will also split the binary buddy system of 2491 * the leaves, if required, and bubble new values within the 2492 * dmapctl tree, if required. similarly, dbJoin() will join 2493 * the binary buddy system of leaves and bubble new values up 2494 * the dmapctl tree as required by the new leaf value. 2495 */ 2496 if (alloc) { 2497 /* check if we are in the middle of a binary buddy 2498 * system. this happens when we are performing the 2499 * first allocation out of an allocation group that 2500 * is part (not the first part) of a larger binary 2501 * buddy system. if we are in the middle, back split 2502 * the system prior to calling dbSplit() which assumes 2503 * that it is at the front of a binary buddy system. 2504 */ 2505 if (oldval == NOFREE) { 2506 rc = dbBackSplit((dmtree_t *) dcp, leafno); 2507 if (rc) { 2508 release_metapage(mp); 2509 return rc; 2510 } 2511 oldval = dcp->stree[ti]; 2512 } 2513 dbSplit((dmtree_t *) dcp, leafno, dcp->budmin, newval); 2514 } else { 2515 rc = dbJoin((dmtree_t *) dcp, leafno, newval); 2516 if (rc) { 2517 release_metapage(mp); 2518 return rc; 2519 } 2520 } 2521 2522 /* check if the root of the current dmap control page changed due 2523 * to the update and if the current dmap control page is not at 2524 * the current top level (i.e. L0, L1, L2) of the map. if so (i.e. 2525 * root changed and this is not the top level), call this routine 2526 * again (recursion) for the next higher level of the mapping to 2527 * reflect the change in root for the current dmap control page. 2528 */ 2529 if (dcp->stree[ROOT] != oldroot) { 2530 /* are we below the top level of the map. if so, 2531 * bubble the root up to the next higher level. 2532 */ 2533 if (level < bmp->db_maxlevel) { 2534 /* bubble up the new root of this dmap control page to 2535 * the next level. 2536 */ 2537 if ((rc = 2538 dbAdjCtl(bmp, blkno, dcp->stree[ROOT], alloc, 2539 level + 1))) { 2540 /* something went wrong in bubbling up the new 2541 * root value, so backout the changes to the 2542 * current dmap control page. 2543 */ 2544 if (alloc) { 2545 dbJoin((dmtree_t *) dcp, leafno, 2546 oldval); 2547 } else { 2548 /* the dbJoin() above might have 2549 * caused a larger binary buddy system 2550 * to form and we may now be in the 2551 * middle of it. if this is the case, 2552 * back split the buddies. 2553 */ 2554 if (dcp->stree[ti] == NOFREE) 2555 dbBackSplit((dmtree_t *) 2556 dcp, leafno); 2557 dbSplit((dmtree_t *) dcp, leafno, 2558 dcp->budmin, oldval); 2559 } 2560 2561 /* release the buffer and return the error. 2562 */ 2563 release_metapage(mp); 2564 return (rc); 2565 } 2566 } else { 2567 /* we're at the top level of the map. update 2568 * the bmap control page to reflect the size 2569 * of the maximum free buddy system. 2570 */ 2571 assert(level == bmp->db_maxlevel); 2572 if (bmp->db_maxfreebud != oldroot) { 2573 jfs_error(bmp->db_ipbmap->i_sb, 2574 "the maximum free buddy is not the old root\n"); 2575 } 2576 bmp->db_maxfreebud = dcp->stree[ROOT]; 2577 } 2578 } 2579 2580 /* write the buffer. 2581 */ 2582 write_metapage(mp); 2583 2584 return (0); 2585 } 2586 2587 2588 /* 2589 * NAME: dbSplit() 2590 * 2591 * FUNCTION: update the leaf of a dmtree with a new value, splitting 2592 * the leaf from the binary buddy system of the dmtree's 2593 * leaves, as required. 2594 * 2595 * PARAMETERS: 2596 * tp - pointer to the tree containing the leaf. 2597 * leafno - the number of the leaf to be updated. 2598 * splitsz - the size the binary buddy system starting at the leaf 2599 * must be split to, specified as the log2 number of blocks. 2600 * newval - the new value for the leaf. 2601 * 2602 * RETURN VALUES: none 2603 * 2604 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit; 2605 */ 2606 static void dbSplit(dmtree_t * tp, int leafno, int splitsz, int newval) 2607 { 2608 int budsz; 2609 int cursz; 2610 s8 *leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx); 2611 2612 /* check if the leaf needs to be split. 2613 */ 2614 if (leaf[leafno] > tp->dmt_budmin) { 2615 /* the split occurs by cutting the buddy system in half 2616 * at the specified leaf until we reach the specified 2617 * size. pick up the starting split size (current size 2618 * - 1 in l2) and the corresponding buddy size. 2619 */ 2620 cursz = leaf[leafno] - 1; 2621 budsz = BUDSIZE(cursz, tp->dmt_budmin); 2622 2623 /* split until we reach the specified size. 2624 */ 2625 while (cursz >= splitsz) { 2626 /* update the buddy's leaf with its new value. 2627 */ 2628 dbAdjTree(tp, leafno ^ budsz, cursz); 2629 2630 /* on to the next size and buddy. 2631 */ 2632 cursz -= 1; 2633 budsz >>= 1; 2634 } 2635 } 2636 2637 /* adjust the dmap tree to reflect the specified leaf's new 2638 * value. 2639 */ 2640 dbAdjTree(tp, leafno, newval); 2641 } 2642 2643 2644 /* 2645 * NAME: dbBackSplit() 2646 * 2647 * FUNCTION: back split the binary buddy system of dmtree leaves 2648 * that hold a specified leaf until the specified leaf 2649 * starts its own binary buddy system. 2650 * 2651 * the allocators typically perform allocations at the start 2652 * of binary buddy systems and dbSplit() is used to accomplish 2653 * any required splits. in some cases, however, allocation 2654 * may occur in the middle of a binary system and requires a 2655 * back split, with the split proceeding out from the middle of 2656 * the system (less efficient) rather than the start of the 2657 * system (more efficient). the cases in which a back split 2658 * is required are rare and are limited to the first allocation 2659 * within an allocation group which is a part (not first part) 2660 * of a larger binary buddy system and a few exception cases 2661 * in which a previous join operation must be backed out. 2662 * 2663 * PARAMETERS: 2664 * tp - pointer to the tree containing the leaf. 2665 * leafno - the number of the leaf to be updated. 2666 * 2667 * RETURN VALUES: none 2668 * 2669 * serialization: IREAD_LOCK(ipbmap) or IWRITE_LOCK(ipbmap) held on entry/exit; 2670 */ 2671 static int dbBackSplit(dmtree_t * tp, int leafno) 2672 { 2673 int budsz, bud, w, bsz, size; 2674 int cursz; 2675 s8 *leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx); 2676 2677 /* leaf should be part (not first part) of a binary 2678 * buddy system. 2679 */ 2680 assert(leaf[leafno] == NOFREE); 2681 2682 /* the back split is accomplished by iteratively finding the leaf 2683 * that starts the buddy system that contains the specified leaf and 2684 * splitting that system in two. this iteration continues until 2685 * the specified leaf becomes the start of a buddy system. 2686 * 2687 * determine maximum possible l2 size for the specified leaf. 2688 */ 2689 size = 2690 LITOL2BSZ(leafno, le32_to_cpu(tp->dmt_l2nleafs), 2691 tp->dmt_budmin); 2692 2693 /* determine the number of leaves covered by this size. this 2694 * is the buddy size that we will start with as we search for 2695 * the buddy system that contains the specified leaf. 2696 */ 2697 budsz = BUDSIZE(size, tp->dmt_budmin); 2698 2699 /* back split. 2700 */ 2701 while (leaf[leafno] == NOFREE) { 2702 /* find the leftmost buddy leaf. 2703 */ 2704 for (w = leafno, bsz = budsz;; bsz <<= 1, 2705 w = (w < bud) ? w : bud) { 2706 if (bsz >= le32_to_cpu(tp->dmt_nleafs)) { 2707 jfs_err("JFS: block map error in dbBackSplit"); 2708 return -EIO; 2709 } 2710 2711 /* determine the buddy. 2712 */ 2713 bud = w ^ bsz; 2714 2715 /* check if this buddy is the start of the system. 2716 */ 2717 if (leaf[bud] != NOFREE) { 2718 /* split the leaf at the start of the 2719 * system in two. 2720 */ 2721 cursz = leaf[bud] - 1; 2722 dbSplit(tp, bud, cursz, cursz); 2723 break; 2724 } 2725 } 2726 } 2727 2728 if (leaf[leafno] != size) { 2729 jfs_err("JFS: wrong leaf value in dbBackSplit"); 2730 return -EIO; 2731 } 2732 return 0; 2733 } 2734 2735 2736 /* 2737 * NAME: dbJoin() 2738 * 2739 * FUNCTION: update the leaf of a dmtree with a new value, joining 2740 * the leaf with other leaves of the dmtree into a multi-leaf 2741 * binary buddy system, as required. 2742 * 2743 * PARAMETERS: 2744 * tp - pointer to the tree containing the leaf. 2745 * leafno - the number of the leaf to be updated. 2746 * newval - the new value for the leaf. 2747 * 2748 * RETURN VALUES: none 2749 */ 2750 static int dbJoin(dmtree_t * tp, int leafno, int newval) 2751 { 2752 int budsz, buddy; 2753 s8 *leaf; 2754 2755 /* can the new leaf value require a join with other leaves ? 2756 */ 2757 if (newval >= tp->dmt_budmin) { 2758 /* pickup a pointer to the leaves of the tree. 2759 */ 2760 leaf = tp->dmt_stree + le32_to_cpu(tp->dmt_leafidx); 2761 2762 /* try to join the specified leaf into a large binary 2763 * buddy system. the join proceeds by attempting to join 2764 * the specified leafno with its buddy (leaf) at new value. 2765 * if the join occurs, we attempt to join the left leaf 2766 * of the joined buddies with its buddy at new value + 1. 2767 * we continue to join until we find a buddy that cannot be 2768 * joined (does not have a value equal to the size of the 2769 * last join) or until all leaves have been joined into a 2770 * single system. 2771 * 2772 * get the buddy size (number of words covered) of 2773 * the new value. 2774 */ 2775 budsz = BUDSIZE(newval, tp->dmt_budmin); 2776 2777 /* try to join. 2778 */ 2779 while (budsz < le32_to_cpu(tp->dmt_nleafs)) { 2780 /* get the buddy leaf. 2781 */ 2782 buddy = leafno ^ budsz; 2783 2784 /* if the leaf's new value is greater than its 2785 * buddy's value, we join no more. 2786 */ 2787 if (newval > leaf[buddy]) 2788 break; 2789 2790 /* It shouldn't be less */ 2791 if (newval < leaf[buddy]) 2792 return -EIO; 2793 2794 /* check which (leafno or buddy) is the left buddy. 2795 * the left buddy gets to claim the blocks resulting 2796 * from the join while the right gets to claim none. 2797 * the left buddy is also eligible to participate in 2798 * a join at the next higher level while the right 2799 * is not. 2800 * 2801 */ 2802 if (leafno < buddy) { 2803 /* leafno is the left buddy. 2804 */ 2805 dbAdjTree(tp, buddy, NOFREE); 2806 } else { 2807 /* buddy is the left buddy and becomes 2808 * leafno. 2809 */ 2810 dbAdjTree(tp, leafno, NOFREE); 2811 leafno = buddy; 2812 } 2813 2814 /* on to try the next join. 2815 */ 2816 newval += 1; 2817 budsz <<= 1; 2818 } 2819 } 2820 2821 /* update the leaf value. 2822 */ 2823 dbAdjTree(tp, leafno, newval); 2824 2825 return 0; 2826 } 2827 2828 2829 /* 2830 * NAME: dbAdjTree() 2831 * 2832 * FUNCTION: update a leaf of a dmtree with a new value, adjusting 2833 * the dmtree, as required, to reflect the new leaf value. 2834 * the combination of any buddies must already be done before 2835 * this is called. 2836 * 2837 * PARAMETERS: 2838 * tp - pointer to the tree to be adjusted. 2839 * leafno - the number of the leaf to be updated. 2840 * newval - the new value for the leaf. 2841 * 2842 * RETURN VALUES: none 2843 */ 2844 static void dbAdjTree(dmtree_t * tp, int leafno, int newval) 2845 { 2846 int lp, pp, k; 2847 int max; 2848 2849 /* pick up the index of the leaf for this leafno. 2850 */ 2851 lp = leafno + le32_to_cpu(tp->dmt_leafidx); 2852 2853 /* is the current value the same as the old value ? if so, 2854 * there is nothing to do. 2855 */ 2856 if (tp->dmt_stree[lp] == newval) 2857 return; 2858 2859 /* set the new value. 2860 */ 2861 tp->dmt_stree[lp] = newval; 2862 2863 /* bubble the new value up the tree as required. 2864 */ 2865 for (k = 0; k < le32_to_cpu(tp->dmt_height); k++) { 2866 /* get the index of the first leaf of the 4 leaf 2867 * group containing the specified leaf (leafno). 2868 */ 2869 lp = ((lp - 1) & ~0x03) + 1; 2870 2871 /* get the index of the parent of this 4 leaf group. 2872 */ 2873 pp = (lp - 1) >> 2; 2874 2875 /* determine the maximum of the 4 leaves. 2876 */ 2877 max = TREEMAX(&tp->dmt_stree[lp]); 2878 2879 /* if the maximum of the 4 is the same as the 2880 * parent's value, we're done. 2881 */ 2882 if (tp->dmt_stree[pp] == max) 2883 break; 2884 2885 /* parent gets new value. 2886 */ 2887 tp->dmt_stree[pp] = max; 2888 2889 /* parent becomes leaf for next go-round. 2890 */ 2891 lp = pp; 2892 } 2893 } 2894 2895 2896 /* 2897 * NAME: dbFindLeaf() 2898 * 2899 * FUNCTION: search a dmtree_t for sufficient free blocks, returning 2900 * the index of a leaf describing the free blocks if 2901 * sufficient free blocks are found. 2902 * 2903 * the search starts at the top of the dmtree_t tree and 2904 * proceeds down the tree to the leftmost leaf with sufficient 2905 * free space. 2906 * 2907 * PARAMETERS: 2908 * tp - pointer to the tree to be searched. 2909 * l2nb - log2 number of free blocks to search for. 2910 * leafidx - return pointer to be set to the index of the leaf 2911 * describing at least l2nb free blocks if sufficient 2912 * free blocks are found. 2913 * 2914 * RETURN VALUES: 2915 * 0 - success 2916 * -ENOSPC - insufficient free blocks. 2917 */ 2918 static int dbFindLeaf(dmtree_t * tp, int l2nb, int *leafidx) 2919 { 2920 int ti, n = 0, k, x = 0; 2921 2922 /* first check the root of the tree to see if there is 2923 * sufficient free space. 2924 */ 2925 if (l2nb > tp->dmt_stree[ROOT]) 2926 return -ENOSPC; 2927 2928 /* sufficient free space available. now search down the tree 2929 * starting at the next level for the leftmost leaf that 2930 * describes sufficient free space. 2931 */ 2932 for (k = le32_to_cpu(tp->dmt_height), ti = 1; 2933 k > 0; k--, ti = ((ti + n) << 2) + 1) { 2934 /* search the four nodes at this level, starting from 2935 * the left. 2936 */ 2937 for (x = ti, n = 0; n < 4; n++) { 2938 /* sufficient free space found. move to the next 2939 * level (or quit if this is the last level). 2940 */ 2941 if (l2nb <= tp->dmt_stree[x + n]) 2942 break; 2943 } 2944 2945 /* better have found something since the higher 2946 * levels of the tree said it was here. 2947 */ 2948 assert(n < 4); 2949 } 2950 2951 /* set the return to the leftmost leaf describing sufficient 2952 * free space. 2953 */ 2954 *leafidx = x + n - le32_to_cpu(tp->dmt_leafidx); 2955 2956 return (0); 2957 } 2958 2959 2960 /* 2961 * NAME: dbFindBits() 2962 * 2963 * FUNCTION: find a specified number of binary buddy free bits within a 2964 * dmap bitmap word value. 2965 * 2966 * this routine searches the bitmap value for (1 << l2nb) free 2967 * bits at (1 << l2nb) alignments within the value. 2968 * 2969 * PARAMETERS: 2970 * word - dmap bitmap word value. 2971 * l2nb - number of free bits specified as a log2 number. 2972 * 2973 * RETURN VALUES: 2974 * starting bit number of free bits. 2975 */ 2976 static int dbFindBits(u32 word, int l2nb) 2977 { 2978 int bitno, nb; 2979 u32 mask; 2980 2981 /* get the number of bits. 2982 */ 2983 nb = 1 << l2nb; 2984 assert(nb <= DBWORD); 2985 2986 /* complement the word so we can use a mask (i.e. 0s represent 2987 * free bits) and compute the mask. 2988 */ 2989 word = ~word; 2990 mask = ONES << (DBWORD - nb); 2991 2992 /* scan the word for nb free bits at nb alignments. 2993 */ 2994 for (bitno = 0; mask != 0; bitno += nb, mask >>= nb) { 2995 if ((mask & word) == mask) 2996 break; 2997 } 2998 2999 ASSERT(bitno < 32); 3000 3001 /* return the bit number. 3002 */ 3003 return (bitno); 3004 } 3005 3006 3007 /* 3008 * NAME: dbMaxBud(u8 *cp) 3009 * 3010 * FUNCTION: determine the largest binary buddy string of free 3011 * bits within 32-bits of the map. 3012 * 3013 * PARAMETERS: 3014 * cp - pointer to the 32-bit value. 3015 * 3016 * RETURN VALUES: 3017 * largest binary buddy of free bits within a dmap word. 3018 */ 3019 static int dbMaxBud(u8 * cp) 3020 { 3021 signed char tmp1, tmp2; 3022 3023 /* check if the wmap word is all free. if so, the 3024 * free buddy size is BUDMIN. 3025 */ 3026 if (*((uint *) cp) == 0) 3027 return (BUDMIN); 3028 3029 /* check if the wmap word is half free. if so, the 3030 * free buddy size is BUDMIN-1. 3031 */ 3032 if (*((u16 *) cp) == 0 || *((u16 *) cp + 1) == 0) 3033 return (BUDMIN - 1); 3034 3035 /* not all free or half free. determine the free buddy 3036 * size thru table lookup using quarters of the wmap word. 3037 */ 3038 tmp1 = max(budtab[cp[2]], budtab[cp[3]]); 3039 tmp2 = max(budtab[cp[0]], budtab[cp[1]]); 3040 return (max(tmp1, tmp2)); 3041 } 3042 3043 3044 /* 3045 * NAME: cnttz(uint word) 3046 * 3047 * FUNCTION: determine the number of trailing zeros within a 32-bit 3048 * value. 3049 * 3050 * PARAMETERS: 3051 * value - 32-bit value to be examined. 3052 * 3053 * RETURN VALUES: 3054 * count of trailing zeros 3055 */ 3056 static int cnttz(u32 word) 3057 { 3058 int n; 3059 3060 for (n = 0; n < 32; n++, word >>= 1) { 3061 if (word & 0x01) 3062 break; 3063 } 3064 3065 return (n); 3066 } 3067 3068 3069 /* 3070 * NAME: cntlz(u32 value) 3071 * 3072 * FUNCTION: determine the number of leading zeros within a 32-bit 3073 * value. 3074 * 3075 * PARAMETERS: 3076 * value - 32-bit value to be examined. 3077 * 3078 * RETURN VALUES: 3079 * count of leading zeros 3080 */ 3081 static int cntlz(u32 value) 3082 { 3083 int n; 3084 3085 for (n = 0; n < 32; n++, value <<= 1) { 3086 if (value & HIGHORDER) 3087 break; 3088 } 3089 return (n); 3090 } 3091 3092 3093 /* 3094 * NAME: blkstol2(s64 nb) 3095 * 3096 * FUNCTION: convert a block count to its log2 value. if the block 3097 * count is not a l2 multiple, it is rounded up to the next 3098 * larger l2 multiple. 3099 * 3100 * PARAMETERS: 3101 * nb - number of blocks 3102 * 3103 * RETURN VALUES: 3104 * log2 number of blocks 3105 */ 3106 static int blkstol2(s64 nb) 3107 { 3108 int l2nb; 3109 s64 mask; /* meant to be signed */ 3110 3111 mask = (s64) 1 << (64 - 1); 3112 3113 /* count the leading bits. 3114 */ 3115 for (l2nb = 0; l2nb < 64; l2nb++, mask >>= 1) { 3116 /* leading bit found. 3117 */ 3118 if (nb & mask) { 3119 /* determine the l2 value. 3120 */ 3121 l2nb = (64 - 1) - l2nb; 3122 3123 /* check if we need to round up. 3124 */ 3125 if (~mask & nb) 3126 l2nb++; 3127 3128 return (l2nb); 3129 } 3130 } 3131 assert(0); 3132 return 0; /* fix compiler warning */ 3133 } 3134 3135 3136 /* 3137 * NAME: dbAllocBottomUp() 3138 * 3139 * FUNCTION: alloc the specified block range from the working block 3140 * allocation map. 3141 * 3142 * the blocks will be alloc from the working map one dmap 3143 * at a time. 3144 * 3145 * PARAMETERS: 3146 * ip - pointer to in-core inode; 3147 * blkno - starting block number to be freed. 3148 * nblocks - number of blocks to be freed. 3149 * 3150 * RETURN VALUES: 3151 * 0 - success 3152 * -EIO - i/o error 3153 */ 3154 int dbAllocBottomUp(struct inode *ip, s64 blkno, s64 nblocks) 3155 { 3156 struct metapage *mp; 3157 struct dmap *dp; 3158 int nb, rc; 3159 s64 lblkno, rem; 3160 struct inode *ipbmap = JFS_SBI(ip->i_sb)->ipbmap; 3161 struct bmap *bmp = JFS_SBI(ip->i_sb)->bmap; 3162 3163 IREAD_LOCK(ipbmap, RDWRLOCK_DMAP); 3164 3165 /* block to be allocated better be within the mapsize. */ 3166 ASSERT(nblocks <= bmp->db_mapsize - blkno); 3167 3168 /* 3169 * allocate the blocks a dmap at a time. 3170 */ 3171 mp = NULL; 3172 for (rem = nblocks; rem > 0; rem -= nb, blkno += nb) { 3173 /* release previous dmap if any */ 3174 if (mp) { 3175 write_metapage(mp); 3176 } 3177 3178 /* get the buffer for the current dmap. */ 3179 lblkno = BLKTODMAP(blkno, bmp->db_l2nbperpage); 3180 mp = read_metapage(ipbmap, lblkno, PSIZE, 0); 3181 if (mp == NULL) { 3182 IREAD_UNLOCK(ipbmap); 3183 return -EIO; 3184 } 3185 dp = (struct dmap *) mp->data; 3186 3187 /* determine the number of blocks to be allocated from 3188 * this dmap. 3189 */ 3190 nb = min(rem, BPERDMAP - (blkno & (BPERDMAP - 1))); 3191 3192 /* allocate the blocks. */ 3193 if ((rc = dbAllocDmapBU(bmp, dp, blkno, nb))) { 3194 release_metapage(mp); 3195 IREAD_UNLOCK(ipbmap); 3196 return (rc); 3197 } 3198 } 3199 3200 /* write the last buffer. */ 3201 write_metapage(mp); 3202 3203 IREAD_UNLOCK(ipbmap); 3204 3205 return (0); 3206 } 3207 3208 3209 static int dbAllocDmapBU(struct bmap * bmp, struct dmap * dp, s64 blkno, 3210 int nblocks) 3211 { 3212 int rc; 3213 int dbitno, word, rembits, nb, nwords, wbitno, agno; 3214 s8 oldroot; 3215 struct dmaptree *tp = (struct dmaptree *) & dp->tree; 3216 3217 /* save the current value of the root (i.e. maximum free string) 3218 * of the dmap tree. 3219 */ 3220 oldroot = tp->stree[ROOT]; 3221 3222 /* determine the bit number and word within the dmap of the 3223 * starting block. 3224 */ 3225 dbitno = blkno & (BPERDMAP - 1); 3226 word = dbitno >> L2DBWORD; 3227 3228 /* block range better be within the dmap */ 3229 assert(dbitno + nblocks <= BPERDMAP); 3230 3231 /* allocate the bits of the dmap's words corresponding to the block 3232 * range. not all bits of the first and last words may be contained 3233 * within the block range. if this is the case, we'll work against 3234 * those words (i.e. partial first and/or last) on an individual basis 3235 * (a single pass), allocating the bits of interest by hand and 3236 * updating the leaf corresponding to the dmap word. a single pass 3237 * will be used for all dmap words fully contained within the 3238 * specified range. within this pass, the bits of all fully contained 3239 * dmap words will be marked as free in a single shot and the leaves 3240 * will be updated. a single leaf may describe the free space of 3241 * multiple dmap words, so we may update only a subset of the actual 3242 * leaves corresponding to the dmap words of the block range. 3243 */ 3244 for (rembits = nblocks; rembits > 0; rembits -= nb, dbitno += nb) { 3245 /* determine the bit number within the word and 3246 * the number of bits within the word. 3247 */ 3248 wbitno = dbitno & (DBWORD - 1); 3249 nb = min(rembits, DBWORD - wbitno); 3250 3251 /* check if only part of a word is to be allocated. 3252 */ 3253 if (nb < DBWORD) { 3254 /* allocate (set to 1) the appropriate bits within 3255 * this dmap word. 3256 */ 3257 dp->wmap[word] |= cpu_to_le32(ONES << (DBWORD - nb) 3258 >> wbitno); 3259 3260 word++; 3261 } else { 3262 /* one or more dmap words are fully contained 3263 * within the block range. determine how many 3264 * words and allocate (set to 1) the bits of these 3265 * words. 3266 */ 3267 nwords = rembits >> L2DBWORD; 3268 memset(&dp->wmap[word], (int) ONES, nwords * 4); 3269 3270 /* determine how many bits */ 3271 nb = nwords << L2DBWORD; 3272 word += nwords; 3273 } 3274 } 3275 3276 /* update the free count for this dmap */ 3277 le32_add_cpu(&dp->nfree, -nblocks); 3278 3279 /* reconstruct summary tree */ 3280 dbInitDmapTree(dp); 3281 3282 BMAP_LOCK(bmp); 3283 3284 /* if this allocation group is completely free, 3285 * update the highest active allocation group number 3286 * if this allocation group is the new max. 3287 */ 3288 agno = blkno >> bmp->db_agl2size; 3289 if (agno > bmp->db_maxag) 3290 bmp->db_maxag = agno; 3291 3292 /* update the free count for the allocation group and map */ 3293 bmp->db_agfree[agno] -= nblocks; 3294 bmp->db_nfree -= nblocks; 3295 3296 BMAP_UNLOCK(bmp); 3297 3298 /* if the root has not changed, done. */ 3299 if (tp->stree[ROOT] == oldroot) 3300 return (0); 3301 3302 /* root changed. bubble the change up to the dmap control pages. 3303 * if the adjustment of the upper level control pages fails, 3304 * backout the bit allocation (thus making everything consistent). 3305 */ 3306 if ((rc = dbAdjCtl(bmp, blkno, tp->stree[ROOT], 1, 0))) 3307 dbFreeBits(bmp, dp, blkno, nblocks); 3308 3309 return (rc); 3310 } 3311 3312 3313 /* 3314 * NAME: dbExtendFS() 3315 * 3316 * FUNCTION: extend bmap from blkno for nblocks; 3317 * dbExtendFS() updates bmap ready for dbAllocBottomUp(); 3318 * 3319 * L2 3320 * | 3321 * L1---------------------------------L1 3322 * | | 3323 * L0---------L0---------L0 L0---------L0---------L0 3324 * | | | | | | 3325 * d0,...,dn d0,...,dn d0,...,dn d0,...,dn d0,...,dn d0,.,dm; 3326 * L2L1L0d0,...,dnL0d0,...,dnL0d0,...,dnL1L0d0,...,dnL0d0,...,dnL0d0,..dm 3327 * 3328 * <---old---><----------------------------extend-----------------------> 3329 */ 3330 int dbExtendFS(struct inode *ipbmap, s64 blkno, s64 nblocks) 3331 { 3332 struct jfs_sb_info *sbi = JFS_SBI(ipbmap->i_sb); 3333 int nbperpage = sbi->nbperpage; 3334 int i, i0 = true, j, j0 = true, k, n; 3335 s64 newsize; 3336 s64 p; 3337 struct metapage *mp, *l2mp, *l1mp = NULL, *l0mp = NULL; 3338 struct dmapctl *l2dcp, *l1dcp, *l0dcp; 3339 struct dmap *dp; 3340 s8 *l0leaf, *l1leaf, *l2leaf; 3341 struct bmap *bmp = sbi->bmap; 3342 int agno, l2agsize, oldl2agsize; 3343 s64 ag_rem; 3344 3345 newsize = blkno + nblocks; 3346 3347 jfs_info("dbExtendFS: blkno:%Ld nblocks:%Ld newsize:%Ld", 3348 (long long) blkno, (long long) nblocks, (long long) newsize); 3349 3350 /* 3351 * initialize bmap control page. 3352 * 3353 * all the data in bmap control page should exclude 3354 * the mkfs hidden dmap page. 3355 */ 3356 3357 /* update mapsize */ 3358 bmp->db_mapsize = newsize; 3359 bmp->db_maxlevel = BMAPSZTOLEV(bmp->db_mapsize); 3360 3361 /* compute new AG size */ 3362 l2agsize = dbGetL2AGSize(newsize); 3363 oldl2agsize = bmp->db_agl2size; 3364 3365 bmp->db_agl2size = l2agsize; 3366 bmp->db_agsize = 1 << l2agsize; 3367 3368 /* compute new number of AG */ 3369 agno = bmp->db_numag; 3370 bmp->db_numag = newsize >> l2agsize; 3371 bmp->db_numag += ((u32) newsize % (u32) bmp->db_agsize) ? 1 : 0; 3372 3373 /* 3374 * reconfigure db_agfree[] 3375 * from old AG configuration to new AG configuration; 3376 * 3377 * coalesce contiguous k (newAGSize/oldAGSize) AGs; 3378 * i.e., (AGi, ..., AGj) where i = k*n and j = k*(n+1) - 1 to AGn; 3379 * note: new AG size = old AG size * (2**x). 3380 */ 3381 if (l2agsize == oldl2agsize) 3382 goto extend; 3383 k = 1 << (l2agsize - oldl2agsize); 3384 ag_rem = bmp->db_agfree[0]; /* save agfree[0] */ 3385 for (i = 0, n = 0; i < agno; n++) { 3386 bmp->db_agfree[n] = 0; /* init collection point */ 3387 3388 /* coalesce contiguous k AGs; */ 3389 for (j = 0; j < k && i < agno; j++, i++) { 3390 /* merge AGi to AGn */ 3391 bmp->db_agfree[n] += bmp->db_agfree[i]; 3392 } 3393 } 3394 bmp->db_agfree[0] += ag_rem; /* restore agfree[0] */ 3395 3396 for (; n < MAXAG; n++) 3397 bmp->db_agfree[n] = 0; 3398 3399 /* 3400 * update highest active ag number 3401 */ 3402 3403 bmp->db_maxag = bmp->db_maxag / k; 3404 3405 /* 3406 * extend bmap 3407 * 3408 * update bit maps and corresponding level control pages; 3409 * global control page db_nfree, db_agfree[agno], db_maxfreebud; 3410 */ 3411 extend: 3412 /* get L2 page */ 3413 p = BMAPBLKNO + nbperpage; /* L2 page */ 3414 l2mp = read_metapage(ipbmap, p, PSIZE, 0); 3415 if (!l2mp) { 3416 jfs_error(ipbmap->i_sb, "L2 page could not be read\n"); 3417 return -EIO; 3418 } 3419 l2dcp = (struct dmapctl *) l2mp->data; 3420 3421 /* compute start L1 */ 3422 k = blkno >> L2MAXL1SIZE; 3423 l2leaf = l2dcp->stree + CTLLEAFIND + k; 3424 p = BLKTOL1(blkno, sbi->l2nbperpage); /* L1 page */ 3425 3426 /* 3427 * extend each L1 in L2 3428 */ 3429 for (; k < LPERCTL; k++, p += nbperpage) { 3430 /* get L1 page */ 3431 if (j0) { 3432 /* read in L1 page: (blkno & (MAXL1SIZE - 1)) */ 3433 l1mp = read_metapage(ipbmap, p, PSIZE, 0); 3434 if (l1mp == NULL) 3435 goto errout; 3436 l1dcp = (struct dmapctl *) l1mp->data; 3437 3438 /* compute start L0 */ 3439 j = (blkno & (MAXL1SIZE - 1)) >> L2MAXL0SIZE; 3440 l1leaf = l1dcp->stree + CTLLEAFIND + j; 3441 p = BLKTOL0(blkno, sbi->l2nbperpage); 3442 j0 = false; 3443 } else { 3444 /* assign/init L1 page */ 3445 l1mp = get_metapage(ipbmap, p, PSIZE, 0); 3446 if (l1mp == NULL) 3447 goto errout; 3448 3449 l1dcp = (struct dmapctl *) l1mp->data; 3450 3451 /* compute start L0 */ 3452 j = 0; 3453 l1leaf = l1dcp->stree + CTLLEAFIND; 3454 p += nbperpage; /* 1st L0 of L1.k */ 3455 } 3456 3457 /* 3458 * extend each L0 in L1 3459 */ 3460 for (; j < LPERCTL; j++) { 3461 /* get L0 page */ 3462 if (i0) { 3463 /* read in L0 page: (blkno & (MAXL0SIZE - 1)) */ 3464 3465 l0mp = read_metapage(ipbmap, p, PSIZE, 0); 3466 if (l0mp == NULL) 3467 goto errout; 3468 l0dcp = (struct dmapctl *) l0mp->data; 3469 3470 /* compute start dmap */ 3471 i = (blkno & (MAXL0SIZE - 1)) >> 3472 L2BPERDMAP; 3473 l0leaf = l0dcp->stree + CTLLEAFIND + i; 3474 p = BLKTODMAP(blkno, 3475 sbi->l2nbperpage); 3476 i0 = false; 3477 } else { 3478 /* assign/init L0 page */ 3479 l0mp = get_metapage(ipbmap, p, PSIZE, 0); 3480 if (l0mp == NULL) 3481 goto errout; 3482 3483 l0dcp = (struct dmapctl *) l0mp->data; 3484 3485 /* compute start dmap */ 3486 i = 0; 3487 l0leaf = l0dcp->stree + CTLLEAFIND; 3488 p += nbperpage; /* 1st dmap of L0.j */ 3489 } 3490 3491 /* 3492 * extend each dmap in L0 3493 */ 3494 for (; i < LPERCTL; i++) { 3495 /* 3496 * reconstruct the dmap page, and 3497 * initialize corresponding parent L0 leaf 3498 */ 3499 if ((n = blkno & (BPERDMAP - 1))) { 3500 /* read in dmap page: */ 3501 mp = read_metapage(ipbmap, p, 3502 PSIZE, 0); 3503 if (mp == NULL) 3504 goto errout; 3505 n = min(nblocks, (s64)BPERDMAP - n); 3506 } else { 3507 /* assign/init dmap page */ 3508 mp = read_metapage(ipbmap, p, 3509 PSIZE, 0); 3510 if (mp == NULL) 3511 goto errout; 3512 3513 n = min_t(s64, nblocks, BPERDMAP); 3514 } 3515 3516 dp = (struct dmap *) mp->data; 3517 *l0leaf = dbInitDmap(dp, blkno, n); 3518 3519 bmp->db_nfree += n; 3520 agno = le64_to_cpu(dp->start) >> l2agsize; 3521 bmp->db_agfree[agno] += n; 3522 3523 write_metapage(mp); 3524 3525 l0leaf++; 3526 p += nbperpage; 3527 3528 blkno += n; 3529 nblocks -= n; 3530 if (nblocks == 0) 3531 break; 3532 } /* for each dmap in a L0 */ 3533 3534 /* 3535 * build current L0 page from its leaves, and 3536 * initialize corresponding parent L1 leaf 3537 */ 3538 *l1leaf = dbInitDmapCtl(l0dcp, 0, ++i); 3539 write_metapage(l0mp); 3540 l0mp = NULL; 3541 3542 if (nblocks) 3543 l1leaf++; /* continue for next L0 */ 3544 else { 3545 /* more than 1 L0 ? */ 3546 if (j > 0) 3547 break; /* build L1 page */ 3548 else { 3549 /* summarize in global bmap page */ 3550 bmp->db_maxfreebud = *l1leaf; 3551 release_metapage(l1mp); 3552 release_metapage(l2mp); 3553 goto finalize; 3554 } 3555 } 3556 } /* for each L0 in a L1 */ 3557 3558 /* 3559 * build current L1 page from its leaves, and 3560 * initialize corresponding parent L2 leaf 3561 */ 3562 *l2leaf = dbInitDmapCtl(l1dcp, 1, ++j); 3563 write_metapage(l1mp); 3564 l1mp = NULL; 3565 3566 if (nblocks) 3567 l2leaf++; /* continue for next L1 */ 3568 else { 3569 /* more than 1 L1 ? */ 3570 if (k > 0) 3571 break; /* build L2 page */ 3572 else { 3573 /* summarize in global bmap page */ 3574 bmp->db_maxfreebud = *l2leaf; 3575 release_metapage(l2mp); 3576 goto finalize; 3577 } 3578 } 3579 } /* for each L1 in a L2 */ 3580 3581 jfs_error(ipbmap->i_sb, "function has not returned as expected\n"); 3582 errout: 3583 if (l0mp) 3584 release_metapage(l0mp); 3585 if (l1mp) 3586 release_metapage(l1mp); 3587 release_metapage(l2mp); 3588 return -EIO; 3589 3590 /* 3591 * finalize bmap control page 3592 */ 3593 finalize: 3594 3595 return 0; 3596 } 3597 3598 3599 /* 3600 * dbFinalizeBmap() 3601 */ 3602 void dbFinalizeBmap(struct inode *ipbmap) 3603 { 3604 struct bmap *bmp = JFS_SBI(ipbmap->i_sb)->bmap; 3605 int actags, inactags, l2nl; 3606 s64 ag_rem, actfree, inactfree, avgfree; 3607 int i, n; 3608 3609 /* 3610 * finalize bmap control page 3611 */ 3612 //finalize: 3613 /* 3614 * compute db_agpref: preferred ag to allocate from 3615 * (the leftmost ag with average free space in it); 3616 */ 3617 //agpref: 3618 /* get the number of active ags and inactive ags */ 3619 actags = bmp->db_maxag + 1; 3620 inactags = bmp->db_numag - actags; 3621 ag_rem = bmp->db_mapsize & (bmp->db_agsize - 1); /* ??? */ 3622 3623 /* determine how many blocks are in the inactive allocation 3624 * groups. in doing this, we must account for the fact that 3625 * the rightmost group might be a partial group (i.e. file 3626 * system size is not a multiple of the group size). 3627 */ 3628 inactfree = (inactags && ag_rem) ? 3629 ((inactags - 1) << bmp->db_agl2size) + ag_rem 3630 : inactags << bmp->db_agl2size; 3631 3632 /* determine how many free blocks are in the active 3633 * allocation groups plus the average number of free blocks 3634 * within the active ags. 3635 */ 3636 actfree = bmp->db_nfree - inactfree; 3637 avgfree = (u32) actfree / (u32) actags; 3638 3639 /* if the preferred allocation group has not average free space. 3640 * re-establish the preferred group as the leftmost 3641 * group with average free space. 3642 */ 3643 if (bmp->db_agfree[bmp->db_agpref] < avgfree) { 3644 for (bmp->db_agpref = 0; bmp->db_agpref < actags; 3645 bmp->db_agpref++) { 3646 if (bmp->db_agfree[bmp->db_agpref] >= avgfree) 3647 break; 3648 } 3649 if (bmp->db_agpref >= bmp->db_numag) { 3650 jfs_error(ipbmap->i_sb, 3651 "cannot find ag with average freespace\n"); 3652 } 3653 } 3654 3655 /* 3656 * compute db_aglevel, db_agheight, db_width, db_agstart: 3657 * an ag is covered in aglevel dmapctl summary tree, 3658 * at agheight level height (from leaf) with agwidth number of nodes 3659 * each, which starts at agstart index node of the smmary tree node 3660 * array; 3661 */ 3662 bmp->db_aglevel = BMAPSZTOLEV(bmp->db_agsize); 3663 l2nl = 3664 bmp->db_agl2size - (L2BPERDMAP + bmp->db_aglevel * L2LPERCTL); 3665 bmp->db_agheight = l2nl >> 1; 3666 bmp->db_agwidth = 1 << (l2nl - (bmp->db_agheight << 1)); 3667 for (i = 5 - bmp->db_agheight, bmp->db_agstart = 0, n = 1; i > 0; 3668 i--) { 3669 bmp->db_agstart += n; 3670 n <<= 2; 3671 } 3672 3673 } 3674 3675 3676 /* 3677 * NAME: dbInitDmap()/ujfs_idmap_page() 3678 * 3679 * FUNCTION: initialize working/persistent bitmap of the dmap page 3680 * for the specified number of blocks: 3681 * 3682 * at entry, the bitmaps had been initialized as free (ZEROS); 3683 * The number of blocks will only account for the actually 3684 * existing blocks. Blocks which don't actually exist in 3685 * the aggregate will be marked as allocated (ONES); 3686 * 3687 * PARAMETERS: 3688 * dp - pointer to page of map 3689 * nblocks - number of blocks this page 3690 * 3691 * RETURNS: NONE 3692 */ 3693 static int dbInitDmap(struct dmap * dp, s64 Blkno, int nblocks) 3694 { 3695 int blkno, w, b, r, nw, nb, i; 3696 3697 /* starting block number within the dmap */ 3698 blkno = Blkno & (BPERDMAP - 1); 3699 3700 if (blkno == 0) { 3701 dp->nblocks = dp->nfree = cpu_to_le32(nblocks); 3702 dp->start = cpu_to_le64(Blkno); 3703 3704 if (nblocks == BPERDMAP) { 3705 memset(&dp->wmap[0], 0, LPERDMAP * 4); 3706 memset(&dp->pmap[0], 0, LPERDMAP * 4); 3707 goto initTree; 3708 } 3709 } else { 3710 le32_add_cpu(&dp->nblocks, nblocks); 3711 le32_add_cpu(&dp->nfree, nblocks); 3712 } 3713 3714 /* word number containing start block number */ 3715 w = blkno >> L2DBWORD; 3716 3717 /* 3718 * free the bits corresponding to the block range (ZEROS): 3719 * note: not all bits of the first and last words may be contained 3720 * within the block range. 3721 */ 3722 for (r = nblocks; r > 0; r -= nb, blkno += nb) { 3723 /* number of bits preceding range to be freed in the word */ 3724 b = blkno & (DBWORD - 1); 3725 /* number of bits to free in the word */ 3726 nb = min(r, DBWORD - b); 3727 3728 /* is partial word to be freed ? */ 3729 if (nb < DBWORD) { 3730 /* free (set to 0) from the bitmap word */ 3731 dp->wmap[w] &= cpu_to_le32(~(ONES << (DBWORD - nb) 3732 >> b)); 3733 dp->pmap[w] &= cpu_to_le32(~(ONES << (DBWORD - nb) 3734 >> b)); 3735 3736 /* skip the word freed */ 3737 w++; 3738 } else { 3739 /* free (set to 0) contiguous bitmap words */ 3740 nw = r >> L2DBWORD; 3741 memset(&dp->wmap[w], 0, nw * 4); 3742 memset(&dp->pmap[w], 0, nw * 4); 3743 3744 /* skip the words freed */ 3745 nb = nw << L2DBWORD; 3746 w += nw; 3747 } 3748 } 3749 3750 /* 3751 * mark bits following the range to be freed (non-existing 3752 * blocks) as allocated (ONES) 3753 */ 3754 3755 if (blkno == BPERDMAP) 3756 goto initTree; 3757 3758 /* the first word beyond the end of existing blocks */ 3759 w = blkno >> L2DBWORD; 3760 3761 /* does nblocks fall on a 32-bit boundary ? */ 3762 b = blkno & (DBWORD - 1); 3763 if (b) { 3764 /* mark a partial word allocated */ 3765 dp->wmap[w] = dp->pmap[w] = cpu_to_le32(ONES >> b); 3766 w++; 3767 } 3768 3769 /* set the rest of the words in the page to allocated (ONES) */ 3770 for (i = w; i < LPERDMAP; i++) 3771 dp->pmap[i] = dp->wmap[i] = cpu_to_le32(ONES); 3772 3773 /* 3774 * init tree 3775 */ 3776 initTree: 3777 return (dbInitDmapTree(dp)); 3778 } 3779 3780 3781 /* 3782 * NAME: dbInitDmapTree()/ujfs_complete_dmap() 3783 * 3784 * FUNCTION: initialize summary tree of the specified dmap: 3785 * 3786 * at entry, bitmap of the dmap has been initialized; 3787 * 3788 * PARAMETERS: 3789 * dp - dmap to complete 3790 * blkno - starting block number for this dmap 3791 * treemax - will be filled in with max free for this dmap 3792 * 3793 * RETURNS: max free string at the root of the tree 3794 */ 3795 static int dbInitDmapTree(struct dmap * dp) 3796 { 3797 struct dmaptree *tp; 3798 s8 *cp; 3799 int i; 3800 3801 /* init fixed info of tree */ 3802 tp = &dp->tree; 3803 tp->nleafs = cpu_to_le32(LPERDMAP); 3804 tp->l2nleafs = cpu_to_le32(L2LPERDMAP); 3805 tp->leafidx = cpu_to_le32(LEAFIND); 3806 tp->height = cpu_to_le32(4); 3807 tp->budmin = BUDMIN; 3808 3809 /* init each leaf from corresponding wmap word: 3810 * note: leaf is set to NOFREE(-1) if all blocks of corresponding 3811 * bitmap word are allocated. 3812 */ 3813 cp = tp->stree + le32_to_cpu(tp->leafidx); 3814 for (i = 0; i < LPERDMAP; i++) 3815 *cp++ = dbMaxBud((u8 *) & dp->wmap[i]); 3816 3817 /* build the dmap's binary buddy summary tree */ 3818 return (dbInitTree(tp)); 3819 } 3820 3821 3822 /* 3823 * NAME: dbInitTree()/ujfs_adjtree() 3824 * 3825 * FUNCTION: initialize binary buddy summary tree of a dmap or dmapctl. 3826 * 3827 * at entry, the leaves of the tree has been initialized 3828 * from corresponding bitmap word or root of summary tree 3829 * of the child control page; 3830 * configure binary buddy system at the leaf level, then 3831 * bubble up the values of the leaf nodes up the tree. 3832 * 3833 * PARAMETERS: 3834 * cp - Pointer to the root of the tree 3835 * l2leaves- Number of leaf nodes as a power of 2 3836 * l2min - Number of blocks that can be covered by a leaf 3837 * as a power of 2 3838 * 3839 * RETURNS: max free string at the root of the tree 3840 */ 3841 static int dbInitTree(struct dmaptree * dtp) 3842 { 3843 int l2max, l2free, bsize, nextb, i; 3844 int child, parent, nparent; 3845 s8 *tp, *cp, *cp1; 3846 3847 tp = dtp->stree; 3848 3849 /* Determine the maximum free string possible for the leaves */ 3850 l2max = le32_to_cpu(dtp->l2nleafs) + dtp->budmin; 3851 3852 /* 3853 * configure the leaf levevl into binary buddy system 3854 * 3855 * Try to combine buddies starting with a buddy size of 1 3856 * (i.e. two leaves). At a buddy size of 1 two buddy leaves 3857 * can be combined if both buddies have a maximum free of l2min; 3858 * the combination will result in the left-most buddy leaf having 3859 * a maximum free of l2min+1. 3860 * After processing all buddies for a given size, process buddies 3861 * at the next higher buddy size (i.e. current size * 2) and 3862 * the next maximum free (current free + 1). 3863 * This continues until the maximum possible buddy combination 3864 * yields maximum free. 3865 */ 3866 for (l2free = dtp->budmin, bsize = 1; l2free < l2max; 3867 l2free++, bsize = nextb) { 3868 /* get next buddy size == current buddy pair size */ 3869 nextb = bsize << 1; 3870 3871 /* scan each adjacent buddy pair at current buddy size */ 3872 for (i = 0, cp = tp + le32_to_cpu(dtp->leafidx); 3873 i < le32_to_cpu(dtp->nleafs); 3874 i += nextb, cp += nextb) { 3875 /* coalesce if both adjacent buddies are max free */ 3876 if (*cp == l2free && *(cp + bsize) == l2free) { 3877 *cp = l2free + 1; /* left take right */ 3878 *(cp + bsize) = -1; /* right give left */ 3879 } 3880 } 3881 } 3882 3883 /* 3884 * bubble summary information of leaves up the tree. 3885 * 3886 * Starting at the leaf node level, the four nodes described by 3887 * the higher level parent node are compared for a maximum free and 3888 * this maximum becomes the value of the parent node. 3889 * when all lower level nodes are processed in this fashion then 3890 * move up to the next level (parent becomes a lower level node) and 3891 * continue the process for that level. 3892 */ 3893 for (child = le32_to_cpu(dtp->leafidx), 3894 nparent = le32_to_cpu(dtp->nleafs) >> 2; 3895 nparent > 0; nparent >>= 2, child = parent) { 3896 /* get index of 1st node of parent level */ 3897 parent = (child - 1) >> 2; 3898 3899 /* set the value of the parent node as the maximum 3900 * of the four nodes of the current level. 3901 */ 3902 for (i = 0, cp = tp + child, cp1 = tp + parent; 3903 i < nparent; i++, cp += 4, cp1++) 3904 *cp1 = TREEMAX(cp); 3905 } 3906 3907 return (*tp); 3908 } 3909 3910 3911 /* 3912 * dbInitDmapCtl() 3913 * 3914 * function: initialize dmapctl page 3915 */ 3916 static int dbInitDmapCtl(struct dmapctl * dcp, int level, int i) 3917 { /* start leaf index not covered by range */ 3918 s8 *cp; 3919 3920 dcp->nleafs = cpu_to_le32(LPERCTL); 3921 dcp->l2nleafs = cpu_to_le32(L2LPERCTL); 3922 dcp->leafidx = cpu_to_le32(CTLLEAFIND); 3923 dcp->height = cpu_to_le32(5); 3924 dcp->budmin = L2BPERDMAP + L2LPERCTL * level; 3925 3926 /* 3927 * initialize the leaves of current level that were not covered 3928 * by the specified input block range (i.e. the leaves have no 3929 * low level dmapctl or dmap). 3930 */ 3931 cp = &dcp->stree[CTLLEAFIND + i]; 3932 for (; i < LPERCTL; i++) 3933 *cp++ = NOFREE; 3934 3935 /* build the dmap's binary buddy summary tree */ 3936 return (dbInitTree((struct dmaptree *) dcp)); 3937 } 3938 3939 3940 /* 3941 * NAME: dbGetL2AGSize()/ujfs_getagl2size() 3942 * 3943 * FUNCTION: Determine log2(allocation group size) from aggregate size 3944 * 3945 * PARAMETERS: 3946 * nblocks - Number of blocks in aggregate 3947 * 3948 * RETURNS: log2(allocation group size) in aggregate blocks 3949 */ 3950 static int dbGetL2AGSize(s64 nblocks) 3951 { 3952 s64 sz; 3953 s64 m; 3954 int l2sz; 3955 3956 if (nblocks < BPERDMAP * MAXAG) 3957 return (L2BPERDMAP); 3958 3959 /* round up aggregate size to power of 2 */ 3960 m = ((u64) 1 << (64 - 1)); 3961 for (l2sz = 64; l2sz >= 0; l2sz--, m >>= 1) { 3962 if (m & nblocks) 3963 break; 3964 } 3965 3966 sz = (s64) 1 << l2sz; 3967 if (sz < nblocks) 3968 l2sz += 1; 3969 3970 /* agsize = roundupSize/max_number_of_ag */ 3971 return (l2sz - L2MAXAG); 3972 } 3973 3974 3975 /* 3976 * NAME: dbMapFileSizeToMapSize() 3977 * 3978 * FUNCTION: compute number of blocks the block allocation map file 3979 * can cover from the map file size; 3980 * 3981 * RETURNS: Number of blocks which can be covered by this block map file; 3982 */ 3983 3984 /* 3985 * maximum number of map pages at each level including control pages 3986 */ 3987 #define MAXL0PAGES (1 + LPERCTL) 3988 #define MAXL1PAGES (1 + LPERCTL * MAXL0PAGES) 3989 3990 /* 3991 * convert number of map pages to the zero origin top dmapctl level 3992 */ 3993 #define BMAPPGTOLEV(npages) \ 3994 (((npages) <= 3 + MAXL0PAGES) ? 0 : \ 3995 ((npages) <= 2 + MAXL1PAGES) ? 1 : 2) 3996 3997 s64 dbMapFileSizeToMapSize(struct inode * ipbmap) 3998 { 3999 struct super_block *sb = ipbmap->i_sb; 4000 s64 nblocks; 4001 s64 npages, ndmaps; 4002 int level, i; 4003 int complete, factor; 4004 4005 nblocks = ipbmap->i_size >> JFS_SBI(sb)->l2bsize; 4006 npages = nblocks >> JFS_SBI(sb)->l2nbperpage; 4007 level = BMAPPGTOLEV(npages); 4008 4009 /* At each level, accumulate the number of dmap pages covered by 4010 * the number of full child levels below it; 4011 * repeat for the last incomplete child level. 4012 */ 4013 ndmaps = 0; 4014 npages--; /* skip the first global control page */ 4015 /* skip higher level control pages above top level covered by map */ 4016 npages -= (2 - level); 4017 npages--; /* skip top level's control page */ 4018 for (i = level; i >= 0; i--) { 4019 factor = 4020 (i == 2) ? MAXL1PAGES : ((i == 1) ? MAXL0PAGES : 1); 4021 complete = (u32) npages / factor; 4022 ndmaps += complete * ((i == 2) ? LPERCTL * LPERCTL : 4023 ((i == 1) ? LPERCTL : 1)); 4024 4025 /* pages in last/incomplete child */ 4026 npages = (u32) npages % factor; 4027 /* skip incomplete child's level control page */ 4028 npages--; 4029 } 4030 4031 /* convert the number of dmaps into the number of blocks 4032 * which can be covered by the dmaps; 4033 */ 4034 nblocks = ndmaps << L2BPERDMAP; 4035 4036 return (nblocks); 4037 } 4038