1 /* -*- mode: c; c-basic-offset: 8; -*- 2 * vim: noexpandtab sw=8 ts=8 sts=0: 3 * 4 * alloc.c 5 * 6 * Extent allocs and frees 7 * 8 * Copyright (C) 2002, 2004 Oracle. All rights reserved. 9 * 10 * This program is free software; you can redistribute it and/or 11 * modify it under the terms of the GNU General Public 12 * License as published by the Free Software Foundation; either 13 * version 2 of the License, or (at your option) any later version. 14 * 15 * This program is distributed in the hope that it will be useful, 16 * but WITHOUT ANY WARRANTY; without even the implied warranty of 17 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU 18 * General Public License for more details. 19 * 20 * You should have received a copy of the GNU General Public 21 * License along with this program; if not, write to the 22 * Free Software Foundation, Inc., 59 Temple Place - Suite 330, 23 * Boston, MA 021110-1307, USA. 24 */ 25 26 #include <linux/fs.h> 27 #include <linux/types.h> 28 #include <linux/slab.h> 29 #include <linux/highmem.h> 30 #include <linux/swap.h> 31 #include <linux/quotaops.h> 32 33 #define MLOG_MASK_PREFIX ML_DISK_ALLOC 34 #include <cluster/masklog.h> 35 36 #include "ocfs2.h" 37 38 #include "alloc.h" 39 #include "aops.h" 40 #include "blockcheck.h" 41 #include "dlmglue.h" 42 #include "extent_map.h" 43 #include "inode.h" 44 #include "journal.h" 45 #include "localalloc.h" 46 #include "suballoc.h" 47 #include "sysfile.h" 48 #include "file.h" 49 #include "super.h" 50 #include "uptodate.h" 51 #include "xattr.h" 52 53 #include "buffer_head_io.h" 54 55 56 /* 57 * Operations for a specific extent tree type. 58 * 59 * To implement an on-disk btree (extent tree) type in ocfs2, add 60 * an ocfs2_extent_tree_operations structure and the matching 61 * ocfs2_init_<thingy>_extent_tree() function. That's pretty much it 62 * for the allocation portion of the extent tree. 63 */ 64 struct ocfs2_extent_tree_operations { 65 /* 66 * last_eb_blk is the block number of the right most leaf extent 67 * block. Most on-disk structures containing an extent tree store 68 * this value for fast access. The ->eo_set_last_eb_blk() and 69 * ->eo_get_last_eb_blk() operations access this value. They are 70 * both required. 71 */ 72 void (*eo_set_last_eb_blk)(struct ocfs2_extent_tree *et, 73 u64 blkno); 74 u64 (*eo_get_last_eb_blk)(struct ocfs2_extent_tree *et); 75 76 /* 77 * The on-disk structure usually keeps track of how many total 78 * clusters are stored in this extent tree. This function updates 79 * that value. new_clusters is the delta, and must be 80 * added to the total. Required. 81 */ 82 void (*eo_update_clusters)(struct ocfs2_extent_tree *et, 83 u32 new_clusters); 84 85 /* 86 * If ->eo_insert_check() exists, it is called before rec is 87 * inserted into the extent tree. It is optional. 88 */ 89 int (*eo_insert_check)(struct ocfs2_extent_tree *et, 90 struct ocfs2_extent_rec *rec); 91 int (*eo_sanity_check)(struct ocfs2_extent_tree *et); 92 93 /* 94 * -------------------------------------------------------------- 95 * The remaining are internal to ocfs2_extent_tree and don't have 96 * accessor functions 97 */ 98 99 /* 100 * ->eo_fill_root_el() takes et->et_object and sets et->et_root_el. 101 * It is required. 102 */ 103 void (*eo_fill_root_el)(struct ocfs2_extent_tree *et); 104 105 /* 106 * ->eo_fill_max_leaf_clusters sets et->et_max_leaf_clusters if 107 * it exists. If it does not, et->et_max_leaf_clusters is set 108 * to 0 (unlimited). Optional. 109 */ 110 void (*eo_fill_max_leaf_clusters)(struct ocfs2_extent_tree *et); 111 }; 112 113 114 /* 115 * Pre-declare ocfs2_dinode_et_ops so we can use it as a sanity check 116 * in the methods. 117 */ 118 static u64 ocfs2_dinode_get_last_eb_blk(struct ocfs2_extent_tree *et); 119 static void ocfs2_dinode_set_last_eb_blk(struct ocfs2_extent_tree *et, 120 u64 blkno); 121 static void ocfs2_dinode_update_clusters(struct ocfs2_extent_tree *et, 122 u32 clusters); 123 static int ocfs2_dinode_insert_check(struct ocfs2_extent_tree *et, 124 struct ocfs2_extent_rec *rec); 125 static int ocfs2_dinode_sanity_check(struct ocfs2_extent_tree *et); 126 static void ocfs2_dinode_fill_root_el(struct ocfs2_extent_tree *et); 127 static struct ocfs2_extent_tree_operations ocfs2_dinode_et_ops = { 128 .eo_set_last_eb_blk = ocfs2_dinode_set_last_eb_blk, 129 .eo_get_last_eb_blk = ocfs2_dinode_get_last_eb_blk, 130 .eo_update_clusters = ocfs2_dinode_update_clusters, 131 .eo_insert_check = ocfs2_dinode_insert_check, 132 .eo_sanity_check = ocfs2_dinode_sanity_check, 133 .eo_fill_root_el = ocfs2_dinode_fill_root_el, 134 }; 135 136 static void ocfs2_dinode_set_last_eb_blk(struct ocfs2_extent_tree *et, 137 u64 blkno) 138 { 139 struct ocfs2_dinode *di = et->et_object; 140 141 BUG_ON(et->et_ops != &ocfs2_dinode_et_ops); 142 di->i_last_eb_blk = cpu_to_le64(blkno); 143 } 144 145 static u64 ocfs2_dinode_get_last_eb_blk(struct ocfs2_extent_tree *et) 146 { 147 struct ocfs2_dinode *di = et->et_object; 148 149 BUG_ON(et->et_ops != &ocfs2_dinode_et_ops); 150 return le64_to_cpu(di->i_last_eb_blk); 151 } 152 153 static void ocfs2_dinode_update_clusters(struct ocfs2_extent_tree *et, 154 u32 clusters) 155 { 156 struct ocfs2_inode_info *oi = cache_info_to_inode(et->et_ci); 157 struct ocfs2_dinode *di = et->et_object; 158 159 le32_add_cpu(&di->i_clusters, clusters); 160 spin_lock(&oi->ip_lock); 161 oi->ip_clusters = le32_to_cpu(di->i_clusters); 162 spin_unlock(&oi->ip_lock); 163 } 164 165 static int ocfs2_dinode_insert_check(struct ocfs2_extent_tree *et, 166 struct ocfs2_extent_rec *rec) 167 { 168 struct ocfs2_inode_info *oi = cache_info_to_inode(et->et_ci); 169 struct ocfs2_super *osb = OCFS2_SB(oi->vfs_inode.i_sb); 170 171 BUG_ON(oi->ip_dyn_features & OCFS2_INLINE_DATA_FL); 172 mlog_bug_on_msg(!ocfs2_sparse_alloc(osb) && 173 (oi->ip_clusters != le32_to_cpu(rec->e_cpos)), 174 "Device %s, asking for sparse allocation: inode %llu, " 175 "cpos %u, clusters %u\n", 176 osb->dev_str, 177 (unsigned long long)oi->ip_blkno, 178 rec->e_cpos, oi->ip_clusters); 179 180 return 0; 181 } 182 183 static int ocfs2_dinode_sanity_check(struct ocfs2_extent_tree *et) 184 { 185 struct ocfs2_dinode *di = et->et_object; 186 187 BUG_ON(et->et_ops != &ocfs2_dinode_et_ops); 188 BUG_ON(!OCFS2_IS_VALID_DINODE(di)); 189 190 return 0; 191 } 192 193 static void ocfs2_dinode_fill_root_el(struct ocfs2_extent_tree *et) 194 { 195 struct ocfs2_dinode *di = et->et_object; 196 197 et->et_root_el = &di->id2.i_list; 198 } 199 200 201 static void ocfs2_xattr_value_fill_root_el(struct ocfs2_extent_tree *et) 202 { 203 struct ocfs2_xattr_value_buf *vb = et->et_object; 204 205 et->et_root_el = &vb->vb_xv->xr_list; 206 } 207 208 static void ocfs2_xattr_value_set_last_eb_blk(struct ocfs2_extent_tree *et, 209 u64 blkno) 210 { 211 struct ocfs2_xattr_value_buf *vb = et->et_object; 212 213 vb->vb_xv->xr_last_eb_blk = cpu_to_le64(blkno); 214 } 215 216 static u64 ocfs2_xattr_value_get_last_eb_blk(struct ocfs2_extent_tree *et) 217 { 218 struct ocfs2_xattr_value_buf *vb = et->et_object; 219 220 return le64_to_cpu(vb->vb_xv->xr_last_eb_blk); 221 } 222 223 static void ocfs2_xattr_value_update_clusters(struct ocfs2_extent_tree *et, 224 u32 clusters) 225 { 226 struct ocfs2_xattr_value_buf *vb = et->et_object; 227 228 le32_add_cpu(&vb->vb_xv->xr_clusters, clusters); 229 } 230 231 static struct ocfs2_extent_tree_operations ocfs2_xattr_value_et_ops = { 232 .eo_set_last_eb_blk = ocfs2_xattr_value_set_last_eb_blk, 233 .eo_get_last_eb_blk = ocfs2_xattr_value_get_last_eb_blk, 234 .eo_update_clusters = ocfs2_xattr_value_update_clusters, 235 .eo_fill_root_el = ocfs2_xattr_value_fill_root_el, 236 }; 237 238 static void ocfs2_xattr_tree_fill_root_el(struct ocfs2_extent_tree *et) 239 { 240 struct ocfs2_xattr_block *xb = et->et_object; 241 242 et->et_root_el = &xb->xb_attrs.xb_root.xt_list; 243 } 244 245 static void ocfs2_xattr_tree_fill_max_leaf_clusters(struct ocfs2_extent_tree *et) 246 { 247 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci); 248 et->et_max_leaf_clusters = 249 ocfs2_clusters_for_bytes(sb, OCFS2_MAX_XATTR_TREE_LEAF_SIZE); 250 } 251 252 static void ocfs2_xattr_tree_set_last_eb_blk(struct ocfs2_extent_tree *et, 253 u64 blkno) 254 { 255 struct ocfs2_xattr_block *xb = et->et_object; 256 struct ocfs2_xattr_tree_root *xt = &xb->xb_attrs.xb_root; 257 258 xt->xt_last_eb_blk = cpu_to_le64(blkno); 259 } 260 261 static u64 ocfs2_xattr_tree_get_last_eb_blk(struct ocfs2_extent_tree *et) 262 { 263 struct ocfs2_xattr_block *xb = et->et_object; 264 struct ocfs2_xattr_tree_root *xt = &xb->xb_attrs.xb_root; 265 266 return le64_to_cpu(xt->xt_last_eb_blk); 267 } 268 269 static void ocfs2_xattr_tree_update_clusters(struct ocfs2_extent_tree *et, 270 u32 clusters) 271 { 272 struct ocfs2_xattr_block *xb = et->et_object; 273 274 le32_add_cpu(&xb->xb_attrs.xb_root.xt_clusters, clusters); 275 } 276 277 static struct ocfs2_extent_tree_operations ocfs2_xattr_tree_et_ops = { 278 .eo_set_last_eb_blk = ocfs2_xattr_tree_set_last_eb_blk, 279 .eo_get_last_eb_blk = ocfs2_xattr_tree_get_last_eb_blk, 280 .eo_update_clusters = ocfs2_xattr_tree_update_clusters, 281 .eo_fill_root_el = ocfs2_xattr_tree_fill_root_el, 282 .eo_fill_max_leaf_clusters = ocfs2_xattr_tree_fill_max_leaf_clusters, 283 }; 284 285 static void ocfs2_dx_root_set_last_eb_blk(struct ocfs2_extent_tree *et, 286 u64 blkno) 287 { 288 struct ocfs2_dx_root_block *dx_root = et->et_object; 289 290 dx_root->dr_last_eb_blk = cpu_to_le64(blkno); 291 } 292 293 static u64 ocfs2_dx_root_get_last_eb_blk(struct ocfs2_extent_tree *et) 294 { 295 struct ocfs2_dx_root_block *dx_root = et->et_object; 296 297 return le64_to_cpu(dx_root->dr_last_eb_blk); 298 } 299 300 static void ocfs2_dx_root_update_clusters(struct ocfs2_extent_tree *et, 301 u32 clusters) 302 { 303 struct ocfs2_dx_root_block *dx_root = et->et_object; 304 305 le32_add_cpu(&dx_root->dr_clusters, clusters); 306 } 307 308 static int ocfs2_dx_root_sanity_check(struct ocfs2_extent_tree *et) 309 { 310 struct ocfs2_dx_root_block *dx_root = et->et_object; 311 312 BUG_ON(!OCFS2_IS_VALID_DX_ROOT(dx_root)); 313 314 return 0; 315 } 316 317 static void ocfs2_dx_root_fill_root_el(struct ocfs2_extent_tree *et) 318 { 319 struct ocfs2_dx_root_block *dx_root = et->et_object; 320 321 et->et_root_el = &dx_root->dr_list; 322 } 323 324 static struct ocfs2_extent_tree_operations ocfs2_dx_root_et_ops = { 325 .eo_set_last_eb_blk = ocfs2_dx_root_set_last_eb_blk, 326 .eo_get_last_eb_blk = ocfs2_dx_root_get_last_eb_blk, 327 .eo_update_clusters = ocfs2_dx_root_update_clusters, 328 .eo_sanity_check = ocfs2_dx_root_sanity_check, 329 .eo_fill_root_el = ocfs2_dx_root_fill_root_el, 330 }; 331 332 static void __ocfs2_init_extent_tree(struct ocfs2_extent_tree *et, 333 struct inode *inode, 334 struct buffer_head *bh, 335 ocfs2_journal_access_func access, 336 void *obj, 337 struct ocfs2_extent_tree_operations *ops) 338 { 339 et->et_ops = ops; 340 et->et_root_bh = bh; 341 et->et_ci = INODE_CACHE(inode); 342 et->et_root_journal_access = access; 343 if (!obj) 344 obj = (void *)bh->b_data; 345 et->et_object = obj; 346 347 et->et_ops->eo_fill_root_el(et); 348 if (!et->et_ops->eo_fill_max_leaf_clusters) 349 et->et_max_leaf_clusters = 0; 350 else 351 et->et_ops->eo_fill_max_leaf_clusters(et); 352 } 353 354 void ocfs2_init_dinode_extent_tree(struct ocfs2_extent_tree *et, 355 struct inode *inode, 356 struct buffer_head *bh) 357 { 358 __ocfs2_init_extent_tree(et, inode, bh, ocfs2_journal_access_di, 359 NULL, &ocfs2_dinode_et_ops); 360 } 361 362 void ocfs2_init_xattr_tree_extent_tree(struct ocfs2_extent_tree *et, 363 struct inode *inode, 364 struct buffer_head *bh) 365 { 366 __ocfs2_init_extent_tree(et, inode, bh, ocfs2_journal_access_xb, 367 NULL, &ocfs2_xattr_tree_et_ops); 368 } 369 370 void ocfs2_init_xattr_value_extent_tree(struct ocfs2_extent_tree *et, 371 struct inode *inode, 372 struct ocfs2_xattr_value_buf *vb) 373 { 374 __ocfs2_init_extent_tree(et, inode, vb->vb_bh, vb->vb_access, vb, 375 &ocfs2_xattr_value_et_ops); 376 } 377 378 void ocfs2_init_dx_root_extent_tree(struct ocfs2_extent_tree *et, 379 struct inode *inode, 380 struct buffer_head *bh) 381 { 382 __ocfs2_init_extent_tree(et, inode, bh, ocfs2_journal_access_dr, 383 NULL, &ocfs2_dx_root_et_ops); 384 } 385 386 static inline void ocfs2_et_set_last_eb_blk(struct ocfs2_extent_tree *et, 387 u64 new_last_eb_blk) 388 { 389 et->et_ops->eo_set_last_eb_blk(et, new_last_eb_blk); 390 } 391 392 static inline u64 ocfs2_et_get_last_eb_blk(struct ocfs2_extent_tree *et) 393 { 394 return et->et_ops->eo_get_last_eb_blk(et); 395 } 396 397 static inline void ocfs2_et_update_clusters(struct ocfs2_extent_tree *et, 398 u32 clusters) 399 { 400 et->et_ops->eo_update_clusters(et, clusters); 401 } 402 403 static inline int ocfs2_et_root_journal_access(handle_t *handle, 404 struct ocfs2_extent_tree *et, 405 int type) 406 { 407 return et->et_root_journal_access(handle, et->et_ci, et->et_root_bh, 408 type); 409 } 410 411 static inline int ocfs2_et_insert_check(struct ocfs2_extent_tree *et, 412 struct ocfs2_extent_rec *rec) 413 { 414 int ret = 0; 415 416 if (et->et_ops->eo_insert_check) 417 ret = et->et_ops->eo_insert_check(et, rec); 418 return ret; 419 } 420 421 static inline int ocfs2_et_sanity_check(struct ocfs2_extent_tree *et) 422 { 423 int ret = 0; 424 425 if (et->et_ops->eo_sanity_check) 426 ret = et->et_ops->eo_sanity_check(et); 427 return ret; 428 } 429 430 static void ocfs2_free_truncate_context(struct ocfs2_truncate_context *tc); 431 static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt *ctxt, 432 struct ocfs2_extent_block *eb); 433 434 /* 435 * Structures which describe a path through a btree, and functions to 436 * manipulate them. 437 * 438 * The idea here is to be as generic as possible with the tree 439 * manipulation code. 440 */ 441 struct ocfs2_path_item { 442 struct buffer_head *bh; 443 struct ocfs2_extent_list *el; 444 }; 445 446 #define OCFS2_MAX_PATH_DEPTH 5 447 448 struct ocfs2_path { 449 int p_tree_depth; 450 ocfs2_journal_access_func p_root_access; 451 struct ocfs2_path_item p_node[OCFS2_MAX_PATH_DEPTH]; 452 }; 453 454 #define path_root_bh(_path) ((_path)->p_node[0].bh) 455 #define path_root_el(_path) ((_path)->p_node[0].el) 456 #define path_root_access(_path)((_path)->p_root_access) 457 #define path_leaf_bh(_path) ((_path)->p_node[(_path)->p_tree_depth].bh) 458 #define path_leaf_el(_path) ((_path)->p_node[(_path)->p_tree_depth].el) 459 #define path_num_items(_path) ((_path)->p_tree_depth + 1) 460 461 static int ocfs2_find_path(struct ocfs2_caching_info *ci, 462 struct ocfs2_path *path, u32 cpos); 463 static void ocfs2_adjust_rightmost_records(struct inode *inode, 464 handle_t *handle, 465 struct ocfs2_path *path, 466 struct ocfs2_extent_rec *insert_rec); 467 /* 468 * Reset the actual path elements so that we can re-use the structure 469 * to build another path. Generally, this involves freeing the buffer 470 * heads. 471 */ 472 static void ocfs2_reinit_path(struct ocfs2_path *path, int keep_root) 473 { 474 int i, start = 0, depth = 0; 475 struct ocfs2_path_item *node; 476 477 if (keep_root) 478 start = 1; 479 480 for(i = start; i < path_num_items(path); i++) { 481 node = &path->p_node[i]; 482 483 brelse(node->bh); 484 node->bh = NULL; 485 node->el = NULL; 486 } 487 488 /* 489 * Tree depth may change during truncate, or insert. If we're 490 * keeping the root extent list, then make sure that our path 491 * structure reflects the proper depth. 492 */ 493 if (keep_root) 494 depth = le16_to_cpu(path_root_el(path)->l_tree_depth); 495 else 496 path_root_access(path) = NULL; 497 498 path->p_tree_depth = depth; 499 } 500 501 static void ocfs2_free_path(struct ocfs2_path *path) 502 { 503 if (path) { 504 ocfs2_reinit_path(path, 0); 505 kfree(path); 506 } 507 } 508 509 /* 510 * All the elements of src into dest. After this call, src could be freed 511 * without affecting dest. 512 * 513 * Both paths should have the same root. Any non-root elements of dest 514 * will be freed. 515 */ 516 static void ocfs2_cp_path(struct ocfs2_path *dest, struct ocfs2_path *src) 517 { 518 int i; 519 520 BUG_ON(path_root_bh(dest) != path_root_bh(src)); 521 BUG_ON(path_root_el(dest) != path_root_el(src)); 522 BUG_ON(path_root_access(dest) != path_root_access(src)); 523 524 ocfs2_reinit_path(dest, 1); 525 526 for(i = 1; i < OCFS2_MAX_PATH_DEPTH; i++) { 527 dest->p_node[i].bh = src->p_node[i].bh; 528 dest->p_node[i].el = src->p_node[i].el; 529 530 if (dest->p_node[i].bh) 531 get_bh(dest->p_node[i].bh); 532 } 533 } 534 535 /* 536 * Make the *dest path the same as src and re-initialize src path to 537 * have a root only. 538 */ 539 static void ocfs2_mv_path(struct ocfs2_path *dest, struct ocfs2_path *src) 540 { 541 int i; 542 543 BUG_ON(path_root_bh(dest) != path_root_bh(src)); 544 BUG_ON(path_root_access(dest) != path_root_access(src)); 545 546 for(i = 1; i < OCFS2_MAX_PATH_DEPTH; i++) { 547 brelse(dest->p_node[i].bh); 548 549 dest->p_node[i].bh = src->p_node[i].bh; 550 dest->p_node[i].el = src->p_node[i].el; 551 552 src->p_node[i].bh = NULL; 553 src->p_node[i].el = NULL; 554 } 555 } 556 557 /* 558 * Insert an extent block at given index. 559 * 560 * This will not take an additional reference on eb_bh. 561 */ 562 static inline void ocfs2_path_insert_eb(struct ocfs2_path *path, int index, 563 struct buffer_head *eb_bh) 564 { 565 struct ocfs2_extent_block *eb = (struct ocfs2_extent_block *)eb_bh->b_data; 566 567 /* 568 * Right now, no root bh is an extent block, so this helps 569 * catch code errors with dinode trees. The assertion can be 570 * safely removed if we ever need to insert extent block 571 * structures at the root. 572 */ 573 BUG_ON(index == 0); 574 575 path->p_node[index].bh = eb_bh; 576 path->p_node[index].el = &eb->h_list; 577 } 578 579 static struct ocfs2_path *ocfs2_new_path(struct buffer_head *root_bh, 580 struct ocfs2_extent_list *root_el, 581 ocfs2_journal_access_func access) 582 { 583 struct ocfs2_path *path; 584 585 BUG_ON(le16_to_cpu(root_el->l_tree_depth) >= OCFS2_MAX_PATH_DEPTH); 586 587 path = kzalloc(sizeof(*path), GFP_NOFS); 588 if (path) { 589 path->p_tree_depth = le16_to_cpu(root_el->l_tree_depth); 590 get_bh(root_bh); 591 path_root_bh(path) = root_bh; 592 path_root_el(path) = root_el; 593 path_root_access(path) = access; 594 } 595 596 return path; 597 } 598 599 static struct ocfs2_path *ocfs2_new_path_from_path(struct ocfs2_path *path) 600 { 601 return ocfs2_new_path(path_root_bh(path), path_root_el(path), 602 path_root_access(path)); 603 } 604 605 static struct ocfs2_path *ocfs2_new_path_from_et(struct ocfs2_extent_tree *et) 606 { 607 return ocfs2_new_path(et->et_root_bh, et->et_root_el, 608 et->et_root_journal_access); 609 } 610 611 /* 612 * Journal the buffer at depth idx. All idx>0 are extent_blocks, 613 * otherwise it's the root_access function. 614 * 615 * I don't like the way this function's name looks next to 616 * ocfs2_journal_access_path(), but I don't have a better one. 617 */ 618 static int ocfs2_path_bh_journal_access(handle_t *handle, 619 struct ocfs2_caching_info *ci, 620 struct ocfs2_path *path, 621 int idx) 622 { 623 ocfs2_journal_access_func access = path_root_access(path); 624 625 if (!access) 626 access = ocfs2_journal_access; 627 628 if (idx) 629 access = ocfs2_journal_access_eb; 630 631 return access(handle, ci, path->p_node[idx].bh, 632 OCFS2_JOURNAL_ACCESS_WRITE); 633 } 634 635 /* 636 * Convenience function to journal all components in a path. 637 */ 638 static int ocfs2_journal_access_path(struct ocfs2_caching_info *ci, 639 handle_t *handle, 640 struct ocfs2_path *path) 641 { 642 int i, ret = 0; 643 644 if (!path) 645 goto out; 646 647 for(i = 0; i < path_num_items(path); i++) { 648 ret = ocfs2_path_bh_journal_access(handle, ci, path, i); 649 if (ret < 0) { 650 mlog_errno(ret); 651 goto out; 652 } 653 } 654 655 out: 656 return ret; 657 } 658 659 /* 660 * Return the index of the extent record which contains cluster #v_cluster. 661 * -1 is returned if it was not found. 662 * 663 * Should work fine on interior and exterior nodes. 664 */ 665 int ocfs2_search_extent_list(struct ocfs2_extent_list *el, u32 v_cluster) 666 { 667 int ret = -1; 668 int i; 669 struct ocfs2_extent_rec *rec; 670 u32 rec_end, rec_start, clusters; 671 672 for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) { 673 rec = &el->l_recs[i]; 674 675 rec_start = le32_to_cpu(rec->e_cpos); 676 clusters = ocfs2_rec_clusters(el, rec); 677 678 rec_end = rec_start + clusters; 679 680 if (v_cluster >= rec_start && v_cluster < rec_end) { 681 ret = i; 682 break; 683 } 684 } 685 686 return ret; 687 } 688 689 enum ocfs2_contig_type { 690 CONTIG_NONE = 0, 691 CONTIG_LEFT, 692 CONTIG_RIGHT, 693 CONTIG_LEFTRIGHT, 694 }; 695 696 697 /* 698 * NOTE: ocfs2_block_extent_contig(), ocfs2_extents_adjacent() and 699 * ocfs2_extent_contig only work properly against leaf nodes! 700 */ 701 static int ocfs2_block_extent_contig(struct super_block *sb, 702 struct ocfs2_extent_rec *ext, 703 u64 blkno) 704 { 705 u64 blk_end = le64_to_cpu(ext->e_blkno); 706 707 blk_end += ocfs2_clusters_to_blocks(sb, 708 le16_to_cpu(ext->e_leaf_clusters)); 709 710 return blkno == blk_end; 711 } 712 713 static int ocfs2_extents_adjacent(struct ocfs2_extent_rec *left, 714 struct ocfs2_extent_rec *right) 715 { 716 u32 left_range; 717 718 left_range = le32_to_cpu(left->e_cpos) + 719 le16_to_cpu(left->e_leaf_clusters); 720 721 return (left_range == le32_to_cpu(right->e_cpos)); 722 } 723 724 static enum ocfs2_contig_type 725 ocfs2_extent_contig(struct inode *inode, 726 struct ocfs2_extent_rec *ext, 727 struct ocfs2_extent_rec *insert_rec) 728 { 729 u64 blkno = le64_to_cpu(insert_rec->e_blkno); 730 731 /* 732 * Refuse to coalesce extent records with different flag 733 * fields - we don't want to mix unwritten extents with user 734 * data. 735 */ 736 if (ext->e_flags != insert_rec->e_flags) 737 return CONTIG_NONE; 738 739 if (ocfs2_extents_adjacent(ext, insert_rec) && 740 ocfs2_block_extent_contig(inode->i_sb, ext, blkno)) 741 return CONTIG_RIGHT; 742 743 blkno = le64_to_cpu(ext->e_blkno); 744 if (ocfs2_extents_adjacent(insert_rec, ext) && 745 ocfs2_block_extent_contig(inode->i_sb, insert_rec, blkno)) 746 return CONTIG_LEFT; 747 748 return CONTIG_NONE; 749 } 750 751 /* 752 * NOTE: We can have pretty much any combination of contiguousness and 753 * appending. 754 * 755 * The usefulness of APPEND_TAIL is more in that it lets us know that 756 * we'll have to update the path to that leaf. 757 */ 758 enum ocfs2_append_type { 759 APPEND_NONE = 0, 760 APPEND_TAIL, 761 }; 762 763 enum ocfs2_split_type { 764 SPLIT_NONE = 0, 765 SPLIT_LEFT, 766 SPLIT_RIGHT, 767 }; 768 769 struct ocfs2_insert_type { 770 enum ocfs2_split_type ins_split; 771 enum ocfs2_append_type ins_appending; 772 enum ocfs2_contig_type ins_contig; 773 int ins_contig_index; 774 int ins_tree_depth; 775 }; 776 777 struct ocfs2_merge_ctxt { 778 enum ocfs2_contig_type c_contig_type; 779 int c_has_empty_extent; 780 int c_split_covers_rec; 781 }; 782 783 static int ocfs2_validate_extent_block(struct super_block *sb, 784 struct buffer_head *bh) 785 { 786 int rc; 787 struct ocfs2_extent_block *eb = 788 (struct ocfs2_extent_block *)bh->b_data; 789 790 mlog(0, "Validating extent block %llu\n", 791 (unsigned long long)bh->b_blocknr); 792 793 BUG_ON(!buffer_uptodate(bh)); 794 795 /* 796 * If the ecc fails, we return the error but otherwise 797 * leave the filesystem running. We know any error is 798 * local to this block. 799 */ 800 rc = ocfs2_validate_meta_ecc(sb, bh->b_data, &eb->h_check); 801 if (rc) { 802 mlog(ML_ERROR, "Checksum failed for extent block %llu\n", 803 (unsigned long long)bh->b_blocknr); 804 return rc; 805 } 806 807 /* 808 * Errors after here are fatal. 809 */ 810 811 if (!OCFS2_IS_VALID_EXTENT_BLOCK(eb)) { 812 ocfs2_error(sb, 813 "Extent block #%llu has bad signature %.*s", 814 (unsigned long long)bh->b_blocknr, 7, 815 eb->h_signature); 816 return -EINVAL; 817 } 818 819 if (le64_to_cpu(eb->h_blkno) != bh->b_blocknr) { 820 ocfs2_error(sb, 821 "Extent block #%llu has an invalid h_blkno " 822 "of %llu", 823 (unsigned long long)bh->b_blocknr, 824 (unsigned long long)le64_to_cpu(eb->h_blkno)); 825 return -EINVAL; 826 } 827 828 if (le32_to_cpu(eb->h_fs_generation) != OCFS2_SB(sb)->fs_generation) { 829 ocfs2_error(sb, 830 "Extent block #%llu has an invalid " 831 "h_fs_generation of #%u", 832 (unsigned long long)bh->b_blocknr, 833 le32_to_cpu(eb->h_fs_generation)); 834 return -EINVAL; 835 } 836 837 return 0; 838 } 839 840 int ocfs2_read_extent_block(struct ocfs2_caching_info *ci, u64 eb_blkno, 841 struct buffer_head **bh) 842 { 843 int rc; 844 struct buffer_head *tmp = *bh; 845 846 rc = ocfs2_read_block(ci, eb_blkno, &tmp, 847 ocfs2_validate_extent_block); 848 849 /* If ocfs2_read_block() got us a new bh, pass it up. */ 850 if (!rc && !*bh) 851 *bh = tmp; 852 853 return rc; 854 } 855 856 857 /* 858 * How many free extents have we got before we need more meta data? 859 */ 860 int ocfs2_num_free_extents(struct ocfs2_super *osb, 861 struct ocfs2_extent_tree *et) 862 { 863 int retval; 864 struct ocfs2_extent_list *el = NULL; 865 struct ocfs2_extent_block *eb; 866 struct buffer_head *eb_bh = NULL; 867 u64 last_eb_blk = 0; 868 869 mlog_entry_void(); 870 871 el = et->et_root_el; 872 last_eb_blk = ocfs2_et_get_last_eb_blk(et); 873 874 if (last_eb_blk) { 875 retval = ocfs2_read_extent_block(et->et_ci, last_eb_blk, 876 &eb_bh); 877 if (retval < 0) { 878 mlog_errno(retval); 879 goto bail; 880 } 881 eb = (struct ocfs2_extent_block *) eb_bh->b_data; 882 el = &eb->h_list; 883 } 884 885 BUG_ON(el->l_tree_depth != 0); 886 887 retval = le16_to_cpu(el->l_count) - le16_to_cpu(el->l_next_free_rec); 888 bail: 889 brelse(eb_bh); 890 891 mlog_exit(retval); 892 return retval; 893 } 894 895 /* expects array to already be allocated 896 * 897 * sets h_signature, h_blkno, h_suballoc_bit, h_suballoc_slot, and 898 * l_count for you 899 */ 900 static int ocfs2_create_new_meta_bhs(handle_t *handle, 901 struct ocfs2_extent_tree *et, 902 int wanted, 903 struct ocfs2_alloc_context *meta_ac, 904 struct buffer_head *bhs[]) 905 { 906 int count, status, i; 907 u16 suballoc_bit_start; 908 u32 num_got; 909 u64 first_blkno; 910 struct ocfs2_super *osb = 911 OCFS2_SB(ocfs2_metadata_cache_get_super(et->et_ci)); 912 struct ocfs2_extent_block *eb; 913 914 mlog_entry_void(); 915 916 count = 0; 917 while (count < wanted) { 918 status = ocfs2_claim_metadata(osb, 919 handle, 920 meta_ac, 921 wanted - count, 922 &suballoc_bit_start, 923 &num_got, 924 &first_blkno); 925 if (status < 0) { 926 mlog_errno(status); 927 goto bail; 928 } 929 930 for(i = count; i < (num_got + count); i++) { 931 bhs[i] = sb_getblk(osb->sb, first_blkno); 932 if (bhs[i] == NULL) { 933 status = -EIO; 934 mlog_errno(status); 935 goto bail; 936 } 937 ocfs2_set_new_buffer_uptodate(et->et_ci, bhs[i]); 938 939 status = ocfs2_journal_access_eb(handle, et->et_ci, 940 bhs[i], 941 OCFS2_JOURNAL_ACCESS_CREATE); 942 if (status < 0) { 943 mlog_errno(status); 944 goto bail; 945 } 946 947 memset(bhs[i]->b_data, 0, osb->sb->s_blocksize); 948 eb = (struct ocfs2_extent_block *) bhs[i]->b_data; 949 /* Ok, setup the minimal stuff here. */ 950 strcpy(eb->h_signature, OCFS2_EXTENT_BLOCK_SIGNATURE); 951 eb->h_blkno = cpu_to_le64(first_blkno); 952 eb->h_fs_generation = cpu_to_le32(osb->fs_generation); 953 eb->h_suballoc_slot = cpu_to_le16(osb->slot_num); 954 eb->h_suballoc_bit = cpu_to_le16(suballoc_bit_start); 955 eb->h_list.l_count = 956 cpu_to_le16(ocfs2_extent_recs_per_eb(osb->sb)); 957 958 suballoc_bit_start++; 959 first_blkno++; 960 961 /* We'll also be dirtied by the caller, so 962 * this isn't absolutely necessary. */ 963 status = ocfs2_journal_dirty(handle, bhs[i]); 964 if (status < 0) { 965 mlog_errno(status); 966 goto bail; 967 } 968 } 969 970 count += num_got; 971 } 972 973 status = 0; 974 bail: 975 if (status < 0) { 976 for(i = 0; i < wanted; i++) { 977 brelse(bhs[i]); 978 bhs[i] = NULL; 979 } 980 } 981 mlog_exit(status); 982 return status; 983 } 984 985 /* 986 * Helper function for ocfs2_add_branch() and ocfs2_shift_tree_depth(). 987 * 988 * Returns the sum of the rightmost extent rec logical offset and 989 * cluster count. 990 * 991 * ocfs2_add_branch() uses this to determine what logical cluster 992 * value should be populated into the leftmost new branch records. 993 * 994 * ocfs2_shift_tree_depth() uses this to determine the # clusters 995 * value for the new topmost tree record. 996 */ 997 static inline u32 ocfs2_sum_rightmost_rec(struct ocfs2_extent_list *el) 998 { 999 int i; 1000 1001 i = le16_to_cpu(el->l_next_free_rec) - 1; 1002 1003 return le32_to_cpu(el->l_recs[i].e_cpos) + 1004 ocfs2_rec_clusters(el, &el->l_recs[i]); 1005 } 1006 1007 /* 1008 * Change range of the branches in the right most path according to the leaf 1009 * extent block's rightmost record. 1010 */ 1011 static int ocfs2_adjust_rightmost_branch(handle_t *handle, 1012 struct inode *inode, 1013 struct ocfs2_extent_tree *et) 1014 { 1015 int status; 1016 struct ocfs2_path *path = NULL; 1017 struct ocfs2_extent_list *el; 1018 struct ocfs2_extent_rec *rec; 1019 1020 path = ocfs2_new_path_from_et(et); 1021 if (!path) { 1022 status = -ENOMEM; 1023 return status; 1024 } 1025 1026 status = ocfs2_find_path(et->et_ci, path, UINT_MAX); 1027 if (status < 0) { 1028 mlog_errno(status); 1029 goto out; 1030 } 1031 1032 status = ocfs2_extend_trans(handle, path_num_items(path) + 1033 handle->h_buffer_credits); 1034 if (status < 0) { 1035 mlog_errno(status); 1036 goto out; 1037 } 1038 1039 status = ocfs2_journal_access_path(INODE_CACHE(inode), handle, path); 1040 if (status < 0) { 1041 mlog_errno(status); 1042 goto out; 1043 } 1044 1045 el = path_leaf_el(path); 1046 rec = &el->l_recs[le32_to_cpu(el->l_next_free_rec) - 1]; 1047 1048 ocfs2_adjust_rightmost_records(inode, handle, path, rec); 1049 1050 out: 1051 ocfs2_free_path(path); 1052 return status; 1053 } 1054 1055 /* 1056 * Add an entire tree branch to our inode. eb_bh is the extent block 1057 * to start at, if we don't want to start the branch at the dinode 1058 * structure. 1059 * 1060 * last_eb_bh is required as we have to update it's next_leaf pointer 1061 * for the new last extent block. 1062 * 1063 * the new branch will be 'empty' in the sense that every block will 1064 * contain a single record with cluster count == 0. 1065 */ 1066 static int ocfs2_add_branch(struct ocfs2_super *osb, 1067 handle_t *handle, 1068 struct inode *inode, 1069 struct ocfs2_extent_tree *et, 1070 struct buffer_head *eb_bh, 1071 struct buffer_head **last_eb_bh, 1072 struct ocfs2_alloc_context *meta_ac) 1073 { 1074 int status, new_blocks, i; 1075 u64 next_blkno, new_last_eb_blk; 1076 struct buffer_head *bh; 1077 struct buffer_head **new_eb_bhs = NULL; 1078 struct ocfs2_extent_block *eb; 1079 struct ocfs2_extent_list *eb_el; 1080 struct ocfs2_extent_list *el; 1081 u32 new_cpos, root_end; 1082 1083 mlog_entry_void(); 1084 1085 BUG_ON(!last_eb_bh || !*last_eb_bh); 1086 1087 if (eb_bh) { 1088 eb = (struct ocfs2_extent_block *) eb_bh->b_data; 1089 el = &eb->h_list; 1090 } else 1091 el = et->et_root_el; 1092 1093 /* we never add a branch to a leaf. */ 1094 BUG_ON(!el->l_tree_depth); 1095 1096 new_blocks = le16_to_cpu(el->l_tree_depth); 1097 1098 eb = (struct ocfs2_extent_block *)(*last_eb_bh)->b_data; 1099 new_cpos = ocfs2_sum_rightmost_rec(&eb->h_list); 1100 root_end = ocfs2_sum_rightmost_rec(et->et_root_el); 1101 1102 /* 1103 * If there is a gap before the root end and the real end 1104 * of the righmost leaf block, we need to remove the gap 1105 * between new_cpos and root_end first so that the tree 1106 * is consistent after we add a new branch(it will start 1107 * from new_cpos). 1108 */ 1109 if (root_end > new_cpos) { 1110 mlog(0, "adjust the cluster end from %u to %u\n", 1111 root_end, new_cpos); 1112 status = ocfs2_adjust_rightmost_branch(handle, inode, et); 1113 if (status) { 1114 mlog_errno(status); 1115 goto bail; 1116 } 1117 } 1118 1119 /* allocate the number of new eb blocks we need */ 1120 new_eb_bhs = kcalloc(new_blocks, sizeof(struct buffer_head *), 1121 GFP_KERNEL); 1122 if (!new_eb_bhs) { 1123 status = -ENOMEM; 1124 mlog_errno(status); 1125 goto bail; 1126 } 1127 1128 status = ocfs2_create_new_meta_bhs(handle, et, new_blocks, 1129 meta_ac, new_eb_bhs); 1130 if (status < 0) { 1131 mlog_errno(status); 1132 goto bail; 1133 } 1134 1135 /* Note: new_eb_bhs[new_blocks - 1] is the guy which will be 1136 * linked with the rest of the tree. 1137 * conversly, new_eb_bhs[0] is the new bottommost leaf. 1138 * 1139 * when we leave the loop, new_last_eb_blk will point to the 1140 * newest leaf, and next_blkno will point to the topmost extent 1141 * block. */ 1142 next_blkno = new_last_eb_blk = 0; 1143 for(i = 0; i < new_blocks; i++) { 1144 bh = new_eb_bhs[i]; 1145 eb = (struct ocfs2_extent_block *) bh->b_data; 1146 /* ocfs2_create_new_meta_bhs() should create it right! */ 1147 BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb)); 1148 eb_el = &eb->h_list; 1149 1150 status = ocfs2_journal_access_eb(handle, INODE_CACHE(inode), bh, 1151 OCFS2_JOURNAL_ACCESS_CREATE); 1152 if (status < 0) { 1153 mlog_errno(status); 1154 goto bail; 1155 } 1156 1157 eb->h_next_leaf_blk = 0; 1158 eb_el->l_tree_depth = cpu_to_le16(i); 1159 eb_el->l_next_free_rec = cpu_to_le16(1); 1160 /* 1161 * This actually counts as an empty extent as 1162 * c_clusters == 0 1163 */ 1164 eb_el->l_recs[0].e_cpos = cpu_to_le32(new_cpos); 1165 eb_el->l_recs[0].e_blkno = cpu_to_le64(next_blkno); 1166 /* 1167 * eb_el isn't always an interior node, but even leaf 1168 * nodes want a zero'd flags and reserved field so 1169 * this gets the whole 32 bits regardless of use. 1170 */ 1171 eb_el->l_recs[0].e_int_clusters = cpu_to_le32(0); 1172 if (!eb_el->l_tree_depth) 1173 new_last_eb_blk = le64_to_cpu(eb->h_blkno); 1174 1175 status = ocfs2_journal_dirty(handle, bh); 1176 if (status < 0) { 1177 mlog_errno(status); 1178 goto bail; 1179 } 1180 1181 next_blkno = le64_to_cpu(eb->h_blkno); 1182 } 1183 1184 /* This is a bit hairy. We want to update up to three blocks 1185 * here without leaving any of them in an inconsistent state 1186 * in case of error. We don't have to worry about 1187 * journal_dirty erroring as it won't unless we've aborted the 1188 * handle (in which case we would never be here) so reserving 1189 * the write with journal_access is all we need to do. */ 1190 status = ocfs2_journal_access_eb(handle, INODE_CACHE(inode), *last_eb_bh, 1191 OCFS2_JOURNAL_ACCESS_WRITE); 1192 if (status < 0) { 1193 mlog_errno(status); 1194 goto bail; 1195 } 1196 status = ocfs2_et_root_journal_access(handle, et, 1197 OCFS2_JOURNAL_ACCESS_WRITE); 1198 if (status < 0) { 1199 mlog_errno(status); 1200 goto bail; 1201 } 1202 if (eb_bh) { 1203 status = ocfs2_journal_access_eb(handle, INODE_CACHE(inode), eb_bh, 1204 OCFS2_JOURNAL_ACCESS_WRITE); 1205 if (status < 0) { 1206 mlog_errno(status); 1207 goto bail; 1208 } 1209 } 1210 1211 /* Link the new branch into the rest of the tree (el will 1212 * either be on the root_bh, or the extent block passed in. */ 1213 i = le16_to_cpu(el->l_next_free_rec); 1214 el->l_recs[i].e_blkno = cpu_to_le64(next_blkno); 1215 el->l_recs[i].e_cpos = cpu_to_le32(new_cpos); 1216 el->l_recs[i].e_int_clusters = 0; 1217 le16_add_cpu(&el->l_next_free_rec, 1); 1218 1219 /* fe needs a new last extent block pointer, as does the 1220 * next_leaf on the previously last-extent-block. */ 1221 ocfs2_et_set_last_eb_blk(et, new_last_eb_blk); 1222 1223 eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data; 1224 eb->h_next_leaf_blk = cpu_to_le64(new_last_eb_blk); 1225 1226 status = ocfs2_journal_dirty(handle, *last_eb_bh); 1227 if (status < 0) 1228 mlog_errno(status); 1229 status = ocfs2_journal_dirty(handle, et->et_root_bh); 1230 if (status < 0) 1231 mlog_errno(status); 1232 if (eb_bh) { 1233 status = ocfs2_journal_dirty(handle, eb_bh); 1234 if (status < 0) 1235 mlog_errno(status); 1236 } 1237 1238 /* 1239 * Some callers want to track the rightmost leaf so pass it 1240 * back here. 1241 */ 1242 brelse(*last_eb_bh); 1243 get_bh(new_eb_bhs[0]); 1244 *last_eb_bh = new_eb_bhs[0]; 1245 1246 status = 0; 1247 bail: 1248 if (new_eb_bhs) { 1249 for (i = 0; i < new_blocks; i++) 1250 brelse(new_eb_bhs[i]); 1251 kfree(new_eb_bhs); 1252 } 1253 1254 mlog_exit(status); 1255 return status; 1256 } 1257 1258 /* 1259 * adds another level to the allocation tree. 1260 * returns back the new extent block so you can add a branch to it 1261 * after this call. 1262 */ 1263 static int ocfs2_shift_tree_depth(struct ocfs2_super *osb, 1264 handle_t *handle, 1265 struct inode *inode, 1266 struct ocfs2_extent_tree *et, 1267 struct ocfs2_alloc_context *meta_ac, 1268 struct buffer_head **ret_new_eb_bh) 1269 { 1270 int status, i; 1271 u32 new_clusters; 1272 struct buffer_head *new_eb_bh = NULL; 1273 struct ocfs2_extent_block *eb; 1274 struct ocfs2_extent_list *root_el; 1275 struct ocfs2_extent_list *eb_el; 1276 1277 mlog_entry_void(); 1278 1279 status = ocfs2_create_new_meta_bhs(handle, et, 1, meta_ac, 1280 &new_eb_bh); 1281 if (status < 0) { 1282 mlog_errno(status); 1283 goto bail; 1284 } 1285 1286 eb = (struct ocfs2_extent_block *) new_eb_bh->b_data; 1287 /* ocfs2_create_new_meta_bhs() should create it right! */ 1288 BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb)); 1289 1290 eb_el = &eb->h_list; 1291 root_el = et->et_root_el; 1292 1293 status = ocfs2_journal_access_eb(handle, INODE_CACHE(inode), new_eb_bh, 1294 OCFS2_JOURNAL_ACCESS_CREATE); 1295 if (status < 0) { 1296 mlog_errno(status); 1297 goto bail; 1298 } 1299 1300 /* copy the root extent list data into the new extent block */ 1301 eb_el->l_tree_depth = root_el->l_tree_depth; 1302 eb_el->l_next_free_rec = root_el->l_next_free_rec; 1303 for (i = 0; i < le16_to_cpu(root_el->l_next_free_rec); i++) 1304 eb_el->l_recs[i] = root_el->l_recs[i]; 1305 1306 status = ocfs2_journal_dirty(handle, new_eb_bh); 1307 if (status < 0) { 1308 mlog_errno(status); 1309 goto bail; 1310 } 1311 1312 status = ocfs2_et_root_journal_access(handle, et, 1313 OCFS2_JOURNAL_ACCESS_WRITE); 1314 if (status < 0) { 1315 mlog_errno(status); 1316 goto bail; 1317 } 1318 1319 new_clusters = ocfs2_sum_rightmost_rec(eb_el); 1320 1321 /* update root_bh now */ 1322 le16_add_cpu(&root_el->l_tree_depth, 1); 1323 root_el->l_recs[0].e_cpos = 0; 1324 root_el->l_recs[0].e_blkno = eb->h_blkno; 1325 root_el->l_recs[0].e_int_clusters = cpu_to_le32(new_clusters); 1326 for (i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++) 1327 memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec)); 1328 root_el->l_next_free_rec = cpu_to_le16(1); 1329 1330 /* If this is our 1st tree depth shift, then last_eb_blk 1331 * becomes the allocated extent block */ 1332 if (root_el->l_tree_depth == cpu_to_le16(1)) 1333 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno)); 1334 1335 status = ocfs2_journal_dirty(handle, et->et_root_bh); 1336 if (status < 0) { 1337 mlog_errno(status); 1338 goto bail; 1339 } 1340 1341 *ret_new_eb_bh = new_eb_bh; 1342 new_eb_bh = NULL; 1343 status = 0; 1344 bail: 1345 brelse(new_eb_bh); 1346 1347 mlog_exit(status); 1348 return status; 1349 } 1350 1351 /* 1352 * Should only be called when there is no space left in any of the 1353 * leaf nodes. What we want to do is find the lowest tree depth 1354 * non-leaf extent block with room for new records. There are three 1355 * valid results of this search: 1356 * 1357 * 1) a lowest extent block is found, then we pass it back in 1358 * *lowest_eb_bh and return '0' 1359 * 1360 * 2) the search fails to find anything, but the root_el has room. We 1361 * pass NULL back in *lowest_eb_bh, but still return '0' 1362 * 1363 * 3) the search fails to find anything AND the root_el is full, in 1364 * which case we return > 0 1365 * 1366 * return status < 0 indicates an error. 1367 */ 1368 static int ocfs2_find_branch_target(struct ocfs2_super *osb, 1369 struct ocfs2_extent_tree *et, 1370 struct buffer_head **target_bh) 1371 { 1372 int status = 0, i; 1373 u64 blkno; 1374 struct ocfs2_extent_block *eb; 1375 struct ocfs2_extent_list *el; 1376 struct buffer_head *bh = NULL; 1377 struct buffer_head *lowest_bh = NULL; 1378 1379 mlog_entry_void(); 1380 1381 *target_bh = NULL; 1382 1383 el = et->et_root_el; 1384 1385 while(le16_to_cpu(el->l_tree_depth) > 1) { 1386 if (le16_to_cpu(el->l_next_free_rec) == 0) { 1387 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci), 1388 "Owner %llu has empty " 1389 "extent list (next_free_rec == 0)", 1390 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci)); 1391 status = -EIO; 1392 goto bail; 1393 } 1394 i = le16_to_cpu(el->l_next_free_rec) - 1; 1395 blkno = le64_to_cpu(el->l_recs[i].e_blkno); 1396 if (!blkno) { 1397 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci), 1398 "Owner %llu has extent " 1399 "list where extent # %d has no physical " 1400 "block start", 1401 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci), i); 1402 status = -EIO; 1403 goto bail; 1404 } 1405 1406 brelse(bh); 1407 bh = NULL; 1408 1409 status = ocfs2_read_extent_block(et->et_ci, blkno, &bh); 1410 if (status < 0) { 1411 mlog_errno(status); 1412 goto bail; 1413 } 1414 1415 eb = (struct ocfs2_extent_block *) bh->b_data; 1416 el = &eb->h_list; 1417 1418 if (le16_to_cpu(el->l_next_free_rec) < 1419 le16_to_cpu(el->l_count)) { 1420 brelse(lowest_bh); 1421 lowest_bh = bh; 1422 get_bh(lowest_bh); 1423 } 1424 } 1425 1426 /* If we didn't find one and the fe doesn't have any room, 1427 * then return '1' */ 1428 el = et->et_root_el; 1429 if (!lowest_bh && (el->l_next_free_rec == el->l_count)) 1430 status = 1; 1431 1432 *target_bh = lowest_bh; 1433 bail: 1434 brelse(bh); 1435 1436 mlog_exit(status); 1437 return status; 1438 } 1439 1440 /* 1441 * Grow a b-tree so that it has more records. 1442 * 1443 * We might shift the tree depth in which case existing paths should 1444 * be considered invalid. 1445 * 1446 * Tree depth after the grow is returned via *final_depth. 1447 * 1448 * *last_eb_bh will be updated by ocfs2_add_branch(). 1449 */ 1450 static int ocfs2_grow_tree(struct inode *inode, handle_t *handle, 1451 struct ocfs2_extent_tree *et, int *final_depth, 1452 struct buffer_head **last_eb_bh, 1453 struct ocfs2_alloc_context *meta_ac) 1454 { 1455 int ret, shift; 1456 struct ocfs2_extent_list *el = et->et_root_el; 1457 int depth = le16_to_cpu(el->l_tree_depth); 1458 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 1459 struct buffer_head *bh = NULL; 1460 1461 BUG_ON(meta_ac == NULL); 1462 1463 shift = ocfs2_find_branch_target(osb, et, &bh); 1464 if (shift < 0) { 1465 ret = shift; 1466 mlog_errno(ret); 1467 goto out; 1468 } 1469 1470 /* We traveled all the way to the bottom of the allocation tree 1471 * and didn't find room for any more extents - we need to add 1472 * another tree level */ 1473 if (shift) { 1474 BUG_ON(bh); 1475 mlog(0, "need to shift tree depth (current = %d)\n", depth); 1476 1477 /* ocfs2_shift_tree_depth will return us a buffer with 1478 * the new extent block (so we can pass that to 1479 * ocfs2_add_branch). */ 1480 ret = ocfs2_shift_tree_depth(osb, handle, inode, et, 1481 meta_ac, &bh); 1482 if (ret < 0) { 1483 mlog_errno(ret); 1484 goto out; 1485 } 1486 depth++; 1487 if (depth == 1) { 1488 /* 1489 * Special case: we have room now if we shifted from 1490 * tree_depth 0, so no more work needs to be done. 1491 * 1492 * We won't be calling add_branch, so pass 1493 * back *last_eb_bh as the new leaf. At depth 1494 * zero, it should always be null so there's 1495 * no reason to brelse. 1496 */ 1497 BUG_ON(*last_eb_bh); 1498 get_bh(bh); 1499 *last_eb_bh = bh; 1500 goto out; 1501 } 1502 } 1503 1504 /* call ocfs2_add_branch to add the final part of the tree with 1505 * the new data. */ 1506 mlog(0, "add branch. bh = %p\n", bh); 1507 ret = ocfs2_add_branch(osb, handle, inode, et, bh, last_eb_bh, 1508 meta_ac); 1509 if (ret < 0) { 1510 mlog_errno(ret); 1511 goto out; 1512 } 1513 1514 out: 1515 if (final_depth) 1516 *final_depth = depth; 1517 brelse(bh); 1518 return ret; 1519 } 1520 1521 /* 1522 * This function will discard the rightmost extent record. 1523 */ 1524 static void ocfs2_shift_records_right(struct ocfs2_extent_list *el) 1525 { 1526 int next_free = le16_to_cpu(el->l_next_free_rec); 1527 int count = le16_to_cpu(el->l_count); 1528 unsigned int num_bytes; 1529 1530 BUG_ON(!next_free); 1531 /* This will cause us to go off the end of our extent list. */ 1532 BUG_ON(next_free >= count); 1533 1534 num_bytes = sizeof(struct ocfs2_extent_rec) * next_free; 1535 1536 memmove(&el->l_recs[1], &el->l_recs[0], num_bytes); 1537 } 1538 1539 static void ocfs2_rotate_leaf(struct ocfs2_extent_list *el, 1540 struct ocfs2_extent_rec *insert_rec) 1541 { 1542 int i, insert_index, next_free, has_empty, num_bytes; 1543 u32 insert_cpos = le32_to_cpu(insert_rec->e_cpos); 1544 struct ocfs2_extent_rec *rec; 1545 1546 next_free = le16_to_cpu(el->l_next_free_rec); 1547 has_empty = ocfs2_is_empty_extent(&el->l_recs[0]); 1548 1549 BUG_ON(!next_free); 1550 1551 /* The tree code before us didn't allow enough room in the leaf. */ 1552 BUG_ON(el->l_next_free_rec == el->l_count && !has_empty); 1553 1554 /* 1555 * The easiest way to approach this is to just remove the 1556 * empty extent and temporarily decrement next_free. 1557 */ 1558 if (has_empty) { 1559 /* 1560 * If next_free was 1 (only an empty extent), this 1561 * loop won't execute, which is fine. We still want 1562 * the decrement above to happen. 1563 */ 1564 for(i = 0; i < (next_free - 1); i++) 1565 el->l_recs[i] = el->l_recs[i+1]; 1566 1567 next_free--; 1568 } 1569 1570 /* 1571 * Figure out what the new record index should be. 1572 */ 1573 for(i = 0; i < next_free; i++) { 1574 rec = &el->l_recs[i]; 1575 1576 if (insert_cpos < le32_to_cpu(rec->e_cpos)) 1577 break; 1578 } 1579 insert_index = i; 1580 1581 mlog(0, "ins %u: index %d, has_empty %d, next_free %d, count %d\n", 1582 insert_cpos, insert_index, has_empty, next_free, le16_to_cpu(el->l_count)); 1583 1584 BUG_ON(insert_index < 0); 1585 BUG_ON(insert_index >= le16_to_cpu(el->l_count)); 1586 BUG_ON(insert_index > next_free); 1587 1588 /* 1589 * No need to memmove if we're just adding to the tail. 1590 */ 1591 if (insert_index != next_free) { 1592 BUG_ON(next_free >= le16_to_cpu(el->l_count)); 1593 1594 num_bytes = next_free - insert_index; 1595 num_bytes *= sizeof(struct ocfs2_extent_rec); 1596 memmove(&el->l_recs[insert_index + 1], 1597 &el->l_recs[insert_index], 1598 num_bytes); 1599 } 1600 1601 /* 1602 * Either we had an empty extent, and need to re-increment or 1603 * there was no empty extent on a non full rightmost leaf node, 1604 * in which case we still need to increment. 1605 */ 1606 next_free++; 1607 el->l_next_free_rec = cpu_to_le16(next_free); 1608 /* 1609 * Make sure none of the math above just messed up our tree. 1610 */ 1611 BUG_ON(le16_to_cpu(el->l_next_free_rec) > le16_to_cpu(el->l_count)); 1612 1613 el->l_recs[insert_index] = *insert_rec; 1614 1615 } 1616 1617 static void ocfs2_remove_empty_extent(struct ocfs2_extent_list *el) 1618 { 1619 int size, num_recs = le16_to_cpu(el->l_next_free_rec); 1620 1621 BUG_ON(num_recs == 0); 1622 1623 if (ocfs2_is_empty_extent(&el->l_recs[0])) { 1624 num_recs--; 1625 size = num_recs * sizeof(struct ocfs2_extent_rec); 1626 memmove(&el->l_recs[0], &el->l_recs[1], size); 1627 memset(&el->l_recs[num_recs], 0, 1628 sizeof(struct ocfs2_extent_rec)); 1629 el->l_next_free_rec = cpu_to_le16(num_recs); 1630 } 1631 } 1632 1633 /* 1634 * Create an empty extent record . 1635 * 1636 * l_next_free_rec may be updated. 1637 * 1638 * If an empty extent already exists do nothing. 1639 */ 1640 static void ocfs2_create_empty_extent(struct ocfs2_extent_list *el) 1641 { 1642 int next_free = le16_to_cpu(el->l_next_free_rec); 1643 1644 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0); 1645 1646 if (next_free == 0) 1647 goto set_and_inc; 1648 1649 if (ocfs2_is_empty_extent(&el->l_recs[0])) 1650 return; 1651 1652 mlog_bug_on_msg(el->l_count == el->l_next_free_rec, 1653 "Asked to create an empty extent in a full list:\n" 1654 "count = %u, tree depth = %u", 1655 le16_to_cpu(el->l_count), 1656 le16_to_cpu(el->l_tree_depth)); 1657 1658 ocfs2_shift_records_right(el); 1659 1660 set_and_inc: 1661 le16_add_cpu(&el->l_next_free_rec, 1); 1662 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec)); 1663 } 1664 1665 /* 1666 * For a rotation which involves two leaf nodes, the "root node" is 1667 * the lowest level tree node which contains a path to both leafs. This 1668 * resulting set of information can be used to form a complete "subtree" 1669 * 1670 * This function is passed two full paths from the dinode down to a 1671 * pair of adjacent leaves. It's task is to figure out which path 1672 * index contains the subtree root - this can be the root index itself 1673 * in a worst-case rotation. 1674 * 1675 * The array index of the subtree root is passed back. 1676 */ 1677 static int ocfs2_find_subtree_root(struct ocfs2_extent_tree *et, 1678 struct ocfs2_path *left, 1679 struct ocfs2_path *right) 1680 { 1681 int i = 0; 1682 1683 /* 1684 * Check that the caller passed in two paths from the same tree. 1685 */ 1686 BUG_ON(path_root_bh(left) != path_root_bh(right)); 1687 1688 do { 1689 i++; 1690 1691 /* 1692 * The caller didn't pass two adjacent paths. 1693 */ 1694 mlog_bug_on_msg(i > left->p_tree_depth, 1695 "Owner %llu, left depth %u, right depth %u\n" 1696 "left leaf blk %llu, right leaf blk %llu\n", 1697 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci), 1698 left->p_tree_depth, right->p_tree_depth, 1699 (unsigned long long)path_leaf_bh(left)->b_blocknr, 1700 (unsigned long long)path_leaf_bh(right)->b_blocknr); 1701 } while (left->p_node[i].bh->b_blocknr == 1702 right->p_node[i].bh->b_blocknr); 1703 1704 return i - 1; 1705 } 1706 1707 typedef void (path_insert_t)(void *, struct buffer_head *); 1708 1709 /* 1710 * Traverse a btree path in search of cpos, starting at root_el. 1711 * 1712 * This code can be called with a cpos larger than the tree, in which 1713 * case it will return the rightmost path. 1714 */ 1715 static int __ocfs2_find_path(struct ocfs2_caching_info *ci, 1716 struct ocfs2_extent_list *root_el, u32 cpos, 1717 path_insert_t *func, void *data) 1718 { 1719 int i, ret = 0; 1720 u32 range; 1721 u64 blkno; 1722 struct buffer_head *bh = NULL; 1723 struct ocfs2_extent_block *eb; 1724 struct ocfs2_extent_list *el; 1725 struct ocfs2_extent_rec *rec; 1726 1727 el = root_el; 1728 while (el->l_tree_depth) { 1729 if (le16_to_cpu(el->l_next_free_rec) == 0) { 1730 ocfs2_error(ocfs2_metadata_cache_get_super(ci), 1731 "Owner %llu has empty extent list at " 1732 "depth %u\n", 1733 (unsigned long long)ocfs2_metadata_cache_owner(ci), 1734 le16_to_cpu(el->l_tree_depth)); 1735 ret = -EROFS; 1736 goto out; 1737 1738 } 1739 1740 for(i = 0; i < le16_to_cpu(el->l_next_free_rec) - 1; i++) { 1741 rec = &el->l_recs[i]; 1742 1743 /* 1744 * In the case that cpos is off the allocation 1745 * tree, this should just wind up returning the 1746 * rightmost record. 1747 */ 1748 range = le32_to_cpu(rec->e_cpos) + 1749 ocfs2_rec_clusters(el, rec); 1750 if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range) 1751 break; 1752 } 1753 1754 blkno = le64_to_cpu(el->l_recs[i].e_blkno); 1755 if (blkno == 0) { 1756 ocfs2_error(ocfs2_metadata_cache_get_super(ci), 1757 "Owner %llu has bad blkno in extent list " 1758 "at depth %u (index %d)\n", 1759 (unsigned long long)ocfs2_metadata_cache_owner(ci), 1760 le16_to_cpu(el->l_tree_depth), i); 1761 ret = -EROFS; 1762 goto out; 1763 } 1764 1765 brelse(bh); 1766 bh = NULL; 1767 ret = ocfs2_read_extent_block(ci, blkno, &bh); 1768 if (ret) { 1769 mlog_errno(ret); 1770 goto out; 1771 } 1772 1773 eb = (struct ocfs2_extent_block *) bh->b_data; 1774 el = &eb->h_list; 1775 1776 if (le16_to_cpu(el->l_next_free_rec) > 1777 le16_to_cpu(el->l_count)) { 1778 ocfs2_error(ocfs2_metadata_cache_get_super(ci), 1779 "Owner %llu has bad count in extent list " 1780 "at block %llu (next free=%u, count=%u)\n", 1781 (unsigned long long)ocfs2_metadata_cache_owner(ci), 1782 (unsigned long long)bh->b_blocknr, 1783 le16_to_cpu(el->l_next_free_rec), 1784 le16_to_cpu(el->l_count)); 1785 ret = -EROFS; 1786 goto out; 1787 } 1788 1789 if (func) 1790 func(data, bh); 1791 } 1792 1793 out: 1794 /* 1795 * Catch any trailing bh that the loop didn't handle. 1796 */ 1797 brelse(bh); 1798 1799 return ret; 1800 } 1801 1802 /* 1803 * Given an initialized path (that is, it has a valid root extent 1804 * list), this function will traverse the btree in search of the path 1805 * which would contain cpos. 1806 * 1807 * The path traveled is recorded in the path structure. 1808 * 1809 * Note that this will not do any comparisons on leaf node extent 1810 * records, so it will work fine in the case that we just added a tree 1811 * branch. 1812 */ 1813 struct find_path_data { 1814 int index; 1815 struct ocfs2_path *path; 1816 }; 1817 static void find_path_ins(void *data, struct buffer_head *bh) 1818 { 1819 struct find_path_data *fp = data; 1820 1821 get_bh(bh); 1822 ocfs2_path_insert_eb(fp->path, fp->index, bh); 1823 fp->index++; 1824 } 1825 static int ocfs2_find_path(struct ocfs2_caching_info *ci, 1826 struct ocfs2_path *path, u32 cpos) 1827 { 1828 struct find_path_data data; 1829 1830 data.index = 1; 1831 data.path = path; 1832 return __ocfs2_find_path(ci, path_root_el(path), cpos, 1833 find_path_ins, &data); 1834 } 1835 1836 static void find_leaf_ins(void *data, struct buffer_head *bh) 1837 { 1838 struct ocfs2_extent_block *eb =(struct ocfs2_extent_block *)bh->b_data; 1839 struct ocfs2_extent_list *el = &eb->h_list; 1840 struct buffer_head **ret = data; 1841 1842 /* We want to retain only the leaf block. */ 1843 if (le16_to_cpu(el->l_tree_depth) == 0) { 1844 get_bh(bh); 1845 *ret = bh; 1846 } 1847 } 1848 /* 1849 * Find the leaf block in the tree which would contain cpos. No 1850 * checking of the actual leaf is done. 1851 * 1852 * Some paths want to call this instead of allocating a path structure 1853 * and calling ocfs2_find_path(). 1854 * 1855 * This function doesn't handle non btree extent lists. 1856 */ 1857 int ocfs2_find_leaf(struct ocfs2_caching_info *ci, 1858 struct ocfs2_extent_list *root_el, u32 cpos, 1859 struct buffer_head **leaf_bh) 1860 { 1861 int ret; 1862 struct buffer_head *bh = NULL; 1863 1864 ret = __ocfs2_find_path(ci, root_el, cpos, find_leaf_ins, &bh); 1865 if (ret) { 1866 mlog_errno(ret); 1867 goto out; 1868 } 1869 1870 *leaf_bh = bh; 1871 out: 1872 return ret; 1873 } 1874 1875 /* 1876 * Adjust the adjacent records (left_rec, right_rec) involved in a rotation. 1877 * 1878 * Basically, we've moved stuff around at the bottom of the tree and 1879 * we need to fix up the extent records above the changes to reflect 1880 * the new changes. 1881 * 1882 * left_rec: the record on the left. 1883 * left_child_el: is the child list pointed to by left_rec 1884 * right_rec: the record to the right of left_rec 1885 * right_child_el: is the child list pointed to by right_rec 1886 * 1887 * By definition, this only works on interior nodes. 1888 */ 1889 static void ocfs2_adjust_adjacent_records(struct ocfs2_extent_rec *left_rec, 1890 struct ocfs2_extent_list *left_child_el, 1891 struct ocfs2_extent_rec *right_rec, 1892 struct ocfs2_extent_list *right_child_el) 1893 { 1894 u32 left_clusters, right_end; 1895 1896 /* 1897 * Interior nodes never have holes. Their cpos is the cpos of 1898 * the leftmost record in their child list. Their cluster 1899 * count covers the full theoretical range of their child list 1900 * - the range between their cpos and the cpos of the record 1901 * immediately to their right. 1902 */ 1903 left_clusters = le32_to_cpu(right_child_el->l_recs[0].e_cpos); 1904 if (!ocfs2_rec_clusters(right_child_el, &right_child_el->l_recs[0])) { 1905 BUG_ON(right_child_el->l_tree_depth); 1906 BUG_ON(le16_to_cpu(right_child_el->l_next_free_rec) <= 1); 1907 left_clusters = le32_to_cpu(right_child_el->l_recs[1].e_cpos); 1908 } 1909 left_clusters -= le32_to_cpu(left_rec->e_cpos); 1910 left_rec->e_int_clusters = cpu_to_le32(left_clusters); 1911 1912 /* 1913 * Calculate the rightmost cluster count boundary before 1914 * moving cpos - we will need to adjust clusters after 1915 * updating e_cpos to keep the same highest cluster count. 1916 */ 1917 right_end = le32_to_cpu(right_rec->e_cpos); 1918 right_end += le32_to_cpu(right_rec->e_int_clusters); 1919 1920 right_rec->e_cpos = left_rec->e_cpos; 1921 le32_add_cpu(&right_rec->e_cpos, left_clusters); 1922 1923 right_end -= le32_to_cpu(right_rec->e_cpos); 1924 right_rec->e_int_clusters = cpu_to_le32(right_end); 1925 } 1926 1927 /* 1928 * Adjust the adjacent root node records involved in a 1929 * rotation. left_el_blkno is passed in as a key so that we can easily 1930 * find it's index in the root list. 1931 */ 1932 static void ocfs2_adjust_root_records(struct ocfs2_extent_list *root_el, 1933 struct ocfs2_extent_list *left_el, 1934 struct ocfs2_extent_list *right_el, 1935 u64 left_el_blkno) 1936 { 1937 int i; 1938 1939 BUG_ON(le16_to_cpu(root_el->l_tree_depth) <= 1940 le16_to_cpu(left_el->l_tree_depth)); 1941 1942 for(i = 0; i < le16_to_cpu(root_el->l_next_free_rec) - 1; i++) { 1943 if (le64_to_cpu(root_el->l_recs[i].e_blkno) == left_el_blkno) 1944 break; 1945 } 1946 1947 /* 1948 * The path walking code should have never returned a root and 1949 * two paths which are not adjacent. 1950 */ 1951 BUG_ON(i >= (le16_to_cpu(root_el->l_next_free_rec) - 1)); 1952 1953 ocfs2_adjust_adjacent_records(&root_el->l_recs[i], left_el, 1954 &root_el->l_recs[i + 1], right_el); 1955 } 1956 1957 /* 1958 * We've changed a leaf block (in right_path) and need to reflect that 1959 * change back up the subtree. 1960 * 1961 * This happens in multiple places: 1962 * - When we've moved an extent record from the left path leaf to the right 1963 * path leaf to make room for an empty extent in the left path leaf. 1964 * - When our insert into the right path leaf is at the leftmost edge 1965 * and requires an update of the path immediately to it's left. This 1966 * can occur at the end of some types of rotation and appending inserts. 1967 * - When we've adjusted the last extent record in the left path leaf and the 1968 * 1st extent record in the right path leaf during cross extent block merge. 1969 */ 1970 static void ocfs2_complete_edge_insert(handle_t *handle, 1971 struct ocfs2_path *left_path, 1972 struct ocfs2_path *right_path, 1973 int subtree_index) 1974 { 1975 int ret, i, idx; 1976 struct ocfs2_extent_list *el, *left_el, *right_el; 1977 struct ocfs2_extent_rec *left_rec, *right_rec; 1978 struct buffer_head *root_bh = left_path->p_node[subtree_index].bh; 1979 1980 /* 1981 * Update the counts and position values within all the 1982 * interior nodes to reflect the leaf rotation we just did. 1983 * 1984 * The root node is handled below the loop. 1985 * 1986 * We begin the loop with right_el and left_el pointing to the 1987 * leaf lists and work our way up. 1988 * 1989 * NOTE: within this loop, left_el and right_el always refer 1990 * to the *child* lists. 1991 */ 1992 left_el = path_leaf_el(left_path); 1993 right_el = path_leaf_el(right_path); 1994 for(i = left_path->p_tree_depth - 1; i > subtree_index; i--) { 1995 mlog(0, "Adjust records at index %u\n", i); 1996 1997 /* 1998 * One nice property of knowing that all of these 1999 * nodes are below the root is that we only deal with 2000 * the leftmost right node record and the rightmost 2001 * left node record. 2002 */ 2003 el = left_path->p_node[i].el; 2004 idx = le16_to_cpu(left_el->l_next_free_rec) - 1; 2005 left_rec = &el->l_recs[idx]; 2006 2007 el = right_path->p_node[i].el; 2008 right_rec = &el->l_recs[0]; 2009 2010 ocfs2_adjust_adjacent_records(left_rec, left_el, right_rec, 2011 right_el); 2012 2013 ret = ocfs2_journal_dirty(handle, left_path->p_node[i].bh); 2014 if (ret) 2015 mlog_errno(ret); 2016 2017 ret = ocfs2_journal_dirty(handle, right_path->p_node[i].bh); 2018 if (ret) 2019 mlog_errno(ret); 2020 2021 /* 2022 * Setup our list pointers now so that the current 2023 * parents become children in the next iteration. 2024 */ 2025 left_el = left_path->p_node[i].el; 2026 right_el = right_path->p_node[i].el; 2027 } 2028 2029 /* 2030 * At the root node, adjust the two adjacent records which 2031 * begin our path to the leaves. 2032 */ 2033 2034 el = left_path->p_node[subtree_index].el; 2035 left_el = left_path->p_node[subtree_index + 1].el; 2036 right_el = right_path->p_node[subtree_index + 1].el; 2037 2038 ocfs2_adjust_root_records(el, left_el, right_el, 2039 left_path->p_node[subtree_index + 1].bh->b_blocknr); 2040 2041 root_bh = left_path->p_node[subtree_index].bh; 2042 2043 ret = ocfs2_journal_dirty(handle, root_bh); 2044 if (ret) 2045 mlog_errno(ret); 2046 } 2047 2048 static int ocfs2_rotate_subtree_right(handle_t *handle, 2049 struct ocfs2_extent_tree *et, 2050 struct ocfs2_path *left_path, 2051 struct ocfs2_path *right_path, 2052 int subtree_index) 2053 { 2054 int ret, i; 2055 struct buffer_head *right_leaf_bh; 2056 struct buffer_head *left_leaf_bh = NULL; 2057 struct buffer_head *root_bh; 2058 struct ocfs2_extent_list *right_el, *left_el; 2059 struct ocfs2_extent_rec move_rec; 2060 2061 left_leaf_bh = path_leaf_bh(left_path); 2062 left_el = path_leaf_el(left_path); 2063 2064 if (left_el->l_next_free_rec != left_el->l_count) { 2065 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci), 2066 "Inode %llu has non-full interior leaf node %llu" 2067 "(next free = %u)", 2068 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci), 2069 (unsigned long long)left_leaf_bh->b_blocknr, 2070 le16_to_cpu(left_el->l_next_free_rec)); 2071 return -EROFS; 2072 } 2073 2074 /* 2075 * This extent block may already have an empty record, so we 2076 * return early if so. 2077 */ 2078 if (ocfs2_is_empty_extent(&left_el->l_recs[0])) 2079 return 0; 2080 2081 root_bh = left_path->p_node[subtree_index].bh; 2082 BUG_ON(root_bh != right_path->p_node[subtree_index].bh); 2083 2084 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path, 2085 subtree_index); 2086 if (ret) { 2087 mlog_errno(ret); 2088 goto out; 2089 } 2090 2091 for(i = subtree_index + 1; i < path_num_items(right_path); i++) { 2092 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, 2093 right_path, i); 2094 if (ret) { 2095 mlog_errno(ret); 2096 goto out; 2097 } 2098 2099 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, 2100 left_path, i); 2101 if (ret) { 2102 mlog_errno(ret); 2103 goto out; 2104 } 2105 } 2106 2107 right_leaf_bh = path_leaf_bh(right_path); 2108 right_el = path_leaf_el(right_path); 2109 2110 /* This is a code error, not a disk corruption. */ 2111 mlog_bug_on_msg(!right_el->l_next_free_rec, "Inode %llu: Rotate fails " 2112 "because rightmost leaf block %llu is empty\n", 2113 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci), 2114 (unsigned long long)right_leaf_bh->b_blocknr); 2115 2116 ocfs2_create_empty_extent(right_el); 2117 2118 ret = ocfs2_journal_dirty(handle, right_leaf_bh); 2119 if (ret) { 2120 mlog_errno(ret); 2121 goto out; 2122 } 2123 2124 /* Do the copy now. */ 2125 i = le16_to_cpu(left_el->l_next_free_rec) - 1; 2126 move_rec = left_el->l_recs[i]; 2127 right_el->l_recs[0] = move_rec; 2128 2129 /* 2130 * Clear out the record we just copied and shift everything 2131 * over, leaving an empty extent in the left leaf. 2132 * 2133 * We temporarily subtract from next_free_rec so that the 2134 * shift will lose the tail record (which is now defunct). 2135 */ 2136 le16_add_cpu(&left_el->l_next_free_rec, -1); 2137 ocfs2_shift_records_right(left_el); 2138 memset(&left_el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec)); 2139 le16_add_cpu(&left_el->l_next_free_rec, 1); 2140 2141 ret = ocfs2_journal_dirty(handle, left_leaf_bh); 2142 if (ret) { 2143 mlog_errno(ret); 2144 goto out; 2145 } 2146 2147 ocfs2_complete_edge_insert(handle, left_path, right_path, 2148 subtree_index); 2149 2150 out: 2151 return ret; 2152 } 2153 2154 /* 2155 * Given a full path, determine what cpos value would return us a path 2156 * containing the leaf immediately to the left of the current one. 2157 * 2158 * Will return zero if the path passed in is already the leftmost path. 2159 */ 2160 static int ocfs2_find_cpos_for_left_leaf(struct super_block *sb, 2161 struct ocfs2_path *path, u32 *cpos) 2162 { 2163 int i, j, ret = 0; 2164 u64 blkno; 2165 struct ocfs2_extent_list *el; 2166 2167 BUG_ON(path->p_tree_depth == 0); 2168 2169 *cpos = 0; 2170 2171 blkno = path_leaf_bh(path)->b_blocknr; 2172 2173 /* Start at the tree node just above the leaf and work our way up. */ 2174 i = path->p_tree_depth - 1; 2175 while (i >= 0) { 2176 el = path->p_node[i].el; 2177 2178 /* 2179 * Find the extent record just before the one in our 2180 * path. 2181 */ 2182 for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) { 2183 if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) { 2184 if (j == 0) { 2185 if (i == 0) { 2186 /* 2187 * We've determined that the 2188 * path specified is already 2189 * the leftmost one - return a 2190 * cpos of zero. 2191 */ 2192 goto out; 2193 } 2194 /* 2195 * The leftmost record points to our 2196 * leaf - we need to travel up the 2197 * tree one level. 2198 */ 2199 goto next_node; 2200 } 2201 2202 *cpos = le32_to_cpu(el->l_recs[j - 1].e_cpos); 2203 *cpos = *cpos + ocfs2_rec_clusters(el, 2204 &el->l_recs[j - 1]); 2205 *cpos = *cpos - 1; 2206 goto out; 2207 } 2208 } 2209 2210 /* 2211 * If we got here, we never found a valid node where 2212 * the tree indicated one should be. 2213 */ 2214 ocfs2_error(sb, 2215 "Invalid extent tree at extent block %llu\n", 2216 (unsigned long long)blkno); 2217 ret = -EROFS; 2218 goto out; 2219 2220 next_node: 2221 blkno = path->p_node[i].bh->b_blocknr; 2222 i--; 2223 } 2224 2225 out: 2226 return ret; 2227 } 2228 2229 /* 2230 * Extend the transaction by enough credits to complete the rotation, 2231 * and still leave at least the original number of credits allocated 2232 * to this transaction. 2233 */ 2234 static int ocfs2_extend_rotate_transaction(handle_t *handle, int subtree_depth, 2235 int op_credits, 2236 struct ocfs2_path *path) 2237 { 2238 int credits = (path->p_tree_depth - subtree_depth) * 2 + 1 + op_credits; 2239 2240 if (handle->h_buffer_credits < credits) 2241 return ocfs2_extend_trans(handle, credits); 2242 2243 return 0; 2244 } 2245 2246 /* 2247 * Trap the case where we're inserting into the theoretical range past 2248 * the _actual_ left leaf range. Otherwise, we'll rotate a record 2249 * whose cpos is less than ours into the right leaf. 2250 * 2251 * It's only necessary to look at the rightmost record of the left 2252 * leaf because the logic that calls us should ensure that the 2253 * theoretical ranges in the path components above the leaves are 2254 * correct. 2255 */ 2256 static int ocfs2_rotate_requires_path_adjustment(struct ocfs2_path *left_path, 2257 u32 insert_cpos) 2258 { 2259 struct ocfs2_extent_list *left_el; 2260 struct ocfs2_extent_rec *rec; 2261 int next_free; 2262 2263 left_el = path_leaf_el(left_path); 2264 next_free = le16_to_cpu(left_el->l_next_free_rec); 2265 rec = &left_el->l_recs[next_free - 1]; 2266 2267 if (insert_cpos > le32_to_cpu(rec->e_cpos)) 2268 return 1; 2269 return 0; 2270 } 2271 2272 static int ocfs2_leftmost_rec_contains(struct ocfs2_extent_list *el, u32 cpos) 2273 { 2274 int next_free = le16_to_cpu(el->l_next_free_rec); 2275 unsigned int range; 2276 struct ocfs2_extent_rec *rec; 2277 2278 if (next_free == 0) 2279 return 0; 2280 2281 rec = &el->l_recs[0]; 2282 if (ocfs2_is_empty_extent(rec)) { 2283 /* Empty list. */ 2284 if (next_free == 1) 2285 return 0; 2286 rec = &el->l_recs[1]; 2287 } 2288 2289 range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec); 2290 if (cpos >= le32_to_cpu(rec->e_cpos) && cpos < range) 2291 return 1; 2292 return 0; 2293 } 2294 2295 /* 2296 * Rotate all the records in a btree right one record, starting at insert_cpos. 2297 * 2298 * The path to the rightmost leaf should be passed in. 2299 * 2300 * The array is assumed to be large enough to hold an entire path (tree depth). 2301 * 2302 * Upon succesful return from this function: 2303 * 2304 * - The 'right_path' array will contain a path to the leaf block 2305 * whose range contains e_cpos. 2306 * - That leaf block will have a single empty extent in list index 0. 2307 * - In the case that the rotation requires a post-insert update, 2308 * *ret_left_path will contain a valid path which can be passed to 2309 * ocfs2_insert_path(). 2310 */ 2311 static int ocfs2_rotate_tree_right(handle_t *handle, 2312 struct ocfs2_extent_tree *et, 2313 enum ocfs2_split_type split, 2314 u32 insert_cpos, 2315 struct ocfs2_path *right_path, 2316 struct ocfs2_path **ret_left_path) 2317 { 2318 int ret, start, orig_credits = handle->h_buffer_credits; 2319 u32 cpos; 2320 struct ocfs2_path *left_path = NULL; 2321 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci); 2322 2323 *ret_left_path = NULL; 2324 2325 left_path = ocfs2_new_path_from_path(right_path); 2326 if (!left_path) { 2327 ret = -ENOMEM; 2328 mlog_errno(ret); 2329 goto out; 2330 } 2331 2332 ret = ocfs2_find_cpos_for_left_leaf(sb, right_path, &cpos); 2333 if (ret) { 2334 mlog_errno(ret); 2335 goto out; 2336 } 2337 2338 mlog(0, "Insert: %u, first left path cpos: %u\n", insert_cpos, cpos); 2339 2340 /* 2341 * What we want to do here is: 2342 * 2343 * 1) Start with the rightmost path. 2344 * 2345 * 2) Determine a path to the leaf block directly to the left 2346 * of that leaf. 2347 * 2348 * 3) Determine the 'subtree root' - the lowest level tree node 2349 * which contains a path to both leaves. 2350 * 2351 * 4) Rotate the subtree. 2352 * 2353 * 5) Find the next subtree by considering the left path to be 2354 * the new right path. 2355 * 2356 * The check at the top of this while loop also accepts 2357 * insert_cpos == cpos because cpos is only a _theoretical_ 2358 * value to get us the left path - insert_cpos might very well 2359 * be filling that hole. 2360 * 2361 * Stop at a cpos of '0' because we either started at the 2362 * leftmost branch (i.e., a tree with one branch and a 2363 * rotation inside of it), or we've gone as far as we can in 2364 * rotating subtrees. 2365 */ 2366 while (cpos && insert_cpos <= cpos) { 2367 mlog(0, "Rotating a tree: ins. cpos: %u, left path cpos: %u\n", 2368 insert_cpos, cpos); 2369 2370 ret = ocfs2_find_path(et->et_ci, left_path, cpos); 2371 if (ret) { 2372 mlog_errno(ret); 2373 goto out; 2374 } 2375 2376 mlog_bug_on_msg(path_leaf_bh(left_path) == 2377 path_leaf_bh(right_path), 2378 "Owner %llu: error during insert of %u " 2379 "(left path cpos %u) results in two identical " 2380 "paths ending at %llu\n", 2381 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci), 2382 insert_cpos, cpos, 2383 (unsigned long long) 2384 path_leaf_bh(left_path)->b_blocknr); 2385 2386 if (split == SPLIT_NONE && 2387 ocfs2_rotate_requires_path_adjustment(left_path, 2388 insert_cpos)) { 2389 2390 /* 2391 * We've rotated the tree as much as we 2392 * should. The rest is up to 2393 * ocfs2_insert_path() to complete, after the 2394 * record insertion. We indicate this 2395 * situation by returning the left path. 2396 * 2397 * The reason we don't adjust the records here 2398 * before the record insert is that an error 2399 * later might break the rule where a parent 2400 * record e_cpos will reflect the actual 2401 * e_cpos of the 1st nonempty record of the 2402 * child list. 2403 */ 2404 *ret_left_path = left_path; 2405 goto out_ret_path; 2406 } 2407 2408 start = ocfs2_find_subtree_root(et, left_path, right_path); 2409 2410 mlog(0, "Subtree root at index %d (blk %llu, depth %d)\n", 2411 start, 2412 (unsigned long long) right_path->p_node[start].bh->b_blocknr, 2413 right_path->p_tree_depth); 2414 2415 ret = ocfs2_extend_rotate_transaction(handle, start, 2416 orig_credits, right_path); 2417 if (ret) { 2418 mlog_errno(ret); 2419 goto out; 2420 } 2421 2422 ret = ocfs2_rotate_subtree_right(handle, et, left_path, 2423 right_path, start); 2424 if (ret) { 2425 mlog_errno(ret); 2426 goto out; 2427 } 2428 2429 if (split != SPLIT_NONE && 2430 ocfs2_leftmost_rec_contains(path_leaf_el(right_path), 2431 insert_cpos)) { 2432 /* 2433 * A rotate moves the rightmost left leaf 2434 * record over to the leftmost right leaf 2435 * slot. If we're doing an extent split 2436 * instead of a real insert, then we have to 2437 * check that the extent to be split wasn't 2438 * just moved over. If it was, then we can 2439 * exit here, passing left_path back - 2440 * ocfs2_split_extent() is smart enough to 2441 * search both leaves. 2442 */ 2443 *ret_left_path = left_path; 2444 goto out_ret_path; 2445 } 2446 2447 /* 2448 * There is no need to re-read the next right path 2449 * as we know that it'll be our current left 2450 * path. Optimize by copying values instead. 2451 */ 2452 ocfs2_mv_path(right_path, left_path); 2453 2454 ret = ocfs2_find_cpos_for_left_leaf(sb, right_path, &cpos); 2455 if (ret) { 2456 mlog_errno(ret); 2457 goto out; 2458 } 2459 } 2460 2461 out: 2462 ocfs2_free_path(left_path); 2463 2464 out_ret_path: 2465 return ret; 2466 } 2467 2468 static int ocfs2_update_edge_lengths(handle_t *handle, 2469 struct ocfs2_extent_tree *et, 2470 int subtree_index, struct ocfs2_path *path) 2471 { 2472 int i, idx, ret; 2473 struct ocfs2_extent_rec *rec; 2474 struct ocfs2_extent_list *el; 2475 struct ocfs2_extent_block *eb; 2476 u32 range; 2477 2478 /* 2479 * In normal tree rotation process, we will never touch the 2480 * tree branch above subtree_index and ocfs2_extend_rotate_transaction 2481 * doesn't reserve the credits for them either. 2482 * 2483 * But we do have a special case here which will update the rightmost 2484 * records for all the bh in the path. 2485 * So we have to allocate extra credits and access them. 2486 */ 2487 ret = ocfs2_extend_trans(handle, 2488 handle->h_buffer_credits + subtree_index); 2489 if (ret) { 2490 mlog_errno(ret); 2491 goto out; 2492 } 2493 2494 ret = ocfs2_journal_access_path(et->et_ci, handle, path); 2495 if (ret) { 2496 mlog_errno(ret); 2497 goto out; 2498 } 2499 2500 /* Path should always be rightmost. */ 2501 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data; 2502 BUG_ON(eb->h_next_leaf_blk != 0ULL); 2503 2504 el = &eb->h_list; 2505 BUG_ON(le16_to_cpu(el->l_next_free_rec) == 0); 2506 idx = le16_to_cpu(el->l_next_free_rec) - 1; 2507 rec = &el->l_recs[idx]; 2508 range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec); 2509 2510 for (i = 0; i < path->p_tree_depth; i++) { 2511 el = path->p_node[i].el; 2512 idx = le16_to_cpu(el->l_next_free_rec) - 1; 2513 rec = &el->l_recs[idx]; 2514 2515 rec->e_int_clusters = cpu_to_le32(range); 2516 le32_add_cpu(&rec->e_int_clusters, -le32_to_cpu(rec->e_cpos)); 2517 2518 ocfs2_journal_dirty(handle, path->p_node[i].bh); 2519 } 2520 out: 2521 return ret; 2522 } 2523 2524 static void ocfs2_unlink_path(handle_t *handle, 2525 struct ocfs2_extent_tree *et, 2526 struct ocfs2_cached_dealloc_ctxt *dealloc, 2527 struct ocfs2_path *path, int unlink_start) 2528 { 2529 int ret, i; 2530 struct ocfs2_extent_block *eb; 2531 struct ocfs2_extent_list *el; 2532 struct buffer_head *bh; 2533 2534 for(i = unlink_start; i < path_num_items(path); i++) { 2535 bh = path->p_node[i].bh; 2536 2537 eb = (struct ocfs2_extent_block *)bh->b_data; 2538 /* 2539 * Not all nodes might have had their final count 2540 * decremented by the caller - handle this here. 2541 */ 2542 el = &eb->h_list; 2543 if (le16_to_cpu(el->l_next_free_rec) > 1) { 2544 mlog(ML_ERROR, 2545 "Inode %llu, attempted to remove extent block " 2546 "%llu with %u records\n", 2547 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci), 2548 (unsigned long long)le64_to_cpu(eb->h_blkno), 2549 le16_to_cpu(el->l_next_free_rec)); 2550 2551 ocfs2_journal_dirty(handle, bh); 2552 ocfs2_remove_from_cache(et->et_ci, bh); 2553 continue; 2554 } 2555 2556 el->l_next_free_rec = 0; 2557 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec)); 2558 2559 ocfs2_journal_dirty(handle, bh); 2560 2561 ret = ocfs2_cache_extent_block_free(dealloc, eb); 2562 if (ret) 2563 mlog_errno(ret); 2564 2565 ocfs2_remove_from_cache(et->et_ci, bh); 2566 } 2567 } 2568 2569 static void ocfs2_unlink_subtree(handle_t *handle, 2570 struct ocfs2_extent_tree *et, 2571 struct ocfs2_path *left_path, 2572 struct ocfs2_path *right_path, 2573 int subtree_index, 2574 struct ocfs2_cached_dealloc_ctxt *dealloc) 2575 { 2576 int i; 2577 struct buffer_head *root_bh = left_path->p_node[subtree_index].bh; 2578 struct ocfs2_extent_list *root_el = left_path->p_node[subtree_index].el; 2579 struct ocfs2_extent_list *el; 2580 struct ocfs2_extent_block *eb; 2581 2582 el = path_leaf_el(left_path); 2583 2584 eb = (struct ocfs2_extent_block *)right_path->p_node[subtree_index + 1].bh->b_data; 2585 2586 for(i = 1; i < le16_to_cpu(root_el->l_next_free_rec); i++) 2587 if (root_el->l_recs[i].e_blkno == eb->h_blkno) 2588 break; 2589 2590 BUG_ON(i >= le16_to_cpu(root_el->l_next_free_rec)); 2591 2592 memset(&root_el->l_recs[i], 0, sizeof(struct ocfs2_extent_rec)); 2593 le16_add_cpu(&root_el->l_next_free_rec, -1); 2594 2595 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data; 2596 eb->h_next_leaf_blk = 0; 2597 2598 ocfs2_journal_dirty(handle, root_bh); 2599 ocfs2_journal_dirty(handle, path_leaf_bh(left_path)); 2600 2601 ocfs2_unlink_path(handle, et, dealloc, right_path, 2602 subtree_index + 1); 2603 } 2604 2605 static int ocfs2_rotate_subtree_left(handle_t *handle, 2606 struct ocfs2_extent_tree *et, 2607 struct ocfs2_path *left_path, 2608 struct ocfs2_path *right_path, 2609 int subtree_index, 2610 struct ocfs2_cached_dealloc_ctxt *dealloc, 2611 int *deleted) 2612 { 2613 int ret, i, del_right_subtree = 0, right_has_empty = 0; 2614 struct buffer_head *root_bh, *et_root_bh = path_root_bh(right_path); 2615 struct ocfs2_extent_list *right_leaf_el, *left_leaf_el; 2616 struct ocfs2_extent_block *eb; 2617 2618 *deleted = 0; 2619 2620 right_leaf_el = path_leaf_el(right_path); 2621 left_leaf_el = path_leaf_el(left_path); 2622 root_bh = left_path->p_node[subtree_index].bh; 2623 BUG_ON(root_bh != right_path->p_node[subtree_index].bh); 2624 2625 if (!ocfs2_is_empty_extent(&left_leaf_el->l_recs[0])) 2626 return 0; 2627 2628 eb = (struct ocfs2_extent_block *)path_leaf_bh(right_path)->b_data; 2629 if (ocfs2_is_empty_extent(&right_leaf_el->l_recs[0])) { 2630 /* 2631 * It's legal for us to proceed if the right leaf is 2632 * the rightmost one and it has an empty extent. There 2633 * are two cases to handle - whether the leaf will be 2634 * empty after removal or not. If the leaf isn't empty 2635 * then just remove the empty extent up front. The 2636 * next block will handle empty leaves by flagging 2637 * them for unlink. 2638 * 2639 * Non rightmost leaves will throw -EAGAIN and the 2640 * caller can manually move the subtree and retry. 2641 */ 2642 2643 if (eb->h_next_leaf_blk != 0ULL) 2644 return -EAGAIN; 2645 2646 if (le16_to_cpu(right_leaf_el->l_next_free_rec) > 1) { 2647 ret = ocfs2_journal_access_eb(handle, et->et_ci, 2648 path_leaf_bh(right_path), 2649 OCFS2_JOURNAL_ACCESS_WRITE); 2650 if (ret) { 2651 mlog_errno(ret); 2652 goto out; 2653 } 2654 2655 ocfs2_remove_empty_extent(right_leaf_el); 2656 } else 2657 right_has_empty = 1; 2658 } 2659 2660 if (eb->h_next_leaf_blk == 0ULL && 2661 le16_to_cpu(right_leaf_el->l_next_free_rec) == 1) { 2662 /* 2663 * We have to update i_last_eb_blk during the meta 2664 * data delete. 2665 */ 2666 ret = ocfs2_et_root_journal_access(handle, et, 2667 OCFS2_JOURNAL_ACCESS_WRITE); 2668 if (ret) { 2669 mlog_errno(ret); 2670 goto out; 2671 } 2672 2673 del_right_subtree = 1; 2674 } 2675 2676 /* 2677 * Getting here with an empty extent in the right path implies 2678 * that it's the rightmost path and will be deleted. 2679 */ 2680 BUG_ON(right_has_empty && !del_right_subtree); 2681 2682 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path, 2683 subtree_index); 2684 if (ret) { 2685 mlog_errno(ret); 2686 goto out; 2687 } 2688 2689 for(i = subtree_index + 1; i < path_num_items(right_path); i++) { 2690 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, 2691 right_path, i); 2692 if (ret) { 2693 mlog_errno(ret); 2694 goto out; 2695 } 2696 2697 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, 2698 left_path, i); 2699 if (ret) { 2700 mlog_errno(ret); 2701 goto out; 2702 } 2703 } 2704 2705 if (!right_has_empty) { 2706 /* 2707 * Only do this if we're moving a real 2708 * record. Otherwise, the action is delayed until 2709 * after removal of the right path in which case we 2710 * can do a simple shift to remove the empty extent. 2711 */ 2712 ocfs2_rotate_leaf(left_leaf_el, &right_leaf_el->l_recs[0]); 2713 memset(&right_leaf_el->l_recs[0], 0, 2714 sizeof(struct ocfs2_extent_rec)); 2715 } 2716 if (eb->h_next_leaf_blk == 0ULL) { 2717 /* 2718 * Move recs over to get rid of empty extent, decrease 2719 * next_free. This is allowed to remove the last 2720 * extent in our leaf (setting l_next_free_rec to 2721 * zero) - the delete code below won't care. 2722 */ 2723 ocfs2_remove_empty_extent(right_leaf_el); 2724 } 2725 2726 ret = ocfs2_journal_dirty(handle, path_leaf_bh(left_path)); 2727 if (ret) 2728 mlog_errno(ret); 2729 ret = ocfs2_journal_dirty(handle, path_leaf_bh(right_path)); 2730 if (ret) 2731 mlog_errno(ret); 2732 2733 if (del_right_subtree) { 2734 ocfs2_unlink_subtree(handle, et, left_path, right_path, 2735 subtree_index, dealloc); 2736 ret = ocfs2_update_edge_lengths(handle, et, subtree_index, 2737 left_path); 2738 if (ret) { 2739 mlog_errno(ret); 2740 goto out; 2741 } 2742 2743 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data; 2744 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno)); 2745 2746 /* 2747 * Removal of the extent in the left leaf was skipped 2748 * above so we could delete the right path 2749 * 1st. 2750 */ 2751 if (right_has_empty) 2752 ocfs2_remove_empty_extent(left_leaf_el); 2753 2754 ret = ocfs2_journal_dirty(handle, et_root_bh); 2755 if (ret) 2756 mlog_errno(ret); 2757 2758 *deleted = 1; 2759 } else 2760 ocfs2_complete_edge_insert(handle, left_path, right_path, 2761 subtree_index); 2762 2763 out: 2764 return ret; 2765 } 2766 2767 /* 2768 * Given a full path, determine what cpos value would return us a path 2769 * containing the leaf immediately to the right of the current one. 2770 * 2771 * Will return zero if the path passed in is already the rightmost path. 2772 * 2773 * This looks similar, but is subtly different to 2774 * ocfs2_find_cpos_for_left_leaf(). 2775 */ 2776 static int ocfs2_find_cpos_for_right_leaf(struct super_block *sb, 2777 struct ocfs2_path *path, u32 *cpos) 2778 { 2779 int i, j, ret = 0; 2780 u64 blkno; 2781 struct ocfs2_extent_list *el; 2782 2783 *cpos = 0; 2784 2785 if (path->p_tree_depth == 0) 2786 return 0; 2787 2788 blkno = path_leaf_bh(path)->b_blocknr; 2789 2790 /* Start at the tree node just above the leaf and work our way up. */ 2791 i = path->p_tree_depth - 1; 2792 while (i >= 0) { 2793 int next_free; 2794 2795 el = path->p_node[i].el; 2796 2797 /* 2798 * Find the extent record just after the one in our 2799 * path. 2800 */ 2801 next_free = le16_to_cpu(el->l_next_free_rec); 2802 for(j = 0; j < le16_to_cpu(el->l_next_free_rec); j++) { 2803 if (le64_to_cpu(el->l_recs[j].e_blkno) == blkno) { 2804 if (j == (next_free - 1)) { 2805 if (i == 0) { 2806 /* 2807 * We've determined that the 2808 * path specified is already 2809 * the rightmost one - return a 2810 * cpos of zero. 2811 */ 2812 goto out; 2813 } 2814 /* 2815 * The rightmost record points to our 2816 * leaf - we need to travel up the 2817 * tree one level. 2818 */ 2819 goto next_node; 2820 } 2821 2822 *cpos = le32_to_cpu(el->l_recs[j + 1].e_cpos); 2823 goto out; 2824 } 2825 } 2826 2827 /* 2828 * If we got here, we never found a valid node where 2829 * the tree indicated one should be. 2830 */ 2831 ocfs2_error(sb, 2832 "Invalid extent tree at extent block %llu\n", 2833 (unsigned long long)blkno); 2834 ret = -EROFS; 2835 goto out; 2836 2837 next_node: 2838 blkno = path->p_node[i].bh->b_blocknr; 2839 i--; 2840 } 2841 2842 out: 2843 return ret; 2844 } 2845 2846 static int ocfs2_rotate_rightmost_leaf_left(handle_t *handle, 2847 struct ocfs2_extent_tree *et, 2848 struct ocfs2_path *path) 2849 { 2850 int ret; 2851 struct buffer_head *bh = path_leaf_bh(path); 2852 struct ocfs2_extent_list *el = path_leaf_el(path); 2853 2854 if (!ocfs2_is_empty_extent(&el->l_recs[0])) 2855 return 0; 2856 2857 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, path, 2858 path_num_items(path) - 1); 2859 if (ret) { 2860 mlog_errno(ret); 2861 goto out; 2862 } 2863 2864 ocfs2_remove_empty_extent(el); 2865 2866 ret = ocfs2_journal_dirty(handle, bh); 2867 if (ret) 2868 mlog_errno(ret); 2869 2870 out: 2871 return ret; 2872 } 2873 2874 static int __ocfs2_rotate_tree_left(handle_t *handle, 2875 struct ocfs2_extent_tree *et, 2876 int orig_credits, 2877 struct ocfs2_path *path, 2878 struct ocfs2_cached_dealloc_ctxt *dealloc, 2879 struct ocfs2_path **empty_extent_path) 2880 { 2881 int ret, subtree_root, deleted; 2882 u32 right_cpos; 2883 struct ocfs2_path *left_path = NULL; 2884 struct ocfs2_path *right_path = NULL; 2885 struct super_block *sb = ocfs2_metadata_cache_get_super(et->et_ci); 2886 2887 BUG_ON(!ocfs2_is_empty_extent(&(path_leaf_el(path)->l_recs[0]))); 2888 2889 *empty_extent_path = NULL; 2890 2891 ret = ocfs2_find_cpos_for_right_leaf(sb, path, &right_cpos); 2892 if (ret) { 2893 mlog_errno(ret); 2894 goto out; 2895 } 2896 2897 left_path = ocfs2_new_path_from_path(path); 2898 if (!left_path) { 2899 ret = -ENOMEM; 2900 mlog_errno(ret); 2901 goto out; 2902 } 2903 2904 ocfs2_cp_path(left_path, path); 2905 2906 right_path = ocfs2_new_path_from_path(path); 2907 if (!right_path) { 2908 ret = -ENOMEM; 2909 mlog_errno(ret); 2910 goto out; 2911 } 2912 2913 while (right_cpos) { 2914 ret = ocfs2_find_path(et->et_ci, right_path, right_cpos); 2915 if (ret) { 2916 mlog_errno(ret); 2917 goto out; 2918 } 2919 2920 subtree_root = ocfs2_find_subtree_root(et, left_path, 2921 right_path); 2922 2923 mlog(0, "Subtree root at index %d (blk %llu, depth %d)\n", 2924 subtree_root, 2925 (unsigned long long) 2926 right_path->p_node[subtree_root].bh->b_blocknr, 2927 right_path->p_tree_depth); 2928 2929 ret = ocfs2_extend_rotate_transaction(handle, subtree_root, 2930 orig_credits, left_path); 2931 if (ret) { 2932 mlog_errno(ret); 2933 goto out; 2934 } 2935 2936 /* 2937 * Caller might still want to make changes to the 2938 * tree root, so re-add it to the journal here. 2939 */ 2940 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, 2941 left_path, 0); 2942 if (ret) { 2943 mlog_errno(ret); 2944 goto out; 2945 } 2946 2947 ret = ocfs2_rotate_subtree_left(handle, et, left_path, 2948 right_path, subtree_root, 2949 dealloc, &deleted); 2950 if (ret == -EAGAIN) { 2951 /* 2952 * The rotation has to temporarily stop due to 2953 * the right subtree having an empty 2954 * extent. Pass it back to the caller for a 2955 * fixup. 2956 */ 2957 *empty_extent_path = right_path; 2958 right_path = NULL; 2959 goto out; 2960 } 2961 if (ret) { 2962 mlog_errno(ret); 2963 goto out; 2964 } 2965 2966 /* 2967 * The subtree rotate might have removed records on 2968 * the rightmost edge. If so, then rotation is 2969 * complete. 2970 */ 2971 if (deleted) 2972 break; 2973 2974 ocfs2_mv_path(left_path, right_path); 2975 2976 ret = ocfs2_find_cpos_for_right_leaf(sb, left_path, 2977 &right_cpos); 2978 if (ret) { 2979 mlog_errno(ret); 2980 goto out; 2981 } 2982 } 2983 2984 out: 2985 ocfs2_free_path(right_path); 2986 ocfs2_free_path(left_path); 2987 2988 return ret; 2989 } 2990 2991 static int ocfs2_remove_rightmost_path(handle_t *handle, 2992 struct ocfs2_extent_tree *et, 2993 struct ocfs2_path *path, 2994 struct ocfs2_cached_dealloc_ctxt *dealloc) 2995 { 2996 int ret, subtree_index; 2997 u32 cpos; 2998 struct ocfs2_path *left_path = NULL; 2999 struct ocfs2_extent_block *eb; 3000 struct ocfs2_extent_list *el; 3001 3002 3003 ret = ocfs2_et_sanity_check(et); 3004 if (ret) 3005 goto out; 3006 /* 3007 * There's two ways we handle this depending on 3008 * whether path is the only existing one. 3009 */ 3010 ret = ocfs2_extend_rotate_transaction(handle, 0, 3011 handle->h_buffer_credits, 3012 path); 3013 if (ret) { 3014 mlog_errno(ret); 3015 goto out; 3016 } 3017 3018 ret = ocfs2_journal_access_path(et->et_ci, handle, path); 3019 if (ret) { 3020 mlog_errno(ret); 3021 goto out; 3022 } 3023 3024 ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci), 3025 path, &cpos); 3026 if (ret) { 3027 mlog_errno(ret); 3028 goto out; 3029 } 3030 3031 if (cpos) { 3032 /* 3033 * We have a path to the left of this one - it needs 3034 * an update too. 3035 */ 3036 left_path = ocfs2_new_path_from_path(path); 3037 if (!left_path) { 3038 ret = -ENOMEM; 3039 mlog_errno(ret); 3040 goto out; 3041 } 3042 3043 ret = ocfs2_find_path(et->et_ci, left_path, cpos); 3044 if (ret) { 3045 mlog_errno(ret); 3046 goto out; 3047 } 3048 3049 ret = ocfs2_journal_access_path(et->et_ci, handle, left_path); 3050 if (ret) { 3051 mlog_errno(ret); 3052 goto out; 3053 } 3054 3055 subtree_index = ocfs2_find_subtree_root(et, left_path, path); 3056 3057 ocfs2_unlink_subtree(handle, et, left_path, path, 3058 subtree_index, dealloc); 3059 ret = ocfs2_update_edge_lengths(handle, et, subtree_index, 3060 left_path); 3061 if (ret) { 3062 mlog_errno(ret); 3063 goto out; 3064 } 3065 3066 eb = (struct ocfs2_extent_block *)path_leaf_bh(left_path)->b_data; 3067 ocfs2_et_set_last_eb_blk(et, le64_to_cpu(eb->h_blkno)); 3068 } else { 3069 /* 3070 * 'path' is also the leftmost path which 3071 * means it must be the only one. This gets 3072 * handled differently because we want to 3073 * revert the root back to having extents 3074 * in-line. 3075 */ 3076 ocfs2_unlink_path(handle, et, dealloc, path, 1); 3077 3078 el = et->et_root_el; 3079 el->l_tree_depth = 0; 3080 el->l_next_free_rec = 0; 3081 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec)); 3082 3083 ocfs2_et_set_last_eb_blk(et, 0); 3084 } 3085 3086 ocfs2_journal_dirty(handle, path_root_bh(path)); 3087 3088 out: 3089 ocfs2_free_path(left_path); 3090 return ret; 3091 } 3092 3093 /* 3094 * Left rotation of btree records. 3095 * 3096 * In many ways, this is (unsurprisingly) the opposite of right 3097 * rotation. We start at some non-rightmost path containing an empty 3098 * extent in the leaf block. The code works its way to the rightmost 3099 * path by rotating records to the left in every subtree. 3100 * 3101 * This is used by any code which reduces the number of extent records 3102 * in a leaf. After removal, an empty record should be placed in the 3103 * leftmost list position. 3104 * 3105 * This won't handle a length update of the rightmost path records if 3106 * the rightmost tree leaf record is removed so the caller is 3107 * responsible for detecting and correcting that. 3108 */ 3109 static int ocfs2_rotate_tree_left(handle_t *handle, 3110 struct ocfs2_extent_tree *et, 3111 struct ocfs2_path *path, 3112 struct ocfs2_cached_dealloc_ctxt *dealloc) 3113 { 3114 int ret, orig_credits = handle->h_buffer_credits; 3115 struct ocfs2_path *tmp_path = NULL, *restart_path = NULL; 3116 struct ocfs2_extent_block *eb; 3117 struct ocfs2_extent_list *el; 3118 3119 el = path_leaf_el(path); 3120 if (!ocfs2_is_empty_extent(&el->l_recs[0])) 3121 return 0; 3122 3123 if (path->p_tree_depth == 0) { 3124 rightmost_no_delete: 3125 /* 3126 * Inline extents. This is trivially handled, so do 3127 * it up front. 3128 */ 3129 ret = ocfs2_rotate_rightmost_leaf_left(handle, et, path); 3130 if (ret) 3131 mlog_errno(ret); 3132 goto out; 3133 } 3134 3135 /* 3136 * Handle rightmost branch now. There's several cases: 3137 * 1) simple rotation leaving records in there. That's trivial. 3138 * 2) rotation requiring a branch delete - there's no more 3139 * records left. Two cases of this: 3140 * a) There are branches to the left. 3141 * b) This is also the leftmost (the only) branch. 3142 * 3143 * 1) is handled via ocfs2_rotate_rightmost_leaf_left() 3144 * 2a) we need the left branch so that we can update it with the unlink 3145 * 2b) we need to bring the root back to inline extents. 3146 */ 3147 3148 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data; 3149 el = &eb->h_list; 3150 if (eb->h_next_leaf_blk == 0) { 3151 /* 3152 * This gets a bit tricky if we're going to delete the 3153 * rightmost path. Get the other cases out of the way 3154 * 1st. 3155 */ 3156 if (le16_to_cpu(el->l_next_free_rec) > 1) 3157 goto rightmost_no_delete; 3158 3159 if (le16_to_cpu(el->l_next_free_rec) == 0) { 3160 ret = -EIO; 3161 ocfs2_error(ocfs2_metadata_cache_get_super(et->et_ci), 3162 "Owner %llu has empty extent block at %llu", 3163 (unsigned long long)ocfs2_metadata_cache_owner(et->et_ci), 3164 (unsigned long long)le64_to_cpu(eb->h_blkno)); 3165 goto out; 3166 } 3167 3168 /* 3169 * XXX: The caller can not trust "path" any more after 3170 * this as it will have been deleted. What do we do? 3171 * 3172 * In theory the rotate-for-merge code will never get 3173 * here because it'll always ask for a rotate in a 3174 * nonempty list. 3175 */ 3176 3177 ret = ocfs2_remove_rightmost_path(handle, et, path, 3178 dealloc); 3179 if (ret) 3180 mlog_errno(ret); 3181 goto out; 3182 } 3183 3184 /* 3185 * Now we can loop, remembering the path we get from -EAGAIN 3186 * and restarting from there. 3187 */ 3188 try_rotate: 3189 ret = __ocfs2_rotate_tree_left(handle, et, orig_credits, path, 3190 dealloc, &restart_path); 3191 if (ret && ret != -EAGAIN) { 3192 mlog_errno(ret); 3193 goto out; 3194 } 3195 3196 while (ret == -EAGAIN) { 3197 tmp_path = restart_path; 3198 restart_path = NULL; 3199 3200 ret = __ocfs2_rotate_tree_left(handle, et, orig_credits, 3201 tmp_path, dealloc, 3202 &restart_path); 3203 if (ret && ret != -EAGAIN) { 3204 mlog_errno(ret); 3205 goto out; 3206 } 3207 3208 ocfs2_free_path(tmp_path); 3209 tmp_path = NULL; 3210 3211 if (ret == 0) 3212 goto try_rotate; 3213 } 3214 3215 out: 3216 ocfs2_free_path(tmp_path); 3217 ocfs2_free_path(restart_path); 3218 return ret; 3219 } 3220 3221 static void ocfs2_cleanup_merge(struct ocfs2_extent_list *el, 3222 int index) 3223 { 3224 struct ocfs2_extent_rec *rec = &el->l_recs[index]; 3225 unsigned int size; 3226 3227 if (rec->e_leaf_clusters == 0) { 3228 /* 3229 * We consumed all of the merged-from record. An empty 3230 * extent cannot exist anywhere but the 1st array 3231 * position, so move things over if the merged-from 3232 * record doesn't occupy that position. 3233 * 3234 * This creates a new empty extent so the caller 3235 * should be smart enough to have removed any existing 3236 * ones. 3237 */ 3238 if (index > 0) { 3239 BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0])); 3240 size = index * sizeof(struct ocfs2_extent_rec); 3241 memmove(&el->l_recs[1], &el->l_recs[0], size); 3242 } 3243 3244 /* 3245 * Always memset - the caller doesn't check whether it 3246 * created an empty extent, so there could be junk in 3247 * the other fields. 3248 */ 3249 memset(&el->l_recs[0], 0, sizeof(struct ocfs2_extent_rec)); 3250 } 3251 } 3252 3253 static int ocfs2_get_right_path(struct ocfs2_extent_tree *et, 3254 struct ocfs2_path *left_path, 3255 struct ocfs2_path **ret_right_path) 3256 { 3257 int ret; 3258 u32 right_cpos; 3259 struct ocfs2_path *right_path = NULL; 3260 struct ocfs2_extent_list *left_el; 3261 3262 *ret_right_path = NULL; 3263 3264 /* This function shouldn't be called for non-trees. */ 3265 BUG_ON(left_path->p_tree_depth == 0); 3266 3267 left_el = path_leaf_el(left_path); 3268 BUG_ON(left_el->l_next_free_rec != left_el->l_count); 3269 3270 ret = ocfs2_find_cpos_for_right_leaf(ocfs2_metadata_cache_get_super(et->et_ci), 3271 left_path, &right_cpos); 3272 if (ret) { 3273 mlog_errno(ret); 3274 goto out; 3275 } 3276 3277 /* This function shouldn't be called for the rightmost leaf. */ 3278 BUG_ON(right_cpos == 0); 3279 3280 right_path = ocfs2_new_path_from_path(left_path); 3281 if (!right_path) { 3282 ret = -ENOMEM; 3283 mlog_errno(ret); 3284 goto out; 3285 } 3286 3287 ret = ocfs2_find_path(et->et_ci, right_path, right_cpos); 3288 if (ret) { 3289 mlog_errno(ret); 3290 goto out; 3291 } 3292 3293 *ret_right_path = right_path; 3294 out: 3295 if (ret) 3296 ocfs2_free_path(right_path); 3297 return ret; 3298 } 3299 3300 /* 3301 * Remove split_rec clusters from the record at index and merge them 3302 * onto the beginning of the record "next" to it. 3303 * For index < l_count - 1, the next means the extent rec at index + 1. 3304 * For index == l_count - 1, the "next" means the 1st extent rec of the 3305 * next extent block. 3306 */ 3307 static int ocfs2_merge_rec_right(struct ocfs2_path *left_path, 3308 handle_t *handle, 3309 struct ocfs2_extent_tree *et, 3310 struct ocfs2_extent_rec *split_rec, 3311 int index) 3312 { 3313 int ret, next_free, i; 3314 unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters); 3315 struct ocfs2_extent_rec *left_rec; 3316 struct ocfs2_extent_rec *right_rec; 3317 struct ocfs2_extent_list *right_el; 3318 struct ocfs2_path *right_path = NULL; 3319 int subtree_index = 0; 3320 struct ocfs2_extent_list *el = path_leaf_el(left_path); 3321 struct buffer_head *bh = path_leaf_bh(left_path); 3322 struct buffer_head *root_bh = NULL; 3323 3324 BUG_ON(index >= le16_to_cpu(el->l_next_free_rec)); 3325 left_rec = &el->l_recs[index]; 3326 3327 if (index == le16_to_cpu(el->l_next_free_rec) - 1 && 3328 le16_to_cpu(el->l_next_free_rec) == le16_to_cpu(el->l_count)) { 3329 /* we meet with a cross extent block merge. */ 3330 ret = ocfs2_get_right_path(et, left_path, &right_path); 3331 if (ret) { 3332 mlog_errno(ret); 3333 goto out; 3334 } 3335 3336 right_el = path_leaf_el(right_path); 3337 next_free = le16_to_cpu(right_el->l_next_free_rec); 3338 BUG_ON(next_free <= 0); 3339 right_rec = &right_el->l_recs[0]; 3340 if (ocfs2_is_empty_extent(right_rec)) { 3341 BUG_ON(next_free <= 1); 3342 right_rec = &right_el->l_recs[1]; 3343 } 3344 3345 BUG_ON(le32_to_cpu(left_rec->e_cpos) + 3346 le16_to_cpu(left_rec->e_leaf_clusters) != 3347 le32_to_cpu(right_rec->e_cpos)); 3348 3349 subtree_index = ocfs2_find_subtree_root(et, left_path, 3350 right_path); 3351 3352 ret = ocfs2_extend_rotate_transaction(handle, subtree_index, 3353 handle->h_buffer_credits, 3354 right_path); 3355 if (ret) { 3356 mlog_errno(ret); 3357 goto out; 3358 } 3359 3360 root_bh = left_path->p_node[subtree_index].bh; 3361 BUG_ON(root_bh != right_path->p_node[subtree_index].bh); 3362 3363 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path, 3364 subtree_index); 3365 if (ret) { 3366 mlog_errno(ret); 3367 goto out; 3368 } 3369 3370 for (i = subtree_index + 1; 3371 i < path_num_items(right_path); i++) { 3372 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, 3373 right_path, i); 3374 if (ret) { 3375 mlog_errno(ret); 3376 goto out; 3377 } 3378 3379 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, 3380 left_path, i); 3381 if (ret) { 3382 mlog_errno(ret); 3383 goto out; 3384 } 3385 } 3386 3387 } else { 3388 BUG_ON(index == le16_to_cpu(el->l_next_free_rec) - 1); 3389 right_rec = &el->l_recs[index + 1]; 3390 } 3391 3392 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, left_path, 3393 path_num_items(left_path) - 1); 3394 if (ret) { 3395 mlog_errno(ret); 3396 goto out; 3397 } 3398 3399 le16_add_cpu(&left_rec->e_leaf_clusters, -split_clusters); 3400 3401 le32_add_cpu(&right_rec->e_cpos, -split_clusters); 3402 le64_add_cpu(&right_rec->e_blkno, 3403 -ocfs2_clusters_to_blocks(ocfs2_metadata_cache_get_super(et->et_ci), 3404 split_clusters)); 3405 le16_add_cpu(&right_rec->e_leaf_clusters, split_clusters); 3406 3407 ocfs2_cleanup_merge(el, index); 3408 3409 ret = ocfs2_journal_dirty(handle, bh); 3410 if (ret) 3411 mlog_errno(ret); 3412 3413 if (right_path) { 3414 ret = ocfs2_journal_dirty(handle, path_leaf_bh(right_path)); 3415 if (ret) 3416 mlog_errno(ret); 3417 3418 ocfs2_complete_edge_insert(handle, left_path, right_path, 3419 subtree_index); 3420 } 3421 out: 3422 if (right_path) 3423 ocfs2_free_path(right_path); 3424 return ret; 3425 } 3426 3427 static int ocfs2_get_left_path(struct ocfs2_extent_tree *et, 3428 struct ocfs2_path *right_path, 3429 struct ocfs2_path **ret_left_path) 3430 { 3431 int ret; 3432 u32 left_cpos; 3433 struct ocfs2_path *left_path = NULL; 3434 3435 *ret_left_path = NULL; 3436 3437 /* This function shouldn't be called for non-trees. */ 3438 BUG_ON(right_path->p_tree_depth == 0); 3439 3440 ret = ocfs2_find_cpos_for_left_leaf(ocfs2_metadata_cache_get_super(et->et_ci), 3441 right_path, &left_cpos); 3442 if (ret) { 3443 mlog_errno(ret); 3444 goto out; 3445 } 3446 3447 /* This function shouldn't be called for the leftmost leaf. */ 3448 BUG_ON(left_cpos == 0); 3449 3450 left_path = ocfs2_new_path_from_path(right_path); 3451 if (!left_path) { 3452 ret = -ENOMEM; 3453 mlog_errno(ret); 3454 goto out; 3455 } 3456 3457 ret = ocfs2_find_path(et->et_ci, left_path, left_cpos); 3458 if (ret) { 3459 mlog_errno(ret); 3460 goto out; 3461 } 3462 3463 *ret_left_path = left_path; 3464 out: 3465 if (ret) 3466 ocfs2_free_path(left_path); 3467 return ret; 3468 } 3469 3470 /* 3471 * Remove split_rec clusters from the record at index and merge them 3472 * onto the tail of the record "before" it. 3473 * For index > 0, the "before" means the extent rec at index - 1. 3474 * 3475 * For index == 0, the "before" means the last record of the previous 3476 * extent block. And there is also a situation that we may need to 3477 * remove the rightmost leaf extent block in the right_path and change 3478 * the right path to indicate the new rightmost path. 3479 */ 3480 static int ocfs2_merge_rec_left(struct ocfs2_path *right_path, 3481 handle_t *handle, 3482 struct ocfs2_extent_tree *et, 3483 struct ocfs2_extent_rec *split_rec, 3484 struct ocfs2_cached_dealloc_ctxt *dealloc, 3485 int index) 3486 { 3487 int ret, i, subtree_index = 0, has_empty_extent = 0; 3488 unsigned int split_clusters = le16_to_cpu(split_rec->e_leaf_clusters); 3489 struct ocfs2_extent_rec *left_rec; 3490 struct ocfs2_extent_rec *right_rec; 3491 struct ocfs2_extent_list *el = path_leaf_el(right_path); 3492 struct buffer_head *bh = path_leaf_bh(right_path); 3493 struct buffer_head *root_bh = NULL; 3494 struct ocfs2_path *left_path = NULL; 3495 struct ocfs2_extent_list *left_el; 3496 3497 BUG_ON(index < 0); 3498 3499 right_rec = &el->l_recs[index]; 3500 if (index == 0) { 3501 /* we meet with a cross extent block merge. */ 3502 ret = ocfs2_get_left_path(et, right_path, &left_path); 3503 if (ret) { 3504 mlog_errno(ret); 3505 goto out; 3506 } 3507 3508 left_el = path_leaf_el(left_path); 3509 BUG_ON(le16_to_cpu(left_el->l_next_free_rec) != 3510 le16_to_cpu(left_el->l_count)); 3511 3512 left_rec = &left_el->l_recs[ 3513 le16_to_cpu(left_el->l_next_free_rec) - 1]; 3514 BUG_ON(le32_to_cpu(left_rec->e_cpos) + 3515 le16_to_cpu(left_rec->e_leaf_clusters) != 3516 le32_to_cpu(split_rec->e_cpos)); 3517 3518 subtree_index = ocfs2_find_subtree_root(et, left_path, 3519 right_path); 3520 3521 ret = ocfs2_extend_rotate_transaction(handle, subtree_index, 3522 handle->h_buffer_credits, 3523 left_path); 3524 if (ret) { 3525 mlog_errno(ret); 3526 goto out; 3527 } 3528 3529 root_bh = left_path->p_node[subtree_index].bh; 3530 BUG_ON(root_bh != right_path->p_node[subtree_index].bh); 3531 3532 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path, 3533 subtree_index); 3534 if (ret) { 3535 mlog_errno(ret); 3536 goto out; 3537 } 3538 3539 for (i = subtree_index + 1; 3540 i < path_num_items(right_path); i++) { 3541 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, 3542 right_path, i); 3543 if (ret) { 3544 mlog_errno(ret); 3545 goto out; 3546 } 3547 3548 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, 3549 left_path, i); 3550 if (ret) { 3551 mlog_errno(ret); 3552 goto out; 3553 } 3554 } 3555 } else { 3556 left_rec = &el->l_recs[index - 1]; 3557 if (ocfs2_is_empty_extent(&el->l_recs[0])) 3558 has_empty_extent = 1; 3559 } 3560 3561 ret = ocfs2_path_bh_journal_access(handle, et->et_ci, right_path, 3562 path_num_items(right_path) - 1); 3563 if (ret) { 3564 mlog_errno(ret); 3565 goto out; 3566 } 3567 3568 if (has_empty_extent && index == 1) { 3569 /* 3570 * The easy case - we can just plop the record right in. 3571 */ 3572 *left_rec = *split_rec; 3573 3574 has_empty_extent = 0; 3575 } else 3576 le16_add_cpu(&left_rec->e_leaf_clusters, split_clusters); 3577 3578 le32_add_cpu(&right_rec->e_cpos, split_clusters); 3579 le64_add_cpu(&right_rec->e_blkno, 3580 ocfs2_clusters_to_blocks(ocfs2_metadata_cache_get_super(et->et_ci), 3581 split_clusters)); 3582 le16_add_cpu(&right_rec->e_leaf_clusters, -split_clusters); 3583 3584 ocfs2_cleanup_merge(el, index); 3585 3586 ret = ocfs2_journal_dirty(handle, bh); 3587 if (ret) 3588 mlog_errno(ret); 3589 3590 if (left_path) { 3591 ret = ocfs2_journal_dirty(handle, path_leaf_bh(left_path)); 3592 if (ret) 3593 mlog_errno(ret); 3594 3595 /* 3596 * In the situation that the right_rec is empty and the extent 3597 * block is empty also, ocfs2_complete_edge_insert can't handle 3598 * it and we need to delete the right extent block. 3599 */ 3600 if (le16_to_cpu(right_rec->e_leaf_clusters) == 0 && 3601 le16_to_cpu(el->l_next_free_rec) == 1) { 3602 3603 ret = ocfs2_remove_rightmost_path(handle, et, 3604 right_path, 3605 dealloc); 3606 if (ret) { 3607 mlog_errno(ret); 3608 goto out; 3609 } 3610 3611 /* Now the rightmost extent block has been deleted. 3612 * So we use the new rightmost path. 3613 */ 3614 ocfs2_mv_path(right_path, left_path); 3615 left_path = NULL; 3616 } else 3617 ocfs2_complete_edge_insert(handle, left_path, 3618 right_path, subtree_index); 3619 } 3620 out: 3621 if (left_path) 3622 ocfs2_free_path(left_path); 3623 return ret; 3624 } 3625 3626 static int ocfs2_try_to_merge_extent(handle_t *handle, 3627 struct ocfs2_extent_tree *et, 3628 struct ocfs2_path *path, 3629 int split_index, 3630 struct ocfs2_extent_rec *split_rec, 3631 struct ocfs2_cached_dealloc_ctxt *dealloc, 3632 struct ocfs2_merge_ctxt *ctxt) 3633 { 3634 int ret = 0; 3635 struct ocfs2_extent_list *el = path_leaf_el(path); 3636 struct ocfs2_extent_rec *rec = &el->l_recs[split_index]; 3637 3638 BUG_ON(ctxt->c_contig_type == CONTIG_NONE); 3639 3640 if (ctxt->c_split_covers_rec && ctxt->c_has_empty_extent) { 3641 /* 3642 * The merge code will need to create an empty 3643 * extent to take the place of the newly 3644 * emptied slot. Remove any pre-existing empty 3645 * extents - having more than one in a leaf is 3646 * illegal. 3647 */ 3648 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc); 3649 if (ret) { 3650 mlog_errno(ret); 3651 goto out; 3652 } 3653 split_index--; 3654 rec = &el->l_recs[split_index]; 3655 } 3656 3657 if (ctxt->c_contig_type == CONTIG_LEFTRIGHT) { 3658 /* 3659 * Left-right contig implies this. 3660 */ 3661 BUG_ON(!ctxt->c_split_covers_rec); 3662 3663 /* 3664 * Since the leftright insert always covers the entire 3665 * extent, this call will delete the insert record 3666 * entirely, resulting in an empty extent record added to 3667 * the extent block. 3668 * 3669 * Since the adding of an empty extent shifts 3670 * everything back to the right, there's no need to 3671 * update split_index here. 3672 * 3673 * When the split_index is zero, we need to merge it to the 3674 * prevoius extent block. It is more efficient and easier 3675 * if we do merge_right first and merge_left later. 3676 */ 3677 ret = ocfs2_merge_rec_right(path, handle, et, split_rec, 3678 split_index); 3679 if (ret) { 3680 mlog_errno(ret); 3681 goto out; 3682 } 3683 3684 /* 3685 * We can only get this from logic error above. 3686 */ 3687 BUG_ON(!ocfs2_is_empty_extent(&el->l_recs[0])); 3688 3689 /* The merge left us with an empty extent, remove it. */ 3690 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc); 3691 if (ret) { 3692 mlog_errno(ret); 3693 goto out; 3694 } 3695 3696 rec = &el->l_recs[split_index]; 3697 3698 /* 3699 * Note that we don't pass split_rec here on purpose - 3700 * we've merged it into the rec already. 3701 */ 3702 ret = ocfs2_merge_rec_left(path, handle, et, rec, 3703 dealloc, split_index); 3704 3705 if (ret) { 3706 mlog_errno(ret); 3707 goto out; 3708 } 3709 3710 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc); 3711 /* 3712 * Error from this last rotate is not critical, so 3713 * print but don't bubble it up. 3714 */ 3715 if (ret) 3716 mlog_errno(ret); 3717 ret = 0; 3718 } else { 3719 /* 3720 * Merge a record to the left or right. 3721 * 3722 * 'contig_type' is relative to the existing record, 3723 * so for example, if we're "right contig", it's to 3724 * the record on the left (hence the left merge). 3725 */ 3726 if (ctxt->c_contig_type == CONTIG_RIGHT) { 3727 ret = ocfs2_merge_rec_left(path, handle, et, 3728 split_rec, dealloc, 3729 split_index); 3730 if (ret) { 3731 mlog_errno(ret); 3732 goto out; 3733 } 3734 } else { 3735 ret = ocfs2_merge_rec_right(path, handle, 3736 et, split_rec, 3737 split_index); 3738 if (ret) { 3739 mlog_errno(ret); 3740 goto out; 3741 } 3742 } 3743 3744 if (ctxt->c_split_covers_rec) { 3745 /* 3746 * The merge may have left an empty extent in 3747 * our leaf. Try to rotate it away. 3748 */ 3749 ret = ocfs2_rotate_tree_left(handle, et, path, 3750 dealloc); 3751 if (ret) 3752 mlog_errno(ret); 3753 ret = 0; 3754 } 3755 } 3756 3757 out: 3758 return ret; 3759 } 3760 3761 static void ocfs2_subtract_from_rec(struct super_block *sb, 3762 enum ocfs2_split_type split, 3763 struct ocfs2_extent_rec *rec, 3764 struct ocfs2_extent_rec *split_rec) 3765 { 3766 u64 len_blocks; 3767 3768 len_blocks = ocfs2_clusters_to_blocks(sb, 3769 le16_to_cpu(split_rec->e_leaf_clusters)); 3770 3771 if (split == SPLIT_LEFT) { 3772 /* 3773 * Region is on the left edge of the existing 3774 * record. 3775 */ 3776 le32_add_cpu(&rec->e_cpos, 3777 le16_to_cpu(split_rec->e_leaf_clusters)); 3778 le64_add_cpu(&rec->e_blkno, len_blocks); 3779 le16_add_cpu(&rec->e_leaf_clusters, 3780 -le16_to_cpu(split_rec->e_leaf_clusters)); 3781 } else { 3782 /* 3783 * Region is on the right edge of the existing 3784 * record. 3785 */ 3786 le16_add_cpu(&rec->e_leaf_clusters, 3787 -le16_to_cpu(split_rec->e_leaf_clusters)); 3788 } 3789 } 3790 3791 /* 3792 * Do the final bits of extent record insertion at the target leaf 3793 * list. If this leaf is part of an allocation tree, it is assumed 3794 * that the tree above has been prepared. 3795 */ 3796 static void ocfs2_insert_at_leaf(struct ocfs2_extent_rec *insert_rec, 3797 struct ocfs2_extent_list *el, 3798 struct ocfs2_insert_type *insert, 3799 struct inode *inode) 3800 { 3801 int i = insert->ins_contig_index; 3802 unsigned int range; 3803 struct ocfs2_extent_rec *rec; 3804 3805 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0); 3806 3807 if (insert->ins_split != SPLIT_NONE) { 3808 i = ocfs2_search_extent_list(el, le32_to_cpu(insert_rec->e_cpos)); 3809 BUG_ON(i == -1); 3810 rec = &el->l_recs[i]; 3811 ocfs2_subtract_from_rec(inode->i_sb, insert->ins_split, rec, 3812 insert_rec); 3813 goto rotate; 3814 } 3815 3816 /* 3817 * Contiguous insert - either left or right. 3818 */ 3819 if (insert->ins_contig != CONTIG_NONE) { 3820 rec = &el->l_recs[i]; 3821 if (insert->ins_contig == CONTIG_LEFT) { 3822 rec->e_blkno = insert_rec->e_blkno; 3823 rec->e_cpos = insert_rec->e_cpos; 3824 } 3825 le16_add_cpu(&rec->e_leaf_clusters, 3826 le16_to_cpu(insert_rec->e_leaf_clusters)); 3827 return; 3828 } 3829 3830 /* 3831 * Handle insert into an empty leaf. 3832 */ 3833 if (le16_to_cpu(el->l_next_free_rec) == 0 || 3834 ((le16_to_cpu(el->l_next_free_rec) == 1) && 3835 ocfs2_is_empty_extent(&el->l_recs[0]))) { 3836 el->l_recs[0] = *insert_rec; 3837 el->l_next_free_rec = cpu_to_le16(1); 3838 return; 3839 } 3840 3841 /* 3842 * Appending insert. 3843 */ 3844 if (insert->ins_appending == APPEND_TAIL) { 3845 i = le16_to_cpu(el->l_next_free_rec) - 1; 3846 rec = &el->l_recs[i]; 3847 range = le32_to_cpu(rec->e_cpos) 3848 + le16_to_cpu(rec->e_leaf_clusters); 3849 BUG_ON(le32_to_cpu(insert_rec->e_cpos) < range); 3850 3851 mlog_bug_on_msg(le16_to_cpu(el->l_next_free_rec) >= 3852 le16_to_cpu(el->l_count), 3853 "inode %lu, depth %u, count %u, next free %u, " 3854 "rec.cpos %u, rec.clusters %u, " 3855 "insert.cpos %u, insert.clusters %u\n", 3856 inode->i_ino, 3857 le16_to_cpu(el->l_tree_depth), 3858 le16_to_cpu(el->l_count), 3859 le16_to_cpu(el->l_next_free_rec), 3860 le32_to_cpu(el->l_recs[i].e_cpos), 3861 le16_to_cpu(el->l_recs[i].e_leaf_clusters), 3862 le32_to_cpu(insert_rec->e_cpos), 3863 le16_to_cpu(insert_rec->e_leaf_clusters)); 3864 i++; 3865 el->l_recs[i] = *insert_rec; 3866 le16_add_cpu(&el->l_next_free_rec, 1); 3867 return; 3868 } 3869 3870 rotate: 3871 /* 3872 * Ok, we have to rotate. 3873 * 3874 * At this point, it is safe to assume that inserting into an 3875 * empty leaf and appending to a leaf have both been handled 3876 * above. 3877 * 3878 * This leaf needs to have space, either by the empty 1st 3879 * extent record, or by virtue of an l_next_rec < l_count. 3880 */ 3881 ocfs2_rotate_leaf(el, insert_rec); 3882 } 3883 3884 static void ocfs2_adjust_rightmost_records(struct inode *inode, 3885 handle_t *handle, 3886 struct ocfs2_path *path, 3887 struct ocfs2_extent_rec *insert_rec) 3888 { 3889 int ret, i, next_free; 3890 struct buffer_head *bh; 3891 struct ocfs2_extent_list *el; 3892 struct ocfs2_extent_rec *rec; 3893 3894 /* 3895 * Update everything except the leaf block. 3896 */ 3897 for (i = 0; i < path->p_tree_depth; i++) { 3898 bh = path->p_node[i].bh; 3899 el = path->p_node[i].el; 3900 3901 next_free = le16_to_cpu(el->l_next_free_rec); 3902 if (next_free == 0) { 3903 ocfs2_error(inode->i_sb, 3904 "Dinode %llu has a bad extent list", 3905 (unsigned long long)OCFS2_I(inode)->ip_blkno); 3906 ret = -EIO; 3907 return; 3908 } 3909 3910 rec = &el->l_recs[next_free - 1]; 3911 3912 rec->e_int_clusters = insert_rec->e_cpos; 3913 le32_add_cpu(&rec->e_int_clusters, 3914 le16_to_cpu(insert_rec->e_leaf_clusters)); 3915 le32_add_cpu(&rec->e_int_clusters, 3916 -le32_to_cpu(rec->e_cpos)); 3917 3918 ret = ocfs2_journal_dirty(handle, bh); 3919 if (ret) 3920 mlog_errno(ret); 3921 3922 } 3923 } 3924 3925 static int ocfs2_append_rec_to_path(struct inode *inode, handle_t *handle, 3926 struct ocfs2_extent_rec *insert_rec, 3927 struct ocfs2_path *right_path, 3928 struct ocfs2_path **ret_left_path) 3929 { 3930 int ret, next_free; 3931 struct ocfs2_extent_list *el; 3932 struct ocfs2_path *left_path = NULL; 3933 3934 *ret_left_path = NULL; 3935 3936 /* 3937 * This shouldn't happen for non-trees. The extent rec cluster 3938 * count manipulation below only works for interior nodes. 3939 */ 3940 BUG_ON(right_path->p_tree_depth == 0); 3941 3942 /* 3943 * If our appending insert is at the leftmost edge of a leaf, 3944 * then we might need to update the rightmost records of the 3945 * neighboring path. 3946 */ 3947 el = path_leaf_el(right_path); 3948 next_free = le16_to_cpu(el->l_next_free_rec); 3949 if (next_free == 0 || 3950 (next_free == 1 && ocfs2_is_empty_extent(&el->l_recs[0]))) { 3951 u32 left_cpos; 3952 3953 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, right_path, 3954 &left_cpos); 3955 if (ret) { 3956 mlog_errno(ret); 3957 goto out; 3958 } 3959 3960 mlog(0, "Append may need a left path update. cpos: %u, " 3961 "left_cpos: %u\n", le32_to_cpu(insert_rec->e_cpos), 3962 left_cpos); 3963 3964 /* 3965 * No need to worry if the append is already in the 3966 * leftmost leaf. 3967 */ 3968 if (left_cpos) { 3969 left_path = ocfs2_new_path_from_path(right_path); 3970 if (!left_path) { 3971 ret = -ENOMEM; 3972 mlog_errno(ret); 3973 goto out; 3974 } 3975 3976 ret = ocfs2_find_path(INODE_CACHE(inode), left_path, 3977 left_cpos); 3978 if (ret) { 3979 mlog_errno(ret); 3980 goto out; 3981 } 3982 3983 /* 3984 * ocfs2_insert_path() will pass the left_path to the 3985 * journal for us. 3986 */ 3987 } 3988 } 3989 3990 ret = ocfs2_journal_access_path(INODE_CACHE(inode), handle, right_path); 3991 if (ret) { 3992 mlog_errno(ret); 3993 goto out; 3994 } 3995 3996 ocfs2_adjust_rightmost_records(inode, handle, right_path, insert_rec); 3997 3998 *ret_left_path = left_path; 3999 ret = 0; 4000 out: 4001 if (ret != 0) 4002 ocfs2_free_path(left_path); 4003 4004 return ret; 4005 } 4006 4007 static void ocfs2_split_record(struct inode *inode, 4008 struct ocfs2_path *left_path, 4009 struct ocfs2_path *right_path, 4010 struct ocfs2_extent_rec *split_rec, 4011 enum ocfs2_split_type split) 4012 { 4013 int index; 4014 u32 cpos = le32_to_cpu(split_rec->e_cpos); 4015 struct ocfs2_extent_list *left_el = NULL, *right_el, *insert_el, *el; 4016 struct ocfs2_extent_rec *rec, *tmprec; 4017 4018 right_el = path_leaf_el(right_path); 4019 if (left_path) 4020 left_el = path_leaf_el(left_path); 4021 4022 el = right_el; 4023 insert_el = right_el; 4024 index = ocfs2_search_extent_list(el, cpos); 4025 if (index != -1) { 4026 if (index == 0 && left_path) { 4027 BUG_ON(ocfs2_is_empty_extent(&el->l_recs[0])); 4028 4029 /* 4030 * This typically means that the record 4031 * started in the left path but moved to the 4032 * right as a result of rotation. We either 4033 * move the existing record to the left, or we 4034 * do the later insert there. 4035 * 4036 * In this case, the left path should always 4037 * exist as the rotate code will have passed 4038 * it back for a post-insert update. 4039 */ 4040 4041 if (split == SPLIT_LEFT) { 4042 /* 4043 * It's a left split. Since we know 4044 * that the rotate code gave us an 4045 * empty extent in the left path, we 4046 * can just do the insert there. 4047 */ 4048 insert_el = left_el; 4049 } else { 4050 /* 4051 * Right split - we have to move the 4052 * existing record over to the left 4053 * leaf. The insert will be into the 4054 * newly created empty extent in the 4055 * right leaf. 4056 */ 4057 tmprec = &right_el->l_recs[index]; 4058 ocfs2_rotate_leaf(left_el, tmprec); 4059 el = left_el; 4060 4061 memset(tmprec, 0, sizeof(*tmprec)); 4062 index = ocfs2_search_extent_list(left_el, cpos); 4063 BUG_ON(index == -1); 4064 } 4065 } 4066 } else { 4067 BUG_ON(!left_path); 4068 BUG_ON(!ocfs2_is_empty_extent(&left_el->l_recs[0])); 4069 /* 4070 * Left path is easy - we can just allow the insert to 4071 * happen. 4072 */ 4073 el = left_el; 4074 insert_el = left_el; 4075 index = ocfs2_search_extent_list(el, cpos); 4076 BUG_ON(index == -1); 4077 } 4078 4079 rec = &el->l_recs[index]; 4080 ocfs2_subtract_from_rec(inode->i_sb, split, rec, split_rec); 4081 ocfs2_rotate_leaf(insert_el, split_rec); 4082 } 4083 4084 /* 4085 * This function only does inserts on an allocation b-tree. For tree 4086 * depth = 0, ocfs2_insert_at_leaf() is called directly. 4087 * 4088 * right_path is the path we want to do the actual insert 4089 * in. left_path should only be passed in if we need to update that 4090 * portion of the tree after an edge insert. 4091 */ 4092 static int ocfs2_insert_path(struct inode *inode, 4093 handle_t *handle, 4094 struct ocfs2_extent_tree *et, 4095 struct ocfs2_path *left_path, 4096 struct ocfs2_path *right_path, 4097 struct ocfs2_extent_rec *insert_rec, 4098 struct ocfs2_insert_type *insert) 4099 { 4100 int ret, subtree_index; 4101 struct buffer_head *leaf_bh = path_leaf_bh(right_path); 4102 4103 if (left_path) { 4104 int credits = handle->h_buffer_credits; 4105 4106 /* 4107 * There's a chance that left_path got passed back to 4108 * us without being accounted for in the 4109 * journal. Extend our transaction here to be sure we 4110 * can change those blocks. 4111 */ 4112 credits += left_path->p_tree_depth; 4113 4114 ret = ocfs2_extend_trans(handle, credits); 4115 if (ret < 0) { 4116 mlog_errno(ret); 4117 goto out; 4118 } 4119 4120 ret = ocfs2_journal_access_path(et->et_ci, handle, left_path); 4121 if (ret < 0) { 4122 mlog_errno(ret); 4123 goto out; 4124 } 4125 } 4126 4127 /* 4128 * Pass both paths to the journal. The majority of inserts 4129 * will be touching all components anyway. 4130 */ 4131 ret = ocfs2_journal_access_path(et->et_ci, handle, right_path); 4132 if (ret < 0) { 4133 mlog_errno(ret); 4134 goto out; 4135 } 4136 4137 if (insert->ins_split != SPLIT_NONE) { 4138 /* 4139 * We could call ocfs2_insert_at_leaf() for some types 4140 * of splits, but it's easier to just let one separate 4141 * function sort it all out. 4142 */ 4143 ocfs2_split_record(inode, left_path, right_path, 4144 insert_rec, insert->ins_split); 4145 4146 /* 4147 * Split might have modified either leaf and we don't 4148 * have a guarantee that the later edge insert will 4149 * dirty this for us. 4150 */ 4151 if (left_path) 4152 ret = ocfs2_journal_dirty(handle, 4153 path_leaf_bh(left_path)); 4154 if (ret) 4155 mlog_errno(ret); 4156 } else 4157 ocfs2_insert_at_leaf(insert_rec, path_leaf_el(right_path), 4158 insert, inode); 4159 4160 ret = ocfs2_journal_dirty(handle, leaf_bh); 4161 if (ret) 4162 mlog_errno(ret); 4163 4164 if (left_path) { 4165 /* 4166 * The rotate code has indicated that we need to fix 4167 * up portions of the tree after the insert. 4168 * 4169 * XXX: Should we extend the transaction here? 4170 */ 4171 subtree_index = ocfs2_find_subtree_root(et, left_path, 4172 right_path); 4173 ocfs2_complete_edge_insert(handle, left_path, right_path, 4174 subtree_index); 4175 } 4176 4177 ret = 0; 4178 out: 4179 return ret; 4180 } 4181 4182 static int ocfs2_do_insert_extent(struct inode *inode, 4183 handle_t *handle, 4184 struct ocfs2_extent_tree *et, 4185 struct ocfs2_extent_rec *insert_rec, 4186 struct ocfs2_insert_type *type) 4187 { 4188 int ret, rotate = 0; 4189 u32 cpos; 4190 struct ocfs2_path *right_path = NULL; 4191 struct ocfs2_path *left_path = NULL; 4192 struct ocfs2_extent_list *el; 4193 4194 el = et->et_root_el; 4195 4196 ret = ocfs2_et_root_journal_access(handle, et, 4197 OCFS2_JOURNAL_ACCESS_WRITE); 4198 if (ret) { 4199 mlog_errno(ret); 4200 goto out; 4201 } 4202 4203 if (le16_to_cpu(el->l_tree_depth) == 0) { 4204 ocfs2_insert_at_leaf(insert_rec, el, type, inode); 4205 goto out_update_clusters; 4206 } 4207 4208 right_path = ocfs2_new_path_from_et(et); 4209 if (!right_path) { 4210 ret = -ENOMEM; 4211 mlog_errno(ret); 4212 goto out; 4213 } 4214 4215 /* 4216 * Determine the path to start with. Rotations need the 4217 * rightmost path, everything else can go directly to the 4218 * target leaf. 4219 */ 4220 cpos = le32_to_cpu(insert_rec->e_cpos); 4221 if (type->ins_appending == APPEND_NONE && 4222 type->ins_contig == CONTIG_NONE) { 4223 rotate = 1; 4224 cpos = UINT_MAX; 4225 } 4226 4227 ret = ocfs2_find_path(et->et_ci, right_path, cpos); 4228 if (ret) { 4229 mlog_errno(ret); 4230 goto out; 4231 } 4232 4233 /* 4234 * Rotations and appends need special treatment - they modify 4235 * parts of the tree's above them. 4236 * 4237 * Both might pass back a path immediate to the left of the 4238 * one being inserted to. This will be cause 4239 * ocfs2_insert_path() to modify the rightmost records of 4240 * left_path to account for an edge insert. 4241 * 4242 * XXX: When modifying this code, keep in mind that an insert 4243 * can wind up skipping both of these two special cases... 4244 */ 4245 if (rotate) { 4246 ret = ocfs2_rotate_tree_right(handle, et, type->ins_split, 4247 le32_to_cpu(insert_rec->e_cpos), 4248 right_path, &left_path); 4249 if (ret) { 4250 mlog_errno(ret); 4251 goto out; 4252 } 4253 4254 /* 4255 * ocfs2_rotate_tree_right() might have extended the 4256 * transaction without re-journaling our tree root. 4257 */ 4258 ret = ocfs2_et_root_journal_access(handle, et, 4259 OCFS2_JOURNAL_ACCESS_WRITE); 4260 if (ret) { 4261 mlog_errno(ret); 4262 goto out; 4263 } 4264 } else if (type->ins_appending == APPEND_TAIL 4265 && type->ins_contig != CONTIG_LEFT) { 4266 ret = ocfs2_append_rec_to_path(inode, handle, insert_rec, 4267 right_path, &left_path); 4268 if (ret) { 4269 mlog_errno(ret); 4270 goto out; 4271 } 4272 } 4273 4274 ret = ocfs2_insert_path(inode, handle, et, left_path, right_path, 4275 insert_rec, type); 4276 if (ret) { 4277 mlog_errno(ret); 4278 goto out; 4279 } 4280 4281 out_update_clusters: 4282 if (type->ins_split == SPLIT_NONE) 4283 ocfs2_et_update_clusters(et, 4284 le16_to_cpu(insert_rec->e_leaf_clusters)); 4285 4286 ret = ocfs2_journal_dirty(handle, et->et_root_bh); 4287 if (ret) 4288 mlog_errno(ret); 4289 4290 out: 4291 ocfs2_free_path(left_path); 4292 ocfs2_free_path(right_path); 4293 4294 return ret; 4295 } 4296 4297 static enum ocfs2_contig_type 4298 ocfs2_figure_merge_contig_type(struct inode *inode, struct ocfs2_path *path, 4299 struct ocfs2_extent_list *el, int index, 4300 struct ocfs2_extent_rec *split_rec) 4301 { 4302 int status; 4303 enum ocfs2_contig_type ret = CONTIG_NONE; 4304 u32 left_cpos, right_cpos; 4305 struct ocfs2_extent_rec *rec = NULL; 4306 struct ocfs2_extent_list *new_el; 4307 struct ocfs2_path *left_path = NULL, *right_path = NULL; 4308 struct buffer_head *bh; 4309 struct ocfs2_extent_block *eb; 4310 4311 if (index > 0) { 4312 rec = &el->l_recs[index - 1]; 4313 } else if (path->p_tree_depth > 0) { 4314 status = ocfs2_find_cpos_for_left_leaf(inode->i_sb, 4315 path, &left_cpos); 4316 if (status) 4317 goto out; 4318 4319 if (left_cpos != 0) { 4320 left_path = ocfs2_new_path_from_path(path); 4321 if (!left_path) 4322 goto out; 4323 4324 status = ocfs2_find_path(INODE_CACHE(inode), 4325 left_path, left_cpos); 4326 if (status) 4327 goto out; 4328 4329 new_el = path_leaf_el(left_path); 4330 4331 if (le16_to_cpu(new_el->l_next_free_rec) != 4332 le16_to_cpu(new_el->l_count)) { 4333 bh = path_leaf_bh(left_path); 4334 eb = (struct ocfs2_extent_block *)bh->b_data; 4335 ocfs2_error(inode->i_sb, 4336 "Extent block #%llu has an " 4337 "invalid l_next_free_rec of " 4338 "%d. It should have " 4339 "matched the l_count of %d", 4340 (unsigned long long)le64_to_cpu(eb->h_blkno), 4341 le16_to_cpu(new_el->l_next_free_rec), 4342 le16_to_cpu(new_el->l_count)); 4343 status = -EINVAL; 4344 goto out; 4345 } 4346 rec = &new_el->l_recs[ 4347 le16_to_cpu(new_el->l_next_free_rec) - 1]; 4348 } 4349 } 4350 4351 /* 4352 * We're careful to check for an empty extent record here - 4353 * the merge code will know what to do if it sees one. 4354 */ 4355 if (rec) { 4356 if (index == 1 && ocfs2_is_empty_extent(rec)) { 4357 if (split_rec->e_cpos == el->l_recs[index].e_cpos) 4358 ret = CONTIG_RIGHT; 4359 } else { 4360 ret = ocfs2_extent_contig(inode, rec, split_rec); 4361 } 4362 } 4363 4364 rec = NULL; 4365 if (index < (le16_to_cpu(el->l_next_free_rec) - 1)) 4366 rec = &el->l_recs[index + 1]; 4367 else if (le16_to_cpu(el->l_next_free_rec) == le16_to_cpu(el->l_count) && 4368 path->p_tree_depth > 0) { 4369 status = ocfs2_find_cpos_for_right_leaf(inode->i_sb, 4370 path, &right_cpos); 4371 if (status) 4372 goto out; 4373 4374 if (right_cpos == 0) 4375 goto out; 4376 4377 right_path = ocfs2_new_path_from_path(path); 4378 if (!right_path) 4379 goto out; 4380 4381 status = ocfs2_find_path(INODE_CACHE(inode), right_path, right_cpos); 4382 if (status) 4383 goto out; 4384 4385 new_el = path_leaf_el(right_path); 4386 rec = &new_el->l_recs[0]; 4387 if (ocfs2_is_empty_extent(rec)) { 4388 if (le16_to_cpu(new_el->l_next_free_rec) <= 1) { 4389 bh = path_leaf_bh(right_path); 4390 eb = (struct ocfs2_extent_block *)bh->b_data; 4391 ocfs2_error(inode->i_sb, 4392 "Extent block #%llu has an " 4393 "invalid l_next_free_rec of %d", 4394 (unsigned long long)le64_to_cpu(eb->h_blkno), 4395 le16_to_cpu(new_el->l_next_free_rec)); 4396 status = -EINVAL; 4397 goto out; 4398 } 4399 rec = &new_el->l_recs[1]; 4400 } 4401 } 4402 4403 if (rec) { 4404 enum ocfs2_contig_type contig_type; 4405 4406 contig_type = ocfs2_extent_contig(inode, rec, split_rec); 4407 4408 if (contig_type == CONTIG_LEFT && ret == CONTIG_RIGHT) 4409 ret = CONTIG_LEFTRIGHT; 4410 else if (ret == CONTIG_NONE) 4411 ret = contig_type; 4412 } 4413 4414 out: 4415 if (left_path) 4416 ocfs2_free_path(left_path); 4417 if (right_path) 4418 ocfs2_free_path(right_path); 4419 4420 return ret; 4421 } 4422 4423 static void ocfs2_figure_contig_type(struct inode *inode, 4424 struct ocfs2_insert_type *insert, 4425 struct ocfs2_extent_list *el, 4426 struct ocfs2_extent_rec *insert_rec, 4427 struct ocfs2_extent_tree *et) 4428 { 4429 int i; 4430 enum ocfs2_contig_type contig_type = CONTIG_NONE; 4431 4432 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0); 4433 4434 for(i = 0; i < le16_to_cpu(el->l_next_free_rec); i++) { 4435 contig_type = ocfs2_extent_contig(inode, &el->l_recs[i], 4436 insert_rec); 4437 if (contig_type != CONTIG_NONE) { 4438 insert->ins_contig_index = i; 4439 break; 4440 } 4441 } 4442 insert->ins_contig = contig_type; 4443 4444 if (insert->ins_contig != CONTIG_NONE) { 4445 struct ocfs2_extent_rec *rec = 4446 &el->l_recs[insert->ins_contig_index]; 4447 unsigned int len = le16_to_cpu(rec->e_leaf_clusters) + 4448 le16_to_cpu(insert_rec->e_leaf_clusters); 4449 4450 /* 4451 * Caller might want us to limit the size of extents, don't 4452 * calculate contiguousness if we might exceed that limit. 4453 */ 4454 if (et->et_max_leaf_clusters && 4455 (len > et->et_max_leaf_clusters)) 4456 insert->ins_contig = CONTIG_NONE; 4457 } 4458 } 4459 4460 /* 4461 * This should only be called against the righmost leaf extent list. 4462 * 4463 * ocfs2_figure_appending_type() will figure out whether we'll have to 4464 * insert at the tail of the rightmost leaf. 4465 * 4466 * This should also work against the root extent list for tree's with 0 4467 * depth. If we consider the root extent list to be the rightmost leaf node 4468 * then the logic here makes sense. 4469 */ 4470 static void ocfs2_figure_appending_type(struct ocfs2_insert_type *insert, 4471 struct ocfs2_extent_list *el, 4472 struct ocfs2_extent_rec *insert_rec) 4473 { 4474 int i; 4475 u32 cpos = le32_to_cpu(insert_rec->e_cpos); 4476 struct ocfs2_extent_rec *rec; 4477 4478 insert->ins_appending = APPEND_NONE; 4479 4480 BUG_ON(le16_to_cpu(el->l_tree_depth) != 0); 4481 4482 if (!el->l_next_free_rec) 4483 goto set_tail_append; 4484 4485 if (ocfs2_is_empty_extent(&el->l_recs[0])) { 4486 /* Were all records empty? */ 4487 if (le16_to_cpu(el->l_next_free_rec) == 1) 4488 goto set_tail_append; 4489 } 4490 4491 i = le16_to_cpu(el->l_next_free_rec) - 1; 4492 rec = &el->l_recs[i]; 4493 4494 if (cpos >= 4495 (le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters))) 4496 goto set_tail_append; 4497 4498 return; 4499 4500 set_tail_append: 4501 insert->ins_appending = APPEND_TAIL; 4502 } 4503 4504 /* 4505 * Helper function called at the begining of an insert. 4506 * 4507 * This computes a few things that are commonly used in the process of 4508 * inserting into the btree: 4509 * - Whether the new extent is contiguous with an existing one. 4510 * - The current tree depth. 4511 * - Whether the insert is an appending one. 4512 * - The total # of free records in the tree. 4513 * 4514 * All of the information is stored on the ocfs2_insert_type 4515 * structure. 4516 */ 4517 static int ocfs2_figure_insert_type(struct inode *inode, 4518 struct ocfs2_extent_tree *et, 4519 struct buffer_head **last_eb_bh, 4520 struct ocfs2_extent_rec *insert_rec, 4521 int *free_records, 4522 struct ocfs2_insert_type *insert) 4523 { 4524 int ret; 4525 struct ocfs2_extent_block *eb; 4526 struct ocfs2_extent_list *el; 4527 struct ocfs2_path *path = NULL; 4528 struct buffer_head *bh = NULL; 4529 4530 insert->ins_split = SPLIT_NONE; 4531 4532 el = et->et_root_el; 4533 insert->ins_tree_depth = le16_to_cpu(el->l_tree_depth); 4534 4535 if (el->l_tree_depth) { 4536 /* 4537 * If we have tree depth, we read in the 4538 * rightmost extent block ahead of time as 4539 * ocfs2_figure_insert_type() and ocfs2_add_branch() 4540 * may want it later. 4541 */ 4542 ret = ocfs2_read_extent_block(et->et_ci, 4543 ocfs2_et_get_last_eb_blk(et), 4544 &bh); 4545 if (ret) { 4546 mlog_exit(ret); 4547 goto out; 4548 } 4549 eb = (struct ocfs2_extent_block *) bh->b_data; 4550 el = &eb->h_list; 4551 } 4552 4553 /* 4554 * Unless we have a contiguous insert, we'll need to know if 4555 * there is room left in our allocation tree for another 4556 * extent record. 4557 * 4558 * XXX: This test is simplistic, we can search for empty 4559 * extent records too. 4560 */ 4561 *free_records = le16_to_cpu(el->l_count) - 4562 le16_to_cpu(el->l_next_free_rec); 4563 4564 if (!insert->ins_tree_depth) { 4565 ocfs2_figure_contig_type(inode, insert, el, insert_rec, et); 4566 ocfs2_figure_appending_type(insert, el, insert_rec); 4567 return 0; 4568 } 4569 4570 path = ocfs2_new_path_from_et(et); 4571 if (!path) { 4572 ret = -ENOMEM; 4573 mlog_errno(ret); 4574 goto out; 4575 } 4576 4577 /* 4578 * In the case that we're inserting past what the tree 4579 * currently accounts for, ocfs2_find_path() will return for 4580 * us the rightmost tree path. This is accounted for below in 4581 * the appending code. 4582 */ 4583 ret = ocfs2_find_path(et->et_ci, path, le32_to_cpu(insert_rec->e_cpos)); 4584 if (ret) { 4585 mlog_errno(ret); 4586 goto out; 4587 } 4588 4589 el = path_leaf_el(path); 4590 4591 /* 4592 * Now that we have the path, there's two things we want to determine: 4593 * 1) Contiguousness (also set contig_index if this is so) 4594 * 4595 * 2) Are we doing an append? We can trivially break this up 4596 * into two types of appends: simple record append, or a 4597 * rotate inside the tail leaf. 4598 */ 4599 ocfs2_figure_contig_type(inode, insert, el, insert_rec, et); 4600 4601 /* 4602 * The insert code isn't quite ready to deal with all cases of 4603 * left contiguousness. Specifically, if it's an insert into 4604 * the 1st record in a leaf, it will require the adjustment of 4605 * cluster count on the last record of the path directly to it's 4606 * left. For now, just catch that case and fool the layers 4607 * above us. This works just fine for tree_depth == 0, which 4608 * is why we allow that above. 4609 */ 4610 if (insert->ins_contig == CONTIG_LEFT && 4611 insert->ins_contig_index == 0) 4612 insert->ins_contig = CONTIG_NONE; 4613 4614 /* 4615 * Ok, so we can simply compare against last_eb to figure out 4616 * whether the path doesn't exist. This will only happen in 4617 * the case that we're doing a tail append, so maybe we can 4618 * take advantage of that information somehow. 4619 */ 4620 if (ocfs2_et_get_last_eb_blk(et) == 4621 path_leaf_bh(path)->b_blocknr) { 4622 /* 4623 * Ok, ocfs2_find_path() returned us the rightmost 4624 * tree path. This might be an appending insert. There are 4625 * two cases: 4626 * 1) We're doing a true append at the tail: 4627 * -This might even be off the end of the leaf 4628 * 2) We're "appending" by rotating in the tail 4629 */ 4630 ocfs2_figure_appending_type(insert, el, insert_rec); 4631 } 4632 4633 out: 4634 ocfs2_free_path(path); 4635 4636 if (ret == 0) 4637 *last_eb_bh = bh; 4638 else 4639 brelse(bh); 4640 return ret; 4641 } 4642 4643 /* 4644 * Insert an extent into an inode btree. 4645 * 4646 * The caller needs to update fe->i_clusters 4647 */ 4648 int ocfs2_insert_extent(struct ocfs2_super *osb, 4649 handle_t *handle, 4650 struct inode *inode, 4651 struct ocfs2_extent_tree *et, 4652 u32 cpos, 4653 u64 start_blk, 4654 u32 new_clusters, 4655 u8 flags, 4656 struct ocfs2_alloc_context *meta_ac) 4657 { 4658 int status; 4659 int uninitialized_var(free_records); 4660 struct buffer_head *last_eb_bh = NULL; 4661 struct ocfs2_insert_type insert = {0, }; 4662 struct ocfs2_extent_rec rec; 4663 4664 mlog(0, "add %u clusters at position %u to inode %llu\n", 4665 new_clusters, cpos, (unsigned long long)OCFS2_I(inode)->ip_blkno); 4666 4667 memset(&rec, 0, sizeof(rec)); 4668 rec.e_cpos = cpu_to_le32(cpos); 4669 rec.e_blkno = cpu_to_le64(start_blk); 4670 rec.e_leaf_clusters = cpu_to_le16(new_clusters); 4671 rec.e_flags = flags; 4672 status = ocfs2_et_insert_check(et, &rec); 4673 if (status) { 4674 mlog_errno(status); 4675 goto bail; 4676 } 4677 4678 status = ocfs2_figure_insert_type(inode, et, &last_eb_bh, &rec, 4679 &free_records, &insert); 4680 if (status < 0) { 4681 mlog_errno(status); 4682 goto bail; 4683 } 4684 4685 mlog(0, "Insert.appending: %u, Insert.Contig: %u, " 4686 "Insert.contig_index: %d, Insert.free_records: %d, " 4687 "Insert.tree_depth: %d\n", 4688 insert.ins_appending, insert.ins_contig, insert.ins_contig_index, 4689 free_records, insert.ins_tree_depth); 4690 4691 if (insert.ins_contig == CONTIG_NONE && free_records == 0) { 4692 status = ocfs2_grow_tree(inode, handle, et, 4693 &insert.ins_tree_depth, &last_eb_bh, 4694 meta_ac); 4695 if (status) { 4696 mlog_errno(status); 4697 goto bail; 4698 } 4699 } 4700 4701 /* Finally, we can add clusters. This might rotate the tree for us. */ 4702 status = ocfs2_do_insert_extent(inode, handle, et, &rec, &insert); 4703 if (status < 0) 4704 mlog_errno(status); 4705 else if (et->et_ops == &ocfs2_dinode_et_ops) 4706 ocfs2_extent_map_insert_rec(inode, &rec); 4707 4708 bail: 4709 brelse(last_eb_bh); 4710 4711 mlog_exit(status); 4712 return status; 4713 } 4714 4715 /* 4716 * Allcate and add clusters into the extent b-tree. 4717 * The new clusters(clusters_to_add) will be inserted at logical_offset. 4718 * The extent b-tree's root is specified by et, and 4719 * it is not limited to the file storage. Any extent tree can use this 4720 * function if it implements the proper ocfs2_extent_tree. 4721 */ 4722 int ocfs2_add_clusters_in_btree(struct ocfs2_super *osb, 4723 struct inode *inode, 4724 u32 *logical_offset, 4725 u32 clusters_to_add, 4726 int mark_unwritten, 4727 struct ocfs2_extent_tree *et, 4728 handle_t *handle, 4729 struct ocfs2_alloc_context *data_ac, 4730 struct ocfs2_alloc_context *meta_ac, 4731 enum ocfs2_alloc_restarted *reason_ret) 4732 { 4733 int status = 0; 4734 int free_extents; 4735 enum ocfs2_alloc_restarted reason = RESTART_NONE; 4736 u32 bit_off, num_bits; 4737 u64 block; 4738 u8 flags = 0; 4739 4740 BUG_ON(!clusters_to_add); 4741 4742 if (mark_unwritten) 4743 flags = OCFS2_EXT_UNWRITTEN; 4744 4745 free_extents = ocfs2_num_free_extents(osb, et); 4746 if (free_extents < 0) { 4747 status = free_extents; 4748 mlog_errno(status); 4749 goto leave; 4750 } 4751 4752 /* there are two cases which could cause us to EAGAIN in the 4753 * we-need-more-metadata case: 4754 * 1) we haven't reserved *any* 4755 * 2) we are so fragmented, we've needed to add metadata too 4756 * many times. */ 4757 if (!free_extents && !meta_ac) { 4758 mlog(0, "we haven't reserved any metadata!\n"); 4759 status = -EAGAIN; 4760 reason = RESTART_META; 4761 goto leave; 4762 } else if ((!free_extents) 4763 && (ocfs2_alloc_context_bits_left(meta_ac) 4764 < ocfs2_extend_meta_needed(et->et_root_el))) { 4765 mlog(0, "filesystem is really fragmented...\n"); 4766 status = -EAGAIN; 4767 reason = RESTART_META; 4768 goto leave; 4769 } 4770 4771 status = __ocfs2_claim_clusters(osb, handle, data_ac, 1, 4772 clusters_to_add, &bit_off, &num_bits); 4773 if (status < 0) { 4774 if (status != -ENOSPC) 4775 mlog_errno(status); 4776 goto leave; 4777 } 4778 4779 BUG_ON(num_bits > clusters_to_add); 4780 4781 /* reserve our write early -- insert_extent may update the tree root */ 4782 status = ocfs2_et_root_journal_access(handle, et, 4783 OCFS2_JOURNAL_ACCESS_WRITE); 4784 if (status < 0) { 4785 mlog_errno(status); 4786 goto leave; 4787 } 4788 4789 block = ocfs2_clusters_to_blocks(osb->sb, bit_off); 4790 mlog(0, "Allocating %u clusters at block %u for inode %llu\n", 4791 num_bits, bit_off, (unsigned long long)OCFS2_I(inode)->ip_blkno); 4792 status = ocfs2_insert_extent(osb, handle, inode, et, 4793 *logical_offset, block, 4794 num_bits, flags, meta_ac); 4795 if (status < 0) { 4796 mlog_errno(status); 4797 goto leave; 4798 } 4799 4800 status = ocfs2_journal_dirty(handle, et->et_root_bh); 4801 if (status < 0) { 4802 mlog_errno(status); 4803 goto leave; 4804 } 4805 4806 clusters_to_add -= num_bits; 4807 *logical_offset += num_bits; 4808 4809 if (clusters_to_add) { 4810 mlog(0, "need to alloc once more, wanted = %u\n", 4811 clusters_to_add); 4812 status = -EAGAIN; 4813 reason = RESTART_TRANS; 4814 } 4815 4816 leave: 4817 mlog_exit(status); 4818 if (reason_ret) 4819 *reason_ret = reason; 4820 return status; 4821 } 4822 4823 static void ocfs2_make_right_split_rec(struct super_block *sb, 4824 struct ocfs2_extent_rec *split_rec, 4825 u32 cpos, 4826 struct ocfs2_extent_rec *rec) 4827 { 4828 u32 rec_cpos = le32_to_cpu(rec->e_cpos); 4829 u32 rec_range = rec_cpos + le16_to_cpu(rec->e_leaf_clusters); 4830 4831 memset(split_rec, 0, sizeof(struct ocfs2_extent_rec)); 4832 4833 split_rec->e_cpos = cpu_to_le32(cpos); 4834 split_rec->e_leaf_clusters = cpu_to_le16(rec_range - cpos); 4835 4836 split_rec->e_blkno = rec->e_blkno; 4837 le64_add_cpu(&split_rec->e_blkno, 4838 ocfs2_clusters_to_blocks(sb, cpos - rec_cpos)); 4839 4840 split_rec->e_flags = rec->e_flags; 4841 } 4842 4843 static int ocfs2_split_and_insert(struct inode *inode, 4844 handle_t *handle, 4845 struct ocfs2_path *path, 4846 struct ocfs2_extent_tree *et, 4847 struct buffer_head **last_eb_bh, 4848 int split_index, 4849 struct ocfs2_extent_rec *orig_split_rec, 4850 struct ocfs2_alloc_context *meta_ac) 4851 { 4852 int ret = 0, depth; 4853 unsigned int insert_range, rec_range, do_leftright = 0; 4854 struct ocfs2_extent_rec tmprec; 4855 struct ocfs2_extent_list *rightmost_el; 4856 struct ocfs2_extent_rec rec; 4857 struct ocfs2_extent_rec split_rec = *orig_split_rec; 4858 struct ocfs2_insert_type insert; 4859 struct ocfs2_extent_block *eb; 4860 4861 leftright: 4862 /* 4863 * Store a copy of the record on the stack - it might move 4864 * around as the tree is manipulated below. 4865 */ 4866 rec = path_leaf_el(path)->l_recs[split_index]; 4867 4868 rightmost_el = et->et_root_el; 4869 4870 depth = le16_to_cpu(rightmost_el->l_tree_depth); 4871 if (depth) { 4872 BUG_ON(!(*last_eb_bh)); 4873 eb = (struct ocfs2_extent_block *) (*last_eb_bh)->b_data; 4874 rightmost_el = &eb->h_list; 4875 } 4876 4877 if (le16_to_cpu(rightmost_el->l_next_free_rec) == 4878 le16_to_cpu(rightmost_el->l_count)) { 4879 ret = ocfs2_grow_tree(inode, handle, et, 4880 &depth, last_eb_bh, meta_ac); 4881 if (ret) { 4882 mlog_errno(ret); 4883 goto out; 4884 } 4885 } 4886 4887 memset(&insert, 0, sizeof(struct ocfs2_insert_type)); 4888 insert.ins_appending = APPEND_NONE; 4889 insert.ins_contig = CONTIG_NONE; 4890 insert.ins_tree_depth = depth; 4891 4892 insert_range = le32_to_cpu(split_rec.e_cpos) + 4893 le16_to_cpu(split_rec.e_leaf_clusters); 4894 rec_range = le32_to_cpu(rec.e_cpos) + 4895 le16_to_cpu(rec.e_leaf_clusters); 4896 4897 if (split_rec.e_cpos == rec.e_cpos) { 4898 insert.ins_split = SPLIT_LEFT; 4899 } else if (insert_range == rec_range) { 4900 insert.ins_split = SPLIT_RIGHT; 4901 } else { 4902 /* 4903 * Left/right split. We fake this as a right split 4904 * first and then make a second pass as a left split. 4905 */ 4906 insert.ins_split = SPLIT_RIGHT; 4907 4908 ocfs2_make_right_split_rec(inode->i_sb, &tmprec, insert_range, 4909 &rec); 4910 4911 split_rec = tmprec; 4912 4913 BUG_ON(do_leftright); 4914 do_leftright = 1; 4915 } 4916 4917 ret = ocfs2_do_insert_extent(inode, handle, et, &split_rec, &insert); 4918 if (ret) { 4919 mlog_errno(ret); 4920 goto out; 4921 } 4922 4923 if (do_leftright == 1) { 4924 u32 cpos; 4925 struct ocfs2_extent_list *el; 4926 4927 do_leftright++; 4928 split_rec = *orig_split_rec; 4929 4930 ocfs2_reinit_path(path, 1); 4931 4932 cpos = le32_to_cpu(split_rec.e_cpos); 4933 ret = ocfs2_find_path(et->et_ci, path, cpos); 4934 if (ret) { 4935 mlog_errno(ret); 4936 goto out; 4937 } 4938 4939 el = path_leaf_el(path); 4940 split_index = ocfs2_search_extent_list(el, cpos); 4941 goto leftright; 4942 } 4943 out: 4944 4945 return ret; 4946 } 4947 4948 static int ocfs2_replace_extent_rec(struct inode *inode, 4949 handle_t *handle, 4950 struct ocfs2_path *path, 4951 struct ocfs2_extent_list *el, 4952 int split_index, 4953 struct ocfs2_extent_rec *split_rec) 4954 { 4955 int ret; 4956 4957 ret = ocfs2_path_bh_journal_access(handle, INODE_CACHE(inode), path, 4958 path_num_items(path) - 1); 4959 if (ret) { 4960 mlog_errno(ret); 4961 goto out; 4962 } 4963 4964 el->l_recs[split_index] = *split_rec; 4965 4966 ocfs2_journal_dirty(handle, path_leaf_bh(path)); 4967 out: 4968 return ret; 4969 } 4970 4971 /* 4972 * Mark part or all of the extent record at split_index in the leaf 4973 * pointed to by path as written. This removes the unwritten 4974 * extent flag. 4975 * 4976 * Care is taken to handle contiguousness so as to not grow the tree. 4977 * 4978 * meta_ac is not strictly necessary - we only truly need it if growth 4979 * of the tree is required. All other cases will degrade into a less 4980 * optimal tree layout. 4981 * 4982 * last_eb_bh should be the rightmost leaf block for any extent 4983 * btree. Since a split may grow the tree or a merge might shrink it, 4984 * the caller cannot trust the contents of that buffer after this call. 4985 * 4986 * This code is optimized for readability - several passes might be 4987 * made over certain portions of the tree. All of those blocks will 4988 * have been brought into cache (and pinned via the journal), so the 4989 * extra overhead is not expressed in terms of disk reads. 4990 */ 4991 static int __ocfs2_mark_extent_written(struct inode *inode, 4992 struct ocfs2_extent_tree *et, 4993 handle_t *handle, 4994 struct ocfs2_path *path, 4995 int split_index, 4996 struct ocfs2_extent_rec *split_rec, 4997 struct ocfs2_alloc_context *meta_ac, 4998 struct ocfs2_cached_dealloc_ctxt *dealloc) 4999 { 5000 int ret = 0; 5001 struct ocfs2_extent_list *el = path_leaf_el(path); 5002 struct buffer_head *last_eb_bh = NULL; 5003 struct ocfs2_extent_rec *rec = &el->l_recs[split_index]; 5004 struct ocfs2_merge_ctxt ctxt; 5005 struct ocfs2_extent_list *rightmost_el; 5006 5007 if (!(rec->e_flags & OCFS2_EXT_UNWRITTEN)) { 5008 ret = -EIO; 5009 mlog_errno(ret); 5010 goto out; 5011 } 5012 5013 if (le32_to_cpu(rec->e_cpos) > le32_to_cpu(split_rec->e_cpos) || 5014 ((le32_to_cpu(rec->e_cpos) + le16_to_cpu(rec->e_leaf_clusters)) < 5015 (le32_to_cpu(split_rec->e_cpos) + le16_to_cpu(split_rec->e_leaf_clusters)))) { 5016 ret = -EIO; 5017 mlog_errno(ret); 5018 goto out; 5019 } 5020 5021 ctxt.c_contig_type = ocfs2_figure_merge_contig_type(inode, path, el, 5022 split_index, 5023 split_rec); 5024 5025 /* 5026 * The core merge / split code wants to know how much room is 5027 * left in this inodes allocation tree, so we pass the 5028 * rightmost extent list. 5029 */ 5030 if (path->p_tree_depth) { 5031 struct ocfs2_extent_block *eb; 5032 5033 ret = ocfs2_read_extent_block(et->et_ci, 5034 ocfs2_et_get_last_eb_blk(et), 5035 &last_eb_bh); 5036 if (ret) { 5037 mlog_exit(ret); 5038 goto out; 5039 } 5040 5041 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data; 5042 rightmost_el = &eb->h_list; 5043 } else 5044 rightmost_el = path_root_el(path); 5045 5046 if (rec->e_cpos == split_rec->e_cpos && 5047 rec->e_leaf_clusters == split_rec->e_leaf_clusters) 5048 ctxt.c_split_covers_rec = 1; 5049 else 5050 ctxt.c_split_covers_rec = 0; 5051 5052 ctxt.c_has_empty_extent = ocfs2_is_empty_extent(&el->l_recs[0]); 5053 5054 mlog(0, "index: %d, contig: %u, has_empty: %u, split_covers: %u\n", 5055 split_index, ctxt.c_contig_type, ctxt.c_has_empty_extent, 5056 ctxt.c_split_covers_rec); 5057 5058 if (ctxt.c_contig_type == CONTIG_NONE) { 5059 if (ctxt.c_split_covers_rec) 5060 ret = ocfs2_replace_extent_rec(inode, handle, 5061 path, el, 5062 split_index, split_rec); 5063 else 5064 ret = ocfs2_split_and_insert(inode, handle, path, et, 5065 &last_eb_bh, split_index, 5066 split_rec, meta_ac); 5067 if (ret) 5068 mlog_errno(ret); 5069 } else { 5070 ret = ocfs2_try_to_merge_extent(handle, et, path, 5071 split_index, split_rec, 5072 dealloc, &ctxt); 5073 if (ret) 5074 mlog_errno(ret); 5075 } 5076 5077 out: 5078 brelse(last_eb_bh); 5079 return ret; 5080 } 5081 5082 /* 5083 * Mark the already-existing extent at cpos as written for len clusters. 5084 * 5085 * If the existing extent is larger than the request, initiate a 5086 * split. An attempt will be made at merging with adjacent extents. 5087 * 5088 * The caller is responsible for passing down meta_ac if we'll need it. 5089 */ 5090 int ocfs2_mark_extent_written(struct inode *inode, 5091 struct ocfs2_extent_tree *et, 5092 handle_t *handle, u32 cpos, u32 len, u32 phys, 5093 struct ocfs2_alloc_context *meta_ac, 5094 struct ocfs2_cached_dealloc_ctxt *dealloc) 5095 { 5096 int ret, index; 5097 u64 start_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys); 5098 struct ocfs2_extent_rec split_rec; 5099 struct ocfs2_path *left_path = NULL; 5100 struct ocfs2_extent_list *el; 5101 5102 mlog(0, "Inode %lu cpos %u, len %u, phys %u (%llu)\n", 5103 inode->i_ino, cpos, len, phys, (unsigned long long)start_blkno); 5104 5105 if (!ocfs2_writes_unwritten_extents(OCFS2_SB(inode->i_sb))) { 5106 ocfs2_error(inode->i_sb, "Inode %llu has unwritten extents " 5107 "that are being written to, but the feature bit " 5108 "is not set in the super block.", 5109 (unsigned long long)OCFS2_I(inode)->ip_blkno); 5110 ret = -EROFS; 5111 goto out; 5112 } 5113 5114 /* 5115 * XXX: This should be fixed up so that we just re-insert the 5116 * next extent records. 5117 * 5118 * XXX: This is a hack on the extent tree, maybe it should be 5119 * an op? 5120 */ 5121 if (et->et_ops == &ocfs2_dinode_et_ops) 5122 ocfs2_extent_map_trunc(inode, 0); 5123 5124 left_path = ocfs2_new_path_from_et(et); 5125 if (!left_path) { 5126 ret = -ENOMEM; 5127 mlog_errno(ret); 5128 goto out; 5129 } 5130 5131 ret = ocfs2_find_path(et->et_ci, left_path, cpos); 5132 if (ret) { 5133 mlog_errno(ret); 5134 goto out; 5135 } 5136 el = path_leaf_el(left_path); 5137 5138 index = ocfs2_search_extent_list(el, cpos); 5139 if (index == -1 || index >= le16_to_cpu(el->l_next_free_rec)) { 5140 ocfs2_error(inode->i_sb, 5141 "Inode %llu has an extent at cpos %u which can no " 5142 "longer be found.\n", 5143 (unsigned long long)OCFS2_I(inode)->ip_blkno, cpos); 5144 ret = -EROFS; 5145 goto out; 5146 } 5147 5148 memset(&split_rec, 0, sizeof(struct ocfs2_extent_rec)); 5149 split_rec.e_cpos = cpu_to_le32(cpos); 5150 split_rec.e_leaf_clusters = cpu_to_le16(len); 5151 split_rec.e_blkno = cpu_to_le64(start_blkno); 5152 split_rec.e_flags = path_leaf_el(left_path)->l_recs[index].e_flags; 5153 split_rec.e_flags &= ~OCFS2_EXT_UNWRITTEN; 5154 5155 ret = __ocfs2_mark_extent_written(inode, et, handle, left_path, 5156 index, &split_rec, meta_ac, 5157 dealloc); 5158 if (ret) 5159 mlog_errno(ret); 5160 5161 out: 5162 ocfs2_free_path(left_path); 5163 return ret; 5164 } 5165 5166 static int ocfs2_split_tree(struct inode *inode, struct ocfs2_extent_tree *et, 5167 handle_t *handle, struct ocfs2_path *path, 5168 int index, u32 new_range, 5169 struct ocfs2_alloc_context *meta_ac) 5170 { 5171 int ret, depth, credits = handle->h_buffer_credits; 5172 struct buffer_head *last_eb_bh = NULL; 5173 struct ocfs2_extent_block *eb; 5174 struct ocfs2_extent_list *rightmost_el, *el; 5175 struct ocfs2_extent_rec split_rec; 5176 struct ocfs2_extent_rec *rec; 5177 struct ocfs2_insert_type insert; 5178 5179 /* 5180 * Setup the record to split before we grow the tree. 5181 */ 5182 el = path_leaf_el(path); 5183 rec = &el->l_recs[index]; 5184 ocfs2_make_right_split_rec(inode->i_sb, &split_rec, new_range, rec); 5185 5186 depth = path->p_tree_depth; 5187 if (depth > 0) { 5188 ret = ocfs2_read_extent_block(et->et_ci, 5189 ocfs2_et_get_last_eb_blk(et), 5190 &last_eb_bh); 5191 if (ret < 0) { 5192 mlog_errno(ret); 5193 goto out; 5194 } 5195 5196 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data; 5197 rightmost_el = &eb->h_list; 5198 } else 5199 rightmost_el = path_leaf_el(path); 5200 5201 credits += path->p_tree_depth + 5202 ocfs2_extend_meta_needed(et->et_root_el); 5203 ret = ocfs2_extend_trans(handle, credits); 5204 if (ret) { 5205 mlog_errno(ret); 5206 goto out; 5207 } 5208 5209 if (le16_to_cpu(rightmost_el->l_next_free_rec) == 5210 le16_to_cpu(rightmost_el->l_count)) { 5211 ret = ocfs2_grow_tree(inode, handle, et, &depth, &last_eb_bh, 5212 meta_ac); 5213 if (ret) { 5214 mlog_errno(ret); 5215 goto out; 5216 } 5217 } 5218 5219 memset(&insert, 0, sizeof(struct ocfs2_insert_type)); 5220 insert.ins_appending = APPEND_NONE; 5221 insert.ins_contig = CONTIG_NONE; 5222 insert.ins_split = SPLIT_RIGHT; 5223 insert.ins_tree_depth = depth; 5224 5225 ret = ocfs2_do_insert_extent(inode, handle, et, &split_rec, &insert); 5226 if (ret) 5227 mlog_errno(ret); 5228 5229 out: 5230 brelse(last_eb_bh); 5231 return ret; 5232 } 5233 5234 static int ocfs2_truncate_rec(struct inode *inode, handle_t *handle, 5235 struct ocfs2_path *path, int index, 5236 struct ocfs2_cached_dealloc_ctxt *dealloc, 5237 u32 cpos, u32 len, 5238 struct ocfs2_extent_tree *et) 5239 { 5240 int ret; 5241 u32 left_cpos, rec_range, trunc_range; 5242 int wants_rotate = 0, is_rightmost_tree_rec = 0; 5243 struct super_block *sb = inode->i_sb; 5244 struct ocfs2_path *left_path = NULL; 5245 struct ocfs2_extent_list *el = path_leaf_el(path); 5246 struct ocfs2_extent_rec *rec; 5247 struct ocfs2_extent_block *eb; 5248 5249 if (ocfs2_is_empty_extent(&el->l_recs[0]) && index > 0) { 5250 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc); 5251 if (ret) { 5252 mlog_errno(ret); 5253 goto out; 5254 } 5255 5256 index--; 5257 } 5258 5259 if (index == (le16_to_cpu(el->l_next_free_rec) - 1) && 5260 path->p_tree_depth) { 5261 /* 5262 * Check whether this is the rightmost tree record. If 5263 * we remove all of this record or part of its right 5264 * edge then an update of the record lengths above it 5265 * will be required. 5266 */ 5267 eb = (struct ocfs2_extent_block *)path_leaf_bh(path)->b_data; 5268 if (eb->h_next_leaf_blk == 0) 5269 is_rightmost_tree_rec = 1; 5270 } 5271 5272 rec = &el->l_recs[index]; 5273 if (index == 0 && path->p_tree_depth && 5274 le32_to_cpu(rec->e_cpos) == cpos) { 5275 /* 5276 * Changing the leftmost offset (via partial or whole 5277 * record truncate) of an interior (or rightmost) path 5278 * means we have to update the subtree that is formed 5279 * by this leaf and the one to it's left. 5280 * 5281 * There are two cases we can skip: 5282 * 1) Path is the leftmost one in our inode tree. 5283 * 2) The leaf is rightmost and will be empty after 5284 * we remove the extent record - the rotate code 5285 * knows how to update the newly formed edge. 5286 */ 5287 5288 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, path, 5289 &left_cpos); 5290 if (ret) { 5291 mlog_errno(ret); 5292 goto out; 5293 } 5294 5295 if (left_cpos && le16_to_cpu(el->l_next_free_rec) > 1) { 5296 left_path = ocfs2_new_path_from_path(path); 5297 if (!left_path) { 5298 ret = -ENOMEM; 5299 mlog_errno(ret); 5300 goto out; 5301 } 5302 5303 ret = ocfs2_find_path(et->et_ci, left_path, 5304 left_cpos); 5305 if (ret) { 5306 mlog_errno(ret); 5307 goto out; 5308 } 5309 } 5310 } 5311 5312 ret = ocfs2_extend_rotate_transaction(handle, 0, 5313 handle->h_buffer_credits, 5314 path); 5315 if (ret) { 5316 mlog_errno(ret); 5317 goto out; 5318 } 5319 5320 ret = ocfs2_journal_access_path(et->et_ci, handle, path); 5321 if (ret) { 5322 mlog_errno(ret); 5323 goto out; 5324 } 5325 5326 ret = ocfs2_journal_access_path(et->et_ci, handle, left_path); 5327 if (ret) { 5328 mlog_errno(ret); 5329 goto out; 5330 } 5331 5332 rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec); 5333 trunc_range = cpos + len; 5334 5335 if (le32_to_cpu(rec->e_cpos) == cpos && rec_range == trunc_range) { 5336 int next_free; 5337 5338 memset(rec, 0, sizeof(*rec)); 5339 ocfs2_cleanup_merge(el, index); 5340 wants_rotate = 1; 5341 5342 next_free = le16_to_cpu(el->l_next_free_rec); 5343 if (is_rightmost_tree_rec && next_free > 1) { 5344 /* 5345 * We skip the edge update if this path will 5346 * be deleted by the rotate code. 5347 */ 5348 rec = &el->l_recs[next_free - 1]; 5349 ocfs2_adjust_rightmost_records(inode, handle, path, 5350 rec); 5351 } 5352 } else if (le32_to_cpu(rec->e_cpos) == cpos) { 5353 /* Remove leftmost portion of the record. */ 5354 le32_add_cpu(&rec->e_cpos, len); 5355 le64_add_cpu(&rec->e_blkno, ocfs2_clusters_to_blocks(sb, len)); 5356 le16_add_cpu(&rec->e_leaf_clusters, -len); 5357 } else if (rec_range == trunc_range) { 5358 /* Remove rightmost portion of the record */ 5359 le16_add_cpu(&rec->e_leaf_clusters, -len); 5360 if (is_rightmost_tree_rec) 5361 ocfs2_adjust_rightmost_records(inode, handle, path, rec); 5362 } else { 5363 /* Caller should have trapped this. */ 5364 mlog(ML_ERROR, "Inode %llu: Invalid record truncate: (%u, %u) " 5365 "(%u, %u)\n", (unsigned long long)OCFS2_I(inode)->ip_blkno, 5366 le32_to_cpu(rec->e_cpos), 5367 le16_to_cpu(rec->e_leaf_clusters), cpos, len); 5368 BUG(); 5369 } 5370 5371 if (left_path) { 5372 int subtree_index; 5373 5374 subtree_index = ocfs2_find_subtree_root(et, left_path, path); 5375 ocfs2_complete_edge_insert(handle, left_path, path, 5376 subtree_index); 5377 } 5378 5379 ocfs2_journal_dirty(handle, path_leaf_bh(path)); 5380 5381 ret = ocfs2_rotate_tree_left(handle, et, path, dealloc); 5382 if (ret) { 5383 mlog_errno(ret); 5384 goto out; 5385 } 5386 5387 out: 5388 ocfs2_free_path(left_path); 5389 return ret; 5390 } 5391 5392 int ocfs2_remove_extent(struct inode *inode, 5393 struct ocfs2_extent_tree *et, 5394 u32 cpos, u32 len, handle_t *handle, 5395 struct ocfs2_alloc_context *meta_ac, 5396 struct ocfs2_cached_dealloc_ctxt *dealloc) 5397 { 5398 int ret, index; 5399 u32 rec_range, trunc_range; 5400 struct ocfs2_extent_rec *rec; 5401 struct ocfs2_extent_list *el; 5402 struct ocfs2_path *path = NULL; 5403 5404 ocfs2_extent_map_trunc(inode, 0); 5405 5406 path = ocfs2_new_path_from_et(et); 5407 if (!path) { 5408 ret = -ENOMEM; 5409 mlog_errno(ret); 5410 goto out; 5411 } 5412 5413 ret = ocfs2_find_path(et->et_ci, path, cpos); 5414 if (ret) { 5415 mlog_errno(ret); 5416 goto out; 5417 } 5418 5419 el = path_leaf_el(path); 5420 index = ocfs2_search_extent_list(el, cpos); 5421 if (index == -1 || index >= le16_to_cpu(el->l_next_free_rec)) { 5422 ocfs2_error(inode->i_sb, 5423 "Inode %llu has an extent at cpos %u which can no " 5424 "longer be found.\n", 5425 (unsigned long long)OCFS2_I(inode)->ip_blkno, cpos); 5426 ret = -EROFS; 5427 goto out; 5428 } 5429 5430 /* 5431 * We have 3 cases of extent removal: 5432 * 1) Range covers the entire extent rec 5433 * 2) Range begins or ends on one edge of the extent rec 5434 * 3) Range is in the middle of the extent rec (no shared edges) 5435 * 5436 * For case 1 we remove the extent rec and left rotate to 5437 * fill the hole. 5438 * 5439 * For case 2 we just shrink the existing extent rec, with a 5440 * tree update if the shrinking edge is also the edge of an 5441 * extent block. 5442 * 5443 * For case 3 we do a right split to turn the extent rec into 5444 * something case 2 can handle. 5445 */ 5446 rec = &el->l_recs[index]; 5447 rec_range = le32_to_cpu(rec->e_cpos) + ocfs2_rec_clusters(el, rec); 5448 trunc_range = cpos + len; 5449 5450 BUG_ON(cpos < le32_to_cpu(rec->e_cpos) || trunc_range > rec_range); 5451 5452 mlog(0, "Inode %llu, remove (cpos %u, len %u). Existing index %d " 5453 "(cpos %u, len %u)\n", 5454 (unsigned long long)OCFS2_I(inode)->ip_blkno, cpos, len, index, 5455 le32_to_cpu(rec->e_cpos), ocfs2_rec_clusters(el, rec)); 5456 5457 if (le32_to_cpu(rec->e_cpos) == cpos || rec_range == trunc_range) { 5458 ret = ocfs2_truncate_rec(inode, handle, path, index, dealloc, 5459 cpos, len, et); 5460 if (ret) { 5461 mlog_errno(ret); 5462 goto out; 5463 } 5464 } else { 5465 ret = ocfs2_split_tree(inode, et, handle, path, index, 5466 trunc_range, meta_ac); 5467 if (ret) { 5468 mlog_errno(ret); 5469 goto out; 5470 } 5471 5472 /* 5473 * The split could have manipulated the tree enough to 5474 * move the record location, so we have to look for it again. 5475 */ 5476 ocfs2_reinit_path(path, 1); 5477 5478 ret = ocfs2_find_path(et->et_ci, path, cpos); 5479 if (ret) { 5480 mlog_errno(ret); 5481 goto out; 5482 } 5483 5484 el = path_leaf_el(path); 5485 index = ocfs2_search_extent_list(el, cpos); 5486 if (index == -1 || index >= le16_to_cpu(el->l_next_free_rec)) { 5487 ocfs2_error(inode->i_sb, 5488 "Inode %llu: split at cpos %u lost record.", 5489 (unsigned long long)OCFS2_I(inode)->ip_blkno, 5490 cpos); 5491 ret = -EROFS; 5492 goto out; 5493 } 5494 5495 /* 5496 * Double check our values here. If anything is fishy, 5497 * it's easier to catch it at the top level. 5498 */ 5499 rec = &el->l_recs[index]; 5500 rec_range = le32_to_cpu(rec->e_cpos) + 5501 ocfs2_rec_clusters(el, rec); 5502 if (rec_range != trunc_range) { 5503 ocfs2_error(inode->i_sb, 5504 "Inode %llu: error after split at cpos %u" 5505 "trunc len %u, existing record is (%u,%u)", 5506 (unsigned long long)OCFS2_I(inode)->ip_blkno, 5507 cpos, len, le32_to_cpu(rec->e_cpos), 5508 ocfs2_rec_clusters(el, rec)); 5509 ret = -EROFS; 5510 goto out; 5511 } 5512 5513 ret = ocfs2_truncate_rec(inode, handle, path, index, dealloc, 5514 cpos, len, et); 5515 if (ret) { 5516 mlog_errno(ret); 5517 goto out; 5518 } 5519 } 5520 5521 out: 5522 ocfs2_free_path(path); 5523 return ret; 5524 } 5525 5526 int ocfs2_remove_btree_range(struct inode *inode, 5527 struct ocfs2_extent_tree *et, 5528 u32 cpos, u32 phys_cpos, u32 len, 5529 struct ocfs2_cached_dealloc_ctxt *dealloc) 5530 { 5531 int ret; 5532 u64 phys_blkno = ocfs2_clusters_to_blocks(inode->i_sb, phys_cpos); 5533 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 5534 struct inode *tl_inode = osb->osb_tl_inode; 5535 handle_t *handle; 5536 struct ocfs2_alloc_context *meta_ac = NULL; 5537 5538 ret = ocfs2_lock_allocators(inode, et, 0, 1, NULL, &meta_ac); 5539 if (ret) { 5540 mlog_errno(ret); 5541 return ret; 5542 } 5543 5544 mutex_lock(&tl_inode->i_mutex); 5545 5546 if (ocfs2_truncate_log_needs_flush(osb)) { 5547 ret = __ocfs2_flush_truncate_log(osb); 5548 if (ret < 0) { 5549 mlog_errno(ret); 5550 goto out; 5551 } 5552 } 5553 5554 handle = ocfs2_start_trans(osb, ocfs2_remove_extent_credits(osb->sb)); 5555 if (IS_ERR(handle)) { 5556 ret = PTR_ERR(handle); 5557 mlog_errno(ret); 5558 goto out; 5559 } 5560 5561 ret = ocfs2_et_root_journal_access(handle, et, 5562 OCFS2_JOURNAL_ACCESS_WRITE); 5563 if (ret) { 5564 mlog_errno(ret); 5565 goto out; 5566 } 5567 5568 vfs_dq_free_space_nodirty(inode, 5569 ocfs2_clusters_to_bytes(inode->i_sb, len)); 5570 5571 ret = ocfs2_remove_extent(inode, et, cpos, len, handle, meta_ac, 5572 dealloc); 5573 if (ret) { 5574 mlog_errno(ret); 5575 goto out_commit; 5576 } 5577 5578 ocfs2_et_update_clusters(et, -len); 5579 5580 ret = ocfs2_journal_dirty(handle, et->et_root_bh); 5581 if (ret) { 5582 mlog_errno(ret); 5583 goto out_commit; 5584 } 5585 5586 ret = ocfs2_truncate_log_append(osb, handle, phys_blkno, len); 5587 if (ret) 5588 mlog_errno(ret); 5589 5590 out_commit: 5591 ocfs2_commit_trans(osb, handle); 5592 out: 5593 mutex_unlock(&tl_inode->i_mutex); 5594 5595 if (meta_ac) 5596 ocfs2_free_alloc_context(meta_ac); 5597 5598 return ret; 5599 } 5600 5601 int ocfs2_truncate_log_needs_flush(struct ocfs2_super *osb) 5602 { 5603 struct buffer_head *tl_bh = osb->osb_tl_bh; 5604 struct ocfs2_dinode *di; 5605 struct ocfs2_truncate_log *tl; 5606 5607 di = (struct ocfs2_dinode *) tl_bh->b_data; 5608 tl = &di->id2.i_dealloc; 5609 5610 mlog_bug_on_msg(le16_to_cpu(tl->tl_used) > le16_to_cpu(tl->tl_count), 5611 "slot %d, invalid truncate log parameters: used = " 5612 "%u, count = %u\n", osb->slot_num, 5613 le16_to_cpu(tl->tl_used), le16_to_cpu(tl->tl_count)); 5614 return le16_to_cpu(tl->tl_used) == le16_to_cpu(tl->tl_count); 5615 } 5616 5617 static int ocfs2_truncate_log_can_coalesce(struct ocfs2_truncate_log *tl, 5618 unsigned int new_start) 5619 { 5620 unsigned int tail_index; 5621 unsigned int current_tail; 5622 5623 /* No records, nothing to coalesce */ 5624 if (!le16_to_cpu(tl->tl_used)) 5625 return 0; 5626 5627 tail_index = le16_to_cpu(tl->tl_used) - 1; 5628 current_tail = le32_to_cpu(tl->tl_recs[tail_index].t_start); 5629 current_tail += le32_to_cpu(tl->tl_recs[tail_index].t_clusters); 5630 5631 return current_tail == new_start; 5632 } 5633 5634 int ocfs2_truncate_log_append(struct ocfs2_super *osb, 5635 handle_t *handle, 5636 u64 start_blk, 5637 unsigned int num_clusters) 5638 { 5639 int status, index; 5640 unsigned int start_cluster, tl_count; 5641 struct inode *tl_inode = osb->osb_tl_inode; 5642 struct buffer_head *tl_bh = osb->osb_tl_bh; 5643 struct ocfs2_dinode *di; 5644 struct ocfs2_truncate_log *tl; 5645 5646 mlog_entry("start_blk = %llu, num_clusters = %u\n", 5647 (unsigned long long)start_blk, num_clusters); 5648 5649 BUG_ON(mutex_trylock(&tl_inode->i_mutex)); 5650 5651 start_cluster = ocfs2_blocks_to_clusters(osb->sb, start_blk); 5652 5653 di = (struct ocfs2_dinode *) tl_bh->b_data; 5654 5655 /* tl_bh is loaded from ocfs2_truncate_log_init(). It's validated 5656 * by the underlying call to ocfs2_read_inode_block(), so any 5657 * corruption is a code bug */ 5658 BUG_ON(!OCFS2_IS_VALID_DINODE(di)); 5659 5660 tl = &di->id2.i_dealloc; 5661 tl_count = le16_to_cpu(tl->tl_count); 5662 mlog_bug_on_msg(tl_count > ocfs2_truncate_recs_per_inode(osb->sb) || 5663 tl_count == 0, 5664 "Truncate record count on #%llu invalid " 5665 "wanted %u, actual %u\n", 5666 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno, 5667 ocfs2_truncate_recs_per_inode(osb->sb), 5668 le16_to_cpu(tl->tl_count)); 5669 5670 /* Caller should have known to flush before calling us. */ 5671 index = le16_to_cpu(tl->tl_used); 5672 if (index >= tl_count) { 5673 status = -ENOSPC; 5674 mlog_errno(status); 5675 goto bail; 5676 } 5677 5678 status = ocfs2_journal_access_di(handle, INODE_CACHE(tl_inode), tl_bh, 5679 OCFS2_JOURNAL_ACCESS_WRITE); 5680 if (status < 0) { 5681 mlog_errno(status); 5682 goto bail; 5683 } 5684 5685 mlog(0, "Log truncate of %u clusters starting at cluster %u to " 5686 "%llu (index = %d)\n", num_clusters, start_cluster, 5687 (unsigned long long)OCFS2_I(tl_inode)->ip_blkno, index); 5688 5689 if (ocfs2_truncate_log_can_coalesce(tl, start_cluster)) { 5690 /* 5691 * Move index back to the record we are coalescing with. 5692 * ocfs2_truncate_log_can_coalesce() guarantees nonzero 5693 */ 5694 index--; 5695 5696 num_clusters += le32_to_cpu(tl->tl_recs[index].t_clusters); 5697 mlog(0, "Coalesce with index %u (start = %u, clusters = %u)\n", 5698 index, le32_to_cpu(tl->tl_recs[index].t_start), 5699 num_clusters); 5700 } else { 5701 tl->tl_recs[index].t_start = cpu_to_le32(start_cluster); 5702 tl->tl_used = cpu_to_le16(index + 1); 5703 } 5704 tl->tl_recs[index].t_clusters = cpu_to_le32(num_clusters); 5705 5706 status = ocfs2_journal_dirty(handle, tl_bh); 5707 if (status < 0) { 5708 mlog_errno(status); 5709 goto bail; 5710 } 5711 5712 bail: 5713 mlog_exit(status); 5714 return status; 5715 } 5716 5717 static int ocfs2_replay_truncate_records(struct ocfs2_super *osb, 5718 handle_t *handle, 5719 struct inode *data_alloc_inode, 5720 struct buffer_head *data_alloc_bh) 5721 { 5722 int status = 0; 5723 int i; 5724 unsigned int num_clusters; 5725 u64 start_blk; 5726 struct ocfs2_truncate_rec rec; 5727 struct ocfs2_dinode *di; 5728 struct ocfs2_truncate_log *tl; 5729 struct inode *tl_inode = osb->osb_tl_inode; 5730 struct buffer_head *tl_bh = osb->osb_tl_bh; 5731 5732 mlog_entry_void(); 5733 5734 di = (struct ocfs2_dinode *) tl_bh->b_data; 5735 tl = &di->id2.i_dealloc; 5736 i = le16_to_cpu(tl->tl_used) - 1; 5737 while (i >= 0) { 5738 /* Caller has given us at least enough credits to 5739 * update the truncate log dinode */ 5740 status = ocfs2_journal_access_di(handle, INODE_CACHE(tl_inode), tl_bh, 5741 OCFS2_JOURNAL_ACCESS_WRITE); 5742 if (status < 0) { 5743 mlog_errno(status); 5744 goto bail; 5745 } 5746 5747 tl->tl_used = cpu_to_le16(i); 5748 5749 status = ocfs2_journal_dirty(handle, tl_bh); 5750 if (status < 0) { 5751 mlog_errno(status); 5752 goto bail; 5753 } 5754 5755 /* TODO: Perhaps we can calculate the bulk of the 5756 * credits up front rather than extending like 5757 * this. */ 5758 status = ocfs2_extend_trans(handle, 5759 OCFS2_TRUNCATE_LOG_FLUSH_ONE_REC); 5760 if (status < 0) { 5761 mlog_errno(status); 5762 goto bail; 5763 } 5764 5765 rec = tl->tl_recs[i]; 5766 start_blk = ocfs2_clusters_to_blocks(data_alloc_inode->i_sb, 5767 le32_to_cpu(rec.t_start)); 5768 num_clusters = le32_to_cpu(rec.t_clusters); 5769 5770 /* if start_blk is not set, we ignore the record as 5771 * invalid. */ 5772 if (start_blk) { 5773 mlog(0, "free record %d, start = %u, clusters = %u\n", 5774 i, le32_to_cpu(rec.t_start), num_clusters); 5775 5776 status = ocfs2_free_clusters(handle, data_alloc_inode, 5777 data_alloc_bh, start_blk, 5778 num_clusters); 5779 if (status < 0) { 5780 mlog_errno(status); 5781 goto bail; 5782 } 5783 } 5784 i--; 5785 } 5786 5787 bail: 5788 mlog_exit(status); 5789 return status; 5790 } 5791 5792 /* Expects you to already be holding tl_inode->i_mutex */ 5793 int __ocfs2_flush_truncate_log(struct ocfs2_super *osb) 5794 { 5795 int status; 5796 unsigned int num_to_flush; 5797 handle_t *handle; 5798 struct inode *tl_inode = osb->osb_tl_inode; 5799 struct inode *data_alloc_inode = NULL; 5800 struct buffer_head *tl_bh = osb->osb_tl_bh; 5801 struct buffer_head *data_alloc_bh = NULL; 5802 struct ocfs2_dinode *di; 5803 struct ocfs2_truncate_log *tl; 5804 5805 mlog_entry_void(); 5806 5807 BUG_ON(mutex_trylock(&tl_inode->i_mutex)); 5808 5809 di = (struct ocfs2_dinode *) tl_bh->b_data; 5810 5811 /* tl_bh is loaded from ocfs2_truncate_log_init(). It's validated 5812 * by the underlying call to ocfs2_read_inode_block(), so any 5813 * corruption is a code bug */ 5814 BUG_ON(!OCFS2_IS_VALID_DINODE(di)); 5815 5816 tl = &di->id2.i_dealloc; 5817 num_to_flush = le16_to_cpu(tl->tl_used); 5818 mlog(0, "Flush %u records from truncate log #%llu\n", 5819 num_to_flush, (unsigned long long)OCFS2_I(tl_inode)->ip_blkno); 5820 if (!num_to_flush) { 5821 status = 0; 5822 goto out; 5823 } 5824 5825 data_alloc_inode = ocfs2_get_system_file_inode(osb, 5826 GLOBAL_BITMAP_SYSTEM_INODE, 5827 OCFS2_INVALID_SLOT); 5828 if (!data_alloc_inode) { 5829 status = -EINVAL; 5830 mlog(ML_ERROR, "Could not get bitmap inode!\n"); 5831 goto out; 5832 } 5833 5834 mutex_lock(&data_alloc_inode->i_mutex); 5835 5836 status = ocfs2_inode_lock(data_alloc_inode, &data_alloc_bh, 1); 5837 if (status < 0) { 5838 mlog_errno(status); 5839 goto out_mutex; 5840 } 5841 5842 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE); 5843 if (IS_ERR(handle)) { 5844 status = PTR_ERR(handle); 5845 mlog_errno(status); 5846 goto out_unlock; 5847 } 5848 5849 status = ocfs2_replay_truncate_records(osb, handle, data_alloc_inode, 5850 data_alloc_bh); 5851 if (status < 0) 5852 mlog_errno(status); 5853 5854 ocfs2_commit_trans(osb, handle); 5855 5856 out_unlock: 5857 brelse(data_alloc_bh); 5858 ocfs2_inode_unlock(data_alloc_inode, 1); 5859 5860 out_mutex: 5861 mutex_unlock(&data_alloc_inode->i_mutex); 5862 iput(data_alloc_inode); 5863 5864 out: 5865 mlog_exit(status); 5866 return status; 5867 } 5868 5869 int ocfs2_flush_truncate_log(struct ocfs2_super *osb) 5870 { 5871 int status; 5872 struct inode *tl_inode = osb->osb_tl_inode; 5873 5874 mutex_lock(&tl_inode->i_mutex); 5875 status = __ocfs2_flush_truncate_log(osb); 5876 mutex_unlock(&tl_inode->i_mutex); 5877 5878 return status; 5879 } 5880 5881 static void ocfs2_truncate_log_worker(struct work_struct *work) 5882 { 5883 int status; 5884 struct ocfs2_super *osb = 5885 container_of(work, struct ocfs2_super, 5886 osb_truncate_log_wq.work); 5887 5888 mlog_entry_void(); 5889 5890 status = ocfs2_flush_truncate_log(osb); 5891 if (status < 0) 5892 mlog_errno(status); 5893 else 5894 ocfs2_init_inode_steal_slot(osb); 5895 5896 mlog_exit(status); 5897 } 5898 5899 #define OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL (2 * HZ) 5900 void ocfs2_schedule_truncate_log_flush(struct ocfs2_super *osb, 5901 int cancel) 5902 { 5903 if (osb->osb_tl_inode) { 5904 /* We want to push off log flushes while truncates are 5905 * still running. */ 5906 if (cancel) 5907 cancel_delayed_work(&osb->osb_truncate_log_wq); 5908 5909 queue_delayed_work(ocfs2_wq, &osb->osb_truncate_log_wq, 5910 OCFS2_TRUNCATE_LOG_FLUSH_INTERVAL); 5911 } 5912 } 5913 5914 static int ocfs2_get_truncate_log_info(struct ocfs2_super *osb, 5915 int slot_num, 5916 struct inode **tl_inode, 5917 struct buffer_head **tl_bh) 5918 { 5919 int status; 5920 struct inode *inode = NULL; 5921 struct buffer_head *bh = NULL; 5922 5923 inode = ocfs2_get_system_file_inode(osb, 5924 TRUNCATE_LOG_SYSTEM_INODE, 5925 slot_num); 5926 if (!inode) { 5927 status = -EINVAL; 5928 mlog(ML_ERROR, "Could not get load truncate log inode!\n"); 5929 goto bail; 5930 } 5931 5932 status = ocfs2_read_inode_block(inode, &bh); 5933 if (status < 0) { 5934 iput(inode); 5935 mlog_errno(status); 5936 goto bail; 5937 } 5938 5939 *tl_inode = inode; 5940 *tl_bh = bh; 5941 bail: 5942 mlog_exit(status); 5943 return status; 5944 } 5945 5946 /* called during the 1st stage of node recovery. we stamp a clean 5947 * truncate log and pass back a copy for processing later. if the 5948 * truncate log does not require processing, a *tl_copy is set to 5949 * NULL. */ 5950 int ocfs2_begin_truncate_log_recovery(struct ocfs2_super *osb, 5951 int slot_num, 5952 struct ocfs2_dinode **tl_copy) 5953 { 5954 int status; 5955 struct inode *tl_inode = NULL; 5956 struct buffer_head *tl_bh = NULL; 5957 struct ocfs2_dinode *di; 5958 struct ocfs2_truncate_log *tl; 5959 5960 *tl_copy = NULL; 5961 5962 mlog(0, "recover truncate log from slot %d\n", slot_num); 5963 5964 status = ocfs2_get_truncate_log_info(osb, slot_num, &tl_inode, &tl_bh); 5965 if (status < 0) { 5966 mlog_errno(status); 5967 goto bail; 5968 } 5969 5970 di = (struct ocfs2_dinode *) tl_bh->b_data; 5971 5972 /* tl_bh is loaded from ocfs2_get_truncate_log_info(). It's 5973 * validated by the underlying call to ocfs2_read_inode_block(), 5974 * so any corruption is a code bug */ 5975 BUG_ON(!OCFS2_IS_VALID_DINODE(di)); 5976 5977 tl = &di->id2.i_dealloc; 5978 if (le16_to_cpu(tl->tl_used)) { 5979 mlog(0, "We'll have %u logs to recover\n", 5980 le16_to_cpu(tl->tl_used)); 5981 5982 *tl_copy = kmalloc(tl_bh->b_size, GFP_KERNEL); 5983 if (!(*tl_copy)) { 5984 status = -ENOMEM; 5985 mlog_errno(status); 5986 goto bail; 5987 } 5988 5989 /* Assuming the write-out below goes well, this copy 5990 * will be passed back to recovery for processing. */ 5991 memcpy(*tl_copy, tl_bh->b_data, tl_bh->b_size); 5992 5993 /* All we need to do to clear the truncate log is set 5994 * tl_used. */ 5995 tl->tl_used = 0; 5996 5997 ocfs2_compute_meta_ecc(osb->sb, tl_bh->b_data, &di->i_check); 5998 status = ocfs2_write_block(osb, tl_bh, INODE_CACHE(tl_inode)); 5999 if (status < 0) { 6000 mlog_errno(status); 6001 goto bail; 6002 } 6003 } 6004 6005 bail: 6006 if (tl_inode) 6007 iput(tl_inode); 6008 brelse(tl_bh); 6009 6010 if (status < 0 && (*tl_copy)) { 6011 kfree(*tl_copy); 6012 *tl_copy = NULL; 6013 } 6014 6015 mlog_exit(status); 6016 return status; 6017 } 6018 6019 int ocfs2_complete_truncate_log_recovery(struct ocfs2_super *osb, 6020 struct ocfs2_dinode *tl_copy) 6021 { 6022 int status = 0; 6023 int i; 6024 unsigned int clusters, num_recs, start_cluster; 6025 u64 start_blk; 6026 handle_t *handle; 6027 struct inode *tl_inode = osb->osb_tl_inode; 6028 struct ocfs2_truncate_log *tl; 6029 6030 mlog_entry_void(); 6031 6032 if (OCFS2_I(tl_inode)->ip_blkno == le64_to_cpu(tl_copy->i_blkno)) { 6033 mlog(ML_ERROR, "Asked to recover my own truncate log!\n"); 6034 return -EINVAL; 6035 } 6036 6037 tl = &tl_copy->id2.i_dealloc; 6038 num_recs = le16_to_cpu(tl->tl_used); 6039 mlog(0, "cleanup %u records from %llu\n", num_recs, 6040 (unsigned long long)le64_to_cpu(tl_copy->i_blkno)); 6041 6042 mutex_lock(&tl_inode->i_mutex); 6043 for(i = 0; i < num_recs; i++) { 6044 if (ocfs2_truncate_log_needs_flush(osb)) { 6045 status = __ocfs2_flush_truncate_log(osb); 6046 if (status < 0) { 6047 mlog_errno(status); 6048 goto bail_up; 6049 } 6050 } 6051 6052 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE); 6053 if (IS_ERR(handle)) { 6054 status = PTR_ERR(handle); 6055 mlog_errno(status); 6056 goto bail_up; 6057 } 6058 6059 clusters = le32_to_cpu(tl->tl_recs[i].t_clusters); 6060 start_cluster = le32_to_cpu(tl->tl_recs[i].t_start); 6061 start_blk = ocfs2_clusters_to_blocks(osb->sb, start_cluster); 6062 6063 status = ocfs2_truncate_log_append(osb, handle, 6064 start_blk, clusters); 6065 ocfs2_commit_trans(osb, handle); 6066 if (status < 0) { 6067 mlog_errno(status); 6068 goto bail_up; 6069 } 6070 } 6071 6072 bail_up: 6073 mutex_unlock(&tl_inode->i_mutex); 6074 6075 mlog_exit(status); 6076 return status; 6077 } 6078 6079 void ocfs2_truncate_log_shutdown(struct ocfs2_super *osb) 6080 { 6081 int status; 6082 struct inode *tl_inode = osb->osb_tl_inode; 6083 6084 mlog_entry_void(); 6085 6086 if (tl_inode) { 6087 cancel_delayed_work(&osb->osb_truncate_log_wq); 6088 flush_workqueue(ocfs2_wq); 6089 6090 status = ocfs2_flush_truncate_log(osb); 6091 if (status < 0) 6092 mlog_errno(status); 6093 6094 brelse(osb->osb_tl_bh); 6095 iput(osb->osb_tl_inode); 6096 } 6097 6098 mlog_exit_void(); 6099 } 6100 6101 int ocfs2_truncate_log_init(struct ocfs2_super *osb) 6102 { 6103 int status; 6104 struct inode *tl_inode = NULL; 6105 struct buffer_head *tl_bh = NULL; 6106 6107 mlog_entry_void(); 6108 6109 status = ocfs2_get_truncate_log_info(osb, 6110 osb->slot_num, 6111 &tl_inode, 6112 &tl_bh); 6113 if (status < 0) 6114 mlog_errno(status); 6115 6116 /* ocfs2_truncate_log_shutdown keys on the existence of 6117 * osb->osb_tl_inode so we don't set any of the osb variables 6118 * until we're sure all is well. */ 6119 INIT_DELAYED_WORK(&osb->osb_truncate_log_wq, 6120 ocfs2_truncate_log_worker); 6121 osb->osb_tl_bh = tl_bh; 6122 osb->osb_tl_inode = tl_inode; 6123 6124 mlog_exit(status); 6125 return status; 6126 } 6127 6128 /* 6129 * Delayed de-allocation of suballocator blocks. 6130 * 6131 * Some sets of block de-allocations might involve multiple suballocator inodes. 6132 * 6133 * The locking for this can get extremely complicated, especially when 6134 * the suballocator inodes to delete from aren't known until deep 6135 * within an unrelated codepath. 6136 * 6137 * ocfs2_extent_block structures are a good example of this - an inode 6138 * btree could have been grown by any number of nodes each allocating 6139 * out of their own suballoc inode. 6140 * 6141 * These structures allow the delay of block de-allocation until a 6142 * later time, when locking of multiple cluster inodes won't cause 6143 * deadlock. 6144 */ 6145 6146 /* 6147 * Describe a single bit freed from a suballocator. For the block 6148 * suballocators, it represents one block. For the global cluster 6149 * allocator, it represents some clusters and free_bit indicates 6150 * clusters number. 6151 */ 6152 struct ocfs2_cached_block_free { 6153 struct ocfs2_cached_block_free *free_next; 6154 u64 free_blk; 6155 unsigned int free_bit; 6156 }; 6157 6158 struct ocfs2_per_slot_free_list { 6159 struct ocfs2_per_slot_free_list *f_next_suballocator; 6160 int f_inode_type; 6161 int f_slot; 6162 struct ocfs2_cached_block_free *f_first; 6163 }; 6164 6165 static int ocfs2_free_cached_blocks(struct ocfs2_super *osb, 6166 int sysfile_type, 6167 int slot, 6168 struct ocfs2_cached_block_free *head) 6169 { 6170 int ret; 6171 u64 bg_blkno; 6172 handle_t *handle; 6173 struct inode *inode; 6174 struct buffer_head *di_bh = NULL; 6175 struct ocfs2_cached_block_free *tmp; 6176 6177 inode = ocfs2_get_system_file_inode(osb, sysfile_type, slot); 6178 if (!inode) { 6179 ret = -EINVAL; 6180 mlog_errno(ret); 6181 goto out; 6182 } 6183 6184 mutex_lock(&inode->i_mutex); 6185 6186 ret = ocfs2_inode_lock(inode, &di_bh, 1); 6187 if (ret) { 6188 mlog_errno(ret); 6189 goto out_mutex; 6190 } 6191 6192 handle = ocfs2_start_trans(osb, OCFS2_SUBALLOC_FREE); 6193 if (IS_ERR(handle)) { 6194 ret = PTR_ERR(handle); 6195 mlog_errno(ret); 6196 goto out_unlock; 6197 } 6198 6199 while (head) { 6200 bg_blkno = ocfs2_which_suballoc_group(head->free_blk, 6201 head->free_bit); 6202 mlog(0, "Free bit: (bit %u, blkno %llu)\n", 6203 head->free_bit, (unsigned long long)head->free_blk); 6204 6205 ret = ocfs2_free_suballoc_bits(handle, inode, di_bh, 6206 head->free_bit, bg_blkno, 1); 6207 if (ret) { 6208 mlog_errno(ret); 6209 goto out_journal; 6210 } 6211 6212 ret = ocfs2_extend_trans(handle, OCFS2_SUBALLOC_FREE); 6213 if (ret) { 6214 mlog_errno(ret); 6215 goto out_journal; 6216 } 6217 6218 tmp = head; 6219 head = head->free_next; 6220 kfree(tmp); 6221 } 6222 6223 out_journal: 6224 ocfs2_commit_trans(osb, handle); 6225 6226 out_unlock: 6227 ocfs2_inode_unlock(inode, 1); 6228 brelse(di_bh); 6229 out_mutex: 6230 mutex_unlock(&inode->i_mutex); 6231 iput(inode); 6232 out: 6233 while(head) { 6234 /* Premature exit may have left some dangling items. */ 6235 tmp = head; 6236 head = head->free_next; 6237 kfree(tmp); 6238 } 6239 6240 return ret; 6241 } 6242 6243 int ocfs2_cache_cluster_dealloc(struct ocfs2_cached_dealloc_ctxt *ctxt, 6244 u64 blkno, unsigned int bit) 6245 { 6246 int ret = 0; 6247 struct ocfs2_cached_block_free *item; 6248 6249 item = kmalloc(sizeof(*item), GFP_NOFS); 6250 if (item == NULL) { 6251 ret = -ENOMEM; 6252 mlog_errno(ret); 6253 return ret; 6254 } 6255 6256 mlog(0, "Insert clusters: (bit %u, blk %llu)\n", 6257 bit, (unsigned long long)blkno); 6258 6259 item->free_blk = blkno; 6260 item->free_bit = bit; 6261 item->free_next = ctxt->c_global_allocator; 6262 6263 ctxt->c_global_allocator = item; 6264 return ret; 6265 } 6266 6267 static int ocfs2_free_cached_clusters(struct ocfs2_super *osb, 6268 struct ocfs2_cached_block_free *head) 6269 { 6270 struct ocfs2_cached_block_free *tmp; 6271 struct inode *tl_inode = osb->osb_tl_inode; 6272 handle_t *handle; 6273 int ret = 0; 6274 6275 mutex_lock(&tl_inode->i_mutex); 6276 6277 while (head) { 6278 if (ocfs2_truncate_log_needs_flush(osb)) { 6279 ret = __ocfs2_flush_truncate_log(osb); 6280 if (ret < 0) { 6281 mlog_errno(ret); 6282 break; 6283 } 6284 } 6285 6286 handle = ocfs2_start_trans(osb, OCFS2_TRUNCATE_LOG_UPDATE); 6287 if (IS_ERR(handle)) { 6288 ret = PTR_ERR(handle); 6289 mlog_errno(ret); 6290 break; 6291 } 6292 6293 ret = ocfs2_truncate_log_append(osb, handle, head->free_blk, 6294 head->free_bit); 6295 6296 ocfs2_commit_trans(osb, handle); 6297 tmp = head; 6298 head = head->free_next; 6299 kfree(tmp); 6300 6301 if (ret < 0) { 6302 mlog_errno(ret); 6303 break; 6304 } 6305 } 6306 6307 mutex_unlock(&tl_inode->i_mutex); 6308 6309 while (head) { 6310 /* Premature exit may have left some dangling items. */ 6311 tmp = head; 6312 head = head->free_next; 6313 kfree(tmp); 6314 } 6315 6316 return ret; 6317 } 6318 6319 int ocfs2_run_deallocs(struct ocfs2_super *osb, 6320 struct ocfs2_cached_dealloc_ctxt *ctxt) 6321 { 6322 int ret = 0, ret2; 6323 struct ocfs2_per_slot_free_list *fl; 6324 6325 if (!ctxt) 6326 return 0; 6327 6328 while (ctxt->c_first_suballocator) { 6329 fl = ctxt->c_first_suballocator; 6330 6331 if (fl->f_first) { 6332 mlog(0, "Free items: (type %u, slot %d)\n", 6333 fl->f_inode_type, fl->f_slot); 6334 ret2 = ocfs2_free_cached_blocks(osb, 6335 fl->f_inode_type, 6336 fl->f_slot, 6337 fl->f_first); 6338 if (ret2) 6339 mlog_errno(ret2); 6340 if (!ret) 6341 ret = ret2; 6342 } 6343 6344 ctxt->c_first_suballocator = fl->f_next_suballocator; 6345 kfree(fl); 6346 } 6347 6348 if (ctxt->c_global_allocator) { 6349 ret2 = ocfs2_free_cached_clusters(osb, 6350 ctxt->c_global_allocator); 6351 if (ret2) 6352 mlog_errno(ret2); 6353 if (!ret) 6354 ret = ret2; 6355 6356 ctxt->c_global_allocator = NULL; 6357 } 6358 6359 return ret; 6360 } 6361 6362 static struct ocfs2_per_slot_free_list * 6363 ocfs2_find_per_slot_free_list(int type, 6364 int slot, 6365 struct ocfs2_cached_dealloc_ctxt *ctxt) 6366 { 6367 struct ocfs2_per_slot_free_list *fl = ctxt->c_first_suballocator; 6368 6369 while (fl) { 6370 if (fl->f_inode_type == type && fl->f_slot == slot) 6371 return fl; 6372 6373 fl = fl->f_next_suballocator; 6374 } 6375 6376 fl = kmalloc(sizeof(*fl), GFP_NOFS); 6377 if (fl) { 6378 fl->f_inode_type = type; 6379 fl->f_slot = slot; 6380 fl->f_first = NULL; 6381 fl->f_next_suballocator = ctxt->c_first_suballocator; 6382 6383 ctxt->c_first_suballocator = fl; 6384 } 6385 return fl; 6386 } 6387 6388 static int ocfs2_cache_block_dealloc(struct ocfs2_cached_dealloc_ctxt *ctxt, 6389 int type, int slot, u64 blkno, 6390 unsigned int bit) 6391 { 6392 int ret; 6393 struct ocfs2_per_slot_free_list *fl; 6394 struct ocfs2_cached_block_free *item; 6395 6396 fl = ocfs2_find_per_slot_free_list(type, slot, ctxt); 6397 if (fl == NULL) { 6398 ret = -ENOMEM; 6399 mlog_errno(ret); 6400 goto out; 6401 } 6402 6403 item = kmalloc(sizeof(*item), GFP_NOFS); 6404 if (item == NULL) { 6405 ret = -ENOMEM; 6406 mlog_errno(ret); 6407 goto out; 6408 } 6409 6410 mlog(0, "Insert: (type %d, slot %u, bit %u, blk %llu)\n", 6411 type, slot, bit, (unsigned long long)blkno); 6412 6413 item->free_blk = blkno; 6414 item->free_bit = bit; 6415 item->free_next = fl->f_first; 6416 6417 fl->f_first = item; 6418 6419 ret = 0; 6420 out: 6421 return ret; 6422 } 6423 6424 static int ocfs2_cache_extent_block_free(struct ocfs2_cached_dealloc_ctxt *ctxt, 6425 struct ocfs2_extent_block *eb) 6426 { 6427 return ocfs2_cache_block_dealloc(ctxt, EXTENT_ALLOC_SYSTEM_INODE, 6428 le16_to_cpu(eb->h_suballoc_slot), 6429 le64_to_cpu(eb->h_blkno), 6430 le16_to_cpu(eb->h_suballoc_bit)); 6431 } 6432 6433 /* This function will figure out whether the currently last extent 6434 * block will be deleted, and if it will, what the new last extent 6435 * block will be so we can update his h_next_leaf_blk field, as well 6436 * as the dinodes i_last_eb_blk */ 6437 static int ocfs2_find_new_last_ext_blk(struct inode *inode, 6438 unsigned int clusters_to_del, 6439 struct ocfs2_path *path, 6440 struct buffer_head **new_last_eb) 6441 { 6442 int next_free, ret = 0; 6443 u32 cpos; 6444 struct ocfs2_extent_rec *rec; 6445 struct ocfs2_extent_block *eb; 6446 struct ocfs2_extent_list *el; 6447 struct buffer_head *bh = NULL; 6448 6449 *new_last_eb = NULL; 6450 6451 /* we have no tree, so of course, no last_eb. */ 6452 if (!path->p_tree_depth) 6453 goto out; 6454 6455 /* trunc to zero special case - this makes tree_depth = 0 6456 * regardless of what it is. */ 6457 if (OCFS2_I(inode)->ip_clusters == clusters_to_del) 6458 goto out; 6459 6460 el = path_leaf_el(path); 6461 BUG_ON(!el->l_next_free_rec); 6462 6463 /* 6464 * Make sure that this extent list will actually be empty 6465 * after we clear away the data. We can shortcut out if 6466 * there's more than one non-empty extent in the 6467 * list. Otherwise, a check of the remaining extent is 6468 * necessary. 6469 */ 6470 next_free = le16_to_cpu(el->l_next_free_rec); 6471 rec = NULL; 6472 if (ocfs2_is_empty_extent(&el->l_recs[0])) { 6473 if (next_free > 2) 6474 goto out; 6475 6476 /* We may have a valid extent in index 1, check it. */ 6477 if (next_free == 2) 6478 rec = &el->l_recs[1]; 6479 6480 /* 6481 * Fall through - no more nonempty extents, so we want 6482 * to delete this leaf. 6483 */ 6484 } else { 6485 if (next_free > 1) 6486 goto out; 6487 6488 rec = &el->l_recs[0]; 6489 } 6490 6491 if (rec) { 6492 /* 6493 * Check it we'll only be trimming off the end of this 6494 * cluster. 6495 */ 6496 if (le16_to_cpu(rec->e_leaf_clusters) > clusters_to_del) 6497 goto out; 6498 } 6499 6500 ret = ocfs2_find_cpos_for_left_leaf(inode->i_sb, path, &cpos); 6501 if (ret) { 6502 mlog_errno(ret); 6503 goto out; 6504 } 6505 6506 ret = ocfs2_find_leaf(INODE_CACHE(inode), path_root_el(path), cpos, &bh); 6507 if (ret) { 6508 mlog_errno(ret); 6509 goto out; 6510 } 6511 6512 eb = (struct ocfs2_extent_block *) bh->b_data; 6513 el = &eb->h_list; 6514 6515 /* ocfs2_find_leaf() gets the eb from ocfs2_read_extent_block(). 6516 * Any corruption is a code bug. */ 6517 BUG_ON(!OCFS2_IS_VALID_EXTENT_BLOCK(eb)); 6518 6519 *new_last_eb = bh; 6520 get_bh(*new_last_eb); 6521 mlog(0, "returning block %llu, (cpos: %u)\n", 6522 (unsigned long long)le64_to_cpu(eb->h_blkno), cpos); 6523 out: 6524 brelse(bh); 6525 6526 return ret; 6527 } 6528 6529 /* 6530 * Trim some clusters off the rightmost edge of a tree. Only called 6531 * during truncate. 6532 * 6533 * The caller needs to: 6534 * - start journaling of each path component. 6535 * - compute and fully set up any new last ext block 6536 */ 6537 static int ocfs2_trim_tree(struct inode *inode, struct ocfs2_path *path, 6538 handle_t *handle, struct ocfs2_truncate_context *tc, 6539 u32 clusters_to_del, u64 *delete_start) 6540 { 6541 int ret, i, index = path->p_tree_depth; 6542 u32 new_edge = 0; 6543 u64 deleted_eb = 0; 6544 struct buffer_head *bh; 6545 struct ocfs2_extent_list *el; 6546 struct ocfs2_extent_rec *rec; 6547 6548 *delete_start = 0; 6549 6550 while (index >= 0) { 6551 bh = path->p_node[index].bh; 6552 el = path->p_node[index].el; 6553 6554 mlog(0, "traveling tree (index = %d, block = %llu)\n", 6555 index, (unsigned long long)bh->b_blocknr); 6556 6557 BUG_ON(le16_to_cpu(el->l_next_free_rec) == 0); 6558 6559 if (index != 6560 (path->p_tree_depth - le16_to_cpu(el->l_tree_depth))) { 6561 ocfs2_error(inode->i_sb, 6562 "Inode %lu has invalid ext. block %llu", 6563 inode->i_ino, 6564 (unsigned long long)bh->b_blocknr); 6565 ret = -EROFS; 6566 goto out; 6567 } 6568 6569 find_tail_record: 6570 i = le16_to_cpu(el->l_next_free_rec) - 1; 6571 rec = &el->l_recs[i]; 6572 6573 mlog(0, "Extent list before: record %d: (%u, %u, %llu), " 6574 "next = %u\n", i, le32_to_cpu(rec->e_cpos), 6575 ocfs2_rec_clusters(el, rec), 6576 (unsigned long long)le64_to_cpu(rec->e_blkno), 6577 le16_to_cpu(el->l_next_free_rec)); 6578 6579 BUG_ON(ocfs2_rec_clusters(el, rec) < clusters_to_del); 6580 6581 if (le16_to_cpu(el->l_tree_depth) == 0) { 6582 /* 6583 * If the leaf block contains a single empty 6584 * extent and no records, we can just remove 6585 * the block. 6586 */ 6587 if (i == 0 && ocfs2_is_empty_extent(rec)) { 6588 memset(rec, 0, 6589 sizeof(struct ocfs2_extent_rec)); 6590 el->l_next_free_rec = cpu_to_le16(0); 6591 6592 goto delete; 6593 } 6594 6595 /* 6596 * Remove any empty extents by shifting things 6597 * left. That should make life much easier on 6598 * the code below. This condition is rare 6599 * enough that we shouldn't see a performance 6600 * hit. 6601 */ 6602 if (ocfs2_is_empty_extent(&el->l_recs[0])) { 6603 le16_add_cpu(&el->l_next_free_rec, -1); 6604 6605 for(i = 0; 6606 i < le16_to_cpu(el->l_next_free_rec); i++) 6607 el->l_recs[i] = el->l_recs[i + 1]; 6608 6609 memset(&el->l_recs[i], 0, 6610 sizeof(struct ocfs2_extent_rec)); 6611 6612 /* 6613 * We've modified our extent list. The 6614 * simplest way to handle this change 6615 * is to being the search from the 6616 * start again. 6617 */ 6618 goto find_tail_record; 6619 } 6620 6621 le16_add_cpu(&rec->e_leaf_clusters, -clusters_to_del); 6622 6623 /* 6624 * We'll use "new_edge" on our way back up the 6625 * tree to know what our rightmost cpos is. 6626 */ 6627 new_edge = le16_to_cpu(rec->e_leaf_clusters); 6628 new_edge += le32_to_cpu(rec->e_cpos); 6629 6630 /* 6631 * The caller will use this to delete data blocks. 6632 */ 6633 *delete_start = le64_to_cpu(rec->e_blkno) 6634 + ocfs2_clusters_to_blocks(inode->i_sb, 6635 le16_to_cpu(rec->e_leaf_clusters)); 6636 6637 /* 6638 * If it's now empty, remove this record. 6639 */ 6640 if (le16_to_cpu(rec->e_leaf_clusters) == 0) { 6641 memset(rec, 0, 6642 sizeof(struct ocfs2_extent_rec)); 6643 le16_add_cpu(&el->l_next_free_rec, -1); 6644 } 6645 } else { 6646 if (le64_to_cpu(rec->e_blkno) == deleted_eb) { 6647 memset(rec, 0, 6648 sizeof(struct ocfs2_extent_rec)); 6649 le16_add_cpu(&el->l_next_free_rec, -1); 6650 6651 goto delete; 6652 } 6653 6654 /* Can this actually happen? */ 6655 if (le16_to_cpu(el->l_next_free_rec) == 0) 6656 goto delete; 6657 6658 /* 6659 * We never actually deleted any clusters 6660 * because our leaf was empty. There's no 6661 * reason to adjust the rightmost edge then. 6662 */ 6663 if (new_edge == 0) 6664 goto delete; 6665 6666 rec->e_int_clusters = cpu_to_le32(new_edge); 6667 le32_add_cpu(&rec->e_int_clusters, 6668 -le32_to_cpu(rec->e_cpos)); 6669 6670 /* 6671 * A deleted child record should have been 6672 * caught above. 6673 */ 6674 BUG_ON(le32_to_cpu(rec->e_int_clusters) == 0); 6675 } 6676 6677 delete: 6678 ret = ocfs2_journal_dirty(handle, bh); 6679 if (ret) { 6680 mlog_errno(ret); 6681 goto out; 6682 } 6683 6684 mlog(0, "extent list container %llu, after: record %d: " 6685 "(%u, %u, %llu), next = %u.\n", 6686 (unsigned long long)bh->b_blocknr, i, 6687 le32_to_cpu(rec->e_cpos), ocfs2_rec_clusters(el, rec), 6688 (unsigned long long)le64_to_cpu(rec->e_blkno), 6689 le16_to_cpu(el->l_next_free_rec)); 6690 6691 /* 6692 * We must be careful to only attempt delete of an 6693 * extent block (and not the root inode block). 6694 */ 6695 if (index > 0 && le16_to_cpu(el->l_next_free_rec) == 0) { 6696 struct ocfs2_extent_block *eb = 6697 (struct ocfs2_extent_block *)bh->b_data; 6698 6699 /* 6700 * Save this for use when processing the 6701 * parent block. 6702 */ 6703 deleted_eb = le64_to_cpu(eb->h_blkno); 6704 6705 mlog(0, "deleting this extent block.\n"); 6706 6707 ocfs2_remove_from_cache(INODE_CACHE(inode), bh); 6708 6709 BUG_ON(ocfs2_rec_clusters(el, &el->l_recs[0])); 6710 BUG_ON(le32_to_cpu(el->l_recs[0].e_cpos)); 6711 BUG_ON(le64_to_cpu(el->l_recs[0].e_blkno)); 6712 6713 ret = ocfs2_cache_extent_block_free(&tc->tc_dealloc, eb); 6714 /* An error here is not fatal. */ 6715 if (ret < 0) 6716 mlog_errno(ret); 6717 } else { 6718 deleted_eb = 0; 6719 } 6720 6721 index--; 6722 } 6723 6724 ret = 0; 6725 out: 6726 return ret; 6727 } 6728 6729 static int ocfs2_do_truncate(struct ocfs2_super *osb, 6730 unsigned int clusters_to_del, 6731 struct inode *inode, 6732 struct buffer_head *fe_bh, 6733 handle_t *handle, 6734 struct ocfs2_truncate_context *tc, 6735 struct ocfs2_path *path) 6736 { 6737 int status; 6738 struct ocfs2_dinode *fe; 6739 struct ocfs2_extent_block *last_eb = NULL; 6740 struct ocfs2_extent_list *el; 6741 struct buffer_head *last_eb_bh = NULL; 6742 u64 delete_blk = 0; 6743 6744 fe = (struct ocfs2_dinode *) fe_bh->b_data; 6745 6746 status = ocfs2_find_new_last_ext_blk(inode, clusters_to_del, 6747 path, &last_eb_bh); 6748 if (status < 0) { 6749 mlog_errno(status); 6750 goto bail; 6751 } 6752 6753 /* 6754 * Each component will be touched, so we might as well journal 6755 * here to avoid having to handle errors later. 6756 */ 6757 status = ocfs2_journal_access_path(INODE_CACHE(inode), handle, path); 6758 if (status < 0) { 6759 mlog_errno(status); 6760 goto bail; 6761 } 6762 6763 if (last_eb_bh) { 6764 status = ocfs2_journal_access_eb(handle, INODE_CACHE(inode), last_eb_bh, 6765 OCFS2_JOURNAL_ACCESS_WRITE); 6766 if (status < 0) { 6767 mlog_errno(status); 6768 goto bail; 6769 } 6770 6771 last_eb = (struct ocfs2_extent_block *) last_eb_bh->b_data; 6772 } 6773 6774 el = &(fe->id2.i_list); 6775 6776 /* 6777 * Lower levels depend on this never happening, but it's best 6778 * to check it up here before changing the tree. 6779 */ 6780 if (el->l_tree_depth && el->l_recs[0].e_int_clusters == 0) { 6781 ocfs2_error(inode->i_sb, 6782 "Inode %lu has an empty extent record, depth %u\n", 6783 inode->i_ino, le16_to_cpu(el->l_tree_depth)); 6784 status = -EROFS; 6785 goto bail; 6786 } 6787 6788 vfs_dq_free_space_nodirty(inode, 6789 ocfs2_clusters_to_bytes(osb->sb, clusters_to_del)); 6790 spin_lock(&OCFS2_I(inode)->ip_lock); 6791 OCFS2_I(inode)->ip_clusters = le32_to_cpu(fe->i_clusters) - 6792 clusters_to_del; 6793 spin_unlock(&OCFS2_I(inode)->ip_lock); 6794 le32_add_cpu(&fe->i_clusters, -clusters_to_del); 6795 inode->i_blocks = ocfs2_inode_sector_count(inode); 6796 6797 status = ocfs2_trim_tree(inode, path, handle, tc, 6798 clusters_to_del, &delete_blk); 6799 if (status) { 6800 mlog_errno(status); 6801 goto bail; 6802 } 6803 6804 if (le32_to_cpu(fe->i_clusters) == 0) { 6805 /* trunc to zero is a special case. */ 6806 el->l_tree_depth = 0; 6807 fe->i_last_eb_blk = 0; 6808 } else if (last_eb) 6809 fe->i_last_eb_blk = last_eb->h_blkno; 6810 6811 status = ocfs2_journal_dirty(handle, fe_bh); 6812 if (status < 0) { 6813 mlog_errno(status); 6814 goto bail; 6815 } 6816 6817 if (last_eb) { 6818 /* If there will be a new last extent block, then by 6819 * definition, there cannot be any leaves to the right of 6820 * him. */ 6821 last_eb->h_next_leaf_blk = 0; 6822 status = ocfs2_journal_dirty(handle, last_eb_bh); 6823 if (status < 0) { 6824 mlog_errno(status); 6825 goto bail; 6826 } 6827 } 6828 6829 if (delete_blk) { 6830 status = ocfs2_truncate_log_append(osb, handle, delete_blk, 6831 clusters_to_del); 6832 if (status < 0) { 6833 mlog_errno(status); 6834 goto bail; 6835 } 6836 } 6837 status = 0; 6838 bail: 6839 brelse(last_eb_bh); 6840 mlog_exit(status); 6841 return status; 6842 } 6843 6844 static int ocfs2_zero_func(handle_t *handle, struct buffer_head *bh) 6845 { 6846 set_buffer_uptodate(bh); 6847 mark_buffer_dirty(bh); 6848 return 0; 6849 } 6850 6851 static void ocfs2_map_and_dirty_page(struct inode *inode, handle_t *handle, 6852 unsigned int from, unsigned int to, 6853 struct page *page, int zero, u64 *phys) 6854 { 6855 int ret, partial = 0; 6856 6857 ret = ocfs2_map_page_blocks(page, phys, inode, from, to, 0); 6858 if (ret) 6859 mlog_errno(ret); 6860 6861 if (zero) 6862 zero_user_segment(page, from, to); 6863 6864 /* 6865 * Need to set the buffers we zero'd into uptodate 6866 * here if they aren't - ocfs2_map_page_blocks() 6867 * might've skipped some 6868 */ 6869 ret = walk_page_buffers(handle, page_buffers(page), 6870 from, to, &partial, 6871 ocfs2_zero_func); 6872 if (ret < 0) 6873 mlog_errno(ret); 6874 else if (ocfs2_should_order_data(inode)) { 6875 ret = ocfs2_jbd2_file_inode(handle, inode); 6876 if (ret < 0) 6877 mlog_errno(ret); 6878 } 6879 6880 if (!partial) 6881 SetPageUptodate(page); 6882 6883 flush_dcache_page(page); 6884 } 6885 6886 static void ocfs2_zero_cluster_pages(struct inode *inode, loff_t start, 6887 loff_t end, struct page **pages, 6888 int numpages, u64 phys, handle_t *handle) 6889 { 6890 int i; 6891 struct page *page; 6892 unsigned int from, to = PAGE_CACHE_SIZE; 6893 struct super_block *sb = inode->i_sb; 6894 6895 BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(sb))); 6896 6897 if (numpages == 0) 6898 goto out; 6899 6900 to = PAGE_CACHE_SIZE; 6901 for(i = 0; i < numpages; i++) { 6902 page = pages[i]; 6903 6904 from = start & (PAGE_CACHE_SIZE - 1); 6905 if ((end >> PAGE_CACHE_SHIFT) == page->index) 6906 to = end & (PAGE_CACHE_SIZE - 1); 6907 6908 BUG_ON(from > PAGE_CACHE_SIZE); 6909 BUG_ON(to > PAGE_CACHE_SIZE); 6910 6911 ocfs2_map_and_dirty_page(inode, handle, from, to, page, 1, 6912 &phys); 6913 6914 start = (page->index + 1) << PAGE_CACHE_SHIFT; 6915 } 6916 out: 6917 if (pages) 6918 ocfs2_unlock_and_free_pages(pages, numpages); 6919 } 6920 6921 static int ocfs2_grab_eof_pages(struct inode *inode, loff_t start, loff_t end, 6922 struct page **pages, int *num) 6923 { 6924 int numpages, ret = 0; 6925 struct super_block *sb = inode->i_sb; 6926 struct address_space *mapping = inode->i_mapping; 6927 unsigned long index; 6928 loff_t last_page_bytes; 6929 6930 BUG_ON(start > end); 6931 6932 BUG_ON(start >> OCFS2_SB(sb)->s_clustersize_bits != 6933 (end - 1) >> OCFS2_SB(sb)->s_clustersize_bits); 6934 6935 numpages = 0; 6936 last_page_bytes = PAGE_ALIGN(end); 6937 index = start >> PAGE_CACHE_SHIFT; 6938 do { 6939 pages[numpages] = grab_cache_page(mapping, index); 6940 if (!pages[numpages]) { 6941 ret = -ENOMEM; 6942 mlog_errno(ret); 6943 goto out; 6944 } 6945 6946 numpages++; 6947 index++; 6948 } while (index < (last_page_bytes >> PAGE_CACHE_SHIFT)); 6949 6950 out: 6951 if (ret != 0) { 6952 if (pages) 6953 ocfs2_unlock_and_free_pages(pages, numpages); 6954 numpages = 0; 6955 } 6956 6957 *num = numpages; 6958 6959 return ret; 6960 } 6961 6962 /* 6963 * Zero the area past i_size but still within an allocated 6964 * cluster. This avoids exposing nonzero data on subsequent file 6965 * extends. 6966 * 6967 * We need to call this before i_size is updated on the inode because 6968 * otherwise block_write_full_page() will skip writeout of pages past 6969 * i_size. The new_i_size parameter is passed for this reason. 6970 */ 6971 int ocfs2_zero_range_for_truncate(struct inode *inode, handle_t *handle, 6972 u64 range_start, u64 range_end) 6973 { 6974 int ret = 0, numpages; 6975 struct page **pages = NULL; 6976 u64 phys; 6977 unsigned int ext_flags; 6978 struct super_block *sb = inode->i_sb; 6979 6980 /* 6981 * File systems which don't support sparse files zero on every 6982 * extend. 6983 */ 6984 if (!ocfs2_sparse_alloc(OCFS2_SB(sb))) 6985 return 0; 6986 6987 pages = kcalloc(ocfs2_pages_per_cluster(sb), 6988 sizeof(struct page *), GFP_NOFS); 6989 if (pages == NULL) { 6990 ret = -ENOMEM; 6991 mlog_errno(ret); 6992 goto out; 6993 } 6994 6995 if (range_start == range_end) 6996 goto out; 6997 6998 ret = ocfs2_extent_map_get_blocks(inode, 6999 range_start >> sb->s_blocksize_bits, 7000 &phys, NULL, &ext_flags); 7001 if (ret) { 7002 mlog_errno(ret); 7003 goto out; 7004 } 7005 7006 /* 7007 * Tail is a hole, or is marked unwritten. In either case, we 7008 * can count on read and write to return/push zero's. 7009 */ 7010 if (phys == 0 || ext_flags & OCFS2_EXT_UNWRITTEN) 7011 goto out; 7012 7013 ret = ocfs2_grab_eof_pages(inode, range_start, range_end, pages, 7014 &numpages); 7015 if (ret) { 7016 mlog_errno(ret); 7017 goto out; 7018 } 7019 7020 ocfs2_zero_cluster_pages(inode, range_start, range_end, pages, 7021 numpages, phys, handle); 7022 7023 /* 7024 * Initiate writeout of the pages we zero'd here. We don't 7025 * wait on them - the truncate_inode_pages() call later will 7026 * do that for us. 7027 */ 7028 ret = do_sync_mapping_range(inode->i_mapping, range_start, 7029 range_end - 1, SYNC_FILE_RANGE_WRITE); 7030 if (ret) 7031 mlog_errno(ret); 7032 7033 out: 7034 if (pages) 7035 kfree(pages); 7036 7037 return ret; 7038 } 7039 7040 static void ocfs2_zero_dinode_id2_with_xattr(struct inode *inode, 7041 struct ocfs2_dinode *di) 7042 { 7043 unsigned int blocksize = 1 << inode->i_sb->s_blocksize_bits; 7044 unsigned int xattrsize = le16_to_cpu(di->i_xattr_inline_size); 7045 7046 if (le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_XATTR_FL) 7047 memset(&di->id2, 0, blocksize - 7048 offsetof(struct ocfs2_dinode, id2) - 7049 xattrsize); 7050 else 7051 memset(&di->id2, 0, blocksize - 7052 offsetof(struct ocfs2_dinode, id2)); 7053 } 7054 7055 void ocfs2_dinode_new_extent_list(struct inode *inode, 7056 struct ocfs2_dinode *di) 7057 { 7058 ocfs2_zero_dinode_id2_with_xattr(inode, di); 7059 di->id2.i_list.l_tree_depth = 0; 7060 di->id2.i_list.l_next_free_rec = 0; 7061 di->id2.i_list.l_count = cpu_to_le16( 7062 ocfs2_extent_recs_per_inode_with_xattr(inode->i_sb, di)); 7063 } 7064 7065 void ocfs2_set_inode_data_inline(struct inode *inode, struct ocfs2_dinode *di) 7066 { 7067 struct ocfs2_inode_info *oi = OCFS2_I(inode); 7068 struct ocfs2_inline_data *idata = &di->id2.i_data; 7069 7070 spin_lock(&oi->ip_lock); 7071 oi->ip_dyn_features |= OCFS2_INLINE_DATA_FL; 7072 di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features); 7073 spin_unlock(&oi->ip_lock); 7074 7075 /* 7076 * We clear the entire i_data structure here so that all 7077 * fields can be properly initialized. 7078 */ 7079 ocfs2_zero_dinode_id2_with_xattr(inode, di); 7080 7081 idata->id_count = cpu_to_le16( 7082 ocfs2_max_inline_data_with_xattr(inode->i_sb, di)); 7083 } 7084 7085 int ocfs2_convert_inline_data_to_extents(struct inode *inode, 7086 struct buffer_head *di_bh) 7087 { 7088 int ret, i, has_data, num_pages = 0; 7089 handle_t *handle; 7090 u64 uninitialized_var(block); 7091 struct ocfs2_inode_info *oi = OCFS2_I(inode); 7092 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 7093 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data; 7094 struct ocfs2_alloc_context *data_ac = NULL; 7095 struct page **pages = NULL; 7096 loff_t end = osb->s_clustersize; 7097 struct ocfs2_extent_tree et; 7098 int did_quota = 0; 7099 7100 has_data = i_size_read(inode) ? 1 : 0; 7101 7102 if (has_data) { 7103 pages = kcalloc(ocfs2_pages_per_cluster(osb->sb), 7104 sizeof(struct page *), GFP_NOFS); 7105 if (pages == NULL) { 7106 ret = -ENOMEM; 7107 mlog_errno(ret); 7108 goto out; 7109 } 7110 7111 ret = ocfs2_reserve_clusters(osb, 1, &data_ac); 7112 if (ret) { 7113 mlog_errno(ret); 7114 goto out; 7115 } 7116 } 7117 7118 handle = ocfs2_start_trans(osb, 7119 ocfs2_inline_to_extents_credits(osb->sb)); 7120 if (IS_ERR(handle)) { 7121 ret = PTR_ERR(handle); 7122 mlog_errno(ret); 7123 goto out_unlock; 7124 } 7125 7126 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh, 7127 OCFS2_JOURNAL_ACCESS_WRITE); 7128 if (ret) { 7129 mlog_errno(ret); 7130 goto out_commit; 7131 } 7132 7133 if (has_data) { 7134 u32 bit_off, num; 7135 unsigned int page_end; 7136 u64 phys; 7137 7138 if (vfs_dq_alloc_space_nodirty(inode, 7139 ocfs2_clusters_to_bytes(osb->sb, 1))) { 7140 ret = -EDQUOT; 7141 goto out_commit; 7142 } 7143 did_quota = 1; 7144 7145 ret = ocfs2_claim_clusters(osb, handle, data_ac, 1, &bit_off, 7146 &num); 7147 if (ret) { 7148 mlog_errno(ret); 7149 goto out_commit; 7150 } 7151 7152 /* 7153 * Save two copies, one for insert, and one that can 7154 * be changed by ocfs2_map_and_dirty_page() below. 7155 */ 7156 block = phys = ocfs2_clusters_to_blocks(inode->i_sb, bit_off); 7157 7158 /* 7159 * Non sparse file systems zero on extend, so no need 7160 * to do that now. 7161 */ 7162 if (!ocfs2_sparse_alloc(osb) && 7163 PAGE_CACHE_SIZE < osb->s_clustersize) 7164 end = PAGE_CACHE_SIZE; 7165 7166 ret = ocfs2_grab_eof_pages(inode, 0, end, pages, &num_pages); 7167 if (ret) { 7168 mlog_errno(ret); 7169 goto out_commit; 7170 } 7171 7172 /* 7173 * This should populate the 1st page for us and mark 7174 * it up to date. 7175 */ 7176 ret = ocfs2_read_inline_data(inode, pages[0], di_bh); 7177 if (ret) { 7178 mlog_errno(ret); 7179 goto out_commit; 7180 } 7181 7182 page_end = PAGE_CACHE_SIZE; 7183 if (PAGE_CACHE_SIZE > osb->s_clustersize) 7184 page_end = osb->s_clustersize; 7185 7186 for (i = 0; i < num_pages; i++) 7187 ocfs2_map_and_dirty_page(inode, handle, 0, page_end, 7188 pages[i], i > 0, &phys); 7189 } 7190 7191 spin_lock(&oi->ip_lock); 7192 oi->ip_dyn_features &= ~OCFS2_INLINE_DATA_FL; 7193 di->i_dyn_features = cpu_to_le16(oi->ip_dyn_features); 7194 spin_unlock(&oi->ip_lock); 7195 7196 ocfs2_dinode_new_extent_list(inode, di); 7197 7198 ocfs2_journal_dirty(handle, di_bh); 7199 7200 if (has_data) { 7201 /* 7202 * An error at this point should be extremely rare. If 7203 * this proves to be false, we could always re-build 7204 * the in-inode data from our pages. 7205 */ 7206 ocfs2_init_dinode_extent_tree(&et, inode, di_bh); 7207 ret = ocfs2_insert_extent(osb, handle, inode, &et, 7208 0, block, 1, 0, NULL); 7209 if (ret) { 7210 mlog_errno(ret); 7211 goto out_commit; 7212 } 7213 7214 inode->i_blocks = ocfs2_inode_sector_count(inode); 7215 } 7216 7217 out_commit: 7218 if (ret < 0 && did_quota) 7219 vfs_dq_free_space_nodirty(inode, 7220 ocfs2_clusters_to_bytes(osb->sb, 1)); 7221 7222 ocfs2_commit_trans(osb, handle); 7223 7224 out_unlock: 7225 if (data_ac) 7226 ocfs2_free_alloc_context(data_ac); 7227 7228 out: 7229 if (pages) { 7230 ocfs2_unlock_and_free_pages(pages, num_pages); 7231 kfree(pages); 7232 } 7233 7234 return ret; 7235 } 7236 7237 /* 7238 * It is expected, that by the time you call this function, 7239 * inode->i_size and fe->i_size have been adjusted. 7240 * 7241 * WARNING: This will kfree the truncate context 7242 */ 7243 int ocfs2_commit_truncate(struct ocfs2_super *osb, 7244 struct inode *inode, 7245 struct buffer_head *fe_bh, 7246 struct ocfs2_truncate_context *tc) 7247 { 7248 int status, i, credits, tl_sem = 0; 7249 u32 clusters_to_del, new_highest_cpos, range; 7250 struct ocfs2_extent_list *el; 7251 handle_t *handle = NULL; 7252 struct inode *tl_inode = osb->osb_tl_inode; 7253 struct ocfs2_path *path = NULL; 7254 struct ocfs2_dinode *di = (struct ocfs2_dinode *)fe_bh->b_data; 7255 7256 mlog_entry_void(); 7257 7258 new_highest_cpos = ocfs2_clusters_for_bytes(osb->sb, 7259 i_size_read(inode)); 7260 7261 path = ocfs2_new_path(fe_bh, &di->id2.i_list, 7262 ocfs2_journal_access_di); 7263 if (!path) { 7264 status = -ENOMEM; 7265 mlog_errno(status); 7266 goto bail; 7267 } 7268 7269 ocfs2_extent_map_trunc(inode, new_highest_cpos); 7270 7271 start: 7272 /* 7273 * Check that we still have allocation to delete. 7274 */ 7275 if (OCFS2_I(inode)->ip_clusters == 0) { 7276 status = 0; 7277 goto bail; 7278 } 7279 7280 /* 7281 * Truncate always works against the rightmost tree branch. 7282 */ 7283 status = ocfs2_find_path(INODE_CACHE(inode), path, UINT_MAX); 7284 if (status) { 7285 mlog_errno(status); 7286 goto bail; 7287 } 7288 7289 mlog(0, "inode->ip_clusters = %u, tree_depth = %u\n", 7290 OCFS2_I(inode)->ip_clusters, path->p_tree_depth); 7291 7292 /* 7293 * By now, el will point to the extent list on the bottom most 7294 * portion of this tree. Only the tail record is considered in 7295 * each pass. 7296 * 7297 * We handle the following cases, in order: 7298 * - empty extent: delete the remaining branch 7299 * - remove the entire record 7300 * - remove a partial record 7301 * - no record needs to be removed (truncate has completed) 7302 */ 7303 el = path_leaf_el(path); 7304 if (le16_to_cpu(el->l_next_free_rec) == 0) { 7305 ocfs2_error(inode->i_sb, 7306 "Inode %llu has empty extent block at %llu\n", 7307 (unsigned long long)OCFS2_I(inode)->ip_blkno, 7308 (unsigned long long)path_leaf_bh(path)->b_blocknr); 7309 status = -EROFS; 7310 goto bail; 7311 } 7312 7313 i = le16_to_cpu(el->l_next_free_rec) - 1; 7314 range = le32_to_cpu(el->l_recs[i].e_cpos) + 7315 ocfs2_rec_clusters(el, &el->l_recs[i]); 7316 if (i == 0 && ocfs2_is_empty_extent(&el->l_recs[i])) { 7317 clusters_to_del = 0; 7318 } else if (le32_to_cpu(el->l_recs[i].e_cpos) >= new_highest_cpos) { 7319 clusters_to_del = ocfs2_rec_clusters(el, &el->l_recs[i]); 7320 } else if (range > new_highest_cpos) { 7321 clusters_to_del = (ocfs2_rec_clusters(el, &el->l_recs[i]) + 7322 le32_to_cpu(el->l_recs[i].e_cpos)) - 7323 new_highest_cpos; 7324 } else { 7325 status = 0; 7326 goto bail; 7327 } 7328 7329 mlog(0, "clusters_to_del = %u in this pass, tail blk=%llu\n", 7330 clusters_to_del, (unsigned long long)path_leaf_bh(path)->b_blocknr); 7331 7332 mutex_lock(&tl_inode->i_mutex); 7333 tl_sem = 1; 7334 /* ocfs2_truncate_log_needs_flush guarantees us at least one 7335 * record is free for use. If there isn't any, we flush to get 7336 * an empty truncate log. */ 7337 if (ocfs2_truncate_log_needs_flush(osb)) { 7338 status = __ocfs2_flush_truncate_log(osb); 7339 if (status < 0) { 7340 mlog_errno(status); 7341 goto bail; 7342 } 7343 } 7344 7345 credits = ocfs2_calc_tree_trunc_credits(osb->sb, clusters_to_del, 7346 (struct ocfs2_dinode *)fe_bh->b_data, 7347 el); 7348 handle = ocfs2_start_trans(osb, credits); 7349 if (IS_ERR(handle)) { 7350 status = PTR_ERR(handle); 7351 handle = NULL; 7352 mlog_errno(status); 7353 goto bail; 7354 } 7355 7356 status = ocfs2_do_truncate(osb, clusters_to_del, inode, fe_bh, handle, 7357 tc, path); 7358 if (status < 0) { 7359 mlog_errno(status); 7360 goto bail; 7361 } 7362 7363 mutex_unlock(&tl_inode->i_mutex); 7364 tl_sem = 0; 7365 7366 ocfs2_commit_trans(osb, handle); 7367 handle = NULL; 7368 7369 ocfs2_reinit_path(path, 1); 7370 7371 /* 7372 * The check above will catch the case where we've truncated 7373 * away all allocation. 7374 */ 7375 goto start; 7376 7377 bail: 7378 7379 ocfs2_schedule_truncate_log_flush(osb, 1); 7380 7381 if (tl_sem) 7382 mutex_unlock(&tl_inode->i_mutex); 7383 7384 if (handle) 7385 ocfs2_commit_trans(osb, handle); 7386 7387 ocfs2_run_deallocs(osb, &tc->tc_dealloc); 7388 7389 ocfs2_free_path(path); 7390 7391 /* This will drop the ext_alloc cluster lock for us */ 7392 ocfs2_free_truncate_context(tc); 7393 7394 mlog_exit(status); 7395 return status; 7396 } 7397 7398 /* 7399 * Expects the inode to already be locked. 7400 */ 7401 int ocfs2_prepare_truncate(struct ocfs2_super *osb, 7402 struct inode *inode, 7403 struct buffer_head *fe_bh, 7404 struct ocfs2_truncate_context **tc) 7405 { 7406 int status; 7407 unsigned int new_i_clusters; 7408 struct ocfs2_dinode *fe; 7409 struct ocfs2_extent_block *eb; 7410 struct buffer_head *last_eb_bh = NULL; 7411 7412 mlog_entry_void(); 7413 7414 *tc = NULL; 7415 7416 new_i_clusters = ocfs2_clusters_for_bytes(osb->sb, 7417 i_size_read(inode)); 7418 fe = (struct ocfs2_dinode *) fe_bh->b_data; 7419 7420 mlog(0, "fe->i_clusters = %u, new_i_clusters = %u, fe->i_size =" 7421 "%llu\n", le32_to_cpu(fe->i_clusters), new_i_clusters, 7422 (unsigned long long)le64_to_cpu(fe->i_size)); 7423 7424 *tc = kzalloc(sizeof(struct ocfs2_truncate_context), GFP_KERNEL); 7425 if (!(*tc)) { 7426 status = -ENOMEM; 7427 mlog_errno(status); 7428 goto bail; 7429 } 7430 ocfs2_init_dealloc_ctxt(&(*tc)->tc_dealloc); 7431 7432 if (fe->id2.i_list.l_tree_depth) { 7433 status = ocfs2_read_extent_block(INODE_CACHE(inode), 7434 le64_to_cpu(fe->i_last_eb_blk), 7435 &last_eb_bh); 7436 if (status < 0) { 7437 mlog_errno(status); 7438 goto bail; 7439 } 7440 eb = (struct ocfs2_extent_block *) last_eb_bh->b_data; 7441 } 7442 7443 (*tc)->tc_last_eb_bh = last_eb_bh; 7444 7445 status = 0; 7446 bail: 7447 if (status < 0) { 7448 if (*tc) 7449 ocfs2_free_truncate_context(*tc); 7450 *tc = NULL; 7451 } 7452 mlog_exit_void(); 7453 return status; 7454 } 7455 7456 /* 7457 * 'start' is inclusive, 'end' is not. 7458 */ 7459 int ocfs2_truncate_inline(struct inode *inode, struct buffer_head *di_bh, 7460 unsigned int start, unsigned int end, int trunc) 7461 { 7462 int ret; 7463 unsigned int numbytes; 7464 handle_t *handle; 7465 struct ocfs2_super *osb = OCFS2_SB(inode->i_sb); 7466 struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data; 7467 struct ocfs2_inline_data *idata = &di->id2.i_data; 7468 7469 if (end > i_size_read(inode)) 7470 end = i_size_read(inode); 7471 7472 BUG_ON(start >= end); 7473 7474 if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) || 7475 !(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL) || 7476 !ocfs2_supports_inline_data(osb)) { 7477 ocfs2_error(inode->i_sb, 7478 "Inline data flags for inode %llu don't agree! " 7479 "Disk: 0x%x, Memory: 0x%x, Superblock: 0x%x\n", 7480 (unsigned long long)OCFS2_I(inode)->ip_blkno, 7481 le16_to_cpu(di->i_dyn_features), 7482 OCFS2_I(inode)->ip_dyn_features, 7483 osb->s_feature_incompat); 7484 ret = -EROFS; 7485 goto out; 7486 } 7487 7488 handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS); 7489 if (IS_ERR(handle)) { 7490 ret = PTR_ERR(handle); 7491 mlog_errno(ret); 7492 goto out; 7493 } 7494 7495 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh, 7496 OCFS2_JOURNAL_ACCESS_WRITE); 7497 if (ret) { 7498 mlog_errno(ret); 7499 goto out_commit; 7500 } 7501 7502 numbytes = end - start; 7503 memset(idata->id_data + start, 0, numbytes); 7504 7505 /* 7506 * No need to worry about the data page here - it's been 7507 * truncated already and inline data doesn't need it for 7508 * pushing zero's to disk, so we'll let readpage pick it up 7509 * later. 7510 */ 7511 if (trunc) { 7512 i_size_write(inode, start); 7513 di->i_size = cpu_to_le64(start); 7514 } 7515 7516 inode->i_blocks = ocfs2_inode_sector_count(inode); 7517 inode->i_ctime = inode->i_mtime = CURRENT_TIME; 7518 7519 di->i_ctime = di->i_mtime = cpu_to_le64(inode->i_ctime.tv_sec); 7520 di->i_ctime_nsec = di->i_mtime_nsec = cpu_to_le32(inode->i_ctime.tv_nsec); 7521 7522 ocfs2_journal_dirty(handle, di_bh); 7523 7524 out_commit: 7525 ocfs2_commit_trans(osb, handle); 7526 7527 out: 7528 return ret; 7529 } 7530 7531 static void ocfs2_free_truncate_context(struct ocfs2_truncate_context *tc) 7532 { 7533 /* 7534 * The caller is responsible for completing deallocation 7535 * before freeing the context. 7536 */ 7537 if (tc->tc_dealloc.c_first_suballocator != NULL) 7538 mlog(ML_NOTICE, 7539 "Truncate completion has non-empty dealloc context\n"); 7540 7541 brelse(tc->tc_last_eb_bh); 7542 7543 kfree(tc); 7544 } 7545