1 /* -*- mode: c; c-basic-offset: 8; -*- 2 * vim: noexpandtab sw=8 ts=8 sts=0: 3 * 4 * dcache.c 5 * 6 * dentry cache handling code 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/namei.h> 30 31 #define MLOG_MASK_PREFIX ML_DCACHE 32 #include <cluster/masklog.h> 33 34 #include "ocfs2.h" 35 36 #include "alloc.h" 37 #include "dcache.h" 38 #include "dlmglue.h" 39 #include "file.h" 40 #include "inode.h" 41 #include "super.h" 42 43 void ocfs2_dentry_attach_gen(struct dentry *dentry) 44 { 45 unsigned long gen = 46 OCFS2_I(dentry->d_parent->d_inode)->ip_dir_lock_gen; 47 BUG_ON(dentry->d_inode); 48 dentry->d_fsdata = (void *)gen; 49 } 50 51 52 static int ocfs2_dentry_revalidate(struct dentry *dentry, 53 struct nameidata *nd) 54 { 55 struct inode *inode = dentry->d_inode; 56 int ret = 0; /* if all else fails, just return false */ 57 struct ocfs2_super *osb = OCFS2_SB(dentry->d_sb); 58 59 mlog_entry("(0x%p, '%.*s')\n", dentry, 60 dentry->d_name.len, dentry->d_name.name); 61 62 /* For a negative dentry - 63 * check the generation number of the parent and compare with the 64 * one stored in the inode. 65 */ 66 if (inode == NULL) { 67 unsigned long gen = (unsigned long) dentry->d_fsdata; 68 unsigned long pgen = 69 OCFS2_I(dentry->d_parent->d_inode)->ip_dir_lock_gen; 70 mlog(0, "negative dentry: %.*s parent gen: %lu " 71 "dentry gen: %lu\n", 72 dentry->d_name.len, dentry->d_name.name, pgen, gen); 73 if (gen != pgen) 74 goto bail; 75 goto valid; 76 } 77 78 BUG_ON(!osb); 79 80 if (inode == osb->root_inode || is_bad_inode(inode)) 81 goto bail; 82 83 spin_lock(&OCFS2_I(inode)->ip_lock); 84 /* did we or someone else delete this inode? */ 85 if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_DELETED) { 86 spin_unlock(&OCFS2_I(inode)->ip_lock); 87 mlog(0, "inode (%llu) deleted, returning false\n", 88 (unsigned long long)OCFS2_I(inode)->ip_blkno); 89 goto bail; 90 } 91 spin_unlock(&OCFS2_I(inode)->ip_lock); 92 93 /* 94 * We don't need a cluster lock to test this because once an 95 * inode nlink hits zero, it never goes back. 96 */ 97 if (inode->i_nlink == 0) { 98 mlog(0, "Inode %llu orphaned, returning false " 99 "dir = %d\n", 100 (unsigned long long)OCFS2_I(inode)->ip_blkno, 101 S_ISDIR(inode->i_mode)); 102 goto bail; 103 } 104 105 /* 106 * If the last lookup failed to create dentry lock, let us 107 * redo it. 108 */ 109 if (!dentry->d_fsdata) { 110 mlog(0, "Inode %llu doesn't have dentry lock, " 111 "returning false\n", 112 (unsigned long long)OCFS2_I(inode)->ip_blkno); 113 goto bail; 114 } 115 116 valid: 117 ret = 1; 118 119 bail: 120 mlog_exit(ret); 121 122 return ret; 123 } 124 125 static int ocfs2_match_dentry(struct dentry *dentry, 126 u64 parent_blkno, 127 int skip_unhashed) 128 { 129 struct inode *parent; 130 131 /* 132 * ocfs2_lookup() does a d_splice_alias() _before_ attaching 133 * to the lock data, so we skip those here, otherwise 134 * ocfs2_dentry_attach_lock() will get its original dentry 135 * back. 136 */ 137 if (!dentry->d_fsdata) 138 return 0; 139 140 if (!dentry->d_parent) 141 return 0; 142 143 if (skip_unhashed && d_unhashed(dentry)) 144 return 0; 145 146 parent = dentry->d_parent->d_inode; 147 /* Negative parent dentry? */ 148 if (!parent) 149 return 0; 150 151 /* Name is in a different directory. */ 152 if (OCFS2_I(parent)->ip_blkno != parent_blkno) 153 return 0; 154 155 return 1; 156 } 157 158 /* 159 * Walk the inode alias list, and find a dentry which has a given 160 * parent. ocfs2_dentry_attach_lock() wants to find _any_ alias as it 161 * is looking for a dentry_lock reference. The downconvert thread is 162 * looking to unhash aliases, so we allow it to skip any that already 163 * have that property. 164 */ 165 struct dentry *ocfs2_find_local_alias(struct inode *inode, 166 u64 parent_blkno, 167 int skip_unhashed) 168 { 169 struct list_head *p; 170 struct dentry *dentry = NULL; 171 172 spin_lock(&dcache_lock); 173 174 list_for_each(p, &inode->i_dentry) { 175 dentry = list_entry(p, struct dentry, d_alias); 176 177 spin_lock(&dentry->d_lock); 178 if (ocfs2_match_dentry(dentry, parent_blkno, skip_unhashed)) { 179 mlog(0, "dentry found: %.*s\n", 180 dentry->d_name.len, dentry->d_name.name); 181 182 dget_locked_dlock(dentry); 183 spin_unlock(&dentry->d_lock); 184 break; 185 } 186 spin_unlock(&dentry->d_lock); 187 188 dentry = NULL; 189 } 190 191 spin_unlock(&dcache_lock); 192 193 return dentry; 194 } 195 196 DEFINE_SPINLOCK(dentry_attach_lock); 197 198 /* 199 * Attach this dentry to a cluster lock. 200 * 201 * Dentry locks cover all links in a given directory to a particular 202 * inode. We do this so that ocfs2 can build a lock name which all 203 * nodes in the cluster can agree on at all times. Shoving full names 204 * in the cluster lock won't work due to size restrictions. Covering 205 * links inside of a directory is a good compromise because it still 206 * allows us to use the parent directory lock to synchronize 207 * operations. 208 * 209 * Call this function with the parent dir semaphore and the parent dir 210 * cluster lock held. 211 * 212 * The dir semaphore will protect us from having to worry about 213 * concurrent processes on our node trying to attach a lock at the 214 * same time. 215 * 216 * The dir cluster lock (held at either PR or EX mode) protects us 217 * from unlink and rename on other nodes. 218 * 219 * A dput() can happen asynchronously due to pruning, so we cover 220 * attaching and detaching the dentry lock with a 221 * dentry_attach_lock. 222 * 223 * A node which has done lookup on a name retains a protected read 224 * lock until final dput. If the user requests and unlink or rename, 225 * the protected read is upgraded to an exclusive lock. Other nodes 226 * who have seen the dentry will then be informed that they need to 227 * downgrade their lock, which will involve d_delete on the 228 * dentry. This happens in ocfs2_dentry_convert_worker(). 229 */ 230 int ocfs2_dentry_attach_lock(struct dentry *dentry, 231 struct inode *inode, 232 u64 parent_blkno) 233 { 234 int ret; 235 struct dentry *alias; 236 struct ocfs2_dentry_lock *dl = dentry->d_fsdata; 237 238 mlog(0, "Attach \"%.*s\", parent %llu, fsdata: %p\n", 239 dentry->d_name.len, dentry->d_name.name, 240 (unsigned long long)parent_blkno, dl); 241 242 /* 243 * Negative dentry. We ignore these for now. 244 * 245 * XXX: Could we can improve ocfs2_dentry_revalidate() by 246 * tracking these? 247 */ 248 if (!inode) 249 return 0; 250 251 if (!dentry->d_inode && dentry->d_fsdata) { 252 /* Converting a negative dentry to positive 253 Clear dentry->d_fsdata */ 254 dentry->d_fsdata = dl = NULL; 255 } 256 257 if (dl) { 258 mlog_bug_on_msg(dl->dl_parent_blkno != parent_blkno, 259 " \"%.*s\": old parent: %llu, new: %llu\n", 260 dentry->d_name.len, dentry->d_name.name, 261 (unsigned long long)parent_blkno, 262 (unsigned long long)dl->dl_parent_blkno); 263 return 0; 264 } 265 266 alias = ocfs2_find_local_alias(inode, parent_blkno, 0); 267 if (alias) { 268 /* 269 * Great, an alias exists, which means we must have a 270 * dentry lock already. We can just grab the lock off 271 * the alias and add it to the list. 272 * 273 * We're depending here on the fact that this dentry 274 * was found and exists in the dcache and so must have 275 * a reference to the dentry_lock because we can't 276 * race creates. Final dput() cannot happen on it 277 * since we have it pinned, so our reference is safe. 278 */ 279 dl = alias->d_fsdata; 280 mlog_bug_on_msg(!dl, "parent %llu, ino %llu\n", 281 (unsigned long long)parent_blkno, 282 (unsigned long long)OCFS2_I(inode)->ip_blkno); 283 284 mlog_bug_on_msg(dl->dl_parent_blkno != parent_blkno, 285 " \"%.*s\": old parent: %llu, new: %llu\n", 286 dentry->d_name.len, dentry->d_name.name, 287 (unsigned long long)parent_blkno, 288 (unsigned long long)dl->dl_parent_blkno); 289 290 mlog(0, "Found: %s\n", dl->dl_lockres.l_name); 291 292 goto out_attach; 293 } 294 295 /* 296 * There are no other aliases 297 */ 298 dl = kmalloc(sizeof(*dl), GFP_NOFS); 299 if (!dl) { 300 ret = -ENOMEM; 301 mlog_errno(ret); 302 return ret; 303 } 304 305 dl->dl_count = 0; 306 /* 307 * Does this have to happen below, for all attaches, in case 308 * the struct inode gets blown away by the downconvert thread? 309 */ 310 dl->dl_inode = igrab(inode); 311 dl->dl_parent_blkno = parent_blkno; 312 ocfs2_dentry_lock_res_init(dl, parent_blkno, inode); 313 314 out_attach: 315 spin_lock(&dentry_attach_lock); 316 dentry->d_fsdata = dl; 317 dl->dl_count++; 318 spin_unlock(&dentry_attach_lock); 319 320 /* 321 * This actually gets us our PRMODE level lock. From now on, 322 * we'll have a notification if one of these names is 323 * destroyed on another node. 324 */ 325 ret = ocfs2_dentry_lock(dentry, 0); 326 if (!ret) 327 ocfs2_dentry_unlock(dentry, 0); 328 else 329 mlog_errno(ret); 330 331 /* 332 * In case of error, manually free the allocation and do the iput(). 333 * We need to do this because error here means no d_instantiate(), 334 * which means iput() will not be called during dput(dentry). 335 */ 336 if (ret < 0 && !alias) { 337 ocfs2_lock_res_free(&dl->dl_lockres); 338 BUG_ON(dl->dl_count != 1); 339 spin_lock(&dentry_attach_lock); 340 dentry->d_fsdata = NULL; 341 spin_unlock(&dentry_attach_lock); 342 kfree(dl); 343 iput(inode); 344 } 345 346 dput(alias); 347 348 return ret; 349 } 350 351 DEFINE_SPINLOCK(dentry_list_lock); 352 353 /* We limit the number of dentry locks to drop in one go. We have 354 * this limit so that we don't starve other users of ocfs2_wq. */ 355 #define DL_INODE_DROP_COUNT 64 356 357 /* Drop inode references from dentry locks */ 358 static void __ocfs2_drop_dl_inodes(struct ocfs2_super *osb, int drop_count) 359 { 360 struct ocfs2_dentry_lock *dl; 361 362 spin_lock(&dentry_list_lock); 363 while (osb->dentry_lock_list && (drop_count < 0 || drop_count--)) { 364 dl = osb->dentry_lock_list; 365 osb->dentry_lock_list = dl->dl_next; 366 spin_unlock(&dentry_list_lock); 367 iput(dl->dl_inode); 368 kfree(dl); 369 spin_lock(&dentry_list_lock); 370 } 371 spin_unlock(&dentry_list_lock); 372 } 373 374 void ocfs2_drop_dl_inodes(struct work_struct *work) 375 { 376 struct ocfs2_super *osb = container_of(work, struct ocfs2_super, 377 dentry_lock_work); 378 379 __ocfs2_drop_dl_inodes(osb, DL_INODE_DROP_COUNT); 380 /* 381 * Don't queue dropping if umount is in progress. We flush the 382 * list in ocfs2_dismount_volume 383 */ 384 spin_lock(&dentry_list_lock); 385 if (osb->dentry_lock_list && 386 !ocfs2_test_osb_flag(osb, OCFS2_OSB_DROP_DENTRY_LOCK_IMMED)) 387 queue_work(ocfs2_wq, &osb->dentry_lock_work); 388 spin_unlock(&dentry_list_lock); 389 } 390 391 /* Flush the whole work queue */ 392 void ocfs2_drop_all_dl_inodes(struct ocfs2_super *osb) 393 { 394 __ocfs2_drop_dl_inodes(osb, -1); 395 } 396 397 /* 398 * ocfs2_dentry_iput() and friends. 399 * 400 * At this point, our particular dentry is detached from the inodes 401 * alias list, so there's no way that the locking code can find it. 402 * 403 * The interesting stuff happens when we determine that our lock needs 404 * to go away because this is the last subdir alias in the 405 * system. This function needs to handle a couple things: 406 * 407 * 1) Synchronizing lock shutdown with the downconvert threads. This 408 * is already handled for us via the lockres release drop function 409 * called in ocfs2_release_dentry_lock() 410 * 411 * 2) A race may occur when we're doing our lock shutdown and 412 * another process wants to create a new dentry lock. Right now we 413 * let them race, which means that for a very short while, this 414 * node might have two locks on a lock resource. This should be a 415 * problem though because one of them is in the process of being 416 * thrown out. 417 */ 418 static void ocfs2_drop_dentry_lock(struct ocfs2_super *osb, 419 struct ocfs2_dentry_lock *dl) 420 { 421 ocfs2_simple_drop_lockres(osb, &dl->dl_lockres); 422 ocfs2_lock_res_free(&dl->dl_lockres); 423 424 /* We leave dropping of inode reference to ocfs2_wq as that can 425 * possibly lead to inode deletion which gets tricky */ 426 spin_lock(&dentry_list_lock); 427 if (!osb->dentry_lock_list && 428 !ocfs2_test_osb_flag(osb, OCFS2_OSB_DROP_DENTRY_LOCK_IMMED)) 429 queue_work(ocfs2_wq, &osb->dentry_lock_work); 430 dl->dl_next = osb->dentry_lock_list; 431 osb->dentry_lock_list = dl; 432 spin_unlock(&dentry_list_lock); 433 } 434 435 void ocfs2_dentry_lock_put(struct ocfs2_super *osb, 436 struct ocfs2_dentry_lock *dl) 437 { 438 int unlock; 439 440 BUG_ON(dl->dl_count == 0); 441 442 spin_lock(&dentry_attach_lock); 443 dl->dl_count--; 444 unlock = !dl->dl_count; 445 spin_unlock(&dentry_attach_lock); 446 447 if (unlock) 448 ocfs2_drop_dentry_lock(osb, dl); 449 } 450 451 static void ocfs2_dentry_iput(struct dentry *dentry, struct inode *inode) 452 { 453 struct ocfs2_dentry_lock *dl = dentry->d_fsdata; 454 455 if (!dl) { 456 /* 457 * No dentry lock is ok if we're disconnected or 458 * unhashed. 459 */ 460 if (!(dentry->d_flags & DCACHE_DISCONNECTED) && 461 !d_unhashed(dentry)) { 462 unsigned long long ino = 0ULL; 463 if (inode) 464 ino = (unsigned long long)OCFS2_I(inode)->ip_blkno; 465 mlog(ML_ERROR, "Dentry is missing cluster lock. " 466 "inode: %llu, d_flags: 0x%x, d_name: %.*s\n", 467 ino, dentry->d_flags, dentry->d_name.len, 468 dentry->d_name.name); 469 } 470 471 goto out; 472 } 473 474 mlog_bug_on_msg(dl->dl_count == 0, "dentry: %.*s, count: %u\n", 475 dentry->d_name.len, dentry->d_name.name, 476 dl->dl_count); 477 478 ocfs2_dentry_lock_put(OCFS2_SB(dentry->d_sb), dl); 479 480 out: 481 iput(inode); 482 } 483 484 /* 485 * d_move(), but keep the locks in sync. 486 * 487 * When we are done, "dentry" will have the parent dir and name of 488 * "target", which will be thrown away. 489 * 490 * We manually update the lock of "dentry" if need be. 491 * 492 * "target" doesn't have it's dentry lock touched - we allow the later 493 * dput() to handle this for us. 494 * 495 * This is called during ocfs2_rename(), while holding parent 496 * directory locks. The dentries have already been deleted on other 497 * nodes via ocfs2_remote_dentry_delete(). 498 * 499 * Normally, the VFS handles the d_move() for the file system, after 500 * the ->rename() callback. OCFS2 wants to handle this internally, so 501 * the new lock can be created atomically with respect to the cluster. 502 */ 503 void ocfs2_dentry_move(struct dentry *dentry, struct dentry *target, 504 struct inode *old_dir, struct inode *new_dir) 505 { 506 int ret; 507 struct ocfs2_super *osb = OCFS2_SB(old_dir->i_sb); 508 struct inode *inode = dentry->d_inode; 509 510 /* 511 * Move within the same directory, so the actual lock info won't 512 * change. 513 * 514 * XXX: Is there any advantage to dropping the lock here? 515 */ 516 if (old_dir == new_dir) 517 goto out_move; 518 519 ocfs2_dentry_lock_put(osb, dentry->d_fsdata); 520 521 dentry->d_fsdata = NULL; 522 ret = ocfs2_dentry_attach_lock(dentry, inode, OCFS2_I(new_dir)->ip_blkno); 523 if (ret) 524 mlog_errno(ret); 525 526 out_move: 527 d_move(dentry, target); 528 } 529 530 const struct dentry_operations ocfs2_dentry_ops = { 531 .d_revalidate = ocfs2_dentry_revalidate, 532 .d_iput = ocfs2_dentry_iput, 533 }; 534