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 #include <cluster/masklog.h> 32 33 #include "ocfs2.h" 34 35 #include "alloc.h" 36 #include "dcache.h" 37 #include "dlmglue.h" 38 #include "file.h" 39 #include "inode.h" 40 #include "super.h" 41 #include "ocfs2_trace.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; 56 int ret = 0; /* if all else fails, just return false */ 57 struct ocfs2_super *osb; 58 59 if (nd && nd->flags & LOOKUP_RCU) 60 return -ECHILD; 61 62 inode = dentry->d_inode; 63 osb = OCFS2_SB(dentry->d_sb); 64 65 trace_ocfs2_dentry_revalidate(dentry, dentry->d_name.len, 66 dentry->d_name.name); 67 68 /* For a negative dentry - 69 * check the generation number of the parent and compare with the 70 * one stored in the inode. 71 */ 72 if (inode == NULL) { 73 unsigned long gen = (unsigned long) dentry->d_fsdata; 74 unsigned long pgen = 75 OCFS2_I(dentry->d_parent->d_inode)->ip_dir_lock_gen; 76 77 trace_ocfs2_dentry_revalidate_negative(dentry->d_name.len, 78 dentry->d_name.name, 79 pgen, gen); 80 if (gen != pgen) 81 goto bail; 82 goto valid; 83 } 84 85 BUG_ON(!osb); 86 87 if (inode == osb->root_inode || is_bad_inode(inode)) 88 goto bail; 89 90 spin_lock(&OCFS2_I(inode)->ip_lock); 91 /* did we or someone else delete this inode? */ 92 if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_DELETED) { 93 spin_unlock(&OCFS2_I(inode)->ip_lock); 94 trace_ocfs2_dentry_revalidate_delete( 95 (unsigned long long)OCFS2_I(inode)->ip_blkno); 96 goto bail; 97 } 98 spin_unlock(&OCFS2_I(inode)->ip_lock); 99 100 /* 101 * We don't need a cluster lock to test this because once an 102 * inode nlink hits zero, it never goes back. 103 */ 104 if (inode->i_nlink == 0) { 105 trace_ocfs2_dentry_revalidate_orphaned( 106 (unsigned long long)OCFS2_I(inode)->ip_blkno, 107 S_ISDIR(inode->i_mode)); 108 goto bail; 109 } 110 111 /* 112 * If the last lookup failed to create dentry lock, let us 113 * redo it. 114 */ 115 if (!dentry->d_fsdata) { 116 trace_ocfs2_dentry_revalidate_nofsdata( 117 (unsigned long long)OCFS2_I(inode)->ip_blkno); 118 goto bail; 119 } 120 121 valid: 122 ret = 1; 123 124 bail: 125 trace_ocfs2_dentry_revalidate_ret(ret); 126 return ret; 127 } 128 129 static int ocfs2_match_dentry(struct dentry *dentry, 130 u64 parent_blkno, 131 int skip_unhashed) 132 { 133 struct inode *parent; 134 135 /* 136 * ocfs2_lookup() does a d_splice_alias() _before_ attaching 137 * to the lock data, so we skip those here, otherwise 138 * ocfs2_dentry_attach_lock() will get its original dentry 139 * back. 140 */ 141 if (!dentry->d_fsdata) 142 return 0; 143 144 if (!dentry->d_parent) 145 return 0; 146 147 if (skip_unhashed && d_unhashed(dentry)) 148 return 0; 149 150 parent = dentry->d_parent->d_inode; 151 /* Negative parent dentry? */ 152 if (!parent) 153 return 0; 154 155 /* Name is in a different directory. */ 156 if (OCFS2_I(parent)->ip_blkno != parent_blkno) 157 return 0; 158 159 return 1; 160 } 161 162 /* 163 * Walk the inode alias list, and find a dentry which has a given 164 * parent. ocfs2_dentry_attach_lock() wants to find _any_ alias as it 165 * is looking for a dentry_lock reference. The downconvert thread is 166 * looking to unhash aliases, so we allow it to skip any that already 167 * have that property. 168 */ 169 struct dentry *ocfs2_find_local_alias(struct inode *inode, 170 u64 parent_blkno, 171 int skip_unhashed) 172 { 173 struct hlist_node *p; 174 struct dentry *dentry; 175 176 spin_lock(&inode->i_lock); 177 hlist_for_each_entry(dentry, p, &inode->i_dentry, d_alias) { 178 spin_lock(&dentry->d_lock); 179 if (ocfs2_match_dentry(dentry, parent_blkno, skip_unhashed)) { 180 trace_ocfs2_find_local_alias(dentry->d_name.len, 181 dentry->d_name.name); 182 183 dget_dlock(dentry); 184 spin_unlock(&dentry->d_lock); 185 spin_unlock(&inode->i_lock); 186 return dentry; 187 } 188 spin_unlock(&dentry->d_lock); 189 } 190 spin_unlock(&inode->i_lock); 191 return NULL; 192 } 193 194 DEFINE_SPINLOCK(dentry_attach_lock); 195 196 /* 197 * Attach this dentry to a cluster lock. 198 * 199 * Dentry locks cover all links in a given directory to a particular 200 * inode. We do this so that ocfs2 can build a lock name which all 201 * nodes in the cluster can agree on at all times. Shoving full names 202 * in the cluster lock won't work due to size restrictions. Covering 203 * links inside of a directory is a good compromise because it still 204 * allows us to use the parent directory lock to synchronize 205 * operations. 206 * 207 * Call this function with the parent dir semaphore and the parent dir 208 * cluster lock held. 209 * 210 * The dir semaphore will protect us from having to worry about 211 * concurrent processes on our node trying to attach a lock at the 212 * same time. 213 * 214 * The dir cluster lock (held at either PR or EX mode) protects us 215 * from unlink and rename on other nodes. 216 * 217 * A dput() can happen asynchronously due to pruning, so we cover 218 * attaching and detaching the dentry lock with a 219 * dentry_attach_lock. 220 * 221 * A node which has done lookup on a name retains a protected read 222 * lock until final dput. If the user requests and unlink or rename, 223 * the protected read is upgraded to an exclusive lock. Other nodes 224 * who have seen the dentry will then be informed that they need to 225 * downgrade their lock, which will involve d_delete on the 226 * dentry. This happens in ocfs2_dentry_convert_worker(). 227 */ 228 int ocfs2_dentry_attach_lock(struct dentry *dentry, 229 struct inode *inode, 230 u64 parent_blkno) 231 { 232 int ret; 233 struct dentry *alias; 234 struct ocfs2_dentry_lock *dl = dentry->d_fsdata; 235 236 trace_ocfs2_dentry_attach_lock(dentry->d_name.len, dentry->d_name.name, 237 (unsigned long long)parent_blkno, dl); 238 239 /* 240 * Negative dentry. We ignore these for now. 241 * 242 * XXX: Could we can improve ocfs2_dentry_revalidate() by 243 * tracking these? 244 */ 245 if (!inode) 246 return 0; 247 248 if (!dentry->d_inode && dentry->d_fsdata) { 249 /* Converting a negative dentry to positive 250 Clear dentry->d_fsdata */ 251 dentry->d_fsdata = dl = NULL; 252 } 253 254 if (dl) { 255 mlog_bug_on_msg(dl->dl_parent_blkno != parent_blkno, 256 " \"%.*s\": old parent: %llu, new: %llu\n", 257 dentry->d_name.len, dentry->d_name.name, 258 (unsigned long long)parent_blkno, 259 (unsigned long long)dl->dl_parent_blkno); 260 return 0; 261 } 262 263 alias = ocfs2_find_local_alias(inode, parent_blkno, 0); 264 if (alias) { 265 /* 266 * Great, an alias exists, which means we must have a 267 * dentry lock already. We can just grab the lock off 268 * the alias and add it to the list. 269 * 270 * We're depending here on the fact that this dentry 271 * was found and exists in the dcache and so must have 272 * a reference to the dentry_lock because we can't 273 * race creates. Final dput() cannot happen on it 274 * since we have it pinned, so our reference is safe. 275 */ 276 dl = alias->d_fsdata; 277 mlog_bug_on_msg(!dl, "parent %llu, ino %llu\n", 278 (unsigned long long)parent_blkno, 279 (unsigned long long)OCFS2_I(inode)->ip_blkno); 280 281 mlog_bug_on_msg(dl->dl_parent_blkno != parent_blkno, 282 " \"%.*s\": old parent: %llu, new: %llu\n", 283 dentry->d_name.len, dentry->d_name.name, 284 (unsigned long long)parent_blkno, 285 (unsigned long long)dl->dl_parent_blkno); 286 287 trace_ocfs2_dentry_attach_lock_found(dl->dl_lockres.l_name, 288 (unsigned long long)parent_blkno, 289 (unsigned long long)OCFS2_I(inode)->ip_blkno); 290 291 goto out_attach; 292 } 293 294 /* 295 * There are no other aliases 296 */ 297 dl = kmalloc(sizeof(*dl), GFP_NOFS); 298 if (!dl) { 299 ret = -ENOMEM; 300 mlog_errno(ret); 301 return ret; 302 } 303 304 dl->dl_count = 0; 305 /* 306 * Does this have to happen below, for all attaches, in case 307 * the struct inode gets blown away by the downconvert thread? 308 */ 309 dl->dl_inode = igrab(inode); 310 dl->dl_parent_blkno = parent_blkno; 311 ocfs2_dentry_lock_res_init(dl, parent_blkno, inode); 312 313 out_attach: 314 spin_lock(&dentry_attach_lock); 315 dentry->d_fsdata = dl; 316 dl->dl_count++; 317 spin_unlock(&dentry_attach_lock); 318 319 /* 320 * This actually gets us our PRMODE level lock. From now on, 321 * we'll have a notification if one of these names is 322 * destroyed on another node. 323 */ 324 ret = ocfs2_dentry_lock(dentry, 0); 325 if (!ret) 326 ocfs2_dentry_unlock(dentry, 0); 327 else 328 mlog_errno(ret); 329 330 /* 331 * In case of error, manually free the allocation and do the iput(). 332 * We need to do this because error here means no d_instantiate(), 333 * which means iput() will not be called during dput(dentry). 334 */ 335 if (ret < 0 && !alias) { 336 ocfs2_lock_res_free(&dl->dl_lockres); 337 BUG_ON(dl->dl_count != 1); 338 spin_lock(&dentry_attach_lock); 339 dentry->d_fsdata = NULL; 340 spin_unlock(&dentry_attach_lock); 341 kfree(dl); 342 iput(inode); 343 } 344 345 dput(alias); 346 347 return ret; 348 } 349 350 DEFINE_SPINLOCK(dentry_list_lock); 351 352 /* We limit the number of dentry locks to drop in one go. We have 353 * this limit so that we don't starve other users of ocfs2_wq. */ 354 #define DL_INODE_DROP_COUNT 64 355 356 /* Drop inode references from dentry locks */ 357 static void __ocfs2_drop_dl_inodes(struct ocfs2_super *osb, int drop_count) 358 { 359 struct ocfs2_dentry_lock *dl; 360 361 spin_lock(&dentry_list_lock); 362 while (osb->dentry_lock_list && (drop_count < 0 || drop_count--)) { 363 dl = osb->dentry_lock_list; 364 osb->dentry_lock_list = dl->dl_next; 365 spin_unlock(&dentry_list_lock); 366 iput(dl->dl_inode); 367 kfree(dl); 368 spin_lock(&dentry_list_lock); 369 } 370 spin_unlock(&dentry_list_lock); 371 } 372 373 void ocfs2_drop_dl_inodes(struct work_struct *work) 374 { 375 struct ocfs2_super *osb = container_of(work, struct ocfs2_super, 376 dentry_lock_work); 377 378 __ocfs2_drop_dl_inodes(osb, DL_INODE_DROP_COUNT); 379 /* 380 * Don't queue dropping if umount is in progress. We flush the 381 * list in ocfs2_dismount_volume 382 */ 383 spin_lock(&dentry_list_lock); 384 if (osb->dentry_lock_list && 385 !ocfs2_test_osb_flag(osb, OCFS2_OSB_DROP_DENTRY_LOCK_IMMED)) 386 queue_work(ocfs2_wq, &osb->dentry_lock_work); 387 spin_unlock(&dentry_list_lock); 388 } 389 390 /* Flush the whole work queue */ 391 void ocfs2_drop_all_dl_inodes(struct ocfs2_super *osb) 392 { 393 __ocfs2_drop_dl_inodes(osb, -1); 394 } 395 396 /* 397 * ocfs2_dentry_iput() and friends. 398 * 399 * At this point, our particular dentry is detached from the inodes 400 * alias list, so there's no way that the locking code can find it. 401 * 402 * The interesting stuff happens when we determine that our lock needs 403 * to go away because this is the last subdir alias in the 404 * system. This function needs to handle a couple things: 405 * 406 * 1) Synchronizing lock shutdown with the downconvert threads. This 407 * is already handled for us via the lockres release drop function 408 * called in ocfs2_release_dentry_lock() 409 * 410 * 2) A race may occur when we're doing our lock shutdown and 411 * another process wants to create a new dentry lock. Right now we 412 * let them race, which means that for a very short while, this 413 * node might have two locks on a lock resource. This should be a 414 * problem though because one of them is in the process of being 415 * thrown out. 416 */ 417 static void ocfs2_drop_dentry_lock(struct ocfs2_super *osb, 418 struct ocfs2_dentry_lock *dl) 419 { 420 ocfs2_simple_drop_lockres(osb, &dl->dl_lockres); 421 ocfs2_lock_res_free(&dl->dl_lockres); 422 423 /* We leave dropping of inode reference to ocfs2_wq as that can 424 * possibly lead to inode deletion which gets tricky */ 425 spin_lock(&dentry_list_lock); 426 if (!osb->dentry_lock_list && 427 !ocfs2_test_osb_flag(osb, OCFS2_OSB_DROP_DENTRY_LOCK_IMMED)) 428 queue_work(ocfs2_wq, &osb->dentry_lock_work); 429 dl->dl_next = osb->dentry_lock_list; 430 osb->dentry_lock_list = dl; 431 spin_unlock(&dentry_list_lock); 432 } 433 434 void ocfs2_dentry_lock_put(struct ocfs2_super *osb, 435 struct ocfs2_dentry_lock *dl) 436 { 437 int unlock; 438 439 BUG_ON(dl->dl_count == 0); 440 441 spin_lock(&dentry_attach_lock); 442 dl->dl_count--; 443 unlock = !dl->dl_count; 444 spin_unlock(&dentry_attach_lock); 445 446 if (unlock) 447 ocfs2_drop_dentry_lock(osb, dl); 448 } 449 450 static void ocfs2_dentry_iput(struct dentry *dentry, struct inode *inode) 451 { 452 struct ocfs2_dentry_lock *dl = dentry->d_fsdata; 453 454 if (!dl) { 455 /* 456 * No dentry lock is ok if we're disconnected or 457 * unhashed. 458 */ 459 if (!(dentry->d_flags & DCACHE_DISCONNECTED) && 460 !d_unhashed(dentry)) { 461 unsigned long long ino = 0ULL; 462 if (inode) 463 ino = (unsigned long long)OCFS2_I(inode)->ip_blkno; 464 mlog(ML_ERROR, "Dentry is missing cluster lock. " 465 "inode: %llu, d_flags: 0x%x, d_name: %.*s\n", 466 ino, dentry->d_flags, dentry->d_name.len, 467 dentry->d_name.name); 468 } 469 470 goto out; 471 } 472 473 mlog_bug_on_msg(dl->dl_count == 0, "dentry: %.*s, count: %u\n", 474 dentry->d_name.len, dentry->d_name.name, 475 dl->dl_count); 476 477 ocfs2_dentry_lock_put(OCFS2_SB(dentry->d_sb), dl); 478 479 out: 480 iput(inode); 481 } 482 483 /* 484 * d_move(), but keep the locks in sync. 485 * 486 * When we are done, "dentry" will have the parent dir and name of 487 * "target", which will be thrown away. 488 * 489 * We manually update the lock of "dentry" if need be. 490 * 491 * "target" doesn't have it's dentry lock touched - we allow the later 492 * dput() to handle this for us. 493 * 494 * This is called during ocfs2_rename(), while holding parent 495 * directory locks. The dentries have already been deleted on other 496 * nodes via ocfs2_remote_dentry_delete(). 497 * 498 * Normally, the VFS handles the d_move() for the file system, after 499 * the ->rename() callback. OCFS2 wants to handle this internally, so 500 * the new lock can be created atomically with respect to the cluster. 501 */ 502 void ocfs2_dentry_move(struct dentry *dentry, struct dentry *target, 503 struct inode *old_dir, struct inode *new_dir) 504 { 505 int ret; 506 struct ocfs2_super *osb = OCFS2_SB(old_dir->i_sb); 507 struct inode *inode = dentry->d_inode; 508 509 /* 510 * Move within the same directory, so the actual lock info won't 511 * change. 512 * 513 * XXX: Is there any advantage to dropping the lock here? 514 */ 515 if (old_dir == new_dir) 516 goto out_move; 517 518 ocfs2_dentry_lock_put(osb, dentry->d_fsdata); 519 520 dentry->d_fsdata = NULL; 521 ret = ocfs2_dentry_attach_lock(dentry, inode, OCFS2_I(new_dir)->ip_blkno); 522 if (ret) 523 mlog_errno(ret); 524 525 out_move: 526 d_move(dentry, target); 527 } 528 529 const struct dentry_operations ocfs2_dentry_ops = { 530 .d_revalidate = ocfs2_dentry_revalidate, 531 .d_iput = ocfs2_dentry_iput, 532 }; 533