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 list_head *p; 174 struct dentry *dentry = NULL; 175 176 spin_lock(&inode->i_lock); 177 list_for_each(p, &inode->i_dentry) { 178 dentry = list_entry(p, struct dentry, d_alias); 179 180 spin_lock(&dentry->d_lock); 181 if (ocfs2_match_dentry(dentry, parent_blkno, skip_unhashed)) { 182 trace_ocfs2_find_local_alias(dentry->d_name.len, 183 dentry->d_name.name); 184 185 dget_dlock(dentry); 186 spin_unlock(&dentry->d_lock); 187 break; 188 } 189 spin_unlock(&dentry->d_lock); 190 191 dentry = NULL; 192 } 193 194 spin_unlock(&inode->i_lock); 195 196 return dentry; 197 } 198 199 DEFINE_SPINLOCK(dentry_attach_lock); 200 201 /* 202 * Attach this dentry to a cluster lock. 203 * 204 * Dentry locks cover all links in a given directory to a particular 205 * inode. We do this so that ocfs2 can build a lock name which all 206 * nodes in the cluster can agree on at all times. Shoving full names 207 * in the cluster lock won't work due to size restrictions. Covering 208 * links inside of a directory is a good compromise because it still 209 * allows us to use the parent directory lock to synchronize 210 * operations. 211 * 212 * Call this function with the parent dir semaphore and the parent dir 213 * cluster lock held. 214 * 215 * The dir semaphore will protect us from having to worry about 216 * concurrent processes on our node trying to attach a lock at the 217 * same time. 218 * 219 * The dir cluster lock (held at either PR or EX mode) protects us 220 * from unlink and rename on other nodes. 221 * 222 * A dput() can happen asynchronously due to pruning, so we cover 223 * attaching and detaching the dentry lock with a 224 * dentry_attach_lock. 225 * 226 * A node which has done lookup on a name retains a protected read 227 * lock until final dput. If the user requests and unlink or rename, 228 * the protected read is upgraded to an exclusive lock. Other nodes 229 * who have seen the dentry will then be informed that they need to 230 * downgrade their lock, which will involve d_delete on the 231 * dentry. This happens in ocfs2_dentry_convert_worker(). 232 */ 233 int ocfs2_dentry_attach_lock(struct dentry *dentry, 234 struct inode *inode, 235 u64 parent_blkno) 236 { 237 int ret; 238 struct dentry *alias; 239 struct ocfs2_dentry_lock *dl = dentry->d_fsdata; 240 241 trace_ocfs2_dentry_attach_lock(dentry->d_name.len, dentry->d_name.name, 242 (unsigned long long)parent_blkno, dl); 243 244 /* 245 * Negative dentry. We ignore these for now. 246 * 247 * XXX: Could we can improve ocfs2_dentry_revalidate() by 248 * tracking these? 249 */ 250 if (!inode) 251 return 0; 252 253 if (!dentry->d_inode && dentry->d_fsdata) { 254 /* Converting a negative dentry to positive 255 Clear dentry->d_fsdata */ 256 dentry->d_fsdata = dl = NULL; 257 } 258 259 if (dl) { 260 mlog_bug_on_msg(dl->dl_parent_blkno != parent_blkno, 261 " \"%.*s\": old parent: %llu, new: %llu\n", 262 dentry->d_name.len, dentry->d_name.name, 263 (unsigned long long)parent_blkno, 264 (unsigned long long)dl->dl_parent_blkno); 265 return 0; 266 } 267 268 alias = ocfs2_find_local_alias(inode, parent_blkno, 0); 269 if (alias) { 270 /* 271 * Great, an alias exists, which means we must have a 272 * dentry lock already. We can just grab the lock off 273 * the alias and add it to the list. 274 * 275 * We're depending here on the fact that this dentry 276 * was found and exists in the dcache and so must have 277 * a reference to the dentry_lock because we can't 278 * race creates. Final dput() cannot happen on it 279 * since we have it pinned, so our reference is safe. 280 */ 281 dl = alias->d_fsdata; 282 mlog_bug_on_msg(!dl, "parent %llu, ino %llu\n", 283 (unsigned long long)parent_blkno, 284 (unsigned long long)OCFS2_I(inode)->ip_blkno); 285 286 mlog_bug_on_msg(dl->dl_parent_blkno != parent_blkno, 287 " \"%.*s\": old parent: %llu, new: %llu\n", 288 dentry->d_name.len, dentry->d_name.name, 289 (unsigned long long)parent_blkno, 290 (unsigned long long)dl->dl_parent_blkno); 291 292 trace_ocfs2_dentry_attach_lock_found(dl->dl_lockres.l_name, 293 (unsigned long long)parent_blkno, 294 (unsigned long long)OCFS2_I(inode)->ip_blkno); 295 296 goto out_attach; 297 } 298 299 /* 300 * There are no other aliases 301 */ 302 dl = kmalloc(sizeof(*dl), GFP_NOFS); 303 if (!dl) { 304 ret = -ENOMEM; 305 mlog_errno(ret); 306 return ret; 307 } 308 309 dl->dl_count = 0; 310 /* 311 * Does this have to happen below, for all attaches, in case 312 * the struct inode gets blown away by the downconvert thread? 313 */ 314 dl->dl_inode = igrab(inode); 315 dl->dl_parent_blkno = parent_blkno; 316 ocfs2_dentry_lock_res_init(dl, parent_blkno, inode); 317 318 out_attach: 319 spin_lock(&dentry_attach_lock); 320 dentry->d_fsdata = dl; 321 dl->dl_count++; 322 spin_unlock(&dentry_attach_lock); 323 324 /* 325 * This actually gets us our PRMODE level lock. From now on, 326 * we'll have a notification if one of these names is 327 * destroyed on another node. 328 */ 329 ret = ocfs2_dentry_lock(dentry, 0); 330 if (!ret) 331 ocfs2_dentry_unlock(dentry, 0); 332 else 333 mlog_errno(ret); 334 335 /* 336 * In case of error, manually free the allocation and do the iput(). 337 * We need to do this because error here means no d_instantiate(), 338 * which means iput() will not be called during dput(dentry). 339 */ 340 if (ret < 0 && !alias) { 341 ocfs2_lock_res_free(&dl->dl_lockres); 342 BUG_ON(dl->dl_count != 1); 343 spin_lock(&dentry_attach_lock); 344 dentry->d_fsdata = NULL; 345 spin_unlock(&dentry_attach_lock); 346 kfree(dl); 347 iput(inode); 348 } 349 350 dput(alias); 351 352 return ret; 353 } 354 355 DEFINE_SPINLOCK(dentry_list_lock); 356 357 /* We limit the number of dentry locks to drop in one go. We have 358 * this limit so that we don't starve other users of ocfs2_wq. */ 359 #define DL_INODE_DROP_COUNT 64 360 361 /* Drop inode references from dentry locks */ 362 static void __ocfs2_drop_dl_inodes(struct ocfs2_super *osb, int drop_count) 363 { 364 struct ocfs2_dentry_lock *dl; 365 366 spin_lock(&dentry_list_lock); 367 while (osb->dentry_lock_list && (drop_count < 0 || drop_count--)) { 368 dl = osb->dentry_lock_list; 369 osb->dentry_lock_list = dl->dl_next; 370 spin_unlock(&dentry_list_lock); 371 iput(dl->dl_inode); 372 kfree(dl); 373 spin_lock(&dentry_list_lock); 374 } 375 spin_unlock(&dentry_list_lock); 376 } 377 378 void ocfs2_drop_dl_inodes(struct work_struct *work) 379 { 380 struct ocfs2_super *osb = container_of(work, struct ocfs2_super, 381 dentry_lock_work); 382 383 __ocfs2_drop_dl_inodes(osb, DL_INODE_DROP_COUNT); 384 /* 385 * Don't queue dropping if umount is in progress. We flush the 386 * list in ocfs2_dismount_volume 387 */ 388 spin_lock(&dentry_list_lock); 389 if (osb->dentry_lock_list && 390 !ocfs2_test_osb_flag(osb, OCFS2_OSB_DROP_DENTRY_LOCK_IMMED)) 391 queue_work(ocfs2_wq, &osb->dentry_lock_work); 392 spin_unlock(&dentry_list_lock); 393 } 394 395 /* Flush the whole work queue */ 396 void ocfs2_drop_all_dl_inodes(struct ocfs2_super *osb) 397 { 398 __ocfs2_drop_dl_inodes(osb, -1); 399 } 400 401 /* 402 * ocfs2_dentry_iput() and friends. 403 * 404 * At this point, our particular dentry is detached from the inodes 405 * alias list, so there's no way that the locking code can find it. 406 * 407 * The interesting stuff happens when we determine that our lock needs 408 * to go away because this is the last subdir alias in the 409 * system. This function needs to handle a couple things: 410 * 411 * 1) Synchronizing lock shutdown with the downconvert threads. This 412 * is already handled for us via the lockres release drop function 413 * called in ocfs2_release_dentry_lock() 414 * 415 * 2) A race may occur when we're doing our lock shutdown and 416 * another process wants to create a new dentry lock. Right now we 417 * let them race, which means that for a very short while, this 418 * node might have two locks on a lock resource. This should be a 419 * problem though because one of them is in the process of being 420 * thrown out. 421 */ 422 static void ocfs2_drop_dentry_lock(struct ocfs2_super *osb, 423 struct ocfs2_dentry_lock *dl) 424 { 425 ocfs2_simple_drop_lockres(osb, &dl->dl_lockres); 426 ocfs2_lock_res_free(&dl->dl_lockres); 427 428 /* We leave dropping of inode reference to ocfs2_wq as that can 429 * possibly lead to inode deletion which gets tricky */ 430 spin_lock(&dentry_list_lock); 431 if (!osb->dentry_lock_list && 432 !ocfs2_test_osb_flag(osb, OCFS2_OSB_DROP_DENTRY_LOCK_IMMED)) 433 queue_work(ocfs2_wq, &osb->dentry_lock_work); 434 dl->dl_next = osb->dentry_lock_list; 435 osb->dentry_lock_list = dl; 436 spin_unlock(&dentry_list_lock); 437 } 438 439 void ocfs2_dentry_lock_put(struct ocfs2_super *osb, 440 struct ocfs2_dentry_lock *dl) 441 { 442 int unlock; 443 444 BUG_ON(dl->dl_count == 0); 445 446 spin_lock(&dentry_attach_lock); 447 dl->dl_count--; 448 unlock = !dl->dl_count; 449 spin_unlock(&dentry_attach_lock); 450 451 if (unlock) 452 ocfs2_drop_dentry_lock(osb, dl); 453 } 454 455 static void ocfs2_dentry_iput(struct dentry *dentry, struct inode *inode) 456 { 457 struct ocfs2_dentry_lock *dl = dentry->d_fsdata; 458 459 if (!dl) { 460 /* 461 * No dentry lock is ok if we're disconnected or 462 * unhashed. 463 */ 464 if (!(dentry->d_flags & DCACHE_DISCONNECTED) && 465 !d_unhashed(dentry)) { 466 unsigned long long ino = 0ULL; 467 if (inode) 468 ino = (unsigned long long)OCFS2_I(inode)->ip_blkno; 469 mlog(ML_ERROR, "Dentry is missing cluster lock. " 470 "inode: %llu, d_flags: 0x%x, d_name: %.*s\n", 471 ino, dentry->d_flags, dentry->d_name.len, 472 dentry->d_name.name); 473 } 474 475 goto out; 476 } 477 478 mlog_bug_on_msg(dl->dl_count == 0, "dentry: %.*s, count: %u\n", 479 dentry->d_name.len, dentry->d_name.name, 480 dl->dl_count); 481 482 ocfs2_dentry_lock_put(OCFS2_SB(dentry->d_sb), dl); 483 484 out: 485 iput(inode); 486 } 487 488 /* 489 * d_move(), but keep the locks in sync. 490 * 491 * When we are done, "dentry" will have the parent dir and name of 492 * "target", which will be thrown away. 493 * 494 * We manually update the lock of "dentry" if need be. 495 * 496 * "target" doesn't have it's dentry lock touched - we allow the later 497 * dput() to handle this for us. 498 * 499 * This is called during ocfs2_rename(), while holding parent 500 * directory locks. The dentries have already been deleted on other 501 * nodes via ocfs2_remote_dentry_delete(). 502 * 503 * Normally, the VFS handles the d_move() for the file system, after 504 * the ->rename() callback. OCFS2 wants to handle this internally, so 505 * the new lock can be created atomically with respect to the cluster. 506 */ 507 void ocfs2_dentry_move(struct dentry *dentry, struct dentry *target, 508 struct inode *old_dir, struct inode *new_dir) 509 { 510 int ret; 511 struct ocfs2_super *osb = OCFS2_SB(old_dir->i_sb); 512 struct inode *inode = dentry->d_inode; 513 514 /* 515 * Move within the same directory, so the actual lock info won't 516 * change. 517 * 518 * XXX: Is there any advantage to dropping the lock here? 519 */ 520 if (old_dir == new_dir) 521 goto out_move; 522 523 ocfs2_dentry_lock_put(osb, dentry->d_fsdata); 524 525 dentry->d_fsdata = NULL; 526 ret = ocfs2_dentry_attach_lock(dentry, inode, OCFS2_I(new_dir)->ip_blkno); 527 if (ret) 528 mlog_errno(ret); 529 530 out_move: 531 d_move(dentry, target); 532 } 533 534 const struct dentry_operations ocfs2_dentry_ops = { 535 .d_revalidate = ocfs2_dentry_revalidate, 536 .d_iput = ocfs2_dentry_iput, 537 }; 538