fs/ocfs2/dcache.c

   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
  42
  43 static int ocfs2_dentry_revalidate(struct dentry *dentry,
  44                                    struct nameidata *nd)
  45 {
  46         struct inode *inode = dentry->d_inode;
  47         int ret = 0;    /* if all else fails, just return false */
  48         struct ocfs2_super *osb = OCFS2_SB(dentry->d_sb);
  49
  50         mlog_entry("(0x%p, '%.*s')\n", dentry,
  51                    dentry->d_name.len, dentry->d_name.name);
  52
  53         /* Never trust a negative dentry - force a new lookup. */
  54         if (inode == NULL) {
  55                 mlog(0, "negative dentry: %.*s\n", dentry->d_name.len,
  56                      dentry->d_name.name);
  57                 goto bail;
  58         }
  59
  60         BUG_ON(!osb);
  61
  62         if (inode == osb->root_inode || is_bad_inode(inode))
  63                 goto bail;
  64
  65         spin_lock(&OCFS2_I(inode)->ip_lock);
  66         /* did we or someone else delete this inode? */
  67         if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_DELETED) {
  68                 spin_unlock(&OCFS2_I(inode)->ip_lock);
  69                 mlog(0, "inode (%llu) deleted, returning false\n",
  70                      (unsigned long long)OCFS2_I(inode)->ip_blkno);
  71                 goto bail;
  72         }
  73         spin_unlock(&OCFS2_I(inode)->ip_lock);
  74
  75         /*
  76          * We don't need a cluster lock to test this because once an
  77          * inode nlink hits zero, it never goes back.
  78          */
  79         if (inode->i_nlink == 0) {
  80                 mlog(0, "Inode %llu orphaned, returning false "
  81                      "dir = %d\n",
  82                      (unsigned long long)OCFS2_I(inode)->ip_blkno,
  83                      S_ISDIR(inode->i_mode));
  84                 goto bail;
  85         }
  86
  87         ret = 1;
  88
  89 bail:
  90         mlog_exit(ret);
  91
  92         return ret;
  93 }
  94
  95 static int ocfs2_match_dentry(struct dentry *dentry,
  96                               u64 parent_blkno,
  97                               int skip_unhashed)
  98 {
  99         struct inode *parent;
 100
 101         /*
 102          * ocfs2_lookup() does a d_splice_alias() _before_ attaching
 103          * to the lock data, so we skip those here, otherwise
 104          * ocfs2_dentry_attach_lock() will get its original dentry
 105          * back.
 106          */
 107         if (!dentry->d_fsdata)
 108                 return 0;
 109
 110         if (!dentry->d_parent)
 111                 return 0;
 112
 113         if (skip_unhashed && d_unhashed(dentry))
 114                 return 0;
 115
 116         parent = dentry->d_parent->d_inode;
 117         /* Negative parent dentry? */
 118         if (!parent)
 119                 return 0;
 120
 121         /* Name is in a different directory. */
 122         if (OCFS2_I(parent)->ip_blkno != parent_blkno)
 123                 return 0;
 124
 125         return 1;
 126 }
 127
 128 /*
 129  * Walk the inode alias list, and find a dentry which has a given
 130  * parent. ocfs2_dentry_attach_lock() wants to find _any_ alias as it
 131  * is looking for a dentry_lock reference. The vote thread is looking
 132  * to unhash aliases, so we allow it to skip any that already have
 133  * that property.
 134  */
 135 struct dentry *ocfs2_find_local_alias(struct inode *inode,
 136                                       u64 parent_blkno,
 137                                       int skip_unhashed)
 138 {
 139         struct list_head *p;
 140         struct dentry *dentry = NULL;
 141
 142         spin_lock(&dcache_lock);
 143
 144         list_for_each(p, &inode->i_dentry) {
 145                 dentry = list_entry(p, struct dentry, d_alias);
 146
 147                 if (ocfs2_match_dentry(dentry, parent_blkno, skip_unhashed)) {
 148                         mlog(0, "dentry found: %.*s\n",
 149                              dentry->d_name.len, dentry->d_name.name);
 150
 151                         dget_locked(dentry);
 152                         break;
 153                 }
 154
 155                 dentry = NULL;
 156         }
 157
 158         spin_unlock(&dcache_lock);
 159
 160         return dentry;
 161 }
 162
 163 DEFINE_SPINLOCK(dentry_attach_lock);
 164
 165 /*
 166  * Attach this dentry to a cluster lock.
 167  *
 168  * Dentry locks cover all links in a given directory to a particular
 169  * inode. We do this so that ocfs2 can build a lock name which all
 170  * nodes in the cluster can agree on at all times. Shoving full names
 171  * in the cluster lock won't work due to size restrictions. Covering
 172  * links inside of a directory is a good compromise because it still
 173  * allows us to use the parent directory lock to synchronize
 174  * operations.
 175  *
 176  * Call this function with the parent dir semaphore and the parent dir
 177  * cluster lock held.
 178  *
 179  * The dir semaphore will protect us from having to worry about
 180  * concurrent processes on our node trying to attach a lock at the
 181  * same time.
 182  *
 183  * The dir cluster lock (held at either PR or EX mode) protects us
 184  * from unlink and rename on other nodes.
 185  *
 186  * A dput() can happen asynchronously due to pruning, so we cover
 187  * attaching and detaching the dentry lock with a
 188  * dentry_attach_lock.
 189  *
 190  * A node which has done lookup on a name retains a protected read
 191  * lock until final dput. If the user requests and unlink or rename,
 192  * the protected read is upgraded to an exclusive lock. Other nodes
 193  * who have seen the dentry will then be informed that they need to
 194  * downgrade their lock, which will involve d_delete on the
 195  * dentry. This happens in ocfs2_dentry_convert_worker().
 196  */
 197 int ocfs2_dentry_attach_lock(struct dentry *dentry,
 198                              struct inode *inode,
 199                              u64 parent_blkno)
 200 {
 201         int ret;
 202         struct dentry *alias;
 203         struct ocfs2_dentry_lock *dl = dentry->d_fsdata;
 204
 205         mlog(0, "Attach \"%.*s\", parent %llu, fsdata: %p\n",
 206              dentry->d_name.len, dentry->d_name.name,
 207              (unsigned long long)parent_blkno, dl);
 208
 209         /*
 210          * Negative dentry. We ignore these for now.
 211          *
 212          * XXX: Could we can improve ocfs2_dentry_revalidate() by
 213          * tracking these?
 214          */
 215         if (!inode)
 216                 return 0;
 217
 218         if (dl) {
 219                 mlog_bug_on_msg(dl->dl_parent_blkno != parent_blkno,
 220                                 " \"%.*s\": old parent: %llu, new: %llu\n",
 221                                 dentry->d_name.len, dentry->d_name.name,
 222                                 (unsigned long long)parent_blkno,
 223                                 (unsigned long long)dl->dl_parent_blkno);
 224                 return 0;
 225         }
 226
 227         alias = ocfs2_find_local_alias(inode, parent_blkno, 0);
 228         if (alias) {
 229                 /*
 230                  * Great, an alias exists, which means we must have a
 231                  * dentry lock already. We can just grab the lock off
 232                  * the alias and add it to the list.
 233                  *
 234                  * We're depending here on the fact that this dentry
 235                  * was found and exists in the dcache and so must have
 236                  * a reference to the dentry_lock because we can't
 237                  * race creates. Final dput() cannot happen on it
 238                  * since we have it pinned, so our reference is safe.
 239                  */
 240                 dl = alias->d_fsdata;
 241                 mlog_bug_on_msg(!dl, "parent %llu, ino %llu\n",
 242                                 (unsigned long long)parent_blkno,
 243                                 (unsigned long long)OCFS2_I(inode)->ip_blkno);
 244
 245                 mlog_bug_on_msg(dl->dl_parent_blkno != parent_blkno,
 246                                 " \"%.*s\": old parent: %llu, new: %llu\n",
 247                                 dentry->d_name.len, dentry->d_name.name,
 248                                 (unsigned long long)parent_blkno,
 249                                 (unsigned long long)dl->dl_parent_blkno);
 250
 251                 mlog(0, "Found: %s\n", dl->dl_lockres.l_name);
 252
 253                 goto out_attach;
 254         }
 255
 256         /*
 257          * There are no other aliases
 258          */
 259         dl = kmalloc(sizeof(*dl), GFP_NOFS);
 260         if (!dl) {
 261                 ret = -ENOMEM;
 262                 mlog_errno(ret);
 263                 return ret;
 264         }
 265
 266         dl->dl_count = 0;
 267         /*
 268          * Does this have to happen below, for all attaches, in case
 269          * the struct inode gets blown away by votes?
 270          */
 271         dl->dl_inode = igrab(inode);
 272         dl->dl_parent_blkno = parent_blkno;
 273         ocfs2_dentry_lock_res_init(dl, parent_blkno, inode);
 274
 275 out_attach:
 276         spin_lock(&dentry_attach_lock);
 277         dentry->d_fsdata = dl;
 278         dl->dl_count++;
 279         spin_unlock(&dentry_attach_lock);
 280
 281         /*
 282          * This actually gets us our PRMODE level lock. From now on,
 283          * we'll have a notification if one of these names is
 284          * destroyed on another node.
 285          */
 286         ret = ocfs2_dentry_lock(dentry, 0);
 287         if (!ret)
 288                 ocfs2_dentry_unlock(dentry, 0);
 289         else
 290                 mlog_errno(ret);
 291
 292         dput(alias);
 293
 294         return ret;
 295 }
 296
 297 /*
 298  * ocfs2_dentry_iput() and friends.
 299  *
 300  * At this point, our particular dentry is detached from the inodes
 301  * alias list, so there's no way that the locking code can find it.
 302  *
 303  * The interesting stuff happens when we determine that our lock needs
 304  * to go away because this is the last subdir alias in the
 305  * system. This function needs to handle a couple things:
 306  *
 307  * 1) Synchronizing lock shutdown with the downconvert threads. This
 308  *    is already handled for us via the lockres release drop function
 309  *    called in ocfs2_release_dentry_lock()
 310  *
 311  * 2) A race may occur when we're doing our lock shutdown and
 312  *    another process wants to create a new dentry lock. Right now we
 313  *    let them race, which means that for a very short while, this
 314  *    node might have two locks on a lock resource. This should be a
 315  *    problem though because one of them is in the process of being
 316  *    thrown out.
 317  */
 318 static void ocfs2_drop_dentry_lock(struct ocfs2_super *osb,
 319                                    struct ocfs2_dentry_lock *dl)
 320 {
 321         ocfs2_simple_drop_lockres(osb, &dl->dl_lockres);
 322         ocfs2_lock_res_free(&dl->dl_lockres);
 323         iput(dl->dl_inode);
 324         kfree(dl);
 325 }
 326
 327 void ocfs2_dentry_lock_put(struct ocfs2_super *osb,
 328                            struct ocfs2_dentry_lock *dl)
 329 {
 330         int unlock = 0;
 331
 332         BUG_ON(dl->dl_count == 0);
 333
 334         spin_lock(&dentry_attach_lock);
 335         dl->dl_count--;
 336         unlock = !dl->dl_count;
 337         spin_unlock(&dentry_attach_lock);
 338
 339         if (unlock)
 340                 ocfs2_drop_dentry_lock(osb, dl);
 341 }
 342
 343 static void ocfs2_dentry_iput(struct dentry *dentry, struct inode *inode)
 344 {
 345         struct ocfs2_dentry_lock *dl = dentry->d_fsdata;
 346
 347         mlog_bug_on_msg(!dl && !(dentry->d_flags & DCACHE_DISCONNECTED),
 348                         "dentry: %.*s\n", dentry->d_name.len,
 349                         dentry->d_name.name);
 350
 351         if (!dl)
 352                 goto out;
 353
 354         mlog_bug_on_msg(dl->dl_count == 0, "dentry: %.*s, count: %u\n",
 355                         dentry->d_name.len, dentry->d_name.name,
 356                         dl->dl_count);
 357
 358         ocfs2_dentry_lock_put(OCFS2_SB(dentry->d_sb), dl);
 359
 360 out:
 361         iput(inode);
 362 }
 363
 364 /*
 365  * d_move(), but keep the locks in sync.
 366  *
 367  * When we are done, "dentry" will have the parent dir and name of
 368  * "target", which will be thrown away.
 369  *
 370  * We manually update the lock of "dentry" if need be.
 371  *
 372  * "target" doesn't have it's dentry lock touched - we allow the later
 373  * dput() to handle this for us.
 374  *
 375  * This is called during ocfs2_rename(), while holding parent
 376  * directory locks. The dentries have already been deleted on other
 377  * nodes via ocfs2_remote_dentry_delete().
 378  *
 379  * Normally, the VFS handles the d_move() for the file sytem, after
 380  * the ->rename() callback. OCFS2 wants to handle this internally, so
 381  * the new lock can be created atomically with respect to the cluster.
 382  */
 383 void ocfs2_dentry_move(struct dentry *dentry, struct dentry *target,
 384                        struct inode *old_dir, struct inode *new_dir)
 385 {
 386         int ret;
 387         struct ocfs2_super *osb = OCFS2_SB(old_dir->i_sb);
 388         struct inode *inode = dentry->d_inode;
 389
 390         /*
 391          * Move within the same directory, so the actual lock info won't
 392          * change.
 393          *
 394          * XXX: Is there any advantage to dropping the lock here?
 395          */
 396         if (old_dir == new_dir)
 397                 goto out_move;
 398
 399         ocfs2_dentry_lock_put(osb, dentry->d_fsdata);
 400
 401         dentry->d_fsdata = NULL;
 402         ret = ocfs2_dentry_attach_lock(dentry, inode, OCFS2_I(new_dir)->ip_blkno);
 403         if (ret)
 404                 mlog_errno(ret);
 405
 406 out_move:
 407         d_move(dentry, target);
 408 }
 409
 410 struct dentry_operations ocfs2_dentry_ops = {
 411         .d_revalidate           = ocfs2_dentry_revalidate,
 412         .d_iput                 = ocfs2_dentry_iput,
 413 };