Whamcloud - gitweb
Land first part of new dcache handling (bug 16654).
[fs/lustre-release.git] / lustre / llite / dcache.c
index fc2ee94..2ef00a9 100644 (file)
@@ -1,22 +1,37 @@
 /* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
  * vim:expandtab:shiftwidth=8:tabstop=8:
  *
- *  Copyright (c) 2001-2003 Cluster File Systems, Inc.
+ * GPL HEADER START
  *
- *   This file is part of Lustre, http://www.lustre.org.
+ * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
- *   Lustre is free software; you can redistribute it and/or
- *   modify it under the terms of version 2 of the GNU General Public
- *   License as published by the Free Software Foundation.
+ * This program is free software; you can redistribute it and/or modify
+ * it under the terms of the GNU General Public License version 2 only,
+ * as published by the Free Software Foundation.
  *
- *   Lustre is distributed in the hope that it will be useful,
- *   but WITHOUT ANY WARRANTY; without even the implied warranty of
- *   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
- *   GNU General Public License for more details.
+ * This program is distributed in the hope that it will be useful, but
+ * WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * General Public License version 2 for more details (a copy is included
+ * in the LICENSE file that accompanied this code).
  *
- *   You should have received a copy of the GNU General Public License
- *   along with Lustre; if not, write to the Free Software
- *   Foundation, Inc., 675 Mass Ave, Cambridge, MA 02139, USA.
+ * You should have received a copy of the GNU General Public License
+ * version 2 along with this program; If not, see
+ * http://www.sun.com/software/products/lustre/docs/GPLv2.pdf
+ *
+ * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
+ * CA 95054 USA or visit www.sun.com if you need additional information or
+ * have any questions.
+ *
+ * GPL HEADER END
+ */
+/*
+ * Copyright  2008 Sun Microsystems, Inc. All rights reserved
+ * Use is subject to license terms.
+ */
+/*
+ * This file is part of Lustre, http://www.lustre.org/
+ * Lustre is a trademark of Sun Microsystems, Inc.
  */
 
 #include <linux/fs.h>
 
 #include "llite_internal.h"
 
+spinlock_t ll_lookup_lock = SPIN_LOCK_UNLOCKED;
+
 /* should NOT be called with the dcache lock, see fs/dcache.c */
-void ll_release(struct dentry *de)
+static void ll_release(struct dentry *de)
 {
         struct ll_dentry_data *lld;
         ENTRY;
@@ -60,7 +77,6 @@ void ll_release(struct dentry *de)
         EXIT;
 }
 
-#ifdef DCACHE_LUSTRE_INVALID
 /* Compare if two dentries are the same.  Don't match if the existing dentry
  * is marked DCACHE_LUSTRE_INVALID.  Returns 1 if different, 0 if the same.
  *
@@ -81,15 +97,22 @@ int ll_dcompare(struct dentry *parent, struct qstr *d_name, struct qstr *name)
 
         /* XXX: d_name must be in-dentry structure */
         dchild = container_of(d_name, struct dentry, d_name); /* ugh */
-        if (dchild->d_flags & DCACHE_LUSTRE_INVALID) {
-                CDEBUG(D_DENTRY,"INVALID dentry %p not matched, was bug 3784\n",
-                       dchild);
+
+        CDEBUG(D_DENTRY,"found name %.*s(%p) - flags %d/%x - refc %d\n",
+               name->len, name->name, dchild,
+               d_mountpoint(dchild), dchild->d_flags & DCACHE_LUSTRE_INVALID,
+               atomic_read(&dchild->d_count));
+
+         /* mountpoint is always valid */
+        if (d_mountpoint(dchild))
+                RETURN(0);
+
+        if (dchild->d_flags & DCACHE_LUSTRE_INVALID)
                 RETURN(1);
-        }
+
 
         RETURN(0);
 }
-#endif
 
 /* should NOT be called with the dcache lock, see fs/dcache.c */
 static int ll_ddelete(struct dentry *de)
@@ -126,7 +149,6 @@ void ll_set_dd(struct dentry *de)
 
                 OBD_ALLOC_PTR(lld);
                 if (likely(lld != NULL)) {
-                        cfs_waitq_init(&lld->lld_waitq);
                         lock_dentry(de);
                         if (likely(de->d_fsdata == NULL))
                                 de->d_fsdata = lld;
@@ -194,11 +216,13 @@ int ll_drop_dentry(struct dentry *dentry)
                 __d_drop(dentry);
                 unlock_dentry(dentry);
                 spin_unlock(&dcache_lock);
+                spin_unlock(&ll_lookup_lock);
                 dput(dentry);
+                spin_lock(&ll_lookup_lock);
                 spin_lock(&dcache_lock);
                 return 1;
         }
-        /* disconected dentry can not be find without lookup, because we 
+        /* disconected dentry can not be find without lookup, because we
          * not need his to unhash or mark invalid. */
         if (dentry->d_flags & DCACHE_DISCONNECTED) {
                 unlock_dentry(dentry);
@@ -242,6 +266,7 @@ void ll_unhash_aliases(struct inode *inode)
                inode->i_ino, inode->i_generation, inode);
 
         head = &inode->i_dentry;
+        spin_lock(&ll_lookup_lock);
         spin_lock(&dcache_lock);
 restart:
         tmp = head;
@@ -258,23 +283,14 @@ restart:
                                "ino=%lu\n", dentry, inode, inode->i_ino);
                         lustre_dump_dentry(dentry, 1);
                         libcfs_debug_dumpstack(NULL);
-                } else if (d_mountpoint(dentry)) {
-                        /* For mountpoints we skip removal of the dentry
-                           which happens solely because we have a lock on it
-                           obtained when this dentry was not a mountpoint yet */
-                        CDEBUG(D_DENTRY, "Skippind mountpoint dentry removal "
-                                         "%.*s (%p) parent %p\n",
-                                          dentry->d_name.len,
-                                          dentry->d_name.name,
-                                          dentry, dentry->d_parent);
-
-                        continue;
                 }
 
                 if (ll_drop_dentry(dentry))
                           goto restart;
         }
         spin_unlock(&dcache_lock);
+        spin_unlock(&ll_lookup_lock);
+
         EXIT;
 }
 
@@ -288,7 +304,7 @@ int ll_revalidate_it_finish(struct ptlrpc_request *request,
         if (!request)
                 RETURN(0);
 
-        if (it_disposition(it, DISP_LOOKUP_NEG)) 
+        if (it_disposition(it, DISP_LOOKUP_NEG))
                 RETURN(-ENOENT);
 
         rc = ll_prep_inode(&de->d_inode, request, NULL);
@@ -308,7 +324,7 @@ void ll_lookup_finish_locks(struct lookup_intent *it, struct dentry *dentry)
                 CDEBUG(D_DLMTRACE, "setting l_data to inode %p (%lu/%u)\n",
                        inode, inode->i_ino, inode->i_generation);
                 md_set_lock_data(sbi->ll_md_exp, &it->d.lustre.it_lock_handle,
-                                 inode);
+                                 inode, NULL);
         }
 
         /* drop lookup or getattr locks immediately */
@@ -325,7 +341,7 @@ void ll_frob_intent(struct lookup_intent **itp, struct lookup_intent *deft)
         struct lookup_intent *it = *itp;
 #ifdef HAVE_VFS_INTENT_PATCHES
         if (it) {
-                LASSERTF(it->it_magic == INTENT_MAGIC, 
+                LASSERTF(it->it_magic == INTENT_MAGIC,
                          "%p has bad intent magic: %x\n",
                          it, it->it_magic);
         }
@@ -372,17 +388,20 @@ int ll_revalidate_it(struct dentry *de, int lookup_flags,
                 GOTO(out_sa, rc);
         }
 
-        exp = ll_i2mdexp(de->d_inode);
-
         /* Never execute intents for mount points.
          * Attributes will be fixed up in ll_inode_revalidate_it */
         if (d_mountpoint(de))
                 GOTO(out_sa, rc = 1);
 
-        /* Root of the lustre tree. Always valid.
-         * Attributes will be fixed up in ll_inode_revalidate_it */
-        if (de == de->d_sb->s_root)
-                GOTO(out_sa, rc = 1);
+        /* need to get attributes in case root got changed from other client */
+        if (de == de->d_sb->s_root) {
+                rc = __ll_inode_revalidate_it(de, it, MDS_INODELOCK_LOOKUP);
+                if (rc == 0)
+                        rc = 1;
+                GOTO(out_sa, rc);
+        }
+
+        exp = ll_i2mdexp(de->d_inode);
 
         OBD_FAIL_TIMEOUT(OBD_FAIL_MDC_REVALIDATE_PAUSE, 5);
         ll_frob_intent(&it, &lookup_it);
@@ -449,14 +468,19 @@ int ll_revalidate_it(struct dentry *de, int lookup_flags,
 
 do_lock:
         it->it_create_mode &= ~current->fs->umask;
-        it->it_flags |= O_CHECK_STALE;
+        it->it_create_mode |= M_CHECK_STALE;
         rc = md_intent_lock(exp, op_data, NULL, 0, it,
                             lookup_flags,
                             &req, ll_md_blocking_ast, 0);
-        it->it_flags &= ~O_CHECK_STALE;
+        it->it_create_mode &= ~M_CHECK_STALE;
         ll_finish_md_op_data(op_data);
         if (it->it_op == IT_GETATTR && !first)
-                ll_statahead_exit(de, rc);
+                /* If there are too many locks on client-side, then some
+                 * locks taken by statahead maybe dropped automatically
+                 * before the real "revalidate" using them. */
+                ll_statahead_exit(de, req == NULL ? rc : 0);
+        else if (first == -EEXIST)
+                ll_statahead_mark(de);
 
         /* If req is NULL, then md_intent_lock only tried to do a lock match;
          * if all was well, it will return 1 if it found locks, 0 otherwise. */
@@ -482,8 +506,8 @@ revalidate_finish:
                 GOTO(out, rc = 0);
         }
 
-        if ((it->it_op & IT_OPEN) && de->d_inode && 
-            !S_ISREG(de->d_inode->i_mode) && 
+        if ((it->it_op & IT_OPEN) && de->d_inode &&
+            !S_ISREG(de->d_inode->i_mode) &&
             !S_ISDIR(de->d_inode->i_mode)) {
                 ll_release_openhandle(de, it);
         }
@@ -491,12 +515,14 @@ revalidate_finish:
 
         /* unfortunately ll_intent_lock may cause a callback and revoke our
          * dentry */
+        spin_lock(&ll_lookup_lock);
         spin_lock(&dcache_lock);
         lock_dentry(de);
         __d_drop(de);
         unlock_dentry(de);
         d_rehash_cond(de, 0);
         spin_unlock(&dcache_lock);
+        spin_unlock(&ll_lookup_lock);
 
 out:
         /* We do not free request as it may be reused during following lookup
@@ -589,7 +615,9 @@ out_sa:
         if (it && it->it_op == IT_GETATTR && rc == 1) {
                 first = ll_statahead_enter(de->d_parent->d_inode, &de, 0);
                 if (!first)
-                        ll_statahead_exit(de, rc);
+                        ll_statahead_exit(de, 1);
+                else if (first == -EEXIST)
+                        ll_statahead_mark(de);
         }
 
         return rc;
@@ -681,7 +709,7 @@ out_sa:
 }
 
 #ifdef HAVE_VFS_INTENT_PATCHES
-static int ll_revalidate_nd(struct dentry *dentry, struct nameidata *nd)
+int ll_revalidate_nd(struct dentry *dentry, struct nameidata *nd)
 {
         int rc;
         ENTRY;
@@ -724,16 +752,20 @@ int ll_revalidate_nd(struct dentry *dentry, struct nameidata *nd)
                                                (struct ptlrpc_request *)
                                                   it->d.lustre.it_data);
                         } else {
-                                struct file *filp;
-
-                                nd->intent.open.file->private_data = it;
-                                filp = lookup_instantiate_filp(nd, dentry,NULL);
 #if (LINUX_VERSION_CODE < KERNEL_VERSION(2,6,17))
 /* 2.6.1[456] have a bug in open_namei() that forgets to check
  * nd->intent.open.file for error, so we need to return it as lookup's result
  * instead */
-                                if (IS_ERR(filp))
-                                        rc = 0;
+                                struct file *filp;
+
+                                nd->intent.open.file->private_data = it;
+                                filp = lookup_instantiate_filp(nd, dentry,NULL);
+                                if (IS_ERR(filp)) {
+                                        rc = PTR_ERR(filp);
+                                }
+#else
+                                nd->intent.open.file->private_data = it;
+                                (void)lookup_instantiate_filp(nd, dentry,NULL);
 #endif
                         }
 #else
@@ -768,53 +800,9 @@ struct dentry_operations ll_d_ops = {
         .d_revalidate = ll_revalidate_nd,
         .d_release = ll_release,
         .d_delete = ll_ddelete,
-#ifdef DCACHE_LUSTRE_INVALID
         .d_compare = ll_dcompare,
-#endif
 #if 0
         .d_pin = ll_pin,
         .d_unpin = ll_unpin,
 #endif
 };
-
-static int ll_fini_revalidate_nd(struct dentry *dentry, struct nameidata *nd)
-{
-        ENTRY;
-        /* need lookup */
-        RETURN(0);
-}
-
-struct dentry_operations ll_fini_d_ops = {
-        .d_revalidate = ll_fini_revalidate_nd,
-        .d_release = ll_release,
-};
-
-/*
- * It is for the following race condition:
- * When someone (maybe statahead thread) adds the dentry to the dentry hash
- * table, the dentry's "d_op" maybe NULL, at the same time, another (maybe
- * "ls -l") process finds such dentry by "do_lookup()" without "do_revalidate()"
- * called. It causes statahead window lost, and maybe other issues. --Fan Yong
- */
-static int ll_init_revalidate_nd(struct dentry *dentry, struct nameidata *nd)
-{
-        struct l_wait_info lwi = { 0 };
-        struct ll_dentry_data *lld;
-        ENTRY;
-
-        ll_set_dd(dentry);
-        lld = ll_d2d(dentry);
-        if (unlikely(lld == NULL))
-                RETURN(-ENOMEM);
-
-        l_wait_event(lld->lld_waitq, dentry->d_op != &ll_init_d_ops, &lwi);
-        if (likely(dentry->d_op == &ll_d_ops))
-                RETURN(ll_revalidate_nd(dentry, nd));
-        else
-                RETURN(dentry->d_op == &ll_fini_d_ops ? 0 : -EINVAL);
-}
-
-struct dentry_operations ll_init_d_ops = {
-        .d_revalidate = ll_init_revalidate_nd,
-        .d_release = ll_release,
-};