Whamcloud - gitweb
LU-3409 llite: silence lockdep warning in ll_md_blocking_ast
[fs/lustre-release.git] / lustre / llite / dcache.c
index b4758e0..029e621 100644 (file)
@@ -1,6 +1,4 @@
-/* -*- mode: c; c-basic-offset: 8; indent-tabs-mode: nil; -*-
- * vim:expandtab:shiftwidth=8:tabstop=8:
- *
+/*
  * GPL HEADER START
  *
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
@@ -28,6 +26,8 @@
 /*
  * Copyright (c) 2002, 2010, Oracle and/or its affiliates. All rights reserved.
  * Use is subject to license terms.
+ *
+ * Copyright (c) 2011, 2013, Intel Corporation.
  */
 /*
  * This file is part of Lustre, http://www.lustre.org/
@@ -36,7 +36,6 @@
 
 #include <linux/fs.h>
 #include <linux/sched.h>
-#include <linux/smp_lock.h>
 #include <linux/quotaops.h>
 
 #define DEBUG_SUBSYSTEM S_LLITE
 #include <lustre_lite.h>
 #include <lustre/lustre_idl.h>
 #include <lustre_dlm.h>
-#include <lustre_mdc.h>
-//#include <lustre_ver.h>
-//#include <lustre_version.h>
 
 #include "llite_internal.h"
 
-cfs_spinlock_t ll_lookup_lock = CFS_SPIN_LOCK_UNLOCKED;
+static void free_dentry_data(struct rcu_head *head)
+{
+       struct ll_dentry_data *lld;
+
+       lld = container_of(head, struct ll_dentry_data, lld_rcu_head);
+       OBD_FREE_PTR(lld);
+}
 
 /* should NOT be called with the dcache lock, see fs/dcache.c */
 static void ll_release(struct dentry *de)
@@ -60,54 +62,67 @@ static void ll_release(struct dentry *de)
         ENTRY;
         LASSERT(de != NULL);
         lld = ll_d2d(de);
-        if (lld == NULL) { /* NFS copies the de->d_op methods (bug 4655) */
-                EXIT;
-                return;
-        }
-#ifndef HAVE_VFS_INTENT_PATCHES
+        if (lld == NULL) /* NFS copies the de->d_op methods (bug 4655) */
+                RETURN_EXIT;
+
         if (lld->lld_it) {
                 ll_intent_release(lld->lld_it);
                 OBD_FREE(lld->lld_it, sizeof(*lld->lld_it));
         }
-#endif
-        LASSERT(lld->lld_cwd_count == 0);
-        LASSERT(lld->lld_mnt_count == 0);
-        OBD_FREE(de->d_fsdata, sizeof(*lld));
+       LASSERT(lld->lld_cwd_count == 0);
+       LASSERT(lld->lld_mnt_count == 0);
+       de->d_fsdata = NULL;
+       call_rcu(&lld->lld_rcu_head, free_dentry_data);
 
-        EXIT;
+       EXIT;
 }
 
 /* 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.
+ * is marked invalid.  Returns 1 if different, 0 if the same.
  *
  * This avoids a race where ll_lookup_it() instantiates a dentry, but we get
  * an AST before calling d_revalidate_it().  The dentry still exists (marked
  * INVALID) so d_lookup() matches it, but we have no lock on it (so
  * lock_match() fails) and we spin around real_lookup(). */
+#ifdef HAVE_D_COMPARE_7ARGS
+int ll_dcompare(const struct dentry *parent, const struct inode *pinode,
+               const struct dentry *dentry, const struct inode *inode,
+               unsigned int len, const char *str, const struct qstr *name)
+#else
 int ll_dcompare(struct dentry *parent, struct qstr *d_name, struct qstr *name)
+#endif
 {
-        struct dentry *dchild;
-        ENTRY;
+#ifdef HAVE_D_COMPARE_7ARGS
+       ENTRY;
 
-        if (d_name->len != name->len)
-                RETURN(1);
+       if (len != name->len)
+               RETURN(1);
 
-        if (memcmp(d_name->name, name->name, name->len))
-                RETURN(1);
+       if (memcmp(str, name->name, len))
+               RETURN(1);
+#else
+       struct dentry *dentry;
+       ENTRY;
 
-        /* XXX: d_name must be in-dentry structure */
-        dchild = container_of(d_name, struct dentry, d_name); /* ugh */
+       if (d_name->len != name->len)
+               RETURN(1);
 
-        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));
+       if (memcmp(d_name->name, name->name, name->len))
+               RETURN(1);
 
-         /* mountpoint is always valid */
-        if (d_mountpoint(dchild))
-                RETURN(0);
+       /* XXX: d_name must be in-dentry structure */
+       dentry = container_of(d_name, struct dentry, d_name); /* ugh */
+#endif
+
+       CDEBUG(D_DENTRY, "found name %.*s(%p) flags %#x refc %d\n",
+              name->len, name->name, dentry, dentry->d_flags,
+              d_refcount(dentry));
+
+       /* mountpoint is always valid */
+       if (d_mountpoint((struct dentry *)dentry))
+               RETURN(0);
 
-        if (dchild->d_flags & DCACHE_LUSTRE_INVALID)
+       if (d_lustre_invalid(dentry))
                 RETURN(1);
 
         RETURN(0);
@@ -115,7 +130,9 @@ int ll_dcompare(struct dentry *parent, struct qstr *d_name, struct qstr *name)
 
 static inline int return_if_equal(struct ldlm_lock *lock, void *data)
 {
-        if (lock->l_flags & LDLM_FL_CANCELING)
+        if ((lock->l_flags &
+             (LDLM_FL_CANCELING | LDLM_FL_DISCARD_DATA)) ==
+            (LDLM_FL_CANCELING | LDLM_FL_DISCARD_DATA))
                 return LDLM_ITER_CONTINUE;
         return LDLM_ITER_STOP;
 }
@@ -126,8 +143,8 @@ static inline int return_if_equal(struct ldlm_lock *lock, void *data)
  *      < 0    error */
 static int find_cbdata(struct inode *inode)
 {
-        struct ll_inode_info *lli = ll_i2info(inode);
-        struct ll_sb_info *sbi = ll_i2sbi(inode);
+       struct ll_sb_info *sbi = ll_i2sbi(inode);
+       struct lov_stripe_md *lsm;
         int rc = 0;
         ENTRY;
 
@@ -137,66 +154,86 @@ static int find_cbdata(struct inode *inode)
         if (rc != 0)
                  RETURN(rc);
 
-        if (lli->lli_smd)
-                rc = obd_find_cbdata(sbi->ll_dt_exp, lli->lli_smd,
-                                     return_if_equal, NULL);
+       lsm = ccc_inode_lsm_get(inode);
+       if (lsm == NULL)
+               RETURN(rc);
 
-        RETURN(rc);
+       rc = obd_find_cbdata(sbi->ll_dt_exp, lsm, return_if_equal, NULL);
+       ccc_inode_lsm_put(inode, lsm);
+
+       RETURN(rc);
 }
 
-/* should NOT be called with the dcache lock, see fs/dcache.c */
-static int ll_ddelete(struct dentry *de)
+/**
+ * Called when last reference to a dentry is dropped and dcache wants to know
+ * whether or not it should cache it:
+ * - return 1 to delete the dentry immediately
+ * - return 0 to cache the dentry
+ * Should NOT be called with the dcache lock, see fs/dcache.c
+ */
+static int ll_ddelete(HAVE_D_DELETE_CONST struct dentry *de)
 {
-        ENTRY;
-        LASSERT(de);
+       ENTRY;
+       LASSERT(de);
 
-        CDEBUG(D_DENTRY, "%s dentry %.*s (%p, parent %p, inode %p) %s%s\n",
-               (de->d_flags & DCACHE_LUSTRE_INVALID ? "deleting" : "keeping"),
-               de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode,
-               d_unhashed(de) ? "" : "hashed,",
-               list_empty(&de->d_subdirs) ? "" : "subdirs");
+       CDEBUG(D_DENTRY, "%s dentry %.*s (%p, parent %p, inode %p) %s%s\n",
+              d_lustre_invalid((struct dentry *)de) ? "deleting" : "keeping",
+              de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode,
+              d_unhashed((struct dentry *)de) ? "" : "hashed,",
+              list_empty(&de->d_subdirs) ? "" : "subdirs");
 
-        /* if not ldlm lock for this inode, set i_nlink to 0 so that
-         * this inode can be recycled later b=20433 */
-        LASSERT(atomic_read(&de->d_count) == 0);
-        if (de->d_inode && !find_cbdata(de->d_inode))
-                de->d_inode->i_nlink = 0;
+#ifdef HAVE_DCACHE_LOCK
+       LASSERT(d_refcount(de) == 0);
+#else
+       /* kernel >= 2.6.38 last refcount is decreased after this function. */
+       LASSERT(d_refcount(de) == 1);
+#endif
 
-        if (de->d_flags & DCACHE_LUSTRE_INVALID)
-                RETURN(1);
+       /* Disable this piece of code temproarily because this is called
+        * inside dcache_lock so it's not appropriate to do lots of work
+        * here. ATTENTION: Before this piece of code enabling, LU-2487 must be
+        * resolved. */
+#if 0
+       /* if not ldlm lock for this inode, set i_nlink to 0 so that
+        * this inode can be recycled later b=20433 */
+       if (de->d_inode && !find_cbdata(de->d_inode))
+               clear_nlink(de->d_inode);
+#endif
 
-        RETURN(0);
+       if (d_lustre_invalid((struct dentry *)de))
+               RETURN(1);
+       RETURN(0);
 }
 
 static int ll_set_dd(struct dentry *de)
 {
-        ENTRY;
-        LASSERT(de != NULL);
-
-        CDEBUG(D_DENTRY, "ldd on dentry %.*s (%p) parent %p inode %p refc %d\n",
-               de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode,
-               atomic_read(&de->d_count));
-
-        if (de->d_fsdata == NULL) {
-                struct ll_dentry_data *lld;
-
-                OBD_ALLOC_PTR(lld);
-                if (likely(lld != NULL)) {
-                        lock_dentry(de);
-                        if (likely(de->d_fsdata == NULL))
-                                de->d_fsdata = lld;
-                        else
-                                OBD_FREE_PTR(lld);
-                        unlock_dentry(de);
-                } else {
-                        RETURN(-ENOMEM);
-                }
-        }
-
-        RETURN(0);
+       ENTRY;
+       LASSERT(de != NULL);
+
+       CDEBUG(D_DENTRY, "ldd on dentry %.*s (%p) parent %p inode %p refc %d\n",
+               de->d_name.len, de->d_name.name, de, de->d_parent, de->d_inode,
+               d_refcount(de));
+
+       if (de->d_fsdata == NULL) {
+               struct ll_dentry_data *lld;
+
+               OBD_ALLOC_PTR(lld);
+               if (likely(lld != NULL)) {
+                       spin_lock(&de->d_lock);
+                       if (likely(de->d_fsdata == NULL))
+                               de->d_fsdata = lld;
+                       else
+                               OBD_FREE_PTR(lld);
+                       spin_unlock(&de->d_lock);
+               } else {
+                       RETURN(-ENOMEM);
+               }
+       }
+
+       RETURN(0);
 }
 
-int ll_dops_init(struct dentry *de, int block)
+int ll_dops_init(struct dentry *de, int block, int init_sa)
 {
         struct ll_dentry_data *lld = ll_d2d(de);
         int rc = 0;
@@ -209,27 +246,42 @@ int ll_dops_init(struct dentry *de, int block)
                 lld = ll_d2d(de);
         }
 
-        if (lld != NULL)
+        if (lld != NULL && init_sa != 0)
                 lld->lld_sa_generation = 0;
 
-        de->d_op = &ll_d_ops;
+#ifdef HAVE_DCACHE_LOCK
+       de->d_op = &ll_d_ops;
+#else
+       /* kernel >= 2.6.38 d_op is set in d_alloc() */
+       LASSERT(de->d_op == &ll_d_ops);
+#endif
         return rc;
 }
 
 void ll_intent_drop_lock(struct lookup_intent *it)
 {
-        struct lustre_handle *handle;
-
         if (it->it_op && it->d.lustre.it_lock_mode) {
-                handle = (struct lustre_handle *)&it->d.lustre.it_lock_handle;
+               struct lustre_handle handle;
+
+               handle.cookie = it->d.lustre.it_lock_handle;
+
                 CDEBUG(D_DLMTRACE, "releasing lock with cookie "LPX64
-                       " from it %p\n", handle->cookie, it);
-                ldlm_lock_decref(handle, it->d.lustre.it_lock_mode);
+                       " from it %p\n", handle.cookie, it);
+                ldlm_lock_decref(&handle, it->d.lustre.it_lock_mode);
 
                 /* bug 494: intent_release may be called multiple times, from
                  * this thread and we don't want to double-decref this lock */
                 it->d.lustre.it_lock_mode = 0;
-        }
+               if (it->d.lustre.it_remote_lock_mode != 0) {
+                       handle.cookie = it->d.lustre.it_remote_lock_handle;
+
+                       CDEBUG(D_DLMTRACE, "releasing remote lock with cookie"
+                              LPX64" from it %p\n", handle.cookie, it);
+                       ldlm_lock_decref(&handle,
+                                        it->d.lustre.it_remote_lock_mode);
+                       it->d.lustre.it_remote_lock_mode = 0;
+               }
+       }
 }
 
 void ll_intent_release(struct lookup_intent *it)
@@ -238,10 +290,6 @@ void ll_intent_release(struct lookup_intent *it)
 
         CDEBUG(D_INFO, "intent %p released\n", it);
         ll_intent_drop_lock(it);
-#ifdef HAVE_VFS_INTENT_PATCHES
-        it->it_magic = 0;
-        it->it_op_release = 0;
-#endif
         /* We are still holding extra reference on a request, need to free it */
         if (it_disposition(it, DISP_ENQ_OPEN_REF))
                  ptlrpc_req_finished(it->d.lustre.it_data); /* ll_file_open */
@@ -256,75 +304,23 @@ void ll_intent_release(struct lookup_intent *it)
         EXIT;
 }
 
-/* Drop dentry if it is not used already, unhash otherwise.
-   Should be called with dcache lock held!
-   Returns: 1 if dentry was dropped, 0 if unhashed. */
-int ll_drop_dentry(struct dentry *dentry)
+void ll_invalidate_aliases(struct inode *inode)
 {
-        lock_dentry(dentry);
-        if (atomic_read(&dentry->d_count) == 0) {
-                CDEBUG(D_DENTRY, "deleting dentry %.*s (%p) parent %p "
-                       "inode %p\n", dentry->d_name.len,
-                       dentry->d_name.name, dentry, dentry->d_parent,
-                       dentry->d_inode);
-                dget_locked(dentry);
-                __d_drop(dentry);
-                unlock_dentry(dentry);
-                spin_unlock(&dcache_lock);
-                cfs_spin_unlock(&ll_lookup_lock);
-                dput(dentry);
-                cfs_spin_lock(&ll_lookup_lock);
-                spin_lock(&dcache_lock);
-                return 1;
-        }
-        /* 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);
-                RETURN (0);
-        }
-
-        if (!(dentry->d_flags & DCACHE_LUSTRE_INVALID)) {
-                CDEBUG(D_DENTRY, "unhashing dentry %.*s (%p) parent %p "
-                       "inode %p refc %d\n", dentry->d_name.len,
-                       dentry->d_name.name, dentry, dentry->d_parent,
-                       dentry->d_inode, atomic_read(&dentry->d_count));
-                /* actually we don't unhash the dentry, rather just
-                 * mark it inaccessible for to __d_lookup(). otherwise
-                 * sys_getcwd() could return -ENOENT -bzzz */
-                dentry->d_flags |= DCACHE_LUSTRE_INVALID;
-                if (!dentry->d_inode || !S_ISDIR(dentry->d_inode->i_mode))
-                        __d_drop(dentry);
-        }
-        unlock_dentry(dentry);
-        return 0;
-}
+       struct dentry *dentry;
+       struct ll_d_hlist_node *p;
+       ENTRY;
 
-void ll_unhash_aliases(struct inode *inode)
-{
-        struct list_head *tmp, *head;
-        ENTRY;
-
-        if (inode == NULL) {
-                CERROR("unexpected NULL inode, tell phil\n");
-                return;
-        }
+       LASSERT(inode != NULL);
 
-        CDEBUG(D_INODE, "marking dentries for ino %lu/%u(%p) invalid\n",
-               inode->i_ino, inode->i_generation, inode);
+       CDEBUG(D_INODE, "marking dentries for ino %lu/%u(%p) invalid\n",
+              inode->i_ino, inode->i_generation, inode);
 
-        head = &inode->i_dentry;
-        cfs_spin_lock(&ll_lookup_lock);
-        spin_lock(&dcache_lock);
-restart:
-        tmp = head;
-        while ((tmp = tmp->next) != head) {
-                struct dentry *dentry = list_entry(tmp, struct dentry, d_alias);
-
-                CDEBUG(D_DENTRY, "dentry in drop %.*s (%p) parent %p "
-                       "inode %p flags %d\n", dentry->d_name.len,
-                       dentry->d_name.name, dentry, dentry->d_parent,
-                       dentry->d_inode, dentry->d_flags);
+       ll_lock_dcache(inode);
+       ll_d_hlist_for_each_entry(dentry, p, &inode->i_dentry, d_alias) {
+               CDEBUG(D_DENTRY, "dentry in drop %.*s (%p) parent %p "
+                      "inode %p flags %d\n", dentry->d_name.len,
+                      dentry->d_name.name, dentry, dentry->d_parent,
+                      dentry->d_inode, dentry->d_flags);
 
                 if (dentry->d_name.len == 1 && dentry->d_name.name[0] == '/') {
                         CERROR("called on root (?) dentry=%p, inode=%p "
@@ -333,11 +329,9 @@ restart:
                         libcfs_debug_dumpstack(NULL);
                 }
 
-                if (ll_drop_dentry(dentry))
-                          goto restart;
-        }
-        spin_unlock(&dcache_lock);
-        cfs_spin_unlock(&ll_lookup_lock);
+               d_lustre_invalidate(dentry, 0);
+       }
+       ll_unlock_dcache(inode);
 
         EXIT;
 }
@@ -355,7 +349,7 @@ int ll_revalidate_it_finish(struct ptlrpc_request *request,
         if (it_disposition(it, DISP_LOOKUP_NEG))
                 RETURN(-ENOENT);
 
-        rc = ll_prep_inode(&de->d_inode, request, NULL);
+        rc = ll_prep_inode(&de->d_inode, request, NULL, it);
 
         RETURN(rc);
 }
@@ -371,8 +365,7 @@ 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, NULL);
+                ll_set_lock_data(sbi->ll_md_exp, inode, it, NULL);
         }
 
         /* drop lookup or getattr locks immediately */
@@ -387,20 +380,10 @@ void ll_lookup_finish_locks(struct lookup_intent *it, struct dentry *dentry)
 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,
-                         "%p has bad intent magic: %x\n",
-                         it, it->it_magic);
-        }
-#endif
 
         if (!it || it->it_op == IT_GETXATTR)
                 it = *itp = deft;
 
-#ifdef HAVE_VFS_INTENT_PATCHES
-        it->it_op_release = ll_intent_release;
-#endif
 }
 
 int ll_revalidate_it(struct dentry *de, int lookup_flags,
@@ -411,13 +394,15 @@ int ll_revalidate_it(struct dentry *de, int lookup_flags,
         struct lookup_intent lookup_it = { .it_op = IT_LOOKUP };
         struct obd_export *exp;
         struct inode *parent = de->d_parent->d_inode;
-        int rc, first = 0;
+        int rc;
 
         ENTRY;
         CDEBUG(D_VFSTRACE, "VFS Op:name=%s,intent=%s\n", de->d_name.name,
                LL_IT2STR(it));
 
         if (de->d_inode == NULL) {
+                __u64 ibits;
+
                 /* We can only use negative dentries if this is stat or lookup,
                    for opens and stuff we do need to query server. */
                 /* If there is IT_CREAT in intent op set, then we must throw
@@ -426,10 +411,11 @@ int ll_revalidate_it(struct dentry *de, int lookup_flags,
                 if (it && (it->it_op & IT_CREAT))
                         RETURN(0);
 
-                if (de->d_flags & DCACHE_LUSTRE_INVALID)
+               if (d_lustre_invalid(de))
                         RETURN(0);
 
-                rc = ll_have_md_lock(parent, MDS_INODELOCK_UPDATE, LCK_MINMODE);
+                ibits = MDS_INODELOCK_UPDATE;
+                rc = ll_have_md_lock(parent, &ibits, LCK_MINMODE);
                 GOTO(out_sa, rc);
         }
 
@@ -452,20 +438,15 @@ int ll_revalidate_it(struct dentry *de, int lookup_flags,
         ll_frob_intent(&it, &lookup_it);
         LASSERT(it);
 
-        if (it->it_op == IT_LOOKUP && !(de->d_flags & DCACHE_LUSTRE_INVALID))
-                GOTO(out_sa, rc = 1);
-
-        op_data = ll_prep_md_op_data(NULL, parent, de->d_inode,
-                                     de->d_name.name, de->d_name.len,
-                                     0, LUSTRE_OPC_ANY, NULL);
-        if (IS_ERR(op_data))
-                RETURN(PTR_ERR(op_data));
+       if (it->it_op == IT_LOOKUP && !d_lustre_invalid(de))
+                RETURN(1);
 
-        if ((it->it_op == IT_OPEN) && de->d_inode) {
-                struct inode *inode = de->d_inode;
-                struct ll_inode_info *lli = ll_i2info(inode);
-                struct obd_client_handle **och_p;
-                __u64 *och_usecount;
+       if (it->it_op == IT_OPEN) {
+               struct inode *inode = de->d_inode;
+               struct ll_inode_info *lli = ll_i2info(inode);
+               struct obd_client_handle **och_p;
+               __u64 *och_usecount;
+               __u64 ibits;
 
                 /*
                  * We used to check for MDS_INODELOCK_OPEN here, but in fact
@@ -489,12 +470,14 @@ int ll_revalidate_it(struct dentry *de, int lookup_flags,
                         och_usecount = &lli->lli_open_fd_read_count;
                 }
                 /* Check for the proper lock. */
-                if (!ll_have_md_lock(inode, MDS_INODELOCK_LOOKUP, LCK_MINMODE))
+                ibits = MDS_INODELOCK_LOOKUP;
+                if (!ll_have_md_lock(inode, &ibits, LCK_MINMODE))
                         goto do_lock;
-                cfs_down(&lli->lli_och_sem);
+               mutex_lock(&lli->lli_och_mutex);
                 if (*och_p) { /* Everything is open already, do nothing */
                         /*(*och_usecount)++;  Do not let them steal our open
                           handle from under us */
+                        SET_BUT_UNUSED(och_usecount);
                         /* XXX The code above was my original idea, but in case
                            we have the handle, but we cannot use it due to later
                            checks (e.g. O_CREAT|O_EXCL flags set), nobody
@@ -502,32 +485,36 @@ int ll_revalidate_it(struct dentry *de, int lookup_flags,
                            hope the lock won't be invalidated in between. But
                            if it would be, we'll reopen the open request to
                            MDS later during file open path */
-                        cfs_up(&lli->lli_och_sem);
-                        ll_finish_md_op_data(op_data);
+                       mutex_unlock(&lli->lli_och_mutex);
                         RETURN(1);
                 } else {
-                        cfs_up(&lli->lli_och_sem);
+                       mutex_unlock(&lli->lli_och_mutex);
                 }
         }
 
-        if (it->it_op == IT_GETATTR)
-                first = ll_statahead_enter(parent, &de, 0);
+        if (it->it_op == IT_GETATTR) {
+                rc = ll_statahead_enter(parent, &de, 0);
+                if (rc == 1)
+                        goto mark;
+                else if (rc != -EAGAIN && rc != 0)
+                        GOTO(out, rc = 0);
+        }
 
 do_lock:
-        it->it_create_mode &= ~cfs_curproc_umask();
+        op_data = ll_prep_md_op_data(NULL, parent, de->d_inode,
+                                     de->d_name.name, de->d_name.len,
+                                     0, LUSTRE_OPC_ANY, NULL);
+        if (IS_ERR(op_data))
+                RETURN(PTR_ERR(op_data));
+
+        if (!IS_POSIXACL(parent) || !exp_connect_umask(exp))
+                it->it_create_mode &= ~cfs_curproc_umask();
         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_create_mode &= ~M_CHECK_STALE;
         ll_finish_md_op_data(op_data);
-        if (it->it_op == IT_GETATTR && !first)
-                /* 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(parent, de, req == NULL ? rc : 0);
-        else if (first == -EEXIST)
-                ll_statahead_mark(parent, 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. */
@@ -560,17 +547,6 @@ revalidate_finish:
         }
         rc = 1;
 
-        /* unfortunately ll_intent_lock may cause a callback and revoke our
-         * dentry */
-        cfs_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);
-        cfs_spin_unlock(&ll_lookup_lock);
-
 out:
         /* We do not free request as it may be reused during following lookup
          * (see comment in mdc/mdc_locks.c::mdc_intent_lock()), request will
@@ -579,21 +555,43 @@ out:
         if (req != NULL && !it_disposition(it, DISP_ENQ_COMPLETE))
                 ptlrpc_req_finished(req);
         if (rc == 0) {
-                ll_unhash_aliases(de->d_inode);
-                /* done in ll_unhash_aliases()
-                   dentry->d_flags |= DCACHE_LUSTRE_INVALID; */
-        } else {
-                CDEBUG(D_DENTRY, "revalidated dentry %.*s (%p) parent %p "
-                       "inode %p refc %d\n", de->d_name.len,
-                       de->d_name.name, de, de->d_parent, de->d_inode,
-                       atomic_read(&de->d_count));
-                if (de->d_flags & DCACHE_LUSTRE_INVALID) {
-                        lock_dentry(de);
-                        de->d_flags &= ~DCACHE_LUSTRE_INVALID;
-                        unlock_dentry(de);
-                }
-                ll_lookup_finish_locks(it, de);
-        }
+               /* mdt may grant layout lock for the newly created file, so
+                * release the lock to avoid leaking */
+               ll_intent_drop_lock(it);
+               ll_invalidate_aliases(de->d_inode);
+       } else {
+               __u64 bits = 0;
+               __u64 matched_bits = 0;
+
+               CDEBUG(D_DENTRY, "revalidated dentry %.*s (%p) parent %p "
+                      "inode %p refc %d\n", de->d_name.len,
+                      de->d_name.name, de, de->d_parent, de->d_inode,
+                      d_refcount(de));
+
+               ll_set_lock_data(exp, de->d_inode, it, &bits);
+
+               /* Note: We have to match both LOOKUP and PERM lock
+                * here to make sure the dentry is valid and no one
+                * changing the permission.
+                * But if the client connects < 2.4 server, which will
+                * only grant LOOKUP lock, so we can only Match LOOKUP
+                * lock for old server */
+               if (exp_connect_flags(ll_i2mdexp(de->d_inode)) &&
+                                                       OBD_CONNECT_LVB_TYPE)
+                       matched_bits =
+                               MDS_INODELOCK_LOOKUP | MDS_INODELOCK_PERM;
+               else
+                       matched_bits = MDS_INODELOCK_LOOKUP;            
+
+               if (((bits & matched_bits) == matched_bits) &&
+                   d_lustre_invalid(de))
+                       d_lustre_revalidate(de);
+               ll_lookup_finish_locks(it, de);
+       }
+
+mark:
+        if (it != NULL && it->it_op == IT_GETATTR && rc > 0)
+                ll_statahead_mark(parent, de);
         RETURN(rc);
 
         /*
@@ -654,122 +652,49 @@ out_sa:
          * For rc == 1 case, should not return directly to prevent losing
          * statahead windows; for rc == 0 case, the "lookup" will be done later.
          */
-        if (it && it->it_op == IT_GETATTR && rc == 1) {
-                first = ll_statahead_enter(parent, &de, 0);
-                if (first >= 0)
-                        ll_statahead_exit(parent, de, 1);
-                else if (first == -EEXIST)
-                        ll_statahead_mark(parent, de);
-        }
-
-        return rc;
+        if (it != NULL && it->it_op == IT_GETATTR && rc == 1)
+                ll_statahead_enter(parent, &de, 1);
+        goto mark;
 }
 
-#if 0
-static void ll_pin(struct dentry *de, struct vfsmount *mnt, int flag)
+#ifdef HAVE_IOP_ATOMIC_OPEN
+/*
+ * Always trust cached dentries. Update statahead window if necessary.
+ */
+int ll_revalidate_nd(struct dentry *dentry, unsigned int flags)
 {
-        struct inode *inode= de->d_inode;
-        struct ll_sb_info *sbi = ll_i2sbi(inode);
-        struct ll_dentry_data *ldd = ll_d2d(de);
-        struct obd_client_handle *handle;
-        struct obd_capa *oc;
-        int rc = 0;
-        ENTRY;
-        LASSERT(ldd);
-
-        cfs_lock_kernel();
-        /* Strictly speaking this introduces an additional race: the
-         * increments should wait until the rpc has returned.
-         * However, given that at present the function is void, this
-         * issue is moot. */
-        if (flag == 1 && (++ldd->lld_mnt_count) > 1) {
-                cfs_unlock_kernel();
-                EXIT;
-                return;
-        }
-
-        if (flag == 0 && (++ldd->lld_cwd_count) > 1) {
-                cfs_unlock_kernel();
-                EXIT;
-                return;
-        }
-        cfs_unlock_kernel();
-
-        handle = (flag) ? &ldd->lld_mnt_och : &ldd->lld_cwd_och;
-        oc = ll_mdscapa_get(inode);
-        rc = obd_pin(sbi->ll_md_exp, ll_inode2fid(inode), oc, handle, flag);
-        capa_put(oc);
-        if (rc) {
-                cfs_lock_kernel();
-                memset(handle, 0, sizeof(*handle));
-                if (flag == 0)
-                        ldd->lld_cwd_count--;
-                else
-                        ldd->lld_mnt_count--;
-                cfs_unlock_kernel();
-        }
+       struct inode *parent = dentry->d_parent->d_inode;
+       int unplug = 0;
 
-        EXIT;
-        return;
-}
-
-static void ll_unpin(struct dentry *de, struct vfsmount *mnt, int flag)
-{
-        struct ll_sb_info *sbi = ll_i2sbi(de->d_inode);
-        struct ll_dentry_data *ldd = ll_d2d(de);
-        struct obd_client_handle handle;
-        int count, rc = 0;
-        ENTRY;
-        LASSERT(ldd);
-
-        cfs_lock_kernel();
-        /* Strictly speaking this introduces an additional race: the
-         * increments should wait until the rpc has returned.
-         * However, given that at present the function is void, this
-         * issue is moot. */
-        handle = (flag) ? ldd->lld_mnt_och : ldd->lld_cwd_och;
-        if (handle.och_magic != OBD_CLIENT_HANDLE_MAGIC) {
-                /* the "pin" failed */
-                cfs_unlock_kernel();
-                EXIT;
-                return;
-        }
+       ENTRY;
+       CDEBUG(D_VFSTRACE, "VFS Op:name=%s,flags=%u\n",
+              dentry->d_name.name, flags);
 
-        if (flag)
-                count = --ldd->lld_mnt_count;
-        else
-                count = --ldd->lld_cwd_count;
-        cfs_unlock_kernel();
+       if (!(flags & (LOOKUP_PARENT|LOOKUP_OPEN|LOOKUP_CREATE)) &&
+           ll_need_statahead(parent, dentry) > 0) {
+               if (flags & LOOKUP_RCU)
+                       RETURN(-ECHILD);
 
-        if (count != 0) {
-                EXIT;
-                return;
-        }
+               if (dentry->d_inode == NULL)
+                       unplug = 1;
+               do_statahead_enter(parent, &dentry, unplug);
+               ll_statahead_mark(parent, dentry);
+       }
 
-        rc = obd_unpin(sbi->ll_md_exp, &handle, flag);
-        EXIT;
-        return;
+       RETURN(1);
 }
-#endif
 
-#ifdef HAVE_VFS_INTENT_PATCHES
+#else /* !HAVE_IOP_ATOMIC_OPEN */
 int ll_revalidate_nd(struct dentry *dentry, struct nameidata *nd)
 {
         int rc;
         ENTRY;
 
-        if (nd && nd->flags & LOOKUP_LAST && !(nd->flags & LOOKUP_LINK_NOTLAST))
-                rc = ll_revalidate_it(dentry, nd->flags, &nd->intent);
-        else
-                rc = ll_revalidate_it(dentry, 0, NULL);
-
-        RETURN(rc);
-}
-#else
-int ll_revalidate_nd(struct dentry *dentry, struct nameidata *nd)
-{
-        int rc;
-        ENTRY;
+#ifndef HAVE_DCACHE_LOCK
+       /* kernel >= 2.6.38 supports rcu-walk, but lustre doesn't. */
+       if (nd->flags & LOOKUP_RCU)
+               return -ECHILD;
+#endif
 
         if (nd && !(nd->flags & (LOOKUP_CONTINUE|LOOKUP_PARENT))) {
                 struct lookup_intent *it;
@@ -789,7 +714,6 @@ int ll_revalidate_nd(struct dentry *dentry, struct nameidata *nd)
 
                 if (rc && (nd->flags & LOOKUP_OPEN) &&
                     it_disposition(it, DISP_OPEN_OPEN)) {/*Open*/
-#ifdef HAVE_FILE_IN_STRUCT_INTENT
 // XXX Code duplication with ll_lookup_nd
                         if (S_ISFIFO(dentry->d_inode->i_mode)) {
                                 // We cannot call open here as it would
@@ -798,25 +722,13 @@ int ll_revalidate_nd(struct dentry *dentry, struct nameidata *nd)
                                                (struct ptlrpc_request *)
                                                   it->d.lustre.it_data);
                         } else {
-#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 */
-                                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
+                               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
-                        ll_release_openhandle(dentry, it);
-#endif /* HAVE_FILE_IN_STRUCT_INTENT */
                 }
                 if (!rc && (nd->flags & LOOKUP_CREATE) &&
                     it_disposition(it, DISP_OPEN_CREATE)) {
@@ -840,14 +752,14 @@ out_it:
 
         RETURN(rc);
 }
-#endif
+#endif /* HAVE_IOP_ATOMIC_OPEN */
 
 void ll_d_iput(struct dentry *de, struct inode *inode)
 {
-        LASSERT(inode);
-        if (!find_cbdata(inode))
-                inode->i_nlink = 0;
-        iput(inode);
+       LASSERT(inode);
+       if (!find_cbdata(inode))
+               clear_nlink(inode);
+       iput(inode);
 }
 
 struct dentry_operations ll_d_ops = {
@@ -856,8 +768,4 @@ struct dentry_operations ll_d_ops = {
         .d_delete  = ll_ddelete,
         .d_iput    = ll_d_iput,
         .d_compare = ll_dcompare,
-#if 0
-        .d_pin = ll_pin,
-        .d_unpin = ll_unpin,
-#endif
 };