Whamcloud - gitweb
not revalidate dentry if it root dentry.
[fs/lustre-release.git] / lustre / llite / dcache.c
index 8b30963..4f5803b 100644 (file)
 
 #define DEBUG_SUBSYSTEM S_LLITE
 
-#include <linux/obd_support.h>
-#include <linux/lustre_lite.h>
-#include <linux/lustre_idl.h>
-#include <linux/lustre_dlm.h>
+#include <obd_support.h>
+#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"
 
@@ -40,19 +43,83 @@ static void ll_release(struct dentry *de)
         ENTRY;
         LASSERT(de != NULL);
         lld = ll_d2d(de);
-        LASSERT(lld != NULL);
+        if (lld == NULL) { /* NFS copies the de->d_op methods (bug 4655) */
+                EXIT;
+                return;
+        }
+#ifndef LUSTRE_KERNEL_VERSION
+        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(struct ll_dentry_data));
+        OBD_FREE(de->d_fsdata, sizeof(*lld));
 
         EXIT;
 }
 
+#ifdef LUSTRE_KERNEL_VERSION
+/* 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.
+ *
+ * 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(). */
+int ll_dcompare(struct dentry *parent, struct qstr *d_name, struct qstr *name)
+{
+        struct dentry *dchild;
+        ENTRY;
+
+        if (d_name->len != name->len)
+                RETURN(1);
+
+        if (memcmp(d_name->name, name->name, name->len))
+                RETURN(1);
+
+        /* 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);
+                RETURN(1);
+        }
+
+        RETURN(0);
+}
+#endif
+
+/* should NOT be called with the dcache lock, see fs/dcache.c */
+static int ll_ddelete(struct dentry *de)
+{
+        ENTRY;
+        LASSERT(de);
+#ifndef DCACHE_LUSTRE_INVALID
+#define DCACHE_LUSTRE_INVALID 0
+#endif
+
+        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");
+#if DCACHE_LUSTRE_INVALID == 0
+#undef DCACHE_LUSTRE_INVALID
+#endif
+
+        RETURN(0);
+}
+
 void 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));
         lock_kernel();
         if (de->d_fsdata == NULL) {
                 OBD_ALLOC(de->d_fsdata, sizeof(struct ll_dentry_data));
@@ -82,70 +149,152 @@ void ll_intent_release(struct lookup_intent *it)
 {
         ENTRY;
 
+        CDEBUG(D_INFO, "intent %p released\n", it);
         ll_intent_drop_lock(it);
+#ifdef LUSTRE_KERNEL_VERSION
         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)) /* open req for llfile_open*/
+                ptlrpc_req_finished(it->d.lustre.it_data);
+        if (it_disposition(it, DISP_ENQ_CREATE_REF)) /* create rec */
+                ptlrpc_req_finished(it->d.lustre.it_data);
+        if (it_disposition(it, DISP_ENQ_COMPLETE)) /* saved req from revalidate
+                                                    * to lookup */
+                ptlrpc_req_finished(it->d.lustre.it_data);
+
         it->d.lustre.it_disposition = 0;
         it->d.lustre.it_data = NULL;
         EXIT;
 }
 
-void ll_unhash_aliases(struct inode *inode)
+/* 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)
 {
-       struct list_head *tmp, *head;
-        struct ll_sb_info *sbi;
-        ENTRY;
+        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);
+                dput(dentry);
+                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);
+        }
 
-        sbi = ll_i2sbi(inode);
+#ifdef LUSTRE_KERNEL_VERSION
+        if (!(dentry->d_flags & DCACHE_LUSTRE_INVALID)) {
+#else
+        if (!d_unhashed(dentry)) {
+#endif
+                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 */
+#ifdef LUSTRE_KERNEL_VERSION
+                dentry->d_flags |= DCACHE_LUSTRE_INVALID;
+#if (LINUX_VERSION_CODE < KERNEL_VERSION(2,5,0))
+                __d_drop(dentry);
+                if (dentry->d_inode) {
+                        /* Put positive dentries to orphan list */
+                        list_add(&dentry->d_hash,
+                                 &ll_i2sbi(dentry->d_inode)->ll_orphan_dentry_list);
+                }
+#else
+                if (!dentry->d_inode || !S_ISDIR(dentry->d_inode->i_mode))
+                        __d_drop(dentry);
+#endif
+#else
+                if (!dentry->d_inode || !S_ISDIR(dentry->d_inode->i_mode))
+                        __d_drop(dentry);
+#endif
 
-        CDEBUG(D_INODE, "marking dentries for ino %lu/%u(%p) invalid\n",
-               inode->i_ino, inode->i_generation, inode);
+        }
+        unlock_dentry(dentry);
+        return 0;
+}
+
+void ll_unhash_aliases(struct inode *inode)
+{
+        struct list_head *tmp, *head;
+        ENTRY;
 
         if (inode == NULL) {
                 CERROR("unexpected NULL inode, tell phil\n");
                 return;
         }
+
+        CDEBUG(D_INODE, "marking dentries for ino %lu/%u(%p) invalid\n",
+               inode->i_ino, inode->i_generation, inode);
+
         head = &inode->i_dentry;
+        spin_lock(&dcache_lock);
 restart:
-       spin_lock(&dcache_lock);
-       tmp = head;
-       while ((tmp = tmp->next) != head) {
-               struct dentry *dentry = list_entry(tmp, struct dentry, d_alias);
-               if (!atomic_read(&dentry->d_count)) {
-                       dget_locked(dentry);
-                       __d_drop(dentry);
-                       spin_unlock(&dcache_lock);
-                       dput(dentry);
-                       goto restart;
-               } else {
-                        hlist_del_init(&dentry->d_hash);
-                        dentry->d_flags |= DCACHE_LUSTRE_INVALID;
-                        hlist_add_head(&dentry->d_hash, 
-                                       &sbi->ll_orphan_dentry_list);
+        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);
+
+                if (dentry->d_name.len == 1 && dentry->d_name.name[0] == '/') {
+                        CERROR("called on root (?) dentry=%p, inode=%p "
+                               "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;
                 }
-       }
-       spin_unlock(&dcache_lock);
+
+                if (ll_drop_dentry(dentry))
+                          goto restart;
+        }
+        spin_unlock(&dcache_lock);
         EXIT;
 }
 
-extern struct dentry *ll_find_alias(struct inode *, struct dentry *);
-
-static int revalidate_it_finish(struct ptlrpc_request *request, int offset,
-                                struct lookup_intent *it,
-                                struct dentry *de)
+int ll_revalidate_it_finish(struct ptlrpc_request *request,
+                            int offset, struct lookup_intent *it,
+                            struct dentry *de)
 {
-        struct ll_sb_info *sbi;
         int rc = 0;
         ENTRY;
 
         if (!request)
                 RETURN(0);
 
-        if (it_disposition(it, DISP_LOOKUP_NEG))
+        if (it_disposition(it, DISP_LOOKUP_NEG)) 
                 RETURN(-ENOENT);
 
-        sbi = ll_i2sbi(de->d_inode);
-        rc = ll_prep_inode(sbi->ll_osc_exp, &de->d_inode, request, offset,NULL);
+        rc = ll_prep_inode(&de->d_inode,
+                           request, offset, NULL);
 
         RETURN(rc);
 }
@@ -157,77 +306,175 @@ void ll_lookup_finish_locks(struct lookup_intent *it, struct dentry *dentry)
 
         if (it->d.lustre.it_lock_mode && dentry->d_inode != NULL) {
                 struct inode *inode = dentry->d_inode;
+                struct ll_sb_info *sbi = ll_i2sbi(dentry->d_inode);
+
                 CDEBUG(D_DLMTRACE, "setting l_data to inode %p (%lu/%u)\n",
                        inode, inode->i_ino, inode->i_generation);
-                mdc_set_lock_data(&it->d.lustre.it_lock_handle, inode);
+                md_set_lock_data(sbi->ll_md_exp, &it->d.lustre.it_lock_handle,
+                                 inode);
         }
 
         /* drop lookup or getattr locks immediately */
-        if (it->it_op == IT_LOOKUP || it->it_op == IT_GETATTR)
+        if (it->it_op == IT_LOOKUP || it->it_op == IT_GETATTR) {
+#if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
+                /* on 2.6 there are situation when several lookups and
+                 * revalidations may be requested during single operation.
+                 * therefore, we don't release intent here -bzzz */
+                ll_intent_drop_lock(it);
+#else
                 ll_intent_release(it);
+#endif
+        }
 }
 
 void ll_frob_intent(struct lookup_intent **itp, struct lookup_intent *deft)
 {
         struct lookup_intent *it = *itp;
-#if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
-        if (it && it->it_magic != INTENT_MAGIC) { 
-                CERROR("WARNING: uninitialized intent\n");
-                LBUG();
+#if defined(LUSTRE_KERNEL_VERSION)&&(LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
+        if (it) {
+                LASSERTF(it->it_magic == INTENT_MAGIC, 
+                         "%p has bad intent magic: %x\n",
+                         it, it->it_magic);
         }
-        if (it && (it->it_op == IT_GETATTR || it->it_op == 0))
-                it->it_op = IT_LOOKUP;
-        
 #endif
+
         if (!it || it->it_op == IT_GETXATTR)
                 it = *itp = deft;
 
+#ifdef LUSTRE_KERNEL_VERSION
         it->it_op_release = ll_intent_release;
+#endif
 }
 
-int ll_revalidate_it(struct dentry *de, int flags, struct lookup_intent *it)
+int ll_revalidate_it(struct dentry *de, int lookup_flags,
+                     struct lookup_intent *it)
 {
         int rc;
-        struct ll_fid pfid, cfid;
-        struct it_cb_data icbd;
-        struct ll_uctxt ctxt;
+        struct md_op_data *op_data;
         struct ptlrpc_request *req = NULL;
         struct lookup_intent lookup_it = { .it_op = IT_LOOKUP };
         struct obd_export *exp;
+        struct inode *parent;
 
         ENTRY;
         CDEBUG(D_VFSTRACE, "VFS Op:name=%s,intent=%s\n", de->d_name.name,
                LL_IT2STR(it));
 
-        /* Cached negative dentries are unsafe for now - look them up again */
-        if (de->d_inode == NULL)
-                RETURN(0);
+        if (de->d_inode == NULL) {
+                /* 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
+                   away this negative dentry and actually do the request to
+                   kernel to create whatever needs to be created (if possible)*/
+                if (it && (it->it_op & IT_CREAT))
+                        RETURN(0);
+
+#ifdef LUSTRE_KERNEL_VERSION
+                if (de->d_flags & DCACHE_LUSTRE_INVALID)
+                        RETURN(0);
+#endif
 
-        exp = ll_i2mdcexp(de->d_inode);
-        ll_inode2fid(&pfid, de->d_parent->d_inode);
-        ll_inode2fid(&cfid, de->d_inode);
-        icbd.icbd_parent = de->d_parent->d_inode;
-        icbd.icbd_childp = &de;
+                rc = ll_have_md_lock(de->d_parent->d_inode, 
+                                     MDS_INODELOCK_UPDATE);
+        
+                RETURN(rc);
+        }
 
-        /* 
-         * never execute intents for mount points
-         * - attrs will be fixed up in ll_revalidate_inode
-         */
+        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))
                 RETURN(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)
+                RETURN(1);
+
+        OBD_FAIL_TIMEOUT(OBD_FAIL_MDC_REVALIDATE_PAUSE, 5);
         ll_frob_intent(&it, &lookup_it);
         LASSERT(it);
 
-        ll_i2uctxt(&ctxt, de->d_parent->d_inode, de->d_inode);
+        parent = de->d_parent->d_inode;
+
+        if (it->it_op & IT_CREAT) {
+                op_data = ll_prep_md_op_data(NULL, parent, NULL,
+                                             de->d_name.name, de->d_name.len,
+                                             0, LUSTRE_OPC_CREATE, NULL);
+        } else {
+                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_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;
+                
+                /*
+                 * We used to check for MDS_INODELOCK_OPEN here, but in fact
+                 * just having LOOKUP lock is enough to justify inode is the
+                 * same. And if inode is the same and we have suitable
+                 * openhandle, then there is no point in doing another OPEN RPC
+                 * just to throw away newly received openhandle.  There are no
+                 * security implications too, if file owner or access mode is
+                 * change, LOOKUP lock is revoked.
+                 */
+
+
+                if (it->it_flags & FMODE_WRITE) {
+                        och_p = &lli->lli_mds_write_och;
+                        och_usecount = &lli->lli_open_fd_write_count;
+                } else if (it->it_flags & FMODE_EXEC) {
+                        och_p = &lli->lli_mds_exec_och;
+                        och_usecount = &lli->lli_open_fd_exec_count;
+                } else {
+                        och_p = &lli->lli_mds_read_och;
+                        och_usecount = &lli->lli_open_fd_read_count;
+                }
+                /* Check for the proper lock. */
+                if (!ll_have_md_lock(inode, MDS_INODELOCK_LOOKUP))
+                        goto do_lock;
+                down(&lli->lli_och_sem);
+                if (*och_p) { /* Everything is open already, do nothing */
+                        /*(*och_usecount)++;  Do not let them steal our open
+                          handle from under us */
+                        /* 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
+                           would decrement counter increased here. So we just
+                           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 */
+                        up(&lli->lli_och_sem);
+                        ll_finish_md_op_data(op_data);
+                        RETURN(1);
+                } else {
+                        up(&lli->lli_och_sem);
+                }
+        }
 
-        rc = mdc_intent_lock(exp, &ctxt, &pfid, de->d_name.name, de->d_name.len,
-                             NULL, 0,
-                             &cfid, it, flags, &req, ll_mdc_blocking_ast);
-        /* If req is NULL, then mdc_intent_lock only tried to do a lock match;
+do_lock:
+        it->it_create_mode &= ~current->fs->umask;
+        it->it_flags |= O_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;
+        ll_finish_md_op_data(op_data);
+        /* 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. */
-        if (req == NULL && rc >= 0)
+        if (req == NULL && rc >= 0) {
+                if (!rc)
+                        goto do_lookup;
                 GOTO(out, rc);
+        }
 
         if (rc < 0) {
                 if (rc != -ESTALE) {
@@ -237,31 +484,113 @@ int ll_revalidate_it(struct dentry *de, int flags, struct lookup_intent *it)
                 GOTO(out, rc = 0);
         }
 
-        rc = revalidate_it_finish(req, 1, it, de);
+revalidate_finish:
+        rc = ll_revalidate_it_finish(req, DLM_REPLY_REC_OFF, it, de);
         if (rc != 0) {
-                ll_intent_release(it);
+                if (rc != -ESTALE && rc != -ENOENT)
+                        ll_intent_release(it);
                 GOTO(out, rc = 0);
         }
+
+        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);
+        }
         rc = 1;
 
         /* unfortunately ll_intent_lock may cause a callback and revoke our
-           dentry */
+         * dentry */
         spin_lock(&dcache_lock);
-        hlist_del_init(&de->d_hash);
+        lock_dentry(de);
+        __d_drop(de);
+        unlock_dentry(de);
         __d_rehash(de, 0);
         spin_unlock(&dcache_lock);
 
- out:
-        if (req != NULL && rc == 1)
+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
+         * be freed in ll_lookup_it or in ll_intent_release. But if
+         * request was not completed, we need to free it. (bug 5154, 9903) */
+        if (req != NULL && !it_disposition(it, DISP_ENQ_COMPLETE))
                 ptlrpc_req_finished(req);
-        if (rc == 0) { 
+        if (rc == 0) {
+#ifdef LUSTRE_KERNEL_VERSION
                 ll_unhash_aliases(de->d_inode);
-                de->d_flags |= DCACHE_LUSTRE_INVALID;
-        } else { 
+                /* done in ll_unhash_aliases()
+                   dentry->d_flags |= DCACHE_LUSTRE_INVALID; */
+#else
+                /* We do not want d_invalidate to kill all child dentries too */
+                d_drop(de);
+#endif
+        } 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));
                 ll_lookup_finish_locks(it, de);
+#ifdef LUSTRE_KERNEL_VERSION
+                lock_dentry(de);
                 de->d_flags &= ~DCACHE_LUSTRE_INVALID;
+                unlock_dentry(de);
+#endif
         }
         RETURN(rc);
+        
+        /*
+         * This part is here to combat evil-evil race in real_lookup on 2.6
+         * kernels.  The race details are: We enter do_lookup() looking for some
+         * name, there is nothing in dcache for this name yet and d_lookup()
+         * returns NULL.  We proceed to real_lookup(), and while we do this,
+         * another process does open on the same file we looking up (most simple
+         * reproducer), open succeeds and the dentry is added. Now back to
+         * us. In real_lookup() we do d_lookup() again and suddenly find the
+         * dentry, so we call d_revalidate on it, but there is no lock, so
+         * without this code we would return 0, but unpatched real_lookup just
+         * returns -ENOENT in such a case instead of retrying the lookup. Once
+         * this is dealt with in real_lookup(), all of this ugly mess can go and
+         * we can just check locks in ->d_revalidate without doing any RPCs
+         * ever.
+         */
+do_lookup:
+        if (it != &lookup_it) {
+                /* MDS_INODELOCK_UPDATE needed for IT_GETATTR case. */
+                if (it->it_op == IT_GETATTR)
+                        lookup_it.it_op = IT_GETATTR;
+                ll_lookup_finish_locks(it, de);
+                it = &lookup_it;
+        }
+        
+        /* Do real lookup here. */
+        op_data = ll_prep_md_op_data(NULL, parent, NULL, de->d_name.name,
+                                     de->d_name.len, 0, (it->it_op & IT_CREAT ?
+                                                         LUSTRE_OPC_CREATE :
+                                                         LUSTRE_OPC_ANY), NULL);
+        if (IS_ERR(op_data))
+                RETURN(PTR_ERR(op_data));
+
+        rc = md_intent_lock(exp, op_data, NULL, 0,  it, 0, &req,
+                            ll_md_blocking_ast, 0);
+        if (rc >= 0) {
+                struct mdt_body *mdt_body = lustre_msg_buf(req->rq_repmsg,
+                                                           DLM_REPLY_REC_OFF,
+                                                           sizeof(*mdt_body));
+                struct lu_fid fid = {.f_seq = 0, .f_oid = 0, .f_ver = 0};
+                
+                if (de->d_inode)
+                        fid = *ll_inode2fid(de->d_inode);
+
+                /* see if we got same inode, if not - return error */
+                if (lu_fid_eq(&fid, &mdt_body->fid1)) {
+                        ll_finish_md_op_data(op_data);
+                        op_data = NULL;
+                        goto revalidate_finish;
+                }
+                ll_intent_release(it);
+        }
+        ll_finish_md_op_data(op_data);
+        GOTO(out, rc = 0);
 }
 
 /*static*/ void ll_pin(struct dentry *de, struct vfsmount *mnt, int flag)
@@ -270,6 +599,7 @@ int ll_revalidate_it(struct dentry *de, int flags, struct lookup_intent *it)
         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);
@@ -293,9 +623,9 @@ int ll_revalidate_it(struct dentry *de, int flags, struct lookup_intent *it)
         unlock_kernel();
 
         handle = (flag) ? &ldd->lld_mnt_och : &ldd->lld_cwd_och;
-        rc = obd_pin(sbi->ll_mdc_exp, inode->i_ino, inode->i_generation,
-                     inode->i_mode & S_IFMT, handle, flag);
-
+        oc = ll_mdscapa_get(inode);
+        rc = obd_pin(sbi->ll_md_exp, ll_inode2fid(inode), oc, handle, flag);
+        capa_put(oc);
         if (rc) {
                 lock_kernel();
                 memset(handle, 0, sizeof(*handle));
@@ -343,25 +673,96 @@ int ll_revalidate_it(struct dentry *de, int flags, struct lookup_intent *it)
                 return;
         }
 
-        rc = obd_unpin(sbi->ll_mdc_exp, &handle, flag);
+        rc = obd_unpin(sbi->ll_md_exp, &handle, flag);
         EXIT;
         return;
 }
 
 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
+#ifdef LUSTRE_KERNEL_VERSION
 static int ll_revalidate_nd(struct dentry *dentry, struct nameidata *nd)
 {
         int rc;
         ENTRY;
 
-        if (nd->flags & LOOKUP_LAST && !(nd->flags & LOOKUP_LINK_NOTLAST))
+        if (nd && nd->flags & LOOKUP_LAST && !(nd->flags & LOOKUP_LINK_NOTLAST))
                 rc = ll_revalidate_it(dentry, nd->flags, &nd->intent);
-        else 
+        else
+                rc = ll_revalidate_it(dentry, 0, NULL);
+
+        RETURN(rc);
+}
+#else
+int ll_revalidate_nd(struct dentry *dentry, struct nameidata *nd)
+{
+        int rc;
+        ENTRY;
+
+        if (nd && !(nd->flags & (LOOKUP_CONTINUE|LOOKUP_PARENT))) {
+                struct lookup_intent *it;
+                it = ll_convert_intent(&nd->intent.open, nd->flags);
+                if (IS_ERR(it))
+                        RETURN(0);
+                if (it->it_op == (IT_OPEN|IT_CREAT))
+                        if (nd->intent.open.flags & O_EXCL) {
+                                CDEBUG(D_VFSTRACE, "create O_EXCL, returning 0\n");
+                                rc = 0;
+                                goto out_it;
+                        }
+
+                rc = ll_revalidate_it(dentry, nd->flags, it);
+
+                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
+                                // deadlock.
+                                ptlrpc_req_finished(
+                                               (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;
+#endif
+                        }
+#else
+                        ll_release_openhandle(dentry, it);
+#endif /* HAVE_FILE_IN_STRUCT_INTENT */
+                }
+                if (!rc && (nd->flags & LOOKUP_CREATE) &&
+                    it_disposition(it, DISP_OPEN_CREATE)) {
+                        /* We created something but we may only return
+                         * negative dentry here, so save request in dentry,
+                         * if lookup will be called later on, it will
+                         * pick the request, otherwise it would be freed
+                         * with dentry */
+                        ll_d2d(dentry)->lld_it = it;
+                        it = NULL; /* avoid freeing */
+                }
+                        
+out_it:
+                if (it) {
+                        ll_intent_release(it);
+                        OBD_FREE(it, sizeof(*it));
+                }
+        } else {
                 rc = ll_revalidate_it(dentry, 0, NULL);
+        }
 
         RETURN(rc);
 }
 #endif
+#endif
 
 struct dentry_operations ll_d_ops = {
 #if (LINUX_VERSION_CODE > KERNEL_VERSION(2,5,0))
@@ -370,6 +771,10 @@ struct dentry_operations ll_d_ops = {
         .d_revalidate_it = ll_revalidate_it,
 #endif
         .d_release = ll_release,
+        .d_delete = ll_ddelete,
+#ifdef LUSTRE_KERNEL_VERSION
+        .d_compare = ll_dcompare,
+#endif
 #if 0
         .d_pin = ll_pin,
         .d_unpin = ll_unpin,