Whamcloud - gitweb
Branch HEAD
[fs/lustre-release.git] / lustre / llite / dcache.c
index 42bce6f..0f7c7e6 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>
@@ -36,6 +51,8 @@
 
 #include "llite_internal.h"
 
+spinlock_t ll_lookup_lock = SPIN_LOCK_UNLOCKED;
+
 /* should NOT be called with the dcache lock, see fs/dcache.c */
 static void ll_release(struct dentry *de)
 {
@@ -120,11 +137,20 @@ void ll_set_dd(struct dentry *de)
         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));
+                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);
+                }
         }
-        unlock_kernel();
 
         EXIT;
 }
@@ -156,8 +182,8 @@ void ll_intent_release(struct lookup_intent *it)
         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_OPEN_REF))
+                 ptlrpc_req_finished(it->d.lustre.it_data); /* ll_file_open */
         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
@@ -184,11 +210,13 @@ 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);
@@ -232,6 +260,7 @@ 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;
@@ -265,6 +294,8 @@ restart:
                           goto restart;
         }
         spin_unlock(&dcache_lock);
+        spin_unlock(&ll_lookup_lock);
+
         EXIT;
 }
 
@@ -278,7 +309,7 @@ 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);
@@ -315,7 +346,7 @@ 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);
         }
@@ -332,12 +363,12 @@ void ll_frob_intent(struct lookup_intent **itp, struct lookup_intent *deft)
 int ll_revalidate_it(struct dentry *de, int lookup_flags,
                      struct lookup_intent *it)
 {
-        int rc;
         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;
+        int rc, first = 0;
 
         ENTRY;
         CDEBUG(D_VFSTRACE, "VFS Op:name=%s,intent=%s\n", de->d_name.name,
@@ -359,8 +390,7 @@ int ll_revalidate_it(struct dentry *de, int lookup_flags,
 
                 rc = ll_have_md_lock(de->d_parent->d_inode,
                                      MDS_INODELOCK_UPDATE);
-        
-                RETURN(rc);
+                GOTO(out_sa, rc);
         }
 
         exp = ll_i2mdexp(de->d_inode);
@@ -368,28 +398,21 @@ int ll_revalidate_it(struct dentry *de, int lookup_flags,
         /* Never execute intents for mount points.
          * Attributes will be fixed up in ll_inode_revalidate_it */
         if (d_mountpoint(de))
-                RETURN(1);
+                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)
-                RETURN(1);
+                GOTO(out_sa, rc = 1);
 
         OBD_FAIL_TIMEOUT(OBD_FAIL_MDC_REVALIDATE_PAUSE, 5);
         ll_frob_intent(&it, &lookup_it);
         LASSERT(it);
-
         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);
-        }
+        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));
 
@@ -398,7 +421,7 @@ int ll_revalidate_it(struct dentry *de, int lookup_flags,
                 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
@@ -442,6 +465,9 @@ int ll_revalidate_it(struct dentry *de, int lookup_flags,
                 }
         }
 
+        if (it->it_op == IT_GETATTR)
+                first = ll_statahead_enter(de->d_parent->d_inode, &de, 0);
+
 do_lock:
         it->it_create_mode &= ~current->fs->umask;
         it->it_flags |= O_CHECK_STALE;
@@ -450,6 +476,14 @@ do_lock:
                             &req, ll_md_blocking_ast, 0);
         it->it_flags &= ~O_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(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. */
         if (req == NULL && rc >= 0) {
@@ -462,6 +496,13 @@ do_lock:
                 if (rc != -ESTALE) {
                         CDEBUG(D_INFO, "ll_intent_lock: rc %d : it->it_status "
                                "%d\n", rc, it->d.lustre.it_status);
+                } else {
+#ifndef HAVE_VFS_INTENT_PATCHES
+                        if (it_disposition(it, DISP_OPEN_OPEN) &&
+                            !it_open_error(DISP_OPEN_OPEN, it))
+                                /* server have valid open - close file first*/
+                                ll_release_openhandle(de, it);
+#endif
                 }
                 GOTO(out, rc = 0);
         }
@@ -474,8 +515,8 @@ 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);
         }
@@ -483,12 +524,14 @@ 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(de, 0);
+        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
@@ -558,7 +601,7 @@ do_lookup:
                 struct mdt_body *mdt_body;
                 struct lu_fid fid = {.f_seq = 0, .f_oid = 0, .f_ver = 0};
                 mdt_body = req_capsule_server_get(&req->rq_pill, &RMF_MDT_BODY);
-                
+
                 if (de->d_inode)
                         fid = *ll_inode2fid(de->d_inode);
 
@@ -572,6 +615,21 @@ do_lookup:
         }
         ll_finish_md_op_data(op_data);
         GOTO(out, rc = 0);
+
+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(de->d_parent->d_inode, &de, 0);
+                if (!first)
+                        ll_statahead_exit(de, 1);
+                else if (first == -EEXIST)
+                        ll_statahead_mark(de);
+        }
+
+        return rc;
 }
 
 /*static*/ void ll_pin(struct dentry *de, struct vfsmount *mnt, int flag)
@@ -703,16 +761,20 @@ 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