ocfs2: Remove i_generation from inode lock names

OCFS2 puts inode meta data in the "lock value block" provided by the DLM.
Typically, i_generation is encoded in the lock name so that a deleted inode
on and a new one in the same block don't share the same lvb.

Unfortunately, that scheme means that the read in ocfs2_read_locked_inode()
is potentially thrown away as soon as the meta data lock is taken - we
cannot encode the lock name without first knowing i_generation, which
requires a disk read.

This patch encodes i_generation in the inode meta data lvb, and removes the
value from the inode meta data lock name. This way, the read can be covered
by a lock, and at the same time we can distinguish between an up to date and
a stale LVB.

This will help cold-cache stat(2) performance in particular.

Since this patch changes the protocol version, we take the opportunity to do
a minor re-organization of two of the LVB fields.

Signed-off-by: Mark Fasheh <mark.fasheh@oracle.com>

10 files changed, 170 insertions, 53 deletions

diff --git a/fs/ocfs2/cluster/tcp_internal.h b/fs/ocfs2/cluster/tcp_internal.h
index da42b515cd1..4b46aac7d24 100644
--- a/fs/ocfs2/cluster/tcp_internal.h
+++ b/fs/ocfs2/cluster/tcp_internal.h
@@ -44,6 +44,9 @@
  * locking semantics of the file system using the protocol.  It should 
  * be somewhere else, I'm sure, but right now it isn't.
  *
+ * New in version 4:
+ *      - Remove i_generation from lock names for better stat performance.
+ *
  * New in version 3:
  *      - Replace dentry votes with a cluster lock
  *
@@ -51,7 +54,7 @@
  *      - full 64 bit i_size in the metadata lock lvbs
  *      - introduction of "rw" lock and pushing meta/data locking down
  */
-#define O2NET_PROTOCOL_VERSION 3ULL
+#define O2NET_PROTOCOL_VERSION 4ULL
 struct o2net_handshake {
         __be64  protocol_version;
         __be64  connector_id;
diff --git a/fs/ocfs2/dlmglue.c b/fs/ocfs2/dlmglue.c
index 6cd84dffbbf..ecb3cba2281 100644
--- a/fs/ocfs2/dlmglue.c
+++ b/fs/ocfs2/dlmglue.c
@@ -320,6 +320,7 @@ void ocfs2_lock_res_init_once(struct ocfs2_lock_res *res)
 
 void ocfs2_inode_lock_res_init(struct ocfs2_lock_res *res,
                                enum ocfs2_lock_type type,
+                               unsigned int generation,
                                struct inode *inode)
 {
         struct ocfs2_lock_res_ops *ops;
@@ -341,7 +342,7 @@ void ocfs2_inode_lock_res_init(struct ocfs2_lock_res *res,
         };
 
         ocfs2_build_lock_name(type, OCFS2_I(inode)->ip_blkno,
-                              inode->i_generation, res->l_name);
+                              generation, res->l_name);
         ocfs2_lock_res_init_common(OCFS2_SB(inode->i_sb), res, type, ops, inode);
 }
 
@@ -1173,17 +1174,19 @@ static void ocfs2_cluster_unlock(struct ocfs2_super *osb,
 
 int ocfs2_create_new_lock(struct ocfs2_super *osb,
                           struct ocfs2_lock_res *lockres,
-                          int ex)
+                          int ex,
+                          int local)
 {
         int level =  ex ? LKM_EXMODE : LKM_PRMODE;
         unsigned long flags;
+        int lkm_flags = local ? LKM_LOCAL : 0;
 
         spin_lock_irqsave(&lockres->l_lock, flags);
         BUG_ON(lockres->l_flags & OCFS2_LOCK_ATTACHED);
         lockres_or_flags(lockres, OCFS2_LOCK_LOCAL);
         spin_unlock_irqrestore(&lockres->l_lock, flags);
 
-        return ocfs2_lock_create(osb, lockres, level, LKM_LOCAL);
+        return ocfs2_lock_create(osb, lockres, level, lkm_flags);
 }
 
 /* Grants us an EX lock on the data and metadata resources, skipping
@@ -1212,19 +1215,23 @@ int ocfs2_create_new_inode_locks(struct inode *inode)
          * on a resource which has an invalid one -- we'll set it
          * valid when we release the EX. */
 
-        ret = ocfs2_create_new_lock(osb, &OCFS2_I(inode)->ip_rw_lockres, 1);
+        ret = ocfs2_create_new_lock(osb, &OCFS2_I(inode)->ip_rw_lockres, 1, 1);
         if (ret) {
                 mlog_errno(ret);
                 goto bail;
         }
 
-        ret = ocfs2_create_new_lock(osb, &OCFS2_I(inode)->ip_meta_lockres, 1);
+        /*
+         * We don't want to use LKM_LOCAL on a meta data lock as they
+         * don't use a generation in their lock names.
+         */
+        ret = ocfs2_create_new_lock(osb, &OCFS2_I(inode)->ip_meta_lockres, 1, 0);
         if (ret) {
                 mlog_errno(ret);
                 goto bail;
         }
 
-        ret = ocfs2_create_new_lock(osb, &OCFS2_I(inode)->ip_data_lockres, 1);
+        ret = ocfs2_create_new_lock(osb, &OCFS2_I(inode)->ip_data_lockres, 1, 1);
         if (ret) {
                 mlog_errno(ret);
                 goto bail;
@@ -1413,6 +1420,16 @@ static void __ocfs2_stuff_meta_lvb(struct inode *inode)
 
         lvb = (struct ocfs2_meta_lvb *) lockres->l_lksb.lvb;
 
+        /*
+         * Invalidate the LVB of a deleted inode - this way other
+         * nodes are forced to go to disk and discover the new inode
+         * status.
+         */
+        if (oi->ip_flags & OCFS2_INODE_DELETED) {
+                lvb->lvb_version = 0;
+                goto out;
+        }
+
         lvb->lvb_version   = OCFS2_LVB_VERSION;
         lvb->lvb_isize     = cpu_to_be64(i_size_read(inode));
         lvb->lvb_iclusters = cpu_to_be32(oi->ip_clusters);
@@ -1429,6 +1446,7 @@ static void __ocfs2_stuff_meta_lvb(struct inode *inode)
         lvb->lvb_iattr    = cpu_to_be32(oi->ip_attr);
         lvb->lvb_igeneration = cpu_to_be32(inode->i_generation);
 
+out:
         mlog_meta_lvb(0, lockres);
 
         mlog_exit_void();
@@ -1727,6 +1745,18 @@ int ocfs2_meta_lock_full(struct inode *inode,
                 wait_event(osb->recovery_event,
                            ocfs2_node_map_is_empty(osb, &osb->recovery_map));
 
+        /*
+         * We only see this flag if we're being called from
+         * ocfs2_read_locked_inode(). It means we're locking an inode
+         * which hasn't been populated yet, so clear the refresh flag
+         * and let the caller handle it.
+         */
+        if (inode->i_state & I_NEW) {
+                status = 0;
+                ocfs2_complete_lock_res_refresh(lockres, 0);
+                goto bail;
+        }
+
         /* This is fun. The caller may want a bh back, or it may
          * not. ocfs2_meta_lock_update definitely wants one in, but
          * may or may not read one, depending on what's in the
diff --git a/fs/ocfs2/dlmglue.h b/fs/ocfs2/dlmglue.h
index 45a74f44b68..4a276938722 100644
--- a/fs/ocfs2/dlmglue.h
+++ b/fs/ocfs2/dlmglue.h
@@ -32,9 +32,9 @@
 #define OCFS2_LVB_VERSION 4
 
 struct ocfs2_meta_lvb {
-        __be16       lvb_reserved0;
-        __u8         lvb_reserved1;
         __u8         lvb_version;
+        __u8         lvb_reserved0;
+        __be16       lvb_reserved1;
         __be32       lvb_iclusters;
         __be32       lvb_iuid;
         __be32       lvb_igid;
@@ -62,13 +62,14 @@ void ocfs2_dlm_shutdown(struct ocfs2_super *osb);
 void ocfs2_lock_res_init_once(struct ocfs2_lock_res *res);
 void ocfs2_inode_lock_res_init(struct ocfs2_lock_res *res,
                                enum ocfs2_lock_type type,
+                               unsigned int generation,
                                struct inode *inode);
 void ocfs2_dentry_lock_res_init(struct ocfs2_dentry_lock *dl,
                                 u64 parent, struct inode *inode);
 void ocfs2_lock_res_free(struct ocfs2_lock_res *res);
 int ocfs2_create_new_inode_locks(struct inode *inode);
 int ocfs2_create_new_lock(struct ocfs2_super *osb,
-                          struct ocfs2_lock_res *lockres, int ex);
+                          struct ocfs2_lock_res *lockres, int ex, int local);
 int ocfs2_drop_inode_locks(struct inode *inode);
 int ocfs2_data_lock_full(struct inode *inode,
                          int write,
diff --git a/fs/ocfs2/export.c b/fs/ocfs2/export.c
index ffcd79749e0..fb91089a60a 100644
--- a/fs/ocfs2/export.c
+++ b/fs/ocfs2/export.c
@@ -58,7 +58,7 @@ static struct dentry *ocfs2_get_dentry(struct super_block *sb, void *vobjp)
                 return ERR_PTR(-ESTALE);
         }
 
-        inode = ocfs2_iget(OCFS2_SB(sb), handle->ih_blkno);
+        inode = ocfs2_iget(OCFS2_SB(sb), handle->ih_blkno, 0);
 
         if (IS_ERR(inode)) {
                 mlog_errno(PTR_ERR(inode));
@@ -115,7 +115,7 @@ static struct dentry *ocfs2_get_parent(struct dentry *child)
                 goto bail_unlock;
         }
 
-        inode = ocfs2_iget(OCFS2_SB(dir->i_sb), blkno);
+        inode = ocfs2_iget(OCFS2_SB(dir->i_sb), blkno, 0);
         if (IS_ERR(inode)) {
                 mlog(ML_ERROR, "Unable to create inode %llu\n",
                      (unsigned long long)blkno);
diff --git a/fs/ocfs2/inode.c b/fs/ocfs2/inode.c
index 66ca7a82b68..69d3db56916 100644
--- a/fs/ocfs2/inode.c
+++ b/fs/ocfs2/inode.c
@@ -54,8 +54,6 @@
 
 #include "buffer_head_io.h"
 
-#define OCFS2_FI_FLAG_NOWAIT    0x1
-#define OCFS2_FI_FLAG_DELETE    0x2
 struct ocfs2_find_inode_args
 {
         u64             fi_blkno;
@@ -109,7 +107,7 @@ struct inode *ocfs2_ilookup_for_vote(struct ocfs2_super *osb,
         return ilookup5(osb->sb, args.fi_ino, ocfs2_find_actor, &args);
 }
 
-struct inode *ocfs2_iget(struct ocfs2_super *osb, u64 blkno)
+struct inode *ocfs2_iget(struct ocfs2_super *osb, u64 blkno, int flags)
 {
         struct inode *inode = NULL;
         struct super_block *sb = osb->sb;
@@ -127,7 +125,7 @@ struct inode *ocfs2_iget(struct ocfs2_super *osb, u64 blkno)
         }
 
         args.fi_blkno = blkno;
-        args.fi_flags = 0;
+        args.fi_flags = flags;
         args.fi_ino = ino_from_blkno(sb, blkno);
 
         inode = iget5_locked(sb, args.fi_ino, ocfs2_find_actor,
@@ -297,15 +295,11 @@ int ocfs2_populate_inode(struct inode *inode, struct ocfs2_dinode *fe,
         OCFS2_I(inode)->ip_orphaned_slot = OCFS2_INVALID_SLOT;
         OCFS2_I(inode)->ip_attr = le32_to_cpu(fe->i_attr);
 
-        if (create_ino)
-                inode->i_ino = ino_from_blkno(inode->i_sb,
-                               le64_to_cpu(fe->i_blkno));
-
-        mlog(0, "blkno = %llu, ino = %lu, create_ino = %s\n",
-             (unsigned long long)fe->i_blkno, inode->i_ino, create_ino ? "true" : "false");
-
         inode->i_nlink = le16_to_cpu(fe->i_links_count);
 
+        if (fe->i_flags & cpu_to_le32(OCFS2_SYSTEM_FL))
+                OCFS2_I(inode)->ip_flags |= OCFS2_INODE_SYSTEM_FILE;
+
         if (fe->i_flags & cpu_to_le32(OCFS2_LOCAL_ALLOC_FL)) {
                 OCFS2_I(inode)->ip_flags |= OCFS2_INODE_BITMAP;
                 mlog(0, "local alloc inode: i_ino=%lu\n", inode->i_ino);
@@ -343,12 +337,28 @@ int ocfs2_populate_inode(struct inode *inode, struct ocfs2_dinode *fe,
                     break;
         }
 
+        if (create_ino) {
+                inode->i_ino = ino_from_blkno(inode->i_sb,
+                               le64_to_cpu(fe->i_blkno));
+
+                /*
+                 * If we ever want to create system files from kernel,
+                 * the generation argument to
+                 * ocfs2_inode_lock_res_init() will have to change.
+                 */
+                BUG_ON(fe->i_flags & cpu_to_le32(OCFS2_SYSTEM_FL));
+
+                ocfs2_inode_lock_res_init(&OCFS2_I(inode)->ip_meta_lockres,
+                                          OCFS2_LOCK_TYPE_META, 0, inode);
+        }
+
         ocfs2_inode_lock_res_init(&OCFS2_I(inode)->ip_rw_lockres,
-                                  OCFS2_LOCK_TYPE_RW, inode);
-        ocfs2_inode_lock_res_init(&OCFS2_I(inode)->ip_meta_lockres,
-                                  OCFS2_LOCK_TYPE_META, inode);
+                                  OCFS2_LOCK_TYPE_RW, inode->i_generation,
+                                  inode);
+
         ocfs2_inode_lock_res_init(&OCFS2_I(inode)->ip_data_lockres,
-                                  OCFS2_LOCK_TYPE_DATA, inode);
+                                  OCFS2_LOCK_TYPE_DATA, inode->i_generation,
+                                  inode);
 
         ocfs2_set_inode_flags(inode);
         inode->i_flags |= S_NOATIME;
@@ -366,15 +376,15 @@ static int ocfs2_read_locked_inode(struct inode *inode,
         struct ocfs2_super *osb;
         struct ocfs2_dinode *fe;
         struct buffer_head *bh = NULL;
-        int status;
-        int sysfile = 0;
+        int status, can_lock;
+        u32 generation = 0;
 
         mlog_entry("(0x%p, 0x%p)\n", inode, args);
 
         status = -EINVAL;
         if (inode == NULL || inode->i_sb == NULL) {
                 mlog(ML_ERROR, "bad inode\n");
-                goto bail;
+                return status;
         }
         sb = inode->i_sb;
         osb = OCFS2_SB(sb);
@@ -382,50 +392,110 @@ static int ocfs2_read_locked_inode(struct inode *inode,
         if (!args) {
                 mlog(ML_ERROR, "bad inode args\n");
                 make_bad_inode(inode);
-                goto bail;
+                return status;
         }
 
-        /* Read the FE off disk. This is safe because the kernel only
-         * does one read_inode2 for a new inode, and if it doesn't
-         * exist yet then nobody can be working on it! */
-        status = ocfs2_read_block(osb, args->fi_blkno, &bh, 0, NULL);
+        /*
+         * To improve performance of cold-cache inode stats, we take
+         * the cluster lock here if possible.
+         *
+         * Generally, OCFS2 never trusts the contents of an inode
+         * unless it's holding a cluster lock, so taking it here isn't
+         * a correctness issue as much as it is a performance
+         * improvement.
+         *
+         * There are three times when taking the lock is not a good idea:
+         *
+         * 1) During startup, before we have initialized the DLM.
+         *
+         * 2) If we are reading certain system files which never get
+         *    cluster locks (local alloc, truncate log).
+         *
+         * 3) If the process doing the iget() is responsible for
+         *    orphan dir recovery. We're holding the orphan dir lock and
+         *    can get into a deadlock with another process on another
+         *    node in ->delete_inode().
+         *
+         * #1 and #2 can be simply solved by never taking the lock
+         * here for system files (which are the only type we read
+         * during mount). It's a heavier approach, but our main
+         * concern is user-accesible files anyway.
+         *
+         * #3 works itself out because we'll eventually take the
+         * cluster lock before trusting anything anyway.
+         */
+        can_lock = !(args->fi_flags & OCFS2_FI_FLAG_SYSFILE)
+                && !(args->fi_flags & OCFS2_FI_FLAG_NOLOCK);
+
+        /*
+         * To maintain backwards compatibility with older versions of
+         * ocfs2-tools, we still store the generation value for system
+         * files. The only ones that actually matter to userspace are
+         * the journals, but it's easier and inexpensive to just flag
+         * all system files similarly.
+         */
+        if (args->fi_flags & OCFS2_FI_FLAG_SYSFILE)
+                generation = osb->fs_generation;
+
+        ocfs2_inode_lock_res_init(&OCFS2_I(inode)->ip_meta_lockres,
+                                  OCFS2_LOCK_TYPE_META,
+                                  generation, inode);
+
+        if (can_lock) {
+                status = ocfs2_meta_lock(inode, NULL, NULL, 0);
+                if (status) {
+                        make_bad_inode(inode);
+                        mlog_errno(status);
+                        return status;
+                }
+        }
+
+        status = ocfs2_read_block(osb, args->fi_blkno, &bh, 0,
+                                  can_lock ? inode : NULL);
         if (status < 0) {
                 mlog_errno(status);
-                make_bad_inode(inode);
                 goto bail;
         }
 
+        status = -EINVAL;
         fe = (struct ocfs2_dinode *) bh->b_data;
         if (!OCFS2_IS_VALID_DINODE(fe)) {
                 mlog(ML_ERROR, "Invalid dinode #%llu: signature = %.*s\n",
                      (unsigned long long)fe->i_blkno, 7, fe->i_signature);
-                make_bad_inode(inode);
                 goto bail;
         }
 
-        if (fe->i_flags & cpu_to_le32(OCFS2_SYSTEM_FL))
-                sysfile = 1;
+        /*
+         * This is a code bug. Right now the caller needs to
+         * understand whether it is asking for a system file inode or
+         * not so the proper lock names can be built.
+         */
+        mlog_bug_on_msg(!!(fe->i_flags & cpu_to_le32(OCFS2_SYSTEM_FL)) !=
+                        !!(args->fi_flags & OCFS2_FI_FLAG_SYSFILE),
+                        "Inode %llu: system file state is ambigous\n",
+                        (unsigned long long)args->fi_blkno);
 
         if (S_ISCHR(le16_to_cpu(fe->i_mode)) ||
             S_ISBLK(le16_to_cpu(fe->i_mode)))
                 inode->i_rdev = huge_decode_dev(le64_to_cpu(fe->id1.dev1.i_rdev));
 
-        status = -EINVAL;
         if (ocfs2_populate_inode(inode, fe, 0) < 0) {
                 mlog(ML_ERROR, "populate failed! i_blkno=%llu, i_ino=%lu\n",
                      (unsigned long long)fe->i_blkno, inode->i_ino);
-                make_bad_inode(inode);
                 goto bail;
         }
 
         BUG_ON(args->fi_blkno != le64_to_cpu(fe->i_blkno));
 
-        if (sysfile)
-               OCFS2_I(inode)->ip_flags |= OCFS2_INODE_SYSTEM_FILE;
-
         status = 0;
 
 bail:
+        if (can_lock)
+                ocfs2_meta_unlock(inode, 0);
+
+        if (status < 0)
+                make_bad_inode(inode);
+
         if (args && bh)
                 brelse(bh);
 
@@ -898,9 +968,15 @@ void ocfs2_delete_inode(struct inode *inode)
                 goto bail_unlock_inode;
         }
 
-        /* Mark the inode as successfully deleted. This is important
-         * for ocfs2_clear_inode as it will check this flag and skip
-         * any checkpointing work */
+        /*
+         * Mark the inode as successfully deleted.
+         *
+         * This is important for ocfs2_clear_inode() as it will check
+         * this flag and skip any checkpointing work
+         *
+         * ocfs2_stuff_meta_lvb() also uses this flag to invalidate
+         * the LVB for other nodes.
+         */
         OCFS2_I(inode)->ip_flags |= OCFS2_INODE_DELETED;
 
 bail_unlock_inode:
diff --git a/fs/ocfs2/inode.h b/fs/ocfs2/inode.h
index 4d1e5399256..9957810fdf8 100644
--- a/fs/ocfs2/inode.h
+++ b/fs/ocfs2/inode.h
@@ -122,7 +122,13 @@ struct buffer_head *ocfs2_bread(struct inode *inode, int block,
 void ocfs2_clear_inode(struct inode *inode);
 void ocfs2_delete_inode(struct inode *inode);
 void ocfs2_drop_inode(struct inode *inode);
-struct inode *ocfs2_iget(struct ocfs2_super *osb, u64 feoff);
+
+/* Flags for ocfs2_iget() */
+#define OCFS2_FI_FLAG_NOWAIT    0x1
+#define OCFS2_FI_FLAG_DELETE    0x2
+#define OCFS2_FI_FLAG_SYSFILE   0x4
+#define OCFS2_FI_FLAG_NOLOCK    0x8
+struct inode *ocfs2_iget(struct ocfs2_super *osb, u64 feoff, int flags);
 struct inode *ocfs2_ilookup_for_vote(struct ocfs2_super *osb,
                                      u64 blkno,
                                      int delete_vote);
diff --git a/fs/ocfs2/journal.c b/fs/ocfs2/journal.c
index f92bf1dd379..fd9734def55 100644
--- a/fs/ocfs2/journal.c
+++ b/fs/ocfs2/journal.c
@@ -1493,7 +1493,8 @@ static int ocfs2_queue_orphans(struct ocfs2_super *osb,
                         if (de->name_len == 2 && !strncmp("..", de->name, 2))
                                 continue;
 
-                        iter = ocfs2_iget(osb, le64_to_cpu(de->inode));
+                        iter = ocfs2_iget(osb, le64_to_cpu(de->inode),
+                                          OCFS2_FI_FLAG_NOLOCK);
                         if (IS_ERR(iter))
                                 continue;
 
diff --git a/fs/ocfs2/namei.c b/fs/ocfs2/namei.c
index 6fa978874c3..849c3b4bb94 100644
--- a/fs/ocfs2/namei.c
+++ b/fs/ocfs2/namei.c
@@ -179,7 +179,7 @@ static struct dentry *ocfs2_lookup(struct inode *dir, struct dentry *dentry,
         if (status < 0)
                 goto bail_add;
 
-        inode = ocfs2_iget(OCFS2_SB(dir->i_sb), blkno);
+        inode = ocfs2_iget(OCFS2_SB(dir->i_sb), blkno, 0);
         if (IS_ERR(inode)) {
                 mlog(ML_ERROR, "Unable to create inode %llu\n",
                      (unsigned long long)blkno);
diff --git a/fs/ocfs2/super.c b/fs/ocfs2/super.c
index 33a6de6fc61..4c29cd7cc8e 100644
--- a/fs/ocfs2/super.c
+++ b/fs/ocfs2/super.c
@@ -202,7 +202,7 @@ static int ocfs2_init_global_system_inodes(struct ocfs2_super *osb)
 
         mlog_entry_void();
 
-        new = ocfs2_iget(osb, osb->root_blkno);
+        new = ocfs2_iget(osb, osb->root_blkno, OCFS2_FI_FLAG_SYSFILE);
         if (IS_ERR(new)) {
                 status = PTR_ERR(new);
                 mlog_errno(status);
@@ -210,7 +210,7 @@ static int ocfs2_init_global_system_inodes(struct ocfs2_super *osb)
         }
         osb->root_inode = new;
 
-        new = ocfs2_iget(osb, osb->system_dir_blkno);
+        new = ocfs2_iget(osb, osb->system_dir_blkno, OCFS2_FI_FLAG_SYSFILE);
         if (IS_ERR(new)) {
                 status = PTR_ERR(new);
                 mlog_errno(status);
diff --git a/fs/ocfs2/sysfile.c b/fs/ocfs2/sysfile.c
index 98435002ac4..5df6e35d09b 100644
--- a/fs/ocfs2/sysfile.c
+++ b/fs/ocfs2/sysfile.c
@@ -115,7 +115,7 @@ static struct inode * _ocfs2_get_system_file_inode(struct ocfs2_super *osb,
                 goto bail;
         }
 
-        inode = ocfs2_iget(osb, blkno);
+        inode = ocfs2_iget(osb, blkno, OCFS2_FI_FLAG_SYSFILE);
         if (IS_ERR(inode)) {
                 mlog_errno(PTR_ERR(inode));
                 inode = NULL;