ocfs2: Add dentry tracking API

Replace the dentry vote mechanism with a cluster lock which covers a set
of dentries. This allows us to force d_delete() only on nodes which actually
care about an unlink.

Every node that does a ->lookup() gets a read only lock on the dentry, until
an unlink during which the unlinking node, will request an exclusive lock,
forcing the other nodes who care about that dentry to d_delete() it. The
effect is that we retain a very lightweight ->d_revalidate(), and at the
same time get to make large improvements to the average case performance of
the ocfs2 unlink and rename operations.

This patch adds the higher level API and the dentry manipulation code.

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

1 files changed, 356 insertions, 19 deletions

diff --git a/fs/ocfs2/dcache.c b/fs/ocfs2/dcache.c
index aea45771894..09efe240e65 100644
--- a/fs/ocfs2/dcache.c
+++ b/fs/ocfs2/dcache.c
@@ -35,15 +35,17 @@
 
 #include "alloc.h"
 #include "dcache.h"
+#include "dlmglue.h"
 #include "file.h"
 #include "inode.h"
 
+
 static int ocfs2_dentry_revalidate(struct dentry *dentry,
                                    struct nameidata *nd)
 {
         struct inode *inode = dentry->d_inode;
         int ret = 0;    /* if all else fails, just return false */
-        struct ocfs2_super *osb;
+        struct ocfs2_super *osb = OCFS2_SB(dentry->d_sb);
 
         mlog_entry("(0x%p, '%.*s')\n", dentry,
                    dentry->d_name.len, dentry->d_name.name);
@@ -55,28 +57,31 @@ static int ocfs2_dentry_revalidate(struct dentry *dentry,
                 goto bail;
         }
 
-        osb = OCFS2_SB(inode->i_sb);
-
         BUG_ON(!osb);
 
-        if (inode != osb->root_inode) {
-                spin_lock(&OCFS2_I(inode)->ip_lock);
-                /* did we or someone else delete this inode? */
-                if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_DELETED) {
-                        spin_unlock(&OCFS2_I(inode)->ip_lock);
-                        mlog(0, "inode (%llu) deleted, returning false\n",
-                             (unsigned long long)OCFS2_I(inode)->ip_blkno);
-                        goto bail;
-                }
+        if (inode == osb->root_inode || is_bad_inode(inode))
+                goto bail;
+
+        spin_lock(&OCFS2_I(inode)->ip_lock);
+        /* did we or someone else delete this inode? */
+        if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_DELETED) {
                 spin_unlock(&OCFS2_I(inode)->ip_lock);
+                mlog(0, "inode (%llu) deleted, returning false\n",
+                     (unsigned long long)OCFS2_I(inode)->ip_blkno);
+                goto bail;
+        }
+        spin_unlock(&OCFS2_I(inode)->ip_lock);
 
-                if (!inode->i_nlink) {
-                        mlog(0, "Inode %llu orphaned, returning false "
-                             "dir = %d\n",
-                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
-                             S_ISDIR(inode->i_mode));
-                        goto bail;
-                }
+        /*
+         * We don't need a cluster lock to test this because once an
+         * inode nlink hits zero, it never goes back.
+         */
+        if (inode->i_nlink == 0) {
+                mlog(0, "Inode %llu orphaned, returning false "
+                     "dir = %d\n",
+                     (unsigned long long)OCFS2_I(inode)->ip_blkno,
+                     S_ISDIR(inode->i_mode));
+                goto bail;
         }
 
         ret = 1;
@@ -87,8 +92,340 @@ bail:
         return ret;
 }
 
+static int ocfs2_match_dentry(struct dentry *dentry,
+                              u64 parent_blkno,
+                              int skip_unhashed)
+{
+        struct inode *parent;
+
+        /*
+         * ocfs2_lookup() does a d_splice_alias() _before_ attaching
+         * to the lock data, so we skip those here, otherwise
+         * ocfs2_dentry_attach_lock() will get its original dentry
+         * back.
+         */
+        if (!dentry->d_fsdata)
+                return 0;
+
+        if (!dentry->d_parent)
+                return 0;
+
+        if (skip_unhashed && d_unhashed(dentry))
+                return 0;
+
+        parent = dentry->d_parent->d_inode;
+        /* Negative parent dentry? */
+        if (!parent)
+                return 0;
+
+        /* Name is in a different directory. */
+        if (OCFS2_I(parent)->ip_blkno != parent_blkno)
+                return 0;
+
+        return 1;
+}
+
+/*
+ * Walk the inode alias list, and find a dentry which has a given
+ * parent. ocfs2_dentry_attach_lock() wants to find _any_ alias as it
+ * is looking for a dentry_lock reference. The vote thread is looking
+ * to unhash aliases, so we allow it to skip any that already have
+ * that property.
+ */
+struct dentry *ocfs2_find_local_alias(struct inode *inode,
+                                      u64 parent_blkno,
+                                      int skip_unhashed)
+{
+        struct list_head *p;
+        struct dentry *dentry = NULL;
+
+        spin_lock(&dcache_lock);
+
+        list_for_each(p, &inode->i_dentry) {
+                dentry = list_entry(p, struct dentry, d_alias);
+
+                if (ocfs2_match_dentry(dentry, parent_blkno, skip_unhashed)) {
+                        mlog(0, "dentry found: %.*s\n",
+                             dentry->d_name.len, dentry->d_name.name);
+
+                        dget_locked(dentry);
+                        break;
+                }
+
+                dentry = NULL;
+        }
+
+        spin_unlock(&dcache_lock);
+
+        return dentry;
+}
+
 DEFINE_SPINLOCK(dentry_attach_lock);
 
+/*
+ * Attach this dentry to a cluster lock.
+ *
+ * Dentry locks cover all links in a given directory to a particular
+ * inode. We do this so that ocfs2 can build a lock name which all
+ * nodes in the cluster can agree on at all times. Shoving full names
+ * in the cluster lock won't work due to size restrictions. Covering
+ * links inside of a directory is a good compromise because it still
+ * allows us to use the parent directory lock to synchronize
+ * operations.
+ *
+ * Call this function with the parent dir semaphore and the parent dir
+ * cluster lock held.
+ *
+ * The dir semaphore will protect us from having to worry about
+ * concurrent processes on our node trying to attach a lock at the
+ * same time.
+ *
+ * The dir cluster lock (held at either PR or EX mode) protects us
+ * from unlink and rename on other nodes.
+ *
+ * The 'create' flag tells us whether we're doing this as a result of
+ * a file creation.
+ *
+ * A dput() can happen asynchronously due to pruning, so we cover
+ * attaching and detaching the dentry lock with a
+ * dentry_attach_lock.
+ *
+ * A node which has done lookup on a name retains a protected read
+ * lock until final dput. If the user requests and unlink or rename,
+ * the protected read is upgraded to an exclusive lock. Other nodes
+ * who have seen the dentry will then be informed that they need to
+ * downgrade their lock, which will involve d_delete on the
+ * dentry. This happens in ocfs2_dentry_convert_worker().
+ */
+int ocfs2_dentry_attach_lock(struct dentry *dentry,
+                             struct inode *inode,
+                             u64 parent_blkno,
+                             int create)
+{
+        int ret;
+        struct dentry *alias;
+        struct ocfs2_dentry_lock *dl = dentry->d_fsdata;
+
+        mlog(0, "Attach \"%.*s\", parent %llu, create %d, fsdata: %p\n",
+             dentry->d_name.len, dentry->d_name.name,
+             (unsigned long long)parent_blkno, create, dl);
+
+        /*
+         * Negative dentry. We ignore these for now.
+         *
+         * XXX: Could we can improve ocfs2_dentry_revalidate() by
+         * tracking these?
+         */
+        if (!inode)
+                return 0;
+
+        if (dl) {
+                mlog_bug_on_msg(dl->dl_parent_blkno != parent_blkno,
+                                " \"%.*s\": old parent: %llu, new: %llu\n",
+                                dentry->d_name.len, dentry->d_name.name,
+                                (unsigned long long)parent_blkno,
+                                (unsigned long long)dl->dl_parent_blkno);
+                return 0;
+        }
+
+        alias = ocfs2_find_local_alias(inode, parent_blkno, 0);
+        if (alias) {
+                /*
+                 * Great, an alias exists, which means we must have a
+                 * dentry lock already. We can just grab the lock off
+                 * the alias and add it to the list.
+                 *
+                 * We're depending here on the fact that this dentry
+                 * was found and exists in the dcache and so must have
+                 * a reference to the dentry_lock because we can't
+                 * race creates. Final dput() cannot happen on it
+                 * since we have it pinned, so our reference is safe.
+                 */
+                dl = alias->d_fsdata;
+                mlog_bug_on_msg(!dl, "parent %llu, ino %llu, create %d\n",
+                                (unsigned long long)parent_blkno,
+                                (unsigned long long)OCFS2_I(inode)->ip_blkno,
+                                create);
+
+                mlog_bug_on_msg(dl->dl_parent_blkno != parent_blkno,
+                                " \"%.*s\": old parent: %llu, new: %llu\n",
+                                dentry->d_name.len, dentry->d_name.name,
+                                (unsigned long long)parent_blkno,
+                                (unsigned long long)dl->dl_parent_blkno);
+
+                mlog(0, "Found: %s\n", dl->dl_lockres.l_name);
+
+                goto out_attach;
+        }
+
+        /*
+         * There are no other aliases
+         */
+        dl = kmalloc(sizeof(*dl), GFP_NOFS);
+        if (!dl) {
+                ret = -ENOMEM;
+                mlog_errno(ret);
+                return ret;
+        }
+
+        dl->dl_count = 0;
+        /*
+         * Does this have to happen below, for all attaches, in case
+         * the struct inode gets blown away by votes?
+         */
+        dl->dl_inode = igrab(inode);
+        dl->dl_parent_blkno = parent_blkno;
+        ocfs2_dentry_lock_res_init(dl, parent_blkno, inode);
+
+out_attach:
+        spin_lock(&dentry_attach_lock);
+        dentry->d_fsdata = dl;
+        dl->dl_count++;
+        spin_unlock(&dentry_attach_lock);
+
+        /*
+         * Creation of a new file means that nobody can possibly have
+         * this name in the system, which means that acquiry of those
+         * locks can easily be optimized.
+         */
+        if (create) {
+                ret = ocfs2_create_new_lock(OCFS2_SB(inode->i_sb),
+                                            &dl->dl_lockres, 0);
+                if (ret)
+                        mlog_errno(ret);
+                goto out;
+        }
+
+        /*
+         * This actually gets us our PRMODE level lock. From now on,
+         * we'll have a notification if one of these names is
+         * destroyed on another node.
+         */
+        ret = ocfs2_dentry_lock(dentry, 0);
+        if (ret) {
+                mlog_errno(ret);
+                goto out;
+        }
+        ocfs2_dentry_unlock(dentry, 0);
+
+out:
+        dput(alias);
+
+        return ret;
+}
+
+/*
+ * ocfs2_dentry_iput() and friends.
+ *
+ * At this point, our particular dentry is detached from the inodes
+ * alias list, so there's no way that the locking code can find it.
+ *
+ * The interesting stuff happens when we determine that our lock needs
+ * to go away because this is the last subdir alias in the
+ * system. This function needs to handle a couple things:
+ *
+ * 1) Synchronizing lock shutdown with the downconvert threads. This
+ *    is already handled for us via the lockres release drop function
+ *    called in ocfs2_release_dentry_lock()
+ *
+ * 2) A race may occur when we're doing our lock shutdown and
+ *    another process wants to create a new dentry lock. Right now we
+ *    let them race, which means that for a very short while, this
+ *    node might have two locks on a lock resource. This should be a
+ *    problem though because one of them is in the process of being
+ *    thrown out.
+ */
+static void ocfs2_drop_dentry_lock(struct ocfs2_super *osb,
+                                   struct ocfs2_dentry_lock *dl)
+{
+        ocfs2_simple_drop_lockres(osb, &dl->dl_lockres);
+        ocfs2_lock_res_free(&dl->dl_lockres);
+        iput(dl->dl_inode);
+        kfree(dl);
+}
+
+void ocfs2_dentry_lock_put(struct ocfs2_super *osb,
+                           struct ocfs2_dentry_lock *dl)
+{
+        int unlock = 0;
+
+        BUG_ON(dl->dl_count == 0);
+
+        spin_lock(&dentry_attach_lock);
+        dl->dl_count--;
+        unlock = !dl->dl_count;
+        spin_unlock(&dentry_attach_lock);
+
+        if (unlock)
+                ocfs2_drop_dentry_lock(osb, dl);
+}
+
+static void ocfs2_dentry_iput(struct dentry *dentry, struct inode *inode)
+{
+        struct ocfs2_dentry_lock *dl = dentry->d_fsdata;
+
+        mlog_bug_on_msg(!dl && !(dentry->d_flags & DCACHE_DISCONNECTED),
+                        "dentry: %.*s\n", dentry->d_name.len,
+                        dentry->d_name.name);
+
+        if (!dl)
+                goto out;
+
+        mlog_bug_on_msg(dl->dl_count == 0, "dentry: %.*s, count: %u\n",
+                        dentry->d_name.len, dentry->d_name.name,
+                        dl->dl_count);
+
+        ocfs2_dentry_lock_put(OCFS2_SB(dentry->d_sb), dl);
+
+out:
+        iput(inode);
+}
+
+/*
+ * d_move(), but keep the locks in sync.
+ *
+ * When we are done, "dentry" will have the parent dir and name of
+ * "target", which will be thrown away.
+ *
+ * We manually update the lock of "dentry" if need be.
+ *
+ * "target" doesn't have it's dentry lock touched - we allow the later
+ * dput() to handle this for us.
+ *
+ * This is called during ocfs2_rename(), while holding parent
+ * directory locks. The dentries have already been deleted on other
+ * nodes via ocfs2_remote_dentry_delete().
+ *
+ * Normally, the VFS handles the d_move() for the file sytem, after
+ * the ->rename() callback. OCFS2 wants to handle this internally, so
+ * the new lock can be created atomically with respect to the cluster.
+ */
+void ocfs2_dentry_move(struct dentry *dentry, struct dentry *target,
+                       struct inode *old_dir, struct inode *new_dir)
+{
+        int ret;
+        struct ocfs2_super *osb = OCFS2_SB(old_dir->i_sb);
+        struct inode *inode = dentry->d_inode;
+
+        /*
+         * Move within the same directory, so the actual lock info won't
+         * change.
+         *
+         * XXX: Is there any advantage to dropping the lock here?
+         */
+        if (old_dir == new_dir)
+                return;
+
+        ocfs2_dentry_lock_put(osb, dentry->d_fsdata);
+
+        dentry->d_fsdata = NULL;
+        ret = ocfs2_dentry_attach_lock(dentry, inode,
+                                       OCFS2_I(new_dir)->ip_blkno, 0);
+        if (ret)
+                mlog_errno(ret);
+}
+
 struct dentry_operations ocfs2_dentry_ops = {
         .d_revalidate           = ocfs2_dentry_revalidate,
+        .d_iput                 = ocfs2_dentry_iput,
 };