From 23bfc2a24ea3d993cc5cc90c9970654e7232502e Mon Sep 17 00:00:00 2001 From: NeilBrown Date: Mon, 13 Oct 2014 15:52:14 -0700 Subject: autofs4: allow RCU-walk to walk through autofs4 This series teaches autofs about RCU-walk so that we don't drop straight into REF-walk when we hit an autofs directory, and so that we avoid spinlocks as much as possible when performing an RCU-walk. This is needed so that the benefits of the recent NFS support for RCU-walk are fully available when NFS filesystems are automounted. Patches have been carefully reviewed and tested both with test suites and in production - thanks a lot to Ian Kent for his support there. This patch (of 6): Any attempt to look up a pathname that passes though an autofs4 mount is currently forced out of RCU-walk into REF-walk. This can significantly hurt performance of many-thread work loads on many-core systems, especially if the automounted filesystem supports RCU-walk but doesn't get to benefit from it. So if autofs4_d_manage is called with rcu_walk set, only fail with -ECHILD if it is necessary to wait longer than a spinlock. Signed-off-by: NeilBrown Reviewed-by: Ian Kent Tested-by: Ian Kent Signed-off-by: Andrew Morton Signed-off-by: Linus Torvalds --- fs/autofs4/root.c | 44 +++++++++++++++++++++++++++++--------------- 1 file changed, 29 insertions(+), 15 deletions(-) (limited to 'fs/autofs4/root.c') diff --git a/fs/autofs4/root.c b/fs/autofs4/root.c index cdb25ebccc4c..2296c8301b66 100644 --- a/fs/autofs4/root.c +++ b/fs/autofs4/root.c @@ -210,7 +210,8 @@ next: return NULL; } -static struct dentry *autofs4_lookup_expiring(struct dentry *dentry) +static struct dentry *autofs4_lookup_expiring(struct dentry *dentry, + bool rcu_walk) { struct autofs_sb_info *sbi = autofs4_sbi(dentry->d_sb); struct dentry *parent = dentry->d_parent; @@ -229,6 +230,11 @@ static struct dentry *autofs4_lookup_expiring(struct dentry *dentry) struct dentry *expiring; struct qstr *qstr; + if (rcu_walk) { + spin_unlock(&sbi->lookup_lock); + return ERR_PTR(-ECHILD); + } + ino = list_entry(p, struct autofs_info, expiring); expiring = ino->dentry; @@ -264,13 +270,15 @@ next: return NULL; } -static int autofs4_mount_wait(struct dentry *dentry) +static int autofs4_mount_wait(struct dentry *dentry, bool rcu_walk) { struct autofs_sb_info *sbi = autofs4_sbi(dentry->d_sb); struct autofs_info *ino = autofs4_dentry_ino(dentry); int status = 0; if (ino->flags & AUTOFS_INF_PENDING) { + if (rcu_walk) + return -ECHILD; DPRINTK("waiting for mount name=%.*s", dentry->d_name.len, dentry->d_name.name); status = autofs4_wait(sbi, dentry, NFY_MOUNT); @@ -280,20 +288,22 @@ static int autofs4_mount_wait(struct dentry *dentry) return status; } -static int do_expire_wait(struct dentry *dentry) +static int do_expire_wait(struct dentry *dentry, bool rcu_walk) { struct dentry *expiring; - expiring = autofs4_lookup_expiring(dentry); + expiring = autofs4_lookup_expiring(dentry, rcu_walk); + if (IS_ERR(expiring)) + return PTR_ERR(expiring); if (!expiring) - return autofs4_expire_wait(dentry); + return autofs4_expire_wait(dentry, rcu_walk); else { /* * If we are racing with expire the request might not * be quite complete, but the directory has been removed * so it must have been successful, just wait for it. */ - autofs4_expire_wait(expiring); + autofs4_expire_wait(expiring, 0); autofs4_del_expiring(expiring); dput(expiring); } @@ -345,7 +355,7 @@ static struct vfsmount *autofs4_d_automount(struct path *path) * and the directory was removed, so just go ahead and try * the mount. */ - status = do_expire_wait(dentry); + status = do_expire_wait(dentry, 0); if (status && status != -EAGAIN) return NULL; @@ -353,7 +363,7 @@ static struct vfsmount *autofs4_d_automount(struct path *path) spin_lock(&sbi->fs_lock); if (ino->flags & AUTOFS_INF_PENDING) { spin_unlock(&sbi->fs_lock); - status = autofs4_mount_wait(dentry); + status = autofs4_mount_wait(dentry, 0); if (status) return ERR_PTR(status); goto done; @@ -394,7 +404,7 @@ static struct vfsmount *autofs4_d_automount(struct path *path) } ino->flags |= AUTOFS_INF_PENDING; spin_unlock(&sbi->fs_lock); - status = autofs4_mount_wait(dentry); + status = autofs4_mount_wait(dentry, 0); spin_lock(&sbi->fs_lock); ino->flags &= ~AUTOFS_INF_PENDING; if (status) { @@ -430,21 +440,25 @@ static int autofs4_d_manage(struct dentry *dentry, bool rcu_walk) return 0; } - /* We need to sleep, so we need pathwalk to be in ref-mode */ - if (rcu_walk) - return -ECHILD; - /* Wait for pending expires */ - do_expire_wait(dentry); + if (do_expire_wait(dentry, rcu_walk) == -ECHILD) + return -ECHILD; /* * This dentry may be under construction so wait on mount * completion. */ - status = autofs4_mount_wait(dentry); + status = autofs4_mount_wait(dentry, rcu_walk); if (status) return status; + if (rcu_walk) + /* it is always safe to return 0 as the worst that + * will happen is we retry in REF-walk mode. + * Better than always taking a lock. + */ + return 0; + spin_lock(&sbi->fs_lock); /* * If the dentry has been selected for expire while we slept -- cgit From ef16cc5909c907586f4eba5455e183fd319637c8 Mon Sep 17 00:00:00 2001 From: NeilBrown Date: Mon, 13 Oct 2014 15:52:22 -0700 Subject: autofs4: d_manage() should return -EISDIR when appropriate in rcu-walk mode. If rcu-walk mode we don't *have* to return -EISDIR for non-mount-traps as we will simply drop into REF-walk and handling DCACHE_NEED_AUTOMOUNT dentrys the slow way. But it is better if we do when possible. In 'oz_mode', use the same condition as ref-walk: if not a mountpoint, then it must be -EISDIR. In regular mode there are most tests needed. Most of them can be performed without taking any spinlocks. If we find a directory that isn't obviously empty, and isn't mounted on, we need to call 'simple_empty()' which does take a spinlock. If this turned out to hurt performance, some other approach could be found to signal when a directory is known to be empty. Signed-off-by: NeilBrown Reviewed-by: Ian Kent Tested-by: Ian Kent Signed-off-by: Andrew Morton Signed-off-by: Linus Torvalds --- fs/autofs4/root.c | 26 ++++++++++++++++++++------ 1 file changed, 20 insertions(+), 6 deletions(-) (limited to 'fs/autofs4/root.c') diff --git a/fs/autofs4/root.c b/fs/autofs4/root.c index 2296c8301b66..d76d083f2f06 100644 --- a/fs/autofs4/root.c +++ b/fs/autofs4/root.c @@ -433,8 +433,6 @@ static int autofs4_d_manage(struct dentry *dentry, bool rcu_walk) /* The daemon never waits. */ if (autofs4_oz_mode(sbi)) { - if (rcu_walk) - return 0; if (!d_mountpoint(dentry)) return -EISDIR; return 0; @@ -452,12 +450,28 @@ static int autofs4_d_manage(struct dentry *dentry, bool rcu_walk) if (status) return status; - if (rcu_walk) - /* it is always safe to return 0 as the worst that - * will happen is we retry in REF-walk mode. - * Better than always taking a lock. + if (rcu_walk) { + /* We don't need fs_lock in rcu_walk mode, + * just testing 'AUTOFS_INFO_NO_RCU' is enough. + * simple_empty() takes a spinlock, so leave it + * to last. + * We only return -EISDIR when certain this isn't + * a mount-trap. */ + struct inode *inode; + if (ino->flags & (AUTOFS_INF_EXPIRING | AUTOFS_INF_NO_RCU)) + return 0; + if (d_mountpoint(dentry)) + return 0; + inode = ACCESS_ONCE(dentry->d_inode); + if (inode && S_ISLNK(inode->i_mode)) + return -EISDIR; + if (list_empty(&dentry->d_subdirs)) + return 0; + if (!simple_empty(dentry)) + return -EISDIR; return 0; + } spin_lock(&sbi->fs_lock); /* -- cgit