Enable lazytime semantic for atime

Linux 4.0 introduces lazytime. The idea is that when we update the atime, we
delay writing it to disk for as long as it is reasonably possible.

When lazytime is enabled, dirty_inode will be called with only I_DIRTY_TIME
flag whenever i_atime is updated. So under such condition, we will set
z_atime_dirty. We will only write it to disk if file is closed, inode is
evicted or setattr is called. Ideally, we should also write it whenever SA
is going to be updated, but it is left for future improvement.

There's one thing that we should take care of now that we allow i_atime to be
dirty. In original implementation, whenever SA is modified, zfs_inode_update
will be called to overwrite every thing in inode. This will cause dirty
i_atime to be discarded. We fix this by don't overwrite i_atime in
zfs_inode_update. We only overwrite i_atime when allocating new inode or doing
zfs_rezget with zfs_inode_update_new.

Signed-off-by: Chunwei Chen <[email protected]>
Signed-off-by: Brian Behlendorf <[email protected]>
Issue #4482

3 files changed, 110 insertions, 72 deletions

diff --git a/module/zfs/zfs_vnops.c b/module/zfs/zfs_vnops.c
index b4ffd8d2a..3c3d587e8 100644
--- a/module/zfs/zfs_vnops.c
+++ b/module/zfs/zfs_vnops.c
@@ -3005,8 +3005,9 @@ top:
 	}
 
 
-	if (mask & ATTR_ATIME) {
-		ZFS_TIME_ENCODE(&vap->va_atime, atime);
+	if ((mask & ATTR_ATIME) || zp->z_atime_dirty) {
+		zp->z_atime_dirty = 0;
+		ZFS_TIME_ENCODE(&ip->i_atime, atime);
 		SA_ADD_BULK_ATTR(bulk, count, SA_ZPL_ATIME(zsb), NULL,
 		    &atime, sizeof (atime));
 	}
@@ -4128,7 +4129,7 @@ zfs_dirty_inode(struct inode *ip, int flags)
 	dmu_tx_t	*tx;
 	uint64_t	mode, atime[2], mtime[2], ctime[2];
 	sa_bulk_attr_t	bulk[4];
-	int		error;
+	int		error = 0;
 	int		cnt = 0;
 
 	if (zfs_is_readonly(zsb) || dmu_objset_is_snapshot(zsb->z_os))
@@ -4137,6 +4138,20 @@ zfs_dirty_inode(struct inode *ip, int flags)
 	ZFS_ENTER(zsb);
 	ZFS_VERIFY_ZP(zp);
 
+#ifdef I_DIRTY_TIME
+	/*
+	 * This is the lazytime semantic indroduced in Linux 4.0
+	 * This flag will only be called from update_time when lazytime is set.
+	 * (Note, I_DIRTY_SYNC will also set if not lazytime)
+	 * Fortunately mtime and ctime are managed within ZFS itself, so we
+	 * only need to dirty atime.
+	 */
+	if (flags == I_DIRTY_TIME) {
+		zp->z_atime_dirty = 1;
+		goto out;
+	}
+#endif
+
 	tx = dmu_tx_create(zsb->z_os);
 
 	dmu_tx_hold_sa(tx, zp->z_sa_hdl, B_FALSE);
@@ -4149,6 +4164,8 @@ zfs_dirty_inode(struct inode *ip, int flags)
 	}
 
 	mutex_enter(&zp->z_lock);
+	zp->z_atime_dirty = 0;
+
 	SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_MODE(zsb), NULL, &mode, 8);
 	SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_ATIME(zsb), NULL, &atime, 16);
 	SA_ADD_BULK_ATTR(bulk, cnt, SA_ZPL_MTIME(zsb), NULL, &mtime, 16);
diff --git a/module/zfs/zfs_znode.c b/module/zfs/zfs_znode.c
index f3662e272..39b2ba07a 100644
--- a/module/zfs/zfs_znode.c
+++ b/module/zfs/zfs_znode.c
@@ -492,6 +492,90 @@ zfs_inode_set_ops(zfs_sb_t *zsb, struct inode *ip)
 	}
 }
 
+void
+zfs_set_inode_flags(znode_t *zp, struct inode *ip)
+{
+	/*
+	 * Linux and Solaris have different sets of file attributes, so we
+	 * restrict this conversion to the intersection of the two.
+	 */
+
+	if (zp->z_pflags & ZFS_IMMUTABLE)
+		ip->i_flags |= S_IMMUTABLE;
+	else
+		ip->i_flags &= ~S_IMMUTABLE;
+
+	if (zp->z_pflags & ZFS_APPENDONLY)
+		ip->i_flags |= S_APPEND;
+	else
+		ip->i_flags &= ~S_APPEND;
+}
+
+/*
+ * Update the embedded inode given the znode.  We should work toward
+ * eliminating this function as soon as possible by removing values
+ * which are duplicated between the znode and inode.  If the generic
+ * inode has the correct field it should be used, and the ZFS code
+ * updated to access the inode.  This can be done incrementally.
+ */
+static void
+zfs_inode_update_impl(znode_t *zp, boolean_t new)
+{
+	zfs_sb_t	*zsb;
+	struct inode	*ip;
+	uint32_t	blksize;
+	u_longlong_t	i_blocks;
+	uint64_t	atime[2], mtime[2], ctime[2];
+
+	ASSERT(zp != NULL);
+	zsb = ZTOZSB(zp);
+	ip = ZTOI(zp);
+
+	/* Skip .zfs control nodes which do not exist on disk. */
+	if (zfsctl_is_node(ip))
+		return;
+
+	sa_lookup(zp->z_sa_hdl, SA_ZPL_ATIME(zsb), &atime, 16);
+	sa_lookup(zp->z_sa_hdl, SA_ZPL_MTIME(zsb), &mtime, 16);
+	sa_lookup(zp->z_sa_hdl, SA_ZPL_CTIME(zsb), &ctime, 16);
+
+	dmu_object_size_from_db(sa_get_db(zp->z_sa_hdl), &blksize, &i_blocks);
+
+	spin_lock(&ip->i_lock);
+	ip->i_generation = zp->z_gen;
+	ip->i_uid = SUID_TO_KUID(zp->z_uid);
+	ip->i_gid = SGID_TO_KGID(zp->z_gid);
+	set_nlink(ip, zp->z_links);
+	ip->i_mode = zp->z_mode;
+	zfs_set_inode_flags(zp, ip);
+	ip->i_blkbits = SPA_MINBLOCKSHIFT;
+	ip->i_blocks = i_blocks;
+
+	/*
+	 * Only read atime from SA if we are newly created inode (or rezget),
+	 * otherwise i_atime might be dirty.
+	 */
+	if (new)
+		ZFS_TIME_DECODE(&ip->i_atime, atime);
+	ZFS_TIME_DECODE(&ip->i_mtime, mtime);
+	ZFS_TIME_DECODE(&ip->i_ctime, ctime);
+
+	i_size_write(ip, zp->z_size);
+	spin_unlock(&ip->i_lock);
+}
+
+static void
+zfs_inode_update_new(znode_t *zp)
+{
+	zfs_inode_update_impl(zp, B_TRUE);
+}
+
+void
+zfs_inode_update(znode_t *zp)
+{
+	zfs_inode_update_impl(zp, B_FALSE);
+}
+
 /*
  * Construct a znode+inode and initialize.
  *
@@ -567,7 +651,7 @@ zfs_znode_alloc(zfs_sb_t *zsb, dmu_buf_t *db, int blksz,
 	}
 
 	ip->i_ino = obj;
-	zfs_inode_update(zp);
+	zfs_inode_update_new(zp);
 	zfs_inode_set_ops(zsb, ip);
 
 	/*
@@ -594,73 +678,6 @@ error:
 	return (NULL);
 }
 
-void
-zfs_set_inode_flags(znode_t *zp, struct inode *ip)
-{
-	/*
-	 * Linux and Solaris have different sets of file attributes, so we
-	 * restrict this conversion to the intersection of the two.
-	 */
-
-	if (zp->z_pflags & ZFS_IMMUTABLE)
-		ip->i_flags |= S_IMMUTABLE;
-	else
-		ip->i_flags &= ~S_IMMUTABLE;
-
-	if (zp->z_pflags & ZFS_APPENDONLY)
-		ip->i_flags |= S_APPEND;
-	else
-		ip->i_flags &= ~S_APPEND;
-}
-
-/*
- * Update the embedded inode given the znode.  We should work toward
- * eliminating this function as soon as possible by removing values
- * which are duplicated between the znode and inode.  If the generic
- * inode has the correct field it should be used, and the ZFS code
- * updated to access the inode.  This can be done incrementally.
- */
-void
-zfs_inode_update(znode_t *zp)
-{
-	zfs_sb_t	*zsb;
-	struct inode	*ip;
-	uint32_t	blksize;
-	u_longlong_t	i_blocks;
-	uint64_t	atime[2], mtime[2], ctime[2];
-
-	ASSERT(zp != NULL);
-	zsb = ZTOZSB(zp);
-	ip = ZTOI(zp);
-
-	/* Skip .zfs control nodes which do not exist on disk. */
-	if (zfsctl_is_node(ip))
-		return;
-
-	sa_lookup(zp->z_sa_hdl, SA_ZPL_ATIME(zsb), &atime, 16);
-	sa_lookup(zp->z_sa_hdl, SA_ZPL_MTIME(zsb), &mtime, 16);
-	sa_lookup(zp->z_sa_hdl, SA_ZPL_CTIME(zsb), &ctime, 16);
-
-	dmu_object_size_from_db(sa_get_db(zp->z_sa_hdl), &blksize, &i_blocks);
-
-	spin_lock(&ip->i_lock);
-	ip->i_generation = zp->z_gen;
-	ip->i_uid = SUID_TO_KUID(zp->z_uid);
-	ip->i_gid = SGID_TO_KGID(zp->z_gid);
-	set_nlink(ip, zp->z_links);
-	ip->i_mode = zp->z_mode;
-	zfs_set_inode_flags(zp, ip);
-	ip->i_blkbits = SPA_MINBLOCKSHIFT;
-	ip->i_blocks = i_blocks;
-
-	ZFS_TIME_DECODE(&ip->i_atime, atime);
-	ZFS_TIME_DECODE(&ip->i_mtime, mtime);
-	ZFS_TIME_DECODE(&ip->i_ctime, ctime);
-
-	i_size_write(ip, zp->z_size);
-	spin_unlock(&ip->i_lock);
-}
-
 /*
  * Safely mark an inode dirty.  Inodes which are part of a read-only
  * file system or snapshot may not be dirtied.
@@ -1220,7 +1237,8 @@ zfs_rezget(znode_t *zp)
 
 	zp->z_unlinked = (zp->z_links == 0);
 	zp->z_blksz = doi.doi_data_block_size;
-	zfs_inode_update(zp);
+	zp->z_atime_dirty = 0;
+	zfs_inode_update_new(zp);
 
 	zfs_znode_hold_exit(zsb, zh);
 
diff --git a/module/zfs/zpl_inode.c b/module/zfs/zpl_inode.c
index cbdab7d30..7cf62783f 100644
--- a/module/zfs/zpl_inode.c
+++ b/module/zfs/zpl_inode.c
@@ -335,6 +335,9 @@ zpl_setattr(struct dentry *dentry, struct iattr *ia)
 	vap->va_mtime = ia->ia_mtime;
 	vap->va_ctime = ia->ia_ctime;
 
+	if (vap->va_mask & ATTR_ATIME)
+		ip->i_atime = ia->ia_atime;
+
 	cookie = spl_fstrans_mark();
 	error = -zfs_setattr(ip, vap, 0, cr);
 	if (!error && (ia->ia_valid & ATTR_MODE))