[XFS] Improve buffered read throughput by removing unnecessary timer calls

that showed in ´kernel profiles.

SGI-PV: 925163
SGI-Modid: xfs-linux:xfs-kern:23861a

Signed-off-by: Nathan Scott <nathans@sgi.com>

4 files changed, 140 insertions, 17 deletions

diff --git a/fs/xfs/linux-2.6/xfs_iops.c b/fs/xfs/linux-2.6/xfs_iops.c
index 77708a8c9f87..68c83ec31143 100644
--- a/fs/xfs/linux-2.6/xfs_iops.c
+++ b/fs/xfs/linux-2.6/xfs_iops.c
@@ -69,6 +69,137 @@
 #include <linux/xattr.h>
 #include <linux/namei.h>
 
+/*
+ * Change the requested timestamp in the given inode.
+ * We don't lock across timestamp updates, and we don't log them but
+ * we do record the fact that there is dirty information in core.
+ *
+ * NOTE -- callers MUST combine XFS_ICHGTIME_MOD or XFS_ICHGTIME_CHG
+ *              with XFS_ICHGTIME_ACC to be sure that access time
+ *              update will take.  Calling first with XFS_ICHGTIME_ACC
+ *              and then XFS_ICHGTIME_MOD may fail to modify the access
+ *              timestamp if the filesystem is mounted noacctm.
+ */
+void
+xfs_ichgtime(
+        xfs_inode_t     *ip,
+        int             flags)
+{
+        struct inode    *inode = LINVFS_GET_IP(XFS_ITOV(ip));
+        timespec_t      tv;
+
+        /*
+         * We're not supposed to change timestamps in readonly-mounted
+         * filesystems.  Throw it away if anyone asks us.
+         */
+        if (unlikely(IS_RDONLY(inode)))
+                return;
+
+        /*
+         * Don't update access timestamps on reads if mounted "noatime".
+         * Throw it away if anyone asks us.
+         */
+        if (unlikely(
+            (ip->i_mount->m_flags & XFS_MOUNT_NOATIME || IS_NOATIME(inode)) &&
+            (flags & (XFS_ICHGTIME_ACC|XFS_ICHGTIME_MOD|XFS_ICHGTIME_CHG)) ==
+                        XFS_ICHGTIME_ACC))
+                return;
+
+        nanotime(&tv);
+        if (flags & XFS_ICHGTIME_MOD) {
+                inode->i_mtime = tv;
+                ip->i_d.di_mtime.t_sec = (__int32_t)tv.tv_sec;
+                ip->i_d.di_mtime.t_nsec = (__int32_t)tv.tv_nsec;
+        }
+        if (flags & XFS_ICHGTIME_ACC) {
+                inode->i_atime = tv;
+                ip->i_d.di_atime.t_sec = (__int32_t)tv.tv_sec;
+                ip->i_d.di_atime.t_nsec = (__int32_t)tv.tv_nsec;
+        }
+        if (flags & XFS_ICHGTIME_CHG) {
+                inode->i_ctime = tv;
+                ip->i_d.di_ctime.t_sec = (__int32_t)tv.tv_sec;
+                ip->i_d.di_ctime.t_nsec = (__int32_t)tv.tv_nsec;
+        }
+
+        /*
+         * We update the i_update_core field _after_ changing
+         * the timestamps in order to coordinate properly with
+         * xfs_iflush() so that we don't lose timestamp updates.
+         * This keeps us from having to hold the inode lock
+         * while doing this.  We use the SYNCHRONIZE macro to
+         * ensure that the compiler does not reorder the update
+         * of i_update_core above the timestamp updates above.
+         */
+        SYNCHRONIZE();
+        ip->i_update_core = 1;
+        if (!(inode->i_state & I_LOCK))
+                mark_inode_dirty_sync(inode);
+}
+
+/*
+ * Variant on the above which avoids querying the system clock
+ * in situations where we know the Linux inode timestamps have
+ * just been updated (and so we can update our inode cheaply).
+ * We also skip the readonly and noatime checks here, they are
+ * also catered for already.
+ */
+void
+xfs_ichgtime_fast(
+        xfs_inode_t     *ip,
+        struct inode    *inode,
+        int             flags)
+{
+        timespec_t      *tvp;
+
+        /*
+         * We're not supposed to change timestamps in readonly-mounted
+         * filesystems.  Throw it away if anyone asks us.
+         */
+        if (unlikely(IS_RDONLY(inode)))
+                return;
+
+        /*
+         * Don't update access timestamps on reads if mounted "noatime".
+         * Throw it away if anyone asks us.
+         */
+        if (unlikely(
+            (ip->i_mount->m_flags & XFS_MOUNT_NOATIME || IS_NOATIME(inode)) &&
+            ((flags & (XFS_ICHGTIME_ACC|XFS_ICHGTIME_MOD|XFS_ICHGTIME_CHG)) ==
+                        XFS_ICHGTIME_ACC)))
+                return;
+
+        if (flags & XFS_ICHGTIME_MOD) {
+                tvp = &inode->i_mtime;
+                ip->i_d.di_mtime.t_sec = (__int32_t)tvp->tv_sec;
+                ip->i_d.di_mtime.t_nsec = (__int32_t)tvp->tv_nsec;
+        }
+        if (flags & XFS_ICHGTIME_ACC) {
+                tvp = &inode->i_atime;
+                ip->i_d.di_atime.t_sec = (__int32_t)tvp->tv_sec;
+                ip->i_d.di_atime.t_nsec = (__int32_t)tvp->tv_nsec;
+        }
+        if (flags & XFS_ICHGTIME_CHG) {
+                tvp = &inode->i_ctime;
+                ip->i_d.di_ctime.t_sec = (__int32_t)tvp->tv_sec;
+                ip->i_d.di_ctime.t_nsec = (__int32_t)tvp->tv_nsec;
+        }
+
+        /*
+         * We update the i_update_core field _after_ changing
+         * the timestamps in order to coordinate properly with
+         * xfs_iflush() so that we don't lose timestamp updates.
+         * This keeps us from having to hold the inode lock
+         * while doing this.  We use the SYNCHRONIZE macro to
+         * ensure that the compiler does not reorder the update
+         * of i_update_core above the timestamp updates above.
+         */
+        SYNCHRONIZE();
+        ip->i_update_core = 1;
+        if (!(inode->i_state & I_LOCK))
+                mark_inode_dirty_sync(inode);
+}
+
 
 /*
  * Pull the link count and size up from the xfs inode to the linux inode
diff --git a/fs/xfs/linux-2.6/xfs_iops.h b/fs/xfs/linux-2.6/xfs_iops.h
index 6a69a62c36b0..415de44b6aed 100644
--- a/fs/xfs/linux-2.6/xfs_iops.h
+++ b/fs/xfs/linux-2.6/xfs_iops.h
@@ -48,4 +48,8 @@ extern void linvfs_unwritten_done(struct buffer_head *, int);
 extern int xfs_ioctl(struct bhv_desc *, struct inode *, struct file *,
                         int, unsigned int, void __user *);
 
+struct xfs_inode;
+extern void xfs_ichgtime(struct xfs_inode *, int);
+extern void xfs_ichgtime_fast(struct xfs_inode *, struct inode *, int);
+
 #endif /* __XFS_IOPS_H__ */
diff --git a/fs/xfs/linux-2.6/xfs_lrw.c b/fs/xfs/linux-2.6/xfs_lrw.c
index 3b5fabe8dae9..c04f0c063c58 100644
--- a/fs/xfs/linux-2.6/xfs_lrw.c
+++ b/fs/xfs/linux-2.6/xfs_lrw.c
@@ -302,7 +302,7 @@ xfs_read(
         xfs_iunlock(ip, XFS_IOLOCK_SHARED);
 
         if (likely(!(ioflags & IO_INVIS)))
-                xfs_ichgtime(ip, XFS_ICHGTIME_ACC);
+                xfs_ichgtime_fast(ip, inode, XFS_ICHGTIME_ACC);
 
 unlock_isem:
         if (unlikely(ioflags & IO_ISDIRECT))
@@ -367,7 +367,7 @@ xfs_sendfile(
                 XFS_STATS_ADD(xs_read_bytes, ret);
 
         if (likely(!(ioflags & IO_INVIS)))
-                xfs_ichgtime(ip, XFS_ICHGTIME_ACC);
+                xfs_ichgtime_fast(ip, LINVFS_GET_IP(vp), XFS_ICHGTIME_ACC);
 
         return ret;
 }
@@ -732,15 +732,10 @@ start:
                 }
         }
 
-        /*
-         * On Linux, generic_file_write updates the times even if
-         * no data is copied in so long as the write had a size.
-         *
-         * We must update xfs' times since revalidate will overcopy xfs.
-         */
-        if (!(ioflags & IO_INVIS)) {
-                xfs_ichgtime(xip, XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
+        if (likely(!(ioflags & IO_INVIS))) {
                 inode_update_time(inode, 1);
+                xfs_ichgtime_fast(xip, inode,
+                                  XFS_ICHGTIME_MOD | XFS_ICHGTIME_CHG);
         }
 
         /*
diff --git a/fs/xfs/linux-2.6/xfs_vnode.h b/fs/xfs/linux-2.6/xfs_vnode.h
index 76a8dddb437a..ef1c65f42fe0 100644
--- a/fs/xfs/linux-2.6/xfs_vnode.h
+++ b/fs/xfs/linux-2.6/xfs_vnode.h
@@ -567,13 +567,6 @@ static __inline__ void vn_flagclr(struct vnode *vp, uint flag)
 }
 
 /*
- * Update modify/access/change times on the vnode
- */
-#define VN_MTIMESET(vp, tvp)    (LINVFS_GET_IP(vp)->i_mtime = *(tvp))
-#define VN_ATIMESET(vp, tvp)    (LINVFS_GET_IP(vp)->i_atime = *(tvp))
-#define VN_CTIMESET(vp, tvp)    (LINVFS_GET_IP(vp)->i_ctime = *(tvp))
-
-/*
  * Dealing with bad inodes
  */
 static inline void vn_mark_bad(struct vnode *vp)