[PATCH-v2,3/5] vfs: don't let the dirty time inodes get more than a day stale
diff mbox

Message ID 1416675267-2191-4-git-send-email-tytso@mit.edu
State Not Applicable
Headers show

Commit Message

Theodore Y. Ts'o Nov. 22, 2014, 4:54 p.m. UTC
Guarantee that the on-disk timestamps will be no more than 24 hours
stale.

Signed-off-by: Theodore Ts'o <tytso@mit.edu>
---
 fs/fs-writeback.c  |  1 +
 fs/inode.c         | 16 +++++++++++++++-
 include/linux/fs.h |  1 +
 3 files changed, 17 insertions(+), 1 deletion(-)

Comments

Rasmus Villemoes Nov. 24, 2014, 12:27 p.m. UTC | #1
On Sat, Nov 22 2014, Theodore Ts'o <tytso@mit.edu> wrote:

> Guarantee that the on-disk timestamps will be no more than 24 hours
> stale.
>
>  static int update_time(struct inode *inode, struct timespec *time, int flags)
>  {
> +	unsigned short days_since_boot = jiffies / (HZ * 86400);
>  	int ret;
>

This seems to wrap every 49 days (assuming 32 bit jiffies and HZ==1000),
so on-disk updates can be delayed indefinitely, assuming just the right
delays between writes.

>  	if (inode->i_op->update_time) {
> @@ -1527,14 +1528,27 @@ static int update_time(struct inode *inode, struct timespec *time, int flags)
>  		if (flags & S_MTIME)
>  			inode->i_mtime = *time;
>  	}
> -	if (inode->i_sb->s_flags & MS_LAZYTIME) {
> +	/*
> +	 * If i_ts_dirty_day is zero, then either we have not deferred
> +	 * timestamp updates, or the system has been up for less than
> +	 * a day (so days_since_boot is zero), so we defer timestamp
> +	 * updates in that case and set the I_DIRTY_TIME flag.  If a
> +	 * day or more has passed, then i_ts_dirty_day will be
> +	 * different from days_since_boot, and then we should update
> +	 * the on-disk inode and then we can clear i_ts_dirty_day.
> +	 */

AFAICT days_since_boot is not actually 0 immediately after boot
due to 

#define INITIAL_JIFFIES ((unsigned long)(unsigned int) (-300*HZ))

On 32 bit platforms, days_since_boot will be 0 shortly after, while on
64 bit it will always be >= 49. Not exactly sure how this affects the
above logic.

Would it make sense to introduce days_since_boot as a global variable
and avoid these issues? This would presumably also make update_time a
few cycles faster (avoiding a division-by-constant), but not sure if
that's important. And something of course needs to update
days_since_boot, but that should be doable.

Rasmus

--
To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Theodore Y. Ts'o Nov. 24, 2014, 5:10 p.m. UTC | #2
On Mon, Nov 24, 2014 at 01:27:21PM +0100, Rasmus Villemoes wrote:
> On Sat, Nov 22 2014, Theodore Ts'o <tytso@mit.edu> wrote:
> 
> > Guarantee that the on-disk timestamps will be no more than 24 hours
> > stale.
> >
> > +	unsigned short days_since_boot = jiffies / (HZ * 86400);
> 
> This seems to wrap every 49 days (assuming 32 bit jiffies and HZ==1000),
> so on-disk updates can be delayed indefinitely, assuming just the right
> delays between writes.

Good point, I'll fix this.

> Would it make sense to introduce days_since_boot as a global variable
> and avoid these issues? This would presumably also make update_time a
> few cycles faster (avoiding a division-by-constant), but not sure if
> that's important. And something of course needs to update
> days_since_boot, but that should be doable.

I can do this fairly simply like this:

	get_monotonic_boottime(&uptime);
	daycode = uptime.tv_sec / (HZ * 86400);

and we only need to do this if lazytime is set, and the inode isn't
marked as I_DIRTY_TIME:

	if ((inode->i_sb->s_flags & MS_LAZYTIME) &&
	    !(flags & S_VERSION)) {
		if (inode->i_state & I_DIRTY_TIME)
			return 0;
		get_monotonic_boottime(&uptime);
		daycode = do_div64(uptime.tv_sec do_div, (HZ * 86400));
		if (!inode->i_ts_dirty_day ||
		    inode->i_ts_dirty_day == daycode) {
			spin_lock(&inode->i_lock);
			inode->i_state |= I_DIRTY_TIME;
			spin_unlock(&inode->i_lock);
			inode->i_ts_dirty_day = daycode;
			return 0;
		}
	}

So I'm not entirely sure it's worth it to create a global variable for
days since boot; I've been runnin with this patch in my laptop, we
wouldn't be triggering the get_monotonic_bootime() function all that
often.  (Since once the dirty_time flg is set, we don't need to check
about whether we need to set it again.)  And if we *did* care, it
would be simple enough to use a static counter which only recalculates
daycode every 30 or 60 minutes.


Cheers,

							- Ted
--
To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Patch
diff mbox

diff --git a/fs/fs-writeback.c b/fs/fs-writeback.c
index ce7de22..eb04277 100644
--- a/fs/fs-writeback.c
+++ b/fs/fs-writeback.c
@@ -1141,6 +1141,7 @@  void __mark_inode_dirty(struct inode *inode, int flags)
 	if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) {
 		trace_writeback_dirty_inode_start(inode, flags);
 
+		inode->i_ts_dirty_day = 0;
 		if (sb->s_op->dirty_inode)
 			sb->s_op->dirty_inode(inode, flags);
 
diff --git a/fs/inode.c b/fs/inode.c
index 11fe81b..0d939a8 100644
--- a/fs/inode.c
+++ b/fs/inode.c
@@ -1511,6 +1511,7 @@  static int relatime_need_update(struct vfsmount *mnt, struct inode *inode,
  */
 static int update_time(struct inode *inode, struct timespec *time, int flags)
 {
+	unsigned short days_since_boot = jiffies / (HZ * 86400);
 	int ret;
 
 	if (inode->i_op->update_time) {
@@ -1527,14 +1528,27 @@  static int update_time(struct inode *inode, struct timespec *time, int flags)
 		if (flags & S_MTIME)
 			inode->i_mtime = *time;
 	}
-	if (inode->i_sb->s_flags & MS_LAZYTIME) {
+	/*
+	 * If i_ts_dirty_day is zero, then either we have not deferred
+	 * timestamp updates, or the system has been up for less than
+	 * a day (so days_since_boot is zero), so we defer timestamp
+	 * updates in that case and set the I_DIRTY_TIME flag.  If a
+	 * day or more has passed, then i_ts_dirty_day will be
+	 * different from days_since_boot, and then we should update
+	 * the on-disk inode and then we can clear i_ts_dirty_day.
+	 */
+	if ((inode->i_sb->s_flags & MS_LAZYTIME) &&
+	    (!inode->i_ts_dirty_day ||
+	     inode->i_ts_dirty_day == days_since_boot)) {
 		if (inode->i_state & I_DIRTY_TIME)
 			return 0;
 		spin_lock(&inode->i_lock);
 		inode->i_state |= I_DIRTY_TIME;
 		spin_unlock(&inode->i_lock);
+		inode->i_ts_dirty_day = days_since_boot;
 		return 0;
 	}
+	inode->i_ts_dirty_day = 0;
 	if (inode->i_op->write_time)
 		return inode->i_op->write_time(inode);
 	mark_inode_dirty_sync(inode);
diff --git a/include/linux/fs.h b/include/linux/fs.h
index 489b2f2..e3574cd 100644
--- a/include/linux/fs.h
+++ b/include/linux/fs.h
@@ -575,6 +575,7 @@  struct inode {
 	struct timespec		i_ctime;
 	spinlock_t		i_lock;	/* i_blocks, i_bytes, maybe i_size */
 	unsigned short          i_bytes;
+	unsigned short		i_ts_dirty_day;
 	unsigned int		i_blkbits;
 	blkcnt_t		i_blocks;