From: Dave Chinner <david@fromorbit.com> Date: 2016-01-04 06:22:33
Hi folks,
I've been looking at implementing the lazytime mount option for XFS,
and I'm struggling to work out what it is supposed to mean.
AFAICT, on ext4, lazytime means that pure timestamp updates are not
journalled and they are only ever written back when the inode is
otherwise dirtied and written, or they are timestamp dirty for 24
hours which triggers writeback.
This poses a couple of problems for XFS:
1. we log every timestamp change, so there is no mechanism
for delayed/deferred update.
2. we track dirty metadata in the journal, not via the VFS
dirty inode lists, so all the infrastructure written for
ext4 to do periodic flushing is useless to us.
These are solvable problems, but what I'm not sure about is exactly
what the intended semantics of lazytime durability are. That is,
exactly what guaranteed are we giving userspace about timestamp
updates when lazytime is used? The guarantees we have to give will
greatly influence the XFS implementation, so I really need to nail
down what we are expected to provide userspace. Can we:
a) just ignore all durability concerns?
b) if not, do we only need to care about the 24 hour
writeback and unmount?
c) if not, are fsync/sync/syncfs/freeze/unmount supposed
to provide durability of all metadata changes?
d) do we have to care about ordering - if we fsync one inode
with 1 hour old timestamps, do we also need to guarantee
that all the inodes with older dirty timestamps also get
made durable?
I really want to completely ignore all ordering and forced
durability requirements for lazytime (i.e. implement only periodic,
optimistic and freeze/unmount writeback), but I haven't found any
documentation of what durability lazytime is supposed to provide and
can really only guess what was intended from the ext4
implementation....
Clarity would be appreciated.
Cheers,
Dave.
--
Dave Chinner
david-FqsqvQoI3Ljby3iVrkZq2A@public.gmane.org
From: Jan Kara <jack@suse.cz> Date: 2016-01-05 17:35:55
Hi,
On Mon 04-01-16 17:22:19, Dave Chinner wrote:
I've been looking at implementing the lazytime mount option for XFS,
and I'm struggling to work out what it is supposed to mean.
AFAICT, on ext4, lazytime means that pure timestamp updates are not
journalled and they are only ever written back when the inode is
otherwise dirtied and written, or they are timestamp dirty for 24
hours which triggers writeback.
This poses a couple of problems for XFS:
1. we log every timestamp change, so there is no mechanism
for delayed/deferred update.
2. we track dirty metadata in the journal, not via the VFS
dirty inode lists, so all the infrastructure written for
ext4 to do periodic flushing is useless to us.
These are solvable problems, but what I'm not sure about is exactly
what the intended semantics of lazytime durability are. That is,
exactly what guaranteed are we giving userspace about timestamp
updates when lazytime is used? The guarantees we have to give will
greatly influence the XFS implementation, so I really need to nail
down what we are expected to provide userspace. Can we:
a) just ignore all durability concerns?
b) if not, do we only need to care about the 24 hour
writeback and unmount?
c) if not, are fsync/sync/syncfs/freeze/unmount supposed
to provide durability of all metadata changes?
d) do we have to care about ordering - if we fsync one inode
with 1 hour old timestamps, do we also need to guarantee
that all the inodes with older dirty timestamps also get
made durable?
So the intended semantics is:
1) fsync / sync / freeze / unmount will write the timestamp updates even
with lazytime. So unless crash happens, timestamps are guaranteed to be
consistent. Also sync / fsync guarantees all changes to get to disk.
2) We periodically write back timestamps (once per 24 hours) to avoid too
big timestamp inconsistencies in case of crash.
Otherwise there are no guarantees wrt durability. I've CCed Ted who
designed this just in case I missed something.
Honza
--
Jan Kara [off-list ref]
SUSE Labs, CR
From: Dave Chinner <david@fromorbit.com> Date: 2016-01-05 22:59:12
On Tue, Jan 05, 2016 at 06:36:04PM +0100, Jan Kara wrote:
Hi,
On Mon 04-01-16 17:22:19, Dave Chinner wrote:
quoted
I've been looking at implementing the lazytime mount option for XFS,
and I'm struggling to work out what it is supposed to mean.
AFAICT, on ext4, lazytime means that pure timestamp updates are not
journalled and they are only ever written back when the inode is
otherwise dirtied and written, or they are timestamp dirty for 24
hours which triggers writeback.
This poses a couple of problems for XFS:
1. we log every timestamp change, so there is no mechanism
for delayed/deferred update.
2. we track dirty metadata in the journal, not via the VFS
dirty inode lists, so all the infrastructure written for
ext4 to do periodic flushing is useless to us.
These are solvable problems, but what I'm not sure about is exactly
what the intended semantics of lazytime durability are. That is,
exactly what guaranteed are we giving userspace about timestamp
updates when lazytime is used? The guarantees we have to give will
greatly influence the XFS implementation, so I really need to nail
down what we are expected to provide userspace. Can we:
a) just ignore all durability concerns?
b) if not, do we only need to care about the 24 hour
writeback and unmount?
c) if not, are fsync/sync/syncfs/freeze/unmount supposed
to provide durability of all metadata changes?
d) do we have to care about ordering - if we fsync one inode
with 1 hour old timestamps, do we also need to guarantee
that all the inodes with older dirty timestamps also get
made durable?
So the intended semantics is:
1) fsync / sync / freeze / unmount will write the timestamp updates even
with lazytime. So unless crash happens, timestamps are guaranteed to be
consistent. Also sync / fsync guarantees all changes to get to disk.
2) We periodically write back timestamps (once per 24 hours) to avoid too
big timestamp inconsistencies in case of crash.
Ok, so it's supposed to be a delayed timestamp update mechanism
without any specific ordering guarantees, not an opportunistic
timestamp update mechanism.
I can work with that.
Cheers,
Dave.
--
Dave Chinner
david-FqsqvQoI3Ljby3iVrkZq2A@public.gmane.org
On Wed, Jan 06, 2016 at 09:59:07AM +1100, Dave Chinner wrote:
quoted
So the intended semantics is:
1) fsync / sync / freeze / unmount will write the timestamp updates even
with lazytime. So unless crash happens, timestamps are guaranteed to be
consistent. Also sync / fsync guarantees all changes to get to disk.
2) We periodically write back timestamps (once per 24 hours) to avoid too
big timestamp inconsistencies in case of crash.
Ok, so it's supposed to be a delayed timestamp update mechanism
without any specific ordering guarantees, not an opportunistic
timestamp update mechanism.
There is an optimization which ext4 has which will update related
timestamps when we write an inode table block, which is
"opportunistic", but there is no guarantee that this will happen.
This is purely optional; other file systems don't have to do this, but
it can be a win in that if related inodes are in the same 4k block,
and we need to update, say, the index file one because we are changing
i_size, but we were also doing non-allocating writes to the data file,
then we might as well write out the timestamps for the data file at
the same time, since this is "free".
- Ted
From: Dave Chinner <david@fromorbit.com> Date: 2016-01-07 02:21:44
On Wed, Jan 06, 2016 at 08:05:06PM -0500, Theodore Ts'o wrote:
On Wed, Jan 06, 2016 at 09:59:07AM +1100, Dave Chinner wrote:
quoted
quoted
So the intended semantics is:
1) fsync / sync / freeze / unmount will write the timestamp updates even
with lazytime. So unless crash happens, timestamps are guaranteed to be
consistent. Also sync / fsync guarantees all changes to get to disk.
2) We periodically write back timestamps (once per 24 hours) to avoid too
big timestamp inconsistencies in case of crash.
Ok, so it's supposed to be a delayed timestamp update mechanism
without any specific ordering guarantees, not an opportunistic
timestamp update mechanism.
There is an optimization which ext4 has which will update related
timestamps when we write an inode table block, which is
"opportunistic", but there is no guarantee that this will happen.
XFS used to do that, too, before we removed all that hackery when we
moved to logging timestamp updates unconditionally a few years ago.
I'm going to have to re-instate some of that code for lazytime, I
think.
This is purely optional; other file systems don't have to do this, but
it can be a win in that if related inodes are in the same 4k block,
and we need to update, say, the index file one because we are changing
i_size, but we were also doing non-allocating writes to the data file,
then we might as well write out the timestamps for the data file at
the same time, since this is "free".
*nod*. Explicit, optimised clustered inode writeback (rather than
purely opportunistic clustering via delayed buffer writeback) was
added to XFS way back in early 1999. :)
Cheers,
Dave.
--
Dave Chinner
david-FqsqvQoI3Ljby3iVrkZq2A@public.gmane.org