I was looking at the /sys/fs/ext4/*/lifetime_write_kbytes counters on
my home server and wondering about how accurate they are. That is most
interesting in the case of flash devices, to get a good idea of the
lifetime writes vs. actual rated drive writes per day.
It looks like s_kbytes_written is only updated on clean unmount
via ext4_commit_super->ext4_update_super() and in a few error handling
codepaths. This means any in-memory updates are typically lost if the
server crashes or loses power (which is typical for long-running servers,
rather than a clean shutdown).
It would be useful to periodically update the superblock with the current
value, maybe once an hour if the value has changed more than some small
margin (to take into account the *previous* update). The superblock used
to be written frequently via ->write_super(), but this has not been the
case since commit v3.5-rc5-19-g4d47603d9703.
Any thoughts/objections to a periodic task calling ext4_update_super()
every hour if there have been any noticeable writes since the last time
it was called? This could potentially be more clever so that it only
writes if the disk is not asleep, and do the writes the next time it wakes,
but I'm not sure how easy/hard that is to detect at the filesystem level.
Cheers, Andreas
PS: there is *also* a function resize.c::ext4_update_super() for added
confusion, but that does something completely different...
On Wed, Nov 02, 2022 at 05:41:07PM -0600, Andreas Dilger wrote:
> I was looking at the /sys/fs/ext4/*/lifetime_write_kbytes counters on
> my home server and wondering about how accurate they are. That is most
> interesting in the case of flash devices, to get a good idea of the
> lifetime writes vs. actual rated drive writes per day.
>
> It looks like s_kbytes_written is only updated on clean unmount
> via ext4_commit_super->ext4_update_super() and in a few error handling
> codepaths. This means any in-memory updates are typically lost if the
> server crashes or loses power (which is typical for long-running servers,
> rather than a clean shutdown).
>
> It would be useful to periodically update the superblock with the current
> value, maybe once an hour if the value has changed more than some small
> margin (to take into account the *previous* update). The superblock used
> to be written frequently via ->write_super(), but this has not been the
> case since commit v3.5-rc5-19-g4d47603d9703.
>
> Any thoughts/objections to a periodic task calling ext4_update_super()
> every hour if there have been any noticeable writes since the last time
> it was called? This could potentially be more clever so that it only
> writes if the disk is not asleep, and do the writes the next time it wakes,
> but I'm not sure how easy/hard that is to detect at the filesystem level.
>
> Cheers, Andreas
>
> PS: there is *also* a function resize.c::ext4_update_super() for added
> confusion, but that does something completely different...
Hi Andreas,
I don't have too much to contribute other than to say I think it's a
good idea. Having the counter be more precise and as such more reliable
is a good thing especially with what looks to me like a litte effort
required.
Thanks!
-Lukas