[RFC,0/3,RESEND] fs: opportunistic high-res file timestamps

  (Apologies for the resend, but I didn't send this with a wide enough
distribution list originally).

A few weeks ago, during one of the discussions around i_version, Dave
Chinner wrote this:

"You've missed the part where I suggested lifting the "nfsd sampled
i_version" state into an inode state flag rather than hiding it in
the i_version field. At that point, we could optimise away the
secondary ctime updates just like you are proposing we do with the
i_version updates.  Further, we could also use that state it to
decide whether we need to use high resolution timestamps when
recording ctime updates - if the nfsd has not sampled the
ctime/i_version, we don't need high res timestamps to be recorded
for ctime...."

While I don't think we can practically optimize away ctime updates
like we do with i_version, I do like the idea of using this scheme to
indicate when we need to use a high-res timestamp.

This patchset is a first stab at a scheme to do this. It declares a new
i_state flag for this purpose and adds two new vfs-layer functions to
implement conditional high-res timestamp fetching. It then converts both
tmpfs and xfs to use it.

This seems to behave fine under xfstests, but I haven't yet done
any performance testing with it. I wouldn't expect it to create huge
regressions though since we're only grabbing high res timestamps after
each query.

I like this scheme because we can potentially convert any filesystem to
use it. No special storage requirements like with i_version field.  I
think it'd potentially improve NFS cache coherency with a whole swath of
exportable filesystems, and helps out NFSv3 too.

This is really just a proof-of-concept. There are a number of things we
could change:

1/ We could use the top bit in the tv_sec field as the flag. That'd give
   us different flags for ctime and mtime. We also wouldn't need to use
   a spinlock.

2/ We could probably optimize away the high-res timestamp fetch in more
   cases. Basically, always do a coarse-grained ts fetch and only fetch
   the high-res ts when the QUERIED flag is set and the existing time
   hasn't changed.

If this approach looks reasonable, I'll plan to start working on
converting more filesystems.

One thing I'm not clear on is how widely available high res timestamps
are. Is this something we need to gate on particular CONFIG_* options?

Thoughts?

Jeff Layton (3):
  fs: add infrastructure for opportunistic high-res ctime/mtime updates
  shmem: mark for high-res timestamps on next update after getattr
  xfs: mark the inode for high-res timestamp update in getattr

 fs/inode.c                      | 40 +++++++++++++++++++++++++++++++--
 fs/stat.c                       | 10 +++++++++
 fs/xfs/libxfs/xfs_trans_inode.c |  2 +-
 fs/xfs/xfs_acl.c                |  2 +-
 fs/xfs/xfs_inode.c              |  2 +-
 fs/xfs/xfs_iops.c               | 15 ++++++++++---
 include/linux/fs.h              |  5 ++++-
 mm/shmem.c                      | 23 ++++++++++---------
 8 files changed, 80 insertions(+), 19 deletions(-)
  

On Tue, Apr 11, 2023 at 10:36:59AM -0400, Jeff Layton wrote:
> (Apologies for the resend, but I didn't send this with a wide enough
> distribution list originally).
> 
> A few weeks ago, during one of the discussions around i_version, Dave
> Chinner wrote this:
> 
> "You've missed the part where I suggested lifting the "nfsd sampled
> i_version" state into an inode state flag rather than hiding it in
> the i_version field. At that point, we could optimise away the
> secondary ctime updates just like you are proposing we do with the
> i_version updates.  Further, we could also use that state it to
> decide whether we need to use high resolution timestamps when
> recording ctime updates - if the nfsd has not sampled the
> ctime/i_version, we don't need high res timestamps to be recorded
> for ctime...."
> 
> While I don't think we can practically optimize away ctime updates
> like we do with i_version, I do like the idea of using this scheme to
> indicate when we need to use a high-res timestamp.
> 
> This patchset is a first stab at a scheme to do this. It declares a new
> i_state flag for this purpose and adds two new vfs-layer functions to
> implement conditional high-res timestamp fetching. It then converts both
> tmpfs and xfs to use it.
> 
> This seems to behave fine under xfstests, but I haven't yet done
> any performance testing with it. I wouldn't expect it to create huge
> regressions though since we're only grabbing high res timestamps after
> each query.
> 
> I like this scheme because we can potentially convert any filesystem to
> use it. No special storage requirements like with i_version field.  I
> think it'd potentially improve NFS cache coherency with a whole swath of
> exportable filesystems, and helps out NFSv3 too.
> 
> This is really just a proof-of-concept. There are a number of things we
> could change:
> 
> 1/ We could use the top bit in the tv_sec field as the flag. That'd give
>    us different flags for ctime and mtime. We also wouldn't need to use
>    a spinlock.
> 
> 2/ We could probably optimize away the high-res timestamp fetch in more
>    cases. Basically, always do a coarse-grained ts fetch and only fetch
>    the high-res ts when the QUERIED flag is set and the existing time
>    hasn't changed.
> 
> If this approach looks reasonable, I'll plan to start working on
> converting more filesystems.

Seems reasonable to me. In terms of testing, I suspect the main
impact is going to be the additionaly overhead of taking a spinlock
in normal stat calls. In which case, testing common tools like giti
would be useful.  e.g. `git status` runs about 170k stat calls on a
typical kernel tree. If anything is going to be noticed by users
that actually care, it'll be workloads like this...

If we manage to elide the spinlock altogether, then I don't think
we're going to be able to measure any sort perf difference on modern
hardware short of high end NFS benchmarks that drive servers to
their CPU usage limits....

> One thing I'm not clear on is how widely available high res timestamps
> are. Is this something we need to gate on particular CONFIG_* options?

Don't think so - the kernel should always provide the highest
resoultion it can through the get_time interfaces - the _coarse
variants simple return what was read from the high res timer at the
last scheduler tick, hence avoiding the hardware timer overhead when
high res timer resolution is not needed.....

Cheers,

Dave.
  

On Tue, Apr 11, 2023 at 5:38 PM Jeff Layton <jlayton@kernel.org> wrote:
>
> (Apologies for the resend, but I didn't send this with a wide enough
> distribution list originally).
>
> A few weeks ago, during one of the discussions around i_version, Dave
> Chinner wrote this:
>
> "You've missed the part where I suggested lifting the "nfsd sampled
> i_version" state into an inode state flag rather than hiding it in
> the i_version field. At that point, we could optimise away the
> secondary ctime updates just like you are proposing we do with the
> i_version updates.  Further, we could also use that state it to
> decide whether we need to use high resolution timestamps when
> recording ctime updates - if the nfsd has not sampled the
> ctime/i_version, we don't need high res timestamps to be recorded
> for ctime...."
>
> While I don't think we can practically optimize away ctime updates
> like we do with i_version, I do like the idea of using this scheme to
> indicate when we need to use a high-res timestamp.
>
> This patchset is a first stab at a scheme to do this. It declares a new
> i_state flag for this purpose and adds two new vfs-layer functions to
> implement conditional high-res timestamp fetching. It then converts both
> tmpfs and xfs to use it.
>
> This seems to behave fine under xfstests, but I haven't yet done
> any performance testing with it. I wouldn't expect it to create huge
> regressions though since we're only grabbing high res timestamps after
> each query.
>
> I like this scheme because we can potentially convert any filesystem to
> use it. No special storage requirements like with i_version field.  I
> think it'd potentially improve NFS cache coherency with a whole swath of
> exportable filesystems, and helps out NFSv3 too.
>
> This is really just a proof-of-concept. There are a number of things we
> could change:
>
> 1/ We could use the top bit in the tv_sec field as the flag. That'd give
>    us different flags for ctime and mtime. We also wouldn't need to use
>    a spinlock.
>
> 2/ We could probably optimize away the high-res timestamp fetch in more
>    cases. Basically, always do a coarse-grained ts fetch and only fetch
>    the high-res ts when the QUERIED flag is set and the existing time
>    hasn't changed.
>
> If this approach looks reasonable, I'll plan to start working on
> converting more filesystems.
>
> One thing I'm not clear on is how widely available high res timestamps
> are. Is this something we need to gate on particular CONFIG_* options?
>
> Thoughts?

Jeff,

Considering that this proposal is pretty uncontroversial,
do you still want to discuss/lead a session on i_version changes in LSF/MM?

I noticed that Chuck listed "timespamt resolution and i_version" as part
of his NFSD BoF topic proposal [1], but I do not think all of these topics
can fit in one 30 minute session.

Dave,

I would like to use this opportunity to invite you and any developers that
are involved in fs development and not going to attend LSF/MM in-person,
to join LSF/MM virtually for some sessions that you may be interested in.
See this lore query [2] for TOPICs proposed this year.

You can let me know privately which sessions you are interested in
attending and your time zone and I will do my best to schedule those
sessions in time slots that would be more convenient for your time zone.

Obviously, I am referring to FS track sessions.
Cross track sessions are going to be harder to accommodate,
but I can try.

Thanks,
Amir.

[1] https://lore.kernel.org/linux-fsdevel/FF0202C3-7500-4BB3-914B-DBAA3E0EA3D7@oracle.com/
[2] https://lore.kernel.org/linux-fsdevel/?q=LSF+TOPIC+-re+d%3A4.months.ago..
  

On Sat, 2023-04-15 at 14:35 +0300, Amir Goldstein wrote:
> On Tue, Apr 11, 2023 at 5:38 PM Jeff Layton <jlayton@kernel.org> wrote:
> > 
> > (Apologies for the resend, but I didn't send this with a wide enough
> > distribution list originally).
> > 
> > A few weeks ago, during one of the discussions around i_version, Dave
> > Chinner wrote this:
> > 
> > "You've missed the part where I suggested lifting the "nfsd sampled
> > i_version" state into an inode state flag rather than hiding it in
> > the i_version field. At that point, we could optimise away the
> > secondary ctime updates just like you are proposing we do with the
> > i_version updates.  Further, we could also use that state it to
> > decide whether we need to use high resolution timestamps when
> > recording ctime updates - if the nfsd has not sampled the
> > ctime/i_version, we don't need high res timestamps to be recorded
> > for ctime...."
> > 
> > While I don't think we can practically optimize away ctime updates
> > like we do with i_version, I do like the idea of using this scheme to
> > indicate when we need to use a high-res timestamp.
> > 
> > This patchset is a first stab at a scheme to do this. It declares a new
> > i_state flag for this purpose and adds two new vfs-layer functions to
> > implement conditional high-res timestamp fetching. It then converts both
> > tmpfs and xfs to use it.
> > 
> > This seems to behave fine under xfstests, but I haven't yet done
> > any performance testing with it. I wouldn't expect it to create huge
> > regressions though since we're only grabbing high res timestamps after
> > each query.
> > 
> > I like this scheme because we can potentially convert any filesystem to
> > use it. No special storage requirements like with i_version field.  I
> > think it'd potentially improve NFS cache coherency with a whole swath of
> > exportable filesystems, and helps out NFSv3 too.
> > 
> > This is really just a proof-of-concept. There are a number of things we
> > could change:
> > 
> > 1/ We could use the top bit in the tv_sec field as the flag. That'd give
> >    us different flags for ctime and mtime. We also wouldn't need to use
> >    a spinlock.
> > 
> > 2/ We could probably optimize away the high-res timestamp fetch in more
> >    cases. Basically, always do a coarse-grained ts fetch and only fetch
> >    the high-res ts when the QUERIED flag is set and the existing time
> >    hasn't changed.
> > 
> > If this approach looks reasonable, I'll plan to start working on
> > converting more filesystems.
> > 
> > One thing I'm not clear on is how widely available high res timestamps
> > are. Is this something we need to gate on particular CONFIG_* options?
> > 
> > Thoughts?
> 
> Jeff,
> 
> Considering that this proposal is pretty uncontroversial,
> do you still want to discuss/lead a session on i_version changes in LSF/MM?
> 
> I noticed that Chuck listed "timespamt resolution and i_version" as part
> of his NFSD BoF topic proposal [1], but I do not think all of these topics
> can fit in one 30 minute session.
> 

Agreed. I think we'll need an hour for the nfsd BoF.

I probably don't need a full 30 min slot for this topic, between the
nfsd BoF and hallway track.

I've started a TOPIC email for this about 5 times now, and keep deleting
it. I think most of the more controversial bits are pretty much settled
at this point, and the rest (crash resilience) is still too embryonic
for discussion.

I might want a lightning talk at some point about what I'd _really_ like
to do long term with the i_version counter (basically: I want to be able
to do a write that is gated on the i_version not having changed).


> Dave,
> 
> I would like to use this opportunity to invite you and any developers that
> are involved in fs development and not going to attend LSF/MM in-person,
> to join LSF/MM virtually for some sessions that you may be interested in.
> See this lore query [2] for TOPICs proposed this year.
> 
> You can let me know privately which sessions you are interested in
> attending and your time zone and I will do my best to schedule those
> sessions in time slots that would be more convenient for your time zone.
> 
> Obviously, I am referring to FS track sessions.
> Cross track sessions are going to be harder to accommodate,
> but I can try.
> 
> Thanks,
> Amir.
> 
> [1] https://lore.kernel.org/linux-fsdevel/FF0202C3-7500-4BB3-914B-DBAA3E0EA3D7@oracle.com/
> [2] https://lore.kernel.org/linux-fsdevel/?q=LSF+TOPIC+-re+d%3A4.months.ago..
  

> On Apr 15, 2023, at 7:35 AM, Amir Goldstein <amir73il@gmail.com> wrote:
> 
> On Tue, Apr 11, 2023 at 5:38 PM Jeff Layton <jlayton@kernel.org> wrote:
>> 
>> (Apologies for the resend, but I didn't send this with a wide enough
>> distribution list originally).
>> 
>> A few weeks ago, during one of the discussions around i_version, Dave
>> Chinner wrote this:
>> 
>> "You've missed the part where I suggested lifting the "nfsd sampled
>> i_version" state into an inode state flag rather than hiding it in
>> the i_version field. At that point, we could optimise away the
>> secondary ctime updates just like you are proposing we do with the
>> i_version updates.  Further, we could also use that state it to
>> decide whether we need to use high resolution timestamps when
>> recording ctime updates - if the nfsd has not sampled the
>> ctime/i_version, we don't need high res timestamps to be recorded
>> for ctime...."
>> 
>> While I don't think we can practically optimize away ctime updates
>> like we do with i_version, I do like the idea of using this scheme to
>> indicate when we need to use a high-res timestamp.
>> 
>> This patchset is a first stab at a scheme to do this. It declares a new
>> i_state flag for this purpose and adds two new vfs-layer functions to
>> implement conditional high-res timestamp fetching. It then converts both
>> tmpfs and xfs to use it.
>> 
>> This seems to behave fine under xfstests, but I haven't yet done
>> any performance testing with it. I wouldn't expect it to create huge
>> regressions though since we're only grabbing high res timestamps after
>> each query.
>> 
>> I like this scheme because we can potentially convert any filesystem to
>> use it. No special storage requirements like with i_version field.  I
>> think it'd potentially improve NFS cache coherency with a whole swath of
>> exportable filesystems, and helps out NFSv3 too.
>> 
>> This is really just a proof-of-concept. There are a number of things we
>> could change:
>> 
>> 1/ We could use the top bit in the tv_sec field as the flag. That'd give
>>   us different flags for ctime and mtime. We also wouldn't need to use
>>   a spinlock.
>> 
>> 2/ We could probably optimize away the high-res timestamp fetch in more
>>   cases. Basically, always do a coarse-grained ts fetch and only fetch
>>   the high-res ts when the QUERIED flag is set and the existing time
>>   hasn't changed.
>> 
>> If this approach looks reasonable, I'll plan to start working on
>> converting more filesystems.
>> 
>> One thing I'm not clear on is how widely available high res timestamps
>> are. Is this something we need to gate on particular CONFIG_* options?
>> 
>> Thoughts?
> 
> Jeff,
> 
> Considering that this proposal is pretty uncontroversial,
> do you still want to discuss/lead a session on i_version changes in LSF/MM?
> 
> I noticed that Chuck listed "timespamt resolution and i_version" as part
> of his NFSD BoF topic proposal [1], but I do not think all of these topics
> can fit in one 30 minute session.

That's fair.

If lumping these topics together doesn't seem sensible,
I'm happy to consider splitting off the major topics,
and then including the remaining in a generic network
filesystem session or relegating them to the hallway
track.

I can suggest something more specific in the LSF TOPIC
thread.


--
Chuck Lever