diff mbox series

[RFC,2/9] timekeeping: new interfaces for multigrain timestamp handing

Message ID 20231018-mgtime-v1-2-4a7a97b1f482@kernel.org (mailing list archive)
State New, archived
Headers show
Series fs: multigrain timestamps (redux) | expand

Commit Message

Jeffrey Layton Oct. 18, 2023, 5:41 p.m. UTC
Multigrain timestamps allow VFS to request fine-grained timestamps when
there has been recent interest in the file's attributes. Unfortunately,
the coarse-grained timestamps can lag the fine-grained ones. A file
stamped with a coarse-grained timestamp after one with a fine-grained
one can appear to have been modified in reverse order. This is
problematic for programs like "make" or "rsync" which depend on being
able to compare timestamps between two different inodes.

One way to prevent this is to ensure that when we stamp a file with a
fine-grained timestamp, that we use that value to establish a floor for
any later timestamp update.

Add a new mg_nsec field to the timekeeper structure for tracking the
nsec value of the most recent multigrain timestamp, along with two new
helper functions for fetching coarse and fine grained timestamps. When
getting a fine-grained timestamp update the mg_nsec with the value that
was fetched via timekeeping_get_ns.

When fetching a coarse-grained timestamp, set the mg_nsec to
TK_MG_NSEC_IGNORE. When fetching a coarse-grained time for a multigrain
timestamp, apply the mg_nsec value if it's not set to TK_MG_NSEC_IGNORE.

Signed-off-by: Jeff Layton <jlayton@kernel.org>
---
 include/linux/timekeeper_internal.h |  2 +
 include/linux/timekeeping.h         |  4 ++
 kernel/time/timekeeping.c           | 80 +++++++++++++++++++++++++++++++++++++
 3 files changed, 86 insertions(+)

Comments

Linus Torvalds Oct. 18, 2023, 7:18 p.m. UTC | #1
On Wed, 18 Oct 2023 at 10:41, Jeff Layton <jlayton@kernel.org> wrote:
>
> One way to prevent this is to ensure that when we stamp a file with a
> fine-grained timestamp, that we use that value to establish a floor for
> any later timestamp update.

I'm very leery of this.

I don't like how it's using a global time - and a global fine-grained
offset - when different filesystems will very naturally have different
granularities. I also don't like how it's no using that global lock.

Yes, yes, since the use of this all is then gated by the 'is_mgtime()'
thing, any filesystem with big granularities will presumably never set
FS_MGTIME in the first time, and that hides the worst pointless cases.
But it still feels iffy to me.

Also, the whole current_ctime() logic seems wrong. Later (in 4/9), you do this:

 static struct timespec64 current_ctime(struct inode *inode)
 {
        if (is_mgtime(inode))
                return current_mgtime(inode);

and current_mgtime() does

        if (nsec & I_CTIME_QUERIED) {
                ktime_get_real_ts64(&now);
                return timestamp_truncate(now, inode);
        }

so once the code has set I_CTIME_QUERIED, it will now use the
expensive fine-grained time - even when it makes no sense.

As far as I can tell, there is *never* a reason to get the
fine-grained time if the old inode ctime is already sufficiently far
away.

IOW, my gut feel is that all this logic should always not only be
guarded by FS_MGTIME (like you do seem to do), *and* by "has anybody
even queried this time" - it should *also* always be guarded by "if I
get the coarse-grained time, is that sufficient?"

So I think the current_ctime() logic should be something along the lines of

    static struct timespec64 current_ctime(struct inode *inode)
    {
        struct timespec64 ts64 = current_time(inode);
        unsigned long old_ctime_sec = inode->i_ctime_sec;
        unsigned int old_ctime_nsec = inode->i_ctime_nsec;

        if (ts64.tv_sec != old_ctime_sec)
                return ts64;

        /*
         * No need to check is_mgtime(inode) - the I_CTIME_QUERIED
         * flag is only set for mgtime filesystems
         */
        if (!(old_ctime_nsec & I_CTIME_QUERIED))
                return ts64;
        old_ctime_nsec &= ~I_CTIME_QUERIED;
        if (ts64.tv_nsec > old_ctime_nsec + inode->i_sb->s_time_gran)
                return ts64;

        /* Ok, only *now* do we do a finegrained value */
        ktime_get_real_ts64(&ts64);
        return timestamp_truncate(ts64);
    }

or whatever. Make it *very* clear that the finegrained timestamp is
the absolute last option, after we have checked that the regular one
isn't possible.

I dunno.

            Linus
Jeffrey Layton Oct. 18, 2023, 8:47 p.m. UTC | #2
On Wed, 2023-10-18 at 12:18 -0700, Linus Torvalds wrote:
> On Wed, 18 Oct 2023 at 10:41, Jeff Layton <jlayton@kernel.org> wrote:
> > 
> > One way to prevent this is to ensure that when we stamp a file with a
> > fine-grained timestamp, that we use that value to establish a floor for
> > any later timestamp update.
> 
> I'm very leery of this.
> 
> I don't like how it's using a global time - and a global fine-grained
> offset - when different filesystems will very naturally have different
> granularities. I also don't like how it's no using that global lock.
> 
> Yes, yes, since the use of this all is then gated by the 'is_mgtime()'
> thing, any filesystem with big granularities will presumably never set
> FS_MGTIME in the first time, and that hides the worst pointless cases.
> But it still feels iffy to me.
> 

Thanks for taking a look!

I'm not too crazy about the global lock either, but the global fine
grained value ensures that when we have mtime changes that occur across
filesystems that they appear to be in the correct order.

We could (hypothetically) track an offset per superblock or something,
but then you could see out-of-order timestamps in inodes across
different filesystems (even of the same type). I think it'd better not
to do that if we can get away with it.


> Also, the whole current_ctime() logic seems wrong. Later (in 4/9), you do this:
> 
>  static struct timespec64 current_ctime(struct inode *inode)
>  {
>         if (is_mgtime(inode))
>                 return current_mgtime(inode);
> 
> and current_mgtime() does
> 
>         if (nsec & I_CTIME_QUERIED) {
>                 ktime_get_real_ts64(&now);
>                 return timestamp_truncate(now, inode);
>         }
> 
> so once the code has set I_CTIME_QUERIED, it will now use the
> expensive fine-grained time - even when it makes no sense.
> 
> As far as I can tell, there is *never* a reason to get the
> fine-grained time if the old inode ctime is already sufficiently far
> away.
> 
> IOW, my gut feel is that all this logic should always not only be
> guarded by FS_MGTIME (like you do seem to do), *and* by "has anybody
> even queried this time" - it should *also* always be guarded by "if I
> get the coarse-grained time, is that sufficient?"
> 
> So I think the current_ctime() logic should be something along the lines of
> 
>     static struct timespec64 current_ctime(struct inode *inode)
>     {
>         struct timespec64 ts64 = current_time(inode);
>         unsigned long old_ctime_sec = inode->i_ctime_sec;
>         unsigned int old_ctime_nsec = inode->i_ctime_nsec;
> 
>         if (ts64.tv_sec != old_ctime_sec)
>                 return ts64;
> 
>         /*
>          * No need to check is_mgtime(inode) - the I_CTIME_QUERIED
>          * flag is only set for mgtime filesystems
>          */
>         if (!(old_ctime_nsec & I_CTIME_QUERIED))
>                 return ts64;
>         old_ctime_nsec &= ~I_CTIME_QUERIED;
>         if (ts64.tv_nsec > old_ctime_nsec + inode->i_sb->s_time_gran)
>                 return ts64;
> 

Does that really do what you expect here? current_time will return a
value that has already been through timestamp_truncate. Regardless, I
don't think this function makes as big a difference as you might think.

>
>         /* Ok, only *now* do we do a finegrained value */
>         ktime_get_real_ts64(&ts64);
>         return timestamp_truncate(ts64);
>     }
> 
> or whatever. Make it *very* clear that the finegrained timestamp is
> the absolute last option, after we have checked that the regular one
> isn't possible.

current_mgtime is calling ktime_get_real_ts64, which is an existing
interface that does not take the global spinlock and won't advance the
global offset. That call should be quite cheap.

The reason we can use that call here is because current_ctime and
current_mgtime are only called from  inode_needs_update_time, which is
only called to check whether we need to get write access to the
inode. What we do is look at the current clock and see whether the
timestamps would perceivably change if we were to do the update right
then.

If so, we get write access and then call inode_set_ctime_current(). That
will fetch its own timestamps and reevaluate what sort of update to do.
That's the only place that fetches an expensive fine-grained timestamp
that advances the offset.

So, I think this set already is only getting the expensive fine-grained
timestamps as a last resort.

This is probably an indicator that I need to document this code better
though. It may also be a good idea to reorganize
inode_needs_update_time, current_ctime and current_mgtime for better
clarity.

Many thanks for the comments!
Linus Torvalds Oct. 18, 2023, 9:31 p.m. UTC | #3
On Wed, 18 Oct 2023 at 13:47, Jeff Layton <jlayton@kernel.org> wrote:
>
> >         old_ctime_nsec &= ~I_CTIME_QUERIED;
> >         if (ts64.tv_nsec > old_ctime_nsec + inode->i_sb->s_time_gran)
> >                 return ts64;
> >
>
> Does that really do what you expect here? current_time will return a
> value that has already been through timestamp_truncate.

Yeah, you're right.

I think you can actually remove the s_time_gran addition. Both the
old_ctime_nsec and the current ts64,tv_nsec are already rounded, so
just doing

        if (ts64.tv_nsec > old_ctime_nsec)
                return ts64;

would already guarantee that it's different enough.

> current_mgtime is calling ktime_get_real_ts64, which is an existing
> interface that does not take the global spinlock and won't advance the
> global offset. That call should be quite cheap.

Ahh, I did indeed mis-read that as the new one with the lock.

I did react to the fact that is_mgtime(inode) itself is horribly
expensive if it's not cached (following three quite possibly cold
pointers), which was part of that whole "look at I_CTIME_QUERIED
instead".

I see the pointer chasing as a huge VFS timesink in all my profiles,
although usually it's the disgusting security pointer (because selinux
creates separate security nodes for each inode, even when the contents
are often identical). So I'm a bit sensitive to "follow several
pointers from 'struct inode'" patterns from looking at too many
instruction profiles.

          Linus
Jeffrey Layton Oct. 18, 2023, 9:52 p.m. UTC | #4
On Wed, 2023-10-18 at 14:31 -0700, Linus Torvalds wrote:
> On Wed, 18 Oct 2023 at 13:47, Jeff Layton <jlayton@kernel.org> wrote:
> > 
> > >         old_ctime_nsec &= ~I_CTIME_QUERIED;
> > >         if (ts64.tv_nsec > old_ctime_nsec + inode->i_sb->s_time_gran)
> > >                 return ts64;
> > > 
> > 
> > Does that really do what you expect here? current_time will return a
> > value that has already been through timestamp_truncate.
> 
> Yeah, you're right.
> 
> I think you can actually remove the s_time_gran addition. Both the
> old_ctime_nsec and the current ts64,tv_nsec are already rounded, so
> just doing
> 
>         if (ts64.tv_nsec > old_ctime_nsec)
>                 return ts64;
> 
> would already guarantee that it's different enough.
> 

Yep, and that's basically what inode_set_ctime_current does (though it
does a timespec64 comparison).

> > current_mgtime is calling ktime_get_real_ts64, which is an existing
> > interface that does not take the global spinlock and won't advance the
> > global offset. That call should be quite cheap.
> 
> Ahh, I did indeed mis-read that as the new one with the lock.
> 
> I did react to the fact that is_mgtime(inode) itself is horribly
> expensive if it's not cached (following three quite possibly cold
> pointers), which was part of that whole "look at I_CTIME_QUERIED
> instead".
>
> I see the pointer chasing as a huge VFS timesink in all my profiles,
> although usually it's the disgusting security pointer (because selinux
> creates separate security nodes for each inode, even when the contents
> are often identical). So I'm a bit sensitive to "follow several
> pointers from 'struct inode'" patterns from looking at too many
> instruction profiles.

That's a very good point. I'll see if I can get rid of that (and maybe
some other mgtime flag checks) before I send the next version. 

Back to your earlier point though:

Is a global offset really a non-starter? I can see about doing something
per-superblock, but ktime_get_mg_coarse_ts64 should be roughly as cheap
as ktime_get_coarse_ts64. I don't see the downside there for the non-
multigrain filesystems to call that.

On another note: maybe I need to put this behind a Kconfig option
initially too?
Christian Brauner Oct. 19, 2023, 9:29 a.m. UTC | #5
> Back to your earlier point though:
> 
> Is a global offset really a non-starter? I can see about doing something
> per-superblock, but ktime_get_mg_coarse_ts64 should be roughly as cheap
> as ktime_get_coarse_ts64. I don't see the downside there for the non-
> multigrain filesystems to call that.

I have to say that this doesn't excite me. This whole thing feels a bit
hackish. I think that a change version is the way more sane way to go.

> 
> On another note: maybe I need to put this behind a Kconfig option
> initially too?

So can we for a second consider not introducing fine-grained timestamps
at all. We let NFSv3 live with the cache problem it's been living with
forever.

And for NFSv4 we actually do introduce a proper i_version for all
filesystems that matter to it.

What filesystems exactly don't expose a proper i_version and what does
prevent them from adding one or fixing it?
Jeffrey Layton Oct. 19, 2023, 11:28 a.m. UTC | #6
On Thu, 2023-10-19 at 11:29 +0200, Christian Brauner wrote:
> > Back to your earlier point though:
> > 
> > Is a global offset really a non-starter? I can see about doing something
> > per-superblock, but ktime_get_mg_coarse_ts64 should be roughly as cheap
> > as ktime_get_coarse_ts64. I don't see the downside there for the non-
> > multigrain filesystems to call that.
> 
> I have to say that this doesn't excite me. This whole thing feels a bit
> hackish. I think that a change version is the way more sane way to go.
> 

What is it about this set that feels so much more hackish to you? Most
of this set is pretty similar to what we had to revert. Is it just the
timekeeper changes? Why do you feel those are a problem?

> > 
> > On another note: maybe I need to put this behind a Kconfig option
> > initially too?
> 
> So can we for a second consider not introducing fine-grained timestamps
> at all. We let NFSv3 live with the cache problem it's been living with
> forever.
> 
> And for NFSv4 we actually do introduce a proper i_version for all
> filesystems that matter to it.
> 
> What filesystems exactly don't expose a proper i_version and what does
> prevent them from adding one or fixing it?

Certainly we can drop this series altogether if that's the consensus.

The main exportable filesystem that doesn't have a suitable change
counter now is XFS. Fixing it will require an on-disk format change to
accommodate a new version counter that doesn't increment on atime
updates. This is something the XFS folks were specifically looking to
avoid, but maybe that's the simpler option.

There is also bcachefs which I don't think has a change attr yet. They'd
also likely need a on-disk format change, but hopefully that's a easier
thing to do there since it's a brand new filesystem.

There are a smattering of lesser-used local filesystems (f2fs, nilfs2,
etc.) that have no i_version support. Multigrain timestamps would make
it simple to add better change attribute support there, but they can (in
principle) all undergo an on-disk format change too if they decide to
add one.

Then there are filesystems like ntfs that are exportable, but where we
can't extend the on-disk format. Those could probably benefit from
multigrain timestamps, but those are much lower priority. Not many
people sharing their NTFS filesystem via NFS anyway.
Thomas Gleixner Oct. 19, 2023, 10 p.m. UTC | #7
Jeff!

On Wed, Oct 18 2023 at 13:41, Jeff Layton wrote:
> +void ktime_get_mg_fine_ts64(struct timespec64 *ts)
> +{
> +	struct timekeeper *tk = &tk_core.timekeeper;
> +	unsigned long flags;
> +	u32 nsecs;
> +
> +	WARN_ON(timekeeping_suspended);
> +
> +	raw_spin_lock_irqsave(&timekeeper_lock, flags);
> +	write_seqcount_begin(&tk_core.seq);

Depending on the usage scenario, this will end up as a scalability issue
which affects _all_ of timekeeping.

The usage of timekeeper_lock and the sequence count has been carefully
crafted to be as non-contended as possible. We went a great length to
optimize that because the ktime_get*() functions are really hotpath all
over the place.

Exposing such an interface which wreckages that is a recipe for disaster
down the road. It might be a non-issue today, but once we hit the
bottleneck of that global lock, we are up the creek without a
paddle. Well not really, but all we can do then is fall back to
ktime_get_real(). So let me ask the obvious question:

     Why don't we do that right away?

Many moons ago when we added ktime_get_real_coarse() the main reason was
that reading the time from the underlying hardware was insanely
expensive.

Many moons later this is not true anymore, except for the stupid case
where the BIOS wreckaged the TSC, but that's a hopeless case for
performance no matter what. Optimizing for that would be beyond stupid.

I'm well aware that ktime_get_real_coarse() is still faster than
ktime_get_real() in micro-benchmarks, i.e. 5ns vs. 15ns on the four
years old laptop I'm writing this.

Many moons ago it was in the ballpark of 40ns vs. 5us due to TSC being
useless and even TSC read was way more expensive (factor 8-10x IIRC) in
comparison. That really mattered for FS, but does todays overhead still
make a difference in the real FS use case scenario?

I'm not in the position of running meaningful FS benchmarks to analyze
that, but I think the delta between ktime_get_real_coarse() and
ktime_get_real() on contemporary hardware is small enough that it
justifies this question.

The point is that both functions have pretty much the same D-cache
pattern because they access the same data in the very same
cacheline. The only difference is the actual TSC read and the extra
conversion, but that's it. The TSC read has been massively optimized by
the CPU vendors. I know that the ARM64 counter has been optimized too,
though I have no idea about PPC64 and S390, but I would be truly
surprised if they didn't optimize the hell out of it because time read
is really used heavily both in kernel and user space.

Does anyone have numbers on contemporary hardware to shed some light on
that in the context of FS and the problem at hand?

Thanks,

        tglx
Dave Chinner Oct. 19, 2023, 10:02 p.m. UTC | #8
On Thu, Oct 19, 2023 at 07:28:48AM -0400, Jeff Layton wrote:
> On Thu, 2023-10-19 at 11:29 +0200, Christian Brauner wrote:
> > > Back to your earlier point though:
> > > 
> > > Is a global offset really a non-starter? I can see about doing something
> > > per-superblock, but ktime_get_mg_coarse_ts64 should be roughly as cheap
> > > as ktime_get_coarse_ts64. I don't see the downside there for the non-
> > > multigrain filesystems to call that.
> > 
> > I have to say that this doesn't excite me. This whole thing feels a bit
> > hackish. I think that a change version is the way more sane way to go.
> > 
> 
> What is it about this set that feels so much more hackish to you? Most
> of this set is pretty similar to what we had to revert. Is it just the
> timekeeper changes? Why do you feel those are a problem?
> 
> > > 
> > > On another note: maybe I need to put this behind a Kconfig option
> > > initially too?
> > 
> > So can we for a second consider not introducing fine-grained timestamps
> > at all. We let NFSv3 live with the cache problem it's been living with
> > forever.
> > 
> > And for NFSv4 we actually do introduce a proper i_version for all
> > filesystems that matter to it.
> > 
> > What filesystems exactly don't expose a proper i_version and what does
> > prevent them from adding one or fixing it?
> 
> Certainly we can drop this series altogether if that's the consensus.
> 
> The main exportable filesystem that doesn't have a suitable change
> counter now is XFS. Fixing it will require an on-disk format change to
> accommodate a new version counter that doesn't increment on atime
> updates. This is something the XFS folks were specifically looking to
> avoid, but maybe that's the simpler option.

And now we have travelled the full circle.

The problem NFS has with atime updates on XFS is a result of
the default behaviour of relatime - it *always* forces a persistent
atime update after mtime has changed. Hence a read-after-write
operation will trigger an atime update because atime is older than
mtime. This is what causes XFS to run a transaction (i.e. a
persistent atime update) and that bumps iversion.

lazytime does not behave this way - it delays all persistent
timestamp updates until the next persistent change or until the
lazytime aggregation period expires (24 hours). Hence with lazytime,
read-after-write operations do not trigger a persistent atime
update, and so XFS does not run a transaction to update atime. Hence
i_version does not get bumped, and NFS behaves as expected.

IOWs, what the NFS server actually wants from the filesytsems is for
lazy timestamp updates to always be used on read operations. It does
not want persistent timestamp updates that change on-disk state. The
recent "redefinition" of when i_version should change effectively
encodes this - i_version should only change when a persistent
metadata or data change is made that also changes [cm]time.

Hence the simple, in-memory solution to this problem is for NFS to
tell the filesysetms that it needs to using lazy (in-memory) atime
updates for the given operation rather than persistent atime updates.

We already need to modify how atime updates work for io_uring -
io_uring needs atime updates to be guaranteed non-blocking similar
to updating mtime in the write IO path. If a persistent timestamp
change needs to be run, then the timestamp update needs to return
-EAGAIN rather than (potentially) blocking so the entire operation
can be punted to a context that can block.

This requires control flags to be passed to the core atime handling
functions.  If a filesystem doesn't understand/support the flags, it
can just ignore it and do the update however it was going to do it.
It won't make anything work incorrectly, just might do something
that is not ideal.

With this new "non-blocking update only" flag for io_uring and a
new "non-persistent update only" flag for NFS, we have a very
similar conditional atime update requirements from two completely
independent in-kernel applications.

IOWs, this can be solved quite simply by having the -application-
define the persistence semantics of the operation being performed.
Add a RWF_LAZYTIME/IOCB_LAZYTIME flag for read IO that is being
issued from the nfs daemon (i.e. passed to vfs_iter_read()) and then
the vfs/filesystem can do exactly the right thing for the IO being
issued.

This is what io_uring does with IOCB_NOWAIT to tell the filesystems
that the IO must be non-blocking, and it's the key we already use
for non-blocking mtime updates and will use to trigger non-blocking
atime updates....

I also know of cases where a per-IO RWF_LAZYTIME flag would be
beneficial - large databases are already using lazytime mount
options so that their data IO doesn't take persistent mtime update
overhead hits on every write IO.....

> There is also bcachefs which I don't think has a change attr yet. They'd
> also likely need a on-disk format change, but hopefully that's a easier
> thing to do there since it's a brand new filesystem.

It's not a "brand new filesystem". It's been out there for quite a
long while, and it has many users that would be impacted by on-disk
format changes at this point in it's life. on-disk format changes
are a fairly major deal for filesystems, and if there is any way we
can avoid them we should.

-Dave.
Jeffrey Layton Oct. 19, 2023, 10:41 p.m. UTC | #9
On Fri, 2023-10-20 at 00:00 +0200, Thomas Gleixner wrote:
> Jeff!
> 
> On Wed, Oct 18 2023 at 13:41, Jeff Layton wrote:
> > +void ktime_get_mg_fine_ts64(struct timespec64 *ts)
> > +{
> > +	struct timekeeper *tk = &tk_core.timekeeper;
> > +	unsigned long flags;
> > +	u32 nsecs;
> > +
> > +	WARN_ON(timekeeping_suspended);
> > +
> > +	raw_spin_lock_irqsave(&timekeeper_lock, flags);
> > +	write_seqcount_begin(&tk_core.seq);
> 
> Depending on the usage scenario, this will end up as a scalability issue
> which affects _all_ of timekeeping.
> 
> The usage of timekeeper_lock and the sequence count has been carefully
> crafted to be as non-contended as possible. We went a great length to
> optimize that because the ktime_get*() functions are really hotpath all
> over the place.
> 

> Exposing such an interface which wreckages that is a recipe for disaster
> down the road. It might be a non-issue today, but once we hit the
> bottleneck of that global lock, we are up the creek without a
> paddle.
> 

Thanks for taking the explanation, Thomas. That's understandable, and
that was my main worry with this set. I'll look at doing this another
way given your feedback. I just started by plumbing this into the
timekeeping code since that seemed like the most obvious place to do it.

I think it's probably still possible to do this by caching the values
returned by the timekeeper at the vfs layer, but there seems to be some
reticence to the basic idea that I don't quite understand yet.

> Well not really, but all we can do then is fall back to
> ktime_get_real(). So let me ask the obvious question:
> 
>      Why don't we do that right away?
> 
> Many moons ago when we added ktime_get_real_coarse() the main reason was
> that reading the time from the underlying hardware was insanely
> expensive.
> 
> Many moons later this is not true anymore, except for the stupid case
> where the BIOS wreckaged the TSC, but that's a hopeless case for
> performance no matter what. Optimizing for that would be beyond stupid.
> 
> I'm well aware that ktime_get_real_coarse() is still faster than
> ktime_get_real() in micro-benchmarks, i.e. 5ns vs. 15ns on the four
> years old laptop I'm writing this.
> 
> Many moons ago it was in the ballpark of 40ns vs. 5us due to TSC being
> useless and even TSC read was way more expensive (factor 8-10x IIRC) in
> comparison. That really mattered for FS, but does todays overhead still
> make a difference in the real FS use case scenario?
> 
> I'm not in the position of running meaningful FS benchmarks to analyze
> that, but I think the delta between ktime_get_real_coarse() and
> ktime_get_real() on contemporary hardware is small enough that it
> justifies this question.
> 
> The point is that both functions have pretty much the same D-cache
> pattern because they access the same data in the very same
> cacheline. The only difference is the actual TSC read and the extra
> conversion, but that's it. The TSC read has been massively optimized by
> the CPU vendors. I know that the ARM64 counter has been optimized too,
> though I have no idea about PPC64 and S390, but I would be truly
> surprised if they didn't optimize the hell out of it because time read
> is really used heavily both in kernel and user space.
> 
> Does anyone have numbers on contemporary hardware to shed some light on
> that in the context of FS and the problem at hand?

That was sort of my suspicion and it's good to have confirmation that
fetching a fine-grained timespec64 from the timekeeper is cheap. It
looked that way when I was poking around in there, but I wasn't sure
whether it was always the case.

It turns out however that the main benefit of using a coarse-grained
timestamp is that it allows the file system to skip a lot of inode
metadata updates.

The way it works today is that when we go to update the timestamp on an
inode, we check whether they have made any visible change, and we dirty
the inode metadata if so. This means that we only really update the
inode on disk once per jiffy or so when an inode is under heavy writes.

The idea with this set is to only use fine-grained timestamps when
someone is actively fetching them via getattr. When the mtime or ctime
is viewed via getattr, we mark the inode and then the following
timestamp update will get a fine-grained timestamp (unless the coarse-
grained clock has already ticked).

That allows us to keep the number of inode updates down to a bare
minimum, but still allows an observer to always see a change in the
timestamp when there have been changes to the inode.

Thanks again for the review! For the next iteration I (probably) won't
need to touch the timekeeper.
Jeffrey Layton Oct. 20, 2023, 12:12 p.m. UTC | #10
On Fri, 2023-10-20 at 09:02 +1100, Dave Chinner wrote:
> On Thu, Oct 19, 2023 at 07:28:48AM -0400, Jeff Layton wrote:
> > On Thu, 2023-10-19 at 11:29 +0200, Christian Brauner wrote:
> > > > Back to your earlier point though:
> > > > 
> > > > Is a global offset really a non-starter? I can see about doing something
> > > > per-superblock, but ktime_get_mg_coarse_ts64 should be roughly as cheap
> > > > as ktime_get_coarse_ts64. I don't see the downside there for the non-
> > > > multigrain filesystems to call that.
> > > 
> > > I have to say that this doesn't excite me. This whole thing feels a bit
> > > hackish. I think that a change version is the way more sane way to go.
> > > 
> > 
> > What is it about this set that feels so much more hackish to you? Most
> > of this set is pretty similar to what we had to revert. Is it just the
> > timekeeper changes? Why do you feel those are a problem?
> > 
> > > > 
> > > > On another note: maybe I need to put this behind a Kconfig option
> > > > initially too?
> > > 
> > > So can we for a second consider not introducing fine-grained timestamps
> > > at all. We let NFSv3 live with the cache problem it's been living with
> > > forever.
> > > 
> > > And for NFSv4 we actually do introduce a proper i_version for all
> > > filesystems that matter to it.
> > > 
> > > What filesystems exactly don't expose a proper i_version and what does
> > > prevent them from adding one or fixing it?
> > 
> > Certainly we can drop this series altogether if that's the consensus.
> > 
> > The main exportable filesystem that doesn't have a suitable change
> > counter now is XFS. Fixing it will require an on-disk format change to
> > accommodate a new version counter that doesn't increment on atime
> > updates. This is something the XFS folks were specifically looking to
> > avoid, but maybe that's the simpler option.
> 
> And now we have travelled the full circle.
> 

LOL, yes!

> The problem NFS has with atime updates on XFS is a result of
> the default behaviour of relatime - it *always* forces a persistent
> atime update after mtime has changed. Hence a read-after-write
> operation will trigger an atime update because atime is older than
> mtime. This is what causes XFS to run a transaction (i.e. a
> persistent atime update) and that bumps iversion.
> 

Those particular atime updates are not a problem. If we're updating the
mtime and ctime anyway, then bumping the change attribute is OK.

The problem is that relatime _also_ does an on-disk update once a day
for just an atime update. On XFS, this means that the change attribute
also gets bumped and the clients invalidate their caches all at once.

That doesn't sound like a big problem at first, but what if you're
sharing a multi-gigabyte r/o file between multiple clients? This sort of
thing is fairly common on render-farm workloads, and all of your clients
will end up invalidating their caches once once a day if you're serving
from XFS.

> lazytime does not behave this way - it delays all persistent
> timestamp updates until the next persistent change or until the
> lazytime aggregation period expires (24 hours). Hence with lazytime,
> read-after-write operations do not trigger a persistent atime
> update, and so XFS does not run a transaction to update atime. Hence
> i_version does not get bumped, and NFS behaves as expected.
> 

Similar problem here. Once a day, NFS clients will invalidate the cache
on any static content served from XFS.

> IOWs, what the NFS server actually wants from the filesytsems is for
> lazy timestamp updates to always be used on read operations. It does
> not want persistent timestamp updates that change on-disk state. The
> recent "redefinition" of when i_version should change effectively
> encodes this - i_version should only change when a persistent
> metadata or data change is made that also changes [cm]time.
>
> Hence the simple, in-memory solution to this problem is for NFS to
> tell the filesysetms that it needs to using lazy (in-memory) atime
> updates for the given operation rather than persistent atime updates.
>
> We already need to modify how atime updates work for io_uring -
> io_uring needs atime updates to be guaranteed non-blocking similar
> to updating mtime in the write IO path. If a persistent timestamp
> change needs to be run, then the timestamp update needs to return
> -EAGAIN rather than (potentially) blocking so the entire operation
> can be punted to a context that can block.
> 
> This requires control flags to be passed to the core atime handling
> functions.  If a filesystem doesn't understand/support the flags, it
> can just ignore it and do the update however it was going to do it.
> It won't make anything work incorrectly, just might do something
> that is not ideal.
> 
> With this new "non-blocking update only" flag for io_uring and a
> new "non-persistent update only" flag for NFS, we have a very
> similar conditional atime update requirements from two completely
> independent in-kernel applications.
> 
> IOWs, this can be solved quite simply by having the -application-
> define the persistence semantics of the operation being performed.
> Add a RWF_LAZYTIME/IOCB_LAZYTIME flag for read IO that is being
> issued from the nfs daemon (i.e. passed to vfs_iter_read()) and then
> the vfs/filesystem can do exactly the right thing for the IO being
> issued.
> 
> This is what io_uring does with IOCB_NOWAIT to tell the filesystems
> that the IO must be non-blocking, and it's the key we already use
> for non-blocking mtime updates and will use to trigger non-blocking
> atime updates....
> 
> I also know of cases where a per-IO RWF_LAZYTIME flag would be
> beneficial - large databases are already using lazytime mount
> options so that their data IO doesn't take persistent mtime update
> overhead hits on every write IO.....
> 

I don't think that trying to do something "special" for activity that is
initiated by the NFS server solves anything. Bear in mind that NFS
clients care about locally-initiated activity too.

The bottom line is that we don't want to be foisting a cache
invalidation on the clients just because someone read a file, or did
some similar activity like a readdir or readlink. The lazytime/relatime
options may mitigate the problem, but they're not a real solution.

What NFS _really_ wants is a proper change counter that doesn't
increment on read(like) operations. In practice, that comes down to just
not incrementing it on atime updates.

btrfs, ext4 and tmpfs have this (now). xfs does not because its change
attribute is incremented when an atime update is logged, and that is
evidently something that cannot be changed without an on-disk format
change.

> > There is also bcachefs which I don't think has a change attr yet. They'd
> > also likely need a on-disk format change, but hopefully that's a easier
> > thing to do there since it's a brand new filesystem.
> 
> It's not a "brand new filesystem". It's been out there for quite a
> long while, and it has many users that would be impacted by on-disk
> format changes at this point in it's life. on-disk format changes
> are a fairly major deal for filesystems, and if there is any way we
> can avoid them we should.


Sure. It's new to me though. It's also not yet merged into mainline.

I'd _really_ like to see a proper change counter added before it's
merged, or at least space in the on-disk inode reserved for one until we
can get it plumbed in. That would suck for the current users, I suppose,
but at least that userbase is small now. Once it's merged, there will be
a lot more people using it and it becomes just that much more difficult.

I suppose bcachefs could try to hold out for the multigrain timestamp
work too, but that may not ever make it in.
Linus Torvalds Oct. 20, 2023, 8:06 p.m. UTC | #11
On Fri, 20 Oct 2023 at 05:12, Jeff Layton <jlayton@kernel.org> wrote:.
>
> I'd _really_ like to see a proper change counter added before it's
> merged, or at least space in the on-disk inode reserved for one until we
> can get it plumbed in.

Hmm. Can we not perhaps just do an in-memory change counter, and try
to initialize it to a random value when instantiating an inode? Do we
even *require* on-disk format changes?

So on reboot, the inode would count as "changed" as far any remote
user is concerned. It would flush client caches, but isn't that what
you'd want anyway? I'd hate to waste lots of memory, but maybe people
would be ok with just a 32-bit random value. And if not...

But I actually came into this whole discussion purely through the
inode timestamp side, so I may *entirely* miss what the change counter
requirements for NFSd actually are. If it needs to be stable across
reboots, my idea is clearly complete garbage.

You can now all jump on me and point out my severe intellectual
limitations. Please use small words when you do ;)

              Linus
Linus Torvalds Oct. 20, 2023, 8:20 p.m. UTC | #12
On Fri, 20 Oct 2023 at 13:06, Linus Torvalds
<torvalds@linux-foundation.org> wrote:
>
> So on reboot, the inode would count as "changed" as far any remote
> user is concerned. [..]

Obviously, not just reboot would do that. Any kind of "it's no longer
cached on the server and gets read back from disk" would do the same
thing.

Again, that may not work for the intended purpose, but if the use-case
is a "same version number means no changes", it might be acceptable?
Even if you then could get spurious version changes when the file
hasn't been accessed in a long time?

Maybe all this together with with some ctime filtering ("old ctime
clealy means that the version number is irrelevant"). After all, the
whole point of fine-grained timestamps was to distinguish *frequent*
changes. An in-memory counter certainly does that even without any
on-disk representation..

               Linus
Jeffrey Layton Oct. 20, 2023, 9:05 p.m. UTC | #13
On Fri, 2023-10-20 at 13:06 -0700, Linus Torvalds wrote:
> On Fri, 20 Oct 2023 at 05:12, Jeff Layton <jlayton@kernel.org> wrote:.
> > 
> > I'd _really_ like to see a proper change counter added before it's
> > merged, or at least space in the on-disk inode reserved for one until we
> > can get it plumbed in.
> 
> Hmm. Can we not perhaps just do an in-memory change counter, and try
> to initialize it to a random value when instantiating an inode? Do we
> even *require* on-disk format changes?
> 
> So on reboot, the inode would count as "changed" as far any remote
> user is concerned. It would flush client caches, but isn't that what
> you'd want anyway? I'd hate to waste lots of memory, but maybe people
> would be ok with just a 32-bit random value. And if not...
> 
> But I actually came into this whole discussion purely through the
> inode timestamp side, so I may *entirely* miss what the change counter
> requirements for NFSd actually are. If it needs to be stable across
> reboots, my idea is clearly complete garbage.
> 
> You can now all jump on me and point out my severe intellectual
> limitations. Please use small words when you do ;)
> 

Much like inode timestamps, we do depend on the change attribute
persisting across reboots. Having to invalidate all of your cached data
just because the server rebooted is particularly awful. That usually
results in the server being hammered with reads from all of the clients
at once, soon after rebooting.
Dave Chinner Oct. 22, 2023, 10:17 p.m. UTC | #14
On Fri, Oct 20, 2023 at 08:12:45AM -0400, Jeff Layton wrote:
> On Fri, 2023-10-20 at 09:02 +1100, Dave Chinner wrote:
> > On Thu, Oct 19, 2023 at 07:28:48AM -0400, Jeff Layton wrote:
> > > On Thu, 2023-10-19 at 11:29 +0200, Christian Brauner wrote:
> > > > > Back to your earlier point though:
> > > > > 
> > > > > Is a global offset really a non-starter? I can see about doing something
> > > > > per-superblock, but ktime_get_mg_coarse_ts64 should be roughly as cheap
> > > > > as ktime_get_coarse_ts64. I don't see the downside there for the non-
> > > > > multigrain filesystems to call that.
> > > > 
> > > > I have to say that this doesn't excite me. This whole thing feels a bit
> > > > hackish. I think that a change version is the way more sane way to go.
> > > > 
> > > 
> > > What is it about this set that feels so much more hackish to you? Most
> > > of this set is pretty similar to what we had to revert. Is it just the
> > > timekeeper changes? Why do you feel those are a problem?
> > > 
> > > > > 
> > > > > On another note: maybe I need to put this behind a Kconfig option
> > > > > initially too?
> > > > 
> > > > So can we for a second consider not introducing fine-grained timestamps
> > > > at all. We let NFSv3 live with the cache problem it's been living with
> > > > forever.
> > > > 
> > > > And for NFSv4 we actually do introduce a proper i_version for all
> > > > filesystems that matter to it.
> > > > 
> > > > What filesystems exactly don't expose a proper i_version and what does
> > > > prevent them from adding one or fixing it?
> > > 
> > > Certainly we can drop this series altogether if that's the consensus.
> > > 
> > > The main exportable filesystem that doesn't have a suitable change
> > > counter now is XFS. Fixing it will require an on-disk format change to
> > > accommodate a new version counter that doesn't increment on atime
> > > updates. This is something the XFS folks were specifically looking to
> > > avoid, but maybe that's the simpler option.
> > 
> > And now we have travelled the full circle.
> > 
> 
> LOL, yes!
> 
> > The problem NFS has with atime updates on XFS is a result of
> > the default behaviour of relatime - it *always* forces a persistent
> > atime update after mtime has changed. Hence a read-after-write
> > operation will trigger an atime update because atime is older than
> > mtime. This is what causes XFS to run a transaction (i.e. a
> > persistent atime update) and that bumps iversion.
> > 
> 
> Those particular atime updates are not a problem. If we're updating the
> mtime and ctime anyway, then bumping the change attribute is OK.
> 
> The problem is that relatime _also_ does an on-disk update once a day
> for just an atime update. On XFS, this means that the change attribute
> also gets bumped and the clients invalidate their caches all at once.
> 
> That doesn't sound like a big problem at first, but what if you're
> sharing a multi-gigabyte r/o file between multiple clients? This sort of
> thing is fairly common on render-farm workloads, and all of your clients
> will end up invalidating their caches once once a day if you're serving
> from XFS.

So we have noatime inode and mount options for such specialised
workloads that cannot tolerate cached ever being invalidated, yes?

> > lazytime does not behave this way - it delays all persistent
> > timestamp updates until the next persistent change or until the
> > lazytime aggregation period expires (24 hours). Hence with lazytime,
> > read-after-write operations do not trigger a persistent atime
> > update, and so XFS does not run a transaction to update atime. Hence
> > i_version does not get bumped, and NFS behaves as expected.
> > 
> 
> Similar problem here. Once a day, NFS clients will invalidate the cache
> on any static content served from XFS.

Lazytime has /proc/sys/vm/dirtytime_expire_seconds to change the
interval that triggers persistent time changes. That could easily be
configured to be longer than a day for workloads that care about
this sort of thing. Indeed, we could just set up timestamps that NFS
says "do not make persistent" to only be persisted when the inode is
removed from server memory rather than be timed out by background
writeback....

-----

All I'm suggesting is that rather than using mount options for
noatime-like behaviour for NFSD accesses, we actually have the nfsd
accesses say "we'd like pure atime updates without iversion, please".

Keep in mind that XFS does actually try to avoid bumping i_version
on pure timestamp updates - we carved that out a long time ago (see
the difference in XFS_ILOG_CORE vs XFS_ILOG_TIMESTAMP in
xfs_vn_update_time() and xfs_trans_log_inode()) so that we could
optimise fdatasync() to ignore timestamp updates that occur as a
result of pure data overwrites.

Hence XFS only bumps i_version for pure timestamp updates if the
iversion queried flag is set. IOWs, XFS it is actually doing exactly
what the VFS iversion implementation is telling it to do with
timestamp updates for non-core inode metadata updates.

That's the fundamental issue here: nfsd has set VFS state that tells
the filesystem to "bump iversion on next persistent inode change",
but the nfsd then runs operations that can change non-critical
persistent inode state in "query-only" operations. It then expects
filesystems to know that it should ignore the iversion queried state
within this context.  However, without external behavioural control
flags, filesystems cannot know that an isolated metadata update has
context specific iversion behavioural constraints.

Hence fixing this is purely a VFS/nfsd i_version implementation
problem - if the nfsd is running a querying operation, it should
tell the filesystem that it should ignore iversion query state. If
nothing the application level cache cares about is being changed
during the query operation, it should tell the filesystem to ignore
iversion query state because it is likely the nfsd query itself will
set it (or have already set it itself in the case of compound
operations).

This does not need XFS on-disk format changes to fix. This does not
need changes to timestamp infrastructure to fix. We just need the
nfsd application to tell us that we should ignore the vfs i_version
query state when we update non-core inode metadata within query
operation contexts.

-Dave.
Jeffrey Layton Oct. 23, 2023, 2:45 p.m. UTC | #15
On Mon, 2023-10-23 at 09:17 +1100, Dave Chinner wrote:
> On Fri, Oct 20, 2023 at 08:12:45AM -0400, Jeff Layton wrote:
> > On Fri, 2023-10-20 at 09:02 +1100, Dave Chinner wrote:
> > > On Thu, Oct 19, 2023 at 07:28:48AM -0400, Jeff Layton wrote:
> > > > On Thu, 2023-10-19 at 11:29 +0200, Christian Brauner wrote:
> > > > > > Back to your earlier point though:
> > > > > > 
> > > > > > Is a global offset really a non-starter? I can see about doing something
> > > > > > per-superblock, but ktime_get_mg_coarse_ts64 should be roughly as cheap
> > > > > > as ktime_get_coarse_ts64. I don't see the downside there for the non-
> > > > > > multigrain filesystems to call that.
> > > > > 
> > > > > I have to say that this doesn't excite me. This whole thing feels a bit
> > > > > hackish. I think that a change version is the way more sane way to go.
> > > > > 
> > > > 
> > > > What is it about this set that feels so much more hackish to you? Most
> > > > of this set is pretty similar to what we had to revert. Is it just the
> > > > timekeeper changes? Why do you feel those are a problem?
> > > > 
> > > > > > 
> > > > > > On another note: maybe I need to put this behind a Kconfig option
> > > > > > initially too?
> > > > > 
> > > > > So can we for a second consider not introducing fine-grained timestamps
> > > > > at all. We let NFSv3 live with the cache problem it's been living with
> > > > > forever.
> > > > > 
> > > > > And for NFSv4 we actually do introduce a proper i_version for all
> > > > > filesystems that matter to it.
> > > > > 
> > > > > What filesystems exactly don't expose a proper i_version and what does
> > > > > prevent them from adding one or fixing it?
> > > > 
> > > > Certainly we can drop this series altogether if that's the consensus.
> > > > 
> > > > The main exportable filesystem that doesn't have a suitable change
> > > > counter now is XFS. Fixing it will require an on-disk format change to
> > > > accommodate a new version counter that doesn't increment on atime
> > > > updates. This is something the XFS folks were specifically looking to
> > > > avoid, but maybe that's the simpler option.
> > > 
> > > And now we have travelled the full circle.
> > > 
> > 
> > LOL, yes!
> > 
> > > The problem NFS has with atime updates on XFS is a result of
> > > the default behaviour of relatime - it *always* forces a persistent
> > > atime update after mtime has changed. Hence a read-after-write
> > > operation will trigger an atime update because atime is older than
> > > mtime. This is what causes XFS to run a transaction (i.e. a
> > > persistent atime update) and that bumps iversion.
> > > 
> > 
> > Those particular atime updates are not a problem. If we're updating the
> > mtime and ctime anyway, then bumping the change attribute is OK.
> > 
> > The problem is that relatime _also_ does an on-disk update once a day
> > for just an atime update. On XFS, this means that the change attribute
> > also gets bumped and the clients invalidate their caches all at once.
> > 
> > That doesn't sound like a big problem at first, but what if you're
> > sharing a multi-gigabyte r/o file between multiple clients? This sort of
> > thing is fairly common on render-farm workloads, and all of your clients
> > will end up invalidating their caches once once a day if you're serving
> > from XFS.
> 
> So we have noatime inode and mount options for such specialised
> workloads that cannot tolerate cached ever being invalidated, yes?
> 
> > > lazytime does not behave this way - it delays all persistent
> > > timestamp updates until the next persistent change or until the
> > > lazytime aggregation period expires (24 hours). Hence with lazytime,
> > > read-after-write operations do not trigger a persistent atime
> > > update, and so XFS does not run a transaction to update atime. Hence
> > > i_version does not get bumped, and NFS behaves as expected.
> > > 
> > 
> > Similar problem here. Once a day, NFS clients will invalidate the cache
> > on any static content served from XFS.
> 
> Lazytime has /proc/sys/vm/dirtytime_expire_seconds to change the
> interval that triggers persistent time changes. That could easily be
> configured to be longer than a day for workloads that care about
> this sort of thing. Indeed, we could just set up timestamps that NFS
> says "do not make persistent" to only be persisted when the inode is
> removed from server memory rather than be timed out by background
> writeback....
> 
> -----
> 
> All I'm suggesting is that rather than using mount options for
> noatime-like behaviour for NFSD accesses, we actually have the nfsd
> accesses say "we'd like pure atime updates without iversion, please".
> 
> Keep in mind that XFS does actually try to avoid bumping i_version
> on pure timestamp updates - we carved that out a long time ago (see
> the difference in XFS_ILOG_CORE vs XFS_ILOG_TIMESTAMP in
> xfs_vn_update_time() and xfs_trans_log_inode()) so that we could
> optimise fdatasync() to ignore timestamp updates that occur as a
> result of pure data overwrites.
> 
> Hence XFS only bumps i_version for pure timestamp updates if the
> iversion queried flag is set. IOWs, XFS it is actually doing exactly
> what the VFS iversion implementation is telling it to do with
> timestamp updates for non-core inode metadata updates.
> 
> That's the fundamental issue here: nfsd has set VFS state that tells
> the filesystem to "bump iversion on next persistent inode change",
> but the nfsd then runs operations that can change non-critical
> persistent inode state in "query-only" operations. It then expects
> filesystems to know that it should ignore the iversion queried state
> within this context.  However, without external behavioural control
> flags, filesystems cannot know that an isolated metadata update has
> context specific iversion behavioural constraints.

> Hence fixing this is purely a VFS/nfsd i_version implementation
> problem - if the nfsd is running a querying operation, it should
> tell the filesystem that it should ignore iversion query state. If
> nothing the application level cache cares about is being changed
> during the query operation, it should tell the filesystem to ignore
> iversion query state because it is likely the nfsd query itself will
> set it (or have already set it itself in the case of compound
> operations).
> 
> This does not need XFS on-disk format changes to fix. This does not
> need changes to timestamp infrastructure to fix. We just need the
> nfsd application to tell us that we should ignore the vfs i_version
> query state when we update non-core inode metadata within query
> operation contexts.
> 


I think you're missing the point of the problem I'm trying to solve.
I'm not necessarily trying to guard nfsd against its own accesses. The
reads that trigger an eventual atime update could come from anywhere --
nfsd, userland accesses, etc.

If you are serving an XFS filesystem, with the (default) relatime mount
option, then you are guaranteed that the clients will invalidate their
cache of a file once per day, assuming that at least one read was issued
against the file during that day.

That read will cause an eventual atime bump to be logged, at which point
the change attribute will change. The client will then assume that it
needs to invalidate its cache when it sees that change.

Changing how nfsd does its own accesses won't fix anything, because the
problematic atime bump can come from any sort of read access.
Dave Chinner Oct. 23, 2023, 11:26 p.m. UTC | #16
On Mon, Oct 23, 2023 at 10:45:21AM -0400, Jeff Layton wrote:
> On Mon, 2023-10-23 at 09:17 +1100, Dave Chinner wrote:
> > All I'm suggesting is that rather than using mount options for
> > noatime-like behaviour for NFSD accesses, we actually have the nfsd
> > accesses say "we'd like pure atime updates without iversion, please".
> > 
> > Keep in mind that XFS does actually try to avoid bumping i_version
> > on pure timestamp updates - we carved that out a long time ago (see
> > the difference in XFS_ILOG_CORE vs XFS_ILOG_TIMESTAMP in
> > xfs_vn_update_time() and xfs_trans_log_inode()) so that we could
> > optimise fdatasync() to ignore timestamp updates that occur as a
> > result of pure data overwrites.
> > 
> > Hence XFS only bumps i_version for pure timestamp updates if the
> > iversion queried flag is set. IOWs, XFS it is actually doing exactly
> > what the VFS iversion implementation is telling it to do with
> > timestamp updates for non-core inode metadata updates.
> > 
> > That's the fundamental issue here: nfsd has set VFS state that tells
> > the filesystem to "bump iversion on next persistent inode change",
> > but the nfsd then runs operations that can change non-critical
> > persistent inode state in "query-only" operations. It then expects
> > filesystems to know that it should ignore the iversion queried state
> > within this context.  However, without external behavioural control
> > flags, filesystems cannot know that an isolated metadata update has
> > context specific iversion behavioural constraints.
> 
> > Hence fixing this is purely a VFS/nfsd i_version implementation
> > problem - if the nfsd is running a querying operation, it should
> > tell the filesystem that it should ignore iversion query state. If
> > nothing the application level cache cares about is being changed
> > during the query operation, it should tell the filesystem to ignore
> > iversion query state because it is likely the nfsd query itself will
> > set it (or have already set it itself in the case of compound
> > operations).
> > 
> > This does not need XFS on-disk format changes to fix. This does not
> > need changes to timestamp infrastructure to fix. We just need the
> > nfsd application to tell us that we should ignore the vfs i_version
> > query state when we update non-core inode metadata within query
> > operation contexts.
> > 
> 
> I think you're missing the point of the problem I'm trying to solve.
> I'm not necessarily trying to guard nfsd against its own accesses. The
> reads that trigger an eventual atime update could come from anywhere --
> nfsd, userland accesses, etc.
>
> If you are serving an XFS filesystem, with the (default) relatime mount
> option, then you are guaranteed that the clients will invalidate their
> cache of a file once per day, assuming that at least one read was issued
> against the file during that day.
>
> That read will cause an eventual atime bump to be logged, at which point
> the change attribute will change. The client will then assume that it
> needs to invalidate its cache when it sees that change.
>
> Changing how nfsd does its own accesses won't fix anything, because the
> problematic atime bump can come from any sort of read access.

I'm not missing the point at all - as I've said in the past I don't
think local vs remote access is in any way relevant to the original
problem that needs to be solved. If the local access is within the
relatime window, it won't cause any persistent metadata change at
all. If it's outside the window, then it's no different to the NFS
client reading data from the server outside the window. If it's the
first access after a NFS client side modification, then it's just
really bad timing but it isn't likely to be a common issue.

Hence I just don't think it matters on bit, and we can address the
24 hour problem separately to the original problem that still needs
to be fixed.

The problem is the first read request after a modification has been
made. That is causing relatime to see mtime > atime and triggering
an atime update. XFS sees this, does an atime update, and in
committing that persistent inode metadata update, it calls
inode_maybe_inc_iversion(force = false) to check if an iversion
update is necessary. The VFS sees I_VERSION_QUERIED, and so it bumps
i_version and tells XFS to persist it.

IOWs, XFS is doing exactly what the VFS is telling it to do with
i_version during the persistent inode metadata update that the VFS
told it to make.

This, however, is not the semantics that the *nfsd application*
wants. It does not want i_version to be updated when it is running a
data read operation despite the fact the VFS is telling the
filesystem it needs to be updated.

What we need to know is when the inode is being accessed by the nfsd
so we can change the in-memory timestamp update behaviour
appropriately.  We really don't need on-disk format changes - we
just need to know that we're supposed to do something special with
pure timestamp updates because i_version needs to be behave in a
manner compatible with the new NFS requirements....

We also don't need generic timestamp infrastructure changes to do
this - the multi-grained timestamp was a neat idea for generic
filesystem support of the nfsd i_version requirements, but it's
collapsed under the weight of complexity.

There are simpler ways individual filesystems can do the right
thing, but to do that we need to know that nfsd has actively
referenced the inode. How we get that information is what I want to
resolve, the filesystem should be able to handle everything else in
memory....

Perhaps we can extract I_VERSION_QUERIED as a proxy for nfsd
activity on the inode rather than need a per-operation context? Is
that going to be reliable enough? Will that cause problems for other
applications that want to use i_version for their own purposes?

Cheers,

Dave.
Linus Torvalds Oct. 24, 2023, 12:18 a.m. UTC | #17
On Mon, 23 Oct 2023 at 13:26, Dave Chinner <david@fromorbit.com> wrote:
>
> The problem is the first read request after a modification has been
> made. That is causing relatime to see mtime > atime and triggering
> an atime update. XFS sees this, does an atime update, and in
> committing that persistent inode metadata update, it calls
> inode_maybe_inc_iversion(force = false) to check if an iversion
> update is necessary. The VFS sees I_VERSION_QUERIED, and so it bumps
> i_version and tells XFS to persist it.

Could we perhaps just have a mode where we don't increment i_version
for just atime updates?

Maybe we don't even need a mode, and could just decide that atime
updates aren't i_version updates at all?

Yes, yes, it's obviously technically a "inode modification", but does
anybody actually *want* atime updates with no actual other changes to
be version events?

Or maybe i_version can update, but callers of getattr() could have two
bits for that STATX_CHANGE_COOKIE, one for "I care about atime" and
one for others, and we'd pass that down to inode_query_version, and
we'd have a I_VERSION_QUERIED and a I_VERSION_QUERIED_STRICT, and the
"I care about atime" case ould set the strict one.

Then inode_maybe_inc_iversion() could - for atome updates - skip the
version update *unless* it sees that I_VERSION_QUERIED_STRICT bit.

Does that sound sane to people?

Because it does sound completely insane to me to say "inode changed"
and have a cache invalidation just for an atime update.

              Linus
Dave Chinner Oct. 24, 2023, 3:40 a.m. UTC | #18
On Mon, Oct 23, 2023 at 02:18:12PM -1000, Linus Torvalds wrote:
> On Mon, 23 Oct 2023 at 13:26, Dave Chinner <david@fromorbit.com> wrote:
> >
> > The problem is the first read request after a modification has been
> > made. That is causing relatime to see mtime > atime and triggering
> > an atime update. XFS sees this, does an atime update, and in
> > committing that persistent inode metadata update, it calls
> > inode_maybe_inc_iversion(force = false) to check if an iversion
> > update is necessary. The VFS sees I_VERSION_QUERIED, and so it bumps
> > i_version and tells XFS to persist it.
> 
> Could we perhaps just have a mode where we don't increment i_version
> for just atime updates?
>
> Maybe we don't even need a mode, and could just decide that atime
> updates aren't i_version updates at all?

We do that already - in memory atime updates don't bump i_version at
all. The issue is the rare persistent atime update requests that
still happen - they are the ones that trigger an i_version bump on
XFS, and one of the relatime heuristics tickle this specific issue.

If we push the problematic persistent atime updates to be in-memory
updates only, then the whole problem with i_version goes away....

> Yes, yes, it's obviously technically a "inode modification", but does
> anybody actually *want* atime updates with no actual other changes to
> be version events?

Well, yes, there was. That's why we defined i_version in the on disk
format this way well over a decade ago. It was part of some deep
dark magical HSM beans that allowed the application to combine
multiple scans for different inode metadata changes into a single
pass. atime changes was one of the things it needed to know about
for tiering and space scavenging purposes....

> Or maybe i_version can update, but callers of getattr() could have two
> bits for that STATX_CHANGE_COOKIE, one for "I care about atime" and
> one for others, and we'd pass that down to inode_query_version, and
> we'd have a I_VERSION_QUERIED and a I_VERSION_QUERIED_STRICT, and the
> "I care about atime" case ould set the strict one.

This makes correct behaviour reliant on the applicaiton using the
query mechanism correctly. I have my doubts that userspace
developers will be able to understand the subtle difference between
the two options and always choose correctly....

And then there's always the issue that we might end up with both
flags set and we get conflicting bug reports about how atime is not
behaving the way the applications want it to behave.

> Then inode_maybe_inc_iversion() could - for atome updates - skip the
> version update *unless* it sees that I_VERSION_QUERIED_STRICT bit.
> 
> Does that sound sane to people?

I'd much prefer we just do the right thing transparently at the
filesystem level; all we need is for the inode to be flagged that it
should be doing in memory atime updates rather than persistent
updates.

Perhaps the nfs server should just set a new S_LAZYTIME flag on
inodes it accesses similar to how we can set S_NOATIME on inodes to
elide atime updates altogether. Once set, the inode will behave that
way until it is reclaimed from memory....

Cheers,

Dave.
Linus Torvalds Oct. 24, 2023, 4:10 a.m. UTC | #19
On Mon, 23 Oct 2023 at 17:40, Dave Chinner <david@fromorbit.com> wrote:
> >
> > Maybe we don't even need a mode, and could just decide that atime
> > updates aren't i_version updates at all?
>
> We do that already - in memory atime updates don't bump i_version at
> all. The issue is the rare persistent atime update requests that
> still happen - they are the ones that trigger an i_version bump on
> XFS, and one of the relatime heuristics tickle this specific issue.

Yes, yes, but that's what I kind of allude to - maybe people still
want the on-disk atime updates, but do they actually want the
i_version updates just because they want more aggressive atime
updates?

> > Or maybe i_version can update, but callers of getattr() could have two
> > bits for that STATX_CHANGE_COOKIE, one for "I care about atime" and
> > one for others, and we'd pass that down to inode_query_version, and
> > we'd have a I_VERSION_QUERIED and a I_VERSION_QUERIED_STRICT, and the
> > "I care about atime" case ould set the strict one.
>
> This makes correct behaviour reliant on the applicaiton using the
> query mechanism correctly. I have my doubts that userspace
> developers will be able to understand the subtle difference between
> the two options and always choose correctly....

I don't think we _have_ a user space interface.

At least the STATX_CHANGE_COOKIE bit isn't exposed to user space at
all. Not in the uapi headers, but not even in xstat():

        /* STATX_CHANGE_COOKIE is kernel-only for now */
        tmp.stx_mask = stat->result_mask & ~STATX_CHANGE_COOKIE;

So the *only* users of STATX_CHANGE_COOKIE seem to be entirely
in-kernel, unless there is something I'm missing where somebody uses
i_version through some other interface (random ioctl?).

End result: splitting STATX_CHANGE_COOKIE into a "I don't care about
atime" and a "give me all change events" shouldn't possibly break
anything that I can see.

The only other uses of inode_query_iversion() seem to be the explicit
directory optimizations (ie the "I'm caching the offset and don't want
to re-check that it's valid unless required, so give me the inode
version for the directory as a way to decide if I need to re-check"
thing).

And those *definitely* don't want i_version updates on any atime updates.

There might be some other use of inode_query_iversion() that I missed,
of course.  But from a quick look, it really looks to me like we can
relax our i_version updates.

              Linus
Amir Goldstein Oct. 24, 2023, 7:08 a.m. UTC | #20
On Tue, Oct 24, 2023 at 6:40 AM Dave Chinner <david@fromorbit.com> wrote:
>
> On Mon, Oct 23, 2023 at 02:18:12PM -1000, Linus Torvalds wrote:
> > On Mon, 23 Oct 2023 at 13:26, Dave Chinner <david@fromorbit.com> wrote:
> > >
> > > The problem is the first read request after a modification has been
> > > made. That is causing relatime to see mtime > atime and triggering
> > > an atime update. XFS sees this, does an atime update, and in
> > > committing that persistent inode metadata update, it calls
> > > inode_maybe_inc_iversion(force = false) to check if an iversion
> > > update is necessary. The VFS sees I_VERSION_QUERIED, and so it bumps
> > > i_version and tells XFS to persist it.
> >
> > Could we perhaps just have a mode where we don't increment i_version
> > for just atime updates?
> >
> > Maybe we don't even need a mode, and could just decide that atime
> > updates aren't i_version updates at all?
>
> We do that already - in memory atime updates don't bump i_version at
> all. The issue is the rare persistent atime update requests that
> still happen - they are the ones that trigger an i_version bump on
> XFS, and one of the relatime heuristics tickle this specific issue.
>
> If we push the problematic persistent atime updates to be in-memory
> updates only, then the whole problem with i_version goes away....
>
> > Yes, yes, it's obviously technically a "inode modification", but does
> > anybody actually *want* atime updates with no actual other changes to
> > be version events?
>
> Well, yes, there was. That's why we defined i_version in the on disk
> format this way well over a decade ago. It was part of some deep
> dark magical HSM beans that allowed the application to combine
> multiple scans for different inode metadata changes into a single
> pass. atime changes was one of the things it needed to know about
> for tiering and space scavenging purposes....
>

But if this is such an ancient mystical program, why do we have to
keep this XFS behavior in the present?
BTW, is this the same HSM whose DMAPI ioctls were deprecated
a few years back?

I mean, I understand that you do not want to change the behavior of
i_version update without an opt-in config or mount option - let the distro
make that choice.
But calling this an "on-disk format change" is a very long stretch.

Does xfs_repair guarantee that changes of atime, or any inode changes
for that matter, update i_version? No, it does not.
So IMO, "atime does not update i_version" is not an "on-disk format change",
it is a runtime behavior change, just like lazytime is.

Thanks,
Amir.
Jeffrey Layton Oct. 24, 2023, 2:24 p.m. UTC | #21
On Tue, 2023-10-24 at 14:40 +1100, Dave Chinner wrote:
> On Mon, Oct 23, 2023 at 02:18:12PM -1000, Linus Torvalds wrote:
> > On Mon, 23 Oct 2023 at 13:26, Dave Chinner <david@fromorbit.com> wrote:
> > > 
> > > The problem is the first read request after a modification has been
> > > made. That is causing relatime to see mtime > atime and triggering
> > > an atime update. XFS sees this, does an atime update, and in
> > > committing that persistent inode metadata update, it calls
> > > inode_maybe_inc_iversion(force = false) to check if an iversion
> > > update is necessary. The VFS sees I_VERSION_QUERIED, and so it bumps
> > > i_version and tells XFS to persist it.
> > 
> > Could we perhaps just have a mode where we don't increment i_version
> > for just atime updates?
> > 
> > Maybe we don't even need a mode, and could just decide that atime
> > updates aren't i_version updates at all?
> 
> We do that already - in memory atime updates don't bump i_version at
> all. The issue is the rare persistent atime update requests that
> still happen - they are the ones that trigger an i_version bump on
> XFS, and one of the relatime heuristics tickle this specific issue.
> 
> If we push the problematic persistent atime updates to be in-memory
> updates only, then the whole problem with i_version goes away....
> 

POSIX (more or less) states that if you're updating the inode for any
reason other than an atime update, then you need to update the ctime.
The NFSv4 spec (more or less) defines the change attribute as something
that should show a change any time the ctime would change.

Now, from mount(8) manpage, in the section on lazytime:

           The on-disk timestamps are updated only when:

           •   the inode needs to be updated for some change unrelated to file timestamps

           •   the application employs fsync(2), syncfs(2), or sync(2)

           •   an undeleted inode is evicted from memory

           •   more than 24 hours have passed since the inode was written to disk.


The first is not a problem for NFS. If we're updating the ctime or mtime
or other attributes, then we _need_ to bump the change attribute
(assuming that something has queried it and can tell a difference).

The second is potentially an issue. If a file has an in-memory atime
update and them someone calls sync(2), XFS will end up bumping the
change attribute.

Ditto for the third. If someone does a getattr and brings an inode back
into core, then you're looking at a cache invalidation on the client.

The fourth is also a problem, once a day your clients will end up
invaliding their caches.

You might think "so what? Once a day is no big deal!", but there are
places that have built up cascading fscache setups across WANs to
distribute large, read-mostly files. This is quite problematic in these
sorts of setups as they end up seeing cache invalidations all over the
place.

noatime is a workaround, but it has its own problems and ugliness, and
it sucks that XFS doesn't "just work" when served by NFS.


> > Yes, yes, it's obviously technically a "inode modification", but does
> > anybody actually *want* atime updates with no actual other changes to
> > be version events?
> 
> Well, yes, there was. That's why we defined i_version in the on disk
> format this way well over a decade ago. It was part of some deep
> dark magical HSM beans that allowed the application to combine
> multiple scans for different inode metadata changes into a single
> pass. atime changes was one of the things it needed to know about
> for tiering and space scavenging purposes....
> 

As Amir points out, is this still behavior that you're required to
preserve? NFS serving is a bit more common than weird HSM stuff. 

Maybe newly created XFS filesystems could be made to update i_version in
a way that nfsd expects? We could add a mkfs.xfs option to allow for the
legacy behavior if required.

> > Or maybe i_version can update, but callers of getattr() could have two
> > bits for that STATX_CHANGE_COOKIE, one for "I care about atime" and
> > one for others, and we'd pass that down to inode_query_version, and
> > we'd have a I_VERSION_QUERIED and a I_VERSION_QUERIED_STRICT, and the
> > "I care about atime" case ould set the strict one.
> 
> This makes correct behaviour reliant on the applicaiton using the
> query mechanism correctly. I have my doubts that userspace
> developers will be able to understand the subtle difference between
> the two options and always choose correctly....
> 
> And then there's always the issue that we might end up with both
> flags set and we get conflicting bug reports about how atime is not
> behaving the way the applications want it to behave.
> 
> > Then inode_maybe_inc_iversion() could - for atome updates - skip the
> > version update *unless* it sees that I_VERSION_QUERIED_STRICT bit.
> > 
> > Does that sound sane to people?
> 
> I'd much prefer we just do the right thing transparently at the
> filesystem level; all we need is for the inode to be flagged that it
> should be doing in memory atime updates rather than persistent
> updates.
> 
> Perhaps the nfs server should just set a new S_LAZYTIME flag on
> inodes it accesses similar to how we can set S_NOATIME on inodes to
> elide atime updates altogether. Once set, the inode will behave that
> way until it is reclaimed from memory....
> 


Yeah, I think adding this sort of extra complexity on the query side is
probably not what's needed.

I'm also not crazy about trying to treat nfsd's accesses as special in
some way. nfsd is really not doing anything special at all, other than
querying for STATX_CHANGE_COOKIE. 

The problem is on the update side: nfsd expects the STATX_CHANGE_COOKIE
to conform to certain behavior, and XFS simply does not (specifically
around updates to the inode that only change the atime).
Jeffrey Layton Oct. 24, 2023, 6:40 p.m. UTC | #22
On Tue, 2023-10-24 at 10:08 +0300, Amir Goldstein wrote:
> On Tue, Oct 24, 2023 at 6:40 AM Dave Chinner <david@fromorbit.com> wrote:
> > 
> > On Mon, Oct 23, 2023 at 02:18:12PM -1000, Linus Torvalds wrote:
> > > On Mon, 23 Oct 2023 at 13:26, Dave Chinner <david@fromorbit.com> wrote:
> > > > 
> > > > The problem is the first read request after a modification has been
> > > > made. That is causing relatime to see mtime > atime and triggering
> > > > an atime update. XFS sees this, does an atime update, and in
> > > > committing that persistent inode metadata update, it calls
> > > > inode_maybe_inc_iversion(force = false) to check if an iversion
> > > > update is necessary. The VFS sees I_VERSION_QUERIED, and so it bumps
> > > > i_version and tells XFS to persist it.
> > > 
> > > Could we perhaps just have a mode where we don't increment i_version
> > > for just atime updates?
> > > 
> > > Maybe we don't even need a mode, and could just decide that atime
> > > updates aren't i_version updates at all?
> > 
> > We do that already - in memory atime updates don't bump i_version at
> > all. The issue is the rare persistent atime update requests that
> > still happen - they are the ones that trigger an i_version bump on
> > XFS, and one of the relatime heuristics tickle this specific issue.
> > 
> > If we push the problematic persistent atime updates to be in-memory
> > updates only, then the whole problem with i_version goes away....
> > 
> > > Yes, yes, it's obviously technically a "inode modification", but does
> > > anybody actually *want* atime updates with no actual other changes to
> > > be version events?
> > 
> > Well, yes, there was. That's why we defined i_version in the on disk
> > format this way well over a decade ago. It was part of some deep
> > dark magical HSM beans that allowed the application to combine
> > multiple scans for different inode metadata changes into a single
> > pass. atime changes was one of the things it needed to know about
> > for tiering and space scavenging purposes....
> > 
> 
> But if this is such an ancient mystical program, why do we have to
> keep this XFS behavior in the present?
> BTW, is this the same HSM whose DMAPI ioctls were deprecated
> a few years back?
> 
> I mean, I understand that you do not want to change the behavior of
> i_version update without an opt-in config or mount option - let the distro
> make that choice.
> But calling this an "on-disk format change" is a very long stretch.
> 
> Does xfs_repair guarantee that changes of atime, or any inode changes
> for that matter, update i_version? No, it does not.
> So IMO, "atime does not update i_version" is not an "on-disk format change",
> it is a runtime behavior change, just like lazytime is.
> 

This would certainly be my preference. I don't want to break any
existing users though.

Perhaps this ought to be a mkfs option? Existing XFS filesystems could
still behave with the legacy behavior, but we could make mkfs.xfs build
filesystems by default that work like NFS requires.
Jeffrey Layton Oct. 24, 2023, 7:06 p.m. UTC | #23
On Mon, 2023-10-23 at 14:18 -1000, Linus Torvalds wrote:
> On Mon, 23 Oct 2023 at 13:26, Dave Chinner <david@fromorbit.com> wrote:
> > 
> > The problem is the first read request after a modification has been
> > made. That is causing relatime to see mtime > atime and triggering
> > an atime update. XFS sees this, does an atime update, and in
> > committing that persistent inode metadata update, it calls
> > inode_maybe_inc_iversion(force = false) to check if an iversion
> > update is necessary. The VFS sees I_VERSION_QUERIED, and so it bumps
> > i_version and tells XFS to persist it.
> 
> Could we perhaps just have a mode where we don't increment i_version
> for just atime updates?
> 
> Maybe we don't even need a mode, and could just decide that atime
> updates aren't i_version updates at all?
> 
> Yes, yes, it's obviously technically a "inode modification", but does
> anybody actually *want* atime updates with no actual other changes to
> be version events?
> 
> Or maybe i_version can update, but callers of getattr() could have two
> bits for that STATX_CHANGE_COOKIE, one for "I care about atime" and
> one for others, and we'd pass that down to inode_query_version, and
> we'd have a I_VERSION_QUERIED and a I_VERSION_QUERIED_STRICT, and the
> "I care about atime" case ould set the strict one.
> 
> Then inode_maybe_inc_iversion() could - for atome updates - skip the
> version update *unless* it sees that I_VERSION_QUERIED_STRICT bit.
> 
> Does that sound sane to people?
> 
> Because it does sound completely insane to me to say "inode changed"
> and have a cache invalidation just for an atime update.
> 


The new flag idea is a good one. The catch though is that there are no
readers of i_version in-kernel other than NFSD and IMA, so there would
be no in-kernel users of I_VERSION_QUERIED_STRICT.

In earlier discussions, I was given to believe that the problem with
changing how this works in XFS involved offline filesystem access tools.
That said, I didn't press for enough details at the time, so I may have
misunderstood Dave's reticence to change how this works.
Linus Torvalds Oct. 24, 2023, 7:40 p.m. UTC | #24
On Tue, 24 Oct 2023 at 09:07, Jeff Layton <jlayton@kernel.org> wrote:
>
> The new flag idea is a good one. The catch though is that there are no
> readers of i_version in-kernel other than NFSD and IMA, so there would
> be no in-kernel users of I_VERSION_QUERIED_STRICT.

I actually see that as an absolute positive.

I think we should *conceptually* do those two flags, but then realize
that there are no users of the STRICT version, and just skip it.

So practically speaking, we'd end up with just a weaker version of
I_VERSION_QUERIED that is that "I don't care about atime" case.

I really can't find any use that would *want* to see i_version updates
for any atime updates. Ever.

We may have had historical user interfaces for i_version, but I can't
find any currently.

But to be very very clear: I've only done some random grepping, and I
may have missed something. I'm not dismissing Dave's worries, and he
may well be entirely correct.

Somebody would need to do a much more careful check than my "I can't
find anything".

             Linus
Jeffrey Layton Oct. 24, 2023, 8:19 p.m. UTC | #25
On Tue, 2023-10-24 at 09:40 -1000, Linus Torvalds wrote:
> On Tue, 24 Oct 2023 at 09:07, Jeff Layton <jlayton@kernel.org> wrote:
> > 
> > The new flag idea is a good one. The catch though is that there are no
> > readers of i_version in-kernel other than NFSD and IMA, so there would
> > be no in-kernel users of I_VERSION_QUERIED_STRICT.
> 
> I actually see that as an absolute positive.
> 
> I think we should *conceptually* do those two flags, but then realize
> that there are no users of the STRICT version, and just skip it.
> 
> So practically speaking, we'd end up with just a weaker version of
> I_VERSION_QUERIED that is that "I don't care about atime" case.
> 

To be clear, this is not kernel-wide behavior. Most filesystems already
don't bump their i_version on atime updates. XFS is the only one that
does. ext4 used to do that too, but we fixed that several months ago.
I did try to just fix XFS in the same way, but the patch was NAK'ed.

> I really can't find any use that would *want* to see i_version updates
> for any atime updates. Ever.
> 
> We may have had historical user interfaces for i_version, but I can't
> find any currently.
> 
> But to be very very clear: I've only done some random grepping, and I
> may have missed something. I'm not dismissing Dave's worries, and he
> may well be entirely correct.
> 
> Somebody would need to do a much more careful check than my "I can't
> find anything".

Exactly. I'm not really an XFS guy, so I took those folks at their word
that this was behavior that they just can't trivially change.

None of the in-kernel callers that look at i_version want it to be
incremented on atime-onlt updates, however. So IIRC, the objection was
due to offline repair/analysis tools that depend this the value being
incremented in a specific way.
Dave Chinner Oct. 25, 2023, 8:05 a.m. UTC | #26
On Tue, Oct 24, 2023 at 02:40:06PM -0400, Jeff Layton wrote:
> On Tue, 2023-10-24 at 10:08 +0300, Amir Goldstein wrote:
> > On Tue, Oct 24, 2023 at 6:40 AM Dave Chinner <david@fromorbit.com> wrote:
> > > 
> > > On Mon, Oct 23, 2023 at 02:18:12PM -1000, Linus Torvalds wrote:
> > > > On Mon, 23 Oct 2023 at 13:26, Dave Chinner <david@fromorbit.com> wrote:
> > > > > 
> > > > > The problem is the first read request after a modification has been
> > > > > made. That is causing relatime to see mtime > atime and triggering
> > > > > an atime update. XFS sees this, does an atime update, and in
> > > > > committing that persistent inode metadata update, it calls
> > > > > inode_maybe_inc_iversion(force = false) to check if an iversion
> > > > > update is necessary. The VFS sees I_VERSION_QUERIED, and so it bumps
> > > > > i_version and tells XFS to persist it.
> > > > 
> > > > Could we perhaps just have a mode where we don't increment i_version
> > > > for just atime updates?
> > > > 
> > > > Maybe we don't even need a mode, and could just decide that atime
> > > > updates aren't i_version updates at all?
> > > 
> > > We do that already - in memory atime updates don't bump i_version at
> > > all. The issue is the rare persistent atime update requests that
> > > still happen - they are the ones that trigger an i_version bump on
> > > XFS, and one of the relatime heuristics tickle this specific issue.
> > > 
> > > If we push the problematic persistent atime updates to be in-memory
> > > updates only, then the whole problem with i_version goes away....
> > > 
> > > > Yes, yes, it's obviously technically a "inode modification", but does
> > > > anybody actually *want* atime updates with no actual other changes to
> > > > be version events?
> > > 
> > > Well, yes, there was. That's why we defined i_version in the on disk
> > > format this way well over a decade ago. It was part of some deep
> > > dark magical HSM beans that allowed the application to combine
> > > multiple scans for different inode metadata changes into a single
> > > pass. atime changes was one of the things it needed to know about
> > > for tiering and space scavenging purposes....
> > > 
> > 
> > But if this is such an ancient mystical program, why do we have to
> > keep this XFS behavior in the present?
> > BTW, is this the same HSM whose DMAPI ioctls were deprecated
> > a few years back?

Drop the attitude, Amir.

That "ancient mystical program" is this:

https://buy.hpe.com/us/en/enterprise-solutions/high-performance-computing-solutions/high-performance-computing-storage-solutions/hpc-storage-solutions/hpe-data-management-framework-7/p/1010144088

Yup, that product is backed by a proprietary descendent of the Irix
XFS code base XFS that is DMAPI enabled and still in use today. It's
called HPE XFS these days....

> > I mean, I understand that you do not want to change the behavior of
> > i_version update without an opt-in config or mount option - let the distro
> > make that choice.
> > But calling this an "on-disk format change" is a very long stretch.

Telling the person who created, defined and implemented the on disk
format that they don't know what constitutes a change of that
on-disk format seems kinda Dunning-Kruger to me....

There are *lots* of ways that di_changecount is now incompatible
with the VFS change counter. That's now defined as "i_version should
only change when [cm]time is changed".

di_changecount is defined to be a count of the number of changes
made to the attributes of the inode.  It's not just atime at issue
here - we bump di_changecount when make any inode change, including
background work that does not otherwise change timestamps. e.g.
allocation at writeback time, unwritten extent conversion, on-disk
EOF extension at IO completion, removal of speculative
pre-allocation beyond EOF, etc.

IOWs, di_changecount was never defined as a linux "i_version"
counter, regardless of the fact we originally we able to implement
i_version with it - all extra bumps to di_changecount were not
important to the users of i_version for about a decade.

Unfortunately, the new i_version definition is very much
incompatible with the existing di_changecount definition and that's
the underlying problem here. i.e. the problem is not that we bump
i_version on atime, it's that di_changecount is now completely
incompatible with the new i_version change semantics.

To implement the new i_version semantics exactly, we need to add a
new field to the inode to hold this information.
If we change the on disk format like this, then the atime
problems go away because the new field would not get updated on
atime updates. We'd still be bumping di_changecount on atime
updates, though, because that's what is required by the on-disk
format.

I'm really trying to avoid changing the on-disk format unless it
is absolutely necessary. If we can get the in-memory timestamp
updates to avoid tripping di_changecount updates then the atime
problems go away.

If we can get [cm]time sufficiently fine grained that we don't need
i_version, then we can turn off i_version in XFS and di_changecount
ends up being entirely internal. That's what was attempted with
generic multi-grain timestamps, but that hasn't worked.

Another options is for XFS to play it's own internal tricks with
[cm]time granularity and turn off i_version. e.g. limit external
timestamp visibility to 1us and use the remaining dozen bits of the
ns field to hold a change counter for updates within a single coarse
timer tick. This guarantees the timestamp changes within a coarse
tick for the purposes of change detection, but we don't expose those
bits to applications so applications that compare timestamps across
inodes won't get things back to front like was happening with the
multi-grain timestamps....

Another option is to work around the visible symptoms of the
semantic mismatch between i_version and di_changecount. The only
visible symptom we currently know about is the atime vs i_version
issue.  If people are happy for us to simply ignore VFS atime
guidelines (i.e. ignore realtime/lazytime) and do completely our own
stuff with timestamp update deferal, then that also solve the
immediate issues.

> > Does xfs_repair guarantee that changes of atime, or any inode changes
> > for that matter, update i_version? No, it does not.
> > So IMO, "atime does not update i_version" is not an "on-disk format change",
> > it is a runtime behavior change, just like lazytime is.
> 
> This would certainly be my preference. I don't want to break any
> existing users though.

That's why I'm trying to get some kind of consensus on what
rules and/or atime configurations people are happy for me to break
to make it look to users like there's a viable working change
attribute being supplied by XFS without needing to change the on
disk format.

> Perhaps this ought to be a mkfs option? Existing XFS filesystems could
> still behave with the legacy behavior, but we could make mkfs.xfs build
> filesystems by default that work like NFS requires.

If we require mkfs to set a flag to change behaviour, then we're
talking about making an explicit on-disk format change to select the
optional behaviour. That's precisely what I want to avoid.

-Dave.
Amir Goldstein Oct. 25, 2023, 10:41 a.m. UTC | #27
On Wed, Oct 25, 2023 at 11:05 AM Dave Chinner <david@fromorbit.com> wrote:
>
> On Tue, Oct 24, 2023 at 02:40:06PM -0400, Jeff Layton wrote:
> > On Tue, 2023-10-24 at 10:08 +0300, Amir Goldstein wrote:
> > > On Tue, Oct 24, 2023 at 6:40 AM Dave Chinner <david@fromorbit.com> wrote:
> > > >
> > > > On Mon, Oct 23, 2023 at 02:18:12PM -1000, Linus Torvalds wrote:
> > > > > On Mon, 23 Oct 2023 at 13:26, Dave Chinner <david@fromorbit.com> wrote:
> > > > > >
> > > > > > The problem is the first read request after a modification has been
> > > > > > made. That is causing relatime to see mtime > atime and triggering
> > > > > > an atime update. XFS sees this, does an atime update, and in
> > > > > > committing that persistent inode metadata update, it calls
> > > > > > inode_maybe_inc_iversion(force = false) to check if an iversion
> > > > > > update is necessary. The VFS sees I_VERSION_QUERIED, and so it bumps
> > > > > > i_version and tells XFS to persist it.
> > > > >
> > > > > Could we perhaps just have a mode where we don't increment i_version
> > > > > for just atime updates?
> > > > >
> > > > > Maybe we don't even need a mode, and could just decide that atime
> > > > > updates aren't i_version updates at all?
> > > >
> > > > We do that already - in memory atime updates don't bump i_version at
> > > > all. The issue is the rare persistent atime update requests that
> > > > still happen - they are the ones that trigger an i_version bump on
> > > > XFS, and one of the relatime heuristics tickle this specific issue.
> > > >
> > > > If we push the problematic persistent atime updates to be in-memory
> > > > updates only, then the whole problem with i_version goes away....
> > > >
> > > > > Yes, yes, it's obviously technically a "inode modification", but does
> > > > > anybody actually *want* atime updates with no actual other changes to
> > > > > be version events?
> > > >
> > > > Well, yes, there was. That's why we defined i_version in the on disk
> > > > format this way well over a decade ago. It was part of some deep
> > > > dark magical HSM beans that allowed the application to combine
> > > > multiple scans for different inode metadata changes into a single
> > > > pass. atime changes was one of the things it needed to know about
> > > > for tiering and space scavenging purposes....
> > > >
> > >
> > > But if this is such an ancient mystical program, why do we have to
> > > keep this XFS behavior in the present?
> > > BTW, is this the same HSM whose DMAPI ioctls were deprecated
> > > a few years back?
>
> Drop the attitude, Amir.
>
> That "ancient mystical program" is this:
>
> https://buy.hpe.com/us/en/enterprise-solutions/high-performance-computing-solutions/high-performance-computing-storage-solutions/hpc-storage-solutions/hpe-data-management-framework-7/p/1010144088
>

Sorry for the attitude Dave, I somehow got the impression that you
were talking about a hypothetical old program that may be out of use.
I believe that Jeff and Linus got the same impression...

> Yup, that product is backed by a proprietary descendent of the Irix
> XFS code base XFS that is DMAPI enabled and still in use today. It's
> called HPE XFS these days....
>

What do you mean?
Do you mean that the HPE product uses patched XFS?
If so, why is that an upstream concern?

Upstream xfs indeed preserves di_dmstate,di_dmevmask, but it does
not change those state members when file changes happen.

So if mounting an HPE XFS disk on with upstream kernel is not
going to record DMAPI state changes, does it matter if upstream
xfs does not update di_changecount on atime change?

Maybe I did not understand the situation w.r.t HPE XFS.

> > > I mean, I understand that you do not want to change the behavior of
> > > i_version update without an opt-in config or mount option - let the distro
> > > make that choice.
> > > But calling this an "on-disk format change" is a very long stretch.
>
> Telling the person who created, defined and implemented the on disk
> format that they don't know what constitutes a change of that
> on-disk format seems kinda Dunning-Kruger to me....
>

OK. I will choose my words more carefully:

I still do not understand, from everything that you have told us
so far, including the mention of the specific product above,
why not updating di_changecount on atime update constitutes
an on-disk format change and not a runtime behavior change.

You also did not address my comment that xfs_repair does not
update di_changecount on any inode changes to the best of my
code reading abilities.

> There are *lots* of ways that di_changecount is now incompatible
> with the VFS change counter. That's now defined as "i_version should
> only change when [cm]time is changed".
>
> di_changecount is defined to be a count of the number of changes
> made to the attributes of the inode.  It's not just atime at issue
> here - we bump di_changecount when make any inode change, including
> background work that does not otherwise change timestamps. e.g.
> allocation at writeback time, unwritten extent conversion, on-disk
> EOF extension at IO completion, removal of speculative
> pre-allocation beyond EOF, etc.
>

I see.
Does xfs update ctime on all those inode block map changes?

> IOWs, di_changecount was never defined as a linux "i_version"
> counter, regardless of the fact we originally we able to implement
> i_version with it - all extra bumps to di_changecount were not
> important to the users of i_version for about a decade.
>
> Unfortunately, the new i_version definition is very much
> incompatible with the existing di_changecount definition and that's
> the underlying problem here. i.e. the problem is not that we bump
> i_version on atime, it's that di_changecount is now completely
> incompatible with the new i_version change semantics.
>
> To implement the new i_version semantics exactly, we need to add a
> new field to the inode to hold this information.
> If we change the on disk format like this, then the atime
> problems go away because the new field would not get updated on
> atime updates. We'd still be bumping di_changecount on atime
> updates, though, because that's what is required by the on-disk
> format.
>

I fully agree with you that we should avoid on-disk format change.
This is exactly the reason that I'm insisting on the point of clarifying
how exactly, this semantic change of di_changecount is going to
break existing applications that run on upstream kernel.

Thanks,
Amir.
Jeffrey Layton Oct. 25, 2023, 12:25 p.m. UTC | #28
On Wed, 2023-10-25 at 19:05 +1100, Dave Chinner wrote:
> On Tue, Oct 24, 2023 at 02:40:06PM -0400, Jeff Layton wrote:
> > On Tue, 2023-10-24 at 10:08 +0300, Amir Goldstein wrote:
> > > On Tue, Oct 24, 2023 at 6:40 AM Dave Chinner <david@fromorbit.com> wrote:
> > > > 
> > > > On Mon, Oct 23, 2023 at 02:18:12PM -1000, Linus Torvalds wrote:
> > > > > On Mon, 23 Oct 2023 at 13:26, Dave Chinner <david@fromorbit.com> wrote:
> > > > > > 
> > > > > > The problem is the first read request after a modification has been
> > > > > > made. That is causing relatime to see mtime > atime and triggering
> > > > > > an atime update. XFS sees this, does an atime update, and in
> > > > > > committing that persistent inode metadata update, it calls
> > > > > > inode_maybe_inc_iversion(force = false) to check if an iversion
> > > > > > update is necessary. The VFS sees I_VERSION_QUERIED, and so it bumps
> > > > > > i_version and tells XFS to persist it.
> > > > > 
> > > > > Could we perhaps just have a mode where we don't increment i_version
> > > > > for just atime updates?
> > > > > 
> > > > > Maybe we don't even need a mode, and could just decide that atime
> > > > > updates aren't i_version updates at all?
> > > > 
> > > > We do that already - in memory atime updates don't bump i_version at
> > > > all. The issue is the rare persistent atime update requests that
> > > > still happen - they are the ones that trigger an i_version bump on
> > > > XFS, and one of the relatime heuristics tickle this specific issue.
> > > > 
> > > > If we push the problematic persistent atime updates to be in-memory
> > > > updates only, then the whole problem with i_version goes away....
> > > > 
> > > > > Yes, yes, it's obviously technically a "inode modification", but does
> > > > > anybody actually *want* atime updates with no actual other changes to
> > > > > be version events?
> > > > 
> > > > Well, yes, there was. That's why we defined i_version in the on disk
> > > > format this way well over a decade ago. It was part of some deep
> > > > dark magical HSM beans that allowed the application to combine
> > > > multiple scans for different inode metadata changes into a single
> > > > pass. atime changes was one of the things it needed to know about
> > > > for tiering and space scavenging purposes....
> > > > 
> > > 
> > > But if this is such an ancient mystical program, why do we have to
> > > keep this XFS behavior in the present?
> > > BTW, is this the same HSM whose DMAPI ioctls were deprecated
> > > a few years back?
> 
> Drop the attitude, Amir.
> 
> That "ancient mystical program" is this:
> 
> https://buy.hpe.com/us/en/enterprise-solutions/high-performance-computing-solutions/high-performance-computing-storage-solutions/hpc-storage-solutions/hpe-data-management-framework-7/p/1010144088
> 
> Yup, that product is backed by a proprietary descendent of the Irix
> XFS code base XFS that is DMAPI enabled and still in use today. It's
> called HPE XFS these days....
> 
> > > I mean, I understand that you do not want to change the behavior of
> > > i_version update without an opt-in config or mount option - let the distro
> > > make that choice.
> > > But calling this an "on-disk format change" is a very long stretch.
> 
> Telling the person who created, defined and implemented the on disk
> format that they don't know what constitutes a change of that
> on-disk format seems kinda Dunning-Kruger to me....
> 
> There are *lots* of ways that di_changecount is now incompatible
> with the VFS change counter. That's now defined as "i_version should
> only change when [cm]time is changed".
> 
> di_changecount is defined to be a count of the number of changes
> made to the attributes of the inode.  It's not just atime at issue
> here - we bump di_changecount when make any inode change, including
> background work that does not otherwise change timestamps. e.g.
> allocation at writeback time, unwritten extent conversion, on-disk
> EOF extension at IO completion, removal of speculative
> pre-allocation beyond EOF, etc.
> 
> IOWs, di_changecount was never defined as a linux "i_version"
> counter, regardless of the fact we originally we able to implement
> i_version with it - all extra bumps to di_changecount were not
> important to the users of i_version for about a decade.
> 
> Unfortunately, the new i_version definition is very much
> incompatible with the existing di_changecount definition and that's
> the underlying problem here. i.e. the problem is not that we bump
> i_version on atime, it's that di_changecount is now completely
> incompatible with the new i_version change semantics.
> 
> To implement the new i_version semantics exactly, we need to add a
> new field to the inode to hold this information.
> If we change the on disk format like this, then the atime
> problems go away because the new field would not get updated on
> atime updates. We'd still be bumping di_changecount on atime
> updates, though, because that's what is required by the on-disk
> format.
> 
> I'm really trying to avoid changing the on-disk format unless it
> is absolutely necessary. If we can get the in-memory timestamp
> updates to avoid tripping di_changecount updates then the atime
> problems go away.
> 
> If we can get [cm]time sufficiently fine grained that we don't need
> i_version, then we can turn off i_version in XFS and di_changecount
> ends up being entirely internal. That's what was attempted with
> generic multi-grain timestamps, but that hasn't worked.
> 
> Another options is for XFS to play it's own internal tricks with
> [cm]time granularity and turn off i_version. e.g. limit external
> timestamp visibility to 1us and use the remaining dozen bits of the
> ns field to hold a change counter for updates within a single coarse
> timer tick. This guarantees the timestamp changes within a coarse
> tick for the purposes of change detection, but we don't expose those
> bits to applications so applications that compare timestamps across
> inodes won't get things back to front like was happening with the
> multi-grain timestamps....
> 
> Another option is to work around the visible symptoms of the
> semantic mismatch between i_version and di_changecount. The only
> visible symptom we currently know about is the atime vs i_version
> issue.  If people are happy for us to simply ignore VFS atime
> guidelines (i.e. ignore realtime/lazytime) and do completely our own
> stuff with timestamp update deferal, then that also solve the
> immediate issues.
> 
> > > Does xfs_repair guarantee that changes of atime, or any inode changes
> > > for that matter, update i_version? No, it does not.
> > > So IMO, "atime does not update i_version" is not an "on-disk format change",
> > > it is a runtime behavior change, just like lazytime is.
> > 
> > This would certainly be my preference. I don't want to break any
> > existing users though.
> 
> That's why I'm trying to get some kind of consensus on what
> rules and/or atime configurations people are happy for me to break
> to make it look to users like there's a viable working change
> attribute being supplied by XFS without needing to change the on
> disk format.
> 

I agree that the only bone of contention is whether to count atime
updates against the change attribute. I think we have consensus that all
in-kernel users do _not_ want atime updates counted against the change
attribute. The only real question is these "legacy" users of
di_changecount.

> > Perhaps this ought to be a mkfs option? Existing XFS filesystems could
> > still behave with the legacy behavior, but we could make mkfs.xfs build
> > filesystems by default that work like NFS requires.
> 
> If we require mkfs to set a flag to change behaviour, then we're
> talking about making an explicit on-disk format change to select the
> optional behaviour. That's precisely what I want to avoid.
> 

Right. The on-disk di_changecount would have a (subtly) different
meaning at that point.

It's not a change that requires drastic retooling though. If we were to
do this, we wouldn't need to grow the on-disk inode. Booting to an older
kernel would cause the behavior to revert. That's sub-optimal, but not
fatal.

What I don't quite understand is how these tools are accessing
di_changecount?

XFS only accesses the di_changecount to propagate the value to and from
the i_version, and there is nothing besides NFSD and IMA that queries
the i_version value in-kernel. So, this must be done via some sort of
userland tool that is directly accessing the block device (or some 3rd
party kernel module).

In earlier discussions you alluded to some repair and/or analysis tools
that depended on this counter. I took a quick look in xfsprogs, but I
didn't see anything there. Is there a library or something that these
tools use to get at this value?
Dave Chinner Oct. 26, 2023, 2:20 a.m. UTC | #29
On Wed, Oct 25, 2023 at 08:25:35AM -0400, Jeff Layton wrote:
> On Wed, 2023-10-25 at 19:05 +1100, Dave Chinner wrote:
> > On Tue, Oct 24, 2023 at 02:40:06PM -0400, Jeff Layton wrote:
> > > On Tue, 2023-10-24 at 10:08 +0300, Amir Goldstein wrote:
> > > > On Tue, Oct 24, 2023 at 6:40 AM Dave Chinner <david@fromorbit.com> wrote:
> > > > > 
> > > > > On Mon, Oct 23, 2023 at 02:18:12PM -1000, Linus Torvalds wrote:
> > > > > > On Mon, 23 Oct 2023 at 13:26, Dave Chinner <david@fromorbit.com> wrote:
> > > > Does xfs_repair guarantee that changes of atime, or any inode changes
> > > > for that matter, update i_version? No, it does not.
> > > > So IMO, "atime does not update i_version" is not an "on-disk format change",
> > > > it is a runtime behavior change, just like lazytime is.
> > > 
> > > This would certainly be my preference. I don't want to break any
> > > existing users though.
> > 
> > That's why I'm trying to get some kind of consensus on what
> > rules and/or atime configurations people are happy for me to break
> > to make it look to users like there's a viable working change
> > attribute being supplied by XFS without needing to change the on
> > disk format.
> > 
> 
> I agree that the only bone of contention is whether to count atime
> updates against the change attribute. I think we have consensus that all
> in-kernel users do _not_ want atime updates counted against the change
> attribute. The only real question is these "legacy" users of
> di_changecount.

Please stop refering to "legacy users" of di_changecount. Whether
there are users or not is irrelevant - it is defined by the current
on-disk format specification, and as such there may be applications
we do not know about making use of the current behaviour.

It's like a linux syscall - we can't remove them because there may
be some user we don't know about still using that old syscall. We
simply don't make changes that can potentially break user
applications like that.

The on disk format is the same - there is software out that we don't
know about that expects a certain behaviour based on the
specification. We don't break the on disk format by making silent
behavioural changes - we require a feature flag to indicate
behaviour has changed so that applications can take appropriate
actions with stuff they don't understand.

The example for this is the BIGTIME timestamp format change. The on
disk inode structure is physically unchanged, but the contents of
the timestamp fields are encoded very differently. Sure, the older
kernels can read the timestamp data without any sort of problem
occurring, except for the fact the timestamps now appear to be
completely corrupted.

Changing the meaning of ithe contents of di_changecount is no
different. It might look OK and nothing crashes, but nothing can be
inferred from the value in the field because we don't know how it
has been modified.

Hence we can't just change the meaning, encoding or behaviour of an
on disk field that would result in existing kernels and applications
doing the wrong thing with that field (either read or write) without
adding a feature flag to indicate what behaviour that field should
have.

> > > Perhaps this ought to be a mkfs option? Existing XFS filesystems could
> > > still behave with the legacy behavior, but we could make mkfs.xfs build
> > > filesystems by default that work like NFS requires.
> > 
> > If we require mkfs to set a flag to change behaviour, then we're
> > talking about making an explicit on-disk format change to select the
> > optional behaviour. That's precisely what I want to avoid.
> > 
> 
> Right. The on-disk di_changecount would have a (subtly) different
> meaning at that point.
> 
> It's not a change that requires drastic retooling though. If we were to
> do this, we wouldn't need to grow the on-disk inode. Booting to an older
> kernel would cause the behavior to revert. That's sub-optimal, but not
> fatal.

See above: redefining the contents, behaviour or encoding of an on
disk field is a change of the on-disk format specification.

The rules for on disk format changes that we work to were set in
place long before I started working on XFS.  They are sane, well
thought out rules that have stood the test of time and massive new
feature introductions (CRCs, reflink, rmap, etc). And they only work
because we don't allow anyone to bend them for convenience, short
cuts or expediting their pet project.

> What I don't quite understand is how these tools are accessing
> di_changecount?

As I keep saying: this is largely irrelevant to the problem at hand.

> XFS only accesses the di_changecount to propagate the value to and from
> the i_version,

Yes.  XFS has a strong separation between on-disk structures and
in-memory values, and i_version is simply the in-memory field we use
to store the current di_changecount value.  We force bump i_version
every time we modify the inode core regardless of whether anyone has
queried i_version because that's what di_changecount requires. i.e.
the filesystem controls the contents of i_version, not the VFS.

Now that NFS is using a proper abstraction (i.e. vfs_statx()) to get
the change cookie, we really don't need to expose di_changecount in
i_version at all - we could simply copy an internal di_changecount
value into the statx cookie field in xfs_vn_getattr() and there
would be almost no change of behaviour from the perspective of NFS
and IMA at all.

> and there is nothing besides NFSD and IMA that queries
> the i_version value in-kernel. So, this must be done via some sort of
> userland tool that is directly accessing the block device (or some 3rd
> party kernel module).

Yup, both of those sort of applications exist. e.g. the DMAPI kernel
module allows direct access to inode metadata through a custom
bulkstat formatter implementation - it returns different information
comapred to the standard XFS one in the upstream kernel.

> In earlier discussions you alluded to some repair and/or analysis tools
> that depended on this counter.

Yes, and one of those "tools" is *me*.

I frequently look at the di_changecount when doing forensic and/or
failure analysis on filesystem corpses.  SOE analysis, relative
modification activity, etc all give insight into what happened to
the filesystem to get it into the state it is currently in, and
di_changecount provides information no other metadata in the inode
contains.

> I took a quick look in xfsprogs, but I
> didn't see anything there. Is there a library or something that these
> tools use to get at this value?

xfs_db is the tool I use for this, such as:

$ sudo xfs_db -c "sb 0" -c "a rootino" -c "p v3.change_count" /dev/mapper/fast
v3.change_count = 35
$

The root inode in this filesystem has a change count of 35. The root
inode has 32 dirents in it, which means that no entries have ever
been removed or renamed. This sort of insight into the past history
of inode metadata is largely impossible to get any other way, and
it's been the difference between understanding failure and having no
clue more than once.

Most block device parsing applications simply write their own
decoder that walks the on-disk format. That's pretty trivial to do,
developers can get all the information needed to do this from the
on-disk format specification documentation we keep on kernel.org...

-Dave.
Amir Goldstein Oct. 26, 2023, 5:42 a.m. UTC | #30
On Thu, Oct 26, 2023 at 5:21 AM Dave Chinner <david@fromorbit.com> wrote:
>
> On Wed, Oct 25, 2023 at 08:25:35AM -0400, Jeff Layton wrote:
> > On Wed, 2023-10-25 at 19:05 +1100, Dave Chinner wrote:
> > > On Tue, Oct 24, 2023 at 02:40:06PM -0400, Jeff Layton wrote:
> > > > On Tue, 2023-10-24 at 10:08 +0300, Amir Goldstein wrote:
> > > > > On Tue, Oct 24, 2023 at 6:40 AM Dave Chinner <david@fromorbit.com> wrote:
> > > > > >
> > > > > > On Mon, Oct 23, 2023 at 02:18:12PM -1000, Linus Torvalds wrote:
> > > > > > > On Mon, 23 Oct 2023 at 13:26, Dave Chinner <david@fromorbit.com> wrote:
> > > > > Does xfs_repair guarantee that changes of atime, or any inode changes
> > > > > for that matter, update i_version? No, it does not.
> > > > > So IMO, "atime does not update i_version" is not an "on-disk format change",
> > > > > it is a runtime behavior change, just like lazytime is.
> > > >
> > > > This would certainly be my preference. I don't want to break any
> > > > existing users though.
> > >
> > > That's why I'm trying to get some kind of consensus on what
> > > rules and/or atime configurations people are happy for me to break
> > > to make it look to users like there's a viable working change
> > > attribute being supplied by XFS without needing to change the on
> > > disk format.
> > >
> >
> > I agree that the only bone of contention is whether to count atime
> > updates against the change attribute. I think we have consensus that all
> > in-kernel users do _not_ want atime updates counted against the change
> > attribute. The only real question is these "legacy" users of
> > di_changecount.
>
> Please stop refering to "legacy users" of di_changecount. Whether
> there are users or not is irrelevant - it is defined by the current
> on-disk format specification, and as such there may be applications
> we do not know about making use of the current behaviour.
>
> It's like a linux syscall - we can't remove them because there may
> be some user we don't know about still using that old syscall. We
> simply don't make changes that can potentially break user
> applications like that.
>
> The on disk format is the same - there is software out that we don't
> know about that expects a certain behaviour based on the
> specification. We don't break the on disk format by making silent
> behavioural changes - we require a feature flag to indicate
> behaviour has changed so that applications can take appropriate
> actions with stuff they don't understand.
>
> The example for this is the BIGTIME timestamp format change. The on
> disk inode structure is physically unchanged, but the contents of
> the timestamp fields are encoded very differently. Sure, the older
> kernels can read the timestamp data without any sort of problem
> occurring, except for the fact the timestamps now appear to be
> completely corrupted.
>
> Changing the meaning of ithe contents of di_changecount is no
> different. It might look OK and nothing crashes, but nothing can be
> inferred from the value in the field because we don't know how it
> has been modified.
>

I don't agree that this change is the same as BIGTIME change,
but it is a good queue to ask:
BIGTIME has an on-disk feature bit in super block that can be set on an
existing filesystem (and not cleared?).
BIGTIME also has an on-disk inode flag to specify the format in which a
specific inode timestampts are stored.

If we were to change the xfs on-disk to change the *meaning* (not the
format that the counter is stored) of di_changecount, would the feature
flag need be RO_COMPAT?
Would this require a per-inode on-disk flag that declares the meaning
of di_changecount on that specific inode?

Neither of those changes is going to be very hard to do btw.
Following the footsteps of the BIGTIME conversion, but without the
need for an actual format convertors.

Thanks,
Amir.
Jeffrey Layton Oct. 27, 2023, 10:35 a.m. UTC | #31
On Thu, 2023-10-26 at 13:20 +1100, Dave Chinner wrote:
> On Wed, Oct 25, 2023 at 08:25:35AM -0400, Jeff Layton wrote:
> > On Wed, 2023-10-25 at 19:05 +1100, Dave Chinner wrote:
> > > On Tue, Oct 24, 2023 at 02:40:06PM -0400, Jeff Layton wrote:
> > > > On Tue, 2023-10-24 at 10:08 +0300, Amir Goldstein wrote:
> > > > > On Tue, Oct 24, 2023 at 6:40 AM Dave Chinner <david@fromorbit.com> wrote:
> > > > > > 
> > > > > > On Mon, Oct 23, 2023 at 02:18:12PM -1000, Linus Torvalds wrote:
> > > > > > > On Mon, 23 Oct 2023 at 13:26, Dave Chinner <david@fromorbit.com> wrote:
> > > > > Does xfs_repair guarantee that changes of atime, or any inode changes
> > > > > for that matter, update i_version? No, it does not.
> > > > > So IMO, "atime does not update i_version" is not an "on-disk format change",
> > > > > it is a runtime behavior change, just like lazytime is.
> > > > 
> > > > This would certainly be my preference. I don't want to break any
> > > > existing users though.
> > > 
> > > That's why I'm trying to get some kind of consensus on what
> > > rules and/or atime configurations people are happy for me to break
> > > to make it look to users like there's a viable working change
> > > attribute being supplied by XFS without needing to change the on
> > > disk format.
> > > 
> > 
> > I agree that the only bone of contention is whether to count atime
> > updates against the change attribute. I think we have consensus that all
> > in-kernel users do _not_ want atime updates counted against the change
> > attribute. The only real question is these "legacy" users of
> > di_changecount.
> 
> Please stop refering to "legacy users" of di_changecount. Whether
> there are users or not is irrelevant - it is defined by the current
> on-disk format specification, and as such there may be applications
> we do not know about making use of the current behaviour.
> 
> It's like a linux syscall - we can't remove them because there may
> be some user we don't know about still using that old syscall. We
> simply don't make changes that can potentially break user
> applications like that.
> 
> The on disk format is the same - there is software out that we don't
> know about that expects a certain behaviour based on the
> specification. We don't break the on disk format by making silent
> behavioural changes - we require a feature flag to indicate
> behaviour has changed so that applications can take appropriate
> actions with stuff they don't understand.
> 
> The example for this is the BIGTIME timestamp format change. The on
> disk inode structure is physically unchanged, but the contents of
> the timestamp fields are encoded very differently. Sure, the older
> kernels can read the timestamp data without any sort of problem
> occurring, except for the fact the timestamps now appear to be
> completely corrupted.
> 
> Changing the meaning of ithe contents of di_changecount is no
> different. It might look OK and nothing crashes, but nothing can be
> inferred from the value in the field because we don't know how it
> has been modified.
> 
> Hence we can't just change the meaning, encoding or behaviour of an
> on disk field that would result in existing kernels and applications
> doing the wrong thing with that field (either read or write) without
> adding a feature flag to indicate what behaviour that field should
> have.
> 
> > > > Perhaps this ought to be a mkfs option? Existing XFS filesystems could
> > > > still behave with the legacy behavior, but we could make mkfs.xfs build
> > > > filesystems by default that work like NFS requires.
> > > 
> > > If we require mkfs to set a flag to change behaviour, then we're
> > > talking about making an explicit on-disk format change to select the
> > > optional behaviour. That's precisely what I want to avoid.
> > > 
> > 
> > Right. The on-disk di_changecount would have a (subtly) different
> > meaning at that point.
> > 
> > It's not a change that requires drastic retooling though. If we were to
> > do this, we wouldn't need to grow the on-disk inode. Booting to an older
> > kernel would cause the behavior to revert. That's sub-optimal, but not
> > fatal.
> 
> See above: redefining the contents, behaviour or encoding of an on
> disk field is a change of the on-disk format specification.
> 
> The rules for on disk format changes that we work to were set in
> place long before I started working on XFS.  They are sane, well
> thought out rules that have stood the test of time and massive new
> feature introductions (CRCs, reflink, rmap, etc). And they only work
> because we don't allow anyone to bend them for convenience, short
> cuts or expediting their pet project.
> 
> > What I don't quite understand is how these tools are accessing
> > di_changecount?
> 
> As I keep saying: this is largely irrelevant to the problem at hand.
> 
> > XFS only accesses the di_changecount to propagate the value to and from
> > the i_version,
> 
> Yes.  XFS has a strong separation between on-disk structures and
> in-memory values, and i_version is simply the in-memory field we use
> to store the current di_changecount value.  We force bump i_version
> every time we modify the inode core regardless of whether anyone has
> queried i_version because that's what di_changecount requires. i.e.
> the filesystem controls the contents of i_version, not the VFS.
> 
> Now that NFS is using a proper abstraction (i.e. vfs_statx()) to get
> the change cookie, we really don't need to expose di_changecount in
> i_version at all - we could simply copy an internal di_changecount
> value into the statx cookie field in xfs_vn_getattr() and there
> would be almost no change of behaviour from the perspective of NFS
> and IMA at all.
> 
> > and there is nothing besides NFSD and IMA that queries
> > the i_version value in-kernel. So, this must be done via some sort of
> > userland tool that is directly accessing the block device (or some 3rd
> > party kernel module).
> 
> Yup, both of those sort of applications exist. e.g. the DMAPI kernel
> module allows direct access to inode metadata through a custom
> bulkstat formatter implementation - it returns different information
> comapred to the standard XFS one in the upstream kernel.
> 
> > In earlier discussions you alluded to some repair and/or analysis tools
> > that depended on this counter.
> 
> Yes, and one of those "tools" is *me*.
> 
> I frequently look at the di_changecount when doing forensic and/or
> failure analysis on filesystem corpses.  SOE analysis, relative
> modification activity, etc all give insight into what happened to
> the filesystem to get it into the state it is currently in, and
> di_changecount provides information no other metadata in the inode
> contains.
> 
> > I took a quick look in xfsprogs, but I
> > didn't see anything there. Is there a library or something that these
> > tools use to get at this value?
> 
> xfs_db is the tool I use for this, such as:
> 
> $ sudo xfs_db -c "sb 0" -c "a rootino" -c "p v3.change_count" /dev/mapper/fast
> v3.change_count = 35
> $
> 
> The root inode in this filesystem has a change count of 35. The root
> inode has 32 dirents in it, which means that no entries have ever
> been removed or renamed. This sort of insight into the past history
> of inode metadata is largely impossible to get any other way, and
> it's been the difference between understanding failure and having no
> clue more than once.
> 
> Most block device parsing applications simply write their own
> decoder that walks the on-disk format. That's pretty trivial to do,
> developers can get all the information needed to do this from the
> on-disk format specification documentation we keep on kernel.org...
> 

Fair enough. I'm not here to tell you that you guys that you need to
change how di_changecount works. If it's too valuable to keep it
counting atime-only updates, then so be it.

If that's the case however, and given that the multigrain timestamp work
is effectively dead, then I don't see an alternative to growing the on-
disk inode. Do you?
Dave Chinner Oct. 30, 2023, 10:37 p.m. UTC | #32
On Fri, Oct 27, 2023 at 06:35:58AM -0400, Jeff Layton wrote:
> On Thu, 2023-10-26 at 13:20 +1100, Dave Chinner wrote:
> > On Wed, Oct 25, 2023 at 08:25:35AM -0400, Jeff Layton wrote:
> > > On Wed, 2023-10-25 at 19:05 +1100, Dave Chinner wrote:
> > > > On Tue, Oct 24, 2023 at 02:40:06PM -0400, Jeff Layton wrote:
> > > In earlier discussions you alluded to some repair and/or analysis tools
> > > that depended on this counter.
> > 
> > Yes, and one of those "tools" is *me*.
> > 
> > I frequently look at the di_changecount when doing forensic and/or
> > failure analysis on filesystem corpses.  SOE analysis, relative
> > modification activity, etc all give insight into what happened to
> > the filesystem to get it into the state it is currently in, and
> > di_changecount provides information no other metadata in the inode
> > contains.
> > 
> > > I took a quick look in xfsprogs, but I
> > > didn't see anything there. Is there a library or something that these
> > > tools use to get at this value?
> > 
> > xfs_db is the tool I use for this, such as:
> > 
> > $ sudo xfs_db -c "sb 0" -c "a rootino" -c "p v3.change_count" /dev/mapper/fast
> > v3.change_count = 35
> > $
> > 
> > The root inode in this filesystem has a change count of 35. The root
> > inode has 32 dirents in it, which means that no entries have ever
> > been removed or renamed. This sort of insight into the past history
> > of inode metadata is largely impossible to get any other way, and
> > it's been the difference between understanding failure and having no
> > clue more than once.
> > 
> > Most block device parsing applications simply write their own
> > decoder that walks the on-disk format. That's pretty trivial to do,
> > developers can get all the information needed to do this from the
> > on-disk format specification documentation we keep on kernel.org...
> > 
> 
> Fair enough. I'm not here to tell you that you guys that you need to
> change how di_changecount works. If it's too valuable to keep it
> counting atime-only updates, then so be it.
> 
> If that's the case however, and given that the multigrain timestamp work
> is effectively dead, then I don't see an alternative to growing the on-
> disk inode. Do you?

Yes, I do see alternatives. That's what I've been trying
(unsuccessfully) to describe and get consensus on. I feel like I'm
being ignored and rail-roaded here, because nobody is even
acknowledging that I'm proposing alternatives and keeps insisting
that the only solution is a change of on-disk format.

So, I'll summarise the situation *yet again* in the hope that this
time I won't get people arguing about atime vs i-version and what
constitutes an on-disk format change because that goes nowhere and
does nothing to determine which solution might be acceptible.

The basic situation is this:

If XFS can ignore relatime or lazytime persistent updates for given
situations, then *we don't need to make periodic on-disk updates of
atime*. This makes the whole problem of "persistent atime update bumps
i_version" go away because then we *aren't making persistent atime
updates* except when some other persistent modification that bumps
[cm]time occurs.

But I don't want to do this unconditionally - for systems not
running anything that samples i_version we want relatime/lazytime
to behave as they are supposed to and do periodic persistent updates
as per normal. Principle of least surprise and all that jazz.

So we really need an indication for inodes that we should enable this
mode for the inode. I have asked if we can have per-operation
context flag to trigger this given the needs for io_uring to have
context flags for timestamp updates to be added. 

I have asked if we can have an inode flag set by the VFS or
application code for this. e.g. a flag set by nfsd whenever it accesses a
given inode.

I have asked if this inode flag can just be triggered if we ever see
I_VERSION_QUERIED set or statx is used to retrieve a change cookie,
and whether this is a reliable mechanism for setting such a flag.

I have suggested mechanisms for using masked off bits of timestamps
to encode sub-timestamp granularity change counts and keep them
invisible to userspace and then not using i_version at all for XFS.
This avoids all the problems that the multi-grain timestamp
infrastructure exposed due to variable granularity of user visible
timestamps and ordering across inodes with different granularity.
This is potentially a general solution, too.

So, yeah, there are *lots* of ways we can solve this problem without
needing to change on-disk formats.

-Dave.
Linus Torvalds Oct. 30, 2023, 11:11 p.m. UTC | #33
On Mon, 30 Oct 2023 at 12:37, Dave Chinner <david@fromorbit.com> wrote:
>
> If XFS can ignore relatime or lazytime persistent updates for given
> situations, then *we don't need to make periodic on-disk updates of
> atime*. This makes the whole problem of "persistent atime update bumps
> i_version" go away because then we *aren't making persistent atime
> updates* except when some other persistent modification that bumps
> [cm]time occurs.

Well, I think this should be split into two independent questions:

 (a) are relatime or lazytime atime updates persistent if nothing else changes?

 (b) do atime updates _ever_ update i_version *regardless* of relatime
or lazytime?

and honestly, I think the best answer to (b) would be that "no,
i_version should simply not change for atime updates". And I think
that answer is what it is because no user of i_version seems to want
it.

Now, the reason it's a single question for you is that apparently for
XFS, the only thing that matters is "inode was written to disk" and
that "di_changecount" value is thus related to the persistence of
atime updates, but splitting di_changecount out to be a separate thing
from i_version seems to be on the table, so I think those two things
really could be independent issues.

> But I don't want to do this unconditionally - for systems not
> running anything that samples i_version we want relatime/lazytime
> to behave as they are supposed to and do periodic persistent updates
> as per normal. Principle of least surprise and all that jazz.

Well - see above: I think in a perfect world, we'd simply never change
i_version at all for any atime updates, and relatime/lazytime simply
wouldn't be an issue at all wrt i_version.

Wouldn't _that_ be the trule "least surprising" behavior? Considering
that nobody wants i_version to change for what are otherwise pure
reads (that's kind of the *definition* of atime, after all).

Now, the annoyance here is that *both* (a) and (b) then have that
impact of "i_version no longer tracks di_changecount".

So I don't think the issue here is "i_version" per se. I think in a
vacuum, the best option of i_version is pretty obvious.  But if you
want i_version to track di_changecount, *then* you end up with that
situation where the persistence of atime matters, and i_version needs
to update whenever a (persistent) atime update happens.

> So we really need an indication for inodes that we should enable this
> mode for the inode. I have asked if we can have per-operation
> context flag to trigger this given the needs for io_uring to have
> context flags for timestamp updates to be added.

I really think some kind of new and even *more* complex and
non-intuitive behavior is the worst of both worlds. Having atime
updates be conditionally persistent - on top of already being delayed
by lazytime/relatime - and having the persistence magically change
depending on whether something wants to get i_version updates - sounds
just completely crazy.

Particularly as *none* of the people who want i_version updates
actually want them for atime at all.

So I really think this all boils down to "is xfs really willing to
split bi_changecount from i_version"?

> I have asked if we can have an inode flag set by the VFS or
> application code for this. e.g. a flag set by nfsd whenever it accesses a
> given inode.
>
> I have asked if this inode flag can just be triggered if we ever see
> I_VERSION_QUERIED set or statx is used to retrieve a change cookie,
> and whether this is a reliable mechanism for setting such a flag.

See above: linking this to I_VERSION_QUERIED is horrific. The people
who set that bit do *NOT* want atime updates to change i_version under
any circumstances. It was always a mistake.

This really is all *entirely* an artifact of that "bi_changecount" vs
"i_version" being tied together. You did seem to imply that you'd be
ok with having "bi_changecount" be split from i_version, ie from an
earlier email in this thread:

 "Now that NFS is using a proper abstraction (i.e. vfs_statx()) to get
  the change cookie, we really don't need to expose di_changecount in
  i_version at all - we could simply copy an internal di_changecount
  value into the statx cookie field in xfs_vn_getattr() and there
  would be almost no change of behaviour from the perspective of NFS
  and IMA at all"

but while I suspect *that* part is easy and straightforward, the
problem then becomes one of "what about the persistence of i_version",
and then you'd need a new field for *that* anyway, and would want a
new on-disk format regardless.

           Linus
ronnie sahlberg Oct. 30, 2023, 11:34 p.m. UTC | #34
On Fri, 27 Oct 2023 at 20:36, Jeff Layton <jlayton@kernel.org> wrote:
>
> On Thu, 2023-10-26 at 13:20 +1100, Dave Chinner wrote:
> > On Wed, Oct 25, 2023 at 08:25:35AM -0400, Jeff Layton wrote:
> > > On Wed, 2023-10-25 at 19:05 +1100, Dave Chinner wrote:
> > > > On Tue, Oct 24, 2023 at 02:40:06PM -0400, Jeff Layton wrote:
> > > > > On Tue, 2023-10-24 at 10:08 +0300, Amir Goldstein wrote:
> > > > > > On Tue, Oct 24, 2023 at 6:40 AM Dave Chinner <david@fromorbit.com> wrote:
> > > > > > >
> > > > > > > On Mon, Oct 23, 2023 at 02:18:12PM -1000, Linus Torvalds wrote:
> > > > > > > > On Mon, 23 Oct 2023 at 13:26, Dave Chinner <david@fromorbit.com> wrote:
> > > > > > Does xfs_repair guarantee that changes of atime, or any inode changes
> > > > > > for that matter, update i_version? No, it does not.
> > > > > > So IMO, "atime does not update i_version" is not an "on-disk format change",
> > > > > > it is a runtime behavior change, just like lazytime is.
> > > > >
> > > > > This would certainly be my preference. I don't want to break any
> > > > > existing users though.
> > > >
> > > > That's why I'm trying to get some kind of consensus on what
> > > > rules and/or atime configurations people are happy for me to break
> > > > to make it look to users like there's a viable working change
> > > > attribute being supplied by XFS without needing to change the on
> > > > disk format.
> > > >
> > >
> > > I agree that the only bone of contention is whether to count atime
> > > updates against the change attribute. I think we have consensus that all
> > > in-kernel users do _not_ want atime updates counted against the change
> > > attribute. The only real question is these "legacy" users of
> > > di_changecount.
> >
> > Please stop refering to "legacy users" of di_changecount. Whether
> > there are users or not is irrelevant - it is defined by the current
> > on-disk format specification, and as such there may be applications
> > we do not know about making use of the current behaviour.
> >
> > It's like a linux syscall - we can't remove them because there may
> > be some user we don't know about still using that old syscall. We
> > simply don't make changes that can potentially break user
> > applications like that.
> >
> > The on disk format is the same - there is software out that we don't
> > know about that expects a certain behaviour based on the
> > specification. We don't break the on disk format by making silent
> > behavioural changes - we require a feature flag to indicate
> > behaviour has changed so that applications can take appropriate
> > actions with stuff they don't understand.
> >
> > The example for this is the BIGTIME timestamp format change. The on
> > disk inode structure is physically unchanged, but the contents of
> > the timestamp fields are encoded very differently. Sure, the older
> > kernels can read the timestamp data without any sort of problem
> > occurring, except for the fact the timestamps now appear to be
> > completely corrupted.
> >
> > Changing the meaning of ithe contents of di_changecount is no
> > different. It might look OK and nothing crashes, but nothing can be
> > inferred from the value in the field because we don't know how it
> > has been modified.
> >
> > Hence we can't just change the meaning, encoding or behaviour of an
> > on disk field that would result in existing kernels and applications
> > doing the wrong thing with that field (either read or write) without
> > adding a feature flag to indicate what behaviour that field should
> > have.
> >
> > > > > Perhaps this ought to be a mkfs option? Existing XFS filesystems could
> > > > > still behave with the legacy behavior, but we could make mkfs.xfs build
> > > > > filesystems by default that work like NFS requires.
> > > >
> > > > If we require mkfs to set a flag to change behaviour, then we're
> > > > talking about making an explicit on-disk format change to select the
> > > > optional behaviour. That's precisely what I want to avoid.
> > > >
> > >
> > > Right. The on-disk di_changecount would have a (subtly) different
> > > meaning at that point.
> > >
> > > It's not a change that requires drastic retooling though. If we were to
> > > do this, we wouldn't need to grow the on-disk inode. Booting to an older
> > > kernel would cause the behavior to revert. That's sub-optimal, but not
> > > fatal.
> >
> > See above: redefining the contents, behaviour or encoding of an on
> > disk field is a change of the on-disk format specification.
> >
> > The rules for on disk format changes that we work to were set in
> > place long before I started working on XFS.  They are sane, well
> > thought out rules that have stood the test of time and massive new
> > feature introductions (CRCs, reflink, rmap, etc). And they only work
> > because we don't allow anyone to bend them for convenience, short
> > cuts or expediting their pet project.
> >
> > > What I don't quite understand is how these tools are accessing
> > > di_changecount?
> >
> > As I keep saying: this is largely irrelevant to the problem at hand.
> >
> > > XFS only accesses the di_changecount to propagate the value to and from
> > > the i_version,
> >
> > Yes.  XFS has a strong separation between on-disk structures and
> > in-memory values, and i_version is simply the in-memory field we use
> > to store the current di_changecount value.  We force bump i_version
> > every time we modify the inode core regardless of whether anyone has
> > queried i_version because that's what di_changecount requires. i.e.
> > the filesystem controls the contents of i_version, not the VFS.
> >
> > Now that NFS is using a proper abstraction (i.e. vfs_statx()) to get
> > the change cookie, we really don't need to expose di_changecount in
> > i_version at all - we could simply copy an internal di_changecount
> > value into the statx cookie field in xfs_vn_getattr() and there
> > would be almost no change of behaviour from the perspective of NFS
> > and IMA at all.
> >
> > > and there is nothing besides NFSD and IMA that queries
> > > the i_version value in-kernel. So, this must be done via some sort of
> > > userland tool that is directly accessing the block device (or some 3rd
> > > party kernel module).
> >
> > Yup, both of those sort of applications exist. e.g. the DMAPI kernel
> > module allows direct access to inode metadata through a custom
> > bulkstat formatter implementation - it returns different information
> > comapred to the standard XFS one in the upstream kernel.
> >
> > > In earlier discussions you alluded to some repair and/or analysis tools
> > > that depended on this counter.
> >
> > Yes, and one of those "tools" is *me*.
> >
> > I frequently look at the di_changecount when doing forensic and/or
> > failure analysis on filesystem corpses.  SOE analysis, relative
> > modification activity, etc all give insight into what happened to
> > the filesystem to get it into the state it is currently in, and
> > di_changecount provides information no other metadata in the inode
> > contains.
> >
> > > I took a quick look in xfsprogs, but I
> > > didn't see anything there. Is there a library or something that these
> > > tools use to get at this value?
> >
> > xfs_db is the tool I use for this, such as:
> >
> > $ sudo xfs_db -c "sb 0" -c "a rootino" -c "p v3.change_count" /dev/mapper/fast
> > v3.change_count = 35
> > $
> >
> > The root inode in this filesystem has a change count of 35. The root
> > inode has 32 dirents in it, which means that no entries have ever
> > been removed or renamed. This sort of insight into the past history
> > of inode metadata is largely impossible to get any other way, and
> > it's been the difference between understanding failure and having no
> > clue more than once.
> >
> > Most block device parsing applications simply write their own
> > decoder that walks the on-disk format. That's pretty trivial to do,
> > developers can get all the information needed to do this from the
> > on-disk format specification documentation we keep on kernel.org...
> >
>
> Fair enough. I'm not here to tell you that you guys that you need to
> change how di_changecount works. If it's too valuable to keep it
> counting atime-only updates, then so be it.
>
> If that's the case however, and given that the multigrain timestamp work
> is effectively dead, then I don't see an alternative to growing the on-
> disk inode. Do you?

Would a new mount option be a viable alternative?
A new option that would when used change the semantics of these fields
to what NFS needs?
With the caveat: using this mount option may break other special tools
that depend on the default
semantics.


> --
> Jeff Layton <jlayton@kernel.org>
Dave Chinner Oct. 31, 2023, 1:42 a.m. UTC | #35
On Mon, Oct 30, 2023 at 01:11:56PM -1000, Linus Torvalds wrote:
> On Mon, 30 Oct 2023 at 12:37, Dave Chinner <david@fromorbit.com> wrote:
> >
> > If XFS can ignore relatime or lazytime persistent updates for given
> > situations, then *we don't need to make periodic on-disk updates of
> > atime*. This makes the whole problem of "persistent atime update bumps
> > i_version" go away because then we *aren't making persistent atime
> > updates* except when some other persistent modification that bumps
> > [cm]time occurs.
> 
> Well, I think this should be split into two independent questions:
> 
>  (a) are relatime or lazytime atime updates persistent if nothing else changes?

They only become persistent after 24 hours or, in the case of
relatime, immediately persistent if mtime < atime (i.e. read after a
modification). Those are the only times that the VFS triggers
persistent writeback of atime, and it's the latter case (mtime <
atime) that is the specific trigger that exposed the problem with
atime bumping i_version in the first place.

>  (b) do atime updates _ever_ update i_version *regardless* of relatime
> or lazytime?
>
> and honestly, I think the best answer to (b) would be that "no,
> i_version should simply not change for atime updates". And I think
> that answer is what it is because no user of i_version seems to want
> it.

As I keep repeating: Repeatedly stating that "atime should not bump
i_version" does not address the questions I'm asking *at all*.

> Now, the reason it's a single question for you is that apparently for
> XFS, the only thing that matters is "inode was written to disk" and
> that "di_changecount" value is thus related to the persistence of
> atime updates, but splitting di_changecount out to be a separate thing
> from i_version seems to be on the table, so I think those two things
> really could be independent issues.

Wrong way around - we'd have to split i_version out from
di_changecount. It's i_version that has changed semantics, not
di_changecount, and di_changecount behaviour must remain unchanged.

What I really don't want to do is implement a new i_version field in
the XFS on-disk format. What this redefinition of i_version
semantics has made clear is that i_version is *user defined
metadata*, not internal filesystem metadata that is defined by the
filesystem on-disk format.

User defined persistent metadata *belongs in xattrs*, not in the
core filesystem on-disk formats. If the VFS wants to define and
manage i_version behaviour, smeantics and persistence independently
of the filesystems that manage the persistent storage (as it clearly
does!) then we should treat it just like any other VFS defined inode
metadata (e.g. per inode objects like security constraints, ACLs,
fsverity digests, fscrypt keys, etc). i.e. it should be in a named
xattr, not directly implemented in the filesystem on-disk format
deinfitions.

Then the application can change the meaning of the metadata whenever
and however it likes. Then filesystem developers just don't need
to care about it at all because the VFS specific persistent metadata
is not part of the on-disk format we need to maintain
cross-platform forwards and backwards compatibility for.

> > But I don't want to do this unconditionally - for systems not
> > running anything that samples i_version we want relatime/lazytime
> > to behave as they are supposed to and do periodic persistent updates
> > as per normal. Principle of least surprise and all that jazz.
> 
> Well - see above: I think in a perfect world, we'd simply never change
> i_version at all for any atime updates, and relatime/lazytime simply
> wouldn't be an issue at all wrt i_version.

Right, that's what I'd like, especially as the new definition of
i_version - "only change when [cm]time changes" - means that the VFS
i_version is really now just a glorified timestamp.

> Wouldn't _that_ be the trule "least surprising" behavior? Considering
> that nobody wants i_version to change for what are otherwise pure
> reads (that's kind of the *definition* of atime, after all).

So, if you don't like the idea of us ignoring relatime/lazytime
conditionally, are we allowed to simply ignore them *all the time*
and do all timestamp updates in the manner that causes users the
least amount of pain?

I mean, relatime only exists because atime updates cause users pain.
lazytime only exists because relatime doesn't address the pain that
timestamp updates cause mixed read/write or pure O_DSYNC overwrite
workloads pain. noatime is a pain because it loses all atime
updates.

There is no "one size is right for everyone", so why not just let
filesystems do what is most efficient from an internal IO and
persistence POV whilst still maintaining the majority of "expected"
behaviours?

Keep in mind, though, that this is all moot if we can get rid of
i_version entirely....

> Now, the annoyance here is that *both* (a) and (b) then have that
> impact of "i_version no longer tracks di_changecount".

.... and what is annoying is that that the new i_version just a
glorified ctime change counter. What we should be fixing is ctime -
integrating this change counting into ctime would allow us to make
i_version go away entirely. i.e. We don't need a persistent ctime
change counter if the ctime has sufficient resolution or persistent
encoding that it does not need an external persistent change
counter.

That was reasoning behind the multi-grain timestamps. While the mgts
implementation was flawed, the reasoning behind it certainly isn't.
We should be trying to get rid of i_version by integrating it into
ctime updates, not arguing how atime vs i_version should work.

> So I don't think the issue here is "i_version" per se. I think in a
> vacuum, the best option of i_version is pretty obvious.  But if you
> want i_version to track di_changecount, *then* you end up with that
> situation where the persistence of atime matters, and i_version needs
> to update whenever a (persistent) atime update happens.

Yet I don't want i_version to track di_changecount.

I want to *stop supporting i_version altogether* in XFS.

I want i_version as filesystem internal metadata to die completely.

I don't want to change the on disk format to add a new i_version
field because we'll be straight back in this same siutation when the
next i_version bug is found and semantics get changed yet again.

Hence if we can encode the necessary change attributes into ctime,
we can drop VFS i_version support altogether.  Then the "atime bumps
i_version" problem also goes away because then we *don't use
i_version*.

But if we can't get the VFS to do this with ctime, at least we have
the abstractions available to us (i.e. timestamp granularity and
statx change cookie) to allow XFS to implement this sort of
ctime-with-integrated-change-counter internally to the filesystem
and be able to drop i_version support.... 

[....]

> This really is all *entirely* an artifact of that "bi_changecount" vs
> "i_version" being tied together. You did seem to imply that you'd be
> ok with having "bi_changecount" be split from i_version, ie from an
> earlier email in this thread:
> 
>  "Now that NFS is using a proper abstraction (i.e. vfs_statx()) to get
>   the change cookie, we really don't need to expose di_changecount in
>   i_version at all - we could simply copy an internal di_changecount
>   value into the statx cookie field in xfs_vn_getattr() and there
>   would be almost no change of behaviour from the perspective of NFS
>   and IMA at all"

.... which is what I was talking about here.

i.e. I was not talking about splitting i_version from di_changecount
- I was talking about being able to stop supporting the VFS
i_version counter entirely and still having NFS and IMA work
correctly.

Continually bring the argument back to "atime vs i_version" misses
the bigger issues around this new i_version definition and
implementation, and that the real solution should be to fix ctime
updates to make i_version at the VFS level go away forever.

-Dave.
Amir Goldstein Oct. 31, 2023, 7:03 a.m. UTC | #36
On Tue, Oct 31, 2023 at 3:42 AM Dave Chinner <david@fromorbit.com> wrote:
>
[...]
> .... and what is annoying is that that the new i_version just a
> glorified ctime change counter. What we should be fixing is ctime -
> integrating this change counting into ctime would allow us to make
> i_version go away entirely. i.e. We don't need a persistent ctime
> change counter if the ctime has sufficient resolution or persistent
> encoding that it does not need an external persistent change
> counter.
>
> That was reasoning behind the multi-grain timestamps. While the mgts
> implementation was flawed, the reasoning behind it certainly isn't.
> We should be trying to get rid of i_version by integrating it into
> ctime updates, not arguing how atime vs i_version should work.
>
> > So I don't think the issue here is "i_version" per se. I think in a
> > vacuum, the best option of i_version is pretty obvious.  But if you
> > want i_version to track di_changecount, *then* you end up with that
> > situation where the persistence of atime matters, and i_version needs
> > to update whenever a (persistent) atime update happens.
>
> Yet I don't want i_version to track di_changecount.
>
> I want to *stop supporting i_version altogether* in XFS.
>
> I want i_version as filesystem internal metadata to die completely.
>
> I don't want to change the on disk format to add a new i_version
> field because we'll be straight back in this same siutation when the
> next i_version bug is found and semantics get changed yet again.
>
> Hence if we can encode the necessary change attributes into ctime,
> we can drop VFS i_version support altogether.  Then the "atime bumps
> i_version" problem also goes away because then we *don't use
> i_version*.
>
> But if we can't get the VFS to do this with ctime, at least we have
> the abstractions available to us (i.e. timestamp granularity and
> statx change cookie) to allow XFS to implement this sort of
> ctime-with-integrated-change-counter internally to the filesystem
> and be able to drop i_version support....
>

I don't know if it was mentioned before in one of the many threads,
but there is another benefit of ctime-with-integrated-change-counter
approach - it is the ability to extend the solution with some adaptations
also to mtime.

The "change cookie" is used to know if inode metadata cache should
be invalidated and mtime is often used to know if data cache should
be invalidated, or if data comparison could be skipped (e.g. rsync).

The difference is that mtime can be set by user, so using lower nsec
bits for modification counter would require to truncate the user set
time granularity to 100ns - that is probably acceptable, but only as
an opt-in behavior.

The special value 0 for mtime-change-counter could be reserved for
mtime that was set by the user or for upgrade of existing inode,
where 0 counter means that mtime cannot be trusted as an accurate
data modification-cookie.

This feature is going to be useful for the vfs HSM implementation [1]
that I am working on and it actually rhymes with the XFS DMAPI
patches that were never fully merged upstream.

Speaking on behalf of my employer, we would love to see the data
modification-cookie feature implemented, whether in vfs or in xfs.

*IF* the result on this thread is that the chosen solution is
ctime-with-change-counter in XFS
*AND* if there is agreement among XFS developers to extend it with
an opt-in mkfs/mount option to 100ns-mtime-with-change-counter in XFS
*THEN* I think I will be able to allocate resources to drive this xfs work.

Thanks,
Amir.

[1] https://github.com/amir73il/fsnotify-utils/wiki/Hierarchical-Storage-Management-API
Christian Brauner Oct. 31, 2023, 10:26 a.m. UTC | #37
On Thu, Oct 19, 2023 at 07:28:48AM -0400, Jeff Layton wrote:
> On Thu, 2023-10-19 at 11:29 +0200, Christian Brauner wrote:
> > > Back to your earlier point though:
> > > 
> > > Is a global offset really a non-starter? I can see about doing something
> > > per-superblock, but ktime_get_mg_coarse_ts64 should be roughly as cheap
> > > as ktime_get_coarse_ts64. I don't see the downside there for the non-
> > > multigrain filesystems to call that.
> > 
> > I have to say that this doesn't excite me. This whole thing feels a bit
> > hackish. I think that a change version is the way more sane way to go.
> > 
> 
> What is it about this set that feels so much more hackish to you? Most
> of this set is pretty similar to what we had to revert. Is it just the
> timekeeper changes? Why do you feel those are a problem?

So I think that the multi-grain timestamp work was well intended but it
was ultimately a mistake. Because we added code that complicated
timestamp timestamp handling in the vfs to a point where the costs
clearly outweighed the benefits.

And I don't think that this direction is worth going into. This whole
thread ultimately boils down to complicating generic infrastructure
quite extensively for nfs to handle exposing xfs without forcing an
on-disk format change. That's even fine.

That's not a problem but in the same way I don't think the solution is
just stuffing this complexity into the vfs. IOW, if we make this a vfs
problem then at the lowest possible cost and not by changing how
timestamps work for everyone even if it's just internal.
Christian Brauner Oct. 31, 2023, 10:30 a.m. UTC | #38
On Tue, Oct 31, 2023 at 09:03:57AM +0200, Amir Goldstein wrote:
> On Tue, Oct 31, 2023 at 3:42 AM Dave Chinner <david@fromorbit.com> wrote:
> >
> [...]
> > .... and what is annoying is that that the new i_version just a
> > glorified ctime change counter. What we should be fixing is ctime -
> > integrating this change counting into ctime would allow us to make
> > i_version go away entirely. i.e. We don't need a persistent ctime
> > change counter if the ctime has sufficient resolution or persistent
> > encoding that it does not need an external persistent change
> > counter.
> >
> > That was reasoning behind the multi-grain timestamps. While the mgts
> > implementation was flawed, the reasoning behind it certainly isn't.
> > We should be trying to get rid of i_version by integrating it into
> > ctime updates, not arguing how atime vs i_version should work.
> >
> > > So I don't think the issue here is "i_version" per se. I think in a
> > > vacuum, the best option of i_version is pretty obvious.  But if you
> > > want i_version to track di_changecount, *then* you end up with that
> > > situation where the persistence of atime matters, and i_version needs
> > > to update whenever a (persistent) atime update happens.
> >
> > Yet I don't want i_version to track di_changecount.
> >
> > I want to *stop supporting i_version altogether* in XFS.
> >
> > I want i_version as filesystem internal metadata to die completely.
> >
> > I don't want to change the on disk format to add a new i_version
> > field because we'll be straight back in this same siutation when the
> > next i_version bug is found and semantics get changed yet again.
> >
> > Hence if we can encode the necessary change attributes into ctime,
> > we can drop VFS i_version support altogether.  Then the "atime bumps
> > i_version" problem also goes away because then we *don't use
> > i_version*.
> >
> > But if we can't get the VFS to do this with ctime, at least we have
> > the abstractions available to us (i.e. timestamp granularity and
> > statx change cookie) to allow XFS to implement this sort of
> > ctime-with-integrated-change-counter internally to the filesystem
> > and be able to drop i_version support....
> >
> 
> I don't know if it was mentioned before in one of the many threads,
> but there is another benefit of ctime-with-integrated-change-counter
> approach - it is the ability to extend the solution with some adaptations
> also to mtime.
> 
> The "change cookie" is used to know if inode metadata cache should
> be invalidated and mtime is often used to know if data cache should
> be invalidated, or if data comparison could be skipped (e.g. rsync).
> 
> The difference is that mtime can be set by user, so using lower nsec
> bits for modification counter would require to truncate the user set
> time granularity to 100ns - that is probably acceptable, but only as
> an opt-in behavior.
> 
> The special value 0 for mtime-change-counter could be reserved for
> mtime that was set by the user or for upgrade of existing inode,
> where 0 counter means that mtime cannot be trusted as an accurate
> data modification-cookie.
> 
> This feature is going to be useful for the vfs HSM implementation [1]
> that I am working on and it actually rhymes with the XFS DMAPI
> patches that were never fully merged upstream.
> 
> Speaking on behalf of my employer, we would love to see the data
> modification-cookie feature implemented, whether in vfs or in xfs.
> 
> *IF* the result on this thread is that the chosen solution is
> ctime-with-change-counter in XFS
> *AND* if there is agreement among XFS developers to extend it with
> an opt-in mkfs/mount option to 100ns-mtime-with-change-counter in XFS
> *THEN* I think I will be able to allocate resources to drive this xfs work.

If it can be solved within XFS then this would be preferable.
Jeffrey Layton Oct. 31, 2023, 11:04 a.m. UTC | #39
On Tue, 2023-10-31 at 09:37 +1100, Dave Chinner wrote:
> On Fri, Oct 27, 2023 at 06:35:58AM -0400, Jeff Layton wrote:
> > On Thu, 2023-10-26 at 13:20 +1100, Dave Chinner wrote:
> > > On Wed, Oct 25, 2023 at 08:25:35AM -0400, Jeff Layton wrote:
> > > > On Wed, 2023-10-25 at 19:05 +1100, Dave Chinner wrote:
> > > > > On Tue, Oct 24, 2023 at 02:40:06PM -0400, Jeff Layton wrote:
> > > > In earlier discussions you alluded to some repair and/or analysis tools
> > > > that depended on this counter.
> > > 
> > > Yes, and one of those "tools" is *me*.
> > > 
> > > I frequently look at the di_changecount when doing forensic and/or
> > > failure analysis on filesystem corpses.  SOE analysis, relative
> > > modification activity, etc all give insight into what happened to
> > > the filesystem to get it into the state it is currently in, and
> > > di_changecount provides information no other metadata in the inode
> > > contains.
> > > 
> > > > I took a quick look in xfsprogs, but I
> > > > didn't see anything there. Is there a library or something that these
> > > > tools use to get at this value?
> > > 
> > > xfs_db is the tool I use for this, such as:
> > > 
> > > $ sudo xfs_db -c "sb 0" -c "a rootino" -c "p v3.change_count" /dev/mapper/fast
> > > v3.change_count = 35
> > > $
> > > 
> > > The root inode in this filesystem has a change count of 35. The root
> > > inode has 32 dirents in it, which means that no entries have ever
> > > been removed or renamed. This sort of insight into the past history
> > > of inode metadata is largely impossible to get any other way, and
> > > it's been the difference between understanding failure and having no
> > > clue more than once.
> > > 
> > > Most block device parsing applications simply write their own
> > > decoder that walks the on-disk format. That's pretty trivial to do,
> > > developers can get all the information needed to do this from the
> > > on-disk format specification documentation we keep on kernel.org...
> > > 
> > 
> > Fair enough. I'm not here to tell you that you guys that you need to
> > change how di_changecount works. If it's too valuable to keep it
> > counting atime-only updates, then so be it.
> > 
> > If that's the case however, and given that the multigrain timestamp work
> > is effectively dead, then I don't see an alternative to growing the on-
> > disk inode. Do you?
> 
> Yes, I do see alternatives. That's what I've been trying
> (unsuccessfully) to describe and get consensus on. I feel like I'm
> being ignored and rail-roaded here, because nobody is even
> acknowledging that I'm proposing alternatives and keeps insisting
> that the only solution is a change of on-disk format.
> 
> So, I'll summarise the situation *yet again* in the hope that this
> time I won't get people arguing about atime vs i-version and what
> constitutes an on-disk format change because that goes nowhere and
> does nothing to determine which solution might be acceptible.
> 
> The basic situation is this:
> 
> If XFS can ignore relatime or lazytime persistent updates for given
> situations, then *we don't need to make periodic on-disk updates of
> atime*. This makes the whole problem of "persistent atime update bumps
> i_version" go away because then we *aren't making persistent atime
> updates* except when some other persistent modification that bumps
> [cm]time occurs.
> 
> But I don't want to do this unconditionally - for systems not
> running anything that samples i_version we want relatime/lazytime
> to behave as they are supposed to and do periodic persistent updates
> as per normal. Principle of least surprise and all that jazz.
> 
> So we really need an indication for inodes that we should enable this
> mode for the inode. I have asked if we can have per-operation
> context flag to trigger this given the needs for io_uring to have
> context flags for timestamp updates to be added. 
> 
> I have asked if we can have an inode flag set by the VFS or
> application code for this. e.g. a flag set by nfsd whenever it accesses a
> given inode.
> 
> I have asked if this inode flag can just be triggered if we ever see
> I_VERSION_QUERIED set or statx is used to retrieve a change cookie,
> and whether this is a reliable mechanism for setting such a flag.
> 

Ok, so to make sure I understand what you're proposing:

This would be a new inode flag that would be set in conjunction with
I_VERSION_QUERIED (but presumably is never cleared)? When XFS sees this
flag set, it would skip sending the atime to disk.

Given that you want to avoid on-disk changes, I assume this flag will
not be stored on disk. What happens after the NFS server reboots?

Consider:

1/ NFS server queries for the i_version and we set the
I_NO_ATIME_UPDATES_ON_DISK flag (or whatever) in conjunction with
I_VERSION_QUERIED. Some atime updates occur and the i_version isn't
bumped (as you'd expect).

2/ The server then reboots.

3/ Server comes back up, and some local task issues a read against the
inode. I_NO_ATIME_UPDATES_ON_DISK never had a chance to be set after the
reboot, so that atime update ends up incrementing the i_version counter.

4/ client cache invalidation occurs even though there was no write to
the file

This might reduce some of the spurious i_version bumps, but I don't see
how it can eliminate them entirely.

> I have suggested mechanisms for using masked off bits of timestamps
> to encode sub-timestamp granularity change counts and keep them
> invisible to userspace and then not using i_version at all for XFS.
> This avoids all the problems that the multi-grain timestamp
> infrastructure exposed due to variable granularity of user visible
> timestamps and ordering across inodes with different granularity.
> This is potentially a general solution, too.
> 

I don't really understand this at all, but trying to do anything with
fine-grained timestamps will just run into a lot of the same problems we
hit with the multigrain work. If you still see this as a path forward,
maybe you can describe it more detail?


> So, yeah, there are *lots* of ways we can solve this problem without
> needing to change on-disk formats.
>
Jeffrey Layton Oct. 31, 2023, 11:26 a.m. UTC | #40
On Tue, 2023-10-31 at 12:42 +1100, Dave Chinner wrote:
> On Mon, Oct 30, 2023 at 01:11:56PM -1000, Linus Torvalds wrote:
> > On Mon, 30 Oct 2023 at 12:37, Dave Chinner <david@fromorbit.com> wrote:
> > > 
> > > If XFS can ignore relatime or lazytime persistent updates for given
> > > situations, then *we don't need to make periodic on-disk updates of
> > > atime*. This makes the whole problem of "persistent atime update bumps
> > > i_version" go away because then we *aren't making persistent atime
> > > updates* except when some other persistent modification that bumps
> > > [cm]time occurs.
> > 
> > Well, I think this should be split into two independent questions:
> > 
> >  (a) are relatime or lazytime atime updates persistent if nothing else changes?
> 
> They only become persistent after 24 hours or, in the case of
> relatime, immediately persistent if mtime < atime (i.e. read after a
> modification). Those are the only times that the VFS triggers
> persistent writeback of atime, and it's the latter case (mtime <
> atime) that is the specific trigger that exposed the problem with
> atime bumping i_version in the first place.
> 
> >  (b) do atime updates _ever_ update i_version *regardless* of relatime
> > or lazytime?
> > 
> > and honestly, I think the best answer to (b) would be that "no,
> > i_version should simply not change for atime updates". And I think
> > that answer is what it is because no user of i_version seems to want
> > it.
> 
> As I keep repeating: Repeatedly stating that "atime should not bump
> i_version" does not address the questions I'm asking *at all*.
> 
> > Now, the reason it's a single question for you is that apparently for
> > XFS, the only thing that matters is "inode was written to disk" and
> > that "di_changecount" value is thus related to the persistence of
> > atime updates, but splitting di_changecount out to be a separate thing
> > from i_version seems to be on the table, so I think those two things
> > really could be independent issues.
> 
> Wrong way around - we'd have to split i_version out from
> di_changecount. It's i_version that has changed semantics, not
> di_changecount, and di_changecount behaviour must remain unchanged.
> 

I have to take issue with your characterization of this. The
requirements for NFS's change counter have not changed. Clearly there
was a breakdown in communications when it was first implemented in Linux
that caused atime updates to get counted in the i_version value, but
that was never intentional and never by design.

I'm simply trying to correct this historical mistake.

> What I really don't want to do is implement a new i_version field in
> the XFS on-disk format. What this redefinition of i_version
> semantics has made clear is that i_version is *user defined
> metadata*, not internal filesystem metadata that is defined by the
> filesystem on-disk format.
> 
> User defined persistent metadata *belongs in xattrs*, not in the
> core filesystem on-disk formats. If the VFS wants to define and
> manage i_version behaviour, smeantics and persistence independently
> of the filesystems that manage the persistent storage (as it clearly
> does!) then we should treat it just like any other VFS defined inode
> metadata (e.g. per inode objects like security constraints, ACLs,
> fsverity digests, fscrypt keys, etc). i.e. it should be in a named
> xattr, not directly implemented in the filesystem on-disk format
> deinfitions.
> 
> Then the application can change the meaning of the metadata whenever
> and however it likes. Then filesystem developers just don't need
> to care about it at all because the VFS specific persistent metadata
> is not part of the on-disk format we need to maintain
> cross-platform forwards and backwards compatibility for.
> 
> > > But I don't want to do this unconditionally - for systems not
> > > running anything that samples i_version we want relatime/lazytime
> > > to behave as they are supposed to and do periodic persistent updates
> > > as per normal. Principle of least surprise and all that jazz.
> > 
> > Well - see above: I think in a perfect world, we'd simply never change
> > i_version at all for any atime updates, and relatime/lazytime simply
> > wouldn't be an issue at all wrt i_version.
> 
> Right, that's what I'd like, especially as the new definition of
> i_version - "only change when [cm]time changes" - means that the VFS
> i_version is really now just a glorified timestamp.
> 
> > Wouldn't _that_ be the trule "least surprising" behavior? Considering
> > that nobody wants i_version to change for what are otherwise pure
> > reads (that's kind of the *definition* of atime, after all).
> 
> So, if you don't like the idea of us ignoring relatime/lazytime
> conditionally, are we allowed to simply ignore them *all the time*
> and do all timestamp updates in the manner that causes users the
> least amount of pain?
> 
> I mean, relatime only exists because atime updates cause users pain.
> lazytime only exists because relatime doesn't address the pain that
> timestamp updates cause mixed read/write or pure O_DSYNC overwrite
> workloads pain. noatime is a pain because it loses all atime
> updates.
> 
> There is no "one size is right for everyone", so why not just let
> filesystems do what is most efficient from an internal IO and
> persistence POV whilst still maintaining the majority of "expected"
> behaviours?
> 
> Keep in mind, though, that this is all moot if we can get rid of
> i_version entirely....
> 
> > Now, the annoyance here is that *both* (a) and (b) then have that
> > impact of "i_version no longer tracks di_changecount".
> 
> .... and what is annoying is that that the new i_version just a
> glorified ctime change counter. What we should be fixing is ctime -
> integrating this change counting into ctime would allow us to make
> i_version go away entirely. i.e. We don't need a persistent ctime
> change counter if the ctime has sufficient resolution or persistent
> encoding that it does not need an external persistent change
> counter.
> 
> That was reasoning behind the multi-grain timestamps. While the mgts
> implementation was flawed, the reasoning behind it certainly isn't.
> We should be trying to get rid of i_version by integrating it into
> ctime updates, not arguing how atime vs i_version should work.
> 
> > So I don't think the issue here is "i_version" per se. I think in a
> > vacuum, the best option of i_version is pretty obvious.  But if you
> > want i_version to track di_changecount, *then* you end up with that
> > situation where the persistence of atime matters, and i_version needs
> > to update whenever a (persistent) atime update happens.
> 
> Yet I don't want i_version to track di_changecount.
> 
> I want to *stop supporting i_version altogether* in XFS.
> 
> I want i_version as filesystem internal metadata to die completely.
> 
> I don't want to change the on disk format to add a new i_version
> field because we'll be straight back in this same siutation when the
> next i_version bug is found and semantics get changed yet again.
> 
> Hence if we can encode the necessary change attributes into ctime,
> we can drop VFS i_version support altogether.  Then the "atime bumps
> i_version" problem also goes away because then we *don't use
> i_version*.
> 
> But if we can't get the VFS to do this with ctime, at least we have
> the abstractions available to us (i.e. timestamp granularity and
> statx change cookie) to allow XFS to implement this sort of
> ctime-with-integrated-change-counter internally to the filesystem
> and be able to drop i_version support.... 
> 
> [....]

> > This really is all *entirely* an artifact of that "bi_changecount" vs
> > "i_version" being tied together. You did seem to imply that you'd be
> > ok with having "bi_changecount" be split from i_version, ie from an
> > earlier email in this thread:
> > 
> >  "Now that NFS is using a proper abstraction (i.e. vfs_statx()) to get
> >   the change cookie, we really don't need to expose di_changecount in
> >   i_version at all - we could simply copy an internal di_changecount
> >   value into the statx cookie field in xfs_vn_getattr() and there
> >   would be almost no change of behaviour from the perspective of NFS
> >   and IMA at all"
> 
> .... which is what I was talking about here.
> 
> i.e. I was not talking about splitting i_version from di_changecount
> - I was talking about being able to stop supporting the VFS
> i_version counter entirely and still having NFS and IMA work
> correctly.
> 
> Continually bring the argument back to "atime vs i_version" misses
> the bigger issues around this new i_version definition and
> implementation, and that the real solution should be to fix ctime
> updates to make i_version at the VFS level go away forever.
> 

Ok, so your proposal is to keep using coarse grained timestamps, but use
the "insignificant" bits in them to store a change counter?

That sounds complex and fraught with peril, but I'm willing to listen to
some specifics about how that would work.
Jeffrey Layton Oct. 31, 2023, 11:29 a.m. UTC | #41
On Tue, 2023-10-31 at 09:03 +0200, Amir Goldstein wrote:
> On Tue, Oct 31, 2023 at 3:42 AM Dave Chinner <david@fromorbit.com> wrote:
> > 
> [...]
> > .... and what is annoying is that that the new i_version just a
> > glorified ctime change counter. What we should be fixing is ctime -
> > integrating this change counting into ctime would allow us to make
> > i_version go away entirely. i.e. We don't need a persistent ctime
> > change counter if the ctime has sufficient resolution or persistent
> > encoding that it does not need an external persistent change
> > counter.
> > 
> > That was reasoning behind the multi-grain timestamps. While the mgts
> > implementation was flawed, the reasoning behind it certainly isn't.
> > We should be trying to get rid of i_version by integrating it into
> > ctime updates, not arguing how atime vs i_version should work.
> > 
> > > So I don't think the issue here is "i_version" per se. I think in a
> > > vacuum, the best option of i_version is pretty obvious.  But if you
> > > want i_version to track di_changecount, *then* you end up with that
> > > situation where the persistence of atime matters, and i_version needs
> > > to update whenever a (persistent) atime update happens.
> > 
> > Yet I don't want i_version to track di_changecount.
> > 
> > I want to *stop supporting i_version altogether* in XFS.
> > 
> > I want i_version as filesystem internal metadata to die completely.
> > 
> > I don't want to change the on disk format to add a new i_version
> > field because we'll be straight back in this same siutation when the
> > next i_version bug is found and semantics get changed yet again.
> > 
> > Hence if we can encode the necessary change attributes into ctime,
> > we can drop VFS i_version support altogether.  Then the "atime bumps
> > i_version" problem also goes away because then we *don't use
> > i_version*.
> > 
> > But if we can't get the VFS to do this with ctime, at least we have
> > the abstractions available to us (i.e. timestamp granularity and
> > statx change cookie) to allow XFS to implement this sort of
> > ctime-with-integrated-change-counter internally to the filesystem
> > and be able to drop i_version support....
> > 
> 
> I don't know if it was mentioned before in one of the many threads,
> but there is another benefit of ctime-with-integrated-change-counter
> approach - it is the ability to extend the solution with some adaptations
> also to mtime.
> 
> The "change cookie" is used to know if inode metadata cache should
> be invalidated and mtime is often used to know if data cache should
> be invalidated, or if data comparison could be skipped (e.g. rsync).
> 
> The difference is that mtime can be set by user, so using lower nsec
> bits for modification counter would require to truncate the user set
> time granularity to 100ns - that is probably acceptable, but only as
> an opt-in behavior.
> 
> The special value 0 for mtime-change-counter could be reserved for
> mtime that was set by the user or for upgrade of existing inode,
> where 0 counter means that mtime cannot be trusted as an accurate
> data modification-cookie.
> 
> This feature is going to be useful for the vfs HSM implementation [1]
> that I am working on and it actually rhymes with the XFS DMAPI
> patches that were never fully merged upstream.
> 
> Speaking on behalf of my employer, we would love to see the data
> modification-cookie feature implemented, whether in vfs or in xfs.
> 
> *IF* the result on this thread is that the chosen solution is
> ctime-with-change-counter in XFS
> *AND* if there is agreement among XFS developers to extend it with
> an opt-in mkfs/mount option to 100ns-mtime-with-change-counter in XFS
> *THEN* I think I will be able to allocate resources to drive this xfs work.
> 
> Thanks,
> Amir.
> 
> [1] https://github.com/amir73il/fsnotify-utils/wiki/Hierarchical-Storage-Management-API

I agree with Christian that if you can do this inside of XFS altogether,
then that's a preferable solution. I don't quite understand how ctime-
with-change-counter is intended to work however, so I'll have to reserve
judgement there.
Jan Kara Oct. 31, 2023, 12:22 p.m. UTC | #42
On Tue 31-10-23 07:04:53, Jeff Layton wrote:
> On Tue, 2023-10-31 at 09:37 +1100, Dave Chinner wrote:
> > I have suggested mechanisms for using masked off bits of timestamps
> > to encode sub-timestamp granularity change counts and keep them
> > invisible to userspace and then not using i_version at all for XFS.
> > This avoids all the problems that the multi-grain timestamp
> > infrastructure exposed due to variable granularity of user visible
> > timestamps and ordering across inodes with different granularity.
> > This is potentially a general solution, too.
> 
> I don't really understand this at all, but trying to do anything with
> fine-grained timestamps will just run into a lot of the same problems we
> hit with the multigrain work. If you still see this as a path forward,
> maybe you can describe it more detail?

Dave explained a bit more details here [1] like:

Another options is for XFS to play it's own internal tricks with
[cm]time granularity and turn off i_version. e.g. limit external
timestamp visibility to 1us and use the remaining dozen bits of the
ns field to hold a change counter for updates within a single coarse
timer tick. This guarantees the timestamp changes within a coarse
tick for the purposes of change detection, but we don't expose those
bits to applications so applications that compare timestamps across
inodes won't get things back to front like was happening with the
multi-grain timestamps....
-

So as far as I understand Dave wants to effectively persist counter in low
bits of ctime and expose ctime+counter as its change cookie. I guess that
could work and what makes the complexity manageable compared to full
multigrain timestamps is the fact that we have one filesystem, one on-disk
format etc. The only slight trouble could be that if we previously handed
out something in low bits of ctime for XFS, we need to keep handing the
same thing out until the inode changes (i.e., no rounding until the moment
inode changes) as the old timestamp could be stored somewhere externally
and compared.

								Honza

[1] https://lore.kernel.org/all/ZTjMRRqmlJ+fTys2@dread.disaster.area/
Jeffrey Layton Oct. 31, 2023, 12:55 p.m. UTC | #43
On Tue, 2023-10-31 at 13:22 +0100, Jan Kara wrote:
> On Tue 31-10-23 07:04:53, Jeff Layton wrote:
> > On Tue, 2023-10-31 at 09:37 +1100, Dave Chinner wrote:
> > > I have suggested mechanisms for using masked off bits of timestamps
> > > to encode sub-timestamp granularity change counts and keep them
> > > invisible to userspace and then not using i_version at all for XFS.
> > > This avoids all the problems that the multi-grain timestamp
> > > infrastructure exposed due to variable granularity of user visible
> > > timestamps and ordering across inodes with different granularity.
> > > This is potentially a general solution, too.
> > 
> > I don't really understand this at all, but trying to do anything with
> > fine-grained timestamps will just run into a lot of the same problems we
> > hit with the multigrain work. If you still see this as a path forward,
> > maybe you can describe it more detail?
> 
> Dave explained a bit more details here [1] like:
> 
> Another options is for XFS to play it's own internal tricks with
> [cm]time granularity and turn off i_version. e.g. limit external
> timestamp visibility to 1us and use the remaining dozen bits of the
> ns field to hold a change counter for updates within a single coarse
> timer tick. This guarantees the timestamp changes within a coarse
> tick for the purposes of change detection, but we don't expose those
> bits to applications so applications that compare timestamps across
> inodes won't get things back to front like was happening with the
> multi-grain timestamps....
> -
> 
> So as far as I understand Dave wants to effectively persist counter in low
> bits of ctime and expose ctime+counter as its change cookie. I guess that
> could work and what makes the complexity manageable compared to full
> multigrain timestamps is the fact that we have one filesystem, one on-disk
> format etc. The only slight trouble could be that if we previously handed
> out something in low bits of ctime for XFS, we need to keep handing the
> same thing out until the inode changes (i.e., no rounding until the moment
> inode changes) as the old timestamp could be stored somewhere externally
> and compared.
> 
> 								Honza
> 
> [1] https://lore.kernel.org/all/ZTjMRRqmlJ+fTys2@dread.disaster.area/
> 
> 

Got it. That makes sense and could probably be made to work.
Doing that all in XFS won't be simple though. You'll need to reimplement
stuff like file_modified() and file_update_time(). Those get called from
deep within the VFS and from page fault handlers.

FWIW, that's the main reason the multigrain work was so invasive, even
though it was a filesystem-specific feature.
Jeffrey Layton Oct. 31, 2023, 1:55 p.m. UTC | #44
On Tue, 2023-10-31 at 11:26 +0100, Christian Brauner wrote:
> On Thu, Oct 19, 2023 at 07:28:48AM -0400, Jeff Layton wrote:
> > On Thu, 2023-10-19 at 11:29 +0200, Christian Brauner wrote:
> > > > Back to your earlier point though:
> > > > 
> > > > Is a global offset really a non-starter? I can see about doing something
> > > > per-superblock, but ktime_get_mg_coarse_ts64 should be roughly as cheap
> > > > as ktime_get_coarse_ts64. I don't see the downside there for the non-
> > > > multigrain filesystems to call that.
> > > 
> > > I have to say that this doesn't excite me. This whole thing feels a bit
> > > hackish. I think that a change version is the way more sane way to go.
> > > 
> > 
> > What is it about this set that feels so much more hackish to you? Most
> > of this set is pretty similar to what we had to revert. Is it just the
> > timekeeper changes? Why do you feel those are a problem?
> 
> So I think that the multi-grain timestamp work was well intended but it
> was ultimately a mistake. Because we added code that complicated
> timestamp timestamp handling in the vfs to a point where the costs
> clearly outweighed the benefits.
> 
> And I don't think that this direction is worth going into. This whole
> thread ultimately boils down to complicating generic infrastructure
> quite extensively for nfs to handle exposing xfs without forcing an
> on-disk format change. That's even fine.
> 
> That's not a problem but in the same way I don't think the solution is
> just stuffing this complexity into the vfs. IOW, if we make this a vfs
> problem then at the lowest possible cost and not by changing how
> timestamps work for everyone even if it's just internal.

I'll point out that this last posting I did was an RFC. It was invasive
to the timekeeping code, but I don't think that's a hard requirement for
doing this.

I do appreciate the feedback on this version of the series (particularly
from Thomas who gave a great technical reason why this approach was
wrong), but I don't think we necessarily have to give up on the whole
idea because this particular implementation was too costly.

The core idea for fixing the problem with the original series is sane,
IMO. There is nothing wrong with simply making it that when we stamp a
file with a fine-grained timestamp that we consider that a floor for all
later timestamp updates. The only real question is how to keep that
(global) fine-grained floor offset at a low cost. I think that's a
solvable problem.

I also believe that real, measurable fine-grained timestamp differences
are worthwhile for other use cases beyond NFS. Everyone was pointing out
the problems with lagging timestamps vs. make and rsync, but that's a
double-edged sword. With the current always coarse-grained timestamps,
the ordering of files written within the same jiffy can't be determined
since their timestamps will be identical. We could conceivably change
that with this series.

That said, if this has no chance of ever being merged, then I won't
bother working on it further, and we can try to pursue something that is
(maybe) XFS-specific.

Let me know, either way.
--
Jeff Layton <jlayton@kernel.org>
John Stoffel Oct. 31, 2023, 7:43 p.m. UTC | #45
>>>>> "Jeff" == Jeff Layton <jlayton@kernel.org> writes:

> On Tue, 2023-10-31 at 12:42 +1100, Dave Chinner wrote:
>> On Mon, Oct 30, 2023 at 01:11:56PM -1000, Linus Torvalds wrote:
>> > On Mon, 30 Oct 2023 at 12:37, Dave Chinner <david@fromorbit.com> wrote:
>> > > 
>> > > If XFS can ignore relatime or lazytime persistent updates for given
>> > > situations, then *we don't need to make periodic on-disk updates of
>> > > atime*. This makes the whole problem of "persistent atime update bumps
>> > > i_version" go away because then we *aren't making persistent atime
>> > > updates* except when some other persistent modification that bumps
>> > > [cm]time occurs.
>> > 
>> > Well, I think this should be split into two independent questions:
>> > 
>> >  (a) are relatime or lazytime atime updates persistent if nothing else changes?
>> 
>> They only become persistent after 24 hours or, in the case of
>> relatime, immediately persistent if mtime < atime (i.e. read after a
>> modification). Those are the only times that the VFS triggers
>> persistent writeback of atime, and it's the latter case (mtime <
>> atime) that is the specific trigger that exposed the problem with
>> atime bumping i_version in the first place.
>> 
>> >  (b) do atime updates _ever_ update i_version *regardless* of relatime
>> > or lazytime?
>> > 
>> > and honestly, I think the best answer to (b) would be that "no,
>> > i_version should simply not change for atime updates". And I think
>> > that answer is what it is because no user of i_version seems to want
>> > it.
>> 
>> As I keep repeating: Repeatedly stating that "atime should not bump
>> i_version" does not address the questions I'm asking *at all*.
>> 
>> > Now, the reason it's a single question for you is that apparently for
>> > XFS, the only thing that matters is "inode was written to disk" and
>> > that "di_changecount" value is thus related to the persistence of
>> > atime updates, but splitting di_changecount out to be a separate thing
>> > from i_version seems to be on the table, so I think those two things
>> > really could be independent issues.
>> 
>> Wrong way around - we'd have to split i_version out from
>> di_changecount. It's i_version that has changed semantics, not
>> di_changecount, and di_changecount behaviour must remain unchanged.
>> 

> I have to take issue with your characterization of this. The
> requirements for NFS's change counter have not changed. Clearly there
> was a breakdown in communications when it was first implemented in Linux
> that caused atime updates to get counted in the i_version value, but
> that was never intentional and never by design.

This has been bugging me, but all the references to NFS really mean
NFSv4.1 or newer, correct?  I can't see how any of this affects NFSv3
at all, and that's probably the still dominant form of NFS, right?  

John
Dave Chinner Oct. 31, 2023, 9:57 p.m. UTC | #46
On Tue, Oct 31, 2023 at 07:29:18AM -0400, Jeff Layton wrote:
> On Tue, 2023-10-31 at 09:03 +0200, Amir Goldstein wrote:
> > On Tue, Oct 31, 2023 at 3:42 AM Dave Chinner <david@fromorbit.com> wrote:
> > > 
> > [...]
> > > .... and what is annoying is that that the new i_version just a
> > > glorified ctime change counter. What we should be fixing is ctime -
> > > integrating this change counting into ctime would allow us to make
> > > i_version go away entirely. i.e. We don't need a persistent ctime
> > > change counter if the ctime has sufficient resolution or persistent
> > > encoding that it does not need an external persistent change
> > > counter.
> > > 
> > > That was reasoning behind the multi-grain timestamps. While the mgts
> > > implementation was flawed, the reasoning behind it certainly isn't.
> > > We should be trying to get rid of i_version by integrating it into
> > > ctime updates, not arguing how atime vs i_version should work.
> > > 
> > > > So I don't think the issue here is "i_version" per se. I think in a
> > > > vacuum, the best option of i_version is pretty obvious.  But if you
> > > > want i_version to track di_changecount, *then* you end up with that
> > > > situation where the persistence of atime matters, and i_version needs
> > > > to update whenever a (persistent) atime update happens.
> > > 
> > > Yet I don't want i_version to track di_changecount.
> > > 
> > > I want to *stop supporting i_version altogether* in XFS.
> > > 
> > > I want i_version as filesystem internal metadata to die completely.
> > > 
> > > I don't want to change the on disk format to add a new i_version
> > > field because we'll be straight back in this same siutation when the
> > > next i_version bug is found and semantics get changed yet again.
> > > 
> > > Hence if we can encode the necessary change attributes into ctime,
> > > we can drop VFS i_version support altogether.  Then the "atime bumps
> > > i_version" problem also goes away because then we *don't use
> > > i_version*.
> > > 
> > > But if we can't get the VFS to do this with ctime, at least we have
> > > the abstractions available to us (i.e. timestamp granularity and
> > > statx change cookie) to allow XFS to implement this sort of
> > > ctime-with-integrated-change-counter internally to the filesystem
> > > and be able to drop i_version support....
> > > 
> > 
> > I don't know if it was mentioned before in one of the many threads,
> > but there is another benefit of ctime-with-integrated-change-counter
> > approach - it is the ability to extend the solution with some adaptations
> > also to mtime.
> > 
> > The "change cookie" is used to know if inode metadata cache should
> > be invalidated and mtime is often used to know if data cache should
> > be invalidated, or if data comparison could be skipped (e.g. rsync).
> > 
> > The difference is that mtime can be set by user, so using lower nsec
> > bits for modification counter would require to truncate the user set
> > time granularity to 100ns - that is probably acceptable, but only as
> > an opt-in behavior.
> > 
> > The special value 0 for mtime-change-counter could be reserved for
> > mtime that was set by the user or for upgrade of existing inode,
> > where 0 counter means that mtime cannot be trusted as an accurate
> > data modification-cookie.
> > 
> > This feature is going to be useful for the vfs HSM implementation [1]
> > that I am working on and it actually rhymes with the XFS DMAPI
> > patches that were never fully merged upstream.
> > 
> > Speaking on behalf of my employer, we would love to see the data
> > modification-cookie feature implemented, whether in vfs or in xfs.
> > 
> > *IF* the result on this thread is that the chosen solution is
> > ctime-with-change-counter in XFS
> > *AND* if there is agreement among XFS developers to extend it with
> > an opt-in mkfs/mount option to 100ns-mtime-with-change-counter in XFS
> > *THEN* I think I will be able to allocate resources to drive this xfs work.
> > 
> > Thanks,
> > Amir.
> > 
> > [1] https://github.com/amir73il/fsnotify-utils/wiki/Hierarchical-Storage-Management-API
> 
> I agree with Christian that if you can do this inside of XFS altogether,
> then that's a preferable solution. I don't quite understand how ctime-
> with-change-counter is intended to work however, so I'll have to reserve
> judgement there.

Yeah, this is pretty straight forward to do in XFS, I think. We were
talking about it on #xfs yesterday afternoon, and it all we really
need to do is this:

1. remove SB_I_VERSION from the superblock.

	Now nothing external to the filesystem will access
	inode->i_version  - it's value is undefined for external
	access and cannot be used for any purpose except filesystem
	internal information.

2. Set sb->s_time_gran to a power of 2 between 2^10 and 2^20 (TBD).

	This now gives the timestamp granularity that we expose to
	the VFS and userspace of 1us to 1ms. It gives us the lower
	10-20 bits for the persistent timestamp change counter to be
	encoded into for on-disk storage and change attribute
	cookies.

	Using the s_time_gran field like this with an external
	change counter means nothing in the VFS is aware that we
	store more information in the timestamp fields on disk.
	Everything should just work as it currently does (e.g.
	timestamp comparisons).

	It also means userspace never sees this internal change
	counter and so all the problems with incremental changes to
	timestamps within a timer tick across multiple files (i.e.
	the mgts problem) don't occur with this setup.

3. Every time the timestamp is touched and the timestamp doesn't
change, we bump inode->i_version. If the timestamp changes, then we
zero inode->i_version.

	THis gives us a per-coarse-timer-tick change counter. When
	the timestamp itself changes, we reset the change counter
	because the high bits in the timestamp have changed and that
	prevents the change coutner from wrapping arbitrarily and
	any on-disk timestamp from going backwards due to change
	counter overflow.

4. When statx asks for the change attribute, we copy the timestamp
into it and fold in the current change counter value.

	This provides the uniqueness that is required for the change
	counter.

5. When-ever the inode is persisted, the timestamp is copied to the
on-disk structure and the current change counter is folded in.

	This means the on-disk structure always contains the latest
	change attribute that has been persisted, just like we
	currently do with i_version now.

6. When-ever we read the inode off disk, we split the change counter
from the timestamp and update the appropriate internal structures
with this information.

	This ensures that the VFS and userspace never see the change
	counter implementation in the inode timestamps.

7. We need to be able to override inode_needs_update_time()
behaviour, as it will skip timestamp updates if timestamps are
equal.

	When timestamps are equal in this case, we need to bump the
	change counter rather than "do nothing". Hence we need
	different logic here, or a new method to be provided so the
	filesystem can decide how/when timestamps get updated.


This last piece of the puzzle is the problematic one.

Ideally, I'd like to see the ->update_time() code path be inverted.
Instead of only being called at the bottom once the VFS has decided
what timestamps need to be changed, it gets called directly from the
high level timestamp modification code and uses VFS helpers to
determine if updates are needed. 

e.g. the current code flow for an atime update is:

touch_atime()
  atime_needs_update()
  <freeze/write protection>
  inode_update_time(S_ATIME)
    ->update_time(S_ATIME)
      <filesystem atime update>

I'd much prefer this to be:

touch_atime()
  if (->update_time(S_ATIME)) {
    ->update_time(S_ATIME)
      xfs_inode_update_time(S_ATIME)
        if (atime_needs_update())
	  <filesystem atime update>
  } else {
    /* run the existing code */
  }

Similarly we'd turn file_modified()/file_update_time() inside out,
and this then allows the filesystem to add custom timestamp update
checks alongside the VFS timestamp update checks.

It would also enable us to untangle the mess that is lazytime, where
we have to implement ->update_time to catch lazytime updates and
punt them back to generic_update_time(), which then has to check for
lazytime again to determine how to dirty and queue the inode.
Of course, generic_update_time() also does timespec_equal checks on
timestamps to determine if times should be updated, and so we'd
probably need overrides on that, too.

Sorting the lazytime mess for internal change counters really needs
for all the timestamp updates to be handled in the filesystem, not
bounced back and forward between the filesystem and VFS helpers as
it currently is, hence I think we need to rework ->update_time to
make this all work cleanly.

-Dave.
Darrick J. Wong Oct. 31, 2023, 11:02 p.m. UTC | #47
On Wed, Nov 01, 2023 at 08:57:09AM +1100, Dave Chinner wrote:
> On Tue, Oct 31, 2023 at 07:29:18AM -0400, Jeff Layton wrote:
> > On Tue, 2023-10-31 at 09:03 +0200, Amir Goldstein wrote:
> > > On Tue, Oct 31, 2023 at 3:42 AM Dave Chinner <david@fromorbit.com> wrote:
> > > > 
> > > [...]
> > > > .... and what is annoying is that that the new i_version just a
> > > > glorified ctime change counter. What we should be fixing is ctime -
> > > > integrating this change counting into ctime would allow us to make
> > > > i_version go away entirely. i.e. We don't need a persistent ctime
> > > > change counter if the ctime has sufficient resolution or persistent
> > > > encoding that it does not need an external persistent change
> > > > counter.
> > > > 
> > > > That was reasoning behind the multi-grain timestamps. While the mgts
> > > > implementation was flawed, the reasoning behind it certainly isn't.
> > > > We should be trying to get rid of i_version by integrating it into
> > > > ctime updates, not arguing how atime vs i_version should work.
> > > > 
> > > > > So I don't think the issue here is "i_version" per se. I think in a
> > > > > vacuum, the best option of i_version is pretty obvious.  But if you
> > > > > want i_version to track di_changecount, *then* you end up with that
> > > > > situation where the persistence of atime matters, and i_version needs
> > > > > to update whenever a (persistent) atime update happens.
> > > > 
> > > > Yet I don't want i_version to track di_changecount.
> > > > 
> > > > I want to *stop supporting i_version altogether* in XFS.
> > > > 
> > > > I want i_version as filesystem internal metadata to die completely.
> > > > 
> > > > I don't want to change the on disk format to add a new i_version
> > > > field because we'll be straight back in this same siutation when the
> > > > next i_version bug is found and semantics get changed yet again.
> > > > 
> > > > Hence if we can encode the necessary change attributes into ctime,
> > > > we can drop VFS i_version support altogether.  Then the "atime bumps
> > > > i_version" problem also goes away because then we *don't use
> > > > i_version*.
> > > > 
> > > > But if we can't get the VFS to do this with ctime, at least we have
> > > > the abstractions available to us (i.e. timestamp granularity and
> > > > statx change cookie) to allow XFS to implement this sort of
> > > > ctime-with-integrated-change-counter internally to the filesystem
> > > > and be able to drop i_version support....
> > > > 
> > > 
> > > I don't know if it was mentioned before in one of the many threads,
> > > but there is another benefit of ctime-with-integrated-change-counter
> > > approach - it is the ability to extend the solution with some adaptations
> > > also to mtime.
> > > 
> > > The "change cookie" is used to know if inode metadata cache should
> > > be invalidated and mtime is often used to know if data cache should
> > > be invalidated, or if data comparison could be skipped (e.g. rsync).
> > > 
> > > The difference is that mtime can be set by user, so using lower nsec
> > > bits for modification counter would require to truncate the user set
> > > time granularity to 100ns - that is probably acceptable, but only as
> > > an opt-in behavior.
> > > 
> > > The special value 0 for mtime-change-counter could be reserved for
> > > mtime that was set by the user or for upgrade of existing inode,
> > > where 0 counter means that mtime cannot be trusted as an accurate
> > > data modification-cookie.
> > > 
> > > This feature is going to be useful for the vfs HSM implementation [1]
> > > that I am working on and it actually rhymes with the XFS DMAPI
> > > patches that were never fully merged upstream.
> > > 
> > > Speaking on behalf of my employer, we would love to see the data
> > > modification-cookie feature implemented, whether in vfs or in xfs.
> > > 
> > > *IF* the result on this thread is that the chosen solution is
> > > ctime-with-change-counter in XFS
> > > *AND* if there is agreement among XFS developers to extend it with
> > > an opt-in mkfs/mount option to 100ns-mtime-with-change-counter in XFS
> > > *THEN* I think I will be able to allocate resources to drive this xfs work.
> > > 
> > > Thanks,
> > > Amir.
> > > 
> > > [1] https://github.com/amir73il/fsnotify-utils/wiki/Hierarchical-Storage-Management-API
> > 
> > I agree with Christian that if you can do this inside of XFS altogether,
> > then that's a preferable solution. I don't quite understand how ctime-
> > with-change-counter is intended to work however, so I'll have to reserve
> > judgement there.
> 
> Yeah, this is pretty straight forward to do in XFS, I think. We were
> talking about it on #xfs yesterday afternoon, and it all we really
> need to do is this:
> 
> 1. remove SB_I_VERSION from the superblock.
> 
> 	Now nothing external to the filesystem will access
> 	inode->i_version  - it's value is undefined for external
> 	access and cannot be used for any purpose except filesystem
> 	internal information.
> 
> 2. Set sb->s_time_gran to a power of 2 between 2^10 and 2^20 (TBD).
> 
> 	This now gives the timestamp granularity that we expose to
> 	the VFS and userspace of 1us to 1ms. It gives us the lower
> 	10-20 bits for the persistent timestamp change counter to be
> 	encoded into for on-disk storage and change attribute
> 	cookies.
> 
> 	Using the s_time_gran field like this with an external
> 	change counter means nothing in the VFS is aware that we
> 	store more information in the timestamp fields on disk.
> 	Everything should just work as it currently does (e.g.
> 	timestamp comparisons).
> 
> 	It also means userspace never sees this internal change
> 	counter and so all the problems with incremental changes to
> 	timestamps within a timer tick across multiple files (i.e.
> 	the mgts problem) don't occur with this setup.
> 
> 3. Every time the timestamp is touched and the timestamp doesn't
> change, we bump inode->i_version. If the timestamp changes, then we
> zero inode->i_version.
> 
> 	THis gives us a per-coarse-timer-tick change counter. When
> 	the timestamp itself changes, we reset the change counter
> 	because the high bits in the timestamp have changed and that
> 	prevents the change coutner from wrapping arbitrarily and
> 	any on-disk timestamp from going backwards due to change
> 	counter overflow.
> 
> 4. When statx asks for the change attribute, we copy the timestamp
> into it and fold in the current change counter value.
> 
> 	This provides the uniqueness that is required for the change
> 	counter.
> 
> 5. When-ever the inode is persisted, the timestamp is copied to the
> on-disk structure and the current change counter is folded in.
> 
> 	This means the on-disk structure always contains the latest
> 	change attribute that has been persisted, just like we
> 	currently do with i_version now.
> 
> 6. When-ever we read the inode off disk, we split the change counter
> from the timestamp and update the appropriate internal structures
> with this information.
> 
> 	This ensures that the VFS and userspace never see the change
> 	counter implementation in the inode timestamps.
> 
> 7. We need to be able to override inode_needs_update_time()
> behaviour, as it will skip timestamp updates if timestamps are
> equal.
> 
> 	When timestamps are equal in this case, we need to bump the
> 	change counter rather than "do nothing". Hence we need
> 	different logic here, or a new method to be provided so the
> 	filesystem can decide how/when timestamps get updated.
> 
> 
> This last piece of the puzzle is the problematic one.
> 
> Ideally, I'd like to see the ->update_time() code path be inverted.
> Instead of only being called at the bottom once the VFS has decided
> what timestamps need to be changed, it gets called directly from the
> high level timestamp modification code and uses VFS helpers to
> determine if updates are needed.

I like the approach of hiding an i_version counter in the unused bits of
inode.i_ctime.tv_nsec.  This would work for /any/ filesystem that can
handle nanosecond-precision timestamps without the need for the explicit
I_VERSION ondisk counter.

> e.g. the current code flow for an atime update is:
> 
> touch_atime()
>   atime_needs_update()
>   <freeze/write protection>
>   inode_update_time(S_ATIME)
>     ->update_time(S_ATIME)
>       <filesystem atime update>
> 
> I'd much prefer this to be:
> 
> touch_atime()
>   if (->update_time(S_ATIME)) {
>     ->update_time(S_ATIME)
>       xfs_inode_update_time(S_ATIME)
>         if (atime_needs_update())
> 	  <filesystem atime update>
>   } else {
>     /* run the existing code */
>   }
> 
> Similarly we'd turn file_modified()/file_update_time() inside out,
> and this then allows the filesystem to add custom timestamp update
> checks alongside the VFS timestamp update checks.
> 
> It would also enable us to untangle the mess that is lazytime, where
> we have to implement ->update_time to catch lazytime updates and
> punt them back to generic_update_time(), which then has to check for
> lazytime again to determine how to dirty and queue the inode.
> Of course, generic_update_time() also does timespec_equal checks on
> timestamps to determine if times should be updated, and so we'd
> probably need overrides on that, too.

Hmm.  So would the VFS update the incore timestamps of struct inode in
whatever manner it wants?  Could that include incrementing the lower
bits of i_ctime.tv_nsec for filesystems that advertise a non-1nsec
granularity but also set a flag that effectively means "but you can use
the lower tv_nsec bits if you want"?

And perhaps after all that, the VFS should decide if a timestamp update
needs to be persisted (e.g. lazytime/nodiratime/poniesatime) and if so,
call ->update_time or __mark_inode_dirty?  Then XFS doesn't have to know
about all the timestamp persistence rules, it just has to follow
whatever the VFS tells it.

> Sorting the lazytime mess for internal change counters really needs
> for all the timestamp updates to be handled in the filesystem, not
> bounced back and forward between the filesystem and VFS helpers as
> it currently is, hence I think we need to rework ->update_time to
> make this all work cleanly.

(Oh, I guess I proposed sort of the opposite of what you just said.)

> 
> -Dave.
> 
> -- 
> Dave Chinner
> david@fromorbit.com
Darrick J. Wong Oct. 31, 2023, 11:12 p.m. UTC | #48
On Tue, Oct 31, 2023 at 09:03:57AM +0200, Amir Goldstein wrote:
> On Tue, Oct 31, 2023 at 3:42 AM Dave Chinner <david@fromorbit.com> wrote:
> >
> [...]
> > .... and what is annoying is that that the new i_version just a
> > glorified ctime change counter. What we should be fixing is ctime -
> > integrating this change counting into ctime would allow us to make
> > i_version go away entirely. i.e. We don't need a persistent ctime
> > change counter if the ctime has sufficient resolution or persistent
> > encoding that it does not need an external persistent change
> > counter.
> >
> > That was reasoning behind the multi-grain timestamps. While the mgts
> > implementation was flawed, the reasoning behind it certainly isn't.
> > We should be trying to get rid of i_version by integrating it into
> > ctime updates, not arguing how atime vs i_version should work.
> >
> > > So I don't think the issue here is "i_version" per se. I think in a
> > > vacuum, the best option of i_version is pretty obvious.  But if you
> > > want i_version to track di_changecount, *then* you end up with that
> > > situation where the persistence of atime matters, and i_version needs
> > > to update whenever a (persistent) atime update happens.
> >
> > Yet I don't want i_version to track di_changecount.
> >
> > I want to *stop supporting i_version altogether* in XFS.
> >
> > I want i_version as filesystem internal metadata to die completely.
> >
> > I don't want to change the on disk format to add a new i_version
> > field because we'll be straight back in this same siutation when the
> > next i_version bug is found and semantics get changed yet again.
> >
> > Hence if we can encode the necessary change attributes into ctime,
> > we can drop VFS i_version support altogether.  Then the "atime bumps
> > i_version" problem also goes away because then we *don't use
> > i_version*.
> >
> > But if we can't get the VFS to do this with ctime, at least we have
> > the abstractions available to us (i.e. timestamp granularity and
> > statx change cookie) to allow XFS to implement this sort of
> > ctime-with-integrated-change-counter internally to the filesystem
> > and be able to drop i_version support....
> >
> 
> I don't know if it was mentioned before in one of the many threads,
> but there is another benefit of ctime-with-integrated-change-counter
> approach - it is the ability to extend the solution with some adaptations
> also to mtime.
> 
> The "change cookie" is used to know if inode metadata cache should
> be invalidated and mtime is often used to know if data cache should
> be invalidated, or if data comparison could be skipped (e.g. rsync).
> 
> The difference is that mtime can be set by user, so using lower nsec
> bits for modification counter would require to truncate the user set
> time granularity to 100ns - that is probably acceptable, but only as
> an opt-in behavior.
> 
> The special value 0 for mtime-change-counter could be reserved for
> mtime that was set by the user or for upgrade of existing inode,
> where 0 counter means that mtime cannot be trusted as an accurate
> data modification-cookie.

What about write faults on an mmap region?  The first ro->rw transition
results in an mtime update, but not again until the page gets cleaned.

> This feature is going to be useful for the vfs HSM implementation [1]
> that I am working on and it actually rhymes with the XFS DMAPI
> patches that were never fully merged upstream.

Kudos, I cannot figure out a non-pejorative word that rhymes with
"**API". ;)

--D

> Speaking on behalf of my employer, we would love to see the data
> modification-cookie feature implemented, whether in vfs or in xfs.
> 
> *IF* the result on this thread is that the chosen solution is
> ctime-with-change-counter in XFS
> *AND* if there is agreement among XFS developers to extend it with
> an opt-in mkfs/mount option to 100ns-mtime-with-change-counter in XFS
> *THEN* I think I will be able to allocate resources to drive this xfs work.
> 
> Thanks,
> Amir.
> 
> [1] https://github.com/amir73il/fsnotify-utils/wiki/Hierarchical-Storage-Management-API
>
Dave Chinner Oct. 31, 2023, 11:47 p.m. UTC | #49
On Tue, Oct 31, 2023 at 04:02:42PM -0700, Darrick J. Wong wrote:
> On Wed, Nov 01, 2023 at 08:57:09AM +1100, Dave Chinner wrote:
> > On Tue, Oct 31, 2023 at 07:29:18AM -0400, Jeff Layton wrote:
> > > On Tue, 2023-10-31 at 09:03 +0200, Amir Goldstein wrote:
> > > > On Tue, Oct 31, 2023 at 3:42 AM Dave Chinner <david@fromorbit.com> wrote:
> > e.g. the current code flow for an atime update is:
> > 
> > touch_atime()
> >   atime_needs_update()
> >   <freeze/write protection>
> >   inode_update_time(S_ATIME)
> >     ->update_time(S_ATIME)
> >       <filesystem atime update>
> > 
> > I'd much prefer this to be:
> > 
> > touch_atime()
> >   if (->update_time(S_ATIME)) {
> >     ->update_time(S_ATIME)
> >       xfs_inode_update_time(S_ATIME)
> >         if (atime_needs_update())
> > 	  <filesystem atime update>
> >   } else {
> >     /* run the existing code */
> >   }
> > 
> > Similarly we'd turn file_modified()/file_update_time() inside out,
> > and this then allows the filesystem to add custom timestamp update
> > checks alongside the VFS timestamp update checks.
> > 
> > It would also enable us to untangle the mess that is lazytime, where
> > we have to implement ->update_time to catch lazytime updates and
> > punt them back to generic_update_time(), which then has to check for
> > lazytime again to determine how to dirty and queue the inode.
> > Of course, generic_update_time() also does timespec_equal checks on
> > timestamps to determine if times should be updated, and so we'd
> > probably need overrides on that, too.
> 
> Hmm.  So would the VFS update the incore timestamps of struct inode in
> whatever manner it wants?

That's kind of what I want to avoid - i want the filesystem to
direct the VFS as to the type of checks and modifications it can
make.

e.g. the timestamp comparisons and actions taken need to be
different for a timestamp-with-integrated-change-counter setup. It
doesn't fold neatly into inode_needs_update_time() - it becomes a
branchy, unreadable mess trying to handle all the different
situations.

Hence the VFS could provide two helpers - one for the existing
timestamp format and one for the new integrated change counter
timestamp. The filesystem can then select the right one to call.

And, further, filesystems that have lazytime enabled should be
checking that early to determine what to do. Lazytime specific
helpers would be useful here.

> Could that include incrementing the lower
> bits of i_ctime.tv_nsec for filesystems that advertise a non-1nsec
> granularity but also set a flag that effectively means "but you can use
> the lower tv_nsec bits if you want"?

Certainly. Similar to multi-grain timestamps, I don't see anything
filesystem specific about this mechanism. I think that anyone saying
"it's ok if it's internal to XFS" is still missing the point that
i_version as a VFS construct needs to die.

At most, i_version is only needed for filesystems that don't have
nanosecond timestamp resolution in their on-disk format and so need
some kind of external ctime change counter to provide fine-grained,
sub-timestamp granularity change recording.

> And perhaps after all that, the VFS should decide if a timestamp update
> needs to be persisted (e.g. lazytime/nodiratime/poniesatime) and if so,
> call ->update_time or __mark_inode_dirty?  Then XFS doesn't have to know
> about all the timestamp persistence rules, it just has to follow
> whatever the VFS tells it.

Sure. I'm not suggesting that the filesystem duplicate and encode
all these rules itself.

I'm just saying that it seems completely backwards that the VFS
encode all this generic logic to handle all these separate cases in
a single code path and then provides a callout that allows the
filesystem to override it's decisions (e.g. lazytime) and do
something else.

The filesystem already knows exactly what specific subset of checks
and updates need to be done so call ou tinto the filesystem first
and then run the VFS helpers that do exactly what is needed for
relatime, lazytime, using timestamps with integrated change
counters, etc.

> > Sorting the lazytime mess for internal change counters really needs
> > for all the timestamp updates to be handled in the filesystem, not
> > bounced back and forward between the filesystem and VFS helpers as
> > it currently is, hence I think we need to rework ->update_time to
> > make this all work cleanly.
> 
> (Oh, I guess I proposed sort of the opposite of what you just said.)

Not really, just seems you're thinking about how to code all the
VFS helpers we'd need a bit differently...

Cheers,

Dav.e
Amir Goldstein Nov. 1, 2023, 8:08 a.m. UTC | #50
On Wed, Nov 1, 2023 at 1:12 AM Darrick J. Wong <djwong@kernel.org> wrote:
>
> On Tue, Oct 31, 2023 at 09:03:57AM +0200, Amir Goldstein wrote:
> > On Tue, Oct 31, 2023 at 3:42 AM Dave Chinner <david@fromorbit.com> wrote:
> > >
> > [...]
> > > .... and what is annoying is that that the new i_version just a
> > > glorified ctime change counter. What we should be fixing is ctime -
> > > integrating this change counting into ctime would allow us to make
> > > i_version go away entirely. i.e. We don't need a persistent ctime
> > > change counter if the ctime has sufficient resolution or persistent
> > > encoding that it does not need an external persistent change
> > > counter.
> > >
> > > That was reasoning behind the multi-grain timestamps. While the mgts
> > > implementation was flawed, the reasoning behind it certainly isn't.
> > > We should be trying to get rid of i_version by integrating it into
> > > ctime updates, not arguing how atime vs i_version should work.
> > >
> > > > So I don't think the issue here is "i_version" per se. I think in a
> > > > vacuum, the best option of i_version is pretty obvious.  But if you
> > > > want i_version to track di_changecount, *then* you end up with that
> > > > situation where the persistence of atime matters, and i_version needs
> > > > to update whenever a (persistent) atime update happens.
> > >
> > > Yet I don't want i_version to track di_changecount.
> > >
> > > I want to *stop supporting i_version altogether* in XFS.
> > >
> > > I want i_version as filesystem internal metadata to die completely.
> > >
> > > I don't want to change the on disk format to add a new i_version
> > > field because we'll be straight back in this same siutation when the
> > > next i_version bug is found and semantics get changed yet again.
> > >
> > > Hence if we can encode the necessary change attributes into ctime,
> > > we can drop VFS i_version support altogether.  Then the "atime bumps
> > > i_version" problem also goes away because then we *don't use
> > > i_version*.
> > >
> > > But if we can't get the VFS to do this with ctime, at least we have
> > > the abstractions available to us (i.e. timestamp granularity and
> > > statx change cookie) to allow XFS to implement this sort of
> > > ctime-with-integrated-change-counter internally to the filesystem
> > > and be able to drop i_version support....
> > >
> >
> > I don't know if it was mentioned before in one of the many threads,
> > but there is another benefit of ctime-with-integrated-change-counter
> > approach - it is the ability to extend the solution with some adaptations
> > also to mtime.
> >
> > The "change cookie" is used to know if inode metadata cache should
> > be invalidated and mtime is often used to know if data cache should
> > be invalidated, or if data comparison could be skipped (e.g. rsync).
> >
> > The difference is that mtime can be set by user, so using lower nsec
> > bits for modification counter would require to truncate the user set
> > time granularity to 100ns - that is probably acceptable, but only as
> > an opt-in behavior.
> >
> > The special value 0 for mtime-change-counter could be reserved for
> > mtime that was set by the user or for upgrade of existing inode,
> > where 0 counter means that mtime cannot be trusted as an accurate
> > data modification-cookie.
>
> What about write faults on an mmap region?  The first ro->rw transition
> results in an mtime update, but not again until the page gets cleaned.
>

That problem is out of scope. As is the issue of whether mtime is updated
before or after the data modification.
For all practical matter, HSM (or any backup/sync software) could fsync
data of file before testing its mtime.
I am working on an additional mechanism (sb_start_write_srcu) which
HSM can use to close the rest of the possible TOCTOU races.

The problem that modification-cookie needs to solve is the coarse grained
mtime timestamp, very much like change-cookie for ctime and additionally
it needs to solve the problem that HSM needs to know if the mtime
timestamp was generated by the filesystem or set by the user.

Thanks,
Amir.
Jan Kara Nov. 1, 2023, 10:16 a.m. UTC | #51
On Wed 01-11-23 08:57:09, Dave Chinner wrote:
> 5. When-ever the inode is persisted, the timestamp is copied to the
> on-disk structure and the current change counter is folded in.
> 
> 	This means the on-disk structure always contains the latest
> 	change attribute that has been persisted, just like we
> 	currently do with i_version now.
> 
> 6. When-ever we read the inode off disk, we split the change counter
> from the timestamp and update the appropriate internal structures
> with this information.
> 
> 	This ensures that the VFS and userspace never see the change
> 	counter implementation in the inode timestamps.

OK, but is this compatible with the current XFS behavior? AFAICS currently
XFS sets sb->s_time_gran to 1 so timestamps currently stored on disk will
have some mostly random garbage in low bits of the ctime. Now if you look
at such inode with a kernel using this new scheme, stat(2) will report
ctime with low bits zeroed-out so if the ctime fetched in the old kernel was
stored in some external database and compared to the newly fetched ctime, it
will appear that ctime has gone backwards... Maybe we don't care but it is
a user visible change that can potentially confuse something.

								Honza
Amir Goldstein Nov. 1, 2023, 11:38 a.m. UTC | #52
On Wed, Nov 1, 2023 at 12:16 PM Jan Kara <jack@suse.cz> wrote:
>
> On Wed 01-11-23 08:57:09, Dave Chinner wrote:
> > 5. When-ever the inode is persisted, the timestamp is copied to the
> > on-disk structure and the current change counter is folded in.
> >
> >       This means the on-disk structure always contains the latest
> >       change attribute that has been persisted, just like we
> >       currently do with i_version now.
> >
> > 6. When-ever we read the inode off disk, we split the change counter
> > from the timestamp and update the appropriate internal structures
> > with this information.
> >
> >       This ensures that the VFS and userspace never see the change
> >       counter implementation in the inode timestamps.
>
> OK, but is this compatible with the current XFS behavior? AFAICS currently
> XFS sets sb->s_time_gran to 1 so timestamps currently stored on disk will
> have some mostly random garbage in low bits of the ctime. Now if you look
> at such inode with a kernel using this new scheme, stat(2) will report
> ctime with low bits zeroed-out so if the ctime fetched in the old kernel was
> stored in some external database and compared to the newly fetched ctime, it
> will appear that ctime has gone backwards... Maybe we don't care but it is
> a user visible change that can potentially confuse something.

See xfs_inode_has_bigtime() and auto-upgrade of inode format in
xfs_inode_item_precommit().

In the case of BIGTIME inode format, admin needs to opt-in to
BIGTIME format conversion by setting an INCOMPAT_BIGTIME
sb feature flag.

I imagine that something similar (inode flag + sb flag) would need
to be done for the versioned-timestamp, but I think that in that case,
the feature flag could be RO_COMPAT - there is no harm in exposing
made-up nsec lower bits if fs would be mounted read-only on an old
kernel, is there?

The same RO_COMPAT feature flag could also be used to determine
s_time_gran, because IIUC, s_time_gran for timestamp updates
is uniform across all inodes.

I know that Dave said he wants to avoid changing on-disk format,
but I am hoping that this well defined and backward compat with
lazy upgrade, on-disk format change may be acceptable?

Thanks,
Amir.
Linus Torvalds Nov. 1, 2023, 8:10 p.m. UTC | #53
On Wed, 1 Nov 2023 at 00:16, Jan Kara <jack@suse.cz> wrote:
>
> OK, but is this compatible with the current XFS behavior? AFAICS currently
> XFS sets sb->s_time_gran to 1 so timestamps currently stored on disk will
> have some mostly random garbage in low bits of the ctime.

I really *really* don't think we can use ctime as a "i_version"
replacement. The whole fine-granularity patches were well-intentioned,
but I do think they were broken.

Note that we can't use ctime as a "i_version" replacement for other
reasons too - you have filesystems like FAT - which people do want to
export - that have a single-second (or is it 2s?) granularity in
reality, even though they report a 1ns value in s_time_gran.

But here's a suggestion that people may hate, but that might just work
in practice:

 - get rid of i_version entirely

 - use the "known good" part of ctime as the upper bits of the change
counter (and by "known good" I mean tv_sec - or possibly even "tv_sec
/ 2" if that dim FAT memory of mine is right)

 - make the rule be that ctime is *never* updated for atime updates
(maybe that's already true, I didn't check - maybe it needs a new
mount flag for nfsd)

 - have a per-inode in-memory and vfs-internal (entirely invisible to
filesystems) "ctime modification counter" that is *NOT* a timestamp,
and is *NOT* i_version

 - make the rule be that the "ctime modification counter" is always
zero, *EXCEPT* if
    (a) I_VERSION_QUERIED is set
   AND
    (b) the ctime modification doesn't modify the "known good" part of ctime

so how the "statx change cookie" ends up being "high bits tv_sec of
ctime, low bits ctime modification cookie", and the end result of that
is:

 - if all the reads happen after the last write (common case), then
the low bits will be zero, because I_VERSION_QUERIED wasn't set when
ctime was modified

 - if you do a write *after* a modification, the ctime cookie is
guaranteed to change, because either the known good (sec/2sec) part of
ctime is new, *or* the counter gets updated

 - if the nfs server reboots, the in-memory counter will be cleared
again, and so the change cookie will cause client cache invalidations,
but *only* for those "ctime changed in the same second _after_
somebody did a read".

 - any long-time caches of files that don't get modified are all fine,
because they will have those low bits zero and depend on just the
stable part of ctime that works across filesystems. So there should be
no nasty thundering herd issues on long-lived caches on lots of
clients if the server reboots, or atime updates every 24 hours or
anything like that.

and note that *NONE* of this requires any filesystem involvement
(except for the rule of "no atime changes ever impact ctime", which
may or may not already be true).

The filesystem does *not* know about that modification counter,
there's no new on-disk stable information.

It's entirely possible that I'm missing something obvious, but the
above sounds to me like the only time you'd have stale invalidations
is really the (unusual) case of having writes after cached reads, and
then a reboot.

We'd get rid of "inode_maybe_inc_iversion()" entirely, and instead
replace it with logic in inode_set_ctime_current() that basically does

 - if the stable part of ctime changes, clear the new 32-bit counter

 - if I_VERSION_QUERIED isn't set, clear the new 32-bit counter

 - otherwise, increment the new 32-bit counter

and then the STATX_CHANGE_COOKIE code basically just returns

   (stable part of ctime << 32) + new 32-bit counter

(and again, the "stable part of ctime" is either just tv_sec, or it's
"tv_sec >> 1" or whatever).

The above does not expose *any* changes to timestamps to users, and
should work across a wide variety of filesystems, without requiring
any special code from the filesystem itself.

And now please all jump on me and say "No, Linus, that won't work, because XYZ".

Because it is *entirely* possible that I missed something truly
fundamental, and the above is completely broken for some obvious
reason that I just didn't think of.

             Linus
Trond Myklebust Nov. 1, 2023, 9:34 p.m. UTC | #54
On Wed, 2023-11-01 at 10:10 -1000, Linus Torvalds wrote:
> On Wed, 1 Nov 2023 at 00:16, Jan Kara <jack@suse.cz> wrote:
> > 
> > OK, but is this compatible with the current XFS behavior? AFAICS
> > currently
> > XFS sets sb->s_time_gran to 1 so timestamps currently stored on
> > disk will
> > have some mostly random garbage in low bits of the ctime.
> 
> I really *really* don't think we can use ctime as a "i_version"
> replacement. The whole fine-granularity patches were well-
> intentioned,
> but I do think they were broken.
> 
> Note that we can't use ctime as a "i_version" replacement for other
> reasons too - you have filesystems like FAT - which people do want to
> export - that have a single-second (or is it 2s?) granularity in
> reality, even though they report a 1ns value in s_time_gran.
> 
> But here's a suggestion that people may hate, but that might just
> work
> in practice:
> 
>  - get rid of i_version entirely
> 
>  - use the "known good" part of ctime as the upper bits of the change
> counter (and by "known good" I mean tv_sec - or possibly even "tv_sec
> / 2" if that dim FAT memory of mine is right)
> 
>  - make the rule be that ctime is *never* updated for atime updates
> (maybe that's already true, I didn't check - maybe it needs a new
> mount flag for nfsd)
> 
>  - have a per-inode in-memory and vfs-internal (entirely invisible to
> filesystems) "ctime modification counter" that is *NOT* a timestamp,
> and is *NOT* i_version
> 
>  - make the rule be that the "ctime modification counter" is always
> zero, *EXCEPT* if
>     (a) I_VERSION_QUERIED is set
>    AND
>     (b) the ctime modification doesn't modify the "known good" part
> of ctime
> 
> so how the "statx change cookie" ends up being "high bits tv_sec of
> ctime, low bits ctime modification cookie", and the end result of
> that
> is:
> 
>  - if all the reads happen after the last write (common case), then
> the low bits will be zero, because I_VERSION_QUERIED wasn't set when
> ctime was modified
> 
>  - if you do a write *after* a modification, the ctime cookie is
> guaranteed to change, because either the known good (sec/2sec) part
> of
> ctime is new, *or* the counter gets updated
> 
>  - if the nfs server reboots, the in-memory counter will be cleared
> again, and so the change cookie will cause client cache
> invalidations,
> but *only* for those "ctime changed in the same second _after_
> somebody did a read".
> 
>  - any long-time caches of files that don't get modified are all
> fine,
> because they will have those low bits zero and depend on just the
> stable part of ctime that works across filesystems. So there should
> be
> no nasty thundering herd issues on long-lived caches on lots of
> clients if the server reboots, or atime updates every 24 hours or
> anything like that.
> 
> and note that *NONE* of this requires any filesystem involvement
> (except for the rule of "no atime changes ever impact ctime", which
> may or may not already be true).
> 
> The filesystem does *not* know about that modification counter,
> there's no new on-disk stable information.
> 
> It's entirely possible that I'm missing something obvious, but the
> above sounds to me like the only time you'd have stale invalidations
> is really the (unusual) case of having writes after cached reads, and
> then a reboot.
> 
> We'd get rid of "inode_maybe_inc_iversion()" entirely, and instead
> replace it with logic in inode_set_ctime_current() that basically
> does
> 
>  - if the stable part of ctime changes, clear the new 32-bit counter
> 
>  - if I_VERSION_QUERIED isn't set, clear the new 32-bit counter
> 
>  - otherwise, increment the new 32-bit counter
> 
> and then the STATX_CHANGE_COOKIE code basically just returns
> 
>    (stable part of ctime << 32) + new 32-bit counter
> 
> (and again, the "stable part of ctime" is either just tv_sec, or it's
> "tv_sec >> 1" or whatever).
> 
> The above does not expose *any* changes to timestamps to users, and
> should work across a wide variety of filesystems, without requiring
> any special code from the filesystem itself.
> 
> And now please all jump on me and say "No, Linus, that won't work,
> because XYZ".
> 
> Because it is *entirely* possible that I missed something truly
> fundamental, and the above is completely broken for some obvious
> reason that I just didn't think of.
> 

My client writes to the file and immediately reads the ctime. A 3rd
party client then writes immediately after my ctime read.
A reboot occurs (maybe minutes later), then I re-read the ctime, and
get the same value as before the 3rd party write.

Yes, most of the time that is better than the naked ctime, but not
across a reboot.
Linus Torvalds Nov. 1, 2023, 10:23 p.m. UTC | #55
On Wed, Nov 1, 2023, 11:35 Trond Myklebust <trondmy@hammerspace.com> wrote:
>
> My client writes to the file and immediately reads the ctime. A 3rd
> party client then writes immediately after my ctime read.
> A reboot occurs (maybe minutes later), then I re-read the ctime, and
> get the same value as before the 3rd party write.
>
> Yes, most of the time that is better than the naked ctime, but not
> across a reboot.

Ahh, I knew I was missing something.

But I think it's fixable, with an additional rule:

 - when generating STATX_CHANGE_COOKIE, if the ctime matches the
current time and the ctime counter is zero, set the ctime counter to
1.

That means that you will have *spurious* cache invalidations of such
cached data after a reboot, but only for reads that happened right
after the file was written.

Now, it's obviously not unheard of to finish writing a file, and then
immediately reading the results again.

But at least those caches should be somewhat limited (and the problem
then only happens when the nfs server is rebooted).

I *assume* that the whole thundering herd issue with lots of clients
tends to be for stable files, not files that were just written and
then immediately cached?

I dunno. I'm sure there's still some thinko here.

             Linus
Trond Myklebust Nov. 1, 2023, 10:45 p.m. UTC | #56
On Wed, 2023-11-01 at 12:23 -1000, Linus Torvalds wrote:
> On Wed, Nov 1, 2023, 11:35 Trond Myklebust <trondmy@hammerspace.com>
> wrote:
> > 
> > My client writes to the file and immediately reads the ctime. A 3rd
> > party client then writes immediately after my ctime read.
> > A reboot occurs (maybe minutes later), then I re-read the ctime,
> > and
> > get the same value as before the 3rd party write.
> > 
> > Yes, most of the time that is better than the naked ctime, but not
> > across a reboot.
> 
> Ahh, I knew I was missing something.
> 
> But I think it's fixable, with an additional rule:
> 
>  - when generating STATX_CHANGE_COOKIE, if the ctime matches the
> current time and the ctime counter is zero, set the ctime counter to
> 1.
> 
> That means that you will have *spurious* cache invalidations of such
> cached data after a reboot, but only for reads that happened right
> after the file was written.

Presumably it will also happen if the file gets kicked out of cache on
the server, since that will cause the I_VERSION_QUERIED flag and any
other in-memory metadata to be lost.

> 
> Now, it's obviously not unheard of to finish writing a file, and then
> immediately reading the results again.
> 
> But at least those caches should be somewhat limited (and the problem
> then only happens when the nfs server is rebooted).
> 
> I *assume* that the whole thundering herd issue with lots of clients
> tends to be for stable files, not files that were just written and
> then immediately cached?
> 
> I dunno. I'm sure there's still some thinko here.

Close-to-open cache consistency means that the client is usually
expected to check the change attribute (or ctime) on file close and
file open. So it is not uncommon for it to have to revalidate the cache
not long after finishing writing the file. Of course, it is rare to
have another client interject with another write to the same file just
a few microseconds after it was closed, however it is extremely common
for that sort of behaviour to occur with directories.
Dave Chinner Nov. 1, 2023, 11:29 p.m. UTC | #57
On Wed, Nov 01, 2023 at 09:34:57PM +0000, Trond Myklebust wrote:
> On Wed, 2023-11-01 at 10:10 -1000, Linus Torvalds wrote:
> > The above does not expose *any* changes to timestamps to users, and
> > should work across a wide variety of filesystems, without requiring
> > any special code from the filesystem itself.
> > 
> > And now please all jump on me and say "No, Linus, that won't work,
> > because XYZ".
> > 
> > Because it is *entirely* possible that I missed something truly
> > fundamental, and the above is completely broken for some obvious
> > reason that I just didn't think of.
> > 
> 
> My client writes to the file and immediately reads the ctime. A 3rd
> party client then writes immediately after my ctime read.
> A reboot occurs (maybe minutes later), then I re-read the ctime, and
> get the same value as before the 3rd party write.
>
> Yes, most of the time that is better than the naked ctime, but not
> across a reboot.

This sort of "crash immediately after 3rd party data write" scenario
has never worked properly, even with i_version.

The issue is that 3rd party (local) buffered writes or metadata
changes do not require any integrity or metadata stability
operations to be performed by the filesystem unless O_[D]SYNC is set
on the fd, RWF_[D]SYNC is set on the IO, or f{data}sync() is
performed on the file.

Hence no local filesystem currently persists i_version or ctime
outside of operations with specific data integrity semantics.

nfsd based modifications have application specific persistence
requirements and that is triggered by the nfsd calling
->commit_metadata prior to returning the operation result to the
client. This is what persists i_version/timestamp changes that were
made during the nfsd operation - this persistence behaviour is not
driven by the local filesystem.

IOWs, this "change attribute failure" scenario is an existing
problem with the current i_version implementation.  It has always
been flawed in this way but this didn't matter a decade ago because
it's only purpose (and user) was nfsd and that had the required
persistence semantics to hide these flaws within the application's
context.

Now that we are trying to expose i_version as a "generic change
attribute", these persistence flaws get exposed because local
filesystem operations do not have the same enforced persistence
semantics as the NFS server.

This is another reason I want i_version to die.

What we need is a clear set of well defined semantics around statx
change attribute sampling. Correct crash-recovery/integrity behaviour
requires this rule:

  If the change attribute has been sampled, then the next
  modification to the filesystem that bumps change attribute *must*
  persist the change attribute modification atomically with the
  modification that requires it to change, or submit and complete
  persistence of the change attribute modification before the
  modification that requires it starts.

e.g. a truncate can bump the change attribute atomically with the
metadata changes in a transaction-based filesystem (ext4, XFS,
btrfs, bcachefs, etc).

Data writes are much harder, though. Some filesysetm structures can
write data and metadata in a single update e.g. log structured or
COW filesystems that can mix data and metadata like btrfs.
Journalling filesystems require ordering between journal writes and
the data writes to guarantee the change attribute is persistent
before we write the data. Non-journalling filesystems require inode
vs data write ordering.

Hence I strongly doubt that a persistent change attribute is best
implemented at the VFS - optimal, efficient implementations are
highly filesystem specific regardless of how the change attribute is
encoded in filesysetm metadata.

This is another reason I want to change how the inode timestamp code
is structured to call into the filesystem first rather than last.
Different filesystems will need to do different things to persist
a "ctime change counter" attribute correctly and efficiently -
it's not a one-size fits all situation....

-Dave.
Jeffrey Layton Nov. 2, 2023, 10:15 a.m. UTC | #58
On Wed, 2023-11-01 at 10:10 -1000, Linus Torvalds wrote:
> On Wed, 1 Nov 2023 at 00:16, Jan Kara <jack@suse.cz> wrote:
> > 
> > OK, but is this compatible with the current XFS behavior? AFAICS currently
> > XFS sets sb->s_time_gran to 1 so timestamps currently stored on disk will
> > have some mostly random garbage in low bits of the ctime.
> 
> I really *really* don't think we can use ctime as a "i_version"
> replacement. The whole fine-granularity patches were well-intentioned,
> but I do think they were broken.
> 

I have to take some issue here. I still the basic concept is sound. The
original implementation was flawed but I think I have a scheme that
could address the problems with the multigrain series.

That said, everyone seems to be haring off after other solutions. I
don't much care which one we end up with, as long as the problem gets
fixed.

> Note that we can't use ctime as a "i_version" replacement for other
> reasons too - you have filesystems like FAT - which people do want to
> export - that have a single-second (or is it 2s?) granularity in
> reality, even though they report a 1ns value in s_time_gran.
> 
> But here's a suggestion that people may hate, but that might just work
> in practice:
> 
>  - get rid of i_version entirely
> 
>  - use the "known good" part of ctime as the upper bits of the change
> counter (and by "known good" I mean tv_sec - or possibly even "tv_sec
> / 2" if that dim FAT memory of mine is right)
> 
>  - make the rule be that ctime is *never* updated for atime updates
> (maybe that's already true, I didn't check - maybe it needs a new
> mount flag for nfsd)
> 
>  - have a per-inode in-memory and vfs-internal (entirely invisible to
> filesystems) "ctime modification counter" that is *NOT* a timestamp,
> and is *NOT* i_version
> 
>  - make the rule be that the "ctime modification counter" is always
> zero, *EXCEPT* if
>     (a) I_VERSION_QUERIED is set
>    AND
>     (b) the ctime modification doesn't modify the "known good" part of ctime
> 
> so how the "statx change cookie" ends up being "high bits tv_sec of
> ctime, low bits ctime modification cookie", and the end result of that
> is:
> 
>  - if all the reads happen after the last write (common case), then
> the low bits will be zero, because I_VERSION_QUERIED wasn't set when
> ctime was modified
> 
>  - if you do a write *after* a modification, the ctime cookie is
> guaranteed to change, because either the known good (sec/2sec) part of
> ctime is new, *or* the counter gets updated
> 
>  - if the nfs server reboots, the in-memory counter will be cleared
> again, and so the change cookie will cause client cache invalidations,
> but *only* for those "ctime changed in the same second _after_
> somebody did a read".
> 
>  - any long-time caches of files that don't get modified are all fine,
> because they will have those low bits zero and depend on just the
> stable part of ctime that works across filesystems. So there should be
> no nasty thundering herd issues on long-lived caches on lots of
> clients if the server reboots, or atime updates every 24 hours or
> anything like that.
> 
> and note that *NONE* of this requires any filesystem involvement
> (except for the rule of "no atime changes ever impact ctime", which
> may or may not already be true).
> 
> The filesystem does *not* know about that modification counter,
> there's no new on-disk stable information.
> 
> It's entirely possible that I'm missing something obvious, but the
> above sounds to me like the only time you'd have stale invalidations
> is really the (unusual) case of having writes after cached reads, and
> then a reboot.
> 
> We'd get rid of "inode_maybe_inc_iversion()" entirely, and instead
> replace it with logic in inode_set_ctime_current() that basically does
> 
>  - if the stable part of ctime changes, clear the new 32-bit counter
> 
>  - if I_VERSION_QUERIED isn't set, clear the new 32-bit counter
> 
>  - otherwise, increment the new 32-bit counter
> 
> and then the STATX_CHANGE_COOKIE code basically just returns
> 
>    (stable part of ctime << 32) + new 32-bit counter
> 
> (and again, the "stable part of ctime" is either just tv_sec, or it's
> "tv_sec >> 1" or whatever).
> 
> The above does not expose *any* changes to timestamps to users, and
> should work across a wide variety of filesystems, without requiring
> any special code from the filesystem itself.
> 
> And now please all jump on me and say "No, Linus, that won't work, because XYZ".
> 
> Because it is *entirely* possible that I missed something truly
> fundamental, and the above is completely broken for some obvious
> reason that I just didn't think of.
> 

Yeah, I think this scheme is problematic for the reasons Trond pointed
out. I also don't quite see the advantage of this over what Dave Chinner
is proposing (using low-order bits of the ctime nsec field to hold a
change counter).
Jeffrey Layton Nov. 2, 2023, 10:17 a.m. UTC | #59
On Wed, 2023-11-01 at 13:38 +0200, Amir Goldstein wrote:
> On Wed, Nov 1, 2023 at 12:16 PM Jan Kara <jack@suse.cz> wrote:
> > 
> > On Wed 01-11-23 08:57:09, Dave Chinner wrote:
> > > 5. When-ever the inode is persisted, the timestamp is copied to the
> > > on-disk structure and the current change counter is folded in.
> > > 
> > >       This means the on-disk structure always contains the latest
> > >       change attribute that has been persisted, just like we
> > >       currently do with i_version now.
> > > 
> > > 6. When-ever we read the inode off disk, we split the change counter
> > > from the timestamp and update the appropriate internal structures
> > > with this information.
> > > 
> > >       This ensures that the VFS and userspace never see the change
> > >       counter implementation in the inode timestamps.
> > 
> > OK, but is this compatible with the current XFS behavior? AFAICS currently
> > XFS sets sb->s_time_gran to 1 so timestamps currently stored on disk will
> > have some mostly random garbage in low bits of the ctime. Now if you look
> > at such inode with a kernel using this new scheme, stat(2) will report
> > ctime with low bits zeroed-out so if the ctime fetched in the old kernel was
> > stored in some external database and compared to the newly fetched ctime, it
> > will appear that ctime has gone backwards... Maybe we don't care but it is
> > a user visible change that can potentially confuse something.
> 
> See xfs_inode_has_bigtime() and auto-upgrade of inode format in
> xfs_inode_item_precommit().
> 
> In the case of BIGTIME inode format, admin needs to opt-in to
> BIGTIME format conversion by setting an INCOMPAT_BIGTIME
> sb feature flag.
> 
> I imagine that something similar (inode flag + sb flag) would need
> to be done for the versioned-timestamp, but I think that in that case,
> the feature flag could be RO_COMPAT - there is no harm in exposing
> made-up nsec lower bits if fs would be mounted read-only on an old
> kernel, is there?
> 
> The same RO_COMPAT feature flag could also be used to determine
> s_time_gran, because IIUC, s_time_gran for timestamp updates
> is uniform across all inodes.
> 
> I know that Dave said he wants to avoid changing on-disk format,
> but I am hoping that this well defined and backward compat with
> lazy upgrade, on-disk format change may be acceptable?


With the ctime, we're somewhat saved by the fact that it's not settable
by users, so we don't need to worry as much about returning specific
values there, I think.

With the scheme Dave is proposing, booting to a new kernel vs. an old
kernel might show a different ctime on an inode though. That might be
enough to justify needing a way to opt-in to the change on existing
filesystems.
Jeffrey Layton Nov. 2, 2023, 10:29 a.m. UTC | #60
On Thu, 2023-11-02 at 10:29 +1100, Dave Chinner wrote:
> On Wed, Nov 01, 2023 at 09:34:57PM +0000, Trond Myklebust wrote:
> > On Wed, 2023-11-01 at 10:10 -1000, Linus Torvalds wrote:
> > > The above does not expose *any* changes to timestamps to users, and
> > > should work across a wide variety of filesystems, without requiring
> > > any special code from the filesystem itself.
> > > 
> > > And now please all jump on me and say "No, Linus, that won't work,
> > > because XYZ".
> > > 
> > > Because it is *entirely* possible that I missed something truly
> > > fundamental, and the above is completely broken for some obvious
> > > reason that I just didn't think of.
> > > 
> > 
> > My client writes to the file and immediately reads the ctime. A 3rd
> > party client then writes immediately after my ctime read.
> > A reboot occurs (maybe minutes later), then I re-read the ctime, and
> > get the same value as before the 3rd party write.
> > 
> > Yes, most of the time that is better than the naked ctime, but not
> > across a reboot.
> 
> This sort of "crash immediately after 3rd party data write" scenario
> has never worked properly, even with i_version.
> 
> The issue is that 3rd party (local) buffered writes or metadata
> changes do not require any integrity or metadata stability
> operations to be performed by the filesystem unless O_[D]SYNC is set
> on the fd, RWF_[D]SYNC is set on the IO, or f{data}sync() is
> performed on the file.
> 
> Hence no local filesystem currently persists i_version or ctime
> outside of operations with specific data integrity semantics.
> 
> nfsd based modifications have application specific persistence
> requirements and that is triggered by the nfsd calling
> ->commit_metadata prior to returning the operation result to the
> client. This is what persists i_version/timestamp changes that were
> made during the nfsd operation - this persistence behaviour is not
> driven by the local filesystem.
> 
> IOWs, this "change attribute failure" scenario is an existing
> problem with the current i_version implementation.  It has always
> been flawed in this way but this didn't matter a decade ago because
> it's only purpose (and user) was nfsd and that had the required
> persistence semantics to hide these flaws within the application's
> context.
>
> Now that we are trying to expose i_version as a "generic change
> attribute", these persistence flaws get exposed because local
> filesystem operations do not have the same enforced persistence
> semantics as the NFS server.
> 
> This is another reason I want i_version to die.
> 
> What we need is a clear set of well defined semantics around statx
> change attribute sampling. Correct crash-recovery/integrity behaviour
> requires this rule:
> 
>   If the change attribute has been sampled, then the next
>   modification to the filesystem that bumps change attribute *must*
>   persist the change attribute modification atomically with the
>   modification that requires it to change, or submit and complete
>   persistence of the change attribute modification before the
>   modification that requires it starts.
> 
> e.g. a truncate can bump the change attribute atomically with the
> metadata changes in a transaction-based filesystem (ext4, XFS,
> btrfs, bcachefs, etc).
> 
> Data writes are much harder, though. Some filesysetm structures can
> write data and metadata in a single update e.g. log structured or
> COW filesystems that can mix data and metadata like btrfs.
> Journalling filesystems require ordering between journal writes and
> the data writes to guarantee the change attribute is persistent
> before we write the data. Non-journalling filesystems require inode
> vs data write ordering.
> 
> Hence I strongly doubt that a persistent change attribute is best
> implemented at the VFS - optimal, efficient implementations are
> highly filesystem specific regardless of how the change attribute is
> encoded in filesysetm metadata.
> 
> This is another reason I want to change how the inode timestamp code
> is structured to call into the filesystem first rather than last.
> Different filesystems will need to do different things to persist
> a "ctime change counter" attribute correctly and efficiently -
> it's not a one-size fits all situation....

FWIW, the big danger for nfsd is is i_version rollback after a crash:

We can end up handing out an i_version value to the client before it
ever makes it to disk. If that happens, and the server crashes before it
ever makes it to disk, then the client can see the old i_version when it
queries it again (assuming the earlier write was lost).

That, in an of itself, is not a _huge_ problem for NFS clients. They'll
typically just invalidate their cache if that occurs and reread any data
they need.

The real danger is that you can have a write that occurs after the
reboot that is different from the earlier one and hand out a change
attribute that is a duplicate of the one viewed earlier. Now you have
the same change attribute that refers to two different states of the
file (and potential data corruption).

We mitigate that today by factoring in the ctime on regular files when
generating the change attribute (see nfsd4_change_attribute()). In
theory, i_version rolling back + a clock jump backward could generate
change attr collisions, even with that, but that's a bit harder to
contrive so we mostly don't worry about it.

I'm all for coming up with a way to make this more resilient though. If
we can offer the guarantee that you're proposing above, then that would
be a very nice thing.
diff mbox series

Patch

diff --git a/include/linux/timekeeper_internal.h b/include/linux/timekeeper_internal.h
index 84ff2844df2a..6583b06e7d08 100644
--- a/include/linux/timekeeper_internal.h
+++ b/include/linux/timekeeper_internal.h
@@ -55,6 +55,7 @@  struct tk_read_base {
  * @tai_offset:		The current UTC to TAI offset in seconds
  * @clock_was_set_seq:	The sequence number of clock was set events
  * @cs_was_changed_seq:	The sequence number of clocksource change events
+ * @mg_nsec:		Nanosecond delta for multigrain timestamps
  * @next_leap_ktime:	CLOCK_MONOTONIC time value of a pending leap-second
  * @raw_sec:		CLOCK_MONOTONIC_RAW  time in seconds
  * @monotonic_to_boot:	CLOCK_MONOTONIC to CLOCK_BOOTTIME offset
@@ -110,6 +111,7 @@  struct timekeeper {
 	u64			xtime_interval;
 	s64			xtime_remainder;
 	u64			raw_interval;
+	u64			mg_nsec;
 	/* The ntp_tick_length() value currently being used.
 	 * This cached copy ensures we consistently apply the tick
 	 * length for an entire tick, as ntp_tick_length may change
diff --git a/include/linux/timekeeping.h b/include/linux/timekeeping.h
index fe1e467ba046..5dc0ad619d42 100644
--- a/include/linux/timekeeping.h
+++ b/include/linux/timekeeping.h
@@ -44,6 +44,10 @@  extern void ktime_get_real_ts64(struct timespec64 *tv);
 extern void ktime_get_coarse_ts64(struct timespec64 *ts);
 extern void ktime_get_coarse_real_ts64(struct timespec64 *ts);
 
+/* multigrain timestamp support */
+void ktime_get_mg_fine_ts64(struct timespec64 *ts);
+void ktime_get_mg_coarse_ts64(struct timespec64 *ts);
+
 void getboottime64(struct timespec64 *ts);
 
 /*
diff --git a/kernel/time/timekeeping.c b/kernel/time/timekeeping.c
index 266d02809dbb..7c20c98b1ea8 100644
--- a/kernel/time/timekeeping.c
+++ b/kernel/time/timekeeping.c
@@ -33,6 +33,9 @@ 
 #define TK_MIRROR		(1 << 1)
 #define TK_CLOCK_WAS_SET	(1 << 2)
 
+/* When mg_nsec is set to this, ignore it */
+#define TK_MG_NSEC_IGNORE	(~(0ULL))
+
 enum timekeeping_adv_mode {
 	/* Update timekeeper when a tick has passed */
 	TK_ADV_TICK,
@@ -139,6 +142,7 @@  static void tk_set_xtime(struct timekeeper *tk, const struct timespec64 *ts)
 {
 	tk->xtime_sec = ts->tv_sec;
 	tk->tkr_mono.xtime_nsec = (u64)ts->tv_nsec << tk->tkr_mono.shift;
+	tk->mg_nsec = TK_MG_NSEC_IGNORE;
 }
 
 static void tk_xtime_add(struct timekeeper *tk, const struct timespec64 *ts)
@@ -146,6 +150,7 @@  static void tk_xtime_add(struct timekeeper *tk, const struct timespec64 *ts)
 	tk->xtime_sec += ts->tv_sec;
 	tk->tkr_mono.xtime_nsec += (u64)ts->tv_nsec << tk->tkr_mono.shift;
 	tk_normalize_xtime(tk);
+	tk->mg_nsec = TK_MG_NSEC_IGNORE;
 }
 
 static void tk_set_wall_to_mono(struct timekeeper *tk, struct timespec64 wtm)
@@ -1664,6 +1669,7 @@  void __init timekeeping_init(void)
 	tk_setup_internals(tk, clock);
 
 	tk_set_xtime(tk, &wall_time);
+	tk->mg_nsec = TK_MG_NSEC_IGNORE;
 	tk->raw_sec = 0;
 
 	tk_set_wall_to_mono(tk, wall_to_mono);
@@ -2283,6 +2289,80 @@  void ktime_get_coarse_ts64(struct timespec64 *ts)
 }
 EXPORT_SYMBOL(ktime_get_coarse_ts64);
 
+/**
+ * ktime_get_mg_fine_ts64 - Returns a fine-grained time of day in a timespec64
+ * @ts: pointer to the timespec64 to be set
+ *
+ * Multigrain filesystems use a scheme where they use coarse-grained
+ * timestamps most of the time, but will use a fine-grained timestamp
+ * if the previous timestamp was queried and the coarse-grained clock
+ * hasn't ticked yet.
+ *
+ * For this case, the caller is requesting a fine-grained timestamp,
+ * so we must update the sliding mg_nsec value before returning it. This
+ * ensures that any coarse-grained timestamp updates that occur after
+ * this won't appear to have occurred before.
+ *
+ * Fills in @ts with the current fine-grained time of day, suitable for
+ * a file timestamp.
+ */
+void ktime_get_mg_fine_ts64(struct timespec64 *ts)
+{
+	struct timekeeper *tk = &tk_core.timekeeper;
+	unsigned long flags;
+	u32 nsecs;
+
+	WARN_ON(timekeeping_suspended);
+
+	raw_spin_lock_irqsave(&timekeeper_lock, flags);
+	write_seqcount_begin(&tk_core.seq);
+
+	ts->tv_sec = tk->xtime_sec;
+	nsecs = timekeeping_get_ns(&tk->tkr_mono);
+	tk->mg_nsec = nsecs;
+
+	write_seqcount_end(&tk_core.seq);
+	raw_spin_unlock_irqrestore(&timekeeper_lock, flags);
+
+	ts->tv_nsec = 0;
+	timespec64_add_ns(ts, nsecs);
+}
+
+/**
+ * ktime_get_mg_coarse_ts64 - Returns a coarse-grained time of day in a timespec64
+ * @ts: pointer to the timespec64 to be set
+ *
+ * Multigrain filesystems use a scheme where they use coarse-grained
+ * timestamps most of the time, but will use a fine-grained timestamp
+ * if the previous timestamp was queried and the coarse-grained clock
+ * hasn't ticked yet.
+ *
+ * For this case, the caller is requesting a coarse-grained timestamp,
+ * which will always be equal to or later than the latest fine-grained
+ * timestamp that has been handed out.
+ *
+ * Fills in @ts with the current coarse-grained time of day, suitable
+ * for a file timestamp.
+ */
+void ktime_get_mg_coarse_ts64(struct timespec64 *ts)
+{
+	struct timekeeper *tk = &tk_core.timekeeper;
+	unsigned int seq;
+	u64 nsec;
+
+	do {
+		seq = read_seqcount_begin(&tk_core.seq);
+
+		*ts = tk_xtime(tk);
+		nsec = tk->mg_nsec;
+	} while (read_seqcount_retry(&tk_core.seq, seq));
+
+	if (nsec != TK_MG_NSEC_IGNORE) {
+		ts->tv_nsec = 0;
+		timespec64_add_ns(ts, nsec);
+	}
+}
+
 /*
  * Must hold jiffies_lock
  */