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

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Fri, 27 Oct 2023 at 20:36, Jeff Layton <jlayton@xxxxxxxxxx> 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@xxxxxxxxxxxxx> 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@xxxxxxxxxxxxx> 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@xxxxxxxxxx>




[Index of Archives]     [XFS Filesystem Development (older mail)]     [Linux Filesystem Development]     [Linux Audio Users]     [Yosemite Trails]     [Linux Kernel]     [Linux RAID]     [Linux SCSI]


  Powered by Linux