lists.openwall.net   lists  /  announce  owl-users  owl-dev  john-users  john-dev  passwdqc-users  yescrypt  popa3d-users  /  oss-security  kernel-hardening  musl  sabotage  tlsify  passwords  /  crypt-dev  xvendor  /  Bugtraq  Full-Disclosure  linux-kernel  linux-netdev  linux-ext4  linux-hardening  linux-cve-announce  PHC 
Open Source and information security mailing list archives
 
Hash Suite for Android: free password hash cracker in your pocket
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20231031230242.GC1205143@frogsfrogsfrogs>
Date:   Tue, 31 Oct 2023 16:02:42 -0700
From:   "Darrick J. Wong" <djwong@...nel.org>
To:     Dave Chinner <david@...morbit.com>
Cc:     Jeff Layton <jlayton@...nel.org>,
        Amir Goldstein <amir73il@...il.com>,
        Linus Torvalds <torvalds@...ux-foundation.org>,
        Kent Overstreet <kent.overstreet@...ux.dev>,
        Christian Brauner <brauner@...nel.org>,
        Alexander Viro <viro@...iv.linux.org.uk>,
        John Stultz <jstultz@...gle.com>,
        Thomas Gleixner <tglx@...utronix.de>,
        Stephen Boyd <sboyd@...nel.org>,
        Chandan Babu R <chandan.babu@...cle.com>,
        Theodore Ts'o <tytso@....edu>,
        Andreas Dilger <adilger.kernel@...ger.ca>,
        Chris Mason <clm@...com>, Josef Bacik <josef@...icpanda.com>,
        David Sterba <dsterba@...e.com>,
        Hugh Dickins <hughd@...gle.com>,
        Andrew Morton <akpm@...ux-foundation.org>,
        Jan Kara <jack@...e.de>, David Howells <dhowells@...hat.com>,
        linux-fsdevel@...r.kernel.org, linux-kernel@...r.kernel.org,
        linux-xfs@...r.kernel.org, linux-ext4@...r.kernel.org,
        linux-btrfs@...r.kernel.org, linux-mm@...ck.org,
        linux-nfs@...r.kernel.org
Subject: Re: [PATCH RFC 2/9] timekeeping: new interfaces for multigrain
 timestamp handing

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@...morbit.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@...morbit.com

Powered by blists - more mailing lists

Powered by Openwall GNU/*/Linux Powered by OpenVZ