Re: [PATCH v8 3/5] loop: Add support for provision requests

[Dave Chinner]

On Tue, Oct 10, 2023 at 03:43:10PM -0700, Sarthak Kukreti wrote:
> On Sun, Oct 8, 2023 at 4:37 PM Dave Chinner <david@xxxxxxxxxxxxx> wrote:
> >
> > On Fri, Oct 06, 2023 at 06:28:15PM -0700, Sarthak Kukreti wrote:
> > > Add support for provision requests to loopback devices.
> > > Loop devices will configure provision support based on
> > > whether the underlying block device/file can support
> > > the provision request and upon receiving a provision bio,
> > > will map it to the backing device/storage. For loop devices
> > > over files, a REQ_OP_PROVISION request will translate to
> > > an fallocate mode 0 call on the backing file.
> > >
> > > Signed-off-by: Sarthak Kukreti <sarthakkukreti@xxxxxxxxxxxx>
> > > Signed-off-by: Mike Snitzer <snitzer@xxxxxxxxxx>
> >
> >
> > Hmmmm.
> >
> > This doesn't actually implement the required semantics of
> > REQ_PROVISION. Yes, it passes the command to the filesystem
> > fallocate() implementation, but fallocate() at the filesystem level
> > does not have the same semantics as REQ_PROVISION.
> >
> > i.e. at the filesystem level, fallocate() only guarantees the next
> > write to the provisioned range will succeed without ENOSPC, it does
> > not guarantee *every* write to the range will succeed without
> > ENOSPC. If someone clones the loop file while it is in use (i.e.
> > snapshots it via cp --reflink) then all guarantees that the next
> > write to a provisioned LBA range will succeed without ENOSPC are
> > voided.
> >
> > So while this will work for basic testing that the filesystem is
> > issuing REQ_PROVISION based IO correctly, it can't actually be used
> > for hosting production filesystems that need full REQ_PROVISION
> > guarantees when the loop device backing file is independently
> > shapshotted via FICLONE....
> >
> > At minimuim, this set of implementation constraints needs tobe
> > documented somewhere...
> >
> Fair point. I wanted to have a separate fallocate() mode
> (FALLOC_FL_PROVISION) in the earlier series of the patchset so that we
> can distinguish between a provision request and a regular fallocate()
> call; I dropped it from the series after feedback that the default
> case should suffice. But this might be one of the cases where we need
> an explicit intent that we want to provision space.

ISTR that I commented that filesystems like XFS can't implement
REQ_PROVISION semantics for extents without on-disk format
changes. Hence that needs to happen before we expose a new API to
userspace....

> Given a separate FALLOC_FL_PROVISION mode in the scenario you
> mentioned, the filesystem could copy previously 'provisioned' blocks
> to new blocks (which implicitly provisions them) or reserve blocks for
> use (and passing through REQ_OP_PROVISION below). That also means that
> the filesystem should track 'provisioned' blocks and take appropriate
> actions to ensure the provisioning guarantees.

Yes, tracking provisioned ranges persistently and the reservations
they require needs on-disk filesytem format changes compared to just
preallocating space.  None of this functionality currently exists in
any filesystem that supports shared extents, and it's a fairly
significant chunk of development work to support it.

Nobody has planned to do this sort of complex surgery to XFS at
this point in time. I doubt that anyone on the btrfs side of
things is really even following this discussion because this is
largely for block device thinp and snapshot support
and btrfs just doesn't care about that.

> For filesystems without copy-on-write semantics (eg. ext4),
> REQ_OP_PROVISION should still be equivalent to mode == 0.

Well, yes. This is the same situation as "for non-sparse block
devices, REQ_PROVISION can just be ignored." This is not an
interesting use case, nor a use case that the functionality or APIs
should be designed around.

-Dave.
-- 
Dave Chinner
david@xxxxxxxxxxxxx