Re: [PATCH v7 0/5] Introduce provisioning primitives

[Mike Snitzer]

On Mon, Jun 05 2023 at  5:14P -0400,
Sarthak Kukreti <sarthakkukreti@xxxxxxxxxxxx> wrote:

> On Sat, Jun 3, 2023 at 8:57 AM Mike Snitzer <snitzer@xxxxxxxxxx> wrote:
> >
> > We all just need to focus on your proposal and Joe's dm-thin
> > reservation design...
> >
> > [Sarthak: FYI, this implies that it doesn't really make sense to add
> > dm-thinp support before Joe's design is implemented.  Otherwise we'll
> > have 2 different responses to REQ_OP_PROVISION.  The one that is
> > captured in your patchset isn't adequate to properly handle ensuring
> > upper layer (like XFS) can depend on the space being available across
> > snapshot boundaries.]
> >
> Ack. Would it be premature for the rest of the series to go through
> (REQ_OP_PROVISION + support for loop and non-dm-thinp device-mapper
> targets)? I'd like to start using this as a reference to suggest
> additions to the virtio-spec for virtio-blk support and start looking
> at what an ext4 implementation would look like.

Please drop the dm-thin.c and dm-snap.c changes.  dm-snap.c would need
more work to provide the type of guarantee XFS requires across
snapshot boundaries. I'm inclined to _not_ add dm-snap.c support
because it is best to just use dm-thin.

And FYI even your dm-thin patch will be the starting point for the
dm-thin support (we'll keep attribution to you for all the code in a
separate patch).

> Fair points, I certainly don't want to derail this conversation; I'd
> be happy to see this work merged sooner rather than later.

Once those dm target changes are dropped I think the rest of the
series is fine to go upstream now.  Feel free to post a v8.

> For posterity, I'll distill what I said above into the following: I'd like
> a capability for userspace to create thin snapshots that ignore the
> thin volume's provisioned areas. IOW, an opt-in flag which makes
> snapshots fallback to what they do today to provide flexibility to
> userspace to decide the space requirements for the above mentioned
> scenarios, and at the same time, not adding separate corner case
> handling for filesystems. But to reiterate, my intent isn't to pile
> this onto the work you, Mike and Joe have planned; just some insight
> into why I'm in favor of ideas that reduce the snapshot size.

I think it'd be useful to ignore a thin device's reservation for
read-only snapshots.  Adding the ability to create read-only thin
snapshots could make sense (later activations don't necessarily need
to impose read-only, doing so would require some additional work).

Mike