Re: [PATCH RFC 1/3] vfs: add copy_file_range syscall and vfs helper
From: Zach Brown
Date: Tue Apr 14 2015 - 14:54:35 EST
On Tue, Apr 14, 2015 at 02:29:06PM -0400, J. Bruce Fields wrote:
> On Tue, Apr 14, 2015 at 11:22:41AM -0700, Zach Brown wrote:
> > On Tue, Apr 14, 2015 at 02:19:11PM -0400, J. Bruce Fields wrote:
> > > On Tue, Apr 14, 2015 at 01:16:13PM -0400, Anna Schumaker wrote:
> > > > On 04/14/2015 12:53 PM, Christoph Hellwig wrote:
> > > > > On Sat, Apr 11, 2015 at 09:04:02AM -0400, Jeff Layton wrote:
> > > > >> Yuck! How the heck do you clean up the mess if that happens? I
> > > > >> guess you're just stuck redoing the copy with normal READ/WRITE?
> > > > >>
> > > > >> Maybe we need to have the interface return a hard error in that
> > > > >> case and not try to give back any sort of offset?
> > > > >
> > > > > The NFSv4.2 COPY interface is a train wreck. At least for Linux I'd
> > > > > expect us to simply ignore it and only implement my new CLONE
> > > > > operation with sane semantics. That is unless someone can show some
> > > > > real life use case for the inter server copy, in which case we'll
> > > > > have to deal with that mess. But getting that one right at the VFS
> > > > > level will be a nightmare anyway.
> > > > >
> > > > > Make this a vote from me to not support partial copies and just
> > > > > return and error in that case.
> > > >
> > > > Agreed. Looking at the v4.2 spec, COPY does take ca_consecutive and a
> > > > ca_synchronous flags that let the client state if the copy should be
> > > > done consecutively or synchronously. I expected to always set
> > > > consecutive to "true" for the Linux client.
> > >
> > > That's supposed to mean results are well-defined in the partial-copy
> > > case, but I think Christoph's suggesting eliminating the partial-copy
> > > case entirely?
> > >
> > > Which would be fine with me.
> > >
> > > It might actually have been me advocating for partial copies. But that
> > > was only because a partial-copy-handling-loop seemed simpler to me than
> > > progress callbacks if we were going to support long-running copies.
> > >
> > > I'm happy enough not to have it at all.
> > Ah, OK, that's great news.
> > I thought at one point we were worried about very long running RPCs on
> > the server. Are we not worried about that now?
> > Is the client expected to cut the work up into arbitrarily managable
> > chunks? Is the server expected to fail COPY/CLONE requests that it
> > thinks would take way too long? Something else?
> Christoph is proposing a CLONE rpc that's required to be atomic:
> "The CLONE operation is atomic, that is either all changes or no
> changes are seen by the client or other clients."
> So that couldn't be really long-running (or the server is nuts).
> So that'd mean Anna would rip out the server-side copy loop and we'd
> initially just support btrfs or whatever.
Is this relying on btrfs range cloning being atomic? It certainly
doesn't look atomic. It can modify items across an arbitrarily large
number of leaf blocks. It can make the changes across multiple
transactions which could introduce partial modification on reboot after
crashes. It can fail (the dynamic duo: enomem, eio) and leave the
desintation partially modified.
> I mean the server-side copy loop may also be useful but I'm all for
> wiring up the obvious case first.
Sure, I'm all for wiring up the simple version that doesn't return
partial progress. If that'll work for you guys.
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/