Re: [PATCH v1 0/8] VFS: In-kernel copy system call

[Darrick J. Wong]

On Tue, Sep 08, 2015 at 02:45:39PM -0700, Andy Lutomirski wrote:
> On Tue, Sep 8, 2015 at 2:29 PM, Darrick J. Wong <address@hidden> wrote:
> > On Tue, Sep 08, 2015 at 09:03:09PM +0100, Pádraig Brady wrote:
> >> On 08/09/15 20:10, Andy Lutomirski wrote:
> >> > On Tue, Sep 8, 2015 at 11:23 AM, Anna Schumaker
> >> > <address@hidden> wrote:
> >> >> On 09/08/2015 11:21 AM, Pádraig Brady wrote:
> >> >>> I see copy_file_range() is a reflink() on BTRFS?
> >> >>> That's a bit surprising, as it avoids the copy completely.
> >> >>> cp(1) for example considered doing a BTRFS clone by default,
> >> >>> but didn't due to expectations that users actually wanted
> >> >>> the data duplicated on disk for resilience reasons,
> >> >>> and for performance reasons so that write latencies were
> >> >>> restricted to the copy operation, rather than being
> >> >>> introduced at usage time as the dest file is CoW'd.
> >> >>>
> >> >>> If reflink() is a possibility for copy_file_range()
> >> >>> then could it be done optionally with a flag?
> >> >>
> >> >> The idea is that filesystems get to choose how to handle copies in the
> >> >> default case.  BTRFS could do a reflink, but NFS could do a server side
> >
> > Eww, different default behaviors depending on the filesystem. :)
> >
> >> >> copy instead.  I can change the default behavior to only do a data copy
> >> >> (unless the reflink flag is specified) instead, if that is desirable.
> >> >>
> >> >> What does everybody think?
> >> >
> >> > I think the best you could do is to have a hint asking politely for
> >> > the data to be deep-copied.  After all, some filesystems reserve the
> >> > right to transparently deduplicate.
> >> >
> >> > Also, on a true COW filesystem (e.g. btrfs sometimes), there may be no
> >> > advantage to deep copying unless you actually want two copies for
> >> > locality reasons.
> >>
> >> Agreed. The relink and server side copy are separate things.
> >> There's no advantage to not doing a server side copy,
> >> but as mentioned there may be advantages to doing deep copies on BTRFS
> >> (another reason not previous mentioned in this thread, would be
> >> to avoid ENOSPC errors at some time in the future).
> >>
> >> So having control over the deep copy seems useful.
> >> It's debatable whether ALLOW_REFLINK should be on/off by default
> >> for copy_file_range().  I'd be inclined to have such a setting off by 
> >> default,
> >> but cp(1) at least will work with whatever is chosen.
> >
> > So far it looks like people are interested in at least these "make data 
> > appear
> > in this other place" filesystem operations:
> >
> > 1. reflink
> > 2. reflink, but only if the contents are the same (dedupe)
> 
> What I meant by this was: if you ask for "regular copy", you may end
> up with a reflink anyway.  Anyway, how can you reflink a range and
> have the contents *not* be the same?

reflink forcibly remaps fd_dest's range to fd_src's range.  If they didn't
match before, they will afterwards.

dedupe remaps fd_dest's range to fd_src's range only if they match, of course.

Perhaps I should have said "...if the contents are the same before the call"?

> 
> > 3. regular copy
> > 4. regular copy, but make the hardware do it for us
> > 5. regular copy, but require a second copy on the media (no-dedupe)
> 
> If this comes from me, I have no desire to ever use this as a flag.

I meant (5) as a "disable auto-dedupe for this operation" flag, not as
a "reallocate all the shared blocks now" op...

> If someone wants to use chattr or some new operation to say "make this
> range of this file belong just to me for purpose of optimizing future
> writes", then sure, go for it, with the understanding that there are
> plenty of filesystems for which that doesn't even make sense.

"Unshare these blocks" sounds more like something fallocate could do.

So far in my XFS reflink playground, it seems that using the defrag tool to
un-cow a file makes most sense.  AFAICT the XFS and ext4 defraggers copy a
fragmented file's data to a second file and use a 'swap extents' operation,
after which the donor file is unlinked.

Hey, if this syscall turns into a more generic "do something involving two
(fd:off:len) (fd:off:len) tuples" call, I guess we could throw in "swap
extents" as a 7th operation, to refactor the ioctls.  <smirk>

> 
> > 6. regular copy, but don't CoW (eatmyothercopies) (joke)
> >
> > (Please add whatever ops I missed.)
> >
> > I think I can see a case for letting (4) fall back to (3) since (4) is an
> > optimization of (3).
> >
> > However, I particularly don't like the idea of (1) falling back to (3-5).
> > Either the kernel can satisfy a request or it can't, but let's not just
> > assume that we should transmogrify one type of request into another.  
> > Userspace
> > should decide if a reflink failure should turn into one of the copy 
> > variants,
> > depending on whether the user wants to spread allocation costs over 
> > rewrites or
> > pay it all up front.  Also, if we allow reflink to fall back to copy, how do
> > programs find out what actually took place?  Or do we simply not allow them 
> > to
> > find out?
> >
> > Also, programs that expect reflink either to finish or fail quickly might be
> > surprised if it's possible for reflink to take a longer time than usual and
> > with the side effect that a deep(er) copy was made.
> >
> > I guess if someone asks for both (1) and (3) we can do the fallback in the
> > kernel, like how we handle it right now.
> >
> 
> I think we should focus on what the actual legit use cases might be.
> Certainly we want to support a mode that's "reflink or fail".  We
> could have these flags:
> 
> COPY_FILE_RANGE_ALLOW_REFLINK
> COPY_FILE_RANGE_ALLOW_COPY
> 
> Setting neither gets -EINVAL.  Setting both works as is.  Setting just
> ALLOW_REFLINK will fail if a reflink can't be supported.  Setting just
> ALLOW_COPY will make a best-effort attempt not to reflink but
> expressly permits reflinking in cases where either (a) plain old
> write(2) might also result in a reflink or (b) there is no advantage
> to not reflinking.

I don't agree with having a 'copy' flag that can reflink when we also have a
'reflink' flag.  I guess I just don't like having a flag with different
meanings depending on context.

Users should be able to get the default behavior by passing '0' for flags, so
provide FORBID_REFLINK and FORBID_COPY flags to turn off those behaviors, with
an admonishment that one should only use them if they have a goooood reason.
Passing neither gets you reflink-xor-copy, which is what I think we both want
in the general case.

FORBID_REFLINK = 1
FORBID_COPY = 2
CHECK_SAME = 4
HW_COPY = 8

DEDUPE = (FORBID_COPY | CHECK_SAME)

What do you say to that?

> An example of (b) would be a filesystem backed by deduped
> thinly-provisioned storage that can't do anything about ENOSPC because
> it doesn't control it in the first place.
> 
> Another option would be to split up the copy case into "I expect to
> overwrite a lot of the target file soon, so (c) try to commit space
> for that or (d) try to make it time-efficient".  Of course, (d) is
> irrelevant on filesystems with no random access (nvdimms, for
> example).
> 
> I guess the tl;dr is that I'm highly skeptical of any use for
> disallowing reflinking other than forcibly committing space in cases
> where committing space actually means something.

That's more or less where I was going too. :)

--D