Re: [PATCH RERESEND 00/11] splice(file<>pipe) I/O on file as-if O_NONBLOCK

[Ahelenia Ziemiańska]

On Thu, Dec 14, 2023 at 12:06:57PM -0700, Jens Axboe wrote:
> On 12/14/23 11:44 AM, Ahelenia Ziemiańska wrote:
> > This does that, effectively making splice(file -> pipe)
> > request (and require) O_NONBLOCK on reads fron the file:
> > this doesn't affect splicing from regular files and blockdevs,
> > since they're always non-blocking
> > (and requesting the stronger "no kernel sleep" IOCB_NOWAIT is non-sensical),
> Not sure how you got the idea that regular files or block devices is
> always non-blocking, this is certainly not true without IOCB_NOWAIT.
> Without IOCB_NOWAIT, you can certainly be waiting for previous IO to
> complete.
Maybe "always non-blocking" is an abuse of the term, but the terminology
is lost on me. By this I mean that O_NONBLOCK files/blockdevs have the
same semantics as non-O_NONBLOCK files/blockdevs ‒ they may block for a
bit while the I/O queue drains, but are guaranteed to complete within
a relatively narrow bounded time; any contending writer/opener
will be blocked for a short bit but will always wake up.

This is in contrast to pipes/sockets/ttys/&c., which wait for a peer to
send some data, and block until there is some; any contending writer/opener
will be blocked potentially ad infinitum.

Or, the way I see it, splice(socket -> pipe) can trivially be used to
lock the pipe forever, whereas I don't think splice(regfile -> pipe) can,
regardless of IOCB_NOWAIT, so the specific semantic IOCB_NOWAIT provides
is immaterial here, so not specifying IOCB_NOWAIT in filemap_splice_read()
provides semantics consistent to "file is read as-if it had O_NONBLOCK set".

> > but always returns -EINVAL for ttys.
> > Sockets behave as expected from O_NONBLOCK reads:
> > splice if there's data available else -EAGAIN.
> > 
> > This should all pretty much behave as-expected.
> Should it? Seems like there's a very high risk of breaking existing use
> cases here.
If something wants to splice from a socket to a pipe and doesn't
degrade to read/write if it gets EAGAIN then it will either retry and
hotloop in the splice or error out, yeah.

I don't think this is surmountable.

> Have you at all looked into the approach of enabling splice to/from
> _without_ holding the pipe lock? That, to me, would seem like a much
> saner approach, with the caveat that I have not looked into that at all
> so there may indeed be reasons why this is not feasible.
IIUC Linus prepared a patch on security@ in
  <CAHk-=whPmrWvXBqcK6ey_mnd-0fz_HNUHEfz3NX97mqoCCcwtA@xxxxxxxxxxxxxx>
(you're in To:) and an evolution of this is in
  https://lore.kernel.org/lkml/CAHk-=wgmLd78uSLU9A9NspXyTM9s6C23OVDiN2YjA-d8_S0zRg@xxxxxxxxxxxxxx/t/#u
(you're in Cc:) that does this.

He summarises it below as 
> So while fixing your NULL pointer check should be trivial, I think
> that first patch is actually fundamentally broken wrt pipe resizing,
> and I see no really sane way to fix it. We could add a new lock just
> for that, but I don't think it's worth it.
and
> But it is possible that we need to just bite the bullet and say
> "copy_splice_read() needs to use a non-blocking kiocb for the IO".
so that's what I did.

If Linus, who drew up and maintained this code for ~30 years,
didn't arrive at a satisfactory approach, I, after ~30 minutes,
won't either.

It would be very sane to both not change the semantic and fix the lock
by just not locking but at the top of that thread Christian said
> Splice would have to be
> refactored to not rely on pipe_lock(). That's likely major work with a
> good portion of regressions if the past is any indication.
and clearly this is true, so lacking the ability and capability
to do that and any reasonable other ideas
(You could either limit splices to a proportion of the pipe size,
 but then just doing five splices gets you where we are rn
 (or, as Linus construed it, into "write() returns -EBUSY" territory,
  which is just as bad but at least the writers aren't unkillable),
 and it reduces the I/O per splice significantly.

 You could limit each pipe to one outstanding splice and always leave
 Some space for normal I/O. This falls into "another lock just for this"
 territory I think, and it also sucks for the 99% of normal users.)
I did this because Linus vaguely sanxioned it.

It's probably feasible, but alas it isn't feasible for me.

Attachment: signature.asc
Description: PGP signature