Re: Pending splice(file -> FIFO) excludes all other FIFO operations forever (was: ... always blocks read(FIFO), regardless of O_NONBLOCK on read side?)

[Ahelenia Ziemiańska]

On Mon, Jul 10, 2023 at 12:33:07AM +0200, Ahelenia Ziemiańska wrote:
> On Sun, Jul 09, 2023 at 03:03:22AM +0200, Ahelenia Ziemiańska wrote:
> > On Sat, Jul 08, 2023 at 01:06:56PM -0700, Linus Torvalds wrote:
> > > I guess combined with something like
> > > 
> > >         if (!(in->f_mode & FMODE_NOWAIT))
> > >                 return -EINVAL;
> > > 
> > > it might all work.
> > Yes, that makes the splice instantly -EINVAL for ttys, but doesn't
> > affect the socketpair() case above, which still blocks forever.
> This also triggers for regular file -> pipe splices,
> which is probably a no-go.
Actually, that's only the case for regular files on some filesystems?
I originally tested on tmpfs, and now on vfat, ramfs, procfs, and sysfs,
and none have FMODE_NOWAIT set.

Conversely, ext4 and XFS files both have FMODE_NOWAIT set,
and behave like blockdevs incl. the filemap_splice_read() oddities below.

Indeed, it looks like Some filesystems
(btrfs/ext4/f2fs/ocfs2/xfs, blockdevs, /dev/{null,zero,random,urandom},
 pipes, tun/tap)
set FMODE_NOWAIT, but that's by far not All of them, so maybe
  /* File is capable of returning -EAGAIN if I/O will block */
is not the right check for regular files.

> filemap_get_pages() does use and inspect IOCB_NOWAIT if set in
> filemap_splice_read(), but it appears to not really make much sense,
> inasmuch it returns EAGAIN for the first splice() from a
> blockdev and then instantly return with data on the next go-around.
Indeed, this is inconsistent to both:
  * readv2(off=-1, RWF_NOWAIT), which always returns EAGAIN, and
  * fcntl(0, F_SETFL, O_NONBLOCK), read(), which always reads.

Thus,
> This doesn't really make much sense ‒ and open(2) says blockdevs
> don't have O_NONBLOCK semantics, so I'm assuming this is not supposed
> to be exposed to userspace ‒ so I'm not setting it in the diff.
not specifying IOCB_NOWAIT in filemap_splice_read() provides consistent
semantics to "file is read as-if it had O_NONBLOCK set".


I've tentatively updated the check to
		if (!((in->f_mode & FMODE_NOWAIT) || S_ISREG(in->f_inode->i_mode)))
(with the reasoning, as previously, that regular files don't /have/ a
 distinct O_NONBLOCK mode, because they always behave non-blockingly)
and with that
> I've tested this for:
  * regular file: works as expected

Attachment: signature.asc
Description: PGP signature