Re: [PATCH v2] Introduce a version6 of autofs interface, to fix design error.

From: Linus Torvalds <torvalds@linux-foundation.org>
To: Michael Tokarev <mjt@tls.msk.ru>
Cc: Alan Cox <alan@lxorguk.ukuu.org.uk>,
	Thomas Meyer <thomas@m3y3r.de>,
	David Miller <davem@davemloft.net>,
	linux-kernel@vger.kernel.org, autofs@vger.kernel.org,
	raven@themaw.net, stable@kernel.org
Subject: Re: [PATCH v2] Introduce a version6 of autofs interface, to fix design error.
Date: Sun, 29 Apr 2012 00:19:50 -0700	[thread overview]
Message-ID: <CA+55aFyS9VYT50JAXeGrLrtFVWSGr+DMTA1hhrcPYzmjrSV5+A@mail.gmail.com> (raw)
In-Reply-To: <4F9CE18C.2060806@msgid.tls.msk.ru>

On Sat, Apr 28, 2012 at 11:37 PM, Michael Tokarev <mjt@tls.msk.ru> wrote:
>
> And no, it does not quite work.  Strace shows that systemd correctly
> reads first 300 bytes, but next read returns 4 bytes, so it reads
> these and waits for next 300-4 = 296 bytes of data.
>
> /proc/1/fdinfo/16  -- the autofs pipe fd# -- shows flags=02004000
> which is O_CLOEXEC|O_NONBLOCK, but this is the other end of the
> pipe..  shouldn't the READ side of the pipe have O_DIRECT flag now?

Gaah, it should, but it won't.

I bet my original patch worked fine, because the pipe has only one
inode and pipe structure. But it has *two* 'struct file's associated
with it, and autofs only ever sees the writing side, and never gets to
mark the reading side O_DIRECT. So yeah, the reading side won't do the
proper packetized read.

Duh. So we need to put the flag back in the pipe_inode_info. Too bad -
because the O_DIRECT approach was so nice in other ways.

                     Linus