Re: regression, bisected: openpty fails from 3.7 onwards without devpts

[Jiri Slaby <jslaby@suse.cz>]

On 01/10/2013 05:29 PM, Florian Westphal wrote:
> Alan Cox <alan@linux.intel.com> wrote:
>> > On Thu, 10 Jan 2013 15:46:26 +0100
>> > Florian Westphal <fw@strlen.de> wrote:
>>> > > Frank Lichtenheld discovered that openpty() doesn't work anymore when
>>> > > /dev/pts is not present.
>>> > > 
>>> > > We bisected this down to
>>> > > 
>>> > > commit bbb63c514a3464342967237a51a21ea8f61ab951
>>> > > Author: Wanlong Gao <gaowanlong@cn.fujitsu.com>
>>> > > Subject: drivers:tty:fix up ENOIOCTLCMD error handling
> [..]
> 
>>> > > #include <stdio.h>
>>> > > #include <pty.h>
>>> > > int main(void) {
>>> > >  int pty_fd, tty_fd;
>>> > >  if (openpty(&pty_fd, &tty_fd, NULL, NULL, NULL) != 0) {
>>> > >   perror("openpty");
>>> > >   return 1;
>>> > >  }
>>> > > return 0;
>>> > > }
>>> > > ----
>>> > > [ compile with cc -lutil pty.c -o pty ]
>>> > > 
>>> > > If devpts is available or above commit reverted openpty works again.
>> > 
>> > The commit is fairly general - what we need to do here is to figure out
>> > which specific thing trips up openpty so we can put the error on that
>> > back as it was (or find a better way) so it still works.
>> > 
>> > Can you attach an strace of the working/failing cases without /dev/pts 
> Sure, attached.  /dev/pts is not present.
> Both traces are from the same machine, with same kernel version (except
> above commit reverted).

> execve("./ptytest", ["./ptytest"], [/* 7 vars */]) = 0
...
> open("/dev/ptmx", O_RDWR)         = -1 ENOSPC (No space left on device)
> open("/dev/ptyp0", O_RDWR)        = 3
> ioctl(3, SNDCTL_TMR_TIMEBASE or TCGETS, {B38400 opost isig icanon echo ...}) = 0
> ioctl(3, TIOCGPTN, [0])           = -1 ENOTTY (Inappropriate ioctl for device)


getptsname expects EINVAL on failure to fall back to /dev/ttyp*... The
same as unlockpt. We should definitely revert now and can teach glibc to
accept also ENOTTY. After some years, we can try again :).

thanks,
-- 
js
suse labs