qemu-system-sparc loses serial console data on EAGAIN

[Andreas Gustafsson]

When running a guest OS with a serial console under
"qemu-system-sparc -nographic", parts of the serial console output
are sometimes lost.

This happens when a write() to standard output by qemu returns EAGAIN,
as may be the case when the guest is generating console output faster
than the tty (or pty/pipe/socket, etc.) connected to qemu's standard
output accepts it. The bug affects all releases of qemu since 1.5,
which was the first version to set stdout to O_NONBLOCK mode. Version
1.4.2 and earlier work correctly.

To reproduce the bug, you will need a guest OS configured with a
serial console, and a host with a slow tty. The attached shell script
"sparc-test.sh" does this by using Aboriginal Linux as the serial
console guest, and a pty controlled by a Python script and the
"pexpect" Python module as the slow tty. A "seq" command is sent
to the guest to generate 100,000 lines of output containing sequential
integers, and the output is checked for gaps. The script limits the
tty output rate by occasionally sleeping for 1/10 of a second.

This bug was originally reported against qemu-system-i386 as
bug #1335444, and has since been fixed in qemu-system-i386,
but remains in qemu-system-sparc as of today's git sources
(d00e6cddc220de993573dfb5fd160ac72ccd49ab). I am opening
this separate bug for the sparc case because I was asked
to do so by Paolo Bonzini in #1335444.

[Andreas Gustafsson (gson)]

The automated test infrastructure of the NetBSD project is based on
qemu, and runs some 100 CPU-hours per day of full system tests of
NetBSD-current on emulated i386, amd64, and sparc systems.

This is all still running on qemu 0.15 (!). The main obstacle to
upgrading to a current version of qemu is this three-year-old regression
which breaks the sparc tests by losing some of the serial console output.

It would be really nice if this could be fixed.

[Mark Cave-Ayland (mark-cave-ayland)]

Hi Andreas,

I've just done a local test using git master qemu-system-sparc here and your test script passes (i.e. reaches 100000 iterations) without any failures locally. I know that Daniel has done a lot of work in this area in the past 2 releases so perhaps this has fixed the issue. Please can you test locally with git master and report back?

Many thanks,

Mark.

[Andreas Gustafsson (gson)]

Hi Mark,

I tested the git master and the 2.7.0 release, and both successfully executed the scripted NetBSD/sparc install that had been failing since 1.5.

So it does indeed look like the bug has been fixed. I will still run a few more tests to be sure, and if those also pass, this bug report can be closed. Thank you.

[Andreas Gustafsson (gson)]

I have now run some more tests, and I'm still seeing occasional failures in the scripted NetBSD installs that look like console data loss when using qemu 2.7.0, with maybe one in ten installs failing. I have seen no failures so far using the git master. So it looks like the bug is fixed in git, but not yet in 2.7.0.

[Mark Cave-Ayland (mark-cave-ayland)]

Hi Andreas,

Does the new 2.8 release pass your tests? If so, I will close this bug report.

ATB,

Mark.

[Andreas Gustafsson (gson)]

Hi Mark,

I have now upgraded to qemu 2.8 and successfully run more than 20 scripted installs of NetBSD/sparc over a serial console without failures, so it does indeed look like the bug is now fixed, and the bug report can be closed.

[Mark Cave-Ayland (mark-cave-ayland)]

Fix confirmed in QEMU 2.8 release.