Features/ChardevFlowControl

Summary

The chardev layer in QEMU did not have any mechanism for flow-control. This meant that a fast guest could pump data to the host (via virtio-serial, for example), and if the host-side reader was slow to consume data, or didn't consume data at all, the guest would freeze. This is resolved by converting the char layer to use the GLib IO loop.

Status

Patches are still under review and discussion, but intended to be merged for the 1.5 release. The main code exists in Anthony's 'char-flow' branches (latest is v3): https://github.com/aliguori/qemu/commits/char-flow.3.

Testing

To reproduce a guest freeze on releases without the patches for this feature, invoke QEMU and perform the tests as noted in the two cases below. The examples use the unix backend, but other backends, like udp, tcp, pty exhibit similar behaviour, and should be tested as well.

Start QEMU with UNIX chardev in server mode

./x86_64-softmmu/qemu-system-x86_64 -m 512 \
 /guests/f14-nolvm.qcow2 -snapshot -enable-kvm   \
 -chardev socket,path=/tmp/foo,server,nowait,id=foo \
 -device virtio-serial -device virtserialport,chardev=foo,id=test0,nr=2 \
 -monitor stdio

On the host, open the chardev but don't read from it, e.g.

$ python
Python 2.7.3 (default, Aug  9 2012, 17:23:57) 
[GCC 4.7.1 20120720 (Red Hat 4.7.1-5)] on linux2
Type "help", "copyright", "credits" or "license" for more information.
>>> import socket
>>> 
>>> sock = socket.socket(socket.AF_UNIX)
>>> sock.connect("/tmp/foo")
>>> 

and in the guest, write to the virtio-serial port, e.g.

# dd if=/dev/zero of=/dev/virtio-ports/test0

Start QEMU with UNIX chardev in client mode

./x86_64-softmmu/qemu-system-x86_64 -m 512 /guests/f14-nolvm.qcow2 \
  -snapshot -enable-kvm   -chardev socket,path=/tmp/foo,id=foo \
  -device virtio-serial -device virtserialport,chardev=foo,id=test0,nr=2 \
  -monitor stdio

on the host, start a listening process, like:

$ socat UNIX-LISTEN:/tmp/foo  -

and in the guest,

# echo foo > /dev/virtio-ports/test0

Named pipe backend

• Create named pipe

  [host]# mkfifo /tmp/serial0

• Boot up a single-CPU guest with a virtio-serial device and named-pipe chardev backend

  [host]# qemu-kvm -enable-kvm -m 1024 -nographic \
            -device virtio-serial-pci -chardev pipe,id=ch0,path=/tmp/serial0 
            -device virtserialport,chardev=ch0,name=serial0 \
            -drive file=disk.img,if=virtio,boot=on

• Run a background work on the guest

  [guest]# while :;do sleep 1; date; done &

• Write lots of data to the virtio-serial port; the guest should not freeze.

  [guest]# cat /proc/kallsyms > /dev/virtio-ports/serial0

• Confirm the guest does not freeze even if there's no reader process in the host

• Read the named-pipe file on the host

  [host]# cat /tmp/serial0

Known bugs

These are the bugs in the implementation known so far:

• QEMU doesn't catch -EPIPE for short-lived guest processes and disconnected host chardevs. This is also described in https://bugzilla.redhat.com/show_bug.cgi?id=621484. To reproduce this issue, use the following steps:

1. start qemu, boot guest
2. connect netcat to the unix socket
3. echo some chars to /dev/virtio-ports/$name
4. watch them appear in netcat
5. stop netcat, restart it
6. echo some chars again
7. watch them *not* appear

Gerd's open-write-close test:

for seq in $(seq 1 100); do echo "test $seq" > /dev/virtio-ports/org.kraxel.test; sleep 1; done

Gerd's keep-device-open test:

(for seq in $(seq 1 100); do echo "test $seq"; sleep 1; done) > /dev/virtio-ports/org.kraxel.test

The bug triggers only with open-write-close.