DNS activity in slirp (user networking) mode quickly depletes file descriptors and crashes qemu
Affects | Status | Importance | Assigned to | Milestone | |
---|---|---|---|---|---|
QEMU |
Expired
|
Undecided
|
Unassigned |
Bug Description
Hi, we have encountered quite some trouble with filedescriptor depletion of the qemu process. We have figured out that it can be demonstrated easily by doing a lot of DNS queries inside the VM -- in our real world scenario this is caused by running centos network install with a fast mirror.
This situation is further problematic because qemu can't handle fd depletion very well:
1) if ulimit is 1024 then qemu hangs in infinite loop whenever it tries to open the 1025th fd
2) setting ulimit >1024 does not help that much because qemu uses select and max. fd set size is 1024 per default => qemu crashes because of buffer overflow in select()
3) setting ulimit > 1024 AND recompiling with large enough fd set size AND disabling gcc's fortify source seems to work, but that's really just a hot-fix
The problem can be replicated quite easily by running something like
while :; do echo >/dev/udp/
inside a Linux VM -- crash comes very soon.
This problem is present in current qemu (1.2.0) and in earlier as well.
Triaging old bug tickets ... can you still reproduce this problem with the latest version of QEMU (currently v2.9 or a release candidate of 2.10)?