Re: segmentation fault in qemu-kvm-0.14.0

[Jan Kiszka <jan.kiszka@web.de>]

[-- Attachment #1: Type: text/plain, Size: 2069 bytes --]

On 2011-03-09 11:16, Peter Lieven wrote:
> 
> Am 09.03.2011 um 08:37 schrieb Jan Kiszka:
> 
>> On 2011-03-08 23:53, Peter Lieven wrote:
>>> Hi,
>>>
>>> during testing of qemu-kvm-0.14.0 i can reproduce the following segfault. i have seen similar crash already in 0.13.0, but had no time to debug.
>>> my guess is that this segfault is related to the threaded vnc server which was introduced in qemu 0.13.0. the bug is only triggerable if a vnc
>>> client is attached. it might also be connected to a resolution change in the guest. i have a backtrace attached. the debugger is still running if someone
>>> needs more output
>>>
>>
>> ...
>>
>>> Thread 1 (Thread 0x7ffff7ff0700 (LWP 29038)):
>>> #0  0x0000000000000000 in ?? ()
>>> No symbol table info available.
>>> #1  0x000000000041d669 in main_loop_wait (nonblocking=0)
>>>    at /usr/src/qemu-kvm-0.14.0/vl.c:1388
>>
>> So we are calling a IOHandlerRecord::fd_write handler that is NULL.
>> Looking at qemu_set_fd_handler2, this may happen if that function is
>> called for an existing io-handler entry with non-NULL write handler,
>> passing a NULL write and a non-NULL read handler. And all this without
>> the global mutex held.
>>
>> And there are actually calls in vnc_client_write_plain and
>> vnc_client_write_locked (in contrast to vnc_write) that may generate
>> this pattern. It's probably worth validating that the iothread lock is
>> always held when qemu_set_fd_handler2 is invoked to confirm this race
>> theory, adding something like
>>
>> assert(pthread_mutex_trylock(&qemu_mutex) != 0);
>> (that's for qemu-kvm only)
> 
> qemu_mutex doesn't exists (anymore). i can only find qemu_global_mutex and qemu_fair_mutex.

In qemu-kvm, qemu-kvm.c contains that lock. In upstream, it's
qemu_global_mutex, but you will have to enable --enable-io-thread and
export it.

> 
> i try with qemu_global_mutex now. if thats not correct please tell.

I think we knokw know the call path, so you could also wait for Corentin
to send a patch and test that one.

Jan


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 259 bytes --]