[Qt-interest] SIGSEGV with single-line gdb backtrace

[Josiah Bryan]

Bill King wrote:
> Bill - thanks for the quick response.
>   
>> helgrind isn't working (according to valgrind) in my current version of 
>> valgrind. memcheck reports various lost memory, but no help finding a 
>> SIGSEGV. In fact, the SIGSEGV didn't even occur during any valgrind 
>> sessions I just ran, but the program *did* run very very slow. Which 
>> leads me to believe that the SIGSEGV is being caused by some wierd 
>> thread interaction.
>>
>> But how to you debug *that* beast?
>>
>> Thanks,
>> -josiah
>>
>>   
>>     
> The dreaded heisenbugs :) Update your valgrind if possible. There's a
> good chance that helgrind will catch it, now... Are you locking access
> to the variables that are likely to be accessed via another thread?
> Definately sounds like a race condition.
>   
Alright, I narrowed it down using the good old-fashioned 
"comment-out-lines-until-it-stops-crashing" method - and narrowed it 
down to an obscure place in the destructor of my video decoder QThread: 
At one point, I free all the libav stuff allocated, and one of the lines 
does this:

av_free(m_av_frame);

Where m_av_frame was simply allocated earlier by "m_av_frame = 
avcodec_alloc_frame();"

So, bottom line, somehow multithreading libavcodec et al is crashing a 
QThread. The odd thing is, I've got two other calls to av_free for other 
structures right above the free on m_av_frame - all are member variables 
of the thread, none should be accessed by any other thread. Access to 
the video frame is thru a mutex-locked buffer - the frame is converted 
from the video's internal format to an rgb frame, a QImage allocated and 
data copied into its bits(), and stuffed into the buffer with a mutex 
lock. Any other thread (GUI thread) that wants a frame uses the buffer, 
not even touching the decoder thread itself.

How, then, is this rotten piece of code causing a race condition and a 
bloody SIGSEGV?

(BTW, I'm trying to update valgrind now - recompiling now. All the above 
ranting was done based on source tracing. Any ideas from that? Or do I 
still need to dredge up valgrind?)

Thanks,
-josiah