On 24/03/11 19:43, Ulf Magnusson wrote:
> On Thu, Mar 24, 2011 at 7:06 PM, Ulf Magnusson <ulfali...@gmail.com> wrote:
>> Hi,
>>
>> My DirectFB application kept hanging on shutdown whenever I used the
>> Linux Input driver, so I did some investigation to figure out why.
>> Here's what happens:
>>
>> 1. On shutdown, the threads processing /dev/event/X are all
>> pthread_cancel()'ed.
>> 2. For some reason[1] one or more of the threads reach the D_PERROR
>> ("linux_input thread died\n") at the end of linux_input_EventThread().
>> 3. One of the threads acquires the log->lock mutex in
>> direct_log_printf(), writes to log->fd, and then dies before releasing
>> the lock as write() is a cancellation point.
>> 4. The next thread that tries to write to the log gets stuck on log->lock.
>>
>> Here's my proposed fix, which temporarily changes the cancellation
>> state of the thread inside direct_log_printf() to prevent it from
>> being canceled while holding the lock (I'm by means a Pthreads guru,
>> so there might very well be a better solution):
>>
>> --- a/lib/direct/log.c 2010-11-15 22:12:08.000000000 +0100
>> +++ b/lib/direct/log.c 2011-03-24 17:58:38.259808355 +0100
>> @@ -167,14 +167,19 @@
>> else {
>> int len;
>> char buf[512];
>> + int old_cancellation_state;
>>
>> len = vsnprintf( buf, sizeof(buf), format, args );
>>
>> - pthread_mutex_lock( &log->lock );
>> + /* Ensure the thread does not get canceled at the write(), which
>> + * would prevent the log lock from being released. */
>> + pthread_setcancelstate( PTHREAD_CANCEL_DISABLE,
>> &old_cancellation_state );
>>
>> + pthread_mutex_lock( &log->lock );
>> write( log->fd, buf, len );
>> -
>> pthread_mutex_unlock( &log->lock );
>> +
>> + pthread_setcancelstate( old_cancellation_state, NULL );
>> }
>>
>> va_end( args );
>>
>> With the above patch my application no longer hangs on shutdown.
>>
>> You probably also ought to make sure a thread can never die inside a
>> direct_log_lock()/direct_log_unlock() pair.
>>
>> [1] This seems to be due to a uClibc bug that causes the select() on
>> linux_input.c:902 (DirectFB 1.4.11) to return -1 while errno remains
>> 0. Seems to depend on subtle details in how the application was
>> compiled.
>
> To clarify: In this case the error was due to uClibc, but the same
> thing would happen any time a thread is pthread_cancel()'ed and then
> writes to the log, which seems like a bug.
Thanks, we'll integrate the changes!
Not sure if we should prevent threads from logging when they are cancelled.
--
Best regards,
Denis Oliver Kropp
.------------------------------------------.
| DirectFB - Hardware accelerated graphics |
| http://www.directfb.org/ |
"------------------------------------------"
_______________________________________________
directfb-dev mailing list
directfb-dev@directfb.org
http://mail.directfb.org/cgi-bin/mailman/listinfo/directfb-dev
