l?rdag 06 august 2005, 21:08, skrev Johannes Stezenbach: > Kenneth Aafl?y wrote: > > I've been searching for the reason why I was _loosing_ filters > > in the demuxer, and I've found out the subtle reason. > > > > In dmxdev.c, the dvb_dmxdev_release is calling dvb_dmxdev_filter_free. > > This function uses down_interruptible to wait for two mutexes, but > > if those are interrupted by a signal, the filter will be _lost_. > > This is easily reproducible with a program that creates a few threads, > > which open and close the filters regulary, and then running/interrupting > > this program some number of times. > > > > Is changing all down_interruptible (that protects freeing a structure) > > to down(), or will this have other unseen side-effects? > > Generally speaking, a driver should only sleep uninterruptibly > if you can make absolutely sure it won't sleep forever, even > in an obscure error case. If you can guarantee that from > reading the code, then it's best to change the down() to > down_interruptible(), otherwise it's better to handle the > error from down_interruptible() correctly. > > (If a process is stuck in uninterruptible sleep you usually > have to reboot to get rid of it, which is bad.) Ok, but do you think it is better interrupt freeing a resource that can't be reclaimed later, or _maybe_ break some other part because of deadlocks? Kenneth
