On Sunday, October 17, 2021 12:29:02 PM CEST Phillip Potter wrote: > On Sat, Oct 16, 2021 at 08:53:15PM +0200, Fabio M. De Francesco wrote: > > On Saturday, October 16, 2021 8:13:43 PM CEST Martin Kaiser wrote: > > > At the moment, our command thread can be killed by user space. > > > > > > [root@host ]# kill `pidof RTW_CMD_THREAD` > > > > > > The driver will then stop working until the module is unloaded > > > and reloaded. > > > > > > Don't process SIGTERM in the command thread. Other drivers that have a > > > command thread don't process SIGTERM either. > > > > Hi Martin, > > > > This is _really_ interesting :) > > > > May be that you have had time to read my last email in reply to a message of > > Phillip P. Soon after writing of the arguments in favor of using > > wait_for_completion_killable() (in patch 2/3 of the series I sent today), I > > read your patch. > > > > If you are right (and I think you are) I'll have to send a v2 that replaces > > the killable wait with an uninterruptible one. > > > > Unfortunately I have not the needed experience to decide whether or not to > > ack your patch, even if I'm strongly tempted to do it. > > > > Let's wait for more experienced people. > > > > Thanks, > > > > Fabio > > > > So I myself am a little confused on this one :-) > > Based on my understanding, so correct me if I'm wrong, a process (kthread or > otherwise) can still be killed if marked TASK_KILLABLE, even if ignoring > SIGTERM. Indeed, from a userspace perspective, SIGKILL is unblockable > anyway - although of course kernel code can choose how to respond to it. Correct. > > So in other words, the kthread could still be killed while waiting > in the wait_for_completion_killable() call, even if we are ignoring > SIGTERM. From that perspective I guess, it is therefore not 'incorrect' as > such - if indeed we wanted that behaviour. No. This misunderstandings is my fault. :( In Martin's patch I read "SIGTERM" but for some reason I thought he was talking of "SIGKILL". At the moment, without Martin's patch, the kthread can be terminated by the command "kill -TERM <PID>". If we try "kill -KILL <PID>", nothing happens. This is because only "allow_signal(SIGTERM);" is present in the code. I think that kthreads must also allow SIGKILL with "allow_signal(SIGKILL);" for allowing root to make them terminate. For what relates to my patch, it doesn't matter if I either leave wait_for_completion_killable() as-is or change it to wait_for_completion(). This is because at the moment SIGKILL cannot kill rtw_cmd_thread(), while SIGTERM can. However, for consistency, I should better change it to the uninterruptible version. @Martin: Please correct me if I'm wrong. Thanks. Regards, Fabio > > That said, killing it would still cause the behaviour Martin mentions - > I guess we don't want it to be either killable or interruptible based on > that logic? > > Regards, > Phil >
