On 06/03, David Rientjes wrote: > > On Fri, 4 Jun 2010, Oleg Nesterov wrote: > > > > > > > Currently, PF_EXITING check is completely broken. because 1) It only > > > > > > care main-thread and ignore sub-threads > > > > > > > > > > Then check the subthreads. > > > > > > > > > > > Did you want to respond to this? > > > > Please explain what you mean. There were already a lot of discussions > > about mt issues, I do not know what you have in mind. > > Can you check the subthreads to see if they are not PF_EXITING? To detect the process with the dead group leader? Yes, we can. We already discussed this. Probably it is better to check PF_EXITING and signal_group_exit(). > > > I'm guessing at the relevancy here because the changelog is extremely > > > poorly worded (if I were Andrew I would have no idea how important this > > > patch is based on the description other than the alarmist words of "... is > > > completely broken)", but if we're concerned about the coredumper not being > > > able to find adequate resources to allocate memory from, we can give it > > > access to reserves specifically, > > > > I don't think so. If oom-kill wants to kill the task which dumps the > > code, it should stop the coredumping and exit. > > That's a coredump change, not an oom killer change. Yes. do_coredump() should be fixed. This is not trivial (and needs the subtle changes outside of fs/exec.c), we are looking for the simple fix for now. > If the coredumper > needs memory and runs into the oom killer, this PF_EXITING check, which > you want to remove, gives it access to memory reserves by setting > TIF_MEMDIE so it can quickly finish and die. This allows it to exit > without oom killing anything else because the tasklist scan in the oom > killer is not preempted by finding a TIF_MEMDIE task. David, sorry. I already tried to explain (at least twice) that TIF_MEMDIE (or SIGKILL even if do_coredump() was interruptible) can not help unless you find the right thread, this is not trivial even if we forget about CLONE_VM tasks. And personally I disagree that it should use memory reserves, but this doesn't matter. Let's stop this. You shouldn't convince me. I am not the author of this patch, and I said many times that I do not pretend I understand oom-kill needs. I jumped into this discussion because your initial objection (fatal_signal_pending() should fix the problems) was technically wrong. Oleg. -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxxx For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>