Prev: oom: give current access to memory reserves if it has been killed
Next: oom: filter tasks not sharing the same cpuset
From: Oleg Nesterov on 8 Jun 2010 14:50 On 06/08, KOSAKI Motohiro wrote: > > Oleg pointed out oom_kill.c has force_sig() abuse. force_sig() mean > ignore signal mask. but SIGKILL itself is not maskable. Yes. And we have other reasons to avoid force_sig(). It should be used only for synchronous signals. But, > @@ -399,7 +399,7 @@ static int __oom_kill_process(struct task_struct *p, struct mem_cgroup *mem) > p->rt.time_slice = HZ; > set_tsk_thread_flag(p, TIF_MEMDIE); > > - force_sig(SIGKILL, p); > + send_sig(SIGKILL, p, 1); This is not right, we need send_sig(SIGKILL, p, 0). Better yet, send_sig_info(SIGKILL, SEND_SIG_NOINFO). I think send_sig() should die. The reason is that si_fromuser() must be true, otherwise we can't kill the SIGNAL_UNKILLABLE (sub-namespace inits) tasks. Oh. This reminds me, we really need the trivial (but annoying) cleanups here. The usage of SEND_SIG_ constants is messy, and they should be renamed at least. And in fact, we need the new one which acts like SEND_SIG_FORCED but si_fromuser(). We do not want to allocate the memory when the caller is oom_kill or zap_pid_ns_processes(). OK. I'll send the simple patch which adds the new helper with the comment. send_sigkill() or kernel_kill_task(), or do you see a better name? Oleg. -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo(a)vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/
From: KOSAKI Motohiro on 13 Jun 2010 07:30
> On 06/08, KOSAKI Motohiro wrote: > > > > Oleg pointed out oom_kill.c has force_sig() abuse. force_sig() mean > > ignore signal mask. but SIGKILL itself is not maskable. > > Yes. And we have other reasons to avoid force_sig(). It should be used > only for synchronous signals. > > But, > > > @@ -399,7 +399,7 @@ static int __oom_kill_process(struct task_struct *p, struct mem_cgroup *mem) > > p->rt.time_slice = HZ; > > set_tsk_thread_flag(p, TIF_MEMDIE); > > > > - force_sig(SIGKILL, p); > > + send_sig(SIGKILL, p, 1); > > This is not right, we need send_sig(SIGKILL, p, 0). Better yet, > send_sig_info(SIGKILL, SEND_SIG_NOINFO). I think send_sig() should > die. > > The reason is that si_fromuser() must be true, otherwise we can't kill > the SIGNAL_UNKILLABLE (sub-namespace inits) tasks. Thanks. I am not signal expert. To be honest, current special siginfo arguments have a bit unclear meanings to me ;) current definition (following) doesn't teach anything. sched.h ===================== /* These can be the second arg to send_sig_info/send_group_sig_info. */ #define SEND_SIG_NOINFO ((struct siginfo *) 0) #define SEND_SIG_PRIV ((struct siginfo *) 1) #define SEND_SIG_FORCED ((struct siginfo *) 2) If anyone write exact meanings, I'm really really glad. > Oh. This reminds me, we really need the trivial (but annoying) cleanups > here. The usage of SEND_SIG_ constants is messy, and they should be > renamed at least. > > And in fact, we need the new one which acts like SEND_SIG_FORCED but > si_fromuser(). We do not want to allocate the memory when the caller > is oom_kill or zap_pid_ns_processes(). > > OK. I'll send the simple patch which adds the new helper with the > comment. send_sigkill() or kernel_kill_task(), or do you see a > better name? Very thanks. both name are pretty good to me. -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo(a)vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/ |