On 11/13, Michal Hocko wrote: > > > > > > > > > To fix this, make de_thread() freezable. It looks safe and works fine. > > > > > > It's been some time since I have looked into this code so bear with me. > > > One thing is not really clear to me. Why does it help to exclude this > > > particular task from the freezer > > > > we don't exclude it, > > > > > when it is not sleeping in the freezer. > > > > Yes, it is not sleeping in __refrigerator(), but it does > > > > schedule(); > > freezer_count(); > > > > so it will enter __refrigerator() right after wakeup. If it won't be woken > > up we do not care, we can consider it "frozen". > > Right, but this is just silencing the freezing code to exclude this > task, right? Well yes... but I'd say this tells the freezing code that the caller is frozen, because it can do nothing till thaw_processes(). Except it can actually call __refrigerator() if, say, it is killed. > > > I can see how other threads need to be zapped and TASK_WAKEKILL doesn't > > > do that but shouldn't we fix that instead? > > > > Not sure I understand, but unlikely we can (or want) to make __refrigerator() > > killable. > > Why would that be a problem. If the kill is fatal then why to keep the > killed task in the fridge? This is the question to Rafael, but I think that uninterruptible fridge makes sense. Because the exiting task can do a lot of things, say IO. So at least we need to ensure that nobody can be killed after try_to_freeze_tasks() succeeds, and this needs the changes in kernel/power/process.c and can lead to other problems. And it is not clear to me why would we want to do this. > > Otherwise, how can we fix that? > > We can mark all threads PF_NOFREEZE and wake them up. We can't mark them PF_NOFREEZE but of course we could do something else for de_thread() in particular, see the 1st version of Chanho's fix: https://lore.kernel.org/lkml/1541671796-8725-1-git-send-email-chanho.min@xxxxxxx/ > This would require > some more changes of course Yes, > but wouldn't that be a more appropriate > solution? Do we want to block exec for ever just because some threads > are in the fridge? Why not? ------------------------------------------------------------------------------ To clarify. speaking of de_thread() in particular, this change can not solve all problems with freezer because de_thread() is called with cred_guard_mutex held. And this obviously means that try_to_freeze_tasks() still can fail if another task waits for this mutex. But. freezable_schedule() doesn't make the thing worse, we have a lot more problems (deadlocks) exactly because de_thread() sleeps wating for other threads with this mutex held. So I didn't even mention this problem, we need to narrow the scope of this mutex in any case, so imo this has nothing to do with s/schedule/freezable_schedule/. Oleg.
