On Mon, Sep 20, 2021 at 10:30:12PM +0200, Michal Hocko wrote: > We usually tend to provide Fixes where there has been something fixed. > It seems just confusing if it is used for non functional changes, > cleanups etc. Thera are gray zones of course. Got it, thanks. So no tag would be used in such a case? > I am not sure I follow. My understanding is that we need to make sure > oom_reaper is not running after the quiescent state as it is changing > user space address space. For that I believe we need to freeze the > kthread at a proper moment. That is currently the entry point and maybe > we can extend that even to the reaping loop but I haven't really > explored that. PF_FREEZER_SKIP would skip over the reaper and that could > result in it racing with the snapshotting no? Kthreads cannot be implicitly frozen; it's not like PREEMPT. From what I've read in the freezer code, two things must occur for a kthread to freeze: the kthread must have PF_NOFREEZE unset (via set_freezable(), as is done in the patch I've submitted here), and the kthread must have an explicit call into the freezer, such as via wait_event_freezable(). Right now, oom_reaper is totally ignored by the freezer because PF_NOFREEZE is set by default in all kthreads. As such, oom_reaper can keep running while system-wide freezing occurs. If you think this can mangle snapshots, then perhaps there is a real bug here after all. It sounds like you don't want oom_reaper to slow down system-wide freezing, but at the same time, you want oom_reaper to participate in system-wide freezing? I'm not sure how you could achieve that, aside from maybe inserting a call into the freezer while iterating through each vma, akin to adding a cond_resched(). My PF_FREEZER_SKIP suggestion was just to emphasize that oom_reaper is currently skipping the freezer anyway due to PF_NOFREEZE, and that you could set PF_FREEZER_SKIP to make it skip the freezer a little faster if you wanted. > Is this something to really worry about? I'm trying to emphasize that the current usage of wait_event_freezable() in oom_repear behaves *exactly* like wait_event_interruptible() but with some extra overhead. Sultan
