Johannes Weiner wrote: > On Wed, Mar 09, 2016 at 03:08:53PM -0800, Andrew Morton wrote: > > On Wed, 9 Mar 2016 17:48:29 -0500 Johannes Weiner <[email protected]> > > wrote: > > > > > However, I disagree with your changelog. > > > > What text would you prefer? > > I'd just keep the one you had initially. Or better, this modified > version: > > When the OOM killer scans tasks and encounters a PF_EXITING one, it > force-selects that task regardless of the score. The problem is that > if that task got stuck waiting for some state the allocation site is > holding, the OOM reaper can not move on to the next best victim. >
There is no guarantee that the OOM reaper is waken up. There are shortcuts which I don't like. > Frankly, I don't even know why we check for exiting tasks in the OOM > killer. We've tried direct reclaim at least 15 times by the time we > decide the system is OOM, there was plenty of time to exit and free > memory; and a task might exit voluntarily right after we issue a kill. > This is testing pure noise. Remove it. > My concern is what an optimistic idea it is to wait for task_will_free_mem() or TIF_MEMDIE task forever blindly ( http://lkml.kernel.org/r/[email protected] ). We have do_exit() { exit_signals(); /* sets PF_EXITING */ /* (1) start */ exit_mm() { mm_release() { exit_robust_list() { get_user() { __do_page_fault() { /* (1) end */ down_read(¤t->mm->mmap_sem); handle_mm_fault() { kmalloc(GFP_KERNEL) { out_of_memory() { if (current->mm && (fatal_signal_pending(current) || task_will_free_mem(current))) { mark_oom_victim(current); /* sets TIF_MEMDIE */ return true; } } } } up_read(¤t->mm->mmap_sem); /* (2) start */ } } } } /* (2) end */ down_read(¤t->mm->mmap_sem); up_read(¤t->mm->mmap_sem); current->mm = NULL; exit_oom_victim(); } } sequence. We will hit silent OOM livelock if somebody sharing the mm does down_write_killable(¤t->mm->mmap_sem) and kmalloc(GFP_KERNEL) for mmap() etc. at (1) or (2) due to failing to send SIGKILL to somebody doing/done down_write_killable(¤t->mm->mmap_sem) and returning OOM_SCAN_ABORT without testing whether down_read(&victim->mm->mmap_sem) will succeed. Since the OOM reaper is not invoked when shortcut is used, nobody can unlock. Doing - if (task_will_free_mem(task) && !is_sysrq_oom(oc)) + if (task_will_free_mem(task) && !is_sysrq_oom(oc) && can_lock_mm_for_read(task)) return OOM_SCAN_ABORT; and if (test_tsk_thread_flag(task, TIF_MEMDIE)) { - if (!is_sysrq_oom(oc)) + if (!is_sysrq_oom(oc) && can_lock_mm_for_read(task)) return OOM_SCAN_ABORT; } is a too fast decision because can_lock_mm_for_read(task) might become true if if we waited for a moment. Doing - if (task_will_free_mem(task) && !is_sysrq_oom(oc)) + if (task_will_free_mem(task) && !is_sysrq_oom(oc) && we_havent_waited_enough_period(task)) return OOM_SCAN_ABORT; and if (test_tsk_thread_flag(task, TIF_MEMDIE)) { - if (!is_sysrq_oom(oc)) + if (!is_sysrq_oom(oc) && we_havent_waited_enough_period(task)) return OOM_SCAN_ABORT; } is a timeout based unlocking which Michal does not like. Doing - if (task_will_free_mem(task) && !is_sysrq_oom(oc)) + if (task_will_free_mem(task) && !is_sysrq_oom(oc) && should_oom_scan_abort(task)) return OOM_SCAN_ABORT; and if (test_tsk_thread_flag(task, TIF_MEMDIE)) { - if (!is_sysrq_oom(oc)) + if (!is_sysrq_oom(oc) && should_oom_scan_abort(task)) return OOM_SCAN_ABORT; } is a counter based unlocking which I don't know what Michal thinks. This situation is similar to when to declare OOM in OOM detection rework.
