----- On Feb 16, 2018, at 11:53 AM, Mark Rutland mark.rutl...@arm.com wrote:
> Hi, > > On Thu, Feb 15, 2018 at 10:08:56PM +0000, Mathieu Desnoyers wrote: >> My current theory: do_exit() gets preempted after having set current->mm >> to NULL, and after having issued mmput(), which brings the mm_count down >> to 0. >> >> Unfortunately, if the scheduler switches from a userspace thread >> to a kernel thread, context_switch() loads prev->active_mm which still >> points to the now-freed mm, mmgrab the mm, and eventually does mmdrop >> in finish_task_switch(). > > For this to happen, we need to get to the mmput() in exit_mm() with: > > mm->mm_count == 1 > mm->mm_users == 1 > mm == active_mm > > ... but AFAICT, this cannot happen. > > If there's no context_switch between clearing current->mm and the > mmput(), then mm->mm_count >= 2, thanks to the prior mmgrab() and the > active_mm reference (in mm_count) that context_switch+finish_task_switch > manage. > > If there is a context_switch between the two, then AFAICT, either: > > a) The task re-inherits its old mm as active_mm, and mm_count >= 2. In > context_switch we mmgrab() the active_mm to inherit it, and in > finish_task_switch() we drop the oldmm, balancing the mmgrab() with > an mmput(). > > e.g we go task -> kernel_task -> task > > b) At some point, another user task is scheduled, and we switch to its > mm. We don't mmgrab() the active_mm, but we mmdrop() the oldmm, which > means mm_count >= 1. Since we witched to a new mm, if we switch back > to the first task, it cannot have its own mm as active_mm. > > e.g. we go task -> other_task -> task > > I suspect we have a bogus mmdrop or mmput elsewhere, and do_exit() and > finish_task_switch() aren't to blame. Currently reviewing: fs/proc/base.c: __set_oom_adj() /* * Make sure we will check other processes sharing the mm if this is * not vfrok which wants its own oom_score_adj. * pin the mm so it doesn't go away and get reused after task_unlock */ if (!task->vfork_done) { struct task_struct *p = find_lock_task_mm(task); if (p) { if (atomic_read(&p->mm->mm_users) > 1) { mm = p->mm; mmgrab(mm); } task_unlock(p); } } Considering that mmput() done by exit_mm() is done outside of the task_lock critical section, I wonder how the mm_users load is synchronized ? Thanks, Mathieu -- Mathieu Desnoyers EfficiOS Inc. http://www.efficios.com