On 2018/10/30 20:39, Michal Hocko wrote: > On Tue 30-10-18 18:47:43, Tetsuo Handa wrote: >> On 2018/10/30 15:31, Michal Hocko wrote: >>> On Tue 30-10-18 13:45:22, Tetsuo Handa wrote: >>>> Michal Hocko wrote: >>>>> @@ -3156,6 +3166,13 @@ void exit_mmap(struct mm_struct *mm) >>>>> vma = remove_vma(vma); >>>>> } >>>>> vm_unacct_memory(nr_accounted); >>>>> + >>>>> + /* >>>>> + * Now that the full address space is torn down, make sure the >>>>> + * OOM killer skips over this task >>>>> + */ >>>>> + if (oom) >>>>> + set_bit(MMF_OOM_SKIP, &mm->flags); >>>>> } >>>>> >>>>> /* Insert vm structure into process list sorted by address >>>> >>>> I don't like setting MMF_OOF_SKIP after remove_vma() loop. 50 users might >>>> call vma->vm_ops->close() from remove_vma(). Some of them are doing fs >>>> writeback, some of them might be doing GFP_KERNEL allocation from >>>> vma->vm_ops->open() with a lock also held by vma->vm_ops->close(). >>>> >>>> I don't think that waiting for completion of remove_vma() loop is safe. >>> >>> What do you mean by 'safe' here? >>> >> >> safe = "Does not cause OOM lockup." >> >> remove_vma() is allowed to sleep, and some users might depend on memory >> allocation when the OOM killer is waiting for remove_vma() to complete. > > But MMF_OOF_SKIP is set after we are done with remove_vma. In fact it is > the very last thing in exit_mmap. So I do not follow what you mean. > So what? Think the worst case. Quite obvious bug here. What happens if memory reclaimed by up to __free_pgtables() was consumed by somebody else, and then some vma->vm_ops->close() started waiting for memory allocation due to dependency? It is called "OOM lockup" because the OOM killer cannot be enabled because MMF_OOM_SKIP cannot be set because vma->vm_ops->close() is waiting for the OOM killer due to memory allocation dependency in vma->vm_ops->close() from remove_vma()...
