On Fri, Jul 15, 2022 at 07:24:55AM +0900, Tetsuo Handa wrote: > On 2022/07/14 23:18, Jan Kara wrote: > > Hello, > > > > so this lockdep report looks real but is more related to OOM handling than > > to ext4 as such. The immediate problem I can see is that > > mem_cgroup_print_oom_meminfo() which is called under oom_lock calls > > memory_stat_format() which does GFP_KERNEL allocations to allocate buffers > > for dumping of MM statistics. This creates oom_lock -> fs reclaim > > dependency and because OOM can be hit (and thus oom_lock acquired) in > > practically any allocation (regardless of GFP_NOFS) this has a potential of > > creating real deadlock cycles. > > > > So should mem_cgroup_print_oom_meminfo() be using > > memalloc_nofs_save/restore() to avoid such deadlocks? Or perhaps someone > > sees another solution? Generally allocating memory to report OOM looks a > > bit dangerous to me ;). mem_cgroup_print_oom_meminfo() is called only for memcg OOMs. So, the situaion would be dangerous only if the system is also OOM at that time. > > > > Honza > > I think mem_cgroup_print_oom_meminfo() should use GFP_ATOMIC, for it will fall into > infinite loop if kmalloc(GFP_NOFS) under oom_lock reached __alloc_pages_may_oom() path. I would prefer GFP_NOWAIT. This is printing info for memcg OOMs and if the system is low on memory then memcg OOMs has lower importance than the system state.
