On Sun, 14 Aug 2016, Michal Hocko wrote:
> On Sat 13-08-16 13:34:29, Mikulas Patocka wrote: > > > > > > On Fri, 12 Aug 2016, Michal Hocko wrote: > > > > > On Thu 04-08-16 14:49:41, Mikulas Patocka wrote: > > > > > > > On Wed, 3 Aug 2016, Michal Hocko wrote: > > > > > > > > > But the device congestion is not the only condition required for the > > > > > throttling. The pgdat has also be marked congested which means that > > > > > the > > > > > LRU page scanner bumped into dirty/writeback/pg_reclaim pages at the > > > > > tail of the LRU. That should only happen if we are rotating LRUs too > > > > > quickly. AFAIU the reclaim shouldn't allow free ticket scanning in > > > > > that > > > > > situation. > > > > > > > > The obvious problem here is that mempool allocations should sleep in > > > > mempool_alloc() on &pool->wait (until someone returns some entries into > > > > the mempool), they should not sleep inside the page allocator. > > > > > > I agree that mempool_alloc should _primarily_ sleep on their own > > > throttling mechanism. I am not questioning that. I am just saying that > > > the page allocator has its own throttling which it relies on and that > > > cannot be just ignored because that might have other undesirable side > > > effects. So if the right approach is really to never throttle certain > > > requests then we have to bail out from a congested nodes/zones as soon > > > as the congestion is detected. > > > > > > Now, I would like to see that something like that is _really_ necessary. > > > > Currently, it is not a problem - device mapper reports the device as > > congested only if the underlying physical disks are congested. > > > > But once we change it so that device mapper reports congested state on its > > own (when it has too many bios in progress), this starts being a problem. > > OK, can we wait until it starts becoming a real problem and solve it > appropriately then? > > I will repost the patch which removes thottle_vm_pageout in the meantime > as it doesn't seem to be needed anymore. > > -- > Michal Hocko > SUSE Labs Hi Michal So, here Google developers hit a stacktrace where a block device driver is being throttled in the memory management: https://www.redhat.com/archives/dm-devel/2016-November/msg00158.html dm-bufio layer is something like a buffer cache, used by block device drivers. Unlike the real buffer cache, dm-bufio guarantees forward progress even if there is no memory free. dm-bufio does something similar like a mempool allocation, it tries an allocation with GFP_NOIO | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN (just like a mempool) and if it fails, it will reuse some existing buffer. Here, they caught it being throttled in the memory management: Workqueue: kverityd verity_prefetch_io __switch_to+0x9c/0xa8 __schedule+0x440/0x6d8 schedule+0x94/0xb4 schedule_timeout+0x204/0x27c schedule_timeout_uninterruptible+0x44/0x50 wait_iff_congested+0x9c/0x1f0 shrink_inactive_list+0x3a0/0x4cc shrink_lruvec+0x418/0x5cc shrink_zone+0x88/0x198 try_to_free_pages+0x51c/0x588 __alloc_pages_nodemask+0x648/0xa88 __get_free_pages+0x34/0x7c alloc_buffer+0xa4/0x144 __bufio_new+0x84/0x278 dm_bufio_prefetch+0x9c/0x154 verity_prefetch_io+0xe8/0x10c process_one_work+0x240/0x424 worker_thread+0x2fc/0x424 kthread+0x10c/0x114 Will you consider removing vm throttling for __GFP_NORETRY allocations? Mikulas