On Fri, 12 Aug 2016, Michal Hocko wrote:

On Thu 04-08-16 14:49:41, Mikulas Patocka wrote:

quoted

On Wed, 3 Aug 2016, Michal Hocko wrote:

quoted

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.

I would add PF_NO_THROTTLE or __GFP_NO_THROTTLE to mempool_alloc.

Or - we can prevent the memory reclaim from throttling if we see both 
__GFP_NOMEMALLOC and __GFP_NORETRY - that would be sufficient to detect 
mempool_alloc usage and it wouldn't hurt other __GFP_NORETRY users.

Mikulas

I believe that we should simply start with easier part and get rid of
throttle_vm_writeout because that seems like a left over from the past.
If that turns out unsatisfactory and we have clear picture when the
throttling is harmful/suboptimal then we can move on with a more complex
solution. Does this sound like a way forward?

-- 
Michal Hocko
SUSE Labs

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>