On Tue 25-02-20 16:09:29, David Hildenbrand wrote:

On 25.02.20 15:58, Michal Hocko wrote:

quoted

On Thu 12-12-19 18:11:36, David Hildenbrand wrote:

quoted

We already have a way to trigger reclaiming of all reclaimable slab objects
from user space (echo 2 > /proc/sys/vm/drop_caches). Let's allow drivers
to also trigger this when they really want to make progress and know what
they are doing.

I cannot say I would be fan of this. This is a global action with user
visible performance impact. I am worried that we will find out that all
sorts of drivers have a very good idea that dropping slab caches is
going to help their problem whatever it is. We have seen the same patter
in the userspace already and that is the reason we are logging the usage
to the log and count invocations in the counter.

Yeah, I decided to hold back patch 11-13 for the v1 (which I am planning
to post in March after more testing). What we really want is to make
memory offlining an alloc_contig_range() work better with reclaimable
objects.

quoted

quoted

virtio-mem wants to use these functions when it failed to unplug memory
for quite some time (e.g., after 30 minutes). It will then try to
free up reclaimable objects by dropping the slab caches every now and
then (e.g., every 30 minutes) as long as necessary. There will be a way to
disable this feature and info messages will be logged.

In the future, we want to have a drop_slab_range() functionality
instead. Memory offlining code has similar demands and also other
alloc_contig_range() users (e.g., gigantic pages) could make good use of
this feature. Adding it, however, requires more work/thought.

We already do have a memory_notify(MEM_GOING_OFFLINE) for that purpose
and slab allocator implements a callback (slab_mem_going_offline_callback).
The callback is quite dumb and it doesn't really try to free objects
from the given memory range or even try to drop active objects which
might turn out to be hard but this sounds like a more robust way to
achieve what you want.

Two things:

1. memory_notify(MEM_GOING_OFFLINE) is called after trying to isolate
the page range and checking if we only have movable pages. Won't help
much I guess.

You are right, I have missed that. Can we reorder those two calls?

2. alloc_contig_range() won't benefit from that.

True.

Something like drop_slab_range() would be better, and calling it from
the right spots in the core (e.g., set_migratetype_isolate() see below).

Especially, have a look at mm/page_isolation.c:set_migratetype_isolate()

"FIXME: Now, memory hotplug doesn't call shrink_slab() by itself. We
just check MOVABLE pages."

shrink_slab is really a large hammer for this purpose. The notifier
mechanism sounds more appropriate to me. If that means to move it
outside of its current position then let's try to experiment with that.
But there is a long route to have per pfn range reclaim.
-- 
Michal Hocko
SUSE Labs