Re: [RFC PATCH] mm: have kswapd only reclaiming use min protection on memcg

On Wed, Oct 27, 2021 at 8:31 PM Michal Hocko [off-list ref] wrote:

On Wed 27-10-21 20:05:30, Zhaoyang Huang wrote:

quoted

On Wed, Oct 27, 2021 at 7:52 PM Michal Hocko [off-list ref] wrote:

quoted

On Wed 27-10-21 17:19:56, Zhaoyang Huang wrote:

quoted

On Wed, Oct 27, 2021 at 4:26 PM Michal Hocko [off-list ref] wrote:

quoted

On Wed 27-10-21 15:46:19, Zhaoyang Huang wrote:

quoted

On Wed, Oct 27, 2021 at 3:20 PM Michal Hocko [off-list ref] wrote:

quoted

On Wed 27-10-21 15:01:50, Huangzhaoyang wrote:

quoted

From: Zhaoyang Huang <redacted>

For the kswapd only reclaiming, there is no chance to try again on
this group while direct reclaim has. fix it by judging gfp flag.

There is no problem description (same as in your last submissions. Have
you looked at the patch submission documentation as recommended
previously?).

Also this patch doesn't make any sense. Both direct reclaim and kswapd
use a gfp mask which contains __GFP_DIRECT_RECLAIM (see balance_pgdat
for the kswapd part)..

ok, but how does the reclaiming try with memcg's min protection on the
alloc without __GFP_DIRECT_RECLAIM?

I do not follow. There is no need to protect memcg if the allocation
request doesn't have __GFP_DIRECT_RECLAIM because that would fail the
charge if a hard limit is reached, see try_charge_memcg and
gfpflags_allow_blocking check.

Background reclaim, on the other hand never breaches reclaim protection.

What is the actual problem you want to solve?

Imagine there is an allocation with gfp_mask & ~GFP_DIRECT_RECLAIM and
all processes are under cgroups. Kswapd is the only hope here which
however has a low efficiency of get_scan_count. I would like to have
kswapd work as direct reclaim in 2nd round which will have
protection=memory.min.

Do you have an example where this would be a practical problem? Atomic
allocations should be rather rare.

Please find below for the search result of '~__GFP_DIRECT_RECLAIM'
which shows some drivers and net prefer to behave like that.
Furthermore, the allocations are always together with high order.

And what is the _practical_ problem you are seeing or trying to solve?

We do have out of tree code behave like this and want to make the
mechanics more robust

--
Michal Hocko
SUSE Labs