[Intel-gfx] [PATCH v3] mm, drm/i915: mark pinned shmemfs pages as unevictable

[Chris Wilson]

Quoting Chris Wilson (2018-10-31 09:41:55)
> Quoting Kuo-Hsin Yang (2018-10-31 08:19:45)
> > The i915 driver uses shmemfs to allocate backing storage for gem
> > objects. These shmemfs pages can be pinned (increased ref count) by
> > shmem_read_mapping_page_gfp(). When a lot of pages are pinned, vmscan
> > wastes a lot of time scanning these pinned pages. In some extreme case,
> > all pages in the inactive anon lru are pinned, and only the inactive
> > anon lru is scanned due to inactive_ratio, the system cannot swap and
> > invokes the oom-killer. Mark these pinned pages as unevictable to speed
> > up vmscan.
> > 
> > Add check_move_lru_page() to move page to appropriate lru list.
> > 
> > This patch was inspired by Chris Wilson's change [1].
> > 
> > [1]: https://patchwork.kernel.org/patch/9768741/
> > 
> > Cc: Chris Wilson <chris at chris-wilson.co.uk>
> > Cc: Michal Hocko <mhocko at suse.com>
> > Cc: Joonas Lahtinen <joonas.lahtinen at linux.intel.com>
> > Cc: Peter Zijlstra <peterz at infradead.org>
> > Cc: Andrew Morton <akpm at linux-foundation.org>
> > Cc: Dave Hansen <dave.hansen at intel.com>
> > Signed-off-by: Kuo-Hsin Yang <vovoy at chromium.org>
> > ---
> > The previous mapping_set_unevictable patch is worse on gem_syslatency
> > because it defers to vmscan to move these pages to the unevictable list
> > and the test measures latency to allocate 2MiB pages. This performance
> > impact can be solved by explicit moving pages to the unevictable list in
> > the i915 function.
> > 
> > Chris, can you help to run the "igt/benchmarks/gem_syslatency -t 120 -b -m"
> > test with this patch on your testing machine? I tried to run the test on
> > a Celeron N4000, 4GB Ram machine. The mean value with this patch is
> > similar to that with the mlock patch.
> 
> Will do. As you are confident, I'll try a few different machines. :)

I had one anomalous result with Ivybridge, but 3/4 different machines
confirm this is effective. I normalized the latency results from each
such that 0 is the baseline median latency (no i915 activity) and 1 is
the median latency with i915 running drm-tip.

    N           Min           Max        Median           Avg        Stddev
ivb 120      0.701641       2.79209       1.24469     1.3333911    0.40871825
byt 120     -0.108194     0.0777012     0.0485302    0.01343581   0.061524734
bxt 120     -0.262057       6.27002     0.0801667    0.15963388    0.63528121
kbl 120    -0.0891262       1.22326    -0.0245336   0.041492506    0.14929689

Just need to go back and check on ivb, perhaps running on a few older 
chipsets as well. But the evidence so far indicates that this eliminates
the impact of i915 activity on the performance of shrink_page_list,
reducing the amount of crippling stalls under mempressure and often
preventing them.
-Chris