On Sun, Nov 29, 2020 at 11:56:55AM +0100, Mike Galbraith wrote:

On Sun, 2020-11-29 at 10:21 +0100, Mike Galbraith wrote:

quoted

On Sun, 2020-11-29 at 08:48 +0100, Mike Galbraith wrote:

quoted

On Sun, 2020-11-29 at 07:41 +0100, Mike Galbraith wrote:

quoted

On Sat, 2020-11-28 at 15:27 +0100, Oleksandr Natalenko wrote:

quoted

quoted

quoted

Shouldn't the list manipulation be protected with
local_lock+this_cpu_ptr instead of get_cpu_ptr+spin_lock?

Totally untested:

Hrm, the thing doesn't seem to care deeply about preemption being
disabled, so adding another lock may be overkill.  It looks like you
could get the job done via migrate_disable()+this_cpu_ptr().

There is however an ever so tiny chance that I'm wrong about that :)

Or not, your local_lock+this_cpu_ptr version exploded too.

Perhaps there's a bit of non-rt related racy racy going on in zswap
thingy that makes swap an even less wonderful idea for RT than usual.

Raciness seems to be restricted to pool compressor.  "zbud" seems to be
solid, virgin "zsmalloc" explodes, as does "z3fold" regardless which of
us puts his grubby fingerprints on it.

Exploding compressors survived zero runs of runltp -f mm, I declared
zbud to be at least kinda sorta stable after box survived five runs.

Ummm so do compressors explode under non-rt kernel in your tests as
well, or it is just -rt that triggers this?

I've never seen that on both -rt and non-rt, thus asking.

-- 
  Oleksandr Natalenko (post-factum)