Re: [PATCH] mm/vmscan: respect cpuset policy during page demotion

["Huang, Ying" <ying.huang@intel.com>]

Michal Hocko <mhocko@suse.com> writes:

> On Thu 27-10-22 15:39:00, Huang, Ying wrote:
>> Michal Hocko <mhocko@suse.com> writes:
>> 
>> > On Thu 27-10-22 14:47:22, Huang, Ying wrote:
>> >> Michal Hocko <mhocko@suse.com> writes:
>> > [...]
>> >> > I can imagine workloads which wouldn't like to get their memory demoted
>> >> > for some reason but wouldn't it be more practical to tell that
>> >> > explicitly (e.g. via prctl) rather than configuring cpusets/memory
>> >> > policies explicitly?
>> >> 
>> >> If my understanding were correct, prctl() configures the process or
>> >> thread.
>> >
>> > Not necessarily. There are properties which are per adddress space like
>> > PR_[GS]ET_THP_DISABLE. This could be very similar.
>> >
>> >> How can we get process/thread configuration at demotion time?
>> >
>> > As already pointed out in previous emails. You could hook into
>> > folio_check_references path, more specifically folio_referenced_one
>> > where you have all that you need already - all vmas mapping the page and
>> > then it is trivial to get the corresponding vm_mm. If at least one of
>> > them has the flag set then the demotion is not allowed (essentially the
>> > same model as VM_LOCKED).
>> 
>> Got it!  Thanks for detailed explanation.
>> 
>> One bit may be not sufficient.  For example, if we want to avoid or
>> control cross-socket demotion and still allow demoting to slow memory
>> nodes in local socket, we need to specify a node mask to exclude some
>> NUMA nodes from demotion targets.
>
> Isn't this something to be configured on the demotion topology side? Or
> do you expect there will be per process/address space usecases? I mean
> different processes running on the same topology, one requesting local
> demotion while other ok with the whole demotion topology?

I think that it's possible for different processes have different
requirements.

- Some processes don't care about where the memory is placed, prefer
  local, then fall back to remote if no free space.

- Some processes want to avoid cross-socket traffic, bind to nodes of
  local socket.

- Some processes want to avoid to use slow memory, bind to fast memory
  node only.

>> >From overhead point of view, this appears similar as that of VMA/task
>> memory policy?  We can make mm->owner available for memory tiers
>> (CONFIG_NUMA && CONFIG_MIGRATION).  The advantage is that we don't need
>> to introduce new ABI.  I guess users may prefer to use `numactl` than a
>> new ABI?
>
> mm->owner is a wrong direction. It doesn't have a strong meaning because
> there is no one task explicitly responsible for the mm so there is no
> real owner (our clone() semantic is just to permissive for that). The
> memcg::owner is a crude and ugly hack and it should go away over time
> rather than build new uses.
>
> Besides that, and as I have already tried to explain, per task demotion
> policy is what makes this whole thing expensive. So this better be a per
> mm or per vma property. Whether it is a on/off knob like PR_[GS]ET_THP_DISABLE
> or there are explicit requirements for fine grain control on the vma
> level I dunno. I haven't seen those usecases yet and it is really easy
> to overengineer this.
>
> To be completely honest I would much rather wait for those usecases
> before adding a more complex APIs.  PR_[GS]_DEMOTION_DISABLED sounds
> like a reasonable first step. Should we have more fine grained
> requirements wrt address space I would follow the MADV_{NO}HUGEPAGE
> lead.
>
> If we really need/want to give a fine grained control over demotion
> nodemask then we would have to go with vma->mempolicy interface. In
> any case a per process on/off knob sounds like a reasonable first step
> before we learn more about real usecases.

Yes.  Per-mm or per-vma property is much better than per-task property.
Another possibility, how about add a new flag to set_mempolicy() system
call to set the per-mm mempolicy?  `numactl` can use that by default.

Best Regards,
Huang, Ying