On Sat, Oct 29, 2022 at 01:23:53AM +0800, Yang Shi wrote:
> On Thu, Oct 27, 2022 at 10:55 PM Huang, Ying <ying.huang@xxxxxxxxx> wrote:
> >
> > Feng Tang <feng.tang@xxxxxxxxx> writes:
> >
> > > On Thu, Oct 27, 2022 at 10:55:58AM -0700, Yang Shi wrote:
> > >> On Thu, Oct 27, 2022 at 12:12 AM Feng Tang <feng.tang@xxxxxxxxx> wrote:
> > >> >
> > >> > On Thu, Oct 27, 2022 at 01:57:52AM +0800, Yang Shi wrote:
> > >> > > On Wed, Oct 26, 2022 at 8:59 AM Michal Hocko <mhocko@xxxxxxxx> wrote:
> > >> > [...]
> > >> > > > > > This all can get quite expensive so the primary question is, does the
> > >> > > > > > existing behavior generates any real issues or is this more of an
> > >> > > > > > correctness exercise? I mean it certainly is not great to demote to an
> > >> > > > > > incompatible numa node but are there any reasonable configurations when
> > >> > > > > > the demotion target node is explicitly excluded from memory
> > >> > > > > > policy/cpuset?
> > >> > > > >
> > >> > > > > We haven't got customer report on this, but there are quite some customers
> > >> > > > > use cpuset to bind some specific memory nodes to a docker (You've helped
> > >> > > > > us solve a OOM issue in such cases), so I think it's practical to respect
> > >> > > > > the cpuset semantics as much as we can.
> > >> > > >
> > >> > > > Yes, it is definitely better to respect cpusets and all local memory
> > >> > > > policies. There is no dispute there. The thing is whether this is really
> > >> > > > worth it. How often would cpusets (or policies in general) go actively
> > >> > > > against demotion nodes (i.e. exclude those nodes from their allowes node
> > >> > > > mask)?
> > >> > > >
> > >> > > > 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?
> > >> > > >
> > >> > > > > Your concern about the expensive cost makes sense! Some raw ideas are:
> > >> > > > > * if the shrink_folio_list is called by kswapd, the folios come from
> > >> > > > > the same per-memcg lruvec, so only one check is enough
> > >> > > > > * if not from kswapd, like called form madvise or DAMON code, we can
> > >> > > > > save a memcg cache, and if the next folio's memcg is same as the
> > >> > > > > cache, we reuse its result. And due to the locality, the real
> > >> > > > > check is rarely performed.
> > >> > > >
> > >> > > > memcg is not the expensive part of the thing. You need to get from page
> > >> > > > -> all vmas::vm_policy -> mm -> task::mempolicy
> > >> > >
> > >> > > Yeah, on the same page with Michal. Figuring out mempolicy from page
> > >> > > seems quite expensive and the correctness can't be guranteed since the
> > >> > > mempolicy could be set per-thread and the mm->task depends on
> > >> > > CONFIG_MEMCG so it doesn't work for !CONFIG_MEMCG.
> > >> >
> > >> > Yes, you are right. Our "working" psudo code for mem policy looks like
> > >> > what Michal mentioned, and it can't work for all cases, but try to
> > >> > enforce it whenever possible:
> > >> >
> > >> > static bool __check_mpol_demotion(struct folio *folio, struct vm_area_struct *vma,
> > >> > unsigned long addr, void *arg)
> > >> > {
> > >> > bool *skip_demotion = arg;
> > >> > struct mempolicy *mpol;
> > >> > int nid, dnid;
> > >> > bool ret = true;
> > >> >
> > >> > mpol = __get_vma_policy(vma, addr);
> > >> > if (!mpol) {
> > >> > struct task_struct *task;
> > >> > if (vma->vm_mm)
> > >> > task = vma->vm_mm->owner;
> > >>
> > >> But this task may not be the task you want IIUC. For example, the
> > >> process has two threads, A and B. They have different mempolicy. The
> > >> vmscan is trying to demote a page belonging to thread A, but the task
> > >> may point to thread B, so you actually get the wrong mempolicy IIUC.
> > >
> > > Yes, this is a valid concern! We don't have good solution for this.
> > > For memory policy, we may only handle the per-vma policy for now whose
> > > cost is relatively low, as a best-effort try.
> >
> > Yes. The solution isn't perfect, especially for multiple-thread
> > processes with thread specific memory policy. But the proposed code
> > above can support the most common cases at least, that is, run workload
> > with `numactl`.
>
> Not only multi threads, but also may be broken for shared pages. When
> you do rmap walk, you may get multiple contradict mempolicy, which one
> would you like to obey?
In our test code, it follows the stricter policy, that if the rmap
walk meets a mempolicy disallowing the demotion, it will stop the walk
and return with 'skip_demotion' flag set.
Thanks,
Feng
