Re: [PATCH][RFC] mm: Introduce kernelcore=reliable option

[Kamezawa Hiroyuki]

On 2015/10/09 19:36, Xishi Qiu wrote:
> On 2015/10/9 17:24, Kamezawa Hiroyuki wrote:
>
> > On 2015/10/09 15:46, Xishi Qiu wrote:
> > > On 2015/10/9 22:56, Taku Izumi wrote:
> > >
> > > > Xeon E7 v3 based systems supports Address Range Mirroring
> > > > and UEFI BIOS complied with UEFI spec 2.5 can notify which
> > > > ranges are reliable (mirrored) via EFI memory map.
> > > > Now Linux kernel utilize its information and allocates
> > > > boot time memory from reliable region.
> > > >
> > > > My requirement is:
> > > >     - allocate kernel memory from reliable region
> > > >     - allocate user memory from non-reliable region
> > > >
> > > > In order to meet my requirement, ZONE_MOVABLE is useful.
> > > > By arranging non-reliable range into ZONE_MOVABLE,
> > > > reliable memory is only used for kernel allocations.
> > >
> > > Hi Taku,
> > >
> > > You mean set non-mirrored memory to movable zone, and set
> > > mirrored memory to normal zone, right? So kernel allocations
> > > will use mirrored memory in normal zone, and user allocations
> > > will use non-mirrored memory in movable zone.
> > >
> > > My question is:
> > > 1) do we need to change the fallback function?
> >
> > For *our* requirement, it's not required. But if someone want to prevent
> > user's memory allocation from NORMAL_ZONE, we need some change in zonelist
> > walking.
>
> Hi Kame,
>
> So we assume kernel will only use normal zone(mirrored), and users use movable
> zone(non-mirrored) first if the memory is not enough, then use normal zone too.

Yes.

> > > 2) the mirrored region should locate at the start of normal
> > > zone, right?
> >
> > Precisely, "not-reliable" range of memory are handled by ZONE_MOVABLE.
> > This patch does only that.
>
> I mean the mirrored region can not at the middle or end of the zone,
> BIOS should report the memory like this,
>
> e.g.
> BIOS
> node0: 0-4G mirrored, 4-8G mirrored, 8-16G non-mirrored
> node1: 16-24G mirrored, 24-32G non-mirrored
>
> OS
> node0: DMA DMA32 are both mirrored, NORMAL(4-8G), MOVABLE(8-16G)
> node1: NORMAL(16-24G), MOVABLE(24-32G)

I think zones can be overlapped even while they are aligned to MAX_ORDER.


> > > I remember Kame has already suggested this idea. In my opinion,
> > > I still think it's better to add a new migratetype or a new zone,
> > > so both user and kernel could use mirrored memory.
> >
> > Hi, Xishi.
> >
> > I and Izumi-san discussed the implementation much and found using "zone"
> > is better approach.
> >
> > The biggest reason is that zone is a unit of vmscan and all statistics and
> > handling the range of memory for a purpose. We can reuse all vmscan and
> > information codes by making use of zones. Introdcing other structure will be messy.
>
> Yes, add a new zone is better, but it will change much code, so reuse ZONE_MOVABLE
> is simpler and easier, right?

I think so. If someone feels difficulty with ZONE_MOVABLE, adding zone will be another job.
(*)Taku-san's bootoption is to specify kernelcore to be placed into reliable memory and
   doesn't specify anything about users.


> > His patch is very simple.
>
> The following plan sounds good to me. Shall we rename the zone name when it is
> used for mirrored memory, "movable" is a little confusion.

Maybe. I think it should be another discussion. With this patch and his fake-reliable-memory
patch, everyone can give a try.


> > For your requirements. I and Izumi-san are discussing following plan.
> >
> >   - Add a flag to show the zone is reliable or not, then, mark ZONE_MOVABLE as not-reliable.
> >   - Add __GFP_RELIABLE. This will allow alloc_pages() to skip not-reliable zone.
> >   - Add madivse() MADV_RELIABLE and modify page fault code's gfp flag with that flag.
>
> like this?
> user: madvise()/mmap()/or others -> add vma_reliable flag -> add gfp_reliable flag -> alloc_pages
> kernel: use __GFP_RELIABLE flag in buddy allocation/slab/vmalloc...
yes.

> Also we can introduce some interfaces in procfs or sysfs, right?

It's based on your use case. I think madvise() will be the 1st choice.

Thanks,
-kame





--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/