Re: [PATCH 1/2] numa: Set default distance map if needed

[Igor Mammedov]

On Wed, 13 Oct 2021 10:32:18 +1100
Gavin Shan <gshan@redhat.com> wrote:

> Hi Igor,
> 
> On 10/13/21 12:53 AM, Igor Mammedov wrote:
> > On Tue, 12 Oct 2021 15:13:08 +0200
> > Andrew Jones <drjones@redhat.com> wrote:   
> >> On Tue, Oct 12, 2021 at 02:27:54PM +0200, Igor Mammedov wrote:  
> >>> On Tue, 12 Oct 2021 12:37:54 +0200
> >>> Andrew Jones <drjones@redhat.com> wrote:     
> >>>> On Tue, Oct 12, 2021 at 11:40:16AM +0200, Igor Mammedov wrote:  
> >>>>> On Wed,  6 Oct 2021 18:22:08 +0800
> >>>>> Gavin Shan <gshan@redhat.com> wrote:
> >>>>>        
> >>>>>> The following option is used to specify the distance map. It's
> >>>>>> possible the option isn't provided by user. In this case, the
> >>>>>> distance map isn't populated and exposed to platform. On the
> >>>>>> other hand, the empty NUMA node, where no memory resides, is
> >>>>>> allowed on ARM64 virt platform. For these empty NUMA nodes,
> >>>>>> their corresponding device-tree nodes aren't populated, but
> >>>>>> their NUMA IDs should be included in the "/distance-map"
> >>>>>> device-tree node, so that kernel can probe them properly if
> >>>>>> device-tree is used.
> >>>>>>
> >>>>>>    -numa,dist,src=<numa_id>,dst=<numa_id>,val=<distance>
> >>>>>>
> >>>>>> So when user doesn't specify distance map, we need to generate
> >>>>>> the default distance map, where the local and remote distances
> >>>>>> are 10 and 20 separately. This adds an extra parameter to the
> >>>>>> exiting complete_init_numa_distance() to generate the default
> >>>>>> distance map for this case.
> >>>>>>
> >>>>>> Signed-off-by: Gavin Shan <gshan@redhat.com>  
> >>>>>
> >>>>>
> >>>>> how about error-ing out if distance map is required but
> >>>>> not provided by user explicitly and asking user to fix
> >>>>> command line?
> >>>>>
> >>>>> Reasoning behind this that defaults are hard to maintain
> >>>>> and will require compat hacks and being raod blocks down
> >>>>> the road.
> >>>>> Approach I was taking with generic NUMA code, is deprecating
> >>>>> defaults and replacing them with sanity checks, which bail
> >>>>> out on incorrect configuration and ask user to correct command line.
> >>>>> Hence I dislike approach taken in this patch.
> >>>>>
> >>>>> If you really wish to provide default, push it out of
> >>>>> generic code into ARM specific one
> >>>>> (then I won't oppose it that much (I think PPC does
> >>>>> some magic like this))
> >>>>> Also behavior seems to be ARM specific so generic
> >>>>> NUMA code isn't a place for it anyways  
> >>>>
> >>>> The distance-map DT node and the default 10/20 distance-map values
> >>>> aren't arch-specific. RISCV is using it too.
> >>>>
> >>>> I'm on the fence with this. I see erroring-out to require users
> >>>> to provide explicit command lines as a good thing, but I also
> >>>> see it as potentially an unnecessary burden for those that want
> >>>> the default map anyway. The optional nature of the distance-map
> >>>> node and the specification of the default map is here [1]
> >>>>
> >>>> [1] Linux source: Documentation/devicetree/bindings/numa.txt  
> >>>
> >>> Looking at proposed linux patches [ https://lkml.org/lkml/2021/9/27/31 ],
> >>> using optional distance table as source for numa-node-ids,
> >>> looks like a hack around kernel's inability to fish them out
> >>> from CPU &| PCI nodes (using those nodes as source should
> >>> cover memory-less node use-case).
> >>>
> >>> I consider including optional node as a policy decision.
> >>> So user shall include it explicitly on QEMU command line
> >>> if necessary (that works just fine for x86), or guest OS
> >>> can make up defaults on its own in absence of data.  
> >>
> >> OK, so erroring-out on configs that must provide distance-maps, rather
> >> than automatically generating them for all configs is better.
> >>  
> >>>      
> >>>> So, my r-b stands for this patch, but I also wouldn't complain
> >>>> about respinning it to error out instead.  
> >>>      
> >>>> I would complain about
> >>>> moving the logic to Arm specific code, though, since RISCV would
> >>>> then need to duplicate it.  
> >>>
> >>> Instead of putting workaround in QEMU and then making them generic,
> >>> I'd prefer to:
> >>>   1. make QEMU to be able generate DT with memory-less nodes  
> >>
> >> How? DT syntax doesn't allow this, because each node needs a unique
> >> name which is derived from its base address, which an empty numa  
> > you are talking about memory@foo nodes, aren't you?
> >   
> >> node doesn't have.  
> > 
> > Looking at Documentation/devicetree/bindings/numa.txt
> > 
> > mem/cpu/pci nodes also contain numa-node-id attribute,
> > so idea is to collect IDs from all present sources
> > instead of abusing distance map.
> >   
> > That would allow QEMU to skip memory@foo elements for
> > memory-less nodes because they obviously do not exist
> > and there is no way to describe them using 'memory' nodes.
> >   
> 
> I don't think it's feasible because it's hard to elaborate NUMA node IDs
> from this sort of sources. Apart from mem/cpu/pci, the NUMA node IDs
> can be included into platform devices, which could be vendor specific
> sometimes. Other type of devices, which I don't know, could include
> NUMA node IDs either.

Most likely mem/cpu(/pci) nodes are sufficient to get node ids
(as they can be node forming entities, not sure about PCI)
So forcing QEMU to provide the yet another optional node
to describe numa-ids that could be figured out from already
present nodes doesn't look like a good idea. Sure thing it's
a simple one from guest kernel pov (though I doubt that it's
any harder to parse cpu nodes in addition to memory ones
to get numa-ids), but otherwise it puts unnecessary
restrictions on QEMU. As for hotplug see comment below.

 
> Besides, things become more complicated when hotplug is considered.
> For example, the hot-added CPU is associated with a non-existing
> NUMA node. The CPU hot-add fails until the associated NUMA node
> is initialized. This means CPU/mem hotplug have to be twisted.

Is hotplug even a thing with device tree (I don't thinks so)?
With QEMU we use ACPI for hotplug, so from arm/virt machine
point of view we probably do not care about theoretical
hotplug with device-tree.


> So the point is to elaborate the NUMA node IDs from the limited
> source: mem/cpu/distance-map. The distance-map is optional in
> current Linux implementation.
> 
> >>>   2. fix guest to get numa-node-id from CPU/PCI nodes if
> >>>      memory node isn't present,  
> >>
> >> I'm not sure that's possible with DT. If it is, then proposing it
> >> upstream to Linux DT maintainers would be the next step.  
> > Added Rob to CC.
> >   
> 
> As explained above.
> 
> >>  
> >>> or use ACPI tables which can
> >>>      describe memory-less NUMA nodes if fixing how DT is
> >>>      parsed unfeasible.  
> >>
> >> We use ACPI already for our guests, but we also generate a DT (which
> >> edk2 consumes). We can't generate a valid DT when empty numa nodes  
> > does edk2 actually uses numa info from QEMU?
> >   
> >> are put on the command line unless we follow a DT spec saying how
> >> to do that. The current spec says we should have a distance-map
> >> that contains those nodes.  
> > 
> > can you point out to the spec and place within it, pls?
> >  
> 
> https://git.kernel.org/pub/scm/linux/kernel/git/next/linux-next.git/commit/?h=next-20211012&id=58ae0b51506802713aa0e9956d1853ba4c722c98
> ("Documentation, dt, numa: Add note to empty NUMA node")

that's not something set in stone, it's description of
a possible implementation introduced for the sake of
this patch. And can be changed.

> Thanks,
> Gavin
>