Re: [PATCH 1/2] arm64: avoid alloc memory on offline node

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



Lorenzo Pieralisi <lorenzo.pieralisi@xxxxxxx> writes:

> On Tue, Jun 19, 2018 at 01:52:16PM +0100, Punit Agrawal wrote:
>> Michal Hocko <mhocko@xxxxxxxxxx> writes:
>> 
>> > On Tue 19-06-18 20:03:07, Xie XiuQi wrote:
>> > [...]
>> >> I tested on a arm board with 128 cores 4 numa nodes, but I set CONFIG_NR_CPUS=72.
>> >> Then node 3 is not be created, because node 3 has no memory, and no cpu.
>> >> But some pci device may related to node 3, which be set in ACPI table.
>> >
>> > Could you double check that zonelists for node 3 are generated
>> > correctly?
>> 
>> The cpus in node 3 aren't onlined and there's no memory attached - I
>> suspect that no zonelists are built for this node.
>> 
>> We skip creating a node, if the number of SRAT entries parsed exceeds
>> NR_CPUS[0]. This in turn prevents onlining the numa node and so no
>> zonelists will be created for it.
>> 
>> I think the problem will go away if the cpus are restricted via the
>> kernel command line by setting nr_cpus.
>> 
>> Xie, can you try the below patch on top of the one enabling memoryless
>> nodes? I'm not sure this is the right solution but at least it'll
>> confirm the problem.
>
> This issue looks familiar (or at least related):
>
> git log d3bd058826aa

Indeed. Thanks for digging into this.

>
> The reason why the NR_CPUS guard is there is to avoid overflowing
> the early_node_cpu_hwid array.

Ah right... I missed that. The below patch is definitely not what we
want.

> IA64 does something different in
> that respect compared to x86, we have to have a look into this.
>
> Regardless, AFAICS the proximity domains to nodes mappings should not
> depend on CONFIG_NR_CPUS, it seems that there is something wrong in that
> in ARM64 ACPI SRAT parsing.

Not only SRAT parsing but it looks like there is a similar restriction
while parsing the ACPI MADT in acpi_map_gic_cpu_interface().

The incomplete parsing introduces a dependency on the ordering of
entries being aligned between SRAT and MADT when NR_CPUS is
restricted. We want to parse the entire table in both cases so that the
code is robust to reordering of entries.

In terms of $SUBJECT, I wonder if it's worth taking the original patch
as a temporary fix (it'll also be easier to backport) while we work on
fixing these other issues and enabling memoryless nodes.



[Index of Archives]     [DMA Engine]     [Linux Coverity]     [Linux USB]     [Video for Linux]     [Linux Audio Users]     [Yosemite News]     [Linux Kernel]     [Linux SCSI]     [Greybus]

  Powered by Linux