lkml.org 
[lkml]   [2014]   [Aug]   [18]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [RFC Patch V1 00/30] Enable memoryless node on x86 platforms
Hi Gerry,

On 25.07.2014 [09:50:01 +0800], Jiang Liu wrote:
>
>
> On 2014/7/25 7:32, Nishanth Aravamudan wrote:
> > On 23.07.2014 [16:20:24 +0800], Jiang Liu wrote:
> >>
> >>
> >> On 2014/7/22 1:57, Nishanth Aravamudan wrote:
> >>> On 21.07.2014 [10:41:59 -0700], Tony Luck wrote:
> >>>> On Mon, Jul 21, 2014 at 10:23 AM, Nishanth Aravamudan
> >>>> <nacc@linux.vnet.ibm.com> wrote:
> >>>>> It seems like the issue is the order of onlining of resources on a
> >>>>> specific x86 platform?
> >>>>
> >>>> Yes. When we online a node the BIOS hits us with some ACPI hotplug events:
> >>>>
> >>>> First: Here are some new cpus
> >>>
> >>> Ok, so during this period, you might get some remote allocations. Do you
> >>> know the topology of these CPUs? That is they belong to a
> >>> (soon-to-exist) NUMA node? Can you online that currently offline NUMA
> >>> node at this point (so that NODE_DATA()) resolves, etc.)?
> >> Hi Nishanth,
> >> We have method to get the NUMA information about the CPU, and
> >> patch "[RFC Patch V1 30/30] x86, NUMA: Online node earlier when doing
> >> CPU hot-addition" tries to solve this issue by onlining NUMA node
> >> as early as possible. Actually we are trying to enable memoryless node
> >> as you have suggested.
> >
> > Ok, it seems like you have two sets of patches then? One is to fix the
> > NUMA information timing (30/30 only). The rest of the patches are
> > general discussions about where cpu_to_mem() might be used instead of
> > cpu_to_node(). However, based upon Tejun's feedback, it seems like
> > rather than force all callers to use cpu_to_mem(), we should be looking
> > at the core VM to ensure fallback is occuring appropriately when
> > memoryless nodes are present.
> >
> > Do you have a specific situation, once you've applied 30/30, where
> > kmalloc_node() leads to an Oops?
> Hi Nishanth,
> After following the two threads related to support of memoryless
> node and digging more code, I realized my first version path set is an
> overkill. As Tejun has pointed out, we shouldn't expose the detail of
> memoryless node to normal user, but there are still some special users
> who need the detail. So I have tried to summarize it as:
> 1) Arch code should online corresponding NUMA node before onlining any
> CPU or memory, otherwise it may cause invalid memory access when
> accessing NODE_DATA(nid).

I think that's reasonable.

A related caveat is that NUMA topology information should be stored as
early as possible in boot for *all* CPUs [I think only cpu_to_* is used,
at least for now], not just the boot CPU, etc. This is because (at least
on my examination) pre-SMP initcalls are not prevented from using
cpu_to_node, which will falsely return 0 for all CPUs until
set_cpu_numa_node() is called.

> 2) For normal memory allocations without __GFP_THISNODE setting in the
> gfp_flags, we should prefer numa_node_id()/cpu_to_node() instead of
> numa_mem_id()/cpu_to_mem() because the latter loses hardware topology
> information as pointed out by Tejun:
> A - B - X - C - D
> Where X is the memless node. numa_mem_id() on X would return
> either B or C, right? If B or C can't satisfy the allocation,
> the allocator would fallback to A from B and D for C, both of
> which aren't optimal. It should first fall back to C or B
> respectively, which the allocator can't do anymoe because the
> information is lost when the caller side performs numa_mem_id().

Yes, this seems like a very good description of the reasoning.

> 3) For memory allocation with __GFP_THISNODE setting in gfp_flags,
> numa_node_id()/cpu_to_node() should be used if caller only wants to
> allocate from local memory, otherwise numa_mem_id()/cpu_to_mem()
> should be used if caller wants to allocate from the nearest node.
>
> 4) numa_mem_id()/cpu_to_mem() should be used if caller wants to check
> whether a page is allocated from the nearest node.

I'm less clear on what you mean here, I'll look at your v2 patches. I
mean, numa_node_id()/cpu_to_node() should be used to indicate node-local
preference with appropriate failure handling. But I don't know why one
would prefer to use numa_node_id() to numa_mem_id() in such a path? The
only time they differ is if memoryless nodes are present, which is what
your local memory allocation would ideally be for those nodes anyways?

Thanks,
Nish



\
 
 \ /
  Last update: 2014-08-19 01:41    [W:0.102 / U:0.268 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site