linux-efi.vger.kernel.org archive mirror
 help / color / mirror / Atom feed
From: Gavin Shan <gshan@redhat.com>
To: Rob Herring <robh@kernel.org>
Cc: Ard Biesheuvel <ardb@kernel.org>,
	"open list:OPEN FIRMWARE AND FLATTENED DEVICE TREE BINDINGS" 
	<devicetree@vger.kernel.org>,
	"linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
	linux-arm-kernel <linux-arm-kernel@lists.infradead.org>,
	linux-efi <linux-efi@vger.kernel.org>,
	"open list:KERNEL VIRTUAL MACHINE FOR ARM64 (KVM/arm64)" 
	<kvmarm@lists.cs.columbia.edu>,
	Randy Dunlap <rdunlap@infradead.org>,
	Andrew Jones <drjones@redhat.com>, Will Deacon <will@kernel.org>,
	Marc Zyngier <maz@kernel.org>,
	Catalin Marinas <catalin.marinas@arm.com>,
	shan.gavin@gmail.com
Subject: Re: [PATCH] Documentation, dt, numa: Add note to empty NUMA node
Date: Mon, 27 Sep 2021 11:16:30 +1000	[thread overview]
Message-ID: <6aa8dd66-7910-4227-709a-8d0bbc2939ee@redhat.com> (raw)
In-Reply-To: <CAL_JsqLRH_Vn7u7b_yXgC=jBE8sZ2ax_yyueA3umbhQtbjW0TA@mail.gmail.com>

Hi Rob,

> On 9/24/21 1:17 AM, Rob Herring wrote:
> On Thu, Sep 23, 2021 at 1:32 AM Gavin Shan <gshan@redhat.com> wrote:
>> On 9/22/21 9:05 PM, Ard Biesheuvel wrote:
>>> On Tue, 21 Sept 2021 at 21:45, Rob Herring <robh@kernel.org> wrote:
>>>> On Sun, Sep 5, 2021 at 11:16 PM Gavin Shan <gshan@redhat.com> wrote:
>>>>>
>>>>> The empty memory nodes, where no memory resides in, are allowed.
>>>>> For these empty memory nodes, the 'len' of 'reg' property is zero.
>>>>> The NUMA node IDs are still valid and parsed, but memory may be
>>>>> added to them through hotplug afterwards. Currently, QEMU fails
>>>>> to boot when multiple empty memory nodes are specified. It's
>>>>> caused by device-tree population failure and duplicated memory
>>>>> node names.
>>>
>>> Those memory regions are known in advance, right? So wouldn't it be
>>> better to use something like 'status = "disabled"' here?
>>>
>>
>> Yes, these memory regions are known in advance. For the empty nodes,
>> their 'len' property is zero and it's equal to disabled state.
>>
>>>>
>>>> I still don't like the fake addresses. I can't really give suggestions
>>>> on alternative ways to fix this with you just presenting a solution.
>>>>
>>>
>>> Agreed. Please try to explain what the problem is, and why this is the
>>> best way to solve it. Please include other solutions that were
>>> considered and rejected if any exist.
>>>
>>>> What is the failure you see? Can we relax the kernel's expectations?
>>>> What about UEFI boot as the memory nodes aren't used (or maybe they
>>>> are for NUMA?) How does this work with ACPI?
>>>>
>>>
>>> The EFI memory map only needs to describe the memory that was present
>>> at boot. More memory can be represented as ACPI objects, including
>>> coldplugged memory that is already present at boot. None of this
>>> involves the memory nodes in DT.
>>>
>>
>> I'm using the following command line to start a virtual machine (VM).
>> There are 4 NUMA nodes specified, but the last two are empty. In QEMU,
>> the device-tree nodes are populated to represent these 4 NUMA nodes.
>> Unfortunately, QEMU fails to start because of the conflicting names
>> for the empty node are found, as the following error message indicates.
>>
>>      /home/gavin/sandbox/qemu.main/build/qemu-system-aarch64        \
>>      -accel kvm -machine virt,gic-version=host                      \
>>      -cpu host -smp 4,sockets=2,cores=2,threads=1                   \
>>      -m 1024M,slots=16,maxmem=64G                                   \
>>      -object memory-backend-ram,id=mem0,size=512M                   \
>>      -object memory-backend-ram,id=mem1,size=512M                   \
>>      -numa node,nodeid=0,cpus=0-1,memdev=mem0                       \
>>      -numa node,nodeid=1,cpus=2-3,memdev=mem1                       \
>>      -numa node,nodeid=2                                            \
>>      -numa node,nodeid=3                                            \
>>        :
>>      -device virtio-balloon-pci,id=balloon0,free-page-reporting=yes
>>        :
>>        :
>>      qemu-system-aarch64: FDT: Failed to create subnode /memory@80000000: FDT_ERR_EXISTS
>>
>> According to device-tree specification, the memory device-tree node's
>> name is following the format 'memory@base-address'. For the empty
>> NUMA nodes, their base addresses aren't determined. The device-tree
>> specification doesn't indicate what 'base-address' should be assigned
>> for the empty nodes. So I proposed this patch because I think the
>> linux device-tree binding documentation is best place to get this
>> documented.
> 
> Why even create the node?
> 
> What does IBM pSeries do here. AIUI, those platforms create/remove
> nodes for hotplug. That's the reason CONFIG_OF_DYNAMIC existed
> originally. Unfortunately, that's the extent of my knowledge on that.
> 

It has been long time that I didn't read pSeries related code. I
spent some time on that and you're correct. On pSeries, the device-tree
node is added dynamically. However, the IBM private nodes or properties,
whose names start with "ibm", are used for the memory hotplug. So
ARM64 can't follow without adaption.

I agree on your suggestion, which will be reflected in v2:

(a) the memory device-tree nodes are added and removed on memory hot
     add and removal.
(b) the supported NUMA nodes, including the empty ones, are identified
     through "numa-distance-map" compatible device-tree node. It's
     exactly same to what you suggested below.

In this way, we won't have the issue of the conflicting memory node
names, introduced by the empty NUMA nodes.

>> ACPI is different story. The NUMA nodes are represented by SRAT
>> (System Resource Affinity Table). In the above example, there are
>> 4 SRATs. We needn't assign names to the tables and we don't have
>> the conflicting names as we do in device-tree case.
>>
>> By the way, QEMU currently prevents to expose SRATs for empty NUMA
>> nodes. I need submit QEMU patch to break the limitation in future.
>> With the limitation, the hot-added memory is always put into the
>> last NUMA node and it's not exactly customer wants.
>>
>>>>> As device-tree specification indicates, the 'unit-address' of
>>>>> these empty memory nodes, part of their names, are the equivalents
>>>>> to 'base-address'. Unfortunately, I finds difficulty to get where
>>>>> the assignment of 'base-address' is properly documented for these
>>>>> empty memory nodes. So lets add a section for empty memory nodes
>>>>> to cover this in NUMA binding document. The 'unit-address',
>>>>> equivalent to 'base-address' in the 'reg' property of these empty
>>>>> memory nodes is specified to be the summation of highest memory
>>>>> address plus the NUMA node ID.
>>>>>
>>>>> Signed-off-by: Gavin Shan <gshan@redhat.com>
>>>>> Acked-by: Randy Dunlap <rdunlap@infradead.org>
>>>>> ---
>>>>>    Documentation/devicetree/bindings/numa.txt | 60 +++++++++++++++++++++-
>>>>>    1 file changed, 59 insertions(+), 1 deletion(-)
>>>>>
>>>>> diff --git a/Documentation/devicetree/bindings/numa.txt b/Documentation/devicetree/bindings/numa.txt
>>>>> index 21b35053ca5a..82f047bc8dd6 100644
>>>>> --- a/Documentation/devicetree/bindings/numa.txt
>>>>> +++ b/Documentation/devicetree/bindings/numa.txt
>>>>> @@ -103,7 +103,65 @@ Example:
>>>>>                   };
>>>>>
>>>>>    ==============================================================================
>>>>> -4 - Example dts
>>>>> +4 - Empty memory nodes
>>>>> +==============================================================================
>>>>> +
>>>>> +Empty memory nodes, which no memory resides in, are allowed. The 'length'
>>>>> +field of the 'reg' property is zero. However, the 'base-address' is a
>>>>> +dummy and invalid address, which is the summation of highest memory address
>>>>> +plus the NUMA node ID. The NUMA node IDs and distance maps are still valid
>>>>> +and memory may be added into them through hotplug afterwards.
>>>>> +
>>>>> +Example:
>>>>> +
>>>>> +       memory@0 {
>>>>> +               device_type = "memory";
>>>>> +               reg = <0x0 0x0 0x0 0x80000000>;
>>>>> +               numa-node-id = <0>;
>>>>> +       };
>>>>> +
>>>>> +       memory@80000000 {
>>>>> +               device_type = "memory";
>>>>> +               reg = <0x0 0x80000000 0x0 0x80000000>;
>>>>> +               numa-node-id = <1>;
>>>>> +       };
>>>>> +
>>>>> +       /* Empty memory node */
>>>>> +       memory@100000002 {
>>>>> +               device_type = "memory";
>>>>> +               reg = <0x1 0x2 0x0 0x0>;
>>>>> +               numa-node-id = <2>;
>>>>> +       };
>>>>> +
>>>>> +       /* Empty memory node */
>>>>> +       memory@100000003 {
>>>>> +               device_type = "memory";
>>>>> +               reg = <0x1 0x3 0x0 0x0>;
>>>>> +               numa-node-id = <3>;
>>>>> +       };
>>>>
>>>> Do you really need the memory nodes here or just some way to define
>>>> numa node id's 2 and 3 as valid?
>>>>
>>
>> It's the way to define NUMA node IDs are valid. Besides, the 'reg'
>> property provides 'base-address', which is part of the device-tree
>> node's name, as described in this patch.
> 
> The distance-matrix already lists all possible NUMA node IDs. That
> should be enough information for the kernel. If not, fix the kernel.
> 

Thanks, Rob. An extra patch will address the issue in v2.

Thanks,
Gavin


      reply	other threads:[~2021-09-27  1:16 UTC|newest]

Thread overview: 6+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2021-09-06  4:14 Gavin Shan
2021-09-21 19:44 ` Rob Herring
2021-09-22 11:05   ` Ard Biesheuvel
2021-09-23  6:32     ` Gavin Shan
2021-09-23 15:17       ` Rob Herring
2021-09-27  1:16         ` Gavin Shan [this message]

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=6aa8dd66-7910-4227-709a-8d0bbc2939ee@redhat.com \
    --to=gshan@redhat.com \
    --cc=ardb@kernel.org \
    --cc=catalin.marinas@arm.com \
    --cc=devicetree@vger.kernel.org \
    --cc=drjones@redhat.com \
    --cc=kvmarm@lists.cs.columbia.edu \
    --cc=linux-arm-kernel@lists.infradead.org \
    --cc=linux-efi@vger.kernel.org \
    --cc=linux-kernel@vger.kernel.org \
    --cc=maz@kernel.org \
    --cc=rdunlap@infradead.org \
    --cc=robh@kernel.org \
    --cc=shan.gavin@gmail.com \
    --cc=will@kernel.org \
    --subject='Re: [PATCH] Documentation, dt, numa: Add note to empty NUMA node' \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for NNTP newsgroup(s).