Re: [Qemu-devel] [PATCH v3 07/10] hw/arm/virt: Introduce opt-in feature "fdt"

[Laszlo Ersek]

On 04/03/19 11:49, Igor Mammedov wrote:
> On Tue, 2 Apr 2019 17:38:26 +0200
> Laszlo Ersek <address@hidden> wrote:
> 
>> On 04/02/19 17:29, Auger Eric wrote:
>>> Hi Laszlo,
>>>
>>> On 4/1/19 3:07 PM, Laszlo Ersek wrote:  
>>>> On 03/29/19 14:56, Auger Eric wrote:  
>>>>> Hi Ard,
>>>>>
>>>>> On 3/29/19 2:14 PM, Ard Biesheuvel wrote:  
>>>>>> On Fri, 29 Mar 2019 at 14:12, Auger Eric <address@hidden> wrote:  
>>>>>>>
>>>>>>> Hi Shameer,
>>>>>>>
>>>>>>> On 3/29/19 10:59 AM, Shameerali Kolothum Thodi wrote:  
>>>>>>>>
>>>>>>>>  
>>>>>>>>> -----Original Message-----
>>>>>>>>> From: Auger Eric [mailto:address@hidden
>>>>>>>>> Sent: 29 March 2019 09:32
>>>>>>>>> To: Shameerali Kolothum Thodi <address@hidden>;
>>>>>>>>> address@hidden; address@hidden; address@hidden;
>>>>>>>>> address@hidden; address@hidden;
>>>>>>>>> address@hidden; address@hidden
>>>>>>>>> Cc: Linuxarm <address@hidden>; xuwei (O) <address@hidden>;
>>>>>>>>> Laszlo Ersek <address@hidden>; Ard Biesheuvel
>>>>>>>>> <address@hidden>; Leif Lindholm <address@hidden>
>>>>>>>>> Subject: Re: [PATCH v3 07/10] hw/arm/virt: Introduce opt-in feature 
>>>>>>>>> "fdt"
>>>>>>>>>
>>>>>>>>> Hi Shameer,
>>>>>>>>>
>>>>>>>>> [ + Laszlo, Ard, Leif ]
>>>>>>>>>
>>>>>>>>> On 3/21/19 11:47 AM, Shameer Kolothum wrote:  
>>>>>>>>>> This is to disable/enable populating DT nodes in case
>>>>>>>>>> any conflict with acpi tables. The default is "off".  
>>>>>>>>> The name of the option sounds misleading to me. Also we don't really
>>>>>>>>> know the scope of the disablement. At the moment this just aims to
>>>>>>>>> prevent the hotpluggable dt nodes from being added if we boot in ACPI 
>>>>>>>>> mode.
>>>>>>>>>  
>>>>>>>>>>
>>>>>>>>>> This will be used in subsequent patch where cold plug
>>>>>>>>>> device-memory support is added for DT boot.  
>>>>>>>>> I am concerned about the fact that in dt mode, by default, you won't 
>>>>>>>>> see
>>>>>>>>> any PCDIMM nodes.  
>>>>>>>>>>
>>>>>>>>>> If DT memory node support is added for cold-plugged device
>>>>>>>>>> memory, those memory will be visible to Guest kernel via
>>>>>>>>>> UEFI GetMemoryMap() and gets treated as early boot memory.  
>>>>>>>>> Don't we have an issue in UEFI then. Normally the SRAT indicates 
>>>>>>>>> whether
>>>>>>>>> the slots are hotpluggable or not. Shouldn't the UEFI code look at 
>>>>>>>>> this
>>>>>>>>> info.  
>>>>>>>>
>>>>>>>> Sorry I missed this part. Yes, that will be a more cleaner solution.
>>>>>>>>
>>>>>>>> Also, to be more clear on what happens,
>>>>>>>>
>>>>>>>> Guest ACPI boot with "fdt=on" ,
>>>>>>>>
>>>>>>>> From kernel log,
>>>>>>>>
>>>>>>>> [    0.000000] Early memory node ranges
>>>>>>>> [    0.000000]   node   0: [mem 0x0000000040000000-0x00000000bbf5ffff]
>>>>>>>> [    0.000000]   node   0: [mem 0x00000000bbf60000-0x00000000bbffffff]
>>>>>>>> [    0.000000]   node   0: [mem 0x00000000bc000000-0x00000000bc02ffff]
>>>>>>>> [    0.000000]   node   0: [mem 0x00000000bc030000-0x00000000bc36ffff]
>>>>>>>> [    0.000000]   node   0: [mem 0x00000000bc370000-0x00000000bf64ffff]
>>>>>>>> [    0.000000]   node   0: [mem 0x00000000bf650000-0x00000000bf6dffff]
>>>>>>>> [    0.000000]   node   0: [mem 0x00000000bf6e0000-0x00000000bf6effff]
>>>>>>>> [    0.000000]   node   0: [mem 0x00000000bf6f0000-0x00000000bf80ffff]
>>>>>>>> [    0.000000]   node   0: [mem 0x00000000bf810000-0x00000000bfffffff]
>>>>>>>> [    0.000000]   node   0: [mem 0x00000000c0000000-0x00000000ffffffff] 
>>>>>>>>  --> This is the hotpluggable memory node from DT.
>>>>>>>> [    0.000000] Zeroed struct page in unavailable ranges: 1040 pages
>>>>>>>> [    0.000000] Initmem setup node 0 [mem 
>>>>>>>> 0x0000000040000000-0x00000000ffffffff]
>>>>>>>>
>>>>>>>>
>>>>>>>> Guest ACPI boot with "fdt=off" ,
>>>>>>>>
>>>>>>>> [    0.000000] Movable zone start for each node
>>>>>>>> [    0.000000] Early memory node ranges
>>>>>>>> [    0.000000]   node   0: [mem 0x0000000040000000-0x00000000bbf5ffff]
>>>>>>>> [    0.000000]   node   0: [mem 0x00000000bbf60000-0x00000000bbffffff]
>>>>>>>> [    0.000000]   node   0: [mem 0x00000000bc000000-0x00000000bc02ffff]
>>>>>>>> [    0.000000]   node   0: [mem 0x00000000bc030000-0x00000000bc36ffff]
>>>>>>>> [    0.000000]   node   0: [mem 0x00000000bc370000-0x00000000bf64ffff]
>>>>>>>> [    0.000000]   node   0: [mem 0x00000000bf650000-0x00000000bf6dffff]
>>>>>>>> [    0.000000]   node   0: [mem 0x00000000bf6e0000-0x00000000bf6effff]
>>>>>>>> [    0.000000]   node   0: [mem 0x00000000bf6f0000-0x00000000bf80ffff]
>>>>>>>> [    0.000000]   node   0: [mem 0x00000000bf810000-0x00000000bfffffff]
>>>>>>>> [    0.000000] Zeroed struct page in unavailable ranges: 1040 pages
>>>>>>>> [    0.000000] Initmem setup node 0 [mem 
>>>>>>>> 0x0000000040000000-0x00000000bfffffff
>>>>>>>>
>>>>>>>> The hotpluggable memory node is absent from early memory nodes here.  
>>>>>>>
>>>>>>> OK thank you for the example illustrating the concern.  
>>>>>>>>
>>>>>>>> As you said, it could be possible to detect this node using SRAT in 
>>>>>>>> UEFI.  
>>>>>>>
>>>>>>> Let's wait for EDK2 experts on this.
>>>>>>>  
>>>>>>
>>>>>> Happy to chime in, but I need a bit more context here.
>>>>>>
>>>>>> What is the problem, how does this path try to solve it, and why is
>>>>>> that a bad idea?
>>>>>>  
>>>>> Sure, sorry.
>>>>>
>>>>> This series:
>>>>> - [PATCH v3 00/10] ARM virt: ACPI memory hotplug support,
>>>>> https://patchwork.kernel.org/cover/10863301/
>>>>>
>>>>> aims to introduce PCDIMM support in qemu. In ACPI mode, it builds the
>>>>> SRAT and DSDT parts and relies on GED to trigger the hotplug.
>>>>>
>>>>> We noticed that if we build the hotpluggable memory dt nodes on top of
>>>>> the above ACPI tables, the DIMM slots are interpreted as not
>>>>> hotpluggable memory slots (at least we think so).
>>>>>
>>>>> We think the EDK2 GetMemoryMap() uses the dt node info and ignores the
>>>>> fact that those slots are exposed as hotpluggable in the SRAT for example.
>>>>>
>>>>> So in this series, we are forced to not generate the hotpluggable memory
>>>>> dt nodes if we want the DIMM slots to be effectively recognized as
>>>>> hotpluggable.
>>>>>
>>>>> Could you confirm we have a correct understanding of the EDK2 behaviour
>>>>> and if so, would there be any solution for EDK2 to absorb both the DT
>>>>> nodes and the relevant SRAT/DSDT tables and make the slots hotpluggable.
>>>>>
>>>>> At qemu level, detecting we are booting in ACPI mode and purposely
>>>>> removing the above mentioned DT nodes does not look straightforward.  
>>>>
>>>> The firmware is not enlightened about the ACPI content that comes from
>>>> QEMU / fw_cfg. That ACPI content is *blindly* processed by the firmware,
>>>> as instructed through the ACPI linker/loader script, in order to install
>>>> the ACPI content for the OS. No actual information is consumed by the
>>>> firmware from the ACPI payload -- and that's a feature.
>>>>
>>>> The firmware does consume DT:
>>>>
>>>> - If you start QEMU *with* "-no-acpi", then the DT is both consumed by
>>>> the firmware (for its own information needs), and passed on to the OS.
>>>>
>>>> - If you start QEMU *without* "-no-acpi" (the default), then the DT is
>>>> consumed only by the firmware (for its own information needs), and the
>>>> DT is hidden from the OS. The OS gets only the ACPI content
>>>> (processed/prepared as described above).
>>>>
>>>>
>>>> In the firmware, the "ArmVirtPkg/HighMemDxe" driver iterates over the
>>>> base/size pairs in all the memory nodes in the DT. For each such base
>>>> address that is currently tracked as "nonexistent" in the GCD memory
>>>> space map, the driver currently adds the base/size range as "system
>>>> memory". This in turn is reflected by the UEFI memmap that the OS gets
>>>> to see as "conventional memory".
>>>>
>>>> If you need some memory ranges to show up as "special" in the UEFI
>>>> memmap, then you need to distinguish them somehow from the "regular"
>>>> memory areas, in the DT. And then extend "ArmVirtPkg/HighMemDxe" in the
>>>> firmware, so that it act upon the discriminator that you set in the DT.
>>>>
>>>>
>>>> Now... from a brief look at the Platform Init and UEFI specs, my
>>>> impression is that the hotpluggable (but presently not plugged) DIMM
>>>> ranges should simply be *absent* from the UEFI memmap; is that correct?
>>>> (I didn't check the ACPI spec, maybe it specifies the expected behavior
>>>> in full.) If my impression is correct, then two options (alternatives)
>>>> exist:
>>>>
>>>> (1) Hide the affected memory nodes -- or at least the affected base/size
>>>> pairs -- from the DT, in case you boot without "-no-acpi" but with an
>>>> external firmware loaded. Then the firmware will not expose those ranges
>>>> as "conventional memory" in the UEFI memmap. This approach requires no
>>>> changes to edk2.
>>>>
>>>> This option is precisely what Eric described up-thread, at
>>>> <http://mid.mail-archive.com/address@hidden>:
>>>>  
>>>>> in machvirt_init, there is firmware_loaded that tells you whether you
>>>>> have a FW image. If this one is not set, you can induce dt. But if
>>>>> there is a FW it can be either DT or ACPI booted. You also have the
>>>>> acpi_enabled knob.  
>>>>
>>>> (The "-no-acpi" cmdline option clears the "acpi_enabled" variable in
>>>> "vl.c").
>>>>
>>>> So, the condition for hiding the hotpluggable memory nodes in question
>>>> from the DT is:  
>>>   
>>>>
>>>>   (aarch64 && firmware_loaded && acpi_enabled)  
>>>
>>> Thanks a lot for all those inputs!
>>>
>>> I don't get why we test aarch64 in above condition (this was useful for
>>> high ECAM range as the aarch32 FW was not supporting it but here, is it
>>> still meaningful?)  
>>
>> Sorry, I should have clarified that. Yes, it is meaningful:
>>
>> While UEFI has bindings for both 32-bit and 64-bit ARM, ACPI has a
>> 64-bit-only binding for ARM. (And you can have UEFI without ACPI, but
>> not the reverse, on ARM.) So if you run the 32-bit build of the
>> ArmVirtQemu firmware, you get no ACPI at all; all you can rely on with
>> the OS is the DT.
>>
>> This "bitness distinction" is implemented in the firmware already. If
>> you hid the memory nodes from the DT under the condition
>>
>>   (!aarch64 && firmware_loaded && acpi_enabled)
>>
>> then the nodes would not be seen by the OS at all (because
>> "acpi_enabled" is irrelevant for the 32-bit build of ArmVirtQemu, and
>> all the OS can ever get is DT).
> 
> It's getting tricky and I don't like a bit that we are trying to carter
> 64 bit only UEFI build (or any other build) on QEMU side. Also Peter has
> a valid about guessing on QEMU side (that's usually a source of problem
> in the future).
> 
> Perhaps we should reconsider and think about marking hotplugbbale RAM
> in DT and let firmware to exclude it from memory map.

I'm fine either way.

(I'm glad to continue discussing either option; that shouldn't be taken
as a preference on my end.)

With option (2), please consider the new version dependency between QEMU
and the firmware -- this may or may not affect migration. (Thinking
about migration is difficult, so I'll leave that to you all :) )

Thanks
Laszlo

>>>> (2) Invent and set an "ignore me, firmware" property for the
>>>> hotpluggable memory nodes in the DT, and update the firmware to honor
>>>> that property.
>>>>
>>>> Thanks
>>>> Laszlo
>>>>  
>>
>