Re: [PATCH v2 2/3] mm/memory_hotplug: Introduce MHP_NO_FIRMWARE_MEMMAP
From: David Hildenbrand
Date: Fri May 01 2020 - 17:11:18 EST
On 01.05.20 22:12, Dan Williams wrote:
> On Fri, May 1, 2020 at 12:18 PM David Hildenbrand <david@xxxxxxxxxx> wrote:
>>
>> On 01.05.20 20:43, Dan Williams wrote:
>>> On Fri, May 1, 2020 at 11:14 AM David Hildenbrand <david@xxxxxxxxxx> wrote:
>>>>
>>>> On 01.05.20 20:03, Dan Williams wrote:
>>>>> On Fri, May 1, 2020 at 10:51 AM David Hildenbrand <david@xxxxxxxxxx> wrote:
>>>>>>
>>>>>> On 01.05.20 19:45, David Hildenbrand wrote:
>>>>>>> On 01.05.20 19:39, Dan Williams wrote:
>>>>>>>> On Fri, May 1, 2020 at 10:21 AM David Hildenbrand <david@xxxxxxxxxx> wrote:
>>>>>>>>>
>>>>>>>>> On 01.05.20 18:56, Dan Williams wrote:
>>>>>>>>>> On Fri, May 1, 2020 at 2:34 AM David Hildenbrand <david@xxxxxxxxxx> wrote:
>>>>>>>>>>>
>>>>>>>>>>> On 01.05.20 00:24, Andrew Morton wrote:
>>>>>>>>>>>> On Thu, 30 Apr 2020 20:43:39 +0200 David Hildenbrand <david@xxxxxxxxxx> wrote:
>>>>>>>>>>>>
>>>>>>>>>>>>>>
>>>>>>>>>>>>>> Why does the firmware map support hotplug entries?
>>>>>>>>>>>>>
>>>>>>>>>>>>> I assume:
>>>>>>>>>>>>>
>>>>>>>>>>>>> The firmware memmap was added primarily for x86-64 kexec (and still, is
>>>>>>>>>>>>> mostly used on x86-64 only IIRC). There, we had ACPI hotplug. When DIMMs
>>>>>>>>>>>>> get hotplugged on real HW, they get added to e820. Same applies to
>>>>>>>>>>>>> memory added via HyperV balloon (unless memory is unplugged via
>>>>>>>>>>>>> ballooning and you reboot ... the the e820 is changed as well). I assume
>>>>>>>>>>>>> we wanted to be able to reflect that, to make kexec look like a real reboot.
>>>>>>>>>>>>>
>>>>>>>>>>>>> This worked for a while. Then came dax/kmem. Now comes virtio-mem.
>>>>>>>>>>>>>
>>>>>>>>>>>>>
>>>>>>>>>>>>> But I assume only Andrew can enlighten us.
>>>>>>>>>>>>>
>>>>>>>>>>>>> @Andrew, any guidance here? Should we really add all memory to the
>>>>>>>>>>>>> firmware memmap, even if this contradicts with the existing
>>>>>>>>>>>>> documentation? (especially, if the actual firmware memmap will *not*
>>>>>>>>>>>>> contain that memory after a reboot)
>>>>>>>>>>>>
>>>>>>>>>>>> For some reason that patch is misattributed - it was authored by
>>>>>>>>>>>> Shaohui Zheng <shaohui.zheng@xxxxxxxxx>, who hasn't been heard from in
>>>>>>>>>>>> a decade. I looked through the email discussion from that time and I'm
>>>>>>>>>>>> not seeing anything useful. But I wasn't able to locate Dave Hansen's
>>>>>>>>>>>> review comments.
>>>>>>>>>>>
>>>>>>>>>>> Okay, thanks for checking. I think the documentation from 2008 is pretty
>>>>>>>>>>> clear what has to be done here. I will add some of these details to the
>>>>>>>>>>> patch description.
>>>>>>>>>>>
>>>>>>>>>>> Also, now that I know that esp. kexec-tools already don't consider
>>>>>>>>>>> dax/kmem memory properly (memory will not get dumped via kdump) and
>>>>>>>>>>> won't really suffer from a name change in /proc/iomem, I will go back to
>>>>>>>>>>> the MHP_DRIVER_MANAGED approach and
>>>>>>>>>>> 1. Don't create firmware memmap entries
>>>>>>>>>>> 2. Name the resource "System RAM (driver managed)"
>>>>>>>>>>> 3. Flag the resource via something like IORESOURCE_MEM_DRIVER_MANAGED.
>>>>>>>>>>>
>>>>>>>>>>> This way, kernel users and user space can figure out that this memory
>>>>>>>>>>> has different semantics and handle it accordingly - I think that was
>>>>>>>>>>> what Eric was asking for.
>>>>>>>>>>>
>>>>>>>>>>> Of course, open for suggestions.
>>>>>>>>>>
>>>>>>>>>> I'm still more of a fan of this being communicated by "System RAM"
>>>>>>>>>
>>>>>>>>> I was mentioning somewhere in this thread that "System RAM" inside a
>>>>>>>>> hierarchy (like dax/kmem) will already be basically ignored by
>>>>>>>>> kexec-tools. So, placing it inside a hierarchy already makes it look
>>>>>>>>> special already.
>>>>>>>>>
>>>>>>>>> But after all, as we have to change kexec-tools either way, we can
>>>>>>>>> directly go ahead and flag it properly as special (in case there will
>>>>>>>>> ever be other cases where we could no longer distinguish it).
>>>>>>>>>
>>>>>>>>>> being parented especially because that tells you something about how
>>>>>>>>>> the memory is driver-managed and which mechanism might be in play.
>>>>>>>>>
>>>>>>>>> The could be communicated to some degree via the resource hierarchy.
>>>>>>>>>
>>>>>>>>> E.g.,
>>>>>>>>>
>>>>>>>>> [root@localhost ~]# cat /proc/iomem
>>>>>>>>> ...
>>>>>>>>> 140000000-33fffffff : Persistent Memory
>>>>>>>>> 140000000-1481fffff : namespace0.0
>>>>>>>>> 150000000-33fffffff : dax0.0
>>>>>>>>> 150000000-33fffffff : System RAM (driver managed)
>>>>>>>>>
>>>>>>>>> vs.
>>>>>>>>>
>>>>>>>>> :/# cat /proc/iomem
>>>>>>>>> [...]
>>>>>>>>> 140000000-333ffffff : virtio-mem (virtio0)
>>>>>>>>> 140000000-147ffffff : System RAM (driver managed)
>>>>>>>>> 148000000-14fffffff : System RAM (driver managed)
>>>>>>>>> 150000000-157ffffff : System RAM (driver managed)
>>>>>>>>>
>>>>>>>>> Good enough for my taste.
>>>>>>>>>
>>>>>>>>>> What about adding an optional /sys/firmware/memmap/X/parent attribute.
>>>>>>>>>
>>>>>>>>> I really don't want any firmware memmap entries for something that is
>>>>>>>>> not part of the firmware provided memmap. In addition,
>>>>>>>>> /sys/firmware/memmap/ is still a fairly x86_64 specific thing. Only mips
>>>>>>>>> and two arm configs enable it at all.
>>>>>>>>>
>>>>>>>>> So, IMHO, /sys/firmware/memmap/ is definitely not the way to go.
>>>>>>>>
>>>>>>>> I think that's a policy decision and policy decisions do not belong in
>>>>>>>> the kernel. Give the tooling the opportunity to decide whether System
>>>>>>>> RAM stays that way over a kexec. The parenthetical reference otherwise
>>>>>>>> looks out of place to me in the /proc/iomem output. What makes it
>>>>>>>> "driver managed" is how the kernel handles it, not how the kernel
>>>>>>>> names it.
>>>>>>>
>>>>>>> At least, virtio-mem is different. It really *has to be handled* by the
>>>>>>> driver. This is not a policy. It's how it works.
>>>>>
>>>>> ...but that's not necessarily how dax/kmem works.
>>>>>
>>>>
>>>> Yes, and user space could still take that memory and add it to the
>>>> firmware memmap if it really wants to. It knows that it is special. It
>>>> can figure out that it belongs to a dax device using /proc/iomem.
>>>>
>>>>>>>
>>>>>>
>>>>>> Oh, and I don't see why "System RAM (driver managed)" would hinder any
>>>>>> policy in user case to still do what it thinks is the right thing to do
>>>>>> (e.g., for dax).
>>>>>>
>>>>>> "System RAM (driver managed)" would mean: Memory is not part of the raw
>>>>>> firmware memmap. It was detected and added by a driver. Handle with
>>>>>> care, this is special.
>>>>>
>>>>> Oh, no, I was more reacting to your, "don't update
>>>>> /sys/firmware/memmap for the (driver managed) range" choice as being a
>>>>> policy decision. It otherwise feels to me "System RAM (driver
>>>>> managed)" adds confusion for casual users of /proc/iomem and for clued
>>>>> in tools they have the parent association to decide policy.
>>>>
>>>> Not sure if I understand correctly, so bear with me :).
>>>>
>>>> Adding or not adding stuff to /sys/firmware/memmap is not a policy
>>>> decision. If it's not part of the raw firmware-provided memmap, it has
>>>> nothing to do in /sys/firmware/memmap. That's what the documentation
>>>> from 2008 tells us.
>>>
>>> It just occurs to me that there are valid cases for both wanting to
>>> start over with driver managed memory with a kexec and keeping it in
>>> the map.
>>
>> Yes, there might be valid cases. My gut feeling is that in the general
>> case, you want to let the kexec kernel implement a policy/ let the user
>> in the new system decide.
>>
>> But as I said, you can implement in kexec-tools whatever policy you
>> want. It has access to all information.
>
> Right, so why is a new type needed if all the information is there by
> other means?
You mean "System RAM (driver managed)" in /proc/iomem? See below for more.
>
>>> Consider the case of EFI Special Purpose (SP) Memory that is
>>> marked EFI Conventional Memory with the SP attribute. In that case the
>>> firmware memory map marked it as conventional RAM, but the kernel
>>> optionally marks it as System RAM vs Soft Reserved. The 2008 patch
>>> simply does not consider that case. I'm not sure strict textualism
>>> works for coding decisions.
>>
>> I am no expert on that matter (esp EFI). But looking at the users of
>> firmware_map_add_early(), the single user is in arch/x86/kernel/e820.c
>> . So the single source of /sys/firmware/memmap is (besides hotplug) e820.
>>
>> "'e820_table_firmware': the original firmware version passed to us by
>> the bootloader - not modified by the kernel. ... inform the user about
>> the firmware's notion of memory layout via /sys/firmware/memmap"
>> (arch/x86/kernel/e820.c)
>>
>> How is the EFI Special Purpose (SP) Memory represented in e820?
>> /sys/firmware/memmap is really simple: just dump in e820. No policies IIUC.
>
> e820 now has a Soft Reserved translation for this which means "try to
> reserve, but treat as System RAM is ok too". It seems generically
> useful to me that the toggle for determining whether Soft Reserved or
> System RAM shows up /sys/firmware/memmap is a determination that
> policy can make. The kernel need not preemptively block it.
So, I think I have to clarify something here. We do have two ways to kexec
1. kexec_load(): User space (kexec-tools) crafts the memmap (e.g., using
/sys/firmware/memmap on x86-64) and selects memory where to place the
kexec images (e.g., using /proc/iomem)
2. kexec_file_load(): The kernel reuses the (basically) raw firmware
memmap and selects memory where to place kexec images.
We are talking about changing 1, to behave like 2 in regards to
dax/kmem. 2. does currently not add any hotplugged memory to the
fixed-up e820, and it should be fixed regarding hotplugged DIMMs that
would appear in e820 after a reboot.
Now, all these policy discussions are nice and fun, but I don't really
see a good reason to (ab)use /sys/firmware/memmap for that (e.g., parent
properties). If you want to be able to make this configurable, then
e.g., add a way to configure this in the kernel (for example along with
kmem) to make 1. and 2. behave the same way. Otherwise, you really only
can change 1.
Now, let's clarify what I want regarding virtio-mem:
1. kexec should not add virtio-mem memory to the initial firmware
memmap. The driver has to be in charge as discussed.
2. kexec should not place kexec images onto virtio-mem memory. That
would end badly.
3. kexec should still dump virtio-mem memory via kdump.
This has to work when using kexec_load() or kexec_file_load(). This has
to theoretically work on different architectures (especially, without
/sys/firmware/memmap). kexec-tools has to have access to that
information to figure out what to do.
Regarding 1:
- kexec_file_load(): works out of the box currently.
- kexec_load(): Don't create entries in /sys/firmware/memmap (for
reasons discussed)
Regarding 2:
- kexec_file_load(): tag the resources as IORESOURCE_MEM_DRIVER_MANAGED
(inspired by Eric)
- kexec_load(): indicate the memory as "System RAM (driver managed)"
Regarding 3:
- Same as 2. kexec-tools need to be thought to properly consider the
memory during kdump.
Now, you are asking, "why System RAM (driver managed)". I don't think
it's strictly needed right now, but it feels cleaner. E.g., for
virtio-mem the current plan is to have /proc/iomem look like
:/# cat /proc/iomem
[...]
140000000-333ffffff : virtio-mem (virtio0)
140000000-147ffffff : System RAM (driver managed)
148000000-14fffffff : System RAM (driver managed)
150000000-157ffffff : System RAM (driver managed)
One could judge by looking at the hierarchy, that this memory is
special. kexec-tools will skip it currently in either form.
If we all agree here, that we can drop it, then let's drop it,
especially if it would allow dax/kmem to use the same mechanism I am
proposing here for virtio-mem.
Now, it would be fairly simple to add a config option for dax/kmem,
making it configurable in the kernel, whether to add memory via
MHP_DRIVER_MANAGED or just as we do now. It would contradict with the
"raw firmware/prov..." description of /sys/firmware/memmap, but hey,
somebody explicitly configured it, so it can't be wrong.
--
Thanks,
David / dhildenb