Re: [Xen-devel] [PATCH v3] xen/balloon: Mark unallocated host memory as UNUSABLE

[Boris Ostrovsky]

On 11/26/18 2:57 PM, Igor Druzhinin wrote:
> On 26/11/2018 19:42, Boris Ostrovsky wrote:
>> On 11/26/18 12:10 PM, Igor Druzhinin wrote:
>>> On 26/11/2018 16:25, Boris Ostrovsky wrote:
>>>> On 11/25/18 8:00 PM, Igor Druzhinin wrote:
>>>>> On 20/12/2017 14:05, Boris Ostrovsky wrote:
>>>>>> Commit f5775e0b6116 ("x86/xen: discard RAM regions above the maximum
>>>>>> reservation") left host memory not assigned to dom0 as available for
>>>>>> memory hotplug.
>>>>>>
>>>>>> Unfortunately this also meant that those regions could be used by
>>>>>> others. Specifically, commit fa564ad96366 ("x86/PCI: Enable a 64bit BAR
>>>>>> on AMD Family 15h (Models 00-1f, 30-3f, 60-7f)") may try to map those
>>>>>> addresses as MMIO.
>>>>>>
>>>>>> To prevent this mark unallocated host memory as E820_TYPE_UNUSABLE (thus
>>>>>> effectively reverting f5775e0b6116) and keep track of that region as
>>>>>> a hostmem resource that can be used for the hotplug.
>>>>>>
>>>>>> Signed-off-by: Boris Ostrovsky <boris.ostrovsky@xxxxxxxxxx>
>>>>> This commit breaks Xen balloon memory hotplug for us in Dom0 with
>>>>> "hoplug_unpopulated" set to 1. The issue is that the common kernel
>>>>> memory onlining procedures require "System RAM" resource to be 1-st
>>>>> level. That means by inserting it under "Unusable memory" as the commit
>>>>> above does (intentionally or not) we make it 2-nd level and break memory
>>>>> onlining.
>>>> What do you mean by 1st and 2nd level?
>>>>
>>> I mean the level of a resource in IOMEM tree (the one that's printed
>>> from /proc/iomem). 1-st level means its parent is root and so on.
>> Ah, OK. Doesn't
>> additional_memory_resource()->insert_resource(iomem_resource) place the
>> RAM at 1st level? And if not, can we make it so?
>>
> That'd mean splitting "Unusable memory" resource. Since it's allocated
> from bootmem it has proven to be quite difficult but there are seem to
> be special functions available particularly for memory resource
> management operations that I've not yet experimented with. So the answer
> is probably - maybe yes but not straightforward.
>
>>>>> There are multiple ways to fix it depending on what was the intention of
>>>>> original commit and what exactly it tried to workaround. It seems it
>>>>> does several things at once:
>>>>> 1) Marks non-Dom0 host memory "Unusable memory" in resource tree.
>>>>> 2) Keeps track of all the areas safe for hotplug in Dom0
>>>>> 3) Changes allocation algorithms itself in balloon driver to use those areas
>>>> Pretty much. (3) is true in the sense that memory is first allocated
>>>> from hostmem_resource (which is non-dom0 RAM).
>>>>
>>>>> Are all the things above necessary to cover the issue in fa564ad96366
>>>>> ("x86/PCI: Enable a 64bit BAR on AMD Family 15h (Models 00-1f, 30-3f,
>>>>> 60-7f)")?
>>>> Not anymore, as far as that particular commit is concerned, but that's
>>>> because of 03a551734 ("x86/PCI: Move and shrink AMD 64-bit window to
>>>> avoid conflict") which was introduced after balloon patch. IIRC there
>>>> were some issues with fa564ad96366 unrelated to balloon.
>>>>
>>> If it's not a problem anymore IIUC, can we revert the change as it still
>>> breaks "hotplug_unpopulated=1" for the reasons I described above?
>> Since this seems to have broken existing feature this would be an
>> option. But before going that route I'd like to see if we can fix the patch.
>>
>> I have been unable to reproduce your problem. Can you describe what you did?
>>
> It doesn't happen on all configurations as sometimes the memory is
> successfully hotplugged to a hole depending on the size of Dom0 memory.
> But we reproduced it quite reliably with small Dom0 sizes like 752MB.
>
> XenServer is using this feature to hotplug additional memory for grant
> table operations so we started a VM and observed a stable hang.
>
>>>>> Can we remove "Unusable memory" resources as soon as we finished
>>>>> booting? Is removing on-demand is preferable over "shoot them all" in
>>>>> that case?
>>>> The concern is that in principle nothing prevents someone else to do
>>>> exact same thing fa564ad96366 did, which is grab something from right
>>>> above end of RAM as the kernel sees it. And that can be done at any point.
>>>>
>>> Nothing prevents - true, but that's plainly wrong from OS point of view
>>> to grab physical ranges for something without knowing what's actually
>>> behind on that platform. 
>> I am not sure I agree that this is plainly wrong. If not for BIOS issues
>> that 03a551734cf mentions I think what the original implementation of
>> fa564ad963 did was perfectly reasonable. Which is why I would prefer to
>> keep keep the hostmem resource *if possible*.
>>
> Exactly, those *are* BIOS issues and are not supposed to be workarounded
> by the OS. And as the next commit showed even the workaround didn't
> quite helped with it.
>
> I agree that having hotmem as a precaution is fine but only if there is
> a non-cringy way to keep things working with it which I'm not sure does
> exist.

We have most of the interfaces in the resource framework to do what we
want. I put together a semi-working prototype but the tricky part is
resource locking --- we need to remove a chunk from hostmem (which will
cause hostmem to be resized and possibly split), and insert this chunk
to iomem's top level as System RAM, all while holding resource_lock.

I haven't been able to come up with an acceptable interface for that.

Given that we are actually broken I guess I am OK with reverting the
patch, but please make sure this works on AMD boxes (I think family 15h
is what needs to be tested).

-boris



>
> Igor
>
>> -boris
>>
>>
>>> I think we shouldn't consider this as a valid
>>> thing to do and don't try to workaround initially incorrect code.
>>>
>>>> -boris
>>>>
>>>>> Does it even make sense to remove the 1-st level only restriction in
>>>>> kernel/resource.c ?
>>>>
> _______________________________________________
> Xen-devel mailing list
> Xen-devel@xxxxxxxxxxxxxxxxxxxx
> https://lists.xenproject.org/mailman/listinfo/xen-devel