Re: [Xen-devel] [PATCH] mmap_vmcore: skip non-ram pages reported by hypervisors

[Vitaly Kuznetsov <vkuznets@redhat.com>]

David Vrabel <david.vrabel@citrix.com> writes:

> On 07/07/14 21:33, Andrew Morton wrote:
>> On Mon,  7 Jul 2014 17:05:49 +0200 Vitaly Kuznetsov <vkuznets@redhat.com> wrote:
>> 
>>> we have a special check in read_vmcore() handler to check if the page was
>>> reported as ram or not by the hypervisor (pfn_is_ram()). However, when
>>> vmcore is read with mmap() no such check is performed. That can lead to
>>> unpredictable results, e.g. when running Xen PVHVM guest memcpy() after
>>> mmap() on /proc/vmcore will hang processing HVMMEM_mmio_dm pages creating
>>> enormous load in both DomU and Dom0.
>
> Does make forward progress though?  Or is it ending up in a repeatedly
> retrying the same instruction?

If memcpy is using SSE2 optimization 16-byte 'movdqu' instruction never
finishes (repeatedly retrying to issue two 8-byte requests to
qemu-dm). qemu-dm decides that it's hitting 'Neither RAM nor known MMIO
space' and returns 8 0xff bytes for both of this requests (I was testing
with qemu-traditional).

>
> Is it failing on a ballooned page in a RAM region? Or is mapping non-RAM
> regions as well?

I wasn't using ballooning, it happens that oldmem has several (two in my
test) pages which are HVMMEM_mmio_dm but qemu-dm considers them being
neither ram nor mmio.

>
>>> Fix the issue by mapping each non-ram page to the zero page. Keep direct
>>> path with remap_oldmem_pfn_range() to avoid looping through all pages on
>>> bare metal.
>>>
>>> The issue can also be solved by overriding remap_oldmem_pfn_range() in
>>> xen-specific code, as remap_oldmem_pfn_range() was been designed for.
>>> That, however, would involve non-obvious xen code path for all x86 builds
>>> with CONFIG_XEN_PVHVM=y and would prevent all other hypervisor-specific
>>> code on x86 arch from doing the same override.
>
> The oldmem_pfn_is_ram() is Xen-specific but this problem (ballooned
> pages) must be common to KVM.  How does KVM handle this?

Is far as I'm concearned the issue was never hit with KVM. I *think* the
issue has something to do with the conjunction of 16-byte 'movdqu'
emulation for io pages in xen hypervisor, 8-byte event channel requests
and qemu-traditional. But even if it gets fixed on hypervisor side I
believe fixing the issue kernel-side still worth it as there are
non-fixed hypervisors out there (e.g. AWS EC2).

>
> David

-- 
  Vitaly