Re: [PATCH 0/3] recover hardware corrupted page by virtio balloon

[David Hildenbrand]

On 01.06.22 04:17, zhenwei pi wrote:
> On 5/31/22 12:08, Jue Wang wrote:
>> On Mon, May 30, 2022 at 8:49 AM Peter Xu <peterx@xxxxxxxxxx> wrote:
>>>
>>> On Mon, May 30, 2022 at 07:33:35PM +0800, zhenwei pi wrote:
>>>> A VM uses RAM of 2M huge page. Once a MCE(@HVAy in [HVAx,HVAz)) occurs, the
>>>> 2M([HVAx,HVAz)) of hypervisor becomes unaccessible, but the guest poisons 4K
>>>> (@GPAy in [GPAx, GPAz)) only, it may hit another 511 MCE ([GPAx, GPAz)
>>>> except GPAy). This is the worse case, so I want to add
>>>>   '__le32 corrupted_pages' in struct virtio_balloon_config, it is used in the
>>>> next step: reporting 512 * 4K 'corrupted_pages' to the guest, the guest has
>>>> a chance to isolate the other 511 pages ahead of time. And the guest
>>>> actually loses 2M, fixing 512*4K seems to help significantly.
>>>
>>> It sounds hackish to teach a virtio device to assume one page will always
>>> be poisoned in huge page granule.  That's only a limitation to host kernel
>>> not virtio itself.
>>>
>>> E.g. there're upstream effort ongoing with enabling doublemap on hugetlbfs
>>> pages so hugetlb pages can be mapped in 4k with it.  It provides potential
>>> possibility to do page poisoning with huge pages in 4k too.  When that'll
>>> be ready the assumption can go away, and that does sound like a better
>>> approach towards this problem.
>>
>> +1.
>>
>> A hypervisor should always strive to minimize the guest memory loss.
>>
>> The HugeTLB double mapping enlightened memory poisoning behavior (only
>> poison 4K out of a 2MB huge page and 4K in guest) is a much better
>> solution here. To be completely transparent, it's not _strictly_
>> required to poison the page (whatever the granularity it is) on the
>> host side, as long as the following are true:
>>
>> 1. A hypervisor can emulate the _minimized_ (e.g., 4K) the poison to the guest.
>> 2. The host page with the UC error is "isolated" (could be PG_HWPOISON
>> or in some other way) and prevented from being reused by other
>> processes.
>>
>> For #2, PG_HWPOISON and HugeTLB double mapping enlightened memory
>> poisoning is a good solution.
>>
>>>
>>>>
>>>>>
>>>>> I assume when talking about "the performance memory drops a lot", you
>>>>> imply that this patch set can mitigate that performance drop?
>>>>>
>>>>> But why do you see a performance drop? Because we might lose some
>>>>> possible THP candidates (in the host or the guest) and you want to plug
>>>>> does holes? I assume you'll see a performance drop simply because
>>>>> poisoning memory is expensive, including migrating pages around on CE.
>>>>>
>>>>> If you have some numbers to share, especially before/after this change,
>>>>> that would be great.
>>>>>
>>>>
>>>> The CE storm leads 2 problems I have even seen:
>>>> 1, the memory bandwidth slows down to 10%~20%, and the cycles per
>>>> instruction of CPU increases a lot.
>>>> 2, the THR (/proc/interrupts) interrupts frequently, the CPU has to use a
>>>> lot time to handle IRQ.
>>>
>>> Totally no good knowledge on CMCI, but if 2) is true then I'm wondering
>>> whether it's necessary to handle the interrupts that frequently.  When I
>>> was reading the Intel CMCI vector handler I stumbled over this comment:
>>>
>>> /*
>>>   * The interrupt handler. This is called on every event.
>>>   * Just call the poller directly to log any events.
>>>   * This could in theory increase the threshold under high load,
>>>   * but doesn't for now.
>>>   */
>>> static void intel_threshold_interrupt(void)
>>>
>>> I think that matches with what I was thinking..  I mean for 2) not sure
>>> whether it can be seen as a CMCI problem and potentially can be optimized
>>> by adjust the cmci threshold dynamically.
>>
>> The CE storm caused performance drop is caused by the extra cycles
>> spent by the ECC steps in memory controller, not in CMCI handling.
>> This is observed in the Google fleet as well. A good solution is to
>> monitor the CE rate closely in user space via /dev/mcelog and migrate
>> all VMs to another host once the CE rate exceeds some threshold.
>>
>> CMCI is a _background_ interrupt that is not handled in the process
>> execution context and its handler is setup to switch to poll (1 / 5
>> min) mode if there are more than ~ a dozen CEs reported via CMCI per
>> second.
>>>
>>> --
>>> Peter Xu
>>>
> 
> Hi, Andrew, David, Naoya
> 
> According to the suggestions, I'd give up the improvement of memory 
> failure on huge page in this series.
> 
> Is it worth recovering corrupted pages for the guest kernel? I'd follow 
> your decision.

Well, as I said, I am not sure if we really need/want this for a handful
of 4k poisoned pages in a VM. As I suspected, doing so might primarily
be interesting for some sort of de-fragmentation (allow again a higher
order page to be placed at the affected PFNs), not because of the slight
reduction of available memory. A simple VM reboot would get the job
similarly done.

As the poisoning refcount code is already a bit shaky as I learned
recently in the context of memory offlining, I do wonder if we really
want to expose the unpoisoning code outside of debugfs (hwpoison) usage.

Interestingly, unpoison_memory() documents: "This is only done on the
software-level, so it only works for linux injected failures, not real
hardware failures" -- ehm?

-- 
Thanks,

David / dhildenb