Avi Kivity wrote:
> Anthony Liguori wrote:
>> Avi Kivity wrote:
>>
>>> Anthony Liguori wrote:
>>>
>>>>>> static int mmu_topup_memory_caches(struct kvm_vcpu *vcpu)
>>>>>> {
>>>>>> int r;
>>>>>>
>>>>>> kvm_mmu_free_some_pages(vcpu);
>>>>>> r = mmu_topup_memory_cache(&vcpu->mmu_pte_chain_cache,
>>>>>> pte_chain_cache, 4);
>>>>>> if (r)
>>>>>> goto out;
>>>>>> r = mmu_topup_memory_cache(&vcpu->mmu_rmap_desc_cache,
>>>>>> rmap_desc_cache, 1);
>>>>>> if (r)
>>>>>> goto out;
>>>>>> r = mmu_topup_memory_cache_page(&vcpu->mmu_page_cache, 8);
>>>>>> if (r)
>>>>>> goto out;
>>>>>> r = mmu_topup_memory_cache(&vcpu->mmu_page_header_cache,
>>>>>> mmu_page_header_cache, 4);
>>>>>> out:
>>>>>> return r;
>>>>>> }
>>>>>>
>>>>> These are the (4, 1, 8, 4) values in the call to
>>>>> mmu_topup_memory_cache. Perhaps one of them is too low.
>>>>>
>>>> Sure. Would this be affected at all by your tpr patch?
>>> I believe not, but the code doesn't care what I believe.
>>>
>>>
>>>> IIUC, if this is the problem, it should be reproducible with the
>>>> latest git too?
>>>>
>>> One difference is that the tpr patch disables nx. That causes
>>> Windows to go into 32-bit paging mode (nice that it has both pae and
>>> nonpae in the same kernel), which may change things.
>>>
>>> You can try booting your host with nx disabled to get the same
>>> effect (or disable nx cpuid in kvm).
>>>
>>
>> I've disabled NX in KVM and that didn't reproduce the issue in the
>> current git.
>>
>> If I double all of the memory caches, I can't seem to reproduce.
>> However, as soon as I reduce rmap_desc_cache down to 1, I can reproduce.
>>
>> I'll try to see if just setting the rmap_desc_cache line to 2 is
>> enough to make the problem go away.
>>
>> How can the guest provoke this BUG() based on the cache size? Should
>> the cache size only affect performance?
>>
>>
>
> The memory caches are a little misnamed; they're really preallocation
> buffers.
>
> They serve two purposes: to avoid allocation in atomic contexts
> (that's no longer needed since preempt notifiers) and to avoid complex
> error recovery paths. We make sure there are enough objects to
> satisfy worst case behavior and assume all allocations will work.
FWIW, two people in IRC just reported the same BUG in kvm-48 and the
nightly snapshot. I asked them both to post more thorough descriptions
here. That leads me to suspect that the problem wasn't introduced by
your TPR optimization.
Regards,
Anthony Liguori
> Regarding the rmap memory cache failure, I can't think of a reason why
> we'll need to add more than one rmap entry per fault.
>
-------------------------------------------------------------------------
This SF.net email is sponsored by: Splunk Inc.
Still grepping through log files to find problems? Stop.
Now Search log events and configuration files using AJAX and a browser.
Download your FREE copy of Splunk now >> http://get.splunk.com/
_______________________________________________
kvm-devel mailing list
kvm-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/kvm-devel
