Re: [announce] [patch] KVM paravirtualization for Linux

[Avi Kivity <avi@qumranet.com>]

Ingo Molnar wrote:

 

>> This is a little too good to be true.  Were both runs with the same 
>> KVM_NUM_MMU_PAGES?
>>     
>
> yes, both had the same elevated KVM_NUM_MMU_PAGES of 2048. The 'trunk' 
> run should have been labeled as: 'cr3 tree with paravirt turned off'. 
> That's not completely 'trunk' but close to it, and all other changes 
> (like elimination of unnecessary TLB flushes) are fairly applied to 
> both.
>   

Ok.  I guess there's a switch/switch back pattern in there.

> i also did a run with much less MMU cache pages of 256, and hackbench 1 
> stayed the same, while hackbench 5 numbers started fluctuating badly (i 
> think that workload if trashing the MMU cache badly).
>   

Yes, 256 is too low.

>   
>>> -    u64 *pae_root;
>>> +    u64 *pae_root[KVM_CR3_CACHE_SIZE];
>>>       
>> hmm.  wouldn't it be simpler to have pae_root always point at the 
>> current root?
>>     
>
> does that guarantee that it's available? I wanted to 'pin' the root 
> itself this way, to make sure that if a guest switches to it via the 
> cache, that it's truly available and a valid root. cr3 addresses are 
> non-virtual so this is the only mechanism available to guarantee that 
> the host-side memory truly contains a root pagetable.
>
>   

I meant

   u64 *pae_root_cache;
   u64 *pae_root; /* == pae_root_cache + 4*cache_index */

so that the rest of the code need not worry about the cache.

>>> +            vcpu->mmu.pae_root[j][i] = INVALID_PAGE;
>>> +        }
>>>     }
>>>     vcpu->mmu.root_hpa = INVALID_PAGE;
>>> }
>>>       
>> You keep the page directories pinned here. [...]
>>     
>
> yes.
>
>   
>> [...]  This can be a problem if a guest frees a page directory, and 
>> then starts using it as a regular page.  kvm sometimes chooses not to 
>> emulate a write to a guest page table, but instead to zap it, which is 
>> impossible when the page is freed.  You need to either unpin the page 
>> when that happens, or add a hypercall to let kvm know when a page 
>> directory is freed.
>>     
>
> the cache is zapped upon pagefaults anyway, so unpinning ought to be 
> possible. Which one would you prefer?
>   

It's zapped by the equivalent of mmu_free_roots(), right?  That's 
effectively unpinning it (by zeroing ->root_count).

However, kvm takes pagefaults even for silly things like setting (in 
hardware) or clearing (in software) the dirty bit.

>
>   
>>> +#define KVM_API_MAGIC 0x87654321
>>> +
>>>       
>> <linux/kvm.h> is the vmm userspace interface.  The guest/host 
>> interface should probably go somewhere else.
>>     
>
> yeah. kvm_para.h?
>
>   

Sounds good.

-- 
Do not meddle in the internals of kernels, for they are subtle and quick to panic.