On 21/02/2018 13:22, Zhuocheng Ding wrote:
> On 2018年02月21日 05:29, Paolo Bonzini wrote:
>> On 17/02/2018 19:01, 丁卓成 wrote:
>>> In `set_spte`, where we prepare a new spte entry, we have this:
>>>> if (pte_access & ACC_WRITE_MASK) {
>>>> kvm_vcpu_mark_page_dirty(vcpu, gfn);
>>>> spte |= spte_shadow_dirty_mask(spte);
>>>> }
>>> And in `__direct_map`, we have this:
>>>> emulate = mmu_set_spte(vcpu, iterator.sptep, ACC_ALL,
>>>> write, level, gfn, pfn, prefault,
>>>> map_writable);
>>> If Guest issue a read to a page, and the corresponding spte is not present,
>>> `__direct_map` will be called to build the spte (and maybe parent sps as well)
>>> in EPT mode and SPT nonpaging mode. Since we pass ACC_ALL to `set_spte` in this
>>> case, even a read fault will cause the page to be marked dirty in dirty bitmap,
>>> and it will also set the D bit in spte.
>>
>> Hi, the "ACC_ALL" is fixed here in set_spte:
>>
>> if (host_writable)
>> spte |= SPTE_HOST_WRITEABLE;
>> else
>> pte_access &= ~ACC_WRITE_MASK;
>>
>> where host_writable is the last argument to mmu_set_spte.
>
> Ultimately, host_writable is set in try_async_pf (as the last argument), and
> it would be false only if:
>
> 1) Memory slot is readonly, or
>
> 2) Host page is readonly.
>
> So if the gfn is backed by writable host memory, a read fault to this gfn will
> still cause it to be marked dirty.
Indeed, this makes little difference most of the time (and improves
performance) but it may hurt performance when e.g. you have swapping of
guest memory.
Paolo
