Re: [PATCH 09/10 RFC] KVM: x86: MMU: Move parent_pte handling from kvm_mmu_get_page() to link_shadow_page()

From: Paolo Bonzini <pbonzini@redhat.com>
To: Takuya Yoshikawa <yoshikawa_takuya_b1@lab.ntt.co.jp>
Cc: kvm@vger.kernel.org, linux-kernel@vger.kernel.org
Subject: Re: [PATCH 09/10 RFC] KVM: x86: MMU: Move parent_pte handling from kvm_mmu_get_page() to link_shadow_page()
Date: Fri, 13 Nov 2015 11:51:06 +0100	[thread overview]
Message-ID: <5645C09A.4000907@redhat.com> (raw)
In-Reply-To: <564547AA.20002@lab.ntt.co.jp>

On 13/11/2015 03:15, Takuya Yoshikawa wrote:
> Actually, I don't understand why this is named kvm_mmu_put_page() for
> just removing parent_pte pointer from the sp->parent_ptes pointer chain.

Because it undoes kvm_mmu_get_page, I guess. :)

> 
>> On to kvm_mmu_get_page...
>>
>>          if (!direct) {
>>                  if (rmap_write_protect(vcpu, gfn))
>>                          kvm_flush_remote_tlbs(vcpu->kvm);
>>                  if (level > PT_PAGE_TABLE_LEVEL && need_sync)
>>                          kvm_sync_pages(vcpu, gfn);
>>
>> This seems fishy.
>>
>> need_sync is set if sp->unsync, but then the parents have not been
>> unsynced yet.
> 
> Reaching here means that kvm_mmu_get_page() could not return sp
> from inside the for_each_gfn_sp() loop above, so even without
> this patch, mark_unsync() has not been called.

You're right.

> Here, sp holds the new page allocated by kvm_mmu_alloc_page().
> One confusing thing is that hlist_add_head() right before this
> "if (!direct)" line has already added the new sp to the hash
> list, so it will be found by for_each_gfn_indirect_valid_sp()
> in kvm_sync_pages().
> 
> Because this sp is new and sp->unsync is not set,  kvm_sync_pages()
> will just skip it and look for other sp's whose ->unsync were found
> to be set in the for_each_gfn_sp() loop.
> 
> I'm not 100% sure if the existence of the parent_pte pointer in the
> newly created sp->parent_ptes chain alone makes any difference:

No, I don't think so.  Nothing needs the parent_ptes at this point:

- kvm_mmu_mark_parents_unsync, even in the existing code, it's called
before the new SPTE is created.

- as you said, kvm_mmu_prepare_zap_page can be called by kvm_sync_pages
but it will not operate on this page because its ->unsync is zero.

> So, "bool accessed" needs to be passed to kvm_mmu_get_page(). 

The "bool accessed" parameter is not necessary, I think.  It is only
false in the nested EPT case, and there's no reason not to set the
accessed bit *in the shadow page* if the host supports EPT
accessed/dirty bits.  I'll test and send a patch to remove the argument.

> But any way, we need to understand if mmu_page_add_parent_pte()
> really needs to be placed before the "if (!direct)" block.

No, I don't think so anymore.

I think these patches are fine as a starting point for further cleanups,
I'll push them to kvm/queue very soon.

Paolo