Subject: Re: [PATCH 05/10] KVM: x86: MMU: Use for_each_rmap_spte macro instead
 of pte_list_walk()
To: Marcelo Tosatti <mtosatti@redhat.com>
References: <20151112204849.ba920599a8426d7196a0df73@lab.ntt.co.jp>
 <20151112205245.6bd773737cfa78422dac5a79@lab.ntt.co.jp>
 <20151113214727.GA29103@amt.cnet> <20151114092047.GA25627@amt.cnet>
Cc: pbonzini@redhat.com, kvm@vger.kernel.org, linux-kernel@vger.kernel.org
From: Takuya Yoshikawa <yoshikawa_takuya_b1@lab.ntt.co.jp>
Message-ID: <564944A3.5080005@lab.ntt.co.jp>
Date: Mon, 16 Nov 2015 11:51:15 +0900
User-Agent: Mozilla/5.0 (Windows NT 6.1; WOW64; rv:38.0) Gecko/20100101
 Thunderbird/38.3.0
MIME-Version: 1.0
In-Reply-To: <20151114092047.GA25627@amt.cnet>
Content-Type: text/plain; charset=windows-1252; format=flowed
Content-Transfer-Encoding: 7bit
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 2589
Lines: 79

On 2015/11/14 18:20, Marcelo Tosatti wrote:

> The actual issue is this: a higher level page that had, under its children,
> no out of sync pages, now, due to your addition, a child that is unsync:
>
> initial state:
> 	level1
>
> final state:
>
> 	level1 -x-> level2 -x-> level3
>
> Where -x-> are the links created by this pagefault fixing round.
>
> If _any_ page under you is unsync (not necessarily the ones this
> pagefault is accessing), you have to mark parents unsync.

I understand this, but I don't think my patch will break this.

What kvm_mmu_mark_parents_unsync() does is:

   for each p_i in sp->parent_ptes rmap chain
     mark_unsync(p_i);

Then, mark_unsync() finds the parent sp including that p_i to
set ->unsync_child_bitmap and increment ->unsync_children if
necessary.  It may also call kvm_mmu_mark_parents_unsync()
recursively.

I understand we need to tell the parents "you have an unsync
child/descendant" until this information reaches the top level
by that recursive calls.

But since these recursive calls cannot come back to the starting sp,
the child->parent graph has no loop, each mark_unsync(p_i) will not
be affected by other parents in that sp->parent_ptes rmap chain,
from which we started the recursive calls.


As the following code shows, my patch does mark_unsync(parent_pte)
separately, and then mmu_page_add_parent_pte(vcpu, sp, parent_pte):

> -		} else if (sp->unsync)
> +			if (parent_pte)
> +				mark_unsync(parent_pte);
> +		} else if (sp->unsync) {
>  			kvm_mmu_mark_parents_unsync(sp);
> +			if (parent_pte)
> +				mark_unsync(parent_pte);
> +		}
> +		mmu_page_add_parent_pte(vcpu, sp, parent_pte);

So, as you worried, during each mark_unsync(p_i) is processed,
this parent_pte does not exist in that sp->parent_ptes rmap chain.

But as I explained above, this does not change anything about what
each mark_unsync(p_i) call does, so keeps the original behaviour.


By the way, I think "kvm_mmu_mark_parents_unsync" and "mark_unsync"
do not tell what they actually do well. When I first saw the names,
I thought they would just set the parents' sp->unsync.

To reflect the following meaning better, it should be
propagate_unsync(_to_parents) or something:

   Tell the parents "you have an unsync child/descendant"
   until this unsync information reaches the top level


Thanks,
   Takuya


--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/