Re: [PATCH 3/4] KVM: Count the number of dirty pages for dirty logging

[Takuya Yoshikawa <takuya.yoshikawa@gmail.com>]

On Fri, 23 Dec 2011 09:14:51 -0200
Marcelo Tosatti <mtosatti@redhat.com> wrote:

> > >btw mark_page_dirty() itself seems to assume mmu_lock protection that
> > >doesn't exist.  Marcelo?
> > >
> 
> Not mmu_lock protection, kvm->srcu protection.

But it just protects slot readers against updates and two, or more, threads
can call mark_page_dirty() concurrently?

What I am worring about here is the atomicity of bitmap updates.

	commit c8240bd6f0b4b1b21ffd36dd44114d05c7afe0c0
	Author: Alexander Graf <agraf@suse.de>
	Date:   Fri Oct 30 05:47:26 2009 +0000

		Use Little Endian for Dirty Bitmap

has changed set_bit() to the non-atomic version and nothing protects
dirty bits if mmu_lock is not held.

	The changelog has no explanation why using non-atomic version is safe.
	Some comment in the code may be worthwhile if it is really safe.

I want to see some clear reasoning now if possible.

	Takuya

> 
> > I want to hear the answer for this question.
> > 
> > Though I myself is reading the code, I cannot understand it thoroughly yet.
> > I wish if there were mmu_lock entry in locking.txt ...
> 
> Agreed.
> 

-- 
Takuya Yoshikawa <takuya.yoshikawa@gmail.com>