Re: [PATCH 01/11] Add shared memory hypercall to PV Linux guest.

[Avi Kivity]

On 11/03/2009 09:16 AM, Gleb Natapov wrote:
> > > I have both! Do you want me to drop version?
> > >        
> > Yes.  Once a kernel is released you can't realistically change the version.
> >
> >      
> Why not? If version doesn't match apf will not be used.
>    

Then you cause a large performance regression (assuming apf is any 
good).  So there will be a lot of pressure to modify things 
incrementally via feature bits.

>    
> > > > Some documentation for this?
> > > >
> > > > Also, the name should reflect the pv pagefault use.  For other uses
> > > > we can register other areas.
> > > >
> > > >          
> > > I wanted it to be generic, but I am fine with making it apf specific.
> > > It will allow to make it smaller too.
> > >        
> > Maybe we can squeeze it into the page-fault error code?
> >
> >      
> apf has to pass two things into a guest kernel:
>   - event type (page not present/wake up)
>   - unique token
> Error code has 32 bits and at least 1 of them should indicate that this
> is apf another one should indicate event type so this leaves us 30 bits
> for a token. 12 bits of a token is used to store vcpu id this leaves 18
> bits for unique per vcpu id. Yes this may be enough. I don't think it is
> realistic to have more then 200000 outstanding apfs per vcpu. Alternately
> we can use CR2 to pass a token.
>    

Or a combination of pfec and cr2, yes.

> > > > would solve this.  I prefer using put_user() though than a permanent
> > > > get_user_pages().
> > > >
> > > >          
> > > I want to prevent it from been swapped out.
> > >        
> > Since you don't prevent the page fault handler or code from being
> > swapped out, you don't get anything out of it.
> >
> >      
> Performance. Currently it is accessed on each page fault and to access
> it gup+kmap should be done each and every time.
>    

put_user() is just as fast as a kmap, and don't prevent page migration 
or defragmentation.

Note we still have to mark_page_dirty() unless we want to chase live 
migration bugs.

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

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