Re: [Qemu-devel] [PATCH 3/4] exec, memory: Call to xen_modified_memory.

[Avi Kivity]

On 07/17/2012 09:36 PM, Stefano Stabellini wrote:
> On Tue, 17 Jul 2012, Avi Kivity wrote:
>> On 07/17/2012 04:59 PM, Anthony PERARD wrote:
>> >>
>> >> This is pretty ugly.  An alternative is to set up a periodic bitmap scan
>> >> that looks at the qemu dirty bitmap and calls xen_modified_memory() for
>> >> dirty page ranges, and clears the bitmap for the next pass.  Is it
>> >> workable?
>> > 
>> > I don't think a periodic scan can do anything useful, unfortunately.
>> 
>> Why not?
> 
> I vaguely remember that we used to have a bitmap years ago, but, aside from
> making the code much more complicated, it caused blue screens on
> intensive disk accesses.

Surely it was some bug, not the scan itself.

> 
> 
>> >> (is xen_modified_memory a hypercall, or does it maintain an in-memory
>> >> structure?)
>> > 
>> > It's an hypercall. The function do something (call the hypercall) only
>> > during migration, otherwise it return immediately.
>> 
>> I see.  I guess it isn't expensive for you because there isn't much dma
>> done by qemu usually with xen (unlike kvm where pv block devices are
>> implemented in qemu).
>> 
>> How about pushing the call into cpu_physical_memory_set_dirty_flags()?
>> Would that reduce the number of call sites?
> 
> Pushing the calls to cpu_physical_memory_set_dirty_flags and
> cpu_physical_memory_set_dirty_range would make the code much nicer.
> However being these functions in exec-obsolete.h, are they at risk of
> removal?

exec-obsolete.h just means don't add new call sites.  The functions
won't be removed, instead they'll be absorbed into the memory code with
different names and different implementations.

-- 
error compiling committee.c: too many arguments to function