Re: [Qemu-devel] [PATCH v2 0/6] proposal to make hostmem listener RAM unplug safe

[Paolo Bonzini]

Il 06/05/2013 03:42, liu ping fan ha scritto:
> On Sat, May 4, 2013 at 5:53 PM, Paolo Bonzini <address@hidden> wrote:
>> Il 03/05/2013 04:45, Liu Ping Fan ha scritto:
>>> v1->v2:
>>>   1.split RCU prepared style update and monitor the RAM-Device refcnt into 
>>> two patches (patch 2,4)
>>>   2.introduce AddrSpaceMem, which is similar to HostMem, but based on 
>>> address space, while
>>>     the original HostMem only server system memory address space
>>
>> This looks suspiciously similar to FlatView, doesn't it?
>>
> FlatView is used for all the listeners, including for mmio dispatching,
> which aims to mapping from hwaddr to DeviceState for dispatching service.
> While here, we mapping from hwaddr to hva.
> 
>> Perhaps the right thing to do is to add the appropriate locking and
>> RCU-style updating to address_space_update_topology and
> 
> RCU implementation is data struct related,  and each listener has its
> local table, so I think it is more reasonable to implement them
> separately.

I mentioned address_space_update_topology simply because it is where the
FlatView is replaced.

>> memory_region_find.   (And replacing flatview_destroy with ref/unref
>> similar to HostMem in your patch 2).  Then just switch dataplane to use
>> memory_region_find...
>>
> In fact, I think, HostMem listener can be an substitute for
> cpu_physical_memory_map(),  the main issue can be the migration
> support.  But before getting big patches, I hope to have this smaller
> and simpler one.

I think replacing HostMem with FlatView is a smaller patch than these
ones.  I'll try to make a prototype.

Paolo