Re: [Qemu-devel] [PATCH v4 03/11] dataplane: add host memory mapping code

[Michael S. Tsirkin]

On Thu, Nov 29, 2012 at 01:45:19PM +0100, Stefan Hajnoczi wrote:
> On Thu, Nov 29, 2012 at 02:33:11PM +0200, Michael S. Tsirkin wrote:
> > On Thu, Nov 22, 2012 at 04:16:44PM +0100, Stefan Hajnoczi wrote:
> > > The data plane thread needs to map guest physical addresses to host
> > > pointers.  Normally this is done with cpu_physical_memory_map() but the
> > > function assumes the global mutex is held.  The data plane thread does
> > > not touch the global mutex and therefore needs a thread-safe memory
> > > mapping mechanism.
> > > 
> > > Hostmem registers a MemoryListener similar to how vhost collects and
> > > pushes memory region information into the kernel.  There is a
> > > fine-grained lock on the regions list which is held during lookup and
> > > when installing a new regions list.
> > > 
> > > When the physical memory map changes the MemoryListener callbacks are
> > > invoked.  They build up a new list of memory regions which is finally
> > > installed when the list has been completed.
> > > 
> > > Note that this approach is not safe across memory hotplug because mapped
> > > pointers may still be in used across memory unplug.  However, this is
> > > currently a problem for QEMU in general and needs to be addressed in the
> > > future.
> > 
> > Sounds like a serious problem.
> > I'm not sure I understand - do you say this currently a problem for QEMU
> > virtio? Coul you give an example please?
> 
> This is a limitation of the memory API but cannot be triggered by users
> today since we don't support memory hot unplug.  I'm just explaining
> that virtio-blk-data-plane has the same issue as hw/virtio-blk.c or any
> other device emulation code here.
> 
> Some more detail:
> 
> The issue is that hw/virtio-blk.c submits an asynchronous I/O request on
> the host with the guest buffer.  Then virtio-blk emulation returns back
> to the caller and continues QEMU execution.
> 
> It is unsafe to unplug memory while the I/O request is pending since
> there's no mechanism (e.g. refcount) to wait until the guest memory is
> no longer in use.
> 
> This is a known issue.  There's no way to trigger a problem today but we
> need to eventually enhance QEMU's memory API to handle this case.
> 
> Stefan

For this problem we would simply need to flush outstanding aio
before freeing memory for unplug, no refcount necessary.

This patch however introduces the issue in the frontend
and it looks like there won't be any way to solve
it without changing the API.

-- 
MST