Re: [Qemu-devel] [PATCH] qxl: async I/O

[Alon Levy <alevy@redhat.com>]

On Thu, Jul 07, 2011 at 09:47:52AM +0200, Gerd Hoffmann wrote:
>   Hi,
> 
> >+void qxl_spice_update_area_async(PCIQXLDevice *qxl, uint32_t surface_id,
> >+                           struct QXLRect *area, struct QXLRect *dirty_rects,
> >+                           uint32_t num_dirty_rects, uint32_t clear_dirty_region,
> >+                           int async)
> >+{
> >+    qemu_mutex_lock(&qxl->ssd.wlock);
> >+    if (async) {
> >+        qxl->ssd.worker->update_area_async(qxl->ssd.worker, surface_id, area, dirty_rects,
> >+                                 num_dirty_rects, clear_dirty_region);
> >+    } else {
> >+        qxl->ssd.worker->update_area(qxl->ssd.worker, surface_id, area, dirty_rects,
> >+                                 num_dirty_rects, clear_dirty_region);
> >+    }
> >+    qemu_mutex_unlock(&qxl->ssd.wlock);
> >+}
> 
> We need a plan to handle backward compatibility here.  Older
> libspice-server versions don't have the update_area_async op.
> Option one is to just not support async mode with older libraries.
> Option two is to handle the request syncronously even though the
> guest has asked for async.  I'd tend to pick option one, that makes
> things easier with S3 support because we just can't do that in any
> way with an older libspice-server.

will do.

> 
> >      switch (io_port) {
> >+    case QXL_IO_UPDATE_AREA_ASYNC:
> >+    case QXL_IO_NOTIFY_OOM_ASYNC:
> >+    case QXL_IO_MEMSLOT_ADD_ASYNC:
> >+    case QXL_IO_CREATE_PRIMARY_ASYNC:
> >+    case QXL_IO_DESTROY_PRIMARY_ASYNC:
> >+    case QXL_IO_DESTROY_SURFACE_ASYNC:
> >+    case QXL_IO_DESTROY_ALL_SURFACES_ASYNC:
> >+        async = 1;
> >+        if (d->current_async != QXL_UNDEFINED_IO) {
> >+            qxl_guest_bug(d, "%d async started before last (%d) complete\n",
> >+                io_port, d->current_async);
> 
> Better return here, ignoring the invalid request?

I wasn't sure. Both are bound to be buggy. But yeah, I guess returning at least
makes qemu correct (especially if I don't use the cookies to remember all outstanding
operations).

> 
> cheers,
>   Gerd