| Message ID | 1309348641-20061-14-git-send-email-alevy@redhat.com |
|---|---|
| State | New |
| Headers | show |
Hi, > + case QXL_IO_DESTROY_ALL_SURFACES_ASYNC: > + d->mode = QXL_MODE_UNDEFINED; Should go to the async thread. cheers, Gerd
On Wed, Jun 29, 2011 at 03:09:36PM +0200, Gerd Hoffmann wrote: > Hi, > > >+ case QXL_IO_DESTROY_ALL_SURFACES_ASYNC: > >+ d->mode = QXL_MODE_UNDEFINED; > > Should go to the async thread. doesn't it make more sense to do all state changes from the vcpu thread? async thread can run much later, if you have a QXL_IO_DESTROY_ALL_SURFACES_ASYNC followed by a QXL_IO_CREATE_PRIMARY_ASYNC where the driver did not wait for the completion of the ASYNC first, I would still like to support that, but it won't work if I move this to the async thread. > > cheers, > Gerd
On 06/29/11 16:29, Alon Levy wrote: > On Wed, Jun 29, 2011 at 03:09:36PM +0200, Gerd Hoffmann wrote: >> Hi, >> >>> + case QXL_IO_DESTROY_ALL_SURFACES_ASYNC: + d->mode = >>> QXL_MODE_UNDEFINED; >> >> Should go to the async thread. > > doesn't it make more sense to do all state changes from the vcpu > thread? async thread can run much later, if you have a > QXL_IO_DESTROY_ALL_SURFACES_ASYNC followed by a > QXL_IO_CREATE_PRIMARY_ASYNC where the driver did not wait for the > completion of the ASYNC first, I would still like to support that, > but it won't work if I move this to the async thread. I think we should disallow doing any I/O ops while one is in progress (except maybe QXL_IO_LOG). Most I/O commands are I/O commands because they either needed for device setup or must be synchronous anyway. QXL_IO_CREATE_PRIMARY_ASYNC wasn't exactly clever designed I think. Would have been better to enter native mode with another I/O command, then send the create request for the primary through the rings like all other surface commands. But given it is a rare event it isn't that a big issue either. cheers, Gerd
On Wed, Jun 29, 2011 at 05:00:20PM +0200, Gerd Hoffmann wrote: > On 06/29/11 16:29, Alon Levy wrote: > >On Wed, Jun 29, 2011 at 03:09:36PM +0200, Gerd Hoffmann wrote: > >>Hi, > >> > >>>+ case QXL_IO_DESTROY_ALL_SURFACES_ASYNC: + d->mode = > >>>QXL_MODE_UNDEFINED; > >> > >>Should go to the async thread. > > > >doesn't it make more sense to do all state changes from the vcpu > >thread? async thread can run much later, if you have a > >QXL_IO_DESTROY_ALL_SURFACES_ASYNC followed by a > >QXL_IO_CREATE_PRIMARY_ASYNC where the driver did not wait for the > >completion of the ASYNC first, I would still like to support that, > >but it won't work if I move this to the async thread. > > I think we should disallow doing any I/O ops while one is in > progress (except maybe QXL_IO_LOG). Most I/O commands are I/O > commands because they either needed for device setup or must be > synchronous anyway. > > QXL_IO_CREATE_PRIMARY_ASYNC wasn't exactly clever designed I think. > Would have been better to enter native mode with another I/O > command, then send the create request for the primary through the > rings like all other surface commands. But given it is a rare event > it isn't that a big issue either. > So - leave or change? I prefer to leave.. Maybe add a "pending async" flag to catch such occasions? > cheers, > Gerd > >
Hi,
> Maybe add a "pending async" flag to catch such occasions?
Yes, good idea.
cheers,
Gerd
diff --git a/hw/qxl.c b/hw/qxl.c index f158d45..b794b2c 100644 --- a/hw/qxl.c +++ b/hw/qxl.c @@ -1245,6 +1245,7 @@ static void ioport_write(void *opaque, uint32_t addr, uint32_t val) qxl_spice_destroy_surface_wait(d, val); break; case QXL_IO_DESTROY_ALL_SURFACES: + d->mode = QXL_MODE_UNDEFINED; qxl_spice_destroy_surfaces(d); break; case QXL_IO_FLUSH_SURFACES: @@ -1302,9 +1303,10 @@ static void ioport_write(void *opaque, uint32_t addr, uint32_t val) async->update_area = d->ram->update_area; async->update_surface = d->ram->update_surface; goto async_common; + case QXL_IO_DESTROY_ALL_SURFACES_ASYNC: + d->mode = QXL_MODE_UNDEFINED; case QXL_IO_NOTIFY_OOM_ASYNC: case QXL_IO_DESTROY_SURFACE_ASYNC: - case QXL_IO_DESTROY_ALL_SURFACES_ASYNC: async = qemu_mallocz(sizeof(*async)); async_common: async->port = io_port;