On Thu, 02/12 15:29, Paolo Bonzini wrote: > > > On 12/02/2015 06:21, Fam Zheng wrote: > > Before processing a request, virtio-scsi dataplane will check if the > > backend runs on the same context with it. If not, it has to be moved, > > with bdrv_set_aio_context. > > > > However this function is unsafe to be called from IOThread outside BQL. > > The reason is that it calls bdrv_drain_all(), to acquire and drain all > > existing BDS. Therefore there is a deadlock problem. > > > > Fix it by offloading the bdrv_set_aio_context to main loop thread, > > through a BH (#1). This main loop BH will set the context, then notify > > the calling thread with another BH (#2). In BH (#2), we can continue the > > virtio-scsi request processing. > > > > A queue is added to VirtIOSCSI for tracking the pending requests, so in > > virtio_scsi_dataplane_stop, we have to drain them for migration. > > > > Signed-off-by: Fam Zheng <f...@redhat.com> > > Could you just do set_aio_context for all devices when starting > dataplane? For example with a new scsi_device_set_aio_context function > (and a new method in SCSIDeviceClass). > > Maybe I'm missing the obvious. :) >
Per VQ IOThread stills needs it, in the case that guest switches the VQ of a disk on the fly. Fam