On Wed, May 11 2022, Jason Wang [off-list ref] wrote:

On Tue, May 10, 2022 at 7:32 PM Michael S. Tsirkin [off-list ref] wrote:

quoted

On Sat, May 07, 2022 at 03:19:53PM +0800, Jason Wang wrote:

quoted

diff --git a/include/linux/virtio_config.h b/include/linux/virtio_config.h
index d8a2340f928e..23f1694cdbd5 100644
--- a/include/linux/virtio_config.h
+++ b/include/linux/virtio_config.h

@@ -256,6 +256,18 @@ void virtio_device_ready(struct virtio_device *dev)
      unsigned status = dev->config->get_status(dev);

      BUG_ON(status & VIRTIO_CONFIG_S_DRIVER_OK);
+
+     /*
+      * The virtio_synchronize_cbs() makes sure vring_interrupt()
+      * will see the driver specific setup if it sees vq->broken
+      * as false.
+      */
+     virtio_synchronize_cbs(dev);

since you mention vq->broken above, maybe add
        "set vq->broken to false"

Ok.

quoted

quoted

+     __virtio_unbreak_device(dev);
+     /*
+      * The transport is expected ensure the visibility of

to ensure

Will fix.

quoted

quoted

+      * vq->broken

let's add: "visibility by vq callbacks"

Sure.

quoted

quoted

before setting VIRTIO_CONFIG_S_DRIVER_OK.
+      */

Can I see some analysis of existing transports showing
this is actually the case for them?

Yes.

quoted

And maybe add a comment near set_status to document the
requirement.

For PCI and MMIO, we can quote the memory-barriers.txt or explain that
wmb() is not needed before the MMIO writel().
For CCW, it looks not obvious, it looks to me the IO was submitted via
__ssch() which has an inline assembly.  Cornelia and Hali, could you
help me to understand if and how did virtio_ccw_set_status() can
ensure the visibility of the previous driver setup and vq->broken
here?

I'm not sure I completely understand the question here, but let me try:

virtio_ccw_set_status() uses a channel command to set the status, with
the interesting stuff done inside ccw_io_helper(). That function
- takes the subchannel lock, disabling interrupts
- does the ssch; this instruction will fail if there's already another
  I/O in progress, or an interrupt is pending for the subchannel; on
  success, it is guaranteed that we'll get an interrupt eventually
- unlock the subchannel, and wait for the interupt handler to eventually
  process the interrupt, so I guess it should see the vq->broken value?

If the I/O fails, virtio_ccw_set_status() will revert its internal
status to the old value.

Thanks

quoted

quoted

      dev->config->set_status(dev, status | VIRTIO_CONFIG_S_DRIVER_OK);
 }

_______________________________________________
Virtualization mailing list
Virtualization@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/virtualization