On Sun, 2011-08-14 at 12:23 +0930, Rusty Russell wrote:

On Sat, 13 Aug 2011 11:51:01 +0300, Sasha Levin [off-list ref] wrote:

quoted

The MAC of a virtio-net device is located at the first field of the device
specific header. This header is located at offset 20 if the device doesn't
support MSI-X or offset 24 if it does.

Erk.  This means, in general, we have to do virtio_find_single_vq or
config->find_vqs before we examine any config options.

Look at virtio_blk, which has the same error.

Solutions in order of best to worst:
(1) Enable MSI-X before calling device probe.  This means reserving two
    vectors in virtio_pci_probe to ensure we *can* do this, I think.  Michael?

Do you mean reserving the vectors even before we probed the device for
MSI-X support? Wouldn't we need 3 vectors then? (config, input, output).

(2) Ensure ordering of "find_vqs then access config space" statically.  This
    probably means handing the vqs array to virtio_config_val, so noone
    can call it before they have their virtqueues.

Just noticed that only virtio-blk uses virtio_config_val(), while the
others are still doing 'if(virtio_has_feature()) vdev->config->get()',
I'll send patches to fix that regardless of what we end up doing here.

Did you want to pass the vq array to virtio_config_val() just to check
that they were already found? 

(3) Ensure ordering dynamically, ie. BUG_ON() if they haven't done
    find_vqs when they call the config accessors.

If (1) is too invasive for -stable, then we should rearrange the drivers
in separate patches (and cc: -stable), then fix it properly.

Good catch Sasha!

Cheers,
Rusty.

-- 

Sasha.