CCW_CMD_READ_STATUS was introduced with revision 2 of virtio-ccw,
and drivers should only rely on it being implemented when they
negotiated at least that revision with the device.
However, virtio_ccw_get_status() issued READ_STATUS for any
device operating at least at revision 1. If the device accepts
READ_STATUS regardless of the negotiated revision (which it is
free to do), everything works as intended; a device rejecting the
command should also be handled gracefully. For correctness, we
should really limit the command to revision 2 or higher, though.
We also negotiated the revision to at most 1, as we never bumped
the maximum revision; let's do that now.
Fixes: 7d3ce5ab9430 ("virtio/s390: support READ_STATUS command for virtio-ccw")
Signed-off-by: Cornelia Huck <cohuck@redhat.com>
---
QEMU does not fence off READ_STATUS for revisions < 2, which is probably
why we never noticed this. I'm not aware of other hypervisors that do
fence it off, nor any that cannot deal properly with an unknown command.
Not sure whether this is stable worthy?
---
drivers/s390/virtio/virtio_ccw.c | 4 ++--
1 file changed, 2 insertions(+), 2 deletions(-)
On Fri, 12 Feb 2021 18:04:11 +0100
Cornelia Huck [off-list ref] wrote:
CCW_CMD_READ_STATUS was introduced with revision 2 of virtio-ccw,
and drivers should only rely on it being implemented when they
negotiated at least that revision with the device.
However, virtio_ccw_get_status() issued READ_STATUS for any
device operating at least at revision 1. If the device accepts
READ_STATUS regardless of the negotiated revision (which it is
free to do),
So, looking at the standard again, the device is actually required to
reject the READ_STATUS if only rev 1 had been negotiated... regardless
of that, I don't think we should change QEMU's behaviour, as it would
affect existing guests (they would lose access to the status bits as
observed by the device, including DEVICE_NEEDS_RESET.)
everything works as intended; a device rejecting the
command should also be handled gracefully. For correctness, we
should really limit the command to revision 2 or higher, though.
We also negotiated the revision to at most 1, as we never bumped
the maximum revision; let's do that now.
Fixes: 7d3ce5ab9430 ("virtio/s390: support READ_STATUS command for virtio-ccw")
Signed-off-by: Cornelia Huck <cohuck@redhat.com>
---
QEMU does not fence off READ_STATUS for revisions < 2, which is probably
why we never noticed this. I'm not aware of other hypervisors that do
fence it off, nor any that cannot deal properly with an unknown command.
Not sure whether this is stable worthy?
Maybe it is, given the MUST reject clause in the standard?
From: Halil Pasic <pasic@linux.ibm.com> Date: 2021-02-15 18:53:01
On Mon, 15 Feb 2021 12:47:02 +0100
Cornelia Huck [off-list ref] wrote:
On Fri, 12 Feb 2021 18:04:11 +0100
Cornelia Huck [off-list ref] wrote:
quoted
CCW_CMD_READ_STATUS was introduced with revision 2 of virtio-ccw,
and drivers should only rely on it being implemented when they
negotiated at least that revision with the device.
However, virtio_ccw_get_status() issued READ_STATUS for any
device operating at least at revision 1. If the device accepts
READ_STATUS regardless of the negotiated revision (which it is
free to do),
So, looking at the standard again, the device is actually required to
reject the READ_STATUS if only rev 1 had been negotiated... regardless
of that, I don't think we should change QEMU's behaviour, as it would
affect existing guests (they would lose access to the status bits as
observed by the device, including DEVICE_NEEDS_RESET.)
Not only that, without READ_STATUS, we can't do device reset which
is a prerequisite for a proper cleanup, as required by the spec.
You certainly remember, the driver has may not assume the reset
was performed (and thus virtqueues are not live) until it reads
back status 0. But without READ_STATUS virtio_ccw_get_status() will
keep returning the status the driver last set via
virtio_ccw_set_status(). And CCW_CMD_VDEV_RESET is of course
revision 1 material. This looks ugly!
quoted
everything works as intended; a device rejecting the
command should also be handled gracefully. For correctness, we
should really limit the command to revision 2 or higher, though.
We also negotiated the revision to at most 1, as we never bumped
the maximum revision; let's do that now.
Fixes: 7d3ce5ab9430 ("virtio/s390: support READ_STATUS command for virtio-ccw")
Signed-off-by: Cornelia Huck <cohuck@redhat.com>
---
QEMU does not fence off READ_STATUS for revisions < 2, which is probably
why we never noticed this. I'm not aware of other hypervisors that do
fence it off, nor any that cannot deal properly with an unknown command.
Not sure whether this is stable worthy?
Maybe it is, given the MUST reject clause in the standard?
Yes, IMHO this must be backported. A device that ain't violating the
spec would currently reject READ_STATUS. Which would break RESET_VDEV
like I described above.
Can we change that MUST to should? There are now good reasons for not
doing like the spec says in case of READ_STATUS.
I don't think our faking of the status read (i.e. returning the old one)
is contributing to spec compliance. Especially not if the inability to
READ is not transient.
Also return old_status; would tell the story better, but on the
other hand, that would be an unrelated cosmetic change. Maybe
a separate patch?
Reviewed-by: Halil Pasic <pasic@linux.ibm.com>
Regards,
Halil
On Mon, 15 Feb 2021 19:51:44 +0100
Halil Pasic [off-list ref] wrote:
On Mon, 15 Feb 2021 12:47:02 +0100
Cornelia Huck [off-list ref] wrote:
quoted
On Fri, 12 Feb 2021 18:04:11 +0100
Cornelia Huck [off-list ref] wrote:
quoted
CCW_CMD_READ_STATUS was introduced with revision 2 of virtio-ccw,
and drivers should only rely on it being implemented when they
negotiated at least that revision with the device.
However, virtio_ccw_get_status() issued READ_STATUS for any
device operating at least at revision 1. If the device accepts
READ_STATUS regardless of the negotiated revision (which it is
free to do),
So, looking at the standard again, the device is actually required to
reject the READ_STATUS if only rev 1 had been negotiated... regardless
of that, I don't think we should change QEMU's behaviour, as it would
affect existing guests (they would lose access to the status bits as
observed by the device, including DEVICE_NEEDS_RESET.)
Not only that, without READ_STATUS, we can't do device reset which
is a prerequisite for a proper cleanup, as required by the spec.
You certainly remember, the driver has may not assume the reset
was performed (and thus virtqueues are not live) until it reads
back status 0. But without READ_STATUS virtio_ccw_get_status() will
keep returning the status the driver last set via
virtio_ccw_set_status(). And CCW_CMD_VDEV_RESET is of course
revision 1 material. This looks ugly!
Yes, that problem kind of cascades down.
quoted
quoted
everything works as intended; a device rejecting the
command should also be handled gracefully. For correctness, we
should really limit the command to revision 2 or higher, though.
We also negotiated the revision to at most 1, as we never bumped
the maximum revision; let's do that now.
Fixes: 7d3ce5ab9430 ("virtio/s390: support READ_STATUS command for virtio-ccw")
Signed-off-by: Cornelia Huck <cohuck@redhat.com>
---
QEMU does not fence off READ_STATUS for revisions < 2, which is probably
why we never noticed this. I'm not aware of other hypervisors that do
fence it off, nor any that cannot deal properly with an unknown command.
Not sure whether this is stable worthy?
Maybe it is, given the MUST reject clause in the standard?
Yes, IMHO this must be backported. A device that ain't violating the
spec would currently reject READ_STATUS. Which would break RESET_VDEV
like I described above.
Can we change that MUST to should? There are now good reasons for not
doing like the spec says in case of READ_STATUS.
Yes. I'm not so sure forcing the device to reject the command was such
a good idea anyway, and relaxing the requirement keeps existing
implementations in compliance.
I've opened https://github.com/oasis-tcs/virtio-spec/issues/96 and will
send a patch for the spec later.
I don't think our faking of the status read (i.e. returning the old one)
is contributing to spec compliance. Especially not if the inability to
READ is not transient.
Also return old_status; would tell the story better, but on the
other hand, that would be an unrelated cosmetic change. Maybe
a separate patch?
We would also need to actively check for success or failure of the
channel program in that case.
I'm currently looking at the virtio-ccw code anyway, so I can put that
on my list as well.
Reviewed-by: Halil Pasic <pasic@linux.ibm.com>
Thanks!
I'll do a v2 with a tweaked commit message and cc:stable.