From: Si-Wei Liu <hidden> Date: 2021-02-06 13:18:26
When feature VIRTIO_NET_F_MTU is negotiated on mlx5_vdpa,
22 extra bytes worth of MTU length is shown in guest.
This is because the mlx5_query_port_max_mtu API returns
the "hardware" MTU value, which does not just contain the
Ethernet payload, but includes extra lengths starting
from the Ethernet header up to the FCS altogether.
Fix the MTU so packets won't get dropped silently.
Signed-off-by: Si-Wei Liu <redacted>
---
drivers/vdpa/mlx5/core/mlx5_vdpa.h | 4 ++++
drivers/vdpa/mlx5/net/mlx5_vnet.c | 15 ++++++++++++++-
2 files changed, 18 insertions(+), 1 deletion(-)
From: Si-Wei Liu <hidden> Date: 2021-02-06 13:18:26
While virtq is stopped, get_vq_state() is supposed to
be called to get sync'ed with the latest internal
avail_index from device. The saved avail_index is used
to restate the virtq once device is started. Commit
b35ccebe3ef7 introduced the clear_virtqueues() routine
to reset the saved avail_index, however, the index
gets cleared a bit earlier before get_vq_state() tries
to read it. This would cause consistency problems when
virtq is restarted, e.g. through a series of link down
and link up events. We could defer the clearing of
avail_index to until the device is to be started,
i.e. until VIRTIO_CONFIG_S_DRIVER_OK is set again in
set_status().
Fixes: b35ccebe3ef7 ("vdpa/mlx5: Restore the hardware used index after change map")
Signed-off-by: Si-Wei Liu <redacted>
---
drivers/vdpa/mlx5/net/mlx5_vnet.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
From: Si-Wei Liu <hidden> Date: 2021-02-06 13:18:26
The mlx_features denotes the capability for which
set of virtio features is supported by device. In
principle, this field needs not be cleared during
virtio device reset, as this capability is static
and does not change across reset.
In fact, the current code may have the assumption
that mlx_features can be reloaded from firmware
via the .get_features ops after device is reset
(via the .set_status ops), which is unfortunately
not true. The userspace VMM might save a copy
of backend capable features and won't call into
kernel again to get it on reset. This causes all
virtio features getting disabled on newly created
virtqs after device reset, while guest would hold
mismatched view of available features. For e.g.,
the guest may still assume tx checksum offload
is available after reset and feature negotiation,
causing frames with bogus (incomplete) checksum
transmitted on the wire.
Signed-off-by: Si-Wei Liu <redacted>
---
drivers/vdpa/mlx5/net/mlx5_vnet.c | 1 -
1 file changed, 1 deletion(-)
From: Jason Wang <hidden> Date: 2021-02-08 04:39:28
On 2021/2/6 下午8:29, Si-Wei Liu wrote:
The mlx_features denotes the capability for which
set of virtio features is supported by device. In
principle, this field needs not be cleared during
virtio device reset, as this capability is static
and does not change across reset.
In fact, the current code may have the assumption
that mlx_features can be reloaded from firmware
via the .get_features ops after device is reset
(via the .set_status ops), which is unfortunately
not true. The userspace VMM might save a copy
of backend capable features and won't call into
kernel again to get it on reset.
This is not the behavior of Qemu but it's valid.
This causes all
virtio features getting disabled on newly created
virtqs after device reset, while guest would hold
mismatched view of available features. For e.g.,
the guest may still assume tx checksum offload
is available after reset and feature negotiation,
causing frames with bogus (incomplete) checksum
transmitted on the wire.
Signed-off-by: Si-Wei Liu <redacted>
From: Jason Wang <hidden> Date: 2021-02-08 04:40:18
On 2021/2/6 下午8:29, Si-Wei Liu wrote:
While virtq is stopped, get_vq_state() is supposed to
be called to get sync'ed with the latest internal
avail_index from device. The saved avail_index is used
to restate the virtq once device is started. Commit
b35ccebe3ef7 introduced the clear_virtqueues() routine
to reset the saved avail_index, however, the index
gets cleared a bit earlier before get_vq_state() tries
to read it. This would cause consistency problems when
virtq is restarted, e.g. through a series of link down
and link up events. We could defer the clearing of
avail_index to until the device is to be started,
i.e. until VIRTIO_CONFIG_S_DRIVER_OK is set again in
set_status().
Fixes: b35ccebe3ef7 ("vdpa/mlx5: Restore the hardware used index after change map")
Signed-off-by: Si-Wei Liu <redacted>
From: Jason Wang <hidden> Date: 2021-02-08 04:41:01
On 2021/2/6 下午8:29, Si-Wei Liu wrote:
When feature VIRTIO_NET_F_MTU is negotiated on mlx5_vdpa,
22 extra bytes worth of MTU length is shown in guest.
This is because the mlx5_query_port_max_mtu API returns
the "hardware" MTU value, which does not just contain the
Ethernet payload, but includes extra lengths starting
from the Ethernet header up to the FCS altogether.
Fix the MTU so packets won't get dropped silently.
Signed-off-by: Si-Wei Liu <redacted>
From: Eli Cohen <hidden> Date: 2021-02-08 05:36:08
On Sat, Feb 06, 2021 at 04:29:23AM -0800, Si-Wei Liu wrote:
quoted hunk
The mlx_features denotes the capability for which
set of virtio features is supported by device. In
principle, this field needs not be cleared during
virtio device reset, as this capability is static
and does not change across reset.
In fact, the current code may have the assumption
that mlx_features can be reloaded from firmware
via the .get_features ops after device is reset
(via the .set_status ops), which is unfortunately
not true. The userspace VMM might save a copy
of backend capable features and won't call into
kernel again to get it on reset. This causes all
virtio features getting disabled on newly created
virtqs after device reset, while guest would hold
mismatched view of available features. For e.g.,
the guest may still assume tx checksum offload
is available after reset and feature negotiation,
causing frames with bogus (incomplete) checksum
transmitted on the wire.
Signed-off-by: Si-Wei Liu <redacted>
---
drivers/vdpa/mlx5/net/mlx5_vnet.c | 1 -
1 file changed, 1 deletion(-)
Since we assume that device capabilities don't change, I think I would
get the features through a call done in mlx5v_probe after the netdev
object is created and change mlx5_vdpa_get_features() to just return
ndev->mvdev.mlx_features.
Did you actually see this issue in action? If you did, can you share
with us how you trigerred this?
From: Eli Cohen <hidden> Date: 2021-02-08 05:37:18
On Sat, Feb 06, 2021 at 04:29:22AM -0800, Si-Wei Liu wrote:
When feature VIRTIO_NET_F_MTU is negotiated on mlx5_vdpa,
22 extra bytes worth of MTU length is shown in guest.
This is because the mlx5_query_port_max_mtu API returns
the "hardware" MTU value, which does not just contain the
Ethernet payload, but includes extra lengths starting
from the Ethernet header up to the FCS altogether.
Fix the MTU so packets won't get dropped silently.
Signed-off-by: Si-Wei Liu <redacted>
From: Eli Cohen <hidden> Date: 2021-02-08 05:49:32
On Sat, Feb 06, 2021 at 04:29:24AM -0800, Si-Wei Liu wrote:
While virtq is stopped, get_vq_state() is supposed to
be called to get sync'ed with the latest internal
avail_index from device. The saved avail_index is used
to restate the virtq once device is started. Commit
b35ccebe3ef7 introduced the clear_virtqueues() routine
to reset the saved avail_index, however, the index
gets cleared a bit earlier before get_vq_state() tries
to read it. This would cause consistency problems when
virtq is restarted, e.g. through a series of link down
and link up events. We could defer the clearing of
avail_index to until the device is to be started,
i.e. until VIRTIO_CONFIG_S_DRIVER_OK is set again in
set_status().
Not sure I understand the scenario. You are talking about reset of the
device followed by up/down events on the interface. How can you trigger
this?
quoted hunk
Fixes: b35ccebe3ef7 ("vdpa/mlx5: Restore the hardware used index after change map")
Signed-off-by: Si-Wei Liu <redacted>
---
drivers/vdpa/mlx5/net/mlx5_vnet.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
From: Si-Wei Liu <hidden> Date: 2021-02-09 01:22:54
On 2/7/2021 9:35 PM, Eli Cohen wrote:
On Sat, Feb 06, 2021 at 04:29:23AM -0800, Si-Wei Liu wrote:
quoted
The mlx_features denotes the capability for which
set of virtio features is supported by device. In
principle, this field needs not be cleared during
virtio device reset, as this capability is static
and does not change across reset.
In fact, the current code may have the assumption
that mlx_features can be reloaded from firmware
via the .get_features ops after device is reset
(via the .set_status ops), which is unfortunately
not true. The userspace VMM might save a copy
of backend capable features and won't call into
kernel again to get it on reset. This causes all
virtio features getting disabled on newly created
virtqs after device reset, while guest would hold
mismatched view of available features. For e.g.,
the guest may still assume tx checksum offload
is available after reset and feature negotiation,
causing frames with bogus (incomplete) checksum
transmitted on the wire.
Signed-off-by: Si-Wei Liu <redacted>
---
drivers/vdpa/mlx5/net/mlx5_vnet.c | 1 -
1 file changed, 1 deletion(-)
Since we assume that device capabilities don't change, I think I would
get the features through a call done in mlx5v_probe after the netdev
object is created and change mlx5_vdpa_get_features() to just return
ndev->mvdev.mlx_features.
Yep, it makes sense. Will post a revised patch. If vdpa tool allows
reconfiguration post probing, the code has to be reconciled then.
Did you actually see this issue in action? If you did, can you share
with us how you trigerred this?
Issue is indeed seen in action. The mismatched tx-checksum offload as
described in the commit message was one of such examples. You would need
a guest reboot though (triggering device reset via the .set_status ops
and zero'ed mlx_features was loaded to deduce new actual_features for
creating the h/w virtq object) for repro.
-Siwei
From: Si-Wei Liu <hidden> Date: 2021-02-09 01:42:57
On 2/7/2021 9:48 PM, Eli Cohen wrote:
On Sat, Feb 06, 2021 at 04:29:24AM -0800, Si-Wei Liu wrote:
quoted
While virtq is stopped, get_vq_state() is supposed to
be called to get sync'ed with the latest internal
avail_index from device. The saved avail_index is used
to restate the virtq once device is started. Commit
b35ccebe3ef7 introduced the clear_virtqueues() routine
to reset the saved avail_index, however, the index
gets cleared a bit earlier before get_vq_state() tries
to read it. This would cause consistency problems when
virtq is restarted, e.g. through a series of link down
and link up events. We could defer the clearing of
avail_index to until the device is to be started,
i.e. until VIRTIO_CONFIG_S_DRIVER_OK is set again in
set_status().
Not sure I understand the scenario. You are talking about reset of the
device followed by up/down events on the interface. How can you trigger
this?
Currently it's not possible to trigger link up/down events with upstream
QEMU due lack of config/control interrupt implementation. And live
migration could be another scenario that cannot be satisfied because of
inconsistent queue state. They share the same root of cause as captured
here.
-Siwei
quoted
Fixes: b35ccebe3ef7 ("vdpa/mlx5: Restore the hardware used index after change map")
Signed-off-by: Si-Wei Liu <redacted>
---
drivers/vdpa/mlx5/net/mlx5_vnet.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
From: Jason Wang <hidden> Date: 2021-02-09 03:49:11
On 2021/2/6 下午8:29, Si-Wei Liu wrote:
quoted hunk
While virtq is stopped, get_vq_state() is supposed to
be called to get sync'ed with the latest internal
avail_index from device. The saved avail_index is used
to restate the virtq once device is started. Commit
b35ccebe3ef7 introduced the clear_virtqueues() routine
to reset the saved avail_index, however, the index
gets cleared a bit earlier before get_vq_state() tries
to read it. This would cause consistency problems when
virtq is restarted, e.g. through a series of link down
and link up events. We could defer the clearing of
avail_index to until the device is to be started,
i.e. until VIRTIO_CONFIG_S_DRIVER_OK is set again in
set_status().
Fixes: b35ccebe3ef7 ("vdpa/mlx5: Restore the hardware used index after change map")
Signed-off-by: Si-Wei Liu <redacted>
---
drivers/vdpa/mlx5/net/mlx5_vnet.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
Rethink about this. As mentioned in another thread, this in fact breaks
set_vq_state(). (See vhost_virtqueue_start() ->
vhost_vdpa_set_vring_base() in qemu codes).
The issue is that the avail idx is forgot, we need keep it.
Thanks
err = setup_driver(ndev);
if (err) {
mlx5_vdpa_warn(mvdev, "failed to setup driver\n");
From: Si-Wei Liu <hidden> Date: 2021-02-10 01:59:14
On 2/8/2021 7:37 PM, Jason Wang wrote:
On 2021/2/6 下午8:29, Si-Wei Liu wrote:
quoted
While virtq is stopped, get_vq_state() is supposed to
be called to get sync'ed with the latest internal
avail_index from device. The saved avail_index is used
to restate the virtq once device is started. Commit
b35ccebe3ef7 introduced the clear_virtqueues() routine
to reset the saved avail_index, however, the index
gets cleared a bit earlier before get_vq_state() tries
to read it. This would cause consistency problems when
virtq is restarted, e.g. through a series of link down
and link up events. We could defer the clearing of
avail_index to until the device is to be started,
i.e. until VIRTIO_CONFIG_S_DRIVER_OK is set again in
set_status().
Fixes: b35ccebe3ef7 ("vdpa/mlx5: Restore the hardware used index
after change map")
Signed-off-by: Si-Wei Liu <redacted>
---
drivers/vdpa/mlx5/net/mlx5_vnet.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/drivers/vdpa/mlx5/net/mlx5_vnet.c
b/drivers/vdpa/mlx5/net/mlx5_vnet.c
index aa6f8cd..444ab58 100644
vdpa_device *vdev, u8 status)
if ((status ^ ndev->mvdev.status) & VIRTIO_CONFIG_S_DRIVER_OK) {
if (status & VIRTIO_CONFIG_S_DRIVER_OK) {
+ clear_virtqueues(ndev);
Rethink about this. As mentioned in another thread, this in fact
breaks set_vq_state(). (See vhost_virtqueue_start() ->
vhost_vdpa_set_vring_base() in qemu codes).
I assume that the clearing for vhost-vdpa would be done via (qemu code),
vhost_dev_start()->vhost_vdpa_dev_start()->vhost_vdpa_call(status |
VIRTIO_CONFIG_S_DRIVER_OK)
which is _after_ vhost_virtqueue_start() gets called to restore the
avail_idx to h/w in vhost_dev_start(). What am I missing here?
-Siwei
The issue is that the avail idx is forgot, we need keep it.
Thanks
quoted
err = setup_driver(ndev);
if (err) {
mlx5_vdpa_warn(mvdev, "failed to setup driver\n");
From: Jason Wang <hidden> Date: 2021-02-10 04:02:31
On 2021/2/10 上午8:26, Si-Wei Liu wrote:
On 2/8/2021 7:37 PM, Jason Wang wrote:
quoted
On 2021/2/6 下午8:29, Si-Wei Liu wrote:
quoted
While virtq is stopped, get_vq_state() is supposed to
be called to get sync'ed with the latest internal
avail_index from device. The saved avail_index is used
to restate the virtq once device is started. Commit
b35ccebe3ef7 introduced the clear_virtqueues() routine
to reset the saved avail_index, however, the index
gets cleared a bit earlier before get_vq_state() tries
to read it. This would cause consistency problems when
virtq is restarted, e.g. through a series of link down
and link up events. We could defer the clearing of
avail_index to until the device is to be started,
i.e. until VIRTIO_CONFIG_S_DRIVER_OK is set again in
set_status().
Fixes: b35ccebe3ef7 ("vdpa/mlx5: Restore the hardware used index
after change map")
Signed-off-by: Si-Wei Liu <redacted>
---
drivers/vdpa/mlx5/net/mlx5_vnet.c | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/drivers/vdpa/mlx5/net/mlx5_vnet.c
b/drivers/vdpa/mlx5/net/mlx5_vnet.c
index aa6f8cd..444ab58 100644
vdpa_device *vdev, u8 status)
if ((status ^ ndev->mvdev.status) &
VIRTIO_CONFIG_S_DRIVER_OK) {
if (status & VIRTIO_CONFIG_S_DRIVER_OK) {
+ clear_virtqueues(ndev);
Rethink about this. As mentioned in another thread, this in fact
breaks set_vq_state(). (See vhost_virtqueue_start() ->
vhost_vdpa_set_vring_base() in qemu codes).
I assume that the clearing for vhost-vdpa would be done via (qemu code),
vhost_dev_start()->vhost_vdpa_dev_start()->vhost_vdpa_call(status |
VIRTIO_CONFIG_S_DRIVER_OK)
which is _after_ vhost_virtqueue_start() gets called to restore the
avail_idx to h/w in vhost_dev_start(). What am I missing here?
-Siwei
I think not. I thought clear_virtqueues() will clear hardware index but
looks not. (I guess we need a better name other than clear_virtqueues(),
e.g from the name it looks like the it will clear the hardware states)
Thanks
quoted
The issue is that the avail idx is forgot, we need keep it.
Thanks
quoted
err = setup_driver(ndev);
if (err) {
mlx5_vdpa_warn(mvdev, "failed to setup driver\n");
From: "Michael S. Tsirkin" <mst@redhat.com> Date: 2021-02-10 12:33:04
On Mon, Feb 08, 2021 at 05:20:11PM -0800, Si-Wei Liu wrote:
On 2/7/2021 9:35 PM, Eli Cohen wrote:
quoted
On Sat, Feb 06, 2021 at 04:29:23AM -0800, Si-Wei Liu wrote:
quoted
The mlx_features denotes the capability for which
set of virtio features is supported by device. In
principle, this field needs not be cleared during
virtio device reset, as this capability is static
and does not change across reset.
In fact, the current code may have the assumption
that mlx_features can be reloaded from firmware
via the .get_features ops after device is reset
(via the .set_status ops), which is unfortunately
not true. The userspace VMM might save a copy
of backend capable features and won't call into
kernel again to get it on reset. This causes all
virtio features getting disabled on newly created
virtqs after device reset, while guest would hold
mismatched view of available features. For e.g.,
the guest may still assume tx checksum offload
is available after reset and feature negotiation,
causing frames with bogus (incomplete) checksum
transmitted on the wire.
Signed-off-by: Si-Wei Liu <redacted>
---
drivers/vdpa/mlx5/net/mlx5_vnet.c | 1 -
1 file changed, 1 deletion(-)
Since we assume that device capabilities don't change, I think I would
get the features through a call done in mlx5v_probe after the netdev
object is created and change mlx5_vdpa_get_features() to just return
ndev->mvdev.mlx_features.
Yep, it makes sense. Will post a revised patch.
So I'm waiting for v2 of this patchset. Please make sure to post a cover letter
with an overall description.
If vdpa tool allows
reconfiguration post probing, the code has to be reconciled then.
quoted
Did you actually see this issue in action? If you did, can you share
with us how you trigerred this?
Issue is indeed seen in action. The mismatched tx-checksum offload as
described in the commit message was one of such examples. You would need a
guest reboot though (triggering device reset via the .set_status ops and
zero'ed mlx_features was loaded to deduce new actual_features for creating
the h/w virtq object) for repro.
-Siwei