[libvirt PATCH v2 0/1] Ignore EPERM on attempts to clear a VF VLAN ID

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



SmartNIC DPUs may not expose some privileged eswitch operations
to the hypervisor hosts. For example, this happens with Bluefield
devices running in the ECPF (default) mode [1] for security reasons. While
VF MAC address programming is possible via an RTM_SETLINK operation,
trying to set a VLAN ID in the same operation will fail with EPERM.

In the kernel a relevant call chain may look like

do_setlink -> do_setvfinfo -> dev->netdev_ops->set_vf_vlan

which calls a driver-specific function like [2] eventually.

The equivalent ip link commands below provide an illustration:

1. This works:

sudo ip link set enp130s0f0 vf 2 mac de:ad:be:ef:ca:fe

2. Setting (or clearing) a VLAN fails with EPERM:

sudo ip link set enp130s0f0 vf 2 vlan 0
RTNETLINK answers: Operation not permitted

3. This is what Libvirt attempts to do today (when trying to clear a
   VF VLAN at the same time as programming a VF MAC).

sudo ip link set enp130s0f0 vf 2 vlan 0 mac de:ad:be:ef:ca:fe
RTNETLINK answers: Operation not permitted

If setting an explicit VLAN ID results in an EPERM, clearing a VLAN
(setting a VLAN ID to 0) can be handled gracefully by ignoring the
EPERM error with the rationale being that if we cannot set this state
in the first place, we cannot clear it either.

Thus, virNetDevSetVfConfig is split into two distinct functions. If
clearing a VLAN ID fails with EPERM, the error is simply ignored.

Both new functions rely virNetDevSendVfSetLinkRequest that implements
common functionality related to formatting a request, sending it and
handling error conditions and returns 0 or an error since in both cases
the payload is either NLMSG_DONE (no error) or NLMSG_ERROR where an
error message is needed by the caller to handle known cases
appropriately. This function allows the conditional code to be unit tested.

An alternative to this could be providing a higher level control plane
mechanism that would provide metadata about a device being remotely
managed in which case Libvirt would avoid trying to set or clear a
VLAN ID. This would be more complicated since other software (like Nova
in the OpenStack case) would have to annotate every guest device with an
attribute indicating whether a device is remotely managed or not based
on operator provided configuration so that Libvirt can act on this and
avoid VLAN programming.

https://gitlab.com/dmitriis/libvirt/-/pipelines/406528913

[1] https://docs.mellanox.com/display/BlueFieldSWv35111601/Modes+of+Operation#ModesofOperation-SmartNICmode
[2] https://github.com/torvalds/linux/blob/v5.15/drivers/net/ethernet/mellanox/mlx5/core/esw/legacy.c#L427-L434

Dmitrii Shcherbakov (1):
  Ignore EPERM on attempts to clear VF VLAN ID

 src/libvirt_private.syms |   7 ++
 src/util/virnetdev.c     | 199 ++++++++++++++++++++++-------------
 src/util/virnetdevpriv.h |  44 ++++++++
 tests/virnetdevtest.c    | 222 ++++++++++++++++++++++++++++++++++++++-
 4 files changed, 396 insertions(+), 76 deletions(-)
 create mode 100644 src/util/virnetdevpriv.h

-- 
2.32.0





[Index of Archives]     [Virt Tools]     [Libvirt Users]     [Lib OS Info]     [Fedora Users]     [Fedora Desktop]     [Fedora SELinux]     [Big List of Linux Books]     [Yosemite News]     [KDE Users]     [Fedora Tools]

  Powered by Linux