Add a test that checks that the created filters do actually trigger on
matching traffic.

Exercising all the rules would be a very lengthy process. Instead, take a
log2 subset of rules. The logic behind picking log2 rules is that then
every bit of the instantiated item's number is exercised. This should catch
issues whether they happen at the high end, low end, or somewhere in
between.
Signed-off-by: Petr Machata <petrm@nvidia.com>
Reviewed-by: Amit Cohen <amcohen@nvidia.com>
Signed-off-by: Ido Schimmel <idosch@nvidia.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

The scale tests are verifying behavior of mlxsw when number of instances of
some resource reaches the ASIC capacity. The number of instances is
referred to as "target" number.

No scale tests so far needed to know this target number to clean up. E.g.
the tc_flower simply removes the clsact qdisc that all the tested filters
are hooked onto, and that takes care of collecting all the filters.

However, for the RIF counter test, which is being added in a future patch,
VLAN netdevices are created. These are created as part of the test, but of
course the cleanup needs to undo them again. For that it needs to know how
many there were. To support this usage, pass the target number to the
cleanup callback.
Signed-off-by: Petr Machata <petrm@nvidia.com>
Reviewed-by: Amit Cohen <amcohen@nvidia.com>
Signed-off-by: Ido Schimmel <idosch@nvidia.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

The scale tests are currently testing two things: that some number of
instances of a given resource can actually be created; and that when an
attempt is made to create more than the supported amount, the failures are
noted and handled gracefully.

Sometimes the scale test depends on more than one resource. In particular,
a following patch will add a RIF counter scale test, which depends on the
number of RIF counters that can be bound, and also on the number of RIFs
that can be created.

When the test is limited by the auxiliary resource and not by the primary
one, there's no point trying to run the overflow test, because it would be
testing exhaustion of the wrong resource.

To support this use case, when the $test_get_target yields 0, skip the test
instead.
Signed-off-by: Petr Machata <petrm@nvidia.com>
Reviewed-by: Amit Cohen <amcohen@nvidia.com>
Signed-off-by: Ido Schimmel <idosch@nvidia.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

The scale tests are currently testing two things: that some number of
instances of a given resource can actually be created; and that when an
attempt is made to create more than the supported amount, the failures are
noted and handled gracefully.

However the ability to allocate the resource does not mean that the
resource actually works when passing traffic. For that, make it possible
for a given scale to also test traffic.

Traffic test is only run on the positive leg of the scale test (no point
trying to pass traffic when the expected outcome is that the resource will
not be allocated). Traffic tests are opt-in, if a given test does not
expose it, it is not run.

To this end, delay the test cleanup until after the traffic test is run.
Signed-off-by: Petr Machata <petrm@nvidia.com>
Reviewed-by: Amit Cohen <amcohen@nvidia.com>
Signed-off-by: Ido Schimmel <idosch@nvidia.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

The scale of each resource is tested in the following manner:

1. The scale target is queried.
2. The test setup is prepared.
3. The test is invoked.

In some cases, the occupancy of a resource changes as part of the second
step, requiring the test to return a scale target that takes this change
into account.

Make this more robust by re-querying the scale target after the second
step.

Another possible solution is to swap the first and second steps, but
when a test needs to be skipped (i.e., scale target is zero), the setup
would have been in vain.
Signed-off-by: Ido Schimmel <idosch@nvidia.com>
Reviewed-by: Petr Machata <petrm@nvidia.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Using mlxsw driver, the configurations are offloaded just in case that
there is a physical port which is enslaved to the virtual device
(e.g., to a bridge). In 'mirror_gre_bridge_1q_lag' test, the bridge gets an
address and route before there are ports in the bridge. It means that these
configurations are not offloaded.

Till now the test passes with mlxsw driver even that the RIF of the
bridge is not in the hardware, because the ARP packets are trapped in
layer 2 and also mirrored, so there is no real need of the RIF in hardware.
The previous patch changed the traps 'ARP_REQUEST' and 'ARP_RESPONSE' to
be done at layer 3 instead of layer 2. With this change the ARP packets are
not trapped during the test, as the RIF is not in the hardware because of
the order of configurations.

Reorder the configurations to make them to be offloaded, then the test will
pass with the change of the traps.
Signed-off-by: Amit Cohen <amcohen@nvidia.com>
Reviewed-by: Petr Machata <petrm@nvidia.com>
Signed-off-by: Ido Schimmel <idosch@nvidia.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

The Spectrum ASIC has a limit on how many L3 devices (called RIFs) can be
created. The limit depends on the ASIC and FW revision, and mlxsw reads it
from the FW. In order to communicate both the number of RIFs that there can
be, and how many are taken now (i.e. occupancy), introduce a corresponding
devlink resource.
Signed-off-by: Petr Machata <petrm@nvidia.com>
Reviewed-by: Amit Cohen <amcohen@nvidia.com>
Signed-off-by: Ido Schimmel <idosch@nvidia.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

In order to expose number of RIFs as a resource, it is going to be handy
to have the number of currently-allocated RIFs as a single number.
Introduce such.
Signed-off-by: Petr Machata <petrm@nvidia.com>
Reviewed-by: Amit Cohen <amcohen@nvidia.com>
Signed-off-by: Ido Schimmel <idosch@nvidia.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Currently, the traps 'ARP_REQUEST' and 'ARP_RESPONSE' occur at layer 2.
To allow the packets to be flooded, they are configured with the action
'MIRROR_TO_CPU' which means that the CPU receives a replica of the packet.

Today, Spectrum ASICs also support trapping ARP packets at layer 3. This
behavior is better, then the packets can just be trapped and there is no
need to mirror them. An additional motivation is that using the traps at
layer 2, the ARP packets are dropped in the router as they do not have an
IP header, then they are counted as error packets, which might confuse
users.

Add the relevant traps for layer 3 and use them instead of the existing
traps. There is no visible change to user space.
Signed-off-by: Amit Cohen <amcohen@nvidia.com>
Reviewed-by: Petr Machata <petrm@nvidia.com>
Signed-off-by: Ido Schimmel <idosch@nvidia.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Eric Dumazet says:

====================
tcp: final (?) round of mem pressure fixes

While working on prior patch series (e10b02ee "Merge branch
'net-reduce-tcp_memory_allocated-inflation'"), I found that we
could still have frozen TCP flows under memory pressure.

I thought we had solved this in 2015, but the fix was not complete.

v2: deal with zerocopy tx paths.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>

Blamed commit only dealt with applications issuing small writes.

Issue here is that we allow to force memory schedule for the sk_buff
allocation, but we have no guarantee that sendmsg() is able to
copy some payload in it.

In this patch, I make sure the socket can use up to tcp_wmem[0] bytes.

For example, if we consider tcp_wmem[0] = 4096 (default on x86),
and initial skb->truesize being 1280, tcp_sendmsg() is able to
copy up to 2816 bytes under memory pressure.

Before this patch a sendmsg() sending more than 2816 bytes
would either block forever (if persistent memory pressure),
or return -EAGAIN.

For bigger MTU networks, it is advised to increase tcp_wmem[0]
to avoid sending too small packets.

v2: deal with zero copy paths.

Fixes: 8e4d980a ("tcp: fix behavior for epoll edge trigger")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Wei Wang <weiwan@google.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Blamed commit only dealt with applications issuing small writes.

Issue here is that we allow to force memory schedule for the sk_buff
allocation, but we have no guarantee that sendmsg() is able to
copy some payload in it.

In this patch, I make sure the socket can use up to tcp_wmem[0] bytes.

For example, if we consider tcp_wmem[0] = 4096 (default on x86),
and initial skb->truesize being 1280, tcp_sendmsg() is able to
copy up to 2816 bytes under memory pressure.

Before this patch a sendmsg() sending more than 2816 bytes
would either block forever (if persistent memory pressure),
or return -EAGAIN.

For bigger MTU networks, it is advised to increase tcp_wmem[0]
to avoid sending too small packets.

v2: deal with zero copy paths.

Fixes: 8e4d980a ("tcp: fix behavior for epoll edge trigger")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Wei Wang <weiwan@google.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

sk_forced_mem_schedule() has a bug similar to ones fixed
in commit 7c80b038 ("net: fix sk_wmem_schedule() and
sk_rmem_schedule() errors")

While this bug has little chance to trigger in old kernels,
we need to fix it before the following patch.

Fixes: d83769a5 ("tcp: fix possible deadlock in tcp_send_fin()")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Shakeel Butt <shakeelb@google.com>
Reviewed-by: Wei Wang <weiwan@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Raju Lakkaraju says:

====================
net: lan743x: PCI11010 / PCI11414 devices Enhancements

This patch series continues with the addition of supported features
for the Ethernet function of the PCI11010 / PCI11414 devices to
the LAN743x driver.
====================

Link: https://lore.kernel.org/r/20220616041226.26996-1-Raju.Lakkaraju@microchip.comSigned-off-by: Jakub Kicinski <kuba@kernel.org>

Add support to Master-Slave configuration and state
Signed-off-by: Raju Lakkaraju <Raju.Lakkaraju@microchip.com>
Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

Add SGMII access read and write functions
Add support to SGMII 1G and 2.5G for PCI11010/PCI11414 chips
Signed-off-by: Raju Lakkaraju <Raju.Lakkaraju@microchip.com>
Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

Add support to Magic Packet Detection with Secure-ON for PCI11010/PCI11414 chips
Signed-off-by: Raju Lakkaraju <Raju.Lakkaraju@microchip.com>
Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

Add support to LAN743x common register dump
Signed-off-by: Raju Lakkaraju <Raju.Lakkaraju@microchip.com>
Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

Alvin Šipraga says:

====================
net: dsa: realtek: rtl8365mb: improve handling of PHY modes

This series introduces some minor cleanup of the driver and improves the
handling of PHY interface modes to break the assumption that CPU ports
are always over an external interface, and the assumption that user
ports are always using an internal PHY.
====================

Link: https://lore.kernel.org/r/20220615225116.432283-1-alvin@pqrs.dkSigned-off-by: Jakub Kicinski <kuba@kernel.org>

Realtek switches in the rtl8365mb family always have at least one port
with a so-called external interface, supporting PHY interface modes such
as RGMII or SGMII. The purpose of this patch is to improve the driver's
handling of these ports.

A new struct rtl8365mb_chip_info is introduced together with a static
array of such structs. An instance of this struct is added for each
supported switch, distinguished by its chip ID and version. Embedded in
each chip_info struct is an array of struct rtl8365mb_extint, describing
the external interfaces available. This is more specific than the old
rtl8365mb_extint_port_map, which was only valid for switches with up to
6 ports.

The struct rtl8365mb_extint also contains a bitmask of supported PHY
interface modes, which allows the driver to distinguish which ports
support RGMII. This corrects a previous mistake in the driver whereby it
was assumed that any port with an external interface supports RGMII.
This is not actually the case: for example, the RTL8367S has two
external interfaces, only the second of which supports RGMII. The first
supports only SGMII and HSGMII. This new design will make it easier to
add support for other interface modes.

Finally, rtl8365mb_phylink_get_caps() is fixed up to return supported
capabilities based on the external interface properties described above.
This addresses Vladimir's point in the linked thread that the
capabilities are not actually a function of the DSA port type: Although
most typical applications will treat the ports with internal PHY as user
ports, there is no actual hardware limitation preventing one from using
them as a CPU port. Equally, ports with external interface(s) may well
be treated as user ports, even though it is typical to use those ports
as CPU ports.

Link: https://lore.kernel.org/netdev/20220510192301.5djdt3ghoavxulhl@bang-olufsen.dk/Signed-off-by: Alvin Šipraga <alsi@bang-olufsen.dk>
Acked-by: Russell King (Oracle) <rmk+kernel@armlinux.org.uk>
Acked-by: Luiz Angelo Daros de Luca <luizluca@gmail.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

The variable is just assigned the value of a macro, so it can be
removed.
Signed-off-by: Alvin Šipraga <alsi@bang-olufsen.dk>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

The maximum number of ports is actually 11, according to two
observations:

1. The highest port ID used in the vendor driver is 10. Since port IDs
   are indexed from 0, and since DSA follows the same numbering system,
   this means up to 11 ports are to be presumed.

2. The registers with port mask fields always amount to a maximum port
   mask of 0x7FF, corresponding to a maximum 11 ports.

In view of this, I also deleted the comment.
Signed-off-by: Alvin Šipraga <alsi@bang-olufsen.dk>
Reviewed-by: Luiz Angelo Daros de Luca <luizluca@gmail.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

There is no real need for this variable: the line change interrupt mask
is sufficiently masked out when getting linkup_ind and linkdown_ind in
the interrupt handler.
Signed-off-by: Alvin Šipraga <alsi@bang-olufsen.dk>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

The official name of this switch is RTL8367RB-VB, not RTL8367RB. There
is also an RTL8367RB-VC which is rather different. Change the name of
the CHIP_ID/_VER macros for reasons of consistency.
Signed-off-by: Alvin Šipraga <alsi@bang-olufsen.dk>
Reviewed-by: Luiz Angelo Daros de Luca <luizluca@gmail.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

Alex Elder says:

====================
net: ipa: more multi-channel event ring work

This series makes a little more progress toward supporting multiple
channels with a single event ring.  The first removes the assumption
that consecutive events are associated with the same RX channel.

The second derives the channel associated with an event from the
event itself, and the next does a small cleanup enabled by that.

The fourth causes updates to occur for every event processed (rather
once).  And the final patch does a little more rework to make TX
completion have more in common with RX completion.
====================

Link: https://lore.kernel.org/r/20220615165929.5924-1-elder@linaro.orgSigned-off-by: Jakub Kicinski <kuba@kernel.org>

Move the processing done for TX channels in gsi_channel_update()
into gsi_evt_ring_rx_update().  The called function is called for
both RX and TX channels, so rename it to be gsi_evt_ring_update().
As a result, this code no longer assumes events in an event ring are
associated with just one channel.

Because all events in a ring are handled in that function, we can
move the call to gsi_trans_move_complete() there, and can ring the
event ring doorbell there as well after all new events in the ring
have been processed.
Signed-off-by: Alex Elder <elder@linaro.org>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

When an RX transaction completes, we update the trans->len field to
contain the actual number of bytes received.  This is done in a loop
in gsi_evt_ring_rx_update().

Change that function so it checks the data transfer direction
recorded in the transaction, and only updates trans->len for RX
transfers.

Then call it unconditionally.  This means events for TX endpoints
will run through the loop without otherwise doing anything, but
this will change shortly.
Signed-off-by: Alex Elder <elder@linaro.org>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

The only reason the event ring's channel pointer is needed in
gsi_evt_ring_rx_update() is so we can get at its GSI pointer.

We can pass the GSI pointer as an argument, along with the event
ring ID, and thereby avoid using the event ring channel pointer.
This is another step toward no longer assuming an event ring
services a single channel.
Signed-off-by: Alex Elder <elder@linaro.org>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

Change gsi_channel_trans_map() so it derives the channel used from
the transaction.  Pass the index of the *first* TRE used by the
transaction, and have the called function account for the fact that
the last one used is what's important.
Signed-off-by: Alex Elder <elder@linaro.org>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

In gsi_evt_ring_rx_update(), use gsi_event_trans() repeatedly
to find the transaction associated with an event, rather than
assuming consecutive events are associated with the same channel.
This removes the only caller of gsi_trans_pool_next(), so get rid
of it.
Signed-off-by: Alex Elder <elder@linaro.org>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

Rasmus Villemoes says:

====================
dt-bindings: dp83867: add binding for io_impedance_ctrl nvmem cell

We have a board where measurements indicate that the current three
options - leaving IO_IMPEDANCE_CTRL at the reset value (which is
factory calibrated to a value corresponding to approximately 50 ohms)
or using one of the two boolean properties to set it to the min/max
value - are too coarse.

This series adds a device tree binding for an nvmem cell which can be
populated during production with a suitable value calibrated for each
board, and corresponding support in the driver. The second patch adds
a trivial phy wrapper for dev_err_probe(), used in the third.
====================

Link: https://lore.kernel.org/r/20220614084612.325229-1-linux@rasmusvillemoes.dkSigned-off-by: Jakub Kicinski <kuba@kernel.org>

We have a board where measurements indicate that the current three
options - leaving IO_IMPEDANCE_CTRL at the (factory calibrated) reset
value or using one of the two boolean properties to set it to the
min/max value - are too coarse.

Implement support for the newly added binding allowing device tree to
specify an nvmem cell containing an appropriate value for this
specific board.
Signed-off-by: Rasmus Villemoes <linux@rasmusvillemoes.dk>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

The dev_err_probe() function is quite useful to avoid boilerplate
related to -EPROBE_DEFER handling. Add a phydev_err_probe() helper to
simplify making use of that from phy drivers which otherwise use the
phydev_* helpers.
Signed-off-by: Rasmus Villemoes <linux@rasmusvillemoes.dk>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

We have a board where measurements indicate that the current three
options - leaving IO_IMPEDANCE_CTRL at the reset value (which is
factory calibrated to a value corresponding to approximately 50 ohms)
or using one of the two boolean properties to set it to the min/max
value - are too coarse.

There is no fixed mapping from register values to values in the range
35-70 ohms; it varies from chip to chip, and even that target range is
approximate. So add a DT binding for an nvmem cell which can be
populated during production with a value suitable for each specific
board.
Reviewed-by: Rob Herring <robh@kernel.org>
Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Signed-off-by: Rasmus Villemoes <linux@rasmusvillemoes.dk>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

No conflicts.
Signed-off-by: Jakub Kicinski <kuba@kernel.org>

Pull networking fixes from Jakub Kicinski:
 "Mostly driver fixes.

  Current release - regressions:

   - Revert "net: Add a second bind table hashed by port and address",
     needs more work

   - amd-xgbe: use platform_irq_count(), static setup of IRQ resources
     had been removed from DT core

   - dts: at91: ksz9477_evb: add phy-mode to fix port/phy validation

  Current release - new code bugs:

   - hns3: modify the ring param print info

  Previous releases - always broken:

   - axienet: make the 64b addressable DMA depends on 64b architectures

   - iavf: fix issue with MAC address of VF shown as zero

   - ice: fix PTP TX timestamp offset calculation

   - usb: ax88179_178a needs FLAG_SEND_ZLP

  Misc:

   - document some net.sctp.* sysctls"

* tag 'net-5.19-rc3' of git://git.kernel.org/pub/scm/linux/kernel/git/netdev/net: (31 commits)
  net: axienet: add missing error return code in axienet_probe()
  Revert "net: Add a second bind table hashed by port and address"
  net: ax25: Fix deadlock caused by skb_recv_datagram in ax25_recvmsg
  net: usb: ax88179_178a needs FLAG_SEND_ZLP
  MAINTAINERS: add include/dt-bindings/net to NETWORKING DRIVERS
  ARM: dts: at91: ksz9477_evb: fix port/phy validation
  net: bgmac: Fix an erroneous kfree() in bgmac_remove()
  ice: Fix memory corruption in VF driver
  ice: Fix queue config fail handling
  ice: Sync VLAN filtering features for DVM
  ice: Fix PTP TX timestamp offset calculation
  mlxsw: spectrum_cnt: Reorder counter pools
  docs: networking: phy: Fix a typo
  amd-xgbe: Use platform_irq_count()
  octeontx2-vf: Add support for adaptive interrupt coalescing
  xilinx:  Fix build on x86.
  net: axienet: Use iowrite64 to write all 64b descriptor pointers
  net: axienet: make the 64b addresable DMA depends on 64b archectures
  net: hns3: fix tm port shapping of fibre port is incorrect after driver initialization
  net: hns3: fix PF rss size initialization bug
  ...

It should return error code in error path in axienet_probe().

Fixes: 00be43a7 ("net: axienet: make the 64b addresable DMA depends on 64b archectures")
Reported-by: Hulk Robot <hulkci@huawei.com>
Signed-off-by: Yang Yingliang <yangyingliang@huawei.com>
Link: https://lore.kernel.org/r/20220616062917.3601-1-yangyingliang@huawei.comSigned-off-by: Jakub Kicinski <kuba@kernel.org>

This reverts:

commit d5a42de8 ("net: Add a second bind table hashed by port and address")
commit 538aaf9b ("selftests: Add test for timing a bind request to a port with a populated bhash entry")
Link: https://lore.kernel.org/netdev/20220520001834.2247810-1-kuba@kernel.org/

There are a few things that need to be fixed here:
* Updating bhash2 in cases where the socket's rcv saddr changes
* Adding bhash2 hashbucket locks

Links to syzbot reports:
https://lore.kernel.org/netdev/00000000000022208805e0df247a@google.com/
https://lore.kernel.org/netdev/0000000000003f33bc05dfaf44fe@google.com/

Fixes: d5a42de8 ("net: Add a second bind table hashed by port and address")
Reported-by: syzbot+015d756bbd1f8b5c8f09@syzkaller.appspotmail.com
Reported-by: syzbot+98fd2d1422063b0f8c44@syzkaller.appspotmail.com
Reported-by: syzbot+0a847a982613c6438fba@syzkaller.appspotmail.com
Signed-off-by: Joanne Koong <joannelkoong@gmail.com>
Link: https://lore.kernel.org/r/20220615193213.2419568-1-joannelkoong@gmail.comSigned-off-by: Jakub Kicinski <kuba@kernel.org>

Haiyang Zhang says:

====================
net: mana: Add PF and XDP_REDIRECT support

The patch set adds PF and XDP_REDIRECT support.
====================

Link: https://lore.kernel.org/r/1655238535-19257-1-git-send-email-haiyangz@microsoft.comSigned-off-by: Paolo Abeni <pabeni@redhat.com>

Add a handler of the XDP_REDIRECT return code from a XDP program. The
packets will be flushed at the end of each RX/CQ NAPI poll cycle.
ndo_xdp_xmit() is implemented by sharing the code in mana_xdp_tx().
Ethtool per queue counters are added for XDP redirect and xmit operations.
Signed-off-by: Haiyang Zhang <haiyangz@microsoft.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>