Adds a Python selftest using the YNL devlink API to verify the devlink
rate ops. The test requires a bond device given in the config as NETIF
containing two PFs. Test setup will then create 1 VF on each PF and
verify the various rate commands.
./devlink_rate_cross_esw.py
TAP version 13
1..3
ok 1 devlink_rate_cross_esw.test_same_esw_parent
ok 2 devlink_rate_cross_esw.test_cross_esw_parent
ok 3 devlink_rate_cross_esw.test_tx_rates_on_cross_esw
Tests will be skipped when the preconditions aren't met, when the
devlink API is too old or when the devices don't appear to support
cross-esw scheduling (detected via EOPNOTSUPP).
Signed-off-by: Cosmin Ratiu <cratiu@nvidia.com>
Reviewed-by: Carolina Jubran <cjubran@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/20260701073254.754518-14-tariqt@nvidia.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Up to now, rate groups could only contain vports from the same E-Switch.
This patch relaxes that restriction if the device supports it
(HCA_CAP.esw_cross_esw_sched == true) and the right conditions are met:
- Link Aggregation (LAG) is enabled.
- The E-Switches are from the same shared devlink device.
Signed-off-by: Cosmin Ratiu <cratiu@nvidia.com>
Reviewed-by: Carolina Jubran <cjubran@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/20260701073254.754518-13-tariqt@nvidia.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
E-Switch QoS domains were added with the intention of eventually
implementing shared qos domains to support cross-esw scheduling in the
previous approach ([1]), but they are no longer necessary in the new
approach.
Remove QoS domains and switch to using the shd lock for protecting
against concurrent QoS modifications.
Enable the supported_cross_device_rate_nodes devlink ops attribute so
that all calls originating from devlink rate acquire the shd lock. Only
the additional entry points into QoS need to acquire the shd lock.
The wrinkle is that since shd can be NULL (e.g. on older HW without
serial number available), there needs to be a fallback locking
mechanism. The devlink instance lock cannot be used, as some code paths
into QoS (get, set & modify vport rate) happen with RTNL held, and the
existing devlink -> RTNL order prevents devlink lock usage there.
The other two options are either esw->state_lock or a new lock as
fallback when shd is NULL. This patch adds esw->state_lock, which
implies:
- 3 new lock/unlock helper pairs to acquire/release the missing lock:
- esw_qos_{,un}lock: acquire/release esw->state_lock when shd is NULL.
- esw_qos_shd_{,un}lock: when esw->state_lock is already held.
- esw_qos_devlink_{,un}lock: when shd is already held.
- esw_assert_qos_lock_held now asserts esw->state_lock is held when shd
is NULL.
Use the corresponding lock/unlock function in all places where either
shd or state_lock would need to be acquired.
Document all of this trickery next to esw_assert_qos_lock_held.
Enabling supported_cross_device_rate_nodes now is safe, because
mlx5_esw_qos_vport_update_parent rejects cross-esw parent updates.
This will change in the next patch.
[1]
https://lore.kernel.org/netdev/20250213180134.323929-1-tariqt@nvidia.com/
Signed-off-by: Cosmin Ratiu <cratiu@nvidia.com>
Reviewed-by: Carolina Jubran <cjubran@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/20260701073254.754518-12-tariqt@nvidia.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
In commit [1] the concept of the root node in the qos hierarchy was
removed due to a bug with how tx_share worked. The side effect is that
in many places, there are now corner cases related to parent handling.
However, since that change, support for tc_bw was added and now, with
upcoming cross-esw support, the code is about to become even more
complicated, increasing the number of such corner cases.
Bring back the concept of the root node, to which all esw vports and
nodes are connected to. This benefits multiple operations which can
assume there's always a valid parent and don't have to do ternary
gymnastics to determine the correct esw to talk to.
As side effect, there's no longer a need to store the groups in the
qos domain, since normalization can simply iterate over all children of
the root node. Normalization gets simplified as a result.
There should be no functionality changes as a result of this change.
[1] commit 330f0f6713 ("net/mlx5: Remove default QoS group and attach
vports directly to root TSAR")
Signed-off-by: Cosmin Ratiu <cratiu@nvidia.com>
Reviewed-by: Carolina Jubran <cjubran@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/20260701073254.754518-11-tariqt@nvidia.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Qos cleanup is a complex affair, because of the two modes of operation
(legacy and switchdev).
Leaf QoS is removed:
1. In legacy mode by esw_vport_cleanup() -> mlx5_esw_qos_vport_disable()
2. In switchdev mode by mlx5_esw_offloads_devlink_port_unregister() ->
mlx5_esw_qos_vport_update_parent(). A little later in the same flow, the
calls in 1 happen but they are noops.
Zooming out a bit, from both mlx5_eswitch_disable_locked() and
mlx5_eswitch_disable_sriov() the leaves are destroyed before the nodes,
which is the reverse of what should be.
For SFs there's no devl_rate_nodes_destroy() call to unparent the
affected leaf.
Sanitize all of this by:
1. Destroying nodes before leaves in both legacy and switchdev mode.
2. Only removing vport qos from esw_vport_cleanup(), reachable from both
legacy and switchdev and also reachable by SF removal.
3. Unexpose mlx5_esw_qos_vport_update_parent(), which becomes internal
to qos.
4. Remove the WARN in mlx5_esw_qos_vport_disable().
This also takes care of a theoretical corner case, when
mlx5_esw_qos_vport_update_parent() tried to reattach the vport to
the original parent on failure, which can fail as well, leaving the
vport in a broken state.
Signed-off-by: Cosmin Ratiu <cratiu@nvidia.com>
Reviewed-by: Carolina Jubran <cjubran@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/20260701073254.754518-10-tariqt@nvidia.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Previously, the master device of the uplink netdev was queried for its
maximum link speed from the QoS layer, requiring the uplink_netdev mutex
and possibly the RTNL (if the call originated from the TC matchall
layer).
Acquiring these locks here is risky, as lock cycles could form. The
locking for the QoS layer is about to change, so to avoid issues,
replace the code querying the LAG's max link speed with the existing
infrastructure added in commit [1].
This simplifies this part and avoids potential lock cycles.
One caveat is that there's a new edge case, when the bond device is not
fully formed to represent the LAG device, the speed isn't calculated and
is left at 0. This now handled explicitly.
[1] commit f0b2fde980 ("net/mlx5: Add support for querying bond
speed")
Signed-off-by: Cosmin Ratiu <cratiu@nvidia.com>
Reviewed-by: Carolina Jubran <cjubran@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/20260701073254.754518-9-tariqt@nvidia.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
This commit makes use of the building blocks previously added to
implement cross-device rate nodes.
A new 'supported_cross_device_rate_nodes' bool is added to devlink_ops
which lets drivers advertise support for cross-device rate objects.
If enabled and if there is a common shared devlink instance, then:
- all rate objects will be stored in the top-most common nested instance
and
- rate objects can have parents from other devices sharing the same
common instance.
Storing rates in the common shared ancestor is safe, because it is
reference counted by its nested devlink instances, so it's guaranteed to
outlive them. Furthermore, the shared devlink infra guarantees a given
nested devlink hierarchy is managed by the same driver.
The parent devlink from info->ctx is not locked, so none of its mutable
fields can be used. But parent setting only requires comparing devlink
pointer comparisons. Additionally, since the shared devlink is locked,
other rate operations cannot concurrently happen.
Signed-off-by: Cosmin Ratiu <cratiu@nvidia.com>
Reviewed-by: Carolina Jubran <cjubran@nvidia.com>
Reviewed-by: Jiri Pirko <jiri@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/20260701073254.754518-8-tariqt@nvidia.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Currently, a devlink rate's parent device is assumed to be the same as
the one where the devlink rate is created.
This patch changes that to allow rate commands to accept an additional
argument that specifies the parent dev. This will allow devlink rate
groups with leafs from other devices.
Example of the new usage with ynl:
Creating a group on pci/0000:08:00.1 with a parent to an already
existing pci/0000:08:00.1/group1:
./tools/net/ynl/pyynl/cli.py --spec \
Documentation/netlink/specs/devlink.yaml --do rate-new --json '{
"bus-name": "pci",
"dev-name": "0000:08:00.1",
"rate-node-name": "group2",
"rate-parent-node-name": "group1",
"parent-dev": {
"bus-name": "pci",
"dev-name": "0000:08:00.1"
}
}'
Setting the parent of leaf node pci/0000:08:00.1/65537 to
pci/0000:08:00.0/group1:
./tools/net/ynl/pyynl/cli.py --spec \
Documentation/netlink/specs/devlink.yaml --do rate-set --json '{
"bus-name": "pci",
"dev-name": "0000:08:00.1",
"port-index": 65537,
"parent-dev": {
"bus-name": "pci",
"dev-name": "0000:08:00.0"
},
"rate-parent-node-name": "group1"
}'
Signed-off-by: Cosmin Ratiu <cratiu@nvidia.com>
Reviewed-by: Carolina Jubran <cjubran@nvidia.com>
Reviewed-by: Jiri Pirko <jiri@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/20260701073254.754518-7-tariqt@nvidia.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Upcoming changes to the rate commands need the parent devlink specified.
This change adds a nested 'parent-dev' attribute to the API and helpers
to obtain and put a reference to the parent devlink instance in
info->ctx.
To avoid deadlocks, the parent devlink is unlocked before obtaining the
main devlink instance that is the target of the request.
A reference to the parent is kept until the end of the request to avoid
it suddenly disappearing.
This means that this reference is of limited use without additional
protection.
Signed-off-by: Cosmin Ratiu <cratiu@nvidia.com>
Reviewed-by: Carolina Jubran <cjubran@nvidia.com>
Reviewed-by: Jiri Pirko <jiri@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/20260701073254.754518-6-tariqt@nvidia.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Devlink rate leafs and nodes were stored in their respective devlink
objects pointed to by devlink_rate->devlink.
This patch removes that association by introducing the concept of
'rate node devlink', which is where all rates that could link to each
other are stored. For now this is the same as devlink_rate->devlink.
After this patch, the devlink rates stored in this devlink instance
could potentially be from multiple other devlink instances. So all rate
node manipulation code was updated to:
- correctly compare the actual devlink object during iteration.
- maybe acquire additional locks (noop for now).
Signed-off-by: Cosmin Ratiu <cratiu@nvidia.com>
Reviewed-by: Carolina Jubran <cjubran@nvidia.com>
Reviewed-by: Jiri Pirko <jiri@nvidia.com>
Signed-off-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://patch.msgid.link/20260701073254.754518-5-tariqt@nvidia.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Ivan Vecera says:
====================
dpll: add NCO pin type and zl3073x support
Add a new DPLL_PIN_TYPE_INT_NCO pin type for virtual pins representing
the NCO mode of a DPLL and implement support for it in the zl3073x driver.
Patch 1 adds a STATE_CONNECTED_OVERRIDE pin capability for pins that
can override input selection in any DPLL mode, with a WARN_ON check
in dpll_pin_register() that STATE_CAN_CHANGE is also set.
Patch 2 adds the new INT_NCO pin type to the DPLL netlink spec and
UAPI header.
Patch 3 replaces the single 2s poll timeout with per-operation timeouts
based on Microchip proprietary source code and own measurement.
Patch 4 adds a per-DPLL serialization mutex taken by all DPLL callbacks
and the periodic worker, establishing a single lock that protects all
per-channel state. The chan_state_update() call is moved under this lock.
Patch 5 adds a virtual NCO input pin to the zl3073x driver that allows
userspace to switch a DPLL channel into NCO mode. The pin reports
connected/active state when the channel is in NCO mode and handles
the hardware-specific details of mode transitions including automatic
df_offset capture and 1PPS phase preservation.
Link: https://lore.kernel.org/netdev/20260531194423.383366-1-ivecera@redhat.com/
====================
Link: https://patch.msgid.link/20260630125536.720717-1-ivecera@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Add a virtual NCO (Numerically Controlled Oscillator) input pin that
lets userspace switch a DPLL channel into NCO mode. A single NCO pin
is shared across all DPLL channels - each channel has its own
independent connection state. The NCO pin is registered with the new
DPLL_PIN_TYPE_INT_NCO type and reports DPLL_PIN_STATE_CONNECTED /
DPLL_PIN_OPERSTATE_ACTIVE when the channel is in NCO mode.
At NCO pin registration the following bits are configured in
dpll_ctrl_x:
- nco_auto_read: auto-capture tracking offset on NCO entry
- tod_step_reset: apply negated ToD step accumulator on NCO exit
- tie_clear: PPS DPLLs set 1 to re-align outputs on NCO exit,
EEC DPLLs keep 0 to prevent an unwanted TIE write
Before switching to NCO mode, dpll_df_read_x is configured with
ref_ofst=0 and cmd=ACC_I so that nco_auto_read captures the
accumulated I-part offset relative to the master clock. Without
this, the captured df_offset would be near zero (offset relative
to the input reference after lock).
On NCO entry the df_offset captured by nco_auto_read is read from
the register. Per the datasheet, nco_auto_read only captures a valid
offset when entering NCO from reflock, auto or holdover mode; from
freerun the captured value is not meaningful and df_offset is marked as
ZL_DPLL_DF_OFFSET_UNKNOWN. The same sentinel is set in
chan_state_update() when the channel is not locked, and both FFO
consumers (NCO pin and input pin) guard against it.
Disconnecting the NCO pin switches to freerun rather than holdover
because holdover averaging is not updated during NCO mode.
When connecting the NCO pin displaces a previously connected input
pin (reflock mode), a change notification is sent for that input pin.
Input reference pins are now always registered regardless of the
initial DPLL mode. Previously they were skipped when the DPLL was
in NCO mode, but the NCO pin provides the proper mechanism for
mode transitions.
Reviewed-by: Petr Oros <poros@redhat.com>
Tested-by: Chris du Quesnay <Chris.duQuesnay@microchip.com>
Signed-off-by: Ivan Vecera <ivecera@redhat.com>
Link: https://patch.msgid.link/20260630125536.720717-6-ivecera@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Add a per-DPLL mutex that serializes all operations on a given DPLL
channel across DPLL netlink callbacks, the periodic kthread worker,
and (in subsequent patches) PTP clock callbacks.
All DPLL pin and device callbacks that access mutable state take the
lock as the first operation. The periodic worker holds it for the
entire check cycle of each channel, deferring change notifications
until after the lock is released to avoid ABBA deadlock with
dpll_lock. This establishes the lock ordering:
dpll_lock (subsystem, outer) -> zldpll->lock (driver, inner).
Move zl3073x_chan_state_update() from the per-device
zl3073x_dev_chan_states_update() loop into the per-DPLL
zl3073x_dpll_changes_check() so it runs under zldpll->lock.
This serializes df_offset writes with all readers and
eliminates the need for separate df_offset synchronization.
Change pin->freq_offset from atomic64_t to plain s64 since all
readers and writers are now serialized by zldpll->lock, making
atomic access unnecessary.
Signed-off-by: Ivan Vecera <ivecera@redhat.com>
Reviewed-by: Petr Oros <poros@redhat.com>
Link: https://patch.msgid.link/20260630125536.720717-5-ivecera@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Replace the single 2s timeout in zl3073x_poll_zero_u8() with a
per-caller timeout parameter. Different HW operations have different
expected completion times so using per-operation timeouts improves
error detection. The timeout values are based on proprietary source
code provided by Microchip and own measurement.
Signed-off-by: Ivan Vecera <ivecera@redhat.com>
Reviewed-by: Petr Oros <poros@redhat.com>
Link: https://patch.msgid.link/20260630125536.720717-4-ivecera@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Add DPLL_PIN_TYPE_INT_NCO pin type for virtual pins representing
the NCO mode of a DPLL. When connected as a DPLL input, the DPLL
enters NCO mode where the output frequency is adjusted by the host
via the PTP clock interface.
Update the fractional-frequency-offset and fractional-frequency-
offset-ppt attribute documentation to note that for INT_NCO pins
these attributes represent the DPLL's current output frequency
offset from its nominal frequency.
Reviewed-by: Jiri Pirko <jiri@nvidia.com>
Signed-off-by: Ivan Vecera <ivecera@redhat.com>
Link: https://patch.msgid.link/20260630125536.720717-3-ivecera@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Add DPLL_PIN_CAPABILITIES_STATE_CONNECTED_OVERRIDE capability flag
that indicates a pin can be set to connected regardless of the
current DPLL device mode, overriding the active input selection.
This is useful for automatic-only DPLL devices where mode cannot
be switched to manual, allowing userspace to directly connect
such pin from automatic mode.
The capability requires STATE_CAN_CHANGE to be set as well;
dpll_pin_register() warns if a driver violates this.
Document the new capability in the Pin selection section of
Documentation/driver-api/dpll.rst.
Signed-off-by: Ivan Vecera <ivecera@redhat.com>
Link: https://patch.msgid.link/20260630125536.720717-2-ivecera@redhat.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Add test cases to reproduce scenarios fixed recently [1] where
multiqueue and taprio forced their children into enqueueing an skb to
gso_skb (during peek), but failed to dequeue from gso_skb because they
called the child's dequeue callback directly. This causes a desync in the
child's qlen/backlog and results in an eventual null-ptr-deref (with a
qfq or dualpi2 child).
Test cases are the following:
- Force multiq to dequeue from its child's gso_skb with qfq leaf (fb6c)
- Force multiq to dequeue from its child's gso_skb with dualpi2 leaf (1922)
- Force taprio to dequeue from its child's gso_skb with qfq leaf (476f)
- Force taprio to dequeue from its child's gso_skb with dualpi2 leaf (0235)
[1] https://lore.kernel.org/netdev/20260625-b4-disp-31bcb279-v1-0-85c40b83c529@proton.me/
Signed-off-by: Victor Nogueira <victor@mojatatu.com>
Reviewed-by: Pedro Tammela <pctammela@mojatatu.com>
Link: https://patch.msgid.link/20260630153651.249752-1-victor@mojatatu.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Update amt_fill_info() to run under RCU read lock instead of RTNL.
The AMT device configuration fields (mode, relay_port, gw_port, local_ip,
discovery_ip, max_tunnels) and stream_dev pointer are initialized during
device creation (amt_newlink) and are immutable. Accessing them locklessly
is safe. The stream_dev net_device structure is protected from being freed
by RCU.
The only field that can change concurrently is amt->remote_ip, which is
updated in the packet receive path (amt_advertisement_handler) and
workqueue (amt_req_work). Add READ_ONCE()/WRITE_ONCE() annotations
around amt->remote_ip to prevent data races.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@google.com>
Link: https://patch.msgid.link/20260701125016.3650708-1-edumazet@google.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Add ethtool support to query and configure the PFC (Priority Flow Control)
storm prevention timeout. When TX continuously sends PFC frames, the peer
end is suppressed from sending packets. If this persists, a PFC frame storm
may occur.
This feature allows configuring a timeout to prevent such storms.
Signed-off-by: Hao Chen <chenhao418@huawei.com>
Signed-off-by: Jijie Shao <shaojijie@huawei.com>
Link: https://patch.msgid.link/20260630134043.1532431-1-shaojijie@huawei.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Florian Westphal says:
====================
netfilter: updates for net-next
The following patchset contains Netfilter updates for *net-next*.
1) Update nfnetlink_hook to dump the individual NAT type chains
instead of the nat base chains to userspace. From Phil Sutter.
2) Replace strlcpy/strlcat() with snprintf() in x_tables, from Ian Bridges.
3) Start replacing u_int8_t and u_int16t with u8 and u16 in netfilter.
From Carlos Grillet.
4) Replace strcpy() with strscpy() in netfilter, from David Laight.
5) Remove redundant NULL check before kvfree().
6) Add parameter validation to xt_tcpmss. Ensure mss_min <= mss_max and
invert <= 1. From Feng Wu.
7) Add checkentry for xt_dscp 'tos' match. Implement tos_mt_check() to reject
invalid invert values. Also from Feng Wu.
8) Stop hashing nf_conntrack_helper by tuple. Switch to hashing by name and
L4 protocol.
9) Remove tuples from conntrack helper definitions and port usage from
broadcast helpers. Add netlink policy validation to prevent protocol
number truncation.
10) Remove obsolete netfilter conntrack module parameters.
11) Bound num_counters in ebtables: do_replace() by MAX_EBT_ENTRIES to prevent
oversized vmalloc_array() allocations. From Jiayuan Chen.
12) Make expectations created via nft_ct rules work with NAT.
netfilter pull request nf-next-26-07-02
* tag 'nf-next-26-07-02' of https://git.kernel.org/pub/scm/linux/kernel/git/netfilter/nf-next:
netfilter: nft_ct: support expectation creation for natted flows
netfilter: ebtables: bound num_counters like nentries in do_replace()
netfilter: conntrack: remove obsolete module parameters
netfilter: conntrack: get rid of tuple in helper definitions
netfilter: nf_conntrack_helper: do not hash by tuple
netfilter: xt_dscp: add checkentry for tos match
netfilter: xt_tcpmss: add checkentry for parameter validation
netfilter: remove redundant null check before kvfree()
netfilter: avoid strcpy usage
netfilter: replace u_int8_t and u_int16t with u8 and u16
netfilter: x_tables: replace strlcat() with snprintf()
netfilter: nfnetlink_hook: Dump nat type chains
====================
Link: https://patch.msgid.link/20260702105003.13550-1-fw@strlen.de
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Eric Dumazet says:
====================
macvlan: RTNL-less macvlan_fill_info()
This series removes the RTNL lock dependency from macvlan_fill_info(),
allowing it to run under RCU read lock.
The first patch annotates data races on 'mode' and 'flags' fields which
are accessed locklessly in the RX/TX paths.
The second patch transitions macvlan_fill_info() to RCU, adding necessary
annotations for other fields and handling concurrent updates to the MAC
address list by computing the count dynamically.
====================
Link: https://patch.msgid.link/20260701082214.2974946-1-edumazet@google.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Add READ_ONCE()/WRITE_ONCE() annotations on vlan->mode, vlan->flags,
vlan->bc_queue_len_req and port->bc_cutoff.
Fill IFLA_MACVLAN_MACADDR_DATA nested attribute and compute
on the fly the precise number of elements we put in it,
to fill an accurate IFLA_MACVLAN_MACADDR_COUNT attribute
as some user space applications could depend on its value
and the attributes order.
Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Nikolay Aleksandrov <razor@blackwall.org>
Reviewed-by: Kuniyuki Iwashima <kuniyu@google.com>
Link: https://patch.msgid.link/20260701082214.2974946-3-edumazet@google.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Jakub Kicinski says:
====================
tools: ynl: pyynl: pull the --family resolution logic into the lib
When packaging YNL as a system level utility we added a --family
argument which auto-resolves the full spec path from a well known
path in /usr/share. Spelling out full YAML spec files is at this
point only done in-tree, for example in the selftests which need
the very latest YAML. But the selftests have their own wrapping
classes for each family so test authors aren't really bothered
by having to spell the paths out.
Afford the same ease of use to the Python library users.
Move the path resolution from the CLI code to the library.
This simplifies the pyynl use by a lot:
from pyynl import YnlFamily
ynl = YnlFamily(family="netdev")
Unless I'm missing a trick, resolving the /usr/share path
is hard enough for most users to lean towards shelling out
to ynl CLI with --output-json, which is sad.
v1: https://lore.kernel.org/20260630001432.2204298-3-kuba@kernel.org
====================
Link: https://patch.msgid.link/20260701021751.3234681-1-kuba@kernel.org
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
When packaging YNL as a system level utility we added a --family
argument which auto-resolves the full spec path from a well known
path in /usr/share. Spelling out full YAML spec files is at this
point only done in-tree, for example in the selftests which need
the very latest YAML. But the selftests have their own wrapping
classes for each family so test authors aren't really bothered
by having to spell the paths out.
Afford the same ease of use to the Python library users.
Move the path resolution from the CLI code to the library.
This simplifies the pyynl use by a lot:
from pyynl import YnlFamily
ynl = YnlFamily(family="netdev")
Unless I'm missing a trick, resolving the /usr/share path
is hard enough for most users to lean towards shelling out
to ynl CLI with --output-json, which is sad.
The ethtool script can now use family= instead of
resolving the path (the helpers are removed from cli.py
so this isn't just a cleanup).
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
Reviewed-by: Donald Hunter <donald.hunter@gmail.com>
Link: https://patch.msgid.link/20260701021751.3234681-3-kuba@kernel.org
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Yuyang Huang says:
====================
net: report multicast group user count
RTM_GETMULTICAST reports IPv4 and IPv6 multicast group membership, but
does not include the per-group user count. Userspace therefore still has
to parse /proc/net/igmp and /proc/net/igmp6 to obtain the Users column.
In particular, this prevents iproute2 from moving "ip maddr show"
entirely from procfs to rtnetlink.
Add IFA_MC_USERS to carry the user count in RTM_GETMULTICAST dumps and
RTM_NEWMULTICAST / RTM_DELMULTICAST notifications for both address
families. Update the rt-addr YNL specification and extend the rtnetlink
selftest to verify that two joins increase the reported count by two.
====================
Link: https://patch.msgid.link/20260630110207.37841-1-sigefriedhyy@gmail.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Extend the RTM_GETMULTICAST dump test to verify IFA_MC_USERS for both
IPv4 and IPv6 multicast groups.
Run each protocol test in a fresh network namespace to avoid changing
host-network state or racing with unrelated multicast users. Join a
fixed multicast group twice using separate sockets and check that the
reported user count increases by two.
Signed-off-by: Yuyang Huang <sigefriedhyy@gmail.com>
Reviewed-by: Vadim Fedorenko <vadim.fedorenko@linux.dev>
Link: https://patch.msgid.link/20260630110207.37841-4-sigefriedhyy@gmail.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
The previous patch added IFA_MC_USERS and emits it for IPv4 multicast
groups. Add the same snapshot attribute to IPv6 RTM_GETMULTICAST
replies and entry-lifecycle notifications, carrying
ifmcaddr6::mca_users.
This makes the multicast rtnetlink ABI symmetric across IPv4 and IPv6
and gives userspace the same user count that /proc/net/igmp6 exposes.
Signed-off-by: Yuyang Huang <sigefriedhyy@gmail.com>
Reviewed-by: Vadim Fedorenko <vadim.fedorenko@linux.dev>
Reviewed-by: Ido Schimmel <idosch@nvidia.com>
Link: https://patch.msgid.link/20260630110207.37841-3-sigefriedhyy@gmail.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
RTM_GETMULTICAST has been part of the rtnetlink ABI for a long time
and already reports IPv4 multicast group membership through
IFA_MULTICAST and IFA_CACHEINFO. It does not report how many consumers
hold each membership, so userspace still has to parse /proc/net/igmp to
get the Users column.
Add IFA_MC_USERS as a u32 attribute carrying ip_mc_list::users in
RTM_GETMULTICAST replies and entry-lifecycle notifications.
This gives iproute2 enough information to migrate the IPv4 part of
"ip maddr show" from procfs parsing to rtnetlink.
Signed-off-by: Yuyang Huang <sigefriedhyy@gmail.com>
Reviewed-by: Vadim Fedorenko <vadim.fedorenko@linux.dev>
Reviewed-by: Ido Schimmel <idosch@nvidia.com>
Link: https://patch.msgid.link/20260630110207.37841-2-sigefriedhyy@gmail.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
inet_dump_fib() saves its progress in cb->args[1] as a positional
index within the current hash chain. Between batches, a concurrent
fib_new_table() can insert a new table at the chain head, shifting
all existing entries. On resume the saved index lands on a different
table, causing already-dumped tables to be re-dumped and the
originally suspended table to restart from the beginning.
Fix by storing tb->tb_id in cb->args[1] instead of a positional
index, mirroring the fix applied to inet6_dump_fib() in commit
9facb861dc ("ipv6: fib6: fix NULL deref in fib6_walk_continue()
on multi-batch dump").
Signed-off-by: Pengfei Zhang <zhangfeionline@gmail.com>
Reviewed-by: Ido Schimmel <idosch@nvidia.com>
Link: https://patch.msgid.link/20260630084220.2711025-1-zhangfeionline@gmail.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Cross-merge networking fixes after downstream PR (net-7.2-rc2).
No conflicts.
Adjacent changes:
MAINTAINERS:
56114690ff ("MAINTAINERS: Update Marvell octeontx2 driver maintainers")
eb56577ae9 ("ehea: remove the ehea driver")
net/core/netpoll.c:
45f1458a85 ("netpoll: fix a use-after-free on shutdown path")
84c0ff1efb ("netpoll: do not warn when the best-effort pool refill fails")
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Pull networking fixes from Paolo Abeni:
"Including fixes from netfilter and batman-adv.
Current release - new code bugs:
- netfilter: cthelper: cap to maximum number of expectation per master
Previous releases - regressions:
- netpoll: fix a use-after-free on shutdown path
- tcp: restore RCU grace period in tcp_ao_destroy_sock
- ipv6: fix NULL deref in fib6_walk_continiue() on multi-batch dump
- batman-adv: dat: ensure accessible eth_hdr proto field
- eth:
- virtio_net: disable cb when NAPI is busy-polled
- lan743x: Initialize eth_syslock spinlock before use
Previous releases - always broken:
- netfilter:
- nft_set_pipapo: don't leak bad clone into future transaction
- sched:
- sch_teql: Introduce slaves_lock to avoid race condition and UAF
- replace direct dequeue call with peek and qdisc_dequeue_peeked
- sctp: add INIT verification after cookie unpacking
- tipc: fix out-of-bounds read in broadcast Gap ACK blocks
- seg6: validate SRH length before reading fixed fields
- eth:
- mlx5e: fix use-after-free of metadata_dst on RX SC delete
- enetc: check the number of BDs needed for xdp_frame
- fbnic: don't cache shinfo across skb realloc"
* tag 'net-7.2-rc2' of git://git.kernel.org/pub/scm/linux/kernel/git/netdev/net: (58 commits)
net/mlx5: HWS, fix matcher leak on resize target setup failure
net/sched: hhf: clear heavy-hitter state on reset
net/sched: dualpi2: clear stale classification on filter miss
net/sched: act_bpf: use rcu_dereference_bh() to read the filter
selftests: drv-net: tso: don't touch dangerous feature bits
cxgb4: Fix decode strings dump for T6 adapters
virtio_net: disable cb when NAPI is busy-polled
sctp: fix addr_wq_timer race in sctp_free_addr_wq()
selftests: net: bump default cmd() timeout to 20 seconds
bridge: stp: Fix a potential use-after-free when deleting a bridge
net/sched: sch_teql: Introduce slaves_lock to avoid race condition and UAF
net: gianfar: dispose irq mappings on probe failure and device removal
net: lan743x: Initialize eth_syslock spinlock before use
net: libwx: fix VMDQ mask for 1-queue mode
net: airoha: fix max receive size configuration
fsl/fman: Free init resources on KeyGen failure in fman_init()
netfilter: nftables: restrict checkum update offset
netfilter: nftables: restrict linklayer and network header writes
netfilter: nfnetlink_queue: restrict writes to network header
netfilter: nft_fib: reject fib expression on the netdev egress hook
...
Pull MFD fix from Lee Jones:
- Add MFD mailing list to MAINTAINERS
* tag 'mfd-fixes-7.2' of git://git.kernel.org/pub/scm/linux/kernel/git/lee/mfd:
MAINTAINERS: Add a mailing list entry to MFD
Kiran Kumar says:
====================
octeontx2-af: NPC parser and RSS improvements
This series extends the Marvell OcteonTX2 admin-function driver with two
improvements to the NPC (Network Parser CAM) block. The NPC parses packets
received by or transmitted from the NIX, and its matching CAM (MCAM)
selects which VFs, queues, or output ports handle each packet.
Patch 1 reserves a new range of PKINDs (46-53) to support configurable
L2 skip-size parsing. Packets arriving with variable length L2 headers
or a CPT (Cryptographic Accelerator Unit) pre-header can be steered
to one of four skip-size PKINDs so the NPC advances past the right number
of bytes before starting protocol classification. The mbox interface is
extended with a skip_size field so PF/VF drivers can program the desired
L2 offset at run time without rebuilding firmware.
Patch 2 adds a NIX_FLOW_KEY_TYPE_ROCEV2 flow-key type so the RSS engine
can distribute RoCEv2 traffic across receive queues using the destination
Queue Pair (QP) field. Without this, all RoCEv2 flows hash the same
way and land on a single queue.
Both changes target the admin-function driver to improve overall hardware
parsing infrastructure.
====================
Link: https://patch.msgid.link/20260630062145.2533816-1-nshettyj@marvell.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
The NPC block uses PKINDs to determine how incoming packets are
parsed. Reserve PKINDs 46-49 (NPC_RX_SKIP_SIZE_PKIND) for
configurable L2 skip-size use in the first pass, and PKINDs 50-53
(NPC_RX_CPT_SKIP_SIZE_PKIND) for the second pass where packets
carry a CPT (Cryptographic Accelerator Unit) header.
Add npc_set_skip_size_pkind() to program NPC_AF_PKINDX_ACTION0
for these reserved PKINDs with a user-supplied ptr_advance value
representing the L2 size to skip. For the corresponding CPT PKINDs
(pkind + 4), additionally configure the var_len_offset, var_len_mask,
var_len_shift, and var_len_right fields so the NPC can extract the
inner payload length from the CPT header.
Update rvu_npc_set_parse_mode() to accept a new skip_size argument
and dispatch to npc_set_skip_size_pkind() when the requested PKIND
falls in the newly reserved range. Extend the npc_set_pkind mbox
message struct with a skip_size field so PF/VF drivers can supply
this value at run time.
Advance NPC_UNRESERVED_PKIND_COUNT to NPC_RX_SKIP_SIZE_PKIND to
reflect the updated reservation boundary.
Signed-off-by: Kiran Kumar K <kirankumark@marvell.com>
Signed-off-by: Nitin Shetty J <nshettyj@marvell.com>
Link: https://patch.msgid.link/20260630062145.2533816-2-nshettyj@marvell.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
The mvneta driver uses the hardware Buffer Manager (BM) for RX buffer
allocation. During suspend, mvneta disables its clock, causing BM to
lose all buffer address state. On resume, mvneta_bm_port_init() re-
attaches the BM pool to the NIC, but BM hardware returns stale/garbage
buffer addresses. When NAPI poll processes these buffers, DMA cache
sync hits an invalid virtual address causing a kernel panic:
Unable to handle kernel paging request at virtual address b0000080
PC is at v7_dma_inv_range
Call trace:
v7_dma_inv_range from arch_sync_dma_for_cpu+0x94/0x158
arch_sync_dma_for_cpu from __dma_sync_single_for_cpu+0xc4/0x15c
__dma_sync_single_for_cpu from mvneta_rx_swbm+0x6c8/0xf48
mvneta_rx_swbm from mvneta_poll+0x6fc/0x70c
mvneta_poll from __napi_poll.constprop.0+0x2c/0x1e0
__napi_poll.constprop.0 from net_rx_action+0x160/0x2c4
net_rx_action from handle_softirqs+0xd8/0x2b8
handle_softirqs from run_ksoftirqd+0x30/0x94
run_ksoftirqd from smpboot_thread_fn+0x100/0x204
smpboot_thread_fn from kthread+0xf4/0x110
kthread from ret_from_fork+0x14/0x28
Fix by adding suspend/resume callbacks to the BM driver:
- suspend: drain all buffers (with DMA unmapping), free the BPPE
regions, and reset pool state to FREE before stopping BM and gating
the clock.
- resume: enable the clock, reinitialize BM defaults, and restore pool
read/write pointers and size registers. Pool allocation and buffer
refill are handled by mvneta_resume() through the normal
mvneta_bm_port_init() path, which sees pools as FREE and performs
full initialization identical to probe.
Add a device_link (DL_FLAG_AUTOREMOVE_CONSUMER) in mvneta_probe to
guarantee BM resumes before mvneta and suspends after mvneta. If the
link cannot be created, fall back to SW buffer management to avoid a
potential crash on resume due to unordered PM transitions.
Signed-off-by: Yun Zhou <yun.zhou@windriver.com>
Link: https://patch.msgid.link/20260630060311.4072140-1-yun.zhou@windriver.com
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Breno Leitao says:
====================
net: convert UDP getsockopt to sockopt_t
The leaf proto_ops getsockopt callbacks have been moving to the new
getsockopt_iter()/sockopt_t interface.
I was trying to get SMC into getsockop and retire .getsockopt, but,
I found the best approach is to keep converting other protocols.
This series starts the same conversion one layer down, at the struct proto
getsockopt path, beginning with UDP.
Example of the current code.
static int udp_getsockopt(struct sock *sk, int level, int optname,
char __user *optval, int __user *optlen)
{
if (level == SOL_UDP)
return udp_lib_getsockopt(sk, level, optname, optval, optlen);
return ip_getsockopt(sk, level, optname, optval, optlen);
}
We want udp_getsockopt to go to .getsockopt_iter, and there are two
approaches in this case:
1) Create a patchset that moves both of them to getsockopt_iter, which
is will be a huge change (ip_getsockopt() is used in many places)
2) Break this down, and transform from bottoms up. First
udp_lib_getsockopt() up to the point we can easily convert
others, such as ip_getsockopt().
I am taking the approach 2), so, the intermediate code will be something
like:
static int udp_getsockopt(struct sock *sk, int level, int optname,
char __user *optval, int __user *optlen)
{
sockopt_t opt;
int err;
if (level != SOL_UDP)
return ip_getsockopt(sk, level, optname, optval, optlen);
// Convert optlen/optval in sockopt // (first patch)
err = udp_lib_getsockopt(sk, level, optname, &opt);
}
The work is bottom-up and mergeable in small steps: a protocol's inner
getsockopt helper is switched to sockopt_t behind its existing thin
__user wrapper, one patch at a time.
Once every inner helper speaks sockopt_t, a later series flips the shared
struct proto.getsockopt and inet_connection_sock_af_ops.getsockopt signatures
and drops the transitional wrappers.
Signed-off-by: Breno Leitao <leitao@debian.org>
====================
Link: https://patch.msgid.link/20260630-getsockopt_phase2-v2-0-193335f3d4d1@debian.org
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Exercise the raw getsockopt path now backed by sockopt_t. ICMP_FILTER
returns a fixed-size struct and, unlike the int/u64 options already
covered, clamps the length down to the user buffer on a short read
instead of failing, so check that semantic explicitly along with the
exact and oversized cases, the -EOPNOTSUPP path on a non-ICMP raw
socket, and an unknown optname.
Signed-off-by: Breno Leitao <leitao@debian.org>
Acked-by: Stanislav Fomichev <sdf@fomichev.me>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Link: https://patch.msgid.link/20260630-getsockopt_phase2-v2-4-193335f3d4d1@debian.org
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Continue converting the proto-layer getsockopt callbacks to the sockopt_t
interface, switching do_raw_getsockopt() and its raw_geticmpfilter()
helper to take a sockopt_t.
The thin raw_getsockopt() wrapper keeps its __user signature for now: it
builds a user-backed sockopt_t with sockopt_init_user(), calls the helper,
and writes the returned length back to optlen. The helper uses
copy_to_iter() instead of copy_to_user(). No functional change.
Signed-off-by: Breno Leitao <leitao@debian.org>
Acked-by: Stanislav Fomichev <sdf@fomichev.me>
Acked-by: Willem de Bruijn <willemb@google.com>
Link: https://patch.msgid.link/20260630-getsockopt_phase2-v2-3-193335f3d4d1@debian.org
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
