platform/kernel/linux-rpi.git
6 years agosch_cake: Conditionally split GSO segments
Toke Høiland-Jørgensen [Fri, 6 Jul 2018 15:37:19 +0000 (17:37 +0200)]
sch_cake: Conditionally split GSO segments

At lower bandwidths, the transmission time of a single GSO segment can add
an unacceptable amount of latency due to HOL blocking. Furthermore, with a
software shaper, any tuning mechanism employed by the kernel to control the
maximum size of GSO segments is thrown off by the artificial limit on
bandwidth. For this reason, we split GSO segments into their individual
packets iff the shaper is active and configured to a bandwidth <= 1 Gbps.

Signed-off-by: Toke Høiland-Jørgensen <toke@toke.dk>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agosch_cake: Add overhead compensation support to the rate shaper
Toke Høiland-Jørgensen [Fri, 6 Jul 2018 15:37:19 +0000 (17:37 +0200)]
sch_cake: Add overhead compensation support to the rate shaper

This commit adds configurable overhead compensation support to the rate
shaper. With this feature, userspace can configure the actual bottleneck
link overhead and encapsulation mode used, which will be used by the shaper
to calculate the precise duration of each packet on the wire.

This feature is needed because CAKE is often deployed one or two hops
upstream of the actual bottleneck (which can be, e.g., inside a DSL or
cable modem). In this case, the link layer characteristics and overhead
reported by the kernel does not match the actual bottleneck. Being able to
set the actual values in use makes it possible to configure the shaper rate
much closer to the actual bottleneck rate (our experience shows it is
possible to get with 0.1% of the actual physical bottleneck rate), thus
keeping latency low without sacrificing bandwidth.

The overhead compensation has three tunables: A fixed per-packet overhead
size (which, if set, will be accounted from the IP packet header), a
minimum packet size (MPU) and a framing mode supporting either ATM or PTM
framing. We include a set of common keywords in TC to help users configure
the right parameters. If no overhead value is set, the value reported by
the kernel is used.

Signed-off-by: Toke Høiland-Jørgensen <toke@toke.dk>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agosch_cake: Add DiffServ handling
Toke Høiland-Jørgensen [Fri, 6 Jul 2018 15:37:19 +0000 (17:37 +0200)]
sch_cake: Add DiffServ handling

This adds support for DiffServ-based priority queueing to CAKE. If the
shaper is in use, each priority tier gets its own virtual clock, which
limits that tier's rate to a fraction of the overall shaped rate, to
discourage trying to game the priority mechanism.

CAKE defaults to a simple, three-tier mode that interprets most code points
as "best effort", but places CS1 traffic into a low-priority "bulk" tier
which is assigned 1/16 of the total rate, and a few code points indicating
latency-sensitive or control traffic (specifically TOS4, VA, EF, CS6, CS7)
into a "latency sensitive" high-priority tier, which is assigned 1/4 rate.
The other supported DiffServ modes are a 4-tier mode matching the 802.11e
precedence rules, as well as two 8-tier modes, one of which implements
strict precedence of the eight priority levels.

This commit also adds an optional DiffServ 'wash' mode, which will zero out
the DSCP fields of any packet passing through CAKE. While this can
technically be done with other mechanisms in the kernel, having the feature
available in CAKE significantly decreases configuration complexity; and the
implementation cost is low on top of the other DiffServ-handling code.

Filters and applications can set the skb->priority field to override the
DSCP-based classification into tiers. If TC_H_MAJ(skb->priority) matches
CAKE's qdisc handle, the minor number will be interpreted as a priority
tier if it is less than or equal to the number of configured priority
tiers.

Signed-off-by: Toke Høiland-Jørgensen <toke@toke.dk>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agosch_cake: Add NAT awareness to packet classifier
Toke Høiland-Jørgensen [Fri, 6 Jul 2018 15:37:19 +0000 (17:37 +0200)]
sch_cake: Add NAT awareness to packet classifier

When CAKE is deployed on a gateway that also performs NAT (which is a
common deployment mode), the host fairness mechanism cannot distinguish
internal hosts from each other, and so fails to work correctly.

To fix this, we add an optional NAT awareness mode, which will query the
kernel conntrack mechanism to obtain the pre-NAT addresses for each packet
and use that in the flow and host hashing.

When the shaper is enabled and the host is already performing NAT, the cost
of this lookup is negligible. However, in unlimited mode with no NAT being
performed, there is a significant CPU cost at higher bandwidths. For this
reason, the feature is turned off by default.

Cc: netfilter-devel@vger.kernel.org
Signed-off-by: Toke Høiland-Jørgensen <toke@toke.dk>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonetfilter: Add nf_ct_get_tuple_skb global lookup function
Toke Høiland-Jørgensen [Fri, 6 Jul 2018 15:37:19 +0000 (17:37 +0200)]
netfilter: Add nf_ct_get_tuple_skb global lookup function

This adds a global netfilter function to extract a conntrack tuple from an
skb. The function uses a new function added to nf_ct_hook, which will try
to get the tuple from skb->_nfct, and do a full lookup if that fails. This
makes it possible to use the lookup function before the skb has passed
through the conntrack init hooks (e.g., in an ingress qdisc). The tuple is
copied to the caller to avoid issues with reference counting.

The function returns false if conntrack is not loaded, allowing it to be
used without incurring a module dependency on conntrack. This is used by
the NAT mode in sch_cake.

Cc: netfilter-devel@vger.kernel.org
Signed-off-by: Toke Høiland-Jørgensen <toke@toke.dk>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agosch_cake: Add optional ACK filter
Toke Høiland-Jørgensen [Fri, 6 Jul 2018 15:37:19 +0000 (17:37 +0200)]
sch_cake: Add optional ACK filter

The ACK filter is an optional feature of CAKE which is designed to improve
performance on links with very asymmetrical rate limits. On such links
(which are unfortunately quite prevalent, especially for DSL and cable
subscribers), the downstream throughput can be limited by the number of
ACKs capable of being transmitted in the *upstream* direction.

Filtering ACKs can, in general, have adverse effects on TCP performance
because it interferes with ACK clocking (especially in slow start), and it
reduces the flow's resiliency to ACKs being dropped further along the path.
To alleviate these drawbacks, the ACK filter in CAKE tries its best to
always keep enough ACKs queued to ensure forward progress in the TCP flow
being filtered. It does this by only filtering redundant ACKs. In its
default 'conservative' mode, the filter will always keep at least two
redundant ACKs in the queue, while in 'aggressive' mode, it will filter
down to a single ACK.

The ACK filter works by inspecting the per-flow queue on every packet
enqueue. Starting at the head of the queue, the filter looks for another
eligible packet to drop (so the ACK being dropped is always closer to the
head of the queue than the packet being enqueued). An ACK is eligible only
if it ACKs *fewer* bytes than the new packet being enqueued, including any
SACK options. This prevents duplicate ACKs from being filtered, to avoid
interfering with retransmission logic. In addition, we check TCP header
options and only drop those that are known to not interfere with sender
state. In particular, packets with unknown option codes are never dropped.

In aggressive mode, an eligible packet is always dropped, while in
conservative mode, at least two ACKs are kept in the queue. Only pure ACKs
(with no data segments) are considered eligible for dropping, but when an
ACK with data segments is enqueued, this can cause another pure ACK to
become eligible for dropping.

The approach described above ensures that this ACK filter avoids most of
the drawbacks of a naive filtering mechanism that only keeps flow state but
does not inspect the queue. This is the rationale for including the ACK
filter in CAKE itself rather than as separate module (as the TC filter, for
instance).

Our performance evaluation has shown that on a 30/1 Mbps link with a
bidirectional traffic test (RRUL), turning on the ACK filter on the
upstream link improves downstream throughput by ~20% (both modes) and
upstream throughput by ~12% in conservative mode and ~40% in aggressive
mode, at the cost of ~5ms of inter-flow latency due to the increased
congestion.

In *really* pathological cases, the effect can be a lot more; for instance,
the ACK filter increases the achievable downstream throughput on a link
with 100 Kbps in the upstream direction by an order of magnitude (from ~2.5
Mbps to ~25 Mbps).

Finally, even though we consider the ACK filter to be safer than most, we
do not recommend turning it on everywhere: on more symmetrical link
bandwidths the effect is negligible at best.

Cc: Yuchung Cheng <ycheng@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Signed-off-by: Toke Høiland-Jørgensen <toke@toke.dk>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agosch_cake: Add ingress mode
Toke Høiland-Jørgensen [Fri, 6 Jul 2018 15:37:19 +0000 (17:37 +0200)]
sch_cake: Add ingress mode

The ingress mode is meant to be enabled when CAKE runs downlink of the
actual bottleneck (such as on an IFB device). The mode changes the shaper
to also account dropped packets to the shaped rate, as these have already
traversed the bottleneck.

Enabling ingress mode will also tune the AQM to always keep at least two
packets queued *for each flow*. This is done by scaling the minimum queue
occupancy level that will disable the AQM by the number of active bulk
flows. The rationale for this is that retransmits are more expensive in
ingress mode, since dropped packets have to traverse the bottleneck again
when they are retransmitted; thus, being more lenient and keeping a minimum
number of packets queued will improve throughput in cases where the number
of active flows are so large that they saturate the bottleneck even at
their minimum window size.

This commit also adds a separate switch to enable ingress mode rate
autoscaling. If enabled, the autoscaling code will observe the actual
traffic rate and adjust the shaper rate to match it. This can help avoid
latency increases in the case where the actual bottleneck rate decreases
below the shaped rate. The scaling filters out spikes by an EWMA filter.

Signed-off-by: Toke Høiland-Jørgensen <toke@toke.dk>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agosched: Add Common Applications Kept Enhanced (cake) qdisc
Toke Høiland-Jørgensen [Fri, 6 Jul 2018 15:37:19 +0000 (17:37 +0200)]
sched: Add Common Applications Kept Enhanced (cake) qdisc

sch_cake targets the home router use case and is intended to squeeze the
most bandwidth and latency out of even the slowest ISP links and routers,
while presenting an API simple enough that even an ISP can configure it.

Example of use on a cable ISP uplink:

tc qdisc add dev eth0 cake bandwidth 20Mbit nat docsis ack-filter

To shape a cable download link (ifb and tc-mirred setup elided)

tc qdisc add dev ifb0 cake bandwidth 200mbit nat docsis ingress wash

CAKE is filled with:

* A hybrid Codel/Blue AQM algorithm, "Cobalt", tied to an FQ_Codel
  derived Flow Queuing system, which autoconfigures based on the bandwidth.
* A novel "triple-isolate" mode (the default) which balances per-host
  and per-flow FQ even through NAT.
* An deficit based shaper, that can also be used in an unlimited mode.
* 8 way set associative hashing to reduce flow collisions to a minimum.
* A reasonable interpretation of various diffserv latency/loss tradeoffs.
* Support for zeroing diffserv markings for entering and exiting traffic.
* Support for interacting well with Docsis 3.0 shaper framing.
* Extensive support for DSL framing types.
* Support for ack filtering.
* Extensive statistics for measuring, loss, ecn markings, latency
  variation.

A paper describing the design of CAKE is available at
https://arxiv.org/abs/1804.07617, and will be published at the 2018 IEEE
International Symposium on Local and Metropolitan Area Networks (LANMAN).

This patch adds the base shaper and packet scheduler, while subsequent
commits add the optional (configurable) features. The full userspace API
and most data structures are included in this commit, but options not
understood in the base version will be ignored.

Various versions baking have been available as an out of tree build for
kernel versions going back to 3.10, as the embedded router world has been
running a few years behind mainline Linux. A stable version has been
generally available on lede-17.01 and later.

sch_cake replaces a combination of iptables, tc filter, htb and fq_codel
in the sqm-scripts, with sane defaults and vastly simpler configuration.

CAKE's principal author is Jonathan Morton, with contributions from
Kevin Darbyshire-Bryant, Toke Høiland-Jørgensen, Sebastian Moeller,
Ryan Mounce, Tony Ambardar, Dean Scarff, Nils Andreas Svee, Dave Täht,
and Loganaden Velvindron.

Testing from Pete Heist, Georgios Amanakis, and the many other members of
the cake@lists.bufferbloat.net mailing list.

tc -s qdisc show dev eth2
 qdisc cake 8017: root refcnt 2 bandwidth 1Gbit diffserv3 triple-isolate split-gso rtt 100.0ms noatm overhead 38 mpu 84
 Sent 51504294511 bytes 37724591 pkt (dropped 6, overlimits 64958695 requeues 12)
  backlog 0b 0p requeues 12
  memory used: 1053008b of 15140Kb
  capacity estimate: 970Mbit
  min/max network layer size:           28 /    1500
  min/max overhead-adjusted size:       84 /    1538
  average network hdr offset:           14
                    Bulk  Best Effort        Voice
   thresh      62500Kbit        1Gbit      250Mbit
   target          5.0ms        5.0ms        5.0ms
   interval      100.0ms      100.0ms      100.0ms
   pk_delay          5us          5us          6us
   av_delay          3us          2us          2us
   sp_delay          2us          1us          1us
   backlog            0b           0b           0b
   pkts          3164050     25030267      9530280
   bytes      3227519915  35396974782  12879808898
   way_inds            0            8            0
   way_miss           21          366           25
   way_cols            0            0            0
   drops               5            0            1
   marks               0            0            0
   ack_drop            0            0            0
   sp_flows            1            3            0
   bk_flows            0            1            1
   un_flows            0            0            0
   max_len         68130        68130        68130

Tested-by: Pete Heist <peteheist@gmail.com>
Tested-by: Georgios Amanakis <gamanakis@gmail.com>
Signed-off-by: Dave Taht <dave.taht@gmail.com>
Signed-off-by: Toke Høiland-Jørgensen <toke@toke.dk>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: Use __u32 in uapi net_stamp.h
Jesus Sanchez-Palencia [Mon, 9 Jul 2018 23:20:56 +0000 (16:20 -0700)]
net: Use __u32 in uapi net_stamp.h

We are not supposed to use u32 in uapi, so change the flags member of
struct sock_txtime from u32 to __u32 instead.

Fixes: 80b14dee2bea ("net: Add a new socket option for a future transmit time")
Reported-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Jesus Sanchez-Palencia <jesus.sanchez-palencia@intel.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agoMerge branch 'mlxsw-More-Spectrum-2-preparations'
David S. Miller [Mon, 9 Jul 2018 23:24:18 +0000 (16:24 -0700)]
Merge branch 'mlxsw-More-Spectrum-2-preparations'

aIdo Schimmel says:

====================
mlxsw: More Spectrum-2 preparations

This is the second and last set of preparations towards initial
Spectrum-2 support in mlxsw. It mainly re-arranges parts of the code
that need to work with both ASICs, but somewhat differ.

The first three patches allow different ASICs to register different set
of operations for KVD linear (KVDL) management. In Spectrum-2 there is
no linear memory and instead entries that reside there in Spectrum
(e.g., nexthops) are hashed and inserted to the hash-based KVD memory.

The fourth patch does a similar restructuring in the low-level multicast
router code. This is necessary because multicast routing is implemented
using regular circuit TCAM (C-TCAM) in Spectrum, whereas Spectrum-2 uses
an algorithmic TCAM (A-TCAM).

Next six patches prepare the ACL code for the introduction of A-TCAM in
follow-up patch sets.

Last two patches allow different ASICs to require different firmware
versions and add two resources that need to be queried from firmware by
Spectrum-2 specific code.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agomlxsw: resources: Add couple of Spectrum-2 KVD resources
Jiri Pirko [Sun, 8 Jul 2018 20:51:27 +0000 (23:51 +0300)]
mlxsw: resources: Add couple of Spectrum-2 KVD resources

These resources are needed for Spectrum-2 KVD linear management
implementation.

Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agomlxsw: spectrum: Prepare for multiple FW versions for Spectrum and Spectrum-2
Jiri Pirko [Sun, 8 Jul 2018 20:51:26 +0000 (23:51 +0300)]
mlxsw: spectrum: Prepare for multiple FW versions for Spectrum and Spectrum-2

Prepare for Spectrum-2 FW version checking and
make mlxsw_sp_fw_rev_validate() per-ASIC as well as required FW revision
and FW filename.

Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agomlxsw: spectrum_acl: Implement priority setting for rules inserted to TCAM
Jiri Pirko [Sun, 8 Jul 2018 20:51:25 +0000 (23:51 +0300)]
mlxsw: spectrum_acl: Implement priority setting for rules inserted to TCAM

For Spectrum-2, we need to insert priority to C-TCAM because HW
needs that info in order to correctly process scenarios where rules
are in both C-TCAM and A-TCAM.

So extend the mlxsw_sp_acl_ctcam_entry_add() args to accept indication
if priority needs to be filled up and implement the priority
computation and fill-up.

Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agomlxsw: reg: Add priority field for PTCEV2 register
Jiri Pirko [Sun, 8 Jul 2018 20:51:24 +0000 (23:51 +0300)]
mlxsw: reg: Add priority field for PTCEV2 register

This is going to be needed for Spectrum-2 C-TCAM implementation.

Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agomlxsw: spectrum_acl: Move block items encoding into Spectrum op
Jiri Pirko [Sun, 8 Jul 2018 20:51:23 +0000 (23:51 +0300)]
mlxsw: spectrum_acl: Move block items encoding into Spectrum op

Since Spectrum-2 encodes blocks into different HW layout, push this
code into Spectrum-specific op.

Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agomlxsw: spectrum_acl: Convert mlxsw_afk_create args to ops
Jiri Pirko [Sun, 8 Jul 2018 20:51:22 +0000 (23:51 +0300)]
mlxsw: spectrum_acl: Convert mlxsw_afk_create args to ops

Since the flex keys for Spectrum-2 differ not only in blocks definitions
but also in encoding layout, prepare for the implementation and pass
Spectrum/Spectrum-2 specific ops down to mlxsw_afk_create.

Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agomlxsw: spectrum_acl: Add tcam init/fini ops
Jiri Pirko [Sun, 8 Jul 2018 20:51:21 +0000 (23:51 +0300)]
mlxsw: spectrum_acl: Add tcam init/fini ops

Add ops to be called on driver instance init and fini.
This is needed in order to be possible to do Spectrum-2 specific init
and fini work.

Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agomlxsw: spectrum_acl: Split TCAM handling 3 ways
Jiri Pirko [Sun, 8 Jul 2018 20:51:20 +0000 (23:51 +0300)]
mlxsw: spectrum_acl: Split TCAM handling 3 ways

To allow easy and clean Spectrum-2 implementation for things that differ
from Spectrum, split the existing ACL TCAM code 3 ways:
1) common code that calls Spectrum/Spectrum-2 specific ops
2) Spectrum ops implementations
3) common C-TCAM code that is going to be shared between Spectrum and
   Spectrum-2 implementations

Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agomlxsw: spectrum_mr_tcam: Push Spectrum-specific operations into a separate file
Jiri Pirko [Sun, 8 Jul 2018 20:51:19 +0000 (23:51 +0300)]
mlxsw: spectrum_mr_tcam: Push Spectrum-specific operations into a separate file

Since Spectrum-2 has different handling of TCAM, push Spectrum MR TCAM
bits to a separate file accessible by ops which allows to implement
Spectrum-2 specific ops.

Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agomlxsw: spectrum_kvdl: Pass entry_count to free function
Jiri Pirko [Sun, 8 Jul 2018 20:51:18 +0000 (23:51 +0300)]
mlxsw: spectrum_kvdl: Pass entry_count to free function

For the Spectrum-2 KVD linear manager implementation, entry_count will be
needed even for the free function. So pass it down.

Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agomlxsw: spectrum_kvdl: Pass entry type to alloc/free
Jiri Pirko [Sun, 8 Jul 2018 20:51:17 +0000 (23:51 +0300)]
mlxsw: spectrum_kvdl: Pass entry type to alloc/free

Future Spectrum-2 KVD linear manager implementation needs to know type
of the entry to alloc and free. So define the types in an enum and
pass it down to alloc and free functions. Once the entry type
is passed down, KVDL common part knows sizes of each entry types,
so replace size function arg with entry count.

Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agomlxsw: spectrum_kvdl: Push out KVD linear management into ops
Jiri Pirko [Sun, 8 Jul 2018 20:51:16 +0000 (23:51 +0300)]
mlxsw: spectrum_kvdl: Push out KVD linear management into ops

In Spectrum-2 there is a different implementation of KVD linear
management. Unlike in Spectrum where there is a single index space,
in Spectrum-2 the indexes are per-resource. Also there is need to
explicitly tell HW that an entry is no longer used.
So push out the existing implementation into spectrum1_kvdl.c and
prepare ops infrastructure to allow new implementation in a follow-up.

Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet/mlx5: Use 2-factor allocator calls
Kees Cook [Wed, 4 Jul 2018 17:28:47 +0000 (10:28 -0700)]
net/mlx5: Use 2-factor allocator calls

This restores the use of 2-factor allocation helpers that were already
fixed treewide. Please do not use open-coded multiplication; prefer,
instead, using 2-factor allocation helpers.

Signed-off-by: Kees Cook <keescook@chromium.org>
Reviewed-by: Leon Romanovsky <leonro@mellanox.com>
Reviewed-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agotcp: remove SG-related comment in tcp_sendmsg()
Julian Wiedmann [Mon, 9 Jul 2018 07:45:14 +0000 (09:45 +0200)]
tcp: remove SG-related comment in tcp_sendmsg()

Since commit 74d4a8f8d378 ("tcp: remove sk_can_gso() use"), the code
doesn't care whether the interface supports SG.

Signed-off-by: Julian Wiedmann <jwi@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agoMerge branch 'fix-use-after-free-bugs-in-skb-list-processing'
David S. Miller [Mon, 9 Jul 2018 21:55:54 +0000 (14:55 -0700)]
Merge branch 'fix-use-after-free-bugs-in-skb-list-processing'

Edward Cree says:

====================
fix use-after-free bugs in skb list processing

A couple of bugs in skb list handling were spotted by Dan Carpenter, with
 the help of Smatch; following up on them I found a couple more similar
 cases.  This series fixes them by changing the relevant loops to use the
 dequeue-enqueue model (rather than in-place list modification).

v3: fixed another similar bug in __netif_receive_skb_list_core().

v2: dropped patch #3 (new list.h helper), per DaveM's request.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: core: fix use-after-free in __netif_receive_skb_list_core
Edward Cree [Mon, 9 Jul 2018 17:10:19 +0000 (18:10 +0100)]
net: core: fix use-after-free in __netif_receive_skb_list_core

__netif_receive_skb_core can free the skb, so we have to use the dequeue-
 enqueue model when calling it from __netif_receive_skb_list_core.

Fixes: 88eb1944e18c ("net: core: propagate SKB lists through packet_type lookup")
Signed-off-by: Edward Cree <ecree@solarflare.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonetfilter: fix use-after-free in NF_HOOK_LIST
Edward Cree [Mon, 9 Jul 2018 17:10:02 +0000 (18:10 +0100)]
netfilter: fix use-after-free in NF_HOOK_LIST

nf_hook() can free the skb, so we need to remove it from the list before
 calling, and add passed skbs to a sublist afterwards.

Fixes: 17266ee93984 ("net: ipv4: listified version of ip_rcv")
Reported-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Edward Cree <ecree@solarflare.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: core: fix uses-after-free in list processing
Edward Cree [Mon, 9 Jul 2018 17:09:54 +0000 (18:09 +0100)]
net: core: fix uses-after-free in list processing

In netif_receive_skb_list_internal(), all of skb_defer_rx_timestamp(),
 do_xdp_generic() and enqueue_to_backlog() can lead to kfree(skb).  Thus,
 we cannot wait until after they return to remove the skb from the list;
 instead, we remove it first and, in the pass case, add it to a sublist
 afterwards.
In the case of enqueue_to_backlog() we have already decided not to pass
 when we call the function, so we do not need a sublist.

Fixes: 7da517a3bc52 ("net: core: Another step of skb receive list processing")
Reported-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Edward Cree <ecree@solarflare.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agotcp: remove redundant SOCK_DONE checks
Eric Dumazet [Sun, 8 Jul 2018 06:15:56 +0000 (23:15 -0700)]
tcp: remove redundant SOCK_DONE checks

In both tcp_splice_read() and tcp_recvmsg(), we already test
sock_flag(sk, SOCK_DONE) right before evaluating sk->sk_state,
so "!sock_flag(sk, SOCK_DONE)" is always true.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agoMerge branch 'mlxsw-Spectrum2-acl-prep'
David S. Miller [Sun, 8 Jul 2018 08:05:20 +0000 (17:05 +0900)]
Merge branch 'mlxsw-Spectrum2-acl-prep'

Ido Schimmel says:

====================
mlxsw: Spectrum-2 small ACL preparations

This is the first set of changes towards Spectrum-2 support in the mlxsw
driver. It contains small changes that prepare the code for the later
introduction of Spectrum-2 support.

The Spectrum-2 ASIC uses an algorithmic TCAM (A-TCAM) instead of a
circuit TCAM (C-TCAM) as Spectrum, and thus most of the changes are
around the ACL code.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agomlxsw: core_acl_flex_actions: Fix helper to get the first KVD linear index
Jiri Pirko [Sun, 8 Jul 2018 07:00:21 +0000 (10:00 +0300)]
mlxsw: core_acl_flex_actions: Fix helper to get the first KVD linear index

The helper should return always KVD linear index of the second set.
It is unused now, but going to be used soon.

Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agomlxsw: core_acl_flex_actions: Allow the first set to be dummy
Jiri Pirko [Sun, 8 Jul 2018 07:00:20 +0000 (10:00 +0300)]
mlxsw: core_acl_flex_actions: Allow the first set to be dummy

In Spectrum-2, the real action sets are always in KVD linear. The first
set is always empty and contains only pointer to the first real set in
KVD linear. So provide possibility to specify the first set is the dummy
one.

Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agomlxsw: spectrum: Put pointer to flex action ops to mlxsw_sp
Jiri Pirko [Sun, 8 Jul 2018 07:00:19 +0000 (10:00 +0300)]
mlxsw: spectrum: Put pointer to flex action ops to mlxsw_sp

Spectrum-2 need a slightly different handling of flexible actions. So
put an ops pointer in mlxsw_sp struct and rename it.

Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agomlxsw: core_acl_flex_keys: Change SRC_SYS_PORT flex key element size
Jiri Pirko [Sun, 8 Jul 2018 07:00:18 +0000 (10:00 +0300)]
mlxsw: core_acl_flex_keys: Change SRC_SYS_PORT flex key element size

The SRC_SYS_PORT is passed as 8 bit value down to hw anyway, so cap it
in the driver as well. Also, in Spectrum-2 the FW iface for SRC_SYS_PORT
is only 8 bits, so prepare for it.

Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agomlxsw: core_acl_flex_keys: Split MAC and IP address flex key elements
Jiri Pirko [Sun, 8 Jul 2018 07:00:17 +0000 (10:00 +0300)]
mlxsw: core_acl_flex_keys: Split MAC and IP address flex key elements

Since in Spectrum-2, MACs are split and IP addresses are split as well,
in order to use the same elements for Spectrum and Spectrum-2 split them
now.

Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agomlxsw: spectrum_acl: Ignore always-zeroed bits in tp->prio
Jiri Pirko [Sun, 8 Jul 2018 07:00:16 +0000 (10:00 +0300)]
mlxsw: spectrum_acl: Ignore always-zeroed bits in tp->prio

The lowest 16 bits of tp->prio are always zero, so ignore them with a
shift.

Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agomlxsw: reg: Introduce Flex2 key type for PTAR register
Jiri Pirko [Sun, 8 Jul 2018 07:00:15 +0000 (10:00 +0300)]
mlxsw: reg: Introduce Flex2 key type for PTAR register

Introduce Flex2 key type for PTAR register which is used in Spectrum-2.
Also, extend mlxsw_reg_ptar_pack() to set the value according to the
caller.

Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agomlxsw: spectrum: Change name of mlxsw_sp_afk_blocks to mlxsw_sp1_afk_blocks
Jiri Pirko [Sun, 8 Jul 2018 07:00:14 +0000 (10:00 +0300)]
mlxsw: spectrum: Change name of mlxsw_sp_afk_blocks to mlxsw_sp1_afk_blocks

This is specific for Spectrum as Spectrum-2 has completely different key
blocks.

Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: sched: Fix warnings from xchg() on RCU'd cookie pointer.
David S. Miller [Sun, 8 Jul 2018 08:02:59 +0000 (17:02 +0900)]
net: sched: Fix warnings from xchg() on RCU'd cookie pointer.

The kbuild test robot reports:

>> net/sched/act_api.c:71:15: sparse: incorrect type in initializer (different address spaces) @@    expected struct tc_cookie [noderef] <asn:4>*__ret @@    got [noderef] <asn:4>*__ret @@
   net/sched/act_api.c:71:15:    expected struct tc_cookie [noderef] <asn:4>*__ret
   net/sched/act_api.c:71:15:    got struct tc_cookie *new_cookie
>> net/sched/act_api.c:71:13: sparse: incorrect type in assignment (different address spaces) @@    expected struct tc_cookie *old @@    got struct tc_cookie [noderef] <struct tc_cookie *old @@
   net/sched/act_api.c:71:13:    expected struct tc_cookie *old
   net/sched/act_api.c:71:13:    got struct tc_cookie [noderef] <asn:4>*[assigned] __ret
>> net/sched/act_api.c:132:48: sparse: dereference of noderef expression

Handle this in the usual way by force casting away the __rcu annotation
when we are using xchg() on it.

Fixes: eec94fdb0480 ("net: sched: use rcu for action cookie update")
Reported-by: kbuild test robot <lkp@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agoMerge branch 'Modify-action-API-for-implementing-lockless-actions'
David S. Miller [Sun, 8 Jul 2018 03:42:29 +0000 (12:42 +0900)]
Merge branch 'Modify-action-API-for-implementing-lockless-actions'

Vlad Buslov says:

====================
Modify action API for implementing lockless actions

Currently, all netlink protocol handlers for updating rules, actions and
qdiscs are protected with single global rtnl lock which removes any
possibility for parallelism. This patch set is a first step to remove
rtnl lock dependency from TC rules update path.

Recently, new rtnl registration flag RTNL_FLAG_DOIT_UNLOCKED was added.
Handlers registered with this flag are called without RTNL taken. End
goal is to have rule update handlers(RTM_NEWTFILTER, RTM_DELTFILTER,
etc.) to be registered with UNLOCKED flag to allow parallel execution.
However, there is no intention to completely remove or split rtnl lock
itself. This patch set addresses specific problems in action API that
prevents it from being executed concurrently. This patch set does not
completely unlock rules or actions update path. Additional patch sets
are required to refactor individual actions and filters update for
parallel execution.

As a preparation for executing TC rules update handlers without rtnl
lock, action API code was audited to determine areas that assume
external synchronization with rtnl lock and must be changed to allow
safe concurrent access with following results:

1. Action idr is already protected with spinlock. However, some code
   paths assume that idr state is not changes between several
   consecutive tcf_idr_* function calls.
2. tc_action reference and bind counters are implemented as plain
   integers. They purpose was to allow single actions to be shared
   between multiple filters, not to provide means for concurrent
   modification.
3. tc_action 'cookie' pointer field is not protected against
   modification.
4. Action API functions, that work with set of actions, use intrusive
   linked list, which cannot be used concurrently without additional
   synchronization.
5. Action API functions don't take reference to actions while using
   them, assuming external synchronization with rtnl lock.

Following solutions to these problems are implemented:

1. To remove assumption that idr state doesn't change between tcf_idr_*
   calls, implement new functions that atomically perform several
   operations on idr without releasing idr spinlock. (function to
   atomically lookup and delete action by index, function to atomically
   check if action exists and allocate new one if necessary, etc.)
2. Use atomic operations on counters to make them suitable for
   concurrent get/put operations.
3. Data that 'cookie' points to is never modified, so it enough to
   refactor it to rcu pointer to prevent concurrent de-allocation.
4. Action API doesn't actually use any linked list specific operations
   on actions intrusive linked list, so it can be refactored to array in
   straightforward manner.
5. Always take reference to action while accessing it in action API.
   tcf_idr_search function modified to take reference to action before
   returning it, so there is no way to lookup an action without
   incrementing its reference counter. All users of this function are
   modified to release the reference, after they done using action. With
   all users using reference counting, it is now safe to concurrently
   delete actions.

Additionally, actions init function signature was expanded with
'rtnl_held' argument, that allows actions that have internal dependency
on rtnl lock to take/release it when necessary.

Since only shared state in action API module are actions themselves and
action idr, these changes are sufficient to not to rely on global rtnl
lock for protection of internal action API data structures.

Changes from V5 to V6:
- Rebase on current net-next
- When action is deleted, set pointer in actions array to NULL to
  prevent double freeing.

Changes from V4 to V5:
- Change action delete API to track actions that were deleted, to
  prevent releasing them on error.

Changes from V3 to V4:
- Expand cover letter.
- Reduce actions array size in tcf_action_init_1.
- Rebase on latest net-next.

Changes from V2 to V3:
- Re-send with changelog copied to individual patches.

Changes from V1 to V2:
- Removed redundant actions ops lookup during delete.
- Merge action ops delete definition and implementation.
- Assume all actions have delete implemented and don't check for it
  explicitly.
- Resplit action lookup/release code to prevent memory leaks in
  individual patches.
- Make __tcf_idr_check function static
- Remove unique idr insertion function. Change original idr insert to do
  the same thing.
- Merge changes that take reference to action when performing lookup and
  changes that account for this additional reference when dumping action
  to user space into single patch.
- Change convoluted commit message.
- Rename "unlocked" to "rtnl_held" for clarity.
- Remove estimator lock add patch.
- Refactor action check-alloc code into standalone function.
- Rename tcf_idr_find_delete to tcf_idr_delete_index.
- Rearrange variable definitions in tc_action_delete.
- Add patch that refactors action API code to use array of pointers to
  actions instead of intrusive linked list.
- Expand cover letter.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: sched: change action API to use array of pointers to actions
Vlad Buslov [Thu, 5 Jul 2018 14:24:33 +0000 (17:24 +0300)]
net: sched: change action API to use array of pointers to actions

Act API used linked list to pass set of actions to functions. It is
intrusive data structure that stores list nodes inside action structure
itself, which means it is not safe to modify such list concurrently.
However, action API doesn't use any linked list specific operations on this
set of actions, so it can be safely refactored into plain pointer array.

Refactor action API to use array of pointers to tc_actions instead of
linked list. Change argument 'actions' type of exported action init,
destroy and dump functions.

Acked-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: Vlad Buslov <vladbu@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: sched: atomically check-allocate action
Vlad Buslov [Thu, 5 Jul 2018 14:24:32 +0000 (17:24 +0300)]
net: sched: atomically check-allocate action

Implement function that atomically checks if action exists and either takes
reference to it, or allocates idr slot for action index to prevent
concurrent allocations of actions with same index. Use EBUSY error pointer
to indicate that idr slot is reserved.

Implement cleanup helper function that removes temporary error pointer from
idr. (in case of error between idr allocation and insertion of newly
created action to specified index)

Refactor all action init functions to insert new action to idr using this
API.

Reviewed-by: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Signed-off-by: Vlad Buslov <vladbu@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: sched: use reference counting action init
Vlad Buslov [Thu, 5 Jul 2018 14:24:31 +0000 (17:24 +0300)]
net: sched: use reference counting action init

Change action API to assume that action init function always takes
reference to action, even when overwriting existing action. This is
necessary because action API continues to use action pointer after init
function is done. At this point action becomes accessible for concurrent
modifications, so user must always hold reference to it.

Implement helper put list function to atomically release list of actions
after action API init code is done using them.

Reviewed-by: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Signed-off-by: Vlad Buslov <vladbu@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: sched: don't release reference on action overwrite
Vlad Buslov [Thu, 5 Jul 2018 14:24:30 +0000 (17:24 +0300)]
net: sched: don't release reference on action overwrite

Return from action init function with reference to action taken,
even when overwriting existing action.

Action init API initializes its fourth argument (pointer to pointer to tc
action) to either existing action with same index or newly created action.
In case of existing index(and bind argument is zero), init function returns
without incrementing action reference counter. Caller of action init then
proceeds working with action, without actually holding reference to it.
This means that action could be deleted concurrently.

Change action init behavior to always take reference to action before
returning successfully, in order to protect from concurrent deletion.

Reviewed-by: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Signed-off-by: Vlad Buslov <vladbu@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: sched: implement reference counted action release
Vlad Buslov [Thu, 5 Jul 2018 14:24:29 +0000 (17:24 +0300)]
net: sched: implement reference counted action release

Implement helper delete function that uses new action ops 'delete', instead
of destroying action directly. This is required so act API could delete
actions by index, without holding any references to action that is being
deleted.

Implement function __tcf_action_put() that releases reference to action and
frees it, if necessary. Refactor action deletion code to use new put
function and not to rely on rtnl lock. Remove rtnl lock assertions that are
no longer needed.

Reviewed-by: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Signed-off-by: Vlad Buslov <vladbu@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: sched: add 'delete' function to action ops
Vlad Buslov [Thu, 5 Jul 2018 14:24:28 +0000 (17:24 +0300)]
net: sched: add 'delete' function to action ops

Extend action ops with 'delete' function. Each action type to implements
its own delete function that doesn't depend on rtnl lock.

Implement delete function that is required to delete actions without
holding rtnl lock. Use action API function that atomically deletes action
only if it is still in action idr. This implementation prevents concurrent
threads from deleting same action twice.

Reviewed-by: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Signed-off-by: Vlad Buslov <vladbu@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: sched: implement action API that deletes action by index
Vlad Buslov [Thu, 5 Jul 2018 14:24:27 +0000 (17:24 +0300)]
net: sched: implement action API that deletes action by index

Implement new action API function that atomically finds and deletes action
from idr by index. Intended to be used by lockless actions that do not rely
on rtnl lock.

Reviewed-by: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Signed-off-by: Vlad Buslov <vladbu@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: sched: always take reference to action
Vlad Buslov [Thu, 5 Jul 2018 14:24:26 +0000 (17:24 +0300)]
net: sched: always take reference to action

Without rtnl lock protection it is no longer safe to use pointer to tc
action without holding reference to it. (it can be destroyed concurrently)

Remove unsafe action idr lookup function. Instead of it, implement safe tcf
idr check function that atomically looks up action in idr and increments
its reference and bind counters. Implement both action search and check
using new safe function

Reference taken by idr check is temporal and should not be accounted by
userspace clients (both logically and to preserver current API behavior).
Subtract temporal reference when dumping action to userspace using existing
tca_get_fill function arguments.

Reviewed-by: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Signed-off-by: Vlad Buslov <vladbu@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: sched: implement unlocked action init API
Vlad Buslov [Thu, 5 Jul 2018 14:24:25 +0000 (17:24 +0300)]
net: sched: implement unlocked action init API

Add additional 'rtnl_held' argument to act API init functions. It is
required to implement actions that need to release rtnl lock before loading
kernel module and reacquire if afterwards.

Reviewed-by: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Signed-off-by: Vlad Buslov <vladbu@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: sched: change type of reference and bind counters
Vlad Buslov [Thu, 5 Jul 2018 14:24:24 +0000 (17:24 +0300)]
net: sched: change type of reference and bind counters

Change type of action reference counter to refcount_t.

Change type of action bind counter to atomic_t.
This type is used to allow decrementing bind counter without testing
for 0 result.

Reviewed-by: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Signed-off-by: Vlad Buslov <vladbu@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: sched: use rcu for action cookie update
Vlad Buslov [Thu, 5 Jul 2018 14:24:23 +0000 (17:24 +0300)]
net: sched: use rcu for action cookie update

Implement functions to atomically update and free action cookie
using rcu mechanism.

Reviewed-by: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Signed-off-by: Vlad Buslov <vladbu@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agoopenvswitch: kernel datapath clone action
Yifeng Sun [Mon, 2 Jul 2018 15:18:03 +0000 (08:18 -0700)]
openvswitch: kernel datapath clone action

Add 'clone' action to kernel datapath by using existing functions.
When actions within clone don't modify the current flow, the flow
key is not cloned before executing clone actions.

This is a follow up patch for this incomplete work:
https://patchwork.ozlabs.org/patch/722096/

v1 -> v2:
Refactor as advised by reviewer.

Signed-off-by: Yifeng Sun <pkusunyifeng@gmail.com>
Signed-off-by: Andy Zhou <azhou@ovn.org>
Acked-by: Pravin B Shelar <pshelar@ovn.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agoisdn/capi: fix defined but not used warnings
Randy Dunlap [Sat, 7 Jul 2018 15:31:15 +0000 (08:31 -0700)]
isdn/capi: fix defined but not used warnings

Fix build warnings in drivers/isdn/capi/ when CONFIG_PROC_FS is not
enabled by marking the unused functions as __maybe_unused.

../drivers/isdn/capi/capi.c:1324:12: warning: 'capi20_proc_show' defined but not used [-Wunused-function]
../drivers/isdn/capi/capi.c:1347:12: warning: 'capi20ncci_proc_show' defined but not used [-Wunused-function]
../drivers/isdn/capi/capidrv.c:2454:12: warning: 'capidrv_proc_show' defined but not used [-Wunused-function]

Signed-off-by: Randy Dunlap <rdunlap@infradead.org>
Cc: Karsten Keil <isdn@linux-pingi.de>
Cc: isdn4linux@listserv.isdn4linux.de (subscribers-only)
Cc: netdev@vger.kernel.org
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agoconnector: fix defined but not used warning
Randy Dunlap [Sat, 7 Jul 2018 15:27:53 +0000 (08:27 -0700)]
connector: fix defined but not used warning

Fix a build warning in connector.c when CONFIG_PROC_FS is not enabled
by marking the unused function as __maybe_unused.

../drivers/connector/connector.c:242:12: warning: 'cn_proc_show' defined but not used [-Wunused-function]

Signed-off-by: Randy Dunlap <rdunlap@infradead.org>
Cc: Evgeniy Polyakov <zbr@ioremap.net>
Cc: netdev@vger.kernel.org
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agodrivers: net: lmc: remove redundant variable next_rx
Colin Ian King [Sat, 7 Jul 2018 12:47:06 +0000 (13:47 +0100)]
drivers: net: lmc: remove redundant variable next_rx

Variable next_rx is being assigned but is never used hence it is
redundant and can be removed.

Cleans up clang warning:
warning: variable 'next_rx' set but not used [-Wunused-but-set-variable]

Signed-off-by: Colin Ian King <colin.king@canonical.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agoMerge branch 'cpsw-allow-PTP-224.0.0.107-to-be-timestamped'
David S. Miller [Sat, 7 Jul 2018 12:22:57 +0000 (21:22 +0900)]
Merge branch 'cpsw-allow-PTP-224.0.0.107-to-be-timestamped'

Ivan Khoronzhuk says:

====================
allow PTP 224.0.0.107 to be timestamped

Allows packets with 224.0.0.107 mcas address for PTP packets to be timestamped.
Only for cpsw version > v1.15.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: ethernet: ti: cpsw: allow PTP 224.0.0.107 to be timestamped
Ivan Khoronzhuk [Fri, 6 Jul 2018 18:44:45 +0000 (21:44 +0300)]
net: ethernet: ti: cpsw: allow PTP 224.0.0.107 to be timestamped

Tested on AM572x with cpsw v1.15 for PTP sync and delay_req messages.
It doesn't work on cpsw v1.12, so added only for cpsw v > 1.15.

Command for testing:
ptp4l -P -4 -H -i eth0 -l 6 -m -q -p /dev/ptp0 -f ptp.cfg
where ptp.cfg:

[global]
tx_timestamp_timeout     20

Signed-off-by: Ivan Khoronzhuk <ivan.khoronzhuk@linaro.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: ethernet: ti: cpsw: use BIT macro
Ivan Khoronzhuk [Fri, 6 Jul 2018 18:44:44 +0000 (21:44 +0300)]
net: ethernet: ti: cpsw: use BIT macro

It's needed to avoid checkpatch warnings for farther changes.

Signed-off-by: Ivan Khoronzhuk <ivan.khoronzhuk@linaro.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agostmmac: fix signed 64-bit division
Arnd Bergmann [Fri, 6 Jul 2018 13:36:07 +0000 (15:36 +0200)]
stmmac: fix signed 64-bit division

I link error on 32-bit ARM points to yet another arithmetic bug:

drivers/net/ethernet/stmicro/stmmac/stmmac_tc.o: In function `tc_setup_cbs':
stmmac_tc.c:(.text+0x148): undefined reference to `__aeabi_uldivmod'
stmmac_tc.c:(.text+0x1fc): undefined reference to `__aeabi_uldivmod'
stmmac_tc.c:(.text+0x308): undefined reference to `__aeabi_uldivmod'
stmmac_tc.c:(.text+0x320): undefined reference to `__aeabi_uldivmod'
stmmac_tc.c:(.text+0x33c): undefined reference to `__aeabi_uldivmod'
drivers/net/ethernet/stmicro/stmmac/stmmac_tc.o:stmmac_tc.c:(.text+0x3a4): more undefined references to `__aeabi_uldivmod' follow

I observe that the last change to add the 'ul' prefix was incorrect,
as it did not turn the result of the multiplication into a 64-bit
expression on 32-bit architectures. Further, it seems that the
do_div() macro gets confused by the fact that we pass a signed
variable rather than unsigned into it.

This changes the code to instead use the div_s64() helper that is
meant for signed division, along with changing the constant suffix
to 'll' to actually make it a 64-bit argument everywhere, fixing
both of the issues I pointed out.

I'm not completely convinced that this makes the code correct, but
I'm fairly sure that we have two problems less than before.

Fixes: 1f705bc61aee ("net: stmmac: Add support for CBS QDISC")
Fixes: c18a9c096683 ("net: stmmac_tc: use 64-bit arithmetic instead of 32-bit")
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Acked-by: Gustavo A. R. Silva <gustavo@embeddedor.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agotipc: extend link reset criteria for stale packet retransmission
Jon Maloy [Fri, 6 Jul 2018 13:22:36 +0000 (15:22 +0200)]
tipc: extend link reset criteria for stale packet retransmission

Currently a link is declared stale and reset if there has been 100
repeated attempts to retransmit the same packet. However, in certain
infrastructures we see that packet (NACK) duplicates and delays may
cause such retransmit attempts to occur at a high rate, so that the
peer doesn't have a reasonable chance to acknowledge the reception
before the 100-limit is hit. This may take much less than the
stipulated link tolerance time, and despite that probe/probe replies
otherwise go through as normal.

We now extend the criteria for link reset to also being time based.
I.e., we don't reset the link until the link tolerance time is passed
AND we have made 100 retransmissions attempts.

Acked-by: Ying Xue <ying.xue@windriver.com>
Signed-off-by: Jon Maloy <jon.maloy@ericsson.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agoMerge branch 'Introduce-matching-on-double-vlan-QinQ-headers-for-TC-flower'
David S. Miller [Sat, 7 Jul 2018 11:51:53 +0000 (20:51 +0900)]
Merge branch 'Introduce-matching-on-double-vlan-QinQ-headers-for-TC-flower'

Jianbo Liu says:

====================
Introduce matching on double vlan/QinQ headers for TC flower

Currently TC flower supports only one vlan tag, it doesn't match on both outer
and inner vlan headers for QinQ. To do this, we add support to get both outer
and inner vlan headers for flow dissector, and then TC flower do matching on
those information.

We also plan to extend TC command to support this feature. We add new
cvlan_id/cvlan_prio/cvlan_ethtype keywords for inner vlan header. The existing
vlan_id/vlan_prio/vlan_ethtype are for outer vlan header, and vlan_ethtype must
be 802.1q or 802.1ad.

The examples for command and output are as the following.
        flower vlan_id 1000 vlan_ethtype 802.1q \
        cvlan_id 100 cvlan_ethtype ipv4 \
        action vlan pop \
        action vlan pop \
        action mirred egress redirect dev ens1f1_0

filter protocol 802.1ad pref 33 flower chain 0
filter protocol 802.1ad pref 33 flower chain 0 handle 0x1
  vlan_id 1000
  vlan_ethtype 802.1Q
  cvlan_id 100
  cvlan_ethtype ip
  eth_type ipv4
  in_hw
    ...

v2:
  fix sparse warning.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet/sched: flower: Add supprt for matching on QinQ vlan headers
Jianbo Liu [Fri, 6 Jul 2018 05:38:16 +0000 (05:38 +0000)]
net/sched: flower: Add supprt for matching on QinQ vlan headers

As support dissecting of QinQ inner and outer vlan headers, user can
add rules to match on QinQ vlan headers.

Signed-off-by: Jianbo Liu <jianbol@mellanox.com>
Acked-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet/sched: flower: Dump the ethertype encapsulated in vlan
Jianbo Liu [Fri, 6 Jul 2018 05:38:15 +0000 (05:38 +0000)]
net/sched: flower: Dump the ethertype encapsulated in vlan

Currently the encapsulated ethertype is not dumped as it's the same as
TCA_FLOWER_KEY_ETH_TYPE keyvalue. But the dumping result is inconsistent
with input, we add dumping it with TCA_FLOWER_KEY_VLAN_ETH_TYPE.

Signed-off-by: Jianbo Liu <jianbol@mellanox.com>
Acked-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet/flow_dissector: Add support for QinQ dissection
Jianbo Liu [Fri, 6 Jul 2018 05:38:14 +0000 (05:38 +0000)]
net/flow_dissector: Add support for QinQ dissection

Dissect the QinQ packets to get both outer and inner vlan information,
then store to the extended flow keys.

Signed-off-by: Jianbo Liu <jianbol@mellanox.com>
Acked-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet/sched: flower: Add support for matching on vlan ethertype
Jianbo Liu [Fri, 6 Jul 2018 05:38:13 +0000 (05:38 +0000)]
net/sched: flower: Add support for matching on vlan ethertype

As flow dissector stores vlan ethertype, tc flower now can match on that.
It is to make preparation for supporting QinQ.

Signed-off-by: Jianbo Liu <jianbol@mellanox.com>
Acked-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet/flow_dissector: Save vlan ethertype from headers
Jianbo Liu [Fri, 6 Jul 2018 05:38:12 +0000 (05:38 +0000)]
net/flow_dissector: Save vlan ethertype from headers

Change vlan dissector key to save vlan tpid to support both 802.1Q
and 802.1AD ethertype.

Signed-off-by: Jianbo Liu <jianbol@mellanox.com>
Acked-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agofarsync: remove redundant variable txq_length
Colin Ian King [Sat, 7 Jul 2018 11:27:30 +0000 (12:27 +0100)]
farsync: remove redundant variable txq_length

Variable txq_length is being assigned but is never used hence it is
redundant and can be removed.

Cleans up clang warning:
warning: variable 'txq_length' set but not used [-Wunused-but-set-variable]

Signed-off-by: Colin Ian King <colin.king@canonical.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agoMerge branch 'Add-a-mlxsw-specific-test-for-mirror-to-gretap'
David S. Miller [Sat, 7 Jul 2018 11:29:34 +0000 (20:29 +0900)]
Merge branch 'Add-a-mlxsw-specific-test-for-mirror-to-gretap'

Petr Machata says:

====================
Add a mlxsw-specific test for mirror-to-gretap

Some configurations of mirror-to-gretap are impossible for mlxsw to
offload. Add a test that checks that these out-of-domain conditions are
handled properly by mlxsw.

In patch #1, fix mirror_gre_lib.sh and mirror_gre_topo_lib.sh so that
they can be imported from directories other than forwarding/.

In patch #2, add a test to check handling of several scenarios that
mlxsw is expected to fail to offload.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agoselftests: mlxsw: Add mlxsw-specific test for mirror to gretap
Petr Machata [Thu, 5 Jul 2018 19:09:04 +0000 (21:09 +0200)]
selftests: mlxsw: Add mlxsw-specific test for mirror to gretap

Test several aspects of offloading mirror to gretap and ip6gretap
netdevices that are specific to mlxsw, such as requirements for TTL and
TOS values.

Signed-off-by: Petr Machata <petrm@mellanox.com>
Reviewed-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agoselftests: forwarding: Allow importing dependent libraries
Petr Machata [Thu, 5 Jul 2018 19:10:59 +0000 (21:10 +0200)]
selftests: forwarding: Allow importing dependent libraries

The next patch introduces a new mlxsw-specific test that uses
mirror_gre_lib.sh and mirror_gre_topo_lib.sh.

However when sourcing their own deps, these libraries assume that the
test that's running is in the same directory. That's not the case for
driver-specific tests.

So change the libraries to source their deps through $relative_path.
That variable is set up by lib.sh, which should be imported by the test
in question in any case.

Signed-off-by: Petr Machata <petrm@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agoMerge branch 'vxlan-fix-default-fdb-entry-user-space-notify-ordering-race'
David S. Miller [Sat, 7 Jul 2018 11:20:46 +0000 (20:20 +0900)]
Merge branch 'vxlan-fix-default-fdb-entry-user-space-notify-ordering-race'

Roopa Prabhu says:

====================
vxlan: fix default fdb entry user-space notify ordering/race

Problem:
In vxlan_newlink, a default fdb entry is added before register_netdev.
The default fdb creation function notifies user-space of the
fdb entry on the vxlan device which user-space does not know about yet.
(RTM_NEWNEIGH goes before RTM_NEWLINK for the same ifindex).

This series fixes the user-space netlink notification ordering issue
with the following changes:
- decouple fdb notify from fdb create.
- Move fdb notify after register_netdev.
- modify rtnl_configure_link to allow configuring a link early.
- Call rtnl_configure_link in vxlan newlink handler to notify
userspace about the newlink before fdb notify and
hence avoiding the user-space race.
====================

Fixes: afbd8bae9c79 ("vxlan: add implicit fdb entry for default destination")
Signed-off-by: Roopa Prabhu <roopa@cumulusnetworks.com>
6 years agovxlan: fix default fdb entry netlink notify ordering during netdev create
Roopa Prabhu [Wed, 4 Jul 2018 23:46:32 +0000 (16:46 -0700)]
vxlan: fix default fdb entry netlink notify ordering during netdev create

Problem:
In vxlan_newlink, a default fdb entry is added before register_netdev.
The default fdb creation function also notifies user-space of the
fdb entry on the vxlan device which user-space does not know about yet.
(RTM_NEWNEIGH goes before RTM_NEWLINK for the same ifindex).

This patch fixes the user-space netlink notification ordering issue
with the following changes:
- decouple fdb notify from fdb create.
- Move fdb notify after register_netdev.
- Call rtnl_configure_link in vxlan newlink handler to notify
userspace about the newlink before fdb notify and
hence avoiding the user-space race.

Fixes: afbd8bae9c79 ("vxlan: add implicit fdb entry for default destination")
Signed-off-by: Roopa Prabhu <roopa@cumulusnetworks.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agovxlan: make netlink notify in vxlan_fdb_destroy optional
Roopa Prabhu [Wed, 4 Jul 2018 23:46:31 +0000 (16:46 -0700)]
vxlan: make netlink notify in vxlan_fdb_destroy optional

Add a new option do_notify to vxlan_fdb_destroy to make
sending netlink notify optional. Used by a later patch.

Signed-off-by: Roopa Prabhu <roopa@cumulusnetworks.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agovxlan: add new fdb alloc and create helpers
Roopa Prabhu [Wed, 4 Jul 2018 23:46:30 +0000 (16:46 -0700)]
vxlan: add new fdb alloc and create helpers

- Add new vxlan_fdb_alloc helper
- rename existing vxlan_fdb_create into vxlan_fdb_update:
        because it really creates or updates an existing
        fdb entry
- move new fdb creation into a separate vxlan_fdb_create

Main motivation for this change is to introduce the ability
to decouple vxlan fdb creation and notify, used in a later patch.

Signed-off-by: Roopa Prabhu <roopa@cumulusnetworks.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agortnetlink: add rtnl_link_state check in rtnl_configure_link
Roopa Prabhu [Wed, 4 Jul 2018 23:46:29 +0000 (16:46 -0700)]
rtnetlink: add rtnl_link_state check in rtnl_configure_link

rtnl_configure_link sets dev->rtnl_link_state to
RTNL_LINK_INITIALIZED and unconditionally calls
__dev_notify_flags to notify user-space of dev flags.

current call sequence for rtnl_configure_link
rtnetlink_newlink
    rtnl_link_ops->newlink
    rtnl_configure_link (unconditionally notifies userspace of
                         default and new dev flags)

If a newlink handler wants to call rtnl_configure_link
early, we will end up with duplicate notifications to
user-space.

This patch fixes rtnl_configure_link to check rtnl_link_state
and call __dev_notify_flags with gchanges = 0 if already
RTNL_LINK_INITIALIZED.

Later in the series, this patch will help the following sequence
where a driver implementing newlink can call rtnl_configure_link
to initialize the link early.

makes the following call sequence work:
rtnetlink_newlink
    rtnl_link_ops->newlink (vxlan) -> rtnl_configure_link (initializes
                                                link and notifies
                                                user-space of default
                                                dev flags)
    rtnl_configure_link (updates dev flags if requested by user ifm
                         and notifies user-space of new dev flags)

Signed-off-by: Roopa Prabhu <roopa@cumulusnetworks.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agodevlink: fix incorrect return statement
Arnd Bergmann [Fri, 6 Jul 2018 12:58:51 +0000 (14:58 +0200)]
devlink: fix incorrect return statement

A newly added dummy helper function tries to return a failure from a "void"
function:

In file included from include/net/dsa.h:24,
                 from arch/arm/plat-orion/common.c:21:
include/net/devlink.h: In function 'devlink_param_value_changed':
include/net/devlink.h:771:9: error: 'return' with a value, in function returning void [-Werror]
  return -EOPNOTSUPP;

This fixes it by removing the bogus statement.

Fixes: ea601e170988 ("devlink: Add devlink notifications support for params")
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agomlxsw: spectrum_router: avoid uninitialized variable access
Arnd Bergmann [Fri, 6 Jul 2018 12:44:45 +0000 (14:44 +0200)]
mlxsw: spectrum_router: avoid uninitialized variable access

When CONFIG_BRIDGE_VLAN_FILTERING is disabled, gcc correctly points out
that the 'vid' variable is uninitialized whenever br_vlan_get_pvid
returns an error:

drivers/net/ethernet/mellanox/mlxsw/spectrum_router.c: In function 'mlxsw_sp_rif_vlan_fid_get':
drivers/net/ethernet/mellanox/mlxsw/spectrum_router.c:6881:6: error: 'vid' may be used uninitialized in this function [-Werror=maybe-uninitialized]

This changes the condition check to always return -EINVAL here,
which I guess is what the author intended here.

Fixes: e6f1960ae6c7 ("mlxsw: spectrum_router: Allocate FID according to PVID")
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Reviewed-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agoMerge branch 'hns3-misc-fixes'
David S. Miller [Sat, 7 Jul 2018 02:13:07 +0000 (11:13 +0900)]
Merge branch 'hns3-misc-fixes'

Salil Mehta says:

====================
Misc. bug fixes & cleanups for HNS3 driver

This patch-set presents some miscellaneous bug fixes and cleanups
for the HNS3 driver.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: hns3: Prevent sending command during global or core reset
Huazhong Tan [Fri, 6 Jul 2018 10:28:04 +0000 (11:28 +0100)]
net: hns3: Prevent sending command during global or core reset

According to hardware's description, driver should not send command to
IMP while hardware doing global or core reset.

Signed-off-by: Huazhong Tan <tanhuazhong@huawei.com>
Signed-off-by: Peng Li <lipeng321@huawei.com>
Signed-off-by: Salil Mehta <salil.mehta@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: hns3: Remove the warning when clear reset cause
Peng Li [Fri, 6 Jul 2018 10:28:03 +0000 (11:28 +0100)]
net: hns3: Remove the warning when clear reset cause

Only the core/global/IMP reset need clear cause, other type does not
need do it. The warning may be treated as error as it is normal. This
patch removes the warning.

Signed-off-by: Peng Li <lipeng321@huawei.com>
Signed-off-by: Salil Mehta <salil.mehta@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: hns3: Fix get_vector ops in hclgevf_main module
Yunsheng Lin [Fri, 6 Jul 2018 10:28:02 +0000 (11:28 +0100)]
net: hns3: Fix get_vector ops in hclgevf_main module

The hclgevf_free_vector function expects the caller to pass
the vector_id to it, and hclgevf_put_vector pass vector to
it now, which will cause vector allocation problem.

This patch fixes it by converting vector into vector_id before
calling hclgevf_free_vector.

Fixes: e2cb1dec9779 ("net: hns3: Add HNS3 VF HCL(Hardware Compatibility Layer) Support")
Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>
Signed-off-by: Peng Li <lipeng321@huawei.com>
Signed-off-by: Salil Mehta <salil.mehta@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: hns3: Fix warning bug when doing lp selftest
Yunsheng Lin [Fri, 6 Jul 2018 10:28:01 +0000 (11:28 +0100)]
net: hns3: Fix warning bug when doing lp selftest

The napi_alloc_skb is excepted to be called under the
non-preemptible code path when it is called by hns3_clean_rx_ring
during loopback selftest, otherwise the below warning will be
logged:

[   92.420780] BUG: using smp_processor_id() in preemptible
[00000000] code: ethtool/1873
<SNIP>
[   92.463202]  check_preemption_disabled+0xf8/0x100
[   92.467893]  debug_smp_processor_id+0x1c/0x28
[   92.472239]  __napi_alloc_skb+0x30/0x130
[   92.476158]  hns3_clean_rx_ring+0x118/0x5f0 [hns3]
[   92.480941]  hns3_self_test+0x32c/0x4d0 [hns3]
[   92.485375]  ethtool_self_test+0xdc/0x1e8
[   92.489372]  dev_ethtool+0x1020/0x1da8
[   92.493109]  dev_ioctl+0x188/0x3a0
[   92.496499]  sock_do_ioctl+0xf4/0x208
[   92.500148]  sock_ioctl+0x228/0x3e8
[   92.503626]  do_vfs_ioctl+0xc4/0x880
[   92.507189]  SyS_ioctl+0x94/0xa8
[   92.510404]  el0_svc_naked+0x30/0x34

This patch fix it by disabling preemption when calling
hns3_clean_rx_ring during loopback selftest.

Fixes: c39c4d98dc65 ("net: hns3: Add mac loopback selftest support in hns3 driver")
Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>
Signed-off-by: Peng Li <lipeng321@huawei.com>
Signed-off-by: Salil Mehta <salil.mehta@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: hns3: Add configure for mac minimal frame size
Jian Shen [Fri, 6 Jul 2018 10:28:00 +0000 (11:28 +0100)]
net: hns3: Add configure for mac minimal frame size

When change the mtu, the minimal frame size of mac will be set
to zero, it is incorrect. This patch fixes it by set it to the
default value.

Signed-off-by: Jian Shen <shenjian15@huawei.com>
Signed-off-by: Peng Li <lipeng321@huawei.com>
Signed-off-by: Salil Mehta <salil.mehta@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: hns3: Fix for mailbox message truncated problem
Fuyun Liang [Fri, 6 Jul 2018 10:27:59 +0000 (11:27 +0100)]
net: hns3: Fix for mailbox message truncated problem

The payload of mailbox message is 16 byte and the value of
HCLGE_MBX_MAX_ARQ_MSG_SIZE is 8. A message truncated problem will
happen when mailbox message is converted to ARQ message. This patch
replaces HCLGE_MBX_MAX_ARQ_MSG_SIZE with the size of ARQ message in
hclgevf_mbx_handler to fix this problem.

Fixes: b11a0bb231f3 ("net: hns3: Add mailbox support to VF driver")
Signed-off-by: Fuyun Liang <liangfuyun1@huawei.com>
Signed-off-by: Peng Li <lipeng321@huawei.com>
Signed-off-by: Salil Mehta <salil.mehta@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: hns3: Fix for l4 checksum offload bug
Yunsheng Lin [Fri, 6 Jul 2018 10:27:58 +0000 (11:27 +0100)]
net: hns3: Fix for l4 checksum offload bug

Hardware only support tcp/udp/sctp l4 checksum offload, but
the driver currently tell hardware to do l4 checksum offlad when
l3 is IPv4 or IPv6, which may cause checksumm error.

This patch fixes it by only enabling the l4 offload when l4 is
tcp/udp/sctp.

Fixes: 76ad4f0ee747 ("net: hns3: Add support of HNS3 Ethernet Driver for hip08 SoC")
Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>
Signed-off-by: Peng Li <lipeng321@huawei.com>
Signed-off-by: Salil Mehta <salil.mehta@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: hns3: Fix for waterline not setting correctly
Yunsheng Lin [Fri, 6 Jul 2018 10:27:57 +0000 (11:27 +0100)]
net: hns3: Fix for waterline not setting correctly

The HCLGE_RX_PRIV_EN_B is used to tell the firmware whether
to update the specific waterline value, if the is not set,
the firmware will ignore the value.

This patch fixes by setting the HCLGE_RX_PRIV_EN_B even if
the updated value is zero.

Fixes: 46a3df9f9718 ("net: hns3: Add HNS3 Acceleration Engine & Compatibility Layer Support")
Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>
Signed-off-by: Peng Li <lipeng321@huawei.com>
Signed-off-by: Salil Mehta <salil.mehta@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: hns3: Fix for mac pause not disable in pfc mode
Yunsheng Lin [Fri, 6 Jul 2018 10:27:56 +0000 (11:27 +0100)]
net: hns3: Fix for mac pause not disable in pfc mode

When pfc pause mode is enable, the mac pause mode need to be
disabled, otherwise the pfc pause packet will not be sent when
congestion happens.

This patch fixes by disabling the mac pause when pfc pause is
enabled.

Fixes: 848440544b41 ("net: hns3: Add support of TX Scheduler & Shaper to HNS3 driver")
Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>
Signed-off-by: Peng Li <lipeng321@huawei.com>
Signed-off-by: Salil Mehta <salil.mehta@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: hns3: Fix tc setup when netdev is first up
Yunsheng Lin [Fri, 6 Jul 2018 10:27:55 +0000 (11:27 +0100)]
net: hns3: Fix tc setup when netdev is first up

Currently, tc related configuration is not setup when the
netdev is first up, which cause the stack only using tc 0
problem.

This patch fixes it by setting the tc related configuration
using the info from NCL_CONFIG when netdev is first up.

Fixes: 76ad4f0ee747 ("net: hns3: Add support of HNS3 Ethernet Driver for hip08 SoC")
Signed-off-by: Yunsheng Lin <linyunsheng@huawei.com>
Signed-off-by: Peng Li <lipeng321@huawei.com>
Signed-off-by: Salil Mehta <salil.mehta@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agoliquidio: fix kernel panic when NIC firmware is older than 1.7.2
Rick Farrington [Fri, 6 Jul 2018 18:27:07 +0000 (11:27 -0700)]
liquidio: fix kernel panic when NIC firmware is older than 1.7.2

Pre-1.7.2 NIC firmware does not support (and does not respond to) the "get
speed" command which is sent by the 1.7.2 driver (for CN23XX-225 cards
only) during modprobe.  Due to a bug in older firmware (with respect to
unknown commands), this unsupported command causes a cascade of errors that
ends in a kernel panic.

Fix it by making the sending of the "get speed" command conditional on the
firmware version.

Signed-off-by: Rick Farrington <ricardo.farrington@cavium.com>
Signed-off-by: Felix Manlunas <felix.manlunas@cavium.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agoMerge branch 'sock-cookie-initializers'
David S. Miller [Sat, 7 Jul 2018 01:58:50 +0000 (10:58 +0900)]
Merge branch 'sock-cookie-initializers'

Willem de Bruijn says:

====================
sock cookie initializers

Recent UDP GSO and SO_TXTIME features added new fields to cookie
structs.

When adding a field, all sites where a struct is initialized have to
be updated, which is a lot of boilerplate. Alternatively, a field can
be initialized selectively, but this is fragile. I introduced a bug
in udp gso where an uninitialized field was read. See also fix commit
("9887cba19978 ip: limit use of gso_size to udp").

Introduce initializers for structs ipcm(6)_cookie and sockc_cookie.

patch 1..3 do exactly this.
patch 4..5 make ipv4 and ipv6 handle cookies the same way and
           remove some boilerplate in doing so.
patch 6    removes the udp gso branch that needed the above fix
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agoip: unconditionally set cork gso_size
Willem de Bruijn [Fri, 6 Jul 2018 14:12:59 +0000 (10:12 -0400)]
ip: unconditionally set cork gso_size

Now that ipc(6)->gso_size is correctly initialized in all callers of
ip(6)_setup_cork, it is safe to unconditionally pass it to the cork.

Link: http://lkml.kernel.org/r/20180619164752.143249-1-willemdebruijn.kernel@gmail.com
Signed-off-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agoip: remove tx_flags from ipcm_cookie and use same logic for v4 and v6
Willem de Bruijn [Fri, 6 Jul 2018 14:12:58 +0000 (10:12 -0400)]
ip: remove tx_flags from ipcm_cookie and use same logic for v4 and v6

skb_shinfo(skb)->tx_flags is derived from sk->sk_tsflags, possibly
after modification by __sock_cmsg_send, by calling sock_tx_timestamp.

The IPv4 and IPv6 paths do this conversion differently. In IPv4, the
individual protocols that support tx timestamps call this function
and store the result in ipc.tx_flags. In IPv6, sock_tx_timestamp is
called in __ip6_append_data.

There is no need to store both tx_flags and ts_flags in the cookie
as one is derived from the other. Convert when setting up the cork
and remove the redundant field. This is similar to IPv6, only have
the conversion happen only once per datagram, in ip(6)_setup_cork.

Also change __ip6_append_data to match __ip_append_data. Only update
tskey if timestamping is enabled with OPT_ID. The SOCK_.. test is
redundant: only valid protocols can have non-zero cork->tx_flags.

After this change the IPv4 and IPv6 logic is the same.

Signed-off-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agoipv6: fold sockcm_cookie into ipcm6_cookie
Willem de Bruijn [Fri, 6 Jul 2018 14:12:57 +0000 (10:12 -0400)]
ipv6: fold sockcm_cookie into ipcm6_cookie

ipcm_cookie includes sockcm_cookie. Do the same for ipcm6_cookie.

This reduces the number of arguments that need to be passed around,
applies ipcm6_init to all cookie fields at once and reduces code
differentiation between ipv4 and ipv6.

Signed-off-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agosock: sockc cookie initializer
Willem de Bruijn [Fri, 6 Jul 2018 14:12:56 +0000 (10:12 -0400)]
sock: sockc cookie initializer

Initialize the cookie in one location to reduce code duplication and
avoid bugs from inconsistent initialization, such as that fixed in
commit 9887cba19978 ("ip: limit use of gso_size to udp").

Signed-off-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agoipv6: ipcm6_cookie initializer
Willem de Bruijn [Fri, 6 Jul 2018 14:12:55 +0000 (10:12 -0400)]
ipv6: ipcm6_cookie initializer

Initialize the cookie in one location to reduce code duplication and
avoid bugs from inconsistent initialization, such as that fixed in
commit 9887cba19978 ("ip: limit use of gso_size to udp").

Signed-off-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agoipv4: ipcm_cookie initializers
Willem de Bruijn [Fri, 6 Jul 2018 14:12:54 +0000 (10:12 -0400)]
ipv4: ipcm_cookie initializers

Initialize the cookie in one location to reduce code duplication and
avoid bugs from inconsistent initialization, such as that fixed in
commit 9887cba19978 ("ip: limit use of gso_size to udp").

Signed-off-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agoMerge branch 'IP-listification-follow-ups'
David S. Miller [Fri, 6 Jul 2018 02:19:08 +0000 (11:19 +0900)]
Merge branch 'IP-listification-follow-ups'

Edward Cree says:

====================
IP listification follow-ups

While working on IPv6 list processing, I found another bug in the IPv4
 version.  So this patch series has that fix, and the IPv6 version with
 both fixes incorporated.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: ipv6: listify ipv6_rcv() and ip6_rcv_finish()
Edward Cree [Thu, 5 Jul 2018 14:49:42 +0000 (15:49 +0100)]
net: ipv6: listify ipv6_rcv() and ip6_rcv_finish()

Essentially the same as the ipv4 equivalents.

Signed-off-by: Edward Cree <ecree@solarflare.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: ipv4: fix list processing on L3 slave devices
Edward Cree [Thu, 5 Jul 2018 14:47:39 +0000 (15:47 +0100)]
net: ipv4: fix list processing on L3 slave devices

If we have an L3 master device, l3mdev_ip_rcv() will steal the skb, but
 we were returning NET_RX_SUCCESS from ip_rcv_finish_core() which meant
 that ip_list_rcv_finish() would keep it on the list.  Instead let's
 move the l3mdev_ip_rcv() call into the caller, so that our response to
 a steal can be different in the single packet path (return
 NET_RX_SUCCESS) and the list path (forget this packet and continue).

Fixes: 5fa12739a53d ("net: ipv4: listify ip_rcv_finish")
Signed-off-by: Edward Cree <ecree@solarflare.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: core: filter: mark expected switch fall-through
Gustavo A. R. Silva [Wed, 4 Jul 2018 22:34:37 +0000 (17:34 -0500)]
net: core: filter: mark expected switch fall-through

In preparation to enabling -Wimplicit-fallthrough, mark switch cases
where we are expecting to fall through.

Warning level 2 was used: -Wimplicit-fallthrough=2

Signed-off-by: Gustavo A. R. Silva <gustavo@embeddedor.com>
Signed-off-by: David S. Miller <davem@davemloft.net>