openwrt/staging/blogic.git
6 years agoMerge branch 'devlink-port-flavours-and-phys_port_name'
David S. Miller [Sat, 19 May 2018 20:30:39 +0000 (16:30 -0400)]
Merge branch 'devlink-port-flavours-and-phys_port_name'

Jiri Pirko says:

====================
devlink: introduce port flavours and common phys_port_name generation

This patchset resolves 2 issues we have right now:
1) There are many netdevices / ports in the system, for port, pf, vf
   represenatation but the user has no way to see which is which
2) The ndo_get_phys_port_name is implemented in each driver separatelly,
   which may lead to inconsistent names between drivers.

This patchset introduces port flavours which should address the first
problem. In this initial patchset, I focus on DSA and their port
flavours. As a follow-up, I plan to add PF and VF representor flavours.
However, that needs additional dependencies in drivers (nfp, mlx5).

The common phys_port_name generation is used by mlxsw. An example output
for mlxsw looks like this:

...
pci/0000:03:00.0/59: type eth netdev enp3s0np4 flavour physical number 4
pci/0000:03:00.0/61: type eth netdev enp3s0np1 flavour physical number 1
pci/0000:03:00.0/63: type eth netdev enp3s0np2 flavour physical number 2
pci/0000:03:00.0/49: type eth netdev enp3s0np8s0 flavour physical number 8 split_group 8 subport 0
pci/0000:03:00.0/50: type eth netdev enp3s0np8s1 flavour physical number 8 split_group 8 subport 1
pci/0000:03:00.0/51: type eth netdev enp3s0np8s2 flavour physical number 8 split_group 8 subport 2
pci/0000:03:00.0/52: type eth netdev enp3s0np8s3 flavour physical number 8 split_group 8 subport 3

As you can see, the netdev names are generated according to the flavour
and port number. In case the port is split, the split subnumber is also
included.

An example output for dsa_loop testing module looks like this:
mdio_bus/fixed-0:1f/0: type eth netdev lan1 flavour physical number 0
mdio_bus/fixed-0:1f/1: type eth netdev lan2 flavour physical number 1
mdio_bus/fixed-0:1f/2: type eth netdev lan3 flavour physical number 2
mdio_bus/fixed-0:1f/3: type eth netdev lan4 flavour physical number 3
mdio_bus/fixed-0:1f/4: type notset
mdio_bus/fixed-0:1f/5: type notset flavour cpu number 5
mdio_bus/fixed-0:1f/6: type notset
mdio_bus/fixed-0:1f/7: type notset
mdio_bus/fixed-0:1f/8: type notset
mdio_bus/fixed-0:1f/9: type notset
mdio_bus/fixed-0:1f/10: type notset
mdio_bus/fixed-0:1f/11: type notset

---
RFC->v1:
-removed nfp patches, removed DSA patch that used name generation helper
-patch 1:
 - Reduced the nfp change just to simply use newly created attr_set func
-patch 2:
 - rebased
 - removed pf/vf reps flavours
-patch 3:
 - rebased
-patch 4:
 - added missing break pointed out by Andrew
====================

Reviewed-by: Florian Fainelli <f.fainelli@gmail.com>
Tested-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agomlxsw: use devlink helper to generate physical port name
Jiri Pirko [Fri, 18 May 2018 07:29:04 +0000 (09:29 +0200)]
mlxsw: use devlink helper to generate physical port name

Since devlink knows the info needed to generate the physical port name
in a generic way for all devlink users, use the helper to do the job.

Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agodsa: set devlink port attrs for dsa ports
Jiri Pirko [Fri, 18 May 2018 07:29:03 +0000 (09:29 +0200)]
dsa: set devlink port attrs for dsa ports

Set the attrs and allow to expose port flavour to user via devlink.

Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agodevlink: introduce a helper to generate physical port names
Jiri Pirko [Fri, 18 May 2018 07:29:02 +0000 (09:29 +0200)]
devlink: introduce a helper to generate physical port names

Each driver implements physical port name generation by itself. However
as devlink has all needed info, it can easily do the job for all its
users. So implement this helper in devlink.

Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agodevlink: extend attrs_set for setting port flavours
Jiri Pirko [Fri, 18 May 2018 07:29:01 +0000 (09:29 +0200)]
devlink: extend attrs_set for setting port flavours

Devlink ports can have specific flavour according to the purpose of use.
This patch extend attrs_set so the driver can say which flavour port
has. Initial flavours are:
physical, cpu, dsa
User can query this to see right away what is the purpose of each port.

Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agodevlink: introduce devlink_port_attrs_set
Jiri Pirko [Fri, 18 May 2018 07:29:00 +0000 (09:29 +0200)]
devlink: introduce devlink_port_attrs_set

Change existing setter for split port information into more generic
attrs setter. Alongside with that, allow to set port number and subport
number for split ports.

Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: stmmac: Populate missing callbacks in HWIF initialization
Jose Abreu [Fri, 18 May 2018 15:54:38 +0000 (16:54 +0100)]
net: stmmac: Populate missing callbacks in HWIF initialization

Some HW specific setups, like sun8i, do not populate all the necessary
callbacks, which is what HWIF helpers were expecting.

Fix this by always trying to get the generic helpers and populate them
if they were not previously populated by HW specific setup.

Signed-off-by: Jose Abreu <joabreu@synopsys.com>
Fixes: 5f0456b43140 ("net: stmmac: Implement logic to automatically
select HW Interface")
Reported-by: Corentin Labbe <clabbe.montjoie@gmail.com>
Tested-by: Corentin Labbe <clabbe.montjoie@gmail.com>
Cc: Corentin Labbe <clabbe.montjoie@gmail.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: Joao Pinto <jpinto@synopsys.com>
Cc: Giuseppe Cavallaro <peppe.cavallaro@st.com>
Cc: Alexandre Torgue <alexandre.torgue@st.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agocxgb4: collect SGE PF/VF queue map
Rahul Lakkireddy [Fri, 18 May 2018 13:42:53 +0000 (19:12 +0530)]
cxgb4: collect SGE PF/VF queue map

For T6, collect info on queue mapping to corresponding PF/VF in SGE.

Signed-off-by: Rahul Lakkireddy <rahul.lakkireddy@chelsio.com>
Signed-off-by: Ganesh Goudar <ganeshgr@chelsio.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: mvpp2: typo and cosmetic fixes
Antoine Tenart [Fri, 18 May 2018 12:34:51 +0000 (14:34 +0200)]
net: mvpp2: typo and cosmetic fixes

This patch on the Marvell PPv2 driver is only cosmetic. Two typos are
removed as well as other cosmetic fixes, such as extra new lines or tabs
vs spaces.

Suggested-by: Stefan Chulski <stefanc@marvell.com>
Signed-off-by: Antoine Tenart <antoine.tenart@bootlin.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agohippi: fix spelling mistake: "Framming" -> "Framing"
Colin Ian King [Fri, 18 May 2018 10:09:22 +0000 (11:09 +0100)]
hippi: fix spelling mistake: "Framming" -> "Framing"

Trivial fix to spelling mistake in printk message text

Signed-off-by: Colin Ian King <colin.king@canonical.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agotcp: tcp_rack_reo_wnd() can be static
kbuild test robot [Fri, 18 May 2018 05:14:23 +0000 (13:14 +0800)]
tcp: tcp_rack_reo_wnd() can be static

Fixes: 20b654dfe1be ("tcp: support DUPACK threshold in RACK")
Signed-off-by: kbuild test robot <fengguang.wu@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agoMerge branch 'net-smc-cleanups'
David S. Miller [Fri, 18 May 2018 17:15:02 +0000 (13:15 -0400)]
Merge branch 'net-smc-cleanups'

Ursula Braun says:

====================
net/smc: cleanups 2018-05-18

here are SMC patches for net-next providing restructuring and cleanup
in different areas.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet/smc: restructure client and server code in af_smc
Hans Wippel [Fri, 18 May 2018 07:34:18 +0000 (09:34 +0200)]
net/smc: restructure client and server code in af_smc

This patch splits up the functions smc_connect_rdma and smc_listen_work
into smaller functions.

Signed-off-by: Hans Wippel <hwippel@linux.ibm.com>
Signed-off-by: Ursula Braun <ubraun@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet/smc: change smc_buf_free function parameters
Hans Wippel [Fri, 18 May 2018 07:34:17 +0000 (09:34 +0200)]
net/smc: change smc_buf_free function parameters

This patch changes the function smc_buf_free to use the SMC link group
instead of the link as function parameter. Also, it changes the order of
the other two parameters.

Signed-off-by: Hans Wippel <hwippel@linux.ibm.com>
Signed-off-by: Ursula Braun <ubraun@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet/smc: do a few smc_core.c cleanups
Hans Wippel [Fri, 18 May 2018 07:34:16 +0000 (09:34 +0200)]
net/smc: do a few smc_core.c cleanups

This patch consists of Christmas tree fixes and removal of an unneeded
function parameter.

Signed-off-by: Hans Wippel <hwippel@linux.ibm.com>
Signed-off-by: Ursula Braun <ubraun@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet/smc: restructure CDC message reception
Hans Wippel [Fri, 18 May 2018 07:34:15 +0000 (09:34 +0200)]
net/smc: restructure CDC message reception

This patch moves a CDC sanity check from smc_cdc_msg_recv_action() to
the other sanity checks in smc_cdc_rx_handler(). While doing this, it
simplifies smc_cdc_msg_recv() and removes unneeded function parameters.

Signed-off-by: Hans Wippel <hwippel@linux.ibm.com>
Signed-off-by: Ursula Braun <ubraun@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet/smc: move smc_core specific code from smc.h to smc_core
Hans Wippel [Fri, 18 May 2018 07:34:14 +0000 (09:34 +0200)]
net/smc: move smc_core specific code from smc.h to smc_core

SMC connection and buffer handling belong to smc_core. So, this patch
moves this code from smc.h to smc_core.

Signed-off-by: Hans Wippel <hwippel@linux.ibm.com>
Signed-off-by: Ursula Braun <ubraun@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet/smc: calculate write offset in RMB only once per connection
Hans Wippel [Fri, 18 May 2018 07:34:13 +0000 (09:34 +0200)]
net/smc: calculate write offset in RMB only once per connection

Currently, the write offset within the RMB is calculated on each write
operation although it is fixed for each connection. With this patch, the
offset is calculated once and stored in a connection specific variable.

Signed-off-by: Hans Wippel <hwippel@linux.ibm.com>
Signed-off-by: Ursula Braun <ubraun@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet/smc: rename connection index to RMBE index
Hans Wippel [Fri, 18 May 2018 07:34:12 +0000 (09:34 +0200)]
net/smc: rename connection index to RMBE index

The connection index is actually a RMBE index. So, this patch changes
the name accordingly.

Signed-off-by: Hans Wippel <hwippel@linux.ibm.com>
Signed-off-by: Ursula Braun <ubraun@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet/smc: move link group list to smc_core
Hans Wippel [Fri, 18 May 2018 07:34:11 +0000 (09:34 +0200)]
net/smc: move link group list to smc_core

This patch moves the global link group list to smc_core where the link
group functions are. To make this work, it moves code in af_smc and
smc_ib that operates on the link group list to smc_core as well.

While at it, the link group counter is integrated into the list
structure and initialized to zero.

Signed-off-by: Hans Wippel <hwippel@linux.ibm.com>
Signed-off-by: Ursula Braun <ubraun@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet/smc: add common buffer size in send and receive buffer descriptors
Hans Wippel [Fri, 18 May 2018 07:34:10 +0000 (09:34 +0200)]
net/smc: add common buffer size in send and receive buffer descriptors

In addition to the buffer references, SMC currently stores the sizes of
the receive and send buffers in each connection as separate variables.
This patch introduces a buffer length variable in the common buffer
descriptor and uses this length instead.

Signed-off-by: Hans Wippel <hwippel@linux.ibm.com>
Signed-off-by: Ursula Braun <ubraun@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agoMerge tag 'mlx5e-updates-2018-05-17' of git://git.kernel.org/pub/scm/linux/kernel...
David S. Miller [Fri, 18 May 2018 17:00:43 +0000 (13:00 -0400)]
Merge tag 'mlx5e-updates-2018-05-17' of git://git./linux/kernel/git/saeed/linux

Saeed Mahameed says:

====================
mlx5e-updates-2018-05-17

From: Or Gerlitz <ogerlitz@mellanox.com>

This series addresses a regression introduced by the
shared block TC changes [1]. Currently, for VF->VF and uplink->VF rules, the
TC core (cls_api) attempts to offload the same flow multiple times into
the driver, as a side effect of the mlx5 registration to the egdev callback.

We use the flow cookie to ignore attempts to add such flows, we can't
reject them (return error), b/c this will fail the offload attempt, so we
ignore that.

The last patch of the series deals with exposing HW stats counters through
ethtool for the vport reps.

Dave - the regression that we are addressing was introduced in 4.15 [1] and applies
to nfp and mlx5. Jiri suggested to push driver side fixes to net-next, this is
already done for nfp [2][3]. Once this is upstream, we will submit a small/point
single patch fix for the TC core code which can serve for net and stable, but not
carried into net-next, b/c it might limit some future use-cases.

[1] 208c0f4b5237 "net: sched: use tc_setup_cb_call to call per-block callbacks"
[2] c50647d "nfp: flower: ignore duplicate cb requests for same rule"
[3] 54a4a03 "nfp: flower: support offloading multiple rules with same cookie"
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agoMerge tag 'mlx5-updates-2018-05-17' of git://git.kernel.org/pub/scm/linux/kernel...
David S. Miller [Fri, 18 May 2018 17:00:08 +0000 (13:00 -0400)]
Merge tag 'mlx5-updates-2018-05-17' of git://git./linux/kernel/git/mellanox/linux

Saeed Mahameed says:

====================
mlx5-updates-2018-05-17

mlx5 core dirver updates for both net-next and rdma-next branches.

From Christophe JAILLET, first three patche to use kvfree where needed.

From: Or Gerlitz <ogerlitz@mellanox.com>

Next six patches from Roi and Co adds support for merged
sriov e-switch which comes to serve cases where both PFs, VFs set
on them and both uplinks are to be used in single v-switch SW model.
When merged e-switch is supported, the per-port e-switch is logically
merged into one e-switch that spans both physical ports and all the VFs.

This model allows to offload TC eswitch rules between VFs belonging
to different PFs (and hence have different eswitch affinity), it also
sets the some of the foundations needed for uplink LAG support.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agoMerge branch 'tcp-implement-SACK-compression'
David S. Miller [Fri, 18 May 2018 15:40:28 +0000 (11:40 -0400)]
Merge branch 'tcp-implement-SACK-compression'

Eric Dumazet says:

====================
tcp: implement SACK compression

When TCP receives an out-of-order packet, it immediately sends
a SACK packet, generating network load but also forcing the
receiver to send 1-MSS pathological packets, increasing its
RTX queue length/depth, and thus processing time.

Wifi networks suffer from this aggressive behavior, but generally
speaking, all these SACK packets add fuel to the fire when networks
are under congestion.

This patch series adds SACK compression, but the infrastructure
could be leveraged to also compress ACK in the future.

v2: Addressed Neal feedback.
    Added two sysctls to allow fine tuning, or even disabling the feature.

v3: take rtt = min(srtt, rcv_rtt) as Yuchung suggested, because rcv_rtt
    can be over estimated for RPC (or sender limited)
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agotcp: add tcp_comp_sack_nr sysctl
Eric Dumazet [Thu, 17 May 2018 21:47:29 +0000 (14:47 -0700)]
tcp: add tcp_comp_sack_nr sysctl

This per netns sysctl allows for TCP SACK compression fine-tuning.

This limits number of SACK that can be compressed.
Using 0 disables SACK compression.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agotcp: add tcp_comp_sack_delay_ns sysctl
Eric Dumazet [Thu, 17 May 2018 21:47:28 +0000 (14:47 -0700)]
tcp: add tcp_comp_sack_delay_ns sysctl

This per netns sysctl allows for TCP SACK compression fine-tuning.

Its default value is 1,000,000, or 1 ms to meet TSO autosizing period.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agotcp: add TCPAckCompressed SNMP counter
Eric Dumazet [Thu, 17 May 2018 21:47:27 +0000 (14:47 -0700)]
tcp: add TCPAckCompressed SNMP counter

This counter tracks number of ACK packets that the host has not sent,
thanks to ACK compression.

Sample output :

$ nstat -n;sleep 1;nstat|egrep "IpInReceives|IpOutRequests|TcpInSegs|TcpOutSegs|TcpExtTCPAckCompressed"
IpInReceives                    123250             0.0
IpOutRequests                   3684               0.0
TcpInSegs                       123251             0.0
TcpOutSegs                      3684               0.0
TcpExtTCPAckCompressed          119252             0.0

Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agotcp: add SACK compression
Eric Dumazet [Thu, 17 May 2018 21:47:26 +0000 (14:47 -0700)]
tcp: add SACK compression

When TCP receives an out-of-order packet, it immediately sends
a SACK packet, generating network load but also forcing the
receiver to send 1-MSS pathological packets, increasing its
RTX queue length/depth, and thus processing time.

Wifi networks suffer from this aggressive behavior, but generally
speaking, all these SACK packets add fuel to the fire when networks
are under congestion.

This patch adds a high resolution timer and tp->compressed_ack counter.

Instead of sending a SACK, we program this timer with a small delay,
based on RTT and capped to 1 ms :

delay = min ( 5 % of RTT, 1 ms)

If subsequent SACKs need to be sent while the timer has not yet
expired, we simply increment tp->compressed_ack.

When timer expires, a SACK is sent with the latest information.
Whenever an ACK is sent (if data is sent, or if in-order
data is received) timer is canceled.

Note that tcp_sack_new_ofo_skb() is able to force a SACK to be sent
if the sack blocks need to be shuffled, even if the timer has not
expired.

A new SNMP counter is added in the following patch.

Two other patches add sysctls to allow changing the 1,000,000 and 44
values that this commit hard-coded.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Yuchung Cheng <ycheng@google.com>
Acked-by: Toke Høiland-Jørgensen <toke@toke.dk>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agotcp: do not force quickack when receiving out-of-order packets
Eric Dumazet [Thu, 17 May 2018 21:47:25 +0000 (14:47 -0700)]
tcp: do not force quickack when receiving out-of-order packets

As explained in commit 9f9843a751d0 ("tcp: properly handle stretch
acks in slow start"), TCP stacks have to consider how many packets
are acknowledged in one single ACK, because of GRO, but also
because of ACK compression or losses.

We plan to add SACK compression in the following patch, we
must therefore not call tcp_enter_quickack_mode()

Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Acked-by: Soheil Hassas Yeganeh <soheil@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agotcp: use __sock_put() instead of sock_put() in tcp_clear_xmit_timers()
Eric Dumazet [Thu, 17 May 2018 21:47:24 +0000 (14:47 -0700)]
tcp: use __sock_put() instead of sock_put() in tcp_clear_xmit_timers()

Socket can not disappear under us.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: mscc: Add SPDX identifier
Alexandre Belloni [Thu, 17 May 2018 19:23:05 +0000 (21:23 +0200)]
net: mscc: Add SPDX identifier

ocelot_qsys.h is missing the SPDX identfier, fix that.

Signed-off-by: Alexandre Belloni <alexandre.belloni@bootlin.com>
Reviewed-by: Allan W. Nielsen <allan.nielsen@microsemi.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agoMerge branch 'stmmac-Clean-up-and-tune-up'
David S. Miller [Fri, 18 May 2018 15:00:16 +0000 (11:00 -0400)]
Merge branch 'stmmac-Clean-up-and-tune-up'

Jose Abreu says:

====================
net: stmmac: Clean-up and tune-up

This targets to uniformize the handling of the different GMAC versions in
stmmac_main.c file and also tune-up the HW.

Currently there are some if/else conditions in the main source file which
calls different callbacks depending on the ID of GMAC.

With the introducion of a generic HW interface handling which automatically
selects the GMAC callbacks to be used, it is now unpleasant to see if
conditions in the main code because this should be completely agnostic of the
GMAC version.

This series removes most of these conditions. There are some if conditions
that remain untouched but the callbacks handling are now uniformized.

Tested in GMAC5, hope I didn't break any previous versions.

Please check [1] for performance analisys of patches 3-12.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: stmmac: Remove if condition by taking advantage of hwif return code
Jose Abreu [Fri, 18 May 2018 13:56:09 +0000 (14:56 +0100)]
net: stmmac: Remove if condition by taking advantage of hwif return code

We can remove the if condition and check if return code is different
than -EINVAL, meaning callback is present.

Signed-off-by: Jose Abreu <joabreu@synopsys.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: Joao Pinto <jpinto@synopsys.com>
Cc: Vitor Soares <soares@synopsys.com>
Cc: Giuseppe Cavallaro <peppe.cavallaro@st.com>
Cc: Alexandre Torgue <alexandre.torgue@st.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: stmmac: Let descriptor code get skbuff address
Jose Abreu [Fri, 18 May 2018 13:56:08 +0000 (14:56 +0100)]
net: stmmac: Let descriptor code get skbuff address

Stop using if conditions depending on the GMAC version for getting the
descriptor skbuff address and use instead a helper implemented in the
descriptor files.

Signed-off-by: Jose Abreu <joabreu@synopsys.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: Joao Pinto <jpinto@synopsys.com>
Cc: Vitor Soares <soares@synopsys.com>
Cc: Giuseppe Cavallaro <peppe.cavallaro@st.com>
Cc: Alexandre Torgue <alexandre.torgue@st.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: stmmac: Uniformize set_rx_owner()
Jose Abreu [Fri, 18 May 2018 13:56:07 +0000 (14:56 +0100)]
net: stmmac: Uniformize set_rx_owner()

Currently an if condition is used to select the correct callback to set
rx_onwer in descriptor. Lets keep this simple and always use the same
callback.

Signed-off-by: Jose Abreu <joabreu@synopsys.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: Joao Pinto <jpinto@synopsys.com>
Cc: Vitor Soares <soares@synopsys.com>
Cc: Giuseppe Cavallaro <peppe.cavallaro@st.com>
Cc: Alexandre Torgue <alexandre.torgue@st.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: stmmac: Remove uneeded check for GMAC version in stmmac_xmit
Jose Abreu [Fri, 18 May 2018 13:56:06 +0000 (14:56 +0100)]
net: stmmac: Remove uneeded check for GMAC version in stmmac_xmit

We either have .enable_dma_transmission or .set_tx_tail_ptr in the HW
table callbacks, we can never have both so there is no need to check for
GMAC version.

Signed-off-by: Jose Abreu <joabreu@synopsys.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: Joao Pinto <jpinto@synopsys.com>
Cc: Vitor Soares <soares@synopsys.com>
Cc: Giuseppe Cavallaro <peppe.cavallaro@st.com>
Cc: Alexandre Torgue <alexandre.torgue@st.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: stmmac: Uniformize the use of dma_init_* callbacks
Jose Abreu [Fri, 18 May 2018 13:56:05 +0000 (14:56 +0100)]
net: stmmac: Uniformize the use of dma_init_* callbacks

Instead of relying on the GMAC version for choosing if we need to use
dma_init or dma_init_{rx/tx}_chan callback, lets uniformize this and
always use the dma_init_{rx/tx}_chan callbacks.

While at it, fix the use of dma_init_chan callback, which shall be
called for as many channels as the max of rx/tx channels.

Signed-off-by: Jose Abreu <joabreu@synopsys.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: Joao Pinto <jpinto@synopsys.com>
Cc: Vitor Soares <soares@synopsys.com>
Cc: Giuseppe Cavallaro <peppe.cavallaro@st.com>
Cc: Alexandre Torgue <alexandre.torgue@st.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: stmmac: Move PTP and MMC base address calculation to hwif.c
Jose Abreu [Fri, 18 May 2018 13:56:04 +0000 (14:56 +0100)]
net: stmmac: Move PTP and MMC base address calculation to hwif.c

PTP and MMC modules base address can depend on the GMAC version. As this
is HW specific lets move this base address calculation to hwif.c. Also,
add an entry in the HW table so that we can specify the module offset.
This can later be extended to more modules, if deemed necessary.

Signed-off-by: Jose Abreu <joabreu@synopsys.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: Joao Pinto <jpinto@synopsys.com>
Cc: Vitor Soares <soares@synopsys.com>
Cc: Giuseppe Cavallaro <peppe.cavallaro@st.com>
Cc: Alexandre Torgue <alexandre.torgue@st.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: stmmac: Remove uneeded checks for GMAC version
Jose Abreu [Fri, 18 May 2018 13:56:03 +0000 (14:56 +0100)]
net: stmmac: Remove uneeded checks for GMAC version

With the introducion of callbacks check in hwif.h we only call the
callback if HW supports it so there is no longer need to check for GMAC
version.

Signed-off-by: Jose Abreu <joabreu@synopsys.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: Joao Pinto <jpinto@synopsys.com>
Cc: Vitor Soares <soares@synopsys.com>
Cc: Giuseppe Cavallaro <peppe.cavallaro@st.com>
Cc: Alexandre Torgue <alexandre.torgue@st.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: stmmac: Uniformize the use of dma_{rx/tx}_mode callbacks
Jose Abreu [Fri, 18 May 2018 13:56:02 +0000 (14:56 +0100)]
net: stmmac: Uniformize the use of dma_{rx/tx}_mode callbacks

Instead of relying on the GMAC version for choosing if we need to use
dma_{rx/tx}_mode or just dma_mode callback lets uniformize this and
always use the dma_{rx/tx}_mode callbacks.

Signed-off-by: Jose Abreu <joabreu@synopsys.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: Joao Pinto <jpinto@synopsys.com>
Cc: Vitor Soares <soares@synopsys.com>
Cc: Giuseppe Cavallaro <peppe.cavallaro@st.com>
Cc: Alexandre Torgue <alexandre.torgue@st.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: stmmac: Let descriptor code clear the descriptor
Jose Abreu [Fri, 18 May 2018 13:56:01 +0000 (14:56 +0100)]
net: stmmac: Let descriptor code clear the descriptor

Stop using if conditions depending on the GMAC version for clearing the
descriptor and use instead a helper implemented in the descriptor files.

Signed-off-by: Jose Abreu <joabreu@synopsys.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: Joao Pinto <jpinto@synopsys.com>
Cc: Vitor Soares <soares@synopsys.com>
Cc: Giuseppe Cavallaro <peppe.cavallaro@st.com>
Cc: Alexandre Torgue <alexandre.torgue@st.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: stmmac: Let descriptor code set skbuff address
Jose Abreu [Fri, 18 May 2018 13:56:00 +0000 (14:56 +0100)]
net: stmmac: Let descriptor code set skbuff address

Stop using if conditions depending on the GMAC version for setting the
the descriptor skbuff address and use instead a helper implemented in
the descriptor files.

Signed-off-by: Jose Abreu <joabreu@synopsys.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: Joao Pinto <jpinto@synopsys.com>
Cc: Vitor Soares <soares@synopsys.com>
Cc: Giuseppe Cavallaro <peppe.cavallaro@st.com>
Cc: Alexandre Torgue <alexandre.torgue@st.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: stmmac: Do not keep rearming the coalesce timer in stmmac_xmit
Jose Abreu [Fri, 18 May 2018 13:55:59 +0000 (14:55 +0100)]
net: stmmac: Do not keep rearming the coalesce timer in stmmac_xmit

This is cutting down performance. Once the timer is armed it should run
after the time expires for the first packet sent and not the last one.

After this change, running iperf, the performance gain is +/- 24%.

Signed-off-by: Jose Abreu <joabreu@synopsys.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: Joao Pinto <jpinto@synopsys.com>
Cc: Vitor Soares <soares@synopsys.com>
Cc: Giuseppe Cavallaro <peppe.cavallaro@st.com>
Cc: Alexandre Torgue <alexandre.torgue@st.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: stmmac: Enable OSP for GMAC4
Jose Abreu [Fri, 18 May 2018 13:55:58 +0000 (14:55 +0100)]
net: stmmac: Enable OSP for GMAC4

This enables OSP (Operate on Second Packet) for GMAC4. The feature
allows DMA to fetch second descriptor while its still processing the
first one.

Running iperf, the performance gain is +/- 38%.

Signed-off-by: Jose Abreu <joabreu@synopsys.com>
Cc: David S. Miller <davem@davemloft.net>
Cc: Joao Pinto <jpinto@synopsys.com>
Cc: Vitor Soares <soares@synopsys.com>
Cc: Giuseppe Cavallaro <peppe.cavallaro@st.com>
Cc: Alexandre Torgue <alexandre.torgue@st.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet/mlx5e: Add HW vport counters to representor ethtool stats
Or Gerlitz [Mon, 30 Apr 2018 08:42:26 +0000 (11:42 +0300)]
net/mlx5e: Add HW vport counters to representor ethtool stats

Currently the representor only report the SW (slow-path) traffic
counters.

Add packet/bytes reporting of the HW counters, which account for the
total amount of traffic that was handled by the vport, both slow and
fast (offloaded) paths. The newly exposed counters are named
vport_rx/tx_packets/bytes.

Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Adi Nissim <adin@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
6 years agonet/mlx5e: Ignore attempts to offload multiple times a TC flow
Or Gerlitz [Tue, 10 Apr 2018 16:24:51 +0000 (19:24 +0300)]
net/mlx5e: Ignore attempts to offload multiple times a TC flow

For VF->VF and uplink->VF rules, the TC core (cls_api) attempts
to offload the same flow multiple times into the driver, b/c we
registered to the egdev callback.

Use the flow cookie to ignore attempts to add such flows, we can't
reject them (return error), b/c this will fail the offload attempt,
so we ignore that. We indentify wrong stat/del calls using the flow
ingress/egress flags, here we do return error to the core.

Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Reviewed-by: Paul Blakey <paulb@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
6 years agonet/mlx5e: Use shared table for offloaded TC eswitch flows
Or Gerlitz [Tue, 10 Apr 2018 15:34:36 +0000 (18:34 +0300)]
net/mlx5e: Use shared table for offloaded TC eswitch flows

Currently, each representor netdev use their own hash table to keep
the mapping from TC flow (f->cookie) to the driver offloaded instance.
The table is the one which originally was added for offloading TC NIC
(not eswitch) rules.

This scheme breaks when the core TC code calls us to add the same flow
twice, (e.g under egdev use case) since we don't spot that and offload
a 2nd flow into the HW with the wrong source vport.

As a pre-step to solve that, we move to use a single table which keeps
all offloaded TC eswitch flows. The table is located at the eswitch
uplink representor object.

Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Reviewed-by: Paul Blakey <paulb@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
6 years agonet/mlx5e: Prepare for shared table to keep TC eswitch flows
Or Gerlitz [Tue, 10 Apr 2018 11:27:43 +0000 (14:27 +0300)]
net/mlx5e: Prepare for shared table to keep TC eswitch flows

This is a refactoring step to be able and store the hash table which
keeps track of offloaded TC flows in a different location for NIC
vs e-switch rules.

Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Reviewed-by: Paul Blakey <paulb@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
6 years agonet/mlx5e: Add ingress/egress indication for offloaded TC flows
Or Gerlitz [Wed, 18 Apr 2018 10:45:11 +0000 (13:45 +0300)]
net/mlx5e: Add ingress/egress indication for offloaded TC flows

When an e-switch TC rule is offloaded through the egdev (egress
device) mechanism, we treat this as egress, all other cases (NIC
and e-switch) are considred ingress.

This is preparation step that will allow us to  identify "wrong"
stat/del offload calls made by the TC core on egdev based flows and
ignore them.

Signed-off-by: Or Gerlitz <ogerlitz@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Reviewed-by: Paul Blakey <paulb@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
6 years agonet/mlx5e: Offload TC eswitch rules for VFs belonging to different PFs
Rabie Loulou [Thu, 11 Jan 2018 08:34:32 +0000 (10:34 +0200)]
net/mlx5e: Offload TC eswitch rules for VFs belonging to different PFs

When the merged eswitch capability is supported, allow offloading rules
between VFs which belong to different PFs (and hence have different
eswitch affinity).

Signed-off-by: Rabie Loulou <rabiel@mellanox.com>
Reviewed-by: Or Gerlitz <ogerlitz@mellanox.com>
Reviewed-by: Roi Dayan <roid@mellanox.com>
Reviewed-by: Shahar Klein <shahark@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
6 years agoMerge tag 'mlx5-updates-2018-05-17' of git://git.kernel.org/pub/scm/linux/kernel...
Saeed Mahameed [Fri, 18 May 2018 00:47:09 +0000 (17:47 -0700)]
Merge tag 'mlx5-updates-2018-05-17' of git://git./linux/kernel/git/mellanox/linux

mlx5-updates-2018-05-17

mlx5 core dirver updates for both net-next and rdma-next branches.

From Christophe JAILLET, first three patches to use kvfree where needed.

From: Or Gerlitz <ogerlitz@mellanox.com>

Next six patches from Roi and Co adds support for merged
sriov e-switch which comes to serve cases where both PFs, VFs set
on them and both uplinks are to be used in single v-switch SW model.
When merged e-switch is supported, the per-port e-switch is logically
merged into one e-switch that spans both physical ports and all the VFs.

This model allows to offload TC eswitch rules between VFs belonging
to different PFs (and hence have different eswitch affinity), it also
sets the some of the foundations needed for uplink LAG support.

Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
6 years agonet/mlx5e: Explicitly set source e-switch in offloaded TC rules
Shahar Klein [Sun, 18 Mar 2018 07:03:49 +0000 (09:03 +0200)]
net/mlx5e: Explicitly set source e-switch in offloaded TC rules

Set a specific source e-switch when setting a rule that matches on the
ingress port.

Signed-off-by: Shahar Klein <shahark@mellanox.com>
Reviewed-by: Or Gerlitz <ogerlitz@mellanox.com>
Reviewed-by: Roi Dayan <roid@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
6 years agonet/mlx5: Add source e-switch owner
Shahar Klein [Sun, 18 Mar 2018 07:02:06 +0000 (09:02 +0200)]
net/mlx5: Add source e-switch owner

The source e-switch owner allows a vport on one e-switch port be associated
with a rule defined on the second port e-switch.

The role of the source eswitch owner valid bit in the flow group is to
allow the firmware fail driver attempts to wild card the source eswitch
match field. If this bit is not set, the firmware ignores the source
eswitch owner field totally.

Signed-off-by: Shahar Klein <shahark@mellanox.com>
Reviewed-by: Or Gerlitz <ogerlitz@mellanox.com>
Reviewed-by: Roi Dayan <roid@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
6 years agonet/mlx5e: Explicitly set destination e-switch in FDB rules
Rabie Loulou [Sun, 18 Mar 2018 06:29:04 +0000 (08:29 +0200)]
net/mlx5e: Explicitly set destination e-switch in FDB rules

Set a specific destination e-switch when setting a destination vport.

Signed-off-by: Rabie Loulou <rabiel@mellanox.com>
Reviewed-by: Or Gerlitz <ogerlitz@mellanox.com>
Reviewed-by: Roi Dayan <roid@mellanox.com>
Reviewed-by: Shahar Klein <shahark@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
6 years agonet/mlx5: Add destination e-switch owner
Shahar Klein [Thu, 22 Mar 2018 10:32:12 +0000 (12:32 +0200)]
net/mlx5: Add destination e-switch owner

The destination e-switch owner allows a rule in namespace of one e-switch
owner to point to a vport that is natively associated with another
e-switch owner.

Signed-off-by: Shahar Klein <shahark@mellanox.com>
Reviewed-by: Or Gerlitz <ogerlitz@mellanox.com>
Reviewed-by: Roi Dayan <roid@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
6 years agonet/mlx5: Properly handle a vport destination when setting FTE
Shahar Klein [Thu, 22 Mar 2018 09:56:51 +0000 (11:56 +0200)]
net/mlx5: Properly handle a vport destination when setting FTE

When creating FTE, properly distinguish between destination being vport
or tir. The previous code just worked accidentally b/c of both dest being
in the same offset within a union.

Signed-off-by: Shahar Klein <shahark@mellanox.com>
Reviewed-by: Or Gerlitz <ogerlitz@mellanox.com>
Reviewed-by: Roi Dayan <roid@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
6 years agonet/mlx5: Add merged e-switch cap
Roi Dayan [Tue, 5 Dec 2017 08:38:58 +0000 (10:38 +0200)]
net/mlx5: Add merged e-switch cap

When merged e-switch is supported, the per-port e-switch is logically
merged into one e-switch that spans both physical ports and all the VFs.
Under merged eswitch, both the matching on source vport and setting
destination vport can have a 2nd attribute which is the vhca id of the
eswitch owner.

For example:
esw0: {match: <src vport=1 owner=0> action: fwd to <dst vport=7, owner=1>}
is a flow set on eswitch0 matching on source vport=1 from his eswitch
and the action being fwd to dest vport=7 of eswitch1.

Signed-off-by: Roi Dayan <roid@mellanox.com>
Reviewed-by: Shahar Klein <shahark@mellanox.com>
Reviewed-by: Or Gerlitz Klein <ogerlitz@mellanox.com>
Signed-off-by: Saeed Mahameed <saeedm@mellanox.com>
6 years agoMerge branch 'net-Allow-more-drivers-with-COMPILE_TEST'
David S. Miller [Thu, 17 May 2018 21:11:07 +0000 (17:11 -0400)]
Merge branch 'net-Allow-more-drivers-with-COMPILE_TEST'

Florian Fainelli says:

====================
net: Allow more drivers with COMPILE_TEST

This patch series includes more drivers to be build tested with COMPILE_TEST
enabled. This helps cover some of the issues I just ran into with missing
a driver *sigh*.

Chanves in v3:

- drop the TI Keystone NETCP driver from the COMPILE_TEST additions

Changes in v2:

- allow FEC to build outside of CONFIG_ARM/ARM64 by defining a layout of
  registers, this is not meant to run, so this is not a real issue if we
  are not matching the correct register layout
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: phy: Allow MDIO_MOXART and MDIO_SUN4I with COMPILE_TEST
Florian Fainelli [Thu, 17 May 2018 20:07:45 +0000 (13:07 -0700)]
net: phy: Allow MDIO_MOXART and MDIO_SUN4I with COMPILE_TEST

Those drivers build just fine with COMPILE_TEST, so make that possible.

Signed-off-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: ethernet: freescale: Allow FEC with COMPILE_TEST
Florian Fainelli [Thu, 17 May 2018 20:07:44 +0000 (13:07 -0700)]
net: ethernet: freescale: Allow FEC with COMPILE_TEST

The Freescale FEC driver builds fine with COMPILE_TEST, so make that
possible.

Signed-off-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: ethernet: ti: Allow most drivers with COMPILE_TEST
Florian Fainelli [Thu, 17 May 2018 20:07:43 +0000 (13:07 -0700)]
net: ethernet: ti: Allow most drivers with COMPILE_TEST

Most of the TI drivers build just fine with COMPILE_TEST, cpmac (AR7) is
the exception because it uses a header file from
arch/mips/include/asm/mach-ar7/ar7.h and keystone netcp which requires
help from drivers/soc/ti/ for queue management helpers.

Signed-off-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agovlan: Add extack messages for link create
David Ahern [Thu, 17 May 2018 19:29:47 +0000 (12:29 -0700)]
vlan: Add extack messages for link create

Add informative messages for error paths related to adding a
VLAN to a device.

Signed-off-by: David Ahern <dsahern@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agoqede: Add build_skb() support.
Manish Chopra [Thu, 17 May 2018 19:05:00 +0000 (12:05 -0700)]
qede: Add build_skb() support.

This patch makes use of build_skb() throughout in driver's receieve
data path [HW gro flow and non HW gro flow]. With this, driver can
build skb directly from the page segments which are already mapped
to the hardware instead of allocating new SKB via netdev_alloc_skb()
and memcpy the data which is quite costly.

This really improves performance (keeping same or slight gain in rx
throughput) in terms of CPU utilization which is significantly reduced
[almost half] in non HW gro flow where for every incoming MTU sized
packet driver had to allocate skb, memcpy headers etc. Additionally
in that flow, it also gets rid of bunch of additional overheads
[eth_get_headlen() etc.] to split headers and data in the skb.

Tested with:
system: 2 sockets, 4 cores per socket, hyperthreading, 2x4x2=16 cores
iperf [server]: iperf -s
iperf [client]: iperf -c <server_ip> -t 500 -i 10 -P 32

HW GRO off – w/o build_skb(), throughput: 36.8 Gbits/sec

Average:     CPU    %usr   %nice    %sys %iowait    %irq   %soft  %steal  %guest   %idle
Average:     all    0.59    0.00   32.93    0.00    0.00   43.07    0.00    0.00   23.42

HW GRO off - with build_skb(), throughput: 36.9 Gbits/sec

Average:     CPU    %usr   %nice    %sys %iowait    %irq   %soft  %steal  %guest   %idle
Average:     all    0.70    0.00   31.70    0.00    0.00   25.68    0.00    0.00   41.92

HW GRO on - w/o build_skb(), throughput: 36.9 Gbits/sec

Average:     CPU    %usr   %nice    %sys %iowait    %irq   %soft  %steal  %guest   %idle
Average:     all    0.86    0.00   24.14    0.00    0.00    6.59    0.00    0.00   68.41

HW GRO on - with build_skb(), throughput: 37.5 Gbits/sec

Average:     CPU    %usr   %nice    %sys %iowait    %irq   %soft  %steal  %guest   %idle
Average:     all    0.87    0.00   23.75    0.00    0.00    6.19    0.00    0.00   69.19

Signed-off-by: Ariel Elior <ariel.elior@cavium.com>
Signed-off-by: Manish Chopra <manish.chopra@cavium.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agoMerge branch '10GbE' of git://git.kernel.org/pub/scm/linux/kernel/git/jkirsher/next...
David S. Miller [Thu, 17 May 2018 21:02:57 +0000 (17:02 -0400)]
Merge branch '10GbE' of git://git./linux/kernel/git/jkirsher/next-queue

Jeff Kirsher says:

====================
10GbE Intel Wired LAN Driver Updates 2018-05-17

This series contains updates to ixgbe, ixgbevf and ice drivers.

Cathy Zhou resolves sparse warnings by using the force attribute.

Mauro S M Rodrigues fixes a bug where IRQs were not freed if a PCI error
recovery system opts to remove the device which causes
ixgbe_io_error_detected() to return PCI_ERS_RESULT_DISCONNECT before
calling ixgbe_close_suspend() which results in IRQs not freed and
crashing when the remove handler calls pci_disable_device().  Resolved
this by calling ixgbe_close_suspend() before evaluating the PCI channel
state.

Pavel Tatashin releases the rtnl_lock during the call to
ixgbe_close_suspend() to allow scaling if device_shutdown() is
multi-threaded.

Emil modifies ixgbe to not validate the MAC address during a reset,
unless the MAC was set on the host so that the VF will get a new MAC
address every time it reloads.  Also updates ixgbevf to set
hw->mac.perm_addr in order to retain the custom MAC on a reset.

Anirudh updates the ice NVM read/erase/update AQ commands to align with
the latest specification.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agotc-testing: fixed copy-pasting error in ife tests
Roman Mashak [Thu, 17 May 2018 13:12:49 +0000 (09:12 -0400)]
tc-testing: fixed copy-pasting error in ife tests

Reported-by: Vlad Buslov <vladbu@mellanox.com>
Reported-by: Davide Caratti <dcaratti@redhat.com>
Signed-off-by: Roman Mashak <mrv@mojatatu.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet/ncsi: prevent a couple array underflows
Dan Carpenter [Thu, 17 May 2018 12:33:36 +0000 (15:33 +0300)]
net/ncsi: prevent a couple array underflows

We recently refactored this code and introduced a static checker
warning.  Smatch complains that if cmd->index is zero then we would
underflow the arrays.  That's obviously true.

The question is whether we prevent cmd->index from being zero at a
different level.  I've looked at the code and I don't immediately see
a check for that.

Fixes: 062b3e1b6d4f ("net/ncsi: Refactor MAC, VLAN filters")
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet/smc: init conn.tx_work & conn.send_lock sooner
Eric Dumazet [Thu, 17 May 2018 10:54:21 +0000 (03:54 -0700)]
net/smc: init conn.tx_work & conn.send_lock sooner

syzkaller found that following program crashes the host :

{
  int fd = socket(AF_SMC, SOCK_STREAM, 0);
  int val = 1;

  listen(fd, 0);
  shutdown(fd, SHUT_RDWR);
  setsockopt(fd, 6, TCP_NODELAY, &val, 4);
}

Simply initialize conn.tx_work & conn.send_lock at socket creation,
rather than deeper in the stack.

ODEBUG: assert_init not available (active state 0) object type: timer_list hint:           (null)
WARNING: CPU: 1 PID: 13988 at lib/debugobjects.c:329 debug_print_object+0x16a/0x210 lib/debugobjects.c:326
Kernel panic - not syncing: panic_on_warn set ...

CPU: 1 PID: 13988 Comm: syz-executor0 Not tainted 4.17.0-rc4+ #46
Hardware name: Google Google Compute Engine/Google Compute Engine, BIOS Google 01/01/2011
Call Trace:
 __dump_stack lib/dump_stack.c:77 [inline]
 dump_stack+0x1b9/0x294 lib/dump_stack.c:113
 panic+0x22f/0x4de kernel/panic.c:184
 __warn.cold.8+0x163/0x1b3 kernel/panic.c:536
 report_bug+0x252/0x2d0 lib/bug.c:186
 fixup_bug arch/x86/kernel/traps.c:178 [inline]
 do_error_trap+0x1de/0x490 arch/x86/kernel/traps.c:296
 do_invalid_op+0x1b/0x20 arch/x86/kernel/traps.c:315
 invalid_op+0x14/0x20 arch/x86/entry/entry_64.S:992
RIP: 0010:debug_print_object+0x16a/0x210 lib/debugobjects.c:326
RSP: 0018:ffff880197a37880 EFLAGS: 00010086
RAX: 0000000000000061 RBX: 0000000000000005 RCX: ffffc90001ed0000
RDX: 0000000000004aaf RSI: ffffffff8160f6f1 RDI: 0000000000000001
RBP: ffff880197a378c0 R08: ffff8801aa7a0080 R09: ffffed003b5e3eb2
R10: ffffed003b5e3eb2 R11: ffff8801daf1f597 R12: 0000000000000001
R13: ffffffff88d96980 R14: ffffffff87fa19a0 R15: ffffffff81666ec0
 debug_object_assert_init+0x309/0x500 lib/debugobjects.c:692
 debug_timer_assert_init kernel/time/timer.c:724 [inline]
 debug_assert_init kernel/time/timer.c:776 [inline]
 del_timer+0x74/0x140 kernel/time/timer.c:1198
 try_to_grab_pending+0x439/0x9a0 kernel/workqueue.c:1223
 mod_delayed_work_on+0x91/0x250 kernel/workqueue.c:1592
 mod_delayed_work include/linux/workqueue.h:541 [inline]
 smc_setsockopt+0x387/0x6d0 net/smc/af_smc.c:1367
 __sys_setsockopt+0x1bd/0x390 net/socket.c:1903
 __do_sys_setsockopt net/socket.c:1914 [inline]
 __se_sys_setsockopt net/socket.c:1911 [inline]
 __x64_sys_setsockopt+0xbe/0x150 net/socket.c:1911
 do_syscall_64+0x1b1/0x800 arch/x86/entry/common.c:287
 entry_SYSCALL_64_after_hwframe+0x49/0xbe

Fixes: 01d2f7e2cdd3 ("net/smc: sockopts TCP_NODELAY and TCP_CORK")
Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Ursula Braun <ubraun@linux.ibm.com>
Cc: linux-s390@vger.kernel.org
Reported-by: syzbot <syzkaller@googlegroups.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonfp: flower: fix error path during representor creation
Jiri Pirko [Thu, 17 May 2018 10:06:43 +0000 (12:06 +0200)]
nfp: flower: fix error path during representor creation

Don't store repr pointer to reprs array until the representor is
successfully created. This avoids message about "representor
destruction" even when it was never created. Also it cleans-up the flow.
Also, check return value after port alloc.

Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Reviewed-by: Jakub Kicinski <jakub.kicinski@netronome.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agoMerge branch 'mvpp2-small-improvements'
David S. Miller [Thu, 17 May 2018 20:18:55 +0000 (16:18 -0400)]
Merge branch 'mvpp2-small-improvements'

Antoine Tenart says:

====================
net: mvpp2: small improvements

Those 3 patches are small improvements to the Marvell PPv2 driver. The
series does not conflict with the one sent about phylink and
1000/2500baseX support, so the two series can live in parallel.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: mvpp2: print rx error with rate-limit
Yan Markman [Thu, 17 May 2018 08:34:27 +0000 (10:34 +0200)]
net: mvpp2: print rx error with rate-limit

Prevent flood of RX error prints during heavy traffic with weak signal
in link by checking net_ratelimit() before using netdev_err().

Signed-off-by: Yan Markman <ymarkman@marvell.com>
[Antoine: small rework, commit message]
Signed-off-by: Antoine Tenart <antoine.tenart@bootlin.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: mvpp2: set mac address does not require the stop/start sequence
Yan Markman [Thu, 17 May 2018 08:34:26 +0000 (10:34 +0200)]
net: mvpp2: set mac address does not require the stop/start sequence

Remove special stop/start handling from the set_mac_address callback.
All this special care is not needed, and can be removed. It also
simplifies the up/down status in the driver and helps avoiding possible
link status mismatch issues.

Signed-off-by: Yan Markman <ymarkman@marvell.com>
[Antoine: commit message]
Signed-off-by: Antoine Tenart <antoine.tenart@bootlin.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: mvpp2: avoid checking for free aggregated descriptors twice
Yan Markman [Thu, 17 May 2018 08:34:25 +0000 (10:34 +0200)]
net: mvpp2: avoid checking for free aggregated descriptors twice

Avoid repeating the check for free aggregated descriptors when it
already failed at the beginning of the function.

Signed-off-by: Yan Markman <ymarkman@marvell.com>
[Antoine: commit message]
Signed-off-by: Antoine Tenart <antoine.tenart@bootlin.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agoMerge branch 'mvpp2-phylink-conversion'
David S. Miller [Thu, 17 May 2018 20:11:40 +0000 (16:11 -0400)]
Merge branch 'mvpp2-phylink-conversion'

Antoine Tenart says:

====================
net: mvpp2: phylink conversion

This series convert the Marvell PPv2 driver to phylink (models the MAC
to PHY link).

One important point is the PPv2 driver supports two probe modes: device
tree and ACPI. This series only brings phylink support for the device
tree mode, as the ACPI one will need further work. Still, the driver
should be working as before when using ACPI. This split should be
temporary, and was discussed with Marcin (in Cc.) who added ACPI support
to the driver.

Also as the SFP cages on both DB boards can be considered as non-wired.
We thus chose not to describe those SFP cages and we use fixed-link.

The rest of the series uses phylink to add support for 1000BaseX and
2500BaseX modes in the PPv2 driver. To do this, two patches are needed
in the common PHY framework (patches 3 and 4). The last 4 patches modify
the device tree to use the new PPv2 functionalities.

The series has been tested for the device tree mode on the 7040-db,
8040-db and 8040-mcbin boards, to ensure all the interface where working
as expected.

@Dave: patches 7 to 10 should go through the mvebu tree (Gregory in
Cc.) to avoid any conflict with the other mvebu dt patches taken during
this cycle.

The series is based on today's net-next.

Since v2:
  - Removed the SFP description from the DB boards, as their SFP cages
    are wired properly. We now use fixed-link.
  - Because of this rework, split the series in two, so that the SFP
    part is reviewed separately.
  - Small fixes in the phylink patch.
  - Rebased on the latest net-next branch.

Since v1:
  - Chose a different approach to the SFP changes, as the previous ones
    weren't valid and reworked both BD boards device trees.
  - Misc fixes.
  - Added Kishon's acked-by on one patch.
  - Rebaed on latest net-next branch.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: mvpp2: 2500baseX support
Antoine Tenart [Thu, 17 May 2018 08:29:35 +0000 (10:29 +0200)]
net: mvpp2: 2500baseX support

This patch adds the 2500Base-X PHY mode support in the Marvell PPv2
driver. 2500Base-X is quite close to 1000Base-X and SGMII modes and uses
nearly the same code path.

Signed-off-by: Antoine Tenart <antoine.tenart@bootlin.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: mvpp2: 1000baseX support
Antoine Tenart [Thu, 17 May 2018 08:29:34 +0000 (10:29 +0200)]
net: mvpp2: 1000baseX support

This patch adds the 1000Base-X PHY mode support in the Marvell PPv2
driver. 1000Base-X is quite close the SGMII and uses nearly the same
code path.

Signed-off-by: Antoine Tenart <antoine.tenart@bootlin.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agophy: cp110-comphy: 2.5G SGMII mode
Antoine Tenart [Thu, 17 May 2018 08:29:33 +0000 (10:29 +0200)]
phy: cp110-comphy: 2.5G SGMII mode

This patch allow the CP110 comphy to configure some lanes in the
2.5G SGMII mode. This mode is quite close to SGMII and uses nearly the
same code path.

Signed-off-by: Antoine Tenart <antoine.tenart@bootlin.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agophy: add 2.5G SGMII mode to the phy_mode enum
Antoine Tenart [Thu, 17 May 2018 08:29:32 +0000 (10:29 +0200)]
phy: add 2.5G SGMII mode to the phy_mode enum

This patch adds one more generic PHY mode to the phy_mode enum, to allow
configuring generic PHYs to the 2.5G SGMII mode by using the set_mode
callback.

Signed-off-by: Antoine Tenart <antoine.tenart@bootlin.com>
Acked-by: Kishon Vijay Abraham I <kishon@ti.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: mvpp2: phylink support
Antoine Tenart [Thu, 17 May 2018 08:29:31 +0000 (10:29 +0200)]
net: mvpp2: phylink support

Convert the PPv2 driver to implement phylink helpers, and use phylink in
DT mode. The other mode supported is ACPI, which will need further work
in order to be entirely compatible with phylink.

The MAC and GoP configuration functions were completely moved to fit
into the phylink helpers. When a PHY is always present between the MAC
and the physical port, phylink only is used, but when this is not the
case (the MAC directly is connected to the physical port) the link IRQ
is used to detect changes in the link state and call phylink_mac_change.

The ACPI mode do not uses phylink as of now, and the changes shouldn't
impact its use.

Signed-off-by: Antoine Tenart <antoine.tenart@bootlin.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: mvpp2: align the ethtool ops definition
Antoine Tenart [Thu, 17 May 2018 08:29:30 +0000 (10:29 +0200)]
net: mvpp2: align the ethtool ops definition

Cosmetic patch to align the ethtool functions to ops definitions. This
patch does not change in any way the driver's behaviour.

Signed-off-by: Antoine Tenart <antoine.tenart@bootlin.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agoMerge tag 'wireless-drivers-next-for-davem-2018-05-17' of git://git.kernel.org/pub...
David S. Miller [Thu, 17 May 2018 20:00:40 +0000 (16:00 -0400)]
Merge tag 'wireless-drivers-next-for-davem-2018-05-17' of git://git./linux/kernel/git/kvalo/wireless-drivers-next

Kalle Valo says:

====================
wireless-drivers-next patches for 4.18

The first pull request for 4.18. As usual new features and bug fixes
but nothing really special.

I also merged wireless-drivers due to an iwlwifi patch dependency.

Major changes:

iwlwifi

* implement Traffic Condition Monitor and use it for scan, BT coex and
  to detect when the AP doesn't support UAPSD properly

* some more work for the 22000 family of devices;

* introduce AMSDU rate control offload

qtnfmac

* DFS offload support

rsi

* roaming enhancements

* increase max supported aggregation subframes

* don't advertise 5 GHz support if the device doesn't support it

brcmfmac

* add support for BCM4366E chipset

* add support for bcm43364 wireless chipset

ath10k

* enable temperature reads for QCA6174 and QCA9377

* add firmware memory dump support for QCA9984

* continue adding WCN3990 support via SNOC bus
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agovmxnet3: Replace msleep(1) with usleep_range()
YueHaibing [Thu, 17 May 2018 03:46:41 +0000 (11:46 +0800)]
vmxnet3: Replace msleep(1) with usleep_range()

As documented in Documentation/timers/timers-howto.txt,
replace msleep(1) with usleep_range().

Signed-off-by: YueHaibing <yuehaibing@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agobonding: introduce link change helper
Tonghao Zhang [Thu, 17 May 2018 02:09:23 +0000 (19:09 -0700)]
bonding: introduce link change helper

Introduce an new common helper to avoid redundancy.

Signed-off-by: Tonghao Zhang <xiangxia.m.yue@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agoMerge branch 'tcp-default-RACK-loss-recovery'
David S. Miller [Thu, 17 May 2018 19:41:30 +0000 (15:41 -0400)]
Merge branch 'tcp-default-RACK-loss-recovery'

Yuchung Cheng says:

====================
tcp: default RACK loss recovery

This patch set implements the features correspond to the
draft-ietf-tcpm-rack-03 version of the RACK draft.
https://datatracker.ietf.org/meeting/101/materials/slides-101-tcpm-update-on-tcp-rack-00

1. SACK: implement equivalent DUPACK threshold heuristic in RACK to
   replace existing RFC6675 recovery (tcp_mark_head_lost).

2. Non-SACK: simplify RFC6582 NewReno implementation

3. RTO: apply RACK's time-based approach to avoid spuriouly
   marking very recently sent packets lost.

4. with (1)(2)(3), make RACK the exclusive fast recovery mechanism to
   mark losses based on time on S/ACK. Tail loss probe and F-RTO remain
   enabled by default as complementary mechanisms to send probes in
   CA_Open and CA_Loss states. The probes would solicit S/ACKs to trigger
   RACK time-based loss detection.

All Google web and internal servers have been running RACK-only mode
(4) for a while now. a/b experiments indicate RACK/TLP on average
reduces recovery latency by 10% compared to RFC6675. RFC6675
is default-off now but can be enabled by disabling RACK (sysctl
net.ipv4.tcp_recovery=0) for unseen issues.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agotcp: don't mark recently sent packets lost on RTO
Yuchung Cheng [Wed, 16 May 2018 23:40:17 +0000 (16:40 -0700)]
tcp: don't mark recently sent packets lost on RTO

An RTO event indicates the head has not been acked for a long time
after its last (re)transmission. But the other packets are not
necessarily lost if they have been only sent recently (for example
due to application limit). This patch would prohibit marking packets
sent within an RTT to be lost on RTO event, using similar logic in
TCP RACK detection.

Normally the head (SND.UNA) would be marked lost since RTO should
fire strictly after the head was sent. An exception is when the
most recent RACK RTT measurement is larger than the (previous)
RTO. To address this exception the head is always marked lost.

Congestion control interaction: since we may not mark every packet
lost, the congestion window may be more than 1 (inflight plus 1).
But only one packet will be retransmitted after RTO, since
tcp_retransmit_timer() calls tcp_retransmit_skb(...,segs=1). The
connection still performs slow start from one packet (with Cubic
congestion control).

This commit was tested in an A/B test with Google web servers,
and showed a reduction of 2% in (spurious) retransmits post
timeout (SlowStartRetrans), and correspondingly reduced DSACKs
(DSACKIgnoredOld) by 7%.

Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Priyaranjan Jha <priyarjha@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agotcp: new helper tcp_rack_skb_timeout
Yuchung Cheng [Wed, 16 May 2018 23:40:16 +0000 (16:40 -0700)]
tcp: new helper tcp_rack_skb_timeout

Create and export a new helper tcp_rack_skb_timeout and move tcp_is_rack
to prepare the final RTO change.

Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Priyaranjan Jha <priyarjha@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agotcp: separate loss marking and state update on RTO
Yuchung Cheng [Wed, 16 May 2018 23:40:15 +0000 (16:40 -0700)]
tcp: separate loss marking and state update on RTO

Previously when TCP times out, it first updates cwnd and ssthresh,
marks packets lost, and then updates congestion state again. This
was fine because everything not yet delivered is marked lost,
so the inflight is always 0 and cwnd can be safely set to 1 to
retransmit one packet on timeout.

But the inflight may not always be 0 on timeout if TCP changes to
mark packets lost based on packet sent time. Therefore we must
first mark the packet lost, then set the cwnd based on the
(updated) inflight.

This is not a pure refactor. Congestion control may potentially
break if it uses (not yet updated) inflight to compute ssthresh.
Fortunately all existing congestion control modules does not do that.
Also it changes the inflight when CA_LOSS_EVENT is called, and only
westwood processes such an event but does not use inflight.

This change has two other minor side benefits:
1) consistent with Fast Recovery s.t. the inflight is updated
   first before tcp_enter_recovery flips state to CA_Recovery.

2) avoid intertwining loss marking with state update, making the
   code more readable.

Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Priyaranjan Jha <priyarjha@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agotcp: new helper tcp_timeout_mark_lost
Yuchung Cheng [Wed, 16 May 2018 23:40:14 +0000 (16:40 -0700)]
tcp: new helper tcp_timeout_mark_lost

Refactor using a new helper, tcp_timeout_mark_loss(), that marks packets
lost upon RTO.

Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Priyaranjan Jha <priyarjha@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agotcp: account lost retransmit after timeout
Yuchung Cheng [Wed, 16 May 2018 23:40:13 +0000 (16:40 -0700)]
tcp: account lost retransmit after timeout

The previous approach for the lost and retransmit bits was to
wipe the slate clean: zero all the lost and retransmit bits,
correspondingly zero the lost_out and retrans_out counters, and
then add back the lost bits (and correspondingly increment lost_out).

The new approach is to treat this very much like marking packets
lost in fast recovery. We don’t wipe the slate clean. We just say
that for all packets that were not yet marked sacked or lost, we now
mark them as lost in exactly the same way we do for fast recovery.

This fixes the lost retransmit accounting at RTO time and greatly
simplifies the RTO code by sharing much of the logic with Fast
Recovery.

Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Priyaranjan Jha <priyarjha@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agotcp: simpler NewReno implementation
Yuchung Cheng [Wed, 16 May 2018 23:40:12 +0000 (16:40 -0700)]
tcp: simpler NewReno implementation

This is a rewrite of NewReno loss recovery implementation that is
simpler and standalone for readability and better performance by
using less states.

Note that NewReno refers to RFC6582 as a modification to the fast
recovery algorithm. It is used only if the connection does not
support SACK in Linux. It should not to be confused with the Reno
(AIMD) congestion control.

Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Priyaranjan Jha <priyarjha@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agotcp: disable RFC6675 loss detection
Yuchung Cheng [Wed, 16 May 2018 23:40:11 +0000 (16:40 -0700)]
tcp: disable RFC6675 loss detection

This patch disables RFC6675 loss detection and make sysctl
net.ipv4.tcp_recovery = 1 controls a binary choice between RACK
(1) or RFC6675 (0).

Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Priyaranjan Jha <priyarjha@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agotcp: support DUPACK threshold in RACK
Yuchung Cheng [Wed, 16 May 2018 23:40:10 +0000 (16:40 -0700)]
tcp: support DUPACK threshold in RACK

This patch adds support for the classic DUPACK threshold rule
(#DupThresh) in RACK.

When the number of packets SACKed is greater or equal to the
threshold, RACK sets the reordering window to zero which would
immediately mark all the unsacked packets below the highest SACKed
sequence lost. Since this approach is known to not work well with
reordering, RACK only uses it if no reordering has been observed.

The DUPACK threshold rule is a particularly useful extension to the
fast recoveries triggered by RACK reordering timer. For example
data-center transfers where the RTT is much smaller than a timer
tick, or high RTT path where the default RTT/4 may take too long.

Note that this patch differs slightly from RFC6675. RFC6675
considers a packet lost when at least #DupThresh higher-sequence
packets are SACKed.

With RACK, for connections that have seen reordering, RACK
continues to use a dynamically-adaptive time-based reordering
window to detect losses. But for connections on which we have not
yet seen reordering, this patch considers a packet lost when at
least one higher sequence packet is SACKed and the total number
of SACKed packets is at least DupThresh. For example, suppose a
connection has not seen reordering, and sends 10 packets, and
packets 3, 5, 7 are SACKed. RFC6675 considers packets 1 and 2
lost. RACK considers packets 1, 2, 4, 6 lost.

There is some small risk of spurious retransmits here due to
reordering. However, this is mostly limited to the first flight of
a connection on which the sender receives SACKs from reordering.
And RFC 6675 and FACK loss detection have a similar risk on the
first flight with reordering (it's just that the risk of spurious
retransmits from reordering was slightly narrower for those older
algorithms due to the margin of 3*MSS).

Also the minimum reordering window is reduced from 1 msec to 0
to recover quicker on short RTT transfers. Therefore RACK is more
aggressive in marking packets lost during recovery to reduce the
reordering window timeouts.

Signed-off-by: Yuchung Cheng <ycheng@google.com>
Signed-off-by: Neal Cardwell <ncardwell@google.com>
Reviewed-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Soheil Hassas Yeganeh <soheil@google.com>
Reviewed-by: Priyaranjan Jha <priyarjha@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agonet: ethernet: ti: cpsw: disable mq feature for "AM33xx ES1.0" devices
Ivan Khoronzhuk [Wed, 16 May 2018 22:21:45 +0000 (01:21 +0300)]
net: ethernet: ti: cpsw: disable mq feature for "AM33xx ES1.0" devices

The early versions of am33xx devices, related to ES1.0 SoC revision
have errata limiting mq support. That's the same errata as
commit 7da1160002f1 ("drivers: net: cpsw: add am335x errata workarround for
interrutps")

AM33xx Errata [1] Advisory 1.0.9
http://www.ti.com/lit/er/sprz360f/sprz360f.pdf

After additional investigation were found that drivers w/a is
propagated on all AM33xx SoCs and on DM814x. But the errata exists
only for ES1.0 of AM33xx family, limiting mq support for revisions
after ES1.0. So, disable mq support only for related SoCs and use
separate polls for revisions allowing mq.

Signed-off-by: Ivan Khoronzhuk <ivan.khoronzhuk@linaro.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agoMerge branch 'sched-refactor-NOLOCK-qdiscs'
David S. Miller [Thu, 17 May 2018 16:46:55 +0000 (12:46 -0400)]
Merge branch 'sched-refactor-NOLOCK-qdiscs'

Paolo Abeni says:

====================
sched: refactor NOLOCK qdiscs

With the introduction of NOLOCK qdiscs, pfifo_fast performances in the
uncontended scenario degraded measurably, especially after the commit
eb82a9944792 ("net: sched, fix OOO packets with pfifo_fast").

This series restore the pfifo_fast performances in such scenario back the
previous level, mainly reducing the number of atomic operations required to
perform the qdisc_run() call. Even performances in the contended scenario
increase measurably.

Note: This series is on top of:

sched: manipulate __QDISC_STATE_RUNNING in qdisc_run_* helpers
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agopfifo_fast: drop unneeded additional lock on dequeue
Paolo Abeni [Tue, 15 May 2018 14:24:37 +0000 (16:24 +0200)]
pfifo_fast: drop unneeded additional lock on dequeue

After the previous patch, for NOLOCK qdiscs, q->seqlock is
always held when the dequeue() is invoked, we can drop
any additional locking to protect such operation.

Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agosched: replace __QDISC_STATE_RUNNING bit with a spin lock
Paolo Abeni [Tue, 15 May 2018 14:24:36 +0000 (16:24 +0200)]
sched: replace __QDISC_STATE_RUNNING bit with a spin lock

So that we can use lockdep on it.
The newly introduced sequence lock has the same scope of busylock,
so it shares the same lockdep annotation, but it's only used for
NOLOCK qdiscs.

With this changeset we acquire such lock in the control path around
flushing operation (qdisc reset), to allow more NOLOCK qdisc perf
improvement in the next patch.

Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
6 years agoice: Update NVM AQ command functions
Anirudh Venkataramanan [Mon, 16 Apr 2018 16:55:36 +0000 (09:55 -0700)]
ice: Update NVM AQ command functions

This patch updates the NVM read/erase/update AQ commands to align with
the latest specification.

Signed-off-by: Anirudh Venkataramanan <anirudh.venkataramanan@intel.com>
Tested-by: Tony Brelinski <tonyx.brelinski@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
6 years agoixgbevf: fix MAC address changes through ixgbevf_set_mac()
Emil Tantilov [Mon, 14 May 2018 18:16:16 +0000 (11:16 -0700)]
ixgbevf: fix MAC address changes through ixgbevf_set_mac()

Set hw->mac.perm_addr in ixgbevf_set_mac() in order to avoid losing the
custom MAC on reset. This can happen in the following case:

>ip link set $vf address $mac
>ethtool -r $vf

Signed-off-by: Emil Tantilov <emil.s.tantilov@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
6 years agoixgbe: force VF to grab new MAC on driver reload
Emil Tantilov [Mon, 14 May 2018 18:16:11 +0000 (11:16 -0700)]
ixgbe: force VF to grab new MAC on driver reload

Do not validate the MAC address during a reset, unless the MAC was set on
the host. This way the VF will get a new MAC address every time it reloads.

Remove the "no MAC address assigned" message since it will get spammed on
reset and it doesn't help much as the MAC on the VF is randomly generated.

Signed-off-by: Emil Tantilov <emil.s.tantilov@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
6 years agoixgbe: release lock for the duration of ixgbe_suspend_close()
Pavel Tatashin [Thu, 3 May 2018 03:59:30 +0000 (23:59 -0400)]
ixgbe: release lock for the duration of ixgbe_suspend_close()

Currently, during device_shutdown() ixgbe holds rtnl_lock for the duration
of lengthy ixgbe_close_suspend(). On machines with multiple ixgbe cards
this lock prevents scaling if device_shutdown() function is multi-threaded.

It is not necessary to hold this lock during ixgbe_close_suspend()
as it is not held when ixgbe_close() is called also during shutdown but for
kexec case.

Signed-off-by: Pavel Tatashin <pasha.tatashin@oracle.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
6 years agoixgbe/ixgbevf: Free IRQ when PCI error recovery removes the device
Mauro S M Rodrigues [Wed, 2 May 2018 20:26:28 +0000 (17:26 -0300)]
ixgbe/ixgbevf: Free IRQ when PCI error recovery removes the device

Since commit f7f37e7ff2b9 ("ixgbe: handle close/suspend race with
netif_device_detach/present") ixgbe_close_suspend is called, from
ixgbe_close, only if the device is present, i.e. if it isn't detached.
That exposed a situation where IRQs weren't freed if a PCI error
recovery system opts to remove the device. For such case the pci channel
state is set to pci_channel_io_perm_failure and ixgbe_io_error_detected
was returning PCI_ERS_RESULT_DISCONNECT before calling
ixgbe_close_suspend consequentially not freeing IRQ and crashing when
the remove handler calls pci_disable_device, hitting a BUG_ON at
free_msi_irqs, which asserts that there is no non-free IRQ associated
with the device to be removed:

BUG_ON(irq_has_action(entry->irq + i));

The issue is fixed by calling the ixgbe_close_suspend before evaluate
the pci channel state.

Reported-by: Naresh Bannoth <nbannoth@in.ibm.com>
Reported-by: Abdul Haleem <abdhalee@in.ibm.com>
Signed-off-by: Mauro S M Rodrigues <maurosr@linux.vnet.ibm.com>
Reviewed-by: Alexander Duyck <alexander.h.duyck@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>