cascardo/linux.git
8 years agomlxsw: spectrum: Add IEEE 802.1Qaz ETS support
Ido Schimmel [Wed, 6 Apr 2016 15:10:10 +0000 (17:10 +0200)]
mlxsw: spectrum: Add IEEE 802.1Qaz ETS support

Implement the appropriate DCB ops and allow a user to configure:
* Priority to traffic class (TC) mapping with a total of 8
  supported TCs
* Transmission selection algorithm (TSA) for each TC and the
  corresponding weights in case of weighted round robin (WRR)

As previously explained, we treat the priority group (PG) buffer in the
port's headroom as the ingress counterpart of the egress TC. Therefore,
when a certain priority to TC mapping is configured, we also configure
the port's headroom buffer.

Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agomlxsw: spectrum: Introduce support for Data Center Bridging (DCB)
Ido Schimmel [Wed, 6 Apr 2016 15:10:09 +0000 (17:10 +0200)]
mlxsw: spectrum: Introduce support for Data Center Bridging (DCB)

Introduce basic infrastructure for DCB and add the missing ops in
following patches.

Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agomlxsw: spectrum: Initialize egress scheduling
Ido Schimmel [Wed, 6 Apr 2016 15:10:08 +0000 (17:10 +0200)]
mlxsw: spectrum: Initialize egress scheduling

Before introducing support for DCB ops we should first make sure we
initialize the relevant parts in the device correctly. Specifically, the
egress scheduling.

The device supports a superset of the 802.1Qaz standard with 4 hierarchy
levels that can be linked to each other in multiple ways and with
different transmission selection algorithms (TSA) employed between them.

However, since we only intend to support the 802.1Qaz standard we
flatten the hierarchies and let the user configure via DCB ops the TSA
and max rate shaper at the subgroup hierarchy (see figure below) and the
mapping between switch priority to traffic class. By default, all switch
priorities are mapped to traffic class 0, strict priority is employed
and max shaper is disabled.

Default configuration:

         switch priority 0      ...         switch priority 7
                 +                                  +
                 |                                  |
                 +----------------------------------+
                 |
              +--v--+                          +-----+
Traffic Class |     |                          |     |
  Hierarchy   | TC0 |           ...            | TC7 |
              |     |                          |     |
              +--+--+                          +--+--+
                 |                                |
              +--v--+                          +--v--+
  Subgroup    | SG0 |                          | SG7 |
  Hierarchy   |     |                          |     |
              +-----+                          +-----+
              | TSA |                          | TSA |
              +-----+           ...            +-----+
              | MAX |                          | MAX |
              +--+--+                          +--+--+
                 |                                |
                 +---------------+----------------+
                                 |
                              +--v--+
                      Group   |     |
                    Hierarchy | GR0 |
                              |     |
                              +--+--+
                                 |
                              +--v--+
                      Port    |     |
                    Hierarchy | PR0 |
                              |     |
                              +-----+

Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agomlxsw: reg: Add QoS Switch Traffic Class Table register
Ido Schimmel [Wed, 6 Apr 2016 15:10:07 +0000 (17:10 +0200)]
mlxsw: reg: Add QoS Switch Traffic Class Table register

As part of DCB ops we'll have to configure the priority to traffic class
mapping of a port.

Add the QoS Switch Traffic Class Table (QTCT) register, which configures
the mapping between the packet switch priority and traffic class on the
transmit port.

Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agomlxsw: reg: Add QoS ETS Element Configuration register
Ido Schimmel [Wed, 6 Apr 2016 15:10:06 +0000 (17:10 +0200)]
mlxsw: reg: Add QoS ETS Element Configuration register

We are going to introduce support for DCB, so we need to be able to
configure the traffic selection algorithm (TSA) used by each traffic
class (TC), as well as the bandwidth percentage allocated to each TC in
case of ETS.

Add the QoS ETS Element Configuration register, which controls the
above parameters.

Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agomlxsw: spectrum: Set port's shared buffer size to 0
Ido Schimmel [Wed, 6 Apr 2016 15:10:05 +0000 (17:10 +0200)]
mlxsw: spectrum: Set port's shared buffer size to 0

In addition to the priority group (PG) buffers in the headroom, the
device enables the allocation of headroom shared buffer, which can
be shared between different PGs.

However, we are not going to use the headroom shared buffer and instead
allow the user to use its size for PGs or the switch's shared buffer.

Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agomlxsw: reg: Use correct PBMC register length
Ido Schimmel [Wed, 6 Apr 2016 15:10:04 +0000 (17:10 +0200)]
mlxsw: reg: Use correct PBMC register length

The last field of the PBMC register is at offset 0x64 and its size is
0x8, so the correct register's length is 0x6C bytes.

Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agomlxsw: spectrum: Correctly configure headroom size
Ido Schimmel [Wed, 6 Apr 2016 15:10:03 +0000 (17:10 +0200)]
mlxsw: spectrum: Correctly configure headroom size

When packets ingress the switch they are assigned a switch priority and
directed to the corresponding priority group (PG) buffer in the port's
headroom buffer.

Since we now map all switch priorities to priority group 0 (PG0) by
default, there is no need to allocate the other priority groups during
initialization. The only exception is PG9, which is used for control
traffic.

At minimum, the PG should be able to store the currently classified
packet (pipeline latency isn't 0) and also the packets arriving during
the classification time. However, an incoming packet will not be
buffered if there is no available MTU-sized buffer space for storing it.

The buffer needed to accommodate for pipeline latency is variable and
needs to take into account both the current link speed and current
latency of the pipeline, which is time-dependent. Testing showed that
setting the PG's size to twice the current MTU is optimal.

Since PG9 is used strictly for control packets and not subject to flow
control, we are not going to resize it according to user configuration,
so we simply set it according to worst case scenario, which is twice the
maximum MTU.

In any case, later patches in the series will allow a user to direct
lossless flows to other PGs than PG0 and set their size to accommodate
for round-trip propagation delay.

The above change also requires us to resize the PG buffer whenever the
port's MTU is changed.

Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agomlxsw: spectrum: Add bytes to cells helper
Ido Schimmel [Wed, 6 Apr 2016 15:10:02 +0000 (17:10 +0200)]
mlxsw: spectrum: Add bytes to cells helper

Buffers in the switch store packets in units called buffer cells. Add a
helper to convert from bytes to cells, so that the actual number of
cells required (result is round up) is returned.

Also, drop the SB (shared buffer) acronym from the BYTES_PER_CELL macro,
as this unit is also used in the ports' buffers and not only the
switch's shared buffer.

Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agomlxsw: spectrum: Map all switch priorities to priority group 0
Ido Schimmel [Wed, 6 Apr 2016 15:10:01 +0000 (17:10 +0200)]
mlxsw: spectrum: Map all switch priorities to priority group 0

During transmission, the skb's priority is used to map the skb to a
traffic class, where the idea is to group priorities with similar
characteristics (e.g. lossy, lossless) to the same traffic class. By
default, all priorities are mapped to traffic class 0.

In the device, we model the skb's priority as the switch priority, which
is assigned to a packet according to its PCP value and ingress port
(untagged packets are assigned the port's default switch priority - 0).

At ingress, the packet is directed to a priority group (PG) buffer in
the port's headroom buffer according to the packet's switch priority and
switch priority to buffer mapping.

While it's possible to configure the egress mapping between skb's
priority (switch priority) and traffic class, there is no mechanism to
configure the ingress mapping to a PG.

In order to keep things simple and since grouping certain priorities into
a traffic class at egress also implies they should be grouped the same
at ingress, treat a PG as the ingress counterpart of an egress traffic
class.

Having established the above, during initialization map all the switch
priorities to PG0 in accordance with the Linux defaults for traffic
class mapping.

Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agomlxsw: reg: Add Port Prio To Buffer register
Ido Schimmel [Wed, 6 Apr 2016 15:10:00 +0000 (17:10 +0200)]
mlxsw: reg: Add Port Prio To Buffer register

When packets ingress the switch they are assigned a switch priority
number that dictates the packet's priority group (PG) buffer in the
port's headroom buffer.

Add the Port Prio To Buffer (PPTB) register, which configures the switch
priority to PG mapping.

Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoMerge branch '40GbE' of git://git.kernel.org/pub/scm/linux/kernel/git/jkirsher/next...
David S. Miller [Wed, 6 Apr 2016 21:03:35 +0000 (17:03 -0400)]
Merge branch '40GbE' of git://git./linux/kernel/git/jkirsher/next-queue

Jeff Kirsher says:

====================
40GbE Intel Wired LAN Driver Updates 2016-04-05

This series contains updates to i40e and i40evf only.

Colin Ian King cleaned up a redundant NULL check which was found by static
analysis.

Anjali enables geneve receive offload for XL710/X710 devices.

Mitch cleans up unused variable in i40e_vc_get_vf_resources_msg().
Fixed the driver to actually be able to adjust VLAN tagging features
through ethtool, as expected.  Fixed a problem where VF resets would
get lost by the PF preventing the VF driver from initializing.  Also
put users mind at ease by lowering some message levels since many of
these conditions can happen any time VFs are enabled or disabled and
are not really indicative a fatal problems, unless they happen
continuously.

Shannon disables the link polling to lessen the admin queue traffic
especially since the link event mask usage has been fixed recently.

Alex Duyck fixes the i40e and i40evf drivers to correctly update
checksums for frames up to 16776960 in length which should be more than
large enough for all possible TSO frames in the near future.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoMerge branch 'vxlan-gpe'
David S. Miller [Wed, 6 Apr 2016 20:50:33 +0000 (16:50 -0400)]
Merge branch 'vxlan-gpe'

Jiri Benc says:

====================
vxlan: implement Generic Protocol Extension (GPE)

v3: just rebased on top of the current net-next, no changes

This patchset implements VXLAN-GPE. It follows the same model as the tun/tap
driver: depending on the chosen mode, the vxlan interface is created either
as ARPHRD_ETHER (non-GPE) or ARPHRD_NONE (GPE).

Note that the internal fdb control plane cannot be used together with
VXLAN-GPE and attempt to configure it will be rejected by the driver. In
fact, COLLECT_METADATA is required to be set for now. This can be relaxed in
the future by adding support for static PtP configuration; it will be
backward compatible and won't affect existing users.

The previous version of the patchset supported two GPE modes, L2 and L3. The
L2 mode (now called "ether mode" in the code) was removed from this version.
It can be easily added later if there's demand. The L3 mode is now called
"raw mode" and supports also encapsulated Ethernet headers (via ETH_P_TEB).

The only limitation of not having "ether mode" for GPE is for ip route based
encapsulation: with such setup, only IP packets can be encapsulated. Meaning
no Ethernet encapsulation. It seems there's not much use for this, though.
If it turns out to be useful, we'll add it.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agovxlan: implement GPE
Jiri Benc [Tue, 5 Apr 2016 12:47:13 +0000 (14:47 +0200)]
vxlan: implement GPE

Implement VXLAN-GPE. Only COLLECT_METADATA is supported for now (it is
possible to support static configuration, too, if there is demand for it).

The GPE header parsing has to be moved before iptunnel_pull_header, as we
need to know the protocol.

v2: Removed what was called "L2 mode" in v1 of the patchset. Only "L3 mode"
    (now called "raw mode") is added by this patch. This mode does not allow
    Ethernet header to be encapsulated in VXLAN-GPE when using ip route to
    specify the encapsulation, IP header is encapsulated instead. The patch
    does support Ethernet to be encapsulated, though, using ETH_P_TEB in
    skb->protocol. This will be utilized by other COLLECT_METADATA users
    (openvswitch in particular).

    If there is ever demand for Ethernet encapsulation with VXLAN-GPE using
    ip route, it's easy to add a new flag switching the interface to
    "Ethernet mode" (called "L2 mode" in v1 of this patchset). For now,
    leave this out, it seems we don't need it.

    Disallowed more flag combinations, especially RCO with GPE.
    Added comment explaining that GBP and GPE cannot be set together.

Signed-off-by: Jiri Benc <jbenc@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoip_tunnel: implement __iptunnel_pull_header
Jiri Benc [Tue, 5 Apr 2016 12:47:12 +0000 (14:47 +0200)]
ip_tunnel: implement __iptunnel_pull_header

Allow calling of iptunnel_pull_header without special casing ETH_P_TEB inner
protocol.

Signed-off-by: Jiri Benc <jbenc@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agovxlan: move fdb code to common location in vxlan_xmit
Jiri Benc [Tue, 5 Apr 2016 12:47:11 +0000 (14:47 +0200)]
vxlan: move fdb code to common location in vxlan_xmit

Handle VXLAN_F_COLLECT_METADATA before VXLAN_F_PROXY. The latter does not
make sense with the former, as it needs populated fdb which does not happen
in metadata mode.

After this cleanup, the fdb code in vxlan_xmit is moved to a common location
and can be later skipped for VXLAN-GPE which does not necessarily carry
inner Ethernet header.

v2: changed commit description to not reference L3 mode

Signed-off-by: Jiri Benc <jbenc@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agovxlan: move Ethernet initialization to a separate function
Jiri Benc [Tue, 5 Apr 2016 12:47:10 +0000 (14:47 +0200)]
vxlan: move Ethernet initialization to a separate function

This will allow to initialize vxlan in ARPHRD_NONE mode based on the passed
rtnl attributes.

v2: renamed "l2mode" to "ether".

Signed-off-by: Jiri Benc <jbenc@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agocxgb4/cxgb4vf: Deprecate module parameter dflt_msg_enable
Hariprasad Shenai [Tue, 5 Apr 2016 04:22:21 +0000 (09:52 +0530)]
cxgb4/cxgb4vf: Deprecate module parameter dflt_msg_enable

Message level can be set through ethtool, so deprecate module parameter
which is used to set the same.

Signed-off-by: Hariprasad Shenai <hariprasad@chelsio.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoi40e/i40evf: Fix TSO checksum pseudo-header adjustment
Alexander Duyck [Fri, 18 Mar 2016 23:06:47 +0000 (16:06 -0700)]
i40e/i40evf: Fix TSO checksum pseudo-header adjustment

With IPv4 and IPv6 now using the same format for checksums based on the
length of the frame we need to update the i40e and i40evf drivers so that
they correctly account for lengths greater than or equal to 64K.

With this patch the driver should now correctly update checksums for frames
up to 16776960 in length which should be more than large enough for all
possible TSO frames in the near future.

Signed-off-by: Alexander Duyck <aduyck@mirantis.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agoi40e/i40evf: Bump patch from 1.5.1 to 1.5.2
Avinash Dayanand [Fri, 18 Mar 2016 19:18:16 +0000 (12:18 -0700)]
i40e/i40evf: Bump patch from 1.5.1 to 1.5.2

Signed-off-by: Avinash Dayanand <avinash.dayanand@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agoi40e: Request PHY media event at reset time
Shannon Nelson [Fri, 18 Mar 2016 19:18:15 +0000 (12:18 -0700)]
i40e: Request PHY media event at reset time

Add the Media Not Available flag to the link event mask.  It seems
that event comes first if you have a DA cable pulled out, but there's no
follow-up event for Link Down; if you're not looking for MEDIA_NA you will
get no event, even though there's now no Link.

Change-ID: cb3340a2849805bb881f64f6f2ae810eef46eba7
Signed-off-by: Shannon Nelson <shannon.nelson@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agoi40e: Lower some message levels
Mitch Williams [Fri, 18 Mar 2016 19:18:14 +0000 (12:18 -0700)]
i40e: Lower some message levels

These conditions can happen any time VFs are enabled or disabled and are
not really indicative of fatal problems unless they happen continuously.

Lower the log level so that people don't get scared.

Change-ID: I1ceb4adbd10d03cbeed54d1f5b7f20d60328351d
Signed-off-by: Mitch Williams <mitch.a.williams@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agoi40e: Fix for supported link modes in 10GBaseT PHY's
Avinash Dayanand [Fri, 18 Mar 2016 19:18:13 +0000 (12:18 -0700)]
i40e: Fix for supported link modes in 10GBaseT PHY's

100baseT/Full is now listed and supported link mode for 10GBaseT PHY.
This is a fix to list all the supported link modes of 10GBaseT PHY.

Change-ID: If2be3212ef0fef85fd5d6e4550c7783de2f915e9
Signed-off-by: Avinash Dayanand <avinash.dayanand@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agoi40evf: Fix get_rss_aq
Catherine Sullivan [Fri, 18 Mar 2016 19:18:12 +0000 (12:18 -0700)]
i40evf: Fix get_rss_aq

We were passing in the seed where we should just be passing false
because we want the VSI table not the pf table.

Change-ID: I9b633ab06eb59468087f0c0af8539857e99f9495
Signed-off-by: Catherine Sullivan <catherine.sullivan@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agoi40e: Disable link polling
Shannon Nelson [Fri, 18 Mar 2016 19:18:11 +0000 (12:18 -0700)]
i40e: Disable link polling

Periodic link polling was added when the link events were found not to be
trustworthy.  This was the case early on, but was likely because the link
event mask was being used incorrectly.  As this has been fixed in recent
code, we can disable the link polling to lessen the AQ traffic.

Change-ID: Id890b5ee3c2d04381fc76ffa434777644f5d8eb0
Signed-off-by: Shannon Nelson <shannon.nelson@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agoi40evf: Add longer wait after remove module
Mitch Williams [Fri, 18 Mar 2016 19:18:10 +0000 (12:18 -0700)]
i40evf: Add longer wait after remove module

Upon module remove, wait a little longer after requesting a reset before
checking to see if the firmware responded. This change prevents double
resets when the firmware is busy.

Change-ID: Ieedc988ee82fac1f32a074bf4d9e4dba426bfa58
Signed-off-by: Mitch Williams <mitch.a.williams@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agoi40e: Make VF resets more reliable
Mitch Williams [Fri, 18 Mar 2016 19:18:09 +0000 (12:18 -0700)]
i40e: Make VF resets more reliable

Clear the VFLR bit immediately after triggering a reset instead of
waiting until after cleanup is complete. Make sure to trigger a reset
every time, not just if the PF is up.

These changes fix a problem where VF resets would get lost by the PF,
preventing the VF driver from initializing.

Change-ID: I5945cf2884095b7b0554867c64df8617e71d9d29
Signed-off-by: Mitch Williams <mitch.a.williams@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agoi40e: Add new device ID for X722
Catherine Sullivan [Fri, 18 Mar 2016 19:18:08 +0000 (12:18 -0700)]
i40e: Add new device ID for X722

The new device ID is 0x37D3 and it should follow the same flows and
branding string as for 0x37D0.

Change-ID: Ia5ad4a1910268c4666a3fd46a7afffbec55b4fc2
Signed-off-by: Catherine Sullivan <catherine.sullivan@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agoi40evf: Fix VLAN features
Mitch Williams [Fri, 18 Mar 2016 19:18:07 +0000 (12:18 -0700)]
i40evf: Fix VLAN features

Users of ethtool were being given the mistaken impression that this
driver was able to change its VLAN tagging features, and were
disappointed that this was not actually the case. Implement
ndo_fix_features method so that we can adjust these flags as needed to
avoid false impressions.

Change-ID: I08584f103a4fa73d6a4128d472e4ef44dcfda57f
Signed-off-by: Mitch Williams <mitch.a.williams@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agoi40e: Remove unused variable
Mitch Williams [Fri, 18 Mar 2016 19:18:06 +0000 (12:18 -0700)]
i40e: Remove unused variable

This variable is vestigial, a remnant of the primordial code from which
this driver spawned. We can safely remove it.

Change-ID: I24e0fe338e7c7c50d27dc5515564f33caefbb93a
Signed-off-by: Mitch Williams <mitch.a.williams@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agoi40e: Enable Geneve offload for FW API ver > 1.4 for XL710/X710 devices
Anjali Singhai Jain [Fri, 18 Mar 2016 19:18:05 +0000 (12:18 -0700)]
i40e: Enable Geneve offload for FW API ver > 1.4 for XL710/X710 devices

This patch enables the Capability for XL710/X710 devices with FW API
version higher than 1.4 to do geneve Rx offload.

Change-ID: I9a8f87772c48d7d67dc85e3701d2e0b845034c0b
Signed-off-by: Anjali Singhai Jain <anjali.singhai@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agoi40e: remove redundant check on vsi->active_vlans
Colin King [Sat, 13 Feb 2016 23:57:16 +0000 (23:57 +0000)]
i40e: remove redundant check on vsi->active_vlans

active_vlans is an unsigned long array, hence a null check on this
array is superfluous and can be removed.

Detected with static analysis by smatch:

drivers/net/ethernet/intel/i40e/i40e_debugfs.c:386
  i40e_dbg_dump_vsi_seid() warn: this array is probably
  non-NULL. 'vsi->active_vlans'

Signed-off-by: Colin Ian King <colin.king@canonical.com>
Acked-by: Shannon Nelson <shannon.nelson@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agonet/core/dev: Warn on a too-short GRO frame
Aaron Conole [Sat, 2 Apr 2016 19:26:43 +0000 (15:26 -0400)]
net/core/dev: Warn on a too-short GRO frame

When signaling that a GRO frame is ready to be processed, the network stack
correctly checks length and aborts processing when a frame is less than 14
bytes. However, such a condition is really indicative of a broken driver,
and should be loudly signaled, rather than silently dropped as the case is
today.

Convert the condition to use net_warn_ratelimited() to ensure the stack
loudly complains about such broken drivers.

Signed-off-by: Aaron Conole <aconole@bytheb.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoibmvnic: enable RX checksum offload
Thomas Falcon [Fri, 1 Apr 2016 22:20:35 +0000 (17:20 -0500)]
ibmvnic: enable RX checksum offload

Enable RX Checksum offload feature in the ibmvnic driver.

Signed-off-by: Thomas Falcon <tlfalcon@linux.vnet.ibm.com>
Cc: John Allen <jallen@linux.vnet.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoibmvnic: map L2/L3/L4 header descriptors to firmware
Thomas Falcon [Fri, 1 Apr 2016 22:20:34 +0000 (17:20 -0500)]
ibmvnic: map L2/L3/L4 header descriptors to firmware

Allow the VNIC driver to provide descriptors containing
L2/L3/L4 headers to firmware.  This feature is needed
for greater hardware compatibility and enablement of checksum
and TCP offloading features.

A new function is included for the hypervisor call,
H_SEND_SUBCRQ_INDIRECT, allowing a DMA-mapped array of SCRQ
descriptor elements to be sent to the VNIC server.

These additions will help fully enable checksum offloading as
well as other features as they are included later.

Signed-off-by: Thomas Falcon <tlfalcon@linux.vnet.ibm.com>
Cc: John Allen <jallen@linux.vnet.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoMerge branch 'bcmgenet-cleanups'
David S. Miller [Tue, 5 Apr 2016 23:33:52 +0000 (19:33 -0400)]
Merge branch 'bcmgenet-cleanups'

Petri Gynther says:

====================
bcmgenet cleanups

Three cleanup patches for bcmgenet.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agonet: bcmgenet: cleanup for dmadesc_set()
Petri Gynther [Tue, 5 Apr 2016 21:00:01 +0000 (14:00 -0700)]
net: bcmgenet: cleanup for dmadesc_set()

dmadesc_set() is used for setting the Tx buffer DMA address, length,
and status bits on a Tx ring descriptor when a frame is being Tx'ed.

Always set the Tx buffer DMA address first, before updating the length
and status bits, i.e. giving the Tx descriptor to the hardware.

The reason this is a cleanup rather than a fix is that the hardware
won't transmit anything from a Tx ring until the TDMA producer index
has been incremented. As long as the dmadesc_set() writes complete
before the TDMA producer index write, life is good.

Signed-off-by: Petri Gynther <pgynther@google.com>
Acked-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agonet: bcmgenet: cleanup for bcmgenet_xmit_frag()
Petri Gynther [Tue, 5 Apr 2016 21:00:00 +0000 (14:00 -0700)]
net: bcmgenet: cleanup for bcmgenet_xmit_frag()

Add frag_size = skb_frag_size(frag) and use it when needed.

Signed-off-by: Petri Gynther <pgynther@google.com>
Acked-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agonet: bcmgenet: cleanup for bcmgenet_xmit()
Petri Gynther [Tue, 5 Apr 2016 20:59:59 +0000 (13:59 -0700)]
net: bcmgenet: cleanup for bcmgenet_xmit()

1. Readability: Move nr_frags assignment a few lines down in order
   to bundle index -> ring -> txq calculations together.
2. Readability: Add parentheses around nr_frags + 1.
3. Minor fix: Stop the Tx queue and throw the error message only if
   the Tx queue hasn't already been stopped.

Signed-off-by: Petri Gynther <pgynther@google.com>
Acked-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoMerge branch 'udp-peek'
David S. Miller [Tue, 5 Apr 2016 20:29:38 +0000 (16:29 -0400)]
Merge branch 'udp-peek'

Willem de Bruijn says:

====================
udp: support SO_PEEK_OFF

Support peeking at a non-zero offset for UDP sockets. Match the
existing behavior on Unix datagram sockets.

1/3 makes the sk_peek_offset functions safe to use outside locks
2/3 removes udp headers before enqueue, to simplify offset arithmetic
3/3 introduces SO_PEEK_OFFSET support, with Unix socket peek semantics.

Changes
  v1->v2
    - squash patches 3 and 4
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoudp: enable MSG_PEEK at non-zero offset
samanthakumar [Tue, 5 Apr 2016 16:41:16 +0000 (12:41 -0400)]
udp: enable MSG_PEEK at non-zero offset

Enable peeking at UDP datagrams at the offset specified with socket
option SOL_SOCKET/SO_PEEK_OFF. Peek at any datagram in the queue, up
to the end of the given datagram.

Implement the SO_PEEK_OFF semantics introduced in commit ef64a54f6e55
("sock: Introduce the SO_PEEK_OFF sock option"). Increase the offset
on peek, decrease it on regular reads.

When peeking, always checksum the packet immediately, to avoid
recomputation on subsequent peeks and final read.

The socket lock is not held for the duration of udp_recvmsg, so
peek and read operations can run concurrently. Only the last store
to sk_peek_off is preserved.

Signed-off-by: Sam Kumar <samanthakumar@google.com>
Signed-off-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoudp: remove headers from UDP packets before queueing
samanthakumar [Tue, 5 Apr 2016 16:41:15 +0000 (12:41 -0400)]
udp: remove headers from UDP packets before queueing

Remove UDP transport headers before queueing packets for reception.
This change simplifies a follow-up patch to add MSG_PEEK support.

Signed-off-by: Sam Kumar <samanthakumar@google.com>
Signed-off-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agosock: convert sk_peek_offset functions to WRITE_ONCE
Willem de Bruijn [Tue, 5 Apr 2016 16:41:14 +0000 (12:41 -0400)]
sock: convert sk_peek_offset functions to WRITE_ONCE

Make the peek offset interface safe to use in lockless environments.
Use READ_ONCE and WRITE_ONCE to avoid race conditions between testing
and updating the peek offset.

Suggested-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Willem de Bruijn <willemb@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoMerge branch '40GbE' of git://git.kernel.org/pub/scm/linux/kernel/git/jkirsher/next...
David S. Miller [Tue, 5 Apr 2016 20:26:31 +0000 (16:26 -0400)]
Merge branch '40GbE' of git://git./linux/kernel/git/jkirsher/next-queue

Jeff Kirsher says:

====================
40GbE Intel Wired LAN Driver Updates 2016-04-05

This series contains updates to i40e and i40evf only.

Stefan converts dev_close() to ndo_stop() for ethtool offline self test,
since dev_close() causes IFF_UP to be cleared which will remove the
interface routes and addresses.

Alex bumps up the size of the transmit data buffer to 12K rather than 8K,
which provides a gain in throughput and a reduction in overhead for
putting together the frame.  Fixed an issue in the polling routines where
we were using bitwise operators to avoid the side effects of the
logical operators.  Then added support for bulk transmit clean for skbs.

Jesse fixed a sparse issue in the type casting in the transmit code and
fixed i40e_aq_set_phy_debug() to use i40e_status as a return code.

Catherine cleans up duplicated code.

Shannon fixed the cleaning up of the interrupt handling to clean up the
IRQs only if we actually got them set up.  Also fixed up the error
scenarios where we were trying to remove a non-existent timer or
worktask, which causes the kernel heartburn.

Mitch changes the notification of resets to the reset interrupt handler,
instead of the actual reset initiation code.  This allows the VFs to get
properly notified for all resets, including resets initiated by different
PFs on the same physical device.  Also moved the clearing of VFLR bit
after reset processing, instead of before which could lead to double
resets on VF init.  Fixed code comment to match the actual function name.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoMerge branch 'bnxt_en-next'
David S. Miller [Tue, 5 Apr 2016 20:20:49 +0000 (16:20 -0400)]
Merge branch 'bnxt_en-next'

Michael Chan says:

====================
bnxt_en: Updates for net-next.

Update to latest firmware interface, add EEE feature, unsupported SFP+
module warning, and ethtool -s improvements.

v2: Removed the GEEPROM patch and added more comments to the get_eee patch.
====================

8 years agobnxt_en: Improve ethtool .get_settings().
Michael Chan [Tue, 5 Apr 2016 18:09:03 +0000 (14:09 -0400)]
bnxt_en: Improve ethtool .get_settings().

If autoneg is off, we should always report the speed and duplex settings
even if it is link down so the user knows the current settings.  The
unknown speed and duplex should only be used for autoneg when link is
down.

Signed-off-by: Michael Chan <michael.chan@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agobnxt_en: Check for valid forced speed during ethtool -s.
Michael Chan [Tue, 5 Apr 2016 18:09:02 +0000 (14:09 -0400)]
bnxt_en: Check for valid forced speed during ethtool -s.

Check that the forced speed is a valid speed supported by firmware.
If not supported, return -EINVAL.

Signed-off-by: Michael Chan <michael.chan@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agobnxt_en: Add unsupported SFP+ module warnings.
Michael Chan [Tue, 5 Apr 2016 18:09:01 +0000 (14:09 -0400)]
bnxt_en: Add unsupported SFP+ module warnings.

Add the PORT_CONN_NOT_ALLOWED async event handling logic.  The driver
will print an appropriate warning to reflect the SFP+ module enforcement
policy done in the firmware.

Signed-off-by: Michael Chan <michael.chan@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agobnxt_en: Set async event bits when registering with the firmware.
Michael Chan [Tue, 5 Apr 2016 18:09:00 +0000 (14:09 -0400)]
bnxt_en: Set async event bits when registering with the firmware.

Currently, the driver only sets bit 0 of the async_event_fwd fields.
To be compatible with the latest spec, we need to set the
appropriate event bits handled by the driver.  We should be handling
link change and PF driver unload events, so these 2 bits should be
set.

Signed-off-by: Michael Chan <michael.chan@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agobnxt_en: Add get_eee() and set_eee() ethtool support.
Michael Chan [Tue, 5 Apr 2016 18:08:59 +0000 (14:08 -0400)]
bnxt_en: Add get_eee() and set_eee() ethtool support.

Allow users to get|set EEE parameters.

v2: Added comment for preserving the tx_lpi_timer value in get_eee.

Signed-off-by: Michael Chan <michael.chan@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agobnxt_en: Add EEE setup code.
Michael Chan [Tue, 5 Apr 2016 18:08:58 +0000 (14:08 -0400)]
bnxt_en: Add EEE setup code.

1. Add bnxt_hwrm_set_eee() function to setup EEE firmware parameters based
on the bp->eee settings.
2. The new function bnxt_eee_config_ok() will check if EEE parameters need
to be modified due to autoneg changes.
3. bnxt_hwrm_set_link() has added a new parameter to update EEE.  If the
parameter is set, it will call bnxt_hwrm_set_eee().

Signed-off-by: Michael Chan <michael.chan@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agobnxt_en: Add basic EEE support.
Michael Chan [Tue, 5 Apr 2016 18:08:57 +0000 (14:08 -0400)]
bnxt_en: Add basic EEE support.

Get EEE capability and the initial EEE settings from firmware.
Add "EEE is active | not active" to link up dmesg.

Signed-off-by: Michael Chan <michael.chan@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agobnxt_en: Improve flow control autoneg with Firmware 1.2.1 interface.
Michael Chan [Tue, 5 Apr 2016 18:08:56 +0000 (14:08 -0400)]
bnxt_en: Improve flow control autoneg with Firmware 1.2.1 interface.

Make use of the new AUTONEG_PAUSE bit in the new interface to better
control autoneg flow control settings, independent of RX and TX
advertisement settings.

Signed-off-by: Michael Chan <michael.chan@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agobnxt_en: Update to Firmware 1.2.2 spec.
Michael Chan [Tue, 5 Apr 2016 18:08:55 +0000 (14:08 -0400)]
bnxt_en: Update to Firmware 1.2.2 spec.

Use new field names in API structs and stop using deprecated fields
auto_link_speed and auto_duplex in phy_cfg/phy_qcfg structs.

Update copyright year to 2016.

Signed-off-by: Michael Chan <michael.chan@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoMerge branch '100GbE' of git://git.kernel.org/pub/scm/linux/kernel/git/jkirsher/next...
David S. Miller [Tue, 5 Apr 2016 20:08:02 +0000 (16:08 -0400)]
Merge branch '100GbE' of git://git./linux/kernel/git/jkirsher/next-queue

Jeff Kirsher says:

====================
100GbE Intel Wired LAN Driver Updates 2016-04-05

This series contains updates to fm10k only.

Bruce provides nearly half of the patches in the series, most of which do
general cleanup of the driver.  These include semantic cleanups,
checkpatch.pl fixes, update driver to use BIT() kernel macro, use
BUILD_BUG_ON() where appropriate and use ether_addr_copy() instead of
memcpy().

Jake provides the remaining patches in the series, starting with a fix
for a possible NULL pointer deference.  Next delays initialization of the
service timer and service task until late in probe().  If we do not wait,
failures in probe do not properly cleanup the service timer or service
task items which result in a kernel panic.  Added better reporting during
error conditions.  Fixed another possible kernel panic where we were
clearing the interrupt scheme before we freed the mailbox IRQ.  Added
helper functions for setting strings and data for ethtool stats.  Fixed
comment mis-spelled words.

v2: Dropped patch 3 from the original submission, until a better solution
    can be worked up based on feedback from Joe Perches and David Miller.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agofm10k: use ethtool_rxfh_indir_default for default redirection table
Jacob Keller [Wed, 17 Feb 2016 00:19:24 +0000 (16:19 -0800)]
fm10k: use ethtool_rxfh_indir_default for default redirection table

The fm10k driver used its own code for generating a default indirection
table on device load, which was not the same as the default generated by
ethtool when indir_size of 0 is passed to SRXFH. Take advantage of
ethtool_rxfh_indir_default() and simplify code to write the redirection
table to reduce some code duplication.

Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Tested-by: Krishneil Singh <Krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agofm10k: fix a minor typo in some comments
Jacob Keller [Wed, 10 Feb 2016 22:45:51 +0000 (14:45 -0800)]
fm10k: fix a minor typo in some comments

s/funciton/function to resolve a typo, and cleanup grammar on a few
comments regarding processing the VF mailboxes.

Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Tested-by: Krishneil Singh <Krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agofm10k: correctly clean up when init_queueing_scheme fails
Jacob Keller [Wed, 10 Feb 2016 22:45:50 +0000 (14:45 -0800)]
fm10k: correctly clean up when init_queueing_scheme fails

Fix a kernel panic that occurs during surprise removal. Clear the
interface queue counts upon fm10k_init_msix_capability failure. This
prevents further code (fm10k_update_stats etc.) from attempting to
access unallocated queue vector or ring memory.

[  628.692648] BUG: unable to handle kernel NULL pointer dereference at 0000000000000068
[  628.692805] IP: [<ffffffffa0475caf>] fm10k_update_stats+0x7f/0x2c0 [fm10k]
[  628.693173] PGD 0
[  628.693759] Oops: 0000 [#1] SMP
[  628.699321] CPU: 10 PID: 8164 Comm: kworker/10:0 Tainted: G           OE  ------------   3.10.0-327.el7.x86_64 #1
[  628.700096] Hardware name: Supermicro X9DAi/X9DAi, BIOS 3.2 05/09/2015
[  628.700894] Workqueue: pciehp-1 pciehp_power_thread
[  628.701686] task: ffff88086559c500 ti: ffff8808593c0000 task.ti: ffff8808593c0000
[  628.702493] RIP: 0010:[<ffffffffa0475caf>]  [<ffffffffa0475caf>] fm10k_update_stats+0x7f/0x2c0 [fm10k]
[  628.703310] RSP: 0018:ffff8808593c3b00  EFLAGS: 00010282
[  628.704132] RAX: 0000000000000000 RBX: ffff880860760000 RCX: 0000000000000000
[  628.704963] RDX: ffff880860760b08 RSI: 0000000000000000 RDI: 0000000000000000
[  628.705794] RBP: ffff8808593c3b40 R08: 0000000000000000 R09: 0000000000000000
[  628.706604] R10: 0000000000000000 R11: ffff880860760c40 R12: 0000000000000080
[  628.707420] R13: ffff8808607608c0 R14: ffff880860779ec0 R15: ffff880860779f40
[  628.708238] FS:  0000000000000000(0000) GS:ffff88086f000000(0000) knlGS:0000000000000000
[  628.709071] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[  628.709923] CR2: 0000000000000068 CR3: 000000000194a000 CR4: 00000000001407e0
[  628.710752] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
[  628.711596] DR3: 0000000000000000 DR6: 00000000ffff0ff0 DR7: 0000000000000400
[  628.712438] Stack:
[  628.713255]  ffff880860764458 ffff8808607608c0 ffff880860760000 ffff880860760000
[  628.714088]  0000000000000080 ffff8808607608c0 ffff880860779ec0 ffff880860779f40
[  628.714925]  ffff8808593c3b88 ffffffffa04780c5 ffff880860764458 0000000a8163cb5b
[  628.715752] Call Trace:
[  628.716560]  [<ffffffffa04780c5>] fm10k_down+0x155/0x1f0 [fm10k]
[  628.717367]  [<ffffffffa0479958>] fm10k_close+0x28/0xd0 [fm10k]
[  628.718184]  [<ffffffff81526365>] __dev_close_many+0x85/0xd0
[  628.718986]  [<ffffffff815264d8>] dev_close_many+0x98/0x120
[  628.719764]  [<ffffffff81527ab8>] rollback_registered_many+0xa8/0x230
[  628.720527]  [<ffffffff81527c80>] rollback_registered+0x40/0x70
[  628.721294]  [<ffffffff81529198>] unregister_netdevice_queue+0x48/0x80
[  628.722052]  [<ffffffff815291ec>] unregister_netdev+0x1c/0x30
[  628.722816]  [<ffffffffa04762b8>] fm10k_remove+0xd8/0xe0 [fm10k]
[  628.723581]  [<ffffffff81328c7b>] pci_device_remove+0x3b/0xb0
[  628.724340]  [<ffffffff813f5fbf>] __device_release_driver+0x7f/0xf0
[  628.725088]  [<ffffffff813f6053>] device_release_driver+0x23/0x30
[  628.725814]  [<ffffffff81321fe4>] pci_stop_bus_device+0x94/0xa0
[  628.726535]  [<ffffffff813220d2>] pci_stop_and_remove_bus_device+0x12/0x20
[  628.727249]  [<ffffffff8133de40>] pciehp_unconfigure_device+0xb0/0x1b0
[  628.727964]  [<ffffffff8133d822>] pciehp_disable_slot+0x52/0xd0
[  628.728664]  [<ffffffff8133d98a>] pciehp_power_thread+0xea/0x150
[  628.729358]  [<ffffffff8109d5fb>] process_one_work+0x17b/0x470
[  628.730036]  [<ffffffff8109e3cb>] worker_thread+0x11b/0x400
[  628.730730]  [<ffffffff8109e2b0>] ? rescuer_thread+0x400/0x400
[  628.731385]  [<ffffffff810a5aef>] kthread+0xcf/0xe0
[  628.732036]  [<ffffffff810a5a20>] ? kthread_create_on_node+0x140/0x140
[  628.732674]  [<ffffffff81645858>] ret_from_fork+0x58/0x90
[  628.733289]  [<ffffffff810a5a20>] ? kthread_create_on_node+0x140/0x140
[  628.733883] Code: 83 e8 01 48 8d 97 40 02 00 00 45 31 c0 4c 8d 9c c7 48 02 0
[  628.735202] RIP  [<ffffffffa0475caf>] fm10k_update_stats+0x7f/0x2c0 [fm10k]
[  628.735732]  RSP <ffff8808593c3b00>
[  628.736285] CR2: 0000000000000068
[  628.736846] ---[ end trace 9156088b311aff42 ]---

Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Tested-by: Krishneil Singh <Krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agofm10k: prevent possibly uninitialized variable
Bruce Allan [Wed, 10 Feb 2016 22:45:47 +0000 (14:45 -0800)]
fm10k: prevent possibly uninitialized variable

If 'attr_flag < (1 << (2 * FM10K_TEST_MSG_NESTED))' is ever false, err
will be used uninitialized.

Signed-off-by: Bruce Allan <bruce.w.allan@intel.com>
Tested-by: Krishneil Singh <Krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agofm10k: add helper functions to set strings and data for ethtool stats
Jacob Keller [Fri, 5 Feb 2016 18:43:08 +0000 (10:43 -0800)]
fm10k: add helper functions to set strings and data for ethtool stats

Reduce duplicate code and the amount of indentation by adding
fm10k_add_stat_strings and fm10k_add_ethtool_stats functions which help
add fm10k_stat structures to the ethtool stats callbacks. This helps
increase ease of use for future stat additions, and increases code
readability. Skip handling of the per-queue stats as these will be
reworked in a following patch.

Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Tested-by: Krishneil Singh <Krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agofm10k: free MBX IRQ before clearing interrupt scheme
Jacob Keller [Thu, 4 Feb 2016 18:47:58 +0000 (10:47 -0800)]
fm10k: free MBX IRQ before clearing interrupt scheme

During fm10k_io_error_detected we were clearing the interrupt scheme
before we freed the MBX IRQ. This causes a kernel panic because the MBX
IRQ are assigned after MSI-X initialization. Clearing the interrupt
scheme results in removing the MSI-X entry table. Fix this by freeing
the MBX IRQ before we clear the interrupt scheme, as we do elsewhere in
the driver.

Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Tested-by: Krishneil Singh <Krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agofm10k: print error message when stop_hw fails
Jacob Keller [Thu, 4 Feb 2016 18:47:57 +0000 (10:47 -0800)]
fm10k: print error message when stop_hw fails

fm10k_stop_hw_generic calls fm10k_disable_queues_generic, which may
return an error code indicating that the queues were not stopped within
the time limit. Notify the user by displaying a message in the kernel
message ring, in a similar way to how we notify the user when reset_hw
fails. There isn't much we can do to recover from this error, so
currently nothing else is done.

Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Tested-by: Krishneil Singh <Krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agofm10k: base queue scheme covered by RSS
Jacob Keller [Thu, 4 Feb 2016 18:47:56 +0000 (10:47 -0800)]
fm10k: base queue scheme covered by RSS

In fm10k_set_num_queues, we previously assigned the base template. This
would always be overwritten by either fm10k_set_qos_queues or
fm10k_set_rss_queues. In either case, we don't need the base values, so
we can just remove them.

Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Tested-by: Krishneil Singh <Krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agofm10k: don't initialize service task until later in probe
Jacob Keller [Thu, 4 Feb 2016 18:47:55 +0000 (10:47 -0800)]
fm10k: don't initialize service task until later in probe

Delay initialization of the service timer and service task until late
probe. If we don't wait, failures in probe do not properly cleanup the
service timer or service task items, which results in the kernel panic
below, potentially freezing the whole system. In addition, ensure that
the SERVICE_DISABLE bit is set before we request the MBX IRQ since the
MBX interrupt attempts to schedule the service task otherwise. This
prevents a similar trace from occurring after this change.

We didn't notice this issue before because probe almost always completes
successfully. I discovered it due to a mis-ordered mailbox handler
array, which resulted in the following failure when requesting mailbox
interrupt.

[  555.325619] ------------[ cut here ]------------
[  555.325628] WARNING: CPU: 0 PID: 4941 at lib/list_debug.c:33 __list_add+0xa0/0xd0()
[  555.325631] list_add corruption. prev->next should be next (ffffffff81f46648), but was           (null). (prev=ffff8807fad5d0e8).
<snip>
[  555.325722] CPU: 0 PID: 4941 Comm: insmod Tainted: G           OE   4.0.4-303.fc22.x86_64 #1
[  555.325725] Hardware name: Intel Corporation S2600CO/S2600CO, BIOS SE5C600.86B.02.03.8x23.060520140825 06/05/2014
[  555.325727]  0000000000000000 00000000b4f161b3 ffff88081a21f8e8 ffffffff81783124
[  555.325734]  0000000000000000 ffff88081a21f940 ffff88081a21f928 ffffffff8109c66a
[  555.325740]  0000000064000000 ffff8807fad5d0e8 ffff8807fad5d0e8 ffffffff81f46648
[  555.325746] Call Trace:
[  555.325752]  [<ffffffff81783124>] dump_stack+0x45/0x57
[  555.325757]  [<ffffffff8109c66a>] warn_slowpath_common+0x8a/0xc0
[  555.325759]  [<ffffffff8109c6f5>] warn_slowpath_fmt+0x55/0x70
[  555.325763]  [<ffffffff813ba270>] __list_add+0xa0/0xd0
[  555.325768]  [<ffffffff81102d1d>] __internal_add_timer+0x9d/0x110
[  555.325771]  [<ffffffff81102dbf>] internal_add_timer+0x2f/0xc0
[  555.325774]  [<ffffffff81104e5a>] mod_timer+0x12a/0x230
[  555.325782]  [<ffffffffa03d54ca>] fm10k_probe+0x69a/0xc80 [fm10k]
[  555.325787]  [<ffffffff813e8355>] local_pci_probe+0x45/0xa0
[  555.325791]  [<ffffffff8129cf42>] ? sysfs_do_create_link_sd.isra.2+0x72/0xc0
[  555.325794]  [<ffffffff813e96b9>] pci_device_probe+0xf9/0x150
[  555.325799]  [<ffffffff814d7e73>] driver_probe_device+0xa3/0x400
[  555.325802]  [<ffffffff814d82ab>] __driver_attach+0x9b/0xa0
[  555.325805]  [<ffffffff814d8210>] ? __device_attach+0x40/0x40
[  555.325808]  [<ffffffff814d5bd3>] bus_for_each_dev+0x73/0xc0
[  555.325811]  [<ffffffff814d78ce>] driver_attach+0x1e/0x20
[  555.325815]  [<ffffffff814d7480>] bus_add_driver+0x180/0x250
[  555.325819]  [<ffffffffa03b2000>] ? 0xffffffffa03b2000
[  555.325823]  [<ffffffff814d8aa4>] driver_register+0x64/0xf0
[  555.325826]  [<ffffffff813e7bec>] __pci_register_driver+0x4c/0x50
[  555.325832]  [<ffffffffa03d6ca3>] fm10k_register_pci_driver+0x23/0x30 [fm10k]
[  555.325838]  [<ffffffffa03b2080>] fm10k_init_module+0x80/0x1000 [fm10k]
[  555.325843]  [<ffffffff81002128>] do_one_initcall+0xb8/0x200
[  555.325848]  [<ffffffff811e10d2>] ? __vunmap+0xa2/0x100
[  555.325852]  [<ffffffff811fe239>] ? kmem_cache_alloc_trace+0x1b9/0x240
[  555.325855]  [<ffffffff8178230e>] ? do_init_module+0x28/0x1cb
[  555.325858]  [<ffffffff81782346>] do_init_module+0x60/0x1cb
[  555.325862]  [<ffffffff8112168e>] load_module+0x205e/0x26b0
[  555.325866]  [<ffffffff8111d110>] ? store_uevent+0x70/0x70
[  555.325870]  [<ffffffff812234b0>] ? kernel_read+0x50/0x80
[  555.325873]  [<ffffffff81121f3e>] SyS_finit_module+0xbe/0xf0
[  555.325878]  [<ffffffff81789749>] system_call_fastpath+0x12/0x17
[  555.325880] ---[ end trace 9e0f58d071eafd2a ]---

Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Tested-by: Krishneil Singh <Krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agofm10k: prevent null pointer dereference of msix_entries table
Jacob Keller [Thu, 4 Feb 2016 18:47:54 +0000 (10:47 -0800)]
fm10k: prevent null pointer dereference of msix_entries table

According to the C standard dereferencing a variable before it is
checked invokes undefined behavior, and thus compilers are free to
assume the check for NULL isn't necessary. Prevent this by re-ordering
the NULL check of msix_entries in fm10k_free_mbx_irq.

Signed-off-by: Jacob Keller <jacob.e.keller@intel.com>
Tested-by: Krishneil Singh <Krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agofm10k: use ether_addr_copy to copy MAC address
Bruce Allan [Tue, 29 Dec 2015 02:00:30 +0000 (18:00 -0800)]
fm10k: use ether_addr_copy to copy MAC address

Cleanup the remaining instances of using memcpy() instead of the preferred
ether_addr_copy().

Signed-off-by: Bruce Allan <bruce.w.allan@intel.com>
Tested-by: Krishneil Singh <Krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agofm10k: cleanup SPACE_BEFORE_TAB checkpatch warning
Bruce Allan [Tue, 22 Dec 2015 22:55:26 +0000 (14:55 -0800)]
fm10k: cleanup SPACE_BEFORE_TAB checkpatch warning

Signed-off-by: Bruce Allan <bruce.w.allan@intel.com>
Tested-by: Krishneil Singh <Krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agofm10k: demote BUG_ON() to WARN_ON() where appropriate
Bruce Allan [Tue, 22 Dec 2015 22:55:20 +0000 (14:55 -0800)]
fm10k: demote BUG_ON() to WARN_ON() where appropriate

We don't need to crash the kernel in this instance so just warn about the
condition and play on.

Signed-off-by: Bruce Allan <bruce.w.allan@intel.com>
Tested-by: Krishneil Singh <Krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agofm10k: cleanup remaining right-bit-shifted 1
Bruce Allan [Tue, 22 Dec 2015 21:43:49 +0000 (13:43 -0800)]
fm10k: cleanup remaining right-bit-shifted 1

Use BIT() macro instead.

Signed-off-by: Bruce Allan <bruce.w.allan@intel.com>
Tested-by: Krishneil Singh <Krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agofm10k: Move constants to the right of binary operators
Bruce Allan [Tue, 22 Dec 2015 21:43:44 +0000 (13:43 -0800)]
fm10k: Move constants to the right of binary operators

The semantic patch that makes this change is available
in scripts/coccinelle/misc/compare_const_fl.cocci.

More information about semantic patching is available at
http://coccinelle.lip6.fr/

Signed-off-by: Bruce Allan <bruce.w.allan@intel.com>
Tested-by: Krishneil Singh <Krishneil.k.singh@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agoi40e/i40evf: Bump patch from 1.4.25 to 1.5.1
Catherine Sullivan [Thu, 10 Mar 2016 22:59:51 +0000 (14:59 -0800)]
i40e/i40evf: Bump patch from 1.4.25 to 1.5.1

Signed-off-by: Catherine Sullivan <catherine.sullivan@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agoi40e: Change comment to reflect correct function name
Mitch Williams [Thu, 10 Mar 2016 22:59:50 +0000 (14:59 -0800)]
i40e: Change comment to reflect correct function name

Minor correction in the comment to reflect the correct function name

Signed-off-by: Mitch Williams <mitch.a.williams@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agoi40evf: Add additional check for reset
Mitch Williams [Thu, 10 Mar 2016 22:59:49 +0000 (14:59 -0800)]
i40evf: Add additional check for reset

If the driver happens to read a register during the time in which the
device is undergoing reset, it will receive a value of 0xdeadbeef
instead of a valid value. Unfortunately, the driver may misinterpret
this as a valid value, especially if it's just looking for individual
bits.

Add an explicit check for this value when we are looking for admin queue
errors, and trigger reset recovery if we find it.

Signed-off-by: Mitch Williams <mitch.a.williams@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agoi40e: Change unknown event error msg to ignore message
Shannon Nelson [Thu, 10 Mar 2016 22:59:48 +0000 (14:59 -0800)]
i40e: Change unknown event error msg to ignore message

There's no real error in an unknown event from the Firmware, we're just
posting a useful FYI notice, so this patch simply removes the "Error" word.

Signed-off-by: Shannon Nelson <shannon.nelson@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agoi40e: Added code to prevent double resets
Mitch Williams [Thu, 10 Mar 2016 22:59:47 +0000 (14:59 -0800)]
i40e: Added code to prevent double resets

Clear the VFLR bit after reset processing, instead of before. This
prevents double resets on VF init.

Signed-off-by: Mitch Williams <mitch.a.williams@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agoi40e: Notify VFs of all resets
Mitch Williams [Thu, 10 Mar 2016 22:59:46 +0000 (14:59 -0800)]
i40e: Notify VFs of all resets

Notify VFs in the reset interrupt handler, instead of the actual
reset initiation code. This allows the VFs to get properly notified for
all resets, including resets initiated by different PFs on the same
physical device.

Signed-off-by: Mitch Williams <mitch.a.williams@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agoi40e: Remove timer and task only if created
Shannon Nelson [Thu, 10 Mar 2016 22:59:45 +0000 (14:59 -0800)]
i40e: Remove timer and task only if created

In some error scenarios, we may find ourselves trying to remove a
non-existent timer or worktask.  This causes the kernel some bit
of consternation, so don't do it.

Signed-off-by: Shannon Nelson <shannon.nelson@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agoMerge branch 'mlxsw-next'
David S. Miller [Tue, 5 Apr 2016 19:07:55 +0000 (15:07 -0400)]
Merge branch 'mlxsw-next'

Jiri Pirko says:

====================
mlxsw: small driver update, including switchdev doc update

Ido Schimmel (3):
  mlxsw: spectrum: Reduce number of supported 802.1D bridges
  switchdev: Use switch ID in suggested udev rule
  mlxsw: spectrum: Add support for physical port names
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agomlxsw: spectrum: Add support for physical port names
Ido Schimmel [Tue, 5 Apr 2016 08:20:04 +0000 (10:20 +0200)]
mlxsw: spectrum: Add support for physical port names

Export to userspace the front panel name of the port, so that udev can
rename the ports accordingly. The convention suggested by switchdev
documentation is used:

1) Non-split: pX
2) Split: pXsY

Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoswitchdev: Use switch ID in suggested udev rule
Ido Schimmel [Tue, 5 Apr 2016 08:20:03 +0000 (10:20 +0200)]
switchdev: Use switch ID in suggested udev rule

Since there can be multiple switch ASICs on the same system we should
use the switch ID in order to differentiate between them and set the
switch name (e.g. swX) accordingly.

Also, replace the order of the "Switch ID" and "Port Netdev Naming"
sections following the above change.

Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agomlxsw: spectrum: Reduce number of supported 802.1D bridges
Ido Schimmel [Tue, 5 Apr 2016 08:20:02 +0000 (10:20 +0200)]
mlxsw: spectrum: Reduce number of supported 802.1D bridges

Resources allocated for these bridges at init time cannot be later used
for other purposes. While current number is supported by the device,
it's mostly theoretical with regards to any real use case, which leads
to poor utilization of device's resources. Solve that by reducing the
number.

The long term plan is to make this value (along with others) user
configurable via devlink and write it to NVRAM, so that it can be used
during the next init. Until then we must hardcode such values.

Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoi40e: Assure that adminq is alive in debug mode
Shannon Nelson [Thu, 10 Mar 2016 22:59:44 +0000 (14:59 -0800)]
i40e: Assure that adminq is alive in debug mode

When dropping into debug mode in a failed probe, make sure that
the AdminQ is left alive for possible hand debug of driver and
firmware states.

Move the mutex_init calls earlier in probe so that if init fails,
the admin queue interface is still available for debugging purposes.

Signed-off-by: Shannon Nelson <shannon.nelson@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agoi40e: Remove MSIx only if created
Shannon Nelson [Thu, 10 Mar 2016 22:59:43 +0000 (14:59 -0800)]
i40e: Remove MSIx only if created

When cleaning up the interrupt handling, clean up the IRQs only if
we actually got them set up.  There are a couple of error recovery
paths that were violating this and causing the kernel a bit of
indigestion.

Signed-off-by: Shannon Nelson <shannon.nelson@intel.com>
Reviewed-by: Williams, Mitch A <mitch.a.williams@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agoi40e: Fix up return code
Jesse Brandeburg [Thu, 10 Mar 2016 22:59:42 +0000 (14:59 -0800)]
i40e: Fix up return code

The i40e_common.c typically uses i40e_status as a return code,
but got missed this one case.

Signed-off-by: Jesse Brandeburg <jesse.brandeburg@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agoi40e: Save off VSI resource count when updating VSI
Kevin Scott [Thu, 10 Mar 2016 22:59:41 +0000 (14:59 -0800)]
i40e: Save off VSI resource count when updating VSI

When updating a VSI, save off the number of allocated and unallocated
VSIs as we do when adding a VSI.

Signed-off-by: Kevin Scott <kevin.c.scott@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agoi40e/i40evf: Remove I40E_MAX_USER_PRIORITY define
Catherine Sullivan [Thu, 10 Mar 2016 22:59:40 +0000 (14:59 -0800)]
i40e/i40evf: Remove I40E_MAX_USER_PRIORITY define

This patch removes the duplicate definition of I40E_MAX_USER_PRIORITY
in i40e.h that is not needed.

Signed-off-by: Catherine Sullivan <catherine.sullivan@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agoi40e/i40evf: Fix casting in transmit code
Jesse Brandeburg [Thu, 10 Mar 2016 22:59:39 +0000 (14:59 -0800)]
i40e/i40evf: Fix casting in transmit code

Simple cast to fix a sparse warning.

Fixes: commit 5453205cd097 ("i40e/i40evf: Enable support for
SKB_GSO_UDP_TUNNEL_CSUM")

Signed-off-by: Jesse Brandeburg <jesse.brandeburg@intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agoi40e/i40evf: Add support for bulk free in Tx cleanup
Alexander Duyck [Mon, 7 Mar 2016 17:30:03 +0000 (09:30 -0800)]
i40e/i40evf: Add support for bulk free in Tx cleanup

This patch enables bulk Tx clean for skbs.  In order to enable it we need
to pass the napi_budget value as that is used to determine if we are truly
running in NAPI mode or if we are simply calling the routine from netpoll
with a budget of 0.  In order to avoid adding too many more variables I
thought it best to pass the VSI directly in a fashion similar to what we do
on igb and ixgbe with the q_vector.

Signed-off-by: Alexander Duyck <aduyck@mirantis.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agoi40e/i40evf: Fix handling of boolean logic in polling routines
Alexander Duyck [Mon, 7 Mar 2016 17:29:57 +0000 (09:29 -0800)]
i40e/i40evf: Fix handling of boolean logic in polling routines

In the polling routines for i40e and i40evf we were using bitwise operators
to avoid the side effects of the logical operators, specifically the fact
that if the first case is true with "||" we skip the second case, or if it
is false with "&&" we skip the second case.  This fixes an earlier patch
that converted the bitwise operators over to the logical operators and
instead replaces the entire thing with just an if statement since it should
be more readable what we are trying to do this way.

Fixes: 1a36d7fadd14 ("i40e/i40evf: use logical operators, not bitwise")
Signed-off-by: Alexander Duyck <aduyck@mirantis.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agoi40evf: remove dead code
Alan Cox [Wed, 2 Mar 2016 00:02:15 +0000 (16:02 -0800)]
i40evf: remove dead code

The only error case is when the malloc fails, in which case the clean up
loop does nothing at all, so remove it

Signed-off-by: Alan Cox <alan@linux.intel.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agoi40e/i40evf: Allow up to 12K bytes of data per Tx descriptor instead of 8K
Alexander Duyck [Fri, 19 Feb 2016 20:17:08 +0000 (12:17 -0800)]
i40e/i40evf: Allow up to 12K bytes of data per Tx descriptor instead of 8K

From what I can tell the practical limitation on the size of the Tx data
buffer is the fact that the Tx descriptor is limited to 14 bits.  As such
we cannot use 16K as is typically used on the other Intel drivers.  However
artificially limiting ourselves to 8K can be expensive as this means that
we will consume up to 10 descriptors (1 context, 1 for header, and 9 for
payload, non-8K aligned) in a single send.

I propose that we can reduce this by increasing the maximum data for a 4K
aligned block to 12K.  We can reduce the descriptors used for a 32K aligned
block by 1 by increasing the size like this.  In addition we still have the
4K - 1 of space that is still unused.  We can use this as a bit of extra
padding when dealing with data that is not aligned to 4K.

By aligning the descriptors after the first to 4K we can improve the
efficiency of PCIe accesses as we can avoid using byte enables and can fetch
full TLP transactions after the first fetch of the buffer.  This helps to
improve PCIe efficiency.  Below is the results of testing before and after
with this patch:

Recv   Send   Send                         Utilization      Service Demand
Socket Socket Message  Elapsed             Send     Recv    Send    Recv
Size   Size   Size     Time    Throughput  local    remote  local   remote
bytes  bytes  bytes    secs.   10^6bits/s  % S      % U     us/KB   us/KB
Before:
87380  16384  16384    10.00     33682.24  20.27    -1.00   0.592   -1.00
After:
87380  16384  16384    10.00     34204.08  20.54    -1.00   0.590   -1.00

So the net result of this patch is that we have a small gain in throughput
due to a reduction in overhead for putting together the frame.

Signed-off-by: Alexander Duyck <aduyck@mirantis.com>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agoi40e: call ndo_stop() instead of dev_close() when running offline selftest
Stefan Assmann [Wed, 3 Feb 2016 08:20:47 +0000 (09:20 +0100)]
i40e: call ndo_stop() instead of dev_close() when running offline selftest

Calling dev_close() causes IFF_UP to be cleared which will remove the
interfaces routes and some addresses. That's probably not what the user
intended when running the offline selftest. Besides this does not happen
if the interface is brought down before the test, so the current
behaviour is inconsistent.
Instead call the net_device_ops ndo_stop function directly and avoid
touching IFF_UP at all.

Signed-off-by: Stefan Assmann <sassmann@kpanic.de>
Tested-by: Andrew Bowers <andrewx.bowers@intel.com>
Signed-off-by: Jeff Kirsher <jeffrey.t.kirsher@intel.com>
8 years agoMerge branch 'tcp-udp-misc'
David S. Miller [Tue, 5 Apr 2016 02:11:21 +0000 (22:11 -0400)]
Merge branch 'tcp-udp-misc'

Eric Dumazet says:

====================
net: various udp/tcp changes

First round of patches for linux-4.7

Add a generic facility for sockets to be freed after an RCU grace
period, if they need to.

Then UDP stack is changed to no longer use SLAB_DESTROY_BY_RCU,
in order to speedup rx processing for traffic encapsulated in UDP.
It gives a 17 % speedup for normal UDP reception in stress conditions.

Then TCP listeners are changed to use SOCK_RCU_FREE as well
to avoid touching sk_refcnt in synflood case :
I got up to 30 % performance increase for a mono listener.

Then three patches add SK_MEMINFO_DROPS to sock_diag
and add per socket rx drops accounting to TCP.

Last patch adds rate limiting on ACK sent on behalf of SYN_RECV
to better resist to SYNFLOOD targeting one or few flows.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agotcp: rate limit ACK sent by SYN_RECV request sockets
Eric Dumazet [Fri, 1 Apr 2016 15:52:22 +0000 (08:52 -0700)]
tcp: rate limit ACK sent by SYN_RECV request sockets

Attackers like to use SYNFLOOD targeting one 5-tuple, as they
hit a single RX queue (and cpu) on the victim.

If they use random sequence numbers in their SYN, we detect
they do not match the expected window and send back an ACK.

This patch adds a rate limitation, so that the effect of such
attacks is limited to ingress only.

We roughly double our ability to absorb such attacks.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Cc: Willem de Bruijn <willemb@google.com>
Cc: Neal Cardwell <ncardwell@google.com>
Cc: Maciej Żenczykowski <maze@google.com>
Acked-by: Neal Cardwell <ncardwell@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoipv4: tcp: set SOCK_USE_WRITE_QUEUE for ip_send_unicast_reply()
Eric Dumazet [Fri, 1 Apr 2016 15:52:21 +0000 (08:52 -0700)]
ipv4: tcp: set SOCK_USE_WRITE_QUEUE for ip_send_unicast_reply()

TCP uses per cpu 'sockets' to send some packets :
- RST packets ( tcp_v4_send_reset()) )
- ACK packets for SYN_RECV and TIMEWAIT sockets

By setting SOCK_USE_WRITE_QUEUE flag, we tell sock_wfree()
to not call sk_write_space() since these internal sockets
do not care.

This gives a small performance improvement, merely by allowing
cpu to properly predict the sock_wfree() conditional branch,
and avoiding one atomic operation.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agotcp: increment sk_drops for listeners
Eric Dumazet [Fri, 1 Apr 2016 15:52:20 +0000 (08:52 -0700)]
tcp: increment sk_drops for listeners

Goal: packets dropped by a listener are accounted for.

This adds tcp_listendrop() helper, and clears sk_drops in sk_clone_lock()
so that children do not inherit their parent drop count.

Note that we no longer increment LINUX_MIB_LISTENDROPS counter when
sending a SYNCOOKIE, since the SYN packet generated a SYNACK.
We already have a separate LINUX_MIB_SYNCOOKIESSENT

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agotcp: increment sk_drops for dropped rx packets
Eric Dumazet [Fri, 1 Apr 2016 15:52:19 +0000 (08:52 -0700)]
tcp: increment sk_drops for dropped rx packets

Now ss can report sk_drops, we can instruct TCP to increment
this per socket counter when it drops an incoming frame, to refine
monitoring and debugging.

Following patch takes care of listeners drops.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agosock_diag: add SK_MEMINFO_DROPS
Eric Dumazet [Fri, 1 Apr 2016 15:52:18 +0000 (08:52 -0700)]
sock_diag: add SK_MEMINFO_DROPS

Reporting sk_drops to user space was available for UDP
sockets using /proc interface.

Add this to sock_diag, so that we can have the same information
available to ss users, and we'll be able to add sk_drops
indications for TCP sockets as well.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agotcp/dccp: do not touch listener sk_refcnt under synflood
Eric Dumazet [Fri, 1 Apr 2016 15:52:17 +0000 (08:52 -0700)]
tcp/dccp: do not touch listener sk_refcnt under synflood

When a SYNFLOOD targets a non SO_REUSEPORT listener, multiple
cpus contend on sk->sk_refcnt and sk->sk_wmem_alloc changes.

By letting listeners use SOCK_RCU_FREE infrastructure,
we can relax TCP_LISTEN lookup rules and avoid touching sk_refcnt

Note that we still use SLAB_DESTROY_BY_RCU rules for other sockets,
only listeners are impacted by this change.

Peak performance under SYNFLOOD is increased by ~33% :

On my test machine, I could process 3.2 Mpps instead of 2.4 Mpps

Most consuming functions are now skb_set_owner_w() and sock_wfree()
contending on sk->sk_wmem_alloc when cooking SYNACK and freeing them.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
8 years agoinet: reqsk_alloc() needs to take care of dead listeners
Eric Dumazet [Fri, 1 Apr 2016 15:52:16 +0000 (08:52 -0700)]
inet: reqsk_alloc() needs to take care of dead listeners

We'll soon no longer take a refcount on listeners,
so reqsk_alloc() can not assume a listener refcount is not
zero. We need to use atomic_inc_not_zero()

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>