Russell King says:

====================
SFP updates

Included in this series are a further few updates for SFP support:

- Adding support for Fiberstore's non-standard BiDi modules operating
  at 1310nm/1550nm wavelengths rather than the 1000BASE-BX standard of
  1310nm/1490nm.
- Adding support for negotiating the PHY interface mode with the MAC,
  so that modules supporting faster speeds and Gigabit ethernet work
  with Gigabit-only MACs.
- Adding support for high power (>1W) SFP modules.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>

This patch is the result of work by both Jon Nettleton and Russell King.
Jon wrote the original patch, adding support for SFP modules which
require a power level greater than '1'.

Russell's changes:
- Fix the power levels for big-endian, and make the code flow better.
- Convert to use device_property_read_u8()
- Warn for power levels exceeding host level
  SFF-8431 says:

  "To avoid exceeding system power supply limits and cooling capacity,
   all modules at power up by default shall operate with up to 1.0 W.
   Hosts supporting Power Level II or III operation may enable a Power
   Level II or III module through the 2-wire interface. Power Level II
   or III modules shall assert the power level declaration bit of
   SFF-8472."

  Print a warning for modules that exceed the host power level, and
  leave them operating in power level 1.

- Fix i2c write
  The first byte of any write after the bus address is always the
  device address.  In order to write a value to device D, address I,
  value V, we need to generate on the bus:

    S DDDDDDDD A IIIIIIII A VVVVVVVV A P

  where S = start, R = restart, A = ack, P = stop.  Splitting this
  as two:

    S DDDDDDDD A IIIIIIII A R DDDDDDDD A VVVVVVVV A P

  results in the device's address register being written first by I
  and then by V - the addressed register within the device is not
  written.

- Avoid power mode switching if 0xa2 is not implemented
  Some modules indicate that they support power level II or power level
  III, but do not implement address 0xa2, meaning that the bit to set
  them to high power mode is not accessible.

  These modules appear to have the sff8472_compliance field set to zero,
  and also do not implement diagnostics.  Detect this, but also ensure
  that the module does not require the address switching mode, which we
  do not implement.

- Use mW for power level rather than power level number.

- Fix high power mode transition
  We must not switch to SFP_MOD_PRESENT state until we have finished
  initialising, because the remaining state machines check for that
  state.  Add SFP_MOD_HPOWER as an intermediate state.

- Use definition for I2C register address rather than constant.
Signed-off-by: Jon Nettleton <jon@solid-run.com>
Signed-off-by: Russell King <rmk+kernel@armlinux.org.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>

Add the new maximum power level property to the SFP binding.
Signed-off-by: Russell King <rmk+kernel@armlinux.org.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>

Negotiate the interface format with the MAC rather than requiring it to
be a fixed type specified solely by the SFP module.  This allows modules
that can work with several different interface signalling formats to
select a format compatible with the MAC - for example, a Fiber module
supporing Gigabit ethernet and faster connected to a Gigabit only MAC
needs to select the 1000BASE-X mode.
Signed-off-by: Russell King <rmk+kernel@armlinux.org.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>

Some BiDi modules (eg, FiberStore SFP-GE-BX) are not compliant with
1000BASE-BX as they use different wavelengths from the 1000BASE-BX
standard (eg, 1310nm/1550nm rather than 1310nm/1490nm).  These modules
support 1000BASE-X ethernet, so detect them by a failure to find any
other support, the 8B10B encoding and a bit rate that falls within the
1Gbps window.
Signed-off-by: Russell King <rmk+kernel@armlinux.org.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>

Use extack inside team's enslavement function and also propagate it to
the netdevice notifier to allow enslaved ports to report the failure
reason. Example:

$ teamd -t team0 -d -c '{"runner": {"name": "lacp"}}'
$ ip link set dev lo master team0
Error: Loopback device can't be added as a team port.
Signed-off-by: Ido Schimmel <idosch@mellanox.com>
Acked-by: Jiri Pirko <jiri@mellanox.com>
Acked-by: David Ahern <dsahern@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Saeed Mahameed says:

mlx5-update-2018-02-23 (IB representors)

From: Mark Bloch <markb@mellanox.com>
=========
Add IB representor when in switchdev mode

The following series adds support for an IB (RAW Ethernet only) device
representor which is created when the user switches to switchdev mode.

Today when switching to switchdev mode the only representors which are
created are net devices. Each netdev is a representor of a virtual
function and any data sent via the representor is received on the virtual
function, and any data sent via the virtual function is received by the
representor.

For the mlx5 driver the main use of this functionality is to be able to
use Open vSwitch on the hypervisor in order to manage/control traffic
from/to the virtual functions. Open vSwitch can also work with  DPDK
devices and not just net devices, this series exposes an IB device, which
Mellanox PMD driver uses, which then can be used by Open vSwitch DPDK.

An IB device representor exposes only RAW Ethernet QP capabilities and
the ability to create flow rules to direct traffic to its RX queues. The
state of the IB device (ACTIVE/DOWN etc..) is based on the state of the
corresponding net device representor. No other RDMA/RoCE functionality is
currently supported and no GID table is exposed.
=========
Signed-off-by: David S. Miller <davem@davemloft.net>

Tariq Toukan says:

====================
mlx4_en misc for 4.17

This patchset contains misc enhancements from the team
to the mlx4 Eth driver.

Patch 1 by Eran adds physical layer counters.
Patch 2 by Eran cleans-up a redundant warn print.
Patch 3 combines the checks of two end cases into a single if statement.
Patch 4 takes common code structures out of the #ifdef, following your
comment on a previous patch.

Series generated against net-next commit:
f74290fd Merge git://git.kernel.org/pub/scm/linux/kernel/git/davem/net
====================
Signed-off-by: David S. Miller <davem@davemloft.net>

Pre-define a mask for IP status of a completion, that tests the
MLX4_CQE_STATUS_IPV6 only in case CONFIG_IPV6 is enabled.
Use it for IP status testing upon completion, instead of separating
the datapath into two flows.
This takes common code structures (such as closing parenthesis)
back to their original place, and makes code more readable.
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Suggested-by: David S. Miller <davem@davemloft.net>
Signed-off-by: David S. Miller <davem@davemloft.net>

Combine two end-cases in the same if statement with a single return value.
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

In mlx4_en_reset_config, there was a redundant warn print that was left
from previous versions of this function. No warn is needed anymore.

This warn can be confusing when RX-FCS is changed:
Turn OFF RX-FCS:
  mlx4_en: eth1: Changing device configuration rx filter(0) rx vlan(1)
Turn ON RX-FCS:
  mlx4_en: eth1: Changing device configuration rx filter(0) rx vlan(1)
Signed-off-by: Eran Ben Elisha <eranbe@mellanox.com>
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Add physical RX/TX packets/bytes counters into ethtool output to monitor
all traffic that was received and transmitted on the port. These
counters are available only for none Virtual Function.
Signed-off-by: Eran Ben Elisha <eranbe@mellanox.com>
Signed-off-by: Tariq Toukan <tariqt@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Jiri Pirko says:

====================
mlxsw: Offloading encapsulated SPAN

Petr says:

This patch series introduces support for mirroring with GRE
encapsulation. It offloads tc action mirred mirror from a mlxsw port to
either a gretap or an ip6gretap netdevice.

Spectrum hardware needs to know all the details of the requested
encapsulation: source and destination MAC and IP addresses, details of
VLAN tagging, etc. The only variables are the encapsulated packet
itself, and TOS field, which may be inherited. To that end, mlxsw driver
resolves the route that encapsulated packets would take, queries the
corresponding neighbor, and with that configuration in hand, configures
the mirroring in the hardware.

The driver also hooks into event handlers for netdevice changes, FIB and
neighbor events, and reconsiders the configuration on each such change.
When the new configuration differs from the currently-offloaded one, the
existing offload is removed and replaced with a new one.

It is possible to mirror to {ip6,}gretap from a matchall rule as well as
from a flower match.

** Note that with this patch set, mlxsw build depends on NET_IPGRE and
   IPV6_GRE.

Current limitations:

- There has to be a route that directs packets to an mlxsw port. We
  intend to extend the logic to support other netdevice types in the
  future, but the eventual egress netdevice will have to be an mlxsw
  port in any case.

- Offload reconfiguration due to changes in netdevice configuration
  creates a window of time where packets are not mirrored. Under some
  circumstances this can be prevented by configuring an unused port
  analyzer and migrating mirrors over to that. However that's currently
  not implemented.

- Remote address of a tunnel device needs to be set, there may not be a
  GRE key, checksumming or sequence numbers, and TTL needs to be fixed
  (non-inherit). These are hard requirements imposed by the underlying
  hardware.

- TOS of a tunnel device needs to be "inherit". The hardware supports a
  fixed TOS, but that's currently not implemented.

The series start with two patches, #1 and #2, that publish one function
and add support for querying IPv6 tunnel parameters.

In patches #3 and #4, we introduce helpers to GRE and tunneling code
that we will use later in the patchset from the SPAN code.

Patches #5 and #6 introduce support for encapsulated SPAN in reg.h.

The following seven patches, #7-#13, then prepare the SPAN codebase for
introduction of mirroring to netdevices that don't correspond to front
panel ports.

Then #14 and #15 pull all this together to implement mirroring to
{ip6,}gretap netdevices.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>

Similarly to mirror-to-gretap, this enables mirroring to IPv6 gretap
netdevice.
Signed-off-by: Petr Machata <petrm@mellanox.com>
Reviewed-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

When a user requests mirror from a mlxsw physical port (possibly based
on an ACL match) to a gretap netdevice, the driver needs to resolve the
request to a particular physical port that the mirrored packets will
egress through, and a suite of configuration keys (importantly, IP and
MAC addresses). That means calling into routing and neighbor kernel code
to simulate the decisions made by the system for packets passing through
a gretap netdevice.

Add a new instance of mlxsw_sp_span_entry_ops to support this.
Signed-off-by: Petr Machata <petrm@mellanox.com>
Reviewed-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

The check for whether a mirror port (which is a mlxsw front panel port)
belongs to the same mlxsw instance as the mirrored port, is currently
only done in spectrum_acl, even though it's applicable for the matchall
case as well. Thus move it to mlxsw_sp_span_entry_create().
Signed-off-by: Petr Machata <petrm@mellanox.com>
Reviewed-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

For some netdevices, for which mlxsw offloads mirroring, may have a
complex relationship between the declared intent and low-level
device configuration.

Trying to accurately track which changes might influence offloading
decisions is finicky and error-prone. Instead, this patch introduces a
function mlxsw_sp_span_entry_respin, which re-queries the configuration
anew and, if different, removes the existing offloads and installs new
ones.

Call this function strategically at event handlers that might influence
the mirroring configuration.
Signed-off-by: Petr Machata <petrm@mellanox.com>
Reviewed-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

To support mirroring to different device types, the functions that
partake in configuring the port analyzer need to be extended to admit
non-trivial SPAN types.

Create a structure where all details of SPAN configuration are kept,
struct mlxsw_sp_span_parms. Also create struct mlxsw_sp_span_entry_ops
to keep per-SPAN-type operations.

Instantiate the latter once for MLXSW_REG_MPAT_SPAN_TYPE_LOCAL_ETH, and
once for a suite of NOP callbacks used for invalidated SPAN entry. Put
the formet as a sole member of a new array mlxsw_sp_span_entry_types,
where all known SPAN types are kept. Introduce a new function,
mlxsw_sp_span_entry_ops(), to look up the right ops suite given a
netdevice.

Change mlxsw_sp_span_mirror_add() to use both parms and ops structures.
Change mlxsw_sp_span_entry_get() and mlxsw_sp_span_entry_create() to
take these as arguments. Modify mlxsw_sp_span_entry_configure() and
mlxsw_sp_span_entry_deconfigure() to dispatch to ops.
Signed-off-by: Petr Machata <petrm@mellanox.com>
Reviewed-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Currently the only mirror action supported by mlxsw is mirror to another
mlxsw physical port. Correspondingly, span_entry, which tracks each
mlxsw mirror in the system, currently holds a u8 number of the
destination port.

To extend this system to mirror to gretap and ip6gretap netdevices, have
struct mlxsw_sp_span_entry actually hold the destination netdevice
itself.

This change then trickles down in obvious manner to SPAN module API and
mirror-related interfaces in struct mlxsw_afa_ops.

To prevent use of invalid pointer, NETDEV_UNREGISTER needs to be hooked
and the corresponding SPAN entry invalidated.
Signed-off-by: Petr Machata <petrm@mellanox.com>
Reviewed-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Configuring the hardware for encapsulated SPAN involves more code than
the simple mirroring case. Extract the related code to a separate
function to separate it from the rest of SPAN entry creation. Extract
deconfigure as well for symmetry, even though disablement is the same
regardless of SPAN type.
Signed-off-by: Petr Machata <petrm@mellanox.com>
Reviewed-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

It is known statically ahead of time which SPAN entry will have which
ID. Just initialize it eagerly in mlxsw_sp_span_init(), don't wait until
the entry is actually created. This simplifies some code in
mlxsw_sp_span_entry_create()
Signed-off-by: Petr Machata <petrm@mellanox.com>
Reviewed-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Instead of removing span_entry by the port number, allow removing by
SPAN id. That simplifies some code right here, and for mirroring to soft
netdevices, avoids problems with netdevice pointer invalidation and
reuse.

Rename mlxsw_sp_span_entry_find() to mlxsw_sp_span_entry_find_by_port()
and keep it--follow-up patches will make use of it.
Signed-off-by: Petr Machata <petrm@mellanox.com>
Reviewed-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

To support encapsulated SPAN, extend mlxsw_reg_mpat_pack() with a field
to set the SPAN type.
Signed-off-by: Petr Machata <petrm@mellanox.com>
Reviewed-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

MPAT Register is used to query and configure the Switch Port Analyzer
Table. To configure Port Analyzer to encapsulate mirrored packets,
additional fields need to be specified for the MPAT register.
Signed-off-by: Petr Machata <petrm@mellanox.com>
Reviewed-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Initializing struct flowi4 is useful for drivers that need to emulate
routing decisions made by a tunnel interface. Publish the
function (appropriately renamed) so that the drivers in question don't
need to cut'n'paste it around.
Signed-off-by: Petr Machata <petrm@mellanox.com>
Reviewed-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Determining whether a device is a GRE device is easily done by
inspecting struct net_device.type. However, for the tap variants, the
type is just ARPHRD_ETHER.

Therefore introduce two predicate functions that use netdev_ops to tell
the tap devices.
Signed-off-by: Petr Machata <petrm@mellanox.com>
Reviewed-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

To support mirroring to ip6gretap, the SPAN module needs to be able to
decode IPv6 addresses specified at that tunnel.

Extend mlxsw_sp_ipip_netdev_saddr() and mlxsw_sp_ipip_netdev_daddr() to
support IPv6 addresses. To that end, add and publish a support function
mlxsw_sp_ipip_netdev_parms6().
Signed-off-by: Petr Machata <petrm@mellanox.com>
Reviewed-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Extract the logic for determining whether a given IPv4/IPv6 address is
all-zeroes from mlxsw_sp_ipip_tunnel_complete to a separate function.
Make that function public within the module.
Signed-off-by: Petr Machata <petrm@mellanox.com>
Reviewed-by: Ido Schimmel <idosch@mellanox.com>
Signed-off-by: Jiri Pirko <jiri@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Thomas Falcon says:

====================
ibmvnic: Miscellaneous driver fixes and enhancements

There is not a general theme to this patch set other than that it
fixes a few issues with the ibmvnic driver. I will just give a quick
summary of what each patch does here.

"ibmvnic: Fix TX descriptor tracking again" resolves a race condition
introduced in an earlier fix to track outstanding transmit descriptors.
This condition can throw off the tracking counter to the point that
a transmit queue will halt forever.

"ibmvnic: Allocate statistics buffers during probe" allocates queue
statistics buffers on device probe to avoid a crash when accessing
statistics of an unopened interface.

"ibmvnic: Harden TX/RX pool cleaning" includes additional checks to
avoid a bad access when cleaning RX and TX buffer pools during a device
reset.

"ibmvnic: Report queue stops and restarts as debug output" changes TX
queue state notifications from informational to debug messages. This
information is not necessarily useful to a user and under load can result
in a lot of log output.

"ibmvnic: Do not attempt to login if RX or TX queues are not allocated"
checks that device queues have been allocated successfully before
attempting device login. This resolves a panic that could occur if a
user attempted to configure a device after a failed reset.

Thanks for your attention.
====================
Signed-off-by: David S. Miller <davem@davemloft.net>

If a device reset fails for some reason, TX and RX queue resources
could be released. If a user attempts to open the device in this scenario,
it may result in a kernel panic as the driver tries to access this
memory. To fix this, include a check before device login that TX/RX
queues are still there before enabling the device. In addition, return a
value that can be checked in case of any errors to avoid waiting for a
completion that will never come.
Signed-off-by: Thomas Falcon <tlfalcon@linux.vnet.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

It's not necessary to report each time a queue is stopped and restarted
as an informational message. Change that to be a debug message so that
it can be observed if needed but not printed by default.
Signed-off-by: Thomas Falcon <tlfalcon@linux.vnet.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

If the driver releases resources after a failed reset or some other
error, the driver might attempt to clean up and free memory that
isn't there anymore. Include some additional checks that RX/TX queues
along with their associated structures are still there before cleaning.
Signed-off-by: Thomas Falcon <tlfalcon@linux.vnet.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Currently, buffers holding individual queue statistics are allocated
when the device is opened. If an ibmvnic interface is hotplugged or
initialized but never opened, an attempt to get statistics with
ethtool will result in a kernel panic.

Since the driver allocates a constant number, the maximum supported
queues, of buffers, these can be allocated during device probe and
freed when the device is hot-unplugged or the module is removed.
Signed-off-by: Thomas Falcon <tlfalcon@linux.vnet.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Sorry, the previous change introduced a race condition between
transmit completion processing and tracking TX descriptors. If a
completion is received before the number of descriptors is logged,
the number of descriptors will be add but not removed. After enough
times, this could halt the transmit queue forever.

Log the number of descriptors used by a transmit before sending.
I stress tested the fix on two different systems running over the
weekend without any issues.
Signed-off-by: Thomas Falcon <tlfalcon@linux.vnet.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Niklas Cassel says:

====================
stmmac barrier fixes and cleanup
====================
Signed-off-by: David S. Miller <davem@davemloft.net>

Make dwmac4_release_tx_desc() clear all descriptor fields, not just
TDES2 and TDES3.

I'm suspecting that TDES0 and TDES1 wasn't cleared because the DMA
engine uses them to store the tx hardware timestamp (if PTP is enabled).

However, stmmac_tx_clean() calls stmmac_get_tx_hwtstamp(), which reads
and saves the timestamp, before it calls release_tx_desc(), so this
is not an issue.

stmmac_xmit() and stmmac_tso_xmit() both always overwrite TDES0,
however, stmmac_tso_xmit() sometimes sets TDES1, and since neither
stmmac_xmit() nor stmmac_tso_xmit() explicitly clears TDES1, both
functions might reuse a DMA descriptor with old TDES1 data.

I haven't observed any misbehavior even though TDES1 sometimes
point to an old skb, however, explicitly clearing both TDES0 and TDES1
in dwmac4_release_tx_desc() minimizes the chances of undefined behavior.
Signed-off-by: Niklas Cassel <niklas.cassel@axis.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

According to Documentation/memory-barriers.txt, we need to use a
dma_rmb() after reading the status/own bit, to ensure that all
descriptor fields are read after reading the own bit.

This way, we ensure that the DMA engine is done with the DMA
descriptor before we read the other descriptor fields, e.g. reading
the tx hardware timestamp (if PTP is enabled).
Signed-off-by: Niklas Cassel <niklas.cassel@axis.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

The last memory barrier in stmmac_xmit()/stmmac_tso_xmit() is placed
between a coherent memory write and a MMIO write:

The own bit is written in First Desc (TSO: MSS desc or First Desc).
<barrier>
The DMA engine is started by a write to the tx desc tail pointer/
enable dma transmission register, i.e. a MMIO write.

This barrier cannot be a simple dma_wmb(), since a dma_wmb() is only
used to guarantee the ordering, with respect to other writes,
to cache coherent DMA memory.

To guarantee that the cache coherent memory writes have completed
before we attempt to write to the cache incoherent MMIO region,
we need to use the more heavyweight barrier wmb().
Signed-off-by: Niklas Cassel <niklas.cassel@axis.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

A dma_wmb() is used to guarantee the ordering, with respect to
other writes, to cache coherent DMA memory.

There is a dma_wmb() in prepare_tx_desc()/prepare_tso_tx_desc() which
ensures that TDES0/1/2 is written before TDES3 (which contains the own
bit), for First Desc.

However, in the rare case that MSS changes, there will be a MSS
context descriptor in front of the regular DMA descriptors:

<MSS desc> <- DMA Next Descriptor
<First Desc>
<desc n>
<Last Desc>

Thus, for this special case, we need a dma_wmb()
after prepare_tso_tx_desc()/before writing the own bit to the MSS desc,
so that we flush the write to TDES3 for First Desc,
in order to ensure that the MSS descriptor is the last descriptor to
set the own bit.
Signed-off-by: Niklas Cassel <niklas.cassel@axis.com>
Signed-off-by: David S. Miller <davem@davemloft.net>

Sowmini Varadhan says:

====================
RDS: optimized notification for zerocopy completion

Resending with acked-by additions: previous attempt does not show
up in Patchwork. This time with a new mail Message-Id.

RDS applications use predominantly request-response, transacation
based IPC, so that ingress and egress traffic are well-balanced,
and it is possible/desirable to reduce system-call overhead by
piggybacking the notifications for zerocopy completion response
with data.

Moreover, it has been pointed out that socket functions block
if sk_err is non-zero, thus if the RDS code does not plan/need
to use sk_error_queue path for completion notification, it
is preferable to remove the sk_errror_queue related paths in
RDS.

Both of these goals are implemented in this series.

v2: removed sk_error_queue support
v3: incorporated additional code review comments (details in each patch)
====================
Signed-off-by: David S. Miller <davem@davemloft.net>