platform/kernel/linux-rpi.git
2 years agogro: add support of (hw)gro packets to gro stack
Coco Li [Fri, 30 Sep 2022 22:09:05 +0000 (15:09 -0700)]
gro: add support of (hw)gro packets to gro stack

Current GRO stack only supports incoming packets containing
one frame/MSS.

This patch changes GRO to accept packets that are already GRO.

HW-GRO (aka RSC for some vendors) is very often limited in presence
of interleaved packets. Linux SW GRO stack can complete the job
and provide larger GRO packets, thus reducing rate of ACK packets
and cpu overhead.

This also means BIG TCP can still be used, even if HW-GRO/RSC was
able to cook ~64 KB GRO packets.

v2: fix logic in tcp_gro_receive()

    Only support TCP for the moment (Paolo)

Co-Developed-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: Coco Li <lixiaoyan@google.com>
Acked-by: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2 years agoMerge branch 'mptcp-fastclose'
David S. Miller [Mon, 3 Oct 2022 10:18:53 +0000 (11:18 +0100)]
Merge branch 'mptcp-fastclose'

Mat Martineau says:

====================
mptcp: Fastclose edge cases and error handling

MPTCP has existing code to use the MP_FASTCLOSE option header, which
works like a RST for the MPTCP-level connection (regular RSTs only
affect specific subflows in MPTCP). This series has some improvements
for fastclose.

Patch 1 aligns fastclose socket error handling with TCP RST behavior on
TCP sockets.

Patch 2 adds use of MP_FASTCLOSE in some more edge cases, like file
descriptor close, FIN_WAIT timeout, and when the socket has unread data.

Patch 3 updates the fastclose self tests.

Patch 4 does not change any code, just fixes some outdated comments.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
2 years agomptcp: update misleading comments.
Paolo Abeni [Fri, 30 Sep 2022 15:59:34 +0000 (08:59 -0700)]
mptcp: update misleading comments.

The MPTCP data path is quite complex and hard to understend even
without some foggy comments referring to modified code and/or
completely misleading from the beginning.

Update a few of them to more accurately describing the current
status.

Reviewed-by: Mat Martineau <mathew.j.martineau@linux.intel.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: Mat Martineau <mathew.j.martineau@linux.intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2 years agoselftests: mptcp: update and extend fastclose test-cases
Paolo Abeni [Fri, 30 Sep 2022 15:59:33 +0000 (08:59 -0700)]
selftests: mptcp: update and extend fastclose test-cases

After the previous patches, the MPTCP protocol can generate
fast-closes on both ends of the connection. Rework the relevant
test-case to carefully trigger the fast-close code-path on a
single end at the time, while ensuring than a predictable amount
of data is spooled on both ends.

Additionally add another test-cases for the passive socket
fast-close.

Reviewed-by: Matthieu Baerts <matthieu.baerts@tessares.net>
Reviewed-by: Mat Martineau <mathew.j.martineau@linux.intel.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: Mat Martineau <mathew.j.martineau@linux.intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2 years agomptcp: use fastclose on more edge scenarios
Paolo Abeni [Fri, 30 Sep 2022 15:59:32 +0000 (08:59 -0700)]
mptcp: use fastclose on more edge scenarios

Daire reported a user-space application hang-up when the
peer is forcibly closed before the data transfer completion.

The relevant application expects the peer to either
do an application-level clean shutdown or a transport-level
connection reset.

We can accommodate a such user by extending the fastclose
usage: at fd close time, if the msk socket has some unread
data, and at FIN_WAIT timeout.

Note that at MPTCP close time we must ensure that the TCP
subflows will reset: set the linger socket option to a suitable
value.

Reviewed-by: Matthieu Baerts <matthieu.baerts@tessares.net>
Reviewed-by: Mat Martineau <mathew.j.martineau@linux.intel.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: Mat Martineau <mathew.j.martineau@linux.intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2 years agomptcp: propagate fastclose error
Paolo Abeni [Fri, 30 Sep 2022 15:59:31 +0000 (08:59 -0700)]
mptcp: propagate fastclose error

When an mptcp socket is closed due to an incoming FASTCLOSE
option, so specific sk_err is set and later syscall will
fail usually with EPIPE.

Align the current fastclose error handling with TCP reset,
properly setting the socket error according to the current
msk state and propagating such error.

Additionally sendmsg() is currently not handling properly
the sk_err, always returning EPIPE.

Reviewed-by: Matthieu Baerts <matthieu.baerts@tessares.net>
Reviewed-by: Mat Martineau <mathew.j.martineau@linux.intel.com>
Signed-off-by: Paolo Abeni <pabeni@redhat.com>
Signed-off-by: Mat Martineau <mathew.j.martineau@linux.intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2 years agoMerge branch 'RollBall-Hilink-Turris-10G-copper-SFP-support'
David S. Miller [Mon, 3 Oct 2022 10:08:33 +0000 (11:08 +0100)]
Merge branch 'RollBall-Hilink-Turris-10G-copper-SFP-support'

Marek Behún says:

====================
RollBall / Hilink / Turris 10G copper SFP support

I am resurrecting my attempt to add support for RollBall / Hilink /
Turris 10G copper SFPs modules.

The modules contain Marvell 88X3310 PHY, which can communicate with
the system via sgmii, 2500base-x, 5gbase-r, 10gbase-r or usxgmii mode.

Some of the patches I've taken from Russell King's net-queue [1]
(with some rebasing).

The important change from my previous attempts are:
- I am including the changes needed to phylink and marvell10g driver,
  so that the 88X3310 PHY is configured to use PHY modes supported by
  the host (the PHY defaults to use 10gbase-r only on host's side)
- I have changed the patch that informs phylib about the interfaces
  supported by the host (patch 5 of this series): it now fills in the
  phydev->host_interfaces member only when connecting a PHY that is
  inside a SFP module. This may change in the future.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
2 years agonet: sfp: add support for multigig RollBall transceivers
Marek Behún [Fri, 30 Sep 2022 14:21:10 +0000 (16:21 +0200)]
net: sfp: add support for multigig RollBall transceivers

This adds support for multigig copper SFP modules from RollBall/Hilink.
These modules have a specific way to access clause 45 registers of the
internal PHY.

We also need to wait at least 22 seconds after deasserting TX disable
before accessing the PHY. The code waits for 25 seconds just to be sure.

Signed-off-by: Marek Behún <kabel@kernel.org>
Reviewed-by: Russell King <rmk+kernel@armlinux.org.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>
2 years agonet: phy: mdio-i2c: support I2C MDIO protocol for RollBall SFP modules
Marek Behún [Fri, 30 Sep 2022 14:21:09 +0000 (16:21 +0200)]
net: phy: mdio-i2c: support I2C MDIO protocol for RollBall SFP modules

Some multigig SFPs from RollBall and Hilink do not expose functional
MDIO access to the internal PHY of the SFP via I2C address 0x56
(although there seems to be read-only clause 22 access on this address).

Instead these SFPs PHY can be accessed via I2C via the SFP Enhanced
Digital Diagnostic Interface - I2C address 0x51. The SFP_PAGE has to be
selected to 3 and the password must be filled with 0xff bytes for this
PHY communication to work.

This extends the mdio-i2c driver to support this protocol by adding a
special parameter to mdio_i2c_alloc function via which this RollBall
protocol can be selected.

Signed-off-by: Marek Behún <kabel@kernel.org>
Cc: Andrew Lunn <andrew@lunn.ch>
Cc: Russell King <rmk+kernel@armlinux.org.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>
2 years agonet: sfp: create/destroy I2C mdiobus before PHY probe/after PHY release
Marek Behún [Fri, 30 Sep 2022 14:21:08 +0000 (16:21 +0200)]
net: sfp: create/destroy I2C mdiobus before PHY probe/after PHY release

Instead of configuring the I2C mdiobus when SFP driver is probed,
create/destroy the mdiobus before the PHY is probed for/after it is
released.

This way we can tell the mdio-i2c code which protocol to use for each
SFP transceiver.

Move the code that determines MDIO I2C protocol from
sfp_sm_probe_for_phy() to sfp_sm_mod_probe(), where most of the SFP ID
parsing is done. Don't allocate I2C bus if no PHY is expected.

Signed-off-by: Marek Behún <kabel@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2 years agonet: sfp: Add and use macros for SFP quirks definitions
Marek Behún [Fri, 30 Sep 2022 14:21:07 +0000 (16:21 +0200)]
net: sfp: Add and use macros for SFP quirks definitions

Add macros SFP_QUIRK(), SFP_QUIRK_M() and SFP_QUIRK_F() for defining SFP
quirk table entries. Use them to deduplicate the code a little bit.

Signed-off-by: Marek Behún <kabel@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2 years agonet: phylink: allow attaching phy for SFP modules on 802.3z mode
Marek Behún [Fri, 30 Sep 2022 14:21:06 +0000 (16:21 +0200)]
net: phylink: allow attaching phy for SFP modules on 802.3z mode

Some SFPs may contain an internal PHY which may in some cases want to
connect with the host interface in 1000base-x/2500base-x mode.
Do not fail if such PHY is being attached in one of these PHY interface
modes.

Signed-off-by: Marek Behún <kabel@kernel.org>
Reviewed-by: Russell King <rmk+kernel@armlinux.org.uk>
Reviewed-by: Pali Rohár <pali@kernel.org>
Cc: Andrew Lunn <andrew@lunn.ch>
Signed-off-by: David S. Miller <davem@davemloft.net>
2 years agonet: phy: marvell10g: select host interface configuration
Russell King [Fri, 30 Sep 2022 14:21:05 +0000 (16:21 +0200)]
net: phy: marvell10g: select host interface configuration

Select the host interface configuration according to the capabilities of
the host if the host provided them. This is currently provided only when
connecting PHY that is inside a SFP.

The PHY supports several configurations of host communication:
- always communicate with host in 10gbase-r, even if copper speed is
  lower (rate matching mode),
- the same as above but use xaui/rxaui instead of 10gbase-r,
- switch host SerDes mode between 10gbase-r, 5gbase-r, 2500base-x and
  sgmii according to copper speed,
- the same as above but use xaui/rxaui instead of 10gbase-r.

This mode of host communication, called MACTYPE, is by default selected
by strapping pins, but it can be changed in software.

This adds support for selecting this mode according to which modes are
supported by the host.

This allows the kernel to:
- support SFP modules with 88X33X0 or 88E21X0 inside them

Note: we use mv3310_select_mactype() for both 88X3310 and 88X3340,
although 88X3340 does not support XAUI. This is not a problem because
88X3340 does not declare XAUI in it's supported_interfaces, and so this
function will never choose that MACTYPE.

Signed-off-by: Russell King <rmk+kernel@armlinux.org.uk>
[ rebase, updated, also added support for 88E21X0 ]
Signed-off-by: Marek Behún <kabel@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2 years agonet: phy: marvell10g: Use tabs instead of spaces for indentation
Marek Behún [Fri, 30 Sep 2022 14:21:04 +0000 (16:21 +0200)]
net: phy: marvell10g: Use tabs instead of spaces for indentation

Some register definitions were defined with spaces used for indentation.
Change them to tabs.

Signed-off-by: Marek Behún <kabel@kernel.org>
Reviewed-by: Russell King (Oracle) <rmk+kernel@armlinux.org.uk>
Signed-off-by: David S. Miller <davem@davemloft.net>
2 years agonet: phylink: pass supported host PHY interface modes to phylib for SFP's PHYs
Marek Behún [Fri, 30 Sep 2022 14:21:03 +0000 (16:21 +0200)]
net: phylink: pass supported host PHY interface modes to phylib for SFP's PHYs

Pass the supported PHY interface types to phylib if the PHY we are
connecting is inside a SFP, so that the PHY driver can select an
appropriate host configuration mode for their interface according to
the host capabilities.

For example the Marvell 88X3310 PHY inside RollBall SFP modules
defaults to 10gbase-r mode on host's side, and the marvell10g
driver currently does not change this setting. But a host may not
support 10gbase-r. For example Turris Omnia only supports sgmii,
1000base-x and 2500base-x modes. The PHY can be configured to use
those modes, but in order for the PHY driver to do that, it needs
to know which modes are supported.

Signed-off-by: Marek Behún <kabel@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2 years agonet: phylink: rename phylink_sfp_config()
Russell King (Oracle) [Fri, 30 Sep 2022 14:21:02 +0000 (16:21 +0200)]
net: phylink: rename phylink_sfp_config()

phylink_sfp_config() now only deals with configuring the MAC for a
SFP containing a PHY. Rename it to be specific.

Signed-off-by: Russell King (Oracle) <rmk+kernel@armlinux.org.uk>
Signed-off-by: Marek Behún <kabel@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2 years agonet: phylink: use phy_interface_t bitmaps for optical modules
Russell King [Fri, 30 Sep 2022 14:21:01 +0000 (16:21 +0200)]
net: phylink: use phy_interface_t bitmaps for optical modules

Where a MAC provides a phy_interface_t bitmap, use these bitmaps to
select the operating interface mode for optical SFP modules, rather
than using the linkmode bitmaps.

Signed-off-by: Russell King <rmk+kernel@armlinux.org.uk>
Signed-off-by: Marek Behún <kabel@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2 years agonet: sfp: augment SFP parsing with phy_interface_t bitmap
Russell King [Fri, 30 Sep 2022 14:21:00 +0000 (16:21 +0200)]
net: sfp: augment SFP parsing with phy_interface_t bitmap

We currently parse the SFP EEPROM to a bitmap of ethtool link modes,
and then attempt to convert the link modes to a PHY interface mode.
While this works at present, there are cases where this is sub-optimal.
For example, where a module can operate with several different PHY
interface modes.

To start addressing this, arrange for the SFP EEPROM parsing to also
provide a bitmap of the possible PHY interface modes.

Signed-off-by: Russell King <rmk+kernel@armlinux.org.uk>
Signed-off-by: Marek Behún <kabel@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2 years agonet: phylink: add ability to validate a set of interface modes
Russell King (Oracle) [Fri, 30 Sep 2022 14:20:59 +0000 (16:20 +0200)]
net: phylink: add ability to validate a set of interface modes

Rather than having the ability to validate all supported interface
modes or a single interface mode, introduce the ability to validate
a subset of supported modes.

Signed-off-by: Russell King (Oracle) <rmk+kernel@armlinux.org.uk>
[ rebased on current net-next ]
Signed-off-by: Marek Behún <kabel@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2 years agoMerge branch 'ip_tunnel-netlink-parms'
David S. Miller [Mon, 3 Oct 2022 06:59:06 +0000 (07:59 +0100)]
Merge branch 'ip_tunnel-netlink-parms'

Liu Jian says:

====================
Add helper functions to parse netlink msg of ip_tunnel

v1->v2: Move the implementation of the helper function to ip_tunnel_core.c
v2->v3: Change EXPORT_SYMBOL to EXPORT_SYMBOL_GPL
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
2 years agonet: Add helper function to parse netlink msg of ip_tunnel_parm
Liu Jian [Thu, 29 Sep 2022 13:52:03 +0000 (21:52 +0800)]
net: Add helper function to parse netlink msg of ip_tunnel_parm

Add ip_tunnel_netlink_parms to parse netlink msg of ip_tunnel_parm.
Reduces duplicate code, no actual functional changes.

Signed-off-by: Liu Jian <liujian56@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2 years agonet: Add helper function to parse netlink msg of ip_tunnel_encap
Liu Jian [Thu, 29 Sep 2022 13:52:02 +0000 (21:52 +0800)]
net: Add helper function to parse netlink msg of ip_tunnel_encap

Add ip_tunnel_netlink_encap_parms to parse netlink msg of ip_tunnel_encap.
Reduces duplicate code, no actual functional changes.

Signed-off-by: Liu Jian <liujian56@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2 years agoMerge branch 'master' of git://git.kernel.org/pub/scm/linux/kernel/git/klassert/ipsec...
David S. Miller [Mon, 3 Oct 2022 06:46:08 +0000 (07:46 +0100)]
Merge branch 'master' of git://git./linux/kernel/git/klassert/ipsec-next

Steffen Klassert says:

====================
1) Refactor selftests to use an array of structs in xfrm_fill_key().
   From Gautam Menghani.

2) Drop an unused argument from xfrm_policy_match.
   From Hongbin Wang.

3) Support collect metadata mode for xfrm interfaces.
   From Eyal Birger.

4) Add netlink extack support to xfrm.
   From Sabrina Dubroca.

Please note, there is a merge conflict in:

include/net/dst_metadata.h

between commit:

0a28bfd4971f ("net/macsec: Add MACsec skb_metadata_dst Tx Data path support")

from the net-next tree and commit:

5182a5d48c3d ("net: allow storing xfrm interface metadata in metadata_dst")

from the ipsec-next tree.

Can be solved as done in linux-next.

Please pull or let me know if there are problems.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
2 years agoMerge branch 'tc-bind_class-hook'
David S. Miller [Sun, 2 Oct 2022 15:07:17 +0000 (16:07 +0100)]
Merge branch 'tc-bind_class-hook'

Zhengchao Shao says:

====================
refactor duplicate codes in bind_class hook function

All the bind_class callback duplicate the same logic, so we can refactor
them. First, ensure n arg not empty before call bind_class hook function.
Then, add tc_cls_bind_class() helper. Last, use tc_cls_bind_class() in
filter.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
2 years agonet: sched: use tc_cls_bind_class() in filter
Zhengchao Shao [Tue, 27 Sep 2022 12:48:55 +0000 (20:48 +0800)]
net: sched: use tc_cls_bind_class() in filter

Use tc_cls_bind_class() in filter.

Signed-off-by: Zhengchao Shao <shaozhengchao@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2 years agonet: sched: cls_api: introduce tc_cls_bind_class() helper
Zhengchao Shao [Tue, 27 Sep 2022 12:48:54 +0000 (20:48 +0800)]
net: sched: cls_api: introduce tc_cls_bind_class() helper

All the bind_class callback duplicate the same logic, this patch
introduces tc_cls_bind_class() helper for common usage.

Signed-off-by: Zhengchao Shao <shaozhengchao@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2 years agonet: sched: ensure n arg not empty before call bind_class
Zhengchao Shao [Tue, 27 Sep 2022 12:48:53 +0000 (20:48 +0800)]
net: sched: ensure n arg not empty before call bind_class

All bind_class callbacks are directly returned when n arg is empty.
Therefore, bind_class is invoked only when n arg is not empty.

Signed-off-by: Zhengchao Shao <shaozhengchao@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2 years agoMerge branch 'mlx5-xsk-updates-part3-2022-09-30'
Jakub Kicinski [Sat, 1 Oct 2022 20:30:25 +0000 (13:30 -0700)]
Merge branch 'mlx5-xsk-updates-part3-2022-09-30'

Saeed Mahameed says:

====================
mlx5 xsk updates part3 2022-09-30

The gist of this 4 part series is in this patchset's last patch

This series contains performance optimizations. XSK starts using the
batching allocator, and XSK data path gets separated from the regular
RX, allowing to drop some branches not relevant for non-XSK use cases.
Some minor optimizations for indirect calls and need_wakeup are also
included.

Other than that, this series adds a few features to the mlx5e
implementation of XSK:

1. XDP metadata support on XSK RQs.

2. RSS contexts support for XSK RQs.

3. Some other optimizations

4. Last but not least, change the queuing scheme, so that XSK RQs no longer
use higher indices, but replace the regular RQs.

Maxim Says:
==========

In the initial implementation of XSK in mlx5e, XSK RQs coexisted with
regular RQs in the same channel. The main idea was to allow RSS work the
same for regular traffic, without need to reconfigure RSS to exclude XSK
queues.

However, this scheme didn't prove to be beneficial, mainly because of
incompatibility with other vendors. Some tools don't properly support
using higher indices for XSK queues, some tools get confused with the
double amount of RQs exposed in sysfs. Some use cases are purely XSK,
and allocating the same amount of unused regular RQs is a waste of
resources.

This commit changes the queuing scheme to the standard one, where XSK
RQs replace regular RQs on the channels where XSK sockets are open. Two
RQs still exist in the channel to allow failsafe disable of XSK, but
only one is exposed at a time. The next commit will achieve the desired
memory save by flushing the buffers when the regular RQ is unused.

As the result of this transition:

1. It's possible to use RSS contexts over XSK RQs.

2. It's possible to dedicate all queues to XSK.

3. When XSK RQs coexist with regular RQs, the admin should make sure no
unwanted traffic goes into XSK RQs by either excluding them from RSS or
settings up the XDP program to return XDP_PASS for non-XSK traffic.

4. When using a mixed fleet of mlx5e devices and other netdevs, the same
configuration can be applied. If the application supports the fallback
to copy mode on unsupported drivers, it will work too.

==========

Part 4 will include some final xsk optimizations and minor improvements

part 1: https://lore.kernel.org/netdev/20220927203611.244301-1-saeed@kernel.org/
part 2: https://lore.kernel.org/netdev/20220929072156.93299-1-saeed@kernel.org/
====================

Link: https://lore.kernel.org/r/20220930162903.62262-1-saeed@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agonet/mlx5e: xsk: Use queue indices starting from 0 for XSK queues
Maxim Mikityanskiy [Fri, 30 Sep 2022 16:29:03 +0000 (09:29 -0700)]
net/mlx5e: xsk: Use queue indices starting from 0 for XSK queues

In the initial implementation of XSK in mlx5e, XSK RQs coexisted with
regular RQs in the same channel. The main idea was to allow RSS work the
same for regular traffic, without need to reconfigure RSS to exclude XSK
queues.

However, this scheme didn't prove to be beneficial, mainly because of
incompatibility with other vendors. Some tools don't properly support
using higher indices for XSK queues, some tools get confused with the
double amount of RQs exposed in sysfs. Some use cases are purely XSK,
and allocating the same amount of unused regular RQs is a waste of
resources.

This commit changes the queuing scheme to the standard one, where XSK
RQs replace regular RQs on the channels where XSK sockets are open. Two
RQs still exist in the channel to allow failsafe disable of XSK, but
only one is exposed at a time. The next commit will achieve the desired
memory save by flushing the buffers when the regular RQ is unused.

As the result of this transition:

1. It's possible to use RSS contexts over XSK RQs.

2. It's possible to dedicate all queues to XSK.

3. When XSK RQs coexist with regular RQs, the admin should make sure no
unwanted traffic goes into XSK RQs by either excluding them from RSS or
settings up the XDP program to return XDP_PASS for non-XSK traffic.

4. When using a mixed fleet of mlx5e devices and other netdevs, the same
configuration can be applied. If the application supports the fallback
to copy mode on unsupported drivers, it will work too.

Signed-off-by: Maxim Mikityanskiy <maximmi@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agonet/mlx5e: Introduce the mlx5e_flush_rq function
Maxim Mikityanskiy [Fri, 30 Sep 2022 16:29:02 +0000 (09:29 -0700)]
net/mlx5e: Introduce the mlx5e_flush_rq function

Add a function to flush an RQ: clean up descriptors, release pages and
reset the RQ. This procedure is used by the recovery flow, and it will
also be used in a following commit to free some memory when switching a
channel to the XSK mode.

Signed-off-by: Maxim Mikityanskiy <maximmi@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agonet/mlx5e: xsk: Support XDP metadata on XSK RQs
Maxim Mikityanskiy [Fri, 30 Sep 2022 16:29:01 +0000 (09:29 -0700)]
net/mlx5e: xsk: Support XDP metadata on XSK RQs

Add support for XDP metadata on XSK RQs for cross-program
communication. The driver no longer calls xdp_set_data_meta_invalid and
copies the metadata to a newly allocated SKB on XDP_PASS.

Signed-off-by: Maxim Mikityanskiy <maximmi@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agonet/mlx5e: Optimize RQ page deallocation
Maxim Mikityanskiy [Fri, 30 Sep 2022 16:29:00 +0000 (09:29 -0700)]
net/mlx5e: Optimize RQ page deallocation

mlx5e_free_rx_mpwqe loops over all pages of a MPWQE, calling
mlx5e_page_release for ones that are not scheduled for XDP_TX or
XDP_REDIRECT; and mlx5e_page_release checks whether it's an XSK RQ or a
regular one for each page/XSK frame. This check can be moved outside the
loop to reduce the number of branches.

mlx5e_free_rx_wqe loops over all fragments, calling mlx5e_page_release
for the ones that are last in a page; and mlx5e_page_release checks
whether it's an XSK RQ or a regular one for each fragment. Using the
fact that XSK doesn't support multiple fragments, it can be optimized
for both XSK and regular usages:

1. Make an early check for XSK and call its deallocator directly, saving
3 branches (loop condition, frag->last_in_page and selection of
deallocator).

2. Call the regular deallocator directly in the non-XSK case, saving a
branch per fragment, except the first one.

After the changes, mlx5e_page_release is removed, as there are no
callers left.

Signed-off-by: Maxim Mikityanskiy <maximmi@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agonet/mlx5e: Call mlx5e_page_release_dynamic directly where possible
Maxim Mikityanskiy [Fri, 30 Sep 2022 16:28:59 +0000 (09:28 -0700)]
net/mlx5e: Call mlx5e_page_release_dynamic directly where possible

mlx5e_page_release calls the appropriate deallocator depending on
whether it's an XSK RQ or a regular one. Some flows that call this
function are not compatible with XSK, so they can call the non-XSK
deallocator directly to save a branch.

Signed-off-by: Maxim Mikityanskiy <maximmi@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agonet/mlx5e: Use non-XSK page allocator in SHAMPO
Maxim Mikityanskiy [Fri, 30 Sep 2022 16:28:58 +0000 (09:28 -0700)]
net/mlx5e: Use non-XSK page allocator in SHAMPO

The SHAMPO flow is not compatible with XSK, it can call the page pool
allocator directly to save a branch.

mlx5e_page_alloc is removed, as it's no longer used in any flow.

Signed-off-by: Maxim Mikityanskiy <maximmi@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agonet/mlx5e: xsk: Use xsk_buff_alloc_batch on striding RQ
Maxim Mikityanskiy [Fri, 30 Sep 2022 16:28:57 +0000 (09:28 -0700)]
net/mlx5e: xsk: Use xsk_buff_alloc_batch on striding RQ

XSK provides a function to allocate frames in batches for more efficient
processing. This commit starts using this function on striding RQ and
creates an optimized flow for XSK. A side effect is an opportunity to
optimize the regular RX flow by dropping branching for XSK cases.

Performance improvement is up to 6.4% in the aligned mode and up to 7.5%
in the unaligned mode.

Aligned mode, 2048-byte frames: 12.9 Mpps -> 13.8 Mpps
Aligned mode, 4096-byte frames: 11.8 Mpps -> 12.5 Mpps
Unaligned mode, 2048-byte frames: 11.9 Mpps -> 12.8 Mpps
Unaligned mode, 3072-byte frames: 11.4 Mpps -> 12.1 Mpps
Unaligned mode, 4096-byte frames: 11.0 Mpps -> 11.2 Mpps

CPU: Intel(R) Xeon(R) Gold 6240 CPU @ 2.60GHz

Signed-off-by: Maxim Mikityanskiy <maximmi@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agonet/mlx5e: xsk: Use xsk_buff_alloc_batch on legacy RQ
Maxim Mikityanskiy [Fri, 30 Sep 2022 16:28:56 +0000 (09:28 -0700)]
net/mlx5e: xsk: Use xsk_buff_alloc_batch on legacy RQ

XSK provides a function to allocate frames in batches for more efficient
processing. This commit starts using this function on legacy RQ, adding
a special case for XSK. The new branch introduced basically replaces the
branch that was removed from the same place a few commits before.

A check is made that DMA sync is not needed, because the batching
allocator falls back to returning one frame when DMA sync is needed, and
this is best handled by the loop in the standard case.

Performance improvement is up to 8% in the aligned mode and up to 9% in
the unaligned mode.

Aligned mode, 2048-byte frames: 12.8 Mpps -> 13.5 Mpps
Aligned mode, 4096-byte frames: 11.5 Mpps -> 12.4 Mpps
Unaligned mode, 2048-byte frames: 12.2 Mpps -> 13.4 Mpps
Unaligned mode, 3072-byte frames: 11.6 Mpps -> 12.5 Mpps
Unaligned mode, 4096-byte frames: 11.2 Mpps -> 12.2 Mpps

CPU: Intel(R) Xeon(R) Gold 6240 CPU @ 2.60GHz

Signed-off-by: Maxim Mikityanskiy <maximmi@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agonet/mlx5e: xsk: Split out WQE allocation for legacy XSK RQ
Maxim Mikityanskiy [Fri, 30 Sep 2022 16:28:55 +0000 (09:28 -0700)]
net/mlx5e: xsk: Split out WQE allocation for legacy XSK RQ

Allocation of XSK frames on legacy RQ may be made more efficient with a
specialized routine that relies on certain assumptions, such as there is
only one fragment, allocation units (XSK frames) are not shared among
multiple packets. It reduces the number of branches both in the XSK code
and in the regular RQ, because with this approach there is only a single
check whether it's an XSK or regular RQ.

Signed-off-by: Maxim Mikityanskiy <maximmi@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agonet/mlx5e: Remove the outer loop when allocating legacy RQ WQEs
Maxim Mikityanskiy [Fri, 30 Sep 2022 16:28:54 +0000 (09:28 -0700)]
net/mlx5e: Remove the outer loop when allocating legacy RQ WQEs

Legacy RQ WQEs are allocated in a loop in small batches (8 WQEs). As
partial batches are allowed, there is no point to have a loop in a loop,
so the outer loop is removed, and the batch size is increased up to the
total number of WQEs to allocate, still not smaller than 8.

Signed-off-by: Maxim Mikityanskiy <maximmi@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agonet/mlx5e: xsk: Use partial batches in legacy RQ with XSK
Maxim Mikityanskiy [Fri, 30 Sep 2022 16:28:53 +0000 (09:28 -0700)]
net/mlx5e: xsk: Use partial batches in legacy RQ with XSK

The previous commit allowed allocating WQE batches in legacy RQ
partially, however, XSK still checks whether there are enough frames in
the fill ring. Remove this check to allow to allocate batches partially
also with XSK.

Signed-off-by: Maxim Mikityanskiy <maximmi@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agonet/mlx5e: Use partial batches in legacy RQ
Maxim Mikityanskiy [Fri, 30 Sep 2022 16:28:52 +0000 (09:28 -0700)]
net/mlx5e: Use partial batches in legacy RQ

Legacy RQ allocates WQEs in batches. If the batch allocation fails, the
pages of the allocated part are released. This commit changes this
behavior to allow to use the pages that have been already allocated.

After this change, we need to be careful about indexing rq->wqe.frags[].
The WQ size is a power of two that divides by wqe_bulk (8), and the old
code used whole bulks, which allowed to use indices [8*K; 8*K+7] without
overflowing. Now that the bulks may be partial, the range can start at
any location (not only at 8*K), so we need to wrap them around to avoid
out-of-bounds array access.

Signed-off-by: Maxim Mikityanskiy <maximmi@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agonet/mlx5e: Make the wqe_index_mask calculation more exact
Maxim Mikityanskiy [Fri, 30 Sep 2022 16:28:51 +0000 (09:28 -0700)]
net/mlx5e: Make the wqe_index_mask calculation more exact

The old calculation of wqe_index_mask may give false positives, i.e.
request bulking of pairs of WQEs when not strictly needed, for example,
when the first fragment size is equal to the PAGE_SIZE, bulking is not
needed, even if the number of fragments is odd.

Make the calculation more exact to cut false positives.

Signed-off-by: Maxim Mikityanskiy <maximmi@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agonet/mlx5e: Introduce wqe_index_mask for legacy RQ
Maxim Mikityanskiy [Fri, 30 Sep 2022 16:28:50 +0000 (09:28 -0700)]
net/mlx5e: Introduce wqe_index_mask for legacy RQ

When fragments of different WQEs share the same page, mlx5e_post_rx_wqes
must wait until the old WQE stops using the page, only then the new WQE
can allocate the new page. Essentially, it means that if WQE index i is
still in use, the allocation must stop before `i % bulk`, where bulk is
the number of WQEs that may share the same page.

As bulk is always a power of two, `i % bulk = i & (bulk - 1)`, and the
new wqe_index_mask field will be equal to `bulk - 1`.

At the same time, wqe_bulk remains for optimization purposes and stores
`max(bulk, 8)`, which allows to skip the allocation until we have at
least 8 WQEs free.

Signed-off-by: Maxim Mikityanskiy <maximmi@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agonet/mlx5e: xsk: Drop the check for XSK state in mlx5e_xsk_wakeup
Maxim Mikityanskiy [Fri, 30 Sep 2022 16:28:49 +0000 (09:28 -0700)]
net/mlx5e: xsk: Drop the check for XSK state in mlx5e_xsk_wakeup

The MLX5E_CHANNEL_STATE_XSK flag checked in mlx5e_xsk_wakeup indicates
that XSK queues are open, but not necessarily activated. This check is
not very useful, because:

0. Both XSK setup and netdev state transitions take the same state_lock
mutex, so they can't happen at the same time.

1. If the netdev is up, xsk_is_bound can return true only when
MLX5E_CHANNEL_STATE_XSK is set on the corresponding channel.
mlx5e_xsk_wakeup is only called when xsk_is_bound is true.

2. If the XSK socket is bound, and the netdev is going up or down,
mlx5e_xsk_wakeup can take one of two branches, depending on the return
value of napi_if_scheduled_mark_missed:

2.1. True means one of two things: either NAPI was enabled at this
point, which means MLX5E_CHANNEL_STATE_XSK was also set; or NAPI was
disabled, and nothing really happened.

2.2. False means that NAPI was enabled by this point, which also implies
MLX5E_CHANNEL_STATE_XSK was set. Additionally, mlx5e_xsk_wakeup contains
a following check for MLX5E_SQ_STATE_ENABLED on async_icosq, and this
flag implies MLX5E_CHANNEL_STATE_XSK too on XSK channels.

As checking this flag doesn't cut any flows, remove the check.

Signed-off-by: Maxim Mikityanskiy <maximmi@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agonet/mlx5e: xsk: Use mlx5e_trigger_napi_icosq for XSK wakeup
Maxim Mikityanskiy [Fri, 30 Sep 2022 16:28:48 +0000 (09:28 -0700)]
net/mlx5e: xsk: Use mlx5e_trigger_napi_icosq for XSK wakeup

mlx5e_xsk_wakeup triggers an IRQ by posting a NOP to async_icosq, taking
a spinlock to protect from concurrent access. There is already a
function that does the same: mlx5e_trigger_napi_icosq. Use this function
in mlx5e_xsk_wakeup.

Signed-off-by: Maxim Mikityanskiy <maximmi@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agoMerge branch 'nfp-support-fec-mode-reporting-and-auto-neg'
Jakub Kicinski [Sat, 1 Oct 2022 02:03:49 +0000 (19:03 -0700)]
Merge branch 'nfp-support-fec-mode-reporting-and-auto-neg'

Simon Horman says:

====================
nfp: support FEC mode reporting and auto-neg

this series adds support for the following features to the nfp driver:

* Patch 1/5: Support active FEC mode
* Patch 2/5: Don't halt driver on non-fatal error when interacting with fw
* Patch 3/5: Treat port independence as a firmware rather than port property
* Patch 4/5: Support link auto negotiation
* Patch 5/5: Support restart of link auto negotiation
====================

Link: https://lore.kernel.org/r/20220929085832.622510-1-simon.horman@corigine.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agonfp: add support restart of link auto-negotiation
Fei Qin [Thu, 29 Sep 2022 08:58:32 +0000 (10:58 +0200)]
nfp: add support restart of link auto-negotiation

Add support restart of link auto-negotiation.
This may be initiated using:

  # ethtool -r <intf>

Signed-off-by: Fei Qin <fei.qin@corigine.com>
Signed-off-by: Simon Horman <simon.horman@corigine.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agonfp: add support for link auto negotiation
Yinjun Zhang [Thu, 29 Sep 2022 08:58:31 +0000 (10:58 +0200)]
nfp: add support for link auto negotiation

Report the auto negotiation capability if it's supported
in management firmware, and advertise it if it's enabled.
Changing port speed is not allowed when autoneg is enabled.

The ethtool <intf> command displays the auto-neg capability:

  # ethtool enp1s0np0
  Settings for enp1s0np0:
          Supported ports: [ FIBRE ]
          Supported link modes:   Not reported
          Supported pause frame use: Symmetric
          Supports auto-negotiation: Yes
          Supported FEC modes: None        RS      BASER
          Advertised link modes:  Not reported
          Advertised pause frame use: Symmetric
          Advertised auto-negotiation: Yes
          Advertised FEC modes: None       RS      BASER
          Speed: 25000Mb/s
          Duplex: Full
          Auto-negotiation: on
          Port: FIBRE
          PHYAD: 0
          Transceiver: internal
          Link detected: yes

Signed-off-by: Yinjun Zhang <yinjun.zhang@corigine.com>
Signed-off-by: Simon Horman <simon.horman@corigine.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agonfp: refine the ABI of getting `sp_indiff` info
Yinjun Zhang [Thu, 29 Sep 2022 08:58:30 +0000 (10:58 +0200)]
nfp: refine the ABI of getting `sp_indiff` info

Considering that whether application firmware is indifferent to
port speed is a firmware property instead of port property, now use
a new rtsym to get the property instead of parsing per-port tlv caps.
With this change, relevant code is moved to `nfp_main` layer.

Signed-off-by: Yinjun Zhang <yinjun.zhang@corigine.com>
Signed-off-by: Simon Horman <simon.horman@corigine.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agonfp: avoid halt of driver init process when non-fatal error happens
Yinjun Zhang [Thu, 29 Sep 2022 08:58:29 +0000 (10:58 +0200)]
nfp: avoid halt of driver init process when non-fatal error happens

It's not a fatal error when setting `hwinfo` into management firmware
fails, no need to halt the whole driver initialization process.

Signed-off-by: Yinjun Zhang <yinjun.zhang@corigine.com>
Signed-off-by: Simon Horman <simon.horman@corigine.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agonfp: add support for reporting active FEC mode
Yinjun Zhang [Thu, 29 Sep 2022 08:58:28 +0000 (10:58 +0200)]
nfp: add support for reporting active FEC mode

The latest management firmware can now report the active FEC
mode. Adapt driver accordingly so that user can get the active
FEC mode by running command:

  # ethtool --show-fec <intf>

Also correct use of `fec` field.

Signed-off-by: Yinjun Zhang <yinjun.zhang@corigine.com>
Signed-off-by: Simon Horman <simon.horman@corigine.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agoselftests/tc-testing: update qdisc/cls/action features in config
Zhengchao Shao [Thu, 29 Sep 2022 04:19:09 +0000 (12:19 +0800)]
selftests/tc-testing: update qdisc/cls/action features in config

Since three patchsets "add tc-testing test cases", "refactor duplicate
codes in the tc cls walk function", and "refactor duplicate codes in the
qdisc class walk function" are merged to net-next tree, the list of
supported features needs to be updated in config file.

Signed-off-by: Zhengchao Shao <shaozhengchao@huawei.com>
Reviewed-by: Victor Nogueira <victor@mojatatu.com>
Acked-by: Jamal Hadi Salim <jhs@mojatatu.com>
Link: https://lore.kernel.org/r/20220929041909.83913-1-shaozhengchao@huawei.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agodt-bindings: nfc: marvell,nci: fix reset line polarity in examples
Dmitry Torokhov [Thu, 29 Sep 2022 20:22:55 +0000 (13:22 -0700)]
dt-bindings: nfc: marvell,nci: fix reset line polarity in examples

The reset line is supposed to be "active low" (it even says so in the
description), but examples incorrectly show it as "active high"
(likely because original examples use 0 which is technically "active
high" but in practice often "don't care" if the driver is using legacy
gpio API, as this one does).

Signed-off-by: Dmitry Torokhov <dmitry.torokhov@gmail.com>
Reviewed-by: Krzysztof Kozlowski <krzysztof.kozlowski@linaro.org>
Link: https://lore.kernel.org/r/YzX+nzJolxAKmt+z@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agoMerge branch 'devlink-sanitize-per-port-region-creation-destruction'
Jakub Kicinski [Sat, 1 Oct 2022 01:17:22 +0000 (18:17 -0700)]
Merge branch 'devlink-sanitize-per-port-region-creation-destruction'

Jiri Pirko says:

====================
devlink: sanitize per-port region creation/destruction

Currently the only user of per-port devlink regions is DSA. All drivers
that use regions register them before devlink registration. For DSA,
this was not possible as the internals of struct devlink_port needed for
region creation are initialized during port registration.

This introduced a mismatch in creation flow of devlink and devlink port
regions. As you can see, it causes the DSA driver to make the port
init/exit flow a bit cumbersome.

Fix this by introducing port_init/fini() which could be optionally
called by drivers like DSA, to prepare struct devlink_port to be used
for region creation purposes before devlink port register is called.

Tested by Vladimir on his setup.
====================

Link: https://lore.kernel.org/r/20220929072902.2986539-1-jiri@resnulli.us
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agonet: dsa: remove bool devlink_port_setup
Vladimir Oltean [Thu, 29 Sep 2022 07:29:02 +0000 (09:29 +0200)]
net: dsa: remove bool devlink_port_setup

Since dsa_port_devlink_setup() and dsa_port_devlink_teardown() are
already called from code paths which only execute once per port (due to
the existing bool dp->setup), keeping another dp->devlink_port_setup is
redundant, because we can already manage to balance the calls properly
(and not call teardown when setup was never called, or call setup twice,
or things like that).

Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: Jiri Pirko <jiri@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agonet: dsa: don't do devlink port setup early
Jiri Pirko [Thu, 29 Sep 2022 07:29:01 +0000 (09:29 +0200)]
net: dsa: don't do devlink port setup early

Commit 3122433eb533 ("net: dsa: Register devlink ports before calling DSA driver setup()")
moved devlink port setup to be done early before driver setup()
was called. That is no longer needed, so move the devlink port
initialization back to dsa_port_setup(), as the first thing done there.

Note there is no longer needed to reinit port as unused if
dsa_port_setup() fails, as it unregisters the devlink port instance on
the error path.

Signed-off-by: Jiri Pirko <jiri@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agonet: dsa: don't leave dangling pointers in dp->pl when failing
Vladimir Oltean [Thu, 29 Sep 2022 07:29:00 +0000 (09:29 +0200)]
net: dsa: don't leave dangling pointers in dp->pl when failing

There is a desire to simplify the dsa_port registration path with
devlink, and this involves reworking a bit how user ports which fail to
connect to their PHY (because it's missing) get reinitialized as UNUSED
devlink ports.

The desire is for the change to look something like this; basically
dsa_port_setup() has failed, we just change dp->type and call
dsa_port_setup() again.

-/* Destroy the current devlink port, and create a new one which has the UNUSED
- * flavour.
- */
-static int dsa_port_reinit_as_unused(struct dsa_port *dp)
+static int dsa_port_setup_as_unused(struct dsa_port *dp)
 {
- dsa_port_devlink_teardown(dp);
  dp->type = DSA_PORT_TYPE_UNUSED;
- return dsa_port_devlink_setup(dp);
+ return dsa_port_setup(dp);
 }

For an UNUSED port, dsa_port_setup() mostly only calls dsa_port_devlink_setup()
anyway, so we could get away with calling just that. But if we call the
full blown dsa_port_setup(dp) (which will be needed to properly set
dp->setup = true), the callee will have the tendency to go through this
code block too, and call dsa_port_disable(dp):

switch (dp->type) {
case DSA_PORT_TYPE_UNUSED:
dsa_port_disable(dp);
break;

That is not very good, because dsa_port_disable() has this hidden inside
of it:

if (dp->pl)
phylink_stop(dp->pl);

Fact is, we are not prepared to handle a call to dsa_port_disable() with
a struct dsa_port that came from a previous (and failed) call to
dsa_port_setup(). We do not clean up dp->pl, and this will make the
second call to dsa_port_setup() call phylink_stop() on a dangling dp->pl
pointer.

Solve this by creating an API for phylink destruction which is symmetric
to the phylink creation, and never leave dp->pl set to anything except
NULL or a valid phylink structure.

Signed-off-by: Vladimir Oltean <vladimir.oltean@nxp.com>
Signed-off-by: Jiri Pirko <jiri@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agonet: dsa: move port_setup/teardown to be called outside devlink port registered area
Jiri Pirko [Thu, 29 Sep 2022 07:28:59 +0000 (09:28 +0200)]
net: dsa: move port_setup/teardown to be called outside devlink port registered area

Move port_setup() op to be called before devlink_port_register() and
port_teardown() after devlink_port_unregister().

Note it makes sense to move this alongside the rest of the devlink port
code, the reinit() function also gets much nicer, as clearly the fact that
port_setup()->devlink_port_region_create() was called in dsa_port_setup
did not fit the flow.

Signed-off-by: Jiri Pirko <jiri@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agonet: devlink: add port_init/fini() helpers to allow pre-register/post-unregister...
Jiri Pirko [Thu, 29 Sep 2022 07:28:58 +0000 (09:28 +0200)]
net: devlink: add port_init/fini() helpers to allow pre-register/post-unregister functions

Lifetime of some of the devlink objects, like regions, is currently
forced to be different for devlink instance and devlink port instance
(per-port regions). The reason is that for devlink ports, the internal
structures initialization happens only after devlink_port_register() is
called.

To resolve this inconsistency, introduce new set of helpers to allow
driver to initialize devlink pointer and region list before
devlink_register() is called. That allows port regions to be created
before devlink port registration and destroyed after devlink
port unregistration.

Signed-off-by: Jiri Pirko <jiri@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agonet: devlink: introduce a flag to indicate devlink port being registered
Jiri Pirko [Thu, 29 Sep 2022 07:28:57 +0000 (09:28 +0200)]
net: devlink: introduce a flag to indicate devlink port being registered

Instead of relying on devlink pointer not being initialized, introduce
an extra flag to indicate if devlink port is registered. This is needed
as later on devlink pointer is going to be initialized even in case
devlink port is not registered yet.

Signed-off-by: Jiri Pirko <jiri@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agonet: devlink: introduce port registered assert helper and use it
Jiri Pirko [Thu, 29 Sep 2022 07:28:56 +0000 (09:28 +0200)]
net: devlink: introduce port registered assert helper and use it

Instead of checking devlink_port->devlink pointer for not being NULL
which indicates that devlink port is registered, put this check to new
pair of helpers similar to what we have for devlink and use them in
other functions.

Signed-off-by: Jiri Pirko <jiri@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agor8169: add rtl_disable_rxdvgate()
Chunhao Lin [Wed, 28 Sep 2022 17:13:56 +0000 (01:13 +0800)]
r8169: add rtl_disable_rxdvgate()

rtl_disable_rxdvgate() is used for disable RXDV_GATE. It is opposite function
of rtl_enable_rxdvgate().

Disable RXDV_GATE does not have to delay. So in this patch, also remove the
delay after disale RXDV_GATE.

Signed-off-by: Chunhao Lin <hau@realtek.com>
Link: https://lore.kernel.org/r/20220928171356.3951-1-hau@realtek.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agoMerge tag 'for-net-next-2022-09-30' of git://git.kernel.org/pub/scm/linux/kernel...
Jakub Kicinski [Sat, 1 Oct 2022 01:02:18 +0000 (18:02 -0700)]
Merge tag 'for-net-next-2022-09-30' of git://git./linux/kernel/git/bluetooth/bluetooth-next

Luiz Augusto von Dentz says:

====================
bluetooth-next pull request for net-next

 - Add RTL8761BUV device (Edimax BT-8500)
 - Add a new PID/VID 13d3/3583 for MT7921
 - Add Realtek RTL8852C support ID 0x13D3:0x3592
 - Add VID/PID 0489/e0e0 for MediaTek MT7921
 - Add a new VID/PID 0e8d/0608 for MT7921
 - Add a new PID/VID 13d3/3578 for MT7921
 - Add BT device 0cb8:c549 from RTW8852AE
 - Add support for Intel Magnetor

* tag 'for-net-next-2022-09-30' of git://git.kernel.org/pub/scm/linux/kernel/git/bluetooth/bluetooth-next: (49 commits)
  Bluetooth: hci_sync: Fix not indicating power state
  Bluetooth: L2CAP: Fix user-after-free
  Bluetooth: Call shutdown for HCI_USER_CHANNEL
  Bluetooth: Prevent double register of suspend
  Bluetooth: hci_core: Fix not handling link timeouts propertly
  Bluetooth: hci_event: Make sure ISO events don't affect non-ISO connections
  Bluetooth: hci_debugfs: Fix not checking conn->debugfs
  Bluetooth: hci_sysfs: Fix attempting to call device_add multiple times
  Bluetooth: MGMT: fix zalloc-simple.cocci warnings
  Bluetooth: hci_{ldisc,serdev}: check percpu_init_rwsem() failure
  Bluetooth: use hdev->workqueue when queuing hdev->{cmd,ncmd}_timer works
  Bluetooth: L2CAP: initialize delayed works at l2cap_chan_create()
  Bluetooth: RFCOMM: Fix possible deadlock on socket shutdown/release
  Bluetooth: hci_sync: allow advertise when scan without RPA
  Bluetooth: btusb: Add a new VID/PID 0e8d/0608 for MT7921
  Bluetooth: btusb: Add a new PID/VID 13d3/3583 for MT7921
  Bluetooth: avoid hci_dev_test_and_set_flag() in mgmt_init_hdev()
  Bluetooth: btintel: Mark Intel controller to support LE_STATES quirk
  Bluetooth: btintel: Add support for Magnetor
  Bluetooth: btusb: Add a new PID/VID 13d3/3578 for MT7921
  ...
====================

Link: https://lore.kernel.org/r/20221001004602.297366-1-luiz.dentz@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agogenetlink: reject use of nlmsg_flags for new commands
Jakub Kicinski [Thu, 29 Sep 2022 14:28:09 +0000 (07:28 -0700)]
genetlink: reject use of nlmsg_flags for new commands

Commit 9c5d03d36251 ("genetlink: start to validate reserved header bytes")
introduced extra validation for genetlink headers. We had to gate it
to only apply to new commands, to maintain bug-wards compatibility.
Use this opportunity (before the new checks make it to Linus's tree)
to add more conditions.

Validate that Generic Netlink families do not use nlmsg_flags outside
of the well-understood set.

Link: https://lore.kernel.org/all/20220928073709.1b93b74a@kernel.org/
Reviewed-by: Johannes Berg <johannes@sipsolutions.net>
Acked-by: Nikolay Aleksandrov <razor@blackwall.org>
Reviewed-by: Nicolas Dichtel <nicolas.dichtel@6wind.com>
Reviewed-by: Guillaume Nault <gnault@redhat.com>
Reviewed-by: Jacob Keller <jacob.e.keller@intel.com>
Link: https://lore.kernel.org/r/20220929142809.1167546-1-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agoBluetooth: hci_sync: Fix not indicating power state
Luiz Augusto von Dentz [Fri, 30 Sep 2022 20:12:30 +0000 (13:12 -0700)]
Bluetooth: hci_sync: Fix not indicating power state

When setting power state using legacy/non-mgmt API
(e.g hcitool hci0 up) the likes of mgmt_set_powered_complete won't be
called causing clients of the MGMT API to not be notified of the change
of the state.

Fixes: cf75ad8b41d2 ("Bluetooth: hci_sync: Convert MGMT_SET_POWERED")
Signed-off-by: Luiz Augusto von Dentz <luiz.von.dentz@intel.com>
Tested-by: Tedd Ho-Jeong An <tedd.an@intel.com>
2 years agoMerge tag 'wireless-next-2022-09-30' of git://git.kernel.org/pub/scm/linux/kernel...
Jakub Kicinski [Fri, 30 Sep 2022 17:07:30 +0000 (10:07 -0700)]
Merge tag 'wireless-next-2022-09-30' of git://git./linux/kernel/git/wireless/wireless-next

Kalle Valo says:

====================
wireless-next patches for v6.1

Few stack changes and lots of driver changes in this round. brcmfmac
has more activity as usual and it gets new hardware support. ath11k
improves WCN6750 support and also other smaller features. And of
course changes all over.

Note: in early September wireless tree was merged to wireless-next to
avoid some conflicts with mac80211 patches, this shouldn't cause any
problems but wanted to mention anyway.

Major changes:

mac80211

 - refactoring and preparation for Wi-Fi 7 Multi-Link Operation (MLO)
  feature continues

brcmfmac

 - support CYW43439 SDIO chipset

 - support BCM4378 on Apple platforms

 - support CYW89459 PCIe chipset

rtw89

 - more work to get rtw8852c supported

 - P2P support

 - support for enabling and disabling MSDU aggregation via nl80211

mt76

 - tx status reporting improvements

ath11k

 - cold boot calibration support on WCN6750

 - Target Wake Time (TWT) debugfs support for STA interface

 - support to connect to a non-transmit MBSSID AP profile

 - enable remain-on-channel support on WCN6750

 - implement SRAM dump debugfs interface

 - enable threaded NAPI on all hardware

 - WoW support for WCN6750

 - support to provide transmit power from firmware via nl80211

 - support to get power save duration for each client

 - spectral scan support for 160 MHz

wcn36xx

 - add SNR from a received frame as a source of system entropy

* tag 'wireless-next-2022-09-30' of git://git.kernel.org/pub/scm/linux/kernel/git/wireless/wireless-next: (231 commits)
  wifi: rtl8xxxu: Improve rtl8xxxu_queue_select
  wifi: rtl8xxxu: Fix AIFS written to REG_EDCA_*_PARAM
  wifi: rtl8xxxu: gen2: Enable 40 MHz channel width
  wifi: rtw89: 8852b: configure DLE mem
  wifi: rtw89: check DLE FIFO size with reserved size
  wifi: rtw89: mac: correct register of report IMR
  wifi: rtw89: pci: set power cut closed for 8852be
  wifi: rtw89: pci: add to do PCI auto calibration
  wifi: rtw89: 8852b: implement chip_ops::{enable,disable}_bb_rf
  wifi: rtw89: add DMA busy checking bits to chip info
  wifi: rtw89: mac: define DMA channel mask to avoid unsupported channels
  wifi: rtw89: pci: mask out unsupported TX channels
  iwlegacy: Replace zero-length arrays with DECLARE_FLEX_ARRAY() helper
  ipw2x00: Replace zero-length array with DECLARE_FLEX_ARRAY() helper
  wifi: iwlwifi: Track scan_cmd allocation size explicitly
  brcmfmac: Remove the call to "dtim_assoc" IOVAR
  brcmfmac: increase dcmd maximum buffer size
  brcmfmac: Support 89459 pcie
  brcmfmac: increase default max WOWL patterns to 16
  cw1200: fix incorrect check to determine if no element is found in list
  ...
====================

Link: https://lore.kernel.org/r/20220930150413.A7984C433D6@smtp.kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agoMerge branch 'mlx5-xsk-updates-part2-2022-09-28'
Jakub Kicinski [Fri, 30 Sep 2022 14:55:55 +0000 (07:55 -0700)]
Merge branch 'mlx5-xsk-updates-part2-2022-09-28'

Saeed Mahameed says:

====================
mlx5 xsk updates part2 2022-09-28

XSK buffer improvements, This is part #2 of 4 parts series.

 1) Expose xsk min chunk size to drivers, to allow the driver to adjust to a
   better buffer stride size

 2) Adjust MTT page size to the XSK frame size, to avoid umem overrun in
  certain situations.

 3) Use xsk frame size as the striding RQ page size for XSK RQs

 4) KSM for unaligned XSK, KSM allows arbitrary buffer chunk lengths
    registration in HW, which makes more sense for unaligned XSK.

 4) More cleanups and optimizations in preparation for next improvements
    in part3

part 1: https://lore.kernel.org/netdev/20220927203611.244301-1-saeed@kernel.org/
====================

Link: https://lore.kernel.org/r/20220929072156.93299-1-saeed@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agonet/mlx5e: Clean up and fix error flows in mlx5e_alloc_rq
Maxim Mikityanskiy [Thu, 29 Sep 2022 07:21:56 +0000 (00:21 -0700)]
net/mlx5e: Clean up and fix error flows in mlx5e_alloc_rq

Although mlx5e_rq_free_shampo can be called unconditionally, it belongs
to case MLX5_WQ_TYPE_LINKED_LIST_STRIDING_RQ. Move it there to allow to
add more init/cleanup actions to the striding RQ case.

If xdp_rxq_info_reg_mem_model fails, don't forget to destroy the page
pool.

Signed-off-by: Maxim Mikityanskiy <maximmi@nvidia.com>
Reviewed-by: Saeed Mahameed <saeedm@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agonet/mlx5e: Move repeating clear_bit in mlx5e_rx_reporter_err_rq_cqe_recover
Maxim Mikityanskiy [Thu, 29 Sep 2022 07:21:55 +0000 (00:21 -0700)]
net/mlx5e: Move repeating clear_bit in mlx5e_rx_reporter_err_rq_cqe_recover

The same clear_bit is called in both error and success flows. Move the
call to do it only once and remove the out label.

Signed-off-by: Maxim Mikityanskiy <maximmi@nvidia.com>
Reviewed-by: Saeed Mahameed <saeedm@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agonet/mlx5e: Split out channel (de)activation in rx_res
Maxim Mikityanskiy [Thu, 29 Sep 2022 07:21:54 +0000 (00:21 -0700)]
net/mlx5e: Split out channel (de)activation in rx_res

To decrease the nesting level and reduce duplication of code, create
functions to redirect direct RQTs to the actual RQs or drop_rq, which
are used in the activation and deactivation flows of channels.

Signed-off-by: Maxim Mikityanskiy <maximmi@nvidia.com>
Reviewed-by: Saeed Mahameed <saeedm@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agonet/mlx5e: xsk: Remove mlx5e_xsk_page_alloc_pool
Maxim Mikityanskiy [Thu, 29 Sep 2022 07:21:53 +0000 (00:21 -0700)]
net/mlx5e: xsk: Remove mlx5e_xsk_page_alloc_pool

mlx5e_xsk_page_alloc_pool became a thin wrapper around xsk_buff_alloc.
Drop it and call xsk_buff_alloc directly.

Signed-off-by: Maxim Mikityanskiy <maximmi@nvidia.com>
Reviewed-by: Saeed Mahameed <saeedm@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agonet/mlx5e: Convert struct mlx5e_alloc_unit to a union
Maxim Mikityanskiy [Thu, 29 Sep 2022 07:21:52 +0000 (00:21 -0700)]
net/mlx5e: Convert struct mlx5e_alloc_unit to a union

struct mlx5e_alloc_unit consists of a single union. Convert it to a
union itself to simplify casting it to struct xdp_buff *, which will be
used to implement XSK batching on striding RQ.

Signed-off-by: Maxim Mikityanskiy <maximmi@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agonet/mlx5e: Remove DMA address from mlx5e_alloc_unit
Maxim Mikityanskiy [Thu, 29 Sep 2022 07:21:51 +0000 (00:21 -0700)]
net/mlx5e: Remove DMA address from mlx5e_alloc_unit

mlx5e_alloc_unit stores the DMA address and a pointer to either struct
page (regular RQ) or struct xdp_buff (XSK RQ). This DMA address is
redundant, because when a page or an XSK frame is allocated, the same
address is also stored there. Some flows take the address from struct
mlx5e_alloc_unit, and some take it from struct page or xdp_buff.

This commit removes the address from struct mlx5e_alloc_unit, which
makes it twice as small and improves locality (this struct is used in an
array), also saving on unnecessary stores to the addr field. Almost all
flows know unambiguously whether the DMA address should be taken from
page or from xdp_buff. The exception is the allocation flows, where a
new branch appeared, which will be optimized out in the next commits.

struct mlx5e_alloc_unit used to be called mlx5e_dma_info.

Signed-off-by: Maxim Mikityanskiy <maximmi@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agonet/mlx5e: Rename mlx5e_dma_info to prepare for removal of DMA address
Maxim Mikityanskiy [Thu, 29 Sep 2022 07:21:50 +0000 (00:21 -0700)]
net/mlx5e: Rename mlx5e_dma_info to prepare for removal of DMA address

The next commit will remove the DMA address from the struct currently
called mlx5e_dma_info, because the same value can be retrieved with
page_pool_get_dma_addr(page) in almost all cases, with the notable
exception of SHAMPO (HW GRO implementation) that modifies this address
on the fly, after the initial allocation.

To keep the SHAMPO logic intact, struct mlx5e_dma_info remains in the
SHAMPO code, consisting of addr and page (XSK is not compatible with
SHAMPO). The struct used in all other places is renamed to
mlx5e_alloc_unit, allowing the next commit to remove the addr field
without affecting SHAMPO.

The new name means "allocation unit", and it's more appropriate after
the field with the DMA address gets removed.

Signed-off-by: Maxim Mikityanskiy <maximmi@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agonet/mlx5e: Optimize the page cache reducing its size 2x
Maxim Mikityanskiy [Thu, 29 Sep 2022 07:21:49 +0000 (00:21 -0700)]
net/mlx5e: Optimize the page cache reducing its size 2x

RX page cache stores dma_info structs, that consist of a pointer to
struct page and a DMA address. In fact, the DMA address is extracted
from struct page using page_pool_get_dma_addr when a page is pushed to
the cache. By moving this call to the point when a page is popped from
the cache, we can avoid storing the DMA address in the cache,
effectively reducing its size by two times without losing any
functionality.

Signed-off-by: Maxim Mikityanskiy <maximmi@nvidia.com>
Reviewed-by: Saeed Mahameed <saeedm@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agonet/mlx5e: Fix calculations for ICOSQ size
Maxim Mikityanskiy [Thu, 29 Sep 2022 07:21:48 +0000 (00:21 -0700)]
net/mlx5e: Fix calculations for ICOSQ size

WQEs must not cross page boundaries, they are padded with NOPs if they
don't fit the page. mlx5e_mpwrq_total_umr_wqebbs doesn't take into
account this padding, risking reserving not enough space.

The padding is not straightforward to add to this calculation, because
WQEs of different sizes may be mixed together in the queue. If each page
ends with a big WQE that doesn't fit and requires at most its size minus
1 WQEBB of padding, the total space can be much bigger than in case when
smaller WQEs take advantage of this padding.

Replace the wrong exact calculation by the following estimation. Each
padding can be at most the size of the maximum WQE used in the queue
minus one WQEBB. Let's call the rest of the page "useful space". If we
divide the total size of all needed WQEs by this useful space, rounding
up, we'll get the number of pages, which is enough to contain all these
WQEs. It's correct, because every WQE that appeared on the boundary
between two blocks of useful space would start in the useful space of
one page and end in the padding of the same page, while our estimation
reserved space for its tail in the next space, making the estimation not
smaller than the real space occupied in the queue.

The code actually uses a looser estimation: instead of taking the
maximum size of all used WQE types minus 1 WQEBB, it takes the maximum
hardware size of a WQE. It's made for simplicity and extensibility.

Signed-off-by: Maxim Mikityanskiy <maximmi@nvidia.com>
Reviewed-by: Saeed Mahameed <saeedm@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agoxsk: Remove unused xsk_buff_discard
Maxim Mikityanskiy [Thu, 29 Sep 2022 07:21:47 +0000 (00:21 -0700)]
xsk: Remove unused xsk_buff_discard

The previous commit removed the last usage of xsk_buff_discard in mlx5e,
so the function that is no longer used can be removed.

Signed-off-by: Maxim Mikityanskiy <maximmi@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
CC: "Björn Töpel" <bjorn@kernel.org>
CC: Magnus Karlsson <magnus.karlsson@intel.com>
CC: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agonet/mlx5e: xsk: Use KSM for unaligned XSK
Maxim Mikityanskiy [Thu, 29 Sep 2022 07:21:46 +0000 (00:21 -0700)]
net/mlx5e: xsk: Use KSM for unaligned XSK

UMR MTTs used in striding RQ have certain alignment requirements. While
it's guaranteed to work when UMR pages are aligned to the UMR page size,
in practice it works then UMR pages are aligned to 8 bytes. However,
it's still not enough flexibility for the unaligned mode of XSK. This
patch leverages KSM to map UMR pages without alignment requirements,
when unaligned XSK is active. The downside is that KSM entries are twice
as big as MTTs, which limits the maximum WQE size, so regular RQs and
aligned XSK continue using MTTs.

Signed-off-by: Maxim Mikityanskiy <maximmi@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agonet/mlx5: Add MLX5_FLEXIBLE_INLEN to safely calculate cmd inlen
Maxim Mikityanskiy [Thu, 29 Sep 2022 07:21:45 +0000 (00:21 -0700)]
net/mlx5: Add MLX5_FLEXIBLE_INLEN to safely calculate cmd inlen

Some commands use a flexible array after a common header. Add a macro to
safely calculate the total input length of the command, detecting
overflows and printing errors with specific values when such overflows
happen.

Signed-off-by: Maxim Mikityanskiy <maximmi@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agonet/mlx5e: Keep a separate MKey for striding RQ
Maxim Mikityanskiy [Thu, 29 Sep 2022 07:21:44 +0000 (00:21 -0700)]
net/mlx5e: Keep a separate MKey for striding RQ

Currently, rq->mkey_be keeps a big-endian value of either the PA MKey
(for legacy RQ, no address translation) or MTT MKey (for striding RQ,
direct address translation). Striding RQ stores the same value in
rq->umr_mkey in the native endianness.

The next commit will make striding RQ use KSM MKey (indirect address
translation) for the unaligned mode of XSK, which will require storing
both KSM MKey and PA MKey in the RQ struct. This commit optimizes fields
of mlx5e_rq: umr_mkey is removed (it's redundant), mkey_be always points
to the PA MKey, and mpwqe.umr_mkey_be points to the MTT MKey (or to the
KSM MKey, starting from the next commit).

Signed-off-by: Maxim Mikityanskiy <maximmi@nvidia.com>
Reviewed-by: Saeed Mahameed <saeedm@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agonet/mlx5e: xsk: Use XSK frame size as striding RQ page size
Maxim Mikityanskiy [Thu, 29 Sep 2022 07:21:43 +0000 (00:21 -0700)]
net/mlx5e: xsk: Use XSK frame size as striding RQ page size

XSK RQs support striding RQ linear mode, but the stride size is always
set to PAGE_SIZE. It may be larger than the XSK frame size,
unnecessarily reducing the useful space in a WQE, but more importantly
causing UMEM data corruption in certain cases.

Normally, stride size bigger than XSK frame size is not a problem if the
hardware enforces the MTU. However, traffic between vports skips the
hardware MTU check, and oversized packets may be received.

If an oversized packet is bigger than the XSK frame but not bigger than
the stride, it will cause overwriting of the adjacent UMEM region. If
the packet takes more than one stride, they can be recycled for reuse
so it's not a problem when the XSK frame size matches the stride size.

To reduce the impact of the above issue, attempt to use the MTT page
size for striding RQ that matches the XSK frame size, allowing to safely
use 2048-byte frames on an up-to-date firmware.

Signed-off-by: Maxim Mikityanskiy <maximmi@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agonet/mlx5e: Use runtime page_shift for striding RQ
Maxim Mikityanskiy [Thu, 29 Sep 2022 07:21:42 +0000 (00:21 -0700)]
net/mlx5e: Use runtime page_shift for striding RQ

This commit allows striding RQ to determine MTT page size at runtime,
instead of sticking to the compile-time PAGE_SIZE. This functionality
will be used by a following commit that adjusts the MTT page size to the
XSK frame size.

Stick with PAGE_SIZE for XSK on legacy RQ, as frag_stride is not used in
data path, it only helps calculate how pages are partitioned into
fragments, and PAGE_SIZE will ensure each fragment starts at the
beginning of a new allocation unit (XSK frame).

Signed-off-by: Maxim Mikityanskiy <maximmi@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agoxsk: Expose min chunk size to drivers
Maxim Mikityanskiy [Thu, 29 Sep 2022 07:21:41 +0000 (00:21 -0700)]
xsk: Expose min chunk size to drivers

Drivers should be aware of the range of valid UMEM chunk sizes to be
able to allocate their internal structures of an appropriate size. It
will be used by mlx5e in the following patches.

Signed-off-by: Maxim Mikityanskiy <maximmi@nvidia.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
CC: "Björn Töpel" <bjorn@kernel.org>
CC: Magnus Karlsson <magnus.karlsson@intel.com>
CC: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Signed-off-by: Saeed Mahameed <saeedm@nvidia.com>
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
2 years agoip6_vti:Remove the space before the comma
Hongbin Wang [Thu, 29 Sep 2022 06:12:05 +0000 (02:12 -0400)]
ip6_vti:Remove the space before the comma

There should be no space before the comma

Signed-off-by: Hongbin Wang <wh_bin@126.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2 years agoMerge branch 'Mediatek-mt8188'
David S. Miller [Fri, 30 Sep 2022 12:04:23 +0000 (13:04 +0100)]
Merge branch 'Mediatek-mt8188'

Jianguo Zhang says:

====================
Mediatek ethernet patches for mt8188

Changes in v7:

v7:
1) Add 'Reviewed-by: AngeloGioacchino Del Regno
<angelogioacchino.delregno@collabora.com>' info in commit message of
patch 'dt-bindings: net: snps,dwmac: add new property snps,clk-csr',
'arm64: dts: mediatek: mt2712e: Update the name of property 'clk_csr''
and 'net: stmmac: add a parse for new property 'snps,clk-csr''.

v6:
1) Update commit message of patch 'dt-bindings: net: snps,dwmac: add new property snps,clk-csr'
2) Add a parse for new property 'snps,clk-csr' in patch
'net: stmmac: add a parse for new property 'snps,clk-csr''

v5:
1) Rename the property 'clk_csr' as 'snps,clk-csr' in binding
file as Krzysztof Kozlowski'comment.
2) Add DTS patch 'arm64: dts: mediatek: mt2712e: Update the name of property 'clk_csr''
as Krzysztof Kozlowski'comment.
3) Add driver patch 'net: stmmac: Update the name of property 'clk_csr''
as Krzysztof Kozlowski'comment.

v4:
1) Update the commit message of patch 'dt-bindings: net: snps,dwmac: add clk_csr property'
as Krzysztof Kozlowski'comment.

v3:
1) List the names of SoCs mt8188 and mt8195 in correct order as
AngeloGioacchino Del Regno's comment.
2) Add patch version info as Krzysztof Kozlowski'comment.

v2:
1) Delete patch 'stmmac: dwmac-mediatek: add support for mt8188' as
Krzysztof Kozlowski's comment.
2) Update patch 'dt-bindings: net: mediatek-dwmac: add support for
mt8188' as Krzysztof Kozlowski's comment.
3) Add clk_csr property to fix warning ('clk_csr' was unexpected) when
runnig 'make dtbs_check'.

v1:
1) Add ethernet driver entry for mt8188.
2) Add binding document for ethernet on mt8188.
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
2 years agonet: stmmac: add a parse for new property 'snps,clk-csr'
Jianguo Zhang [Thu, 29 Sep 2022 01:47:58 +0000 (09:47 +0800)]
net: stmmac: add a parse for new property 'snps,clk-csr'

Parse new property 'snps,clk-csr' firstly because the new property
is documented in binding file, if failed, fall back to old property
'clk_csr' for legacy case

Signed-off-by: Jianguo Zhang <jianguo.zhang@mediatek.com>
Reviewed-by: AngeloGioacchino Del Regno <angelogioacchino.delregno@collabora.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2 years agoarm64: dts: mediatek: mt2712e: Update the name of property 'clk_csr'
Jianguo Zhang [Thu, 29 Sep 2022 01:47:57 +0000 (09:47 +0800)]
arm64: dts: mediatek: mt2712e: Update the name of property 'clk_csr'

Update the name of property 'clk_csr' as 'snps,clk-csr' to align with
the property name in the binding file.

Signed-off-by: Jianguo Zhang <jianguo.zhang@mediatek.com>
Reviewed-by: AngeloGioacchino Del Regno <angelogioacchino.delregno@collabora.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2 years agodt-bindings: net: snps,dwmac: add new property snps,clk-csr
Jianguo Zhang [Thu, 29 Sep 2022 01:47:56 +0000 (09:47 +0800)]
dt-bindings: net: snps,dwmac: add new property snps,clk-csr

Add description for new property snps,clk-csr in binding file

Signed-off-by: Jianguo Zhang <jianguo.zhang@mediatek.com>
Reviewed-by: Krzysztof Kozlowski <krzysztof.kozlowski@linaro.org>
Reviewed-by: AngeloGioacchino Del Regno <angelogioacchino.delregno@collabora.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2 years agodt-bindings: net: mediatek-dwmac: add support for mt8188
Jianguo Zhang [Thu, 29 Sep 2022 01:47:55 +0000 (09:47 +0800)]
dt-bindings: net: mediatek-dwmac: add support for mt8188

Add binding document for the ethernet on mt8188

Signed-off-by: Jianguo Zhang <jianguo.zhang@mediatek.com>
Reviewed-by: AngeloGioacchino Del Regno <angelogioacchino.delregno@collabora.com>
Reviewed-by: Krzysztof Kozlowski <krzysztof.kozlowski@linaro.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
2 years agonet/mlx5: Fix spelling mistake "syndrom" -> "syndrome"
Colin Ian King [Wed, 28 Sep 2022 22:07:55 +0000 (23:07 +0100)]
net/mlx5: Fix spelling mistake "syndrom" -> "syndrome"

There is a spelling mistake in a devlink_health_report message. Fix it.

Signed-off-by: Colin Ian King <colin.i.king@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2 years agonet: bna: Fix spelling mistake "muliple" -> "multiple"
Colin Ian King [Wed, 28 Sep 2022 21:54:39 +0000 (22:54 +0100)]
net: bna: Fix spelling mistake "muliple" -> "multiple"

There is a spelling mistake in a literal string in the array
bnad_net_stats_strings. Fix it.

Signed-off-by: Colin Ian King <colin.i.king@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2 years agoibmveth: Ethtool set queue support
Nick Child [Wed, 28 Sep 2022 21:43:50 +0000 (16:43 -0500)]
ibmveth: Ethtool set queue support

Implement channel management functions to allow dynamic addition and
removal of transmit queues. The `ethtool --show-channels` and
`ethtool --set-channels` commands can be used to get and set the
number of queues, respectively. Allow the ability to add as many
transmit queues as available processors but never allow more than the
hard maximum of 16. The number of receive queues is one and cannot be
modified.

Depending on whether the requested number of queues is larger or
smaller than the current value, either allocate or free long term
buffers. Since long term buffer construction and destruction can
occur in two different areas, from either channel set requests or
device open/close, define functions for performing this work. If
allocation of a new buffer fails, then attempt to revert back to the
previous number of queues.

Signed-off-by: Nick Child <nnac123@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2 years agoibmveth: Implement multi queue on xmit
Nick Child [Wed, 28 Sep 2022 21:43:49 +0000 (16:43 -0500)]
ibmveth: Implement multi queue on xmit

The `ndo_start_xmit` function is protected by a spinlock on the tx queue
being used to transmit the skb. Allow concurrent calls to
`ndo_start_xmit` by using more than one tx queue. This allows for
greater throughput when several jobs are trying to transmit data.

Introduce 16 tx queues (leave single rx queue as is) which each
correspond to one DMA mapped long term buffer.

Signed-off-by: Nick Child <nnac123@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2 years agoibmveth: Copy tx skbs into a premapped buffer
Nick Child [Wed, 28 Sep 2022 21:43:48 +0000 (16:43 -0500)]
ibmveth: Copy tx skbs into a premapped buffer

Rather than DMA mapping and unmapping every outgoing skb, copy the skb
into a buffer that was mapped during the drivers open function. Copying
the skb and its frags have proven to be more time efficient than
mapping and unmapping. As an effect, performance increases by 3-5
Gbits/s.

Allocate and DMA map one continuous 64KB buffer at `ndo_open`. This
buffer is maintained until `ibmveth_close` is called. This buffer is
large enough to hold the largest possible linnear skb. During
`ndo_start_xmit`, copy the skb and all of it's frags into the continuous
buffer. By manually linnearizing all the socket buffers, time is saved
during memcpy as well as more efficient handling in FW.
As a result, we no longer need to worry about the firmware limitation
of handling a max of 6 frags. So, we only need to maintain 1 descriptor
instead of 6 and can hardcode 0 for the other 5 descriptors during
h_send_logical_lan.

Since, DMA allocation/mapping issues can no longer arise in xmit
functions, we can further reduce code size by removing the need for a
bounce buffer on DMA errors.

Signed-off-by: Nick Child <nnac123@linux.ibm.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2 years agobnx2: Fix spelling mistake "bufferred" -> "buffered"
Colin Ian King [Wed, 28 Sep 2022 21:37:53 +0000 (22:37 +0100)]
bnx2: Fix spelling mistake "bufferred" -> "buffered"

There are spelling mistakes in two literal strings. Fix these.

Signed-off-by: Colin Ian King <colin.i.king@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2 years agonet: bridge: assign path_cost for 2.5G and 5G link speed
Steven Hsieh [Wed, 28 Sep 2022 17:57:58 +0000 (10:57 -0700)]
net: bridge: assign path_cost for 2.5G and 5G link speed

As 2.5G, 5G ethernet ports are more common and affordable,
these ports are being used in LAN bridge devices.
STP port_cost() is missing path_cost assignment for these link speeds,
causes highest cost 100 being used.
This result in lower speed port being picked
when there is loop between 5G and 1G ports.

Original path_cost: 10G=2, 1G=4, 100m=19, 10m=100
Adjusted path_cost: 10G=2, 5G=3, 2.5G=4, 1G=5, 100m=19, 10m=100
                    speed greater than 10G = 1

Signed-off-by: Steven Hsieh <steven.hsieh@broadcom.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2 years agonet: lan966x: Fix spelling mistake "tarffic" -> "traffic"
Colin Ian King [Wed, 28 Sep 2022 14:36:18 +0000 (15:36 +0100)]
net: lan966x: Fix spelling mistake "tarffic" -> "traffic"

There is a spelling mistake in a netdev_err message. Fix it.

Signed-off-by: Colin Ian King <colin.i.king@gmail.com>
Reviewed-by: Horatiu Vultur <horatiu.vultur@microchip.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2 years agonet-next: skbuff: refactor pskb_pull
Richard Gobert [Wed, 28 Sep 2022 12:55:31 +0000 (14:55 +0200)]
net-next: skbuff: refactor pskb_pull

pskb_may_pull already contains all of the checks performed by
pskb_pull.
Use pskb_may_pull for validation in pskb_pull, eliminating the
duplication and making __pskb_pull obsolete.
Replace __pskb_pull with pskb_pull where applicable.

Signed-off-by: Richard Gobert <richardbgobert@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2 years agonet: bonding: Convert to use sysfs_emit()/sysfs_emit_at() APIs
Wang Yufen [Wed, 28 Sep 2022 12:30:14 +0000 (20:30 +0800)]
net: bonding: Convert to use sysfs_emit()/sysfs_emit_at() APIs

Follow the advice of the Documentation/filesystems/sysfs.rst and show()
should only use sysfs_emit() or sysfs_emit_at() when formatting the value
to be returned to user space.

Signed-off-by: Wang Yufen <wangyufen@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2 years agonet-sysfs: Convert to use sysfs_emit() APIs
Wang Yufen [Wed, 28 Sep 2022 11:49:44 +0000 (19:49 +0800)]
net-sysfs: Convert to use sysfs_emit() APIs

Follow the advice of the Documentation/filesystems/sysfs.rst and show()
should only use sysfs_emit() or sysfs_emit_at() when formatting the value
to be returned to user space.

Signed-off-by: Wang Yufen <wangyufen@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
2 years agonet: tun: Convert to use sysfs_emit() APIs
Wang Yufen [Wed, 28 Sep 2022 11:49:43 +0000 (19:49 +0800)]
net: tun: Convert to use sysfs_emit() APIs

Follow the advice of the Documentation/filesystems/sysfs.rst and show()
should only use sysfs_emit() or sysfs_emit_at() when formatting the value
to be returned to user space.

Signed-off-by: Wang Yufen <wangyufen@huawei.com>
Signed-off-by: David S. Miller <davem@davemloft.net>