linux.git
12 months agoipv6: remove RTNL protection from inet6_dump_fib()
Eric Dumazet [Fri, 29 Mar 2024 18:30:53 +0000 (18:30 +0000)]
ipv6: remove RTNL protection from inet6_dump_fib()

No longer hold RTNL while calling inet6_dump_fib().

Also change return value for a completed dump,
so that NLMSG_DONE can be appended to current skb,
saving one recvmsg() system call.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: David Ahern <dsahern@kernel.org>
Link: https://lore.kernel.org/r/20240329183053.644630-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agoMerge branch 'genetlink-remove-linux-genetlink-h'
Jakub Kicinski [Tue, 2 Apr 2024 04:44:35 +0000 (21:44 -0700)]
Merge branch 'genetlink-remove-linux-genetlink-h'

Jakub Kicinski says:

====================
genetlink: remove linux/genetlink.h

There are two genetlink headers net/genetlink.h and linux/genetlink.h
This is similar to netlink.h, but for netlink.h both contain good
amount of code. For genetlink.h the linux/ version is leftover
from before uAPI headers were split out, it has 10 lines of code.
Move those 10 lines into other appropriate headers and delete
linux/genetlink.h.

I occasionally open the wrong header in the editor when coding,
I guess I'm not the only one.

v2: https://lore.kernel.org/all/20240325173716.2390605-1-kuba@kernel.org/
v1: https://lore.kernel.org/all/20240309183458.3014713-1-kuba@kernel.org
====================

Link: https://lore.kernel.org/r/20240329175710.291749-1-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agogenetlink: remove linux/genetlink.h
Jakub Kicinski [Fri, 29 Mar 2024 17:57:10 +0000 (10:57 -0700)]
genetlink: remove linux/genetlink.h

genetlink.h is a shell of what used to be a combined uAPI
and kernel header over a decade ago. It has fewer than
10 lines of code. Merge it into net/genetlink.h.
In some ways it'd be better to keep the combined header
under linux/ but it would make looking through git history
harder.

Acked-by: Sven Eckelmann <sven@narfation.org>
Link: https://lore.kernel.org/r/20240329175710.291749-4-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agonet: openvswitch: remove unnecessary linux/genetlink.h include
Jakub Kicinski [Fri, 29 Mar 2024 17:57:09 +0000 (10:57 -0700)]
net: openvswitch: remove unnecessary linux/genetlink.h include

The only legit reason I could think of for net/genetlink.h
and linux/genetlink.h to be separate would be if one was
included by other headers and we wanted to keep it lightweight.
That is not the case, net/openvswitch/meter.h includes
linux/genetlink.h but for no apparent reason (for struct genl_family
perhaps? it's not necessary, types of externs do not need
to be known).

Link: https://lore.kernel.org/r/20240329175710.291749-3-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agonetlink: create a new header for internal genetlink symbols
Jakub Kicinski [Fri, 29 Mar 2024 17:57:08 +0000 (10:57 -0700)]
netlink: create a new header for internal genetlink symbols

There are things in linux/genetlink.h which are only used
under net/netlink/. Move them to a new local header.
A new header with just 2 externs isn't great, but alternative
would be to include af_netlink.h in genetlink.c which feels
even worse.

Link: https://lore.kernel.org/r/20240329175710.291749-2-kuba@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agoMerge branch '1GbE' of git://git.kernel.org/pub/scm/linux/kernel/git/tnguy/next-queue
Jakub Kicinski [Tue, 2 Apr 2024 04:44:12 +0000 (21:44 -0700)]
Merge branch '1GbE' of git://git./linux/kernel/git/tnguy/next-queue

Tony Nguyen says:

====================
Intel Wired LAN Driver Updates 2024-03-29 (net: intel)

This series contains updates to most Intel drivers.

Jesse moves declaration of pci_driver struct to remove need for forward
declarations in igb and converts Intel drivers to user newer power
management ops.

Sasha reworks power management flow on igc to avoid using rtnl_lock()
during those flows.

Maciej reorganizes i40e_nvm file to avoid forward declarations.

* '1GbE' of git://git.kernel.org/pub/scm/linux/kernel/git/tnguy/next-queue:
  i40e: avoid forward declarations in i40e_nvm.c
  igc: Refactor runtime power management flow
  net: intel: implement modern PM ops declarations
  igb: simplify pci ops declaration
====================

Link: https://lore.kernel.org/r/20240329175632.211340-1-anthony.l.nguyen@intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agotcp/dccp: do not care about families in inet_twsk_purge()
Eric Dumazet [Fri, 29 Mar 2024 15:32:03 +0000 (15:32 +0000)]
tcp/dccp: do not care about families in inet_twsk_purge()

We lost ability to unload ipv6 module a long time ago.

Instead of calling expensive inet_twsk_purge() twice,
we can handle all families in one round.

Also remove an extra line added in my prior patch,
per Kuniyuki Iwashima feedback.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/netdev/20240327192934.6843-1-kuniyu@amazon.com/
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Link: https://lore.kernel.org/r/20240329153203.345203-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agoinet: preserve const qualifier in inet_csk()
Eric Dumazet [Fri, 29 Mar 2024 14:49:31 +0000 (14:49 +0000)]
inet: preserve const qualifier in inet_csk()

We can change inet_csk() to propagate its argument const qualifier,
thanks to container_of_const().

We have to fix few places that had mistakes, like tcp_bound_rto().

Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/r/20240329144931.295800-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agoMerge branch 'doc-netlink-add-hyperlinks-to-generated-docs'
Jakub Kicinski [Tue, 2 Apr 2024 04:26:03 +0000 (21:26 -0700)]
Merge branch 'doc-netlink-add-hyperlinks-to-generated-docs'

Donald Hunter says:

====================
doc: netlink: Add hyperlinks to generated docs

Extend ynl-gen-rst to generate hyperlinks to definitions, attribute sets
and sub-messages from all the places that reference them.
====================

Link: https://lore.kernel.org/r/20240329135021.52534-1-donald.hunter@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agodoc: netlink: Update tc spec with missing definitions
Donald Hunter [Fri, 29 Mar 2024 13:50:21 +0000 (13:50 +0000)]
doc: netlink: Update tc spec with missing definitions

The tc spec referenced tc-u32-mark and tc-act-police-attrs but did not
define them. The missing definitions were discovered when building the
docs with generated hyperlinks because the hyperlink target labels were
missing.

Add definitions for tc-u32-mark and tc-act-police-attrs.

Signed-off-by: Donald Hunter <donald.hunter@gmail.com>
Link: https://lore.kernel.org/r/20240329135021.52534-4-donald.hunter@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agodoc: netlink: Add hyperlinks to generated Netlink docs
Donald Hunter [Fri, 29 Mar 2024 13:50:20 +0000 (13:50 +0000)]
doc: netlink: Add hyperlinks to generated Netlink docs

Update ynl-gen-rst to generate hyperlinks to definitions, attribute
sets and sub-messages from all the places that reference them.

Note that there is a single label namespace for all of the kernel docs.
Hyperlinks within a single netlink doc need to be qualified by the
family name to avoid collisions.

The label format is 'family-type-name' which gives, for example,
'rt-link-attribute-set-link-attrs' as the link id.

Signed-off-by: Donald Hunter <donald.hunter@gmail.com>
Link: https://lore.kernel.org/r/20240329135021.52534-3-donald.hunter@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agodoc: netlink: Change generated docs to limit TOC to depth 3
Donald Hunter [Fri, 29 Mar 2024 13:50:19 +0000 (13:50 +0000)]
doc: netlink: Change generated docs to limit TOC to depth 3

The tables of contents in the generated Netlink docs include individual
attribute definitions. This can make the contents exceedingly long and
repeats a lot of what is on the rest of the pages. See for example:

https://docs.kernel.org/networking/netlink_spec/tc.html

Add a depth limit to the contents directive in generated .rst files to
limit the contents depth to 3 levels. This reduces the contents to:

 - Family
   - Summary
   - Operations
     - op-one
     - op-two
     - ...
   - Definitions
     - struct-one
     - struct-two
     - enum-one
     - ...
   - Attribute sets
     - attrs-one
     - attrs-two
     - ...

Signed-off-by: Donald Hunter <donald.hunter@gmail.com>
Reviewed-by: Breno Leitao <leitao@debian.org>
Link: https://lore.kernel.org/r/20240329135021.52534-2-donald.hunter@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agoMerge branch 'net-rps-misc'
David S. Miller [Mon, 1 Apr 2024 10:28:32 +0000 (11:28 +0100)]
Merge branch 'net-rps-misc'

Eric Dumazet says:

====================
net: rps: misc changes

Make RPS/RFS a bit more efficient with better cache locality
and heuristics.

Aso shrink include/linux/netdevice.h a bit.

v2: fixed a build issue in patch 6/8 with CONFIG_RPS=n
    (Jakub and kernel build bots)
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
12 months agonet: rps: move received_rps field to a better location
Eric Dumazet [Fri, 29 Mar 2024 15:42:25 +0000 (15:42 +0000)]
net: rps: move received_rps field to a better location

Commit 14d898f3c1b3 ("dev: Move received_rps counter next
to RPS members in softnet data") was unfortunate:

received_rps is dirtied by a cpu and never read by other
cpus in fast path.

Its presence in the hot RPS cache line (shared by many cpus)
is hurting RPS/RFS performance.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
12 months agonet: rps: add rps_input_queue_head_add() helper
Eric Dumazet [Fri, 29 Mar 2024 15:42:24 +0000 (15:42 +0000)]
net: rps: add rps_input_queue_head_add() helper

process_backlog() can batch increments of sd->input_queue_head,
saving some memory bandwidth.

Also add READ_ONCE()/WRITE_ONCE() annotations around
sd->input_queue_head accesses.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
12 months agonet: rps: change input_queue_tail_incr_save()
Eric Dumazet [Fri, 29 Mar 2024 15:42:23 +0000 (15:42 +0000)]
net: rps: change input_queue_tail_incr_save()

input_queue_tail_incr_save() is incrementing the sd queue_tail
and save it in the flow last_qtail.

Two issues here :

- no lock protects the write on last_qtail, we should use appropriate
  annotations.

- We can perform this write after releasing the per-cpu backlog lock,
  to decrease this lock hold duration (move away the cache line miss)

Also move input_queue_head_incr() and rps helpers to include/net/rps.h,
while adding rps_ prefix to better reflect their role.

v2: Fixed a build issue (Jakub and kernel build bots)

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
12 months agonet: enqueue_to_backlog() cleanup
Eric Dumazet [Fri, 29 Mar 2024 15:42:22 +0000 (15:42 +0000)]
net: enqueue_to_backlog() cleanup

We can remove a goto and a label by reversing a condition.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
12 months agonet: make softnet_data.dropped an atomic_t
Eric Dumazet [Fri, 29 Mar 2024 15:42:21 +0000 (15:42 +0000)]
net: make softnet_data.dropped an atomic_t

If under extreme cpu backlog pressure enqueue_to_backlog() has
to drop a packet, it could do this without dirtying a cache line
and potentially slowing down the target cpu.

Move sd->dropped into a separate cache line, and make it atomic.

In non pressure mode, this field is not touched, no need to consume
valuable space in a hot cache line.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
12 months agonet: enqueue_to_backlog() change vs not running device
Eric Dumazet [Fri, 29 Mar 2024 15:42:20 +0000 (15:42 +0000)]
net: enqueue_to_backlog() change vs not running device

If the device attached to the packet given to enqueue_to_backlog()
is not running, we drop the packet.

But we accidentally increase sd->dropped, giving false signals
to admins: sd->dropped should be reserved to cpu backlog pressure,
not to temporary glitches at device dismantles.

While we are at it, perform the netif_running() test before
we get the rps lock, and use REASON_DEV_READY
drop reason instead of NOT_SPECIFIED.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
12 months agonet: move dev_xmit_recursion() helpers to net/core/dev.h
Eric Dumazet [Fri, 29 Mar 2024 15:42:19 +0000 (15:42 +0000)]
net: move dev_xmit_recursion() helpers to net/core/dev.h

Move dev_xmit_recursion() and friends to net/core/dev.h

They are only used from net/core/dev.c and net/core/filter.c.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
12 months agonet: move kick_defer_list_purge() to net/core/dev.h
Eric Dumazet [Fri, 29 Mar 2024 15:42:18 +0000 (15:42 +0000)]
net: move kick_defer_list_purge() to net/core/dev.h

kick_defer_list_purge() is defined in net/core/dev.c
and used from net/core/skubff.c

Because we need softnet_data, include <linux/netdevice.h>
from net/core/dev.h

Signed-off-by: Eric Dumazet <edumazet@google.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
12 months agoMerge branch 'ice-pfcp-filter'
David S. Miller [Mon, 1 Apr 2024 09:49:29 +0000 (10:49 +0100)]
Merge branch 'ice-pfcp-filter'

Alexander Lobakin says:

====================
ice: add PFCP filter support

Add support for creating PFCP filters in switchdev mode. Add pfcp module
that allows to create a PFCP-type netdev. The netdev then can be passed to
tc when creating a filter to indicate that PFCP filter should be created.

To add a PFCP filter, a special netdev must be created and passed to tc
command:

  ip link add pfcp0 type pfcp
  tc filter add dev eth0 ingress prio 1 flower pfcp_opts \
    1:12ab/ff:fffffffffffffff0 skip_hw action mirred egress redirect \
    dev pfcp0

Changes in iproute2 [1] are required to use pfcp_opts in tc.

ICE COMMS package is required as it contains PFCP profiles.

Part of this patchset modifies IP_TUNNEL_*_OPTs, which were previously
stored in a __be16. All possible values have already been used, making
it impossible to add new ones.

* 1-3: add new bitmap_{read,write}(), which is used later in the IP
       tunnel flags code (from Alexander's ARM64 MTE series[2]);
* 4-14: some bitmap code preparations also used later in IP tunnels;
* 15-17: convert IP tunnel flags from __be16 to a bitmap;
* 18-21: add PFCP module and support for it in ice.

[1] https://lore.kernel.org/netdev/20230614091758.11180-1-marcin.szycik@linux.intel.com
[2] https://lore.kernel.org/linux-kernel/20231218124033.551770-1-glider@google.com
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
12 months agoice: Add support for PFCP hardware offload in switchdev
Marcin Szycik [Wed, 27 Mar 2024 15:23:58 +0000 (16:23 +0100)]
ice: Add support for PFCP hardware offload in switchdev

Add support for creating PFCP filters in switchdev mode. Add support
for parsing PFCP-specific tc options: S flag and SEID.

To create a PFCP filter, a special netdev must be created and passed
to tc command:

  ip link add pfcp0 type pfcp
  tc filter add dev eth0 ingress prio 1 flower pfcp_opts \
    1:123/ff:fffffffffffffff0 skip_hw action mirred egress redirect \
    dev pfcp0

Changes in iproute2 [1] are required to be able to use pfcp_opts in tc.

ICE COMMS package is required to create a filter as it contains PFCP
profiles.

Link: https://lore.kernel.org/netdev/20230614091758.11180-1-marcin.szycik@linux.intel.com
Signed-off-by: Marcin Szycik <marcin.szycik@linux.intel.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
12 months agoice: refactor ICE_TC_FLWR_FIELD_ENC_OPTS
Marcin Szycik [Wed, 27 Mar 2024 15:23:57 +0000 (16:23 +0100)]
ice: refactor ICE_TC_FLWR_FIELD_ENC_OPTS

FLOW_DISSECTOR_KEY_ENC_OPTS can be used for multiple headers, but currently
it is treated as GTP-exclusive in ice. Rename ICE_TC_FLWR_FIELD_ENC_OPTS to
ICE_TC_FLWR_FIELD_GTP_OPTS and check for tunnel type earlier. After this
refactor, it is easier to add new headers using FLOW_DISSECTOR_KEY_ENC_OPTS
- instead of checking tunnel type in ice_tc_count_lkups() and
ice_tc_fill_tunnel_outer(), it needs to be checked only once, in
ice_parse_tunnel_attr().

Signed-off-by: Marcin Szycik <marcin.szycik@linux.intel.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
12 months agopfcp: always set pfcp metadata
Michal Swiatkowski [Wed, 27 Mar 2024 15:23:56 +0000 (16:23 +0100)]
pfcp: always set pfcp metadata

In PFCP receive path set metadata needed by flower code to do correct
classification based on this metadata.

Signed-off-by: Michal Swiatkowski <michal.swiatkowski@linux.intel.com>
Signed-off-by: Marcin Szycik <marcin.szycik@linux.intel.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
12 months agopfcp: add PFCP module
Wojciech Drewek [Wed, 27 Mar 2024 15:23:55 +0000 (16:23 +0100)]
pfcp: add PFCP module

Packet Forwarding Control Protocol (PFCP) is a 3GPP Protocol
used between the control plane and the user plane function.
It is specified in TS 29.244[1].

Note that this module is not designed to support this Protocol
in the kernel space. There is no support for parsing any PFCP messages.
There is no API that could be used by any userspace daemon.
Basically it does not support PFCP. This protocol is sophisticated
and there is no need for implementing it in the kernel. The purpose
of this module is to allow users to setup software and hardware offload
of PFCP packets using tc tool.

When user requests to create a PFCP device, a new socket is created.
The socket is set up with port number 8805 which is specific for
PFCP [29.244 4.2.2]. This allow to receive PFCP request messages,
response messages use other ports.

Note that only one PFCP netdev can be created.

Only IPv4 is supported at this time.

[1] https://portal.3gpp.org/desktopmodules/Specifications/SpecificationDetails.aspx?specificationId=3111

Signed-off-by: Wojciech Drewek <wojciech.drewek@intel.com>
Signed-off-by: Marcin Szycik <marcin.szycik@linux.intel.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
12 months agonet: net_test: add tests for IP tunnel flags conversion helpers
Alexander Lobakin [Wed, 27 Mar 2024 15:23:54 +0000 (16:23 +0100)]
net: net_test: add tests for IP tunnel flags conversion helpers

Now that there are helpers for converting IP tunnel flags between the
old __be16 format and the bitmap format, make sure they work as expected
by adding a couple of tests to the networking testing suite. The helpers
are all inline, so no dependencies on the related CONFIG_* (or a
standalone module) are needed.

Cover three possible cases:

1. No bits past BIT(15) are set, VTI/SIT bits are not set. This
   conversion is almost a direct assignment.
2. No bits past BIT(15) are set, but VTI/SIT bit is set. During the
   conversion, it must be transformed into BIT(16) in the bitmap,
   but still compatible with the __be16 format.
3. The bitmap has bits past BIT(15) set (not the VTI/SIT one). The
   result will be truncated.
   Note that currently __IP_TUNNEL_FLAG_NUM is 17 (incl. special),
   which means that the result of this case is currently
   semi-false-positive. When BIT(17) is finally here, it will be
   adjusted accordingly.

Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
12 months agoip_tunnel: convert __be16 tunnel flags to bitmaps
Alexander Lobakin [Wed, 27 Mar 2024 15:23:53 +0000 (16:23 +0100)]
ip_tunnel: convert __be16 tunnel flags to bitmaps

Historically, tunnel flags like TUNNEL_CSUM or TUNNEL_ERSPAN_OPT
have been defined as __be16. Now all of those 16 bits are occupied
and there's no more free space for new flags.
It can't be simply switched to a bigger container with no
adjustments to the values, since it's an explicit Endian storage,
and on LE systems (__be16)0x0001 equals to
(__be64)0x0001000000000000.
We could probably define new 64-bit flags depending on the
Endianness, i.e. (__be64)0x0001 on BE and (__be64)0x00010000... on
LE, but that would introduce an Endianness dependency and spawn a
ton of Sparse warnings. To mitigate them, all of those places which
were adjusted with this change would be touched anyway, so why not
define stuff properly if there's no choice.

Define IP_TUNNEL_*_BIT counterparts as a bit number instead of the
value already coded and a fistful of <16 <-> bitmap> converters and
helpers. The two flags which have a different bit position are
SIT_ISATAP_BIT and VTI_ISVTI_BIT, as they were defined not as
__cpu_to_be16(), but as (__force __be16), i.e. had different
positions on LE and BE. Now they both have strongly defined places.
Change all __be16 fields which were used to store those flags, to
IP_TUNNEL_DECLARE_FLAGS() -> DECLARE_BITMAP(__IP_TUNNEL_FLAG_NUM) ->
unsigned long[1] for now, and replace all TUNNEL_* occurrences to
their bitmap counterparts. Use the converters in the places which talk
to the userspace, hardware (NFP) or other hosts (GRE header). The rest
must explicitly use the new flags only. This must be done at once,
otherwise there will be too many conversions throughout the code in
the intermediate commits.
Finally, disable the old __be16 flags for use in the kernel code
(except for the two 'irregular' flags mentioned above), to prevent
any accidental (mis)use of them. For the userspace, nothing is
changed, only additions were made.

Most noticeable bloat-o-meter difference (.text):

vmlinux: 307/-1 (306)
gre.ko: 62/0 (62)
ip_gre.ko: 941/-217 (724) [*]
ip_tunnel.ko: 390/-900 (-510) [**]
ip_vti.ko: 138/0 (138)
ip6_gre.ko: 534/-18 (516) [*]
ip6_tunnel.ko: 118/-10 (108)

[*] gre_flags_to_tnl_flags() grew, but still is inlined
[**] ip_tunnel_find() got uninlined, hence such decrease

The average code size increase in non-extreme case is 100-200 bytes
per module, mostly due to sizeof(long) > sizeof(__be16), as
%__IP_TUNNEL_FLAG_NUM is less than %BITS_PER_LONG and the compilers
are able to expand the majority of bitmap_*() calls here into direct
operations on scalars.

Reviewed-by: Simon Horman <horms@kernel.org>
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
12 months agoip_tunnel: use a separate struct to store tunnel params in the kernel
Alexander Lobakin [Wed, 27 Mar 2024 15:23:52 +0000 (16:23 +0100)]
ip_tunnel: use a separate struct to store tunnel params in the kernel

Unlike IPv6 tunnels which use purely-kernel __ip6_tnl_parm structure
to store params inside the kernel, IPv4 tunnel code uses the same
ip_tunnel_parm which is being used to talk with the userspace.
This makes it difficult to alter or add any fields or use a
different format for whatever data.
Define struct ip_tunnel_parm_kern, a 1:1 copy of ip_tunnel_parm for
now, and use it throughout the code. Define the pieces, where the copy
user <-> kernel happens, as standalone functions, and copy the data
there field-by-field, so that the kernel-side structure could be easily
modified later on and the users wouldn't have to care about this.

Reviewed-by: Simon Horman <horms@kernel.org>
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
12 months agolib/bitmap: add compile-time test for __assign_bit() optimization
Alexander Lobakin [Wed, 27 Mar 2024 15:23:51 +0000 (16:23 +0100)]
lib/bitmap: add compile-time test for __assign_bit() optimization

Commit dc34d5036692 ("lib: test_bitmap: add compile-time
optimization/evaluations assertions") initially missed __assign_bit(),
which led to that quite a time passed before I realized it doesn't get
optimized at compilation time. Now that it does, add test for that just
to make sure nothing will break one day.
To make things more interesting, use bitmap_complement() and
bitmap_full(), thus checking their compile-time evaluation as well. And
remove the misleading comment mentioning the workaround removed recently
in favor of adding the whole file to GCov exceptions.

Reviewed-by: Przemek Kitszel <przemyslaw.kitszel@intel.com>
Signed-off-by: Yury Norov <yury.norov@gmail.com>
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
12 months agobitmap: make bitmap_{get,set}_value8() use bitmap_{read,write}()
Alexander Lobakin [Wed, 27 Mar 2024 15:23:50 +0000 (16:23 +0100)]
bitmap: make bitmap_{get,set}_value8() use bitmap_{read,write}()

Now that we have generic bitmap_read() and bitmap_write(), which are
inline and try to take care of non-bound-crossing and aligned cases
to keep them optimized, collapse bitmap_{get,set}_value8() into
simple wrappers around the former ones.
bloat-o-meter shows no difference in vmlinux and -2 bytes for
gpio-pca953x.ko, which says the optimization didn't suffer due to
that change. The converted helpers have the value width embedded
and always compile-time constant and that helps a lot.

Suggested-by: Yury Norov <yury.norov@gmail.com>
Signed-off-by: Yury Norov <yury.norov@gmail.com>
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
12 months agobitmap: introduce generic optimized bitmap_size()
Alexander Lobakin [Wed, 27 Mar 2024 15:23:49 +0000 (16:23 +0100)]
bitmap: introduce generic optimized bitmap_size()

The number of times yet another open coded
`BITS_TO_LONGS(nbits) * sizeof(long)` can be spotted is huge.
Some generic helper is long overdue.

Add one, bitmap_size(), but with one detail.
BITS_TO_LONGS() uses DIV_ROUND_UP(). The latter works well when both
divident and divisor are compile-time constants or when the divisor
is not a pow-of-2. When it is however, the compilers sometimes tend
to generate suboptimal code (GCC 13):

48 83 c0 3f           add    $0x3f,%rax
48 c1 e8 06           shr    $0x6,%rax
48 8d 14 c5 00 00 00 00 lea    0x0(,%rax,8),%rdx

%BITS_PER_LONG is always a pow-2 (either 32 or 64), but GCC still does
full division of `nbits + 63` by it and then multiplication by 8.
Instead of BITS_TO_LONGS(), use ALIGN() and then divide by 8. GCC:

8d 50 3f              lea    0x3f(%rax),%edx
c1 ea 03              shr    $0x3,%edx
81 e2 f8 ff ff 1f     and    $0x1ffffff8,%edx

Now it shifts `nbits + 63` by 3 positions (IOW performs fast division
by 8) and then masks bits[2:0]. bloat-o-meter:

add/remove: 0/0 grow/shrink: 20/133 up/down: 156/-773 (-617)

Clang does it better and generates the same code before/after starting
from -O1, except that with the ALIGN() approach it uses %edx and thus
still saves some bytes:

add/remove: 0/0 grow/shrink: 9/133 up/down: 18/-538 (-520)

Note that we can't expand DIV_ROUND_UP() by adding a check and using
this approach there, as it's used in array declarations where
expressions are not allowed.
Add this helper to tools/ as well.

Reviewed-by: Przemek Kitszel <przemyslaw.kitszel@intel.com>
Acked-by: Yury Norov <yury.norov@gmail.com>
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
12 months agotools: move alignment-related macros to new <linux/align.h>
Alexander Lobakin [Wed, 27 Mar 2024 15:23:48 +0000 (16:23 +0100)]
tools: move alignment-related macros to new <linux/align.h>

Currently, tools have *ALIGN*() macros scattered across the unrelated
headers, as there are only 3 of them and they were added separately
each time on an as-needed basis.
Anyway, let's make it more consistent with the kernel headers and allow
using those macros outside of the mentioned headers. Create
<linux/align.h> inside the tools/ folder and include it where needed.

Signed-off-by: Yury Norov <yury.norov@gmail.com>
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
12 months agobtrfs: rename bitmap_set_bits() -> btrfs_bitmap_set_bits()
Alexander Lobakin [Wed, 27 Mar 2024 15:23:47 +0000 (16:23 +0100)]
btrfs: rename bitmap_set_bits() -> btrfs_bitmap_set_bits()

bitmap_set_bits() does not start with the FS' prefix and may collide
with a new generic helper one day. It operates with the FS-specific
types, so there's no change those two could do the same thing.
Just add the prefix to exclude such possible conflict.

Reviewed-by: Przemek Kitszel <przemyslaw.kitszel@intel.com>
Acked-by: David Sterba <dsterba@suse.com>
Reviewed-by: Yury Norov <yury.norov@gmail.com>
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
12 months agofs/ntfs3: add prefix to bitmap_size() and use BITS_TO_U64()
Alexander Lobakin [Wed, 27 Mar 2024 15:23:46 +0000 (16:23 +0100)]
fs/ntfs3: add prefix to bitmap_size() and use BITS_TO_U64()

bitmap_size() is a pretty generic name and one may want to use it for
a generic bitmap API function. At the same time, its logic is
NTFS-specific, as it aligns to the sizeof(u64), not the sizeof(long)
(although it uses ideologically right ALIGN() instead of division).
Add the prefix 'ntfs3_' used for that FS (not just 'ntfs_' to not mix
it with the legacy module) and use generic BITS_TO_U64() while at it.

Suggested-by: Yury Norov <yury.norov@gmail.com> # BITS_TO_U64()
Reviewed-by: Przemek Kitszel <przemyslaw.kitszel@intel.com>
Reviewed-by: Yury Norov <yury.norov@gmail.com>
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
12 months agos390/cio: rename bitmap_size() -> idset_bitmap_size()
Alexander Lobakin [Wed, 27 Mar 2024 15:23:45 +0000 (16:23 +0100)]
s390/cio: rename bitmap_size() -> idset_bitmap_size()

bitmap_size() is a pretty generic name and one may want to use it for
a generic bitmap API function. At the same time, its logic is not
"generic", i.e. it's not just `nbits -> size of bitmap in bytes`
converter as it would be expected from its name.
Add the prefix 'idset_' used throughout the file where the function
resides.

Reviewed-by: Przemek Kitszel <przemyslaw.kitszel@intel.com>
Acked-by: Peter Oberparleiter <oberpar@linux.ibm.com>
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
12 months agolinkmode: convert linkmode_{test,set,clear,mod}_bit() to macros
Alexander Lobakin [Wed, 27 Mar 2024 15:23:44 +0000 (16:23 +0100)]
linkmode: convert linkmode_{test,set,clear,mod}_bit() to macros

Since commit b03fc1173c0c ("bitops: let optimize out non-atomic bitops
on compile-time constants"), the non-atomic bitops are macros which can
be expanded by the compilers into compile-time expressions, which will
result in better optimized object code. Unfortunately, turned out that
passing `volatile` to those macros discards any possibility of
optimization, as the compilers then don't even try to look whether
the passed bitmap is known at compilation time. In addition to that,
the mentioned linkmode helpers are marked with `inline`, not
`__always_inline`, meaning that it's not guaranteed some compiler won't
uninline them for no reason, which will also effectively prevent them
from being optimized (it's a well-known thing the compilers sometimes
uninline `2 + 2`).
Convert linkmode_*_bit() from inlines to macros. Their calling
convention are 1:1 with the corresponding bitops, so that it's not even
needed to enumerate and map the arguments, only the names. No changes in
vmlinux' object code (compiled by LLVM for x86_64) whatsoever, but that
doesn't necessarily means the change is meaningless.

Reviewed-by: Przemek Kitszel <przemyslaw.kitszel@intel.com>
Acked-by: Jakub Kicinski <kuba@kernel.org>
Acked-by: Yury Norov <yury.norov@gmail.com>
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
12 months agobitops: let the compiler optimize {__,}assign_bit()
Alexander Lobakin [Wed, 27 Mar 2024 15:23:43 +0000 (16:23 +0100)]
bitops: let the compiler optimize {__,}assign_bit()

Since commit b03fc1173c0c ("bitops: let optimize out non-atomic bitops
on compile-time constants"), the compilers are able to expand inline
bitmap operations to compile-time initializers when possible.
However, during the round of replacement if-__set-else-__clear with
__assign_bit() as per Andy's advice, bloat-o-meter showed +1024 bytes
difference in object code size for one module (even one function),
where the pattern:

DECLARE_BITMAP(foo) = { }; // on the stack, zeroed

if (a)
__set_bit(const_bit_num, foo);
if (b)
__set_bit(another_const_bit_num, foo);
...

is heavily used, although there should be no difference: the bitmap is
zeroed, so the second half of __assign_bit() should be compiled-out as
a no-op.
I either missed the fact that __assign_bit() has bitmap pointer marked
as `volatile` (as we usually do for bitops) or was hoping that the
compilers would at least try to look past the `volatile` for
__always_inline functions. Anyhow, due to that attribute, the compilers
were always compiling the whole expression and no mentioned compile-time
optimizations were working.

Convert __assign_bit() to a macro since it's a very simple if-else and
all of the checks are performed inside __set_bit() and __clear_bit(),
thus that wrapper has to be as transparent as possible. After that
change, despite it showing only -20 bytes change for vmlinux (due to
that it's still relatively unpopular), no drastic code size changes
happen when replacing if-set-else-clear for onstack bitmaps with
__assign_bit(), meaning the compiler now expands them to the actual
operations will all the expected optimizations.

Atomic assign_bit() is less affected due to its nature, but let's
convert it to a macro as well to keep the code consistent and not
leave a place for possible suboptimal codegen. Moreover, with certain
kernel configuration it actually gives some saves (x86):

do_ip_setsockopt    4154    4099     -55

Suggested-by: Yury Norov <yury.norov@gmail.com> # assign_bit(), too
Cc: Andy Shevchenko <andriy.shevchenko@linux.intel.com>
Reviewed-by: Przemek Kitszel <przemyslaw.kitszel@intel.com>
Acked-by: Yury Norov <yury.norov@gmail.com>
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
12 months agobitops: make BYTES_TO_BITS() treewide-available
Alexander Lobakin [Wed, 27 Mar 2024 15:23:42 +0000 (16:23 +0100)]
bitops: make BYTES_TO_BITS() treewide-available

Avoid open-coding that simple expression each time by moving
BYTES_TO_BITS() from the probes code to <linux/bitops.h> to export
it to the rest of the kernel.
Simplify the macro while at it. `BITS_PER_LONG / sizeof(long)` always
equals to %BITS_PER_BYTE, regardless of the target architecture.
Do the same for the tools ecosystem as well (incl. its version of
bitops.h). The previous implementation had its implicit type of long,
while the new one is int, so adjust the format literal accordingly in
the perf code.

Suggested-by: Andy Shevchenko <andriy.shevchenko@linux.intel.com>
Reviewed-by: Przemek Kitszel <przemyslaw.kitszel@intel.com>
Acked-by: Yury Norov <yury.norov@gmail.com>
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
12 months agobitops: add missing prototype check
Alexander Lobakin [Wed, 27 Mar 2024 15:23:41 +0000 (16:23 +0100)]
bitops: add missing prototype check

Commit 8238b4579866 ("wait_on_bit: add an acquire memory barrier") added
a new bitop, test_bit_acquire(), with proper wrapping in order to try to
optimize it at compile-time, but missed the list of bitops used for
checking their prototypes a bit below.
The functions added have consistent prototypes, so that no more changes
are required and no functional changes take place.

Fixes: 8238b4579866 ("wait_on_bit: add an acquire memory barrier")
Reviewed-by: Przemek Kitszel <przemyslaw.kitszel@intel.com>
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
12 months agolib/test_bitmap: use pr_info() for non-error messages
Alexander Potapenko [Wed, 27 Mar 2024 15:23:40 +0000 (16:23 +0100)]
lib/test_bitmap: use pr_info() for non-error messages

pr_err() messages may be treated as errors by some log readers, so let
us only use them for test failures. For non-error messages, replace them
with pr_info().

Suggested-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Signed-off-by: Alexander Potapenko <glider@google.com>
Acked-by: Yury Norov <yury.norov@gmail.com>
Signed-off-by: Yury Norov <yury.norov@gmail.com>
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
12 months agolib/test_bitmap: add tests for bitmap_{read,write}()
Alexander Potapenko [Wed, 27 Mar 2024 15:23:39 +0000 (16:23 +0100)]
lib/test_bitmap: add tests for bitmap_{read,write}()

Add basic tests ensuring that values can be added at arbitrary positions
of the bitmap, including those spanning into the adjacent unsigned
longs.

Two new performance tests, test_bitmap_read_perf() and
test_bitmap_write_perf(), can be used to assess future performance
improvements of bitmap_read() and bitmap_write():

[    0.431119][    T1] test_bitmap: Time spent in test_bitmap_read_perf: 615253
[    0.433197][    T1] test_bitmap: Time spent in test_bitmap_write_perf: 916313

(numbers from a Intel(R) Xeon(R) Gold 6154 CPU @ 3.00GHz machine running
QEMU).

Signed-off-by: Alexander Potapenko <glider@google.com>
Reviewed-by: Andy Shevchenko <andriy.shevchenko@linux.intel.com>
Acked-by: Yury Norov <yury.norov@gmail.com>
Signed-off-by: Yury Norov <yury.norov@gmail.com>
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
12 months agolib/bitmap: add bitmap_{read,write}()
Syed Nayyar Waris [Wed, 27 Mar 2024 15:23:38 +0000 (16:23 +0100)]
lib/bitmap: add bitmap_{read,write}()

The two new functions allow reading/writing values of length up to
BITS_PER_LONG bits at arbitrary position in the bitmap.

The code was taken from "bitops: Introduce the for_each_set_clump macro"
by Syed Nayyar Waris with a number of changes and simplifications:
 - instead of using roundup(), which adds an unnecessary dependency
   on <linux/math.h>, we calculate space as BITS_PER_LONG-offset;
 - indentation is reduced by not using else-clauses (suggested by
   checkpatch for bitmap_get_value());
 - bitmap_get_value()/bitmap_set_value() are renamed to bitmap_read()
   and bitmap_write();
 - some redundant computations are omitted.

Cc: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Syed Nayyar Waris <syednwaris@gmail.com>
Signed-off-by: William Breathitt Gray <william.gray@linaro.org>
Link: https://lore.kernel.org/lkml/fe12eedf3666f4af5138de0e70b67a07c7f40338.1592224129.git.syednwaris@gmail.com/
Suggested-by: Yury Norov <yury.norov@gmail.com>
Co-developed-by: Alexander Potapenko <glider@google.com>
Signed-off-by: Alexander Potapenko <glider@google.com>
Reviewed-by: Andy Shevchenko <andriy.shevchenko@linux.intel.com>
Acked-by: Yury Norov <yury.norov@gmail.com>
Signed-off-by: Yury Norov <yury.norov@gmail.com>
Signed-off-by: Alexander Lobakin <aleksander.lobakin@intel.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
12 months agoMerge branch 'add-property-in-dwmac-stm32-documentation'
Jakub Kicinski [Fri, 29 Mar 2024 22:42:14 +0000 (15:42 -0700)]
Merge branch 'add-property-in-dwmac-stm32-documentation'

Christophe Roullier says:

====================
Add property in dwmac-stm32 documentation

Introduce property in dwmac-stm32 documentation

 - st,ext-phyclk: is present since 2020 in driver so need to explain
   it and avoid dtbs check issue : views/kernel/upstream/net-next/arch/arm/boot/dts/st/stm32mp157c-dk2.dtb:
ethernet@5800a000: Unevaluated properties are not allowed
('st,ext-phyclk' was unexpected)
   Furthermore this property will be use in upstream of MP13 dwmac glue. (next step)
====================

Link: https://lore.kernel.org/r/20240328185337.332703-1-christophe.roullier@foss.st.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agodt-bindings: net: dwmac: Document STM32 property st,ext-phyclk
Christophe Roullier [Thu, 28 Mar 2024 18:53:37 +0000 (19:53 +0100)]
dt-bindings: net: dwmac: Document STM32 property st,ext-phyclk

The Linux kernel dwmac-stm32 driver currently supports three DT
properties used to configure whether PHY clock are generated by
the MAC or supplied to the MAC from the PHY.

Originally there were two properties, st,eth-clk-sel and
st,eth-ref-clk-sel, each used to configure MAC clocking in
different bus mode and for different MAC clock frequency.
Since it is possible to determine the MAC 'eth-ck' clock
frequency from the clock subsystem and PHY bus mode from
the 'phy-mode' property, two disparate DT properties are
no longer required to configure MAC clocking.

Linux kernel commit 1bb694e20839 ("net: ethernet: stmmac: simplify phy modes management for stm32")
introduced a third, unified, property st,ext-phyclk. This property
covers both use cases of st,eth-clk-sel and st,eth-ref-clk-sel DT
properties, as well as a new use case for 25 MHz clock generated
by the MAC.

The third property st,ext-phyclk is so far undocumented,
document it.

Below table summarizes the clock requirement and clock sources for
supported PHY interface modes.
 __________________________________________________________________________
|PHY_MODE | Normal | PHY wo crystal|   PHY wo crystal   |No 125Mhz from PHY|
|         |        |      25MHz    |        50MHz       |                  |

---------------------------------------------------------------------------
|  MII    |    -   |     eth-ck    |        n/a         |       n/a        |
|         |        | st,ext-phyclk |                    |                  |

---------------------------------------------------------------------------
|  GMII   |    -   |     eth-ck    |        n/a         |       n/a        |
|         |        | st,ext-phyclk |                    |                  |

---------------------------------------------------------------------------
| RGMII   |    -   |     eth-ck    |        n/a         |      eth-ck      |
|         |        | st,ext-phyclk |                    | st,eth-clk-sel or|
|         |        |               |                    | st,ext-phyclk    |

---------------------------------------------------------------------------
| RMII    |    -   |     eth-ck    |      eth-ck        |       n/a        |
|         |        | st,ext-phyclk | st,eth-ref-clk-sel |                  |
|         |        |               | or st,ext-phyclk   |                  |

---------------------------------------------------------------------------

Reviewed-by: Krzysztof Kozlowski <krzysztof.kozlowski@linaro.org>
Reviewed-by: Marek Vasut <marex@denx.de>
Signed-off-by: Christophe Roullier <christophe.roullier@foss.st.com>
Link: https://lore.kernel.org/r/20240328185337.332703-2-christophe.roullier@foss.st.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agonetlink: introduce type-checking attribute iteration
Johannes Berg [Thu, 28 Mar 2024 19:31:45 +0000 (20:31 +0100)]
netlink: introduce type-checking attribute iteration

There are, especially with multi-attr arrays, many cases
of needing to iterate all attributes of a specific type
in a netlink message or a nested attribute. Add specific
macros to support that case.

Also convert many instances using this spatch:

    @@
    iterator nla_for_each_attr;
    iterator name nla_for_each_attr_type;
    identifier nla;
    expression head, len, rem;
    expression ATTR;
    type T;
    identifier x;
    @@
    -nla_for_each_attr(nla, head, len, rem)
    +nla_for_each_attr_type(nla, ATTR, head, len, rem)
     {
    <... T x; ...>
    -if (nla_type(nla) == ATTR) {
     ...
    -}
     }

    @@
    identifier nla;
    iterator nla_for_each_nested;
    iterator name nla_for_each_nested_type;
    expression attr, rem;
    expression ATTR;
    type T;
    identifier x;
    @@
    -nla_for_each_nested(nla, attr, rem)
    +nla_for_each_nested_type(nla, ATTR, attr, rem)
     {
    <... T x; ...>
    -if (nla_type(nla) == ATTR) {
     ...
    -}
     }

    @@
    iterator nla_for_each_attr;
    iterator name nla_for_each_attr_type;
    identifier nla;
    expression head, len, rem;
    expression ATTR;
    type T;
    identifier x;
    @@
    -nla_for_each_attr(nla, head, len, rem)
    +nla_for_each_attr_type(nla, ATTR, head, len, rem)
     {
    <... T x; ...>
    -if (nla_type(nla) != ATTR) continue;
     ...
     }

    @@
    identifier nla;
    iterator nla_for_each_nested;
    iterator name nla_for_each_nested_type;
    expression attr, rem;
    expression ATTR;
    type T;
    identifier x;
    @@
    -nla_for_each_nested(nla, attr, rem)
    +nla_for_each_nested_type(nla, ATTR, attr, rem)
     {
    <... T x; ...>
    -if (nla_type(nla) != ATTR) continue;
     ...
     }

Although I had to undo one bad change this made, and
I also adjusted some other code for whitespace and to
use direct variable initialization now.

Signed-off-by: Johannes Berg <johannes.berg@intel.com>
Link: https://lore.kernel.org/r/20240328203144.b5a6c895fb80.I1869b44767379f204998ff44dd239803f39c23e0@changeid
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agoMerge branch 'udp-small-changes-on-receive-path'
Jakub Kicinski [Fri, 29 Mar 2024 22:03:14 +0000 (15:03 -0700)]
Merge branch 'udp-small-changes-on-receive-path'

Eric Dumazet says:

====================
udp: small changes on receive path

This series is based on an observation I made in UDP receive path.

The sock_def_readable() costs are pretty high, especially when
epoll is used to generate EPOLLIN events.

First patch annotates races on sk->sk_rcvbuf reads.

Second patch replaces an atomic_add_return()
 with a less expensive atomic_add()

Third patch avoids calling sock_def_readable() when possible.

Fourth patch adds sk_wake_async_rcu() to get better inlining
and code generation.
====================

Link: https://lore.kernel.org/r/20240328144032.1864988-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agonet: add sk_wake_async_rcu() helper
Eric Dumazet [Thu, 28 Mar 2024 14:40:32 +0000 (14:40 +0000)]
net: add sk_wake_async_rcu() helper

While looking at UDP receive performance, I saw sk_wake_async()
was no longer inlined.

This matters at least on AMD Zen1-4 platforms (see SRSO)

This might be because rcu_read_lock() and rcu_read_unlock()
are no longer nops in recent kernels ?

Add sk_wake_async_rcu() variant, which must be called from
contexts already holding rcu lock.

As SOCK_FASYNC is deprecated in modern days, use unlikely()
to give a hint to the compiler.

sk_wake_async_rcu() is properly inlined from
__udp_enqueue_schedule_skb() and sock_def_readable().

Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/r/20240328144032.1864988-5-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agoudp: avoid calling sock_def_readable() if possible
Eric Dumazet [Thu, 28 Mar 2024 14:40:31 +0000 (14:40 +0000)]
udp: avoid calling sock_def_readable() if possible

sock_def_readable() is quite expensive (particularly
when ep_poll_callback() is in the picture).

We must call sk->sk_data_ready() when :

- receive queue was empty, or
- SO_PEEK_OFF is enabled on the socket, or
- sk->sk_data_ready is not sock_def_readable.

We still need to call sk_wake_async().

Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Willem de Bruijn <willemb@google.com>
Acked-by: Paolo Abeni <pabeni@redhat.com>
Link: https://lore.kernel.org/r/20240328144032.1864988-4-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agoudp: relax atomic operation on sk->sk_rmem_alloc
Eric Dumazet [Thu, 28 Mar 2024 14:40:30 +0000 (14:40 +0000)]
udp: relax atomic operation on sk->sk_rmem_alloc

atomic_add_return() is more expensive than atomic_add()
and seems overkill in UDP rx fast path.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/r/20240328144032.1864988-3-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agoudp: annotate data-race in __udp_enqueue_schedule_skb()
Eric Dumazet [Thu, 28 Mar 2024 14:40:29 +0000 (14:40 +0000)]
udp: annotate data-race in __udp_enqueue_schedule_skb()

sk->sk_rcvbuf is read locklessly twice, while other threads
could change its value.

Use a READ_ONCE() to annotate the race.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Link: https://lore.kernel.org/r/20240328144032.1864988-2-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agoMerge branch 'address-remaining-wtautological-constant-out-of-range-compare'
Jakub Kicinski [Fri, 29 Mar 2024 19:47:58 +0000 (12:47 -0700)]
Merge branch 'address-remaining-wtautological-constant-out-of-range-compare'

Arnd Bergmann says:

====================
address remaining -Wtautological-constant-out-of-range-compare

The warning option was introduced a few years ago but left disabled
by default. All of the actual bugs that this has found have been
fixed in the meantime, and this series should address the remaining
false-positives, as tested on arm/arm64/x86 randconfigs as well as
allmodconfig builds for all architectures supported by clang.
====================

Link: https://lore.kernel.org/r/20240328143051.1069575-1-arnd@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agomlx5: stop warning for 64KB pages
Arnd Bergmann [Thu, 28 Mar 2024 14:30:46 +0000 (15:30 +0100)]
mlx5: stop warning for 64KB pages

When building with 64KB pages, clang points out that xsk->chunk_size
can never be PAGE_SIZE:

drivers/net/ethernet/mellanox/mlx5/core/en/xsk/setup.c:19:22: error: result of comparison of constant 65536 with expression of type 'u16' (aka 'unsigned short') is always false [-Werror,-Wtautological-constant-out-of-range-compare]
        if (xsk->chunk_size > PAGE_SIZE ||
            ~~~~~~~~~~~~~~~ ^ ~~~~~~~~~

In older versions of this code, using PAGE_SIZE was the only
possibility, so this would have never worked on 64KB page kernels,
but the patch apparently did not address this case completely.

As Maxim Mikityanskiy suggested, 64KB chunks are really not all that
useful, so just shut up the warning by adding a cast.

Fixes: 282c0c798f8e ("net/mlx5e: Allow XSK frames smaller than a page")
Link: https://lore.kernel.org/netdev/20211013150232.2942146-1-arnd@kernel.org/
Link: https://lore.kernel.org/lkml/a7b27541-0ebb-4f2d-bd06-270a4d404613@app.fastmail.com/
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Acked-by: Maxim Mikityanskiy <maxtram95@gmail.com>
Reviewed-by: Justin Stitt <justinstitt@google.com>
Reviewed-by: Tariq Toukan <tariqt@nvidia.com>
Link: https://lore.kernel.org/r/20240328143051.1069575-9-arnd@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agonet: axienet: Fix kernel doc warnings
Suraj Gupta [Thu, 28 Mar 2024 11:07:13 +0000 (16:37 +0530)]
net: axienet: Fix kernel doc warnings

Add description of mdio enable, mdio disable and mdio wait functions.
Add description of skb pointer in axidma_bd data structure.
Remove 'phy_node' description in axienet local data structure since
it is not a valid struct member.
Correct description of struct axienet_option.

Fix below kernel-doc warnings in drivers/net/ethernet/xilinx/:
1) xilinx_axienet_mdio.c:1: warning: no structured comments found
2) xilinx_axienet.h:379: warning: Function parameter or struct member
'skb' not described in 'axidma_bd'
3) xilinx_axienet.h:538: warning: Excess struct member 'phy_node'
description in 'axienet_local'
4) xilinx_axienet.h:1002: warning: expecting prototype for struct
axiethernet_option. Prototype was for struct axienet_option instead

Signed-off-by: Suraj Gupta <suraj.gupta2@amd.com>
Reviewed-by: Radhey Shyam Pandey <radhey.shyam.pandey@amd.com>
Link: https://lore.kernel.org/r/20240328110713.12885-1-suraj.gupta2@amd.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agoocteontx2-pf: remove unused variables req_hdr and rsp_hdr
Su Hui [Thu, 28 Mar 2024 02:07:24 +0000 (10:07 +0800)]
octeontx2-pf: remove unused variables req_hdr and rsp_hdr

Clang static checker(scan-buid):
drivers/net/ethernet/marvell/octeontx2/nic/otx2_pf.c:503:2: warning:
Value stored to 'rsp_hdr' is never read [deadcode.DeadStores]

Remove these unused variables to save some space.

Signed-off-by: Su Hui <suhui@nfschina.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Link: https://lore.kernel.org/r/20240328020723.4071539-1-suhui@nfschina.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agonfc: st95hf: drop driver owner assignment
Krzysztof Kozlowski [Wed, 27 Mar 2024 17:48:10 +0000 (18:48 +0100)]
nfc: st95hf: drop driver owner assignment

Core in spi_register_driver() already sets the .owner, so driver
does not need to.

Signed-off-by: Krzysztof Kozlowski <krzysztof.kozlowski@linaro.org>
Link: https://lore.kernel.org/r/20240327174810.519676-4-krzysztof.kozlowski@linaro.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agonfc: mrvl: spi: drop driver owner assignment
Krzysztof Kozlowski [Wed, 27 Mar 2024 17:48:09 +0000 (18:48 +0100)]
nfc: mrvl: spi: drop driver owner assignment

Core in spi_register_driver() already sets the .owner, so driver
does not need to.

Signed-off-by: Krzysztof Kozlowski <krzysztof.kozlowski@linaro.org>
Link: https://lore.kernel.org/r/20240327174810.519676-3-krzysztof.kozlowski@linaro.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agonet: wwan: mhi: drop driver owner assignment
Krzysztof Kozlowski [Wed, 27 Mar 2024 17:48:08 +0000 (18:48 +0100)]
net: wwan: mhi: drop driver owner assignment

Core in mhi_driver_register() already sets the .owner, so driver
does not need to.

Signed-off-by: Krzysztof Kozlowski <krzysztof.kozlowski@linaro.org>
Acked-by: Sergey Ryazanov <ryazanov.s.a@gmail.com>
Link: https://lore.kernel.org/r/20240327174810.519676-2-krzysztof.kozlowski@linaro.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agonet: microchip: encx24j600: drop driver owner assignment
Krzysztof Kozlowski [Wed, 27 Mar 2024 17:48:07 +0000 (18:48 +0100)]
net: microchip: encx24j600: drop driver owner assignment

Core in spi_register_driver() already sets the .owner, so driver
does not need to.

Signed-off-by: Krzysztof Kozlowski <krzysztof.kozlowski@linaro.org>
Link: https://lore.kernel.org/r/20240327174810.519676-1-krzysztof.kozlowski@linaro.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agotools/net/ynl: Add extack policy attribute decoding
Donald Hunter [Thu, 28 Mar 2024 15:56:36 +0000 (15:56 +0000)]
tools/net/ynl: Add extack policy attribute decoding

The NLMSGERR_ATTR_POLICY extack attribute has been ignored by ynl up to
now. Extend extack decoding to include _POLICY and the nested
NL_POLICY_TYPE_ATTR_* attributes.

For example:

./tools/net/ynl/cli.py \
  --spec Documentation/netlink/specs/rt_link.yaml \
  --create --do newlink --json '{
    "ifname": "12345678901234567890",
    "linkinfo": {"kind": "bridge"}
    }'
Netlink error: Numerical result out of range
nl_len = 104 (88) nl_flags = 0x300 nl_type = 2
error: -34 extack: {'msg': 'Attribute failed policy validation',
'policy': {'max-length': 15, 'type': 'string'}, 'bad-attr': '.ifname'}

Signed-off-by: Donald Hunter <donald.hunter@gmail.com>
Link: https://lore.kernel.org/r/20240328155636.64688-1-donald.hunter@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agodevlink: use kvzalloc() to allocate devlink instance resources
Jian Wen [Wed, 27 Mar 2024 08:21:28 +0000 (16:21 +0800)]
devlink: use kvzalloc() to allocate devlink instance resources

During live migration of a virtual machine, the SR-IOV VF need to be
re-registered. It may fail when the memory is badly fragmented.

The related log is as follows.

    kernel: hv_netvsc 6045bdaa-c0d1-6045-bdaa-c0d16045bdaa eth0: VF slot 1 added
...
    kernel: kworker/0:0: page allocation failure: order:7, mode:0x40dc0(GFP_KERNEL|__GFP_COMP|__GFP_ZERO), nodemask=(null),cpuset=/,mems_allowed=0
    kernel: CPU: 0 PID: 24006 Comm: kworker/0:0 Tainted: G            E     5.4...x86_64 #1
    kernel: Hardware name: Microsoft Corporation Virtual Machine/Virtual Machine, BIOS 090008  12/07/2018
    kernel: Workqueue: events work_for_cpu_fn
    kernel: Call Trace:
    kernel: dump_stack+0x8b/0xc8
    kernel: warn_alloc+0xff/0x170
    kernel: __alloc_pages_slowpath+0x92c/0xb2b
    kernel: ? get_page_from_freelist+0x1d4/0x1140
    kernel: __alloc_pages_nodemask+0x2f9/0x320
    kernel: alloc_pages_current+0x6a/0xb0
    kernel: kmalloc_order+0x1e/0x70
    kernel: kmalloc_order_trace+0x26/0xb0
    kernel: ? __switch_to_asm+0x34/0x70
    kernel: __kmalloc+0x276/0x280
    kernel: ? _raw_spin_unlock_irqrestore+0x1e/0x40
    kernel: devlink_alloc+0x29/0x110
    kernel: mlx5_devlink_alloc+0x1a/0x20 [mlx5_core]
    kernel: init_one+0x1d/0x650 [mlx5_core]
    kernel: local_pci_probe+0x46/0x90
    kernel: work_for_cpu_fn+0x1a/0x30
    kernel: process_one_work+0x16d/0x390
    kernel: worker_thread+0x1d3/0x3f0
    kernel: kthread+0x105/0x140
    kernel: ? max_active_store+0x80/0x80
    kernel: ? kthread_bind+0x20/0x20
    kernel: ret_from_fork+0x3a/0x50

Signed-off-by: Jian Wen <wenjian1@xiaomi.com>
Link: https://lore.kernel.org/r/20240327082128.942818-1-wenjian1@xiaomi.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agoMerge branch 'enabled-wformat-truncation-for-clang'
Jakub Kicinski [Fri, 29 Mar 2024 19:22:31 +0000 (12:22 -0700)]
Merge branch 'enabled-wformat-truncation-for-clang'

Arnd Bergmann says:

====================
enabled -Wformat-truncation for clang

With randconfig build testing, I found only eight files that produce
warnings with clang when -Wformat-truncation is enabled. This means
we can just turn it on by default rather than only enabling it for
"make W=1".

Unfortunately, gcc produces a lot more warnings when the option
is enabled, so it's not yet possible to turn it on both compilers.
====================

Link: https://lore.kernel.org/r/20240326223825.4084412-1-arnd@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agomlx5: avoid truncating error message
Arnd Bergmann [Tue, 26 Mar 2024 22:38:03 +0000 (23:38 +0100)]
mlx5: avoid truncating error message

clang warns that one error message is too long for its destination buffer:

drivers/net/ethernet/mellanox/mlx5/core/esw/bridge.c:1876:4: error: 'snprintf' will always be truncated; specified size is 80, but format string expands to at least 94 [-Werror,-Wformat-truncation-non-kprintf]

Reword it to be a bit shorter so it always fits.

Fixes: 70f0302b3f20 ("net/mlx5: Bridge, implement mdb offload")
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Reviewed-by: Subbaraya Sundeep <sbhatta@marvell.com>
Link: https://lore.kernel.org/r/20240326223825.4084412-5-arnd@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agoqed: avoid truncating work queue length
Arnd Bergmann [Tue, 26 Mar 2024 22:38:02 +0000 (23:38 +0100)]
qed: avoid truncating work queue length

clang complains that the temporary string for the name passed into
alloc_workqueue() is too short for its contents:

drivers/net/ethernet/qlogic/qed/qed_main.c:1218:3: error: 'snprintf' will always be truncated; specified size is 16, but format string expands to at least 18 [-Werror,-Wformat-truncation]

There is no need for a temporary buffer, and the actual name of a workqueue
is 32 bytes (WQ_NAME_LEN), so just use the interface as intended to avoid
the truncation.

Fixes: 59ccf86fe69a ("qed: Add driver infrastucture for handling mfw requests.")
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Link: https://lore.kernel.org/r/20240326223825.4084412-4-arnd@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agoenetc: avoid truncating error message
Arnd Bergmann [Tue, 26 Mar 2024 22:38:01 +0000 (23:38 +0100)]
enetc: avoid truncating error message

As clang points out, the error message in enetc_setup_xdp_prog()
still does not fit in the buffer and will be truncated:

drivers/net/ethernet/freescale/enetc/enetc.c:2771:3: error: 'snprintf' will always be truncated; specified size is 80, but format string expands to at least 87 [-Werror,-Wformat-truncation]

Replace it with an even shorter message that should fit.

Fixes: f968c56417f0 ("net: enetc: shorten enetc_setup_xdp_prog() error message to fit NETLINK_MAX_FMTMSG_LEN")
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Link: https://lore.kernel.org/r/20240326223825.4084412-3-arnd@kernel.org
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agoocteontx2-af: Increase maximum BPID channels
Radha Mohan Chintakuntla [Tue, 26 Mar 2024 18:45:14 +0000 (11:45 -0700)]
octeontx2-af: Increase maximum BPID channels

Any NIX interface type can have maximum 256 channels. So increased the
backpressure ID count to 256 so that it can cover cn9k and cn10k SoCs that
have different NIX interface types with varied maximum channels.

Signed-off-by: Radha Mohan Chintakuntla <radhac@marvell.com>
Link: https://lore.kernel.org/r/20240326184514.1628284-1-radhac@marvell.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agoMerge branch 'add-ip-port-information-to-udp-drop-tracepoint'
Jakub Kicinski [Fri, 29 Mar 2024 19:18:26 +0000 (12:18 -0700)]
Merge branch 'add-ip-port-information-to-udp-drop-tracepoint'

Balazs Scheidler says:

====================
Add IP/port information to UDP drop tracepoint

In our use-case we would like to recover the properties of dropped UDP
packets. Unfortunately the current udp_fail_queue_rcv_skb tracepoint
only exposes the port number of the receiving socket.

This patch-set will add the source/dest ip/port to the tracepoint, while
keeping the socket's local port as well for compatibility.

Thanks for the review comments by Jason and Kuniyuki, they helped me a lot
and I tried to address all of their comments in this new iteration.
====================

Link: https://lore.kernel.org/r/cover.1711475011.git.balazs.scheidler@axoflow.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agonet: udp: add IP/port data to the tracepoint udp/udp_fail_queue_rcv_skb
Balazs Scheidler [Tue, 26 Mar 2024 18:05:47 +0000 (19:05 +0100)]
net: udp: add IP/port data to the tracepoint udp/udp_fail_queue_rcv_skb

The udp_fail_queue_rcv_skb() tracepoint lacks any details on the source
and destination IP/port whereas this information can be critical in case
of UDP/syslog.

Signed-off-by: Balazs Scheidler <balazs.scheidler@axoflow.com>
Reviewed-by: Jason Xing <kerneljasonxing@gmail.com>
Link: https://lore.kernel.org/r/0c8b3e33dbf679e190be6f4c6736603a76988a20.1711475011.git.balazs.scheidler@axoflow.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agonet: port TP_STORE_ADDR_PORTS_SKB macro to be tcp/udp independent
Balazs Scheidler [Tue, 26 Mar 2024 18:05:46 +0000 (19:05 +0100)]
net: port TP_STORE_ADDR_PORTS_SKB macro to be tcp/udp independent

This patch moves TP_STORE_ADDR_PORTS_SKB() to a common header and removes
the TCP specific implementation details.

Previously the macro assumed the skb passed as an argument is a
TCP packet, the implementation now uses an argument to the L4 header and
uses that to extract the source/destination ports, which happen
to be named the same in "struct tcphdr" and "struct udphdr"

Reviewed-by: Jason Xing <kerneljasonxing@gmail.com>
Signed-off-by: Balazs Scheidler <balazs.scheidler@axoflow.com>
Link: https://lore.kernel.org/r/9e306f78260dfbbdc7353ba5f864cc027a409540.1711475011.git.balazs.scheidler@axoflow.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agonfc: st95hf: Switch to using gpiod API
Andy Shevchenko [Tue, 26 Mar 2024 17:58:36 +0000 (19:58 +0200)]
nfc: st95hf: Switch to using gpiod API

This updates the driver to gpiod API, and removes yet another use of
of_get_named_gpio().

Signed-off-by: Andy Shevchenko <andriy.shevchenko@linux.intel.com>
Reviewed-by: Krzysztof Kozlowski <krzysztof.kozlowski@linaro.org>
Link: https://lore.kernel.org/r/20240326175836.1418718-1-andriy.shevchenko@linux.intel.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agoMerge branch 'add-en8811h-phy-driver-and-devicetree-binding-doc'
Jakub Kicinski [Fri, 29 Mar 2024 19:06:06 +0000 (12:06 -0700)]
Merge branch 'add-en8811h-phy-driver-and-devicetree-binding-doc'

Eric Woudstra says:

====================
Add en8811h phy driver and devicetree binding doc

This patch series adds the driver and the devicetree binding documentation
for the Airoha en8811h PHY.
====================

Link: https://lore.kernel.org/r/20240326162305.303598-1-ericwouds@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agonet: phy: air_en8811h: Add the Airoha EN8811H PHY driver
Eric Woudstra [Tue, 26 Mar 2024 16:23:05 +0000 (17:23 +0100)]
net: phy: air_en8811h: Add the Airoha EN8811H PHY driver

Add the driver for the Airoha EN8811H 2.5 Gigabit PHY. The phy supports
100/1000/2500 Mbps with auto negotiation only.

The driver uses two firmware files, for which updated versions are added to
linux-firmware already.

Note: At phy-address + 8 there is another device on the mdio bus, that
belongs to the EN881H. While the original driver writes to it, Airoha
has confirmed this is not needed. Therefore, communication with this
device is not included in this driver.

Signed-off-by: Eric Woudstra <ericwouds@gmail.com>
Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Link: https://lore.kernel.org/r/20240326162305.303598-3-ericwouds@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agodt-bindings: net: airoha,en8811h: Add en8811h
Eric Woudstra [Tue, 26 Mar 2024 16:23:04 +0000 (17:23 +0100)]
dt-bindings: net: airoha,en8811h: Add en8811h

Add the Airoha EN8811H 2.5 Gigabit PHY.

The en8811h phy can be set with serdes polarity reversed on rx and/or tx.

Signed-off-by: Eric Woudstra <ericwouds@gmail.com>
Reviewed-by: Krzysztof Kozlowski <krzysztof.kozlowski@linaro.org>
Reviewed-by: Andrew Lunn <andrew@lunn.ch>
Link: https://lore.kernel.org/r/20240326162305.303598-2-ericwouds@gmail.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agoi40e: avoid forward declarations in i40e_nvm.c
Maciej Fijalkowski [Wed, 6 Mar 2024 16:30:54 +0000 (17:30 +0100)]
i40e: avoid forward declarations in i40e_nvm.c

Move code around to get rid of forward declarations. No functional
changes.

After a plain code juggling, checkpatch reported:
total: 0 errors, 7 warnings, 12 checks, 1581 lines checked

so while at it let's address old issues as well. Should we ever address
the remaining unnecessary forward declarations within
drivers/net/ethernet/intel/, consider this change as a starting
point/reference.

As reported in [0], there would be a lot more of work to do...if we
care.

[0]: https://lore.kernel.org/intel-wired-lan/Zeh8qadiTGf413YU@boxer/T/#u
Signed-off-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Tested-by: Pucha Himasekhar Reddy <himasekharx.reddy.pucha@intel.com> (A Contingent worker at Intel)
Reviewed-by: Simon Horman <horms@kernel.org>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
12 months agoigc: Refactor runtime power management flow
Sasha Neftin [Sun, 11 Feb 2024 07:30:58 +0000 (09:30 +0200)]
igc: Refactor runtime power management flow

Following the corresponding discussion [1] and [2] refactor the 'igc_open'
method and avoid taking the rtnl_lock() during the 'igc_resume' method.
The rtnl_lock is held by the upper layer and could lead to the deadlock
during resuming from a runtime power management flow. Notify the stack of
the actual queue counts 'netif_set_real_num_*_queues' outside the
'_igc_open' wrapper. This notification doesn't have to be called on each
resume.

Test:
1. Disconnect the ethernet cable
2. Enable the runtime power management via file system:
echo auto > /sys/devices/pci0000\.../power/control
3. Check the device state (lspci -s <device> -vvv | grep -i Status)

Link: https://lore.kernel.org/netdev/20231206113934.8d7819857574.I2deb5804
ef1739a2af307283d320ef7d82456494@changeid/#r [1]
Link: https://lore.kernel.org/netdev/20211125074949.5f897431@kicinski-fedo
ra-pc1c0hjn.dhcp.thefacebook.com/t/ [2]
Signed-off-by: Sasha Neftin <sasha.neftin@intel.com>
Tested-by: Naama Meir <naamax.meir@linux.intel.com>
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
12 months agonet: intel: implement modern PM ops declarations
Jesse Brandeburg [Wed, 6 Mar 2024 02:50:22 +0000 (18:50 -0800)]
net: intel: implement modern PM ops declarations

Switch the Intel networking drivers to use the new power management ops
declaration formats and macros, which allows us to drop __maybe_unused,
as well as a bunch of ifdef checking CONFIG_PM.

This is safe to do because the compiler drops the unused functions,
verified by checking for any of the power management function symbols
being present in System.map for a build without CONFIG_PM.

If a driver has runtime PM, define the ops with pm_ptr(), and if the
driver has Simple PM, use pm_sleep_ptr(), as well as the new versions of
the macros for declaring the members of the pm_ops structs.

Checked with network-enabled allnoconfig, allyesconfig, allmodconfig on
x64_64.

Reviewed-by: Alan Brady <alan.brady@intel.com>
Signed-off-by: Jesse Brandeburg <jesse.brandeburg@intel.com>
Reviewed-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Tested-by: Pucha Himasekhar Reddy <himasekharx.reddy.pucha@intel.com> (A Contingent worker at Intel)
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
12 months agoigb: simplify pci ops declaration
Jesse Brandeburg [Wed, 6 Mar 2024 02:50:21 +0000 (18:50 -0800)]
igb: simplify pci ops declaration

The igb driver was pre-declaring tons of functions just so that it could
have an early declaration of the pci_driver struct.

Delete a bunch of the declarations and move the struct to the bottom of the
file, after all the functions are declared.

Reviewed-by: Alan Brady <alan.brady@intel.com>
Signed-off-by: Jesse Brandeburg <jesse.brandeburg@intel.com>
Reviewed-by: Maciej Fijalkowski <maciej.fijalkowski@intel.com>
Tested-by: Pucha Himasekhar Reddy <himasekharx.reddy.pucha@intel.com> (A Contingent worker at Intel)
Signed-off-by: Tony Nguyen <anthony.l.nguyen@intel.com>
12 months agoMerge branch 'af_unix-rework-gc'
Jakub Kicinski [Fri, 29 Mar 2024 15:28:50 +0000 (08:28 -0700)]
Merge branch 'af_unix-rework-gc'

Kuniyuki Iwashima says:

====================
af_unix: Rework GC.

When we pass a file descriptor to an AF_UNIX socket via SCM_RIGTHS,
the underlying struct file of the inflight fd gets its refcount bumped.
If the fd is of an AF_UNIX socket, we need to track it in case it forms
cyclic references.

Let's say we send a fd of AF_UNIX socket A to B and vice versa and
close() both sockets.

When created, each socket's struct file initially has one reference.
After the fd exchange, both refcounts are bumped up to 2.  Then, close()
decreases both to 1.  From this point on, no one can touch the file/socket.

However, the struct file has one refcount and thus never calls the
release() function of the AF_UNIX socket.

That's why we need to track all inflight AF_UNIX sockets and run garbage
collection.

This series replaces the current GC implementation that locks each inflight
socket's receive queue and requires trickiness in other places.

The new GC does not lock each socket's queue to minimise its effect and
tries to be lightweight if there is no cyclic reference or no update in
the shape of the inflight fd graph.

The new implementation is based on Tarjan's Strongly Connected Components
algorithm, and we will consider each inflight AF_UNIX socket as a vertex
and its file descriptor as an edge in a directed graph.

For the details, please see each patch.

  patch 1  -  3 : Add struct to express inflight socket graphs
  patch       4 : Optimse inflight fd counting
  patch 5  -  6 : Group SCC possibly forming a cycle
  patch 7  -  8 : Support embryo socket
  patch 9  - 11 : Make GC lightweight
  patch 12 - 13 : Detect dead cycle references
  patch      14 : Replace GC algorithm
  patch      15 : selftest

After this series is applied, we can remove the two ugly tricks for race,
scm_fp_dup() in unix_attach_fds() and spin_lock dance in unix_peek_fds()
as done in patch 14/15 of v1.

Also, we will add cond_resched_lock() in __unix_gc() and convert it to
use a dedicated kthread instead of global system workqueue as suggested
by Paolo in a v4 thread.

v4: https://lore.kernel.org/netdev/20240301022243.73908-1-kuniyu@amazon.com/
v3: https://lore.kernel.org/netdev/20240223214003.17369-1-kuniyu@amazon.com/
v2: https://lore.kernel.org/netdev/20240216210556.65913-1-kuniyu@amazon.com/
v1: https://lore.kernel.org/netdev/20240203030058.60750-1-kuniyu@amazon.com/
====================

Link: https://lore.kernel.org/r/20240325202425.60930-1-kuniyu@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agoselftest: af_unix: Test GC for SCM_RIGHTS.
Kuniyuki Iwashima [Mon, 25 Mar 2024 20:24:25 +0000 (13:24 -0700)]
selftest: af_unix: Test GC for SCM_RIGHTS.

This patch adds test cases to verify the new GC.

We run each test for the following cases:

  * SOCK_DGRAM
  * SOCK_STREAM without embryo socket
  * SOCK_STREAM without embryo socket + MSG_OOB
  * SOCK_STREAM with embryo sockets
  * SOCK_STREAM with embryo sockets + MSG_OOB

Before and after running each test case, we ensure that there is
no AF_UNIX socket left in the netns by reading /proc/net/protocols.

We cannot use /proc/net/unix and UNIX_DIAG because the embryo socket
does not show up there.

Each test creates multiple sockets in an array.  We pass sockets in
the even index using the peer sockets in the odd index.

So, send_fd(0, 1) actually sends fd[0] to fd[2] via fd[0 + 1].

  Test 1 : A <-> A
  Test 2 : A <-> B
  Test 3 : A -> B -> C <- D
           ^.___|___.'    ^
                `---------'

Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Acked-by: Paolo Abeni <pabeni@redhat.com>
Link: https://lore.kernel.org/r/20240325202425.60930-16-kuniyu@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agoaf_unix: Replace garbage collection algorithm.
Kuniyuki Iwashima [Mon, 25 Mar 2024 20:24:24 +0000 (13:24 -0700)]
af_unix: Replace garbage collection algorithm.

If we find a dead SCC during iteration, we call unix_collect_skb()
to splice all skb in the SCC to the global sk_buff_head, hitlist.

After iterating all SCC, we unlock unix_gc_lock and purge the queue.

Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Acked-by: Paolo Abeni <pabeni@redhat.com>
Link: https://lore.kernel.org/r/20240325202425.60930-15-kuniyu@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agoaf_unix: Detect dead SCC.
Kuniyuki Iwashima [Mon, 25 Mar 2024 20:24:23 +0000 (13:24 -0700)]
af_unix: Detect dead SCC.

When iterating SCC, we call unix_vertex_dead() for each vertex
to check if the vertex is close()d and has no bridge to another
SCC.

If both conditions are true for every vertex in SCC, we can
execute garbage collection for all skb in the SCC.

The actual garbage collection is done in the following patch,
replacing the old implementation.

Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Acked-by: Paolo Abeni <pabeni@redhat.com>
Link: https://lore.kernel.org/r/20240325202425.60930-14-kuniyu@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agoaf_unix: Assign a unique index to SCC.
Kuniyuki Iwashima [Mon, 25 Mar 2024 20:24:22 +0000 (13:24 -0700)]
af_unix: Assign a unique index to SCC.

The definition of the lowlink in Tarjan's algorithm is the
smallest index of a vertex that is reachable with at most one
back-edge in SCC.  This is not useful for a cross-edge.

If we start traversing from A in the following graph, the final
lowlink of D is 3.  The cross-edge here is one between D and C.

  A -> B -> D   D = (4, 3)  (index, lowlink)
  ^    |    |   C = (3, 1)
  |    V    |   B = (2, 1)
  `--- C <--'   A = (1, 1)

This is because the lowlink of D is updated with the index of C.

In the following patch, we detect a dead SCC by checking two
conditions for each vertex.

  1) vertex has no edge directed to another SCC (no bridge)
  2) vertex's out_degree is the same as the refcount of its file

If 1) is false, there is a receiver of all fds of the SCC and
its ancestor SCC.

To evaluate 1), we need to assign a unique index to each SCC and
assign it to all vertices in the SCC.

This patch changes the lowlink update logic for cross-edge so
that in the example above, the lowlink of D is updated with the
lowlink of C.

  A -> B -> D   D = (4, 1)  (index, lowlink)
  ^    |    |   C = (3, 1)
  |    V    |   B = (2, 1)
  `--- C <--'   A = (1, 1)

Then, all vertices in the same SCC have the same lowlink, and we
can quickly find the bridge connecting to different SCC if exists.

However, it is no longer called lowlink, so we rename it to
scc_index.  (It's sometimes called lowpoint.)

Also, we add a global variable to hold the last index used in DFS
so that we do not reset the initial index in each DFS.

This patch can be squashed to the SCC detection patch but is
split deliberately for anyone wondering why lowlink is not used
as used in the original Tarjan's algorithm and many reference
implementations.

Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Acked-by: Paolo Abeni <pabeni@redhat.com>
Link: https://lore.kernel.org/r/20240325202425.60930-13-kuniyu@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agoaf_unix: Avoid Tarjan's algorithm if unnecessary.
Kuniyuki Iwashima [Mon, 25 Mar 2024 20:24:21 +0000 (13:24 -0700)]
af_unix: Avoid Tarjan's algorithm if unnecessary.

Once a cyclic reference is formed, we need to run GC to check if
there is dead SCC.

However, we do not need to run Tarjan's algorithm if we know that
the shape of the inflight graph has not been changed.

If an edge is added/updated/deleted and the edge's successor is
inflight, we set false to unix_graph_grouped, which means we need
to re-classify SCC.

Once we finalise SCC, we set true to unix_graph_grouped.

While unix_graph_grouped is true, we can iterate the grouped
SCC using vertex->scc_entry in unix_walk_scc_fast().

list_add() and list_for_each_entry_reverse() uses seem weird, but
they are to keep the vertex order consistent and make writing test
easier.

Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Acked-by: Paolo Abeni <pabeni@redhat.com>
Link: https://lore.kernel.org/r/20240325202425.60930-12-kuniyu@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agoaf_unix: Skip GC if no cycle exists.
Kuniyuki Iwashima [Mon, 25 Mar 2024 20:24:20 +0000 (13:24 -0700)]
af_unix: Skip GC if no cycle exists.

We do not need to run GC if there is no possible cyclic reference.
We use unix_graph_maybe_cyclic to decide if we should run GC.

If a fd of an AF_UNIX socket is passed to an already inflight AF_UNIX
socket, they could form a cyclic reference.  Then, we set true to
unix_graph_maybe_cyclic and later run Tarjan's algorithm to group
them into SCC.

Once we run Tarjan's algorithm, we are 100% sure whether cyclic
references exist or not.  If there is no cycle, we set false to
unix_graph_maybe_cyclic and can skip the entire garbage collection
next time.

When finalising SCC, we set true to unix_graph_maybe_cyclic if SCC
consists of multiple vertices.

Even if SCC is a single vertex, a cycle might exist as self-fd passing.
Given the corner case is rare, we detect it by checking all edges of
the vertex and set true to unix_graph_maybe_cyclic.

With this change, __unix_gc() is just a spin_lock() dance in the normal
usage.

Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Acked-by: Paolo Abeni <pabeni@redhat.com>
Link: https://lore.kernel.org/r/20240325202425.60930-11-kuniyu@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agoaf_unix: Save O(n) setup of Tarjan's algo.
Kuniyuki Iwashima [Mon, 25 Mar 2024 20:24:19 +0000 (13:24 -0700)]
af_unix: Save O(n) setup of Tarjan's algo.

Before starting Tarjan's algorithm, we need to mark all vertices
as unvisited.  We can save this O(n) setup by reserving two special
indices (0, 1) and using two variables.

The first time we link a vertex to unix_unvisited_vertices, we set
unix_vertex_unvisited_index to index.

During DFS, we can see that the index of unvisited vertices is the
same as unix_vertex_unvisited_index.

When we finalise SCC later, we set unix_vertex_grouped_index to each
vertex's index.

Then, we can know (i) that the vertex is on the stack if the index
of a visited vertex is >= 2 and (ii) that it is not on the stack and
belongs to a different SCC if the index is unix_vertex_grouped_index.

After the whole algorithm, all indices of vertices are set as
unix_vertex_grouped_index.

Next time we start DFS, we know that all unvisited vertices have
unix_vertex_grouped_index, and we can use unix_vertex_unvisited_index
as the not-on-stack marker.

To use the same variable in __unix_walk_scc(), we can swap
unix_vertex_(grouped|unvisited)_index at the end of Tarjan's
algorithm.

Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Acked-by: Paolo Abeni <pabeni@redhat.com>
Link: https://lore.kernel.org/r/20240325202425.60930-10-kuniyu@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agoaf_unix: Fix up unix_edge.successor for embryo socket.
Kuniyuki Iwashima [Mon, 25 Mar 2024 20:24:18 +0000 (13:24 -0700)]
af_unix: Fix up unix_edge.successor for embryo socket.

To garbage collect inflight AF_UNIX sockets, we must define the
cyclic reference appropriately.  This is a bit tricky if the loop
consists of embryo sockets.

Suppose that the fd of AF_UNIX socket A is passed to D and the fd B
to C and that C and D are embryo sockets of A and B, respectively.
It may appear that there are two separate graphs, A (-> D) and
B (-> C), but this is not correct.

     A --. .-- B
          X
     C <-' `-> D

Now, D holds A's refcount, and C has B's refcount, so unix_release()
will never be called for A and B when we close() them.  However, no
one can call close() for D and C to free skbs holding refcounts of A
and B because C/D is in A/B's receive queue, which should have been
purged by unix_release() for A and B.

So, here's another type of cyclic reference.  When a fd of an AF_UNIX
socket is passed to an embryo socket, the reference is indirectly held
by its parent listening socket.

  .-> A                            .-> B
  |   `- sk_receive_queue          |   `- sk_receive_queue
  |      `- skb                    |      `- skb
  |         `- sk == C             |         `- sk == D
  |            `- sk_receive_queue |           `- sk_receive_queue
  |               `- skb +---------'               `- skb +-.
  |                                                         |
  `---------------------------------------------------------'

Technically, the graph must be denoted as A <-> B instead of A (-> D)
and B (-> C) to find such a cyclic reference without touching each
socket's receive queue.

  .-> A --. .-- B <-.
  |        X        |  ==  A <-> B
  `-- C <-' `-> D --'

We apply this fixup during GC by fetching the real successor by
unix_edge_successor().

When we call accept(), we clear unix_sock.listener under unix_gc_lock
not to confuse GC.

Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Acked-by: Paolo Abeni <pabeni@redhat.com>
Link: https://lore.kernel.org/r/20240325202425.60930-9-kuniyu@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agoaf_unix: Save listener for embryo socket.
Kuniyuki Iwashima [Mon, 25 Mar 2024 20:24:17 +0000 (13:24 -0700)]
af_unix: Save listener for embryo socket.

This is a prep patch for the following change, where we need to
fetch the listening socket from the successor embryo socket
during GC.

We add a new field to struct unix_sock to save a pointer to a
listening socket.

We set it when connect() creates a new socket, and clear it when
accept() is called.

Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Acked-by: Paolo Abeni <pabeni@redhat.com>
Link: https://lore.kernel.org/r/20240325202425.60930-8-kuniyu@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agoaf_unix: Detect Strongly Connected Components.
Kuniyuki Iwashima [Mon, 25 Mar 2024 20:24:16 +0000 (13:24 -0700)]
af_unix: Detect Strongly Connected Components.

In the new GC, we use a simple graph algorithm, Tarjan's Strongly
Connected Components (SCC) algorithm, to find cyclic references.

The algorithm visits every vertex exactly once using depth-first
search (DFS).

DFS starts by pushing an input vertex to a stack and assigning it
a unique number.  Two fields, index and lowlink, are initialised
with the number, but lowlink could be updated later during DFS.

If a vertex has an edge to an unvisited inflight vertex, we visit
it and do the same processing.  So, we will have vertices in the
stack in the order they appear and number them consecutively in
the same order.

If a vertex has a back-edge to a visited vertex in the stack,
we update the predecessor's lowlink with the successor's index.

After iterating edges from the vertex, we check if its index
equals its lowlink.

If the lowlink is different from the index, it shows there was a
back-edge.  Then, we go backtracking and propagate the lowlink to
its predecessor and resume the previous edge iteration from the
next edge.

If the lowlink is the same as the index, we pop vertices before
and including the vertex from the stack.  Then, the set of vertices
is SCC, possibly forming a cycle.  At the same time, we move the
vertices to unix_visited_vertices.

When we finish the algorithm, all vertices in each SCC will be
linked via unix_vertex.scc_entry.

Let's take an example.  We have a graph including five inflight
vertices (F is not inflight):

  A -> B -> C -> D -> E (-> F)
       ^         |
       `---------'

Suppose that we start DFS from C.  We will visit C, D, and B first
and initialise their index and lowlink.  Then, the stack looks like
this:

  > B = (3, 3)  (index, lowlink)
    D = (2, 2)
    C = (1, 1)

When checking B's edge to C, we update B's lowlink with C's index
and propagate it to D.

    B = (3, 1)  (index, lowlink)
  > D = (2, 1)
    C = (1, 1)

Next, we visit E, which has no edge to an inflight vertex.

  > E = (4, 4)  (index, lowlink)
    B = (3, 1)
    D = (2, 1)
    C = (1, 1)

When we leave from E, its index and lowlink are the same, so we
pop E from the stack as single-vertex SCC.  Next, we leave from
B and D but do nothing because their lowlink are different from
their index.

    B = (3, 1)  (index, lowlink)
    D = (2, 1)
  > C = (1, 1)

Then, we leave from C, whose index and lowlink are the same, so
we pop B, D and C as SCC.

Last, we do DFS for the rest of vertices, A, which is also a
single-vertex SCC.

Finally, each unix_vertex.scc_entry is linked as follows:

  A -.  B -> C -> D  E -.
  ^  |  ^         |  ^  |
  `--'  `---------'  `--'

We use SCC later to decide whether we can garbage-collect the
sockets.

Note that we still cannot detect SCC properly if an edge points
to an embryo socket.  The following two patches will sort it out.

Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Acked-by: Paolo Abeni <pabeni@redhat.com>
Link: https://lore.kernel.org/r/20240325202425.60930-7-kuniyu@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agoaf_unix: Iterate all vertices by DFS.
Kuniyuki Iwashima [Mon, 25 Mar 2024 20:24:15 +0000 (13:24 -0700)]
af_unix: Iterate all vertices by DFS.

The new GC will use a depth first search graph algorithm to find
cyclic references.  The algorithm visits every vertex exactly once.

Here, we implement the DFS part without recursion so that no one
can abuse it.

unix_walk_scc() marks every vertex unvisited by initialising index
as UNIX_VERTEX_INDEX_UNVISITED and iterates inflight vertices in
unix_unvisited_vertices and call __unix_walk_scc() to start DFS from
an arbitrary vertex.

__unix_walk_scc() iterates all edges starting from the vertex and
explores the neighbour vertices with DFS using edge_stack.

After visiting all neighbours, __unix_walk_scc() moves the visited
vertex to unix_visited_vertices so that unix_walk_scc() will not
restart DFS from the visited vertex.

Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Acked-by: Paolo Abeni <pabeni@redhat.com>
Link: https://lore.kernel.org/r/20240325202425.60930-6-kuniyu@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agoaf_unix: Bulk update unix_tot_inflight/unix_inflight when queuing skb.
Kuniyuki Iwashima [Mon, 25 Mar 2024 20:24:14 +0000 (13:24 -0700)]
af_unix: Bulk update unix_tot_inflight/unix_inflight when queuing skb.

Currently, we track the number of inflight sockets in two variables.
unix_tot_inflight is the total number of inflight AF_UNIX sockets on
the host, and user->unix_inflight is the number of inflight fds per
user.

We update them one by one in unix_inflight(), which can be done once
in batch.  Also, sendmsg() could fail even after unix_inflight(), then
we need to acquire unix_gc_lock only to decrement the counters.

Let's bulk update the counters in unix_add_edges() and unix_del_edges(),
which is called only for successfully passed fds.

Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Acked-by: Paolo Abeni <pabeni@redhat.com>
Link: https://lore.kernel.org/r/20240325202425.60930-5-kuniyu@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agoaf_unix: Link struct unix_edge when queuing skb.
Kuniyuki Iwashima [Mon, 25 Mar 2024 20:24:13 +0000 (13:24 -0700)]
af_unix: Link struct unix_edge when queuing skb.

Just before queuing skb with inflight fds, we call scm_stat_add(),
which is a good place to set up the preallocated struct unix_vertex
and struct unix_edge in UNIXCB(skb).fp.

Then, we call unix_add_edges() and construct the directed graph
as follows:

  1. Set the inflight socket's unix_sock to unix_edge.predecessor.
  2. Set the receiver's unix_sock to unix_edge.successor.
  3. Set the preallocated vertex to inflight socket's unix_sock.vertex.
  4. Link inflight socket's unix_vertex.entry to unix_unvisited_vertices.
  5. Link unix_edge.vertex_entry to the inflight socket's unix_vertex.edges.

Let's say we pass the fd of AF_UNIX socket A to B and the fd of B
to C.  The graph looks like this:

  +-------------------------+
  | unix_unvisited_vertices | <-------------------------.
  +-------------------------+                           |
  +                                                     |
  |     +--------------+             +--------------+   |         +--------------+
  |     |  unix_sock A | <---. .---> |  unix_sock B | <-|-. .---> |  unix_sock C |
  |     +--------------+     | |     +--------------+   | | |     +--------------+
  | .-+ |    vertex    |     | | .-+ |    vertex    |   | | |     |    vertex    |
  | |   +--------------+     | | |   +--------------+   | | |     +--------------+
  | |                        | | |                      | | |
  | |   +--------------+     | | |   +--------------+   | | |
  | '-> |  unix_vertex |     | | '-> |  unix_vertex |   | | |
  |     +--------------+     | |     +--------------+   | | |
  `---> |    entry     | +---------> |    entry     | +-' | |
        |--------------|     | |     |--------------|     | |
        |    edges     | <-. | |     |    edges     | <-. | |
        +--------------+   | | |     +--------------+   | | |
                           | | |                        | | |
    .----------------------' | | .----------------------' | |
    |                        | | |                        | |
    |   +--------------+     | | |   +--------------+     | |
    |   |   unix_edge  |     | | |   |   unix_edge  |     | |
    |   +--------------+     | | |   +--------------+     | |
    `-> | vertex_entry |     | | `-> | vertex_entry |     | |
        |--------------|     | |     |--------------|     | |
        |  predecessor | +---' |     |  predecessor | +---' |
        |--------------|       |     |--------------|       |
        |   successor  | +-----'     |   successor  | +-----'
        +--------------+             +--------------+

Henceforth, we denote such a graph as A -> B (-> C).

Now, we can express all inflight fd graphs that do not contain
embryo sockets.  We will support the particular case later.

Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Acked-by: Paolo Abeni <pabeni@redhat.com>
Link: https://lore.kernel.org/r/20240325202425.60930-4-kuniyu@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agoaf_unix: Allocate struct unix_edge for each inflight AF_UNIX fd.
Kuniyuki Iwashima [Mon, 25 Mar 2024 20:24:12 +0000 (13:24 -0700)]
af_unix: Allocate struct unix_edge for each inflight AF_UNIX fd.

As with the previous patch, we preallocate to skb's scm_fp_list an
array of struct unix_edge in the number of inflight AF_UNIX fds.

There we just preallocate memory and do not use immediately because
sendmsg() could fail after this point.  The actual use will be in
the next patch.

When we queue skb with inflight edges, we will set the inflight
socket's unix_sock as unix_edge->predecessor and the receiver's
unix_sock as successor, and then we will link the edge to the
inflight socket's unix_vertex.edges.

Note that we set NULL to cloned scm_fp_list.edges in scm_fp_dup()
so that MSG_PEEK does not change the shape of the directed graph.

Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Acked-by: Paolo Abeni <pabeni@redhat.com>
Link: https://lore.kernel.org/r/20240325202425.60930-3-kuniyu@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agoaf_unix: Allocate struct unix_vertex for each inflight AF_UNIX fd.
Kuniyuki Iwashima [Mon, 25 Mar 2024 20:24:11 +0000 (13:24 -0700)]
af_unix: Allocate struct unix_vertex for each inflight AF_UNIX fd.

We will replace the garbage collection algorithm for AF_UNIX, where
we will consider each inflight AF_UNIX socket as a vertex and its file
descriptor as an edge in a directed graph.

This patch introduces a new struct unix_vertex representing a vertex
in the graph and adds its pointer to struct unix_sock.

When we send a fd using the SCM_RIGHTS message, we allocate struct
scm_fp_list to struct scm_cookie in scm_fp_copy().  Then, we bump
each refcount of the inflight fds' struct file and save them in
scm_fp_list.fp.

After that, unix_attach_fds() inexplicably clones scm_fp_list of
scm_cookie and sets it to skb.  (We will remove this part after
replacing GC.)

Here, we add a new function call in unix_attach_fds() to preallocate
struct unix_vertex per inflight AF_UNIX fd and link each vertex to
skb's scm_fp_list.vertices.

When sendmsg() succeeds later, if the socket of the inflight fd is
still not inflight yet, we will set the preallocated vertex to struct
unix_sock.vertex and link it to a global list unix_unvisited_vertices
under spin_lock(&unix_gc_lock).

If the socket is already inflight, we free the preallocated vertex.
This is to avoid taking the lock unnecessarily when sendmsg() could
fail later.

In the following patch, we will similarly allocate another struct
per edge, which will finally be linked to the inflight socket's
unix_vertex.edges.

And then, we will count the number of edges as unix_vertex.out_degree.

Signed-off-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Acked-by: Paolo Abeni <pabeni@redhat.com>
Link: https://lore.kernel.org/r/20240325202425.60930-2-kuniyu@amazon.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agotcp/dccp: bypass empty buckets in inet_twsk_purge()
Eric Dumazet [Wed, 27 Mar 2024 19:12:06 +0000 (19:12 +0000)]
tcp/dccp: bypass empty buckets in inet_twsk_purge()

TCP ehash table is often sparsely populated.

inet_twsk_purge() spends too much time calling cond_resched().

This patch can reduce time spent in inet_twsk_purge() by 20x.

Signed-off-by: Eric Dumazet <edumazet@google.com>
Reviewed-by: Kuniyuki Iwashima <kuniyu@amazon.com>
Link: https://lore.kernel.org/r/20240327191206.508114-1-edumazet@google.com
Signed-off-by: Jakub Kicinski <kuba@kernel.org>
12 months agonet: dsa: hellcreek: Convert to gettimex64()
Kurt Kanzenbach [Tue, 26 Mar 2024 11:16:36 +0000 (12:16 +0100)]
net: dsa: hellcreek: Convert to gettimex64()

As of commit 916444df305e ("ptp: deprecate gettime64() in favor of
gettimex64()") (new) PTP drivers should rather implement gettimex64().

In addition, this variant provides timestamps from the system clock. The
readings have to be recorded right before and after reading the lowest bits
of the PHC timestamp.

Reported-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Kurt Kanzenbach <kurt@linutronix.de>
Acked-by: Richard Cochran <richardcochran@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
12 months agonet/smc: make smc_hash_sk/smc_unhash_sk static
Zhengchao Shao [Tue, 26 Mar 2024 07:29:52 +0000 (15:29 +0800)]
net/smc: make smc_hash_sk/smc_unhash_sk static

smc_hash_sk and smc_unhash_sk are only used in af_smc.c, so make them
static and remove the output symbol. They can be called under the path
.prot->hash()/unhash().

Signed-off-by: Zhengchao Shao <shaozhengchao@huawei.com>
Reviewed-by: Wen Gu <guwen@linux.alibaba.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Signed-off-by: David S. Miller <davem@davemloft.net>
12 months agoMerge branch 'net-sched-skip_sw'
David S. Miller [Fri, 29 Mar 2024 09:46:39 +0000 (09:46 +0000)]
Merge branch 'net-sched-skip_sw'

Asbjørn Sloth Tønnesen says:

====================
make skip_sw actually skip software

During development of flower-route[1], which I
recently presented at FOSDEM[2], I noticed that
CPU usage, would increase the more rules I installed
into the hardware for IP forwarding offloading.

Since we use TC flower offload for the hottest
prefixes, and leave the long tail to the normal (non-TC)
Linux network stack for slow-path IP forwarding.
We therefore need both the hardware and software
datapath to perform well.

I found that skip_sw rules, are quite expensive
in the kernel datapath, since they must be evaluated
and matched upon, before the kernel checks the
skip_sw flag.

This patchset optimizes the case where all rules
are skip_sw, by implementing a TC bypass for these
cases, where TC is only used as a control plane
for the hardware path.

v4:
- Rebased onto net-next, now that net-next is open again

v3: https://lore.kernel.org/netdev/20240306165813.656931-1-ast@fiberby.net/
- Patch 3:
  - Fix source_inline
  - Fix build failure, when CONFIG_NET_CLS without CONFIG_NET_CLS_ACT.

v2: https://lore.kernel.org/netdev/20240305144404.569632-1-ast@fiberby.net/
- Patch 1:
  - Add Reviewed-By from Jiri Pirko
- Patch 2:
  - Move code, to avoid forward declaration (Jiri).
- Patch 3
  - Refactor to use a static key.
  - Add performance data for trapping, or sending
    a packet to a non-existent chain (as suggested by Marcelo).

v1: https://lore.kernel.org/netdev/20240215160458.1727237-1-ast@fiberby.net/

[1] flower-route
    https://github.com/fiberby-dk/flower-route

[2] FOSDEM talk
    https://fosdem.org/2024/schedule/event/fosdem-2024-3337-flying-higher-hardware-offloading-with-bird/
====================

Signed-off-by: David S. Miller <davem@davemloft.net>
12 months agonet: sched: make skip_sw actually skip software
Asbjørn Sloth Tønnesen [Mon, 25 Mar 2024 20:47:36 +0000 (20:47 +0000)]
net: sched: make skip_sw actually skip software

TC filters come in 3 variants:
- no flag (try to process in hardware, but fallback to software))
- skip_hw (do not process filter by hardware)
- skip_sw (do not process filter by software)

However skip_sw is implemented so that the skip_sw
flag can first be checked, after it has been matched.

IMHO it's common when using skip_sw, to use it on all rules.

So if all filters in a block is skip_sw filters, then
we can bail early, we can thus avoid having to match
the filters, just to check for the skip_sw flag.

This patch adds a bypass, for when only TC skip_sw rules
are used. The bypass is guarded by a static key, to avoid
harming other workloads.

There are 3 ways that a packet from a skip_sw ruleset, can
end up in the kernel path. Although the send packets to a
non-existent chain way is only improved a few percents, then
I believe it's worth optimizing the trap and fall-though
use-cases.

 +----------------------------+--------+--------+--------+
 | Test description           | Pre-   | Post-  | Rel.   |
 |                            | kpps   | kpps   | chg.   |
 +----------------------------+--------+--------+--------+
 | basic forwarding + notrack | 3589.3 | 3587.9 |  1.00x |
 | switch to eswitch mode     | 3081.8 | 3094.7 |  1.00x |
 | add ingress qdisc          | 3042.9 | 3063.6 |  1.01x |
 | tc forward in hw / skip_sw |37024.7 |37028.4 |  1.00x |
 | tc forward in sw / skip_hw | 3245.0 | 3245.3 |  1.00x |
 +----------------------------+--------+--------+--------+
 | tests with only skip_sw rules below:                  |
 +----------------------------+--------+--------+--------+
 | 1 non-matching rule        | 2694.7 | 3058.7 |  1.14x |
 | 1 n-m rule, match trap     | 2611.2 | 3323.1 |  1.27x |
 | 1 n-m rule, goto non-chain | 2886.8 | 2945.9 |  1.02x |
 | 5 non-matching rules       | 1958.2 | 3061.3 |  1.56x |
 | 5 n-m rules, match trap    | 1911.9 | 3327.0 |  1.74x |
 | 5 n-m rules, goto non-chain| 2883.1 | 2947.5 |  1.02x |
 | 10 non-matching rules      | 1466.3 | 3062.8 |  2.09x |
 | 10 n-m rules, match trap   | 1444.3 | 3317.9 |  2.30x |
 | 10 n-m rules,goto non-chain| 2883.1 | 2939.5 |  1.02x |
 | 25 non-matching rules      |  838.5 | 3058.9 |  3.65x |
 | 25 n-m rules, match trap   |  824.5 | 3323.0 |  4.03x |
 | 25 n-m rules,goto non-chain| 2875.8 | 2944.7 |  1.02x |
 | 50 non-matching rules      |  488.1 | 3054.7 |  6.26x |
 | 50 n-m rules, match trap   |  484.9 | 3318.5 |  6.84x |
 | 50 n-m rules,goto non-chain| 2884.1 | 2939.7 |  1.02x |
 +----------------------------+--------+--------+--------+

perf top (25 n-m skip_sw rules - pre patch):
  20.39%  [kernel]  [k] __skb_flow_dissect
  16.43%  [kernel]  [k] rhashtable_jhash2
  10.58%  [kernel]  [k] fl_classify
  10.23%  [kernel]  [k] fl_mask_lookup
   4.79%  [kernel]  [k] memset_orig
   2.58%  [kernel]  [k] tcf_classify
   1.47%  [kernel]  [k] __x86_indirect_thunk_rax
   1.42%  [kernel]  [k] __dev_queue_xmit
   1.36%  [kernel]  [k] nft_do_chain
   1.21%  [kernel]  [k] __rcu_read_lock

perf top (25 n-m skip_sw rules - post patch):
   5.12%  [kernel]  [k] __dev_queue_xmit
   4.77%  [kernel]  [k] nft_do_chain
   3.65%  [kernel]  [k] dev_gro_receive
   3.41%  [kernel]  [k] check_preemption_disabled
   3.14%  [kernel]  [k] mlx5e_skb_from_cqe_mpwrq_nonlinear
   2.88%  [kernel]  [k] __netif_receive_skb_core.constprop.0
   2.49%  [kernel]  [k] mlx5e_xmit
   2.15%  [kernel]  [k] ip_forward
   1.95%  [kernel]  [k] mlx5e_tc_restore_tunnel
   1.92%  [kernel]  [k] vlan_gro_receive

Test setup:
 DUT: Intel Xeon D-1518 (2.20GHz) w/ Nvidia/Mellanox ConnectX-6 Dx 2x100G
 Data rate measured on switch (Extreme X690), and DUT connected as
 a router on a stick, with pktgen and pktsink as VLANs.
 Pktgen-dpdk was in range 36.6-37.7 Mpps 64B packets across all tests.
 Full test data at https://files.fiberby.net/ast/2024/tc_skip_sw/v2_tests/

Signed-off-by: Asbjørn Sloth Tønnesen <ast@fiberby.net>
Reviewed-by: Simon Horman <horms@kernel.org>
Reviewed-by: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
12 months agonet: sched: cls_api: add filter counter
Asbjørn Sloth Tønnesen [Mon, 25 Mar 2024 20:47:35 +0000 (20:47 +0000)]
net: sched: cls_api: add filter counter

Maintain a count of filters per block.

Counter updates are protected by cb_lock, which is
also used to protect the offload counters.

Signed-off-by: Asbjørn Sloth Tønnesen <ast@fiberby.net>
Reviewed-by: Simon Horman <horms@kernel.org>
Reviewed-by: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
12 months agonet: sched: cls_api: add skip_sw counter
Asbjørn Sloth Tønnesen [Mon, 25 Mar 2024 20:47:34 +0000 (20:47 +0000)]
net: sched: cls_api: add skip_sw counter

Maintain a count of skip_sw filters.

This counter is protected by the cb_lock, and is updated
at the same time as offloadcnt.

Signed-off-by: Asbjørn Sloth Tønnesen <ast@fiberby.net>
Reviewed-by: Jiri Pirko <jiri@nvidia.com>
Reviewed-by: Simon Horman <horms@kernel.org>
Reviewed-by: Marcelo Ricardo Leitner <marcelo.leitner@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>