net: dsa: initial support for MaxLinear MxL862xx switches
Subsystems (6)
Daniel Golle
maintainerv13 cover net: dsa: initial support for MaxLinear MxL862xx switches
This series adds very basic DSA support for the MaxLinear MxL86252 (5x 2500Base-T PHYs) and MxL86282 (8x 2500Base-T PHYs) switches. In addition to the 2.5G TP ports both switches also come with two SerDes interfaces which can be used either to connect external PHYs or SFP cages, or as CPU port when using the switch with this DSA driver.
MxL862xx integrates a firmware running on an embedded processor (based on Zephyr RTOS). Host interaction uses a simple netlink-like API transported over MDIO/MMD.
This series includes only what's needed to pass traffic between user ports and the CPU port: relayed MDIO to internal PHYs, basic port enable/disable, and CPU-port special tagging.
The SerDes interface of the CPU port is automatically configured by the switch after reset using a board-specific configuration stored together with the firmware in the flash chip attached to the switch, so no action is needed from the driver to setup the interface mode of the CPU port.
Also MAC settings of the PHY ports are automatically configured, which means the driver works fine with phylink_mac_ops being all no-op stubs.
Multiple follow up series will bring support for setting up the other SerDes PCS interface (ie. not used for the CPU port), bridge, VLAN, ... offloading, and support for using an 802.1Q-based special tag instead of the proprietary 8-byte tag.
---
basic DSA selftests were run, results:
* no_forwarding.sh: all tests PASS
* bridge_vlan_unaware.sh: all tests PASS
* bridge_vlan_mcast.sh: all tests PASS
* bridge_vlan_aware.sh: all tests PASS
* local_termination.sh: all tests PASS or XFAIL, except for
TEST: VLAN over vlan_filtering=1 bridged port: Unicast IPv4 to unknown MAC address [FAIL]
reception succeeded, but should have failed
TEST: VLAN over vlan_filtering=1 bridged port: Unicast IPv4 to unknown MAC address, allmulti [FAIL]
reception succeeded, but should have failed
As obviously this is mostly testing the Linux software bridge at this point so I didn't bother to run any of the FDB or MDB related tests.
Some mystery about the relationship between the sub-interface ID and CTP port assignment remains for now, so I cannot yet fully explain the magic numbers when calculating the two.
(Why is it neccesary to allocate 23 CTPs when using port 9 as CPU port while there are at most 17 phyiscal interfaces in total, and only at most 13 when using one of the two SerDes as CPU interface? Why is there an offset of 16 when calculating the sub-interface ID? How will a total of 17 ports be possibly represented in a 4-bit field in the tag RX path?)
At this point this doesn't impact functionality, and I hope for further clarification discussing all that with MaxLinear engineers.
I believe the initial driver is anyway in good shape now, a follow-up series addressing configuring the SerDes interfaces will also have to take care of CTP assignment and sub-interfaces to not hit the mentioned 4-bit cap in the tag driver.
Changes since v12:
Address comments from Vladimir Oltean and add Reviewed-by:-tag given by Russell King (Oracle).
1/4 dt-bindings: net: dsa: add MaxLinear MxL862xx
* 'mdio' node is required
2/4 net: dsa: add tag format for MxL862xx switches
* no changes
3/4 net: mdio: add unlocked mdiodev C45 bus accessors
* add Russell's Reviewed-by:-tag given for v11
4/4 net: dsa: add basic initial driver for MxL862xx switches
* use '#pragma pack(push, 1)' in mxl862xx-api.h
* relax timeout in mxl862xx_busy_wait() from 10000us to 500000us, this turned out to be needed when using the driver on older firmware
* reorder functions
* remove unneeded checks for dp or cpu_dp being !=NULL, add comments
Changes since v11:
Address comments reported by Jakub Kicinski using netdev-ai
https://patchwork.kernel.org/comment/26767905/
1/4 dt-bindings: net: dsa: add MaxLinear MxL862xx
* no changes
2/4 net: dsa: add tag format for MxL862xx switches
* no changes
3/4 net: mdio: add unlocked mdiodev C45 bus accessors
* no changes
4/4 net: dsa: add basic initial driver for MxL862xx switches
* list switch variants in Kconfig starting with dash ('-')
* remove the __packed attribute on structs which are naturally packed according to C rules (ie. one other struct with only a single byte-aligned member)
* log error in mxl862xx_port_disable()
* introduce !NULL check for return value of dsa_to_port() in mxl862xx_add_single_port_bridge
* check cpu_dp being non-NULL before dereferencing dp->cpu_dp
* use non-racy and deterministic name for MII bus
* skip ports without cpu_dp assigned in mxl862xx_setup_cpu_bridge() to avoid potential NULL-pointer dereference
* call dev_set_drvdata() only after dsa_register_switch() has been successfully completed
Changes since v10:
1/4 dt-bindings: net: dsa: add MaxLinear MxL862xx
* no changes
2/4 net: dsa: add tag format for MxL862xx switches
* arrange local variables in reverse xmas tree order (again)
* use MXL862_NAME as .name in struct dsa_device_ops
3/4 net: mdio: add unlocked mdiodev C45 bus accessors
* no changes
4/4 net: dsa: add basic initial driver for MxL862xx switches
* fix fatal last-second changes dsa_port_is_unused vs. dsa_is_unused_port and dsa_port_is_unused vs. dsa_is_unused_port
* use cpu_to_le16 in mxl862xx_configure_ctp_port
* improve formatting of some kerneldoc comments
Changes since v9:
1/4 dt-bindings: net: dsa: add MaxLinear MxL862xx
* add microcontroller port 0
2/4 net: dsa: add tag format for MxL862xx switches
* count ports starting from 0
3/4 net: mdio: add unlocked mdiodev C45 bus accessors
* no changes
4/4 net: dsa: add basic initial driver for MxL862xx switches
* fix indentation in kernel-doc
* remove misleading comment regarding sub-interface IDs
* remove unused macros, set MXL862XX_MAX_PORTS to 17
* count ports starting from 0, 0 being the microcontroller
* improve port setup:
- split tag protocol and CTP port assignment
- use ports cpu_dp instead of relying on single (ie. first) CPU port
- handle CPU port like any other port
- prepare forward DT compatibility with multiple CPU ports
Changes since v8:
1/4 dt-bindings: net: dsa: add MaxLinear MxL862xx
* no changes
2/4 net: dsa: add tag format for MxL862xx switches
* no changes
3/4 net: mdio: add unlocked mdiodev C45 bus accessors
* no changes
4/4 net: dsa: add basic initial driver for MxL862xx switches
* remove practically unused struct hw_info
* lots of kerneldoc improvements in mxl862xx-api.h
* drop .mac_select_pcs() stub
* better handling for firmware error return value
* apply reverse xmas tree in mxl862xx_api_wrap
* guard headers with #ifdef macro
* include net/dsa.h and linux/mdio.h in mxl862xx.h
* call mxl862xx_port_fast_age() only once in .port_setup
* don't create isolation bridges for unused ports
* replace errornous cast with correct range of values denoting firmware errors
Changes since v7
1/4 dt-bindings: net: dsa: add MaxLinear MxL862xx
* no changes
2/4 net: dsa: add tag format for MxL862xx switches
* no changes
3/4 net: mdio: add unlocked mdiodev C45 bus accessors
* no changes
4/4 net: dsa: add basic initial driver for MxL862xx switches
* use little-endian in bridge_port_config API
* remove duplciate assignment of br_port_cfg.bridge_port_id when setting up CPU port
Changes since v6
1/4 dt-bindings: net: dsa: add MaxLinear MxL862xx
* no changes
2/4 net: dsa: add tag format for MxL862xx switches
* no changes
3/4 net: mdio: add unlocked mdiodev C45 bus accessors
* no changes
4/4 net: dsa: add basic initial driver for MxL862xx switches
* fix kerneldoc style
Changes since RFC v5
1/4 dt-bindings: net: dsa: add MaxLinear MxL862xx
* no changes
2/4 net: dsa: add tag format for MxL862xx switches
* remove unnecessary check for skb != NULL
* merge consecutively printed warnings into single dev_warn_ratelimited
3/4 net: mdio: add unlocked mdiodev C45 bus accessors
* no changes
4/4 net: dsa: add basic initial driver for MxL862xx switches
* include bridge and bridgeport API needed to isolate ports
* remove warning in .setup as ports are now isolated
* make ready-after-reset check more robust by adding delay
* sort structs in order of struct definitions
* best effort to sort functions without introducing additional prototypes
* always use enums with kerneldoc comments in mxl862xx-api.h
* remove bogus .phy_read and .phy_write DSA ops as the driver anyway registers a user MDIO bus with Clause-22 and Clause-45 operations
* various small style fixes
Changes since RFC v4
1/4 dt-bindings: net: dsa: add MaxLinear MxL862xx
* no changes
2/4 net: dsa: add tag format for MxL862xx switches
* drop unused precompiler macros
3/4 net: mdio: add unlocked mdiodev C45 bus accessors
* fix indentation
4/4 net: dsa: add basic initial driver for MxL862xx switches
* output warning in .setup regarding unknown pre-configuration
* add comment explaining why CFGGET is used in reset function
Changes since RFC v3
1/4 dt-bindings: net: dsa: add MaxLinear MxL862xx
* remove labels from example
* remove 'bindings for' from commit title
2/4 net: dsa: add tag format for MxL862xx switches
* describe fields and variables with comments
* sub-interface is only 5 bits
* harmonize Kconfig symbol name
* maintain alphabetic order in Kconfig
* fix typo s/beginnig/beginning/
* fix typo s/swtiches/switches/
* arrange local variables in reverse xmas tree order
3/4 net: mdio: add unlocked mdiodev C45 bus accessors
* unchanged
4/4 net: dsa: add basic initial driver for MxL862xx switches
* poll switch readiness after reset
* implement driver shutdown
* added port_fast_aging API call and driver op
* unified port setup in new .port_setup op
* improve comment explaining special handlign for unaligned API read
* various typos and formatting improvements
Changes since RFC v2
1/4, 2/4, 3/4: unchanged
4/4 net: dsa: add basic initial driver for MxL862xx switches
* fix return value being uninitialized on error in mxl862xx_api_wrap()
* add missing description in kerneldoc comment of
struct mxl862xx_ss_sp_tag
Changes since initial RFC
1/4 dt-bindings: net: dsa: add bindings for MaxLinear MxL862xx
* better description in dt-bindings doc
2/4 net: dsa: add tag formats for MxL862xx switches
* make sure all tag fields are initialized
3/4 net: mdio: add unlocked mdiodev C45 bus accessors
* new patch
4/4 net: dsa: add basic initial driver for MxL862xx switches
* make use of struct mdio_device
* add phylink_mac_ops stubs
* drop leftover nonsense from mxl862xx_phylink_get_caps()
* fix endian conversions
* use __le32 instead of enum types in over-the-wire structs
* use existing MDIO_* macros whenever possible
* simplify API constants to be more readable
* use readx_poll_timeout instead of open-coding poll timeout loop
* add mxl862xx_reg_read() and mxl862xx_reg_write() helpers
* demystify error codes returned by the firmware
* add #defines for mxl862xx_ss_sp_tag member values
* move reset to dedicated function, clarify magic number being the reset command ID
Daniel Golle (4):
dt-bindings: net: dsa: add MaxLinear MxL862xx
net: dsa: add tag format for MxL862xx switches
net: mdio: add unlocked mdiodev C45 bus accessors
net: dsa: add basic initial driver for MxL862xx switches
.../bindings/net/dsa/maxlinear,mxl862xx.yaml | 161 +++++ MAINTAINERS | 8 + drivers/net/dsa/Kconfig | 2 + drivers/net/dsa/Makefile | 1 + drivers/net/dsa/mxl862xx/Kconfig | 12 + drivers/net/dsa/mxl862xx/Makefile | 3 + drivers/net/dsa/mxl862xx/mxl862xx-api.h | 680 ++++++++++++++++++ drivers/net/dsa/mxl862xx/mxl862xx-cmd.h | 49 ++ drivers/net/dsa/mxl862xx/mxl862xx-host.c | 245 +++++++ drivers/net/dsa/mxl862xx/mxl862xx-host.h | 12 + drivers/net/dsa/mxl862xx/mxl862xx.c | 476 ++++++++++++ drivers/net/dsa/mxl862xx/mxl862xx.h | 16 + include/linux/mdio.h | 13 + include/net/dsa.h | 2 + net/dsa/Kconfig | 7 + net/dsa/Makefile | 1 + net/dsa/tag_mxl862xx.c | 110 +++ 17 files changed, 1798 insertions(+) create mode 100644 Documentation/devicetree/bindings/net/dsa/maxlinear,mxl862xx.yaml create mode 100644 drivers/net/dsa/mxl862xx/Kconfig create mode 100644 drivers/net/dsa/mxl862xx/Makefile create mode 100644 drivers/net/dsa/mxl862xx/mxl862xx-api.h create mode 100644 drivers/net/dsa/mxl862xx/mxl862xx-cmd.h create mode 100644 drivers/net/dsa/mxl862xx/mxl862xx-host.c create mode 100644 drivers/net/dsa/mxl862xx/mxl862xx-host.h create mode 100644 drivers/net/dsa/mxl862xx/mxl862xx.c create mode 100644 drivers/net/dsa/mxl862xx/mxl862xx.h create mode 100644 net/dsa/tag_mxl862xx.c
Signature · 1 line
2.52.0
Daniel Golle
maintainerPATCH v13 1/4 dt-bindings: net: dsa: add MaxLinear MxL862xx
Add documentation and an example for MaxLinear MxL86282 and MxL86252 switches.
---
v13: 'mdio' node is required
v12: no changes
v11: no changes
v10: add microcontroller port 0
v9: no changes
v8: no changes
v7: no changes
v6: no changes
v5: no changes
RFC v4:
* remove labels from example
* remove 'bindings for' from commit title
RFC v3: no changes
RFC v2: better description in dt-bindings doc
--- .../bindings/net/dsa/maxlinear,mxl862xx.yaml | 161 ++++++++++++++++++ MAINTAINERS | 6 + 2 files changed, 167 insertions(+) create mode 100644 Documentation/devicetree/bindings/net/dsa/maxlinear,mxl862xx.yaml
Documentation/devicetree/bindings/net/dsa/maxlinear,mxl862xx.yaml new file +161 -0
@@ -0,0 +1,161 @@ +# SPDX-License-Identifier: (GPL-2.0-only OR BSD-2-Clause) +%YAML 1.2 +--- +$id: http://devicetree.org/schemas/net/dsa/maxlinear,mxl862xx.yaml# +$schema: http://devicetree.org/meta-schemas/core.yaml# + +title: MaxLinear MxL862xx Ethernet Switch Family + +maintainers: + - Daniel Golle <[email protected]> + +description: + The MaxLinear MxL862xx switch family are multi-port Ethernet switches with + integrated 2.5GE PHYs. The MxL86252 has five PHY ports and the MxL86282 + has eight PHY ports. Both models come with two 10 Gigabit/s SerDes + interfaces to be used to connect external PHYs or SFP cages, or as CPU + port. + +allOf: + - $ref: dsa.yaml#/$defs/ethernet-ports + +properties: + compatible: + enum: + - maxlinear,mxl86252 + - maxlinear,mxl86282 + + reg: + maxItems: 1 + description: MDIO address of the switch + + mdio: + $ref: /schemas/net/mdio.yaml# + unevaluatedProperties: false + +required: + - compatible + - mdio + - reg + +unevaluatedProperties: false + +examples: + - | + mdio { + #address-cells = <1>; + #size-cells = <0>; + + switch@0 { + compatible = "maxlinear,mxl86282"; + reg = <0>; + + ethernet-ports { + #address-cells = <1>; + #size-cells = <0>; + + /* Microcontroller port */ + port@0 { + reg = <0>; + status = "disabled"; + }; + + port@1 { + reg = <1>; + phy-handle = <&phy0>; + phy-mode = "internal"; + }; + + port@2 { + reg = <2>; + phy-handle = <&phy1>; + phy-mode = "internal"; + }; + + port@3 { + reg = <3>; + phy-handle = <&phy2>; + phy-mode = "internal"; + }; + + port@4 { + reg = <4>; + phy-handle = <&phy3>; + phy-mode = "internal"; + }; + + port@5 { + reg = <5>; + phy-handle = <&phy4>; + phy-mode = "internal"; + }; + + port@6 { + reg = <6>; + phy-handle = <&phy5>; + phy-mode = "internal"; + }; + + port@7 { + reg = <7>; + phy-handle = <&phy6>; + phy-mode = "internal"; + }; + + port@8 { + reg = <8>; + phy-handle = <&phy7>; + phy-mode = "internal"; + }; + + port@9 { + reg = <9>; + label = "cpu"; + ethernet = <&gmac0>; + phy-mode = "usxgmii"; + + fixed-link { + speed = <10000>; + full-duplex; + }; + }; + }; + + mdio { + #address-cells = <1>; + #size-cells = <0>; + + phy0: ethernet-phy@0 { + reg = <0>; + }; + + phy1: ethernet-phy@1 { + reg = <1>; + }; + + phy2: ethernet-phy@2 { + reg = <2>; + }; + + phy3: ethernet-phy@3 { + reg = <3>; + }; + + phy4: ethernet-phy@4 { + reg = <4>; + }; + + phy5: ethernet-phy@5 { + reg = <5>; + }; + + phy6: ethernet-phy@6 { + reg = <6>; + }; + + phy7: ethernet-phy@7 { + reg = <7>; + }; + }; + }; + };
MAINTAINERS +6 -0
@@ -15619,6 +15619,12 @@ S: Supported F: drivers/net/phy/mxl-86110.c F: drivers/net/phy/mxl-gpy.c +MAXLINEAR MXL862XX SWITCH DRIVER +M: Daniel Golle <[email protected]> +L: [email protected] +S: Maintained +F: Documentation/devicetree/bindings/net/dsa/maxlinear,mxl862xx.yaml + MCAN DEVICE DRIVER M: Markus Schneider-Pargmann <[email protected]> L: [email protected]
Signature · 1 line
2.52.0
Daniel Golle
maintainerPATCH v13 2/4 net: dsa: add tag format for MxL862xx switches
In reply to Daniel GolleAdd proprietary special tag format for the MaxLinear MXL862xx family of switches. While using the same Ethertype as MaxLinear's GSW1xx switches, the actual tag format differs significantly, hence we need a dedicated tag driver for that.
---
v13: no changes
v12: no changes
v11:
* arrange local variables in reverse xmas tree order (again)
* use MXL862_NAME as .name in struct dsa_device_ops
v10: count ports starting from 0
v9: no changes
v8: no changes
v7: no changes
v6:
* remove unnecessary check for skb != NULL
* merge consecutively printed warnings into single dev_warn_ratelimited
v5:
* remove unused macro definitions
RFC v4:
* describe fields and variables with comments
* sub-interface is only 5 bits
* harmonize Kconfig symbol name
* maintain alphabetic order in Kconfig
* fix typo s/beginnig/beginning/
* fix typo s/swtiches/switches/
* arrange local variables in reverse xmas tree order
RFC v3: no changes
RFC v2: make sure all tag fields are initialized
--- MAINTAINERS | 1 + include/net/dsa.h | 2 + net/dsa/Kconfig | 7 +++ net/dsa/Makefile | 1 + net/dsa/tag_mxl862xx.c | 110 +++++++++++++++++++++++++++++++++++++++++ 5 files changed, 121 insertions(+) create mode 100644 net/dsa/tag_mxl862xx.c
MAINTAINERS +1 -0
@@ -15624,6 +15624,7 @@ M: Daniel Golle <[email protected]> L: [email protected] S: Maintained F: Documentation/devicetree/bindings/net/dsa/maxlinear,mxl862xx.yaml +F: net/dsa/tag_mxl862xx.c MCAN DEVICE DRIVER M: Markus Schneider-Pargmann <[email protected]>
include/net/dsa.h +2 -0
@@ -57,6 +57,7 @@ struct tc_action;#define DSA_TAG_PROTO_BRCM_LEGACY_FCS_VALUE 29#define DSA_TAG_PROTO_YT921X_VALUE 30#define DSA_TAG_PROTO_MXL_GSW1XX_VALUE 31 +#define DSA_TAG_PROTO_MXL862_VALUE 32enum dsa_tag_protocol {DSA_TAG_PROTO_NONE =DSA_TAG_PROTO_NONE_VALUE ,
@@ -91,6 +92,7 @@ enum dsa_tag_protocol {DSA_TAG_PROTO_VSC73XX_8021Q =DSA_TAG_PROTO_VSC73XX_8021Q_VALUE ,DSA_TAG_PROTO_YT921X =DSA_TAG_PROTO_YT921X_VALUE ,DSA_TAG_PROTO_MXL_GSW1XX =DSA_TAG_PROTO_MXL_GSW1XX_VALUE , + DSA_TAG_PROTO_MXL862 = DSA_TAG_PROTO_MXL862_VALUE, };struct dsa_switch ;
net/dsa/Kconfig +7 -0
@@ -104,6 +104,13 @@ config NET_DSA_TAG_MTK Say Y or M if you want to enable support for tagging frames for Mediatek switches. +config NET_DSA_TAG_MXL_862XX + tristate "Tag driver for MxL862xx switches" + help + Say Y or M if you want to enable support for tagging frames for the + Maxlinear MxL86252 and MxL86282 switches using their native 8-byte + tagging protocol. + config NET_DSA_TAG_MXL_GSW1XX tristate "Tag driver for MaxLinear GSW1xx switches" help
net/dsa/Makefile +1 -0
@@ -28,6 +28,7 @@ obj-$(CONFIG_NET_DSA_TAG_HELLCREEK) += tag_hellcreek.o obj-$(CONFIG_NET_DSA_TAG_KSZ) += tag_ksz.o obj-$(CONFIG_NET_DSA_TAG_LAN9303) += tag_lan9303.o obj-$(CONFIG_NET_DSA_TAG_MTK) += tag_mtk.o +obj-$(CONFIG_NET_DSA_TAG_MXL_862XX) += tag_mxl862xx.o obj-$(CONFIG_NET_DSA_TAG_MXL_GSW1XX) += tag_mxl-gsw1xx.o obj-$(CONFIG_NET_DSA_TAG_NONE) += tag_none.o obj-$(CONFIG_NET_DSA_TAG_OCELOT) += tag_ocelot.o
net/dsa/tag_mxl862xx.c new file +110 -0
110 lines not on this page · raw message
Signature · 1 line
2.52.0
Jakub Kicinski
maintainerRe: [net-next,v13,2/4] net: dsa: add tag format for MxL862xx switches
This is an AI-generated review of your patch. The human sending this email has considered the AI review valid, or at least plausible.
For local reproduction steps see https://netdev-ai.bots.linux.dev/ai-local.html
---
net: dsa: add tag format for MxL862xx switches
This patch adds a DSA tag driver for MaxLinear MxL862xx switches. The tag format uses an 8-byte header with an ethertype marker that is shared with GSW1xx switches, though the format itself differs.
Quoted 4 lines
net/dsa/Kconfig +0
-0@@ -104,6 +104,13 @@
[ ... ]
net/dsa/Kconfig config NET_DSA_TAG_MTK · line 111
Quoted 5 lines
net/dsa/Kconfig +5
-0+config NET_DSA_TAG_MXL_862XX + tristate "Tag driver for MxL862xx switches" + help + Say Y or M if you want to enable support for tagging frames for the + Maxlinear MxL86252 and MxL86282 switches using their native 8-byte
^^^^^^^^^
This isn't a bug, but there's inconsistent capitalization here. The help text uses "Maxlinear" while the adjacent NET_DSA_TAG_MXL_GSW1XX entry uses "MaxLinear" (with capital L). The MODULE_DESCRIPTION in tag_mxl862xx.c also uses "MaxLinear".
Quoted 6 lines
net/dsa/Kconfig +1
-0+ tagging protocol.net/dsa/tag_mxl862xx.c new file +0
-0@@ -0,0 +1,110 @@
[ ... ]
net/dsa/tag_mxl862xx.c line 38
Quoted 12 lines
net/dsa/tag_mxl862xx.c new file +12
-0+static struct sk_buff *mxl862_tag_xmit (struct sk_buff *skb, +struct net_device *dev) +{ +struct dsa_port *dp =dsa_user_to_port (dev); +struct dsa_port *cpu_dp = dp->cpu_dp; +unsigned int cpu_port, sub_interface; +__be16 *mxl862_tag; + + cpu_port = cpu_dp->index; + +/* target port sub-interface ID relative to the CPU port */ + sub_interface = dp->index +16 - cpu_port;
^^^^^^^^^^^^^^^^^^^^^^^^
Can sub_interface overflow the 5-bit MXL862_SUBIF_ID field here? With dp->index = 16 and cpu_port = 0, the calculation would yield 32 which exceeds the valid range (0-31) for GENMASK(4, 0). FIELD_PREP would silently truncate this to 0 via masking. Is there a constraint on port configurations that prevents this from occurring?
Quoted 3 lines
net/dsa/tag_mxl862xx.c new file +3
-0+ +/* provide additional space 'MXL862_HEADER_LEN' bytes */ +skb_push (skb,MXL862_HEADER_LEN );
[ ... ]
net/dsa/tag_mxl862xx.c line 49
Quoted 3 lines
net/dsa/tag_mxl862xx.c new file +3
-0+dsa_alloc_etype_header (skb,MXL862_HEADER_LEN ); + +/* special tag ingress */
^^^^^^^
This comment says "ingress" but mxl862_tag_xmit() is the TX/egress path. Should this say "egress" instead?
Quoted 1 line
net/dsa/tag_mxl862xx.c new file +1
-0+ mxl862_tag =dsa_etype_header_pos_tx (skb);
Signature · 1 line
pw-bot: cr
Daniel Golle
maintainerRe: [net-next,v13,2/4] net: dsa: add tag format for MxL862xx switches
Quoted 25 lines · Jakub Kicinski
This is an AI-generated review of your patch. The human sending this email has considered the AI review valid, or at least plausible.
For local reproduction steps see https://netdev-ai.bots.linux.dev/ai-local.html
---
net: dsa: add tag format for MxL862xx switchesThis patch adds a DSA tag driver for MaxLinear MxL862xx switches. The tag format uses an 8-byte header with an ethertype marker that is shared with GSW1xx switches, though the format itself differs.
Quoted 4 lines
net/dsa/Kconfig +0
-0@@ -104,6 +104,13 @@[ ... ]
net/dsa/Kconfig config NET_DSA_TAG_MTK · line 111
Quoted 5 lines
net/dsa/Kconfig +5
-0+config NET_DSA_TAG_MXL_862XX + tristate "Tag driver for MxL862xx switches" + help + Say Y or M if you want to enable support for tagging frames for the + Maxlinear MxL86252 and MxL86282 switches using their native 8-byte^^^^^^^^^
Quoted 4 lines
This isn't a bug, but there's inconsistent capitalization here. The help text uses "Maxlinear" while the adjacent
NET_DSA_TAG_MXL_GSW1XXentry uses "MaxLinear" (with capital L). TheMODULE_DESCRIPTIONintag_mxl862xx.calso uses "MaxLinear".
+1 Will fix that.
Quoted 24 lines
Quoted 7 lines
net/dsa/Kconfig +1
-0+ tagging protocol.net/dsa/tag_mxl862xx.c new file +0
-0@@ -0,0 +1,110 @@[ ... ]
net/dsa/tag_mxl862xx.c line 38
Quoted 12 lines
net/dsa/tag_mxl862xx.c new file +12
-0+static struct sk_buff *mxl862_tag_xmit (struct sk_buff *skb, +struct net_device *dev) +{ +struct dsa_port *dp =dsa_user_to_port (dev); +struct dsa_port *cpu_dp = dp->cpu_dp; +unsigned int cpu_port, sub_interface; +__be16 *mxl862_tag; + + cpu_port = cpu_dp->index; + +/* target port sub-interface ID relative to the CPU port */ + sub_interface = dp->index +16 - cpu_port;^^^^^^^^^^^^^^^^^^^^^^^^
Quoted 5 lines
Can
sub_interfaceoverflow the 5-bitMXL862_SUBIF_IDfield here? Withdp->index= 16 andcpu_port= 0, the calculation would yield 32 which exceeds the valid range (0-31) forGENMASK(4, 0).FIELD_PREPwould silently truncate this to 0 via masking. Is there a constraint on port configurations that prevents this from occurring?
Yes. Only the two SerDes interfaces (port 9 and port 13) can be used as CPU ports. Also, when using a USXGMII-M 4-port PHY, it is recommended to connect the PHY to SerDes 1 (port 9 to 12) while using port 13 as CPU port.
Addressing port 16 is problematic when using the switch with DSA also because the 4-bit EGP/IGP field used in the receive function can only present port 0 to 15 obviously. There is a way to do it by also using the sub-interface ID on switch-to-host tagged frames, but it is worth avoiding that due to the performance hit which comes with the additional lookup and multiple pointer dereference steps required when doing so.
As there are only disadvantages when using SerDes 1 for a USXGMII-M quad-PHY while using SerDes 0 as CPU port it is reasonable to expect that board designers will opt for the convenient and also vendor-recommended way and use SerDes 0 for the quad-PHY while using SerDes 1 as CPU port.
Hence this theoretical limitation will play no role in practise.
Quoted 7 lines
Quoted 4 lines
net/dsa/tag_mxl862xx.c new file +3
-0+ +/* provide additional space 'MXL862_HEADER_LEN' bytes */ +skb_push (skb,MXL862_HEADER_LEN );[ ... ]
Quoted 7 lines
net/dsa/tag_mxl862xx.c line 49
Quoted 3 lines
net/dsa/tag_mxl862xx.c new file +3
-0+dsa_alloc_etype_header (skb,MXL862_HEADER_LEN ); + +/* special tag ingress */ ^^^^^^^This comment says "ingress" but
mxl862_tag_xmit()is the TX/egress path. Should this say "egress" instead?
"ingress" is the perspective of the switch, and how the tag to be used to send frames from the host to the switch is called in the datasheet[1]
[1]: https://www.mouser.com/datasheet/2/146/621181_MxL86282I_DS_Rev1_1-3675315.pdf
page 95
Jakub Kicinski
maintainerRe: [net-next,v13,2/4] net: dsa: add tag format for MxL862xx switches
Quoted 11 lines · On Fri, 6 Feb 2026 03:30:30 +0000 Daniel Golle wrote:
Quoted 11 lines
Quoted 3 lines
net/dsa/tag_mxl862xx.c new file +3
-0+ +/* provide additional space 'MXL862_HEADER_LEN' bytes */ +skb_push (skb,MXL862_HEADER_LEN );[ ... ]
net/dsa/tag_mxl862xx.c line 49
Quoted 3 lines
net/dsa/tag_mxl862xx.c new file +3
-0+dsa_alloc_etype_header (skb,MXL862_HEADER_LEN ); + +/* special tag ingress */ ^^^^^^^
Quoted 5 lines
Quoted 2 lines
This comment says "ingress" but
mxl862_tag_xmit()is the TX/egress path. Should this say "egress" instead?"ingress" is the perspective of the switch, and how the tag to be used to send frames from the host to the switch is called in the datasheet[1]
Maybe expand on the comment a bit? Or remove it?
Cause reading the code I was also consued why this is called ingress
Daniel Golle
maintainerPATCH v13 3/4 net: mdio: add unlocked mdiodev C45 bus accessors
In reply to Daniel GolleAdd helper inline functions __mdiodev_c45_read() and __mdiodev_c45_write(), which are the C45 equivalents of the existing __mdiodev_read() and __mdiodev_write() added by commit e6a45700e7e1 ("net: mdio: add unlocked mdiobus and mdiodev bus accessors")
---
v13: add Russell's Reviewed-by:-tag given for v11
v12: no changes
v11: no changes
v10: no changes
v9: no changes
v8: no changes
v7: no changes
v6: no changes
v5: fix indentation
RFC v4: no changes
RFC v3: no changes
RFC v2: add this patch, initial submission
--- include/linux/mdio.h | 13 +++++++++++++ 1 file changed, 13 insertions(+)
include/linux/mdio.h +13 -0
@@ -647,6 +647,19 @@ static inline int mdiodev_modify_changed(struct mdio_device *mdiodev, mask, set); } +static inline int __mdiodev_c45_read (struct mdio_device *mdiodev,int devad, +u16 regnum) +{ +return __mdiobus_c45_read (mdiodev->bus, mdiodev->addr, devad, regnum); +} + +static inline int __mdiodev_c45_write (struct mdio_device *mdiodev,u32 devad, +u16 regnum,u16 val) +{ +return __mdiobus_c45_write (mdiodev->bus, mdiodev->addr, devad, regnum, + val); +} +static inline int mdiodev_c45_modify (struct mdio_device *mdiodev,int devad,u32 regnum,u16 mask,u16 set) {
Signature · 1 line
2.52.0
Daniel Golle
maintainerPATCH v13 4/4 net: dsa: add basic initial driver for MxL862xx switches
In reply to Daniel GolleAdd very basic DSA driver for MaxLinear's MxL862xx switches.
In contrast to previous MaxLinear switches the MxL862xx has a built-in processor that runs a sophisticated firmware based on Zephyr RTOS. Interaction between the host and the switch hence is organized using a software API of that firmware rather than accessing hardware registers directly.
Add descriptions of the most basic firmware API calls to access the built-in MDIO bus hosting the 2.5GE PHYs, basic port control as well as setting up the CPU port.
Implement a very basic DSA driver using that API which is sufficient to get packets flowing between the user ports and the CPU port.
The firmware offers all features one would expect from a modern switch hardware, they are going to be added one by one in follow-up patch series.
---
v13:
* use '#pragma pack(push, 1)' in mxl862xx-api.h
* relax timeout in mxl862xx_busy_wait() from 10000us to 500000us, this turned out to be needed when using the driver on older firmware
* reorder functions
* remove unneeded checks for dp or cpu_do being !=NULL
v12:
* list switch variants in Kconfig starting with dash ('-')
* remove the __packed attribute on structs which are naturally packed according to C rules (ie. one other struct with only a single byte-aligned member)
* log error in mxl862xx_port_disable()
* introduce !NULL check for return value of dsa_to_port() in mxl862xx_add_single_port_bridge
* check cpu_dp being non-NULL before dereferencing dp->cpu_dp
* use non-racy and deterministic name for MII bus
* skip ports without cpu_dp assigned in mxl862xx_setup_cpu_bridge() to avoid potential NULL-pointer dereference
* call dev_set_drvdata() only after dsa_register_switch() has been successfully completed
v11:
* fix fatal last-second changes dsa_port_is_unused vs. dsa_is_unused_port and dsa_port_is_unused vs. dsa_is_unused_port
* use cpu_to_le16 in mxl862xx_configure_ctp_port
* improve formatting of some kerneldoc comments
v10:
* fix indentation in kernel-doc
* remove misleading comment regarding sub-interface IDs
* remove unused macros, set MXL862XX_MAX_PORTS to 17
* count ports starting from 0, 0 being the microcontroller
* improve port setup:
- split tag protocol and CTP port assignment
- use ports cpu_dp instead of relying on single (ie. first) CPU port
- handle CPU port like any other port
- prepare forward DT compatibility with multiple CPU ports
v9:
* remove practically unused struct hw_info
* lots of kerneldoc improvements in mxl862xx-api.h
* drop .mac_select_pcs() stub
* better handling for firmware error return value
* apply reverse xmas tree in mxl862xx_api_wrap
* guard headers with #ifdef macro
* include net/dsa.h and linux/mdio.h in mxl862xx.h
* call mxl862xx_port_fast_age() only once in .port_setup
* don't create isolation bridges for unused ports
* replace errornous cast with correct range of values denoting firmware errors
v8:
* use le32 endian in bridge_port_config API
* remove duplciate assignment of br_port_cfg.bridge_port_id when setting up CPU port
v7:
* fix kerneldoc style
v6:
* include bridge and bridgeport API needed to isolate ports
* remove warning in .setup as ports are now isolated
* make ready-after-reset check more robust by adding delay
* sort structs in order of struct definitions
* best effort to sort functions without introducing additional prototypes
* always use enums with kerneldoc comments in mxl862xx-api.h
* remove bogus .phy_read and .phy_write DSA ops as the driver anyway registers a user MDIO bus with Clause-22 and Clause-45 operations
* various small style fixes
v5:
* output warning in .setup regarding unknown pre-configuration
* add comment explaining why CFGGET is used in reset function
RFC v4:
* poll switch readiness after reset
* implement driver shutdown
* added port_fast_aging API call and driver op
* unified port setup in new .port_setup op
* improve comment explaining special handlign for unaligned API read
* various typos
RFC v3:
* fix return value being uninitialized on error in mxl862xx_api_wrap()
* add missing descrition in kerneldoc comment of
struct mxl862xx_ss_sp_tag
RFC v2:
* make use of struct mdio_device
* add phylink_mac_ops stubs
* drop leftover nonsense from mxl862xx_phylink_get_caps()
* use __le32 instead of enum types in over-the-wire structs
* use existing MDIO_* macros whenever possible
* simplify API constants to be more readable
* use readx_poll_timeout instead of open-coding poll timeout loop
* add mxl862xx_reg_read() and mxl862xx_reg_write() helpers
* demystify error codes returned by the firmware
* add #defines for mxl862xx_ss_sp_tag member values
* move reset to dedicated function, clarify magic number being the reset command ID
--- MAINTAINERS | 1 + drivers/net/dsa/Kconfig | 2 + drivers/net/dsa/Makefile | 1 + drivers/net/dsa/mxl862xx/Kconfig | 12 + drivers/net/dsa/mxl862xx/Makefile | 3 + drivers/net/dsa/mxl862xx/mxl862xx-api.h | 680 +++++++++++++++++++++++ drivers/net/dsa/mxl862xx/mxl862xx-cmd.h | 49 ++ drivers/net/dsa/mxl862xx/mxl862xx-host.c | 245 ++++++++ drivers/net/dsa/mxl862xx/mxl862xx-host.h | 12 + drivers/net/dsa/mxl862xx/mxl862xx.c | 476 ++++++++++++++++ drivers/net/dsa/mxl862xx/mxl862xx.h | 16 + 11 files changed, 1497 insertions(+) create mode 100644 drivers/net/dsa/mxl862xx/Kconfig create mode 100644 drivers/net/dsa/mxl862xx/Makefile create mode 100644 drivers/net/dsa/mxl862xx/mxl862xx-api.h create mode 100644 drivers/net/dsa/mxl862xx/mxl862xx-cmd.h create mode 100644 drivers/net/dsa/mxl862xx/mxl862xx-host.c create mode 100644 drivers/net/dsa/mxl862xx/mxl862xx-host.h create mode 100644 drivers/net/dsa/mxl862xx/mxl862xx.c create mode 100644 drivers/net/dsa/mxl862xx/mxl862xx.h
MAINTAINERS +1 -0
@@ -15624,6 +15624,7 @@ M: Daniel Golle <[email protected]> L: [email protected] S: Maintained F: Documentation/devicetree/bindings/net/dsa/maxlinear,mxl862xx.yaml +F: drivers/net/dsa/mxl862xx/ F: net/dsa/tag_mxl862xx.c MCAN DEVICE DRIVER
drivers/net/dsa/Kconfig +2 -0
@@ -74,6 +74,8 @@ source "drivers/net/dsa/microchip/Kconfig" source "drivers/net/dsa/mv88e6xxx/Kconfig" +source "drivers/net/dsa/mxl862xx/Kconfig" + source "drivers/net/dsa/ocelot/Kconfig" source "drivers/net/dsa/qca/Kconfig"
drivers/net/dsa/Makefile +1 -0
@@ -20,6 +20,7 @@ obj-y += hirschmann/ obj-y += lantiq/ obj-y += microchip/ obj-y += mv88e6xxx/ +obj-y += mxl862xx/ obj-y += ocelot/ obj-y += qca/ obj-y += realtek/
drivers/net/dsa/mxl862xx/Kconfig new file +12 -0
@@ -0,0 +1,12 @@ +# SPDX-License-Identifier: GPL-2.0-only +config NET_DSA_MXL862 + tristate "MaxLinear MxL862xx" + depends on NET_DSA + select MAXLINEAR_GPHY + select NET_DSA_TAG_MXL_862XX + help + This enables support for the MaxLinear MxL862xx switch family. + These switches have two 10GE SerDes interfaces, one typically + used as CPU port. + - MxL86282 has eight 2.5 Gigabit PHYs + - MxL86252 has five 2.5 Gigabit PHYs
drivers/net/dsa/mxl862xx/Makefile new file +3 -0
@@ -0,0 +1,3 @@ +# SPDX-License-Identifier: GPL-2.0 +obj-$(CONFIG_NET_DSA_MXL862) += mxl862xx_dsa.o +mxl862xx_dsa-y := mxl862xx.o mxl862xx-host.o
drivers/net/dsa/mxl862xx/mxl862xx-api.h new file +680 -0
680 lines not on this page · raw message
drivers/net/dsa/mxl862xx/mxl862xx-cmd.h new file +49 -0
49 lines not on this page · raw message
drivers/net/dsa/mxl862xx/mxl862xx-host.c new file +245 -0
245 lines not on this page · raw message
drivers/net/dsa/mxl862xx/mxl862xx-host.h new file +12 -0
@@ -0,0 +1,12 @@ +/* SPDX-License-Identifier: GPL-2.0-or-later */ + +#ifndef __MXL862XX_HOST_H +#define __MXL862XX_HOST_H + +#include "mxl862xx.h" + +int mxl862xx_api_wrap (struct mxl862xx_priv *priv,u16 cmd,void *data,u16 size, +bool read,bool quiet); +int mxl862xx_reset (struct mxl862xx_priv *priv); + +#endif /* __MXL862XX_HOST_H */
drivers/net/dsa/mxl862xx/mxl862xx.c new file +476 -0
476 lines not on this page · raw message
drivers/net/dsa/mxl862xx/mxl862xx.h new file +16 -0
@@ -0,0 +1,16 @@ +/* SPDX-License-Identifier: GPL-2.0-or-later */ + +#ifndef __MXL862XX_H +#define __MXL862XX_H + +#include <linux/mdio.h> +#include <net/dsa.h> + +#define MXL862XX_MAX_PORTS 17 + +struct mxl862xx_priv { +struct dsa_switch *ds; +struct mdio_device *mdiodev; +}; + +#endif /* __MXL862XX_H */
Signature · 1 line
2.52.0
Jakub Kicinski
maintainerRe: [PATCH net-next v13 4/4] net: dsa: add basic initial driver for MxL862xx switches
drivers/net/dsa/mxl862xx/mxl862xx-api.h line 9
Quoted 2 lines · On Wed, 4 Feb 2026 13:33:19 +0000 Daniel Golle wrote:
drivers/net/dsa/mxl862xx/mxl862xx-api.h new file +2
-0+/* The switch firmware expects all structs to be byte-aligned */ +#pragma pack(push, 1)
"Byte-aligned" means..? Generally aligned means that it starts at an address which is multiple of X. All addresses are multiple of 1
We used you push back against blanket __packed because it's forcing all *host* accesses to also assume that the structures are unaligned. The best practice is to pack only specific structs which need it and add compile_assert()s to make sure that the compiler doesn't add any padding.
There's a couple of AI nitpicks, since I'm already complaining I'll send these out too..
Daniel Golle
maintainerRe: [PATCH net-next v13 4/4] net: dsa: add basic initial driver for MxL862xx switches
Hi Jakub,
thank you for looking into this driver another time.
Quoted 6 lines · Jakub Kicinski
drivers/net/dsa/mxl862xx/mxl862xx-api.h line 9
Quoted 2 lines · On Wed, 4 Feb 2026 13:33:19 +0000 Daniel Golle wrote:
drivers/net/dsa/mxl862xx/mxl862xx-api.h new file +2
-0+/* The switch firmware expects all structs to be byte-aligned */ +#pragma pack(push, 1)"Byte-aligned" means..? Generally aligned means that it starts at an address which is multiple of X. All addresses are multiple of 1
In case fo the firmware running on this switch it means that data types used in structures used as input and output parameters for firmware functions should be aligned to 8 bits, without any additional padding in between.
struct foo {
u8 var1;
__le16 var2;
__le32 var3;
} __packed;It's size is 7 bytes and it looks like this:
.||||||||.||||||||.||||||||.||||||||.||||||||.||||||||.||||||||. | var1 | var2 | var3 |
. . LSB . MSB . LSB MSB .
This is what the firmware on the other end expects, from all data sent to it and what the Linux host has to expect from all data received from it.
Quoted 2 lines
We used you push back against blanket
__packedbecause it's forcing all *host* accesses to also assume that the structures are unaligned.
Understand that in general, and of course know that using packed structs without a hardware requirement of doing so needlessly wastes CPU cycles on each access to struct members.
However, in this case this is a header file which exclusively defines structs which are only used to communicate with the firmware running on the switch. Using them for anything else, such as storing or processing data the driver deals with internally is, very inconvenient because all types are also defined as little-endian, so not only unaligned access, but also endian conversion burdens every access (in the sense that it burdens the programmer on little-endian machines, but the CPU as well on big-endian machines).
tl;dr: This whole file is only API definition. And this is how the firmware API is defined, and that's the only way to deal with that switch.
(I would have preferred if they just exposed the internal 16-bit registers of the switch via MDIO, and have asked MxL for that several times, without success)
Quoted 3 lines
The best practice is to pack only specific structs which need it and add compile_assert()s to make sure that the compiler doesn't add any padding.
Imho checking whether each of these structs is naturally packed (ie. 8-bit aligned without padding between 8-bit aligned members) is prone to human errors which only become visible when testing on the real hardware, and hence complicates maintainance.
Other drivers which operate on similar APIs (many GPU drivers, for example) also use #pragma pack(push, 1) in header files defining external API. Also there all external API definitions are kept in a separate file, away from any of the datastructures used by the driver internally at runtime.
Anyway, if you really really want me to set individual __packed for each struct which isn't naturally packed in this whole file, please tell me clearly that this is what you would like, and I will of course do it despite disagreeing with the reasoning.
Cheers
Daniel
Vladimir Oltean
maintainerRe: [PATCH net-next v13 4/4] net: dsa: add basic initial driver for MxL862xx switches
Quoted 73 lines · Daniel Golle
Hi Jakub,
thank you for looking into this driver another time.
Quoted 6 lines · Jakub Kicinski
drivers/net/dsa/mxl862xx/mxl862xx-api.h line 9
Quoted 2 lines · On Wed, 4 Feb 2026 13:33:19 +0000 Daniel Golle wrote:
drivers/net/dsa/mxl862xx/mxl862xx-api.h new file +2
-0+/* The switch firmware expects all structs to be byte-aligned */ +#pragma pack(push, 1)"Byte-aligned" means..? Generally aligned means that it starts at an address which is multiple of X. All addresses are multiple of 1
In case fo the firmware running on this switch it means that data types used in structures used as input and output parameters for firmware functions should be aligned to 8 bits, without any additional padding in between.
struct foo { u8 var1; __le16 var2; __le32 var3; } __packed;It's size is 7 bytes and it looks like this:
.||||||||.||||||||.||||||||.||||||||.||||||||.||||||||.||||||||. | var1 | var2 | var3 |. . LSB . MSB . LSB MSB .
This is what the firmware on the other end expects, from all data sent to it and what the Linux host has to expect from all data received from it.
Quoted 2 lines
We used you push back against blanket
__packedbecause it's forcing all *host* accesses to also assume that the structures are unaligned.Understand that in general, and of course know that using packed structs without a hardware requirement of doing so needlessly wastes CPU cycles on each access to struct members.
However, in this case this is a header file which exclusively defines structs which are only used to communicate with the firmware running on the switch. Using them for anything else, such as storing or processing data the driver deals with internally is, very inconvenient because all types are also defined as little-endian, so not only unaligned access, but also endian conversion burdens every access (in the sense that it burdens the programmer on little-endian machines, but the CPU as well on big-endian machines).
tl;dr: This whole file is only API definition. And this is how the firmware API is defined, and that's the only way to deal with that switch.
(I would have preferred if they just exposed the internal 16-bit registers of the switch via MDIO, and have asked MxL for that several times, without success)
Quoted 3 lines
The best practice is to pack only specific structs which need it and add compile_assert()s to make sure that the compiler doesn't add any padding.
Imho checking whether each of these structs is naturally packed (ie. 8-bit aligned without padding between 8-bit aligned members) is prone to human errors which only become visible when testing on the real hardware, and hence complicates maintainance.
Other drivers which operate on similar APIs (many GPU drivers, for example) also use #pragma
pack(push, 1)in header files defining external API. Also there all external API definitions are kept in a separate file, away from any of the datastructures used by the driver internally at runtime.
Quoted 4 lines
Anyway, if you really really want me to set individual
__packedfor each struct which isn't naturally packed in this whole file, please tell me clearly that this is what you would like, and I will of course do it despite disagreeing with the reasoning.
When I created the pack_fields() API it was exactly for situations like this. It helps you keep naturally aligned structures in native CPU endianness while adapting to whatever quirks the peripheral you're talking to has.
Daniel Golle
maintainerRe: [PATCH net-next v13 4/4] net: dsa: add basic initial driver for MxL862xx switches
Quoted 79 lines · Vladimir Oltean
Quoted 73 lines · Daniel Golle
Hi Jakub,
thank you for looking into this driver another time.
Quoted 6 lines · Jakub Kicinski
drivers/net/dsa/mxl862xx/mxl862xx-api.h line 9
Quoted 2 lines · On Wed, 4 Feb 2026 13:33:19 +0000 Daniel Golle wrote:
drivers/net/dsa/mxl862xx/mxl862xx-api.h new file +2
-0+/* The switch firmware expects all structs to be byte-aligned */ +#pragma pack(push, 1)"Byte-aligned" means..? Generally aligned means that it starts at an address which is multiple of X. All addresses are multiple of 1
In case fo the firmware running on this switch it means that data types used in structures used as input and output parameters for firmware functions should be aligned to 8 bits, without any additional padding in between.
struct foo { u8 var1; __le16 var2; __le32 var3; } __packed;It's size is 7 bytes and it looks like this:
.||||||||.||||||||.||||||||.||||||||.||||||||.||||||||.||||||||. | var1 | var2 | var3 |. . LSB . MSB . LSB MSB .
This is what the firmware on the other end expects, from all data sent to it and what the Linux host has to expect from all data received from it.
Quoted 2 lines
We used you push back against blanket
__packedbecause it's forcing all *host* accesses to also assume that the structures are unaligned.Understand that in general, and of course know that using packed structs without a hardware requirement of doing so needlessly wastes CPU cycles on each access to struct members.
However, in this case this is a header file which exclusively defines structs which are only used to communicate with the firmware running on the switch. Using them for anything else, such as storing or processing data the driver deals with internally is, very inconvenient because all types are also defined as little-endian, so not only unaligned access, but also endian conversion burdens every access (in the sense that it burdens the programmer on little-endian machines, but the CPU as well on big-endian machines).
tl;dr: This whole file is only API definition. And this is how the firmware API is defined, and that's the only way to deal with that switch.
(I would have preferred if they just exposed the internal 16-bit registers of the switch via MDIO, and have asked MxL for that several times, without success)
Quoted 3 lines
The best practice is to pack only specific structs which need it and add compile_assert()s to make sure that the compiler doesn't add any padding.
Imho checking whether each of these structs is naturally packed (ie. 8-bit aligned without padding between 8-bit aligned members) is prone to human errors which only become visible when testing on the real hardware, and hence complicates maintainance.
Other drivers which operate on similar APIs (many GPU drivers, for example) also use #pragma
pack(push, 1)in header files defining external API. Also there all external API definitions are kept in a separate file, away from any of the datastructures used by the driver internally at runtime.Quoted 4 lines
Anyway, if you really really want me to set individual
__packedfor each struct which isn't naturally packed in this whole file, please tell me clearly that this is what you would like, and I will of course do it despite disagreeing with the reasoning.
Quoted 4 lines
When I created the
pack_fields()API it was exactly for situations like this. It helps you keep naturally aligned structures in native CPU endianness while adapting to whatever quirks the peripheral you're talking to has.
I've spent an hour studying the pack_fields() API and it's (well written) documentation. The only example of it's use in the current kernel I could find is the Intel E800 (ICE) driver. And there it does make sense as it is handling conversion between CPU and hardware formats in the hotpath for DMA descriptors, a total of 3 different structs, each with their individual accessor functions.
Using this approach for this switch driver would require writing a lot of boilerplate code, accessor functions for each and every struct, and a struct definition once unpacked for the host platform and then again using the PACKED_FIELD(...) notation for the hardware format. Surely, most of that could be auto-generated using the existing vendor drivers API definition. Yet (at least to me) it feels like over-engineering and also it would require rewriting most of the driver which has been discussed for almost 2 months now.
Also note that the driver doesn't need the naturally aligned version of all these structs in native CPU endian -- they are not used for further processing anything, you can see that because they aren't ever used as a function parameters, but only ever as exchange formats when communicating with the firmware.
Maybe I'm missing something obvious here and there is a more simple way to use this API, some generic macros using compiler introspection to magically handle everything without needing to write packed and unpacked struct definitions and individual pack/unpack boiler-plate functions for each struct. If so, please provide me with an example or explain how you imagine the pack_fields() API to be used in the context of this driver and it's total of at more than 30 different structs which will be used for all the different firmware function I will need to use in order to implement phylink_pcs as well as the various offloading and VLAN-related functionality the driver should have in the end (ie. the structs you currently see in the mxl862xx-api.h file are just a fraction of what I hope to add there by follow-up series)
Vladimir Oltean
maintainerRe: [PATCH net-next v13 4/4] net: dsa: add basic initial driver for MxL862xx switches
Quoted 16 lines · Daniel Golle
I've spent an hour studying the
pack_fields()API and it's (well written) documentation. The only example of it's use in the current kernel I could find is the Intel E800 (ICE) driver. And there it does make sense as it is handling conversion between CPU and hardware formats in the hotpath for DMA descriptors, a total of 3 different structs, each with their individual accessor functions.Using this approach for this switch driver would require writing a lot of boilerplate code, accessor functions for each and every struct, and a struct definition once unpacked for the host platform and then again using the
PACKED_FIELD(...)notation for the hardware format. Surely, most of that could be auto-generated using the existing vendor drivers API definition. Yet (at least to me) it feels like over-engineering and also it would require rewriting most of the driver which has been discussed for almost 2 months now.
Quoted 5 lines
Also note that the driver doesn't need the naturally aligned version of all these structs in native CPU endian -- they are not used for further processing anything, you can see that because they aren't ever used as a function parameters, but only ever as exchange formats when communicating with the firmware.
OK. If the fields were packed more densely maybe the tradeoff would have looked differently then. But you're talking to an MCU and not to hardware. And you don't need to keep a local direct representation of the data passed through those packed buffers. Your arguments are valid.
Quoted 4 lines
Maybe I'm missing something obvious here and there is a more simple way to use this API, some generic macros using compiler introspection to magically handle everything without needing to write packed and unpacked struct definitions and individual pack/unpack boiler-plate functions for
Quoted 8 lines
each struct. If so, please provide me with an example or explain how you imagine the
pack_fields()API to be used in the context of this driver and it's total of at more than 30 different structs which will be used for all the different firmware function I will need to use in order to implementphylink_pcsas well as the various offloading and VLAN-related functionality the driver should have in the end (ie. the structs you currently see in themxl862xx-api.hfile are just a fraction of what I hope to add there by follow-up series)
No, the API usage example is how you imagine it. There's a structure where you only need to pull in the fields you care about (and in whatever order), rather than every other unrelated tidbit you don't currently need and possibly never will (like ingress_marking_mode, etc etc). And another array of PACKED_FIELD() where you say where each field goes. You probably don't need a pack_fields() and an unpack_fields() call for the same data structure in most cases, just one or the other.
Jakub Kicinski
maintainerRe: [PATCH net-next v13 4/4] net: dsa: add basic initial driver for MxL862xx switches
In reply to Daniel GolleQuoted 5 lines · On Fri, 6 Feb 2026 03:14:26 +0000 Daniel Golle wrote:
Other drivers which operate on similar APIs (many GPU drivers, for example) also use #pragma
pack(push, 1)in header files defining external API. Also there all external API definitions are kept in a separate file, away from any of the datastructures used by the driver internally at runtime.
No pragma pack in networking drivers please. However, well considered your application of it is the next guy will sprinkle it randomly and argue that it's already used in other places.
GPU folks are special, they'd probably write their code in C++ if they could?