RTL83xx devices have two types of receive interrupts for each of its
8 rings. One for packet received and another for ring overflow. When
the switch is flooded with incoming packets the receive handler will
disable the packet receive notification but still keeps the overflow
notification enabled. While the receive path "slowly" processes the
received packets each new packet triggers the overflow IRQ again. The
device becomes unresponsive and eventually produces messages like:
[18441.709764] rcu: Stack dump where RCU GP kthread last ran:
[18441.727892] Sending NMI from CPU 1 to CPUs 0:
[18441.742300] NMI backtrace for cpu 0 skipped: idling at 0x8080e994
[18415.251700] rcu: INFO: rcu_sched detected stalls on CPUs/tasks:
[18415.271350] rcu: 0-...!: (0 ticks this GP) idle=d740/0/0x0 ...
[18415.303046] rcu: (detected by 1, t=6004 jiffies, g=230925, ...
[18415.326095] Sending NMI from CPU 1 to CPUs 0:
[18415.340540] NMI backtrace for cpu 0
Fix this issue by always disabling receive and overflow interrupts at
the same time.
Test with hping3 --udp -p 5021 -d 1400 --flood 192.168.2.72
Before (3sec run):
[183260.324846] rtl838x-eth 1b00a300.ethernet eth0: RX buffer overrun: status 0x101, mask: 0x7ffeff
[183260.340524] rtl838x-eth 1b00a300.ethernet eth0: RX buffer overrun: status 0x1, mask: 0x7ffeff
[183260.345799] net_ratelimit: 489997 callbacks suppressed
After (3 sec run):
[ 373.981479] rtl838x-eth 1b00a300.ethernet eth0: rx ring overrun: status 0x101, mask: 0x7fffff
[ 374.031118] rtl838x-eth 1b00a300.ethernet eth0: rx ring overrun: status 0x101, mask: 0x7fffff
[ 377.919996] net_ratelimit: 34 callbacks suppressed
Signed-off-by: Markus Stockhausen <markus.stockhausen@gmx.de>
Link: https://github.com/openwrt/openwrt/pull/19365
Signed-off-by: Hauke Mehrtens <hauke@hauke-m.de>
The current implementation has several issues:
- it uses the hacky phy_port* macros
- it uses SoC dependent raw pages
- it disables/enables SoC dependent polling
Get rid of these dependencies and access the mdio bus the normal way.
Signed-off-by: Markus Stockhausen <markus.stockhausen@gmx.de>
Link: https://github.com/openwrt/openwrt/pull/19372
Signed-off-by: Hauke Mehrtens <hauke@hauke-m.de>
In rtl931x_led_init, the number of leds per port is not properly set. It
currently uses a hardcoded value of 1 which seems to be taken initially
from a specific device. This hardcoded value assumes any port always has
exactly two leds.
The RTL930x variant - rtl930x_led_init - does a better job at this. So
take it and use it for RTL931x too with the corresponding register.
While at it, rename the function to a proper naming scheme.
Signed-off-by: Jonas Jelonek <jelonek.jonas@gmail.com>
Link: https://github.com/openwrt/openwrt/pull/19241
Signed-off-by: Hauke Mehrtens <hauke@hauke-m.de>
The operating mode of a SerDes must be aligned with the attached PHY or
SFP module. That does not only require to change the protocol (e.g. SGMII,
10Gbase-R, ...) but also the speed (e.g. 1.25G). For this the SerDes must
be re-initialized properly.
- It must be taken into power down
- The PLL speed must be set
- Maybe the CMU (clock management unit) must be resetted
- The new mode must be set
- The state machine must be resetted
- The power must be reactivated
Until now this sequence is bugged. First the driver relies on a clean
setup from U-Boot (rtk network on) and second trying to to change mode
and PLL speeds does not work at all. And not to forget: Currently two
adjacent SerDes cannot drive SGMII/HSGMII at the same time. Fix this by
taking care about the right SerDes/PLL/CMU command init order.
P.S. This code is inspired by the work of Jan Hofmann, who tried to
enable parallel SGMII/HSGMII mode. The only missing bit was a proper CMU
reset sequence.
Signed-off-by: Jan Hoffmann <jan@3e8.eu>
Signed-off-by: Markus Stockhausen <markus.stockhausen@gmx.de>
Link: https://github.com/openwrt/openwrt/pull/19220
Signed-off-by: Robert Marko <robimarko@gmail.com>
During setup the mdio driver decides the polling mode of the 4 smi
busses depending on the DTS phy settings. This works as follows:
- set polling to c45 if at least one phy is ethernet-phy-ieee802.3-c45
- set polling to c22 if all phys are ethernet-phy-ieee802.3-c22
On RTL930x it is not possible to switch to c22 if uboot has set c45
before. Fix this by overwriting the bitfield properly. While we are
here:
- Sort variables according to kernel style (inverse christmas tree)
- Initialize fields properly with = { 0 }
- Use GENMASK() for better readability
- Make use of RTMDIO_MAX_SMI_BUS
Signed-off-by: Markus Stockhausen <markus.stockhausen@gmx.de>
Link: https://github.com/openwrt/openwrt/pull/19161
Signed-off-by: Robert Marko <robimarko@gmail.com>
The drivers for I2C bus and mux for RTL931x have an incorrectly defined
SDA0 pin number, causing an error with correct pin numbers specified in
the device tree.
Using the `show tech-support board` on the vendor firmware of a Netgear
MS510TXM shows the correct pin numbers but they don't work with the
drivers. So fix this.
Signed-off-by: Jonas Jelonek <jelonek.jonas@gmail.com>
Link: https://github.com/openwrt/openwrt/pull/19171
Signed-off-by: Hauke Mehrtens <hauke@hauke-m.de>
Some Realtek switches have been designed with I2C busses that share a
single SCL line. The clock line is used for 2 or more busses. This cannot
be used with the standard i2c-gpio driver that relies on distinct SDA
and SCL pairs.
Provide a derived i2c-gpio-shared driver that can be used instead. This
driver can handle up to 4 busses with only a single clock line.
Signed-off-by: Markus Stockhausen <markus.stockhausen@gmx.de>
Link: https://github.com/openwrt/openwrt/pull/18737
Signed-off-by: Robert Marko <robimarko@gmail.com>
Three different code paths for the same phy model. Now the bus
is prepared to handle c45 (mmd) read/writes correctly. Remove
the custom implementations and let generic kernel functions do
their best. To achieve this
- disable the PHY-mode EEE in rtl821x_config_init() as upstream does
- provide mmd read/write functions that avoid EEE via c45 over c22
While chaning the phy_driver functions sort them alphabetically.
Signed-off-by: Markus Stockhausen <markus.stockhausen@gmx.de>
Link: https://github.com/openwrt/openwrt/pull/18935
Signed-off-by: Robert Marko <robimarko@gmail.com>
rtl8214fc_media_is_fibre() will need to be run when bus lock is held.
Split the function into two versions.
Signed-off-by: Markus Stockhausen <markus.stockhausen@gmx.de>
Link: https://github.com/openwrt/openwrt/pull/18935
Signed-off-by: Robert Marko <robimarko@gmail.com>
As the mdio bus has been hardened and can now handle c45 requests
the DSA driver must honor that as well. For this
- add proper upstreamed bus read_c45 and write_c45 functions
- take over the disabled port mask from upstream bus
Signed-off-by: Markus Stockhausen <markus.stockhausen@gmx.de>
Link: https://github.com/openwrt/openwrt/pull/18935
Signed-off-by: Robert Marko <robimarko@gmail.com>
With the follow up EEE patches the mdio bus will run c22 and c45
accesses during initial scan. Especially when accessing addresses
beyond the CPU port phy requests might fail in a way that cannot
be handled gratefully. Do two things
- do not allow access to addresses starting from cpu port
- set the scan disable bitmask to ports starting from cpu port
Signed-off-by: Markus Stockhausen <markus.stockhausen@gmx.de>
Link: https://github.com/openwrt/openwrt/pull/18935
Signed-off-by: Robert Marko <robimarko@gmail.com>
If mmd and normal phy ops are issued the bus is lost because
of wrong park page settings. Force it to 0x1f as in GPL.
Signed-off-by: Markus Stockhausen <markus.stockhausen@gmx.de>
Link: https://github.com/openwrt/openwrt/pull/18935
Signed-off-by: Robert Marko <robimarko@gmail.com>
The new kernel has relocated the definition of struct platform_device.
Signed-off-by: Markus Stockhausen <markus.stockhausen@gmx.de>
Link: https://github.com/openwrt/openwrt/pull/18935
Signed-off-by: Robert Marko <robimarko@gmail.com>
The DSA driver uses set_mac_eee() for the outside API while
the interal helper is called port_eee_set(). Align that.
Additionally do not call the internal helpers directly by
the function names but use the register assignments.
Signed-off-by: Markus Stockhausen <markus.stockhausen@gmx.de>
Link: https://github.com/openwrt/openwrt/pull/18935
Signed-off-by: Robert Marko <robimarko@gmail.com>
Upstream will get rid of the get_mac_eee() function in the DSA
driver and replace it by a boolean alternative. While we fill a
lot of data here (because of EEE bugs in the Realtek phy layer)
other DSA drivers only return if EEE is available or not for a
port. To make the next kernel upgrade easier follow that design.
Signed-off-by: Markus Stockhausen <markus.stockhausen@gmx.de>
Link: https://github.com/openwrt/openwrt/pull/18935
Signed-off-by: Robert Marko <robimarko@gmail.com>
To avoid unneeded interrupts the R4K timer is deactivated during
secondary cpu initialization. This is currently done during
phase init_secondary(). With the upgrade to 6.12 the kernel runs
a primary/secondary cpu timer/counter synchronization to verify the
proper setup in synchronise_count_slave(). That runs at a later
point in time and expects the secondary counter to be fully
functional. Finding a deactivated counter results in the following
messages:
WARNING: CPU: 1 PID: 0 at arch/mips/kernel/sync-r4k.c:99 check_counter_warp+0x220/0x254
Warning: zero counter calibration delta: 0 [max: 6500000]
Counter synchronization [CPU#0 -> CPU#1]:
Measured 278760029 cycles counter warp between CPUs
Relocate the deactivation to smp_finsh() at the end of the cpu
startup sequence. Additionally polish the startup code and remove
all unneeded parts.
Signed-off-by: Markus Stockhausen <markus.stockhausen@gmx.de>
Link: https://github.com/openwrt/openwrt/pull/18935
Signed-off-by: Robert Marko <robimarko@gmail.com>
Fix minor compilation errors due to kernel changes.
Signed-off-by: Jonas Jelonek <jelonek.jonas@gmail.com>
Signed-off-by: Markus Stockhausen <markus.stockhausen@gmx.de>
Link: https://github.com/openwrt/openwrt/pull/18935
Signed-off-by: Robert Marko <robimarko@gmail.com>
The validate function no longer exists in phylink_mac_ops. Remove
it for the internal ethernet interface. Instead provide some
meaningful mac capabilities.
Signed-off-by: Markus Stockhausen <markus.stockhausen@gmx.de>
Link: https://github.com/openwrt/openwrt/pull/18935
Signed-off-by: Robert Marko <robimarko@gmail.com>
Not only the link but also the mac capabilities are needed here.
Additionally do some alphabetical sorting.
Signed-off-by: Markus Stockhausen <markus.stockhausen@gmx.de>
Link: https://github.com/openwrt/openwrt/pull/18935
Signed-off-by: Robert Marko <robimarko@gmail.com>
The old upstream notation has been changed to something not so racist.
Signed-off-by: Markus Stockhausen <markus.stockhausen@gmx.de>
Link: https://github.com/openwrt/openwrt/pull/18935
Signed-off-by: Robert Marko <robimarko@gmail.com>
EEE functions are now called with ethtool_keee instead of
ethtool_eee. Replace all occurrences. This will fix function
signature checks but still produces compilation errors due
to structure changes.
Signed-off-by: Markus Stockhausen <markus.stockhausen@gmx.de>
Link: https://github.com/openwrt/openwrt/pull/18935
Signed-off-by: Robert Marko <robimarko@gmail.com>
Thanks to Chris Packham this driver got upstreamed. Drop the
downstream files and patches.
Signed-off-by: Markus Stockhausen <markus.stockhausen@gmx.de>
Link: https://github.com/openwrt/openwrt/pull/18935
Signed-off-by: Robert Marko <robimarko@gmail.com>
Automatically generated commit.
Signed-off-by: Markus Stockhausen <markus.stockhausen@gmx.de>
Link: https://github.com/openwrt/openwrt/pull/18935
Signed-off-by: Robert Marko <robimarko@gmail.com>
