On 5/9/19 1:24 AM, Joe Hershberger wrote:

On Tue, May 7, 2019 at 5:15 PM Joe Hershberger joe.hershberger@ni.com wrote:

...

Hi Tom,

The following changes since commit 8d7f06bbbef16f172cd5e9c4923cdcebe16b8980:

I rebased on your master and built for BB Black. DHCP seems to work fine. MLO also now fits again.

Merge branch 'master' of git://git.denx.de/u-boot-sh (2019-05-07 09:38:00 -0400)

are available in the git repository at:

git://git.denx.de/u-boot-net.git master

for you to fetch changes up to 8d0c6858455e89b089222a08d55ff711681ca011:

net: phy: micrel: Find Micrel PHY node correctly (2019-05-07 14:51:55 -0500)

---

Carlo Caione (4): net: phy: Add generic helpers to access MMD PHY registers net: phy: ti: use generic helpers to access MMD registers cmd: mdio: Switch to generic helpers when accessing the registers net: phy: realtek: Introduce quirk to mark RXC not stoppable

James Byrne (2): net: phy: micrel: Use correct skew values on KSZ9021 net: phy: micrel: Find Micrel PHY node correctly

Murali Karicheri (2): ARM: k2g-gp-evm: update to rgmii pinmux configuration ARM: k2g-ice: Add pinmux support for rgmii interface

Pankaj Bansal (1): drivers: net: ldpaa_eth: fix resource leak

Siva Durga Prasad Paladugu (2): net: phy: Reloc next and prev pointers inside phy_drivers net: phy: Fix return value check phy_probe

Valentin-catalin Neacsu (1): net: phy: aquantia: Set only autoneg on in register 4.c441

Vladimir Oltean (6): net: phy: ar803x: Address packet drops at low traffic rate due to SmartEEE feature net: phy: ar803x: Make RGMII Tx delays actually configurable for AR8035 net: phy: ar803x: Use common functions for RGMII internal delays net: phy: ar803x: Clarify the configuration of the CLK_25M output pin net: phy: ar803x: Explicitly disable RGMII delays

Tom, this [1] is the patch that is breaking the evm. It doesn't affect BB Black because it uses an SMSC phy, where as this evm uses an AR8031/AR8033.

Is it possible the device tree [2] is wrong for the board? It lists 'phy-mode = "rgmii-txid";', so that means that with this patch the RX delay is now being disabled.

Any thoughts, Vladimir?

Thanks, -Joe

[1] b3224e0f7e - "net: phy: ar803x: Explicitly disable RGMII delays" [2] arch/arm/dts/am335x-evm.dts

...

   net: phy: ar803x: Clarify the intention of ar8021_config

arch/arm/dts/sama5d3xcm.dtsi | 32 +++--- arch/arm/dts/sama5d3xcm_cmp.dtsi | 32 +++--- arch/arm/dts/socfpga_arria5_socdk.dts | 4 +- arch/arm/dts/socfpga_cyclone5_is1.dts | 4 +- arch/arm/dts/socfpga_cyclone5_socdk.dts | 4 +- arch/arm/dts/socfpga_cyclone5_sockit.dts | 4 +- arch/arm/dts/socfpga_cyclone5_vining_fpga.dts | 4 +- board/ti/ks2_evm/mux-k2g.h | 36 +++---- cmd/mdio.c | 27 +++-- doc/device-tree-bindings/net/micrel-ksz90x1.txt | 27 +++++ drivers/net/ldpaa_eth/ldpaa_eth.c | 1 + drivers/net/phy/Kconfig | 41 ++++++++ drivers/net/phy/aquantia.c | 7 +- drivers/net/phy/atheros.c | 128 ++++++++++++++++------- drivers/net/phy/micrel_ksz90x1.c | 24 ++++- drivers/net/phy/phy.c | 21 +++- drivers/net/phy/realtek.c | 19 ++++ drivers/net/phy/ti.c | 130 +++++------------------- include/phy.h | 70 +++++++++++++ 19 files changed, 394 insertions(+), 221 deletions(-)

Thanks! -Joe _______________________________________________ U-Boot mailing list U-Boot@lists.denx.de https://lists.denx.de/listinfo/u-boot

Hi Joe, Tom,

It sounds like what Joe pointed to (my patch) has a high chance of causing link failure. If the board is relying on RX delays in the Atheros PHY to ensure correct RGMII timing budget, then for sure it was working before and now it is broken. In that case, it was working by mistake; the DT blob is broken and should be corrected. Sorry for the trouble this has caused.

Regards, -Vladimir

On 5/9/19 1:42 AM, Tom Rini wrote:

On Wed, May 08, 2019 at 10:40:57PM +0000, Vladimir Oltean wrote:

...

On 5/9/19 1:24 AM, Joe Hershberger wrote:

...

On Tue, May 7, 2019 at 5:15 PM Joe Hershberger joe.hershberger@ni.com wrote:

...

Hi Tom,

The following changes since commit 8d7f06bbbef16f172cd5e9c4923cdcebe16b8980:

I rebased on your master and built for BB Black. DHCP seems to work fine. MLO also now fits again.

Merge branch 'master' of git://git.denx.de/u-boot-sh (2019-05-07 09:38:00 -0400)

are available in the git repository at:

git://git.denx.de/u-boot-net.git master

for you to fetch changes up to 8d0c6858455e89b089222a08d55ff711681ca011:

net: phy: micrel: Find Micrel PHY node correctly (2019-05-07 14:51:55 -0500)

---

Carlo Caione (4): net: phy: Add generic helpers to access MMD PHY registers net: phy: ti: use generic helpers to access MMD registers cmd: mdio: Switch to generic helpers when accessing the registers net: phy: realtek: Introduce quirk to mark RXC not stoppable

James Byrne (2): net: phy: micrel: Use correct skew values on KSZ9021 net: phy: micrel: Find Micrel PHY node correctly

Murali Karicheri (2): ARM: k2g-gp-evm: update to rgmii pinmux configuration ARM: k2g-ice: Add pinmux support for rgmii interface

Pankaj Bansal (1): drivers: net: ldpaa_eth: fix resource leak

Siva Durga Prasad Paladugu (2): net: phy: Reloc next and prev pointers inside phy_drivers net: phy: Fix return value check phy_probe

Valentin-catalin Neacsu (1): net: phy: aquantia: Set only autoneg on in register 4.c441

Vladimir Oltean (6): net: phy: ar803x: Address packet drops at low traffic rate due to SmartEEE feature net: phy: ar803x: Make RGMII Tx delays actually configurable for AR8035 net: phy: ar803x: Use common functions for RGMII internal delays net: phy: ar803x: Clarify the configuration of the CLK_25M output pin net: phy: ar803x: Explicitly disable RGMII delays

Tom, this [1] is the patch that is breaking the evm. It doesn't affect BB Black because it uses an SMSC phy, where as this evm uses an AR8031/AR8033.

Is it possible the device tree [2] is wrong for the board? It lists 'phy-mode = "rgmii-txid";', so that means that with this patch the RX delay is now being disabled.

Any thoughts, Vladimir?

Thanks, -Joe

[1] b3224e0f7e - "net: phy: ar803x: Explicitly disable RGMII delays" [2] arch/arm/dts/am335x-evm.dts

...

    net: phy: ar803x: Clarify the intention of ar8021_config

arch/arm/dts/sama5d3xcm.dtsi | 32 +++--- arch/arm/dts/sama5d3xcm_cmp.dtsi | 32 +++--- arch/arm/dts/socfpga_arria5_socdk.dts | 4 +- arch/arm/dts/socfpga_cyclone5_is1.dts | 4 +- arch/arm/dts/socfpga_cyclone5_socdk.dts | 4 +- arch/arm/dts/socfpga_cyclone5_sockit.dts | 4 +- arch/arm/dts/socfpga_cyclone5_vining_fpga.dts | 4 +- board/ti/ks2_evm/mux-k2g.h | 36 +++---- cmd/mdio.c | 27 +++-- doc/device-tree-bindings/net/micrel-ksz90x1.txt | 27 +++++ drivers/net/ldpaa_eth/ldpaa_eth.c | 1 + drivers/net/phy/Kconfig | 41 ++++++++ drivers/net/phy/aquantia.c | 7 +- drivers/net/phy/atheros.c | 128 ++++++++++++++++------- drivers/net/phy/micrel_ksz90x1.c | 24 ++++- drivers/net/phy/phy.c | 21 +++- drivers/net/phy/realtek.c | 19 ++++ drivers/net/phy/ti.c | 130 +++++------------------- include/phy.h | 70 +++++++++++++ 19 files changed, 394 insertions(+), 221 deletions(-)

Thanks! -Joe _______________________________________________ U-Boot mailing list U-Boot@lists.denx.de https://lists.denx.de/listinfo/u-boot

Hi Joe, Tom,

It sounds like what Joe pointed to (my patch) has a high chance of causing link failure. If the board is relying on RX delays in the Atheros PHY to ensure correct RGMII timing budget, then for sure it was working before and now it is broken. In that case, it was working by mistake; the DT blob is broken and should be corrected. Sorry for the trouble this has caused.

How is this handled in the Linux kernel and/or why doesn't it fail there?

On the latest net-next, it's handled the same: disable everything and enable only what's specified in phy-mode. I didn't keep track of older versions.

-Vladimir

On 5/9/19 1:48 AM, Tom Rini wrote:

On Wed, May 08, 2019 at 10:45:50PM +0000, Vladimir Oltean wrote:

...

On 5/9/19 1:42 AM, Tom Rini wrote:

...

On Wed, May 08, 2019 at 10:40:57PM +0000, Vladimir Oltean wrote:

...

On 5/9/19 1:24 AM, Joe Hershberger wrote:

...

On Tue, May 7, 2019 at 5:15 PM Joe Hershberger joe.hershberger@ni.com wrote:

...

Hi Tom,

The following changes since commit 8d7f06bbbef16f172cd5e9c4923cdcebe16b8980:

I rebased on your master and built for BB Black. DHCP seems to work fine. MLO also now fits again.

 Merge branch 'master' of git://git.denx.de/u-boot-sh (2019-05-07 09:38:00 -0400)

are available in the git repository at:

 git://git.denx.de/u-boot-net.git master

for you to fetch changes up to 8d0c6858455e89b089222a08d55ff711681ca011:

 net: phy: micrel: Find Micrel PHY node correctly (2019-05-07 14:51:55 -0500)

---

Carlo Caione (4): net: phy: Add generic helpers to access MMD PHY registers net: phy: ti: use generic helpers to access MMD registers cmd: mdio: Switch to generic helpers when accessing the registers net: phy: realtek: Introduce quirk to mark RXC not stoppable

James Byrne (2): net: phy: micrel: Use correct skew values on KSZ9021 net: phy: micrel: Find Micrel PHY node correctly

Murali Karicheri (2): ARM: k2g-gp-evm: update to rgmii pinmux configuration ARM: k2g-ice: Add pinmux support for rgmii interface

Pankaj Bansal (1): drivers: net: ldpaa_eth: fix resource leak

Siva Durga Prasad Paladugu (2): net: phy: Reloc next and prev pointers inside phy_drivers net: phy: Fix return value check phy_probe

Valentin-catalin Neacsu (1): net: phy: aquantia: Set only autoneg on in register 4.c441

Vladimir Oltean (6): net: phy: ar803x: Address packet drops at low traffic rate due to SmartEEE feature net: phy: ar803x: Make RGMII Tx delays actually configurable for AR8035 net: phy: ar803x: Use common functions for RGMII internal delays net: phy: ar803x: Clarify the configuration of the CLK_25M output pin net: phy: ar803x: Explicitly disable RGMII delays

Tom, this [1] is the patch that is breaking the evm. It doesn't affect BB Black because it uses an SMSC phy, where as this evm uses an AR8031/AR8033.

Is it possible the device tree [2] is wrong for the board? It lists 'phy-mode = "rgmii-txid";', so that means that with this patch the RX delay is now being disabled.

Any thoughts, Vladimir?

Thanks, -Joe

[1] b3224e0f7e - "net: phy: ar803x: Explicitly disable RGMII delays" [2] arch/arm/dts/am335x-evm.dts

...

     net: phy: ar803x: Clarify the intention of ar8021_config

arch/arm/dts/sama5d3xcm.dtsi                    |  32 +++---
arch/arm/dts/sama5d3xcm_cmp.dtsi                |  32 +++---
arch/arm/dts/socfpga_arria5_socdk.dts           |   4 +-
arch/arm/dts/socfpga_cyclone5_is1.dts           |   4 +-
arch/arm/dts/socfpga_cyclone5_socdk.dts         |   4 +-
arch/arm/dts/socfpga_cyclone5_sockit.dts        |   4 +-
arch/arm/dts/socfpga_cyclone5_vining_fpga.dts   |   4 +-
board/ti/ks2_evm/mux-k2g.h                      |  36 +++----
cmd/mdio.c                                      |  27 +++--
doc/device-tree-bindings/net/micrel-ksz90x1.txt |  27 +++++
drivers/net/ldpaa_eth/ldpaa_eth.c               |   1 +
drivers/net/phy/Kconfig                         |  41 ++++++++
drivers/net/phy/aquantia.c                      |   7 +-
drivers/net/phy/atheros.c                       | 128 ++++++++++++++++-------
drivers/net/phy/micrel_ksz90x1.c                |  24 ++++-
drivers/net/phy/phy.c                           |  21 +++-
drivers/net/phy/realtek.c                       |  19 ++++
drivers/net/phy/ti.c                            | 130 +++++-------------------
include/phy.h                                   |  70 +++++++++++++
19 files changed, 394 insertions(+), 221 deletions(-)

Thanks! -Joe _______________________________________________ U-Boot mailing list U-Boot@lists.denx.de https://lists.denx.de/listinfo/u-boot

Hi Joe, Tom,

It sounds like what Joe pointed to (my patch) has a high chance of causing link failure. If the board is relying on RX delays in the Atheros PHY to ensure correct RGMII timing budget, then for sure it was working before and now it is broken. In that case, it was working by mistake; the DT blob is broken and should be corrected. Sorry for the trouble this has caused.

How is this handled in the Linux kernel and/or why doesn't it fail there?

On the latest net-next, it's handled the same: disable everything and enable only what's specified in phy-mode. I didn't keep track of older versions.

So you would expect the network to be broken on net-next with this DTS? Because that's not allowed...

Yes, if it's the same problem, the behavior should be the same. I see net-next is only rejecting the bad DT since 2019-02-21: https://patchwork.kernel.org/patch/10819279/ So it depends when you checked it last time.

-Vladimir

On 5/9/19 1:55 AM, Tom Rini wrote:

On Wed, May 08, 2019 at 10:52:28PM +0000, Vladimir Oltean wrote:

...

On 5/9/19 1:48 AM, Tom Rini wrote:

...

On Wed, May 08, 2019 at 10:45:50PM +0000, Vladimir Oltean wrote:

...

On 5/9/19 1:42 AM, Tom Rini wrote:

...

On Wed, May 08, 2019 at 10:40:57PM +0000, Vladimir Oltean wrote:

...

On 5/9/19 1:24 AM, Joe Hershberger wrote: > On Tue, May 7, 2019 at 5:15 PM Joe Hershberger joe.hershberger@ni.com wrote: >> >> Hi Tom, >> >> The following changes since commit 8d7f06bbbef16f172cd5e9c4923cdcebe16b8980: >> >> I rebased on your master and built for BB Black. DHCP seems to work fine. >> MLO also now fits again. >> >> Merge branch 'master' of git://git.denx.de/u-boot-sh (2019-05-07 09:38:00 -0400) >> >> are available in the git repository at: >> >> git://git.denx.de/u-boot-net.git master >> >> for you to fetch changes up to 8d0c6858455e89b089222a08d55ff711681ca011: >> >> net: phy: micrel: Find Micrel PHY node correctly (2019-05-07 14:51:55 -0500) >> >> ---------------------------------------------------------------- >> Carlo Caione (4): >> net: phy: Add generic helpers to access MMD PHY registers >> net: phy: ti: use generic helpers to access MMD registers >> cmd: mdio: Switch to generic helpers when accessing the registers >> net: phy: realtek: Introduce quirk to mark RXC not stoppable >> >> James Byrne (2): >> net: phy: micrel: Use correct skew values on KSZ9021 >> net: phy: micrel: Find Micrel PHY node correctly >> >> Murali Karicheri (2): >> ARM: k2g-gp-evm: update to rgmii pinmux configuration >> ARM: k2g-ice: Add pinmux support for rgmii interface >> >> Pankaj Bansal (1): >> drivers: net: ldpaa_eth: fix resource leak >> >> Siva Durga Prasad Paladugu (2): >> net: phy: Reloc next and prev pointers inside phy_drivers >> net: phy: Fix return value check phy_probe >> >> Valentin-catalin Neacsu (1): >> net: phy: aquantia: Set only autoneg on in register 4.c441 >> >> Vladimir Oltean (6): >> net: phy: ar803x: Address packet drops at low traffic rate due to SmartEEE feature >> net: phy: ar803x: Make RGMII Tx delays actually configurable for AR8035 >> net: phy: ar803x: Use common functions for RGMII internal delays >> net: phy: ar803x: Clarify the configuration of the CLK_25M output pin >> net: phy: ar803x: Explicitly disable RGMII delays > > Tom, this [1] is the patch that is breaking the evm. It doesn't affect > BB Black because it uses an SMSC phy, where as this evm uses an > AR8031/AR8033. > > Is it possible the device tree [2] is wrong for the board? It lists > 'phy-mode = "rgmii-txid";', so that means that with this patch the RX > delay is now being disabled. > > Any thoughts, Vladimir? > > Thanks, > -Joe > > [1] b3224e0f7e - "net: phy: ar803x: Explicitly disable RGMII delays" > [2] arch/arm/dts/am335x-evm.dts > >> net: phy: ar803x: Clarify the intention of ar8021_config >> >> arch/arm/dts/sama5d3xcm.dtsi | 32 +++--- >> arch/arm/dts/sama5d3xcm_cmp.dtsi | 32 +++--- >> arch/arm/dts/socfpga_arria5_socdk.dts | 4 +- >> arch/arm/dts/socfpga_cyclone5_is1.dts | 4 +- >> arch/arm/dts/socfpga_cyclone5_socdk.dts | 4 +- >> arch/arm/dts/socfpga_cyclone5_sockit.dts | 4 +- >> arch/arm/dts/socfpga_cyclone5_vining_fpga.dts | 4 +- >> board/ti/ks2_evm/mux-k2g.h | 36 +++---- >> cmd/mdio.c | 27 +++-- >> doc/device-tree-bindings/net/micrel-ksz90x1.txt | 27 +++++ >> drivers/net/ldpaa_eth/ldpaa_eth.c | 1 + >> drivers/net/phy/Kconfig | 41 ++++++++ >> drivers/net/phy/aquantia.c | 7 +- >> drivers/net/phy/atheros.c | 128 ++++++++++++++++------- >> drivers/net/phy/micrel_ksz90x1.c | 24 ++++- >> drivers/net/phy/phy.c | 21 +++- >> drivers/net/phy/realtek.c | 19 ++++ >> drivers/net/phy/ti.c | 130 +++++------------------- >> include/phy.h | 70 +++++++++++++ >> 19 files changed, 394 insertions(+), 221 deletions(-) >> >> Thanks! >> -Joe >> _______________________________________________ >> U-Boot mailing list >> U-Boot@lists.denx.de >> https://lists.denx.de/listinfo/u-boot >

Hi Joe, Tom,

It sounds like what Joe pointed to (my patch) has a high chance of causing link failure. If the board is relying on RX delays in the Atheros PHY to ensure correct RGMII timing budget, then for sure it was working before and now it is broken. In that case, it was working by mistake; the DT blob is broken and should be corrected. Sorry for the trouble this has caused.

How is this handled in the Linux kernel and/or why doesn't it fail there?

On the latest net-next, it's handled the same: disable everything and enable only what's specified in phy-mode. I didn't keep track of older versions.

So you would expect the network to be broken on net-next with this DTS? Because that's not allowed...

Yes, if it's the same problem, the behavior should be the same. I see net-next is only rejecting the bad DT since 2019-02-21: https://patchwork.kernel.org/patch/10819279/ So it depends when you checked it last time.

I'll see about testing eth in Linux with net-next tomorrow, but, uh, is there some work-around for the DTS so that ethernet works?

Work-around in Linux or in U-boot? U-boot can fix up the DT for Linux and replace "rgmii-txid" with "rgmii-id", whereas neither Linux nor U-boot can do anything for themselves except either load a correct DT blob, or not use (revert, not upgrade to, etc) the enforcing patch.

You cannot break existing DTs

RGMII delay breakages happen all the time. To play the devil's advocate, protecting against them should be done by observing the semantics described in the devicetree/bindings document. If you follow Linux net-next, a breakage for RTL8211F is planned out right as we speak.

But let's wait until we know for sure that this is what's causing the issues.

Regards, -Vladimir

On 09.05.2019 02:05, Vladimir Oltean wrote:

On 5/9/19 1:55 AM, Tom Rini wrote:

...

On Wed, May 08, 2019 at 10:52:28PM +0000, Vladimir Oltean wrote:

...

On 5/9/19 1:48 AM, Tom Rini wrote:

...

On Wed, May 08, 2019 at 10:45:50PM +0000, Vladimir Oltean wrote:

...

On 5/9/19 1:42 AM, Tom Rini wrote:

...

On Wed, May 08, 2019 at 10:40:57PM +0000, Vladimir Oltean wrote: > On 5/9/19 1:24 AM, Joe Hershberger wrote: >> On Tue, May 7, 2019 at 5:15 PM Joe Hershberger joe.hershberger@ni.com wrote: >>> >>> Hi Tom, >>> >>> The following changes since commit 8d7f06bbbef16f172cd5e9c4923cdcebe16b8980: >>> >>> I rebased on your master and built for BB Black. DHCP seems to work fine. >>> MLO also now fits again. >>> >>> Merge branch 'master' of git://git.denx.de/u-boot-sh (2019-05-07 09:38:00 -0400) >>> >>> are available in the git repository at: >>> >>> git://git.denx.de/u-boot-net.git master >>> >>> for you to fetch changes up to 8d0c6858455e89b089222a08d55ff711681ca011: >>> >>> net: phy: micrel: Find Micrel PHY node correctly (2019-05-07 14:51:55 -0500) >>> >>> ---------------------------------------------------------------- >>> Carlo Caione (4): >>> net: phy: Add generic helpers to access MMD PHY registers >>> net: phy: ti: use generic helpers to access MMD registers >>> cmd: mdio: Switch to generic helpers when accessing the registers >>> net: phy: realtek: Introduce quirk to mark RXC not stoppable >>> >>> James Byrne (2): >>> net: phy: micrel: Use correct skew values on KSZ9021 >>> net: phy: micrel: Find Micrel PHY node correctly >>> >>> Murali Karicheri (2): >>> ARM: k2g-gp-evm: update to rgmii pinmux configuration >>> ARM: k2g-ice: Add pinmux support for rgmii interface >>> >>> Pankaj Bansal (1): >>> drivers: net: ldpaa_eth: fix resource leak >>> >>> Siva Durga Prasad Paladugu (2): >>> net: phy: Reloc next and prev pointers inside phy_drivers >>> net: phy: Fix return value check phy_probe >>> >>> Valentin-catalin Neacsu (1): >>> net: phy: aquantia: Set only autoneg on in register 4.c441 >>> >>> Vladimir Oltean (6): >>> net: phy: ar803x: Address packet drops at low traffic rate due to SmartEEE feature >>> net: phy: ar803x: Make RGMII Tx delays actually configurable for AR8035 >>> net: phy: ar803x: Use common functions for RGMII internal delays >>> net: phy: ar803x: Clarify the configuration of the CLK_25M output pin >>> net: phy: ar803x: Explicitly disable RGMII delays >> >> Tom, this [1] is the patch that is breaking the evm. It doesn't affect >> BB Black because it uses an SMSC phy, where as this evm uses an >> AR8031/AR8033. >> >> Is it possible the device tree [2] is wrong for the board? It lists >> 'phy-mode = "rgmii-txid";', so that means that with this patch the RX >> delay is now being disabled. >> >> Any thoughts, Vladimir? >> >> Thanks, >> -Joe >> >> [1] b3224e0f7e - "net: phy: ar803x: Explicitly disable RGMII delays" >> [2] arch/arm/dts/am335x-evm.dts >> >>> net: phy: ar803x: Clarify the intention of ar8021_config >>> >>> arch/arm/dts/sama5d3xcm.dtsi | 32 +++--- >>> arch/arm/dts/sama5d3xcm_cmp.dtsi | 32 +++--- >>> arch/arm/dts/socfpga_arria5_socdk.dts | 4 +- >>> arch/arm/dts/socfpga_cyclone5_is1.dts | 4 +- >>> arch/arm/dts/socfpga_cyclone5_socdk.dts | 4 +- >>> arch/arm/dts/socfpga_cyclone5_sockit.dts | 4 +- >>> arch/arm/dts/socfpga_cyclone5_vining_fpga.dts | 4 +- >>> board/ti/ks2_evm/mux-k2g.h | 36 +++---- >>> cmd/mdio.c | 27 +++-- >>> doc/device-tree-bindings/net/micrel-ksz90x1.txt | 27 +++++ >>> drivers/net/ldpaa_eth/ldpaa_eth.c | 1 + >>> drivers/net/phy/Kconfig | 41 ++++++++ >>> drivers/net/phy/aquantia.c | 7 +- >>> drivers/net/phy/atheros.c | 128 ++++++++++++++++------- >>> drivers/net/phy/micrel_ksz90x1.c | 24 ++++- >>> drivers/net/phy/phy.c | 21 +++- >>> drivers/net/phy/realtek.c | 19 ++++ >>> drivers/net/phy/ti.c | 130 +++++------------------- >>> include/phy.h | 70 +++++++++++++ >>> 19 files changed, 394 insertions(+), 221 deletions(-) >>> >>> Thanks! >>> -Joe >>> _______________________________________________ >>> U-Boot mailing list >>> U-Boot@lists.denx.de >>> https://lists.denx.de/listinfo/u-boot >> > > Hi Joe, Tom, > > It sounds like what Joe pointed to (my patch) has a high chance of > causing link failure. > If the board is relying on RX delays in the Atheros PHY to ensure > correct RGMII timing budget, then for sure it was working before and now > it is broken. In that case, it was working by mistake; the DT blob is > broken and should be corrected. > Sorry for the trouble this has caused.

How is this handled in the Linux kernel and/or why doesn't it fail there?

On the latest net-next, it's handled the same: disable everything and enable only what's specified in phy-mode. I didn't keep track of older versions.

So you would expect the network to be broken on net-next with this DTS? Because that's not allowed...

Yes, if it's the same problem, the behavior should be the same. I see net-next is only rejecting the bad DT since 2019-02-21: https://patchwork.kernel.org/patch/10819279/ So it depends when you checked it last time.

I'll see about testing eth in Linux with net-next tomorrow, but, uh, is there some work-around for the DTS so that ethernet works?

Work-around in Linux or in U-boot? U-boot can fix up the DT for Linux and replace "rgmii-txid" with "rgmii-id", whereas neither Linux nor U-boot can do anything for themselves except either load a correct DT blob, or not use (revert, not upgrade to, etc) the enforcing patch.

...

You cannot break existing DTs

RGMII delay breakages happen all the time. To play the devil's advocate, protecting against them should be done by observing the semantics described in the devicetree/bindings document. If you follow Linux net-next, a breakage for RTL8211F is planned out right as we speak.

But let's wait until we know for sure that this is what's causing the issues.

Regards, -Vladimir

Hi Tom,

I understand now what you mean by "workaround". Yes you can enable RGMII delays by hand in U-boot:

Check if RX delay is enabled (bit 15 of debug register 0): => mdio write eTSEC1 0x1D 0 => mdio read eTSEC1 0x1E => mdio write eTSEC1 0x1E <new value>

Check if TX delay is enabled (bit 8 of debug register 5): => mdio write eTSEC1 0x1D 5 => mdio read eTSEC1 0x1E => mdio write eTSEC1 0x1E <new value>

Replace eTSEC1 with your net device from "mdio list".

Regards, -Vladimir

On Fri, May 10, 2019 at 09:50:45PM +0000, Joe Hershberger wrote:

Hi Vladimir and Tom,

On Thu, May 9, 2019 at 7:51 AM Vladimir Oltean vladimir.oltean@nxp.com wrote:

...

On 09.05.2019 02:05, Vladimir Oltean wrote:

...

On 5/9/19 1:55 AM, Tom Rini wrote:

...

On Wed, May 08, 2019 at 10:52:28PM +0000, Vladimir Oltean wrote:

...

On 5/9/19 1:48 AM, Tom Rini wrote:

...

On Wed, May 08, 2019 at 10:45:50PM +0000, Vladimir Oltean wrote: > On 5/9/19 1:42 AM, Tom Rini wrote: >> On Wed, May 08, 2019 at 10:40:57PM +0000, Vladimir Oltean wrote: >>> On 5/9/19 1:24 AM, Joe Hershberger wrote: >>>> On Tue, May 7, 2019 at 5:15 PM Joe Hershberger joe.hershberger@ni.com wrote: >>>>> >>>>> Hi Tom, >>>>> >>>>> The following changes since commit 8d7f06bbbef16f172cd5e9c4923cdcebe16b8980: >>>>> >>>>> I rebased on your master and built for BB Black. DHCP seems to work fine. >>>>> MLO also now fits again. >>>>> >>>>> Merge branch 'master' of git://git.denx.de/u-boot-sh (2019-05-07 09:38:00 -0400) >>>>> >>>>> are available in the git repository at: >>>>> >>>>> git://git.denx.de/u-boot-net.git master >>>>> >>>>> for you to fetch changes up to 8d0c6858455e89b089222a08d55ff711681ca011: >>>>> >>>>> net: phy: micrel: Find Micrel PHY node correctly (2019-05-07 14:51:55 -0500) >>>>> >>>>> ---------------------------------------------------------------- >>>>> Carlo Caione (4): >>>>> net: phy: Add generic helpers to access MMD PHY registers >>>>> net: phy: ti: use generic helpers to access MMD registers >>>>> cmd: mdio: Switch to generic helpers when accessing the registers >>>>> net: phy: realtek: Introduce quirk to mark RXC not stoppable >>>>> >>>>> James Byrne (2): >>>>> net: phy: micrel: Use correct skew values on KSZ9021 >>>>> net: phy: micrel: Find Micrel PHY node correctly >>>>> >>>>> Murali Karicheri (2): >>>>> ARM: k2g-gp-evm: update to rgmii pinmux configuration >>>>> ARM: k2g-ice: Add pinmux support for rgmii interface >>>>> >>>>> Pankaj Bansal (1): >>>>> drivers: net: ldpaa_eth: fix resource leak >>>>> >>>>> Siva Durga Prasad Paladugu (2): >>>>> net: phy: Reloc next and prev pointers inside phy_drivers >>>>> net: phy: Fix return value check phy_probe >>>>> >>>>> Valentin-catalin Neacsu (1): >>>>> net: phy: aquantia: Set only autoneg on in register 4.c441 >>>>> >>>>> Vladimir Oltean (6): >>>>> net: phy: ar803x: Address packet drops at low traffic rate due to SmartEEE feature >>>>> net: phy: ar803x: Make RGMII Tx delays actually configurable for AR8035 >>>>> net: phy: ar803x: Use common functions for RGMII internal delays >>>>> net: phy: ar803x: Clarify the configuration of the CLK_25M output pin >>>>> net: phy: ar803x: Explicitly disable RGMII delays >>>> >>>> Tom, this [1] is the patch that is breaking the evm. It doesn't affect >>>> BB Black because it uses an SMSC phy, where as this evm uses an >>>> AR8031/AR8033. >>>> >>>> Is it possible the device tree [2] is wrong for the board? It lists >>>> 'phy-mode = "rgmii-txid";', so that means that with this patch the RX >>>> delay is now being disabled. >>>> >>>> Any thoughts, Vladimir? >>>> >>>> Thanks, >>>> -Joe >>>> >>>> [1] b3224e0f7e - "net: phy: ar803x: Explicitly disable RGMII delays" >>>> [2] arch/arm/dts/am335x-evm.dts >>>> >>>>> net: phy: ar803x: Clarify the intention of ar8021_config >>>>> >>>>> arch/arm/dts/sama5d3xcm.dtsi | 32 +++--- >>>>> arch/arm/dts/sama5d3xcm_cmp.dtsi | 32 +++--- >>>>> arch/arm/dts/socfpga_arria5_socdk.dts | 4 +- >>>>> arch/arm/dts/socfpga_cyclone5_is1.dts | 4 +- >>>>> arch/arm/dts/socfpga_cyclone5_socdk.dts | 4 +- >>>>> arch/arm/dts/socfpga_cyclone5_sockit.dts | 4 +- >>>>> arch/arm/dts/socfpga_cyclone5_vining_fpga.dts | 4 +- >>>>> board/ti/ks2_evm/mux-k2g.h | 36 +++---- >>>>> cmd/mdio.c | 27 +++-- >>>>> doc/device-tree-bindings/net/micrel-ksz90x1.txt | 27 +++++ >>>>> drivers/net/ldpaa_eth/ldpaa_eth.c | 1 + >>>>> drivers/net/phy/Kconfig | 41 ++++++++ >>>>> drivers/net/phy/aquantia.c | 7 +- >>>>> drivers/net/phy/atheros.c | 128 ++++++++++++++++------- >>>>> drivers/net/phy/micrel_ksz90x1.c | 24 ++++- >>>>> drivers/net/phy/phy.c | 21 +++- >>>>> drivers/net/phy/realtek.c | 19 ++++ >>>>> drivers/net/phy/ti.c | 130 +++++------------------- >>>>> include/phy.h | 70 +++++++++++++ >>>>> 19 files changed, 394 insertions(+), 221 deletions(-) >>>>> >>>>> Thanks! >>>>> -Joe >>>>> _______________________________________________ >>>>> U-Boot mailing list >>>>> U-Boot@lists.denx.de >>>>> https://lists.denx.de/listinfo/u-boot >>>> >>> >>> Hi Joe, Tom, >>> >>> It sounds like what Joe pointed to (my patch) has a high chance of >>> causing link failure. >>> If the board is relying on RX delays in the Atheros PHY to ensure >>> correct RGMII timing budget, then for sure it was working before and now >>> it is broken. In that case, it was working by mistake; the DT blob is >>> broken and should be corrected. >>> Sorry for the trouble this has caused. >> >> How is this handled in the Linux kernel and/or why doesn't it fail >> there? > > On the latest net-next, it's handled the same: disable everything and > enable only what's specified in phy-mode. I didn't keep track of older > versions.

So you would expect the network to be broken on net-next with this DTS? Because that's not allowed...

Yes, if it's the same problem, the behavior should be the same. I see net-next is only rejecting the bad DT since 2019-02-21: https://patchwork.kernel.org/patch/10819279/ So it depends when you checked it last time.

I'll see about testing eth in Linux with net-next tomorrow, but, uh, is there some work-around for the DTS so that ethernet works?

Work-around in Linux or in U-boot? U-boot can fix up the DT for Linux and replace "rgmii-txid" with "rgmii-id", whereas neither Linux nor U-boot can do anything for themselves except either load a correct DT blob, or not use (revert, not upgrade to, etc) the enforcing patch.

...

You cannot break existing DTs

RGMII delay breakages happen all the time. To play the devil's advocate, protecting against them should be done by observing the semantics described in the devicetree/bindings document. If you follow Linux net-next, a breakage for RTL8211F is planned out right as we speak.

But let's wait until we know for sure that this is what's causing the issues.

Regards, -Vladimir

Hi Tom,

I understand now what you mean by "workaround". Yes you can enable RGMII delays by hand in U-boot:

Check if RX delay is enabled (bit 15 of debug register 0): => mdio write eTSEC1 0x1D 0 => mdio read eTSEC1 0x1E => mdio write eTSEC1 0x1E <new value>

Check if TX delay is enabled (bit 8 of debug register 5): => mdio write eTSEC1 0x1D 5 => mdio read eTSEC1 0x1E => mdio write eTSEC1 0x1E <new value>

Replace eTSEC1 with your net device from "mdio list".

I'm not sure what the follow-on action is here. I'll hold off on Vladimir's series until there is a resolution that everyone is satisfied with.

While I personally think this breaks the "DT is ABI" model, the required change has already been merged into Linus' tree so we just need to re-sync the am335x DTS files again and then we can apply these changes. I'll put it on my TODO list.

On Thu, Jul 25, 2019 at 10:40:45PM +0000, Vladimir Oltean wrote:

On 5/15/19 5:58 PM, Tom Rini wrote:

...

On Fri, May 10, 2019 at 09:50:45PM +0000, Joe Hershberger wrote:

...

Hi Vladimir and Tom,

On Thu, May 9, 2019 at 7:51 AM Vladimir Oltean vladimir.oltean@nxp.com wrote:

...

On 09.05.2019 02:05, Vladimir Oltean wrote:

...

On 5/9/19 1:55 AM, Tom Rini wrote:

...

On Wed, May 08, 2019 at 10:52:28PM +0000, Vladimir Oltean wrote: > On 5/9/19 1:48 AM, Tom Rini wrote: >> On Wed, May 08, 2019 at 10:45:50PM +0000, Vladimir Oltean wrote: >>> On 5/9/19 1:42 AM, Tom Rini wrote: >>>> On Wed, May 08, 2019 at 10:40:57PM +0000, Vladimir Oltean wrote: >>>>> On 5/9/19 1:24 AM, Joe Hershberger wrote: >>>>>> On Tue, May 7, 2019 at 5:15 PM Joe Hershberger joe.hershberger@ni.com wrote: >>>>>>> >>>>>>> Hi Tom, >>>>>>> >>>>>>> The following changes since commit 8d7f06bbbef16f172cd5e9c4923cdcebe16b8980: >>>>>>> >>>>>>> I rebased on your master and built for BB Black. DHCP seems to work fine. >>>>>>> MLO also now fits again. >>>>>>> >>>>>>> Merge branch 'master' of git://git.denx.de/u-boot-sh (2019-05-07 09:38:00 -0400) >>>>>>> >>>>>>> are available in the git repository at: >>>>>>> >>>>>>> git://git.denx.de/u-boot-net.git master >>>>>>> >>>>>>> for you to fetch changes up to 8d0c6858455e89b089222a08d55ff711681ca011: >>>>>>> >>>>>>> net: phy: micrel: Find Micrel PHY node correctly (2019-05-07 14:51:55 -0500) >>>>>>> >>>>>>> ---------------------------------------------------------------- >>>>>>> Carlo Caione (4): >>>>>>> net: phy: Add generic helpers to access MMD PHY registers >>>>>>> net: phy: ti: use generic helpers to access MMD registers >>>>>>> cmd: mdio: Switch to generic helpers when accessing the registers >>>>>>> net: phy: realtek: Introduce quirk to mark RXC not stoppable >>>>>>> >>>>>>> James Byrne (2): >>>>>>> net: phy: micrel: Use correct skew values on KSZ9021 >>>>>>> net: phy: micrel: Find Micrel PHY node correctly >>>>>>> >>>>>>> Murali Karicheri (2): >>>>>>> ARM: k2g-gp-evm: update to rgmii pinmux configuration >>>>>>> ARM: k2g-ice: Add pinmux support for rgmii interface >>>>>>> >>>>>>> Pankaj Bansal (1): >>>>>>> drivers: net: ldpaa_eth: fix resource leak >>>>>>> >>>>>>> Siva Durga Prasad Paladugu (2): >>>>>>> net: phy: Reloc next and prev pointers inside phy_drivers >>>>>>> net: phy: Fix return value check phy_probe >>>>>>> >>>>>>> Valentin-catalin Neacsu (1): >>>>>>> net: phy: aquantia: Set only autoneg on in register 4.c441 >>>>>>> >>>>>>> Vladimir Oltean (6): >>>>>>> net: phy: ar803x: Address packet drops at low traffic rate due to SmartEEE feature >>>>>>> net: phy: ar803x: Make RGMII Tx delays actually configurable for AR8035 >>>>>>> net: phy: ar803x: Use common functions for RGMII internal delays >>>>>>> net: phy: ar803x: Clarify the configuration of the CLK_25M output pin >>>>>>> net: phy: ar803x: Explicitly disable RGMII delays >>>>>> >>>>>> Tom, this [1] is the patch that is breaking the evm. It doesn't affect >>>>>> BB Black because it uses an SMSC phy, where as this evm uses an >>>>>> AR8031/AR8033. >>>>>> >>>>>> Is it possible the device tree [2] is wrong for the board? It lists >>>>>> 'phy-mode = "rgmii-txid";', so that means that with this patch the RX >>>>>> delay is now being disabled. >>>>>> >>>>>> Any thoughts, Vladimir? >>>>>> >>>>>> Thanks, >>>>>> -Joe >>>>>> >>>>>> [1] b3224e0f7e - "net: phy: ar803x: Explicitly disable RGMII delays" >>>>>> [2] arch/arm/dts/am335x-evm.dts >>>>>> >>>>>>> net: phy: ar803x: Clarify the intention of ar8021_config >>>>>>> >>>>>>> arch/arm/dts/sama5d3xcm.dtsi | 32 +++--- >>>>>>> arch/arm/dts/sama5d3xcm_cmp.dtsi | 32 +++--- >>>>>>> arch/arm/dts/socfpga_arria5_socdk.dts | 4 +- >>>>>>> arch/arm/dts/socfpga_cyclone5_is1.dts | 4 +- >>>>>>> arch/arm/dts/socfpga_cyclone5_socdk.dts | 4 +- >>>>>>> arch/arm/dts/socfpga_cyclone5_sockit.dts | 4 +- >>>>>>> arch/arm/dts/socfpga_cyclone5_vining_fpga.dts | 4 +- >>>>>>> board/ti/ks2_evm/mux-k2g.h | 36 +++---- >>>>>>> cmd/mdio.c | 27 +++-- >>>>>>> doc/device-tree-bindings/net/micrel-ksz90x1.txt | 27 +++++ >>>>>>> drivers/net/ldpaa_eth/ldpaa_eth.c | 1 + >>>>>>> drivers/net/phy/Kconfig | 41 ++++++++ >>>>>>> drivers/net/phy/aquantia.c | 7 +- >>>>>>> drivers/net/phy/atheros.c | 128 ++++++++++++++++------- >>>>>>> drivers/net/phy/micrel_ksz90x1.c | 24 ++++- >>>>>>> drivers/net/phy/phy.c | 21 +++- >>>>>>> drivers/net/phy/realtek.c | 19 ++++ >>>>>>> drivers/net/phy/ti.c | 130 +++++------------------- >>>>>>> include/phy.h | 70 +++++++++++++ >>>>>>> 19 files changed, 394 insertions(+), 221 deletions(-) >>>>>>> >>>>>>> Thanks! >>>>>>> -Joe >>>>>>> _______________________________________________ >>>>>>> U-Boot mailing list >>>>>>> U-Boot@lists.denx.de >>>>>>> https://lists.denx.de/listinfo/u-boot >>>>>> >>>>> >>>>> Hi Joe, Tom, >>>>> >>>>> It sounds like what Joe pointed to (my patch) has a high chance of >>>>> causing link failure. >>>>> If the board is relying on RX delays in the Atheros PHY to ensure >>>>> correct RGMII timing budget, then for sure it was working before and now >>>>> it is broken. In that case, it was working by mistake; the DT blob is >>>>> broken and should be corrected. >>>>> Sorry for the trouble this has caused. >>>> >>>> How is this handled in the Linux kernel and/or why doesn't it fail >>>> there? >>> >>> On the latest net-next, it's handled the same: disable everything and >>> enable only what's specified in phy-mode. I didn't keep track of older >>> versions. >> >> So you would expect the network to be broken on net-next with this DTS? >> Because that's not allowed... >> > > Yes, if it's the same problem, the behavior should be the same. > I see net-next is only rejecting the bad DT since 2019-02-21: > https://patchwork.kernel.org/patch/10819279/ > So it depends when you checked it last time.

I'll see about testing eth in Linux with net-next tomorrow, but, uh, is there some work-around for the DTS so that ethernet works?

Work-around in Linux or in U-boot? U-boot can fix up the DT for Linux and replace "rgmii-txid" with "rgmii-id", whereas neither Linux nor U-boot can do anything for themselves except either load a correct DT blob, or not use (revert, not upgrade to, etc) the enforcing patch.

...

You cannot break existing DTs

RGMII delay breakages happen all the time. To play the devil's advocate, protecting against them should be done by observing the semantics described in the devicetree/bindings document. If you follow Linux net-next, a breakage for RTL8211F is planned out right as we speak.

But let's wait until we know for sure that this is what's causing the issues.

Regards, -Vladimir

Hi Tom,

I understand now what you mean by "workaround". Yes you can enable RGMII delays by hand in U-boot:

Check if RX delay is enabled (bit 15 of debug register 0): => mdio write eTSEC1 0x1D 0 => mdio read eTSEC1 0x1E => mdio write eTSEC1 0x1E <new value>

Check if TX delay is enabled (bit 8 of debug register 5): => mdio write eTSEC1 0x1D 5 => mdio read eTSEC1 0x1E => mdio write eTSEC1 0x1E <new value>

Replace eTSEC1 with your net device from "mdio list".

I'm not sure what the follow-on action is here. I'll hold off on Vladimir's series until there is a resolution that everyone is satisfied with.

While I personally think this breaks the "DT is ABI" model, the required change has already been merged into Linus' tree so we just need to re-sync the am335x DTS files again and then we can apply these changes. I'll put it on my TODO list.

Hi Tom, Joe,

It's been a long while - what is the latest news on RGMII delays for the am335x + Atheros PHY? I'd like to know whether Joe can pick up these patches again. As you more-or-less admitted, it's not really ABI breakage if the DT semantics were never properly followed in the first place.

It's ABI breakage that no one cared about. At this point, it's blocked until the am335x DTS files get a resync from the kernel and that's somewhere on my TODO list.

On Sat, Nov 02, 2019 at 02:17:28PM +0100, Michael Walle wrote:

Am 2019-05-15 16:58, schrieb Tom Rini:

...

On Fri, May 10, 2019 at 09:50:45PM +0000, Joe Hershberger wrote:

...

Hi Vladimir and Tom,

On Thu, May 9, 2019 at 7:51 AM Vladimir Oltean vladimir.oltean@nxp.com wrote:

...

On 09.05.2019 02:05, Vladimir Oltean wrote:

...

On 5/9/19 1:55 AM, Tom Rini wrote:

...

On Wed, May 08, 2019 at 10:52:28PM +0000, Vladimir Oltean wrote: > On 5/9/19 1:48 AM, Tom Rini wrote: >> On Wed, May 08, 2019 at 10:45:50PM +0000, Vladimir Oltean wrote: >>> On 5/9/19 1:42 AM, Tom Rini wrote: >>>> On Wed, May 08, 2019 at 10:40:57PM +0000, Vladimir Oltean wrote: >>>>> On 5/9/19 1:24 AM, Joe Hershberger wrote: >>>>>> On Tue, May 7, 2019 at 5:15 PM Joe Hershberger joe.hershberger@ni.com wrote: >>>>>>> >>>>>>> Hi Tom, >>>>>>> >>>>>>> The following changes since commit 8d7f06bbbef16f172cd5e9c4923cdcebe16b8980: >>>>>>> >>>>>>> I rebased on your master and built for BB Black. DHCP seems to work fine. >>>>>>> MLO also now fits again. >>>>>>> >>>>>>> Merge branch 'master' of git://git.denx.de/u-boot-sh (2019-05-07 09:38:00 -0400) >>>>>>> >>>>>>> are available in the git repository at: >>>>>>> >>>>>>> git://git.denx.de/u-boot-net.git master >>>>>>> >>>>>>> for you to fetch changes up to 8d0c6858455e89b089222a08d55ff711681ca011: >>>>>>> >>>>>>> net: phy: micrel: Find Micrel PHY node correctly (2019-05-07 14:51:55 -0500) >>>>>>> >>>>>>> ---------------------------------------------------------------- >>>>>>> Carlo Caione (4): >>>>>>> net: phy: Add generic helpers to access MMD PHY registers >>>>>>> net: phy: ti: use generic helpers to access MMD registers >>>>>>> cmd: mdio: Switch to generic helpers when accessing the registers >>>>>>> net: phy: realtek: Introduce quirk to mark RXC not stoppable >>>>>>> >>>>>>> James Byrne (2): >>>>>>> net: phy: micrel: Use correct skew values on KSZ9021 >>>>>>> net: phy: micrel: Find Micrel PHY node correctly >>>>>>> >>>>>>> Murali Karicheri (2): >>>>>>> ARM: k2g-gp-evm: update to rgmii pinmux configuration >>>>>>> ARM: k2g-ice: Add pinmux support for rgmii interface >>>>>>> >>>>>>> Pankaj Bansal (1): >>>>>>> drivers: net: ldpaa_eth: fix resource leak >>>>>>> >>>>>>> Siva Durga Prasad Paladugu (2): >>>>>>> net: phy: Reloc next and prev pointers inside phy_drivers >>>>>>> net: phy: Fix return value check phy_probe >>>>>>> >>>>>>> Valentin-catalin Neacsu (1): >>>>>>> net: phy: aquantia: Set only autoneg on in register 4.c441 >>>>>>> >>>>>>> Vladimir Oltean (6): >>>>>>> net: phy: ar803x: Address packet drops at low traffic rate due to SmartEEE feature >>>>>>> net: phy: ar803x: Make RGMII Tx delays actually configurable for AR8035 >>>>>>> net: phy: ar803x: Use common functions for RGMII internal delays >>>>>>> net: phy: ar803x: Clarify the configuration of the CLK_25M output pin >>>>>>> net: phy: ar803x: Explicitly disable RGMII delays >>>>>> >>>>>> Tom, this [1] is the patch that is breaking the evm. It doesn't affect >>>>>> BB Black because it uses an SMSC phy, where as this evm uses an >>>>>> AR8031/AR8033. >>>>>> >>>>>> Is it possible the device tree [2] is wrong for the board? It lists >>>>>> 'phy-mode = "rgmii-txid";', so that means that with this patch the RX >>>>>> delay is now being disabled. >>>>>> >>>>>> Any thoughts, Vladimir? >>>>>> >>>>>> Thanks, >>>>>> -Joe >>>>>> >>>>>> [1] b3224e0f7e - "net: phy: ar803x: Explicitly disable RGMII delays" >>>>>> [2] arch/arm/dts/am335x-evm.dts >>>>>> >>>>>>> net: phy: ar803x: Clarify the intention of ar8021_config >>>>>>> >>>>>>> arch/arm/dts/sama5d3xcm.dtsi | 32 +++--- >>>>>>> arch/arm/dts/sama5d3xcm_cmp.dtsi | 32 +++--- >>>>>>> arch/arm/dts/socfpga_arria5_socdk.dts | 4 +- >>>>>>> arch/arm/dts/socfpga_cyclone5_is1.dts | 4 +- >>>>>>> arch/arm/dts/socfpga_cyclone5_socdk.dts | 4 +- >>>>>>> arch/arm/dts/socfpga_cyclone5_sockit.dts | 4 +- >>>>>>> arch/arm/dts/socfpga_cyclone5_vining_fpga.dts | 4 +- >>>>>>> board/ti/ks2_evm/mux-k2g.h | 36 +++---- >>>>>>> cmd/mdio.c | 27 +++-- >>>>>>> doc/device-tree-bindings/net/micrel-ksz90x1.txt | 27 +++++ >>>>>>> drivers/net/ldpaa_eth/ldpaa_eth.c | 1 + >>>>>>> drivers/net/phy/Kconfig | 41 ++++++++ >>>>>>> drivers/net/phy/aquantia.c | 7 +- >>>>>>> drivers/net/phy/atheros.c | 128 ++++++++++++++++------- >>>>>>> drivers/net/phy/micrel_ksz90x1.c | 24 ++++- >>>>>>> drivers/net/phy/phy.c | 21 +++- >>>>>>> drivers/net/phy/realtek.c | 19 ++++ >>>>>>> drivers/net/phy/ti.c | 130 +++++------------------- >>>>>>> include/phy.h | 70 +++++++++++++ >>>>>>> 19 files changed, 394 insertions(+), 221 deletions(-) >>>>>>> >>>>>>> Thanks! >>>>>>> -Joe >>>>>>> _______________________________________________ >>>>>>> U-Boot mailing list >>>>>>> U-Boot@lists.denx.de >>>>>>> https://lists.denx.de/listinfo/u-boot >>>>>> >>>>> >>>>> Hi Joe, Tom, >>>>> >>>>> It sounds like what Joe pointed to (my patch) has a high chance of >>>>> causing link failure. >>>>> If the board is relying on RX delays in the Atheros PHY to ensure >>>>> correct RGMII timing budget, then for sure it was working before and now >>>>> it is broken. In that case, it was working by mistake; the DT blob is >>>>> broken and should be corrected. >>>>> Sorry for the trouble this has caused. >>>> >>>> How is this handled in the Linux kernel and/or why doesn't it fail >>>> there? >>> >>> On the latest net-next, it's handled the same: disable everything and >>> enable only what's specified in phy-mode. I didn't keep track of older >>> versions. >> >> So you would expect the network to be broken on net-next with this DTS? >> Because that's not allowed... >> > > Yes, if it's the same problem, the behavior should be the same. > I see net-next is only rejecting the bad DT since 2019-02-21: > https://patchwork.kernel.org/patch/10819279/ > So it depends when you checked it last time.

I'll see about testing eth in Linux with net-next tomorrow, but, uh, is there some work-around for the DTS so that ethernet works?

Work-around in Linux or in U-boot? U-boot can fix up the DT for Linux and replace "rgmii-txid" with "rgmii-id", whereas neither Linux nor U-boot can do anything for themselves except either load a correct DT blob, or not use (revert, not upgrade to, etc) the enforcing patch.

...

You cannot break existing DTs

RGMII delay breakages happen all the time. To play the devil's advocate, protecting against them should be done by observing the semantics described in the devicetree/bindings document. If you follow Linux net-next, a breakage for RTL8211F is planned out right as we speak.

But let's wait until we know for sure that this is what's causing the issues.

Regards, -Vladimir

Hi Tom,

I understand now what you mean by "workaround". Yes you can enable RGMII delays by hand in U-boot:

Check if RX delay is enabled (bit 15 of debug register 0): => mdio write eTSEC1 0x1D 0 => mdio read eTSEC1 0x1E => mdio write eTSEC1 0x1E <new value>

Check if TX delay is enabled (bit 8 of debug register 5): => mdio write eTSEC1 0x1D 5 => mdio read eTSEC1 0x1E => mdio write eTSEC1 0x1E <new value>

Replace eTSEC1 with your net device from "mdio list".

I'm not sure what the follow-on action is here. I'll hold off on Vladimir's series until there is a resolution that everyone is satisfied with.

While I personally think this breaks the "DT is ABI" model, the required change has already been merged into Linus' tree so we just need to re-sync the am335x DTS files again and then we can apply these changes. I'll put it on my TODO list.

Hi,

sorry for pulling up such an old thread. Are there any news on this? Because I've had the same fix for the atheros PHYs posted to the u-boot mailing list lately (without knowing anything about this thread).

So with that rule "don't break old (and very wrong!) DT blobs" we seem not to be able to fix any serious errors. Because while it works for the most cases eg. phy-mode = rgmii-id, it wont work (and cant be fixed!) for rgmii-txid (or was it rxid? the one where the hardware default is enabled).

I could also post my series without fixing that issue and just adding clock support. But I'd rather see that fixed.

So, there's patches to fix / keep cpsw working in U-Boot. So while I believe that Linux broke the rule of "DT is ABI" I've also gone (virtually) hoarse pointing this out to the people that were supposed to care. So it's on my short list to grab the changes that are needed for the TI platforms and we can grab all of the other changes as well.

On Sat, Nov 02, 2019 at 04:30:11PM +0200, Vladimir Oltean wrote:

On Sat, 2 Nov 2019 at 16:12, Tom Rini trini@konsulko.com wrote:

...

On Sat, Nov 02, 2019 at 02:17:28PM +0100, Michael Walle wrote:

...

Am 2019-05-15 16:58, schrieb Tom Rini:

...

On Fri, May 10, 2019 at 09:50:45PM +0000, Joe Hershberger wrote:

...

Hi Vladimir and Tom,

On Thu, May 9, 2019 at 7:51 AM Vladimir Oltean vladimir.oltean@nxp.com wrote:

...

On 09.05.2019 02:05, Vladimir Oltean wrote: > On 5/9/19 1:55 AM, Tom Rini wrote: >> On Wed, May 08, 2019 at 10:52:28PM +0000, Vladimir Oltean wrote: >>> On 5/9/19 1:48 AM, Tom Rini wrote: >>>> On Wed, May 08, 2019 at 10:45:50PM +0000, Vladimir Oltean wrote: >>>>> On 5/9/19 1:42 AM, Tom Rini wrote: >>>>>> On Wed, May 08, 2019 at 10:40:57PM +0000, Vladimir Oltean wrote: >>>>>>> On 5/9/19 1:24 AM, Joe Hershberger wrote: >>>>>>>> On Tue, May 7, 2019 at 5:15 PM Joe Hershberger joe.hershberger@ni.com wrote: >>>>>>>>> >>>>>>>>> Hi Tom, >>>>>>>>> >>>>>>>>> The following changes since commit 8d7f06bbbef16f172cd5e9c4923cdcebe16b8980: >>>>>>>>> >>>>>>>>> I rebased on your master and built for BB Black. DHCP seems to work fine. >>>>>>>>> MLO also now fits again. >>>>>>>>> >>>>>>>>> Merge branch 'master' of git://git.denx.de/u-boot-sh (2019-05-07 09:38:00 -0400) >>>>>>>>> >>>>>>>>> are available in the git repository at: >>>>>>>>> >>>>>>>>> git://git.denx.de/u-boot-net.git master >>>>>>>>> >>>>>>>>> for you to fetch changes up to 8d0c6858455e89b089222a08d55ff711681ca011: >>>>>>>>> >>>>>>>>> net: phy: micrel: Find Micrel PHY node correctly (2019-05-07 14:51:55 -0500) >>>>>>>>> >>>>>>>>> ---------------------------------------------------------------- >>>>>>>>> Carlo Caione (4): >>>>>>>>> net: phy: Add generic helpers to access MMD PHY registers >>>>>>>>> net: phy: ti: use generic helpers to access MMD registers >>>>>>>>> cmd: mdio: Switch to generic helpers when accessing the registers >>>>>>>>> net: phy: realtek: Introduce quirk to mark RXC not stoppable >>>>>>>>> >>>>>>>>> James Byrne (2): >>>>>>>>> net: phy: micrel: Use correct skew values on KSZ9021 >>>>>>>>> net: phy: micrel: Find Micrel PHY node correctly >>>>>>>>> >>>>>>>>> Murali Karicheri (2): >>>>>>>>> ARM: k2g-gp-evm: update to rgmii pinmux configuration >>>>>>>>> ARM: k2g-ice: Add pinmux support for rgmii interface >>>>>>>>> >>>>>>>>> Pankaj Bansal (1): >>>>>>>>> drivers: net: ldpaa_eth: fix resource leak >>>>>>>>> >>>>>>>>> Siva Durga Prasad Paladugu (2): >>>>>>>>> net: phy: Reloc next and prev pointers inside phy_drivers >>>>>>>>> net: phy: Fix return value check phy_probe >>>>>>>>> >>>>>>>>> Valentin-catalin Neacsu (1): >>>>>>>>> net: phy: aquantia: Set only autoneg on in register 4.c441 >>>>>>>>> >>>>>>>>> Vladimir Oltean (6): >>>>>>>>> net: phy: ar803x: Address packet drops at low traffic rate due to SmartEEE feature >>>>>>>>> net: phy: ar803x: Make RGMII Tx delays actually configurable for AR8035 >>>>>>>>> net: phy: ar803x: Use common functions for RGMII internal delays >>>>>>>>> net: phy: ar803x: Clarify the configuration of the CLK_25M output pin >>>>>>>>> net: phy: ar803x: Explicitly disable RGMII delays >>>>>>>> >>>>>>>> Tom, this [1] is the patch that is breaking the evm. It doesn't affect >>>>>>>> BB Black because it uses an SMSC phy, where as this evm uses an >>>>>>>> AR8031/AR8033. >>>>>>>> >>>>>>>> Is it possible the device tree [2] is wrong for the board? It lists >>>>>>>> 'phy-mode = "rgmii-txid";', so that means that with this patch the RX >>>>>>>> delay is now being disabled. >>>>>>>> >>>>>>>> Any thoughts, Vladimir? >>>>>>>> >>>>>>>> Thanks, >>>>>>>> -Joe >>>>>>>> >>>>>>>> [1] b3224e0f7e - "net: phy: ar803x: Explicitly disable RGMII delays" >>>>>>>> [2] arch/arm/dts/am335x-evm.dts >>>>>>>> >>>>>>>>> net: phy: ar803x: Clarify the intention of ar8021_config >>>>>>>>> >>>>>>>>> arch/arm/dts/sama5d3xcm.dtsi | 32 +++--- >>>>>>>>> arch/arm/dts/sama5d3xcm_cmp.dtsi | 32 +++--- >>>>>>>>> arch/arm/dts/socfpga_arria5_socdk.dts | 4 +- >>>>>>>>> arch/arm/dts/socfpga_cyclone5_is1.dts | 4 +- >>>>>>>>> arch/arm/dts/socfpga_cyclone5_socdk.dts | 4 +- >>>>>>>>> arch/arm/dts/socfpga_cyclone5_sockit.dts | 4 +- >>>>>>>>> arch/arm/dts/socfpga_cyclone5_vining_fpga.dts | 4 +- >>>>>>>>> board/ti/ks2_evm/mux-k2g.h | 36 +++---- >>>>>>>>> cmd/mdio.c | 27 +++-- >>>>>>>>> doc/device-tree-bindings/net/micrel-ksz90x1.txt | 27 +++++ >>>>>>>>> drivers/net/ldpaa_eth/ldpaa_eth.c | 1 + >>>>>>>>> drivers/net/phy/Kconfig | 41 ++++++++ >>>>>>>>> drivers/net/phy/aquantia.c | 7 +- >>>>>>>>> drivers/net/phy/atheros.c | 128 ++++++++++++++++------- >>>>>>>>> drivers/net/phy/micrel_ksz90x1.c | 24 ++++- >>>>>>>>> drivers/net/phy/phy.c | 21 +++- >>>>>>>>> drivers/net/phy/realtek.c | 19 ++++ >>>>>>>>> drivers/net/phy/ti.c | 130 +++++------------------- >>>>>>>>> include/phy.h | 70 +++++++++++++ >>>>>>>>> 19 files changed, 394 insertions(+), 221 deletions(-) >>>>>>>>> >>>>>>>>> Thanks! >>>>>>>>> -Joe >>>>>>>>> _______________________________________________ >>>>>>>>> U-Boot mailing list >>>>>>>>> U-Boot@lists.denx.de >>>>>>>>> https://lists.denx.de/listinfo/u-boot >>>>>>>> >>>>>>> >>>>>>> Hi Joe, Tom, >>>>>>> >>>>>>> It sounds like what Joe pointed to (my patch) has a high chance of >>>>>>> causing link failure. >>>>>>> If the board is relying on RX delays in the Atheros PHY to ensure >>>>>>> correct RGMII timing budget, then for sure it was working before and now >>>>>>> it is broken. In that case, it was working by mistake; the DT blob is >>>>>>> broken and should be corrected. >>>>>>> Sorry for the trouble this has caused. >>>>>> >>>>>> How is this handled in the Linux kernel and/or why doesn't it fail >>>>>> there? >>>>> >>>>> On the latest net-next, it's handled the same: disable everything and >>>>> enable only what's specified in phy-mode. I didn't keep track of older >>>>> versions. >>>> >>>> So you would expect the network to be broken on net-next with this DTS? >>>> Because that's not allowed... >>>> >>> >>> Yes, if it's the same problem, the behavior should be the same. >>> I see net-next is only rejecting the bad DT since 2019-02-21: >>> https://patchwork.kernel.org/patch/10819279/ >>> So it depends when you checked it last time. >> >> I'll see about testing eth in Linux with net-next tomorrow, but, uh, is >> there some work-around for the DTS so that ethernet works? > > Work-around in Linux or in U-boot? > U-boot can fix up the DT for Linux and replace "rgmii-txid" with > "rgmii-id", whereas neither Linux nor U-boot can do anything for > themselves except either load a correct DT blob, or not use (revert, not > upgrade to, etc) the enforcing patch. > >> You cannot break existing DTs > RGMII delay breakages happen all the time. To play the devil's advocate, > protecting against them should be done by observing the semantics > described in the devicetree/bindings document. > If you follow Linux net-next, a breakage for RTL8211F is planned out > right as we speak. > > But let's wait until we know for sure that this is what's causing the > issues. > > Regards, > -Vladimir >

Hi Tom,

I understand now what you mean by "workaround". Yes you can enable RGMII delays by hand in U-boot:

Check if RX delay is enabled (bit 15 of debug register 0): => mdio write eTSEC1 0x1D 0 => mdio read eTSEC1 0x1E => mdio write eTSEC1 0x1E <new value>

Check if TX delay is enabled (bit 8 of debug register 5): => mdio write eTSEC1 0x1D 5 => mdio read eTSEC1 0x1E => mdio write eTSEC1 0x1E <new value>

Replace eTSEC1 with your net device from "mdio list".

I'm not sure what the follow-on action is here. I'll hold off on Vladimir's series until there is a resolution that everyone is satisfied with.

While I personally think this breaks the "DT is ABI" model, the required change has already been merged into Linus' tree so we just need to re-sync the am335x DTS files again and then we can apply these changes. I'll put it on my TODO list.

Hi,

sorry for pulling up such an old thread. Are there any news on this? Because I've had the same fix for the atheros PHYs posted to the u-boot mailing list lately (without knowing anything about this thread).

So with that rule "don't break old (and very wrong!) DT blobs" we seem not to be able to fix any serious errors. Because while it works for the most cases eg. phy-mode = rgmii-id, it wont work (and cant be fixed!) for rgmii-txid (or was it rxid? the one where the hardware default is enabled).

I could also post my series without fixing that issue and just adding clock support. But I'd rather see that fixed.

So, there's patches to fix / keep cpsw working in U-Boot. So while I believe that Linux broke the rule of "DT is ABI" I've also gone (virtually) hoarse pointing this out to the people that were supposed to care. So it's on my short list to grab the changes that are needed for the TI platforms and we can grab all of the other changes as well.

--

Tom,

How come the people who proclaim that DT is ABI (which it is, nobody is saying otherwise) are the same as (or do it in the name of) those who don't follow it?

It's a good question, and I don't know why no one cared.

Had your boards been written with the correct phy-mode in the first place, no breakage would have happened, and maybe the Atheros PHY driver bugs would have been caught earlier.

Yes, it's true that had the problem been caught 6 years ago, a lot of things would have been better. But it wasn't.

I agree with Michael that interpreting the "DT is ABI" phrase as "be compatible with past mistakes" (which are not in the ABI specification, but in its interpretation, at that!) is only going to lead to a cacophony of incoherent interpretations, which at some point will become impossible to reconcile.

Part of the problem here is that I was not one of the people at the time saying that "DT must be ABI". I was in fact worried that there would be a pain to fix broken DT files once they were out in the wild. Other people said it would be fine and that it had been done for a long while. And that future break would be handled in a backwards compatible way.

But here we are.

And for U-Boot, I guess the point is moot, since the DT blob is typically packaged with the bootloader binary anyway.

Right, so that's the whole problem. You can't pass the old DTB to the new kernel and have things work. Since the example in question is an EVM rather than shipping HW platform based on the EVM, we can say "No one shipped the DTB embedded in to the physical HW, just pass along the new DTB!". I am far from convinced that there's not custom hardware based on the EVM design and reference dts file that won't hit a problem here when moving up.

But again, I've given up. I say that the ABI meant that the wrong value was supposed to work since that's what happened and a new version of the binding needed to be used where the right value must be used. Others disagree. I'm not holding U-Boot up on the new changes over this, I just haven't put together the PR. It's on my short list now.