igb: Problems with early PHY reset when using an SFP module with SGMII as phy->addr is unset
Grönke
Hello Aaron, Hello Jeff,
I have problems with the newer versions of the e1000 (igb) driver of current kernels. I stumbled over this with a vendor kernel[*] after they pulled in some updates. When starting on my board, which has several i210 MACs connected to SFP ports. In case I use the hardware flashed for SGMII and SGMII capable SFP modules I see errors like this:
Quoted 2 lines
[ 4776.137522] igb: Intel(R) Gigabit Ethernet Network Driver - version 5.4.0-k [ 4776.137525] igb: Copyright (c) 2007-2014 Intel Corporation.
[...]
Quoted 1 line
[ 4776.337079] igb: probe of 0000:09:00.0 failed with error -3
I tracked this down to the failing reset that was introduced by Aaron[*]:
Quoted 2 lines
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/drivers/net/ethernet/intel/igb/e1000_82575.c?id=182785335447957409282ca745aa5bc3968facee
igb: reset the PHY before reading the PHY ID
The problem here, at least for me, is that during this time in the initialization the phy->addr is '0' and thus 'e1000_write_phy_reg_sgmii_82575' fails. The SGMII PHY will only be discovered later on in 'igb_get_phy_id_82575'.
Quoted 3 lines
[ 4776.336987] igb: IGB: Soft resetting SGMII attached PHY... [ 4776.336999] IGB: PHY I2C Address 0 is out of range. [ 4776.337009] igb: IGB: Error resetting the PHY.
When hacking the driver and presenting the PHY to the correct address, or doing a discovery of the PHY in case SGMII is active and MDIO is not used, it works. However, it seems a bit opposed to your patch. I understand your patch note in the way that in some cases the PHY has to be reset as otherwise the PHY ID can not be discovered and the "scan" would thus fail. May this inly be the case if SGMII is not used and thus 'igb_phy_hw_reset' would normally be called? It seems that in case SGMII is used the phy->addr has to be known to do a reset.
I know too little about all the possible IGB/i210 configuration to judge what is okay here. This is why I wanted to get in touch. I am not sure what would be the right fix. While I can fix it for myself and our product by disabling your patch and have no reset, I would much prefer an upstream fix that I can backport.
Kind regards
Christian
[*] Note: This is still the case in the most current driver on the master branch.
[*] Note2: I first observed this on a L4Linux 4.9 (https://l4linux.org/) running on L4RE. However, I can also reproduce this with a current CentOS 7.5 which has a 3.10++ kernel and a Ubuntu 18.04.
Signature · 4 lines
Christian GR?NKE INFODAS GmbH ~ Rhonestra?e 2 ~ 50765 K?ln ~ Germany Phone: +49 221 7091255 ~ Mail: c.groenke at infodas.de
Quoted 26 lines · Christian
Hello Aaron, Hello Jeff,
I have problems with the newer versions of the e1000 (igb) driver of current kernels. I stumbled over this with a vendor kernel[*] after they pulled in some updates. When starting on my board, which has several i210 MACs connected to SFP ports. In case I use the hardware flashed for SGMII and SGMII capable SFP modules I see errors like this:
Quoted 2 lines
[ 4776.137522] igb: Intel(R) Gigabit Ethernet Network Driver - version 5.4.0-k [ 4776.137525] igb: Copyright (c) 2007-2014 Intel Corporation.[...]
Quoted 1 line
[ 4776.337079] igb: probe of 0000:09:00.0 failed with error -3I tracked this down to the failing reset that was introduced by Aaron[*]:
Quoted 2 lines
https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/drivers/net/ethernet/intel/igb/e1000_82575.c?id=182785335447957409282ca745aa5bc3968facee
igb: reset the PHY before reading the PHY IDThe problem here, at least for me, is that during this time in the initialization the
phy->addris '0' and thus 'e1000_write_phy_reg_sgmii_82575' fails. The SGMII PHY will only be discovered later on in 'igb_get_phy_id_82575'.Quoted 3 lines
[ 4776.336987] igb: IGB: Soft resetting SGMII attached PHY... [ 4776.336999] IGB: PHY I2C Address 0 is out of range. [ 4776.337009] igb: IGB: Error resetting the PHY.When hacking the driver and presenting the PHY to the correct address, or doing a discovery of the PHY in case SGMII is active and MDIO is not used, it works. However, it seems a bit opposed to your patch. I understand your patch note in the way that in some cases the PHY has to be reset as otherwise the PHY ID can not be discovered and the "scan" would thus fail. May this inly be the case if SGMII is not used and thus '
igb_phy_hw_reset' would normally be called? It seems that in case SGMII is used thephy->addrhas to be known to do a reset.I know too little about all the possible IGB/i210 configuration to judge what is okay here. This is why I wanted to get in touch. I am not sure what would be the right fix. While I can fix it for myself and our product by disabling your patch and have no reset, I would much prefer an upstream fix that I can backport.
Kind regards
Christian
Quoted 6 lines
[*] Note: This is still the case in the most current driver on the master branch.
[*] Note2: I first observed this on a L4Linux 4.9 (https://l4linux.org/) running on L4RE. However, I can also reproduce this with a current CentOS 7.5 which has a 3.10++ kernel and a Ubuntu 18.04.--
Christian GR?NKE
Hi Christian,
Looking over the code I would say it is probably acceptable to just revert the patch. In addition you could also probably revert commit 440aeca4b985 "igb: Explicitly select page 0 at initialization". Both of those patches seem to be redundant as this was also fixed via a third commit 4e684f59d760 "igb: Workaround for igb i210 firmware issue" which I feel did a much better job of addressing the issue without introducing other side effects.
If you want to submit the patches to revert the first commit you mentioned, and the one I mentioned I would be willing to ack it and see if we can get it pushed through the tree.
Thanks.
- Alex
On Fri, Jun 22, 2018 at 11:23 AM, Alexander Duyck <[email protected]
Quoted 71 lines
wrote:
Quoted 3 lines · Christian
Hello Aaron, Hello Jeff,
I have problems with the newer versions of the e1000 (igb) driver of
current kernels. I stumbled over this with a vendor kernel[*] after they pulled in some updates. When starting on my board, which has several i210 MACs connected to SFP ports. In case I use the hardware flashed for SGMII and SGMII capable SFP modules I see errors like this:
Quoted 2 lines
Quoted 2 lines
[ 4776.137522] igb: Intel(R) Gigabit Ethernet Network Driver - version5.4.0-k
Quoted 6 lines
Quoted 1 line
[ 4776.137525] igb: Copyright (c) 2007-2014 Intel Corporation.[...]
Quoted 1 line
[ 4776.337079] igb: probe of 0000:09:00.0 failed with error -3I tracked this down to the failing reset that was introduced by Aaron[*]:
linux.git/commit/drivers/net/ethernet/intel/igb/e1000_82575.c?id=
182785335447957409282ca745aa5bc3968faceeQuoted 3 lines
Quoted 1 line
igb: reset the PHY before reading the PHY ID
The problem here, at least for me, is that during this time in the
initialization the
phy->addris '0' and thus 'e1000_write_phy_reg_sgmii_82575' fails. The SGMII PHY will only be discovered later on in 'igb_get_phy_id_82575'.Quoted 6 lines
Quoted 4 lines
[ 4776.336987] igb: IGB: Soft resetting SGMII attached PHY... [ 4776.336999] IGB: PHY I2C Address 0 is out of range. [ 4776.337009] igb: IGB: Error resetting the PHY.When hacking the driver and presenting the PHY to the correct address,
or doing a discovery of the PHY in case SGMII is active and MDIO is not used, it works. However, it seems a bit opposed to your patch. I understand your patch note in the way that in some cases the PHY has to be reset as otherwise the PHY ID can not be discovered and the "scan" would thus fail. May this inly be the case if SGMII is not used and thus '
igb_phy_hw_reset' would normally be called? It seems that in case SGMII is used thephy->addrhas to be known to do a reset.Quoted 2 lines
I know too little about all the possible IGB/i210 configuration to judge
what is okay here. This is why I wanted to get in touch. I am not sure what would be the right fix. While I can fix it for myself and our product by disabling your patch and have no reset, I would much prefer an upstream fix that I can backport.
Quoted 5 lines
Kind regards
Christian[*] Note: This is still the case in the most current driver on the
master branch.
Quoted 1 line
[*] Note2: I first observed this on a L4Linux 4.9 (https://l4linux.org/)
running on L4RE. However, I can also reproduce this with a current CentOS 7.5 which has a 3.10++ kernel and a Ubuntu 18.04.
Quoted 4 lines
--
Christian GR?NKE
Hi Christian,
Looking over the code I would say it is probably acceptable to just revert the patch. In addition you could also probably revert commit
440aeca4b985"igb: Explicitly select page 0 at initialization". Both of those patches seem to be redundant as this was also fixed via a third commit4e684f59d760"igb: Workaround for igb i210 firmware issue" which I feel did a much better job of addressing the issue without introducing other side effects.
Quoted 8 lines
If you want to submit the patches to revert the first commit you mentioned, and the one I mentioned I would be willing to ack it and see if we can get it pushed through the tree.
Thanks.
- Alex
I would be happy to test this patch out. It did solve a specific problem on
the machine I tested this on; so I would not want to regress this fix.
Thanks,
--chris
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.osuosl.org/pipermail/intel-wired-lan/attachments/20180622/916d411e/attachment.html>