Re: R8169: Network lockups in 4.18.{8,9,10} (and 4.19 dev)

[Maciej S. Szmigiero]

On 09.10.2018 22:36, Heiner Kallweit wrote:
> On 09.10.2018 16:40, Chris Clayton wrote:
>> Thanks to Maciej and Heiner for their replies.
>>
>> On 09/10/2018 13:32, Maciej S. Szmigiero wrote:
>>> On 07.10.2018 21:36, Chris Clayton wrote:
>>>> Hi again,
>>>>
>>>> I didn't think there was anything in 4.19-rc7 to fix this regression, but tried it anyway. I can confirm that the
>>>> regression is still present and my network still fails when, after a resume from suspend (to ram or disk), I open my
>>>> browser or my mail client. In both those cases the failure is almost immediate - e.g. my home page doesn't get displayed
>>>> in the browser. Pinging one of my ISPs name servers doesn't fail quite so quickly but the reported time increases from
>>>> 14-15ms to more than 1000ms.
>>>
>>> You can try comparing chip registers (ethtool -d eth0) in the working
>>> state (before a suspend) and in the broken state (after a resume).
>>> Maybe there will be some obvious in the difference.
>>>
>>> The same goes for the PCI configuration (lspci -d :8168 -vv).
>>>
>> Maciej suggested comparing the output from lspci -vv for the ethernet device. They are identical.
>>
>> Both Maciej and Heiner suggested comparing the output from "ethtool -d" pre and post suspend. Again, they are identical.
>> Heiner specifically suggested looking at the RxConfig. The value of that is 0x0002870e both pre and post suspend.
>>
> Hmm, this is very weird, especially taking into account that in your original
> report you state that removing the call to rtl_init_rxcfg() from rtl_hw_start()
> fixes the issue. rtl_init_rxcfg() deals with the RxConfig register only and
> register values seem to be the same before and after resume. So how can the
> chip behave differently?
> So far my best guess is that some chip quirk causes it to accept writes to
> register RxConfig, but to misinterpret or ignore the written value.
> So far your report is the only one (affecting RTL8411), but we don't know
> whether other chip versions are affected too.

Also, it is interesting that even if one removes a call to
rtl_init_rxcfg() from rtl_hw_start() the RxConfig register will still get
written to moments later by rtl_set_rx_mode().

The only chip accesses in the meantime seems to be a write to TxConfig by
rtl_set_tx_config_registers() and then a read of RxConfig plus two writes
to MAR0 earlier in rtl_set_rx_mode().

My proposals are:
1) Try swapping "rtl_init_rxcfg(tp);" and "rtl_set_tx_config_registers(tp);"
in rtl_hw_start().
Maybe the chip does not like sometimes that RxConfig is written before
TxConfig.

2) Check the original value of RxConfig (after a resume) before
rtl_init_rxcfg() overwrites it (compile tested only):
--- r8169.c.ori
+++ r8169.c
@@ -5155,6 +5155,9 @@
 	/* Initially a 10 us delay. Turned it into a PCI commit. - FR */
 	RTL_R8(tp, IntrMask);
 	RTL_W8(tp, ChipCmd, CmdTxEnb | CmdRxEnb);
+
+	pr_notice("RxConfig before init was %.8x\n",
+		(unsigned int)RTL_R32(tp, RxConfig));
 	rtl_init_rxcfg(tp);
 	rtl_set_tx_config_registers(tp);
 

This should be the value that you got when you removed the call to
rtl_init_rxcfg() for testing.
Now, knowing the "right" value you can experiment with what rtl_init_rxcfg()
writes (under the "default:" label for your NIC model).

Hope this helps,
Maciej