Hi Heiner,

this issue was on the back-burner for some time, but I've got some
interesting news now.

On Sat, 18 Jul 2020 14:07:50 +0200
Heiner Kallweit [off-list ref] wrote:

quoted

[...]
Maybe the following gives us an idea:
Please do "ethtool -d <if>" after boot and after resume from suspend,
and check for differences.  

The register dump did not reveal anything of interest - the only
differences were in the physical addresses after a device reopen.

However, knowing that reloading the driver can fix the issue, I copied
the initialization sequence from init_one() to rtl8169_resume() and
gave it a try. That works!

Then I started removing the initialization calls one by one. This
exercise left me with a call to rtl_init_rxcfg(), which simply sets the
RxConfig register. In other words, these is the difference between
5.8.4 and my working version:

--- linux-orig/drivers/net/ethernet/realtek/r8169_main.c	2020-09-02 22:43:09.361951750 +0200
+++ linux/drivers/net/ethernet/realtek/r8169_main.c	2020-09-03 10:36:23.915803703 +0200

@@ -4925,6 +4925,9 @@
 
 	clk_prepare_enable(tp->clk);
 
+	if (tp->mac_version == RTL_GIGA_MAC_VER_37)
+		RTL_W32(tp, RxConfig, RX128_INT_EN | RX_DMA_BURST);
+
 	if (netif_running(tp->dev))
 		__rtl8169_resume(tp);
 

This is quite surprising, at least when the device is managed by
NetworkManager, because then it is closed on wakeup, and the open
method should call rtl_init_rxcfg() anyway. So, it might be a timing
issue, or incorrect order of register writes.

Since I have no idea why the above change fixes my issue, I'm hesitant
to post it as a patch. It might break other people's systems...