Hi all,

01.06.2021 10:20, Arnd Bergmann:
[...]

quoted

What was discussed here 16 yrs ago should sound familiar to you.
https://www.mail-archive.com/linux-kernel@vger.kernel.org/msg92234.html
"It was an option in my BIOS PCI level/edge settings as I posted."
You could check whether you have same/similar option in your BIOS
and play with it.

Yes indeed, this motherboard does have such an option, and it defaulted 
to "Edge", which apparently is not what PCI device normally expects.
Changing it to "Level" made unmodified kernel 2.6.4 work fine.
And 8259A.pl comfirms this, too.

Before:
# ./8259A.pl
irq 0: 00, edge
irq 1: 00, edge
irq 2: 00, edge
irq 3: 00, edge
irq 4: 00, edge
irq 5: 00, edge
irq 6: 00, edge
irq 7: 00, edge
irq 8: 00, edge
irq 9: 00, edge
irq 10: 00, edge
irq 11: 00, edge
irq 12: 00, edge
irq 13: 00, edge
irq 14: 00, edge
irq 15: 00, edge

After:
# ./8259A.pl
irq 0: 00, edge
irq 1: 00, edge
irq 2: 00, edge
irq 3: 00, edge
irq 4: 00, edge
irq 5: 00, edge
irq 6: 00, edge
irq 7: 00, edge
irq 8: 06, edge
irq 9: 06, level
irq 10: 06, level
irq 11: 06, edge
irq 12: 06, edge
irq 13: 06, edge
irq 14: 06, edge
irq 15: 06, edge

So it appears that the interrupt is lost if new TX events come in after the
status register is read, and that checking it again manages to make that
race harder to hit, but maybe not reliably.

It looks like incorrect IRQ triggering mode makes 2 or more IRQs merge 
into one, kind of. However, if I understand this 8139 operation logic 
correctly, the possible max number of signaled events in one go is 
limited by the number of tx/rx descriptors and can not grow beyond it 
while inside the interrupt handler in any case. If so, using the loop 
would seem not that bad, and the limit would be certainly not 20 but 
max(NUM_TX_DESC, CONFIG_8139_RXBUF_IDX) == 4.

The best idea I have for a proper fix would be to move the TX processing
into the poll function as well, making sure that by the end of that function
the driver is either still in napi polling mode, or both RX and TX interrupts
are enabled and acked.

This one is too complicated for me to implement myself, so I'll have to 
wait if someone does this.

Alternatively, maybe it is possible to explicitely request level mode 
from 8259 at the driver startup?


Thank you,

Regards,
Nikolai

          Arnd