On Tue, 2017-02-28 at 10:14 +0000, David Laight wrote:

From: Eric Dumazet

quoted

Sent: 27 February 2017 22:35
On Mon, 2017-02-27 at 14:14 -0800, Stephen Hemminger wrote:

quoted

The original design (as Davem mentioned) was that IRQ's must be disabled
during device polling. If that was true, then the race above
would be impossible.

I would love to see an alternative patch.

Can you test for 'receive data available' after telling the NAPI
logic that you've finished?
You'd then need to force a reschedule.

I think your proposed patch will do a reschedule if any packet arrives
during the receive processing, not just when one arrives right at the end.
You might want to 'unset' the reschedule flag before each check of the
receive ring.

I also wonder about the cost of processing the MSI-X (I guess) interrupts
compared to the cost of posted PCIe writes to disable and/or mask the
interrupt generation.
Clearly you don't want to do PCIe reads.

Have you seen the mlx4 patch I provided ?

Then, I did not want to review 100+ NAPI drivers and provide patches for
them.

This generic solution is basically free. Same number of atomic
operations.

Given it took more than 2 years to even spot the bug, I have no idea how
people on netdev expect me to review all drivers. This is crazy.