Re: [PATCH] [RFT] 2.6.4 - epic100 napi

[Francois Romieu <romieu@xxxxxxxxxxxxx>]

OGAWA Hirofumi <hirofumi@xxxxxxxxxxxxxxxxxx> :
[...]
> Do you care the lost interrupt? If so, I was miss reading it.
> 
> IIRC, PCI spec requires the level-trigger. So devices asserts IRQ
> signal until the driver clears the pending interrupt. No?

<insert usual "correct me if I'm wrong" disclaimer here>

The driver only masks the interruptions which are napi related so
epic_poll() and epic_interrupt() are always racing. I completely agree that
it "wastes" the nice behavior of level-triggered irq. If one wants to avoid
the race, everything should go to the poll() handler. It implies polled
access to the INTSTAT register (as done in epic_rx_err).

I could not find in the archives some general napi-wisdom which suggests
that everything has to go to the poll() handler, be it for simplicity or
for a real gain. So I hesitate to follow the other way and exchange the
polled access for some (locked) traffic between the poll() and the irq
handler.

Comments welcome.

--
Ueimor