request_threaded_irq()

request_threaded_irq()

[Mark Ryden]

Hello netdev,

grepping under net-next-2.6/drivers/net for request_threaded_irq() ,
shows that it appears only in 3 drivers:

can/mcp251x.c
wireless/b43/main.c
wireless/rt2x00/rt2x00pci.c

I was wondering: when thinking about performance, is it worthwhile to use this
API instead of ordinary request_irq() . It seems to me that
request_threaded_irq()  might
be better in some cases than NAPI polling forr network drivers (or at
list it might be so in some systems, maybe multicore ?)


Best,
Mark Ryden

Re: request_threaded_irq()

[Stephen Hemminger]

On Fri, 12 Nov 2010 14:27:11 +0200
Mark Ryden <markryde@gmail.com> wrote:

> Hello netdev,
> 
> grepping under net-next-2.6/drivers/net for request_threaded_irq() ,
> shows that it appears only in 3 drivers:
> 
> can/mcp251x.c
> wireless/b43/main.c
> wireless/rt2x00/rt2x00pci.c
> 
> I was wondering: when thinking about performance, is it worthwhile to use this
> API instead of ordinary request_irq() . It seems to me that
> request_threaded_irq()  might
> be better in some cases than NAPI polling forr network drivers (or at
> list it might be so in some systems, maybe multicore ?)

Threaded irq is not needed for properly written NAPI driver.
A NAPI driver should do minimum work in real irq and the whole NAPI processing
will happen in soft-irq.

There has been discussion of moving NAPI softirq into a high
priority thread. But last time I tried it, the performance was noticably
worse on network intensive benchmarks.