From: Arnd Bergmann <arnd@arndb.de>
To: Stephen Hemminger <shemminger@osdl.org>
Subject: Re: [RFC] HOWTO use NAPI to reduce TX interrupts
Date: Sat, 19 Aug 2006 13:25:18 +0200
User-Agent: KMail/1.9.1
Cc: linuxppc-dev@ozlabs.org, akpm@osdl.org, James K Lewis <jklewis@us.ibm.com>,
       linux-kernel@vger.kernel.org, netdev@vger.kernel.org,
       Jeff Garzik <jgarzik@pobox.com>,
       Jens Osterkamp <Jens.Osterkamp@de.ibm.com>,
       David Miller <davem@davemloft.net>
References: <20060818220700.GG26889@austin.ibm.com> <200608190256.26373.arnd@arndb.de> <44E7BB7F.7030204@osdl.org>
In-Reply-To: <44E7BB7F.7030204@osdl.org>
MIME-Version: 1.0
Content-Type: text/plain;
  charset="iso-8859-1"
Content-Transfer-Encoding: 8BIT
Content-Disposition: inline
Message-Id: <200608191325.19557.arnd@arndb.de>
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 1749
Lines: 45

On Sunday 20 August 2006 03:31, Stephen Hemminger wrote:
> 
> The reason reclaim via poll() is efficient is because it avoid causing a 
> softirq that is
> necessary when skb_free_irq() is done. Instead it reuses the softirq 
> from the poll() routine. 

Ok, I completely missed this point so far, thanks for the info.

> Like all Rx NAPI, using poll() for reclaim means: 
> ? ? + aggregating multiple frames in one irq
> ? ? - increased overhead of twiddling with the IRQ mask
> ? ? - more ways to get driver stuck

What is the best way to treat the IRQ mask for TX interrupts?
I guess it should be roughly:

- off when we expect ->poll() to be called, i.e. after calling
  netif_rx_schedule() or returning after a partial rx from poll().
- off when there are no packets left in the TX queue
- on while RX interrupts are on and we're waiting for packets
  to be transmitted.

> Some drivers do all their irq work in the poll() routine (including PHY 
> handling).
> This is good if reading the IRQ status does an auto mask operation.
> 
> The whole NAPI documentation area is a mess and needs a good writer
> to do some major restructuring. It should also be split into reference 
> information,
> tutorial and guide sections.

I won't be able to do that work, I'm neither a good writer nor a networking
person.

Do you think we should still merge a section like the text I wrote up, even
if it makes the text even less well structured? Should I maybe add it
somewhere else than the appendix?

	Arnd <><
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/