Date:	Wed, 29 Apr 2009 12:01:23 +0200
From:	Philippe De Muyter <phdm@...qel.be>
To:	Matthew Lear <matt@...blegen.co.uk>
Cc:	netdev@...r.kernel.org, uclinux-dev@...inux.org
Subject: Re: fec driver question

Hi Matthew,

[CCing uclinux-dev@...inux.org and netdev@...r.kernel.org]

On Wed, Apr 29, 2009 at 09:48:37AM +0100, Matthew Lear wrote:
> Hi Phillippe - Thanks very much for your reply. some comments below:
> 
> > Hi Matthew,
> > On Wed, Apr 29, 2009 at 08:15:43AM +0100, Matthew Lear wrote:
> >> Hello Philippe,
> >>
> >> I hope you don't mind me emailing you. Basically I have a dev board from
> >> freescale for doing some coldfire development on the mfc54455 device.
> >> I'm
> >> using the fec driver in the kernel. Kernel version is 2.6.23. I'm havng
> >> some problems and I was hoping you might be able to help me.
> >>
> >> It seems that running some quite heavy network throughput tests on the
> >> platform result in the driver dropping packets and the userspace app
> >> running on the dev board to run consume ~ 85% cpu. I'm using netcat as
> >> the
> >> app running on host and the target to do the tests.
> >>
> >> I can appreciate that this question  is somewhat 'open' in that there
> >> could be several causes but I'm fairly certain that a) it's not ksoftirq
> >> related and b) that it's not driver related (because it's mature and has
> >> been used in all sorts of different application/platforms).
> >>
> >> Can you think of any possible causes for this? The fact that the driver
> >> is
> >> dropping packets is surely indicative of there not being enough buffers
> >> to
> >> place the incoming data and/or there are issues with the consumption
> >> (and
> >> subsequent freeing of these buffers) by something else.
> >
> > 1. You could make the same test after increasing the number of receive
> > buffers
> > in the driver.
> >
> > 2. Actually, each incoming packet generates one interrupt so it needs some
> > processing time in the interrupt service routine.  Hence if your receive
> > app itself consumes 85% CPU that's probably normal that at times all
> > buffers
> > are used and that the chip has to drop frames.  Check if you have idle
> > time
> > remaining.
> >
> > 3. It can also be a hardware bug/limitation in the chip itself.  I used
> > mainly the FEC driver with mcf5272 chips at 10 Mbps, because 100 Mbps
> > was not really supported in hardware, although it was possible to ask for
> > it.
> > There is an offical errata for that :)
> 
> I did try to increase the number of buffers and I was surprised at the
> result because it seemed that the cpu utilisation of the user space app
> increases. There are some comments at the top of fec.c regarding keeping
> numbers associated to the buffers as powers of 2. I increased the number
> of buffers to 32 but bizarrely it seemed to make things worse (netcat
> consumed ~ 95% cpu). Not sure what's going on there!

For me, it means that you loose/drop less packets.  I surmise that your
CPU is mmu-less, so packets must be copied from kernel to userspace for
each received packet.  That time passed by the kernel in copying the
packet for the app is counted as app time, I presume.
You could measure memcpy's speed and compute how much time is needed
with your expected throughput.

> 
> When you say "check idle time remaining", do you mean in the driver itself
> or use a profiling tool?

I only meant looking for %id in 'top' header.

> 
> I have seen the scenario of the cpu at ~85% and no packets dropped but
> typically there are overruns and in this case /proc/net/dev indicates that
> there are fifo issues within the driver somehow.
> 
> Yes. One interrupt per packet is what I expected but I also have an SH4
> dev board (though it uses a different ethernet driver). Running the same
> kernel version and exactly the same test with netcat on that platform
> shows seriously contrasting results in that cpu utilisation of netcat on
> the sh4 target is minimal (as it should be).

Could that be that sh4 has a mmu, and that its ethernet driver implement
zero-copy mode ?  I'm not an expert in that area though.

> 
> I'm suspecting that it may be a mm or dma issue with how the buffers are
> relayed to the upper layers. The driver is mature isn't it so I would have

I'm not sure at all that dma is used here, but I could be wrong.

> expected that any problem such as this would have been spotted long before
> now? In this regard, I am of the opinion that it could possible be an
> issue with the device as you say.

It depends on what other people do with the ethernet device on their
board.  Here it is only used for some lightweight communication.
And, when I used it, the driver was already mature, but I still discovered
real bugs in initialisation sequences and error recovery, e.g. when testing
link connection/disconnection.

> 
> The coldfire part I have is specified as supporting 10 and 100 Mbps so I
> assume that there are no issues with it. Interesting though that you
> mention the errata...
> 
> I think it's just a case of trying to find where the cpu is spending its
> time. It is quite frustrating though... :-(

Yes, that's part of our job :)

Best regards

Philippe
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Powered by blists - more mailing lists