Date:	Sun, 07 Nov 2010 14:04:23 +0200
From:	Maxim Levitsky <maximlevitsky@...il.com>
To:	Stefan Richter <stefanr@...6.in-berlin.de>
Cc:	linux1394-devel@...ts.sourceforge.net,
	"netdev@...r.kernel.org" <netdev@...r.kernel.org>
Subject: Re: [PATCH] firewire: net: rate-limit log spam at transmit failure

On Sun, 2010-11-07 at 12:10 +0100, Stefan Richter wrote:
> Maxim Levitsky wrote:
> > On Sun, 2010-11-07 at 00:23 +0100, Stefan Richter wrote:
> >> On  6 Nov, Stefan Richter wrote:
> >>> Then I tried an XIO2213A card in the AMD PC (again the Intel PC as peer)
> >>> and got 243 times "failed: 12" i.e. RCODE_BUSY and 81 times "failed: 10"
> >>> i.e. RCODE_SEND_ERROR during ftp transfer of a >500 MB large file from
> >>> XIO2213A to FW323.
> > 
> > I also am getting strange results (but very good compared to what I had
> > recently).
> > 
> > With all your patches, I get very stable TCP and UDP streams from laptop
> > to desktop at 180~190 Mbits/s.
> > 
> > However, the opposite direction (desktop->laptop) still suffers from
> > tlabel exhaustion.
> > I added some printks, and I see, clearly that netif_stop_queue doesn't
> > always work (probably this is intended?).
> > 
> > If I replace == with >= in inc_queue_packets and similar in
> > dec_queued_packets, then tlabel exhaustion disappears, and I get ~240
> > Mbit/s on TCP and UDP.
> 
> Remind me, is this FireWire 800?  And what controllers in particular?  I get
> about half of your numbers with FireWire 400 connections.
Firewire 400.

Laptop: Ricoh R5C832
Desktop: TI TSB43AB22A


> 
> The == vs. >= is a good hint.  If .ndo_start_xmit can be entered by multiple
> CPUs, the upper limit will clearly exceeded eventually.
> 
> With >= instead of ==, the same test as that quoted above gives 71x RCODE_BUSY
> + 0x RCODE_SEND_ERROR, and 59x RCODE_BUSY + 0x RCODE_SEND_ERROR in a
> repetition.  (0x + 0x in the other direction.)  There were no RCODE_CANCELLED
> occurrences, which I had occasionally in the past.
RCODE_CANCELLED is result of timeout timer which is set to run with 0
jiffies in the future.

> 
> I then tried
> 
> 	if (dev->queued_packets >= FWNET_MAX_QUEUED_PACKETS)
> 		return NETDEV_TX_BUSY;
> 
> at the top of fwnet_tx but it did not change the amount of RCODE_BUSY, which
> is not too surprising.  So next I should have a look at the responder side again.
I see RCODE_BUSY sometimes too.

> 
> BTW, FireWire 400 CardBus controllers usually feature a limitation of max_rec
> = 1024 (maximum size of asynchronous packets they can receive).  Incidentally,
> the VT6306 card that I used in my other tests from yesterday is one of those.
> So, since link fragmentation is quite common due to this kind of cards, I
> should perhaps count queued fragments instead of queued datagrams.
> 
> > UDP transfers work quite well, tested for few minutes.
> > TCP transfers unfortunelly trigger (probably a hardware) bug in notebook
> > OHCI controller (I have seen that meny times so far.)
> > 
> > Transfer just stops, and controller goes south.
> > If I unload the firewire-ohci, then when I load it:
> > 
> > [ 2062.632532] firewire_ohci 0000:07:00.0: PCI INT A -> GSI 20 (level, low) -> IRQ 20
> > [ 2072.650173] firewire_ohci: Failed to reset ohci card.
> > [ 2072.650267] firewire_ohci 0000:07:00.0: PCI INT A disabled
> > [ 2072.650314] firewire_ohci: probe of 0000:07:00.0 failed with error -16
> > 
> > 
> > Only suspend to ram helps bring it back from that state.
> 
> On the bright side, s2ram fixes things for once instead of breaking them...
Well, of course after s2ram I have to reload the firewire-ohci to make
it work, but that is separate issue.

Best regards,
	Maxim Levitsky




--
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