Re: [REGRESSION] asix: Lots of asix_rx_fixup() errors and slow transmissions

From: Dean Jenkins
Date: Tue May 03 2016 - 06:55:03 EST

On 03/05/16 11:04, Guodong Xu wrote:
On 3 May 2016 at 17:23, Dean Jenkins <Dean_Jenkins@xxxxxxxxxx> wrote:
On 03/05/16 05:55, John Stultz wrote:
In testing with HiKey, we found that since commit 3f30b158eba5c60
(asix: On RX avoid creating bad Ethernet frames), we're seeing lots of
noise during network transfers:

[ 239.027993] asix 1-1.1:1.0 eth0: asix_rx_fixup() Data Header
synchronisation was lost, remaining 988
[ 239.037310] asix 1-1.1:1.0 eth0: asix_rx_fixup() Bad Header Length
0x54ebb5ec, offset 4
[ 239.045519] asix 1-1.1:1.0 eth0: asix_rx_fixup() Bad Header Length
0xcdffe7a2, offset 4
[ 239.275044] asix 1-1.1:1.0 eth0: asix_rx_fixup() Data Header
synchronisation was lost, remaining 988
[ 239.284355] asix 1-1.1:1.0 eth0: asix_rx_fixup() Bad Header Length
0x1d36f59d, offset 4
[ 239.292541] asix 1-1.1:1.0 eth0: asix_rx_fixup() Bad Header Length
0xaef3c1e9, offset 4
[ 239.518996] asix 1-1.1:1.0 eth0: asix_rx_fixup() Data Header
synchronisation was lost, remaining 988
[ 239.528300] asix 1-1.1:1.0 eth0: asix_rx_fixup() Bad Header Length
0x2881912, offset 4
[ 239.536413] asix 1-1.1:1.0 eth0: asix_rx_fixup() Bad Header Length
0x5638f7e2, offset 4


And network throughput ends up being pretty bursty and slow with a
overall throughput of at best ~30kB/s.

Looking through the commits since the v4.1 kernel where we didn't see
this, I narrowed the regression down, and reverting the following two
commits seems to avoid the problem:

6a570814cd430fa5ef4f278e8046dcf12ee63f13 asix: Continue processing URB
if no RX netdev buffer
3f30b158eba5c604b6e0870027eef5d19fc9271d asix: On RX avoid creating
bad Ethernet frames

With these reverted, we don't see all the error messages, and we see
better ~1.1MB/s throughput (I've got a mouse plugged in, so I think
the usb host is only running at "full-speed" mode here).

This worries me some, as the patches seem to describe trying to fix
the issue they seem to cause, so I suspect a revert isn't the correct
solution, but am not sure why we're having such trouble and the patch
authors did not. I'd be happy to do further testing of patches if
folks have any ideas.

Originally Reported-by: Yongqin Liu <yongqin.liu@xxxxxxxxxx>

thanks
-john
Hi John,

Some ASIX chipsets span the Ethernet frame over consecutive URBs which
requires successful transfer of 2 URBs.

This means states of a previous URB influences the processing of the next
URB including a dropped URB (causes a discontinuity in the data stream). In
other words synchronisation of the in-band 32-bit header word needs to be
tracked between URBs. Some ASIX chipsets allow the in-band 32-bit header
word to be no longer fixed to the start of the URB buffer so it moves to any
position within the URB buffer.

I understand your point of suggesting it is a "regression" for your device
but the driver was broken for DUB-E100 C1 (small black USB device). So you
cannot revert the commits as this would break DUB-E100 C1 (small black USB
device).

6a570814cd430fa5ef4f278e8046dcf12ee63f13 asix: Continue processing URB
if no RX netdev buffer
This commit is necessary because it avoids a crash when netdev buffer failed
to be allocated for the 1st URB and the 2nd URB containing a spanned
Ethernet frame is processed. The crash happens because the 2nd URB assumed
that the netdev buffer had been allocated.

3f30b158eba5c604b6e0870027eef5d19fc9271d asix: On RX avoid creating
bad Ethernet frames
This commit is necessary to avoid sending bad Ethernet frames into the IP
stack during loss of synchronisation and to dropping good Ethernet frames.
This commit improves the synchronisation recovery mechanism of the in-band
32-bit header word.

The ASIX USB to Ethernet devices these commits were tested on where DUB-E100
C1 (small black USB device). Embedded ARM based systems were used where
memory resources can run out.
I don't have the chance to look into detail yet. But just a caution,
did you test on ARM 64-bit system or ARM 32-bit? I ask because HiKey
is an ARM 64-bit system. I suggest we should be careful on that. I saw
similar issues when transferring to a 64-bit system in other net
drivers.
We used 32-bit ARM and never tested on 64-bit ARM so I suggest that the commits need to be reviewed with 64-bit OS in mind.

Do you have any suggestion on this regard?
Try testing on a Linux PC x86 32-bit OS which has has a kernel containing my ASIX commits. This will help to confirm whether the failure is related to 32-bit or 64-bit OS. Then try with Linux PC x86 64-bit OS, this should fail otherwise it points to something specific in your ARM 64-bit platform.


It could be that for your USB to Ethernet device that the wrong
configuration settings have been used. In other words the ASIX driver is
flexible to support various variants of the ASIX chipsets. For example, does
your device support Ethernet frames spanning multiple URBs (multiple USB
transfers) ?
Would you please suggest how to find out this information? How can I
change my device's configuration settings to support spanning multiple
URBs?

So I doubt my commits are "broken" because we don't see your failures (not
tested your device). It is more likely that your ASIX device needs to be
properly identified and configured to be compatible with the ASIX driver. At
least, I suggest that is the best place to start your investigation.

Of course, your ASIX chipset might have a different behaviour for how the
in-band 32-bit header word operates so perhaps special treatment is needed
for your chipset ?

Please send to the mailing list the output of lsusb for your device so that
people can know the USB product ID and vendor ID for your device. This is
allows people to assist with the investigation. Do you have any links to
websites that sell your device ?
I experienced the same issue, working in the same project with John
actually. My USB ID:
Bus 001 Device 003: ID 0b95:772b ASIX Electronics Corp. AX88772B

Link to purchase: http://item.jd.com/1192582.html (by UGREEN)

John has his own device. And in our lab, there is a third kind of
device which uses the same AX88772B. All purchased from difference
sources with different brand names. And all can reproduce the same
issue.
The D-Link DUB-100 C1 also uses AX88772 (might be a different variant to UGREEN). Next step should be for someone to look at the commits for any 64-bit issues.


Are you using UDP or TCP connections ?
In my tests, I use iperf and transfer in TCP mode.
iperf works by creating a certain length of IP packet. In particular, iperf with IPv6 can cause IPv6 fragmentation to occur causing 2 Ethernet frames (fragmented) to be sent instead of the single original Ethernet frame. This is likely to increase the probability of Ethernet frames spanning URBs.

Try testing iperf with IPv4 and IPv6 using TCP to see whether the issue is worse or better. Also try reducing the length of the iperf IP packet to avoid IPv6 fragmentation eg. to fit within the MTU size.

Sorry, for my quick reply but I don't have time to support you full-time. I will respond to E-mails but it might take some days. Please include my E-mail address in the TO: field (I added it in my reply), thanks.

Best regards,
Dean

-Guodong


--
Dean Jenkins
Embedded Software Engineer
Linux Transportation Solutions
Mentor Embedded Software Division
Mentor Graphics (UK) Ltd.