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

From: Dean Jenkins
Date: Fri May 06 2016 - 12:54:49 EST

On 06/05/16 16:27, Andrew Lunn wrote:
In other words, the full-speed hub is restricting the USB to
Ethernet Adaptor to a 12Mbps (half-duplex) bandwidth to support
Ethernet 100Mbps (full-duplex) traffic. That is not going to work
very well because Ethernet frames (perhaps partial Ethernet frames)
need to be discarded within the USB link.
If that really is true, the design is broken. I would expect the
adaptor to reliably transfer whole frames over USB, and drop whole
frames from its receive queue when the USB is congested. TCP is also
going to see the USB bottleneck as just like any bottleneck in the
network and back off. So TCP streams should not cause major congestion
on the USB link.
The host's USB host controller polls the USB to Ethernet adaptor for more data. The USB to Ethernet adaptor cannot predict when the next poll request comes. The AX88772B can span Ethernet frames across multiple poll requests. This means it is possible get a partial Ethernet frame received in the USB host controller on one poll and it is assumed that the next poll (sometime in the near future) will get the remaining part of the Ethernet frame.

However, the USB to Ethernet adaptor does not contain an infinitely sized RX Ethernet buffer for the incoming Ethernet frames. I believe the USB to Ethernet adaptor is just a pipe and does not directly implement flow control for Ethernet frames so the RX buffer is going to overflow causing loss of whole Ethernet frames. I suspect the IP stack in the host computer implements flow control for Ethernet frames.

Because the AX88772B can span Ethernet frames across multiple poll requests there is a risk that the designers of the device could of implemented a solution to discard the remaining part of the Ethernet frame before the next poll arrives due to the RX buffer overflowing. I don't know the algorithm used in the AX88772B but there will be loss of data due to the mismatch in bandwidths. I agree that dropping whole Ethernet frames would be preferable to dropping partial Ethernet frames which would corrupt the data stream.

My suspicion is that the URB buffers are containing discontinues in the data stream because of lost data due to insufficient bandwidth on the USB link.

Going over a 12Mbps USB link should be no different
to hitting an old Ethernet hub which can only do 10/Half.
Not exactly, because USB is a transport link which is agnostic to the type of data that is flowing. It is up to the layers above USB to manage the data content.

In other words, the USB speed needs to be higher than the Ethernet speed to avoid mismatches in bandwidth.
Therefore please retest with a working high-speed USB hub or remove
the full-speed USB hub from the test environment and directly
connect the USB to Ethernet Adaptor to the root hub of the USB port.
Then repeat the tests to see whether anything improved.

In other words, you need to eliminate the dmesg messages saying "not
running at top speed; connect to a high speed hub".
I would also suggest testing with the Ethernet at 10/half. You should
be able to use Ethtool to set that up. Your USB and Ethernet bandwidth
become more equal. If you still see errors, it suggests a protocol
implementation error somewhere.
I agree with the suggestion but I hope USB high speed (480Mbps) operation was the intended environment rather than the useless USB full speed (12Mbps) operation.

Let's hope that not using the USB hub improves things.



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