[U-Boot] [PATCH 1/1] dwc2 USB controller hangs with lan78xx

[Alexander Graf]

On 06/26/2018 02:47 PM, Peter Robinson wrote:
>>>> This bug is the combination of dwc2 USB controller and lan78xx
>>>> USB ethernet controller, which is the combination in use on
>>>> the Raspberry Pi Model 3 B+.
>>>>
>>>> When the host attempts to receive a packet, but a packet has not
>>>> arrived, the lan78xx controller responds by setting BIR
>>>> (Bulk-In Empty Response) to NAK. Unfortunately, this hangs
>>>> the USB controller and requires the USB controller to
>>>> be reset.
>>>>
>>>> The fix proposed is to have the lan78xx controller respond
>>>> by setting BIR to ZLP.
>>>>
>>>> Signed-off-by: Andrew Thomas <andrew.thomas at oracle.com>
>>> Tested-by: Peter Robinson <pbrobinson at gmail.com>
>>>
>>> Tested on the RPi 3B+ and certainly improves this situation a number
>>> of Fedora users have seen.
>>
>>
>> What exactly have you tested?
>>
>> Even with this patch, I am not reliably to reliably boot into grub. It
>> almost seems as if the packet buffer keeps getting overwritten by newer
>> packets so that by the time we process the old ones, the ones we wanted to
>> see are gone.
> Booting to grub and then a local system, it's seemed improved in terms
> of stability for me and a few other users, at least was more
> consistent in getting to grub without "Rx: failed to receive: -5" but
> then from the various testing different Fedora users have seen prior
> to this patch it seems that the issue is very
> network/cable/environment specific. I'm fairly certain that this patch
> doesn't fix the problem but it at least appears from current reports
> to improve the situation for some users.

Improve, yes, but not solve it to a point where it's fully usable. And I 
don't think we have a full grasp on what's broken yet.

What I'm seeing here is that booting grub stalls within early bootup 
(where it loads modules) until I press any key on serial. Pressing a key 
on a USB keyboard does not help.

Given that there are no interrupts enabled on the RPi at this point, I 
can only think of caches as the culprit for this breakage. But why did 
it work with the normal 3B then?


Alex