Bug#808953: linux: xhci regression in backported patch

[Ron <ron@debian.org>]

Source: linux
Version: 3.16.7-ckt20-1
Severity: important

Hi,

While testing the code from here: http://www.bitbabbler.org I ran into a
problem on recently updated Jessie systems where devices plugged into an
XHCI port would always time out between transfers when the transfer size
was larger than 16kB.  And since it wasn't doing that a few weeks ago, I
went digging :)

All the following was tested on amd64.

I can confirm that 3.16.7-ckt17-1 does not have this problem and it's a
newly introduced regression that's first seen with the ckt20 patch set.
I can also confirm that 4.2.0-0.bpo.1 from jessie-backports does not
have this problem.

The cause appears to be a botched backport of
e210c422b6fdd2dc123bedc588f399aefd8bf9de (from Linus' repo) because I can
also confirm that a locally built kernel from the 3.16.7-ckt20-1 source
with only that patch reverted also does not have this problem.

If it's really important that we have that patch for some reason, then
it looks like (at least) 40a3b775f49c2784c96b19170fd2478e5e5511a1 would
also need to be backported too, and maybe some other changes that effect
this as well.  I haven't analysed that part of it in more detail, but
it's clear this patch needs something more than just being massaged to
fit the context to actually work.

This is a fairly serious regression that makes USB3 ports (even with
USB2 devices in them) pretty close to useless you're heavily into having
10 second latency between every transfer that occurs - so we should
probably revert this unless someone who knows that code well can review
a better backport of it to ensure there won't be unintended ill effects.

  Cheers,
  Ron