On Sun, Nov 10, 2019 at 10:34:02AM -0500, Alan Stern wrote:
> On Sun, 10 Nov 2019, Oliver Neukum wrote:
>
> > Am Freitag, den 08.11.2019, 10:35 -0500 schrieb Alan Stern:
> > > On Fri, 8 Nov 2019, Greg Kroah-Hartman wrote:
> > >
> > > > On Thu, Nov 07, 2019 at 12:32:45PM +0100, Oliver Neukum wrote:
> > > > > Am Dienstag, den 05.11.2019, 17:38 +0100 schrieb Greg Kroah-Hartman:
> > > > > > > > Given this information, perhaps you will decide that the revert is
> > > > > > > > worthwhile.
> > > > > > >
> > > > > > > Damned if I do, damned if I do not.
> > > > > > > Check for usbip and special case it?
> > > > > >
> > > > > > We should be able to do that in the host controller driver for usbip,
> > > > > > right? What is the symptom if you use a UAS device with usbip and this
> > > > > > commit?
> > > > >
> > > > > Yes, that patch should then also be applied. Then it will work.
> > > > > Without it, commands will fail, as transfers will end prematurely.
> > > >
> > > > Ok, I'm confused now. I just checked, and I really have no idea what
> > > > needs to be backported anymore. 3ae62a42090f ("UAS: fix alignment of
> > > > scatter/gather segments") was backported to all of the stable kernels,
> > > > and now we reverted it.
> > > >
> > > > So what else needs to be done here?
> > >
> > > In one sense, nothing needs to be done. 3ae62a42090f was intended to
> > > fix a long-standing problem with USBIP, but people reported a
> >
> > OK, now I am a bit confused. AFAICT 3ae62a42090f actually did fix the
> > issue. So if you simply revert it, the issue will reappear.
>
> Correct. I think.
>
> > > regression in performance. (Admittedly, the report was about the
> > > correponding change to usb-storage, not the change to uas, but it's
> > > reasonable to think the effect would be the same.) So in line with the
> >
> > Yes.
> >
> > > no-regressions policy, we only need to revert the commit -- which you
> > > have already done.
> >
> > But that breaks UAS over USBIP, doesn't it?
>
> It was already broken to start with. The attempted fix caused a
> regression, so the fix must be reverted.
>
> > > On the other hand, the long-standing problem in USBIP can be fixed by
> > > back-porting commit ea44d190764b. But since that commit isn't a
> > > bug-fix (and since it's rather large), you may question whether it is
> > > appropriate for the -stable kernel series.
> >
> > It certainly is large. But without it UAS won't work over USBIP, will
> > it? I think that is the central question we need to answer here.
>
> You are right. If Greg is okay with porting ea44d190764b to the stable
> kernels, I won't object.
Ok, I've backported this to 4.14.y, 4.19.y, and 5.3.y. As superspeed
support was not added to vhci until the 4.13 release, I don't think that
UAS will be an issue on those older kernels.
thanks,
greg k-h
