Re: [PATCH] usb: musb: musb_host: Introduce postponed URB giveback

[Bin Liu]

On Fri, Apr 28, 2017 at 04:15:09PM +0300, Matwey V. Kornilov wrote:
> 2017-04-28 15:43 GMT+03:00 Bin Liu <b-liu@xxxxxx>:
> > On Fri, Apr 28, 2017 at 03:13:55PM +0300, Matwey V. Kornilov wrote:
> >>  which i
> >>
> >> 2017-04-28 14:58 GMT+03:00 Bin Liu <b-liu@xxxxxx>:
> >> > On Fri, Apr 28, 2017 at 10:04:30AM +0300, Matwey V. Kornilov wrote:
> >> >> 2017-04-27 20:13 GMT+03:00 Bin Liu <b-liu@xxxxxx>:
> >> >> > On Thu, Apr 27, 2017 at 07:26:31PM +0300, Matwey V. Kornilov wrote:
> >> >> >> 2017-04-27 18:35 GMT+03:00 Bin Liu <b-liu@xxxxxx>:
> >> >> >> > Hi Matwey,
> >> >> >> >
> >> >> >> > On Thu, Apr 27, 2017 at 01:20:33PM +0300, Matwey V. Kornilov wrote:
> >> >> >> >> This commit changes the order of actions undertaken in
> >> >> >> >> musb_advance_schedule() in order to overcome issue with broken
> >> >> >> >> isochronous transfer [1].
> >> >> >> >>
> >> >> >> >> There is no harm to split musb_giveback into two pieces.  The first
> >> >> >> >> unlinks finished urb, the second givebacks it. The issue here that
> >> >> >> >> givebacking may be quite time-consuming due to urb->complete() call.
> >> >> >> >> As it happens in case of pwc-driven web cameras. It may take about 0.5
> >> >> >> >> ms to call __musb_giveback() that calls urb->callback() internally.
> >> >> >> >> Under specific circumstances setting MUSB_RXCSR_H_REQPKT in subsequent
> >> >> >> >> musb_start_urb() for the next urb will be too late to produce physical
> >> >> >> >> IN packet. Since auto req is not used by this module the exchange
> >> >> >> >> would be as the following:
> >> >> >> >>
> >> >> >> >> [        ]   7.220456 d=  0.000997 [182   +  3.667] [  3] IN   : 4.5
> >> >> >> >> [    T   ]   7.220459 d=  0.000003 [182   +  7.000] [800] DATA0: [skipped]
> >> >> >> >> [        ]   7.222456 d=  0.001997 [184   +  3.667] [  3] IN   : 4.5
> >> >> >> >> [        ]   7.222459 d=  0.000003 [184   +  7.000] [  3] DATA0: 00 00
> >> >> >> >>
> >> >> >> >> It is known that missed IN in isochronous mode makes some
> >> >> >> >> perepherial broken. For instance, pwc-driven or uvc-driven
> >> >> >> >> web cameras.
> >> >> >> >> In order to workaround this issue we postpone calling
> >> >> >> >> urb->callback() after setting MUSB_RXCSR_H_REQPKT for the
> >> >> >> >> next urb if there is the next urb pending in queue.
> >> >> >> >>
> >> >> >> >> [1] https://www.spinics.net/lists/linux-usb/msg145747.html
> >> >> >> >>
> >> >> >> >> Fixes: f551e1352983 ("Revert "usb: musb: musb_host: Enable HCD_BH flag to handle urb return in bottom half"")
> >> >> >> >> Signed-off-by: Matwey V. Kornilov <matwey@xxxxxxxxxx>
> >> >> >> >
> >> >> >> > Thanks for the effort of working on this long standing issue, I know you
> >> >> >> > have spent alot of time on it, but what I am thinking is instead of
> >> >> >> > workaround the problem we need to understand the root cause - why
> >> >> >> > uvc-video takes longer to exec the urb callback, why only am335x
> >> >> >> > reported this issue. This is on my backlog, just seems never got time
> >> >> >> > for it...
> >> >> >>
> >> >> >> Have you tried other SoCs with Invetra MUSB?
> >> >> >
> >> >> > That is the plan, I got an A20 board, but haven't bring it up yet.
> >> >> >
> >> >> >>
> >> >> >> >
> >> >> >> > Ideally MUSB driver should be just using HCD_BH flag and let the core to
> >> >> >> > handle the urb callback, regardless the usb transfer types.
> >> >> >>
> >> >> >> I think the only reason why everything worked before with HCD_BH is
> >> >> >> that execution of urb->callback() was placed after musb_start(). The
> >> >> >> order of operations matters.
> >> >> >> However, you said that something was also wrong with HCD_BH and other
> >> >> >> peripherals.
> >> >> >
> >> >> > HCD_BH flag cause some issues which are docummented in the commit log of
> >> >> > f551e1352983.
> >> >> > But even with HCD_BH flag, it didn't work for uvc webcams, it still misses
> >> >> > IN tokens. It might helps pwc webcams though.
> >> >>
> >> >> pwc webcams work with HCD_BH fine indeed.
> >> >
> >> > yeah, this is what you told long time ago.
> >> >
> >> >>
> >> >> >
> >> >> >> > The MUSB drivers are already messy and complicated enough for
> >> >> >> > maintenance, I'd like to understand the root cause of the delay first
> >> >> >> > before decide how to solve the issue.
> >> >> >> >
> >> >> >>
> >> >> >> I feel from playing with OpenVizsla that REQPKT should be set well in
> >> >> >> advance. So, time window to set the flag is actually smaller than 1
> >> >> >> ms.
> >> >> >> urb->callback() is never takes longer than 0.4 ms for pwc driver, but
> >> >> >> INs are skipped.
> >> >> >
> >> >> > Setting REQPKT in advance might be the solution, but I'd like to
> >> >> > understand why only Isoch transfer shows such issue, and why only am335x
> >> >> > reports this issue. The later concerns me more if this would be a
> >> >> > system issue not only in usb module.
> >> >>
> >> >> 0.4 ms is about 100000 CPU cycles given that CPU is running at 275Mhz
> >> >> (which is the lowest cpufreq). Long time ago, I run pwc webcam with
> >> >> SIS Vortex86 at 200Mhz It worked fine. I would not say that it is
> >> >> extraordinary value.
> >> >> Do you think that somewhere CPU cycles are wasted globally for some reason?
> >> >
> >> > Depends on how to interpret 'wasted', my understanding the issue is the
> >> > core urb giveback holds a spinlock and in uvc case the giveback takes
> >> > longer to finish (I guess the same in pwc case), so the musb driver
> >> > doesn't get a chance to re-program the controller on time, which causes
> >> > missing issuingIN tokens.
> >> >
> >> > The questions are, why uvc takes longer to run urb giveback (which holds
> >> > a spinlock), and is it am335x specific issue...
> >>
> >> It is not clear how does it hold a spinlock on a BeagleBone which is
> >> single-core system.
> >
> > Sorry, I have to take it back, urb giveback doesn't hold a spinlock, but
> > disables irq. It has been a while (a year?) since the last time I looked
> > this issue. Please see the call below flow.
> >
> > musb_giveback() -->
> >   usb_hcd_giveback_urb() -->
> >     __usb_hcd_giveback_urb() --> # it gets here regardless HCD_BH flag
> >       1765         local_irq_save(flags);
> >       1766         urb->complete(urb);
> >       1767         local_irq_restore(flags);
> >
> > so musb driver only gets a chance to re-program the controller after
> > line 1766 returns, which is the urb callback in the class driver (uvc in
> > this case). If urb->complete() takes too long, the controller will miss
> > the IN tokens.
> >
> > HCD_BH flag could help the situation only if urb->complete() doesn't
> > take that long.
> 
> Sure, I think that the question is why urb->complete() itself takes so
> long only (?) at am335x.

This is what I meant. I haven't checked other platforms yet, but at
least it seems the issue only reported on am335x as far as I am aware.

> 
> >
> >> I mean if it is waiting for blocked lock it will be waiting forever,
> >> because it is in interrupt context on single core system.
> >
> > Hope my message above explains it.
> >
> >> Here it is waiting for finite time amount. It is also quite unlikely
> >> that spinlocks are implemented inefficiently for arm architecture.
> >> So, I tested with CONFIG_SMP=y which is default in my distro kernel.
> >> Do you think it is worth to try CONFIG_SMP=n or have you already tried
> >> UP kernel configuration?
> >
> > CONFIG_SMP is irrelevant here. BTY, because of my laziness, I use
> > CONFIG_SMP in most times, if not all.
> >
> > Regards,
> > -Bin.
> >
> 
> 
> 
> -- 
> With best regards,
> Matwey V. Kornilov.
> Sternberg Astronomical Institute, Lomonosov Moscow State University, Russia
> 119234, Moscow, Universitetsky pr-k 13, +7 (495) 9392382