RE: [PATCH v2] usb: dwc2: call dwc2_is_controller_alive() under spinlock

[Paul Zimmerman]

> From: Felipe Balbi [mailto:balbi@xxxxxx]
> Sent: Wednesday, January 14, 2015 2:50 PM
> 
> On Wed, Jan 14, 2015 at 10:45:26PM +0000, Paul Zimmerman wrote:
> > > From: Felipe Balbi [mailto:balbi@xxxxxx]
> > > Sent: Wednesday, January 14, 2015 2:40 PM
> > >
> > > On Wed, Jan 14, 2015 at 10:28:54PM +0000, Paul Zimmerman wrote:
> > > > > From: Felipe Balbi [mailto:balbi@xxxxxx]
> > > > > Sent: Wednesday, January 14, 2015 1:46 PM
> > > > >
> > > > > On Wed, Jan 14, 2015 at 04:41:23PM -0500, Alan Stern wrote:
> > > > > > > > > This is really, really odd. Register accesses are atomic, so the lock
> > > > > > > > > isn't really doing anything. Besides, you're calling
> > > > > > > > > dwc2_is_controller_alive() from within the IRQ handler, so IRQs are
> > > > > > > > > already disabled.
> > > > > > > >
> > > > > > > > Spinlocks sometimes do more than you think.  For instance, here the
> > > > > > > > lock prevents the register access from happening while some other CPU
> > > > > > > > is holding the lock.  If a silicon quirk causes the register access to
> > > > > > > > interfere with other activities, this could be important.
> > > > > > >
> > > > > > > readl() (which is used by dwc2_is_controller_alive()) adds a memory
> > > > > > > barrier to the register accesses, that should force all register
> > > > > > > accesses the be correctly ordered.
> > > > > >
> > > > > > Memory barriers will order accesses that are all made on the same CPU
> > > > > > with respect to each other.  They do not order these accesses against
> > > > > > accesses made from another CPU -- that's why we have spinlocks.  :-)
> > > > >
> > > > > a fair point :-) The register is still read-only, so that shouldn't
> > > > > matter either :-)
> > > > >
> > > > > > >  I fail to see how a silicon quirk
> > > > > > > could cause this and if, indeed, it does, I'd be more comfortable with a
> > > > > > > proper STARS tickect number from synopsys :-s
> > > > > >
> > > > > > Maybe accessing this register somehow resets something else.  I don't
> > > > > > know.  It seems unlikely, but at least it explains how adding a
> > > > > > spinlock could fix the problem.
> > > > >
> > > > > I would really need Paul (or someone at Synopsys) to confirm this
> > > > > somehow. Maybe it has something to do with how the register is
> > > > > implemented, dunno.
> > > > >
> > > > > Paul, do you have any idea what could cause this ? Could the HW into
> > > > > some weird state if we read GSNPSID at random locations or when data is
> > > > > being transferred, or anything like that ?
> > > >
> > > > Only thing I can think of is that there is some silicon bug in Robert's
> > > > platform. But I am not aware of any STARs that mention accesses to the
> > > > GSNPSID register as being problematic.
> > > >
> > > > Funny thing is, this code has been basically the same since at least
> > > > November 2013. So I think some other recent change must have modified
> > > > the timing of the register accesses, or something like that. But that's
> > > > just handwaving, really.
> > >
> > > Alright, I'll apply this patch but for 3.20 with a stable tag as I have
> > > already sent my last pull request to Greg. Unless someone has a really
> > > big complaint about doing things as such.
> >
> > It should go to 3.19-rc shouldn't it? It's a fix, and Robert's platform
> > is broken without it, IIUC.
> 
> It can also be categorized as "has-never-worked-before" before the code
> has been like this forever. Since we don't really have a git bisect
> result pointing to a commit that went in v3.19 merge window, I'm not
> sure how I can convince myself that this absolutely needs to be in
> v3.19.
> 
> At a minimum, I need a proper bisection with a proper commit being
> blamed (even if it's a commit from months ago). From my point of view,
> debugging of this "regression" has not been finalized and we're just
> "assuming" it's caused by GSNPSID because moving that inside the
> spin_lock seems to fix the problem.

On further investigation, I was wrong about "this code has been
basically the same since at least November 2013". Prior to commit
db8178c33db "usb: dwc2: Update common interrupt handler to call gadget
interrupt handler" from November 2014, the gadget interrupt handler
did not read from the GSNPSID register.

So likely the bug in Robert's hardware has been there all along, and
that commit just caused it to manifest itself.

-- 
Paul

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/