[PATCH] drm/bridge: sii902x: fix get edid may fail

[Daniel Vetter]

On Mon, Jan 23, 2017 at 03:06:36PM +0100, Andrea Merello wrote:
> On Mon, Jan 23, 2017 at 1:36 PM, Boris Brezillon
> <boris.brezillon at free-electrons.com> wrote:
> > On Mon, 23 Jan 2017 13:12:12 +0100
> > Andrea Merello <andrea.merello at gmail.com> wrote:
> >
> >> On Mon, Jan 23, 2017 at 12:32 PM, Jani Nikula
> >> <jani.nikula at linux.intel.com> wrote:
> >> > On Mon, 23 Jan 2017, Boris Brezillon <boris.brezillon at free-electrons.com> wrote:
> >> >> Hi Andrea,
> >> >>
> >> >> On Mon, 23 Jan 2017 11:00:02 +0100
> >> >> Andrea Merello <andrea.merello at gmail.com> wrote:
> >> >>
> >> >>> From: Andrea Merello <andrea.merello at gmail.com>
> >> >>>
> >> >>> The standard DRM function to get the edid from the i2c bus performs
> >> >>> (at least) two transfers.
> >> >>>
> >> >>> By experiments it seems that the sii9022a have problems with the
> >> >>> 2nd I2C start, at least unless a wait is introduced detween the
> >> >>
> >> >>                                                     ^ between
> >> >>
> >> >>> two transfers.
> >> >>>
> >> >>> So we perform one single I2C transfer, and if the transfer must be
> >> >>> split, then we introduce a delay.
> >> >>
> >> >> That's not exactly what this patch does: you're introducing a delay
> >> >> between each retry. So, if the transceiver really requires a delay
> >> >> between each transfer, you'll have to retry at least once on the 2nd
> >> >> transfer.
> >> >>
> >> >> I guess a better solution would be to add a delay even in case of
> >> >> success, or maybe modify drm_do_get_edid() to optionally wait for a
> >> >> specified time between each transfer.
> >> >
> >> > Is the problem related specifically to EDID reads, or generally to I2C
> >> > transfers? Perhaps this should be fixed at the adapter master_xfer layer
> >> > instead? Does the master_xfer perhaps return -EAGAIN, have you looked
> >> > into i2c_adapter timeout member? (See __i2c_transfer in i2c-core.c.)
> >> >
> >> > The intention of drm_do_get_edid() is to facilitate *alternative*
> >> > methods to doing I2C, not to workaround quirks like this.
> >
> > Yes, I tend to agree with that.
> >
> >>
> >> This problem seems not related to the I2C master itself. I saw the I2C
> >> master to TX correctly by looking at the I2C bus before the sii9022a
> >> with the scope, and I saw the same I2C transfer to be corrupted after
> >> the sii9022a; actually the sii9022a seems to gate the bus damaging the
> >> I2C start of the 2nd transfer, unless we place this delay.
> >>
> >> This happened with two different I2C master (the one on the SoC as
> >> well as another different one implemented on FPGA).
> >
> > Actually, I faced the same problem on an at91-based platform, and came
> > up with pretty much the same fix (add a delay after each transfer),
> > except mine was uglier and I never took the time to clean it up and
> > submit it.
> >
> >>
> >> So IMHO this quirk should be done in the sii902x dirver. Said that, I
> >> admit that I've not looked at i2c_adapter timeout member or other
> >> details in upper layers.
> >
> > Well, let's sum up the situation: the sii902x device is acting both as
> > a regular i2c device and as an i2c 'pass-through' device. In the
> > pass-through mode, the si902x has a new constraint: the i2c transfers
> > have to be separated by an idle period.
> >
> > Maybe we should expose the pass-through mode as separate i2c adapter,
> > that would be registered when entering pass-through mode and
> > de-registered when leaving this mode.
> > This way, we can implement the ->master_xfer() as we need (add a delay
> > after each xfer), and still use the generic drm_get_edid().
> >
> > What do you think?
> 
> IMHO this could be a good way to go.. Although honestly the current
> quirk doesn't seem so much bad to me :)

I agree with Jani that using drm_do_get_edid is the wrong thing to do
here, it's really for cases where you have a hw edid-fetch engine that
entirely hides the underlying i2c bus. If you do have the i2c bus exposed,
then the right place to fix it is at the i2c_adapter level. There's a pile
of userspace tools that also probe the ddc i2c bus (there's not just an
edid there), and if you don't fix this at the i2c_adapter level those
would still be broken.

Since this seems unclear, can you pls also do a patch for the kerneldoc
for the drm_do_get_edid function, so it's clear when and when not this
should be used?

Thanks, Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch