On Wednesday, January 3, 2018 8:14:47 AM CET Jani Nikula wrote:
On Tue, 02 Jan 2018, Chris Wilson chris@chris-wilson.co.uk wrote:
Quoting Rodrigo Vivi (2018-01-02 19:12:18)
On Sun, Dec 31, 2017 at 10:34:54PM +0000, Stefan Brüns wrote:
edid = drm_get_edid(connector, i2c);if (!edid && !intel_gmbus_is_forced_bit(i2c)) {DRM_DEBUG_KMS("HDMI GMBUS EDID read failed, retry usingGPIO bit-banging\n"); + intel_gmbus_force_bit(i2c, true);
edid = drm_get_edid(connector, i2c);intel_gmbus_force_bit(i2c, false);}Approach seems fine for this case. I just wonder what would be the risks of forcing this bit and edid read when nothing is present on the other end?>
Should be no more risky than using GMBUS as the bit-banging is the underlying HW protocol; it should just be adding an extra delay to the disconnected probe. Offset against the chance that it fixes detection of borderline devices.
I would say that given the explanation above, the question is why not apply it universally? (Bonus points for including the explanation as comments.)
I'm wondering, is gmbus too fast for the adapters, does gmbus generally have different timing for the ack/nak as described in the commit message than bit banging, or are the adapters just plain buggy? Do we have any control over gmbus timings (don't have the time to peruse the bpsec just now)?
I have seen two different behaviours, one on the ~2009 GM965, the other on the ~2013 Haswell. The Haswell provides a 250..500ns hold time, the other does not.
There is a flag in the GMBUS0 register, GMBUS_HOLD_EXT, "300ns hold time, rsvd on Pineview". The driver does not set this flag. Possibly it is always set/ implied on the Haswell (which is post-Pineview), and should be set for anything older than Pineview.
There is another odd fact with the GM965, according to the register setting it should run at 100 kBit/s, but it only runs at 30 kBit/s. The Haswell runs at 100 kBit/s, as specified. As there are also idle periods ever 8 bytes, the EDID read takes 270ms before it fails.
The bitbanging code, running at 45 kBit/s (2 * 20us per clock cycle plus overhead) on the other hand just needs 58 ms, but keeps one core busy (udelay).
Unfortunately I currently have no older system than the Haswell available, so I can not check if the GMBUS_HOLD_EXT flag has any effect.
Kind regards,
Stefan