[PATCH v2 1/3] imx-drm: Add mx6 hdmi transmitter support
Russell King - ARM Linux
linux at arm.linux.org.uk
Wed Oct 16 16:27:50 EDT 2013
On Wed, Oct 16, 2013 at 12:37:42PM -0700, Troy Kisky wrote:
> On 10/16/2013 10:03 AM, Russell King - ARM Linux wrote:
>> On Tue, Oct 15, 2013 at 10:17:07AM -0300, Fabio Estevam wrote:
>>> On Tue, Oct 15, 2013 at 10:10 AM, Russell King - ARM Linux
>>> <linux at arm.linux.org.uk> wrote:
>>>> Another point on patch 1. Sorry, I don't have patch 1 to reply to, it
>>>> seems it was deleted from linux-arm-kernel's moderation queue.
>>>>
>>>> drm_mode_connector_attach_encoder() is called too early, before the
>>>> base.id field in the encoder has been initialised. This causes the
>>>> connectors encoder array to be empty, and userspace KMS to fail.
>>>>
>>>> There's also bugs in the CSC setting too - it runs off the end of the
>>>> array and gcc warns about this. The code was clearly wrong.
>>>>
>>>> You may wish to combine this patch with patch 1 to sort all that out.
>>>> For the patch below:
>>>>
>>>> Signed-off-by: Russell King <rmk+kernel at arm.linux.org.uk>
>>>> Tested-by: Russell King <rmk+kernel at arm.linux.org.uk>
>>> Thanks, Russell.
>>>
>>> Will submit v3 when I am back to the office.
>> Okay, I still have a problem with HDMI: I have a magenta vertical line
>> down the left hand side of the frame, the displayed frame is shifted
>> right by the width of that line and the right hand side is missing some
>> pixels.
>>
>> First off, the hsync position programmed into the HDMI registers appears
>> to be wrong.
>>
>> I'm at a loss why imx-hdmi is obfuscated with a conversion from a
>> drm_display_mode structure to a fb_videomode. This adds additional
>> confusion and additional opportunities for bugs; this is probably
>> exactly why the hsync position is wrong.
>>
>> In CEA-861-B for 720p @60Hz:
>>
>> DE: ^^^^__________^^^^^^^
>> HS: _______^^^___________
>> ^ ^ ^
>> | | 220 clocks
>> | 40 clocks
>> 110 clocks
>>
>> The IMX6 manual says HSYINCINDELAY0 is "Hsync active edge delay. Integer
>> number of pixel clock cycles from de non-active edge". So, this should be
>> 110. Yet it ends up being programmed as 220, leading to a magenta vertical
>> bar down the left hand side of the display.
>>
>> Now, if you look at Documentation/fb/framebuffer.txt, in this case, the
>> right margin should be 110 clocks, hsync len should be 40 and the left
>> margin should be 220 clocks.
>>
>> However, this is not what your conversion to a fb_videomode does. It
>> reverses the left and right margin. A similar confusion also exists
>> in the conversion of the upper/lower margins too.
>>
>> The DRM model is this (for 720p @60Hz):
>>
>> 0 1280 1390 1430 1650
>> |===============================|------------|---|----------|
>> ^ ^ ^ ^ ^
>> start hdisplay hsync_start hsync_end htotal
>>
>> The fb model is the same as the above but rotated:
>>
>> left margin displayed right margin hsync_len
>> |----------|===============================|------------|---|
>>
>> So, the left margin is the bit between hsync_end and htotal, and the
>> right margin is the bit between hdisplay and hsync_start. Exactly
>> the same analysis applies to the upper/lower margins.
>>
>> What I suggest is that the use of fb_videomode is entirely killed off
>> in this driver to remove this layer of confusion and instead the DRM
>> model is stuck to within this DRM driver.
>>
>> Now on to the next problem. HSYNC/VSYNC polarity.
>>
>> So, this is the mode which is set:
>>
>> 1280x720 (0x41) 74.2MHz +HSync +VSync *current +preferred
>> h: width 1280 start 1390 end 1430 total 1650 skew 0 clock 45.0KHz
>> v: height 720 start 725 end 730 total 750 clock 60.0Hz
>>
>> Note the positive HSync and VSync polarity - this is correct, backed
>> up by CEA-861-B.
>>
>> The IPU DI0 is configured thusly in its general control register:
>> 0x2640000 = 0x200006, which is what is set when the requested mode
>> has DRM_MODE_FLAG_PHSYNC and DRM_MODE_FLAG_PVSYNC flags set.
>>
>> However, if we look at the HDMI config: 0x121000 = 0x18. Active low
>> hsync/vsync. This is quite obvious when you look at the code -
>> convert_to_video_timing() does *nothing* at all when converting the
>> sync polarities, which means hdmi_av_composer() doesn't program them
>> appropriately. And yes, poking 0x78 into this register finally fixes
>> the problem.
>>
>> Yet another reason why this silly conversion from one structure form
>> to another is a Very Bad Idea. Until I found this, I was merely going
>> to send a patch to sort out convert_to_video_timing(), but quite frankly
>> I'm going to kill this thing off right now.
>>
>> Another thing:
>>
>> static int imx_hdmi_setup(struct imx_hdmi *hdmi, struct drm_display_mode *mode)
>> {
>> int ret;
>> convert_to_video_timing(&hdmi->fb_mode, mode);
>>
>> hdmi_disable_overflow_interrupts(hdmi);
>> hdmi->vic = 6;
>>
>> It's quite wrong to force every video mode set to be CEA mode 6. IIRC,
>> There is a function in DRM which will tell you the CEA mode number.
>> Again, I'll fix this in my patch rewriting this bit of the driver to
>> be correct.
>>
>> I'm also suspicious of the "HDMI Initialization Step" comments, because
>> they make no sense:
>>
>> /* HDMI Initialization Step B.4 */
>> static void imx_hdmi_enable_video_path(struct imx_hdmi *hdmi)
>> {
>>
>> yet:
>>
>> /* HDMI Initialization Step B.3 */
>> imx_hdmi_enable_video_path(hdmi);
>>
>> One's left wondering whether Step B.3 really is to just call a function
>> with a particular name, but B.4 is to actually do something with the
>> hardware. I'm quite sure that if this is a documented procedure, that
>> it doesn't say that and these comments are wrong (and probably the code
>> too.)
>>
>> Even after all this, I still haven't got rid of that magenta line - in
>> as far as I can tell, nothing has changed as a result of any of these
>> (although reading back the register values, they're now much better.)
>> What I do find is if I poke 0x78 back into the INVIDCONF register
>> (which already contains 0x78) the magenta line disappears.
>>
>> I'm beginning to suspect that there's some ordering dependency that
>> isn't being satisfied - but as the i.MX6Solo/DualLite manual doesn't
>> seem to contain any of these details, I'm running out of ideas.
>>
>> _______________________________________________
>>
> It sound to me like you hit errata "ERR004308 HDMI: 8000504668—The
> arithmetic unit may get wrong video
> timing values although the FC_* registers hold correct values"
>
>
> Also, checkout ERR005173, it says
> "Workarounds:
> The programming flow should be:
> 1. Program all the controller registers including the frame composer
> registers.
> 2. Assert software resets
> 3. Write (3 or 4 times) in FC_INVIDCONF the final value (this will make
> sure that the update
> pulse for the fc_arithlogicunit_div that will update the
> fc_arithlogicunit_div units is generated)
> 4. If frame composer packet queue overflow still occurs, then repeat
> steps 2 and 3"
Hi Troy,
I think it's implementing that - we have this code:
/* Workaround to clear the overflow condition */
static void imx_hdmi_clear_overflow(struct imx_hdmi *hdmi)
{
int count;
u8 val;
val = hdmi_readb(hdmi, HDMI_FC_INVIDCONF);
for (count = 0; count < 5; count++)
hdmi_writeb(hdmi, val, HDMI_FC_INVIDCONF);
/* TMDS software reset */
hdmi_writeb(hdmi, (u8)~HDMI_MC_SWRSTZ_TMDSSWRST_REQ, HDMI_MC_SWRSTZ);
}
However, there's no detection of whether a FC packet queue overflow
occurs: yes, sure, the interrupts are unmasked, but the irq handler
does nothing with it.
However, there is something of interest here. Is there any ordering
requirement between the setup of the IPU and HDMI?
At present, the DRM core does this order:
- encoders: prepare callback
- crtc: prepare callback
- crtc: mode_set
- encoders: mode_set
- crtc: commit callback
- encoders: commit callback
This translates to:
- disable HDMI
- disable IPU
- configure IPU
- configure *and* enable HDMI output
- enable IPU
- reconfigure and enable HDMI output
Now, what I've just discovered is that killing off the 2nd reconfiguration
and enable in the HDMI commit callback gets rid of the magenta line, but
now I'm left with a picture which occasionally blanks. I've tried using
devmem2 to write the INVIDCONF register three times and then hit the
reset bit (./devmem2 0x124002 b 0xfd) but that doesn't cure this.
However, after X goes into DPMS and wakes up, it's fine.
If I comment out the configure and enable in the mode_set, and leave
the one in the commit callback, I get the magenta line back.
I think there's a clue in the errata document:
Each time one writes to some FC registers, and depending on the clock
relation of sfr clk and tmds clk ...
That seems to imply that there's supposed to be various clocks running
here while these registers are accessed. If the IPU is all shut down,
and the HDMI itself is disabled, there's no TMDS clock. That could
explain why the video timings appear to be wrong.
Oh my, and while reading through this, I find:
spin_lock_irqsave(&hdmi->irq_lock, flags);
ret = clk_prepare_enable(hdmi->iahb_clk);
if (ret)
return ret;
That is sooo wrong. Not only is that returning leaving a spin lock in
the locked state, but clk_prepare_enable() _can_ sleep. And then I've
turned my attention to finding out what exactly irq_lock protects?
Would it be preventing the IRQ handler from doing stuff? Well, the
IRQ handler doesn't take the lock at all, so that can't be it. The
only function which takes that lock is imx_hdmi_fb_registered(), which
is only called at driver probe time. What use is this lock? What's
it protecting? As far as I can see, it's doing nothing apart from
disabling interrupts. And that doesn't work if you're running on a
dual-CPU or quad-CPU IMX6.
This driver just seems to be wrong in soo many ways, it's really not
funny.
So, here's the thing: is there any documentation around (apart from
the applications processor reference manual) which gives guidance as
to how the IPU and HDMI are supposed to be configured on the IMX6, or
is that all secret NDA stuff which only Freescale people are allowed
to know? Why? At this point I'm beginning to think that rewriting
at least the HDMI backend from scratch is probably the best solution.
More information about the linux-arm-kernel
mailing list