Hello,
while playing around with FIMD enabled, I noticed that when first using the device a zero division was triggered in fimd_calc_clkdiv(). I remembered that I had a similar issue some time ago.
I added a stub fimd_atomic_check() which shows that vrefresh is zero when fimd_calc_clkdiv() is called.
[ 164.059361] [drm:exynos_plane_mode_set] plane : offset_x/y(0,0), width/height(1366,768) [ 164.067175] [drm:fimd_atomic_check] xres=1366, yres=768, refresh=0, intl=0 [ 164.074198] [drm:drm_atomic_helper_check_planes] [CRTC:24:crtc-0] atomic driver check failed
I went back to the git log and noticed that some time ago in 50bbfbffa5c894def440ce8157dfe53e60960d35 the fimd_mode_fixup() call was removed.
I'm now wondering where exactly vrefresh is set to a sane value. As far as I can see of_get_videomode() is used to fetch the video mode from the DT, and drm_display_mode_from_videomode() is used to convert it. However vrefresh is nowhere set.
So is something broken here, or am I missing something?
With best wishes, Tobias
On Sat, Apr 30, 2016 at 5:37 PM, Tobias Jakobi tjakobi@math.uni-bielefeld.de wrote:
Hello,
while playing around with FIMD enabled, I noticed that when first using the device a zero division was triggered in fimd_calc_clkdiv(). I remembered that I had a similar issue some time ago.
I added a stub fimd_atomic_check() which shows that vrefresh is zero when fimd_calc_clkdiv() is called.
[ 164.059361] [drm:exynos_plane_mode_set] plane : offset_x/y(0,0), width/height(1366,768) [ 164.067175] [drm:fimd_atomic_check] xres=1366, yres=768, refresh=0, intl=0 [ 164.074198] [drm:drm_atomic_helper_check_planes] [CRTC:24:crtc-0] atomic driver check failed
I went back to the git log and noticed that some time ago in 50bbfbffa5c894def440ce8157dfe53e60960d35 the fimd_mode_fixup() call was removed.
I'm now wondering where exactly vrefresh is set to a sane value. As far as I can see of_get_videomode() is used to fetch the video mode from the DT, and drm_display_mode_from_videomode() is used to convert it. However vrefresh is nowhere set.
So is something broken here, or am I missing something?
There is no guarantee that vrefresh is actually set. I think the only way to reliably get it is to call drm_mode_vrefresh().
Alex
Hello Alex,
Alex Deucher wrote:
On Sat, Apr 30, 2016 at 5:37 PM, Tobias Jakobi tjakobi@math.uni-bielefeld.de wrote:
Hello,
while playing around with FIMD enabled, I noticed that when first using the device a zero division was triggered in fimd_calc_clkdiv(). I remembered that I had a similar issue some time ago.
I added a stub fimd_atomic_check() which shows that vrefresh is zero when fimd_calc_clkdiv() is called.
[ 164.059361] [drm:exynos_plane_mode_set] plane : offset_x/y(0,0), width/height(1366,768) [ 164.067175] [drm:fimd_atomic_check] xres=1366, yres=768, refresh=0, intl=0 [ 164.074198] [drm:drm_atomic_helper_check_planes] [CRTC:24:crtc-0] atomic driver check failed
I went back to the git log and noticed that some time ago in 50bbfbffa5c894def440ce8157dfe53e60960d35 the fimd_mode_fixup() call was removed.
I'm now wondering where exactly vrefresh is set to a sane value. As far as I can see of_get_videomode() is used to fetch the video mode from the DT, and drm_display_mode_from_videomode() is used to convert it. However vrefresh is nowhere set.
So is something broken here, or am I missing something?
There is no guarantee that vrefresh is actually set. I think the only way to reliably get it is to call drm_mode_vrefresh().
Well, my impression from studying the code currently is that vrefresh is _never_ set. In particular the Exynos specific DRM code only ever reads vrefresh.
I was just wondering how this is supposed to function at all.
- Tobias
Alex
On Mon, May 02, 2016 at 04:30:37PM +0200, Tobias Jakobi wrote:
Hello Alex,
Alex Deucher wrote:
On Sat, Apr 30, 2016 at 5:37 PM, Tobias Jakobi tjakobi@math.uni-bielefeld.de wrote:
Hello,
while playing around with FIMD enabled, I noticed that when first using the device a zero division was triggered in fimd_calc_clkdiv(). I remembered that I had a similar issue some time ago.
I added a stub fimd_atomic_check() which shows that vrefresh is zero when fimd_calc_clkdiv() is called.
[ 164.059361] [drm:exynos_plane_mode_set] plane : offset_x/y(0,0), width/height(1366,768) [ 164.067175] [drm:fimd_atomic_check] xres=1366, yres=768, refresh=0, intl=0 [ 164.074198] [drm:drm_atomic_helper_check_planes] [CRTC:24:crtc-0] atomic driver check failed
I went back to the git log and noticed that some time ago in 50bbfbffa5c894def440ce8157dfe53e60960d35 the fimd_mode_fixup() call was removed.
I'm now wondering where exactly vrefresh is set to a sane value. As far as I can see of_get_videomode() is used to fetch the video mode from the DT, and drm_display_mode_from_videomode() is used to convert it. However vrefresh is nowhere set.
So is something broken here, or am I missing something?
There is no guarantee that vrefresh is actually set. I think the only way to reliably get it is to call drm_mode_vrefresh().
Well, my impression from studying the code currently is that vrefresh is _never_ set. In particular the Exynos specific DRM code only ever reads vrefresh.
I was just wondering how this is supposed to function at all.
drm_mode_set_crtcinfo() is meant to be used to fill in all the derived values. We might or might not want to have a default call for that in atomic helpers actually (before we call down into any of the driver's check functions for the first time). There's a bunch of flags to control it, but drivers with special needs could simply call it once more and overwrite the computed values. No harm done.
Still need to audit all callers to make sure we don't break a driver. But shouldn't be too much work. -Daniel
Hey Daniel,
Daniel Vetter wrote:
On Mon, May 02, 2016 at 04:30:37PM +0200, Tobias Jakobi wrote:
Hello Alex,
Alex Deucher wrote:
On Sat, Apr 30, 2016 at 5:37 PM, Tobias Jakobi tjakobi@math.uni-bielefeld.de wrote:
Hello,
while playing around with FIMD enabled, I noticed that when first using the device a zero division was triggered in fimd_calc_clkdiv(). I remembered that I had a similar issue some time ago.
I added a stub fimd_atomic_check() which shows that vrefresh is zero when fimd_calc_clkdiv() is called.
[ 164.059361] [drm:exynos_plane_mode_set] plane : offset_x/y(0,0), width/height(1366,768) [ 164.067175] [drm:fimd_atomic_check] xres=1366, yres=768, refresh=0, intl=0 [ 164.074198] [drm:drm_atomic_helper_check_planes] [CRTC:24:crtc-0] atomic driver check failed
I went back to the git log and noticed that some time ago in 50bbfbffa5c894def440ce8157dfe53e60960d35 the fimd_mode_fixup() call was removed.
I'm now wondering where exactly vrefresh is set to a sane value. As far as I can see of_get_videomode() is used to fetch the video mode from the DT, and drm_display_mode_from_videomode() is used to convert it. However vrefresh is nowhere set.
So is something broken here, or am I missing something?
There is no guarantee that vrefresh is actually set. I think the only way to reliably get it is to call drm_mode_vrefresh().
Well, my impression from studying the code currently is that vrefresh is _never_ set. In particular the Exynos specific DRM code only ever reads vrefresh.
I was just wondering how this is supposed to function at all.
drm_mode_set_crtcinfo() is meant to be used to fill in all the derived values. We might or might not want to have a default call for that in atomic helpers actually (before we call down into any of the driver's check functions for the first time). There's a bunch of flags to control it, but drivers with special needs could simply call it once more and overwrite the computed values. No harm done.
I had a look at drm_mode_set_crtcinfo() (from 4.6-rc6) but it also doesn't set the vrefresh field. Maybe you mean a different call?
With best wishes, Tobias
Still need to audit all callers to make sure we don't break a driver. But shouldn't be too much work. -Daniel
On Mon, May 2, 2016 at 10:30 PM, Tobias Jakobi tjakobi@math.uni-bielefeld.de wrote:
drm_mode_set_crtcinfo() is meant to be used to fill in all the derived values. We might or might not want to have a default call for that in atomic helpers actually (before we call down into any of the driver's check functions for the first time). There's a bunch of flags to control it, but drivers with special needs could simply call it once more and overwrite the computed values. No harm done.
I had a look at drm_mode_set_crtcinfo() (from 4.6-rc6) but it also doesn't set the vrefresh field. Maybe you mean a different call?
I was blind - it fills in everything except vrefresh. If you look at drivers most of them use vrefresh purely for debug output or as an informational field in general. Most hw wants the actual pixelclock anyway, so there's not really a need for vrefresh (e.g. i915 only computes vrefresh for the internally used panel modes). The value itself is directly taken from userspace, so probably just a bug in exynos userspace? Of course the driver should still keep working. From a quick look you want to remove any use of vrefresh (except debug output), and especially you shouldn't recompute the actual pixelclock like decon/fimd seem to do (stored into ideal_clock). Instead use mode->clock, that's supposed to be the real value. -Daniel
Hi Tobias, Daniel,
Little late, but maybe helpful anyway...
On 05/02/2016 11:00 PM, Daniel Vetter wrote:
On Mon, May 2, 2016 at 10:30 PM, Tobias Jakobi tjakobi@math.uni-bielefeld.de wrote:
drm_mode_set_crtcinfo() is meant to be used to fill in all the derived values. We might or might not want to have a default call for that in atomic helpers actually (before we call down into any of the driver's check functions for the first time). There's a bunch of flags to control it, but drivers with special needs could simply call it once more and overwrite the computed values. No harm done.
I had a look at drm_mode_set_crtcinfo() (from 4.6-rc6) but it also doesn't set the vrefresh field. Maybe you mean a different call?
I was blind - it fills in everything except vrefresh. If you look at drivers most of them use vrefresh purely for debug output or as an informational field in general. Most hw wants the actual pixelclock anyway, so there's not really a need for vrefresh (e.g. i915 only computes vrefresh for the internally used panel modes). The value itself is directly taken from userspace, so probably just a bug in exynos userspace? Of course the driver should still keep working. From a quick look you want to remove any use of vrefresh (except debug output), and especially you shouldn't recompute the actual pixelclock like decon/fimd seem to do (stored into ideal_clock). Instead use mode->clock, that's supposed to be the real value. -Daniel
Regarding the initial question, mode->vrefresh is set in drm_helper_probe_single_connector_modes. I guess division by zero could be caused by lack of clock in timings provided by dts timing node. As I remember in the past dts timings were passed without provided clock, in such case driver calculated mode->clock for some default value of vrefresh (60Hz for fimd if I remember correctly). Later validation code was added and the clock should be always provided. DTS files were updated after some time and the recalculation code was removed from fimd driver. So it is possible you are a victim of this change.
Regarding usage by drivers, from a quick look it seems to be used for something more than debugging by: - amdgpu, - exynos, - adv7511, - intel_dp(?), - mdp5_cmd_encoder(?), - nouveau, - omap(?), - tilcdc.
Regards Andrzej
On 05/05/2016 11:12 AM, Andrzej Hajda wrote:
Regarding the initial question, mode->vrefresh is set in drm_helper_probe_single_connector_modes. I guess division by zero could be caused by lack of clock in timings provided by dts timing node. As I remember in the past dts timings were passed without provided clock, in such case driver calculated mode->clock for some default value of vrefresh (60Hz for fimd if I remember correctly). Later validation code was added and the clock should be always provided. DTS files were updated after some time and the recalculation code was removed from fimd driver. So it is possible you are a victim of this change.
One more thing, I have looked at Exynos DTS files in mainline. In case of exynos5*.dts files clock-frequency is set to 50000, it is of course insanely low value and vrefresh calculated for such timings will be 0. So please verify if clock-frequency is not too low in your case.
Regards Andrzej
Regarding usage by drivers, from a quick look it seems to be used for something more than debugging by:
- amdgpu,
- exynos,
- adv7511,
- intel_dp(?),
- mdp5_cmd_encoder(?),
- nouveau,
- omap(?),
- tilcdc.
Regards Andrzej _______________________________________________ dri-devel mailing list dri-devel@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/dri-devel
Hello Andrzej,
Andrzej Hajda wrote:
On 05/05/2016 11:12 AM, Andrzej Hajda wrote:
Regarding the initial question, mode->vrefresh is set in drm_helper_probe_single_connector_modes. I guess division by zero could be caused by lack of clock in timings provided by dts timing node. As I remember in the past dts timings were passed without provided clock, in such case driver calculated mode->clock for some default value of vrefresh (60Hz for fimd if I remember correctly). Later validation code was added and the clock should be always provided. DTS files were updated after some time and the recalculation code was removed from fimd driver. So it is possible you are a victim of this change.
One more thing, I have looked at Exynos DTS files in mainline. In case of exynos5*.dts files clock-frequency is set to 50000, it is of course insanely low value and vrefresh calculated for such timings will be 0. So please verify if clock-frequency is not too low in your case.
I'm pretty sure that this is not the case for me.
This is the FIMD entry in my DTS: &fimd { status = "okay";
pinctrl-0 = <&lcd_clk &lcd_data24 &pwm1_out>; pinctrl-names = "default";
display-timings { native-mode = <&timing0>;
timing0: timing { clock-frequency = <85860000>; hactive = <1366>; vactive = <768>; hfront-porch = <48>; hback-porch = <80>; hsync-len = <32>; vback-porch = <14>; vfront-porch = <3>; vsync-len = <5>; pixelclk-active = <1>; hsync-active = <1>; vsync-active = <1>; de-active = <0>; }; }; };
The only peculiarity is that there is nothing attached to the external connector on the board (ODROID-X2).
With regard to use of the vrefresh field by the various drivers, I was made aware of Daniel on IRC that this is incorrect use.
He pointed me to: http://lxr.free-electrons.com/source/include/drm/drm_modes.h#L374 (-> "Not used in a functional way.")
I'm going to send a patch shortly.
With best wishes, Tobias
P.S.: I'm also experiencing another (more serious) issue, but more about that in another thread.
Regards Andrzej
Regarding usage by drivers, from a quick look it seems to be used for something more than debugging by:
- amdgpu,
- exynos,
- adv7511,
- intel_dp(?),
- mdp5_cmd_encoder(?),
- nouveau,
- omap(?),
- tilcdc.
Regards Andrzej _______________________________________________ dri-devel mailing list dri-devel@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/dri-devel
dri-devel@lists.freedesktop.org
-
Alex Deucher -
Andrzej Hajda -
Daniel Vetter -
Tobias Jakobi