Re: Bug#1029850: linux: Driver not loaded for ST Microelectronics LSM6DS3TR-C accelerometer (acpi:SMO8B30:SMO8B30:)

[Jonathan Cameron]

On Sun, 5 Feb 2023 09:50:51 +0100
Hans de Goede <hdegoede@xxxxxxxxxx> wrote:

> Hi,
> 
> On 2/4/23 23:15, Darrell Kavanagh wrote:
> > Yes, I understand that.
> > 
> > What I mean is that the matrix read from the DSDT by Jonathan's
> > amended driver is
> > 
> >  0 -1  0
> >  1  0  0
> >  0  0  1
> > 
> > and the (correct) matrix created with my new hwdb entry is
> > 
> >  0  1  0
> > -1  0  0
> >  0  0  1

May be concidence, but I think that's the inverse of the one we are reading
from ROTM - so represents the transform in the other direction.

The way ROTM is defined is that first row represents the direction of
the x axis in device coordinates - so it's the transform from sensor
to device space.

I wonder if the hwdb matrix is defined from world space to sensor?  Seems
unlikely. 

The IIO ABI docs describe mount matrix as being what you apply to data to
tranform into device space (oh for a diagram in the docs).  Anyhow my reading
is that matches with ROTM definition but maybe I'm reading that wrong...

For extra annoyance, the ROTM matrix on this device isn't a rotation matrix.
It's flipping the handedness of the sensor.  Determinant isn't -1 which it
should be.  I guess the sensor itself might have an axis backwards from
windows convention though *sigh*  I think windows uses left handed convention
and looks like sensor is using right handed (which I think is what Android and
similar use).

> > 
> > which is the algebraic transposition (ie reflection in the diagonal)
> > of the DSDT one.
> > 
> > In other words, though the DST matrix is wrong, it is wrong in a
> > specific way - the rows should be the columns, and vv. I was just
> > wondering if this was a DSDT bug that might have been seen elsewhere
> > before.  
> 
> No this does not ring a bell, but the x and y axis being swapped
> does seem related to the LCD panel being 90° rotated.
> 
> > BTW, there is another matrix in the DSTD, but I can't find the
> > associated HID (10EC5280) anywhere (Linux sysfs or Windows Powershell
> > system data extract). It's not a correct matrix, though - could it be
> > just a bit of redundant code in the DST?  
> 
> Yes that is likely there often is a bunch of dead stuff DSDT leftover
> from other device models.
> 
> Regards,
> 
> Hans
> 
> 
> 
> 
> 
> > 
> > Darrell
> > 
> > On Sat, 4 Feb 2023 at 21:31, Hans de Goede <hdegoede@xxxxxxxxxx> wrote:  
> >>
> >> Hi,
> >>
> >> On 2/4/23 18:09, Darrell Kavanagh wrote:  
> >>> I've just noticed that the working mount matrix that I added to my
> >>> hwdb is the matrix retrieved from the ACPI ROTM call in the amended
> >>> driver, transposed.  
> >>
> >> An other word for the mount matrix would be a rotation matrix,
> >> since it defines how the physical sensor is mounted on the PCB
> >> in a rotated fashion compared to its standard orientation.
> >>
> >> The x, y, z axis relationship underling of course does
> >> not change by the rotation, so yes all mount matrices
> >> are a transposition of the standard:
> >>
> >> 1, 0, 0 : 0, 1, 0 : 0, 0, 1
> >>
> >> matrix, that is expected. Where that to not be the case
> >> then there would be a bug in the accelerometer driver itself
> >> where the driver itself is swapping or inverting axis.
> >>
> >> Regards,
> >>
> >> Hans
> >>
> >>
> >>
> >>  
> >>> On Fri, 3 Feb 2023 at 18:23, Darrell Kavanagh
> >>> <darrell.kavanagh@xxxxxxxxx> wrote:  
> >>>>
> >>>> Finally got a 6.2.0-rc6 kernel built and installed, with the following
> >>>> patch, and everything is working as expected.
> >>>>
> >>>> Moving on now to look at Bastien's suggestion.
> >>>>
> >>>> Thanks,
> >>>> Darrell
> >>>>
> >>>> diff --git a/kernel/drm_panel_orientation_quirks.c
> >>>> b/kernel/linux-6.2-rc6/drivers/gpu/drm/drm_panel_orientation_quirks.c
> >>>> index 3659f04..590bb7b 100644
> >>>> --- a/kernel/drm_panel_orientation_quirks.c
> >>>> +++ b/kernel/linux-6.2-rc6/drivers/gpu/drm/drm_panel_orientation_quirks.c
> >>>> @@ -304,6 +304,12 @@ static const struct dmi_system_id orientation_data[] = {
> >>>>                   DMI_EXACT_MATCH(DMI_PRODUCT_VERSION, "Lenovo ideapad
> >>>> D330-10IGM"),
> >>>>                 },
> >>>>                 .driver_data = (void *)&lcd1200x1920_rightside_up,
> >>>> +       }, {    /* Lenovo IdeaPad Duet 3 10IGL5 */
> >>>> +               .matches = {
> >>>> +                 DMI_EXACT_MATCH(DMI_SYS_VENDOR, "LENOVO"),
> >>>> +                 DMI_EXACT_MATCH(DMI_PRODUCT_VERSION, "IdeaPad Duet 3 10IGL5"),
> >>>> +               },
> >>>> +               .driver_data = (void *)&lcd1200x1920_rightside_up,
> >>>>         }, {    /* Lenovo Ideapad D330-10IGL (HD) */
> >>>>                 .matches = {
> >>>>                   DMI_EXACT_MATCH(DMI_SYS_VENDOR, "LENOVO"),
> >>>>
> >>>> On Wed, 1 Feb 2023 at 17:55, Hans de Goede <hdegoede@xxxxxxxxxx> wrote:  
> >>>>>
> >>>>> Hi,
> >>>>>
> >>>>> On 2/1/23 18:50, Darrell Kavanagh wrote:  
> >>>>>> Thank you. I don't have anything that could be called a big machine.
> >>>>>> The fastest processor I have access to is a Core m3-8100Y - that's in
> >>>>>> a Chromebook with 4GB memory - it can run Linux in a chroot or
> >>>>>> officially in Google's VM. I also have an ancient gen 2 core i5-2410M
> >>>>>> machine which is slower than the m3 in theory, but that has 6GB of
> >>>>>> memory.
> >>>>>>
> >>>>>> Is the kernel build more processor or memory bound?  
> >>>>>
> >>>>> It is mostly processor bound, esp. wtih something like make -j4,
> >>>>> make -j16 will start taking some RAM, but with make -j4 I expect you
> >>>>> to be fully CPU bound.
> >>>>>
> >>>>> Regards,
> >>>>>
> >>>>> Hans
> >>>>>
> >>>>>
> >>>>>
> >>>>>
> >>>>>
> >>>>>
> >>>>>
> >>>>>
> >>>>>
> >>>>>  
> >>>>>> On Wed, 1 Feb 2023 at 16:12, Bastien Nocera <hadess@xxxxxxxxxx> wrote:  
> >>>>>>>
> >>>>>>> On Wed, 2023-02-01 at 12:00 +0100, Hans de Goede wrote:  
> >>>>>>>> Hi,
> >>>>>>>>
> >>>>>>>> On 2/1/23 11:28, Jonathan Cameron wrote:  
> >>>>>>>>> On Wed, 1 Feb 2023 01:40:49 +0000
> >>>>>>>>> Darrell Kavanagh <darrell.kavanagh@xxxxxxxxx> wrote:
> >>>>>>>>>  
> >>>>>>>>>> Hello, all.
> >>>>>>>>>>
> >>>>>>>>>> I've finally reached a conclusion on this, after testing all the
> >>>>>>>>>> combinations of the patches (with and without reading the acpi
> >>>>>>>>>> mounting matrix), window managers (wayland, xorg) and the
> >>>>>>>>>> presence or
> >>>>>>>>>> not of my custom kernel parms.
> >>>>>>>>>>
> >>>>>>>>>> What works well is the full set of patches with the custom kernel
> >>>>>>>>>> parms and a new hwdb entry for the sensor:
> >>>>>>>>>>
> >>>>>>>>>> sensor:modalias:acpi:SMO8B30*:dmi:*:svnLENOVO*:pn82AT:*
> >>>>>>>>>>  ACCEL_MOUNT_MATRIX=0, 1, 0; -1, 0, 0; 0, 0, 1
> >>>>>>>>>>
> >>>>>>>>>> The autorotate then works correctly in wayland and xorg, but for
> >>>>>>>>>> xorg,
> >>>>>>>>>> the settings say the screen is "portrait left" when in actual
> >>>>>>>>>> fact it
> >>>>>>>>>> is in standard laptop landscape orientation. Wayland does not
> >>>>>>>>>> have
> >>>>>>>>>> this problem (I guess because wayland's view of the screen is
> >>>>>>>>>> straight
> >>>>>>>>>> from the kernel).
> >>>>>>>>>>
> >>>>>>>>>> Without the hwdb entry, the orientation is 90 degrees out without
> >>>>>>>>>> using the acpi matrix and 180 degrees out when using it. I could
> >>>>>>>>>> have
> >>>>>>>>>> gone either way here with appropriate hwdb entries, but my view
> >>>>>>>>>> is
> >>>>>>>>>> that we *should* be using the matrix.  
> >>>>>>>>>
> >>>>>>>>> Added Hans de Goede as he has probably run into more of this mess
> >>>>>>>>> than anyone else.  Hans, any thoughts on if we are doing something
> >>>>>>>>> wrong on kernel side?  Or is the matrix just wrong *sigh*  
> >>>>>>>>
> >>>>>>>> I see below that this laptop has a panel which is mounted 90 degrees
> >>>>>>>> rotated, that likely explains why the ACPI matrix does not work.
> >>>>>>>> So the best thing to do here is to just override it with a hwdb
> >>>>>>>> entries.
> >>>>>>>>
> >>>>>>>> IIRC there are already 1 or 2 other hwdb entries which actually
> >>>>>>>> override the ACPI provided matrix because of similar issues.
> >>>>>>>>
> >>>>>>>> Linux userspace expects the matrix in this case to be set so that
> >>>>>>>> it causes e.g. gnome's auto-rotation to put the image upright
> >>>>>>>> even with older gnome versions / mate / xfce which don't know about
> >>>>>>>> the panel being mounted 90 degrees.
> >>>>>>>>
> >>>>>>>> So e.g. "monitor-sensor" will report left-side-up or right-side-up
> >>>>>>>> while the device is actually in normal clamshell mode with the
> >>>>>>>> display up-right.
> >>>>>>>>
> >>>>>>>> This reporting of left-side-up or right-side-up is actually "correct"
> >>>>>>>> looking from the native LCD panel orientation and as mentioned is
> >>>>>>>> done for backward compatibility. This is documented here:
> >>>>>>>>
> >>>>>>>> https://github.com/systemd/systemd/blob/main/hwdb.d/60-sensor.hwdb#L54
> >>>>>>>>
> >>>>>>>> The way we are handling this is likely incompatible with how Windows
> >>>>>>>> handles this special case of 90° rotated screen + ROTM. Or the
> >>>>>>>> matrix in the ACPI tables could be just wrong...
> >>>>>>>>  
> >>>>>>>>> I think 'ROTM' is defined by MS.
> >>>>>>>>> https://learn.microsoft.com/en-us/windows-hardware/drivers/sensors/sensors-acpi-entries  
> >>>>>>>>
> >>>>>>>> Right and as such it would be good if we can still add support to
> >>>>>>>> it to the sensor driver in question. Because the ROTM info usually
> >>>>>>>> is correct and avoids the need for adding more and more hwdb entries.
> >>>>>>>>
> >>>>>>>> Note there already is existing support in some other sensor drivers.
> >>>>>>>>
> >>>>>>>> So we probably need to factor out some helper code for this and share
> >>>>>>>> that between sensor drivers.
> >>>>>>>>
> >>>>>>>>  
> >>>>>>>>>> The only thing that concerns me is the need for custom kernel
> >>>>>>>>>> parms.
> >>>>>>>>>> It would be better if there was a way to avoid this, so that the
> >>>>>>>>>> user
> >>>>>>>>>> didn't have to mess around with their grub config. Though having
> >>>>>>>>>> said
> >>>>>>>>>> that, the sensors fix as we have it doesn't make things worse -
> >>>>>>>>>> under
> >>>>>>>>>> currently released kernels the screen always starts up sideways
> >>>>>>>>>> unless
> >>>>>>>>>> custom parms are added in grub.  
> >>>>>>>>
> >>>>>>>> We actually have a quirk mechanism in the kernel for specifying
> >>>>>>>> the need for: video=DSI-1:panel_orientation=right_side_up  and this
> >>>>>>>> will also automatically fix the fbcon orientation, see:
> >>>>>>>>
> >>>>>>>> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/gpu/drm/drm_panel_orientation_quirks.c
> >>>>>>>>
> >>>>>>>> If you submit a patch for this upstream please Cc me.  
> >>>>>>>
> >>>>>>> And if after that change, and copy/pasting the orientation from the
> >>>>>>> DSDT into hwdb the sensor and screen move in the expected ways, then
> >>>>>>> maybe stealing the BMC150 driver's
> >>>>>>> bmc150_apply_bosc0200_acpi_orientation() might be a good idea.
> >>>>>>>
> >>>>>>> Once exported through "mount_matrix", iio-sensor-proxy should see it
> >>>>>>> and read it without the need for a hwdb entry.
> >>>>>>>
> >>>>>>> Cheers  
> >>>>>>  
> >>>>>  
> >>>  
> >>  
> >   
>