Re: [PATCH 8/9] iio: imu: st_lsm6dsx: add event configurability on a per axis basis

[Jonathan Cameron <jic23@kernel.org>]

On Fri, 21 Nov 2025 15:57:57 +0100
Francesco Lavra <flavra@baylibre.com> wrote:

> On Fri, 2025-11-21 at 11:31 +0200, Andy Shevchenko wrote:
> > On Fri, Nov 21, 2025 at 10:14:06AM +0100, Francesco Lavra wrote:  
> > > On Thu, 2025-11-20 at 20:31 +0200, Andy Shevchenko wrote:  
> > > > On Thu, Nov 20, 2025 at 03:59:18PM +0200, Andy Shevchenko wrote:  
> > > > > On Thu, Nov 20, 2025 at 12:43:09PM +0100, Francesco Lavra wrote:  
> > > > > > On Thu, 2025-11-20 at 10:05 +0100, Andy Shevchenko wrote:  
> > > > > > > On Tue, Nov 18, 2025 at 12:01:57PM +0100, Francesco Lavra
> > > > > > > wrote:  
> > > > > > > > On Tue, 2025-11-18 at 11:44 +0100, Andy Shevchenko wrote:  
> > > > > > > > > On Mon, Nov 17, 2025 at 08:23:35PM +0100, Francesco Lavra
> > > > > > > > > wrote:  
> > > > > > > > > > On Thu, 2025-10-30 at 15:56 +0200, Andy Shevchenko wrote:  
> > > > > > > > > > > On Thu, Oct 30, 2025 at 12:23:19PM +0100, Francesco
> > > > > > > > > > > Lavra
> > > > > > > > > > > wrote:  
> > > > > > > > > > > > On Thu, 2025-10-30 at 10:24 +0200, Andy Shevchenko
> > > > > > > > > > > > wrote:  
> > > > > > > > > > > > > On Thu, Oct 30, 2025 at 08:27:51AM +0100, Francesco
> > > > > > > > > > > > > Lavra
> > > > > > > > > > > > > wrote:  
> > 
> > ...
> >   
> > > > > > > > > > > > > > +       old_enable = hw->enable_event[event];
> > > > > > > > > > > > > > +       new_enable = state ? (old_enable |
> > > > > > > > > > > > > > BIT(axis))
> > > > > > > > > > > > > > :
> > > > > > > > > > > > > > (old_enable
> > > > > > > > > > > > > > &
> > > > > > > > > > > > > > ~BIT(axis));
> > > > > > > > > > > > > > +       if (!!old_enable == !!new_enable)  
> > > > > > > > > > > > > 
> > > > > > > > > > > > > This is an interesting check. So, old_enable and
> > > > > > > > > > > > > new_enable
> > > > > > > > > > > > > are
> > > > > > > > > > > > > _not_
> > > > > > > > > > > > > booleans, right?
> > > > > > > > > > > > > So, this means the check test if _any_ of the bit
> > > > > > > > > > > > > was
> > > > > > > > > > > > > set and
> > > > > > > > > > > > > kept
> > > > > > > > > > > > > set or
> > > > > > > > > > > > > none were set
> > > > > > > > > > > > > and non is going to be set. Correct? I think a
> > > > > > > > > > > > > short
> > > > > > > > > > > > > comment
> > > > > > > > > > > > > would be
> > > > > > > > > > > > > good to have.  
> > > > > > > > > > > > 
> > > > > > > > > > > > old_enable and new_enable are bit masks, but we are
> > > > > > > > > > > > only
> > > > > > > > > > > > interested
> > > > > > > > > > > > in
> > > > > > > > > > > > whether any bit is set, to catch the cases where the
> > > > > > > > > > > > bit
> > > > > > > > > > > > mask
> > > > > > > > > > > > goes
> > > > > > > > > > > > from
> > > > > > > > > > > > zero to non-zero and vice versa. Will add a comment.  
> > > > > > > > > > > 
> > > > > > > > > > > If it's a true bitmask (assuming unsigned long type)
> > > > > > > > > > > then
> > > > > > > > > > > all
> > > > > > > > > > > this
> > > > > > > > > > > can be
> > > > > > > > > > > done
> > > > > > > > > > > via bitmap API calls. Otherwise you can also compare a
> > > > > > > > > > > Hamming
> > > > > > > > > > > weights of
> > > > > > > > > > > them
> > > > > > > > > > > (probably that gives even the same size of the object
> > > > > > > > > > > file,
> > > > > > > > > > > but
> > > > > > > > > > > !!
> > > > > > > > > > > instructions
> > > > > > > > > > >  will be changed to hweight() calls (still a single
> > > > > > > > > > > assembly
> > > > > > > > > > > instr on
> > > > > > > > > > > modern
> > > > > > > > > > >  architectures).  
> > > > > > > > > > 
> > > > > > > > > > These are u8 variables, so we can't use the bitmap API.  
> > > > > > > > > 
> > > > > > > > > OK. But hweight8() can still be used.
> > > > > > > > >   
> > > > > > > > > > And I don't
> > > > > > > > > > understand the reason for using hweight(), given that the
> > > > > > > > > > !!
> > > > > > > > > > operators
> > > > > > > > > > would still be needed.  
> > > > > > > > > 
> > > > > > > > > No, you won't need !! with that.  
> > > > > > > > 
> > > > > > > > I still don't understand. Are you proposing to replace `if
> > > > > > > > (!!old_enable ==
> > > > > > > > !!new_enable)` with `if (hweight8(old_enable) ==
> > > > > > > > hweight8(new_enable))`?
> > > > > > > > That won't work, because we only need to check whether the
> > > > > > > > Hamming
> > > > > > > > weight
> > > > > > > > goes from zero to non-zero and vice versa.  
> > > > > > > 
> > > > > > >        old_enable = hw->enable_event[event];
> > > > > > >        new_enable = state ? (old_enable | BIT(axis)) :
> > > > > > >                             (old_enable & ~BIT(axis));
> > > > > > >        if (!!old_enable == !!new_enable)
> > > > > > >                return 0;
> > > > > > > 
> > > > > > > If I am not mistaken this will do exactly the same in a simpler
> > > > > > > way.
> > > > > > > 
> > > > > > >         old_enable = hw->enable_event[event];
> > > > > > >         if (state)
> > > > > > >                 new_enable = old_enable | BIT(axis);
> > > > > > >         else
> > > > > > >                 new_enable = old_enable & ~BIT(axis);
> > > > > > >         if ((new_enable ^ old_enable) != BIT(axis))
> > > > > > >                 return 0;  
> > > > > > 
> > > > > > This doesn't look right to me, if new_enable differs from
> > > > > > old_enable
> > > > > > by
> > > > > > just one bit (which it does), then the XOR result will always
> > > > > > have
> > > > > > this bit
> > > > > > (and no others) set, so (new_enable ^ old_enable) will always
> > > > > > equal
> > > > > > BIT(axis).
> > > > > > We are not checking if the bit was already set or clear, when
> > > > > > this
> > > > > > code
> > > > > > runs we already know that the bit is changing, what we are
> > > > > > checking
> > > > > > is
> > > > > > whether all bits are zero before or after this change.  
> > > > > 
> > > > > The check I proposed is to have a look for the cases when
> > > > > old_enable
> > > > > was 0 and
> > > > > the BIT(axis) is set and when the BIT(axis) was _the last_ bit in
> > > > > the
> > > > > mask that
> > > > > got reset. If it's not the case, the code will return 0. I think my
> > > > > check is
> > > > > correct.
> > > > > 
> > > > > Should I write a test case?  
> > > > 
> > > > FWIW,
> > > > https://gist.github.com/andy-shev/afe4c0e009cb3008ac613d8384aaa6eb  
> > > 
> > > The code in your gist produces:
> > > 
> > > Initial event: 0x92, new state: True for bit 0x20  
> > 
> > Initial event is 10010010b, we assume that we got in the code when
> > required
> > state is to 'set' (True) with axis = 5 (means 00100000b when powered).
> > 
> > The [-] are special cases just to show the algo.
> >   
> > > [-] 0x00 | 0x20 --> 1: handle  
> > 
> > If initial event is 0 we handle
> > 
> > If _after_ that the bit 5 set (which is NOT the case in _this_
> > iteration),
> > we will stop handling.  
> 
> We have to stop handling not only if bit 5 is set, but also if any other
> bit is set after that.
> 
> > > [0] 0x92 | 0x20 --> 1: handle  
> > 
> > So, it's again step 1. I _assumed_ that your code works and sets the bit.  
> 
> Even if it's again step 1, this is not supposed to be handled, because
> there are bits already set in the current bitmask.
> 
> Enabling an event source for an axis may need two registers to be set:
> 1) an axis-specific enable register
> 2) an event-specific enable register (global for all axes)
> 
> If no events are enabled on any axis, when we enable the event source for
> axis X, we have to do both steps above; if then we enable the same event
> source for axis Y, we have to do just step 1 but not step 2; and that's
> what the (!!old_enable == !!new_enable) check is supposed to do: to check
> if, after setting the axis-specific enable register, we have to set also
> the event-specific enable register. If I replace that check with
> ((new_enable ^ old_enable) != BIT(axis)), then I'm doing both steps for
> every axis, which is unnecessary when enabling the event (because we don't
> need to set again the event-specific register after we set it for the first
> axis), and is wrong when disabling the event (because disabling the event
> for a single axis would inadvertently disable it globally (i.e. for all
> axes).
> 
Obviously I've very late to this thread, but just thought I'd comment
that if the driver was using regcache (which might well make sense) then
I wouldn't bother with the check at all. The write to update the register
to the value already has wouldn't do anything other than a few trivial
operations to check the cache value matches.

Might be worth enabling the regcache part of regmap simply to avoid
having to care about corners like this.

Jonathan