Re: [PATCH] iio: sanity check available_scan_masks array

[Jonathan Cameron]

On Tue, 10 Oct 2023 15:56:22 +0300
Matti Vaittinen <mazziesaccount@xxxxxxxxx> wrote:

> On 10/10/23 13:04, Jonathan Cameron wrote:
> > On Fri, 6 Oct 2023 14:10:16 +0300
> > Matti Vaittinen <mazziesaccount@xxxxxxxxx> wrote:
> >   
> >> Hi Again Jonathan.
> >>
> >> On 10/5/23 18:30, Jonathan Cameron wrote:  
> >>> On Tue, 3 Oct 2023 12:49:45 +0300
> >>> Matti Vaittinen <mazziesaccount@xxxxxxxxx> wrote:
> >>>      
> >>>> When IIO goes through the available scan masks in order to select the
> >>>> best suiting one, it will just accept the first listed subset of channels
> >>>> which meets the user's requirements. If driver lists a mask which is a
> >>>> subset of some of the masks previously in the array of
> >>>> avaliable_scan_masks, then the latter one will never be selected.
> >>>>
> >>>> Add a warning if driver registers masks which can't be used due to the
> >>>> available_scan_masks-array ordering.
> >>>>
> >>>> Suggested-by: Jonathan Cameron <jic23@xxxxxxxxxx>
> >>>> Signed-off-by: Matti Vaittinen <mazziesaccount@xxxxxxxxx>  
> >>> Hi Matti
> >>>
> >>> Thanks for doing this.  A few comments inline + maybe we need to think
> >>> about a unit test for the matching code. I feel we aren't pushing the
> >>> corners of that in any drivers so far so it might bite us later.
> >>>
> >>> Still that's a job for another day.
> >>>
> >>> Jonathan
> >>>      
> >>>>
> >>>> ---
> >>>> The change was suggested by Jonathan here:
> >>>> https://lore.kernel.org/lkml/20230924170726.41443502@jic23-huawei/
> >>>> ---
> >>>>    drivers/iio/industrialio-core.c | 57 +++++++++++++++++++++++++++++++++
> >>>>    1 file changed, 57 insertions(+)
> >>>>
> >>>> diff --git a/drivers/iio/industrialio-core.c b/drivers/iio/industrialio-core.c
> >>>> index c77745b594bd..d4f37f4eeec0 100644
> >>>> --- a/drivers/iio/industrialio-core.c
> >>>> +++ b/drivers/iio/industrialio-core.c
> >>>> @@ -1896,6 +1896,53 @@ static int iio_check_extended_name(const struct iio_dev *indio_dev)  
> >>
> >> ...
> >>  
> >>>> +
> >>>> +	for (num_masks = 0; *av_masks; num_masks++)  
> >>>
> >>> I think we can't just check *av_masks - need bitmap_empty() as first
> >>> long might be 0 but could be bits set in the next one.
> >>>      
> >>>> +		av_masks += longs_per_mask;  
> >>
> >> I did switch this to:
> >> +       for (num_masks = 0; !bitmap_empty(av_masks, masklength);
> >> num_masks++)
> >> +               av_masks += longs_per_mask;
> >>
> >> but this kind of freaks me out.  
> > 
> > Good because I'm fairly sure you need to reduce the masklength int hat
> > bitmap_empty as well.  It is getting a bit complex.  
> 
> Hm. As far as I can say the masklength is constant telling how many bits 
> there is in one mask(?) I don't think we can reduce it. Idea is just to 
> increment the av_masks pointer until we find the zero mask indicating 
> end of an array. This is how we count the amount of masks in the array.
I read it wrong  :(    Thought we were iterating over the longs not the
sets of longs in each mask. oops. I read the same code further down correctly.

I blame spending whole day waiting for a plumber (who hasn't come... sigh)

> 
> Caveat being that if driver has used single long '0' to indicate the end 
> of an array, and if the masklength > 32 - we'll end up reading out of 
> bounds. (as we agreed later in the mail. OTOH, we also agreed there does 
> not seem to be drivers with masklength > 32 utilizing the 
> available_scan_masks - so this is just trying to avoid problems in the 
> future).
> 
> >>
> >> I think in kernel we see two ways of constructing and passing arrays to
> >> frameworks. One is creating a NULL terminated array, the other being an
> >> array which size is given. The available_scan_masks is using the first
> >> approach.
> >>
> >> The array represents bitmasks, which are thought to be of arbitrary
> >> length. The type of array items is longs though. When building an arry
> >> like this, it is easy to just do:
> >>
> >> unsigned long masks[] = {
> >> 	mask1_hi,
> >> 	mask1_lo,
> >> 	mask2_hi,
> >> 	mask2_lo,
> >> 	...
> >> 	maskN_lo,
> >> 	/* sentinel */
> >> 	0
> >> }
> >>
> >> (By the way, I've always hated that 'sentinel' comment as it - in my
> >> opinion - is not worth adding. I think the meaning of 0 should be
> >> obvious, but here I just added it to alleviate the problem).
> >>
> >> Here, if I'm not mistaken, the check I implemented would go reading out
> >> of the array bounds.  
> > 
> > It does indeed.
> > 
> >   
> >>
> >> Knowing how easy it would be slip the above array past my reviewing eyes
> >> - I find this scary. And ugly part of this is that we can't detect this
> >> in the iio-core side, because we have no way of knowing how big the
> >> array and sentinel are. What makes this worse is that the core does:
> >>
> >> for (i = 0; i < indio_dev->num_channels; i++)
> >>                           ml = max(ml, channels[i].scan_index + 1);
> >>                   indio_dev->masklength = ml;
> >>
> >> so, masklength may not be what was set in driver.  
> > 
> > IIRC this is there to allow for sparse scan_index values.  Those are
> > very rare, but I think there are a few drivers doing that because
> > it allowed for slightly simpler code a long time back.  May not even
> > matter today.  Key is that mask_length is big enough to allow the
> > bits at all present scan_index values to be set.
> > 
> > So it should always match with the drivers where this is used to make
> > sure available_scan_masks has the right number of longs per entry,
> > but the drivers may need to be a little clever if they are both
> > doing large numbers of channels and sparse scan_index values.
> > AFAIK there are none doing that. Going further I don't recall any
> > drivers that use the available_scan_masks stuff going beyond 32
> > channels (so needing more than one unsigned long per element).  
> 
> I didn't find one either.
> 
> >> I did quick and dirty grep for "_scan_mask\[" in iio directory and
> >> didn't spot any bigger than a few channels masks. Still, this makes me
> >> worried.
> >>
> >> BTW: I did also:
> >>
> >> Author: Matti Vaittinen <mazziesaccount@xxxxxxxxx>
> >> Date:   Fri Oct 6 13:53:11 2023 +0300
> >>
> >>       iio: buffer: use bitmap_empty() to find last mask
> >>
> >>       When IIO buffer code is scanning the array of available masks for
> >>       matching the user's enable request to channel configuration
> >> supported by
> >>       driver, the code uses a 'check for long 0' as indication of last mask.
> >>       This does not work right for channel masks greater than BITS_PER_LONG.
> >>
> >>       Use bitmap_empty() to find the last element in available_scan_masks
> >>       array.
> >>
> >>       Signed-off-by: Matti Vaittinen <mazziesaccount@xxxxxxxxx>
> >>       ---
> >>       NOTE: This is potentially hazardous change. Please, don't pick without
> >>       thorough check and understanding.
> >>
> >> diff --git a/drivers/iio/industrialio-buffer.c
> >> b/drivers/iio/industrialio-buffer.c
> >> index 176d31d9f9d8..1e59afddcf9a 100644
> >> --- a/drivers/iio/industrialio-buffer.c
> >> +++ b/drivers/iio/industrialio-buffer.c
> >> @@ -413,7 +413,7 @@ static const unsigned long
> >> *iio_scan_mask_match(const unsigned long *av_masks,
> >>    {
> >>           if (bitmap_empty(mask, masklength))
> >>                   return NULL;
> >> -       while (*av_masks) {
> >> +       while (!bitmap_empty(av_masks, masklength)) {
> >>                   if (strict) {
> >>                           if (bitmap_equal(mask, av_masks, masklength))
> >>                                   return av_masks;
> >>
> >> but this is just as fragile - for obvious reasons.  
> > Ah. yes, that is indeed a bug.  I'm not sure your fix is particularly
> > fragile though.  This comes back to us having no drivers that actually use
> > big bitmaps yet.
> > 
> > Key is that we need the available_scan_masks null terminator to be the
> > same length as any other entry - so if multiple unsigned longs needed
> > then multiple 0's should be there.  
> 
> Exactly my thinking, and why I think this fix would be fragile. I am not 
> convinced people would think of adding enough of zeroes.
> 
> > We should definitely document that
> > and ideally add a test case.   We can bulk out the dummy driver
> > to trigger these and provide an example of how available_scan_masks
> > should be set.
> >   
> >>
> >> One way around this would be to have the first bit in the long always
> >> set for a valid mask - and take this into account when going through the
> >> masks. It's probably somewhat more confusing than current code though -
> >> but it would allow using just a single long (with all - or  at least
> >> first - bits zero to indicate end of masks).  
> > 
> > Too complex.
> >   
> 
> I think I agree. At least for as long as we don't actually have any 
> available_scan_masks users with masklength > 32
> 
> >>
> >> Other option I see is to just error out if available_scan_masks array is
> >> given with larger than one 'long' wide masks and worry things when this
> >> breaks.  
> > 
> > That would kick the problem into the long grass.  
> 
> Well, not 100% sure I interpret the idiom correctly ;) In any case, I'd 
> say this would indeed postpone dealing with the problem to the future. 

It does indeed mean that!  Sorry bad habit to use idioms in discussions like
this.

> To the point we actually seem to have a problem. The "long grass" as if 
> hiding the problem is something we can avoid by adding something like:
> 
> if (masklength > 32 && idev->available_scan_masks) {
> 	/*
> 	 * Comment mowing the long grass.
> 	 */
> 	dev_err( ...);
> 	return -EINVAL;
> }
> 
> to the device registration.
> 
> >>
> >> Anyways, I don't like using bitmap_empty() for array of bitmaps which
> >> may be longer than BITS_PER_LONG unless we can sanity check the size of
> >> the array...
> >>
> >> How do you feel about this?  
> > 
> > Agreed it's problematic as that null terminator isn't clearly forced to
> > be big enough.  Hmm. Can we cheat for any drivers that actually need large
> > masks (when they come along) and use an appropriate 2D array.  
> 
> I think we could. Or, maybe develop some mask initialization macro 
> magic. It'd just be cool if one did not need to do things differently 
> for multi long masks.
> 
> > unsigned long available_masks[][2] = {
> > 	{mask0_ll, mask0_hl},
> > 	{mask0_ll, mask0_hl},
> > 	{}
> > };  
> 
> don't know how would it work to have
> unsigned long available_masks[][1] = {
> 	{mask0},
> 	{mask1},
> 	{}
> };
> 
> for regular masks with 1 long / mask as well? At first look it seems 
> horrible to me but at least it would be a standard :) Don't know if 
> there is a sane way to make a macro for it.

Agreed - ugly for one entry so we keep those flat which means we have
to keep our eyes open for the multiple long version.

I've never come up with a sane way to do a macro to set bits
in multiple longs or seen one unfortunately.

> 
> > 	iio_dev->available_scan_masks = (unsigned long *)available_masks;
> > 
> > If we put such an example into the dummy / example driver then that might
> > act to avoid us getting bugs in future + test the fix you have above and
> > related.  
> 
> Well, at least it shouldn't hurt to have some example - although I'm 
> still tempted to use the "long grass" - option ;)

That is probably a good idea for now.  Though we are carrying other infrastructure
to support this eventually and it feels weird to error out on it whilst we have
code to support it (assuming that terminator is long enough).


> 
> Yours,
> 	-- Matti
>