Hi dee Ho peeps!
Disclaimer - I have no HW to test this using real in-tree drivers. If
someone has a device with a variant of bmc150 or adxl372 or - it'd be
nice to see if reading hwfifo_watermark_max or hwfifo_watermark_min
works with the v6.0-rc4. Maybe I am misreading code and have my own
issues - in which case I apologize already now and go to the corner
while being deeply ashamed :)
On 2/15/21 12:40, Alexandru Ardelean wrote:
> This change wraps all buffer attributes into iio_dev_attr objects, and
> assigns a reference to the IIO buffer they belong to.
>
> With the addition of multiple IIO buffers per one IIO device, we need a way
> to know which IIO buffer is being enabled/disabled/controlled.
>
> We know that all buffer attributes are device_attributes.
I think this assumption is slightly unsafe. I see few drivers adding
IIO_CONST_ATTRs in attribute groups. For example the bmc150 and adxl372
add the hwfifo_watermark_min and hwfifo_watermark_max.
Long story short (or the other way around?)
I was developing a driver for ROHM/Kionix KX022A accelerometer. I did a
random pick and chose the bmc150-core as a reference how others have
implemented the accel drivers. During the testing I noticed that using
IIO_CONST_ATTRs for triggered buffers seem to cause access to somewhere
it shouldn't... Oops.
Reading the code allows me to assume the problem is wrapping the
attributes to IIO_DEV_ATTRs.
static struct attribute *iio_buffer_wrap_attr(struct iio_buffer *buffer,
+ struct attribute *attr)
+{
+ struct device_attribute *dattr = to_dev_attr(attr);
+ struct iio_dev_attr *iio_attr;
+
+ iio_attr = kzalloc(sizeof(*iio_attr), GFP_KERNEL);
+ if (!iio_attr)
+ return NULL;
+
+ iio_attr->buffer = buffer;
+ memcpy(&iio_attr->dev_attr, dattr, sizeof(iio_attr->dev_attr));
This copy does assume all attributes are device_attrs, and does not take
into account that IIO_CONST_ATTRS have the string stored in a struct
iio_const_attr which is containing the dev_attr. Eg, copying in the
iio_buffer_wrap_attr() does not copy the string - and later invoking the
'show' callback goes reading something else than the mentioned string
because the pointer is not copied.
I believe problem has been there for a while now - introduced by this:
15097c7a1adc ("iio: buffer: wrap all buffer attributes into
iio_dev_attr") at Feb 2021.
I don't have a fix for you as I am not sure what would be the correct
fix. We _could_ go through all the IIO drivers and ensure none used
IIO_CONST_ATTRs for triggered buffers. This would IMHO be just a
workaround to the real problem - which is assumption that all attributes
in the attribute group are dev_attrs. Such silent assumptions are
fragile as we see :)
If we opt treating all attributes in the group as device_attrs also in
the future, then (in my not always so humble opinion) the APIs should
force this - but I'm not sure how to do that.
Best regards
-- Matti
--
Matti Vaittinen
Linux kernel developer at ROHM Semiconductors
Oulu Finland
~~ When things go utterly wrong vim users can always type :help! ~~
