On 9/25/19 10:51 AM, Eugeniu Rosca wrote:
Hi All,
I've additionally Cc-ed Laurent and Stephen, since their fruitful
discussion in [1] back in 2014 concluded with a useful documentation
update [2] which is precisely related to the interpretation and usage
of the polarity flag in GPIO specifiers.
I've also Cc-ed those people who have participated in reviewing the
previous patch iterations (Geert, Phil, Enrico).
Before leaving this thread in limbo, I would like to attempt clarifying
what it actually tried to accomplish, one more time.
First of all, it stems from the need to implement a specific customer
requirement. Whether this requirement is sane or not, that's actually
a very important question, but I haven't found much discussion around
it the comments posted so far.
To paraphrase what Harish stated in [3], the customer has a list of GPIO
pins which need to be controlled from userspace. Of course, the customer
can set the polarity of those pins from userspace, as pointed out by
Linus in [4] (thanks!). But, keeping track of GPIO polarity in userspace
is seen like a burden. The customer thinks that the right place for this
HW-specific detail is in device trees. Do you think this preference
is ill-formed?
I think the DT should represent the device that's attached to the GPIOs.
That way, there's already a clear way to represent the GPIO polarity, as
described in the document linked by Eugenui in [2] below.
If for some reason that's not possible, then I think keeping track of
the GPIO polarity in user-space is entirely reasonable, and is the
correct approach. To claim that tracking GPIO polarity in user-space is
too much burden, yet to also allow user-space to control GPIOs at all,
and hence to know exactly which GPIOs must be controlled, is an
inconsistent assertion.
Put another way: If a piece of user-space SW controls GPIOs, it must
know which GPIO number to use for each logical purpose. This information
presumably varies on different platforms, so the SW must have a list of
GPIO numbers and GPIO controller IDs per platform. Additionally storing
a polarity bit along with that information seems entirely trivial to me.
Is there some other issue that I'm overlooking?
If the list of GPIO IDs is retrieved from DT by the user-space SW, I
could see an argument for storing the polarity information in DT along
with that list of GPIO IDs. However, I don't believe there's any
standard way of representing "a list of GPIO IDs for user space use" in DT.
If we hog a GPIO pin in DTS (which allows specifying its polarity),
userspace no longer has access to that pin. There isn't a way to define
GPIO polarity by means of DTS without affecting userspace access
(can anybody contradict this statement?).
GPIO hog doesn't seem like the right approach; its intent is to actively
configure the GPIO in a fixed state, which is logically incompatible
with user-space control of the GPIO.
Whether it is obvious or not, the main goal of this series is actually
to provide the possibility of inverting the default ACTIVE_HIGH polarity
for GPIO pin X _via DTS_ while still allowing to operate on that pin
_from userspace_. My two questions are then:
- I hope it is something sane to desire?
- If it is sane, how can this be accomplished, if the functionality
implemented by Harish doesn't pass the community review?
[1] https://marc.info/?l=linux-gpio&m=139204273132477&w=4 ("Correct meaning of the GPIO active low flag")
[2] https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/commit/?id=51e8afc1c43c75 ("gpio: document polarity flag best practices")
[3] https://marc.info/?l=linux-gpio&m=155721267517644&w=2 ("[PATCH V1 1/2] gpio: make it possible to set active-state on GPIO lines")
[4] https://marc.info/?l=linux-gpio&m=155713157122847&w=2 ("[PATCH V1 1/2] gpio: make it possible to set active-state on GPIO lines")
