tree e07d87bba28678aa80d9325a9c48eac0f57a7fe2
parent 962bcbc57aa244eeb1176fa2e9f65ac865cca68a
author Stephen Warren <swarren@nvidia.com> 1330718744 -0700
committer Linus Walleij <linus.walleij@linaro.org> 1330942789 +0100

pinctrl: fix and simplify locking

There are many problems with the current pinctrl locking:

struct pinctrl_dev's gpio_ranges_lock isn't effective;
pinctrl_match_gpio_range() only holds this lock while searching for a gpio
range, but the found range is return and manipulated after releading the
lock. This could allow pinctrl_remove_gpio_range() for that range while it
is in use, and the caller may very well delete the range after removing it,
causing pinctrl code to touch the now-free range object.

Solving this requires the introduction of a higher-level lock, at least
a lock per pin controller, which both gpio range registration and
pinctrl_get()/put() will acquire.

There is missing locking on HW programming; pin controllers may pack the
configuration for different pins/groups/config options/... into one
register, and hence have to read-modify-write the register. This needs to
be protected, but currently isn't. Related, a future change will add a
"complete" op to the pin controller drivers, the idea being that each
state's programming will be programmed into the pinctrl driver followed
by the "complete" call, which may e.g. flush a register cache to HW. For
this to work, it must not be possible to interleave the pinctrl driver
calls for different devices.

As above, solving this requires the introduction of a higher-level lock,
at least a lock per pin controller, which will be held for the duration
of any pinctrl_enable()/disable() call.

However, each pinctrl mapping table entry may affect a different pin
controller if necessary. Hence, with a per-pin-controller lock, almost
any pinctrl API may need to acquire multiple locks, one per controller.
To avoid deadlock, these would need to be acquired in the same order in
all cases. This is extremely difficult to implement in the case of
pinctrl_get(), which doesn't know which pin controllers to lock until it
has parsed the entire mapping table, since it contains somewhat arbitrary
data.

The simplest solution here is to introduce a single lock that covers all
pin controllers at once. This will be acquired by all pinctrl APIs.

This then makes struct pinctrl's mutex irrelevant, since that single lock
will always be held whenever this mutex is currently held.

Signed-off-by: Stephen Warren <swarren@nvidia.com>
Signed-off-by: Linus Walleij <linus.walleij@linaro.org>