1 files changed, 176 insertions, 0 deletions
diff --git a/Documentation/arm/msm/gpiomux.txt b/Documentation/arm/msm/gpiomux.txt
new file mode 100644
index 00000000000..67a81620adf
--- /dev/null
+++ b/Documentation/arm/msm/gpiomux.txt
@@ -0,0 +1,176 @@
+This document provides an overview of the msm_gpiomux interface, which
+is used to provide gpio pin multiplexing and configuration on mach-msm
+targets.
+
+History
+=======
+
+The first-generation API for gpio configuration & multiplexing on msm
+is the function gpio_tlmm_config().  This function has a few notable
+shortcomings, which led to its deprecation and replacement by gpiomux:
+
+The 'disable' parameter:  Setting the second parameter to
+gpio_tlmm_config to GPIO_CFG_DISABLE tells the peripheral
+processor in charge of the subsystem to perform a look-up into a
+low-power table and apply the low-power/sleep setting for the pin.
+As the msm family evolved this became problematic. Not all pins
+have sleep settings, not all peripheral processors will accept requests
+to apply said sleep settings, and not all msm targets have their gpio
+subsystems managed by a peripheral processor. In order to get consistent
+behavior on all targets, drivers are forced to ignore this parameter,
+rendering it useless.
+
+The 'direction' flag: for all mux-settings other than raw-gpio (0),
+the output-enable bit of a gpio is hard-wired to a known
+input (usually VDD or ground).  For those settings, the direction flag
+is meaningless at best, and deceptive at worst.  In addition, using the
+direction flag to change output-enable (OE) directly can cause trouble in
+gpiolib, which has no visibility into gpio direction changes made
+in this way.  Direction control in gpio mode should be made through gpiolib.
+
+Key Features of gpiomux
+=======================
+
+- A consistent interface across all generations of msm.  Drivers can expect
+the same results on every target.
+- gpiomux plays nicely with gpiolib.  Functions that should belong to gpiolib
+are left to gpiolib and not duplicated here.  gpiomux is written with the
+intent that gpio_chips will call gpiomux reference-counting methods
+from their request() and free() hooks, providing full integration.
+- Tabular configuration.  Instead of having to call gpio_tlmm_config
+hundreds of times, gpio configuration is placed in a single table.
+- Per-gpio sleep.  Each gpio is individually reference counted, allowing only
+those lines which are in use to be put in high-power states.
+- 0 means 'do nothing': all flags are designed so that the default memset-zero
+equates to a sensible default of 'no configuration', preventing users
+from having to provide hundreds of 'no-op' configs for unused or
+unwanted lines.
+
+Usage
+=====
+
+To use gpiomux, provide configuration information for relevant gpio lines
+in the msm_gpiomux_configs table.  Since a 0 equates to "unconfigured",
+only those lines to be managed by gpiomux need to be specified.  Here
+is a completely fictional example:
+
+struct msm_gpiomux_config msm_gpiomux_configs[GPIOMUX_NGPIOS] = {
+	[12] = {
+		.active = GPIOMUX_VALID | GPIOMUX_DRV_8MA | GPIOMUX_FUNC_1,
+		.suspended = GPIOMUX_VALID | GPIOMUX_PULL_DOWN,
+	},
+	[34] = {
+		.suspended = GPIOMUX_VALID | GPIOMUX_PULL_DOWN,
+	},
+};
+
+To indicate that a gpio is in use, call msm_gpiomux_get() to increase
+its reference count.  To decrease the reference count, call msm_gpiomux_put().
+
+The effect of this configuration is as follows:
+
+When the system boots, gpios 12 and 34 will be initialized with their
+'suspended' configurations.  All other gpios, which were left unconfigured,
+will not be touched.
+
+When msm_gpiomux_get() is called on gpio 12 to raise its reference count
+above 0, its active configuration will be applied.  Since no other gpio
+line has a valid active configuration, msm_gpiomux_get() will have no
+effect on any other line.
+
+When msm_gpiomux_put() is called on gpio 12 or 34 to drop their reference
+count to 0, their suspended configurations will be applied.
+Since no other gpio line has a valid suspended configuration, no other
+gpio line will be effected by msm_gpiomux_put().  Since gpio 34 has no valid
+active configuration, this is effectively a no-op for gpio 34 as well,
+with one small caveat, see the section "About Output-Enable Settings".
+
+All of the GPIOMUX_VALID flags may seem like unnecessary overhead, but
+they address some important issues.  As unused entries (all those
+except 12 and 34) are zero-filled, gpiomux needs a way to distinguish
+the used fields from the unused.  In addition, the all-zero pattern
+is a valid configuration!  Therefore, gpiomux defines an additional bit
+which is used to indicate when a field is used.  This has the pleasant
+side-effect of allowing calls to msm_gpiomux_write to use '0' to indicate
+that a value should not be changed:
+
+  msm_gpiomux_write(0, GPIOMUX_VALID, 0);
+
+replaces the active configuration of gpio 0 with an all-zero configuration,
+but leaves the suspended configuration as it was.
+
+Static Configurations
+=====================
+
+To install a static configuration, which is applied at boot and does
+not change after that, install a configuration with a suspended component
+but no active component, as in the previous example:
+
+	[34] = {
+		.suspended = GPIOMUX_VALID | GPIOMUX_PULL_DOWN,
+	},
+
+The suspended setting is applied during boot, and the lack of any valid
+active setting prevents any other setting from being applied at runtime.
+If other subsystems attempting to access the line is a concern, one could
+*really* anchor the configuration down by calling msm_gpiomux_get on the
+line at initialization to move the line into active mode.  With the line
+held, it will never be re-suspended, and with no valid active configuration,
+no new configurations will be applied.
+
+But then, if having other subsystems grabbing for the line is truly a concern,
+it should be reserved with gpio_request instead, which carries an implicit
+msm_gpiomux_get.
+
+gpiomux and gpiolib
+===================
+
+It is expected that msm gpio_chips will call msm_gpiomux_get() and
+msm_gpiomux_put() from their request and free hooks, like this fictional
+example:
+
+static int request(struct gpio_chip *chip, unsigned offset)
+{
+        return msm_gpiomux_get(chip->base + offset);
+}
+
+static void free(struct gpio_chip *chip, unsigned offset)
+{
+        msm_gpiomux_put(chip->base + offset);
+}
+
+	...somewhere in a gpio_chip declaration...
+	.request = request,
+	.free    = free,
+
+This provides important functionality:
+- It guarantees that a gpio line will have its 'active' config applied
+  when the line is requested, and will not be suspended while the line
+  remains requested; and
+- It guarantees that gpio-direction settings from gpiolib behave sensibly.
+  See "About Output-Enable Settings."
+
+This mechanism allows for "auto-request" of gpiomux lines via gpiolib
+when it is suitable.  Drivers wishing more exact control are, of course,
+free to also use msm_gpiomux_set and msm_gpiomux_get.
+
+About Output-Enable Settings
+============================
+
+Some msm targets do not have the ability to query the current gpio
+configuration setting.  This means that changes made to the output-enable
+(OE) bit by gpiolib cannot be consistently detected and preserved by gpiomux.
+Therefore, when gpiomux applies a configuration setting, any direction
+settings which may have been applied by gpiolib are lost and the default
+input settings are re-applied.
+
+For this reason, drivers should not assume that gpio direction settings
+continue to hold if they free and then re-request a gpio.  This seems like
+common sense - after all, anybody could have obtained the line in the
+meantime - but it needs saying.
+
+This also means that calls to msm_gpiomux_write will reset the OE bit,
+which means that if the gpio line is held by a client of gpiolib and
+msm_gpiomux_write is called, the direction setting has been lost and
+gpiolib's internal state has been broken.
+Release gpio lines before reconfiguring them.