diff --git a/Documentation/gpio.txt b/Documentation/gpio.txt
new file mode 100644
index 0000000..09dd510
--- /dev/null
+++ b/Documentation/gpio.txt
@@ -0,0 +1,271 @@
+GPIO Interfaces
+
+This provides an overview of GPIO access conventions on Linux.
+
+
+What is a GPIO?
+===============
+A "General Purpose Input/Output" (GPIO) is a flexible software-controlled
+digital signal.  They are provided from many kinds of chip, and are familiar
+to Linux developers working with embedded and custom hardware.  Each GPIO
+represents a bit connected to a particular pin, or "ball" on Ball Grid Array
+(BGA) packages.  Board schematics show which external hardware connects to
+which GPIOs.  Drivers can be written generically, so that board setup code
+passes such pin configuration data to drivers.
+
+System-on-Chip (SOC) processors heavily rely on GPIOs.  In some cases, every
+non-dedicated pin can be configured as a GPIO; and most chips have at least
+several dozen of them.  Programmable logic devices (like FPGAs) can easily
+provide GPIOs; multifunction chips like power managers, and audio codecs
+often have a few such pins to help with pin scarcity on SOCs; and there are
+also "GPIO Expander" chips that connect using the I2C or SPI serial busses.
+Most PC southbridges have a few dozen GPIO-capable pins (with only the BIOS
+firmware knowing how they're used).
+
+The exact capabilities of GPIOs vary between systems.  Common options:
+
+  - Output values are writable (high=1, low=0).  Some chips also have
+    options about how that value is driven, so that for example only one
+    value might be driven ... supporting "wire-OR" and similar schemes
+    for the other value.
+
+  - Input values are likewise readable (1, 0).  Some chips support readback
+    of pins configured as "output", which is very useful in such "wire-OR"
+    cases (to support bidirectional signaling).  GPIO controllers may have
+    input de-glitch logic, sometimes with software controls.
+
+  - Inputs can often be used as IRQ signals, often edge triggered but
+    sometimes level triggered.  Such IRQs may be configurable as system
+    wakeup events, to wake the system from a low power state.
+
+  - Usually a GPIO will be configurable as either input or output, as needed
+    by different product boards; single direction ones exist too.
+
+  - Most GPIOs can be accessed while holding spinlocks, but those accessed
+    through a serial bus normally can't.  Some systems support both types.
+
+On a given board each GPIO is used for one specific purpose like monitoring
+MMC/SD card insertion/removal, detecting card writeprotect status, driving
+a LED, configuring a transceiver, bitbanging a serial bus, poking a hardware
+watchdog, sensing a switch, and so on.
+
+
+GPIO conventions
+================
+Note that this is called a "convention" because you don't need to do it this
+way, and it's no crime if you don't.  There **are** cases where portability
+is not the main issue; GPIOs are often used for the kind of board-specific
+glue logic that may even change between board revisions, and can't ever be
+used on a board that's wired differently.  Only least-common-denominator
+functionality can be very portable.  Other features are platform-specific,
+and that can be critical for glue logic.
+
+Plus, this doesn't define an implementation framework, just an interface.
+One platform might implement it as simple inline functions accessing chip
+registers; another might implement it by delegating through abstractions
+used for several very different kinds of GPIO controller.
+
+That said, if the convention is supported on their platform, drivers should
+use it when possible:
+
+	#include <asm/gpio.h>
+
+If you stick to this convention then it'll be easier for other developers to
+see what your code is doing, and help maintain it.
+
+
+Identifying GPIOs
+-----------------
+GPIOs are identified by unsigned integers in the range 0..MAX_INT.  That
+reserves "negative" numbers for other purposes like marking signals as
+"not available on this board", or indicating faults.
+
+Platforms define how they use those integers, and usually #define symbols
+for the GPIO lines so that board-specific setup code directly corresponds
+to the relevant schematics.  In contrast, drivers should only use GPIO
+numbers passed to them from that setup code, using platform_data to hold
+board-specific pin configuration data (along with other board specific
+data they need).  That avoids portability problems.
+
+So for example one platform uses numbers 32-159 for GPIOs; while another
+uses numbers 0..63 with one set of GPIO controllers, 64-79 with another
+type of GPIO controller, and on one particular board 80-95 with an FPGA.
+The numbers need not be contiguous; either of those platforms could also
+use numbers 2000-2063 to identify GPIOs in a bank of I2C GPIO expanders.
+
+Whether a platform supports multiple GPIO controllers is currently a
+platform-specific implementation issue.
+
+
+Using GPIOs
+-----------
+One of the first things to do with a GPIO, often in board setup code when
+setting up a platform_device using the GPIO, is mark its direction:
+
+	/* set as input or output, returning 0 or negative errno */
+	int gpio_direction_input(unsigned gpio);
+	int gpio_direction_output(unsigned gpio);
+
+The return value is zero for success, else a negative errno.  It should
+be checked, since the get/set calls don't have error returns and since
+misconfiguration is possible.  (These calls could sleep.)
+
+Setting the direction can fail if the GPIO number is invalid, or when
+that particular GPIO can't be used in that mode.  It's generally a bad
+idea to rely on boot firmware to have set the direction correctly, since
+it probably wasn't validated to do more than boot Linux.  (Similarly,
+that board setup code probably needs to multiplex that pin as a GPIO,
+and configure pullups/pulldowns appropriately.)
+
+
+Spinlock-Safe GPIO access
+-------------------------
+Most GPIO controllers can be accessed with memory read/write instructions.
+That doesn't need to sleep, and can safely be done from inside IRQ handlers.
+
+Use these calls to access such GPIOs:
+
+	/* GPIO INPUT:  return zero or nonzero */
+	int gpio_get_value(unsigned gpio);
+
+	/* GPIO OUTPUT */
+	void gpio_set_value(unsigned gpio, int value);
+
+The values are boolean, zero for low, nonzero for high.  When reading the
+value of an output pin, the value returned should be what's seen on the
+pin ... that won't always match the specified output value, because of
+issues including wire-OR and output latencies.
+
+The get/set calls have no error returns because "invalid GPIO" should have
+been reported earlier in gpio_set_direction().  However, note that not all
+platforms can read the value of output pins; those that can't should always
+return zero.  Also, these calls will be ignored for GPIOs that can't safely
+be accessed wihtout sleeping (see below).
+
+Platform-specific implementations are encouraged to optimise the two
+calls to access the GPIO value in cases where the GPIO number (and for
+output, value) are constant.  It's normal for them to need only a couple
+of instructions in such cases (reading or writing a hardware register),
+and not to need spinlocks.  Such optimized calls can make bitbanging
+applications a lot more efficient (in both space and time) than spending
+dozens of instructions on subroutine calls.
+
+
+GPIO access that may sleep
+--------------------------
+Some GPIO controllers must be accessed using message based busses like I2C
+or SPI.  Commands to read or write those GPIO values require waiting to
+get to the head of a queue to transmit a command and get its response.
+This requires sleeping, which can't be done from inside IRQ handlers.
+
+Platforms that support this type of GPIO distinguish them from other GPIOs
+by returning nonzero from this call:
+
+	int gpio_cansleep(unsigned gpio);
+
+To access such GPIOs, a different set of accessors is defined:
+
+	/* GPIO INPUT:  return zero or nonzero, might sleep */
+	int gpio_get_value_cansleep(unsigned gpio);
+
+	/* GPIO OUTPUT, might sleep */
+	void gpio_set_value_cansleep(unsigned gpio, int value);
+
+Other than the fact that these calls might sleep, and will not be ignored
+for GPIOs that can't be accessed from IRQ handlers, these calls act the
+same as the spinlock-safe calls.
+
+
+Claiming and Releasing GPIOs (OPTIONAL)
+---------------------------------------
+To help catch system configuration errors, two calls are defined.
+However, many platforms don't currently support this mechanism.
+
+	/* request GPIO, returning 0 or negative errno.
+	 * non-null labels may be useful for diagnostics.
+	 */
+	int gpio_request(unsigned gpio, const char *label);
+
+	/* release previously-claimed GPIO */
+	void gpio_free(unsigned gpio);
+
+Passing invalid GPIO numbers to gpio_request() will fail, as will requesting
+GPIOs that have already been claimed with that call.  The return value of
+gpio_request() must be checked.  (These calls could sleep.)
+
+These calls serve two basic purposes.  One is marking the signals which
+are actually in use as GPIOs, for better diagnostics; systems may have
+several hundred potential GPIOs, but often only a dozen are used on any
+given board.  Another is to catch conflicts between drivers, reporting
+errors when drivers wrongly think they have exclusive use of that signal.
+
+These two calls are optional because not not all current Linux platforms
+offer such functionality in their GPIO support; a valid implementation
+could return success for all gpio_request() calls.  Unlike the other calls,
+the state they represent doesn't normally match anything from a hardware
+register; it's just a software bitmap which clearly is not necessary for
+correct operation of hardware or (bug free) drivers.
+
+Note that requesting a GPIO does NOT cause it to be configured in any
+way; it just marks that GPIO as in use.  Separate code must handle any
+pin setup (e.g. controlling which pin the GPIO uses, pullup/pulldown).
+
+
+GPIOs mapped to IRQs
+--------------------
+GPIO numbers are unsigned integers; so are IRQ numbers.  These make up
+two logically distinct namespaces (GPIO 0 need not use IRQ 0).  You can
+map between them using calls like:
+
+	/* map GPIO numbers to IRQ numbers */
+	int gpio_to_irq(unsigned gpio);
+
+	/* map IRQ numbers to GPIO numbers */
+	int irq_to_gpio(unsigned irq);
+
+Those return either the corresponding number in the other namespace, or
+else a negative errno code if the mapping can't be done.  (For example,
+some GPIOs can't used as IRQs.)  It is an unchecked error to use a GPIO
+number that hasn't been marked as an input using gpio_set_direction(), or
+to use an IRQ number that didn't originally come from gpio_to_irq().
+
+These two mapping calls are expected to cost on the order of a single
+addition or subtraction.  They're not allowed to sleep.
+
+Non-error values returned from gpio_to_irq() can be passed to request_irq()
+or free_irq().  They will often be stored into IRQ resources for platform
+devices, by the board-specific initialization code.  Note that IRQ trigger
+options are part of the IRQ interface, e.g. IRQF_TRIGGER_FALLING, as are
+system wakeup capabilities.
+
+Non-error values returned from irq_to_gpio() would most commonly be used
+with gpio_get_value().
+
+
+
+What do these conventions omit?
+===============================
+One of the biggest things these conventions omit is pin multiplexing, since
+this is highly chip-specific and nonportable.  One platform might not need
+explicit multiplexing; another might have just two options for use of any
+given pin; another might have eight options per pin; another might be able
+to route a given GPIO to any one of several pins.  (Yes, those examples all
+come from systems that run Linux today.)
+
+Related to multiplexing is configuration and enabling of the pullups or
+pulldowns integrated on some platforms.  Not all platforms support them,
+or support them in the same way; and any given board might use external
+pullups (or pulldowns) so that the on-chip ones should not be used.
+
+There are other system-specific mechanisms that are not specified here,
+like the aforementioned options for input de-glitching and wire-OR output.
+Hardware may support reading or writing GPIOs in gangs, but that's usually
+configuration dependednt:  for GPIOs sharing the same bank.  (GPIOs are
+commonly grouped in banks of 16 or 32, with a given SOC having several such
+banks.)  Code relying on such mechanisms will necessarily be nonportable.
+
+Dynamic definition of GPIOs is not currently supported; for example, as
+a side effect of configuring an add-on board with some GPIO expanders.
+
+These calls are purely for kernel space, but a userspace API could be built
+on top of it.