Originally, the I2C controller supported by the i2c-mv64xxx driver
requires a lot of software support: an interrupt is generated at each
step of an I2C transaction (after the start bit, after sending the
address, etc.) and the driver is in charge of re-programming the I2C
controller to do the next step of the I2C transaction. This explains
the fairly complex state machine that the driver has.

On Marvell Armada XP and later processors (Armada 375, 38x, etc.), the
I2C controller was extended with a part called the "I2C Bridge", which
allows to offload the I2C transaction completely to the
hardware. Initial support for this mechanism was added in commit
930ab3d4 ("i2c: mv64xxx: Add I2C Transaction Generator support").

However, the implementation done in this commit has two related
issues, which this commit fixes by completely changing how the offload
implementation is done:

 * SMBus read transfers, where there is one write to select the
   register immediately followed in the same transaction by one read,
   were making the processor hang. This was easier visible on the
   Marvell Armada XP WRT1900AC platform using a driver for an I2C LED
   controller, or on other Armada XP platforms by using a simple
   'i2cget' command to read an I2C EEPROM.

 * The implementation was based on the fact that the offload engine
   was re-programmed to transfer each message of an I2C xfer: this
   meant that each message sent with the offload engine was starting
   with a normal I2C start sequence. However, the I2C subsystem
   assumes that all messages belonging to the same xfer will use the
   so-called "repeated start" so that the entire I2C xfer is seen as
   one transfer by the I2C devices and cannot be interrupt by other
   I2C masters on the same bus.

In fact, the "I2C Bridge" allows to offload three types of xfer:

 - xfer of one write message
 - xfer of one read message
 - xfer of one write message followed by one read message

For all other situations, we have to fallback to not using the "I2C
Bridge" in order to get proper I2C semantics.

Therefore, this commit reworks the offload implementation to put it
not at the message level, but at the xfer level: in the
mv64xxx_i2c_xfer() function, we decide if the transaction can be
offloaded (in which case it is handled by the
mv64xxx_i2c_offload_xfer() function), or otherwise it is handled by
the slow path (implemented in the existing mv64xxx_i2c_execute_msg()).

This allows to simplify the state machine, which no longer needs to
have any state related to the offload implementation: the offload
implementation is now completely separated from the slow path (with
the exception of the interrupt handler, of course).

In summary:

 - mv64xxx_i2c_can_offload() will analyze an I2C xfer and decided of
   the "I2C Bridge" can be used to offload it or not.

 - mv64xxx_i2c_offload_xfer() will actually program the "I2C Bridge"
   to offload one xfer (of either one or two messages), and block
   using mv64xxx_i2c_wait_for_completion() until the xfer completes.

 - The interrupt handler mv64xxx_i2c_intr() is modified to push the
   offload related code to a separate function,
   mv64xxx_i2c_intr_offload(). It will take care of reading the
   received data if needed.

This commit was tested on:

 - Armada XP OpenBlocks AX3-4 (EEPROM on I2C and RTC on I2C)
 - Armada XP WRT1900AC (LED controller on I2C)
 - Armada XP GP (EEPROM on I2C)

Fixes: 930ab3d4 ("i2c: mv64xxx: Add I2C Transaction Generator support")
Cc: <stable@vger.kernel.org> # v3.12+
Signed-off-by: Thomas Petazzoni <thomas.petazzoni@free-electrons.com>
[wsa: fixed checkpatch warnings]
Signed-off-by: Wolfram Sang <wsa@the-dreams.de>