Re: [PATCH] i2c: s3c2410: fix ABBA deadlock by keeping clock prepared

From: Wolfram Sang
Date: Fri Jan 23 2015 - 09:53:04 EST


On Mon, Jan 19, 2015 at 05:03:33PM +0100, Paul Osmialowski wrote:
> This patch solves deadlock between clock prepare mutex and regmap mutex reported
> by Tomasz Figa in [1] by implementing solution from [2]: "always leave the clock
> of the i2c controller in a prepared state".
>
> [1] https://lkml.org/lkml/2014/7/2/171
> [2] https://lkml.org/lkml/2014/7/2/207
>
> On each i2c transfer handled by s3c24xx_i2c_xfer(), clk_prepare_enable() was
> called, which calls clk_prepare() then clk_enable(). clk_prepare() takes
> prepare_lock mutex before proceeding. Note that i2c transfer functions are
> invoked from many places in kernel, typically with some other additional lock
> held.
>
> It may happen that function on CPU1 (e.g. regmap_update_bits()) has taken a
> mutex (i.e. regmap lock mutex) then it attempts i2c communication in order to
> proceed (so it needs to obtain clock related prepare_lock mutex during transfer
> preparation stage due to clk_prepare() call). At the same time other task on
> CPU0 wants to operate on clock (e.g. to (un)prepare clock for some other reason)
> so it has taken prepare_lock mutex.
>
> CPU0: CPU1:
> clk_disable_unused() regulator_disable()
> clk_prepare_lock() map->lock(map->lock_arg)
> regmap_read() s3c24xx_i2c_xfer()
> map->lock(map->lock_arg) clk_prepare_lock()
>
> Implemented solution from [2] leaves i2c clock prepared. Preparation is done in
> s3c24xx_i2c_probe() function. Without this patch, it is immediately unprepared
> by clk_disable_unprepare() call. I've replaced this call with clk_disable() and
> I've added clk_unprepare() call in s3c24xx_i2c_remove().
>
> The s3c24xx_i2c_xfer() function now uses clk_enable() instead of
> clk_prepare_enable() (and clk_disable() instead of clk_unprepare_disable()).
>
> Signed-off-by: Paul Osmialowski <p.osmialowsk@xxxxxxxxxxx>

Applied to for-current, thanks!

Attachment: signature.asc
Description: Digital signature