Deadlock with I2C based clock providers

[Mike Turquette]

Quoting Martin Fuzzey (2014-03-18 06:26:53)
> Hi Mike,
> 
> I have a clock provide that implements the clock operations over i2c.
> 
> Although this is now possible since commit 533ddeb1e8 clk: allow 
> reentrant calls into the clk framework
> 
> that patch only deals with single threaded reentrancy.
> 
> 
> The issue I am having involves two threads:
> 
> Thread A calls the i2c clock driver
> Thread B is in an unrelated driver for another I2C device on the same 
> I2C bus (audio codec).
> 
> 
> Thread A:
>      clk_set_rate(my_clk)
>          clk_prepare_lock()  <== Takes clock lock
>          clk_change_rate()
>              ops->set_rate()
>                  i2c_transfer()
>                      i2c_lock_adapter()  <== Blocks on I2C bus lock
> 
> 
> 
> Thread B:
>      i2c_transfer()
>          i2c_lock_adapter()    <== Takes I2C bus lock
>          i2c_imx_xfer()
>              i2c_imx_start()
>                  clk_prepare_enable(i2c_clk)
>                      clk_prepare_lock()  <== Blocks on clock lock
> 
> 
> The problem is that the locks are taken in the reverse order leading to 
> a classic deadlock.
> 
> I saw your [RFC] clk: reentrancy via per-clock locking but that seemed 
> to have other problems.
> 
> 
> My current (horrible) workaround involves exposing clk_prepare_lock() 
> and calling it from i2c_core but that obviously isn't the right solution.

A slightly less hacky (but still horrible) solution might be to move the
calls to clk_prepare_enable before calling i2c_lock_adapter().

Alternatively some trylock solution might be necessary to unwind the
locks and yield execution for a while and then try again. The clk
framework reentrancy patches are pretty irrelevant here since those were
only designed to allow reentrancy from within the same task context and
this is not an example of that. This deadlock is a more clear-cut
circular dependency between i2c and the clock framework.

Regards,
Mike

> 
> Regards,
> 
> Martin
>