Re: [PATCH RFC v1 2/5] clk: add missing lock when call clk_core_enable in clk_set_parent

[Stephen Boyd <sboyd@codeaurora.org>]

On 05/04, Dong Aisheng wrote:
> On Thu, Apr 30, 2015 at 12:07:47PM -0700, Stephen Boyd wrote:
> > On 04/15/15 07:26, Dong Aisheng wrote:
> > > clk_core_enable is executed without &enable_clock in clk_set_parent function.
> > > Adding it to avoid potential race condition issue.
> > >
> > > Fixes: 035a61c314eb ("clk: Make clk API return per-user struct clk instances")
> > > Cc: Mike Turquette <mturquette@linaro.org>
> > > Cc: Stephen Boyd <sboyd@codeaurora.org>
> > > Signed-off-by: Dong Aisheng <aisheng.dong@freescale.com>
> > > ---
> > 
> > Can you please describe the race condition? From what I can tell there
> > is not a race condition here and we've gone around on this part of the
> > code before to fix any race conditions.
> > 
> 
> Do you mean we do not need to acquire enable lock when execute clk_core_enable
> in set_parent function? Can you help explain a bit more why?
> 
> The clk doc looks to me says the enable lock should be held across calls to
> the .enable, .disable and .is_enabled operations.
> 
> And before the commit
> 035a61c314eb ("clk: Make clk API return per-user struct clk instances"),
> all the clk_enable/disable in set_parent() is executed with lock.
> 
> A rough thinking of race condition is assuming Thread A calls
> clk_set_parent(x, y) while Thread B calls clk_enable(x), clock x is disabled
> but prepared initially, due to clk_core_enable in set_parent() is not
> executed with enable clock, the clk_core_enable may be reentrant during
> the locking time executed by B.
> Won't this be a race condition?
> 

Ah I see now. The commit text could say something like this:

Before commit 035a61c314eb ("clk: Make clk API return per-user
struct clk instances") we acquired the enable_lock in
__clk_set_parent_{before,after}() by means of calling
clk_enable(). After commit 035a61c314eb we use clk_core_enable()
in place of the clk_enable(), and clk_core_enable() doesn't
acquire the enable_lock. This opens up a race condition between
clk_set_parent() and clk_enable().

I've replaced the commit text and applied it to clk-fixes.

-- 
Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
a Linux Foundation Collaborative Project