On Mon, Mar 18, 2013 at 01:15:51PM -0700, Mike Turquette wrote:

Quoting Ulf Hansson (2013-02-28 01:54:34)

quoted

On 28 February 2013 05:49, Mike Turquette [off-list ref] wrote:

quoted

@@ -703,10 +744,29 @@ int clk_enable(struct clk *clk)
        unsigned long flags;
        int ret;

+       /* this call re-enters if it is from the same context */
+       if (spin_is_locked(&enable_lock) || mutex_is_locked(&prepare_lock)) {
+               if ((void *) atomic_read(&context) == get_current()) {
+                       ret = __clk_enable(clk);
+                       goto out;
+               }
+       }

I beleive the clk_enable|disable code will be racy. What do you think
about this scenario:

1. Thread 1, calls clk_prepare -> clk is not reentrant -> mutex_lock
-> set_context to thread1.
2. Thread 2, calls clk_enable -> above "if" will mean that get_current
returns thread 1 context and then clk_enable continues ->
spin_lock_irqsave -> set_context to thread 2.
3. Thread 1 continues and triggers a reentancy for clk_prepare -> clk
is not reentrant (since thread 2 has set a new context) -> mutex_lock
and we will hang forever.

Do you think above scenario could happen?

I think the solution would be to invent another "static atomic_t
context;" which is used only for fast path functions
(clk_enable|disable). That should do the trick I think.

Ulf,

You are correct.  In fact I have a branch that has two separate context
pointers, one for mutex-protected functions and one for
spinlock-protected functions.  Somehow I managed to discard that change
before settling on the final version that was published.

Err.

Do not forget one very important point.

Any clock which has clk_enable() called on it must have had clk_prepare()
already called _and_ completed.  A second clk_prepare() call on the same
clock should be a no-op other than to increase the prepare reference count
on it.

If you do anything else, you are going to get into sticky problems.