[PATCH] clk: Fix race condition between clk_set_parent and clk_enable()

[Tomasz Figa]

On Tuesday 14 of May 2013 15:46:33 Saravana Kannan wrote:
> On 05/14/2013 03:10 PM, Tomasz Figa wrote:
> > Hi,
> > 
> > On Tuesday 14 of May 2013 11:54:17 Mike Turquette wrote:
> >> Quoting Saravana Kannan (2013-04-30 21:42:08)
> >> 
> >>> Without this patch, the following race conditions are possible.
> >>> 
> >>> Race condition 1:
> >>> * clk-A has two parents - clk-X and clk-Y.
> >>> * All three are disabled and clk-X is current parent.
> >>> * Thread A: clk_set_parent(clk-A, clk-Y).
> >>> * Thread A: <snip execution flow>
> >>> * Thread A: Grabs enable lock.
> >>> * Thread A: Sees enable count of clk-A is 0, so doesn't enable
> >>> clk-Y.
> >>> * Thread A: Updates clk-A SW parent to clk-Y
> >>> * Thread A: Releases enable lock.
> >>> * Thread B: clk_enable(clk-A).
> >>> * Thread B: clk_enable() enables clk-Y, then enabled clk-A and
> >>> returns.
> >>> 
> >>> clk-A is now enabled in software, but not clocking in hardware since
> >>> the hardware parent is still clk-X.
> >>> 
> >>> The only way to avoid race conditions between clk_set_parent() and
> >>> clk_enable/disable() is to ensure that clk_enable/disable() calls
> >>> don't
> >>> require changes to hardware enable state between changes to software
> >>> clock topology and hardware clock topology.
> >>> 
> >>> There are options to achieve the above:
> >>> 1. Grab the enable lock before changing software/hardware topology
> >>> and
> >>> 
> >>>     release it afterwards.
> >>> 
> >>> 2. Keep the clock enabled for the duration of software/hardware
> >>> topology>
> >>> 
> >>>     change so that any additional enable/disable calls don't try to
> >>>     change
> >>>     the hardware state. Once the topology change is complete, the
> >>>     clock
> >>>     can
> >>>     be put back in its original enable state.
> >>> 
> >>> Option (1) is not an acceptable solution since the set_parent() ops
> >>> might need to sleep.
> >>> 
> >>> Therefore, this patch implements option (2).
> >>> 
> >>> This patch doesn't violate any API semantics. clk_disable() doesn't
> >>> guarantee that the clock is actually disabled. So, no clients of a
> >>> clock can assume that a clock is disabled after their last call to
> >>> clk_disable(). So, enabling the clock during a parent change is not
> >>> a
> >>> violation of any API semantics.
> >>> 
> >>> This also has the nice side effect of simplifying the error handling
> >>> code.
> >>> 
> >>> Signed-off-by: Saravana Kannan <skannan at codeaurora.org>
> >> 
> >> I've taken this patch into clk-next for testing.  The code itself
> >> looks
> >> fine.  The only thing that remains to be seen is if any platforms
> >> have a problem with disabled clocks getting turned on during a
> >> reparent operation.
> > 
> > IMHO this behavior should be documented somewhere, with a note that
> > the
> > clock must not be prepared to keep it disabled during reparent
> > operation and possibly also pointing to the CLK_SET_PARENT_GATE flag.
> 
> Reasonable request. I can update the documentation of clk_set_parent()
> to indicate that the clock might get turned on for the duration of the
> call and if they need a guarantee the GATE flag should be used.
> 
> >> On platforms that I have worked on this is OK, but I suppose there
> >> could be some platform out there where a clock is prepared and
> >> disabled, and briefly enabling the clock during the reparent
> >> operation somehow puts the hardware in a bad state.
> > 
> > Well, on any platform where default clock settings are not completely
> > correct this is likely to cause problems, because some device might
> > get
> > too high frequency for some period of time, which might crash it alone
> > as well as the whole system.
> 
> I don't think this is really a problem with this patch. It's present
> even without this patch.
> 
> The patch doesn't switch to some other unspecified parent. It only
> switches between the new/old parent. Even without this patch, if a clock
> is prepared while you reparent it, clk_enable() could be called at
> anytime between the parent switch and the future clock API calls to set
> up the new parent correctly. I think that's just crappy driver code to
> switch to a new parent before setting it up correctly. There's
> absolutely no good reason to do it that way.

This is not exactly what I meant. I was just giving an example problem of 
turning a clock on, if it's not set up correctly yet.

AFAIK most (if not all) of current code either does necessary reparenting 
and initial rate setting early, before clk_prepare(), so it is not a 
problem or already after clk_enable() (in case of reparenting dynamically 
at runtime), so there shouldn't be any problem.

Best regards,
Tomasz