[pulseaudio-discuss] pa_once can run twice?
frederic.dalleau at intel.com
Wed Apr 25 04:30:04 PDT 2012
On Wed, Apr 25, 2012 at 1:09 PM, Dalleau, Frederic
<frederic.dalleau at intel.com> wrote:
> Hi David,
> On Tue, Apr 24, 2012 at 3:01 PM, David Henningsson
> <david.henningsson at canonical.com> wrote:
>> On 04/23/2012 05:03 PM, Dalleau, Frederic wrote:
>>> Hi David,
>>> On Fri, Apr 20, 2012 at 4:44 PM, David Henningsson
>>> <david.henningsson at canonical.com> wrote:
>>>> While researching a bug I came across something that might be a bug in
>>>> pa_once logic, but this stuff is tricky, so I might also be missing
>>>> Imagine this:
>>>> * Thread 1 runs pa_once_begin, succeeds and starts running the payload
>>>> (i e
>>>> the code that should only run once).
>>>> * Thread 2 starts running pa_once_begin, but only the first row. We're
>>>> right *before* pa_atomic_inc(&control->ref) but *after*
>>>> * Thread 1 finishes the payload, runs pa_once_done which sets
>>>> and frees the mutex.
>>>> * Thread 2 continues, pa_once_begin succeeds and the payload is now run
>>>> second time!
>>> After reading your mail, I made some experiments by adding a usleep() call
>>> in Thread 1 between pa_atomic_load(&control->done) and
>>> and that failed once-test 100% of time.
>>> I reverted the usleep and made another experiment using 50000 iterations
>>> once-test and it just failed.
>>> Good catch !
>> I tried to look up implementation/algorithm suggestions, and for the ones I
>> found , there was no freeing of the mutex. Without freeing, the code
>> becomes simpler. The attached patch is a version of that. I've just tried a
>> simple once-test (but do feel free to run it 50000 times :-) ).
>> But of course, now we leak a mutex. But that's what we already do with the
>> static mutexes we use in a few places already, so maybe it doesn't matter
> Your patch has passed the 50000 iteration test, but I'm reluctant in the idea of
> leaking memory since solutions exists.
> I have attached a version which uses reference counting to free the
> mutex. I can't find
> any significant difference in cpu usage, but it must be double checked
> for correctness.
> libatomic_ops documentation suggest the following :
> With a static mutex, it uses a few extra bytes (pthread_mutex_t is 24
> bytes on my box), but it may also offer the best performances of all:
> no allocation
> and no busy loop waiting for mutex creation. I'm giving it a try but
> it needs a bit
> of plumbing.
This might not be feasible at all,
I just forgot that there were other OSes than Linux ;)
More information about the pulseaudio-discuss