[ltt-dev] Userspace rcu using POSIX tls?

[Mathieu Desnoyers]

* pplists at me.com (pplists at me.com) wrote:
>
>
> On 04 May, 2011,at 06:31 AM, Mathieu Desnoyers <compudj at krystal.dyndns.org> wrote:
>  
> We use the pthread primitives for mutexes, thread creation, and rely on
> the TLS provided by gcc. Can you elaborate more on the limitations the
> Android target has with respect to pthread and TLS support ?
>  
> The biggest obstacle is that gcc's __thread specifier isn't supported on Android. 
>
> Currently, the only alternative seems to be the pthread_{get,set}specific() API, see http://pubs.opengroup.org/onlinepubs/009695399/functions/pthread_setspecific.html
>
> The difference between the approaches is that with POSIX TLS we will have to dynamically obtain a pointer to the thread local rcu_reader rather than thread local struct. (See code below)
>
> static inline void _rcu_read_lock(void) {
> 	unsigned long tmp;
> 	cmm_barrier();	/* Ensure the compiler does not reorder us with mutex */
> #ifdef USE_POSIX_TLS
> 	struct rcu_reader *rcu_reader = thread_local_rcu_reader();
> 	tmp = rcu_reader->ctr;

Yep, that should work fine where __thread is not implemented.

> 	if (likely(!(tmp & RCU_GP_CTR_NEST_MASK))) {
> 		_CMM_STORE_SHARED(rcu_reader->ctr, _CMM_LOAD_SHARED(rcu_gp_ctr));
> 		smp_mb_slave(RCU_MB_GROUP);
> 	} else {
> 		_CMM_STORE_SHARED(rcu_reader->ctr, tmp + RCU_GP_COUNT);
> 	}
> #else
> 	tmp = rcu_reader.ctr;
> 	if (likely(!(tmp & RCU_GP_CTR_NEST_MASK))) {
> 		_CMM_STORE_SHARED(rcu_reader.ctr, _CMM_LOAD_SHARED(rcu_gp_ctr));
> 		smp_mb_slave(RCU_MB_GROUP);
> 	} else {
> 		_CMM_STORE_SHARED(rcu_reader.ctr, tmp + RCU_GP_COUNT);
> 	}
> #endif
> }
>
> What worries me a bit is the macro
> CMM_ACCESS_ONCE (inside _CMM_STORE_SHARED) that I have yet to grok.

CMM_ACCESS_ONCE is really just a cast through a volatile, to make sure
the compiler consider the data to store (or load) as being stored
(loaded) into (from) a volatile container. Nothing more, really. You
might want to refer to the Linux kernel comment above ACCESS_ONCE() to
see why it is needed.

Thanks,

Mathieu

>
> /Per

-- 
Mathieu Desnoyers
Operating System Efficiency R&D Consultant
EfficiOS Inc.
http://www.efficios.com