X-Git-Url: http://git.liburcu.org/?p=urcu.git;a=blobdiff_plain;f=urcu%2Fstatic%2Furcu.h;h=6c8b92602234b97d27a4cb462ec0ec06c56a62c7;hp=289827248c8405aa41918b6b7619dc80f0bd7489;hb=882f335739b978d1c55be2faeed077f315afe5d7;hpb=a2e7bf9ce5de5113c7f59c380b0087e291cd603d diff --git a/urcu/static/urcu.h b/urcu/static/urcu.h index 2898272..6c8b926 100644 --- a/urcu/static/urcu.h +++ b/urcu/static/urcu.h @@ -6,8 +6,8 @@ * * Userspace RCU header. * - * TO BE INCLUDED ONLY IN LGPL-COMPATIBLE CODE. See urcu.h for linking - * dynamically with the userspace rcu library. + * TO BE INCLUDED ONLY IN CODE THAT IS TO BE RECOMPILED ON EACH LIBURCU + * RELEASE. See urcu.h for linking dynamically with the userspace rcu library. * * Copyright (c) 2009 Mathieu Desnoyers * Copyright (c) 2009 Paul E. McKenney, IBM Corporation. @@ -31,15 +31,16 @@ #include #include -#include #include +#include #include #include #include #include #include -#include +#include +#include #ifdef __cplusplus extern "C" { @@ -50,9 +51,17 @@ extern "C" { #define RCU_MEMBARRIER #endif +/* + * RCU_MEMBARRIER is only possibly available on Linux. Fallback to RCU_MB + * otherwise. + */ +#if !defined(__linux__) && defined(RCU_MEMBARRIER) +#undef RCU_MEMBARRIER +#define RCU_MB +#endif + #ifdef RCU_MEMBARRIER -#include -#include +#include /* If the headers do not support SYS_membarrier, statically use RCU_MB */ #ifdef SYS_membarrier @@ -87,17 +96,11 @@ extern "C" { #define SIGRCU SIGUSR1 #endif -/* - * If a reader is really non-cooperative and refuses to commit its - * rcu_active_readers count to memory (there is no barrier in the reader - * per-se), kick it after a few loops waiting for it. - */ -#define KICK_READER_LOOPS 10000 - -/* - * Active attempts to check for reader Q.S. before calling futex(). - */ -#define RCU_QS_ACTIVE_ATTEMPTS 100 +enum rcu_state { + RCU_READER_ACTIVE_CURRENT, + RCU_READER_ACTIVE_OLD, + RCU_READER_INACTIVE, +}; #ifdef DEBUG_RCU #define rcu_assert(args...) assert(args) @@ -111,8 +114,8 @@ extern "C" { #include #include -#define YIELD_READ (1 << 0) -#define YIELD_WRITE (1 << 1) +#define RCU_YIELD_READ (1 << 0) +#define RCU_YIELD_WRITE (1 << 1) /* * Updates with RCU_SIGNAL are much slower. Account this in the delay. @@ -124,37 +127,37 @@ extern "C" { #define MAX_SLEEP 50 #endif -extern unsigned int yield_active; -extern unsigned int __thread rand_yield; +extern unsigned int rcu_yield_active; +extern DECLARE_URCU_TLS(unsigned int, rcu_rand_yield); -static inline void debug_yield_read(void) +static inline void rcu_debug_yield_read(void) { - if (yield_active & YIELD_READ) - if (rand_r(&rand_yield) & 0x1) - usleep(rand_r(&rand_yield) % MAX_SLEEP); + if (rcu_yield_active & RCU_YIELD_READ) + if (rand_r(&URCU_TLS(rcu_rand_yield)) & 0x1) + usleep(rand_r(&URCU_TLS(rcu_rand_yield)) % MAX_SLEEP); } -static inline void debug_yield_write(void) +static inline void rcu_debug_yield_write(void) { - if (yield_active & YIELD_WRITE) - if (rand_r(&rand_yield) & 0x1) - usleep(rand_r(&rand_yield) % MAX_SLEEP); + if (rcu_yield_active & RCU_YIELD_WRITE) + if (rand_r(&URCU_TLS(rcu_rand_yield)) & 0x1) + usleep(rand_r(&URCU_TLS(rcu_rand_yield)) % MAX_SLEEP); } -static inline void debug_yield_init(void) +static inline void rcu_debug_yield_init(void) { - rand_yield = time(NULL) ^ pthread_self(); + URCU_TLS(rcu_rand_yield) = time(NULL) ^ (unsigned long) pthread_self(); } #else -static inline void debug_yield_read(void) +static inline void rcu_debug_yield_read(void) { } -static inline void debug_yield_write(void) +static inline void rcu_debug_yield_write(void) { } -static inline void debug_yield_init(void) +static inline void rcu_debug_yield_init(void) { } @@ -176,11 +179,11 @@ static inline void debug_yield_init(void) #define RCU_MB_GROUP MB_GROUP_ALL #ifdef RCU_MEMBARRIER -extern int has_sys_membarrier; +extern int rcu_has_sys_membarrier; static inline void smp_mb_slave(int group) { - if (likely(has_sys_membarrier)) + if (caa_likely(rcu_has_sys_membarrier)) cmm_barrier(); else cmm_smp_mb(); @@ -226,23 +229,23 @@ struct rcu_reader { pthread_t tid; }; -extern struct rcu_reader __thread rcu_reader; +extern DECLARE_URCU_TLS(struct rcu_reader, rcu_reader); -extern int gp_futex; +extern int32_t rcu_gp_futex; /* * Wake-up waiting synchronize_rcu(). Called from many concurrent threads. */ static inline void wake_up_gp(void) { - if (unlikely(uatomic_read(&gp_futex) == -1)) { - uatomic_set(&gp_futex, 0); - futex_async(&gp_futex, FUTEX_WAKE, 1, + if (caa_unlikely(uatomic_read(&rcu_gp_futex) == -1)) { + uatomic_set(&rcu_gp_futex, 0); + futex_async(&rcu_gp_futex, FUTEX_WAKE, 1, NULL, NULL, 0); } } -static inline int rcu_gp_ongoing(unsigned long *ctr) +static inline enum rcu_state rcu_reader_state(unsigned long *ctr) { unsigned long v; @@ -251,53 +254,93 @@ static inline int rcu_gp_ongoing(unsigned long *ctr) * to insure consistency. */ v = CMM_LOAD_SHARED(*ctr); - return (v & RCU_GP_CTR_NEST_MASK) && - ((v ^ rcu_gp_ctr) & RCU_GP_CTR_PHASE); + if (!(v & RCU_GP_CTR_NEST_MASK)) + return RCU_READER_INACTIVE; + if (!((v ^ rcu_gp_ctr) & RCU_GP_CTR_PHASE)) + return RCU_READER_ACTIVE_CURRENT; + return RCU_READER_ACTIVE_OLD; } +/* + * Helper for _rcu_read_lock(). The format of rcu_gp_ctr (as well as + * the per-thread rcu_reader.ctr) has the upper bits containing a count of + * _rcu_read_lock() nesting, and a lower-order bit that contains either zero + * or RCU_GP_CTR_PHASE. The smp_mb_slave() ensures that the accesses in + * _rcu_read_lock() happen before the subsequent read-side critical section. + */ +static inline void _rcu_read_lock_update(unsigned long tmp) +{ + if (caa_likely(!(tmp & RCU_GP_CTR_NEST_MASK))) { + _CMM_STORE_SHARED(URCU_TLS(rcu_reader).ctr, _CMM_LOAD_SHARED(rcu_gp_ctr)); + smp_mb_slave(RCU_MB_GROUP); + } else + _CMM_STORE_SHARED(URCU_TLS(rcu_reader).ctr, tmp + RCU_GP_COUNT); +} + +/* + * Enter an RCU read-side critical section. + * + * The first cmm_barrier() call ensures that the compiler does not reorder + * the body of _rcu_read_lock() with a mutex. + * + * This function and its helper are both less than 10 lines long. The + * intent is that this function meets the 10-line criterion in LGPL, + * allowing this function to be invoked directly from non-LGPL code. + */ static inline void _rcu_read_lock(void) { unsigned long tmp; - cmm_barrier(); /* Ensure the compiler does not reorder us with mutex */ - tmp = rcu_reader.ctr; - /* - * rcu_gp_ctr is - * RCU_GP_COUNT | (~RCU_GP_CTR_PHASE or RCU_GP_CTR_PHASE) - */ - if (likely(!(tmp & RCU_GP_CTR_NEST_MASK))) { - _CMM_STORE_SHARED(rcu_reader.ctr, _CMM_LOAD_SHARED(rcu_gp_ctr)); - /* - * Set active readers count for outermost nesting level before - * accessing the pointer. See smp_mb_master(). - */ + cmm_barrier(); + tmp = URCU_TLS(rcu_reader).ctr; + _rcu_read_lock_update(tmp); +} + +/* + * This is a helper function for _rcu_read_unlock(). + * + * The first smp_mb_slave() call ensures that the critical section is + * seen to precede the store to rcu_reader.ctr. + * The second smp_mb_slave() call ensures that we write to rcu_reader.ctr + * before reading the update-side futex. + */ +static inline void _rcu_read_unlock_update_and_wakeup(unsigned long tmp) +{ + if (caa_likely((tmp & RCU_GP_CTR_NEST_MASK) == RCU_GP_COUNT)) { smp_mb_slave(RCU_MB_GROUP); - } else { - _CMM_STORE_SHARED(rcu_reader.ctr, tmp + RCU_GP_COUNT); - } + _CMM_STORE_SHARED(URCU_TLS(rcu_reader).ctr, URCU_TLS(rcu_reader).ctr - RCU_GP_COUNT); + smp_mb_slave(RCU_MB_GROUP); + wake_up_gp(); + } else + _CMM_STORE_SHARED(URCU_TLS(rcu_reader).ctr, URCU_TLS(rcu_reader).ctr - RCU_GP_COUNT); } +/* + * Exit an RCU read-side crtical section. Both this function and its + * helper are smaller than 10 lines of code, and are intended to be + * usable by non-LGPL code, as called out in LGPL. + */ static inline void _rcu_read_unlock(void) { unsigned long tmp; - tmp = rcu_reader.ctr; - /* - * Finish using rcu before decrementing the pointer. - * See smp_mb_master(). - */ - if (likely((tmp & RCU_GP_CTR_NEST_MASK) == RCU_GP_COUNT)) { - smp_mb_slave(RCU_MB_GROUP); - _CMM_STORE_SHARED(rcu_reader.ctr, rcu_reader.ctr - RCU_GP_COUNT); - /* write rcu_reader.ctr before read futex */ - smp_mb_slave(RCU_MB_GROUP); - wake_up_gp(); - } else { - _CMM_STORE_SHARED(rcu_reader.ctr, rcu_reader.ctr - RCU_GP_COUNT); - } + tmp = URCU_TLS(rcu_reader).ctr; + _rcu_read_unlock_update_and_wakeup(tmp); cmm_barrier(); /* Ensure the compiler does not reorder us with mutex */ } +/* + * Returns whether within a RCU read-side critical section. + * + * This function is less than 10 lines long. The intent is that this + * function meets the 10-line criterion for LGPL, allowing this function + * to be invoked directly from non-LGPL code. + */ +static inline int _rcu_read_ongoing(void) +{ + return URCU_TLS(rcu_reader).ctr & RCU_GP_CTR_NEST_MASK; +} + #ifdef __cplusplus } #endif