X-Git-Url: https://git.liburcu.org/?p=urcu.git;a=blobdiff_plain;f=urcu%2Fstatic%2Furcu-bp.h;h=6e52d84bafe84f75dfd45593a2b58cad6431bcd0;hp=2f365206afad6db4c561654109def617ef559ce5;hb=67ecffc0f530a7b5c4dd5111ea7dd3213da8eb91;hpb=882f335739b978d1c55be2faeed077f315afe5d7 diff --git a/urcu/static/urcu-bp.h b/urcu/static/urcu-bp.h index 2f36520..6e52d84 100644 --- a/urcu/static/urcu-bp.h +++ b/urcu/static/urcu-bp.h @@ -39,6 +39,7 @@ #include #include #include +#include /* * This code section can only be included in LGPL 2.1 compatible source code. @@ -52,70 +53,12 @@ extern "C" { #endif -#ifdef DEBUG_RCU -#define rcu_assert(args...) assert(args) -#else -#define rcu_assert(args...) -#endif - enum rcu_state { RCU_READER_ACTIVE_CURRENT, RCU_READER_ACTIVE_OLD, RCU_READER_INACTIVE, }; -#ifdef DEBUG_YIELD -#include -#include -#include -#include - -#define RCU_YIELD_READ (1 << 0) -#define RCU_YIELD_WRITE (1 << 1) - -/* - * Updates without RCU_MB are much slower. Account this in - * the delay. - */ -/* maximum sleep delay, in us */ -#define MAX_SLEEP 50 - -extern unsigned int rcu_yield_active; -extern DECLARE_URCU_TLS(unsigned int, rcu_rand_yield); - -static inline void rcu_debug_yield_read(void) -{ - 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 rcu_debug_yield_write(void) -{ - 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 rcu_debug_yield_init(void) -{ - URCU_TLS(rcu_rand_yield) = time(NULL) ^ (unsigned long) pthread_self(); -} -#else -static inline void rcu_debug_yield_read(void) -{ -} - -static inline void rcu_debug_yield_write(void) -{ -} - -static inline void rcu_debug_yield_init(void) -{ - -} -#endif - /* * The trick here is that RCU_GP_CTR_PHASE must be a multiple of 8 so we can use a * full 8-bits, 16-bits or 32-bits bitmask for the lower order bits. @@ -130,12 +73,18 @@ static inline void rcu_debug_yield_init(void) */ extern void rcu_bp_register(void); -/* - * Global quiescent period counter with low-order bits unused. - * Using a int rather than a char to eliminate false register dependencies - * causing stalls on some architectures. - */ -extern unsigned long rcu_gp_ctr; +struct rcu_gp { + /* + * Global grace period counter. + * Contains the current RCU_GP_CTR_PHASE. + * Also has a RCU_GP_COUNT of 1, to accelerate the reader fast path. + * Written to only by writer with mutex taken. + * Read by both writer and readers. + */ + unsigned long ctr; +} __attribute__((aligned(CAA_CACHE_LINE_SIZE))); + +extern struct rcu_gp rcu_gp; struct rcu_reader { /* Data used by both reader and synchronize_rcu() */ @@ -153,6 +102,16 @@ struct rcu_reader { */ extern DECLARE_URCU_TLS(struct rcu_reader *, rcu_reader); +extern int urcu_bp_has_sys_membarrier; + +static inline void urcu_bp_smp_mb_slave(void) +{ + if (caa_likely(urcu_bp_has_sys_membarrier)) + cmm_barrier(); + else + cmm_smp_mb(); +} + static inline enum rcu_state rcu_reader_state(unsigned long *ctr) { unsigned long v; @@ -166,13 +125,13 @@ static inline enum rcu_state rcu_reader_state(unsigned long *ctr) v = CMM_LOAD_SHARED(*ctr); if (!(v & RCU_GP_CTR_NEST_MASK)) return RCU_READER_INACTIVE; - if (!((v ^ rcu_gp_ctr) & RCU_GP_CTR_PHASE)) + 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 + * 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 @@ -181,8 +140,8 @@ static inline enum rcu_state rcu_reader_state(unsigned long *ctr) 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)); - cmm_smp_mb(); + _CMM_STORE_SHARED(URCU_TLS(rcu_reader)->ctr, _CMM_LOAD_SHARED(rcu_gp.ctr)); + urcu_bp_smp_mb_slave(); } else _CMM_STORE_SHARED(URCU_TLS(rcu_reader)->ctr, tmp + RCU_GP_COUNT); } @@ -205,6 +164,7 @@ static inline void _rcu_read_lock(void) rcu_bp_register(); /* If not yet registered. */ cmm_barrier(); /* Ensure the compiler does not reorder us with mutex */ tmp = URCU_TLS(rcu_reader)->ctr; + urcu_assert((tmp & RCU_GP_CTR_NEST_MASK) != RCU_GP_CTR_NEST_MASK); _rcu_read_lock_update(tmp); } @@ -215,11 +175,13 @@ static inline void _rcu_read_lock(void) */ static inline void _rcu_read_unlock(void) { - /* - * Finish using rcu before decrementing the pointer. - */ - cmm_smp_mb(); - _CMM_STORE_SHARED(URCU_TLS(rcu_reader)->ctr, URCU_TLS(rcu_reader)->ctr - RCU_GP_COUNT); + unsigned long tmp; + + tmp = URCU_TLS(rcu_reader)->ctr; + urcu_assert(tmp & RCU_GP_CTR_NEST_MASK); + /* Finish using rcu before decrementing the pointer. */ + urcu_bp_smp_mb_slave(); + _CMM_STORE_SHARED(URCU_TLS(rcu_reader)->ctr, tmp - RCU_GP_COUNT); cmm_barrier(); /* Ensure the compiler does not reorder us with mutex */ } @@ -237,7 +199,7 @@ static inline int _rcu_read_ongoing(void) return URCU_TLS(rcu_reader)->ctr & RCU_GP_CTR_NEST_MASK; } -#ifdef __cplusplus +#ifdef __cplusplus } #endif