X-Git-Url: http://git.liburcu.org/?p=urcu.git;a=blobdiff_plain;f=urcu-static.h;h=18e4826a5a31789c7cb408832421cbddb7d83c22;hp=dec8a204166d2f2c9848227a7629c53057164958;hb=acfd099eae9026369f558e73b2a7e35860962d9d;hpb=36bc70a84250927ba68d5096a0a9740aec157f9b diff --git a/urcu-static.h b/urcu-static.h index dec8a20..18e4826 100644 --- a/urcu-static.h +++ b/urcu-static.h @@ -9,7 +9,7 @@ * TO BE INCLUDED ONLY IN LGPL-COMPATIBLE CODE. See urcu.h for linking * dynamically with the userspace rcu library. * - * Copyright (c) 2009 Mathieu Desnoyers + * Copyright (c) 2009 Mathieu Desnoyers * Copyright (c) 2009 Paul E. McKenney, IBM Corporation. * * This library is free software; you can redistribute it and/or @@ -45,6 +45,27 @@ extern "C" { #endif +/* Default is RCU_MEMBARRIER */ +#if !defined(RCU_MEMBARRIER) && !defined(RCU_MB) && !defined(RCU_SIGNAL) +#define RCU_MEMBARRIER +#endif + +#ifdef RCU_MEMBARRIER +#include +#include + +/* If the headers do not support SYS_membarrier, statically use RCU_MB */ +#ifdef SYS_membarrier +# define MEMBARRIER_EXPEDITED (1 << 0) +# define MEMBARRIER_DELAYED (1 << 1) +# define MEMBARRIER_QUERY (1 << 16) +# define membarrier(...) syscall(SYS_membarrier, __VA_ARGS__) +#else +# undef RCU_MEMBARRIER +# define RCU_MB +#endif +#endif + /* * This code section can only be included in LGPL 2.1 compatible source code. * See below for the function call wrappers which can be used in code meant to @@ -55,15 +76,20 @@ extern "C" { /* * The signal number used by the RCU library can be overridden with - * -DSIGURCU= when compiling the library. + * -DSIGRCU= when compiling the library. + * Provide backward compatibility for liburcu 0.3.x SIGURCU. */ -#ifndef SIGURCU -#define SIGURCU SIGUSR1 +#ifdef SIGURCU +#define SIGRCU SIGURCU +#endif + +#ifndef SIGRCU +#define SIGRCU SIGUSR1 #endif /* * If a reader is really non-cooperative and refuses to commit its - * urcu_active_readers count to memory (there is no barrier in the reader + * 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 @@ -89,14 +115,13 @@ extern "C" { #define YIELD_WRITE (1 << 1) /* - * Updates without URCU_MB are much slower. Account this in - * the delay. + * Updates with RCU_SIGNAL are much slower. Account this in the delay. */ -#ifdef URCU_MB +#ifdef RCU_SIGNAL /* maximum sleep delay, in us */ -#define MAX_SLEEP 50 -#else #define MAX_SLEEP 30000 +#else +#define MAX_SLEEP 50 #endif extern unsigned int yield_active; @@ -135,44 +160,73 @@ static inline void debug_yield_init(void) } #endif -#ifdef URCU_MB -static inline void reader_barrier() +/* + * RCU memory barrier broadcast group. Currently, only broadcast to all process + * threads is supported (group 0). + * + * Slave barriers are only guaranteed to be ordered wrt master barriers. + * + * The pair ordering is detailed as (O: ordered, X: not ordered) : + * slave master + * slave X O + * master O O + */ + +#define MB_GROUP_ALL 0 +#define RCU_MB_GROUP MB_GROUP_ALL + +#ifdef RCU_MEMBARRIER +extern int has_sys_membarrier; + +static inline void smp_mb_slave(int group) +{ + if (likely(has_sys_membarrier)) + cmm_barrier(); + else + cmm_smp_mb(); +} +#endif + +#ifdef RCU_MB +static inline void smp_mb_slave(int group) { - smp_mb(); + cmm_smp_mb(); } -#else -static inline void reader_barrier() +#endif + +#ifdef RCU_SIGNAL +static inline void smp_mb_slave(int group) { - barrier(); + cmm_barrier(); } #endif /* - * The trick here is that RCU_GP_CTR_BIT 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. + * 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. */ #define RCU_GP_COUNT (1UL << 0) /* Use the amount of bits equal to half of the architecture long size */ -#define RCU_GP_CTR_BIT (1UL << (sizeof(long) << 2)) -#define RCU_GP_CTR_NEST_MASK (RCU_GP_CTR_BIT - 1) +#define RCU_GP_CTR_PHASE (1UL << (sizeof(unsigned long) << 2)) +#define RCU_GP_CTR_NEST_MASK (RCU_GP_CTR_PHASE - 1) /* * 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 long urcu_gp_ctr; +extern unsigned long rcu_gp_ctr; -struct urcu_reader { +struct rcu_reader { /* Data used by both reader and synchronize_rcu() */ - long ctr; + unsigned long ctr; char need_mb; /* Data used for registry */ - struct list_head head __attribute__((aligned(CACHE_LINE_SIZE))); + struct cds_list_head node __attribute__((aligned(CAA_CACHE_LINE_SIZE))); pthread_t tid; }; -extern struct urcu_reader __thread urcu_reader; +extern struct rcu_reader __thread rcu_reader; extern int gp_futex; @@ -188,60 +242,63 @@ static inline void wake_up_gp(void) } } -static inline int rcu_old_gp_ongoing(long *value) +static inline int rcu_gp_ongoing(unsigned long *ctr) { - long v; + unsigned long v; - if (value == NULL) - return 0; /* * Make sure both tests below are done on the same version of *value * to insure consistency. */ - v = LOAD_SHARED(*value); + v = CMM_LOAD_SHARED(*ctr); return (v & RCU_GP_CTR_NEST_MASK) && - ((v ^ urcu_gp_ctr) & RCU_GP_CTR_BIT); + ((v ^ rcu_gp_ctr) & RCU_GP_CTR_PHASE); } static inline void _rcu_read_lock(void) { - long tmp; + unsigned long tmp; - tmp = urcu_reader.ctr; - /* urcu_gp_ctr = RCU_GP_COUNT | (~RCU_GP_CTR_BIT or RCU_GP_CTR_BIT) */ + 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))) { - _STORE_SHARED(urcu_reader.ctr, _LOAD_SHARED(urcu_gp_ctr)); + _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 force_mb_all_threads(). + * accessing the pointer. See smp_mb_master(). */ - reader_barrier(); + smp_mb_slave(RCU_MB_GROUP); } else { - _STORE_SHARED(urcu_reader.ctr, tmp + RCU_GP_COUNT); + _CMM_STORE_SHARED(rcu_reader.ctr, tmp + RCU_GP_COUNT); } } static inline void _rcu_read_unlock(void) { - long tmp; + unsigned long tmp; - tmp = urcu_reader.ctr; + tmp = rcu_reader.ctr; /* * Finish using rcu before decrementing the pointer. - * See force_mb_all_threads(). + * See smp_mb_master(). */ if (likely((tmp & RCU_GP_CTR_NEST_MASK) == RCU_GP_COUNT)) { - reader_barrier(); - _STORE_SHARED(urcu_reader.ctr, urcu_reader.ctr - RCU_GP_COUNT); - /* write urcu_reader.ctr before read futex */ - reader_barrier(); + 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 { - _STORE_SHARED(urcu_reader.ctr, urcu_reader.ctr - RCU_GP_COUNT); + _CMM_STORE_SHARED(rcu_reader.ctr, rcu_reader.ctr - RCU_GP_COUNT); } + cmm_barrier(); /* Ensure the compiler does not reorder us with mutex */ } -#ifdef __cplusplus +#ifdef __cplusplus } #endif