X-Git-Url: http://git.liburcu.org/?p=urcu.git;a=blobdiff_plain;f=urcu-bp.c;h=8207a4120ffd6b0790939e0ee176e064198704fe;hp=3af7c3b9712879b3b74481900b5ee615cae7f8e0;hb=882f335739b978d1c55be2faeed077f315afe5d7;hpb=79266051c59604c3923944d8003f8a9d2ad3c08b diff --git a/urcu-bp.c b/urcu-bp.c index 3af7c3b..8207a41 100644 --- a/urcu-bp.c +++ b/urcu-bp.c @@ -3,7 +3,7 @@ * * Userspace RCU library, "bulletproof" version. * - * 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 @@ -24,6 +24,7 @@ */ #define _GNU_SOURCE +#define _LGPL_SOURCE #include #include #include @@ -35,21 +36,77 @@ #include #include -#include "urcu-bp-static.h" +#include "urcu/wfcqueue.h" +#include "urcu/map/urcu-bp.h" +#include "urcu/static/urcu-bp.h" +#include "urcu-pointer.h" +#include "urcu/tls-compat.h" + +#include "urcu-die.h" + /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */ +#undef _LGPL_SOURCE #include "urcu-bp.h" +#define _LGPL_SOURCE + +#ifndef MAP_ANONYMOUS +#define MAP_ANONYMOUS MAP_ANON +#endif + +#ifdef __linux__ +static +void *mremap_wrapper(void *old_address, size_t old_size, + size_t new_size, int flags) +{ + return mremap(old_address, old_size, new_size, flags); +} +#else + +#define MREMAP_MAYMOVE 1 +#define MREMAP_FIXED 2 + +/* + * mremap wrapper for non-Linux systems. Maps a RW, anonymous private mapping. + * This is not generic. +*/ +static +void *mremap_wrapper(void *old_address, size_t old_size, + size_t new_size, int flags) +{ + void *new_address; + + assert(flags & MREMAP_MAYMOVE); + assert(!(flags & MREMAP_FIXED)); + new_address = mmap(old_address, new_size, + PROT_READ | PROT_WRITE, + MAP_ANONYMOUS | MAP_PRIVATE, + -1, 0); + if (new_address == MAP_FAILED) + return MAP_FAILED; + if (old_address) { + memcpy(new_address, old_address, old_size); + munmap(old_address, old_size); + } + return new_address; +} +#endif /* Sleep delay in us */ #define RCU_SLEEP_DELAY 1000 #define ARENA_INIT_ALLOC 16 +/* + * Active attempts to check for reader Q.S. before calling sleep(). + */ +#define RCU_QS_ACTIVE_ATTEMPTS 100 + void __attribute__((destructor)) rcu_bp_exit(void); static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER; #ifdef DEBUG_YIELD -unsigned int yield_active; -unsigned int __thread rand_yield; +unsigned int rcu_yield_active; +DEFINE_URCU_TLS(unsigned int, rcu_rand_yield); #endif /* @@ -58,15 +115,15 @@ unsigned int __thread rand_yield; * 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. */ -long rcu_gp_ctr = RCU_GP_COUNT; +unsigned long rcu_gp_ctr = RCU_GP_COUNT; /* * Pointer to registry elements. Written to only by each individual reader. Read * by both the reader and the writers. */ -struct rcu_reader __thread *rcu_reader; +DEFINE_URCU_TLS(struct rcu_reader *, rcu_reader); -static LIST_HEAD(registry); +static CDS_LIST_HEAD(registry); struct registry_arena { void *p; @@ -76,6 +133,9 @@ struct registry_arena { static struct registry_arena registry_arena; +/* Saved fork signal mask, protected by rcu_gp_lock */ +static sigset_t saved_fork_signal_mask; + static void rcu_gc_registry(void); static void mutex_lock(pthread_mutex_t *mutex) @@ -84,22 +144,12 @@ static void mutex_lock(pthread_mutex_t *mutex) #ifndef DISTRUST_SIGNALS_EXTREME ret = pthread_mutex_lock(mutex); - if (ret) { - perror("Error in pthread mutex lock"); - exit(-1); - } + if (ret) + urcu_die(ret); #else /* #ifndef DISTRUST_SIGNALS_EXTREME */ while ((ret = pthread_mutex_trylock(mutex)) != 0) { - if (ret != EBUSY && ret != EINTR) { - printf("ret = %d, errno = %d\n", ret, errno); - perror("Error in pthread mutex lock"); - exit(-1); - } - if (rcu_reader.need_mb) { - smp_mb(); - rcu_reader.need_mb = 0; - smp_mb(); - } + if (ret != EBUSY && ret != EINTR) + urcu_die(ret); poll(NULL,0,10); } #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */ @@ -110,60 +160,62 @@ static void mutex_unlock(pthread_mutex_t *mutex) int ret; ret = pthread_mutex_unlock(mutex); - if (ret) { - perror("Error in pthread mutex unlock"); - exit(-1); - } + if (ret) + urcu_die(ret); } -void update_counter_and_wait(void) +static void wait_for_readers(struct cds_list_head *input_readers, + struct cds_list_head *cur_snap_readers, + struct cds_list_head *qsreaders) { - LIST_HEAD(qsreaders); int wait_loops = 0; struct rcu_reader *index, *tmp; - /* Switch parity: 0 -> 1, 1 -> 0 */ - STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR_PHASE); - /* - * Must commit qparity update to memory before waiting for other parity - * quiescent state. Failure to do so could result in the writer waiting - * forever while new readers are always accessing data (no progress). - * Ensured by STORE_SHARED and LOAD_SHARED. - */ - - /* - * Adding a smp_mb() which is _not_ formally required, but makes the - * model easier to understand. It does not have a big performance impact - * anyway, given this is the write-side. - */ - smp_mb(); - - /* - * Wait for each thread rcu_reader.ctr count to become 0. + * Wait for each thread URCU_TLS(rcu_reader).ctr to either + * indicate quiescence (not nested), or observe the current + * rcu_gp_ctr value. */ for (;;) { wait_loops++; - list_for_each_entry_safe(index, tmp, ®istry, head) { - if (!rcu_old_gp_ongoing(&index->ctr)) - list_move(&index->head, &qsreaders); + cds_list_for_each_entry_safe(index, tmp, input_readers, node) { + switch (rcu_reader_state(&index->ctr)) { + case RCU_READER_ACTIVE_CURRENT: + if (cur_snap_readers) { + cds_list_move(&index->node, + cur_snap_readers); + break; + } + /* Fall-through */ + case RCU_READER_INACTIVE: + cds_list_move(&index->node, qsreaders); + break; + case RCU_READER_ACTIVE_OLD: + /* + * Old snapshot. Leaving node in + * input_readers will make us busy-loop + * until the snapshot becomes current or + * the reader becomes inactive. + */ + break; + } } - if (list_empty(®istry)) { + if (cds_list_empty(input_readers)) { break; } else { if (wait_loops == RCU_QS_ACTIVE_ATTEMPTS) usleep(RCU_SLEEP_DELAY); else - cpu_relax(); + caa_cpu_relax(); } } - /* put back the reader list in the registry */ - list_splice(&qsreaders, ®istry); } void synchronize_rcu(void) { + CDS_LIST_HEAD(cur_snap_readers); + CDS_LIST_HEAD(qsreaders); sigset_t newmask, oldmask; int ret; @@ -174,39 +226,61 @@ void synchronize_rcu(void) mutex_lock(&rcu_gp_lock); - if (list_empty(®istry)) + if (cds_list_empty(®istry)) goto out; /* All threads should read qparity before accessing data structure * where new ptr points to. */ /* Write new ptr before changing the qparity */ - smp_mb(); + cmm_smp_mb(); /* Remove old registry elements */ rcu_gc_registry(); /* - * Wait for previous parity to be empty of readers. + * Wait for readers to observe original parity or be quiescent. + */ + wait_for_readers(®istry, &cur_snap_readers, &qsreaders); + + /* + * Adding a cmm_smp_mb() which is _not_ formally required, but makes the + * model easier to understand. It does not have a big performance impact + * anyway, given this is the write-side. + */ + cmm_smp_mb(); + + /* Switch parity: 0 -> 1, 1 -> 0 */ + CMM_STORE_SHARED(rcu_gp_ctr, rcu_gp_ctr ^ RCU_GP_CTR_PHASE); + + /* + * Must commit qparity update to memory before waiting for other parity + * quiescent state. Failure to do so could result in the writer waiting + * forever while new readers are always accessing data (no progress). + * Ensured by CMM_STORE_SHARED and CMM_LOAD_SHARED. */ - update_counter_and_wait(); /* 0 -> 1, wait readers in parity 0 */ /* - * Adding a smp_mb() which is _not_ formally required, but makes the + * Adding a cmm_smp_mb() which is _not_ formally required, but makes the * model easier to understand. It does not have a big performance impact * anyway, given this is the write-side. */ - smp_mb(); + cmm_smp_mb(); + + /* + * Wait for readers to observe new parity or be quiescent. + */ + wait_for_readers(&cur_snap_readers, NULL, &qsreaders); /* - * Wait for previous parity to be empty of readers. + * Put quiescent reader list back into registry. */ - update_counter_and_wait(); /* 1 -> 0, wait readers in parity 1 */ + cds_list_splice(&qsreaders, ®istry); /* * Finish waiting for reader threads before letting the old ptr being * freed. */ - smp_mb(); + cmm_smp_mb(); out: mutex_unlock(&rcu_gp_lock); ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); @@ -227,6 +301,11 @@ void rcu_read_unlock(void) _rcu_read_unlock(); } +int rcu_read_ongoing(void) +{ + return _rcu_read_ongoing(); +} + /* * only grow for now. */ @@ -240,8 +319,8 @@ static void resize_arena(struct registry_arena *arena, size_t len) MAP_ANONYMOUS | MAP_PRIVATE, -1, 0); else - new_arena = mremap(arena->p, arena->len, - len, MREMAP_MAYMOVE); + new_arena = mremap_wrapper(arena->p, arena->len, + len, MREMAP_MAYMOVE); assert(new_arena != MAP_FAILED); /* @@ -250,7 +329,6 @@ static void resize_arena(struct registry_arena *arena, size_t len) if (new_arena == arena->p) return; - memcpy(new_arena, arena->p, arena->len); bzero(new_arena + arena->len, len - arena->len); arena->p = new_arena; } @@ -263,7 +341,7 @@ static void add_thread(void) if (registry_arena.len < registry_arena.used + sizeof(struct rcu_reader)) resize_arena(®istry_arena, - max(registry_arena.len << 1, ARENA_INIT_ALLOC)); + caa_max(registry_arena.len << 1, ARENA_INIT_ALLOC)); /* * Find a free spot. */ @@ -279,8 +357,8 @@ static void add_thread(void) /* Add to registry */ rcu_reader_reg->tid = pthread_self(); assert(rcu_reader_reg->ctr == 0); - list_add(&rcu_reader_reg->head, ®istry); - rcu_reader = rcu_reader_reg; + cds_list_add(&rcu_reader_reg->node, ®istry); + URCU_TLS(rcu_reader) = rcu_reader_reg; } /* Called with signals off and mutex locked */ @@ -299,7 +377,7 @@ static void rcu_gc_registry(void) ret = pthread_kill(tid, 0); assert(ret != EINVAL); if (ret == ESRCH) { - list_del(&rcu_reader_reg->head); + cds_list_del(&rcu_reader_reg->node); rcu_reader_reg->ctr = 0; rcu_reader_reg->alloc = 0; registry_arena.used -= sizeof(struct rcu_reader); @@ -321,7 +399,7 @@ void rcu_bp_register(void) /* * Check if a signal concurrently registered our thread since * the check in rcu_read_lock(). */ - if (rcu_reader) + if (URCU_TLS(rcu_reader)) goto end; mutex_lock(&rcu_gp_lock); @@ -332,7 +410,78 @@ end: assert(!ret); } -void rcu_bp_exit() +void rcu_bp_exit(void) { - munmap(registry_arena.p, registry_arena.len); + if (registry_arena.p) + munmap(registry_arena.p, registry_arena.len); } + +/* + * Holding the rcu_gp_lock across fork will make sure we fork() don't race with + * a concurrent thread executing with this same lock held. This ensures that the + * registry is in a coherent state in the child. + */ +void rcu_bp_before_fork(void) +{ + sigset_t newmask, oldmask; + int ret; + + ret = sigemptyset(&newmask); + assert(!ret); + ret = pthread_sigmask(SIG_SETMASK, &newmask, &oldmask); + assert(!ret); + mutex_lock(&rcu_gp_lock); + saved_fork_signal_mask = oldmask; +} + +void rcu_bp_after_fork_parent(void) +{ + sigset_t oldmask; + int ret; + + oldmask = saved_fork_signal_mask; + mutex_unlock(&rcu_gp_lock); + ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); + assert(!ret); +} + +void rcu_bp_after_fork_child(void) +{ + sigset_t oldmask; + int ret; + + rcu_gc_registry(); + oldmask = saved_fork_signal_mask; + mutex_unlock(&rcu_gp_lock); + ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); + assert(!ret); +} + +void *rcu_dereference_sym_bp(void *p) +{ + return _rcu_dereference(p); +} + +void *rcu_set_pointer_sym_bp(void **p, void *v) +{ + cmm_wmb(); + uatomic_set(p, v); + return v; +} + +void *rcu_xchg_pointer_sym_bp(void **p, void *v) +{ + cmm_wmb(); + return uatomic_xchg(p, v); +} + +void *rcu_cmpxchg_pointer_sym_bp(void **p, void *old, void *_new) +{ + cmm_wmb(); + return uatomic_cmpxchg(p, old, _new); +} + +DEFINE_RCU_FLAVOR(rcu_flavor); + +#include "urcu-call-rcu-impl.h" +#include "urcu-defer-impl.h"