*
* 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 <mathieu.desnoyers@efficios.com>
* Copyright (c) 2009 Paul E. McKenney, IBM Corporation.
#include <urcu/uatomic.h>
#include <urcu/list.h>
#include <urcu/futex.h>
+#include <urcu/tls-compat.h>
+#include <urcu/rand-compat.h>
+#include <urcu/debug.h>
#ifdef __cplusplus
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 <syscall.h>
-
-/* 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
#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
-
-#ifdef DEBUG_RCU
-#define rcu_assert(args...) assert(args)
-#else
-#define rcu_assert(args...)
-#endif
-
-#ifdef DEBUG_YIELD
-#include <sched.h>
-#include <time.h>
-#include <pthread.h>
-#include <unistd.h>
-
-#define YIELD_READ (1 << 0)
-#define YIELD_WRITE (1 << 1)
-
-/*
- * Updates with RCU_SIGNAL are much slower. Account this in the delay.
- */
-#ifdef RCU_SIGNAL
-/* maximum sleep delay, in us */
-#define MAX_SLEEP 30000
-#else
-#define MAX_SLEEP 50
-#endif
-
-extern unsigned int yield_active;
-extern unsigned int __thread rand_yield;
-
-static inline void debug_yield_read(void)
-{
- if (yield_active & YIELD_READ)
- if (rand_r(&rand_yield) & 0x1)
- usleep(rand_r(&rand_yield) % MAX_SLEEP);
-}
-
-static inline void debug_yield_write(void)
-{
- if (yield_active & YIELD_WRITE)
- if (rand_r(&rand_yield) & 0x1)
- usleep(rand_r(&rand_yield) % MAX_SLEEP);
-}
-
-static inline void debug_yield_init(void)
-{
- rand_yield = time(NULL) ^ (unsigned long) pthread_self();
-}
-#else
-static inline void debug_yield_read(void)
-{
-}
-
-static inline void debug_yield_write(void)
-{
-}
-
-static inline void debug_yield_init(void)
-{
-
-}
-#endif
+enum rcu_state {
+ RCU_READER_ACTIVE_CURRENT,
+ RCU_READER_ACTIVE_OLD,
+ RCU_READER_INACTIVE,
+};
/*
- * 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) :
* master O O
*/
-#define MB_GROUP_ALL 0
-#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)
+static inline void smp_mb_slave(void)
{
- if (likely(has_sys_membarrier))
+ if (caa_likely(rcu_has_sys_membarrier))
cmm_barrier();
else
cmm_smp_mb();
#endif
#ifdef RCU_MB
-static inline void smp_mb_slave(int group)
+static inline void smp_mb_slave(void)
{
cmm_smp_mb();
}
#endif
#ifdef RCU_SIGNAL
-static inline void smp_mb_slave(int group)
+static inline void smp_mb_slave(void)
{
cmm_barrier();
}
#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 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;
+
+ int32_t futex;
+} __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
+
+extern struct rcu_gp rcu_gp;
struct rcu_reader {
/* Data used by both reader and synchronize_rcu() */
/* Data used for registry */
struct cds_list_head node __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
pthread_t tid;
+ /* Reader registered flag, for internal checks. */
+ unsigned int registered:1;
};
-extern struct rcu_reader __thread rcu_reader;
-
-extern int32_t gp_futex;
+extern DECLARE_URCU_TLS(struct rcu_reader, rcu_reader);
/*
* 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,
- NULL, NULL, 0);
+ if (caa_unlikely(uatomic_read(&rcu_gp.futex) == -1)) {
+ uatomic_set(&rcu_gp.futex, 0);
+ /*
+ * Ignoring return value until we can make this function
+ * return something (because urcu_die() is not publicly
+ * exposed).
+ */
+ (void) 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;
* 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();
+ } 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().
- */
- smp_mb_slave(RCU_MB_GROUP);
- } else {
- _CMM_STORE_SHARED(rcu_reader.ctr, tmp + RCU_GP_COUNT);
- }
+ urcu_assert(URCU_TLS(rcu_reader).registered);
+ cmm_barrier();
+ tmp = URCU_TLS(rcu_reader).ctr;
+ urcu_assert((tmp & RCU_GP_CTR_NEST_MASK) != RCU_GP_CTR_NEST_MASK);
+ _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();
+ _CMM_STORE_SHARED(URCU_TLS(rcu_reader).ctr, tmp - RCU_GP_COUNT);
+ smp_mb_slave();
+ wake_up_gp();
+ } else
+ _CMM_STORE_SHARED(URCU_TLS(rcu_reader).ctr, tmp - 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);
- }
+ urcu_assert(URCU_TLS(rcu_reader).registered);
+ tmp = URCU_TLS(rcu_reader).ctr;
+ urcu_assert(tmp & RCU_GP_CTR_NEST_MASK);
+ _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
projects / urcu.git / blobdiff