[urcu.git] / urcu.h

#ifndef _URCU_H
#define _URCU_H

/*
 * urcu.h
 *
 * Userspace RCU header
 *
 * Copyright February 2009 - Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
 *
 * Credits for Paul e. McKenney <paulmck@linux.vnet.ibm.com>
 * for inspiration coming from the Linux kernel RCU and rcu-preempt.
 *
 * The barrier, mb, rmb, wmb, atomic_inc, smp_read_barrier_depends, ACCESS_ONCE
 * and rcu_dereference primitives come from the Linux kernel.
 *
 * Distributed under GPLv2
 */

#define __USE_GNU
#include <stdlib.h>

/* The "volatile" is due to gcc bugs */
#define barrier() __asm__ __volatile__("": : :"memory")

/* x86 32/64 specific */
#define mb()    asm volatile("mfence":::"memory")
#define rmb()   asm volatile("lfence":::"memory")
#define wmb()   asm volatile("sfence" ::: "memory")

static inline void atomic_inc(int *v)
{
	asm volatile("lock; incl %0"
		     : "+m" (*v));
}

#define xchg(ptr, v)							\
	((__typeof__(*(ptr)))__xchg((unsigned long)(v), (ptr), sizeof(*(ptr))))

struct __xchg_dummy {
	unsigned long a[100];
};
#define __xg(x) ((struct __xchg_dummy *)(x))

/*
 * Note: no "lock" prefix even on SMP: xchg always implies lock anyway
 * Note 2: xchg has side effect, so that attribute volatile is necessary,
 *	  but generally the primitive is invalid, *ptr is output argument. --ANK
 */
static inline unsigned long __xchg(unsigned long x, volatile void *ptr,
				   int size)
{
	switch (size) {
	case 1:
		asm volatile("xchgb %b0,%1"
			     : "=q" (x)
			     : "m" (*__xg(ptr)), "0" (x)
			     : "memory");
		break;
	case 2:
		asm volatile("xchgw %w0,%1"
			     : "=r" (x)
			     : "m" (*__xg(ptr)), "0" (x)
			     : "memory");
		break;
	case 4:
		asm volatile("xchgl %0,%1"
			     : "=r" (x)
			     : "m" (*__xg(ptr)), "0" (x)
			     : "memory");
		break;
	}
	return x;
}

/* Nop everywhere except on alpha. */
#define smp_read_barrier_depends()

/*
 * Prevent the compiler from merging or refetching accesses.  The compiler
 * is also forbidden from reordering successive instances of ACCESS_ONCE(),
 * but only when the compiler is aware of some particular ordering.  One way
 * to make the compiler aware of ordering is to put the two invocations of
 * ACCESS_ONCE() in different C statements.
 *
 * This macro does absolutely -nothing- to prevent the CPU from reordering,
 * merging, or refetching absolutely anything at any time.  Its main intended
 * use is to mediate communication between process-level code and irq/NMI
 * handlers, all running on the same CPU.
 */
#define ACCESS_ONCE(x) (*(volatile typeof(x) *)&(x))

/**
 * rcu_dereference - fetch an RCU-protected pointer in an
 * RCU read-side critical section.  This pointer may later
 * be safely dereferenced.
 *
 * Inserts memory barriers on architectures that require them
 * (currently only the Alpha), and, more importantly, documents
 * exactly which pointers are protected by RCU.
 */

#define rcu_dereference(p)     ({ \
				typeof(p) _________p1 = ACCESS_ONCE(p); \
				smp_read_barrier_depends(); \
				(_________p1); \
				})

#define SIGURCU SIGUSR1

#ifdef DEBUG_YIELD
#include <sched.h>
#include <time.h>
#include <pthread.h>

#define YIELD_READ 	(1 << 0)
#define YIELD_WRITE	(1 << 1)

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)
			sched_yield();
}

static inline void debug_yield_write(void)
{
	if (yield_active & YIELD_WRITE)
		if (rand_r(&rand_yield) & 0x1)
			sched_yield();
}

static inline void debug_yield_init(void)
{
	rand_yield = time(NULL) ^ 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

/*
 * Limiting the nesting level to 256 to keep instructions small in the read
 * fast-path.
 */
#define RCU_GP_COUNT		(1U << 0)
#define RCU_GP_CTR_BIT		(1U << 8)
#define RCU_GP_CTR_NEST_MASK	(RCU_GP_CTR_BIT - 1)

/* Global quiescent period counter with low-order bits unused. */
extern int urcu_gp_ctr;

extern int __thread urcu_active_readers;

static inline int rcu_old_gp_ongoing(int *value)
{
	int v;

	if (value == NULL)
		return 0;
	debug_yield_write();
	v = ACCESS_ONCE(*value);
	debug_yield_write();
	return (v & RCU_GP_CTR_NEST_MASK) &&
		 ((v ^ ACCESS_ONCE(urcu_gp_ctr)) & RCU_GP_CTR_BIT);
}

static inline void rcu_read_lock(void)
{
	int tmp;

	debug_yield_read();
	tmp = urcu_active_readers;
	debug_yield_read();
	if (!(tmp & RCU_GP_CTR_NEST_MASK))
		urcu_active_readers = urcu_gp_ctr + RCU_GP_COUNT;
	else
		urcu_active_readers = tmp + RCU_GP_COUNT;
	debug_yield_read();
	/*
	 * Increment active readers count before accessing the pointer.
	 * See force_mb_all_threads().
	 */
	barrier();
	debug_yield_read();
}

static inline void rcu_read_unlock(void)
{
	debug_yield_read();
	barrier();
	debug_yield_read();
	/*
	 * Finish using rcu before decrementing the pointer.
	 * See force_mb_all_threads().
	 */
	urcu_active_readers -= RCU_GP_COUNT;
	debug_yield_read();
}

/**
 * rcu_assign_pointer - assign (publicize) a pointer to a newly
 * initialized structure that will be dereferenced by RCU read-side
 * critical sections.  Returns the value assigned.
 *
 * Inserts memory barriers on architectures that require them
 * (pretty much all of them other than x86), and also prevents
 * the compiler from reordering the code that initializes the
 * structure after the pointer assignment.  More importantly, this
 * call documents which pointers will be dereferenced by RCU read-side
 * code.
 */

#define rcu_assign_pointer(p, v) \
	({ \
		if (!__builtin_constant_p(v) || \
		    ((v) != NULL)) \
			wmb(); \
		(p) = (v); \
	})

#define rcu_xchg_pointer(p, v) \
	({ \
		if (!__builtin_constant_p(v) || \
		    ((v) != NULL)) \
			wmb(); \
		xchg(p, v); \
	})

extern void synchronize_rcu(void);

/*
 * Exchanges the pointer and waits for quiescent state.
 * The pointer returned can be freed.
 */
#define urcu_publish_content(p, v) \
	({ \
		void *oldptr; \
		debug_yield_write(); \
		oldptr = rcu_xchg_pointer(p, v); \
		synchronize_rcu(); \
		oldptr; \
	})

/*
 * Reader thread registration.
 */
extern void urcu_register_thread(void);
extern void urcu_unregister_thread(void);

#endif /* _URCU_H */
Commit	Line	Data
	1	#ifndef _URCU_H
	2	#define _URCU_H
	3
	4	/*
	5	* urcu.h
	6	*
	7	* Userspace RCU header
	8	*
	9	* Copyright February 2009 - Mathieu Desnoyers <mathieu.desnoyers@polymtl.ca>
	10	*
	11	* Credits for Paul e. McKenney <paulmck@linux.vnet.ibm.com>
	12	* for inspiration coming from the Linux kernel RCU and rcu-preempt.
	13	*
	14	* The barrier, mb, rmb, wmb, atomic_inc, smp_read_barrier_depends, ACCESS_ONCE
	15	* and rcu_dereference primitives come from the Linux kernel.
	16	*
	17	* Distributed under GPLv2
	18	*/
	19
	20	#define __USE_GNU
	21	#include <stdlib.h>
	22
	23	/* The "volatile" is due to gcc bugs */
	24	#define barrier() __asm__ __volatile__("": : :"memory")
	25
	26	/* x86 32/64 specific */
	27	#define mb() asm volatile("mfence":::"memory")
	28	#define rmb() asm volatile("lfence":::"memory")
	29	#define wmb() asm volatile("sfence" ::: "memory")
	30
	31	static inline void atomic_inc(int *v)
	32	{
	33	asm volatile("lock; incl %0"
	34	: "+m" (*v));
	35	}
	36
	37	#define xchg(ptr, v) \
	38	((__typeof__((ptr)))__xchg((unsigned long)(v), (ptr), sizeof((ptr))))
	39
	40	struct __xchg_dummy {
	41	unsigned long a[100];
	42	};
	43	#define __xg(x) ((struct __xchg_dummy *)(x))
	44
	45	/*
	46	* Note: no "lock" prefix even on SMP: xchg always implies lock anyway
	47	* Note 2: xchg has side effect, so that attribute volatile is necessary,
	48	* but generally the primitive is invalid, *ptr is output argument. --ANK
	49	*/
	50	static inline unsigned long __xchg(unsigned long x, volatile void *ptr,
	51	int size)
	52	{
	53	switch (size) {
	54	case 1:
	55	asm volatile("xchgb %b0,%1"
	56	: "=q" (x)
	57	: "m" (*__xg(ptr)), "0" (x)
	58	: "memory");
	59	break;
	60	case 2:
	61	asm volatile("xchgw %w0,%1"
	62	: "=r" (x)
	63	: "m" (*__xg(ptr)), "0" (x)
	64	: "memory");
	65	break;
	66	case 4:
	67	asm volatile("xchgl %0,%1"
	68	: "=r" (x)
	69	: "m" (*__xg(ptr)), "0" (x)
	70	: "memory");
	71	break;
	72	}
	73	return x;
	74	}
	75
	76	/* Nop everywhere except on alpha. */
	77	#define smp_read_barrier_depends()
	78
	79	/*
	80	* Prevent the compiler from merging or refetching accesses. The compiler
	81	* is also forbidden from reordering successive instances of ACCESS_ONCE(),
	82	* but only when the compiler is aware of some particular ordering. One way
	83	* to make the compiler aware of ordering is to put the two invocations of
	84	* ACCESS_ONCE() in different C statements.
	85	*
	86	* This macro does absolutely -nothing- to prevent the CPU from reordering,
	87	* merging, or refetching absolutely anything at any time. Its main intended
	88	* use is to mediate communication between process-level code and irq/NMI
	89	* handlers, all running on the same CPU.
	90	*/
	91	#define ACCESS_ONCE(x) ((volatile typeof(x) )&(x))
	92
	93	/**
	94	* rcu_dereference - fetch an RCU-protected pointer in an
	95	* RCU read-side critical section. This pointer may later
	96	* be safely dereferenced.
	97	*
	98	* Inserts memory barriers on architectures that require them
	99	* (currently only the Alpha), and, more importantly, documents
	100	* exactly which pointers are protected by RCU.
	101	*/
	102
	103	#define rcu_dereference(p) ({ \
	104	typeof(p) _________p1 = ACCESS_ONCE(p); \
	105	smp_read_barrier_depends(); \
	106	(_________p1); \
	107	})
	108
	109	#define SIGURCU SIGUSR1
	110
	111	#ifdef DEBUG_YIELD
	112	#include <sched.h>
	113	#include <time.h>
	114	#include <pthread.h>
	115
	116	#define YIELD_READ (1 << 0)
	117	#define YIELD_WRITE (1 << 1)
	118
	119	extern unsigned int yield_active;
	120	extern unsigned int __thread rand_yield;
	121
	122	static inline void debug_yield_read(void)
	123	{
	124	if (yield_active & YIELD_READ)
	125	if (rand_r(&rand_yield) & 0x1)
	126	sched_yield();
	127	}
	128
	129	static inline void debug_yield_write(void)
	130	{
	131	if (yield_active & YIELD_WRITE)
	132	if (rand_r(&rand_yield) & 0x1)
	133	sched_yield();
	134	}
	135
	136	static inline void debug_yield_init(void)
	137	{
	138	rand_yield = time(NULL) ^ pthread_self();
	139	}
	140	#else
	141	static inline void debug_yield_read(void)
	142	{
	143	}
	144
	145	static inline void debug_yield_write(void)
	146	{
	147	}
	148
	149	static inline void debug_yield_init(void)
	150	{
	151
	152	}
	153	#endif
	154
	155	/*
	156	* Limiting the nesting level to 256 to keep instructions small in the read
	157	* fast-path.
	158	*/
	159	#define RCU_GP_COUNT (1U << 0)
	160	#define RCU_GP_CTR_BIT (1U << 8)
	161	#define RCU_GP_CTR_NEST_MASK (RCU_GP_CTR_BIT - 1)
	162
	163	/* Global quiescent period counter with low-order bits unused. */
	164	extern int urcu_gp_ctr;
	165
	166	extern int __thread urcu_active_readers;
	167
	168	static inline int rcu_old_gp_ongoing(int *value)
	169	{
	170	int v;
	171
	172	if (value == NULL)
	173	return 0;
	174	debug_yield_write();
	175	v = ACCESS_ONCE(*value);
	176	debug_yield_write();
	177	return (v & RCU_GP_CTR_NEST_MASK) &&
	178	((v ^ ACCESS_ONCE(urcu_gp_ctr)) & RCU_GP_CTR_BIT);
	179	}
	180
	181	static inline void rcu_read_lock(void)
	182	{
	183	int tmp;
	184
	185	debug_yield_read();
	186	tmp = urcu_active_readers;
	187	debug_yield_read();
	188	if (!(tmp & RCU_GP_CTR_NEST_MASK))
	189	urcu_active_readers = urcu_gp_ctr + RCU_GP_COUNT;
	190	else
	191	urcu_active_readers = tmp + RCU_GP_COUNT;
	192	debug_yield_read();
	193	/*
	194	* Increment active readers count before accessing the pointer.
	195	* See force_mb_all_threads().
	196	*/
	197	barrier();
	198	debug_yield_read();
	199	}
	200
	201	static inline void rcu_read_unlock(void)
	202	{
	203	debug_yield_read();
	204	barrier();
	205	debug_yield_read();
	206	/*
	207	* Finish using rcu before decrementing the pointer.
	208	* See force_mb_all_threads().
	209	*/
	210	urcu_active_readers -= RCU_GP_COUNT;
	211	debug_yield_read();
	212	}
	213
	214	/**
	215	* rcu_assign_pointer - assign (publicize) a pointer to a newly
	216	* initialized structure that will be dereferenced by RCU read-side
	217	* critical sections. Returns the value assigned.
	218	*
	219	* Inserts memory barriers on architectures that require them
	220	* (pretty much all of them other than x86), and also prevents
	221	* the compiler from reordering the code that initializes the
	222	* structure after the pointer assignment. More importantly, this
	223	* call documents which pointers will be dereferenced by RCU read-side
	224	* code.
	225	*/
	226
	227	#define rcu_assign_pointer(p, v) \
	228	({ \
	229	if (!__builtin_constant_p(v) \|\| \
	230	((v) != NULL)) \
	231	wmb(); \
	232	(p) = (v); \
	233	})
	234
	235	#define rcu_xchg_pointer(p, v) \
	236	({ \
	237	if (!__builtin_constant_p(v) \|\| \
	238	((v) != NULL)) \
	239	wmb(); \
	240	xchg(p, v); \
	241	})
	242
	243	extern void synchronize_rcu(void);
	244
	245	/*
	246	* Exchanges the pointer and waits for quiescent state.
	247	* The pointer returned can be freed.
	248	*/
	249	#define urcu_publish_content(p, v) \
	250	({ \
	251	void *oldptr; \
	252	debug_yield_write(); \
	253	oldptr = rcu_xchg_pointer(p, v); \
	254	synchronize_rcu(); \
	255	oldptr; \
	256	})
	257
	258	/*
	259	* Reader thread registration.
	260	*/
	261	extern void urcu_register_thread(void);
	262	extern void urcu_unregister_thread(void);
	263
	264	#endif /* _URCU_H */