fastlock.h

/*
 * fast architecture specific locking
 *
 * $Id$
 *
 * 
 *
 * Copyright (C) 2001-2003 FhG Fokus
 *
 * Permission to use, copy, modify, and distribute this software for any
 * purpose with or without fee is hereby granted, provided that the above
 * copyright notice and this permission notice appear in all copies.
 *
 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
 */
/*
 *
 *History:
 *--------
 *  2002-02-05  created by andrei
 *  2003-01-16  added PPC locking code contributed by Dinos Dorkofikis
 *               <kdor@intranet.gr>
 *  2004-09-12  added MIPS locking for ISA>=2 (>r3000)  (andrei)
 *  2004-12-16  for now use the same locking code for sparc32 as for sparc64
 *               (it will work only if NOSMP is defined) (andrei)
 *  2005-04-27  added alpha locking code (andrei)
 *  2005-05-25  PPC locking code enabled for PPC64; added a lwsync to
 *               the tsl part and replaced the sync with a lwsync for the
 *               unlock part (andrei)
 *  2006-03-08  mips2 NOSMP (skip sync), optimized x86 & mips clobbers and
 *               input/output constraints (andrei)
 *  2006-04-03  optimization: call lock_get memory barrier outside tsl,in the 
 *               calling function, only if the lock operation succeeded
 *               (membar_getlock()) (andrei)
 *              added try_lock(); more x86 optimizations, x86  release_lock
 *               fix (andrei)
 * 2006-04-04  sparc* optimizations, sparc32 smp support, armv6 no smp support,
 *              ppc, mips*, alpha optimizations (andrei)
 * 2006-04-05  ppc fixes (s/stw/stwx/, s/lwz/lwzx), early clobber added
 *             where needed (andrei)
 * 2006-11-22  arm early clobber added: according to the swp instruction 
 *              specification the address register must be != from the other 2
 *              (Julien Blache <jblache@debian.org>)
 *
 */

/*
 * WARNING: the code was not tested on the following architectures:
 *           - arm6  (cross-compiles ok, no test)
 *           - alpha (cross-compiles ok, no test)
 *           - mips64 (cross-compiles ok)
 *           - ppc64 (compiles ok)
 *           - sparc32 (tested on a sparc64)
 */


#ifndef fastlock_h
#define fastlock_h

#include "sched_yield.h"


#define SPIN_OPTIMIZE /* if defined optimize spining on the lock:
                         try first the lock with non-atomic/non memory locking
                         operations, and only if the lock appears to be free
                         switch to the more expensive version */

typedef  volatile int fl_lock_t;



#define init_lock( l ) (l)=0


/* what membar to use (if any) after taking a lock. This
 *  was separated from the lock code to allow better optimizations.
 *  e.g.: use the membar_getlock only after getting the lock and don't use 
 *  it if lock_get fails / when spinning on tsl.
 *  There is no corresponding membar_release_lock (because lock_release
 *  must always include the needed memory barrier).
 *  WARNING: this is intended only for internal fastlock use*/
#if defined(__CPU_i386) || defined(__CPU_x86_64)
#define membar_getlock()   /* not needed on x86 */

#elif defined(__CPU_sparc64)
#ifndef NOSMP
#define membar_getlock() \
	asm volatile ("membar #StoreStore | #StoreLoad \n\t" : : : "memory");
	/* can be either StoreStore|StoreLoad or LoadStore|LoadLoad
	 * since ldstub acts both as a store and as a load */
#else
/* no need for a compiler barrier, that is already included in lock_get/tsl*/
#define membar_getlock() /* not needed if no smp*/
#endif /* NOSMP */

#elif  defined(__CPU_sparc)
#define membar_getlock()/* no need for a compiler barrier, already included */

#elif defined __CPU_arm || defined __CPU_arm6
#ifndef NOSMP
#warning smp not supported on arm* (no membars), try compiling with -DNOSMP
#endif /* NOSMP */
#define membar_getlock() 

#elif defined(__CPU_ppc) || defined(__CPU_ppc64)
#ifndef NOSMP
#define membar_getlock() \
	asm volatile("lwsync \n\t" : : : "memory");
#else
#define membar_getlock() 
#endif /* NOSMP */

#elif defined __CPU_mips2 || defined __CPU_mips64
#ifndef NOSMP
#define membar_getlock() \
	asm volatile("sync \n\t" : : : "memory");
#else
#define membar_getlock() 
#endif /* NOSMP */

#elif defined __CPU_mips
#ifndef NOSMP
#warning smp not supported on mips1 (no membars), try compiling with -DNOSMP
#endif
#define membar_getlock() 

#elif defined __CPU_alpha
#ifndef NOSMP
#define membar_getlock() \
	asm volatile("mb \n\t" : : : "memory");
#else
#define membar_getlock() 
#endif /* NOSMP */

#else /* __CPU_xxx */
#error "unknown architecture"
#endif



/*test and set lock, ret !=0 if lock held by someone else, 0 otherwise
 * WARNING: no memory barriers included, if you use this function directly
 *          (not recommended) and it gets the lock (ret==0), you should call 
 *          membar_getlock() after it */
inline static int tsl(fl_lock_t* lock)
{
	int val;

#if defined(__CPU_i386) || defined(__CPU_x86_64)

#ifdef NOSMP
	asm volatile(
		" xor %0, %0 \n\t"
		" btsl $0, %2 \n\t"
		" setc %b0 \n\t"
		: "=&q" (val), "=m" (*lock) : "m"(*lock) : "memory", "cc"
	);
#else
	asm volatile(
#ifdef SPIN_OPTIMIZE
		" cmpb $0, %2 \n\t"
		" mov $1, %0 \n\t"
		" jnz 1f \n\t"
#else
		" mov $1, %0 \n\t"
#endif
		" xchgb %2, %b0 \n\t"
		"1: \n\t"
		: "=&q" (val), "=m" (*lock) : "m"(*lock) : "memory"
#ifdef SPIN_OPTIMIZE
				, "cc"
#endif
	);
#endif /*NOSMP*/
#elif defined(__CPU_sparc64)
	asm volatile(
#ifdef SPIN_OPTIMIZE
			"   ldub [%2], %0 \n\t"
			"   brnz,a,pn %0, 1f \n\t"
			"   nop \n\t"
#endif
			"   ldstub [%2], %0 \n\t"
			"1: \n\t"
			/* membar_getlock must be  called outside this function */
			: "=&r"(val), "=m"(*lock) : "r"(lock): "memory"
	);
#elif defined(__CPU_sparc)
	asm volatile(
#ifdef SPIN_OPTIMIZE
			"   ldub [%2], %0 \n\t"
			"   tst %0 \n\t"
			"   bne,a  1f \n\t"
			"   nop \n\t"
#endif
			"   ldstub [%2], %0 \n\t"
			"1: \n\t"
			/* membar_getlock must be  called outside this function */
			: "=&r"(val), "=m"(*lock) : "r"(lock): "memory"
#ifdef SPIN_OPTIMIZE
				, "cc"
#endif
	);
#elif defined __CPU_arm 
	asm volatile(
			"swp %0, %2, [%3] \n\t"
			: "=&r" (val), "=m"(*lock) : "r"(1), "r" (lock) : "memory"
	);
#elif defined __CPU_arm6
	asm volatile(
			"   ldrex %0, [%2] \n\t" 
			"   cmp %0, #0 \n\t"
			"   strexeq %0, %3, [%2] \n\t" /* executed only if Z=1 */
			/* if %0!=0 => either it was 1 initially or was 0
			 * and somebody changed it just before the strexeq (so the 
			 * lock is taken) => it's safe to return %0 */
			: "=&r"(val), "=m"(*lock) : "r"(lock), "r"(1) : "cc"
	);
#elif defined(__CPU_ppc) || defined(__CPU_ppc64)
	asm volatile(
			"1: \n\t"
#ifdef SPIN_OPTIMIZE
			"   lwzx %0, 0, %2 \n\t"
			"   cmpwi %0, 0 \n\t"
			"   bne- 2f \n\t" /* predict: not taken */
#endif
			"   lwarx  %0, 0, %2\n\t"
			"   cmpwi  %0, 0\n\t"
			"   bne-    2f\n\t"
			"   stwcx. %3, 0, %2\n\t"
			"   bne-   1b\n\t"
			/* membar_getlock must be  called outside this function */
			"2:\n\t"
			: "=&r" (val), "=m"(*lock) :  "r"(lock), "r"(1) : "memory", "cc"
        );
#elif defined __CPU_mips2 || ( defined __CPU_mips && defined MIPS_HAS_LLSC ) \
	|| defined __CPU_mips64
	long tmp;
	
	asm volatile(
		".set push \n\t"
		".set noreorder\n\t"
		".set mips2 \n\t"
#ifdef SPIN_OPTIMIZE
		"    lw %1, %2 \n\t"
		"    bne %1, $0, 2f \n\t"
		"    nop \n\t"
#endif
		"1:  ll %1, %2   \n\t"
		"    bne %1, $0, 2f \n\t"
		"    li %0, 1 \n\t"  /* delay slot */
		"    sc %0, %2  \n\t"
		"    beqz %0, 1b \n\t"
		"    nop \n\t"
		"2: \n\t"
		/* membar_getlock must be called outside this function */
		".set pop\n\t"
		: "=&r" (tmp), "=&r" (val), "=m" (*lock) 
		: "m" (*lock) 
		: "memory"
	);
#elif defined __CPU_alpha
	long tmp;
	tmp=0;
	/* lock low bit set to 1 when the lock is hold and to 0 otherwise */
	asm volatile(
		"1:  ldl %0, %1   \n\t"
		"    blbs %0, 2f  \n\t"  /* optimization if locked */
		"    ldl_l %0, %1 \n\t"
		"    blbs %0, 2f  \n\t" 
		"    lda %2, 1    \n\t"  /* or: or $31, 1, %2 ??? */
		"    stl_c %2, %1 \n\t"
		"    beq %2, 3f   \n\t" /* back cond. jumps are always predicted to be 
								   taken => make forward jump */
		/* membar_getlock must be called outside this function */
		"2:               \n\t"
		".subsection 2 \n\t"
		"3:  br 1b \n\t"
		".previous \n\t"
		:"=&r" (val), "=m"(*lock), "=&r"(tmp)
		:"m"(*lock) 
		: "memory"
	);
#else
#error "unknown architecture"
#endif
	return val;
}



inline static void get_lock(fl_lock_t* lock)
{
#ifdef ADAPTIVE_WAIT
	int i=ADAPTIVE_WAIT_LOOPS;
#endif
	
	while(tsl(lock)){
#ifdef BUSY_WAIT
#elif defined ADAPTIVE_WAIT
		if (i>0) i--;
		else sched_yield();
#else
		sched_yield();
#endif
	}
	membar_getlock();
}



/* like get_lock, but it doesn't wait. If it gets the lock returns 0,
 *  <0  otherwise (-1) */
inline static int try_lock(fl_lock_t* lock)
{
	if (tsl(lock)){
		return -1;
	}
	membar_getlock();
	return 0;
}



inline static void release_lock(fl_lock_t* lock)
{
#if defined(__CPU_i386) 
#ifdef NOSMP
	asm volatile(
		" movb $0, %0 \n\t" 
		: "=m"(*lock) : : "memory"
	); 
#else /* ! NOSMP */
	int val;
	/* a simple mov $0, (lock) does not force StoreStore ordering on all
	   x86 versions and it doesn't seem to force LoadStore either */
	asm volatile(
		" xchgb %b0, %1 \n\t"
		: "=q" (val), "=m" (*lock) : "0" (0) : "memory"
	);
#endif /* NOSMP */
#elif defined(__CPU_x86_64)
	asm volatile(
		" movb $0, %0 \n\t" /* on amd64 membar StoreStore | LoadStore is 
							   implicit (at least on the same mem. type) */
		: "=m"(*lock) : : "memory"
	);
#elif defined(__CPU_sparc64) || defined(__CPU_sparc)
	asm volatile(
#ifndef NOSMP
#ifdef __CPU_sparc64
			"membar #LoadStore | #StoreStore \n\t"
#else /* __CPU_sparc */
			"stbar \n\t"
#endif /* __CPU_sparc64 */
#endif
			"stb %%g0, [%1] \n\t"
			: "=m"(*lock) : "r" (lock) : "memory"
	);
#elif defined __CPU_arm || defined __CPU_arm6
#ifndef NOSMP
#warning arm* smp mode not supported (no membars), try compiling with -DNOSMP
#endif
	asm volatile(
		" str %1, [%2] \n\r" 
		: "=m"(*lock) : "r"(0), "r"(lock) : "memory"
	);
#elif defined(__CPU_ppc) || defined(__CPU_ppc64)
	asm volatile(
			/* "sync\n\t"  lwsync is faster and will work
			 *             here too
			 *             [IBM Prgramming Environments Manual, D.4.2.2]
			 */
			"lwsync\n\t"
			"stwx %1, 0, %2\n\t"
			: "=m"(*lock) : "r"(0), "r"(lock) : "memory"
	);
#elif defined __CPU_mips2 || ( defined __CPU_mips && defined MIPS_HAS_LLSC ) \
	|| defined __CPU_mips64
	asm volatile(
		".set push \n\t"
		".set noreorder \n\t"
		".set mips2 \n\t"
#ifndef NOSMP
#ifdef __CPU_mips
#warning mips1 smp mode not supported (no membars), try compiling with -DNOSMP
#else
		"    sync \n\t"
#endif
#endif
		"    sw $0, %0 \n\t"
		".set pop \n\t"
		: "=m" (*lock)  : /* no input */ : "memory"
	);
#elif defined __CPU_alpha
	asm volatile(
#ifndef  NOSMP
		"    mb          \n\t"
#endif
		"    stl $31, %0 \n\t"
		: "=m"(*lock) :/* no input*/ : "memory"  /* because of the mb */
	);  
#else
#error "unknown architecture"
#endif
}



#endif