[PATCH] add cpu_relax() in spin loops & clean up barrier()

The patch adds cpu_relax() in the body of some spin loops for 2.6.9.  The
patch also removes redundant barrier() code after cpu_relax() on ia32.

In the PAUSE instruction section, IA32 SDM claims "it is recommended that a
PASUE instruction be placed in all spin-wait loops".  And x86_64 SDM says
that PAUSE instruction is same as legacy mode in IA-32e mode operation.

This patch is against 2.6.9 (kernel.org).  It was tested on ia32 and
x86_64.
Acked-by: Andi Kleen <ak@muc.de>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>

arch/i386/kernel/cpu/mtrr/main.c

@@ -149,10 +149,8 @@ static void ipi_handler(void *info)
 	local_irq_save(flags);
 
 	atomic_dec(&data->count);
-	while(!atomic_read(&data->gate)) {
+	while(!atomic_read(&data->gate))
 		cpu_relax();
-		barrier();
-	}
 
 	/*  The master has cleared me to execute  */
 	if (data->smp_reg != ~0U) 
@@ -162,10 +160,9 @@ static void ipi_handler(void *info)
 		mtrr_if->set_all();
 
 	atomic_dec(&data->count);
-	while(atomic_read(&data->gate)) {
+	while(atomic_read(&data->gate))
 		cpu_relax();
-		barrier();
-	}
+
 	atomic_dec(&data->count);
 	local_irq_restore(flags);
 }
@@ -230,10 +227,9 @@ static void set_mtrr(unsigned int reg, unsigned long base,
 
 	local_irq_save(flags);
 
-	while(atomic_read(&data.count)) {
+	while(atomic_read(&data.count))
 		cpu_relax();
-		barrier();
-	}
+
 	/* ok, reset count and toggle gate */
 	atomic_set(&data.count, num_booting_cpus() - 1);
 	atomic_set(&data.gate,1);
@@ -250,10 +246,9 @@ static void set_mtrr(unsigned int reg, unsigned long base,
 		mtrr_if->set(reg,base,size,type);
 
 	/* wait for the others */
-	while(atomic_read(&data.count)) {
+	while(atomic_read(&data.count))
 		cpu_relax();
-		barrier();
-	}
+
 	atomic_set(&data.count, num_booting_cpus() - 1);
 	atomic_set(&data.gate,0);
 
@@ -261,10 +256,9 @@ static void set_mtrr(unsigned int reg, unsigned long base,
 	 * Wait here for everyone to have seen the gate change
 	 * So we're the last ones to touch 'data'
 	 */
-	while(atomic_read(&data.count)) {
+	while(atomic_read(&data.count))
 		cpu_relax();
-		barrier();
-	}
+
 	local_irq_restore(flags);
 }
 

arch/i386/kernel/smp.c

@@ -538,11 +538,11 @@ int smp_call_function (void (*func) (void *info), void *info, int nonatomic,
 
 	/* Wait for response */
 	while (atomic_read(&data.started) != cpus)
-		barrier();
+		cpu_relax();
 
 	if (wait)
 		while (atomic_read(&data.finished) != cpus)
-			barrier();
+			cpu_relax();
 	spin_unlock(&call_lock);
 
 	return 0;

arch/x86_64/kernel/smp.c

@@ -310,10 +310,10 @@ static void __smp_call_function (void (*func) (void *info), void *info,
 
 	/* Wait for response */
 	while (atomic_read(&data.started) != cpus)
-		barrier();
+		cpu_relax();
 
 	while (atomic_read(&data.finished) != cpus)
-		barrier();
+		cpu_relax();
 }
 
 /*

include/asm-i386/apic.h

@@ -53,7 +53,8 @@ static __inline unsigned long apic_read(unsigned long reg)
 
 static __inline__ void apic_wait_icr_idle(void)
 {
-	do { } while ( apic_read( APIC_ICR ) & APIC_ICR_BUSY );
+	while ( apic_read( APIC_ICR ) & APIC_ICR_BUSY )
+		cpu_relax();
 }
 
 int get_physical_broadcast(void);