From: Jan Glauber <jglauber@redhat.com>
Subject: [RHEL5.1 PATCH] direct yield for spinlocks on s390
Date: Tue, 20 Feb 2007 17:36:34 +0100
Bugzilla: 228869
Message-Id: <1171989394.6031.50.camel@localhost.localdomain>
Changelog: [s390] direct yield for spinlocks on s390


BZ 219871

Add support for direct yield of spinlocks on s390. There is a new
diagnose call (0x9c) that yields the remaining timeslice of the virtual cpu
that tries to acquire a lock to the virtual cpu that is the current holder 
of the lock.

The patch doesn't change the spinlock kabi for s390 since R5 already
contained a dummy field for the lock owner.

With directed yield the very high number of diag calls due to spinlocks 
(diag 0x44) rapidly decreases. That improves the performance on a z/VM
with many Linux guests.

Jan
-- 
jglauber@redhat.com
jang@de.ibm.com

 arch/s390/kernel/head31.S   |   11 ++++++++
 arch/s390/kernel/head64.S   |   11 ++++++++
 arch/s390/lib/spinlock.c    |   54 +++++++++++++++++++++++++-------------------
 include/asm-s390/setup.h    |    1 
 include/asm-s390/spinlock.h |   49 ++++++++++++++++++++++++++++++++-------
 5 files changed, 94 insertions(+), 32 deletions(-)

diff -urpN linux-2.6.18.s390x/arch/s390/kernel/head31.S linux-2.6.18.s390x-i4i/arch/s390/kernel/head31.S
--- linux-2.6.18.s390x/arch/s390/kernel/head31.S	2006-09-20 03:42:06.000000000 +0000
+++ linux-2.6.18.s390x-i4i/arch/s390/kernel/head31.S	2007-01-10 16:29:39.000000000 +0000
@@ -239,6 +239,16 @@ startup_continue:
 	oi	3(%r12),0x80		# set IDTE flag
 .Lchkidte:
 
+#
+# find out if the diag 0x9c is available
+#
+	mvc	__LC_PGM_NEW_PSW(8),.Lpcdiag9c-.LPG1(%r13)
+	stap   __LC_CPUID+4		# store cpu address
+	lh     %r1,__LC_CPUID+4
+	diag   %r1,0,0x9c		# test diag 0x9c
+	oi     2(%r12),1		# set diag9c flag
+.Lchkdiag9c:
+
 	lpsw  .Lentry-.LPG1(13)		# jump to _stext in primary-space,
 					# virtual and never return ...
 	.align	8
@@ -266,6 +276,7 @@ startup_continue:
 .Lpccsp:.long	0x00080000,0x80000000 + .Lchkcsp
 .Lpcmvpg:.long	0x00080000,0x80000000 + .Lchkmvpg
 .Lpcidte:.long	0x00080000,0x80000000 + .Lchkidte
+.Lpcdiag9c:.long 0x00080000,0x80000000 + .Lchkdiag9c
 .Lmemsize:.long memory_size
 .Lmchunk:.long	memory_chunk
 .Lmflags:.long	machine_flags
diff -urpN linux-2.6.18.s390x/arch/s390/kernel/head64.S linux-2.6.18.s390x-i4i/arch/s390/kernel/head64.S
--- linux-2.6.18.s390x/arch/s390/kernel/head64.S	2006-09-20 03:42:06.000000000 +0000
+++ linux-2.6.18.s390x-i4i/arch/s390/kernel/head64.S	2007-01-10 16:29:39.000000000 +0000
@@ -239,6 +239,17 @@ startup_continue:
 	oi	7(%r12),0x80		# set IDTE flag
 0:
 
+#
+# find out if the diag 0x9c is available
+#
+	la     %r1,0f-.LPG1(%r13)	# set program check address
+	stg    %r1,__LC_PGM_NEW_PSW+8
+	stap   __LC_CPUID+4		# store cpu address
+	lh     %r1,__LC_CPUID+4
+	diag   %r1,0,0x9c		# test diag 0x9c
+	oi     6(%r12),1		# set diag9c flag
+0:
+
         lpswe .Lentry-.LPG1(13)         # jump to _stext in primary-space,
                                         # virtual and never return ...
         .align 16
diff -urpN linux-2.6.18.s390x/arch/s390/lib/spinlock.c linux-2.6.18.s390x-i4i/arch/s390/lib/spinlock.c
--- linux-2.6.18.s390x/arch/s390/lib/spinlock.c	2006-09-20 03:42:06.000000000 +0000
+++ linux-2.6.18.s390x-i4i/arch/s390/lib/spinlock.c	2007-01-10 16:29:39.000000000 +0000
@@ -24,57 +24,68 @@ static int __init spin_retry_setup(char 
 }
 __setup("spin_retry=", spin_retry_setup);
 
-static inline void
-_diag44(void)
+static inline void _raw_yield(void)
 {
-#ifdef CONFIG_64BIT
 	if (MACHINE_HAS_DIAG44)
-#endif
 		asm volatile("diag 0,0,0x44");
 }
 
-void
-_raw_spin_lock_wait(raw_spinlock_t *lp, unsigned int pc)
+static inline void _raw_yield_cpu(int cpu)
+{
+	if (MACHINE_HAS_DIAG9C)
+		asm volatile("diag %0,0,0x9c"
+			     : : "d" (__cpu_logical_map[cpu]));
+	else
+		_raw_yield();
+}
+
+void _raw_spin_lock_wait(raw_spinlock_t *lp, unsigned int pc)
 {
 	int count = spin_retry;
+	unsigned int cpu = ~smp_processor_id();
 
 	while (1) {
 		if (count-- <= 0) {
-			_diag44();
+			unsigned int owner = lp->lock;
+			if (owner != 0)
+				_raw_yield_cpu(~owner);
 			count = spin_retry;
 		}
 		if (__raw_spin_is_locked(lp))
 			continue;
-		if (_raw_compare_and_swap(&lp->lock, 0, pc) == 0)
+		if (_raw_compare_and_swap(&lp->lock, 0, cpu) == 0) {
+			lp->owner_pc = pc;
 			return;
+		}
 	}
 }
 EXPORT_SYMBOL(_raw_spin_lock_wait);
 
-int
-_raw_spin_trylock_retry(raw_spinlock_t *lp, unsigned int pc)
+int _raw_spin_trylock_retry(raw_spinlock_t *lp, unsigned int pc)
 {
-	int count = spin_retry;
+	unsigned int cpu = ~smp_processor_id();
+	int count;
 
-	while (count-- > 0) {
+	for (count = spin_retry; count > 0; count--) {
 		if (__raw_spin_is_locked(lp))
 			continue;
-		if (_raw_compare_and_swap(&lp->lock, 0, pc) == 0)
+		if (_raw_compare_and_swap(&lp->lock, 0, cpu) == 0) {
+			lp->owner_pc = pc;
 			return 1;
+		}
 	}
 	return 0;
 }
 EXPORT_SYMBOL(_raw_spin_trylock_retry);
 
-void
-_raw_read_lock_wait(raw_rwlock_t *rw)
+void _raw_read_lock_wait(raw_rwlock_t *rw)
 {
 	unsigned int old;
 	int count = spin_retry;
 
 	while (1) {
 		if (count-- <= 0) {
-			_diag44();
+			_raw_yield();
 			count = spin_retry;
 		}
 		if (!__raw_read_can_lock(rw))
@@ -86,8 +97,7 @@ _raw_read_lock_wait(raw_rwlock_t *rw)
 }
 EXPORT_SYMBOL(_raw_read_lock_wait);
 
-int
-_raw_read_trylock_retry(raw_rwlock_t *rw)
+int _raw_read_trylock_retry(raw_rwlock_t *rw)
 {
 	unsigned int old;
 	int count = spin_retry;
@@ -103,14 +113,13 @@ _raw_read_trylock_retry(raw_rwlock_t *rw
 }
 EXPORT_SYMBOL(_raw_read_trylock_retry);
 
-void
-_raw_write_lock_wait(raw_rwlock_t *rw)
+void _raw_write_lock_wait(raw_rwlock_t *rw)
 {
 	int count = spin_retry;
 
 	while (1) {
 		if (count-- <= 0) {
-			_diag44();
+			_raw_yield();
 			count = spin_retry;
 		}
 		if (!__raw_write_can_lock(rw))
@@ -121,8 +130,7 @@ _raw_write_lock_wait(raw_rwlock_t *rw)
 }
 EXPORT_SYMBOL(_raw_write_lock_wait);
 
-int
-_raw_write_trylock_retry(raw_rwlock_t *rw)
+int _raw_write_trylock_retry(raw_rwlock_t *rw)
 {
 	int count = spin_retry;
 
diff -urpN linux-2.6.18.s390x/include/asm-s390/setup.h linux-2.6.18.s390x-i4i/include/asm-s390/setup.h
--- linux-2.6.18.s390x/include/asm-s390/setup.h	2006-09-20 03:42:06.000000000 +0000
+++ linux-2.6.18.s390x-i4i/include/asm-s390/setup.h	2007-01-10 16:29:39.000000000 +0000
@@ -41,6 +41,7 @@ extern unsigned long machine_flags;
 #define MACHINE_IS_P390		(machine_flags & 4)
 #define MACHINE_HAS_MVPG	(machine_flags & 16)
 #define MACHINE_HAS_IDTE	(machine_flags & 128)
+#define MACHINE_HAS_DIAG9C	(machine_flags & 256)
 
 #ifndef __s390x__
 #define MACHINE_HAS_IEEE	(machine_flags & 2)
diff -urpN linux-2.6.18.s390x/include/asm-s390/spinlock.h linux-2.6.18.s390x-i4i/include/asm-s390/spinlock.h
--- linux-2.6.18.s390x/include/asm-s390/spinlock.h	2006-09-20 03:42:06.000000000 +0000
+++ linux-2.6.18.s390x-i4i/include/asm-s390/spinlock.h	2007-01-10 16:29:39.000000000 +0000
@@ -11,17 +11,38 @@
 #ifndef __ASM_SPINLOCK_H
 #define __ASM_SPINLOCK_H
 
+#include <linux/smp.h>
+
+#if __GNUC__ > 3 || (__GNUC__ == 3 && __GNUC_MINOR__ > 2)
+
 static inline int
 _raw_compare_and_swap(volatile unsigned int *lock,
 		      unsigned int old, unsigned int new)
 {
-	asm volatile ("cs %0,%3,0(%4)"
-		      : "=d" (old), "=m" (*lock)
-		      : "0" (old), "d" (new), "a" (lock), "m" (*lock)
-		      : "cc", "memory" );
+	asm volatile(
+		"	cs	%0,%3,%1"
+		: "=d" (old), "=Q" (*lock)
+		: "0" (old), "d" (new), "Q" (*lock)
+		: "cc", "memory" );
 	return old;
 }
 
+#else /* __GNUC__ */
+
+static inline int
+_raw_compare_and_swap(volatile unsigned int *lock,
+		      unsigned int old, unsigned int new)
+{
+	asm volatile(
+		"	cs	%0,%3,0(%4)"
+		: "=d" (old), "=m" (*lock)
+		: "0" (old), "d" (new), "a" (lock), "m" (*lock)
+		: "cc", "memory" );
+	return old;
+}
+
+#endif /* __GNUC__ */
+
 /*
  * Simple spin lock operations.  There are two variants, one clears IRQ's
  * on the local processor, one does not.
@@ -36,28 +57,38 @@ _raw_compare_and_swap(volatile unsigned 
 #define __raw_spin_unlock_wait(lock) \
 	do { while (__raw_spin_is_locked(lock)) cpu_relax(); } while (0)
 
-extern void _raw_spin_lock_wait(raw_spinlock_t *lp, unsigned int pc);
-extern int _raw_spin_trylock_retry(raw_spinlock_t *lp, unsigned int pc);
+extern void _raw_spin_lock_wait(raw_spinlock_t *, unsigned int pc);
+extern int _raw_spin_trylock_retry(raw_spinlock_t *, unsigned int pc);
 
 static inline void __raw_spin_lock(raw_spinlock_t *lp)
 {
 	unsigned long pc = 1 | (unsigned long) __builtin_return_address(0);
+	int old;
 
-	if (unlikely(_raw_compare_and_swap(&lp->lock, 0, pc) != 0))
-		_raw_spin_lock_wait(lp, pc);
+	old = _raw_compare_and_swap(&lp->lock, 0, ~smp_processor_id());
+	if (likely(old == 0)) {
+		lp->owner_pc = pc;
+		return;
+	}
+	_raw_spin_lock_wait(lp, pc);
 }
 
 static inline int __raw_spin_trylock(raw_spinlock_t *lp)
 {
 	unsigned long pc = 1 | (unsigned long) __builtin_return_address(0);
+	int old;
 
-	if (likely(_raw_compare_and_swap(&lp->lock, 0, pc) == 0))
+	old = _raw_compare_and_swap(&lp->lock, 0, ~smp_processor_id());
+	if (likely(old == 0)) {
+		lp->owner_pc = pc;
 		return 1;
+	}
 	return _raw_spin_trylock_retry(lp, pc);
 }
 
 static inline void __raw_spin_unlock(raw_spinlock_t *lp)
 {
+	lp->owner_pc = 0;
 	_raw_compare_and_swap(&lp->lock, lp->lock, 0);
 }