[PATCH] Directed yield: cpu_relax variants for spinlocks and rw-locks
Martin Schwidefsky [Sun, 1 Oct 2006 06:27:43 +0000 (23:27 -0700)]
On systems running with virtual cpus there is optimization potential in
regard to spinlocks and rw-locks.  If the virtual cpu that has taken a lock
is known to a cpu that wants to acquire the same lock it is beneficial to
yield the timeslice of the virtual cpu in favour of the cpu that has the
lock (directed yield).

With CONFIG_PREEMPT="n" this can be implemented by the architecture without
common code changes.  Powerpc already does this.

With CONFIG_PREEMPT="y" the lock loops are coded with _raw_spin_trylock,
_raw_read_trylock and _raw_write_trylock in kernel/spinlock.c.  If the lock
could not be taken cpu_relax is called.  A directed yield is not possible
because cpu_relax doesn't know anything about the lock.  To be able to
yield the lock in favour of the current lock holder variants of cpu_relax
for spinlocks and rw-locks are needed.  The new _raw_spin_relax,
_raw_read_relax and _raw_write_relax primitives differ from cpu_relax
insofar that they have an argument: a pointer to the lock structure.

Signed-off-by: Martin Schwidefsky <schwidefsky@de.ibm.com>
Cc: Ingo Molnar <mingo@elte.hu>
Cc: Paul Mackerras <paulus@samba.org>
Cc: Haavard Skinnemoen <hskinnemoen@atmel.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>

16 files changed:
include/asm-alpha/spinlock.h
include/asm-arm/spinlock.h
include/asm-cris/arch-v32/spinlock.h
include/asm-i386/spinlock.h
include/asm-ia64/spinlock.h
include/asm-m32r/spinlock.h
include/asm-mips/spinlock.h
include/asm-parisc/spinlock.h
include/asm-powerpc/spinlock.h
include/asm-ppc/spinlock.h
include/asm-s390/spinlock.h
include/asm-sh/spinlock.h
include/asm-sparc/spinlock.h
include/asm-sparc64/spinlock.h
include/asm-x86_64/spinlock.h
kernel/spinlock.c

index 0c294c9..aeeb125 100644 (file)
@@ -166,4 +166,8 @@ static inline void __raw_write_unlock(raw_rwlock_t * lock)
        lock->lock = 0;
 }
 
+#define _raw_spin_relax(lock)  cpu_relax()
+#define _raw_read_relax(lock)  cpu_relax()
+#define _raw_write_relax(lock) cpu_relax()
+
 #endif /* _ALPHA_SPINLOCK_H */
index 01b7c26..861092f 100644 (file)
@@ -218,4 +218,8 @@ static inline int __raw_read_trylock(raw_rwlock_t *rw)
 /* read_can_lock - would read_trylock() succeed? */
 #define __raw_read_can_lock(x)         ((x)->lock < 0x80000000)
 
+#define _raw_spin_relax(lock)  cpu_relax()
+#define _raw_read_relax(lock)  cpu_relax()
+#define _raw_write_relax(lock) cpu_relax()
+
 #endif /* __ASM_SPINLOCK_H */
index 52df72a..5f43df0 100644 (file)
@@ -160,4 +160,8 @@ static __inline__ int is_write_locked(rwlock_t *rw)
        return rw->counter < 0;
 }
 
+#define _raw_spin_relax(lock)  cpu_relax()
+#define _raw_read_relax(lock)  cpu_relax()
+#define _raw_write_relax(lock) cpu_relax()
+
 #endif /* __ASM_ARCH_SPINLOCK_H */
index b0b3043..c18b71f 100644 (file)
@@ -205,4 +205,8 @@ static inline void __raw_write_unlock(raw_rwlock_t *rw)
                                 : "+m" (rw->lock) : : "memory");
 }
 
+#define _raw_spin_relax(lock)  cpu_relax()
+#define _raw_read_relax(lock)  cpu_relax()
+#define _raw_write_relax(lock) cpu_relax()
+
 #endif /* __ASM_SPINLOCK_H */
index 9e83210..ff857e3 100644 (file)
@@ -213,4 +213,8 @@ static inline int __raw_read_trylock(raw_rwlock_t *x)
        return (u32)ia64_cmpxchg4_acq((__u32 *)(x), new.word, old.word) == old.word;
 }
 
+#define _raw_spin_relax(lock)  cpu_relax()
+#define _raw_read_relax(lock)  cpu_relax()
+#define _raw_write_relax(lock) cpu_relax()
+
 #endif /*  _ASM_IA64_SPINLOCK_H */
index f9f9072..f5cfba8 100644 (file)
@@ -316,4 +316,8 @@ static inline int __raw_write_trylock(raw_rwlock_t *lock)
        return 0;
 }
 
+#define _raw_spin_relax(lock)  cpu_relax()
+#define _raw_read_relax(lock)  cpu_relax()
+#define _raw_write_relax(lock) cpu_relax()
+
 #endif /* _ASM_M32R_SPINLOCK_H */
index 4c1a1b5..c8d5587 100644 (file)
@@ -328,4 +328,8 @@ static inline int __raw_write_trylock(raw_rwlock_t *rw)
 }
 
 
+#define _raw_spin_relax(lock)  cpu_relax()
+#define _raw_read_relax(lock)  cpu_relax()
+#define _raw_write_relax(lock) cpu_relax()
+
 #endif /* _ASM_SPINLOCK_H */
index a93960e..e182553 100644 (file)
@@ -152,4 +152,8 @@ static __inline__ int __raw_write_can_lock(raw_rwlock_t *rw)
        return !rw->counter;
 }
 
+#define _raw_spin_relax(lock)  cpu_relax()
+#define _raw_read_relax(lock)  cpu_relax()
+#define _raw_write_relax(lock) cpu_relax()
+
 #endif /* __ASM_SPINLOCK_H */
index c31e438..eaccac8 100644 (file)
@@ -285,5 +285,9 @@ static __inline__ void __raw_write_unlock(raw_rwlock_t *rw)
        rw->lock = 0;
 }
 
+#define _raw_spin_relax(lock)  cpu_relax()
+#define _raw_read_relax(lock)  cpu_relax()
+#define _raw_write_relax(lock) cpu_relax()
+
 #endif /* __KERNEL__ */
 #endif /* __ASM_SPINLOCK_H */
index 5c64b75..fccaf55 100644 (file)
@@ -161,4 +161,8 @@ static __inline__ void __raw_write_unlock(raw_rwlock_t *rw)
        rw->lock = 0;
 }
 
+#define _raw_spin_relax(lock)  cpu_relax()
+#define _raw_read_relax(lock)  cpu_relax()
+#define _raw_write_relax(lock) cpu_relax()
+
 #endif /* __ASM_SPINLOCK_H */
index ce3edf6..5f00fea 100644 (file)
@@ -154,4 +154,8 @@ static inline int __raw_write_trylock(raw_rwlock_t *rw)
        return _raw_write_trylock_retry(rw);
 }
 
+#define _raw_spin_relax(lock)  cpu_relax()
+#define _raw_read_relax(lock)  cpu_relax()
+#define _raw_write_relax(lock) cpu_relax()
+
 #endif /* __ASM_SPINLOCK_H */
index 846322d..54458fd 100644 (file)
@@ -100,4 +100,8 @@ static inline int __raw_write_trylock(raw_rwlock_t *rw)
        return 0;
 }
 
+#define _raw_spin_relax(lock)  cpu_relax()
+#define _raw_read_relax(lock)  cpu_relax()
+#define _raw_write_relax(lock) cpu_relax()
+
 #endif /* __ASM_SH_SPINLOCK_H */
index 1c75474..557d089 100644 (file)
@@ -154,6 +154,10 @@ static inline int __raw_write_trylock(raw_rwlock_t *rw)
 #define __raw_spin_lock_flags(lock, flags) __raw_spin_lock(lock)
 #define __raw_read_trylock(lock) generic__raw_read_trylock(lock)
 
+#define _raw_spin_relax(lock)  cpu_relax()
+#define _raw_read_relax(lock)  cpu_relax()
+#define _raw_write_relax(lock) cpu_relax()
+
 #define __raw_read_can_lock(rw) (!((rw)->lock & 0xff))
 #define __raw_write_can_lock(rw) (!(rw)->lock)
 
index bd5ffc7..0006fe9 100644 (file)
@@ -241,6 +241,10 @@ static int inline __write_trylock(raw_rwlock_t *lock)
 #define __raw_read_can_lock(rw)                (!((rw)->lock & 0x80000000UL))
 #define __raw_write_can_lock(rw)       (!(rw)->lock)
 
+#define _raw_spin_relax(lock)  cpu_relax()
+#define _raw_read_relax(lock)  cpu_relax()
+#define _raw_write_relax(lock) cpu_relax()
+
 #endif /* !(__ASSEMBLY__) */
 
 #endif /* !(__SPARC64_SPINLOCK_H) */
index 3daf5b0..05ef097 100644 (file)
@@ -133,4 +133,8 @@ static inline void __raw_write_unlock(raw_rwlock_t *rw)
                                : "=m" (rw->lock) : : "memory");
 }
 
+#define _raw_spin_relax(lock)  cpu_relax()
+#define _raw_read_relax(lock)  cpu_relax()
+#define _raw_write_relax(lock) cpu_relax()
+
 #endif /* __ASM_SPINLOCK_H */
index d48143e..476c374 100644 (file)
@@ -215,7 +215,7 @@ void __lockfunc _##op##_lock(locktype##_t *lock)                    \
                if (!(lock)->break_lock)                                \
                        (lock)->break_lock = 1;                         \
                while (!op##_can_lock(lock) && (lock)->break_lock)      \
-                       cpu_relax();                                    \
+                       _raw_##op##_relax(&lock->raw_lock);             \
        }                                                               \
        (lock)->break_lock = 0;                                         \
 }                                                                      \
@@ -237,7 +237,7 @@ unsigned long __lockfunc _##op##_lock_irqsave(locktype##_t *lock)   \
                if (!(lock)->break_lock)                                \
                        (lock)->break_lock = 1;                         \
                while (!op##_can_lock(lock) && (lock)->break_lock)      \
-                       cpu_relax();                                    \
+                       _raw_##op##_relax(&lock->raw_lock);             \
        }                                                               \
        (lock)->break_lock = 0;                                         \
        return flags;                                                   \