vserver 1.9.3
[linux-2.6.git] / include / asm-ppc64 / spinlock.h
index d2074c4..45a6f75 100644 (file)
@@ -6,6 +6,8 @@
  *
  * Copyright (C) 2001-2004 Paul Mackerras <paulus@au.ibm.com>, IBM
  * Copyright (C) 2001 Anton Blanchard <anton@au.ibm.com>, IBM
+ * Copyright (C) 2002 Dave Engebretsen <engebret@us.ibm.com>, IBM
+ *     Rework to support virtual processors
  *
  * Type of int is used as a full 64b word is not necessary.
  *
  */
 #include <linux/config.h>
 #include <asm/paca.h>
+#include <asm/hvcall.h>
+#include <asm/iSeries/HvCall.h>
 
 typedef struct {
        volatile unsigned int lock;
 } spinlock_t;
 
+typedef struct {
+       volatile signed int lock;
+} rwlock_t;
+
 #ifdef __KERNEL__
 #define SPIN_LOCK_UNLOCKED     (spinlock_t) { 0 }
 
@@ -34,101 +42,91 @@ static __inline__ void _raw_spin_unlock(spinlock_t *lock)
 }
 
 /*
- * Normally we use the spinlock functions in arch/ppc64/lib/locks.c.
- * For special applications such as profiling, we can have the
- * spinlock functions inline by defining CONFIG_SPINLINE.
- * This is not recommended on partitioned systems with shared
- * processors, since the inline spinlock functions don't include
- * the code for yielding the CPU to the lock holder.
+ * On a system with shared processors (that is, where a physical
+ * processor is multiplexed between several virtual processors),
+ * there is no point spinning on a lock if the holder of the lock
+ * isn't currently scheduled on a physical processor.  Instead
+ * we detect this situation and ask the hypervisor to give the
+ * rest of our timeslice to the lock holder.
+ *
+ * So that we can tell which virtual processor is holding a lock,
+ * we put 0x80000000 | smp_processor_id() in the lock when it is
+ * held.  Conveniently, we have a word in the paca that holds this
+ * value.
  */
 
-#ifndef CONFIG_SPINLINE
-extern int _raw_spin_trylock(spinlock_t *lock);
-extern void _raw_spin_lock(spinlock_t *lock);
-extern void _raw_spin_lock_flags(spinlock_t *lock, unsigned long flags);
+#if defined(CONFIG_PPC_SPLPAR) || defined(CONFIG_PPC_ISERIES)
+/* We only yield to the hypervisor if we are in shared processor mode */
+#define SHARED_PROCESSOR (get_paca()->lppaca.xSharedProc)
+extern void __spin_yield(spinlock_t *lock);
+extern void __rw_yield(rwlock_t *lock);
+#else /* SPLPAR || ISERIES */
+#define __spin_yield(x)        barrier()
+#define __rw_yield(x)  barrier()
+#define SHARED_PROCESSOR       0
+#endif
 extern void spin_unlock_wait(spinlock_t *lock);
 
-#else
-
-static __inline__ int _raw_spin_trylock(spinlock_t *lock)
+/*
+ * This returns the old value in the lock, so we succeeded
+ * in getting the lock if the return value is 0.
+ */
+static __inline__ unsigned long __spin_trylock(spinlock_t *lock)
 {
-       unsigned int tmp, tmp2;
+       unsigned long tmp, tmp2;
 
        __asm__ __volatile__(
-"1:    lwarx           %0,0,%2         # spin_trylock\n\
+"      lwz             %1,%3(13)               # __spin_trylock\n\
+1:     lwarx           %0,0,%2\n\
        cmpwi           0,%0,0\n\
        bne-            2f\n\
-       lwz             %1,%3(13)\n\
        stwcx.          %1,0,%2\n\
        bne-            1b\n\
        isync\n\
-2:"    : "=&r"(tmp), "=&r"(tmp2)
-       : "r"(&lock->lock), "i"(offsetof(struct paca_struct, lock_token))
+2:"    : "=&r" (tmp), "=&r" (tmp2)
+       : "r" (&lock->lock), "i" (offsetof(struct paca_struct, lock_token))
        : "cr0", "memory");
 
-       return tmp == 0;
+       return tmp;
 }
 
-static __inline__ void _raw_spin_lock(spinlock_t *lock)
+static int __inline__ _raw_spin_trylock(spinlock_t *lock)
 {
-       unsigned int tmp;
-
-       __asm__ __volatile__(
-       "b              2f              # spin_lock\n\
-1:"
-       HMT_LOW
-"      lwzx            %0,0,%1\n\
-       cmpwi           0,%0,0\n\
-       bne+            1b\n"
-       HMT_MEDIUM
-"2:    lwarx           %0,0,%1\n\
-       cmpwi           0,%0,0\n\
-       bne-            1b\n\
-       lwz             %0,%2(13)\n\
-       stwcx.          %0,0,%1\n\
-       bne-            2b\n\
-       isync"
-       : "=&r"(tmp)
-       : "r"(&lock->lock), "i"(offsetof(struct paca_struct, lock_token))
-       : "cr0", "memory");
+       return __spin_trylock(lock) == 0;
 }
 
-/*
- * Note: if we ever want to inline the spinlocks on iSeries,
- * we will have to change the irq enable/disable stuff in here.
- */
-static __inline__ void _raw_spin_lock_flags(spinlock_t *lock,
-                                           unsigned long flags)
+static void __inline__ _raw_spin_lock(spinlock_t *lock)
 {
-       unsigned int tmp;
-       unsigned long tmp2;
-
-       __asm__ __volatile__(
-       "b              3f              # spin_lock\n\
-1:     mfmsr           %1\n\
-       mtmsrd          %3,1\n\
-2:"    HMT_LOW
-"      lwzx            %0,0,%2\n\
-       cmpwi           0,%0,0\n\
-       bne+            2b\n"
-       HMT_MEDIUM
-"      mtmsrd          %1,1\n\
-3:     lwarx           %0,0,%2\n\
-       cmpwi           0,%0,0\n\
-       bne-            1b\n\
-       lwz             %1,%4(13)\n\
-       stwcx.          %1,0,%2\n\
-       bne-            3b\n\
-       isync"
-       : "=&r"(tmp), "=&r"(tmp2)
-       : "r"(&lock->lock), "r"(flags),
-         "i" (offsetof(struct paca_struct, lock_token))
-       : "cr0", "memory");
+       while (1) {
+               if (likely(__spin_trylock(lock) == 0))
+                       break;
+               do {
+                       HMT_low();
+                       if (SHARED_PROCESSOR)
+                               __spin_yield(lock);
+               } while (likely(lock->lock != 0));
+               HMT_medium();
+       }
 }
 
-#define spin_unlock_wait(x)    do { cpu_relax(); } while (spin_is_locked(x))
-
-#endif /* CONFIG_SPINLINE */
+static void __inline__ _raw_spin_lock_flags(spinlock_t *lock, unsigned long flags)
+{
+       unsigned long flags_dis;
+
+       while (1) {
+               if (likely(__spin_trylock(lock) == 0))
+                       break;
+               local_save_flags(flags_dis);
+               local_irq_restore(flags);
+               do {
+                       HMT_low();
+                       if (SHARED_PROCESSOR)
+                               __spin_yield(lock);
+               } while (likely(lock->lock != 0));
+               HMT_medium();
+               local_irq_restore(flags_dis);
+       }
+}
 
 /*
  * Read-write spinlocks, allowing multiple readers
@@ -140,10 +138,6 @@ static __inline__ void _raw_spin_lock_flags(spinlock_t *lock,
  * irq-safe write-lock, but readers can get non-irqsafe
  * read-locks.
  */
-typedef struct {
-       volatile signed int lock;
-} rwlock_t;
-
 #define RW_LOCK_UNLOCKED (rwlock_t) { 0 }
 
 #define rwlock_init(x)         do { *(x) = RW_LOCK_UNLOCKED; } while(0)
@@ -165,67 +159,54 @@ static __inline__ void _raw_write_unlock(rwlock_t *rw)
        rw->lock = 0;
 }
 
-#ifndef CONFIG_SPINLINE
-extern int _raw_read_trylock(rwlock_t *rw);
-extern void _raw_read_lock(rwlock_t *rw);
-extern void _raw_read_unlock(rwlock_t *rw);
-extern int _raw_write_trylock(rwlock_t *rw);
-extern void _raw_write_lock(rwlock_t *rw);
-extern void _raw_write_unlock(rwlock_t *rw);
-
-#else
-static __inline__ int _raw_read_trylock(rwlock_t *rw)
+/*
+ * This returns the old value in the lock + 1,
+ * so we got a read lock if the return value is > 0.
+ */
+static long __inline__ __read_trylock(rwlock_t *rw)
 {
-       unsigned int tmp;
-       unsigned int ret;
+       long tmp;
 
        __asm__ __volatile__(
-"1:    lwarx           %0,0,%2         # read_trylock\n\
-       li              %1,0\n\
+"1:    lwarx           %0,0,%1         # read_trylock\n\
        extsw           %0,%0\n\
        addic.          %0,%0,1\n\
        ble-            2f\n\
-       stwcx.          %0,0,%2\n\
+       stwcx.          %0,0,%1\n\
        bne-            1b\n\
-       li              %1,1\n\
        isync\n\
-2:"    : "=&r"(tmp), "=&r"(ret)
-       : "r"(&rw->lock)
-       : "cr0", "memory");
+2:"    : "=&r" (tmp)
+       : "r" (&rw->lock)
+       : "cr0", "xer", "memory");
 
-       return ret;
+       return tmp;
 }
 
-static __inline__ void _raw_read_lock(rwlock_t *rw)
+static int __inline__ _raw_read_trylock(rwlock_t *rw)
 {
-       unsigned int tmp;
+       return __read_trylock(rw) > 0;
+}
 
-       __asm__ __volatile__(
-       "b              2f              # read_lock\n\
-1:"
-       HMT_LOW
-"      lwax            %0,0,%1\n\
-       cmpwi           0,%0,0\n\
-       blt+            1b\n"
-       HMT_MEDIUM
-"2:    lwarx           %0,0,%1\n\
-       extsw           %0,%0\n\
-       addic.          %0,%0,1\n\
-       ble-            1b\n\
-       stwcx.          %0,0,%1\n\
-       bne-            2b\n\
-       isync"
-       : "=&r"(tmp)
-       : "r"(&rw->lock)
-       : "cr0", "memory");
+static void __inline__ _raw_read_lock(rwlock_t *rw)
+{
+       while (1) {
+               if (likely(__read_trylock(rw) > 0))
+                       break;
+               do {
+                       HMT_low();
+                       if (SHARED_PROCESSOR)
+                               __rw_yield(rw);
+               } while (likely(rw->lock < 0));
+               HMT_medium();
+       }
 }
 
-static __inline__ void _raw_read_unlock(rwlock_t *rw)
+static void __inline__ _raw_read_unlock(rwlock_t *rw)
 {
-       unsigned int tmp;
+       long tmp;
 
        __asm__ __volatile__(
-       "lwsync                         # read_unlock\n\
+       "eieio                          # read_unlock\n\
 1:     lwarx           %0,0,%1\n\
        addic           %0,%0,-1\n\
        stwcx.          %0,0,%1\n\
@@ -235,50 +216,47 @@ static __inline__ void _raw_read_unlock(rwlock_t *rw)
        : "cr0", "memory");
 }
 
-static __inline__ int _raw_write_trylock(rwlock_t *rw)
+/*
+ * This returns the old value in the lock,
+ * so we got the write lock if the return value is 0.
+ */
+static __inline__ long __write_trylock(rwlock_t *rw)
 {
-       unsigned int tmp;
-       unsigned int ret;
+       long tmp, tmp2;
 
        __asm__ __volatile__(
-"1:    lwarx           %0,0,%2         # write_trylock\n\
+"      lwz             %1,%3(13)       # write_trylock\n\
+1:     lwarx           %0,0,%2\n\
        cmpwi           0,%0,0\n\
-       li              %1,0\n\
        bne-            2f\n\
-       stwcx.          %3,0,%2\n\
+       stwcx.          %1,0,%2\n\
        bne-            1b\n\
-       li              %1,1\n\
        isync\n\
-2:"    : "=&r"(tmp), "=&r"(ret)
-       : "r"(&rw->lock), "r"(-1)
+2:"    : "=&r" (tmp), "=&r" (tmp2)
+       : "r" (&rw->lock), "i" (offsetof(struct paca_struct, lock_token))
        : "cr0", "memory");
 
-       return ret;
+       return tmp;
 }
 
-static __inline__ void _raw_write_lock(rwlock_t *rw)
+static int __inline__ _raw_write_trylock(rwlock_t *rw)
 {
-       unsigned int tmp;
+       return __write_trylock(rw) == 0;
+}
 
-       __asm__ __volatile__(
-       "b              2f              # write_lock\n\
-1:"
-       HMT_LOW
-       "lwax           %0,0,%1\n\
-       cmpwi           0,%0,0\n\
-       bne+            1b\n"
-       HMT_MEDIUM
-"2:    lwarx           %0,0,%1\n\
-       cmpwi           0,%0,0\n\
-       bne-            1b\n\
-       stwcx.          %2,0,%1\n\
-       bne-            2b\n\
-       isync"
-       : "=&r"(tmp)
-       : "r"(&rw->lock), "r"(-1)
-       : "cr0", "memory");
+static void __inline__ _raw_write_lock(rwlock_t *rw)
+{
+       while (1) {
+               if (likely(__write_trylock(rw) == 0))
+                       break;
+               do {
+                       HMT_low();
+                       if (SHARED_PROCESSOR)
+                               __rw_yield(rw);
+               } while (likely(rw->lock != 0));
+               HMT_medium();
+       }
 }
-#endif /* CONFIG_SPINLINE */
 
 #endif /* __KERNEL__ */
 #endif /* __ASM_SPINLOCK_H */