[PATCH v2 4/6] ARC: LLOCK/SCOND based rwlock

[Vineet Gupta <Vineet.Gupta1@synopsys.com>]

With LLOCK/SCOND, the rwlock counter can be atomically updated w/o need
for a guarding spin lock.

This in turn elides the EXchange instruction based spinning which causes
the cacheline transition to exclusive state and concurrent spinning
across cores would cause the line to keep bouncing around.
LLOCK/SCOND based implementation is superior as spinning on LLOCK keeps
the cacheline in shared state.

Cc: Peter Zijlstra (Intel) <peterz@infradead.org>
Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
---
 arch/arc/include/asm/spinlock.h       | 174 ++++++++++++++++++++++++++++++++--
 arch/arc/include/asm/spinlock_types.h |   2 +
 2 files changed, 166 insertions(+), 10 deletions(-)

diff --git a/arch/arc/include/asm/spinlock.h b/arch/arc/include/asm/spinlock.h
index 4f6c90a0a68a..9fd5a0221671 100644
--- a/arch/arc/include/asm/spinlock.h
+++ b/arch/arc/include/asm/spinlock.h
@@ -75,6 +75,164 @@ static inline void arch_spin_unlock(arch_spinlock_t *lock)
 	smp_mb();
 }
 
+/*
+ * Read-write spinlocks, allowing multiple readers but only one writer.
+ * Unfair locking as Writers could be starved indefinitely by Reader(s)
+ */
+
+static inline void arch_read_lock(arch_rwlock_t *rw)
+{
+	unsigned int val;
+
+	smp_mb();
+
+	/*
+	 * zero means writer holds the lock exclusively, deny Reader.
+	 * Otherwise grant lock to first/subseq reader
+	 *
+	 * 	if (rw->counter > 0) {
+	 *		rw->counter--;
+	 *		ret = 1;
+	 *	}
+	 */
+
+	__asm__ __volatile__(
+	"1:	llock	%[val], [%[rwlock]]	\n"
+	"	brls	%[val], %[WR_LOCKED], 1b\n"	/* <= 0: spin while write locked */
+	"	sub	%[val], %[val], 1	\n"	/* reader lock */
+	"	scond	%[val], [%[rwlock]]	\n"
+	"	bnz	1b			\n"
+	"					\n"
+	: [val]		"=&r"	(val)
+	: [rwlock]	"r"	(&(rw->counter)),
+	  [WR_LOCKED]	"ir"	(0)
+	: "memory", "cc");
+
+	smp_mb();
+}
+
+/* 1 - lock taken successfully */
+static inline int arch_read_trylock(arch_rwlock_t *rw)
+{
+	unsigned int val, got_it = 0;
+
+	smp_mb();
+
+	__asm__ __volatile__(
+	"1:	llock	%[val], [%[rwlock]]	\n"
+	"	brls	%[val], %[WR_LOCKED], 4f\n"	/* <= 0: already write locked, bail */
+	"	sub	%[val], %[val], 1	\n"	/* counter-- */
+	"	scond	%[val], [%[rwlock]]	\n"
+	"	bnz	1b			\n"	/* retry if collided with someone */
+	"	mov	%[got_it], 1		\n"
+	"					\n"
+	"4: ; --- done ---			\n"
+
+	: [val]		"=&r"	(val),
+	  [got_it]	"+&r"	(got_it)
+	: [rwlock]	"r"	(&(rw->counter)),
+	  [WR_LOCKED]	"ir"	(0)
+	: "memory", "cc");
+
+	smp_mb();
+
+	return got_it;
+}
+
+static inline void arch_write_lock(arch_rwlock_t *rw)
+{
+	unsigned int val;
+
+	smp_mb();
+
+	/*
+	 * If reader(s) hold lock (lock < __ARCH_RW_LOCK_UNLOCKED__),
+	 * deny writer. Otherwise if unlocked grant to writer
+	 * Hence the claim that Linux rwlocks are unfair to writers.
+	 * (can be starved for an indefinite time by readers).
+	 *
+	 *	if (rw->counter == __ARCH_RW_LOCK_UNLOCKED__) {
+	 *		rw->counter = 0;
+	 *		ret = 1;
+	 *	}
+	 */
+
+	__asm__ __volatile__(
+	"1:	llock	%[val], [%[rwlock]]	\n"
+	"	brne	%[val], %[UNLOCKED], 1b	\n"	/* while !UNLOCKED spin */
+	"	mov	%[val], %[WR_LOCKED]	\n"
+	"	scond	%[val], [%[rwlock]]	\n"
+	"	bnz	1b			\n"
+	"					\n"
+	: [val]		"=&r"	(val)
+	: [rwlock]	"r"	(&(rw->counter)),
+	  [UNLOCKED]	"ir"	(__ARCH_RW_LOCK_UNLOCKED__),
+	  [WR_LOCKED]	"ir"	(0)
+	: "memory", "cc");
+
+	smp_mb();
+}
+
+/* 1 - lock taken successfully */
+static inline int arch_write_trylock(arch_rwlock_t *rw)
+{
+	unsigned int val, got_it = 0;
+
+	smp_mb();
+
+	__asm__ __volatile__(
+	"1:	llock	%[val], [%[rwlock]]	\n"
+	"	brne	%[val], %[UNLOCKED], 4f	\n"	/* !UNLOCKED, bail */
+	"	mov	%[val], %[WR_LOCKED]	\n"
+	"	scond	%[val], [%[rwlock]]	\n"
+	"	bnz	1b			\n"	/* retry if collided with someone */
+	"	mov	%[got_it], 1		\n"
+	"					\n"
+	"4: ; --- done ---			\n"
+
+	: [val]		"=&r"	(val),
+	  [got_it]	"+&r"	(got_it)
+	: [rwlock]	"r"	(&(rw->counter)),
+	  [UNLOCKED]	"ir"	(__ARCH_RW_LOCK_UNLOCKED__),
+	  [WR_LOCKED]	"ir"	(0)
+	: "memory", "cc");
+
+	smp_mb();
+
+	return got_it;
+}
+
+static inline void arch_read_unlock(arch_rwlock_t *rw)
+{
+	unsigned int val;
+
+	smp_mb();
+
+	/*
+	 * rw->counter++;
+	 */
+	__asm__ __volatile__(
+	"1:	llock	%[val], [%[rwlock]]	\n"
+	"	add	%[val], %[val], 1	\n"
+	"	scond	%[val], [%[rwlock]]	\n"
+	"	bnz	1b			\n"
+	"					\n"
+	: [val]		"=&r"	(val)
+	: [rwlock]	"r"	(&(rw->counter))
+	: "memory", "cc");
+
+	smp_mb();
+}
+
+static inline void arch_write_unlock(arch_rwlock_t *rw)
+{
+	smp_mb();
+
+	rw->counter = __ARCH_RW_LOCK_UNLOCKED__;
+
+	smp_mb();
+}
+
 #else	/* !CONFIG_ARC_HAS_LLSC */
 
 static inline void arch_spin_lock(arch_spinlock_t *lock)
@@ -148,23 +306,14 @@ static inline void arch_spin_unlock(arch_spinlock_t *lock)
 	smp_mb();
 }
 
-#endif
-
 /*
  * Read-write spinlocks, allowing multiple readers but only one writer.
+ * Unfair locking as Writers could be starved indefinitely by Reader(s)
  *
  * The spinlock itself is contained in @counter and access to it is
  * serialized with @lock_mutex.
- *
- * Unfair locking as Writers could be starved indefinitely by Reader(s)
  */
 
-/* Would read_trylock() succeed? */
-#define arch_read_can_lock(x)	((x)->counter > 0)
-
-/* Would write_trylock() succeed? */
-#define arch_write_can_lock(x)	((x)->counter == __ARCH_RW_LOCK_UNLOCKED__)
-
 /* 1 - lock taken successfully */
 static inline int arch_read_trylock(arch_rwlock_t *rw)
 {
@@ -235,6 +384,11 @@ static inline void arch_write_unlock(arch_rwlock_t *rw)
 	arch_spin_unlock(&(rw->lock_mutex));
 }
 
+#endif
+
+#define arch_read_can_lock(x)	((x)->counter > 0)
+#define arch_write_can_lock(x)	((x)->counter == __ARCH_RW_LOCK_UNLOCKED__)
+
 #define arch_read_lock_flags(lock, flags)	arch_read_lock(lock)
 #define arch_write_lock_flags(lock, flags)	arch_write_lock(lock)
 
diff --git a/arch/arc/include/asm/spinlock_types.h b/arch/arc/include/asm/spinlock_types.h
index 662627ced4f2..4e1ef5f650c6 100644
--- a/arch/arc/include/asm/spinlock_types.h
+++ b/arch/arc/include/asm/spinlock_types.h
@@ -26,7 +26,9 @@ typedef struct {
  */
 typedef struct {
 	volatile unsigned int	counter;
+#ifndef CONFIG_ARC_HAS_LLSC
 	arch_spinlock_t		lock_mutex;
+#endif
 } arch_rwlock_t;
 
 #define __ARCH_RW_LOCK_UNLOCKED__	0x01000000
-- 
1.9.1