powerpc/qspinlock: add compile-time tuning adjustments

This adds compile-time options that allow the EH lock hint bit to be
enabled or disabled, and adds some new options that may or may not
help matters. To help with experimentation and tuning.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
Link: https://lore.kernel.org/r/20221126095932.1234527-18-npiggin@gmail.com

2 files changed, 94 insertions, 6 deletions

diff --git a/arch/powerpc/include/asm/qspinlock.h b/arch/powerpc/include/asm/qspinlock.h
index 93b1c976db8a..28a53fb69b38 100644
--- a/arch/powerpc/include/asm/qspinlock.h
+++ b/arch/powerpc/include/asm/qspinlock.h
@@ -6,15 +6,68 @@
 #include <asm/qspinlock_types.h>
 #include <asm/paravirt.h>
 
+#ifdef CONFIG_PPC64
+/*
+ * Use the EH=1 hint for accesses that result in the lock being acquired.
+ * The hardware is supposed to optimise this pattern by holding the lock
+ * cacheline longer, and releasing when a store to the same memory (the
+ * unlock) is performed.
+ */
+#define _Q_SPIN_EH_HINT 1
+#else
+#define _Q_SPIN_EH_HINT 0
+#endif
+
 /*
  * The trylock itself may steal. This makes trylocks slightly stronger, and
- * might make spin locks slightly more efficient when stealing.
+ * makes locks slightly more efficient when stealing.
  *
  * This is compile-time, so if true then there may always be stealers, so the
  * nosteal paths become unused.
  */
 #define _Q_SPIN_TRY_LOCK_STEAL 1
 
+/*
+ * Put a speculation barrier after testing the lock/node and finding it
+ * busy. Try to prevent pointless speculation in slow paths.
+ *
+ * Slows down the lockstorm microbenchmark with no stealing, where locking
+ * is purely FIFO through the queue. May have more benefit in real workload
+ * where speculating into the wrong place could have a greater cost.
+ */
+#define _Q_SPIN_SPEC_BARRIER 0
+
+#ifdef CONFIG_PPC64
+/*
+ * Execute a miso instruction after passing the MCS lock ownership to the
+ * queue head. Miso is intended to make stores visible to other CPUs sooner.
+ *
+ * This seems to make the lockstorm microbenchmark nospin test go slightly
+ * faster on POWER10, but disable for now.
+ */
+#define _Q_SPIN_MISO 0
+#else
+#define _Q_SPIN_MISO 0
+#endif
+
+#ifdef CONFIG_PPC64
+/*
+ * This executes miso after an unlock of the lock word, having ownership
+ * pass to the next CPU sooner. This will slow the uncontended path to some
+ * degree. Not evidence it helps yet.
+ */
+#define _Q_SPIN_MISO_UNLOCK 0
+#else
+#define _Q_SPIN_MISO_UNLOCK 0
+#endif
+
+/*
+ * Seems to slow down lockstorm microbenchmark, suspect queue node just
+ * has to become shared again right afterwards when its waiter spins on
+ * the lock field.
+ */
+#define _Q_SPIN_PREFETCH_NEXT 0
+
 static __always_inline int queued_spin_is_locked(struct qspinlock *lock)
 {
 	return READ_ONCE(lock->val);
@@ -52,7 +105,7 @@ static __always_inline int __queued_spin_trylock_nosteal(struct qspinlock *lock)
 "2:									\n"
 	: "=&r" (prev)
 	: "r" (&lock->val), "r" (new),
-	  "i" (IS_ENABLED(CONFIG_PPC64))
+	  "i" (_Q_SPIN_EH_HINT)
 	: "cr0", "memory");
 
 	return likely(prev == 0);
@@ -76,7 +129,7 @@ static __always_inline int __queued_spin_trylock_steal(struct qspinlock *lock)
 "2:									\n"
 	: "=&r" (prev), "=&r" (tmp)
 	: "r" (&lock->val), "r" (new), "r" (_Q_TAIL_CPU_MASK),
-	  "i" (IS_ENABLED(CONFIG_PPC64))
+	  "i" (_Q_SPIN_EH_HINT)
 	: "cr0", "memory");
 
 	return likely(!(prev & ~_Q_TAIL_CPU_MASK));
@@ -101,6 +154,8 @@ static __always_inline void queued_spin_lock(struct qspinlock *lock)
 static inline void queued_spin_unlock(struct qspinlock *lock)
 {
 	smp_store_release(&lock->locked, 0);
+	if (_Q_SPIN_MISO_UNLOCK)
+		asm volatile("miso" ::: "memory");
 }
 
 #define arch_spin_is_locked(l)		queued_spin_is_locked(l)
diff --git a/arch/powerpc/lib/qspinlock.c b/arch/powerpc/lib/qspinlock.c
index 0f33a07c1d19..1cf5d3e75250 100644
--- a/arch/powerpc/lib/qspinlock.c
+++ b/arch/powerpc/lib/qspinlock.c
@@ -48,6 +48,12 @@ static bool pv_prod_head __read_mostly = false;
 static DEFINE_PER_CPU_ALIGNED(struct qnodes, qnodes);
 static DEFINE_PER_CPU_ALIGNED(u64, sleepy_lock_seen_clock);
 
+#if _Q_SPIN_SPEC_BARRIER == 1
+#define spec_barrier() do { asm volatile("ori 31,31,0" ::: "memory"); } while (0)
+#else
+#define spec_barrier() do { } while (0)
+#endif
+
 static __always_inline bool recently_sleepy(void)
 {
 	/* pv_sleepy_lock is true when this is called */
@@ -137,7 +143,7 @@ static __always_inline u32 trylock_clean_tail(struct qspinlock *lock, u32 tail)
 	: "r" (&lock->val), "r"(tail), "r" (newval),
 	  "i" (_Q_LOCKED_VAL),
 	  "r" (_Q_TAIL_CPU_MASK),
-	  "i" (IS_ENABLED(CONFIG_PPC64))
+	  "i" (_Q_SPIN_EH_HINT)
 	: "cr0", "memory");
 
 	return prev;
@@ -475,6 +481,7 @@ static __always_inline bool try_to_steal_lock(struct qspinlock *lock, bool parav
 		val = READ_ONCE(lock->val);
 		if (val & _Q_MUST_Q_VAL)
 			break;
+		spec_barrier();
 
 		if (unlikely(!(val & _Q_LOCKED_VAL))) {
 			spin_end();
@@ -540,6 +547,7 @@ static __always_inline void queued_spin_lock_mcs_queue(struct qspinlock *lock, b
 
 	qnodesp = this_cpu_ptr(&qnodes);
 	if (unlikely(qnodesp->count >= MAX_NODES)) {
+		spec_barrier();
 		while (!queued_spin_trylock(lock))
 			cpu_relax();
 		return;
@@ -576,9 +584,12 @@ static __always_inline void queued_spin_lock_mcs_queue(struct qspinlock *lock, b
 		/* Wait for mcs node lock to be released */
 		spin_begin();
 		while (!node->locked) {
+			spec_barrier();
+
 			if (yield_to_prev(lock, node, old, paravirt))
 				seen_preempted = true;
 		}
+		spec_barrier();
 		spin_end();
 
 		/* Clear out stale propagated yield_cpu */
@@ -586,6 +597,17 @@ static __always_inline void queued_spin_lock_mcs_queue(struct qspinlock *lock, b
 			node->yield_cpu = -1;
 
 		smp_rmb(); /* acquire barrier for the mcs lock */
+
+		/*
+		 * Generic qspinlocks have this prefetch here, but it seems
+		 * like it could cause additional line transitions because
+		 * the waiter will keep loading from it.
+		 */
+		if (_Q_SPIN_PREFETCH_NEXT) {
+			next = READ_ONCE(node->next);
+			if (next)
+				prefetchw(next);
+		}
 	}
 
 	/* We're at the head of the waitqueue, wait for the lock. */
@@ -597,6 +619,7 @@ again:
 		val = READ_ONCE(lock->val);
 		if (!(val & _Q_LOCKED_VAL))
 			break;
+		spec_barrier();
 
 		if (paravirt && pv_sleepy_lock && maybe_stealers) {
 			if (!sleepy) {
@@ -637,6 +660,7 @@ again:
 			val |= _Q_MUST_Q_VAL;
 		}
 	}
+	spec_barrier();
 	spin_end();
 
 	/* If we're the last queued, must clean up the tail. */
@@ -657,6 +681,7 @@ again:
 			cpu_relax();
 		spin_end();
 	}
+	spec_barrier();
 
 	/*
 	 * Unlock the next mcs waiter node. Release barrier is not required
@@ -668,10 +693,14 @@ again:
 	if (paravirt && pv_prod_head) {
 		int next_cpu = next->cpu;
 		WRITE_ONCE(next->locked, 1);
+		if (_Q_SPIN_MISO)
+			asm volatile("miso" ::: "memory");
 		if (vcpu_is_preempted(next_cpu))
 			prod_cpu(next_cpu);
 	} else {
 		WRITE_ONCE(next->locked, 1);
+		if (_Q_SPIN_MISO)
+			asm volatile("miso" ::: "memory");
 	}
 
 release:
@@ -686,12 +715,16 @@ void queued_spin_lock_slowpath(struct qspinlock *lock)
 	 * is passed as the paravirt argument to the functions.
 	 */
 	if (IS_ENABLED(CONFIG_PARAVIRT_SPINLOCKS) && is_shared_processor()) {
-		if (try_to_steal_lock(lock, true))
+		if (try_to_steal_lock(lock, true)) {
+			spec_barrier();
 			return;
+		}
 		queued_spin_lock_mcs_queue(lock, true);
 	} else {
-		if (try_to_steal_lock(lock, false))
+		if (try_to_steal_lock(lock, false)) {
+			spec_barrier();
 			return;
+		}
 		queued_spin_lock_mcs_queue(lock, false);
 	}
 }