lockmgr: add adaptive spinning

It is very conservative. Only spinning when LK_ADAPTIVE is passed, only on
exclusive lock and never when any waiters are present. buffer cache is remains
not spinning.

This reduces total sleep times during buildworld etc., but it does not shorten
total real time (culprits are contention in the vm subsystem along with slock +
upgrade which is not covered).

For microbenchmarks: open3_processes -t 52 (open/close of the same file for
writing) ops/s:
before: 258845
after: 801638

Reviewed by:	kib
Tested by:	pho
Differential Revision:	https://reviews.freebsd.org/D25753

1 files changed, 111 insertions, 18 deletions

diff --git a/sys/kern/kern_lock.c b/sys/kern/kern_lock.c
index 11396a6e259f..36084f3877bd 100644
--- a/sys/kern/kern_lock.c
+++ b/sys/kern/kern_lock.c
@@ -167,6 +167,12 @@ struct lock_class lock_class_lockmgr = {
 #endif
 };
 
+static __read_mostly bool lk_adaptive = true;
+static SYSCTL_NODE(_debug, OID_AUTO, lockmgr, CTLFLAG_RD, NULL, "lockmgr debugging");
+SYSCTL_BOOL(_debug_lockmgr, OID_AUTO, adaptive_spinning, CTLFLAG_RW, &lk_adaptive,
+    0, "");
+#define lockmgr_delay  locks_delay
+
 struct lockmgr_wait {
 	const char *iwmesg;
 	int ipri;
@@ -515,7 +521,6 @@ lockmgr_slock_try(struct lock *lk, uintptr_t *xp, int flags, bool fp)
 	 * waiters, if we fail to acquire the shared lock
 	 * loop back and retry.
 	 */
-	*xp = lockmgr_read_value(lk);
 	while (LK_CAN_SHARE(*xp, flags, fp)) {
 		if (atomic_fcmpset_acq_ptr(&lk->lk_lock, xp,
 		    *xp + LK_ONE_SHARER)) {
@@ -541,6 +546,38 @@ lockmgr_sunlock_try(struct lock *lk, uintptr_t *xp)
 	return (false);
 }
 
+static bool
+lockmgr_slock_adaptive(struct lock_delay_arg *lda, struct lock *lk, uintptr_t *xp,
+    int flags)
+{
+	struct thread *owner;
+	uintptr_t x;
+
+	x = *xp;
+	MPASS(x != LK_UNLOCKED);
+	owner = (struct thread *)LK_HOLDER(x);
+	for (;;) {
+		MPASS(owner != curthread);
+		if (owner == (struct thread *)LK_KERNPROC)
+			return (false);
+		if ((x & LK_SHARE) && LK_SHARERS(x) > 0)
+			return (false);
+		if (owner == NULL)
+			return (false);
+		if (!TD_IS_RUNNING(owner))
+			return (false);
+		if ((x & LK_ALL_WAITERS) != 0)
+			return (false);
+		lock_delay(lda);
+		x = lockmgr_read_value(lk);
+		if (LK_CAN_SHARE(x, flags, false)) {
+			*xp = x;
+			return (true);
+		}
+		owner = (struct thread *)LK_HOLDER(x);
+	}
+}
+
 static __noinline int
 lockmgr_slock_hard(struct lock *lk, u_int flags, struct lock_object *ilk,
     const char *file, int line, struct lockmgr_wait *lwa)
@@ -557,6 +594,7 @@ lockmgr_slock_hard(struct lock *lk, u_int flags, struct lock_object *ilk,
 	uint64_t waittime = 0;
 	int contested = 0;
 #endif
+	struct lock_delay_arg lda;
 
 	if (KERNEL_PANICKED())
 		goto out;
@@ -566,9 +604,31 @@ lockmgr_slock_hard(struct lock *lk, u_int flags, struct lock_object *ilk,
 	if (LK_CAN_WITNESS(flags))
 		WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER,
 		    file, line, flags & LK_INTERLOCK ? ilk : NULL);
+	lock_delay_arg_init(&lda, &lockmgr_delay);
+	if (!lk_adaptive)
+		flags &= ~LK_ADAPTIVE;
+	x = lockmgr_read_value(lk);
+	/*
+	 * The lock may already be locked exclusive by curthread,
+	 * avoid deadlock.
+	 */
+	if (LK_HOLDER(x) == tid) {
+		LOCK_LOG2(lk,
+		    "%s: %p already held in exclusive mode",
+		    __func__, lk);
+		error = EDEADLK;
+		goto out;
+	}
+
 	for (;;) {
 		if (lockmgr_slock_try(lk, &x, flags, false))
 			break;
+
+		if ((flags & (LK_ADAPTIVE | LK_INTERLOCK)) == LK_ADAPTIVE) {
+			if (lockmgr_slock_adaptive(&lda, lk, &x, flags))
+				continue;
+		}
+
 #ifdef HWPMC_HOOKS
 		PMC_SOFT_CALL( , , lock, failed);
 #endif
@@ -576,18 +636,6 @@ lockmgr_slock_hard(struct lock *lk, u_int flags, struct lock_object *ilk,
 		    &contested, &waittime);
 
 		/*
-		 * If the lock is already held by curthread in
-		 * exclusive way avoid a deadlock.
-		 */
-		if (LK_HOLDER(x) == tid) {
-			LOCK_LOG2(lk,
-			    "%s: %p already held in exclusive mode",
-			    __func__, lk);
-			error = EDEADLK;
-			break;
-		}
-
-		/*
 		 * If the lock is expected to not sleep just give up
 		 * and return.
 		 */
@@ -660,6 +708,7 @@ retry_sleepq:
 		}
 		LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
 		    __func__, lk);
+		x = lockmgr_read_value(lk);
 	}
 	if (error == 0) {
 #ifdef KDTRACE_HOOKS
@@ -682,6 +731,37 @@ out:
 	return (error);
 }
 
+static bool
+lockmgr_xlock_adaptive(struct lock_delay_arg *lda, struct lock *lk, uintptr_t *xp)
+{
+	struct thread *owner;
+	uintptr_t x;
+
+	x = *xp;
+	MPASS(x != LK_UNLOCKED);
+	owner = (struct thread *)LK_HOLDER(x);
+	for (;;) {
+		MPASS(owner != curthread);
+		if (owner == NULL)
+			return (false);
+		if ((x & LK_SHARE) && LK_SHARERS(x) > 0)
+			return (false);
+		if (owner == (struct thread *)LK_KERNPROC)
+			return (false);
+		if (!TD_IS_RUNNING(owner))
+			return (false);
+		if ((x & LK_ALL_WAITERS) != 0)
+			return (false);
+		lock_delay(lda);
+		x = lockmgr_read_value(lk);
+		if (x == LK_UNLOCKED) {
+			*xp = x;
+			return (true);
+		}
+		owner = (struct thread *)LK_HOLDER(x);
+	}
+}
+
 static __noinline int
 lockmgr_xlock_hard(struct lock *lk, u_int flags, struct lock_object *ilk,
     const char *file, int line, struct lockmgr_wait *lwa)
@@ -699,6 +779,7 @@ lockmgr_xlock_hard(struct lock *lk, u_int flags, struct lock_object *ilk,
 	uint64_t waittime = 0;
 	int contested = 0;
 #endif
+	struct lock_delay_arg lda;
 
 	if (KERNEL_PANICKED())
 		goto out;
@@ -747,10 +828,19 @@ lockmgr_xlock_hard(struct lock *lk, u_int flags, struct lock_object *ilk,
 		goto out;
 	}
 
+	x = LK_UNLOCKED;
+	lock_delay_arg_init(&lda, &lockmgr_delay);
+	if (!lk_adaptive)
+		flags &= ~LK_ADAPTIVE;
 	for (;;) {
-		if (lk->lk_lock == LK_UNLOCKED &&
-		    atomic_cmpset_acq_ptr(&lk->lk_lock, LK_UNLOCKED, tid))
-			break;
+		if (x == LK_UNLOCKED) {
+			if (atomic_fcmpset_acq_ptr(&lk->lk_lock, &x, tid))
+				break;
+		}
+		if ((flags & (LK_ADAPTIVE | LK_INTERLOCK)) == LK_ADAPTIVE) {
+			if (lockmgr_xlock_adaptive(&lda, lk, &x))
+				continue;
+		}
 #ifdef HWPMC_HOOKS
 		PMC_SOFT_CALL( , , lock, failed);
 #endif
@@ -853,6 +943,7 @@ retry_sleepq:
 		}
 		LOCK_LOG2(lk, "%s: %p resuming from the sleep queue",
 		    __func__, lk);
+		x = lockmgr_read_value(lk);
 	}
 	if (error == 0) {
 #ifdef KDTRACE_HOOKS
@@ -954,6 +1045,7 @@ lockmgr_lock_flags(struct lock *lk, u_int flags, struct lock_object *ilk,
 			    file, line, flags & LK_INTERLOCK ? ilk : NULL);
 		if (__predict_false(lk->lock_object.lo_flags & LK_NOSHARE))
 			break;
+		x = lockmgr_read_value(lk);
 		if (lockmgr_slock_try(lk, &x, flags, true)) {
 			lockmgr_note_shared_acquire(lk, 0, 0,
 			    file, line, flags);
@@ -1139,12 +1231,13 @@ lockmgr_slock(struct lock *lk, u_int flags, const char *file, int line)
 	if (LK_CAN_WITNESS(flags))
 		WITNESS_CHECKORDER(&lk->lock_object, LOP_NEWORDER,
 		    file, line, NULL);
+	x = lockmgr_read_value(lk);
 	if (__predict_true(lockmgr_slock_try(lk, &x, flags, true))) {
 		lockmgr_note_shared_acquire(lk, 0, 0, file, line, flags);
 		return (0);
 	}
 
-	return (lockmgr_slock_hard(lk, flags, NULL, file, line, NULL));
+	return (lockmgr_slock_hard(lk, flags | LK_ADAPTIVE, NULL, file, line, NULL));
 }
 
 int
@@ -1165,7 +1258,7 @@ lockmgr_xlock(struct lock *lk, u_int flags, const char *file, int line)
 		return (0);
 	}
 
-	return (lockmgr_xlock_hard(lk, flags, NULL, file, line, NULL));
+	return (lockmgr_xlock_hard(lk, flags | LK_ADAPTIVE, NULL, file, line, NULL));
 }
 
 int