syncer_thread: sleep without lbolt

The syncer_thread() uses lbolt to perform periodic execution.  We can
do this without lbolt.

-  Adding a local wakeup(9) channel (syncer_chan) and sleep on it.

- Use a local copy of getnsecuptime() to get 1/hz resolution for time
  measurements.  This is much better than using gettime(9), which is
  wholly unsuitable for this use case.  Measure how long we spend in
  the loop and use this to calculate how long to sleep until the next
  execution.

  NB: getnsecuptime() is probably ready to be moved to kern_tc.c and
  documented.

- Using the system uptime instead of the UTC time avoids issues with
  time jumps.

ok mpi@

1 files changed, 25 insertions, 6 deletions

diff --git a/sys/kern/vfs_sync.c b/sys/kern/vfs_sync.c
index 62693a1d2f0..b7d16fa3b6b 100644
--- a/sys/kern/vfs_sync.c
+++ b/sys/kern/vfs_sync.c
@@ -1,4 +1,4 @@
-/*       $OpenBSD: vfs_sync.c,v 1.64 2020/06/24 22:03:41 cheloha Exp $  */
+/*       $OpenBSD: vfs_sync.c,v 1.65 2021/01/14 03:32:01 cheloha Exp $  */
 
 /*
  *  Portions of this code are:
@@ -48,6 +48,7 @@
 #include <sys/vnode.h>
 #include <sys/lock.h>
 #include <sys/malloc.h>
+#include <sys/time.h>
 
 #include <sys/kernel.h>
 #include <sys/sched.h>
@@ -73,6 +74,7 @@ LIST_HEAD(synclist, vnode);
 static struct synclist *syncer_workitem_pending;
 
 struct proc *syncerproc;
+int syncer_chan;
 
 /*
  * The workitem queue.
@@ -130,19 +132,32 @@ vn_syncer_add_to_worklist(struct vnode *vp, int delay)
 }
 
 /*
+ * TODO Move getnsecuptime() to kern_tc.c and document it when we have
+ * more users in the kernel.
+ */
+static uint64_t
+getnsecuptime(void)
+{
+	struct timespec now;
+
+	getnanouptime(&now);
+	return TIMESPEC_TO_NSEC(&now);
+}
+
+/*
  * System filesystem synchronizer daemon.
  */
 void
 syncer_thread(void *arg)
 {
+	uint64_t elapsed, start;
 	struct proc *p = curproc;
 	struct synclist *slp;
 	struct vnode *vp;
-	time_t starttime;
 	int s;
 
 	for (;;) {
-		starttime = gettime();
+		start = getnsecuptime();
 
 		/*
 		 * Push files whose dirty time has expired.
@@ -220,6 +235,7 @@ syncer_thread(void *arg)
 			rushjob -= 1;
 			continue;
 		}
+
 		/*
 		 * If it has taken us less than a second to process the
 		 * current work, then wait. Otherwise start right over
@@ -228,8 +244,11 @@ syncer_thread(void *arg)
 		 * matter as we are just trying to generally pace the
 		 * filesystem activity.
 		 */
-		if (gettime() == starttime)
-			tsleep_nsec(&lbolt, PPAUSE, "syncer", INFSLP);
+		elapsed = getnsecuptime() - start;
+		if (elapsed < SEC_TO_NSEC(1)) {
+			tsleep_nsec(&syncer_chan, PPAUSE, "syncer",
+			    SEC_TO_NSEC(1) - elapsed);
+		}
 	}
 }
 
@@ -242,7 +261,7 @@ int
 speedup_syncer(void)
 {
 	if (syncerproc)
-		wakeup_proc(syncerproc, &lbolt);
+		wakeup_proc(syncerproc, &syncer_chan);
 	if (rushjob < syncdelay / 2) {
 		rushjob += 1;
 		stat_rush_requests += 1;