From: Matthew Dillon <dillon@apollo.backplane.com>
Date: Wed, 19 Sep 2012 03:55:18 +0000 (-0700)
Subject: kernel - Add usched_dfly algorith, set as default for now (4)
X-Git-Tag: v3.2.0~112
X-Git-Url: https://gitweb.dragonflybsd.org/~nant/dragonfly.git/commitdiff_plain/08bb2a8309b22300155c1fafe69f9a46d3526e44

kernel - Add usched_dfly algorith, set as default for now (4)

* Fix fork regression with usched_dfly.  Most fork/exec sequences involve
  the parent waiting.  The new scheduler was placing the newly forked
  process on another cpu which is non-optimal if the parent is going
  to immediately wait.

  Instead if there is nothing else waiting to run on the current cpu,
  leave the forked process on the current cpu initially.  If the parent
  waits quickly the forked process will get cpu, otherwise it will get
  scheduled away soon enough.  If the parent forks additional children
  then we find there is something on the queue now (the first child) and
  put the additional children on other cpus.

Reported-by: thesjg
---

diff --git a/sys/kern/usched_dfly.c b/sys/kern/usched_dfly.c
index 28848cb420..775849863c 100644
--- a/sys/kern/usched_dfly.c
+++ b/sys/kern/usched_dfly.c
@@ -90,6 +90,7 @@ int dfly_rebalanced;
 TAILQ_HEAD(rq, lwp);
 
 #define lwp_priority	lwp_usdata.dfly.priority
+#define lwp_forked	lwp_usdata.dfly.forked
 #define lwp_rqindex	lwp_usdata.dfly.rqindex
 #define lwp_estcpu	lwp_usdata.dfly.estcpu
 #define lwp_batch	lwp_usdata.dfly.batch
@@ -135,6 +136,9 @@ static void dfly_yield(struct lwp *lp);
 static dfly_pcpu_t dfly_choose_best_queue(struct lwp *lp);
 static dfly_pcpu_t dfly_choose_worst_queue(dfly_pcpu_t dd);
 static dfly_pcpu_t dfly_choose_queue_simple(dfly_pcpu_t dd, struct lwp *lp);
+#if 0
+static void dfly_wakeup_random_helper(dfly_pcpu_t notdd);
+#endif
 #endif
 
 #ifdef SMP
@@ -600,9 +604,25 @@ dfly_setrunqueue(struct lwp *lp)
 	 * target even if the current cpu has no running user thread (for
 	 * example, because the current cpu might be a hyperthread and its
 	 * sibling has a thread assigned).
+	 *
+	 * If we just forked it is most optimal to run the child on the same
+	 * cpu just in case the parent decides to wait for it (thus getting
+	 * off that cpu).  As long as there is nothing else runnable on the
+	 * cpu, that is.  If we did this unconditionally a parent forking
+	 * multiple children before waiting (e.g. make -j N) leaves other
+	 * cpus idle that could be working.
 	 */
 	/*spin_lock(&dfly_spin);*/
-	rdd = dfly_choose_best_queue(lp);
+	if (lp->lwp_forked) {
+		lp->lwp_forked = 0;
+		if (dfly_pcpu[lp->lwp_qcpu].runqcount)
+			rdd = dfly_choose_best_queue(lp);
+		else
+			rdd = &dfly_pcpu[lp->lwp_qcpu];
+		/* dfly_wakeup_random_helper(rdd); */
+	} else {
+		rdd = dfly_choose_best_queue(lp);
+	}
 	rgd = globaldata_find(rdd->cpuid);
 
 	/*
@@ -654,6 +674,50 @@ dfly_setrunqueue(struct lwp *lp)
 	crit_exit();
 }
 
+#if 0
+
+/*
+ * This wakes up a random helper that might have no work on its cpu to do.
+ * The idea is to improve fork/fork-exec/fork-wait/exec and similar
+ * process-spawning sequences by first scheduling the forked process
+ * on the same cpu as the parent, in case the parent is just going to
+ * wait*().  But if the parent does not wait we want another cpu to pick
+ * the forked process up ASAP.
+ *
+ * The ipi/helper-scheduling sequence typically takes a lot longer to run
+ * than a return-from-procedure-call and the parent then entering a
+ * wait*().  There's a race here that we want the parent to win ONLY if
+ * it is going to wait*().
+ *
+ * If a process sticks around for long enough normal scheduling action
+ * will move it to the right place.
+ */
+static
+void
+dfly_wakeup_random_helper(dfly_pcpu_t notdd)
+{
+	cpumask_t tmpmask;
+	cpumask_t mask;
+	int cpuid;
+
+	mask = dfly_rdyprocmask & ~dfly_curprocmask & smp_active_mask &
+	       usched_global_cpumask & ~notdd->cpumask;
+	++dfly_scancpu;
+	cpuid = (dfly_scancpu & 0xFFFF) % ncpus;
+
+	if (mask) {
+		tmpmask = ~(CPUMASK(cpuid) - 1);
+		if (mask & tmpmask)
+			cpuid = BSFCPUMASK(mask & tmpmask);
+		else
+			cpuid = BSFCPUMASK(mask);
+		atomic_clear_cpumask(&dfly_rdyprocmask, CPUMASK(cpuid));
+		wakeup(&dfly_pcpu[cpuid].helper_thread);
+	}
+}
+
+#endif
+
 /*
  * This routine is called from a systimer IPI.  It MUST be MP-safe and
  * the BGL IS NOT HELD ON ENTRY.  This routine is called at ESTCPUFREQ on
@@ -1032,6 +1096,7 @@ dfly_forking(struct lwp *plp, struct lwp *lp)
 	 * (less desireable than the parent).
 	 */
 	lp->lwp_estcpu = ESTCPULIM(plp->lwp_estcpu + ESTCPUPPQ * 4);
+	lp->lwp_forked = 1;
 
 	/*
 	 * The batch status of children always starts out centerline
@@ -1442,6 +1507,7 @@ dfly_choose_queue_simple(dfly_pcpu_t dd, struct lwp *lp)
 	 * Fallback to the original heuristic, select random cpu,
 	 * first checking cpus not currently running a user thread.
 	 */
+	++dfly_scancpu;
 	cpuid = (dfly_scancpu & 0xFFFF) % ncpus;
 	mask = ~dfly_curprocmask & dfly_rdyprocmask & lp->lwp_cpumask &
 	       smp_active_mask & usched_global_cpumask;
diff --git a/sys/sys/usched.h b/sys/sys/usched.h
index c0649c0a33..3b445f1100 100644
--- a/sys/sys/usched.h
+++ b/sys/sys/usched.h
@@ -61,7 +61,7 @@ union usched_data {
     } bsd4;
     struct {
 	short	priority;	/* lower is better */
-	char	unused01;	/* (currently not used) */
+	char	forked;		/* lock cpu during fork */
 	char	rqindex;
 	int	batch;		/* batch mode heuristic */
 	int	estcpu;		/* dynamic priority modification */