Keep track of kernel runq length, only try to steal when != 0.

This improves forexecbench performance on josmp by about 18%. Now steal_nonexec = 0 seems to have no affect or maybe a small negative affect.

Keep track of kernel runq length, only try to steal when != 0.
1f41a725 · Silas Boyd-Wickizer · 4c16796a · 1f41a725
--- a/kernel/sched.cc
+++ b/kernel/sched.cc
@@ -19,6 +19,7 @@ enum { steal_nonexec = 1 };
 struct runq {
  STAILQ_HEAD(queue, proc) q;
  struct spinlock lock;
+  volatile u64 len __mpalign__;
  __padout__;
 };

@@ -119,6 +120,7 @@ addrun(struct proc *p)
  acquire(&q->lock);
  STAILQ_INSERT_HEAD(&q->q, p, runqlink);
  p->runq = q;
+  q->len++;
  release(&q->lock);
 }

@@ -131,6 +133,7 @@ delrun(struct proc *p)
  q = p->runq;
  acquire(&q->lock);
  STAILQ_REMOVE(&q->q, p, proc, runqlink);
+  q->len--;
  release(&q->lock);
 }

@@ -147,6 +150,9 @@ steal(void)
      struct runq *q = &runq[(i+mycpu()->id) % ncpu];
      struct proc *p;

+      if (q->len == 0)
+        continue;
+
      // XXX(sbw) Look for a process to steal.  Acquiring q->lock
      // then p->lock can result in deadlock.  So we acquire
      // q->lock, scan for a process, drop q->lock, acquire p->lock,
@@ -160,6 +166,7 @@ steal(void)
            p->curcycles != 0 && p->curcycles > VICTIMAGE)
        {
          STAILQ_REMOVE(&q->q, p, proc, runqlink);
+          q->len--;
          steal = p;
          break;
        }
@@ -201,8 +208,10 @@ schednext(void)
  q = &runq[mycpu()->id];
  acquire(&q->lock);
  p = STAILQ_LAST(&q->q, proc, runqlink);
-  if (p)
+  if (p) {
    STAILQ_REMOVE(&q->q, p, proc, runqlink);
+    q->len--;
+  }
  release(&q->lock);
  popcli();
  return p;
@@ -216,6 +225,7 @@ initsched(void)
  for (i = 0; i < NCPU; i++) {
    initlock(&runq[i].lock, "runq", LOCKSTAT_SCHED);
    STAILQ_INIT(&runq[i].q);
+    runq[i].len = 0;
  }
 }