A simple/stupid scheduler implementation.

The scheduler using namespaces is broken. For a given CPU, the scheduler always runs the process that was last inserted into the namespace. This leads to starvation and deadlock (in userspace). I want to test the VM implementation, so I wrote a simple per-core runqueue implementation.

A simple/stupid scheduler implementation.
a0620682 · Silas Boyd-Wickizer · 0c791553 · a0620682 · a0620682 · a0620682
--- a/kernel.h
+++ b/kernel.h
@@ -17,6 +17,7 @@ static inline void *p2v(uptr a) { return (void *) a + KBASE; }
 #define cmpswap(ptr, old, new) __sync_bool_compare_and_swap(ptr, old, new)
 #define subfetch(ptr, val)     __sync_sub_and_fetch(ptr, val)
+#define __offsetof             offsetof
 struct spinlock;
 struct condvar;

--- a/proc.h
+++ b/proc.h
@@ -58,6 +58,8 @@ struct proc {
 #if MTRACE
  struct mtrace_stacks mtrace_stacks;
 #endif
+  struct runq *runq;
+  STAILQ_ENTRY(proc) runqlink;
 };
 extern struct ns *nspid;
--- a/sched.c
+++ b/sched.c
@@ -12,9 +12,128 @@
 #include "kmtrace.h"
 #include "vm.h"
 #include "sched.h"
+#include <stddef.h>
 enum { sched_debug = 0 };
+struct runq {
+  STAILQ_HEAD(queue, proc) q;
+  struct spinlock lock;
+  __padout__;
+};
+static struct runq runq[NCPU] __mpalign__;
+void
+addrun(struct proc *p)
+{
+  // Always called with p->lock held
+  struct runq *q;
+  q = &runq[cpunum()];
+  acquire(&q->lock);
+  STAILQ_INSERT_HEAD(&q->q, p, runqlink);
+  p->runq = q;
+  release(&q->lock);
+}
+void
+delrun(struct proc *p)
+{
+  // Always called with p->lock held
+  struct runq *q;
+  q = p->runq;
+  acquire(&q->lock);
+  STAILQ_REMOVE(&q->q, p, proc, runqlink);
+  release(&q->lock);
+}
+int
+steal(void)
+{
+  struct proc *steal;
+  int i;
+  int r = 0;
+  pushcli();
+  for (i = 1; i < NCPU; i++) {
+    struct runq *q = &runq[(i+cpunum()) % NCPU];
+    struct proc *p;
+    // 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,
+    // and then check that it's still ok to steal p.
+    steal = NULL;
+    acquire(&q->lock);
+    STAILQ_FOREACH(p, &q->q, runqlink) {
+      if (p->state == RUNNABLE && !p->cpu_pin && 
+          p->curcycles != 0 && p->curcycles > MINCYCTHRESH)
+      {
+        steal = p;
+        break;
+      }
+    }
+    release(&q->lock);
+    if (steal) {
+      acquire(&steal->lock);
+      if (steal->state == RUNNABLE && !steal->cpu_pin &&
+          steal->curcycles != 0 && steal->curcycles > MINCYCTHRESH)
+      {
+        delrun(steal);
+        steal->curcycles = 0;
+        steal->cpuid = mycpu()->id;
+        addrun(steal);
+        release(&steal->lock);
+        r = 1;
+        break;
+      }
+      release(&steal->lock);
+    }
+  }
+  popcli();
+  return r;
+}
+struct proc *
+schednext(void)
+{
+  // No locks, interrupts enabled
+  struct runq *q;
+  struct proc *p = NULL;
+  pushcli();
+  q = &runq[cpunum()];
+  acquire(&q->lock);
+  p = STAILQ_LAST(&q->q, proc, runqlink);
+  if (p) {
+    STAILQ_REMOVE(&q->q, p, proc, runqlink);
+    STAILQ_INSERT_HEAD(&q->q, p, runqlink);    
+  }
+  release(&q->lock);
+  popcli();
+  return p;
+}
+void
+initsched(void)
+{
+  int i;
+  for (i = 0; i < NCPU; i++) {
+    initlock(&runq[i].lock, "runq");
+    STAILQ_INIT(&runq[i].q);
+  }
+}
+#if 0
+// XXX(sbw) This code is broken.  It essentially implements "LIFO"
+// scheduling, which leads to starvation and deadlock (in userspace).
 struct ns *nsrunq __mpalign__;
 // Mark a process RUNNABLE and add it to the runq
@@ -100,6 +219,9 @@ choose_runnable(void *pp, void *arg)
 struct proc *
 schednext(void)
 {
+  // XXX(sbw) Yikes!  If two processes are runnable on the same
+  // CPU this will always return to process addrun inserted last 
+  // into the ns.
  return ns_enumerate_key(nsrunq, KI(mycpu()->id), choose_runnable, 0);
 }
@@ -110,3 +232,4 @@ initsched(void)
  if (nsrunq == 0)
    panic("pinit runq");
 }
+#endif
--- a/thrtest.c
+++ b/thrtest.c
@@ -7,7 +7,7 @@
 static struct uspinlock l;
 static volatile int tcount;
-enum { nthread = 4 };
+enum { nthread = 8 };
 void
 thr(void *arg)
@@ -26,7 +26,7 @@ main(void)
  printf(1, "thrtest[%d]: start esp %x\n", getpid(), rrsp());
  for(int i = 0; i < nthread; i++) {
-    sbrk(4096);
+    sbrk(8192);
    void *tstack = sbrk(0);
    int tid = forkt(tstack, thr, (void*)(u64)(0xc0ffee00|i));
    printf(1, "thrtest[%d]: child %d\n", getpid(), tid);