First pass at Nickolai's exec hack.

I think i got it right..if it's buggy you can disable with EXECSWITCH 0 in param.h.

First pass at Nickolai's exec hack.
6ca8f6fd · Silas Boyd-Wickizer · a6b4886d · 6ca8f6fd · 6ca8f6fd · 6ca8f6fd
--- a/include/kernel.hh
+++ b/include/kernel.hh
@@ -175,6 +175,7 @@ int             pipewrite(struct pipe*, const char*, int);
 // proc.c
 struct proc*    copyproc(struct proc*);
 void            finishproc(struct proc*, bool removepid = true);
+void            execswitch(proc* p);
 void            exit(void);
 int             fork(int);
 int             growproc(int);
@@ -219,6 +220,10 @@ int             fetchmem(void*, const void*, u64);
 int             putmem(void*, const void*, u64);
 u64             syscall(u64 a0, u64 a1, u64 a2, u64 a3, u64 a4, u64 a5, u64 num);
+// sysfile.cc
+int             doexec(const char* upath,
+                       userptr<userptr<const char> > uargv);
 // string.c
 extern  "C" int  memcmp(const void*, const void*, size_t);
 void*           memmove(void*, const void*, size_t);

--- a/include/proc.hh
+++ b/include/proc.hh
@@ -84,6 +84,9 @@ struct proc : public rcu_freed, public sched_link {
  int run_cpuid_;
  int in_exec_;
  int uaccess_;
+  const char *upath;
+  userptr<userptr<const char> > uargv;
  u8 __cxa_eh_global[16];
  std::atomic<int> exception_inuse;

--- a/kernel/exec.cc
+++ b/kernel/exec.cc
@@ -157,17 +157,6 @@ exec(const char *path, char **argv, void *ascopev)
  cwork* w;
  long sp;
-  myproc()->exec_cpuid_ = mycpuid();
-  mt_ascope *ascope = (mt_ascope*) ascopev;
-  ascope->close();
-  myproc()->in_exec_ = 1;
-  yield();
-  myproc()->in_exec_ = 0;
-  ascope->open("sys_exec2(%s)", path);
-  myproc()->run_cpuid_ = mycpuid();
  if((ip = namei(myproc()->cwd, path)) == 0)
    return -1;

--- a/kernel/proc.cc
+++ b/kernel/proc.cc
@@ -42,6 +42,7 @@ proc::proc(int npid) :
  ftable(0), cwd(0), tsc(0), curcycles(0), cpuid(0), epoch(0),
  cpu_pin(0), oncv(0), cv_wakeup(0),
  user_fs_(0), unmap_tlbreq_(0), in_exec_(0), uaccess_(0),
+  upath(0), uargv(userptr<const char>(nullptr)),
  exception_inuse(0), magic(PROC_MAGIC), state_(EMBRYO)
 {
  snprintf(lockname, sizeof(lockname), "cv:proc:%d", pid);
@@ -204,6 +205,85 @@ freeproc(struct proc *p)
  gc_delayed(p);
 }
+void
+execstub(void)
+{
+  userptr<userptr<const char> > uargv;
+  const char* upath;
+  upath = myproc()->upath;
+  uargv = myproc()->uargv;
+  barrier();
+  myproc()->upath = nullptr;
+  post_swtch();
+  myproc()->run_cpuid_ = mycpuid();
+  long r = doexec(upath, uargv);
+  myproc()->tf->rax = r;
+  // This stuff would have been called in syscall and syscall_c
+  // if we returned from the the previous kstack
+  mtstop(myproc());
+  mtign();
+  if (myproc()->killed) {
+    mtstart(trap, myproc());
+    exit();
+  } 
+}
+static void
+kstackfree(void* kstack)
+{
+  ksfree(slab_stack, kstack);
+}
+void
+execswitch(proc* p)
+{
+  // Alloc a new kernel stack, set it up, and free the old one
+  context* cntxt;
+  trapframe* tf;
+  char* kstack;
+  char* sp;
+  if ((kstack = (char*) ksalloc(slab_stack)) == 0)
+    panic("execswitch: ksalloc");
+  sp = kstack + KSTACKSIZE;
+  sp -= sizeof(*p->tf);
+  tf = (trapframe*)sp;
+  // XXX(sbw) we only need the whole tf if exec fails
+  *tf = *p->tf;
+  sp -= 8;
+  // XXX(sbw) we could use the sysret return path
+  *(u64*)sp = (u64)trapret;
+  sp -= sizeof(*p->context);
+  cntxt = (context*)sp;
+  memset(cntxt, 0, sizeof(*cntxt));
+  cntxt->rip = (uptr)execstub;
+  cwork* w = new cwork();
+  if (w != nullptr) {
+    w->rip = (void*) kstackfree;
+    w->arg0 = p->kstack;
+    if (wqcrit_push(w, myproc()->exec_cpuid_) < 0) {
+      ksfree(slab_stack, p->kstack);
+      delete w;
+    }
+  } else {
+    ksfree(slab_stack, p->kstack);
+  }
+  p->kstack = kstack;
+  p->context = cntxt;
+  p->tf = tf;
+}
 proc*
 proc::alloc(void)
 {

--- a/kernel/sched.cc
+++ b/kernel/sched.cc
@@ -227,7 +227,9 @@ scheddump(void)
 void
 addrun(struct proc* p)
 {
-  ANON_REGION(__func__, &perfgroup);
+  if (p->upath)
+    execswitch(p);
  p->set_state(RUNNABLE);
  schedule_[p->cpuid].enq(p);
 }

--- a/kernel/sysfile.cc
+++ b/kernel/sysfile.cc
@@ -430,9 +430,8 @@ sys_chdir(const char *path)
  return 0;
 }
-//SYSCALL
 int
-sys_exec(const char *upath, userptr<userptr<const char> > uargv)
+doexec(const char* upath, userptr<userptr<const char> > uargv)
 {
  ANON_REGION(__func__, &perfgroup);
  char *argv[MAXARG];
@@ -459,16 +458,33 @@ sys_exec(const char *upath, userptr<userptr<const char> > uargv)
    if (argv[i]==nullptr || fetchstr(argv[i], (char*)uarg, MAXARGLEN)<0)
      goto clean;
  }
  argv[i] = 0;
  r = exec(path, argv, &ascope);
 clean:
-  for (i=i-i; i >= 0; i--)
+  for (i=i-1; i >= 0; i--)
    kmfree(argv[i], MAXARGLEN);
  return r;
 }
 //SYSCALL
 int
+sys_exec(const char *upath, userptr<userptr<const char> > uargv)
+{
+#if EXECSWITCH
+  myproc()->exec_cpuid_ = mycpuid();
+  myproc()->uargv = uargv;
+  barrier();
+  // upath serves as a flag to the scheduler
+  myproc()->upath = upath;
+  yield();
+  myproc()->upath = nullptr;
+#endif
+  return doexec(upath, uargv);
+}
+//SYSCALL
+int
 sys_pipe(int *fd)
 {
  struct file *rf, *wf;

--- a/param.h
+++ b/param.h
@@ -24,7 +24,7 @@
 #define USERWQSIZE    (1 << 14)
 #define USTACKPAGES   8
 #define WQSHIFT       7
-#define CILKENABLE    0
+#define EXECSWITCH    1
 #if defined(HW_qemu)
 #define NCPU          8   // maximum number of CPUs
 #define MTRACE        0