Merge branch 'scale-amd64' of git+ssh://pdos.csail.mit.edu/home/am0/6.828/xv6 into scale-amd64

c3171cdc · Nickolai Zeldovich · 6aa59406 · a995b487 · c3171cdc · c3171cdc
--- a/README
+++ b/README
@@ -56,4 +56,3 @@

  $ apt-get install libjemalloc-dev
  $ make HW=user o.user/utest
-
--- a/bin/Makefrag
+++ b/bin/Makefrag
@@ -33,7 +33,7 @@ UPROGS := $(addprefix $(O)/bin/, $(UPROGS))
 $(O)/bin/%.unstripped: $(O)/bin/%.o $(ULIB)
 	@echo "  LD     $@"
 	$(Q)mkdir -p $(@D)
-	$(Q)$(LD) $(LDFLAGS) -N -e main -Ttext 0x100000 -o $@ $^
+	$(Q)$(LD) $(LDFLAGS) -N -Ttext 0x100000 -o $@ $^

 $(O)/bin/%: $(O)/bin/%.unstripped
 	@echo "  STRIP  $@"

--- a/include/cilk.hh
+++ b/include/cilk.hh
+#if CILKENABLE
+
+template<typename A1>
+static void
+cilk_call(void (*fn)(A1), A1 a1)
+{
+  cilk_push((void(*)(uptr, uptr))fn, (uptr)a1, (uptr)0);
+}
+
+template<typename A1, typename A2>
+static void
+cilk_call(void (*fn)(A1, A2), A1 a1, A2 a2)
+{
+  cilk_push((void(*)(uptr, uptr))fn, (uptr)a1, (uptr)a2);
+}
+
+#else // !CILKENABLE
+
+template<typename A1>
+static void
+cilk_call(void (*fn)(A1), A1 a1)
+{
+  fn(a1);
+}
+
+template<typename A1, typename A2>
+static void
+cilk_call(void (*fn)(A1, A2), A1 a1, A2 a2)
+{
+  fn(a1, a2);
+}
+#endif
--- a/include/kernel.hh
+++ b/include/kernel.hh
@@ -194,11 +194,6 @@ void            yield(void);
 struct proc*    threadalloc(void (*fn)(void*), void *arg);
 struct proc*    threadpin(void (*fn)(void*), void *arg, const char *name, int cpu);

-// prof.c
-extern int profenable;
-void            profreset(void);
-void            profdump(void);
-
 // sampler.c
 void            sampstart(void);
 int             sampintr(struct trapframe*);

--- a/include/prof.hh
+++ b/include/prof.hh
-struct profrec {
-  u64 tot;
-  u64 cnt;
-  __padout__;
-} __mpalign__;
-
-typedef struct profctr {
-  const char *name;
-  struct profrec rec[NCPU] __mpalign__;
-  __padout__;
-} profctr_t;
-
-#if 0   /* not for c++ */
-#define DEFINE_PROFCTR(xname) \
-  profctr_t xname __attribute__((section(".prof"))) = { .name = #xname };
-
-#define prof_start(name) \
-  u64 __prof##name = profenable ? rdtsc() : 0;
-
-#define prof_end(name) do { \
-  if (profenable) { \
-    u64 __eprof##name = rdtsc(); \
-    u64 __profid = mycpu()->id; \
-    name.rec[__profid].tot += __eprof##name - __prof##name; \
-    name.rec[__profid].cnt++; \
-  } \
-} while (0)
-#else
-#define DEFINE_PROFCTR(x)
-#define prof_start(x)
-#define prof_end(x)
-#endif
--- a/kernel/Makefrag
+++ b/kernel/Makefrag
@@ -29,7 +29,6 @@ OBJS = \
 	picirq.o \
 	pipe.o \
 	proc.o \
-	prof.o \
 	gc.o \
 	rnd.o \
 	sampler.o \

--- a/kernel/console.cc
+++ b/kernel/console.cc
@@ -270,18 +270,6 @@ consoleintr(int (*getc)(void))
    case C('W'):
      wq_dump();
      break;
-    case C('L'):  // Prof stats
-      profdump();
-      break;
-    case C('K'):  // Prof enable
-      profreset();
-      cprintf("prof enabled\n");
-      profenable = 1;
-      break;
-    case C('I'):  // Prof disable
-      profenable = 0;
-      cprintf("prof disabled\n");
-      break;
    case C('F'):  // kmem stats
      kmemprint();
      break;

--- a/kernel/exec.cc
+++ b/kernel/exec.cc
@@ -12,13 +12,12 @@
 #include "vm.hh"
 #include "elf.hh"
 #include "cpu.hh"
-#include "prof.hh"
+#include "wq.hh"
+#include "cilk.hh"

 #define USTACKPAGES 2
 #define BRK (USERTOP >> 1)

-static const int odp = 1;
-
 struct eargs {
  struct proc *proc;
  struct inode *ip;
@@ -28,10 +27,8 @@ struct eargs {
 };

 static void
-dosegment(uptr a0, u64 a1)
+dosegment(struct eargs *args, u64 off)
 {
-  struct eargs *args = (eargs*) a0;
-  u64 off = a1;
  struct vmnode *vmn = nullptr;
  struct proghdr ph;
  uptr va_start, va_end;
@@ -39,7 +36,6 @@ dosegment(uptr a0, u64 a1)
  uptr in_sz;
  int npg;

-  prof_start(dosegment_prof);
  if(readi(args->ip, (char*)&ph, off, sizeof(ph)) != sizeof(ph))
    goto bad;
  if(ph.type != ELF_PROG_LOAD)
@@ -55,56 +51,63 @@ dosegment(uptr a0, u64 a1)
  in_sz = ph.filesz + PGOFFSET(ph.vaddr);

  npg = (va_end - va_start) / PGSIZE;
-  if ((vmn = new vmnode(npg, odp ? ONDEMAND : EAGER,
+  if ((vmn = new vmnode(npg, ONDEMAND,
                        args->ip, in_off, in_sz)) == 0)
    goto bad;

  if(args->vmap->insert(vmn, va_start, 1) < 0)
    goto bad;

-  prof_end(dosegment_prof);
  return;
  
 bad:
  cilk_abort(-1);
 }

-static void  __attribute__((unused)) dostack(uptr a0, u64 a1)
+static void
+dostack(struct eargs *args)
 {
  struct vmnode *vmn = nullptr;
-  struct eargs *args = (eargs*) a0;  
-  int argc;
-  uptr sp;
-  uptr ustack[1+MAXARG+1];
+  uptr argstck[1+MAXARG];
  const char *s, *last;
+  s64 argc;
+  uptr sp;
+
+  // User stack should be:
+  //   char argv[argc-1]
+  //   char argv[argc-2]
+  //   ...
+  //   char argv[0]
+  //   char *argv[argc+1]
+  //   u64 argc

-  prof_start(dostack_prof);
-    // Allocate a one-page stack at the top of the (user) address space
+  // Allocate a one-page stack at the top of the (user) address space
  if((vmn = new vmnode(USTACKPAGES)) == 0)
    goto bad;
  if(args->vmap->insert(vmn, USERTOP-(USTACKPAGES*PGSIZE), 1) < 0)
    goto bad;

-  // Push argument strings, prepare rest of stack in ustack.
-  sp = USERTOP;
-  for(argc = 0; args->argv[argc]; argc++) {
+  for (argc = 0; args->argv[argc]; argc++)
    if(argc >= MAXARG)
      goto bad;
-    sp -= strlen(args->argv[argc]) + 1;
+
+  // Push argument strings
+  sp = USERTOP;
+  for(int i = argc-1; i >= 0; i--) {
+    sp -= strlen(args->argv[i]) + 1;
    sp &= ~7;
-    if(args->vmap->copyout(sp, args->argv[argc], strlen(args->argv[argc]) + 1) < 0)
+    if(args->vmap->copyout(sp, args->argv[i], strlen(args->argv[i]) + 1) < 0)
      goto bad;
-    ustack[1+argc] = sp;
+    argstck[i] = sp;
  }
-  ustack[1+argc] = 0;
+  argstck[argc] = 0;

-  //prof_start(exec3_prof);
-  ustack[0] = 0xffffffffffffffffull;  // fake return PC
-  args->proc->tf->rdi = argc;
-  args->proc->tf->rsi = sp - (argc+1)*8;
+  sp -= (argc+1) * 8;
+  if(args->vmap->copyout(sp, argstck, (argc+1)*8) < 0)
+    goto bad;

-  sp -= (1+argc+1) * 8;
-  if(args->vmap->copyout(sp, ustack, (1+argc+1)*8) < 0)
+  sp -= 8;
+  if(args->vmap->copyout(sp, &argc, 8) < 0)
    goto bad;

  // Save program name for debugging.
@@ -116,26 +119,23 @@ static void  __attribute__((unused)) dostack(uptr a0, u64 a1)
  safestrcpy(args->proc->name, last, sizeof(args->proc->name));
  args->proc->tf->rsp = sp;

-  prof_end(dostack_prof);
  return;

 bad:
  cilk_abort(-1);
 }

-static void __attribute__((unused)) doheap(uptr a0, u64 a1)
+static void 
+doheap(struct eargs *args)
 {
  struct vmnode *vmn = nullptr;
-  struct eargs *args = (eargs*) a0;

-  prof_start(doheap_prof);
  // Allocate a vmnode for the heap.
  // XXX pre-allocate 32 pages..
  if((vmn = new vmnode(32)) == 0)
    goto bad;
  if(args->vmap->insert(vmn, BRK, 1) < 0)
    goto bad;
-  prof_end(doheap_prof);

  return;

@@ -154,7 +154,6 @@ exec(const char *path, char **argv)
  int i;
  struct vmap *oldvmap;

-  prof_start(exec_prof);
  if((ip = namei(myproc()->cwd, path)) == 0)
    return -1;

@@ -188,23 +187,14 @@ exec(const char *path, char **argv)
      goto bad;
    if(type != ELF_PROG_LOAD)
      continue;
-    cilk_push(dosegment, (uptr)&args, (uptr)off);
-  }
-
-  if (odp) {
-    // iunlock(ip);
-  } else {
-    // iunlockput(ip);
-    iput(ip);
-    ip = 0;
+    cilk_call(dosegment, &args, off);
  }

-  cilk_push(doheap, (uptr)&args, (uptr)0);
+  cilk_call(doheap, &args);

-  // dostack reads from the user address space.  The wq
-  // stuff doesn't switch to the user address space.
-  //cilk_push(dostack, (uptr)&args, (uptr)0);
-  dostack((uptr)&args, (uptr)0);
+  // dostack reads from the user vm space.  wq workers don't switch 
+  // the user vm.
+  dostack(&args);

  if (cilk_end())
    goto bad;
@@ -219,7 +209,6 @@ exec(const char *path, char **argv)
  oldvmap->decref();

  gc_end_epoch();
-  prof_end(exec_prof);
  return 0;

 bad:

--- a/kernel/prof.cc
+++ b/kernel/prof.cc
-#include "types.h"
-#include "kernel.hh"
-#include "spinlock.h"
-#include "condvar.h"
-#include "fs.h"
-#include "file.hh"
-#include "prof.hh"
-#include "bits.hh"
-#include "amd64.h"
-
-extern profctr_t sprof[];
-extern profctr_t eprof[];
-int profenable;
-
-void
-profreset(void)
-{
-  profctr_t *p = sprof;
-  for (; p != eprof; p++) {
-    memset(p->rec, 0, sizeof(p->rec));
-  }
-}
-
-static void
-profsum(struct profctr *ctr, u64 *tot, u64 *cnt)
-{
-  for (int i = 0; i < NCPU; i++) {
-    *tot += ctr->rec[i].tot;
-    *cnt += ctr->rec[i].cnt;
-  }
-}
-
-void
-profdump(void)
-{
-  profctr_t *p = sprof;
-  for (; p != eprof; p++) {
-    u64 tot = 0, cnt = 0;
-    profsum(p, &tot, &cnt);
-    if (cnt)
-      cprintf("%s %lu\n", p->name, tot/cnt);
-  }
-
-  sampdump();
-}
--- a/lib/Makefrag
+++ b/lib/Makefrag
 ULIB = ulib.o usys.o printf.o umalloc.o uthread.o fmt.o stream.o ipc.o \
-       threads.o
+       threads.o crt.o
 ULIB := $(addprefix $(O)/lib/, $(ULIB))

 .PRECIOUS: $(O)/lib/%.o

--- a/lib/crt.S
+++ b/lib/crt.S
+.code64
+.align 8
+.globl _start
+_start:
+        pop %rdi
+        mov %rsp, %rsi
+        call main
+        call exit