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

Conflicts: bin/Makefrag

Merge branch 'scale-amd64' of git+ssh://pdos.csail.mit.edu/home/am0/6.828/xv6 into scale-amd64
cb928e7a · Robert Morris · 2ab0ec03 · 93f8127a · cb928e7a · cb928e7a
--- a/bin/Makefrag
+++ b/bin/Makefrag
@@ -24,7 +24,8 @@ UPROGS= \
 	scripttest \
 	ftest \
 	wqsh \
-	perf
+	perf \
+	xls
 # pdu
 # pls

--- a/bin/xls.cc
+++ b/bin/xls.cc
@@ -21,6 +21,7 @@
 #include "fs.h"
 #include "lib.h"
 #include "dirit.hh"
+#include "wq.hh"
 #define ST_SIZE(st)  (st).size
 #define ST_TYPE(st)  (st).type
 #define ST_INO(st)   (st).ino

--- a/include/uspinlock.h
+++ b/include/uspinlock.h
@@ -17,3 +17,9 @@ release(struct uspinlock *lk)
 {
  xchg32(&lk->locked, 0);
 }
+static int inline
+tryacquire(struct uspinlock *lk)
+{
+  return xchg32(&lk->locked, 1) == 0;
+}
--- a/include/wq.hh
+++ b/include/wq.hh
@@ -38,7 +38,8 @@ struct cwork : public work {
 #define xmalloc(n)   kmalloc(n) 
 #define xfree(p, sz) kmfree(p, sz)
 #else
-#warning "Unknown wq implementation"
+#define xmalloc(n)     malloc(n)
+#define xfree(p, sz)   free(p)
 #endif
 #include "wqfor.hh"
--- a/include/wqfor.hh
+++ b/include/wqfor.hh
+struct forframe
+{
+  forframe(int v) : v_(v) {}
+  int inc() { return __sync_add_and_fetch(&v_, 1); }
+  int dec() { return __sync_sub_and_fetch(&v_, 1); }
+  bool zero() volatile { return v_ == 0; };
+  volatile int v_;
+};
 template <typename IT, typename BODY>
 struct forwork : public work {
-  typedef std::atomic<int> forframe;
  forwork(IT &it, bool (*cond)(IT &it), BODY &body, forframe &frame) 
    : it_(it), cond_(cond), body_(body), frame_(frame) {}
@@ -10,11 +17,11 @@ struct forwork : public work {
    ++it_;
    if (cond_(it_)) {
      forwork<IT, BODY> *w = new forwork<IT, BODY>(it_, cond_, body_, frame_);
-      ++frame_;
+      frame_.inc();
      wq_push(w);    
    }
    body_(v);
-    --frame_;
+    frame_.dec();
    delete this;
  }
@@ -37,7 +44,7 @@ template <typename IT, typename BODY>
 static inline void
 wq_for(IT &init, bool (*cond)(IT &it), BODY body)
 {
-  typename forwork<IT, BODY>::forframe frame(0);
+  forframe frame(0);
  // XXX(sbw) should be able to coarsen loop
@@ -45,11 +52,11 @@ wq_for(IT &init, bool (*cond)(IT &it), BODY body)
  ++init;
  if (cond(init)) {
    forwork<IT, BODY> *w = new forwork<IT, BODY>(init, cond, body, frame);
-    ++frame;
+    frame.inc();
    wq_push(w);
  }
  body(v);
-  while (frame != 0)
+  while (!frame.zero())
    wq_trywork();
 }
--- a/include/wquser.hh
+++ b/include/wquser.hh
-// XXX(sbw)
+#include "types.h"
+#include "uspinlock.h"
+#include "amd64.h"
+#include "user.h"
+#include "wq.hh"
+#include "pthread.h"
 typedef struct uspinlock wqlock_t;
+static pthread_key_t idkey;
+int
+mycpuid(void)
+{
+  return (int)(u64)pthread_getspecific(idkey);
+}
+static inline void*
+allocwq(void)
+{
+  return malloc(WQSIZE);
+}
+static inline void
+wqlock_acquire(wqlock_t *lock)
+{
+  acquire(lock);
+}
+static inline int
+wqlock_tryacquire(wqlock_t *lock)
+{
+  return tryacquire(lock);
+}
+static inline void
+wqlock_release(wqlock_t *lock)
+{
+  release(lock);
+}
+static inline void
+wqlock_init(wqlock_t *lock) 
+{
+  lock->locked = 0;
+}
+static void
+setaffinity(int c)
+{
+  // XXX(sbw)
+}
+static void*
+workerth(void *x)
+{
+  u64 c = (u64)x;
+  setaffinity(c);
+  pthread_setspecific(idkey, (void*)c);
+  while (1)
+    wq_trywork();
+  return 0;
+}
+static inline void
+wqarch_init(void)
+{
+  pthread_t th;
+  int r;
+  if (pthread_key_create(&idkey, 0))
+    die("wqarch_init: pthread_key_create");
+  pthread_setspecific(idkey, 0);
+  setaffinity(0);
+  for (int i = 1; i < NCPU; i++) {
+    r = pthread_create(&th, 0, workerth, (void*)(u64)i);
+    if (r < 0)
+      die("wqarch_init: pthread_create");
+  }
+}
+#define xprintf      printf 
+#define pushcli()
+#define popcli()
--- a/lib/Makefrag
+++ b/lib/Makefrag
@@ -2,7 +2,7 @@ $(O)/lib/%.o: CFLAGS:=$(CFLAGS)
 $(O)/lib/%.o: CXXFLAGS:=$(CXXFLAGS)
 ULIB = ulib.o usys.o printf.o umalloc.o uthread.o fmt.o stream.o ipc.o \
-       threads.o crt.o
+       threads.o crt.o wq.o
 ULIB := $(addprefix $(O)/lib/, $(ULIB))
 .PRECIOUS: $(O)/lib/%.o

--- a/lib/crt.S
+++ b/lib/crt.S
@@ -2,6 +2,7 @@
 .align 8
 .globl _start
 _start:
+        call usetup
        pop %rdi
        mov %rsp, %rsi
        call main

--- a/lib/ulib.c
+++ b/lib/ulib.c
@@ -150,3 +150,15 @@ open(const char *path, int omode)
 {
  return openat(AT_FDCWD, path, omode);
 }
+extern void __cxa_pure_virtual(void);
+void __cxa_pure_virtual(void)
+{ 
+  die("__cxa_pure_virtual");
+}
+void
+usetup(void)
+{
+  forkt_setup(getpid());
+}
--- a/lib/wq.cc
+++ b/lib/wq.cc
 #if defined(LINUX)
 #include "user/wqlinux.hh"
-#include "percpu.hh"
 #elif defined(XV6_KERNEL)
 #include "wqkernel.hh"
-#include "percpu.hh"
 #else
-#warning "Unknown wq implementation"
+#include "wquser.hh"
 #endif
+#include "percpu.hh"
 #define NSLOTS (1 << WQSHIFT)

--- a/user/wqlinux.hh
+++ b/user/wqlinux.hh
@@ -84,7 +84,5 @@ wqarch_init(void)
 }
 #define xprintf        printf
-#define xmalloc(n)     malloc(n)
-#define xfree(p, sz)   free(p)
 #define pushcli()
 #define popcli()