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

4ce84865 · Silas Boyd-Wickizer · ef3b56a2 · 118bc008 · 4ce84865 · 4ce84865
--- a/bin/Makefrag
+++ b/bin/Makefrag
@@ -11,7 +11,6 @@ UPROGS= \
 	mapbench \
 	maptest \
 	sh \
-	thrtest \
 	halt \
 	time \
 	sleep \

--- a/bin/mapbench.cc
+++ b/bin/mapbench.cc
@@ -4,12 +4,13 @@
 #include "mtrace.h"
 #include "amd64.h"
 #include "uspinlock.h"
+#include "pthread.h"
 enum { readaccess = 0 };
 enum { verbose = 0 };
 enum { npg = 1 };
-void
+void*
 thr(void *arg)
 {
  u64 tid = (u64)arg;
@@ -34,6 +35,7 @@ thr(void *arg)
      exit();
    }
  }
+  return 0;
 }
 int
@@ -50,11 +52,9 @@ main(int ac, char **av)
  // fprintf(1, "mapbench[%d]: start esp %x, nthread=%d\n", getpid(), rrsp(), nthread);
-  for(int i = 0; i < nthread; i++) {
+  for(u64 i = 0; i < nthread; i++) {
-    sbrk(8192);
+    pthread_t tid;
-    void *tstack = sbrk(0);
+    pthread_create(&tid, 0, thr, (void*) i);
-    // fprintf(1, "tstack %lx\n", tstack);
-    int tid = forkt(tstack, (void*) thr, (void *)(u64)i);
    if (0) fprintf(1, "mapbench[%d]: child %d\n", getpid(), tid);
  }

--- a/bin/maptest.cc
+++ b/bin/maptest.cc
@@ -4,6 +4,7 @@
 #include "mtrace.h"
 #include "amd64.h"
 #include "uspinlock.h"
+#include "pthread.h"
 static volatile char *p;
 static struct uspinlock l;
@@ -17,8 +18,8 @@ spin(void)
    ;
 }
-void
+void*
-thr(void)
+thr(void*)
 {
  for (;;) {
    acquire(&l);
@@ -54,8 +55,8 @@ main(void)
    exit();
  }
-  sbrk(4096);
+  pthread_t tid;
-  forkt(sbrk(0), (void*) thr, 0);
+  pthread_create(&tid, 0, thr, 0);
  acquire(&l);
  state = 1;

--- a/bin/thrtest.cc
+++ b/bin/thrtest.cc
-#include "types.h"
-#include "stat.h"
-#include "user.h"
-#include "mtrace.h"
-#include "amd64.h"
-#include "uspinlock.h"
-static struct uspinlock l;
-static volatile int tcount;
-enum { nthread = 8 };
-void
-thr(void *arg)
-{
-  acquire(&l);
-  fprintf(1, "thrtest[%d]: arg 0x%lx rsp %lx\n", getpid(), arg, rrsp());
-  tcount++;
-  release(&l);
-  exit();
-}
-int
-main(void)
-{
-  acquire(&l);
-  fprintf(1, "thrtest[%d]: start esp %x\n", getpid(), rrsp());
-  for(int i = 0; i < nthread; i++) {
-    sbrk(8192);
-    void *tstack = sbrk(0);
-    int tid = forkt(tstack, (void*) thr, (void*)(u64)(0xc0ffee00|i));
-    fprintf(1, "thrtest[%d]: child %d\n", getpid(), tid);
-  }
-  for(;;){
-    int lastc = tcount;
-    fprintf(1, "thrtest[%d]: tcount=%d\n", getpid(), lastc);
-    release(&l);
-    if(lastc==nthread)
-      break;
-    while(tcount==lastc)
-      __asm __volatile("");
-    acquire(&l);
-  }
-  release(&l);
-  fprintf(1, "thrtest[%d]: done\n", getpid());
-  for(int i = 0; i < nthread; i++)
-    wait();
-  exit();
-}
--- a/bin/usertests.cc
+++ b/bin/usertests.cc
@@ -5,6 +5,7 @@
 #include "fcntl.h"
 #include "syscall.h"
 #include "traps.h"
+#include "pthread.h"
 char buf[2048];
 char name[3];
@@ -1600,6 +1601,7 @@ preads(void)
 void
 tls_test(void)
 {
+  printf("tls_test\n");
  u64 buf[128];
  for (int i = 0; i < sizeof(buf) / sizeof(buf[0]); i++)
@@ -1626,6 +1628,49 @@ tls_test(void)
  fprintf(1, "tls_test ok\n");
 }
+static pthread_key_t tkey;
+static pthread_barrier_t bar0, bar1;
+enum { nthread = 8 };
+static void*
+thr(void *arg)
+{
+  pthread_setspecific(tkey, arg);
+  pthread_barrier_wait(&bar0);
+  u64 x = (u64) arg;
+  if ((x >> 8) != 0xc0ffee)
+    fprintf(2, "thr: x 0x%lx\n", x);
+  if (arg != pthread_getspecific(tkey))
+    fprintf(2, "thr: arg %p getspec %p\n", arg, pthread_getspecific(tkey));
+  pthread_barrier_wait(&bar1);
+  return 0;
+}
+void
+thrtest(void)
+{
+  printf("thrtest\n");
+  pthread_key_create(&tkey, 0);
+  pthread_barrier_init(&bar0, 0, nthread);
+  pthread_barrier_init(&bar1, 0, nthread+1);
+  for(int i = 0; i < nthread; i++) {
+    pthread_t tid;
+    pthread_create(&tid, 0, &thr, (void*) (0xc0ffee00ULL | i));
+  }
+  pthread_barrier_wait(&bar1);
+  for(int i = 0; i < nthread; i++)
+    wait();
+  printf("thrtest ok\n");
+}
 int
 main(int argc, char *argv[])
 {
@@ -1674,6 +1719,7 @@ main(int argc, char *argv[])
  exectest();
  tls_test();
+  thrtest();
  exit();
 }
--- a/include/amd64.h
+++ b/include/amd64.h
@@ -133,6 +133,20 @@ ltr(u16 sel)
 }
 static inline void
+writefs(u16 v)
+{
+  __asm volatile("movw %0, %%fs" : : "r" (v));
+}
+static inline u16
+readfs(void)
+{
+  u16 v;
+  __asm volatile("movw %%fs, %0" : "=r" (v));
+  return v;
+}
+static inline void
 writegs(u16 v)
 {
  __asm volatile("movw %0, %%gs" : : "r" (v));

--- a/include/pthread.h
+++ b/include/pthread.h
+#pragma once
+/*
+ * Our minimal version of pthreads
+ */
+#include "atomic.hh"
+typedef int pthread_t;
+typedef int pthread_attr_t;
+typedef int pthread_key_t;
+typedef int pthread_barrierattr_t;
+typedef std::atomic<unsigned> pthread_barrier_t;
+int       pthread_create(pthread_t* tid, const pthread_attr_t* attr,
+                         void* (*start)(void*), void* arg);
+pthread_t pthread_self(void);
+int       pthread_key_create(pthread_key_t* key, void (*destructor)(void*));
+void*     pthread_getspecific(pthread_key_t key);
+int       pthread_setspecific(pthread_key_t key, void* value);
+int       pthread_barrier_init(pthread_barrier_t *b,
+                               const pthread_barrierattr_t *attr,
+                               unsigned count);
+int       pthread_barrier_wait(pthread_barrier_t *b);
--- a/include/user.h
+++ b/include/user.h
@@ -50,6 +50,7 @@ int atoi(const char*);
 // uthread.S
 int forkt(void *sp, void *pc, void *arg);
+void forkt_setup(u64 pid);
 // printf.c
 void printf(const char*, ...);

--- a/kernel/hwvm.cc
+++ b/kernel/hwvm.cc
@@ -127,6 +127,8 @@ switchvm(struct proc *p)
    lcr3(v2p(p->vmap->pml4));  // switch to new address space
  else
    switchkvm();
+  writefs(UDSEG);
  writemsr(MSR_FS_BASE, p->user_fs_);
  popcli();

--- a/lib/Makefrag
+++ b/lib/Makefrag
-ULIB = ulib.o usys.o printf.o umalloc.o uthread.o fmt.o stream.o ipc.o
+ULIB = ulib.o usys.o printf.o umalloc.o uthread.o fmt.o stream.o ipc.o \
+       threads.o
 ULIB := $(addprefix $(O)/lib/, $(ULIB))
 .PRECIOUS: $(O)/lib/%.o

--- a/lib/threads.cc
+++ b/lib/threads.cc
+#include "types.h"
+#include "pthread.h"
+#include "user.h"
+#include "atomic.hh"
+enum { stack_size = 8192 };
+static std::atomic<int> nextkey;
+struct tlsdata {
+  void* buf[128];
+};
+void
+forkt_setup(u64 pid)
+{
+  tlsdata* t = (tlsdata*) sbrk(sizeof(*t));
+  setfs((u64) t);
+}
+int
+pthread_create(pthread_t* tid, const pthread_attr_t* attr,
+               void* (*start)(void*), void* arg)
+{
+  char* base = (char*) sbrk(stack_size);
+  int t = forkt(base + stack_size, (void*) start, arg);
+  if (t < 0)
+    return t;
+  *tid = t;
+  return 0;
+}
+pthread_t
+pthread_self()
+{
+  return getpid();
+}
+int
+pthread_key_create(pthread_key_t *key, void (*destructor)(void*))
+{
+  // Ignore the destructor for now.
+  *key = nextkey++;
+  return 0;
+}
+void*
+pthread_getspecific(pthread_key_t key)
+{
+  u64 v;
+  __asm volatile("movq %%fs:(%1), %0" : "=r" (v) : "r" ((u64) key * 8));
+  return (void*) v;
+}
+int
+pthread_setspecific(pthread_key_t key, void* value)
+{
+  __asm volatile("movq %0, %%fs:(%1)" : : "r" (value), "r" ((u64) key * 8));
+  return 0;
+}
+int
+pthread_barrier_init(pthread_barrier_t *b,
+                     const pthread_barrierattr_t *attr, unsigned count)
+{
+  b->store(count);
+  return 0;
+}
+int
+pthread_barrier_wait(pthread_barrier_t *b)
+{
+  (*b)--;
+  while (*b != 0)
+    ;   // spin
+  return 0;
+}
--- a/lib/uthread.S
+++ b/lib/uthread.S
@@ -8,20 +8,23 @@ forkt:
        pushq %r12
        andq $(~0xf), %rdi      # new stack ptr amd64 ABI alignment 
        movq %rdi, %r12
-        movq %rdx, -16(%r12)    # arg
+        subq $0x10, %r12
-        movq %rsi, -8(%r12)     # function ptr
+        movq %rdx, 0x00(%r12)   # arg
+        movq %rsi, 0x08(%r12)   # function ptr
        movq $1, %rdi           # flag for sys_fork
        movq $SYS_fork, %rax
-        int $T_SYSCALL
+        syscall
        cmpq $0, %rax           # pid/tid
        jne 1f
        # child
-        movq -8(%r12), %rax
-        movq -16(%r12), %rdi
        movq %r12, %rsp
+        movq %rax, %rdi
+        call forkt_setup
+        popq %rdi
+        popq %rax
        call *%rax
        call exit
 1:      # parent