Fix some race conditions in xca6ll/hot/amcssth.

* Use memory-model-aware builtins in GCC and Clang when a memory
  location may be written by one thread and read by another, avoiding
  race conditions due to out-of-order updates on ARM.
* Call dylan_make_wrappers while the test is still single-threaded,
  preventing multiple threads from racing to call it.
* Prevent dylan_init from creating a padding object, as we must not
  have an exact root pointing at a padding object.

Author: gareth-rees

code/amcssth.c

@@ -108,19 +108,22 @@ static mps_res_t area_scan(mps_ss_t ss, void *base, void *limit, void *closure)
 
 static void churn(mps_ap_t ap, size_t roots_count)
 {
-  size_t i;
-  size_t r;
+  size_t i, j, r;
 
   ++objs;
   r = (size_t)rnd();
   if (r & 1) {
+    mps_addr_t root;
     i = (r >> 1) % exactRootsCOUNT;
-    if (exactRoots[i] != objNULL)
-      cdie(dylan_check(exactRoots[i]), "dying root check");
-    exactRoots[i] = make(ap, roots_count);
-    if (exactRoots[(exactRootsCOUNT-1) - i] != objNULL)
-      dylan_write(exactRoots[(exactRootsCOUNT-1) - i],
-                  exactRoots, exactRootsCOUNT);
+    atomic_load(&exactRoots[i], &root);
+    if (root != objNULL)
+      cdie(dylan_check(root), "dying root check");
+    root = make(ap, roots_count);
+    atomic_store(&exactRoots[i], &root);
+    j = exactRootsCOUNT - i - 1;
+    atomic_load(&exactRoots[j], &root);
+    if (root != objNULL)
+      dylan_write(root, exactRoots, exactRootsCOUNT);
   } else {
     i = (r >> 1) % ambigRootsCOUNT;
     ambigRoots[(ambigRootsCOUNT-1) - i] = make(ap, roots_count);
@@ -221,9 +224,11 @@ static void test_pool(const char *name, mps_pool_t pool, size_t roots_count)
                (unsigned long)collections, objs,
                (unsigned long)mps_arena_committed(arena));
 
-        for (i = 0; i < exactRootsCOUNT; ++i)
-          cdie(exactRoots[i] == objNULL || dylan_check(exactRoots[i]),
-               "all roots check");
+        for (i = 0; i < exactRootsCOUNT; ++i) {
+          mps_addr_t root;
+          atomic_load(&exactRoots[i], &root);
+          cdie(root == objNULL || dylan_check(root), "all roots check");
+        }
 
         if (collections >= collectionsCOUNT / 2 && !walked)
         {
@@ -248,9 +253,12 @@ static void test_pool(const char *name, mps_pool_t pool, size_t roots_count)
             ramping = 0;
             /* kill half of the roots */
             for(i = 0; i < exactRootsCOUNT; i += 2) {
-              if (exactRoots[i] != objNULL) {
-                cdie(dylan_check(exactRoots[i]), "ramp kill check");
-                exactRoots[i] = objNULL;
+              mps_addr_t root;
+              atomic_load(&exactRoots[i], &root);
+              if (root != objNULL) {
+                cdie(dylan_check(root), "ramp kill check");
+                root = objNULL;
+                atomic_store(&exactRoots[i], &root);
               }
             }
           }
@@ -298,6 +306,8 @@ static void test_arena(void)
   mps_pool_t amc_pool, amcz_pool;
   void *marker = &marker;
 
+  die(dylan_make_wrappers(), "make wrappers");
+
   MPS_ARGS_BEGIN(args) {
     MPS_ARGS_ADD(args, MPS_KEY_ARENA_SIZE, testArenaSIZE);
     MPS_ARGS_ADD(args, MPS_KEY_ARENA_GRAIN_SIZE, rnd_grain(testArenaSIZE));

code/fmtdytst.c

@@ -68,7 +68,7 @@ mps_res_t dylan_make_wrappers(void)
   return MPS_RES_OK;
 }
 
-/* dylan_init -- turn raw memory into initialised dylan-vector (or pad)
+/* dylan_init -- turn raw memory into initialised dylan-vector
  *
  * If the raw memory is large enough, initialises it to a dylan-vector,
  * whose slots are initialised to either dylan-ints, or valid refs, at
@@ -78,8 +78,6 @@ mps_res_t dylan_make_wrappers(void)
  * and "nr_refs" arguments. If "nr_refs" is 0, all slots are
  * initialized to dylan-ints: this may be useful for making leaf
  * objects.
- *
- * (Makes a pad if the raw memory is too small to hold a dylan-vector)
  */
 
 mps_res_t dylan_init(mps_addr_t addr, size_t size,
@@ -93,8 +91,7 @@ mps_res_t dylan_init(mps_addr_t addr, size_t size,
   if (res != MPS_RES_OK)
     return res;
 
-  /* If there is enough room, make a vector, otherwise just */
-  /* make a padding object. */
+  /* If there is enough room, make a vector. */
   if(size >= sizeof(mps_word_t) * 2) {
     mps_word_t *p = (mps_word_t *)addr;
     mps_word_t i, t = (size / sizeof(mps_word_t)) - 2;
@@ -110,7 +107,7 @@ mps_res_t dylan_init(mps_addr_t addr, size_t size,
         p[2+i] = (mps_word_t)refs[(r >> 1) % nr_refs]; /* random ptr */
     }
   } else {
-    dylan_pad(addr, size);
+    return MPS_RES_UNIMPL;
   }
 
   return MPS_RES_OK;

code/testlib.h

@@ -292,6 +292,25 @@ extern void randomize(int argc, char *argv[]);
 extern void testlib_init(int argc, char *argv[]);
 
 
+/* Memory-model-aware operations */
+
+#if defined(MPS_BUILD_GC) || defined(MPS_BUILD_LL)
+
+/* See <https://gcc.gnu.org/onlinedocs/gcc/_005f_005fatomic-Builtins.html>
+ * and <https://clang.llvm.org/docs/LanguageExtensions.html> */
+#define atomic_load(SRC, DEST) __atomic_load(SRC, DEST, __ATOMIC_ACQUIRE)
+#define atomic_store(DEST, SRC) __atomic_store(DEST, SRC, __ATOMIC_RELEASE)
+
+#elif defined(MPS_BUILD_MV)
+
+/* Microsoft Visual C/C++ does not need memory-model-aware load and store as
+ * loads and stores of register-sized values are atomic on Intel. */
+#define atomic_load(SRC, DEST) (*(DEST) = *(SRC))
+#define atomic_store(DEST, SRC) (*(DEST) = *(SRC))
+
+#endif
+
+
 #endif /* testlib_h */