erts: Fix magic binary alignment on 32-bit

Caused bus error on 32-bit sparc from unaligned 64-bit word in
binary_to_term trap context.

Also add _UNALIGNED_ magic macros to avoid double alignment padding
in NIF resources.

4 files changed, 77 insertions, 20 deletions

diff --git a/erts/emulator/beam/erl_binary.h b/erts/emulator/beam/erl_binary.h
index 8d264d166e..6b96787d40 100644
--- a/erts/emulator/beam/erl_binary.h
+++ b/erts/emulator/beam/erl_binary.h
@@ -194,6 +194,9 @@ ERTS_GLB_INLINE Binary *erts_bin_nrml_alloc(Uint size);
 ERTS_GLB_INLINE Binary *erts_bin_realloc_fnf(Binary *bp, Uint size);
 ERTS_GLB_INLINE Binary *erts_bin_realloc(Binary *bp, Uint size);
 ERTS_GLB_INLINE void erts_bin_free(Binary *bp);
+ERTS_GLB_INLINE Binary *erts_create_magic_binary_x(Uint size,
+                                                  void (*destructor)(Binary *),
+                                                  int unaligned);
 ERTS_GLB_INLINE Binary *erts_create_magic_binary(Uint size,
 						 void (*destructor)(Binary *));
 
@@ -332,21 +335,30 @@ erts_bin_free(Binary *bp)
 }
 
 ERTS_GLB_INLINE Binary *
-erts_create_magic_binary(Uint size, void (*destructor)(Binary *))
+erts_create_magic_binary_x(Uint size, void (*destructor)(Binary *),
+                           int unaligned)
 {
-    Uint bsize = ERTS_MAGIC_BIN_SIZE(size);
+    Uint bsize = unaligned ? ERTS_MAGIC_BIN_UNALIGNED_SIZE(size)
+                           : ERTS_MAGIC_BIN_SIZE(size);
     Binary* bptr = erts_alloc_fnf(ERTS_ALC_T_BINARY, bsize);
     ASSERT(bsize > size);
     if (!bptr)
 	erts_alloc_n_enomem(ERTS_ALC_T2N(ERTS_ALC_T_BINARY), bsize);
     ERTS_CHK_BIN_ALIGNMENT(bptr);
     bptr->flags = BIN_FLAG_MAGIC;
-    bptr->orig_size = ERTS_MAGIC_BIN_ORIG_SIZE(size);
+    bptr->orig_size = unaligned ? ERTS_MAGIC_BIN_UNALIGNED_ORIG_SIZE(size)
+                                : ERTS_MAGIC_BIN_ORIG_SIZE(size);
     erts_refc_init(&bptr->refc, 0);
     ERTS_MAGIC_BIN_DESTRUCTOR(bptr) = destructor;
     return bptr;
 }
 
+ERTS_GLB_INLINE Binary *
+erts_create_magic_binary(Uint size, void (*destructor)(Binary *))
+{
+    return erts_create_magic_binary_x(size, destructor, 0);
+}
+
 #endif /* #if ERTS_GLB_INLINE_INCL_FUNC_DEF */
 
 #endif /* !__ERL_BINARY_H */
diff --git a/erts/emulator/beam/erl_bits.c b/erts/emulator/beam/erl_bits.c
index b8ae93fa58..2e29bf8895 100644
--- a/erts/emulator/beam/erl_bits.c
+++ b/erts/emulator/beam/erl_bits.c
@@ -107,6 +107,14 @@ erts_bits_destroy_state(ERL_BITS_PROTO_0)
 void
 erts_init_bits(void)
 {
+    ERTS_CT_ASSERT(offsetof(Binary,orig_bytes) % 8 == 0);
+    ERTS_CT_ASSERT(offsetof(ErtsMagicBinary,u.aligned.data) % 8 == 0);
+    ERTS_CT_ASSERT(ERTS_MAGIC_BIN_BYTES_TO_ALIGN ==
+                   (offsetof(ErtsMagicBinary,u.aligned.data)
+                    - offsetof(ErtsMagicBinary,u.unaligned.data)));
+    ERTS_CT_ASSERT(offsetof(ErtsBinary,driver.binary.orig_bytes)
+                == offsetof(Binary,orig_bytes));
+
     erts_smp_atomic_init_nob(&bits_bufs_size, 0);
 #if defined(ERTS_SMP)
     /* erl_process.c calls erts_bits_init_state() on all state instances */
diff --git a/erts/emulator/beam/erl_nif.c b/erts/emulator/beam/erl_nif.c
index 426a00304e..f42ccf23c2 100644
--- a/erts/emulator/beam/erl_nif.c
+++ b/erts/emulator/beam/erl_nif.c
@@ -1199,7 +1199,11 @@ typedef struct enif_resource_t
     struct enif_resource_type_t* type;
 #ifdef DEBUG
     erts_refc_t nif_refc;
+# ifdef ARCH_32
+    byte align__[4];
+# endif
 #endif
+
     char data[1];
 }ErlNifResource;
 
@@ -1375,7 +1379,7 @@ static void rollback_opened_resource_types(void)
 
 static void nif_resource_dtor(Binary* bin)
 {
-    ErlNifResource* resource = (ErlNifResource*) ERTS_MAGIC_BIN_DATA(bin);
+    ErlNifResource* resource = (ErlNifResource*) ERTS_MAGIC_BIN_UNALIGNED_DATA(bin);
     ErlNifResourceType* type = resource->type;
     ASSERT(ERTS_MAGIC_BIN_DESTRUCTOR(bin) == &nif_resource_dtor);
 
@@ -1396,8 +1400,10 @@ static void nif_resource_dtor(Binary* bin)
 
 void* enif_alloc_resource(ErlNifResourceType* type, size_t size)
 {
-    Binary* bin = erts_create_magic_binary(SIZEOF_ErlNifResource(size), &nif_resource_dtor);
-    ErlNifResource* resource = ERTS_MAGIC_BIN_DATA(bin);
+    Binary* bin = erts_create_magic_binary_x(SIZEOF_ErlNifResource(size),
+                                             &nif_resource_dtor,
+                                             1); /* unaligned */
+    ErlNifResource* resource = ERTS_MAGIC_BIN_UNALIGNED_DATA(bin);
 
     ASSERT(type->owner && type->next && type->prev); /* not allowed in load/upgrade */
     resource->type = type;
@@ -1412,7 +1418,7 @@ void* enif_alloc_resource(ErlNifResourceType* type, size_t size)
 void enif_release_resource(void* obj)
 {
     ErlNifResource* resource = DATA_TO_RESOURCE(obj);
-    ErtsBinary* bin = ERTS_MAGIC_BIN_FROM_DATA(resource);
+    ErtsBinary* bin = ERTS_MAGIC_BIN_FROM_UNALIGNED_DATA(resource);
 
     ASSERT(ERTS_MAGIC_BIN_DESTRUCTOR(bin) == &nif_resource_dtor);
 #ifdef DEBUG
@@ -1426,7 +1432,7 @@ void enif_release_resource(void* obj)
 void enif_keep_resource(void* obj)
 {
     ErlNifResource* resource = DATA_TO_RESOURCE(obj);
-    ErtsBinary* bin = ERTS_MAGIC_BIN_FROM_DATA(resource);
+    ErtsBinary* bin = ERTS_MAGIC_BIN_FROM_UNALIGNED_DATA(resource);
 
     ASSERT(ERTS_MAGIC_BIN_DESTRUCTOR(bin) == &nif_resource_dtor);
 #ifdef DEBUG
@@ -1438,7 +1444,7 @@ void enif_keep_resource(void* obj)
 ERL_NIF_TERM enif_make_resource(ErlNifEnv* env, void* obj)
 {
     ErlNifResource* resource = DATA_TO_RESOURCE(obj);
-    ErtsBinary* bin = ERTS_MAGIC_BIN_FROM_DATA(resource);
+    ErtsBinary* bin = ERTS_MAGIC_BIN_FROM_UNALIGNED_DATA(resource);
     Eterm* hp = alloc_heap(env,PROC_BIN_SIZE);
     return erts_mk_magic_binary_term(&hp, &MSO(env->proc), &bin->binary);
 }
@@ -1467,7 +1473,7 @@ int enif_get_resource(ErlNifEnv* env, ERL_NIF_TERM term, ErlNifResourceType* typ
 	return 0; / * Or should we allow "resource binaries" as handles? * /
     }*/
     mbin = pb->val;
-    resource = (ErlNifResource*) ERTS_MAGIC_BIN_DATA(mbin);
+    resource = (ErlNifResource*) ERTS_MAGIC_BIN_UNALIGNED_DATA(mbin);
     if (ERTS_MAGIC_BIN_DESTRUCTOR(mbin) != &nif_resource_dtor
 	|| resource->type != type) {	
 	return 0;
@@ -1479,8 +1485,8 @@ int enif_get_resource(ErlNifEnv* env, ERL_NIF_TERM term, ErlNifResourceType* typ
 size_t enif_sizeof_resource(void* obj)
 {
     ErlNifResource* resource = DATA_TO_RESOURCE(obj);
-    Binary* bin = &ERTS_MAGIC_BIN_FROM_DATA(resource)->binary;
-    return ERTS_MAGIC_BIN_DATA_SIZE(bin) - offsetof(ErlNifResource,data);
+    Binary* bin = &ERTS_MAGIC_BIN_FROM_UNALIGNED_DATA(resource)->binary;
+    return ERTS_MAGIC_BIN_UNALIGNED_DATA_SIZE(bin) - offsetof(ErlNifResource,data);
 }
 
 
@@ -2689,6 +2695,8 @@ erts_unload_nif(struct erl_module_nif* lib)
 
 void erl_nif_init()
 {
+    ERTS_CT_ASSERT((offsetof(ErlNifResource,data) % 8) == ERTS_MAGIC_BIN_BYTES_TO_ALIGN);
+
     resource_type_list.next = &resource_type_list;
     resource_type_list.prev = &resource_type_list;
     resource_type_list.dtor = NULL;
diff --git a/erts/emulator/beam/global.h b/erts/emulator/beam/global.h
index 340c7033ab..ee1f70b748 100644
--- a/erts/emulator/beam/global.h
+++ b/erts/emulator/beam/global.h
@@ -230,9 +230,23 @@ typedef struct {
     ERTS_INTERNAL_BINARY_FIELDS
     SWord orig_size;
     void (*destructor)(Binary *);
-    char magic_bin_data[1];
+    union {
+        struct {
+            ERTS_BINARY_STRUCT_ALIGNMENT
+            char data[1];
+        } aligned;
+        struct {
+            char data[1];
+        } unaligned;
+    } u;
 } ErtsMagicBinary;
 
+#ifdef ARCH_32
+#define ERTS_MAGIC_BIN_BYTES_TO_ALIGN 4
+#else
+#define ERTS_MAGIC_BIN_BYTES_TO_ALIGN 0
+#endif
+
 typedef union {
     Binary binary;
     ErtsMagicBinary magic_binary;
@@ -252,15 +266,30 @@ typedef union {
 #define ERTS_MAGIC_BIN_DESTRUCTOR(BP) \
   ((ErtsBinary *) (BP))->magic_binary.destructor
 #define ERTS_MAGIC_BIN_DATA(BP) \
-  ((void *) ((ErtsBinary *) (BP))->magic_binary.magic_bin_data)
-#define ERTS_MAGIC_BIN_DATA_SIZE(BP) \
-  ((BP)->orig_size - sizeof(void (*)(Binary *)))
+  ((void *) ((ErtsBinary *) (BP))->magic_binary.u.aligned.data)
+#define ERTS_MAGIC_DATA_OFFSET \
+  (offsetof(ErtsMagicBinary,u.aligned.data) - offsetof(Binary,orig_bytes))
 #define ERTS_MAGIC_BIN_ORIG_SIZE(Sz) \
-  (sizeof(void (*)(Binary *)) + (Sz))
+  (ERTS_MAGIC_DATA_OFFSET + (Sz))
 #define ERTS_MAGIC_BIN_SIZE(Sz) \
-  (offsetof(ErtsMagicBinary,magic_bin_data) + (Sz))
-#define ERTS_MAGIC_BIN_FROM_DATA(DATA) \
-  ((ErtsBinary*)((char*)(DATA) - offsetof(ErtsMagicBinary,magic_bin_data)))
+  (offsetof(ErtsMagicBinary,u.aligned.data) + (Sz))
+
+/* On 32-bit arch these macro variants will save memory
+   by not forcing 8-byte alignment for the magic payload.
+*/
+#define ERTS_MAGIC_BIN_UNALIGNED_DATA(BP) \
+  ((void *) ((ErtsBinary *) (BP))->magic_binary.u.unaligned.data)
+#define ERTS_MAGIC_UNALIGNED_DATA_OFFSET \
+  (offsetof(ErtsMagicBinary,u.unaligned.data) - offsetof(Binary,orig_bytes))
+#define ERTS_MAGIC_BIN_UNALIGNED_DATA_SIZE(BP) \
+  ((BP)->orig_size - ERTS_MAGIC_UNALIGNED_DATA_OFFSET)
+#define ERTS_MAGIC_BIN_UNALIGNED_ORIG_SIZE(Sz) \
+  (ERTS_MAGIC_UNALIGNED_DATA_OFFSET + (Sz))
+#define ERTS_MAGIC_BIN_UNALIGNED_SIZE(Sz) \
+  (offsetof(ErtsMagicBinary,u.unaligned.data) + (Sz))
+#define ERTS_MAGIC_BIN_FROM_UNALIGNED_DATA(DATA) \
+  ((ErtsBinary*)((char*)(DATA) - offsetof(ErtsMagicBinary,u.unaligned.data)))
+
 
 #define Binary2ErlDrvBinary(B) (&((ErtsBinary *) (B))->driver.binary)
 #define ErlDrvBinary2Binary(D) ((Binary *) \