[ruby-dev:39659] stack area in blocking region with -O0

-O0 でコンパイルした ruby で以下のようにすると、stack の
 mark で変なところをアクセスします。

% valgrind ./ruby -ve 'Thread.new { gets }; 10.times { GC.start }'
==4690== Memcheck, a memory error detector.
==4690== Copyright (C) 2002-2007, and GNU GPL'd, by Julian Seward et al.
==4690== Using LibVEX rev 1854, a library for dynamic binary translation.
==4690== Copyright (C) 2004-2007, and GNU GPL'd, by OpenWorks LLP.
==4690== Using valgrind-3.3.1-Debian, a dynamic binary instrumentation framework.
==4690== Copyright (C) 2000-2007, and GNU GPL'd, by Julian Seward et al.
==4690== For more details, rerun with: -v
==4690== 
ruby 1.9.2dev (2009-11-09 trunk 25699) [i686-linux]
==4690== Invalid read of size 4
==4690==    at 0x80643AD: mark_locations_array (gc.c:1313)
==4690==    by 0x8064575: gc_mark_locations (gc.c:1329)
==4690==    by 0x8066182: rb_gc_mark_machine_stack (gc.c:2221)
==4690==    by 0x814EC56: rb_thread_mark (vm.c:1630)
==4690==    by 0x8064F07: gc_mark_children (gc.c:1713)
==4690==    by 0x8064A11: gc_mark (gc.c:1508)
==4690==    by 0x8064A3B: rb_gc_mark (gc.c:1514)
==4690==    by 0x814E66F: vm_mark_each_thread_func (vm.c:1432)
==4690==    by 0x80F3C2C: st_foreach (st.c:747)
==4690==    by 0x814E6DC: rb_vm_mark (vm.c:1455)
==4690==    by 0x8064F07: gc_mark_children (gc.c:1713)
==4690==    by 0x8064A11: gc_mark (gc.c:1508)
==4690==  Address 0x48d6140 is not stack'd, malloc'd or (recently) free'd
==4690== 
==4690== ERROR SUMMARY: 70 errors from 1 contexts (suppressed: 29 from 1)
==4690== malloc/free: in use at exit: 890,874 bytes in 13,756 blocks.
==4690== malloc/free: 15,651 allocs, 1,895 frees, 2,158,101 bytes allocated.
==4690== For counts of detected errors, rerun with: -v
==4690== searching for pointers to 13,756 not-freed blocks.
==4690== checked 962,284 bytes.
==4690== 
==4690== LEAK SUMMARY:
==4690==    definitely lost: 222,005 bytes in 9,073 blocks.
==4690==      possibly lost: 136 bytes in 1 blocks.
==4690==    still reachable: 668,733 bytes in 4,682 blocks.
==4690==         suppressed: 0 bytes in 0 blocks.
==4690== Rerun with --leak-check=full to see details of leaked memory.

調べてみると、この変なところというのは、スタックが縮んだとこ
ろです。どうも rb_thread_t の machine_stack_end を記録した後
にスタックが縮んでそのタイミングで GC が起きるということのよ
うです。

machine_stack_end の記録からスレッドのスイッチに至るまでにス
タックが縮まないようにして修正できます。

% svn diff --diff-cmd diff -x '-u -p'
Index: thread.c
===================================================================
--- thread.c	(revision 25700)
+++ thread.c	(working copy)
@@ -97,13 +97,17 @@ static void set_unblock_function(rb_thre
 				 struct rb_unblock_callback *old);
 static void reset_unblock_function(rb_thread_t *th, const struct rb_unblock_callback *old);
 
-static inline void blocking_region_begin(rb_thread_t *th, struct rb_blocking_region_buffer *region,
-					 rb_unblock_function_t *func, void *arg);
 static inline void blocking_region_end(rb_thread_t *th, struct rb_blocking_region_buffer *region);
 
+#define RB_GC_SAVE_MACHINE_CONTEXT(th) \
+  do { \
+    rb_gc_save_machine_context(th); \
+    SET_MACHINE_STACK_END(&(th)->machine_stack_end); \
+  } while (0)
+
 #define GVL_UNLOCK_BEGIN() do { \
   rb_thread_t *_th_stored = GET_THREAD(); \
-  rb_gc_save_machine_context(_th_stored); \
+  RB_GC_SAVE_MACHINE_CONTEXT(_th_stored); \
   native_mutex_unlock(&_th_stored->vm->global_vm_lock)
 
 #define GVL_UNLOCK_END() \
@@ -960,7 +964,7 @@ rb_thread_schedule_rec(int sched_depth)
 
 	thread_debug("rb_thread_schedule/switch start\n");
 
-	rb_gc_save_machine_context(th);
+	RB_GC_SAVE_MACHINE_CONTEXT(th);
 	native_mutex_unlock(&th->vm->global_vm_lock);
 	{
 	    native_thread_yield();
@@ -983,18 +987,16 @@ rb_thread_schedule(void)
 }
 
 /* blocking region */
-static inline void
-blocking_region_begin(rb_thread_t *th, struct rb_blocking_region_buffer *region,
-		      rb_unblock_function_t *func, void *arg)
-{
-    region->prev_status = th->status;
-    th->blocking_region_buffer = region;
-    set_unblock_function(th, func, arg, &region->oldubf);
-    th->status = THREAD_STOPPED;
-    thread_debug("enter blocking region (%p)\n", (void *)th);
-    rb_gc_save_machine_context(th);
-    native_mutex_unlock(&th->vm->global_vm_lock);
-}
+#define blocking_region_begin(th, region, func, arg) \
+  do { \
+    (region)->prev_status = (th)->status; \
+    (th)->blocking_region_buffer = (region); \
+    set_unblock_function((th), (func), (arg), &(region)->oldubf); \
+    (th)->status = THREAD_STOPPED; \
+    thread_debug("enter blocking region (%p)\n", (void *)(th)); \
+    RB_GC_SAVE_MACHINE_CONTEXT(th); \
+    native_mutex_unlock(&(th)->vm->global_vm_lock); \
+  } while (0)
 
 static inline void
 blocking_region_end(rb_thread_t *th, struct rb_blocking_region_buffer *region)
@@ -2599,7 +2601,6 @@ rb_gc_set_stack_end(VALUE **stack_end_p)
 void
 rb_gc_save_machine_context(rb_thread_t *th)
 {
-    SET_MACHINE_STACK_END(&th->machine_stack_end);
     FLUSH_REGISTER_WINDOWS;
 #ifdef __ia64
     th->machine_register_stack_end = rb_ia64_bsp();
-- 
[田中 哲][たなか あきら][Tanaka Akira]