Merge pull request #3084 from zhoukangsheng/fix_bthread_id_unlock_failed

fix: fix bthread_id_unlock failed in aarch64 cpu(fix issue #3083)

Author: Huixxi

src/butil/object_pool_inl.h

@@ -415,7 +415,7 @@ class BAIDU_CACHELINE_ALIGNMENT ObjectPool {
     }
 
     inline LocalPool* get_or_new_local_pool() {
-        LocalPool* lp = _local_pool;
+	    LocalPool* lp = BAIDU_GET_VOLATILE_THREAD_LOCAL(_local_pool);
         if (BAIDU_LIKELY(lp != NULL)) {
             return lp;
         }
@@ -424,7 +424,7 @@ class BAIDU_CACHELINE_ALIGNMENT ObjectPool {
             return NULL;
         }
         BAIDU_SCOPED_LOCK(_change_thread_mutex); //avoid race with clear()
-        _local_pool = lp;
+	    BAIDU_SET_VOLATILE_THREAD_LOCAL(_local_pool, lp);
         butil::thread_atexit(LocalPool::delete_local_pool, lp);
         _nlocal.fetch_add(1, butil::memory_order_relaxed);
         return lp;
@@ -523,7 +523,7 @@ class BAIDU_CACHELINE_ALIGNMENT ObjectPool {
 
     static butil::static_atomic<ObjectPool*> _singleton;
     static pthread_mutex_t _singleton_mutex;
-    static BAIDU_THREAD_LOCAL LocalPool* _local_pool;
+    STATIC_MEMBER_BAIDU_VOLATILE_THREAD_LOCAL(LocalPool*, _local_pool);
     static butil::static_atomic<long> _nlocal;
     static butil::static_atomic<size_t> _ngroup;
     static pthread_mutex_t _block_group_mutex;

src/butil/resource_pool_inl.h

@@ -419,7 +419,7 @@ class BAIDU_CACHELINE_ALIGNMENT ResourcePool {
     }
 
     inline LocalPool* get_or_new_local_pool() {
-        LocalPool* lp = _local_pool;
+	    LocalPool* lp = BAIDU_GET_VOLATILE_THREAD_LOCAL(_local_pool);
         if (lp != NULL) {
             return lp;
         }
@@ -428,7 +428,7 @@ class BAIDU_CACHELINE_ALIGNMENT ResourcePool {
             return NULL;
         }
         BAIDU_SCOPED_LOCK(_change_thread_mutex); //avoid race with clear()
-        _local_pool = lp;
+	    BAIDU_SET_VOLATILE_THREAD_LOCAL(_local_pool, lp);
         butil::thread_atexit(LocalPool::delete_local_pool, lp);
         _nlocal.fetch_add(1, butil::memory_order_relaxed);
         return lp;
@@ -524,7 +524,7 @@ class BAIDU_CACHELINE_ALIGNMENT ResourcePool {
 
     static butil::static_atomic<ResourcePool*> _singleton;
     static pthread_mutex_t _singleton_mutex;
-    static BAIDU_THREAD_LOCAL LocalPool* _local_pool;
+    STATIC_MEMBER_BAIDU_VOLATILE_THREAD_LOCAL(LocalPool*, _local_pool);
     static butil::static_atomic<long> _nlocal;
     static butil::static_atomic<size_t> _ngroup;
     static pthread_mutex_t _block_group_mutex;

src/butil/thread_local.h

@@ -46,6 +46,22 @@
     var_name = v;                                                              \
   }
 
+ #define STATIC_MEMBER_BAIDU_VOLATILE_THREAD_LOCAL(type, var_name)              \
+   static BAIDU_THREAD_LOCAL type var_name;                                     \
+   static __attribute__((noinline, unused)) type get_##var_name(void) {         \
+     asm volatile("");                                                          \
+     return var_name;                                                           \
+   }                                                                            \
+   static __attribute__((noinline, unused)) type *get_ptr_##var_name(void) {    \
+     type *ptr = &var_name;                                                     \
+     asm volatile("" : "+rm"(ptr));                                             \
+     return ptr;                                                                \
+   }                                                                            \
+   static __attribute__((noinline, unused)) void set_##var_name(type v) {       \
+     asm volatile("");                                                          \
+     var_name = v;                                                              \
+   }
+
 #if (defined (__aarch64__) && defined (__GNUC__)) || defined(__clang__)
 // GNU compiler under aarch and Clang compiler is incorrectly caching the 
 // address of thread_local variables across a suspend-point. The following