[ovs-dev] [threads 02/11] ovs-thread: Add per-thread data support.

Ben Pfaff blp at nicira.com
Thu Jun 20 20:18:03 UTC 2013


On Wed, Jun 19, 2013 at 01:17:03PM -0700, Ben Pfaff wrote:
> POSIX defines a portable pthread_key_t API for per-thread data.  GCC and
> C11 have two different forms of per-thread data that are generally faster
> than the POSIX API, where they are available.  This commit adds a
> macro-based wrapper, DEFINE_PER_THREAD_DATA, that takes advantage of the
> GCC extension where it is available and falls back to the POSIX API
> otherwise.  (I'm not aware of any compilers that implement the C11 feature,
> so this commit doesn't try to use it.)

Ed Maste pointed out off-list that clang on FreeBSD supports
_Thread_local.  Here's a revised version of the patch that supports
both _Thread_local and __thread.  I've also updated the "reviews"
branch.

--8<--------------------------cut here-------------------------->8--

From: Ben Pfaff <blp at nicira.com>
Date: Thu, 20 Jun 2013 13:11:26 -0700
Subject: [PATCH] ovs-thread: Add per-thread data support.

POSIX defines a portable pthread_key_t API for per-thread data.  GCC and
C11 have two different forms of per-thread data that are generally faster
than the POSIX API, where they are available.  This commit adds a
macro-based wrapper, DEFINE_PER_THREAD_DATA, that takes advantage of these
features where they are available and falls back to the POSIX API
otherwise.

The Clang compiler implements the C11 _Thread_local keyword.

This commit also adds a convenience wrapper for the POSIX API, via the
DEFINE_PER_THREAD_MALLOCED_DATA macro.

Signed-off-by: Ben Pfaff <blp at nicira.com>
---
 configure.ac      |    1 +
 lib/ovs-thread.h  |  194 +++++++++++++++++++++++++++++++++++++++++++++++++++++
 m4/openvswitch.m4 |   37 ++++++++++-
 3 files changed, 231 insertions(+), 1 deletions(-)

diff --git a/configure.ac b/configure.ac
index a691963..6c610c1 100644
--- a/configure.ac
+++ b/configure.ac
@@ -80,6 +80,7 @@ OVS_CHECK_XENSERVER_VERSION
 OVS_CHECK_GROFF
 OVS_CHECK_GNU_MAKE
 OVS_CHECK_CACHE_TIME
+OVS_CHECK__THREAD_LOCAL
 
 OVS_ENABLE_OPTION([-Wall])
 OVS_ENABLE_OPTION([-Wno-sign-compare])
diff --git a/lib/ovs-thread.h b/lib/ovs-thread.h
index cafeedf..d5bf048 100644
--- a/lib/ovs-thread.h
+++ b/lib/ovs-thread.h
@@ -85,5 +85,199 @@ void xpthread_cond_wait(pthread_cond_t *, pthread_mutex_t *mutex)
 void xpthread_key_create(pthread_key_t *, void (*destructor)(void *));
 
 void xpthread_create(pthread_t *, pthread_attr_t *, void *(*)(void *), void *);
+
+/* Per-thread data.
+ *
+ * Multiple forms of per-thread data exist, each with its own pluses and
+ * minuses:
+ *
+ *     - POSIX per-thread data via pthread_key_t is portable to any pthreads
+ *       implementation, and allows a destructor function to be defined.  It
+ *       only (directly) supports per-thread pointers, which are always
+ *       initialized to NULL.  It requires once-only allocation of a
+ *       pthread_key_t value.  It is relatively slow.
+ *
+ *     - The _Thread_local keyword newly defined in C11 works with any data
+ *       type and initializer, and it is fast.  _Thread_local does not require
+ *       once-only initialization like pthread_key_t.  C11 does not define what
+ *       happens if one attempts to access a _Thread_local object from a thread
+ *       other than the one to which that object belongs.  There is no
+ *       provision to call a user-specified destructor when a thread ends.
+ *
+ *     - The __thread keyword is a GCC extension similar to _Thread_local but
+ *       with a longer history.  __thread is not portable to every GCC version
+ *       or environment.  __thread does not restrict the use of a thread-local
+ *       object outside its own thread.
+ *
+ * Here's a handy summary:
+ *
+ *                     pthread_key_t     _Thread_local        __thread
+ *                     -------------     -------------     -------------
+ * portability             high               low              medium
+ * speed                    low              high                high
+ * supports destructors?    yes                no                  no
+ * needs key allocation?    yes                no                  no
+ * arbitrary initializer?    no               yes                 yes
+ * cross-thread access?     yes                no                 yes
+ */
+
+/* DEFINE_PER_THREAD_DATA(TYPE, NAME, INITIALIZER).
+ *
+ * One should prefer to use POSIX per-thread data, via pthread_key_t, when its
+ * performance is acceptable, because of its portability (see the table above).
+ * This macro is an alternatives that takes advantage of _Thread_local (and
+ * __thread, via a config.h macro substitution), for its performance, when it
+ * is available, and falls back to POSIX per-thread data otherwise.
+ *
+ * Defines per-thread variable NAME with the given TYPE, initialized to
+ * INITIALIZER (which must be valid as an initializer for a variable with
+ * static lifetime).
+ *
+ * The public interface to the variable is:
+ *
+ *    TYPE *NAME_get(void)
+ *    TYPE *NAME_get__(void)
+ *
+ *       Returns the address of this thread's instance of NAME.
+ *
+ *       Use NAME_get() in a context where this might be the first use of the
+ *       per-thread variable in the program.  Use NAME_get__(), which avoids a
+ *       conditional test and is thus slightly faster, in a context where one
+ *       knows that NAME_get() has already been called previously.
+ *
+ * There are no "NAME_set()" or "NAME_set__()" functions.  To set the value of
+ * the per-thread variable, dereference the pointer returned by TYPE_get() or
+ * TYPE_get__(), e.g. *TYPE_get() = 0.
+ */
+#if HAVE__THREAD_LOCAL
+#define DEFINE_PER_THREAD_DATA(TYPE, NAME, ...)                 \
+    typedef TYPE NAME##_type;                                   \
+    static _Thread_local NAME##_type NAME##_var = __VA_ARGS__;  \
+                                                                \
+    static NAME##_type *                                        \
+    NAME##_get__(void)                                          \
+    {                                                           \
+        return &NAME##_var;                                     \
+    }                                                           \
+                                                                \
+    static NAME##_type *                                        \
+    NAME##_get(void)                                            \
+    {                                                           \
+        return NAME##_get__();                                  \
+    }
+#else
+#define DEFINE_PER_THREAD_DATA(TYPE, NAME, ...)                         \
+    typedef TYPE NAME##_type;                                           \
+    static pthread_key_t NAME##_key;                                    \
+                                                                        \
+    static NAME##_type *                                                \
+    NAME##_get__(void)                                                  \
+    {                                                                   \
+        return pthread_getspecific(NAME##_key);                         \
+    }                                                                   \
+                                                                        \
+    static void                                                         \
+    NAME##_once_init(void)                                              \
+    {                                                                   \
+        if (pthread_key_create(&NAME##_key, free)) {                    \
+            abort();                                                    \
+        }                                                               \
+    }                                                                   \
+                                                                        \
+    static NAME##_type *                                                \
+    NAME##_get(void)                                                    \
+    {                                                                   \
+        static pthread_once_t once = PTHREAD_ONCE_INIT;                 \
+        NAME##_type *value;                                             \
+                                                                        \
+        pthread_once(&once, NAME##_once_init);                          \
+        value = NAME##_get__();                                         \
+        if (!value) {                                                   \
+            static const NAME##_type initial_value = __VA_ARGS__;       \
+                                                                        \
+            value = xmalloc(sizeof *value);                             \
+            *value = initial_value;                                     \
+            pthread_setspecific(NAME##_key, value);                     \
+        }                                                               \
+        return value;                                                   \
+    }
+#endif
+
+/* DEFINE_PER_THREAD_MALLOCED_DATA(TYPE, NAME).
+ *
+ * This is a simple wrapper around POSIX per-thread data primitives.  It
+ * defines per-thread variable NAME with the given TYPE, which must be a
+ * pointer type.  In each thread, the per-thread variable is initialized to
+ * NULL.  When a thread terminates, the variable is freed with free().
+ *
+ * The public interface to the variable is:
+ *
+ *    TYPE NAME_get(void)
+ *    TYPE NAME_get__(void)
+ *
+ *       Returns the value of per-thread variable NAME in this thread.
+ *
+ *       Use NAME_get() in a context where this might be the first use of the
+ *       per-thread variable in the program.  Use NAME_get__(), which avoids a
+ *       conditional test and is thus slightly faster, in a context where one
+ *       knows that NAME_get() has already been called previously.
+ *
+ *    TYPE NAME_set(TYPE new_value)
+ *    TYPE NAME_set__(TYPE new_value)
+ *
+ *       Sets the value of per-thread variable NAME to 'new_value' in this
+ *       thread, and returns its previous value.
+ *
+ *       Use NAME_set() in a context where this might be the first use of the
+ *       per-thread variable in the program.  Use NAME_set__(), which avoids a
+ *       conditional test and is thus slightly faster, in a context where one
+ *       knows that NAME_set() has already been called previously.
+ */
+#define DEFINE_PER_THREAD_MALLOCED_DATA(TYPE, NAME)     \
+    static pthread_key_t NAME##_key;                    \
+                                                        \
+    static void                                         \
+    NAME##_once_init(void)                              \
+    {                                                   \
+        if (pthread_key_create(&NAME##_key, free)) {    \
+            abort();                                    \
+        }                                               \
+    }                                                   \
+                                                        \
+    static void                                         \
+    NAME##_init(void)                                   \
+    {                                                   \
+        static pthread_once_t once = PTHREAD_ONCE_INIT; \
+        pthread_once(&once, NAME##_once_init);          \
+    }                                                   \
+                                                        \
+    static TYPE                                         \
+    NAME##_get__(void)                                  \
+    {                                                   \
+        return pthread_getspecific(NAME##_key);         \
+    }                                                   \
+                                                        \
+    static OVS_UNUSED TYPE                              \
+    NAME##_get(void)                                    \
+    {                                                   \
+        NAME##_init();                                  \
+        return NAME##_get__();                          \
+    }                                                   \
+                                                        \
+    static TYPE                                         \
+    NAME##_set__(TYPE value)                            \
+    {                                                   \
+        TYPE old_value = NAME##_get__();                \
+        pthread_setspecific(NAME##_key, value);         \
+        return old_value;                               \
+    }                                                   \
+                                                        \
+    static OVS_UNUSED TYPE                              \
+    NAME##_set(TYPE value)                              \
+    {                                                   \
+        NAME##_init();                                  \
+        return NAME##_set__(value);                     \
+    }
+
 
 #endif /* ovs-thread.h */
diff --git a/m4/openvswitch.m4 b/m4/openvswitch.m4
index 12c02c0..3895346 100644
--- a/m4/openvswitch.m4
+++ b/m4/openvswitch.m4
@@ -1,6 +1,6 @@
 # -*- autoconf -*-
 
-# Copyright (c) 2008, 2009, 2010, 2011, 2012 Nicira, Inc.
+# Copyright (c) 2008, 2009, 2010, 2011, 2012, 2013 Nicira, Inc.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
 # you may not use this file except in compliance with the License.
@@ -390,3 +390,38 @@ AC_DEFUN([OVS_CHECK_GROFF],
        ovs_cv_groff=no
      fi])
    AM_CONDITIONAL([HAVE_GROFF], [test "$ovs_cv_groff" = yes])])
+
+dnl OVS_CHECK__THREAD_LOCAL
+dnl
+dnl Checks whether the compiler and linker support the C11
+dnl _Thread_local storage class or the GCC __thread extension, and if
+dnl so define HAVE__THREAD_LOCAL.  If __thread is supported, also #defines
+dnl _Thread_local to __thread to hide the difference.
+AC_DEFUN([OVS_CHECK__THREAD_LOCAL],
+  [AC_CACHE_CHECK(
+     [whether $CC supports _Thread_local],
+     [ovs_cv__Thread_local],
+     [AC_LINK_IFELSE(
+        [AC_LANG_PROGRAM([static _Thread_local var;], [return var;])],
+        [ovs_cv__Thread_local=yes],
+        [ovs_cv__Thread_local=no])])
+   if test $ovs_cv__Thread_local = no; then
+     AC_CACHE_CHECK(
+       [whether $CC supports __thread],
+       [ovs_cv___thread],
+       [AC_LINK_IFELSE(
+          [AC_LANG_PROGRAM([static __thread var;], [return var;])],
+          [ovs_cv___thread=yes],
+          [ovs_cv___thread=no])])
+     if test $ovs_cv___thread = yes; then
+       AC_DEFINE([_Thread_local], [__thread],
+                 [Define _Thread_local to use GCC __thread extension.])
+     fi
+   fi
+   if test $ovs_cv__Thread_local = yes || test $ovs_cv___thread = yes; then
+     AC_DEFINE(
+       [HAVE__THREAD_LOCAL],
+       [1],
+       [Define to 1 if the C compiler and linker support the C11 _Thread_local
+        storage class or the compatible GCC __thread extension.])
+   fi])
-- 
1.7.2.5




More information about the dev mailing list