diff --git a/configure.ac b/configure.ac
index 670863f98..09941599e 100644
--- a/configure.ac
+++ b/configure.ac
@@ -101,6 +101,27 @@ AS_IF([test "x$enable_apple_tsd_optimizations" = "xyes"],
     [Define to use non-portable pthread TSD optimizations for Mac OS X)])]
 )
 
+#
+# Enable __thread based TSD on platforms where it is efficient
+# Allow override based on command line argument to configure
+#
+AC_CANONICAL_TARGET
+AC_ARG_ENABLE([thread-local-storage],
+  [AS_HELP_STRING([--enable-thread-local-storage],
+    [Enable usage of thread local storage via __thread])],,
+  [case $target_os in
+      linux*)
+        enable_thread_local_storage=yes
+	    ;;
+      *)
+        enable_thread_local_storage=no
+   esac]
+)
+AS_IF([test "x$enable_thread_local_storage" = "xyes"],
+  [AC_DEFINE(DISPATCH_USE_THREAD_LOCAL_STORAGE, 1,
+    [Enable usage of thread local storage via __thread])]
+)
+
 AC_USE_SYSTEM_EXTENSIONS
 AM_INIT_AUTOMAKE([foreign no-dependencies subdir-objects])
 LT_INIT([disable-static])
diff --git a/src/init.c b/src/init.c
index 8e538b913..1cbaa7ac7 100644
--- a/src/init.c
+++ b/src/init.c
@@ -65,7 +65,7 @@ void *(*_dispatch_begin_NSAutoReleasePool)(void);
 void (*_dispatch_end_NSAutoReleasePool)(void *);
 #endif
 
-#if !DISPATCH_USE_DIRECT_TSD
+#if !DISPATCH_USE_DIRECT_TSD && !DISPATCH_USE_THREAD_LOCAL_STORAGE
 pthread_key_t dispatch_queue_key;
 pthread_key_t dispatch_sema4_key;
 pthread_key_t dispatch_cache_key;
@@ -77,7 +77,7 @@ pthread_key_t dispatch_introspection_key;
 #elif DISPATCH_PERF_MON
 pthread_key_t dispatch_bcounter_key;
 #endif
-#endif // !DISPATCH_USE_DIRECT_TSD
+#endif // !DISPATCH_USE_DIRECT_TSD && !DISPATCH_USE_THREAD_LOCAL_STORAGE
 
 #if VOUCHER_USE_MACH_VOUCHER
 dispatch_once_t _voucher_task_mach_voucher_pred;
@@ -161,10 +161,6 @@ const struct dispatch_tsd_indexes_s dispatch_tsd_indexes = {
 	.dti_voucher_index = dispatch_voucher_key,
 	.dti_qos_class_index = dispatch_priority_key,
 };
-#else // DISPATCH_USE_DIRECT_TSD
-#ifndef __LINUX_PORT_HDD__
-#error Not implemented on this platform
-#endif
 #endif // DISPATCH_USE_DIRECT_TSD
 
 // 6618342 Contact the team that owns the Instrument DTrace probe before
diff --git a/src/queue.c b/src/queue.c
index db414f69d..d7c4640f9 100644
--- a/src/queue.c
+++ b/src/queue.c
@@ -913,6 +913,60 @@ _dispatch_root_queues_init(void *context DISPATCH_UNUSED)
 
 #define countof(x) (sizeof(x) / sizeof(x[0]))
 
+#if DISPATCH_USE_THREAD_LOCAL_STORAGE
+#include <unistd.h>
+#include <sys/syscall.h>
+
+#ifdef SYS_gettid
+DISPATCH_ALWAYS_INLINE
+static inline pid_t
+gettid(void)
+{
+	return (pid_t) syscall(SYS_gettid);
+}
+#else
+#error "SYS_gettid unavailable on this system"
+#endif
+
+#define _tsd_call_cleanup(k, f)  do { \
+		if (f && tsd->k) ((void(*)(void*))(f))(tsd->k); \
+    } while (0)
+
+static void
+_libdispatch_tsd_cleanup(void *ctx)
+{
+	struct dispatch_tsd *tsd = (struct dispatch_tsd*) ctx;
+
+	_tsd_call_cleanup(dispatch_queue_key, _dispatch_queue_cleanup);
+	_tsd_call_cleanup(dispatch_voucher_key, _voucher_thread_cleanup);
+	_tsd_call_cleanup(dispatch_cache_key, _dispatch_cache_cleanup);
+	_tsd_call_cleanup(dispatch_io_key, NULL);
+	_tsd_call_cleanup(dispatch_apply_key, NULL);
+	_tsd_call_cleanup(dispatch_defaultpriority_key, NULL);
+	_tsd_call_cleanup(dispatch_pthread_root_queue_observer_hooks_key,
+			NULL);
+#if DISPATCH_PERF_MON && !DISPATCH_INTROSPECTION
+	_tsd_call_cleanup(dispatch_bcounter_key, NULL);
+#endif
+#if !DISPATCH_USE_OS_SEMAPHORE_CACHE
+	_tsd_call_cleanup(dispatch_sema4_key,
+			(void (*)(void *))_dispatch_thread_semaphore_dispose);
+#endif
+	tsd->tid = -1;
+	tsd->initialized = false;
+}
+
+DISPATCH_EXPORT DISPATCH_NOINLINE
+void
+libdispatch_tsd_init(void)
+{
+	__dispatch_tsd.tid = gettid();
+	pthread_setspecific(__dispatch_tsd_key, &__dispatch_tsd);
+	__dispatch_tsd.initialized = true;
+}
+
+#endif
+
 DISPATCH_EXPORT DISPATCH_NOTHROW
 void
 libdispatch_init(void)
@@ -951,6 +1005,9 @@ libdispatch_init(void)
 	dispatch_assert(sizeof(struct dispatch_root_queue_context_s) %
 			DISPATCH_CACHELINE_SIZE == 0);
 
+#if DISPATCH_USE_THREAD_LOCAL_STORAGE
+	_dispatch_thread_key_create(&__dispatch_tsd_key, _libdispatch_tsd_cleanup);
+#else
 	_dispatch_thread_key_create(&dispatch_queue_key, _dispatch_queue_cleanup);
 	_dispatch_thread_key_create(&dispatch_voucher_key, _voucher_thread_cleanup);
 	_dispatch_thread_key_create(&dispatch_cache_key, _dispatch_cache_cleanup);
@@ -966,6 +1023,7 @@ libdispatch_init(void)
 	_dispatch_thread_key_create(&dispatch_sema4_key,
 			(void (*)(void *))_dispatch_thread_semaphore_dispose);
 #endif
+#endif
 
 #if DISPATCH_USE_RESOLVERS // rdar://problem/8541707
 	_dispatch_main_q.do_targetq = &_dispatch_root_queues[
diff --git a/src/shims/linux_stubs.c b/src/shims/linux_stubs.c
index e62ee8fbd..bf9af6254 100644
--- a/src/shims/linux_stubs.c
+++ b/src/shims/linux_stubs.c
@@ -53,8 +53,9 @@ char* mach_error_string(mach_msg_return_t x) {
 void mach_vm_deallocate() { LINUX_PORT_ERROR();  }
 
 mach_port_t pthread_mach_thread_np(void) {
-  return (pid_t)syscall(SYS_gettid);
+  return (mach_port_t)pthread_self();
 }
+
 mach_port_t mach_task_self(void) {
   return (mach_port_t)pthread_self();
 }
@@ -72,3 +73,8 @@ unsigned short dispatch_timer__fire_semaphore;
 unsigned short dispatch_timer__configure_semaphore;
 void (*_dispatch_block_special_invoke)(void*);
 struct dispatch_queue_attr_s _dispatch_queue_attr_concurrent;
+
+#if DISPATCH_USE_THREAD_LOCAL_STORAGE
+__thread struct dispatch_tsd __dispatch_tsd;
+pthread_key_t __dispatch_tsd_key;
+#endif
diff --git a/src/shims/tsd.h b/src/shims/tsd.h
index 40c4a583e..6da78ca83 100644
--- a/src/shims/tsd.h
+++ b/src/shims/tsd.h
@@ -73,6 +73,56 @@ _dispatch_thread_key_create(const unsigned long *k, void (*d)(void *))
 	if (!*k || !d) return;
 	dispatch_assert_zero(pthread_key_init_np((int)*k, d));
 }
+#elif DISPATCH_USE_THREAD_LOCAL_STORAGE
+
+DISPATCH_TSD_INLINE
+static inline void
+_dispatch_thread_key_create(pthread_key_t *k, void (*d)(void *))
+{
+	dispatch_assert_zero(pthread_key_create(k, d));
+}
+
+struct dispatch_tsd {
+	bool initialized;
+	pid_t tid;
+	void *dispatch_queue_key;
+	void *dispatch_voucher_key;
+#if DISPATCH_USE_OS_SEMAPHORE_CACHE
+#error "Invalid DISPATCH_USE_OS_SEMAPHORE_CACHE configuration"
+#else
+	void *dispatch_sema4_key;
+#endif
+	void *dispatch_cache_key;
+	void *dispatch_io_key;
+	void *dispatch_apply_key;
+	void *dispatch_defaultpriority_key;
+#if DISPATCH_INTROSPECTION
+	void *dispatch_introspection_key;
+#elif DISPATCH_PERF_MON
+	void *dispatch_bcounter_key;
+#endif
+	void *dispatch_pthread_root_queue_observer_hooks_key;
+};
+
+extern __thread struct dispatch_tsd __dispatch_tsd;
+extern pthread_key_t __dispatch_tsd_key;
+extern void libdispatch_tsd_init(void);
+
+DISPATCH_ALWAYS_INLINE
+static inline struct dispatch_tsd *
+_dispatch_get_tsd_base(void)
+{
+	if (slowpath(!__dispatch_tsd.initialized)) {
+		libdispatch_tsd_init();
+	}
+	return &__dispatch_tsd;
+}
+
+#define _dispatch_thread_getspecific(key) \
+	(_dispatch_get_tsd_base()->key)
+#define _dispatch_thread_setspecific(key, value) \
+	(void)(_dispatch_get_tsd_base()->key = (value))
+
 #else
 extern pthread_key_t dispatch_queue_key;
 extern pthread_key_t dispatch_voucher_key;
@@ -100,7 +150,7 @@ _dispatch_thread_key_create(pthread_key_t *k, void (*d)(void *))
 }
 #endif
 
-#if DISPATCH_USE_TSD_BASE && !DISPATCH_DEBUG
+#if (DISPATCH_USE_TSD_BASE && !DISPATCH_DEBUG) || DISPATCH_USE_THREAD_LOCAL_STORAGE
 #else // DISPATCH_USE_TSD_BASE
 DISPATCH_TSD_INLINE
 static inline void
@@ -134,6 +184,8 @@ _dispatch_thread_getspecific(pthread_key_t k)
 #if DISPATCH_USE_DIRECT_TSD
 #define _dispatch_thread_self() ((uintptr_t)_dispatch_thread_getspecific( \
 		_PTHREAD_TSD_SLOT_PTHREAD_SELF))
+#elif DISPATCH_USE_THREAD_LOCAL_STORAGE
+#define _dispatch_thread_self() ((uintptr_t)pthread_self())
 #else
 #define _dispatch_thread_self() ((uintptr_t)pthread_self())
 #endif
@@ -145,6 +197,8 @@ _dispatch_thread_getspecific(pthread_key_t k)
 #if DISPATCH_USE_DIRECT_TSD
 #define _dispatch_thread_port() ((mach_port_t)_dispatch_thread_getspecific(\
 		_PTHREAD_TSD_SLOT_MACH_THREAD_SELF))
+#elif DISPATCH_USE_THREAD_LOCAL_STORAGE
+#define _dispatch_thread_port() ((mach_port_t)(_dispatch_get_tsd_base()->tid))
 #else
 #define _dispatch_thread_port() (pthread_mach_thread_np(_dispatch_thread_self()))
 #endif