Merged,
Maxim.
On 12/16/2015 16:51, Savolainen, Petri (Nokia - FI/Espoo) wrote:
Ping. Waiting to be merged.
*From:*EXT Bill Fischofer [mailto:[email protected]]
*Sent:* Sunday, December 13, 2015 12:44 AM
*To:* Savolainen, Petri (Nokia - FI/Espoo)
*Cc:* LNG ODP Mailman List
*Subject:* Re: [lng-odp] [API-NEXT PATCH 1/2] api: barrier: added
memory barriers
On Fri, Dec 11, 2015 at 5:30 AM, Petri Savolainen
<[email protected] <mailto:[email protected]>> wrote:
Added new memory barriers. These follow C11 release /
acquire specification and replaces odp_sync_stores().
Used GCC __atomic_thread_fence to implement all three
barriers.
Signed-off-by: Petri Savolainen <[email protected]
<mailto:[email protected]>>
Reviewed-by: Bill Fischofer <[email protected]
<mailto:[email protected]>>
---
include/odp/api/barrier.h | 11 ++++-
include/odp/api/sync.h | 82
++++++++++++++++++++-----------
platform/linux-generic/include/odp/sync.h | 28 +++++++++++
3 files changed, 90 insertions(+), 31 deletions(-)
diff --git a/include/odp/api/barrier.h b/include/odp/api/barrier.h
index 8ca2647..823eae6 100644
--- a/include/odp/api/barrier.h
+++ b/include/odp/api/barrier.h
@@ -18,8 +18,15 @@
extern "C" {
#endif
-/** @defgroup odp_barrier ODP BARRIER
- * Thread excution and memory ordering barriers.
+/**
+ * @defgroup odp_barrier ODP BARRIER
+ * Thread excution and memory ordering barriers.
+ *
+ * @details
+ * <b> Thread execution barrier (odp_barrier_t) </b>
+ *
+ * Thread execution barrier synchronizes a group of threads to
wait on the
+ * barrier until the entire group has reached the barrier.
* @{
*/
diff --git a/include/odp/api/sync.h b/include/odp/api/sync.h
index 6477e74..c6f790c 100644
--- a/include/odp/api/sync.h
+++ b/include/odp/api/sync.h
@@ -8,7 +8,7 @@
/**
* @file
*
- * ODP synchronisation
+ * ODP memory barriers
*/
#ifndef ODP_API_SYNC_H_
@@ -18,42 +18,66 @@
extern "C" {
#endif
-/** @addtogroup odp_barrier
+/**
+ * @addtogroup odp_barrier
+ * @details
+ * <b> Memory barriers </b>
+ *
+ * Memory barriers enforce ordering of memory load and store
operations
+ * specified before and after the barrier. These barriers may
affect both
+ * compiler optimizations and CPU out-of-order execution. All ODP
+ * synchronization mechanisms (e.g. execution barriers, locks,
queues, etc )
+ * include all necessary memory barriers, so these calls are not
needed when
+ * using those. Also ODP atomic operations have memory ordered
versions. These
+ * explicit barriers may be needed when thread synchronization is
based on
+ * a non-ODP defined mechanism. Depending on the HW platform,
heavy usage of
+ * memory barriers may cause significant performance degradation.
+ *
* @{
*/
/**
- * Synchronise stores
+ * Memory barrier for release operations
*
- * Ensures that all CPU store operations that precede the
odp_sync_stores()
- * call are globally visible before any store operation that
follows it.
+ * This memory barrier has release semantics. It synchronizes
with a pairing
+ * barrier for acquire operations. The releasing and acquiring
threads
+ * synchronize through shared memory. The releasing thread must
call this
+ * barrier before signaling the acquiring thread. After the
acquiring thread
+ * receives the signal, it must call odp_mb_acquire() before it
reads the
+ * memory written by the releasing thread.
+ *
+ * This call is not needed when using ODP defined synchronization
mechanisms.
+ *
+ * @see odp_mb_acquire()
*/
-static inline void odp_sync_stores(void)
-{
-#if defined __x86_64__ || defined __i386__
-
- __asm__ __volatile__ ("sfence\n" : : : "memory");
-
-#elif defined(__arm__)
-#if __ARM_ARCH == 6
- __asm__ __volatile__ ("mcr p15, 0, %0, c7, c10, 5" \
- : : "r" (0) : "memory");
-#elif __ARM_ARCH >= 7 || defined __aarch64__
-
- __asm__ __volatile__ ("dmb st" : : : "memory");
-#else
- __asm__ __volatile__ ("" : : : "memory");
-#endif
-
-#elif defined __OCTEON__
-
- __asm__ __volatile__ ("syncws\n" : : : "memory");
+void odp_mb_release(void);
-#else
- __sync_synchronize();
-#endif
-}
+/**
+ * Memory barrier for acquire operations
+ *
+ * This memory barrier has acquire semantics. It synchronizes
with a pairing
+ * barrier for release operations. The releasing and acquiring
threads
+ * synchronize through shared memory. The releasing thread must call
+ * odp_mb_release() before signaling the acquiring thread. After
the acquiring
+ * thread receives the signal, it must call this barrier before
it reads the
+ * memory written by the releasing thread.
+ *
+ * This call is not needed when using ODP defined synchronization
mechanisms.
+ *
+ * @see odp_mb_release()
+ */
+void odp_mb_acquire(void);
+/**
+ * Full memory barrier
+ *
+ * This is a full memory barrier. It guarantees that all load and
store
+ * operations specified before it are visible to other threads before
+ * all load and store operations specified after it.
+ *
+ * This call is not needed when using ODP defined synchronization
mechanisms.
+ */
+void odp_mb_full(void);
/**
* @}
diff --git a/platform/linux-generic/include/odp/sync.h
b/platform/linux-generic/include/odp/sync.h
index bc73083..bfe67ee 100644
--- a/platform/linux-generic/include/odp/sync.h
+++ b/platform/linux-generic/include/odp/sync.h
@@ -17,6 +17,34 @@
extern "C" {
#endif
+/** @ingroup odp_barrier
+ * @{
+ */
+
+static inline void odp_mb_release(void)
+{
+ __atomic_thread_fence(__ATOMIC_RELEASE);
+}
+
+static inline void odp_mb_acquire(void)
+{
+ __atomic_thread_fence(__ATOMIC_ACQUIRE);
+}
+
+static inline void odp_mb_full(void)
+{
+ __atomic_thread_fence(__ATOMIC_SEQ_CST);
+}
+
+static inline void odp_sync_stores(void)
+{
+ __atomic_thread_fence(__ATOMIC_RELEASE);
+}
+
+/**
+ * @}
+ */
+
#include <odp/api/sync.h>
#ifdef __cplusplus
--
2.6.3
_______________________________________________
lng-odp mailing list
[email protected] <mailto:[email protected]>
https://lists.linaro.org/mailman/listinfo/lng-odp
_______________________________________________
lng-odp mailing list
[email protected]
https://lists.linaro.org/mailman/listinfo/lng-odp
