The Tile platform doesn't natively support atomic operations other than
4-byte and 8-byte. Although our runtime can handle subword atomics (by
doing a word-aligned read, inserting the sub-word properly, and trying
compare-and-exchange) it is more efficient to use word-size atomics where
possible. I suspect this may also be true for other non-Intel platforms,
and certainly the top of rsyslog's runtime/atomic.h does say "THESE MACROS
MUST ONLY BE USED WITH WORD-SIZED DATA TYPES!".
The attached patch against 6.1.0 converts msg_t's iRefCount from short to
int, and moves it in the structure so that the neighboring sbool and short
types can be packed more efficiently.
--
Chris Metcalf, Tilera Corp.
http://www.tilera.com
--- rsyslog-6.1.0/runtime/msg.h.orig 2010-11-22 09:42:22.971057000 -0500
+++ rsyslog-6.1.0/runtime/msg.h 2010-11-22 09:43:19.100110000 -0500
@@ -60,8 +60,8 @@
flowControl_t flowCtlType; /**< type of flow control we can apply, for
enqueueing, needs not to be persisted because
once data has entered the queue, this
property is no longer needed. */
pthread_mutex_t mut;
+ int iRefCount; /* reference counter (0 = unused) */
sbool bDoLock; /* use the mutex? */
- short iRefCount; /* reference counter (0 = unused) */
short iSeverity; /* the severity 0..7 */
short iFacility; /* Facility code 0 .. 23*/
short offAfterPRI; /* offset, at which raw message WITHOUT PRI
part starts in pszRawMsg */
_______________________________________________
rsyslog mailing list
http://lists.adiscon.net/mailman/listinfo/rsyslog
http://www.rsyslog.com
