[RFC][PATCH] per-task delay accounting: avoid send without listeners

[RFC][PATCH] per-task delay accounting: avoid send without listeners

[Shailabh Nagar]

Don't send taskstats (per-pid or per-tgid) on thread exit when no one is
listening for such data.

Currently the taskstats interface allocates a structure, fills it in
and calls netlink to send out per-pid and per-tgid stats regardless of whether
a userspace listener for the data exists (netlink layer would check for that
and avoid the multicast).

As a result of this patch, the check for the no-listener case is performed
early, avoiding the redundant allocation and filling up of the taskstats
structures.

Signed-off-by: Balbir Singh <balbir@in.ibm.com>
Signed-off-by: Shailabh Nagar <nagar@watson.ibm.com>

---

 include/linux/taskstats_kern.h |   13 ++++++++++++-
 1 files changed, 12 insertions(+), 1 deletion(-)

Index: linux-2.6.17/include/linux/taskstats_kern.h
===================================================================
--- linux-2.6.17.orig/include/linux/taskstats_kern.h	2006-06-26 16:45:33.000000000 -0400
+++ linux-2.6.17/include/linux/taskstats_kern.h	2006-06-26 16:47:08.000000000 -0400
@@ -9,6 +9,7 @@

 #include <linux/taskstats.h>
 #include <linux/sched.h>
+#include <net/genetlink.h>

 enum {
 	TASKSTATS_MSG_UNICAST,		/* send data only to requester */
@@ -19,9 +20,19 @@ enum {
 extern kmem_cache_t *taskstats_cache;
 extern struct mutex taskstats_exit_mutex;

+static inline int taskstats_has_listeners(void)
+{
+	if (!genl_sock)
+		return 0;
+	return netlink_has_listeners(genl_sock, TASKSTATS_LISTEN_GROUP);
+}
+
+
 static inline void taskstats_exit_alloc(struct taskstats **ptidstats)
 {
-	*ptidstats = kmem_cache_zalloc(taskstats_cache, SLAB_KERNEL);
+	*ptidstats = NULL;
+	if (taskstats_has_listeners())
+		*ptidstats = kmem_cache_zalloc(taskstats_cache, SLAB_KERNEL);
 }

 static inline void taskstats_exit_free(struct taskstats *tidstats)

Re: [RFC][PATCH] per-task delay accounting: avoid send without listeners

[Shailabh Nagar]

Andrew,

This applies on top of the rollup of -mm (as of 5:39, 26 Jun 06), over 2.6.17,
that you have provided at
http://www.zip.com.au/~akpm/linux/patches/stuff/x.bz2

Using this and not 2.6.17-mm2 due to inclusion of other patches that aren't in -mm2.

This concludes the changes I planned to make and should hopefully address all
the overhead concerns.


Jay,

If there are any remaining overhead issues you see, pls let me know asap.

Thanks,
Shailabh

Shailabh Nagar wrote:
> Don't send taskstats (per-pid or per-tgid) on thread exit when no one is
> listening for such data.
> 
> Currently the taskstats interface allocates a structure, fills it in
> and calls netlink to send out per-pid and per-tgid stats regardless of whether
> a userspace listener for the data exists (netlink layer would check for that
> and avoid the multicast).
> 
> As a result of this patch, the check for the no-listener case is performed
> early, avoiding the redundant allocation and filling up of the taskstats
> structures.
> 
> Signed-off-by: Balbir Singh <balbir@in.ibm.com>
> Signed-off-by: Shailabh Nagar <nagar@watson.ibm.com>