Skip to contentSkip to conversation

taskstats-fork: Add a new taskstats command to get notification on fork/clone

PATCH In review Nikanth Karthikesan 3 participants 6 messages Jul 21, 2009 – Aug 31, 2009

Subsystems (4)
lore
Share & tools
lore
#0 lore raw

PATCH 1/3 taskstats-fork: Add a new taskstats command to get notification on fork/clone

Add a new taskstats command to register for notification, whenever a new task forks in the cpumask specified.

---

include/linux/taskstats.h +3 -0
index 341dddb..1623a1f 100644
@@ -188,6 +188,7 @@ enum {
 	TASKSTATS_TYPE_STATS,		/* taskstats structure */
 	TASKSTATS_TYPE_AGGR_PID,	/* contains pid + stats */
 	TASKSTATS_TYPE_AGGR_TGID,	/* contains tgid + stats */
+	TASKSTATS_TYPE_PID_TGID,	/* contains pid + tgid */
 	__TASKSTATS_TYPE_MAX,
 };
 
@@ -199,6 +200,8 @@ enum {
 	TASKSTATS_CMD_ATTR_TGID,
 	TASKSTATS_CMD_ATTR_REGISTER_CPUMASK,
 	TASKSTATS_CMD_ATTR_DEREGISTER_CPUMASK,
+	TASKSTATS_CMD_ATTR_REGISTER_FORK_CPUMASK,
+	TASKSTATS_CMD_ATTR_DEREGISTER_FORK_CPUMASK,
 	__TASKSTATS_CMD_ATTR_MAX,
 };
 
include/linux/taskstats_kern.h +3 -0
index 7e9680f..b727370 100644
@@ -26,9 +26,12 @@ static inline void taskstats_tgid_free(struct signal_struct *sig)
 		kmem_cache_free(taskstats_cache, sig->stats);
 }
 
+extern void taskstats_fork(struct task_struct *);
 extern void taskstats_exit(struct task_struct *, int group_dead);
 extern void taskstats_init_early(void);
 #else
+static inline void taskstats_fork(struct task_struct *tsk)
+{}
 static inline void taskstats_exit(struct task_struct *tsk, int group_dead)
 {}
 static inline void taskstats_tgid_init(struct signal_struct *sig)
kernel/fork.c +1 -0
index 467746b..8ed8c9f 100644
@@ -1447,6 +1447,7 @@ long do_fork(unsigned long clone_flags,
 			freezer_count();
 			tracehook_report_vfork_done(p, nr);
 		}
+		taskstats_fork(p);
 	} else {
 		nr = PTR_ERR(p);
 	}
kernel/taskstats.c +84 -10

158 lines not on this page · raw message

#1 lore raw
Quoted 3 lines · Nikanth Karthikesan

Add a new taskstats command to register for notification, whenever a new task forks in the cpumask specified.

The changelog is sucky.. why do we need this? Why is proc-events not sufficient?

Signature · 1 line
	Balbir
#2 lore raw
Quoted 6 lines · On Tuesday 21 July 2009 11:31:14 Balbir Singh wrote:
Quoted 2 lines · Nikanth Karthikesan

Add a new taskstats command to register for notification, whenever a new task forks in the cpumask specified.

The changelog is sucky.. why do we need this? Why is proc-events not sufficient?

Ah.. proc-events was the exact thing, I was looking for! Thanks. But it seems there is no documentation for it? May be I didn't search properly...

BTW I did mention the need for this, in
http://patchwork.kernel.org/patch/36451/ , but didn't invest time to write a good changelog here, as I suspected something exists already. Sorry. ;-(

[CC-ed iotop/top people, who were CC-ed in the above mail. So that they can consider using proc-events interface, if it was not considered/used already]

Thanks
Nikanth

#3 lore raw
Quoted 1 line · Nikanth Karthikesan

Ah.. proc-events was the exact thing, I was looking for! Thanks.

AFAICT proc-events can only be used by root, so that makes it unsuitable for *top.

Anyway, iotop will always have to iterate over all threads to query their taskstats and other attributes, so the O(nr_threads) complexity is unavoidable.

Thanks.

Signature · 1 line
Guillaume
#4 lore raw
Quoted 10 lines · Guillaume Chazarain
Quoted 1 line · Nikanth Karthikesan

Ah.. proc-events was the exact thing, I was looking for! Thanks.

AFAICT proc-events can only be used by root, so that makes it unsuitable for *top.

Anyway, iotop will always have to iterate over all threads to query their taskstats and other attributes, so the O(nr_threads) complexity is unavoidable.

Unless you want the stats for tgid in signal struct. In some cases we accumulate, in some we don't. But yes irrespective or kernel/user space the complexity exists.

Signature · 1 line
	Balbir
#5 lore raw
In reply to Guillaume Chazarain
Quoted 1 line · On Monday 31 August 2009 01:22:22 Guillaume Chazarain wrote:

On Tue, Jul 21, 2009 at 8:24 AM, Nikanth Karthikesan<[email protected]>

Quoted 9 lines · wrote:
Quoted 1 line

Ah.. proc-events was the exact thing, I was looking for! Thanks.

AFAICT proc-events can only be used by root, so that makes it unsuitable for *top.

Anyway, iotop will always have to iterate over all threads to query their taskstats and other attributes, so the O(nr_threads) complexity is unavoidable.

If it would simply iterate over all the current threads, the complexity would be O(nr_threads). But while iterating through all the current threads, it checks whether it is a new thread or a known thread, so that it can calculate the delta values. This is what I optimized by using taskstats-fork, which could be optimized using proc-events.

Thanks
Nikanth