Skip to contentSkip to conversation

/proc output - add CPU usage stats

PATCH Applied Jan Kiszka 2 participants 6 messages Jul 7, 2006 – Jul 8, 2006

lore
Share & tools
lore

Jan Kiszka

author
#0 lore raw

PATCH 2/2 /proc output - add CPU usage stats

Yet another (hopefully the final) version of my attempt to add simple per-thread CPU usage statistics. This variant now prints the usage in percent directly via /proc/xenomai/stat.

As Philippe suggested to reorganise the /proc section for statistics, I decided to make this probably intermediate version independent of any user-space tools. You can simply type "cat /proc/xenomai/stat" as often as you like to update the stats. Example:

[email protected] :/root# cat /proc/xenomai/stat

CPU  PID    MSW        CSW        PF    STAT       %CPU  NAME
  0  0      0          5836670    0     01400080   78.8  ROOT
  0  0      0          1          0     00000082    0.0  timsPipeReceiver
  0  1029   1          1          0     00c00180    0.0  irqloop
  0  1030   1          7843       0     00c00082    4.2  irqloop
  0  1039   16         140        0     00c00082    0.1  display-1038
  0  0      0          166598     0     00000084   16.9  timerbench

Each output marks the beginning of the following measuring interval, thus when running

while true; do cat /proc/xenomai/stat; sleep 1; clear; done

you effectively get the CPU usage of the last second continuously updated.

Note that the results are not always accurate when threads are added, removed, or migrated. If numbers looks weird, just re-run and things should normalise. This is due to the fact that the measuring only takes place between context switches on each CPU, now longer explicitly on output. This appears least intrusive and SMP-safe to me. Nevertheless, I'm lacking SMP here, real testing is still required.

Jan

---
 include/nucleus/pod.h    |   38 ++++++++++++++++++++++++++++++++++++++
 include/nucleus/thread.h |    2 ++
 ksrc/nucleus/module.c    |   40 +++++++++++++++++++++++++++++++++++-----
 ksrc/nucleus/pod.c       |    8 ++++++++
 ksrc/nucleus/thread.c    |    2 ++
 5 files changed, 85 insertions(+), 5 deletions(-)
nucleus/thread.h renamed from nucleus/thread.h.orig +2 -0
Index: include/nucleus/thread.h
===================================================================
@@ -152,6 +152,8 @@ typedef struct xnthread {
 	unsigned long csw;	/* Context switches (includes
 				   secondary -> primary switches) */
 	unsigned long pf;	/* Number of page faults */
+	xnticks_t exec_time;	/* Accumulated execution time (tsc) */
+	xnticks_t exec_start;	/* Start of execution time accumulation (tsc) */
     } stat;
 #endif /* CONFIG_XENO_OPT_STATS */
 
nucleus/pod.h renamed from nucleus/pod.h.orig +38 -0
Index: include/nucleus/pod.h
===================================================================
@@ -145,6 +145,10 @@ typedef struct xnsched {
 
     xnthread_t rootcb;          /*!< Root thread control block. */
 
+#ifdef CONFIG_XENO_OPT_STATS
+    xnticks_t last_csw;         /*!< Last context switch (ticks). */
+#endif /* CONFIG_XENO_OPT_STATS */
+
 } xnsched_t;
 
 #ifdef CONFIG_SMP
@@ -545,6 +549,40 @@ static inline void xnpod_delete_self (vo
     xnpod_delete_thread(xnpod_current_thread());
 }
 
+#ifdef CONFIG_XENO_OPT_STATS
+static inline void xnpod_acc_exec_time(xnsched_t *sched, xnthread_t *thread)
+{
+    xnticks_t now = xnarch_get_cpu_tsc();
+
+    thread->stat.exec_time += now - sched->last_csw;
+    sched->last_csw = now;
+}
+
+static inline void xnpod_reset_exec_stats(xnthread_t *thread)
+{
+    thread->stat.exec_time = 0;
+    thread->stat.exec_start = xnarch_get_cpu_tsc();
+}
+
+static inline void xnpod_update_csw_date(xnsched_t *sched)
+{
+    sched->last_csw = xnarch_get_cpu_tsc();
+}
+
+#else /* !CONFIG_XENO_OPT_STATS */
+static inline void xnpod_acc_exec_time(xnsched_t *sched, xnthread_t *thread)
+{
+}
+
+static inline void xnpod_reset_exec_stats(xnthread_t *thread)
+{
+}
+
+static inline void xnpod_update_csw_date(xnsched_t *sched)
+{
+}
+#endif /* CONFIG_XENO_OPT_STATS */
+
 #ifdef __cplusplus
 }
 #endif
nucleus/pod.c renamed from nucleus/pod.c.orig +8 -0
Index: ksrc/nucleus/pod.c
===================================================================
@@ -669,6 +669,9 @@ static inline void xnpod_switch_zombie(x
 
 	xnthread_cleanup_tcb(threadout);
 
+	/* no need to update stats of dying thread */
+	xnpod_update_csw_date(sched);
+
 	xnarch_finalize_and_switch(xnthread_archtcb(threadout),
 				   xnthread_archtcb(threadin));
 
@@ -1889,6 +1892,9 @@ int xnpod_migrate_thread(int cpu)
 
 	xnpod_schedule();
 
+	/* Reset execution time stats due to unsync'ed TSCs */
+	xnpod_reset_exec_stats(thread);
+
       unlock_and_exit:
 
 	xnlock_put_irqrestore(&nklock, s);
@@ -2433,6 +2439,7 @@ void xnpod_schedule(void)
 		xnarch_enter_root(xnthread_archtcb(threadin));
 	}
 
+	xnpod_acc_exec_time(sched, threadout);
 	xnthread_inc_csw(threadin);
 
 	xnarch_switch_to(xnthread_archtcb(threadout),
@@ -2604,6 +2611,7 @@ void xnpod_schedule_runnable(xnthread_t 
 		nkpod->schedhook(runthread, XNREADY);
 #endif /* __XENO_SIM__ */
 
+	xnpod_acc_exec_time(sched, runthread);
 	xnthread_inc_csw(threadin);
 
 	xnarch_switch_to(xnthread_archtcb(runthread),
nucleus/module.c renamed from nucleus/module.c.orig +35 -5
Index: ksrc/nucleus/module.c
===================================================================
@@ -269,6 +269,8 @@ struct stat_seq_iterator {
 		unsigned long ssw;
 		unsigned long csw;
 		unsigned long pf;
+		xnticks_t exec_time;
+		xnticks_t exec_period;
 	} stat_info[1];
 };
 
@@ -309,13 +311,27 @@ static void stat_seq_stop(struct seq_fil
 static int stat_seq_show(struct seq_file *seq, void *v)
 {
 	if (v == SEQ_START_TOKEN)
-		seq_printf(seq, "%-3s  %-6s %-10s %-10s %-4s  %-8s  %s\n",
-			   "CPU", "PID", "MSW", "CSW", "PF", "STAT", "NAME");
+		seq_printf(seq, "%-3s  %-6s %-10s %-10s %-4s  %-8s  %5s"
+			   "  %s\n",
+			   "CPU", "PID", "MSW", "CSW", "PF", "STAT", "%CPU",
+			   "NAME");
 	else {
 		struct stat_seq_info *p = (struct stat_seq_info *)v;
-		seq_printf(seq, "%3u  %-6d %-10lu %-10lu %-4lu  %.8lx  %s\n",
+		int usage = 0;
+
+		if (p->exec_period) {
+			while (p->exec_period > 0xFFFFFFFF) {
+				p->exec_time >>= 16;
+				p->exec_period >>= 16;
+			}
+			usage = xnarch_ulldiv(
+				p->exec_time * 1000LL + (p->exec_period >> 1),
+				p->exec_period, NULL);
+		}
+		seq_printf(seq, "%3u  %-6d %-10lu %-10lu %-4lu  %.8lx  %3u.%u"
+			   "  %s\n",
 			   p->cpu, p->pid, p->ssw, p->csw, p->pf, p->status,
-			   p->name);
+			   usage / 10, usage % 10, p->name);
 	}
 
 	return 0;
@@ -370,6 +386,8 @@ static int stat_seq_open(struct inode *i
 
 	while (holder) {
 		xnthread_t *thread;
+		xnsched_t *sched;
+		xnticks_t period;
 		int n;
 
 		xnlock_get_irqsave(&nklock, s);
@@ -381,7 +399,8 @@ static int stat_seq_open(struct inode *i
 		thread = link2thread(holder, glink);
 		n = iter->nentries++;
 
-		iter->stat_info[n].cpu = xnsched_cpu(thread->sched);
+		sched = thread->sched;
+		iter->stat_info[n].cpu = xnsched_cpu(sched);
 		iter->stat_info[n].pid = xnthread_user_pid(thread);
 		memcpy(iter->stat_info[n].name, thread->name,
 		       sizeof(iter->stat_info[n].name));
@@ -390,6 +409,17 @@ static int stat_seq_open(struct inode *i
 		iter->stat_info[n].csw = thread->stat.csw;
 		iter->stat_info[n].pf = thread->stat.pf;
 
+		period = sched->last_csw - thread->stat.exec_start;
+		if (!period && thread == sched->runthread) {
+			iter->stat_info[n].exec_time = 1;
+			iter->stat_info[n].exec_period = 1;
+		} else {
+			iter->stat_info[n].exec_time = thread->stat.exec_time;
+			iter->stat_info[n].exec_period = period;
+		}
+		thread->stat.exec_time = 0;
+		thread->stat.exec_start = sched->last_csw;
+
 		holder = nextq(&nkpod->threadq, holder);
 
 		xnlock_put_irqrestore(&nklock, s);
nucleus/thread.c renamed from nucleus/thread.c.orig +2 -0
Index: ksrc/nucleus/thread.c
===================================================================
@@ -90,6 +90,8 @@ int xnthread_init(xnthread_t *thread,
 	thread->stat.ssw = 0;
 	thread->stat.csw = 0;
 	thread->stat.pf = 0;
+	thread->stat.exec_time = 0;
+	thread->stat.exec_start = 0;
 #endif /* CONFIG_XENO_OPT_STATS */
 
 	/* These will be filled by xnpod_start_thread() */
#1 lore raw
Quoted 3 lines · Jan Kiszka

Yet another (hopefully the final) version of my attempt to add simple per-thread CPU usage statistics. This variant now prints the usage in percent directly via /proc/xenomai/stat.

Merged, thanks.

Signature · 1 line
Philippe.
author
#2 lore raw
Quoted 7 lines · Philippe Gerum
Quoted 3 lines · Jan Kiszka

Yet another (hopefully the final) version of my attempt to add simple per-thread CPU usage statistics. This variant now prints the usage in percent directly via /proc/xenomai/stat.

Merged, thanks.

Tabs seems to got lost on merge (I think my posted patches where ok).

Jan

#3 lore raw
Quoted 10 lines · Jan Kiszka
Quoted 7 lines · Philippe Gerum
Quoted 3 lines · Jan Kiszka

Yet another (hopefully the final) version of my attempt to add simple per-thread CPU usage statistics. This variant now prints the usage in percent directly via /proc/xenomai/stat.

Merged, thanks.

Tabs seems to got lost on merge (I think my posted patches where ok).

Looking at the mail archive on GNA, it seems they were not. I've planned to rerun Lindent over the sources anyway.

Quoted 3 lines

Jan

Signature · 1 line
Philippe.
author
#4 lore raw
Quoted 6 lines · Philippe Gerum
Quoted 5 lines · Jan Kiszka
Quoted 4 lines · Philippe Gerum
Quoted 3 lines · Jan Kiszka

Yet another (hopefully the final) version of my attempt to add simple per-thread CPU usage statistics. This variant now prints the usage in percent directly via /proc/xenomai/stat.

Quoted 6 lines
Quoted 3 lines
Quoted 2 lines

Merged, thanks.

Tabs seems to got lost on merge (I think my posted patches where ok).

Looking at the mail archive on GNA, it seems they were not. I've planned to rerun Lindent over the sources anyway.

Hmm, strange. Checking the versions received via my own subscription (PATCH 1/2 and 2/2), there are tabs at lines where there are non in trunk now.

It's indeed a minor issue, but I wonder if I should continue to post patches inlined or fall back to attachments again.

Jan

#5 lore raw
Quoted 14 lines · Jan Kiszka
Quoted 6 lines · Philippe Gerum
Quoted 5 lines · Jan Kiszka
Quoted 4 lines · Philippe Gerum
Quoted 3 lines · Jan Kiszka

Yet another (hopefully the final) version of my attempt to add simple per-thread CPU usage statistics. This variant now prints the usage in percent directly via /proc/xenomai/stat.

Quoted 6 lines
Quoted 3 lines
Quoted 2 lines

Merged, thanks.

Tabs seems to got lost on merge (I think my posted patches where ok).

Looking at the mail archive on GNA, it seems they were not. I've planned to rerun Lindent over the sources anyway.

Quoted 6 lines

Hmm, strange. Checking the versions received via my own subscription (PATCH 1/2 and 2/2), there are tabs at lines where there are non in trunk now.

It's indeed a minor issue, but I wonder if I should continue to post patches inlined or fall back to attachments again.

I don't know what went wrong, but the space/Tab issue appeared specifically with your latest patches a week ago or so; so far, patching with -l solved the issue, so that's not a big deal, provided a reindentation is done afterwise to suppress unwanted spaces. I've never had any problem using your attachments verbatim, though.

Quoted 3 lines

Jan

Signature · 1 line
Philippe.