public inbox for linux-ia64@vger.kernel.org
 help / color / mirror / Atom feed
* [PATCH] SGI Altix IA64 mmtimer: Eliminate long interval timer
@ 2010-10-07 14:39 Dimitri Sivanich
  2010-10-07 22:42 ` Andrew Morton
  2010-10-08 13:23 ` Dimitri Sivanich
  0 siblings, 2 replies; 3+ messages in thread
From: Dimitri Sivanich @ 2010-10-07 14:39 UTC (permalink / raw)
  To: linux-ia64

This patch for SGI Altix/IA64 eliminates interval long timer holdoffs in
cases where we don't start an interval timer before the expiration time.
This sometimes happens when a number of interval timers on the same shub
with the same interval run simultaneously.

Signed-off-by: Dimitri Sivanich <sivanich@sgi.com>

---
 drivers/char/mmtimer.c |   55 ++++++++++++++++++++++-----------------------
 1 file changed, 27 insertions(+), 28 deletions(-)

Index: linux-2.6/drivers/char/mmtimer.c
=================================--- linux-2.6.orig/drivers/char/mmtimer.c	2010-10-07 09:34:29.837725250 -0500
+++ linux-2.6/drivers/char/mmtimer.c	2010-10-07 09:34:31.609948769 -0500
@@ -174,9 +174,8 @@ static void mmtimer_setup_int_2(int cpu,
  * in order to insure that the setup succeeds in a deterministic time frame.
  * It will check if the interrupt setup succeeded.
  */
-static int mmtimer_setup(int cpu, int comparator, unsigned long expires)
+static int mmtimer_setup(int cpu, int comparator, unsigned long expires, u64 *r)
 {
-
 	switch (comparator) {
 	case 0:
 		mmtimer_setup_int_0(cpu, expires);
@@ -189,7 +188,8 @@ static int mmtimer_setup(int cpu, int co
 		break;
 	}
 	/* We might've missed our expiration time */
-	if (rtc_time() <= expires)
+	*r = rtc_time();
+	if (*r <= expires)
 		return 1;
 
 	/*
@@ -225,6 +225,8 @@ static int mmtimer_disable_int(long nasi
 #define TIMER_OFF	0xbadcabLL	/* Timer is not setup */
 #define TIMER_SET	0		/* Comparator is set for this timer */
 
+#define MMTIMER_INTRVL_RETRY_INCR_DEFAULT 40
+
 /* There is one of these for each timer */
 struct mmtimer {
 	struct rb_node list;
@@ -240,6 +242,10 @@ struct mmtimer_node {
 };
 static struct mmtimer_node *timers;
 
+static unsigned mmtimer_intrvl_retry_incr = MMTIMER_INTRVL_RETRY_INCR_DEFAULT;
+module_param(mmtimer_intrvl_retry_incr, uint, 0644);
+MODULE_PARM_DESC(mmtimer_intrvl_retry_incr,
+	"RTC ticks to add to expiration on interval retry (default 40)");
 
 /*
  * Add a new mmtimer struct to the node's mmtimer list.
@@ -287,7 +293,8 @@ static void mmtimer_set_next_timer(int n
 	struct mmtimer_node *n = &timers[nodeid];
 	struct mmtimer *x;
 	struct k_itimer *t;
-	int o;
+	u64 expires, exp, r;
+	int i;
 
 restart:
 	if (n->next = NULL)
@@ -298,7 +305,7 @@ restart:
 	if (!t->it.mmtimer.incr) {
 		/* Not an interval timer */
 		if (!mmtimer_setup(x->cpu, COMPARATOR,
-					t->it.mmtimer.expires)) {
+					t->it.mmtimer.expires, &r)) {
 			/* Late setup, fire now */
 			tasklet_schedule(&n->tasklet);
 		}
@@ -306,14 +313,21 @@ restart:
 	}
 
 	/* Interval timer */
-	o = 0;
-	while (!mmtimer_setup(x->cpu, COMPARATOR, t->it.mmtimer.expires)) {
-		unsigned long e, e1;
-		struct rb_node *next;
-		t->it.mmtimer.expires += t->it.mmtimer.incr << o;
-		t->it_overrun += 1 << o;
-		o++;
-		if (o > 20) {
+	i = 0;
+	expires = exp = t->it.mmtimer.expires;
+	while (!mmtimer_setup(x->cpu, COMPARATOR, expires, &r)) {
+		int to;
+
+		i++;
+		expires = r + mmtimer_intrvl_retry_incr + (1 << i);
+		/* Calculate overruns as we go. */
+		to = ((u64)(expires - exp) / t->it.mmtimer.incr);
+		if (to) {
+			t->it_overrun += to;
+			t->it.mmtimer.expires += t->it.mmtimer.incr * to;
+			exp = t->it.mmtimer.expires;
+		}
+		if (i > 20) {
 			printk(KERN_ALERT "mmtimer: cannot reschedule timer\n");
 			t->it.mmtimer.clock = TIMER_OFF;
 			n->next = rb_next(&x->list);
@@ -321,21 +335,6 @@ restart:
 			kfree(x);
 			goto restart;
 		}
-
-		e = t->it.mmtimer.expires;
-		next = rb_next(&x->list);
-
-		if (next = NULL)
-			continue;
-
-		e1 = rb_entry(next, struct mmtimer, list)->
-			timer->it.mmtimer.expires;
-		if (e > e1) {
-			n->next = next;
-			rb_erase(&x->list, &n->timer_head);
-			mmtimer_add_list(x);
-			goto restart;
-		}
 	}
 }
 

^ permalink raw reply	[flat|nested] 3+ messages in thread

* Re: [PATCH] SGI Altix IA64 mmtimer: Eliminate long interval timer
  2010-10-07 14:39 [PATCH] SGI Altix IA64 mmtimer: Eliminate long interval timer Dimitri Sivanich
@ 2010-10-07 22:42 ` Andrew Morton
  2010-10-08 13:23 ` Dimitri Sivanich
  1 sibling, 0 replies; 3+ messages in thread
From: Andrew Morton @ 2010-10-07 22:42 UTC (permalink / raw)
  To: linux-ia64

On Thu, 7 Oct 2010 09:39:41 -0500
Dimitri Sivanich <sivanich@sgi.com> wrote:

> This patch for SGI Altix/IA64 eliminates interval long timer holdoffs in
> cases where we don't start an interval timer before the expiration time.
> This sometimes happens when a number of interval timers on the same shub
> with the same interval run simultaneously.
> 

Grumble.

> --- linux-2.6.orig/drivers/char/mmtimer.c	2010-10-07 09:34:29.837725250 -0500
> +++ linux-2.6/drivers/char/mmtimer.c	2010-10-07 09:34:31.609948769 -0500
> @@ -174,9 +174,8 @@ static void mmtimer_setup_int_2(int cpu,
>   * in order to insure that the setup succeeds in a deterministic time frame.
>   * It will check if the interrupt setup succeeded.
>   */
> -static int mmtimer_setup(int cpu, int comparator, unsigned long expires)
> +static int mmtimer_setup(int cpu, int comparator, unsigned long expires, u64 *r)

Choosing a variable's identifier is your last chance to explain to the
reader what that variable does.  And you blew it!

>  
> ...
>
> +static unsigned mmtimer_intrvl_retry_incr = MMTIMER_INTRVL_RETRY_INCR_DEFAULT;
> +module_param(mmtimer_intrvl_retry_incr, uint, 0644);
> +MODULE_PARM_DESC(mmtimer_intrvl_retry_incr,
> +	"RTC ticks to add to expiration on interval retry (default 40)");

And the problem with abbreviations like "intrvl" is that they're hard
to remember.  Which is why in the kernel we much prefer to use the
entire English word.


I'm unable to work out how important this patch is.  I'm presently
assuming 2.6.37.

^ permalink raw reply	[flat|nested] 3+ messages in thread

* Re: [PATCH] SGI Altix IA64 mmtimer: Eliminate long interval timer
  2010-10-07 14:39 [PATCH] SGI Altix IA64 mmtimer: Eliminate long interval timer Dimitri Sivanich
  2010-10-07 22:42 ` Andrew Morton
@ 2010-10-08 13:23 ` Dimitri Sivanich
  1 sibling, 0 replies; 3+ messages in thread
From: Dimitri Sivanich @ 2010-10-08 13:23 UTC (permalink / raw)
  To: linux-ia64

On Thu, Oct 07, 2010 at 03:42:56PM -0700, Andrew Morton wrote:
> On Thu, 7 Oct 2010 09:39:41 -0500
> Dimitri Sivanich <sivanich@sgi.com> wrote:
> 
> > This patch for SGI Altix/IA64 eliminates interval long timer holdoffs in
> > cases where we don't start an interval timer before the expiration time.
> > This sometimes happens when a number of interval timers on the same shub
> > with the same interval run simultaneously.
> > 
> 
> Grumble.

Ditto.

> 
> > --- linux-2.6.orig/drivers/char/mmtimer.c	2010-10-07 09:34:29.837725250 -0500
> > +++ linux-2.6/drivers/char/mmtimer.c	2010-10-07 09:34:31.609948769 -0500
> > @@ -174,9 +174,8 @@ static void mmtimer_setup_int_2(int cpu,
> >   * in order to insure that the setup succeeds in a deterministic time frame.
> >   * It will check if the interrupt setup succeeded.
> >   */
> > -static int mmtimer_setup(int cpu, int comparator, unsigned long expires)
> > +static int mmtimer_setup(int cpu, int comparator, unsigned long expires, u64 *r)
> 
> Choosing a variable's identifier is your last chance to explain to the
> reader what that variable does.  And you blew it!

See the variable identifier in the attached patch.  Better?

> 
> >  
> > ...
> >
> > +static unsigned mmtimer_intrvl_retry_incr = MMTIMER_INTRVL_RETRY_INCR_DEFAULT;
> > +module_param(mmtimer_intrvl_retry_incr, uint, 0644);
> > +MODULE_PARM_DESC(mmtimer_intrvl_retry_incr,
> > +	"RTC ticks to add to expiration on interval retry (default 40)");
> 
> And the problem with abbreviations like "intrvl" is that they're hard
> to remember.  Which is why in the kernel we much prefer to use the
> entire English word.
> 

Again, see the variable names in the attached patch.

> 
> I'm unable to work out how important this patch is.  I'm presently
> assuming 2.6.37.

Good enough.
__________________


This patch for SGI Altix/IA64 eliminates interval long timer holdoffs in
cases where we don't start an interval timer before the expiration time.
This sometimes happens when a number of interval timers on the same shub
with the same interval run simultaneously.

Signed-off-by: Dimitri Sivanich <sivanich@sgi.com>

---
 drivers/char/mmtimer.c |   60 ++++++++++++++++++++++++---------------------
 1 file changed, 32 insertions(+), 28 deletions(-)

Index: linux-2.6/drivers/char/mmtimer.c
=================================--- linux-2.6.orig/drivers/char/mmtimer.c	2010-10-08 08:14:45.971073581 -0500
+++ linux-2.6/drivers/char/mmtimer.c	2010-10-08 08:14:49.027447493 -0500
@@ -174,9 +174,9 @@ static void mmtimer_setup_int_2(int cpu,
  * in order to insure that the setup succeeds in a deterministic time frame.
  * It will check if the interrupt setup succeeded.
  */
-static int mmtimer_setup(int cpu, int comparator, unsigned long expires)
+static int mmtimer_setup(int cpu, int comparator, unsigned long expires,
+	u64 *set_completion_time)
 {
-
 	switch (comparator) {
 	case 0:
 		mmtimer_setup_int_0(cpu, expires);
@@ -189,7 +189,8 @@ static int mmtimer_setup(int cpu, int co
 		break;
 	}
 	/* We might've missed our expiration time */
-	if (rtc_time() <= expires)
+	*set_completion_time = rtc_time();
+	if (*set_completion_time <= expires)
 		return 1;
 
 	/*
@@ -225,6 +226,8 @@ static int mmtimer_disable_int(long nasi
 #define TIMER_OFF	0xbadcabLL	/* Timer is not setup */
 #define TIMER_SET	0		/* Comparator is set for this timer */
 
+#define MMTIMER_INTERVAL_RETRY_INCREMENT_DEFAULT 40
+
 /* There is one of these for each timer */
 struct mmtimer {
 	struct rb_node list;
@@ -240,6 +243,11 @@ struct mmtimer_node {
 };
 static struct mmtimer_node *timers;
 
+static unsigned mmtimer_interval_retry_increment +	MMTIMER_INTERVAL_RETRY_INCREMENT_DEFAULT;
+module_param(mmtimer_interval_retry_increment, uint, 0644);
+MODULE_PARM_DESC(mmtimer_interval_retry_increment,
+	"RTC ticks to add to expiration on interval retry (default 40)");
 
 /*
  * Add a new mmtimer struct to the node's mmtimer list.
@@ -287,7 +295,8 @@ static void mmtimer_set_next_timer(int n
 	struct mmtimer_node *n = &timers[nodeid];
 	struct mmtimer *x;
 	struct k_itimer *t;
-	int o;
+	u64 expires, exp, set_completion_time;
+	int i;
 
 restart:
 	if (n->next = NULL)
@@ -298,7 +307,8 @@ restart:
 	if (!t->it.mmtimer.incr) {
 		/* Not an interval timer */
 		if (!mmtimer_setup(x->cpu, COMPARATOR,
-					t->it.mmtimer.expires)) {
+					t->it.mmtimer.expires,
+					&set_completion_time)) {
 			/* Late setup, fire now */
 			tasklet_schedule(&n->tasklet);
 		}
@@ -306,14 +316,23 @@ restart:
 	}
 
 	/* Interval timer */
-	o = 0;
-	while (!mmtimer_setup(x->cpu, COMPARATOR, t->it.mmtimer.expires)) {
-		unsigned long e, e1;
-		struct rb_node *next;
-		t->it.mmtimer.expires += t->it.mmtimer.incr << o;
-		t->it_overrun += 1 << o;
-		o++;
-		if (o > 20) {
+	i = 0;
+	expires = exp = t->it.mmtimer.expires;
+	while (!mmtimer_setup(x->cpu, COMPARATOR, expires,
+				&set_completion_time)) {
+		int to;
+
+		i++;
+		expires = set_completion_time +
+				mmtimer_interval_retry_increment + (1 << i);
+		/* Calculate overruns as we go. */
+		to = ((u64)(expires - exp) / t->it.mmtimer.incr);
+		if (to) {
+			t->it_overrun += to;
+			t->it.mmtimer.expires += t->it.mmtimer.incr * to;
+			exp = t->it.mmtimer.expires;
+		}
+		if (i > 20) {
 			printk(KERN_ALERT "mmtimer: cannot reschedule timer\n");
 			t->it.mmtimer.clock = TIMER_OFF;
 			n->next = rb_next(&x->list);
@@ -321,21 +340,6 @@ restart:
 			kfree(x);
 			goto restart;
 		}
-
-		e = t->it.mmtimer.expires;
-		next = rb_next(&x->list);
-
-		if (next = NULL)
-			continue;
-
-		e1 = rb_entry(next, struct mmtimer, list)->
-			timer->it.mmtimer.expires;
-		if (e > e1) {
-			n->next = next;
-			rb_erase(&x->list, &n->timer_head);
-			mmtimer_add_list(x);
-			goto restart;
-		}
 	}
 }
 

^ permalink raw reply	[flat|nested] 3+ messages in thread

end of thread, other threads:[~2010-10-08 13:23 UTC | newest]

Thread overview: 3+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2010-10-07 14:39 [PATCH] SGI Altix IA64 mmtimer: Eliminate long interval timer Dimitri Sivanich
2010-10-07 22:42 ` Andrew Morton
2010-10-08 13:23 ` Dimitri Sivanich

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox