Skip to contentSkip to conversation

(1/3) cpufreq_ondemand - 01_ignore-nice.diff

PATCH In review Alexander Clouter 7 participants 14 messages Feb 20, 2005 – Feb 21, 2005 CPU frequency scaling framework

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 (1/3) cpufreq_ondemand - 01_ignore-nice.diff

Adds support to cpufreq_ondemand to ignore 'nice' cpu time

Signature · 8 lines
 ________________________________ 
< A couch is as good as a chair. >
 -------------------------------- 
        \   ^__^
         \  (oo)\_______
            (__)\       )\/\
                ||----w |
                ||     ||
drivers/cpufreq/cpufreq_ondemand.c +62 -4
@@ -79,6 +79,7 @@
 	unsigned int		sampling_down_factor;
 	unsigned int		up_threshold;
 	unsigned int		down_threshold;
+	unsigned int		ignore_nice;
 };
 
 static struct dbs_tuners dbs_tuners_ins = {
@@ -116,6 +117,7 @@
 show_one(sampling_down_factor, sampling_down_factor);
 show_one(up_threshold, up_threshold);
 show_one(down_threshold, down_threshold);
+show_one(ignore_nice, ignore_nice);
 
 static ssize_t store_sampling_down_factor(struct cpufreq_policy *unused, 
 		const char *buf, size_t count)
@@ -194,6 +196,46 @@
 	return count;
 }
 
+static ssize_t store_ignore_nice(struct cpufreq_policy *policy,
+		const char *buf, size_t count)
+{
+	unsigned int input;
+	int ret;
+
+	unsigned int j;
+	
+	ret = sscanf (buf, "%u", &input);
+	if ( ret != 1 )
+		return -EINVAL;
+
+	if ( input > 1 )
+		input = 1;
+	
+	down(&dbs_sem);
+	if ( input == dbs_tuners_ins.ignore_nice ) { /* nothing to do */
+		up(&dbs_sem);
+		return count;
+	}
+	dbs_tuners_ins.ignore_nice = input;
+
+	/* we need to re-evaluate prev_cpu_idle_up and prev_cpu_idle_down */
+	for_each_cpu_mask(j, policy->cpus) {
+		struct cpu_dbs_info_s *j_dbs_info;
+		j_dbs_info = &per_cpu(cpu_dbs_info, j);
+		j_dbs_info->cur_policy = policy;
+
+		j_dbs_info->prev_cpu_idle_up =
+			kstat_cpu(j).cpustat.idle +
+			kstat_cpu(j).cpustat.iowait +
+			( !dbs_tuners_ins.ignore_nice
+			  ? kstat_cpu(j).cpustat.nice : NULL );
+		j_dbs_info->prev_cpu_idle_down = j_dbs_info->prev_cpu_idle_up;
+	}
+	up(&dbs_sem);
+
+	return count;
+}
+
 #define define_one_rw(_name) \
 static struct freq_attr _name = \
 __ATTR(_name, 0644, show_##_name, store_##_name)
@@ -202,6 +244,7 @@
 define_one_rw(sampling_down_factor);
 define_one_rw(up_threshold);
 define_one_rw(down_threshold);
+define_one_rw(ignore_nice);
 
 static struct attribute * dbs_attributes[] = {
 	&sampling_rate_max.attr,
@@ -210,6 +253,7 @@
 	&sampling_down_factor.attr,
 	&up_threshold.attr,
 	&down_threshold.attr,
+	&ignore_nice.attr,
 	NULL
 };
 
@@ -254,6 +298,9 @@
 	/* Check for frequency increase */
 	total_idle_ticks = kstat_cpu(cpu).cpustat.idle +
 		kstat_cpu(cpu).cpustat.iowait;
+	  /* consider 'nice' tasks as 'idle' time too if required */
+	  if (dbs_tuners_ins.ignore_nice == 0)
+		total_idle_ticks += kstat_cpu(cpu).cpustat.nice;
 	idle_ticks = total_idle_ticks -
 		this_dbs_info->prev_cpu_idle_up;
 	this_dbs_info->prev_cpu_idle_up = total_idle_ticks;
@@ -270,6 +317,9 @@
 		/* Check for frequency increase */
 		total_idle_ticks = kstat_cpu(j).cpustat.idle +
 			kstat_cpu(j).cpustat.iowait;
+		  /* consider 'nice' too? */
+		  if (dbs_tuners_ins.ignore_nice == 0)
+			   total_idle_ticks += kstat_cpu(j).cpustat.nice;
 		tmp_idle_ticks = total_idle_ticks -
 			j_dbs_info->prev_cpu_idle_up;
 		j_dbs_info->prev_cpu_idle_up = total_idle_ticks;
@@ -298,6 +348,9 @@
 
 	total_idle_ticks = kstat_cpu(cpu).cpustat.idle +
 		kstat_cpu(cpu).cpustat.iowait;
+	  /* consider 'nice' too? */
+	  if (dbs_tuners_ins.ignore_nice == 0)
+		  total_idle_ticks += kstat_cpu(cpu).cpustat.nice;
 	idle_ticks = total_idle_ticks -
 		this_dbs_info->prev_cpu_idle_down;
 	this_dbs_info->prev_cpu_idle_down = total_idle_ticks;
@@ -313,6 +366,9 @@
 		/* Check for frequency increase */
 		total_idle_ticks = kstat_cpu(j).cpustat.idle +
 			kstat_cpu(j).cpustat.iowait;
+		  /* consider 'nice' too? */
+		  if (dbs_tuners_ins.ignore_nice == 0)
+			total_idle_ticks += kstat_cpu(j).cpustat.nice;
 		tmp_idle_ticks = total_idle_ticks -
 			j_dbs_info->prev_cpu_idle_down;
 		j_dbs_info->prev_cpu_idle_down = total_idle_ticks;
@@ -399,10 +455,11 @@
 		
 			j_dbs_info->prev_cpu_idle_up = 
 				kstat_cpu(j).cpustat.idle +
-				kstat_cpu(j).cpustat.iowait;
-			j_dbs_info->prev_cpu_idle_down = 
-				kstat_cpu(j).cpustat.idle +
-				kstat_cpu(j).cpustat.iowait;
+				kstat_cpu(j).cpustat.iowait +
+				( !dbs_tuners_ins.ignore_nice
+				  ? kstat_cpu(j).cpustat.nice : NULL );
+			j_dbs_info->prev_cpu_idle_down
+				= j_dbs_info->prev_cpu_idle_up;
 		}
 		this_dbs_info->enable = 1;
 		sysfs_create_group(&policy->kobj, &dbs_attr_group);
@@ -422,6 +479,7 @@
 			def_sampling_rate = (latency / 1000) *
 					DEF_SAMPLING_RATE_LATENCY_MULTIPLIER;
 			dbs_tuners_ins.sampling_rate = def_sampling_rate;
+			dbs_tuners_ins.ignore_nice = 0;
 
 			dbs_timer_init();
 		}

_______________________________________________
Cpufreq mailing list
[email protected]
http://lists.linux.org.uk/mailman/listinfo/cpufreq

#1 lore raw
Quoted 1 line · Alexander Clouter

Adds support to cpufreq_ondemand to ignore 'nice' cpu time

this increasingly starts to smell like it wants this policy in userspace.....

#2 lore raw
Quoted 5 lines · Arjan van de Ven
Quoted 1 line · Alexander Clouter

Adds support to cpufreq_ondemand to ignore 'nice' cpu time

this increasingly starts to smell like it wants this policy in userspace.....

There are enough userspace cpufreq daemons already. However, the policy decision back then was to prefer governing CPU frequencies in the following order:
a) hardware
b) kernelspace
c) userspace

http://marc.theaimsgroup.com/?t=103053598900004&r=2&w=2

and two decisive posts:

http://marc.theaimsgroup.com/?l=linux-kernel&m=103056055008566&w=2
http://marc.theaimsgroup.com/?l=linux-kernel&m=103056413012458&w=2

Thanks,
Dominik

#3 lore raw
In reply to Alexander Clouter

Review

Quoted 2 lines · Alexander Clouter
(quoted diff) +2 -0
+			( !dbs_tuners_ins.ignore_nice
+			  ? kstat_cpu(j).cpustat.nice : NULL );

NULL is a pointer, 0 is a number. You most probably want "0" here.

Review

Quoted 2 lines
(quoted diff) +2 -0
+				( !dbs_tuners_ins.ignore_nice
+				  ? kstat_cpu(j).cpustat.nice : NULL );

dito.

Dominik

#4 lore raw

Morning,

Quoted 9 lines · Dominik Brodowski

Review

Quoted 2 lines · Alexander Clouter
(quoted diff) +2 -0
+			( !dbs_tuners_ins.ignore_nice
+			  ? kstat_cpu(j).cpustat.nice : NULL );

NULL is a pointer, 0 is a number. You most probably want "0" here.

Review

Quoted 2 lines
(quoted diff) +2 -0
+				( !dbs_tuners_ins.ignore_nice
+				  ? kstat_cpu(j).cpustat.nice : NULL );

dito.

NULL comes out as a nothing, there is no problem in adding NULL to a variable; the c-compiler, I have been assured, drops this out (or rather it should) as a NOOP. Of course this is what I have been told, by my flatmate whom works with a large number of different compilers for different architectures from different software houses, and lives/breathes C (concerningly :)

Along similar lines, I was told that initialising statics to zero *must* be done and not assumed; friends have told me its a 'gcc specific extension'. Last time I was told this was not necessary...

I'm happy for it to be changed on reasons of clarity, but technically there should be no problems. You're the kernel guys and I of course am wielding my "talking out my ass" disclaimer, so correct me where necessary :)

Cheers

Alex

Signature · 11 lines
 _________________________________________ 
/  A horse breeder has his young colts    \
| bottle-fed after they're three days     |
| old. He heard that a foal and his mummy |
\ are soon parted.                        /
 ----------------------------------------- 
        \   ^__^
         \  (oo)\_______
            (__)\       )\/\
                ||----w |
                ||     ||

_______________________________________________
Cpufreq mailing list
[email protected]
http://lists.linux.org.uk/mailman/listinfo/cpufreq

#5 lore raw

Le Dimanche 20 Février 2005 19.14, Alexander Clouter a écrit :
(snip)

Quoted 3 lines

Along similar lines, I was told that initialising statics to zero *must* be done and not assumed; friends have told me its a 'gcc specific extension'. Last time I was told this was not necessary...

I'm really not sure it is a "gcc specific extension", re-read paragraph 6.7.8 from the ISO 9899:1999 norm, especially point 10. It explicitly say :

10 If an object that has automatic storage duration is not initialized
explicitly, its value is indeterminate. If an object that has static storage duration is not initialized explicitly, then:
- if it has pointer type, it is initialized to a null pointer;
- if it has arithmetic type, it is initialized to (positive or unsigned

      zero;
   - if it is an aggregate, every member is initialized (recursively)
     according to these rules;

- if it is a union, the first named member is initialized (recursively) according to these rules.

Regards,
Pcs

#6 lore raw
In reply to Alexander Clouter
Quoted 12 lines · Alexander Clouter

Morning,

Quoted 9 lines · Dominik Brodowski

Review

Quoted 2 lines · Alexander Clouter
(quoted diff) +2 -0
+			( !dbs_tuners_ins.ignore_nice
+			  ? kstat_cpu(j).cpustat.nice : NULL );

NULL is a pointer, 0 is a number. You most probably want "0" here.

Review

Quoted 2 lines
(quoted diff) +2 -0
+				( !dbs_tuners_ins.ignore_nice
+				  ? kstat_cpu(j).cpustat.nice : NULL );

dito.

Quoted 6 lines

NULL comes out as a nothing, there is no problem in adding NULL to a variable; the c-compiler, I have been assured, drops this out (or rather it should) as a NOOP. Of course this is what I have been told, by my flatmate whom works with a large number of different compilers for different architectures from different software houses, and lives/breathes C (concerningly :)

Please don't assume anything about NULL in kernelspace. It's defined in a header (stddef.h for C99) and standard headers can't be included in kernel space. NULL is defined elsewhere for kernel, but that's another story.

Anyway, you are mixing pointers and integers types. These lead *always* to bugs, especially if you must change to another HW architecture (and sometimes another software environment like DOS).

Quoted 3 lines

Along similar lines, I was told that initialising statics to zero *must* be done and not assumed; friends have told me its a 'gcc specific extension'. Last time I was told this was not necessary...

I don't get it. I assume statics variables being always initialized as said by a previous answer even on pre-ansi C compilers.

Cheers,

Signature · 4 lines
Bruno Ducrot

--  Which is worse:  ignorance or apathy?
--  Don't know.  Don't care.
#7 lore raw
In reply to Alexander Clouter

Alexander Clouter a écrit :

Quoted 2 lines

Adds support to cpufreq_ondemand to ignore 'nice' cpu time

IMO, this policy should be has lightweight as possible. Why not making this feature just hard-coded? Don't count the load that nice'd tasks do, period!

It would make the patch 5 lines instead of 50. Every userspace deamon have this feature enabled by default, so this is a very reasonable trade-off.

Eric

#8 lore raw
Quoted 8 lines · Eric Piel
Quoted 2 lines · Alexander Clouter

Adds support to cpufreq_ondemand to ignore 'nice' cpu time

IMO, this policy should be has lightweight as possible. Why not making this feature just hard-coded? Don't count the load that nice'd tasks do, period!

It would make the patch 5 lines instead of 50. Every userspace deamon have this feature enabled by default, so this is a very reasonable trade-off.

Here is the patch, it does exactly the same functionality but it's completly static and therefore much smaller. Just tested on my computer, it does the trick.

Eric
--
Avoid the cpufreq ondemand governor to count the nice'd tasks as load for the cpu.

 cpufreq_ondemand.c |   17 ++++++++++-------
 1 files changed, 10 insertions(+), 7 deletions(-)

--

drivers/cpufreq/cpufreq_ondemand.c renamed from drivers/cpufreq/cpufreq_ondemand.c.bak +10 -7
@@ -253,7 +253,8 @@ static void dbs_check_cpu(int cpu)
 
 	/* Check for frequency increase */
 	total_idle_ticks = kstat_cpu(cpu).cpustat.idle +
-		kstat_cpu(cpu).cpustat.iowait;
+		kstat_cpu(cpu).cpustat.iowait +
+		kstat_cpu(cpu).cpustat.nice;
 	idle_ticks = total_idle_ticks -
 		this_dbs_info->prev_cpu_idle_up;
 	this_dbs_info->prev_cpu_idle_up = total_idle_ticks;
@@ -269,7 +270,8 @@ static void dbs_check_cpu(int cpu)
 		j_dbs_info = &per_cpu(cpu_dbs_info, j);
 		/* Check for frequency increase */
 		total_idle_ticks = kstat_cpu(j).cpustat.idle +
-			kstat_cpu(j).cpustat.iowait;
+			kstat_cpu(j).cpustat.iowait +
+			kstat_cpu(j).cpustat.nice;
 		tmp_idle_ticks = total_idle_ticks -
 			j_dbs_info->prev_cpu_idle_up;
 		j_dbs_info->prev_cpu_idle_up = total_idle_ticks;
@@ -297,7 +299,8 @@ static void dbs_check_cpu(int cpu)
 		return;
 
 	total_idle_ticks = kstat_cpu(cpu).cpustat.idle +
-		kstat_cpu(cpu).cpustat.iowait;
+		kstat_cpu(cpu).cpustat.iowait +
+		kstat_cpu(cpu).cpustat.nice;
 	idle_ticks = total_idle_ticks -
 		this_dbs_info->prev_cpu_idle_down;
 	this_dbs_info->prev_cpu_idle_down = total_idle_ticks;
@@ -312,7 +315,8 @@ static void dbs_check_cpu(int cpu)
 		j_dbs_info = &per_cpu(cpu_dbs_info, j);
 		/* Check for frequency increase */
 		total_idle_ticks = kstat_cpu(j).cpustat.idle +
-			kstat_cpu(j).cpustat.iowait;
+			kstat_cpu(j).cpustat.iowait +
+			kstat_cpu(j).cpustat.nice;
 		tmp_idle_ticks = total_idle_ticks -
 			j_dbs_info->prev_cpu_idle_down;
 		j_dbs_info->prev_cpu_idle_down = total_idle_ticks;
@@ -398,11 +402,10 @@ static int cpufreq_governor_dbs(struct c
 			j_dbs_info->cur_policy = policy;
 		
 			j_dbs_info->prev_cpu_idle_up = 
-				kstat_cpu(j).cpustat.idle +
-				kstat_cpu(j).cpustat.iowait;
 			j_dbs_info->prev_cpu_idle_down = 
 				kstat_cpu(j).cpustat.idle +
-				kstat_cpu(j).cpustat.iowait;
+				kstat_cpu(j).cpustat.iowait +
+				kstat_cpu(j).cpustat.nice;
 		}
 		this_dbs_info->enable = 1;
 		sysfs_create_group(&policy->kobj, &dbs_attr_group);

_______________________________________________
Cpufreq mailing list
[email protected]
http://lists.linux.org.uk/mailman/listinfo/cpufreq

#9 lore raw
Quoted 5 lines · Eric Piel

Here is the patch, it does exactly the same functionality but it's completly static and therefore much smaller. Just tested on my computer, it does the trick.

I'm happy for this to go ahead as I will want it on always by default, however you pretty much kill the folk who prefer to 'nice' tasks to prevent interuption to other applications rather than using it to say "I do not care how long this task takes".

I know on our proxy server there are a few rebuilds which we run at a nice'd level, not because we do not care how long they take (we want the rebuilds to finish as soon as possible) but rather so that squid's performance is not affected.

Cheers

Alex

Quoted 10 lines

Eric
--
Avoid the cpufreq ondemand governor to count the nice'd tasks as load for the cpu.

 cpufreq_ondemand.c |   17 ++++++++++-------
 1 files changed, 10 insertions(+), 7 deletions(-)

--

Signature · 9 lines
 _______________________________________ 
/ You can drive a horse to water, but a \
\ pencil must be lead.                  /
 --------------------------------------- 
        \   ^__^
         \  (oo)\_______
            (__)\       )\/\
                ||----w |
                ||     ||

_______________________________________________
Cpufreq mailing list
[email protected]
http://lists.linux.org.uk/mailman/listinfo/cpufreq

#10 lore raw

Alexander Clouter a écrit :

Quoted 5 lines · Eric Piel
Quoted 3 lines

Here is the patch, it does exactly the same functionality but it's completly static and therefore much smaller. Just tested on my computer, it does the trick.

Quoted 7 lines

I'm happy for this to go ahead as I will want it on always by default, however you pretty much kill the folk who prefer to 'nice' tasks to prevent interuption to other applications rather than using it to say "I do not care how long this task takes".

I know on our proxy server there are a few rebuilds which we run at a nice'd level, not because we do not care how long they take (we want the rebuilds to finish as soon as possible) but rather so that squid's performance is not affected.

That's very unusual cases. You cannot take care of all unusual case with a governor that tries to be as simple as possible. Probably in the case you cite either you want to use a userspace governor which can be much more complex and sofisticated or maybe just:
renice -5 $(pidof) squid

Eric

#11 lore raw
In reply to Eric Piel
Quoted 9 lines · Eric Piel
Quoted 8 lines · Eric Piel
Quoted 2 lines · Alexander Clouter

Adds support to cpufreq_ondemand to ignore 'nice' cpu time

IMO, this policy should be has lightweight as possible. Why not making this feature just hard-coded? Don't count the load that nice'd tasks do, period!

Quoted 7 lines
Quoted 4 lines

It would make the patch 5 lines instead of 50. Every userspace deamon have this feature enabled by default, so this is a very reasonable trade-off.

Here is the patch, it does exactly the same functionality but it's completly static and therefore much smaller. Just tested on my computer, it does the trick.

IMO it was a wise decision to do this as a tuneable parameter.

Dominik

#12 lore raw
In reply to Eric Piel
Quoted 3 lines · Eric Piel

It would make the patch 5 lines instead of 50. Every userspace deamon have this feature enabled by default, so this is a very reasonable trade-off.

I'm sure that you cannot imagine how much beating i have gotten for making CONSIDER_NICE=no the default for the powersaved on SUSE 9.1... :-) It took man-weeks of debating what could be a sensible default for this. This has to be tweakable.

Signature · 1 line
Stefan Seyfried
#13 lore raw

Stefan Seyfried a écrit :

Quoted 10 lines · Eric Piel
Quoted 3 lines

It would make the patch 5 lines instead of 50. Every userspace deamon have this feature enabled by default, so this is a very reasonable trade-off.

I'm sure that you cannot imagine how much beating i have gotten for making CONSIDER_NICE=no the default for the powersaved on SUSE 9.1... :-) It took man-weeks of debating what could be a sensible default for this. This has to be tweakable.

Well, my point was that if people want different things from the simple and normal behaviour, they use a userspace deamon... for instance powersaved .

Eric