thr3ads.net - Xen devel - [Xen-devel] [PATCH][cpufreq] Xen support for the ondemand governor [2/2] (linux) [Oct 2007]

If this information is useful, please help other people find it:
Share via:

Mark Langsdorf

2007-Oct-23 22:00 UTC

[Xen-devel] [PATCH][cpufreq] Xen support for the ondemand governor [2/2] (linux)

Modify the cpufreq ondemand governor so that it can get idle and
total ticks from the Xen hypervisor.  Linux and Xen have different
ideas of what an idle tick is, so the Xen values for both have to
be returned in the same platform hypercall.  Otherwise, use 
basically the same scheme as native Linux.

Signed-off-by: Mark Langsdorf <mark.langsdorf@amd.com>

iff -r 9bf1ddd0f6bf drivers/cpufreq/cpufreq_ondemand.c
--- a/drivers/cpufreq/cpufreq_ondemand.c	Wed Oct 17 09:24:03 2007 +0100
+++ b/drivers/cpufreq/cpufreq_ondemand.c	Thu Oct 18 16:08:36 2007 -0500
@@ -221,18 +222,14 @@ static struct attribute_group dbs_attr_g
 };
 
 /************************** sysfs end ************************/
-
-static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
-{
-	unsigned int idle_ticks, total_ticks;
-	unsigned int load;
+#ifndef CONFIG_XEN
+static int dbs_calc_load(struct cpu_dbs_info_s *this_dbs_info)
+{
+	struct cpufreq_policy *policy;
 	cputime64_t cur_jiffies;
-
-	struct cpufreq_policy *policy;
+	cputime64_t total_ticks, idle_ticks;
 	unsigned int j;
-
-	if (!this_dbs_info->enable)
-		return;
+	int load;
 
 	policy = this_dbs_info->cur_policy;
 	cur_jiffies = jiffies64_to_cputime64(get_jiffies_64());
@@ -240,7 +237,7 @@ static void dbs_check_cpu(struct cpu_dbs
 			this_dbs_info->prev_cpu_wall);
 	this_dbs_info->prev_cpu_wall = cur_jiffies;
 	if (!total_ticks)
-		return;
+		return 200;
 	/*
 	 * Every sampling_rate, we check, if current idle time is less
 	 * than 20% (default), then we try to increase frequency
@@ -270,6 +267,67 @@ static void dbs_check_cpu(struct cpu_dbs
 			idle_ticks = tmp_idle_ticks;
 	}
 	load = (100 * (total_ticks - idle_ticks)) / total_ticks;
+	return load;
+}
+#else
+
+#include <xen/interface/platform.h>
+static int dbs_calc_load(struct cpu_dbs_info_s *this_dbs_info)
+{
+	int load = 0;
+        struct xen_platform_op op;
+        uint64_t idletime[NR_CPUS];
+	uint64_t totaltime[NR_CPUS];
+	struct cpufreq_policy *policy;
+	unsigned int j;
+
+        op.cmd = XENPF_getidletime;
+        op.u.getidletime.max_cpus = num_online_cpus();
+        set_xen_guest_handle(op.u.getidletime.idletime, idletime);
+        set_xen_guest_handle(op.u.getidletime.totaltime, totaltime);
+        HYPERVISOR_platform_op(&op);
+
+	policy = this_dbs_info->cur_policy;
+	for_each_cpu_mask(j, policy->cpus) {
+		cputime64_t total_idle_ticks, total_wall_ticks;
+		cputime64_t tmp_idle_ticks, tmp_wall_ticks;
+		struct cpu_dbs_info_s *j_dbs_info;
+		unsigned long tmp_load;
+
+		j_dbs_info = &per_cpu(cpu_dbs_info, j);
+		total_idle_ticks = idletime[j];
+		tmp_idle_ticks = cputime64_sub(total_idle_ticks,
+				j_dbs_info->prev_cpu_idle);
+		total_wall_ticks = totaltime[j];
+		tmp_wall_ticks = cputime64_sub(total_wall_ticks,
+				j_dbs_info->prev_cpu_wall);
+		if (tmp_wall_ticks == 0)
+			return 200;
+
+		j_dbs_info->prev_cpu_wall = total_wall_ticks;
+		j_dbs_info->prev_cpu_idle = total_idle_ticks;
+
+		tmp_load = (100 * (tmp_wall_ticks - tmp_idle_ticks)) /
+				tmp_wall_ticks;
+		load = max(load, min(100, (int) tmp_load));
+	}
+	return load;
+}
+#endif
+
+static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
+{
+	int load;
+
+	struct cpufreq_policy *policy;
+
+	if (!this_dbs_info->enable)
+		return;
+
+	policy = this_dbs_info->cur_policy;
+	load = dbs_calc_load(this_dbs_info);
+	if (load > 100) 
+		return;
 
 	/* Check for frequency increase */
 	if (load > dbs_tuners_ins.up_threshold) {
diff -r 9bf1ddd0f6bf include/xen/interface/platform.h
--- a/include/xen/interface/platform.h	Wed Oct 17 09:24:03 2007 +0100
+++ b/include/xen/interface/platform.h	Thu Oct 18 16:08:36 2007 -0500
@@ -179,6 +179,7 @@ struct xenpf_getidletime {
     /* IN variables. */
     uint32_t max_cpus;
     XEN_GUEST_HANDLE(uint64_t) idletime;
+    XEN_GUEST_HANDLE(uint64_t) totaltime;
     /* OUT variables. */
     uint32_t nr_cpus;
 };





_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel

Tian, Kevin

2007-Oct-24 03:18 UTC

head link

RE: [Xen-devel] [PATCH][cpufreq] Xen support for the ondemand governor [2/2] (linux)

>From: Mark Langsdorf
>Sent: 2007年10月24日 6:00
>
>Modify the cpufreq ondemand governor so that it can get idle and
>total ticks from the Xen hypervisor.  Linux and Xen have different
>ideas of what an idle tick is, so the Xen values for both have to
>be returned in the same platform hypercall.  Otherwise, use
>basically the same scheme as native Linux.
>
>Signed-off-by: Mark Langsdorf <mark.langsdorf@amd.com>
>
[...]>+#else
>+
>+#include <xen/interface/platform.h>
>+static int dbs_calc_load(struct cpu_dbs_info_s *this_dbs_info)
>+{
>+	int load = 0;
>+        struct xen_platform_op op;
>+        uint64_t idletime[NR_CPUS];
>+	uint64_t totaltime[NR_CPUS];
>+	struct cpufreq_policy *policy;
>+	unsigned int j;
Note indent.
>+
>+        op.cmd = XENPF_getidletime;
>+        op.u.getidletime.max_cpus = num_online_cpus();
For current getidletime syntax, max_cpus is the max cpu index instead 
of max active cpu numbers. How do you handle cpu-hotplug case?
>+        set_xen_guest_handle(op.u.getidletime.idletime, idletime);
>+        set_xen_guest_handle(op.u.getidletime.totaltime, totaltime);
>+        HYPERVISOR_platform_op(&op);
Check return value.
>+
>+	policy = this_dbs_info->cur_policy;
>+	for_each_cpu_mask(j, policy->cpus) {
>+		cputime64_t total_idle_ticks, total_wall_ticks;
>+		cputime64_t tmp_idle_ticks, tmp_wall_ticks;
Rename "*_ticks" since Xen returns ns-based values?
>+		struct cpu_dbs_info_s *j_dbs_info;
>+		unsigned long tmp_load;
>+
>+		j_dbs_info = &per_cpu(cpu_dbs_info, j);
>+		total_idle_ticks = idletime[j];
>+		tmp_idle_ticks = cputime64_sub(total_idle_ticks,
>+				j_dbs_info->prev_cpu_idle);
>+		total_wall_ticks = totaltime[j];
>+		tmp_wall_ticks = cputime64_sub(total_wall_ticks,
>+				j_dbs_info->prev_cpu_wall);
>+		if (tmp_wall_ticks == 0)
>+			return 200;
As I commented in another mail, you don''t need count on elapsed time 
on other cpus. The agent cpu runs governor and thus the sample period 
is decided by that cpu. Just like the native Linux case, only pick the 
smallest idle delta in this loop.
>+
>+		j_dbs_info->prev_cpu_wall = total_wall_ticks;
>+		j_dbs_info->prev_cpu_idle = total_idle_ticks;
>+
>+		tmp_load = (100 * (tmp_wall_ticks - tmp_idle_ticks)) /
>+				tmp_wall_ticks;
>+		load = max(load, min(100, (int) tmp_load));
I suppose ''min'' here instead of ''max''. and
then init value of load is set
to ~0...
>+	}
>+	return load;
>+}
>+#endif[...]

Thanks,
Kevin

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel

Tian, Kevin

2007-Oct-24 03:23 UTC

head link

RE: [Xen-devel] [PATCH][cpufreq] Xen support for the ondemandgovernor [2/2] (linux)

>From: Tian, Kevin
>Sent: 2007年10月24日 11:19
>>+		load = max(load, min(100, (int) tmp_load));
>
>I suppose ''min'' here instead of ''max''.
and then init value of load is set
>to ~0...
>
Oh, forgive my mindless comment, and yours are correct indeed. :-)

Thanks,
Kevin

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel

Xen devel - Oct 2007 - [PATCH][cpufreq] Xen support for the ondemand governor [2/2] (linux)

[Xen-devel] [PATCH][cpufreq] Xen support for the ondemand governor [2/2] (linux)

RE: [Xen-devel] [PATCH][cpufreq] Xen support for the ondemand governor [2/2] (linux)

RE: [Xen-devel] [PATCH][cpufreq] Xen support for the ondemandgovernor [2/2] (linux)

Possibly Parallel Threads