Split from the series for powerclamp user of powercap idle-inject.
v2
- Build warnings reported by Rui
- Moved the powerclamp documentation to admin guide folder
- Commit log updated as suggested by Rafael and other code suggestion
Srinivas Pandruvada (2):
Documentation:admin-guide: Move intel_powerclamp documentation
thermal/drivers/intel_powerclamp: Add two module parameters
Documentation/admin-guide/index.rst | 1 +
.../thermal/intel_powerclamp.rst | 22 +++
Documentation/driver-api/thermal/index.rst | 1 -
MAINTAINERS | 1 +
drivers/thermal/intel/intel_powerclamp.c | 177 +++++++++++++++---
5 files changed, 180 insertions(+), 22 deletions(-)
rename Documentation/{driver-api => admin-guide}/thermal/intel_powerclamp.rst (93%)
--
2.39.1
Create a folder "thermal" under Documentation/admin-guide and move
intel_powerclamp documentation to this folder.
Signed-off-by: Srinivas Pandruvada <[email protected]>
---
Documentation/admin-guide/index.rst | 1 +
.../{driver-api => admin-guide}/thermal/intel_powerclamp.rst | 0
Documentation/driver-api/thermal/index.rst | 1 -
MAINTAINERS | 1 +
4 files changed, 2 insertions(+), 1 deletion(-)
rename Documentation/{driver-api => admin-guide}/thermal/intel_powerclamp.rst (100%)
diff --git a/Documentation/admin-guide/index.rst b/Documentation/admin-guide/index.rst
index 5bfafcbb9562..c872a8a1ddfa 100644
--- a/Documentation/admin-guide/index.rst
+++ b/Documentation/admin-guide/index.rst
@@ -116,6 +116,7 @@ configure specific aspects of kernel behavior to your liking.
svga
syscall-user-dispatch
sysrq
+ thermal
thunderbolt
ufs
unicode
diff --git a/Documentation/driver-api/thermal/intel_powerclamp.rst b/Documentation/admin-guide/thermal/intel_powerclamp.rst
similarity index 100%
rename from Documentation/driver-api/thermal/intel_powerclamp.rst
rename to Documentation/admin-guide/thermal/intel_powerclamp.rst
diff --git a/Documentation/driver-api/thermal/index.rst b/Documentation/driver-api/thermal/index.rst
index 030306ffa408..a886028014ab 100644
--- a/Documentation/driver-api/thermal/index.rst
+++ b/Documentation/driver-api/thermal/index.rst
@@ -14,7 +14,6 @@ Thermal
exynos_thermal
exynos_thermal_emulation
- intel_powerclamp
nouveau_thermal
x86_pkg_temperature_thermal
intel_dptf
diff --git a/MAINTAINERS b/MAINTAINERS
index b4043f72dfac..28fd62eacaf3 100644
--- a/MAINTAINERS
+++ b/MAINTAINERS
@@ -20714,6 +20714,7 @@ S: Supported
Q: https://patchwork.kernel.org/project/linux-pm/list/
T: git git://git.kernel.org/pub/scm/linux/kernel/git/rafael/linux-pm.git thermal
F: Documentation/ABI/testing/sysfs-class-thermal
+F: Documentation/admin-guide/thermal/
F: Documentation/devicetree/bindings/thermal/
F: Documentation/driver-api/thermal/
F: drivers/thermal/
--
2.39.1
In some use cases, it is desirable to only inject idle on certain set
of CPUs. For example on Alder Lake systems, it is possible that we force
idle only on P-Cores for thermal reasons. Also the idle percent can be
more than 50% if we only choose partial set of CPUs in the system.
Introduce 2 new module parameters for this purpose. They can be only
changed when the cooling device is inactive.
cpumask (Read/Write): A bit mask of CPUs to inject idle. The format of
this bitmask is same as used in other subsystems like in
/proc/irq/*/smp_affinity. The mask is comma separated 32 bit groups.
Each CPU is one bit. For example for 256 CPU system the full mask is:
ffffffff,ffffffff,ffffffff,ffffffff,ffffffff,ffffffff,ffffffff,ffffffff
The leftmost mask is for CPU 0-32.
max_idle (Read/Write): Maximum injected idle time to the total CPU time
ratio in percent range from 1 to 100. Even if the cooling device max_state
is always 100 (100%), this parameter allows to add a max idle percent
limit. The default is 50, to match the current implementation of powerclamp
driver. Also doesn't allow value more than 75, if the cpumask includes
every CPU present in the system.
Also when the cpumask doesn't include every CPU, there is no use of
compensation using package C-state idle counters. Hence don't start
package C-state polling thread even for a single package or a single die
system in this case.
Signed-off-by: Srinivas Pandruvada <[email protected]>
---
.../admin-guide/thermal/intel_powerclamp.rst | 22 +++
drivers/thermal/intel/intel_powerclamp.c | 177 +++++++++++++++---
2 files changed, 178 insertions(+), 21 deletions(-)
diff --git a/Documentation/admin-guide/thermal/intel_powerclamp.rst b/Documentation/admin-guide/thermal/intel_powerclamp.rst
index 3f6dfb0b3ea6..da83b5eefbff 100644
--- a/Documentation/admin-guide/thermal/intel_powerclamp.rst
+++ b/Documentation/admin-guide/thermal/intel_powerclamp.rst
@@ -26,6 +26,8 @@ By:
- Generic Thermal Layer (sysfs)
- Kernel APIs (TBD)
+ (*) Module Parameters
+
INTRODUCTION
============
@@ -318,3 +320,23 @@ device, a PID based userspace thermal controller can manage to
control CPU temperature effectively, when no other thermal influence
is added. For example, a UltraBook user can compile the kernel under
certain temperature (below most active trip points).
+
+Module Parameters
+=================
+
+``cpumask`` (RW)
+ A bit mask of CPUs to inject idle. The format of the bitmask is same as
+ used in other subsystems like in /proc/irq/*/smp_affinity. The mask is
+ comma separated 32 bit groups. Each CPU is one bit. For example for a 256
+ CPU system the full mask is:
+ ffffffff,ffffffff,ffffffff,ffffffff,ffffffff,ffffffff,ffffffff,ffffffff
+
+ The leftmost mask is for CPU 0-32.
+
+``max_idle`` (RW)
+ Maximum injected idle time to the total CPU time ratio in percent range
+ from 1 to 100. Even if the cooling device max_state is always 100 (100%),
+ this parameter allows to add a max idle percent limit. The default is 50,
+ to match the current implementation of powerclamp driver. Also doesn't
+ allow value more than 75, if the cpumask includes every CPU present in
+ the system.
diff --git a/drivers/thermal/intel/intel_powerclamp.c b/drivers/thermal/intel/intel_powerclamp.c
index 1390748706a6..6d00ac597b8a 100644
--- a/drivers/thermal/intel/intel_powerclamp.c
+++ b/drivers/thermal/intel/intel_powerclamp.c
@@ -37,7 +37,7 @@
#include <asm/mwait.h>
#include <asm/cpu_device_id.h>
-#define MAX_TARGET_RATIO (50U)
+#define MAX_TARGET_RATIO (100U)
/* For each undisturbed clamping period (no extra wake ups during idle time),
* we increment the confidence counter for the given target ratio.
* CONFIDENCE_OK defines the level where runtime calibration results are
@@ -105,10 +105,144 @@ static const struct kernel_param_ops duration_ops = {
.get = param_get_int,
};
-
module_param_cb(duration, &duration_ops, &duration, 0644);
MODULE_PARM_DESC(duration, "forced idle time for each attempt in msec.");
+#define DEFAULT_MAX_IDLE 50
+#define MAX_ALL_CPU_IDLE 75
+
+static u8 max_idle = DEFAULT_MAX_IDLE;
+
+static cpumask_var_t idle_injection_cpu_mask;
+
+static int allocate_copy_idle_injection_mask(const struct cpumask *copy_mask)
+{
+ if (cpumask_available(idle_injection_cpu_mask))
+ goto copy_mask;
+
+ /* This mask is allocated only one time and freed during module exit */
+ if (!alloc_cpumask_var(&idle_injection_cpu_mask, GFP_KERNEL))
+ return -ENOMEM;
+
+copy_mask:
+ cpumask_copy(idle_injection_cpu_mask, copy_mask);
+
+ return 0;
+}
+
+/* Return true if the cpumask and idle percent combination is invalid */
+static bool check_invalid(cpumask_var_t mask, u8 idle)
+{
+ if (cpumask_equal(cpu_present_mask, mask) && idle > MAX_ALL_CPU_IDLE)
+ return true;
+
+ return false;
+}
+
+static int cpumask_set(const char *arg, const struct kernel_param *kp)
+{
+ cpumask_var_t new_mask;
+ int ret;
+
+ mutex_lock(&powerclamp_lock);
+
+ /* Can't set mask when cooling device is in use */
+ if (powerclamp_data.clamping) {
+ ret = -EAGAIN;
+ goto skip_cpumask_set;
+ }
+
+ ret = alloc_cpumask_var(&new_mask, GFP_KERNEL);
+ if (!ret)
+ goto skip_cpumask_set;
+
+ ret = bitmap_parse(arg, strlen(arg), cpumask_bits(new_mask),
+ nr_cpumask_bits);
+ if (ret)
+ goto free_cpumask_set;
+
+ if (cpumask_empty(new_mask) || check_invalid(new_mask, max_idle)) {
+ ret = -EINVAL;
+ goto free_cpumask_set;
+ }
+
+ /*
+ * When module parameters are passed from kernel command line
+ * during insmod, the module parameter callback is called
+ * before powerclamp_init(), so we can't assume that some
+ * cpumask can be allocated and copied before here. Also
+ * in this case this cpumask is used as the default mask.
+ */
+ ret = allocate_copy_idle_injection_mask(new_mask);
+
+free_cpumask_set:
+ free_cpumask_var(new_mask);
+skip_cpumask_set:
+ mutex_unlock(&powerclamp_lock);
+
+ return ret;
+}
+
+static int cpumask_get(char *buf, const struct kernel_param *kp)
+{
+ if (!cpumask_available(idle_injection_cpu_mask))
+ return -ENODEV;
+
+ return bitmap_print_to_pagebuf(false, buf, cpumask_bits(idle_injection_cpu_mask),
+ nr_cpumask_bits);
+}
+
+static const struct kernel_param_ops cpumask_ops = {
+ .set = cpumask_set,
+ .get = cpumask_get,
+};
+
+module_param_cb(cpumask, &cpumask_ops, NULL, 0644);
+MODULE_PARM_DESC(cpumask, "Mask of CPUs to use for idle injection.");
+
+static int max_idle_set(const char *arg, const struct kernel_param *kp)
+{
+ u8 new_max_idle;
+ int ret = 0;
+
+ mutex_lock(&powerclamp_lock);
+
+ /* Can't set mask when cooling device is in use */
+ if (powerclamp_data.clamping) {
+ ret = -EAGAIN;
+ goto skip_limit_set;
+ }
+
+ ret = kstrtou8(arg, 10, &new_max_idle);
+ if (ret)
+ goto skip_limit_set;
+
+ if (new_max_idle > MAX_TARGET_RATIO) {
+ ret = -EINVAL;
+ goto skip_limit_set;
+ }
+
+ if (check_invalid(idle_injection_cpu_mask, new_max_idle)) {
+ ret = -EINVAL;
+ goto skip_limit_set;
+ }
+
+ max_idle = new_max_idle;
+
+skip_limit_set:
+ mutex_unlock(&powerclamp_lock);
+
+ return ret;
+}
+
+static const struct kernel_param_ops max_idle_ops = {
+ .set = max_idle_set,
+ .get = param_get_int,
+};
+
+module_param_cb(max_idle, &max_idle_ops, &max_idle, 0644);
+MODULE_PARM_DESC(max_idle, "maximum injected idle time to the total CPU time ratio in percent range:1-100");
+
struct powerclamp_calibration_data {
unsigned long confidence; /* used for calibration, basically a counter
* gets incremented each time a clamping
@@ -460,21 +594,15 @@ static void trigger_idle_injection(void)
*/
static int powerclamp_idle_injection_register(void)
{
- /*
- * The idle inject core will only inject for online CPUs,
- * So we can register for all present CPUs. In this way
- * if some CPU goes online/offline while idle inject
- * is registered, nothing additional calls are required.
- * The same runtime and idle time is applicable for
- * newly onlined CPUs if any.
- *
- * Here cpu_present_mask can be used as is.
- * cast to (struct cpumask *) is required as the
- * cpu_present_mask is const struct cpumask *, otherwise
- * there will be compiler warnings.
- */
- ii_dev = idle_inject_register_full((struct cpumask *)cpu_present_mask,
- idle_inject_update);
+ poll_pkg_cstate_enable = false;
+ if (cpumask_equal(cpu_present_mask, idle_injection_cpu_mask)) {
+ ii_dev = idle_inject_register_full(idle_injection_cpu_mask, idle_inject_update);
+ if (topology_max_packages() == 1 && topology_max_die_per_package() == 1)
+ poll_pkg_cstate_enable = true;
+ } else {
+ ii_dev = idle_inject_register(idle_injection_cpu_mask);
+ }
+
if (!ii_dev) {
pr_err("powerclamp: idle_inject_register failed\n");
return -EAGAIN;
@@ -555,7 +683,7 @@ static int powerclamp_set_cur_state(struct thermal_cooling_device *cdev,
mutex_lock(&powerclamp_lock);
new_target_ratio = clamp(new_target_ratio, 0UL,
- (unsigned long) (MAX_TARGET_RATIO - 1));
+ (unsigned long) (max_idle - 1));
if (!powerclamp_data.target_ratio && new_target_ratio > 0) {
pr_info("Start idle injection to reduce power\n");
powerclamp_data.target_ratio = new_target_ratio;
@@ -646,15 +774,19 @@ static int __init powerclamp_init(void)
/* probe cpu features and ids here */
retval = powerclamp_probe();
+ if (retval)
+ return retval;
+
+ mutex_lock(&powerclamp_lock);
+ retval = allocate_copy_idle_injection_mask(cpu_present_mask);
+ mutex_unlock(&powerclamp_lock);
+
if (retval)
return retval;
/* set default limit, maybe adjusted during runtime based on feedback */
window_size = 2;
- if (topology_max_packages() == 1 && topology_max_die_per_package() == 1)
- poll_pkg_cstate_enable = true;
-
cooling_dev = thermal_cooling_device_register("intel_powerclamp", NULL,
&powerclamp_cooling_ops);
if (IS_ERR(cooling_dev))
@@ -679,6 +811,9 @@ static void __exit powerclamp_exit(void)
cancel_delayed_work_sync(&poll_pkg_cstate_work);
debugfs_remove_recursive(debug_dir);
+
+ if (cpumask_available(idle_injection_cpu_mask))
+ free_cpumask_var(idle_injection_cpu_mask);
}
module_exit(powerclamp_exit);
--
2.39.1
Hi, Srinivas,
First of all, the previous build error is gone.
Second, I found something strange, which may be related with the
scheduler asym-packing, so CC Ricardo.
The test is done with pm linux-intel branch + this patch series on an
ADL system. cpu0~cpu7 are Pcore cpus, cpu8-cpu15 are Ecore cpus, and
intel_powerclamp is register as cooling_device21.
1. run stress -c 16
2. update /sys/module/intel_powerclamp/parameters/cpumask
echo 90 > /sys/module/intel_powerclamp/parameters/max_idle
3. echo 90 > /sys/class/thermal/cooling_device21/cur_state
4. echo 0 > /sys/class/thermal/cooling_device21/cur_state
I use turbostat to monitor the CPU Busy% in all 4 steps.
If 'cpumask' does not include all the Ecore CPUs, all CPUs becomes 100%
busy after idle injection removed in step 4.
If 'cpumask' includes all the Ecore CPUs, i.e. cpumask = FFxy, in some
cases, the Ecore CPUs will drop to an Busy% much lower than 10%, and
then they don't come back to busy after idle injection removed in step
4, although we have 16 stress threads. And this also relates with how
long we stay in idle injection.
Say, when cpumask=fff3, the problem can be triggered occasionally if
there is a 10 second timeout between step 3 and step4, but it is much
easier to reproducible if I increase the timeout to 20 seconds.
It seems that Pcore can always pull tasks from Ecores, but Ecore can
not pull tasks from Pcore HT siblings.
thanks,
rui
On Sat, 2023-02-04 at 18:59 -0800, Srinivas Pandruvada wrote:
> Split from the series for powerclamp user of powercap idle-inject.
>
> v2
> - Build warnings reported by Rui
> - Moved the powerclamp documentation to admin guide folder
> - Commit log updated as suggested by Rafael and other code suggestion
>
> Srinivas Pandruvada (2):
> Documentation:admin-guide: Move intel_powerclamp documentation
> thermal/drivers/intel_powerclamp: Add two module parameters
>
> Documentation/admin-guide/index.rst | 1 +
> .../thermal/intel_powerclamp.rst | 22 +++
> Documentation/driver-api/thermal/index.rst | 1 -
> MAINTAINERS | 1 +
> drivers/thermal/intel/intel_powerclamp.c | 177 +++++++++++++++-
> --
> 5 files changed, 180 insertions(+), 22 deletions(-)
> rename Documentation/{driver-api => admin-
> guide}/thermal/intel_powerclamp.rst (93%)
>
Hi Rui,
On Sun, 2023-02-05 at 15:57 +0000, Zhang, Rui wrote:
> Hi, Srinivas,
>
> First of all, the previous build error is gone.
>
> Second, I found something strange, which may be related with the
> scheduler asym-packing, so CC Ricardo.
>
I thought you disable ITMT before idle injection and reenebale after
removal.
> The test is done with pm linux-intel branch + this patch series on an
> ADL system.
Can you do test on bleeding edge branch of Linux-pm?
> cpu0~cpu7 are Pcore cpus, cpu8-cpu15 are Ecore cpus, and
> intel_powerclamp is register as cooling_device21.
>
> 1. run stress -c 16
> 2. update /sys/module/intel_powerclamp/parameters/cpumask
> echo 90 > /sys/module/intel_powerclamp/parameters/max_idle
> 3. echo 90 > /sys/class/thermal/cooling_device21/cur_state
> 4. echo 0 > /sys/class/thermal/cooling_device21/cur_state
> I use turbostat to monitor the CPU Busy% in all 4 steps.
>
> If 'cpumask' does not include all the Ecore CPUs, all CPUs becomes
> 100%
> busy after idle injection removed in step 4.
>
that should be the case.
> If 'cpumask' includes all the Ecore CPUs, i.e. cpumask = FFxy, in
> some
> cases, the Ecore CPUs will drop to an Busy% much lower than 10%, and
> then they don't come back to busy after idle injection removed in
> step
Do you see that idle injection is removed message in dmesg?
We can also check powercap idle-inejct, if some CPUs still not wake
from play_idle.
> 4, although we have 16 stress threads. And this also relates with how
> long we stay in idle injection.
>
> Say, when cpumask=fff3, the problem can be triggered occasionally if
> there is a 10 second timeout between step 3 and step4, but it is much
> easier to reproducible if I increase the timeout to 20 seconds.
>
> It seems that Pcore can always pull tasks from Ecores, but Ecore can
> not pull tasks from Pcore HT siblings.
>
That will be regular load balance threads should do.
Better to try upsteam kernel first.
Thanks,
Srinivas
> thanks,
> rui
>
> On Sat, 2023-02-04 at 18:59 -0800, Srinivas Pandruvada wrote:
> > Split from the series for powerclamp user of powercap idle-inject.
> >
> > v2
> > - Build warnings reported by Rui
> > - Moved the powerclamp documentation to admin guide folder
> > - Commit log updated as suggested by Rafael and other code
> > suggestion
> >
> > Srinivas Pandruvada (2):
> > Documentation:admin-guide: Move intel_powerclamp documentation
> > thermal/drivers/intel_powerclamp: Add two module parameters
> >
> > Documentation/admin-guide/index.rst | 1 +
> > .../thermal/intel_powerclamp.rst | 22 +++
> > Documentation/driver-api/thermal/index.rst | 1 -
> > MAINTAINERS | 1 +
> > drivers/thermal/intel/intel_powerclamp.c | 177
> > +++++++++++++++-
> > --
> > 5 files changed, 180 insertions(+), 22 deletions(-)
> > rename Documentation/{driver-api => admin-
> > guide}/thermal/intel_powerclamp.rst (93%)
> >
On Sun, 2023-02-05 at 18:45 -0800, srinivas pandruvada wrote:
> Hi Rui,
>
> On Sun, 2023-02-05 at 15:57 +0000, Zhang, Rui wrote:
> > Hi, Srinivas,
> >
> > First of all, the previous build error is gone.
> >
> > Second, I found something strange, which may be related with the
> > scheduler asym-packing, so CC Ricardo.
> >
> I thought you disable ITMT before idle injection and reenebale after
> removal.
No.
I can reproduce this by playing with raw intel_powerclamp sysfs knobs
and ITMT enabled.
>
>
>
> > The test is done with pm linux-intel branch
sorry, I mean linux-next branch.
> > + this patch series on an
> > ADL system.
> Can you do test on bleeding edge branch of Linux-pm?
>
> > cpu0~cpu7 are Pcore cpus, cpu8-cpu15 are Ecore cpus, and
> > intel_powerclamp is register as cooling_device21.
> >
> > 1. run stress -c 16
> > 2. update /sys/module/intel_powerclamp/parameters/cpumask
> > echo 90 > /sys/module/intel_powerclamp/parameters/max_idle
> > 3. echo 90 > /sys/class/thermal/cooling_device21/cur_state
> > 4. echo 0 > /sys/class/thermal/cooling_device21/cur_state
> > I use turbostat to monitor the CPU Busy% in all 4 steps.
> >
> > If 'cpumask' does not include all the Ecore CPUs, all CPUs becomes
> > 100%
> > busy after idle injection removed in step 4.
> >
> that should be the case.
>
> > If 'cpumask' includes all the Ecore CPUs, i.e. cpumask = FFxy, in
> > some
> > cases, the Ecore CPUs will drop to an Busy% much lower than 10%,
> > and
> > then they don't come back to busy after idle injection removed in
> > step
> Do you see that idle injection is removed message in dmesg?
yes.
> We can also check powercap idle-inejct, if some CPUs still not wake
> from play_idle.
"ps" command shows the the idle_injection threads time is not
increasing any more.
>
>
> > 4, although we have 16 stress threads. And this also relates with
> > how
> > long we stay in idle injection.
> >
> > Say, when cpumask=fff3, the problem can be triggered occasionally
> > if
> > there is a 10 second timeout between step 3 and step4, but it is
> > much
> > easier to reproducible if I increase the timeout to 20 seconds.
> >
> > It seems that Pcore can always pull tasks from Ecores, but Ecore
> > can
> > not pull tasks from Pcore HT siblings.
> >
> That will be regular load balance threads should do.
> Better to try upsteam kernel first.
I'm already running with linux-pm tree linux-next branch + this patch
series.
thanks,
rui
>
> Thanks,
> Srinivas
>
>
> > thanks,
> > rui
> >
> > On Sat, 2023-02-04 at 18:59 -0800, Srinivas Pandruvada wrote:
> > > Split from the series for powerclamp user of powercap idle-
> > > inject.
> > >
> > > v2
> > > - Build warnings reported by Rui
> > > - Moved the powerclamp documentation to admin guide folder
> > > - Commit log updated as suggested by Rafael and other code
> > > suggestion
> > >
> > > Srinivas Pandruvada (2):
> > > Documentation:admin-guide: Move intel_powerclamp documentation
> > > thermal/drivers/intel_powerclamp: Add two module parameters
> > >
> > > Documentation/admin-guide/index.rst | 1 +
> > > .../thermal/intel_powerclamp.rst | 22 +++
> > > Documentation/driver-api/thermal/index.rst | 1 -
> > > MAINTAINERS | 1 +
> > > drivers/thermal/intel/intel_powerclamp.c | 177
> > > +++++++++++++++-
> > > --
> > > 5 files changed, 180 insertions(+), 22 deletions(-)
> > > rename Documentation/{driver-api => admin-
> > > guide}/thermal/intel_powerclamp.rst (93%)
> > >
On Mon, 2023-02-06 at 08:05 +0000, Zhang, Rui wrote:
> On Sun, 2023-02-05 at 18:45 -0800, srinivas pandruvada wrote:
> > Hi Rui,
> >
> > On Sun, 2023-02-05 at 15:57 +0000, Zhang, Rui wrote:
> > > Hi, Srinivas,
> > >
> > > First of all, the previous build error is gone.
> > >
> > > Second, I found something strange, which may be related with the
> > > scheduler asym-packing, so CC Ricardo.
> > >
> > I thought you disable ITMT before idle injection and reenebale
> > after
> > removal.
>
> No.
>
> I can reproduce this by playing with raw intel_powerclamp sysfs knobs
> and ITMT enabled.
>
This issue is happening even if ITMT disabled. If the module mask is
composed of P-cores it works or even on servers as expected.
Also if you offline all P-cores then select mask among E-cores, it is
working. Somehow P-core influences E-cores.
Since this patch is module mask related, that is functioning correctly.
We have to debug this interaction with P and E cores separately.
Thanks,
Srinivas
> >
> >
> >
> > > The test is done with pm linux-intel branch
>
> sorry, I mean linux-next branch.
>
> > > + this patch series on an
> > > ADL system.
> > Can you do test on bleeding edge branch of Linux-pm?
> >
> > > cpu0~cpu7 are Pcore cpus, cpu8-cpu15 are Ecore cpus, and
> > > intel_powerclamp is register as cooling_device21.
> > >
> > > 1. run stress -c 16
> > > 2. update /sys/module/intel_powerclamp/parameters/cpumask
> > > echo 90 > /sys/module/intel_powerclamp/parameters/max_idle
> > > 3. echo 90 > /sys/class/thermal/cooling_device21/cur_state
> > > 4. echo 0 > /sys/class/thermal/cooling_device21/cur_state
> > > I use turbostat to monitor the CPU Busy% in all 4 steps.
> > >
> > > If 'cpumask' does not include all the Ecore CPUs, all CPUs
> > > becomes
> > > 100%
> > > busy after idle injection removed in step 4.
> > >
> > that should be the case.
> >
> > > If 'cpumask' includes all the Ecore CPUs, i.e. cpumask = FFxy, in
> > > some
> > > cases, the Ecore CPUs will drop to an Busy% much lower than 10%,
> > > and
> > > then they don't come back to busy after idle injection removed in
> > > step
> > Do you see that idle injection is removed message in dmesg?
>
> yes.
>
> > We can also check powercap idle-inejct, if some CPUs still not wake
> > from play_idle.
>
> "ps" command shows the the idle_injection threads time is not
> increasing any more.
>
> >
> >
> > > 4, although we have 16 stress threads. And this also relates with
> > > how
> > > long we stay in idle injection.
> > >
> > > Say, when cpumask=fff3, the problem can be triggered occasionally
> > > if
> > > there is a 10 second timeout between step 3 and step4, but it is
> > > much
> > > easier to reproducible if I increase the timeout to 20 seconds.
> > >
> > > It seems that Pcore can always pull tasks from Ecores, but Ecore
> > > can
> > > not pull tasks from Pcore HT siblings.
> > >
> > That will be regular load balance threads should do.
> > Better to try upsteam kernel first.
>
> I'm already running with linux-pm tree linux-next branch + this patch
> series.
>
> thanks,
> rui
>
> >
> > Thanks,
> > Srinivas
> >
> >
> > > thanks,
> > > rui
> > >
> > > On Sat, 2023-02-04 at 18:59 -0800, Srinivas Pandruvada wrote:
> > > > Split from the series for powerclamp user of powercap idle-
> > > > inject.
> > > >
> > > > v2
> > > > - Build warnings reported by Rui
> > > > - Moved the powerclamp documentation to admin guide folder
> > > > - Commit log updated as suggested by Rafael and other code
> > > > suggestion
> > > >
> > > > Srinivas Pandruvada (2):
> > > > Documentation:admin-guide: Move intel_powerclamp
> > > > documentation
> > > > thermal/drivers/intel_powerclamp: Add two module parameters
> > > >
> > > > Documentation/admin-guide/index.rst | 1 +
> > > > .../thermal/intel_powerclamp.rst | 22 +++
> > > > Documentation/driver-api/thermal/index.rst | 1 -
> > > > MAINTAINERS | 1 +
> > > > drivers/thermal/intel/intel_powerclamp.c | 177
> > > > +++++++++++++++-
> > > > --
> > > > 5 files changed, 180 insertions(+), 22 deletions(-)
> > > > rename Documentation/{driver-api => admin-
> > > > guide}/thermal/intel_powerclamp.rst (93%)
> > > >
On Mon, Feb 06, 2023 at 02:02:28AM -0800, srinivas pandruvada wrote:
> On Mon, 2023-02-06 at 08:05 +0000, Zhang, Rui wrote:
> > On Sun, 2023-02-05 at 18:45 -0800, srinivas pandruvada wrote:
> > > Hi Rui,
> > >
> > > On Sun, 2023-02-05 at 15:57 +0000, Zhang, Rui wrote:
> > > > Hi, Srinivas,
> > > >
> > > > First of all, the previous build error is gone.
> > > >
> > > > Second, I found something strange, which may be related with the
> > > > scheduler asym-packing, so CC Ricardo.
> > > >
> > > I thought you disable ITMT before idle injection and reenebale
> > > after
> > > removal.
> >
> > No.
> >
> > I can reproduce this by playing with raw intel_powerclamp sysfs knobs
> > and ITMT enabled.
> >
>
> This issue is happening even if ITMT disabled. If the module mask is
> composed of P-cores it works or even on servers as expected.
> Also if you offline all P-cores then select mask among E-cores, it is
> working. Somehow P-core influences E-cores.
>
> Since this patch is module mask related, that is functioning correctly.
> We have to debug this interaction with P and E cores separately.
Currently, when doing asym_packing, ECores will only pull tasks from a
PCore only if both SMT siblings are busy. It will only pull from the
lower-priority sibling. These patches [1] let ECores pull from either
sibling, if both are busy.
I presume that by injecting idle, the scheduler thinks that the CPU is
idle (i.e., idle_cpu() returns true) and it will not do asym_packing from
lower-priority CPUs.
However, in your experiment you have 16 threads. If a Pcore is overloaded,
an ECore should be able to help.
[1]. https://lore.kernel.org/lkml/[email protected]/
On Tue, 2023-02-07 at 05:42 -0800, Ricardo Neri wrote:
> On Mon, Feb 06, 2023 at 02:02:28AM -0800, srinivas pandruvada wrote:
> > On Mon, 2023-02-06 at 08:05 +0000, Zhang, Rui wrote:
> > > On Sun, 2023-02-05 at 18:45 -0800, srinivas pandruvada wrote:
> > > > Hi Rui,
> > > >
> > > > On Sun, 2023-02-05 at 15:57 +0000, Zhang, Rui wrote:
> > > > > Hi, Srinivas,
> > > > >
> > > > > First of all, the previous build error is gone.
> > > > >
> > > > > Second, I found something strange, which may be related with
> > > > > the
> > > > > scheduler asym-packing, so CC Ricardo.
> > > > >
> > > > I thought you disable ITMT before idle injection and reenebale
> > > > after
> > > > removal.
> > >
> > > No.
> > >
> > > I can reproduce this by playing with raw intel_powerclamp sysfs
> > > knobs
> > > and ITMT enabled.
> > >
> >
> > This issue is happening even if ITMT disabled. If the module mask
> > is
> > composed of P-cores it works or even on servers as expected.
> > Also if you offline all P-cores then select mask among E-cores, it
> > is
> > working. Somehow P-core influences E-cores.
> >
> > Since this patch is module mask related, that is functioning
> > correctly.
> > We have to debug this interaction with P and E cores separately.
>
> Currently, when doing asym_packing, ECores will only pull tasks from
> a
> PCore only if both SMT siblings are busy. It will only pull from the
> lower-priority sibling. These patches [1] let ECores pull from either
> sibling, if both are busy.
>
> I presume that by injecting idle, the scheduler thinks that the CPU
> is
> idle (i.e., idle_cpu() returns true) and it will not do asym_packing
> from
> lower-priority CPUs.
>
> However, in your experiment you have 16 threads. If a Pcore is
> overloaded,
> an ECore should be able to help.
This issue happens with or without ITMT and also without any idle
injection active.
Thanks,
Srinivas
>
> [1].
> https://lore.kernel.org/lkml/[email protected]/
>
On Tue, Feb 07, 2023 at 05:51:59AM -0800, srinivas pandruvada wrote:
> On Tue, 2023-02-07 at 05:42 -0800, Ricardo Neri wrote:
> > On Mon, Feb 06, 2023 at 02:02:28AM -0800, srinivas pandruvada wrote:
> > > On Mon, 2023-02-06 at 08:05 +0000, Zhang, Rui wrote:
> > > > On Sun, 2023-02-05 at 18:45 -0800, srinivas pandruvada wrote:
> > > > > Hi Rui,
> > > > >
> > > > > On Sun, 2023-02-05 at 15:57 +0000, Zhang, Rui wrote:
> > > > > > Hi, Srinivas,
> > > > > >
> > > > > > First of all, the previous build error is gone.
> > > > > >
> > > > > > Second, I found something strange, which may be related with
> > > > > > the
> > > > > > scheduler asym-packing, so CC Ricardo.
> > > > > >
> > > > > I thought you disable ITMT before idle injection and reenebale
> > > > > after
> > > > > removal.
> > > >
> > > > No.
> > > >
> > > > I can reproduce this by playing with raw intel_powerclamp sysfs
> > > > knobs
> > > > and ITMT enabled.
> > > >
> > >
> > > This issue is happening even if ITMT disabled. If the module mask
> > > is
> > > composed of P-cores it works or even on servers as expected.
> > > Also if you offline all P-cores then select mask among E-cores, it
> > > is
> > > working. Somehow P-core influences E-cores.
> > >
> > > Since this patch is module mask related, that is functioning
> > > correctly.
> > > We have to debug this interaction with P and E cores separately.
> >
> > Currently, when doing asym_packing, ECores will only pull tasks from
> > a
> > PCore only if both SMT siblings are busy. It will only pull from the
> > lower-priority sibling. These patches [1] let ECores pull from either
> > sibling, if both are busy.
> >
> > I presume that by injecting idle, the scheduler thinks that the CPU
> > is
> > idle (i.e., idle_cpu() returns true) and it will not do asym_packing
> > from
> > lower-priority CPUs.
> >
> > However, in your experiment you have 16 threads. If a Pcore is
> > overloaded,
> > an ECore should be able to help.
> This issue happens with or without ITMT and also without any idle
> injection active.
I was not able to reproduce this issue on my ADL-S system with ITMT. The
described bug is exactly what and old patchset of mine was supposed to
fix [2]. Maybe the CPU priorities in the failing system are such that it
prevents asym_packing from kicking in.
I was able to reproduce the issue without ITMT.
I had found that the scheduler cannot handle load balance between SMT and
non-SMT cores correctly. My patchset [1] includes fixes for this case. I
applied it on top of Rafael's linux-next branch and it fixed the issue for
me in the non-ITMT case. Perhaps patches 5 and 6 are sufficient, but I
applied the whole series.
Thanks and BR,
Ricardo
[1]. https://lore.kernel.org/lkml/[email protected]/
[2]. https://lore.kernel.org/all/[email protected]/