powercap, intel_rapl, implement time window checks
cover powercap, intel_rapl, implement time window checks
Try #2 at this ....
P.
------8<---------
The Intel ASDM provides a maximum time window that can be specified when setting a time window in the RAPL driver. While the ASDM doesn't explicitly provide a minimum time window value, it does provide a minimum time window unit that also can be used as a minimum value.
This patchset implements barrier checking for the time windows, and adds reporting of a known bug in which the maxmimum time window value may be erroneously set to 0, as well as a module parameter to avoid the maximum window checks on broken BIOSes.
[v2]: update 3/3 with minor changes
Prarit Bhargava (3):
powercap, intel_rapl, implement get_max_time_window powercap, intel_rapl, implement check for minimum time window powercap, intel_rapl, Add ignore_max_time_window_check module
parameter for broken BIOSes
drivers/powercap/intel_rapl.c | 50 +++++++++++++++++++++++++++++++++++++++ drivers/powercap/powercap_sys.c | 6 +++-- 2 files changed, 54 insertions(+), 2 deletions(-)
Signature · 1 line
1.7.9.3
PATCH 1/3 powercap, intel_rapl, implement get_max_time_window
The MSR_PKG_POWER_INFO register (Intel ASDM, section 14.9.3 "Package RAPL Domain") provides a maximum time window which the system can support. This window is read-only and is currently not examined when setting the time windows for the package.
This patch implements get_max_time_window_us() and checks the window when a user attempts to set power capping for the package.
Before the patch it was possible to set the window to, for example, 10000 micro seconds:
[root@intel-chiefriver-03 rhel7]# echo 10000 > /sys/devices/virtual/powercap/intel-rapl/intel-rapl\:0/constraint_0_time_window_us; egrep ^ /sys/devices/virtual/powercap/intel-rapl/intel-rapl\:0/constraint_0_time_window_us
/sys/devices/virtual/powercap/intel-rapl/intel-rapl:0/constraint_0_time_window_us:1:9765
but from 'turbostat -d', the package is limited to 976us:
cpu0: MSR_PKG_POWER_INFO: 0x01200168 (45 W TDP, RAPL 36 - 0 W, 0.000977 sec.)
(Note, there appears to be a rounding issue in turbostat which needs to also be fixed. Looking at the values in the register it is clear the value is 1/1024 = 976us.)
After the patch we are limited by the maximum time window:
[root@intel-chiefriver-03 rhel7]# echo 10000 > /sys/devices/virtual/powercap/intel-rapl/intel-rapl\:0/constraint_0_time_window_us; egrep ^ /sys/devices/virtual/powercap/intel-rapl/intel-rapl\:0/constraint_0_time_window_us
-bash: echo: write error: Invalid argument /sys/devices/virtual/powercap/intel-rapl/intel-rapl:0/constraint_0_time_window_us:1:976
--- drivers/powercap/intel_rapl.c | 31 +++++++++++++++++++++++++++++++ drivers/powercap/powercap_sys.c | 6 ++++-- 2 files changed, 35 insertions(+), 2 deletions(-)
drivers/powercap/intel_rapl.c +31 -0
@@ -493,13 +493,42 @@ static int get_current_power_limit(struct powercap_zone *power_zone, int id,return ret; } +static int get_max_time_window (struct powercap_zone *power_zone,int id, +u64 *data) +{ +struct rapl_domain *rd; +int ret =0 ; +u64 val; + +get_online_cpus (); + rd =power_zone_to_rapl_domain (power_zone); + +if (rapl_read_data_raw (rd,MAX_TIME_WINDOW , true, &val)) + ret = -EIO ; +else + *data = val; + +put_online_cpus (); +return ret; +} +static int set_time_window (struct powercap_zone *power_zone,int id,u64 window) {struct rapl_domain *rd;int ret =0 ; +u64 max_window;get_online_cpus (); + ret =get_max_time_window (power_zone, id, &max_window); +if (ret <0 ) + goto out; + +if (window > max_window) { + ret = -EINVAL ; + goto out; + } + rd =power_zone_to_rapl_domain (power_zone);switch (rd->rpl[id].prim_id) {case PL1_ENABLE :
@@ -511,6 +540,7 @@ static int set_time_window(struct powercap_zone *power_zone, int id,default : ret = -EINVAL ; } +out:put_online_cpus ();return ret; }
@@ -590,6 +620,7 @@ static struct powercap_zone_constraint_ops constraint_ops = { .set_time_window_us = set_time_window, .get_time_window_us = get_time_window, .get_max_power_uw = get_max_power, + .get_max_time_window_us = get_max_time_window, .get_name = get_constraint_name, };
drivers/powercap/powercap_sys.c +4 -2
@@ -101,7 +101,7 @@ static ssize_t store_constraint_##_attr(struct device *dev,\int err; \u64 value; \struct powercap_zone *power_zone =to_powercap_zone (dev); \ -int id; \ +int id, ret; \struct powercap_zone_constraint *pconst;\ \if (!sscanf (dev_attr->attr.name,"constraint_%d_" , &id)) \
@@ -113,8 +113,10 @@ static ssize_t store_constraint_##_attr(struct device *dev,\if (err) \return -EINVAL ; \ if (pconst && pconst->ops && pconst->ops->set_##_attr) { \ - if (!pconst->ops->set_##_attr(power_zone, id, value)) \ + ret = pconst->ops->set_##_attr(power_zone, id, value); \ +if (!ret) \return count; \ +return ret; \ } \ \return -ENODATA ; \
Signature · 1 line
1.7.9.3
Re: [PATCH 1/3] powercap, intel_rapl, implement get_max_time_window
Quoted 51 lines · Prarit Bhargava
The
MSR_PKG_POWER_INFOregister (Intel ASDM, section 14.9.3 "Package RAPL Domain") provides a maximum time window which the system can support. This window is read-only and is currently not examined when setting the time windows for the package.This patch implements
get_max_time_window_us()and checks the window when a user attempts to set power capping for the package.Before the patch it was possible to set the window to, for example, 10000 micro seconds:
[root@intel-chiefriver-03 rhel7]# echo 10000 > /sys/devices/virtual/powercap/intel-rapl/intel-rapl\:0/constraint_0_time_window_us; egrep ^ /sys/devices/virtual/powercap/intel-rapl/intel-rapl\:0/constraint_0_time_window_us
/sys/devices/virtual/powercap/intel-rapl/intel-rapl:0/constraint_0_time_window_us:1:9765
but from 'turbostat -d', the package is limited to 976us:
cpu0:
MSR_PKG_POWER_INFO:0x01200168(45 W TDP, RAPL 36 - 0 W, 0.000977 sec.)(Note, there appears to be a rounding issue in turbostat which needs to also be fixed. Looking at the values in the register it is clear the value is 1/1024 = 976us.)
After the patch we are limited by the maximum time window:
[root@intel-chiefriver-03 rhel7]# echo 10000 > /sys/devices/virtual/powercap/intel-rapl/intel-rapl\:0/constraint_0_time_window_us; egrep ^ /sys/devices/virtual/powercap/intel-rapl/intel-rapl\:0/constraint_0_time_window_us
-bash: echo: write error: Invalid argument /sys/devices/virtual/powercap/intel-rapl/intel-rapl:0/constraint_0_time_window_us:1:976
Cc "Rafael J. Wysocki" <[email protected]>Cc Prarit Bhargava <[email protected]>Cc Radivoje Jovanovic <[email protected]>Cc Seiichi Ikarashi <[email protected]>Cc Mathias Krause <[email protected]>Cc Ajay Thomas <[email protected]>Signed-off-by Prarit Bhargava <[email protected]>--- drivers/powercap/intel_rapl.c | 31 +++++++++++++++++++++++++++++++ drivers/powercap/powercap_sys.c | 6 ++++-- 2 files changed, 35 insertions(+), 2 deletions(-)drivers/powercap/intel_rapl.c +0
-0
drivers/powercap/intel_rapl.c static int get_current_power_limit(struct powercap_zone *power_zone, int id, · line 497
Quoted 5 lines
drivers/powercap/intel_rapl.c +2
-0@@ -493,13 +493,42 @@ static int get_current_power_limit(struct powercap_zone *power_zone, int id,return ret; } +static int get_max_time_window (struct powercap_zone *power_zone,int id, +u64 *data)
The 2nd arg "id" is not necessary.
--
Seiichi
Re: [PATCH 1/3] powercap, intel_rapl, implement get_max_time_window
Quoted 55 lines · Seiichi Ikarashi
Quoted 54 lines · Prarit Bhargava
The
MSR_PKG_POWER_INFOregister (Intel ASDM, section 14.9.3 "Package RAPL Domain") provides a maximum time window which the system can support. This window is read-only and is currently not examined when setting the time windows for the package.This patch implements
get_max_time_window_us()and checks the window when a user attempts to set power capping for the package.Before the patch it was possible to set the window to, for example, 10000 micro seconds:
[root@intel-chiefriver-03 rhel7]# echo 10000 > /sys/devices/virtual/powercap/intel-rapl/intel-rapl\:0/constraint_0_time_window_us; egrep ^ /sys/devices/virtual/powercap/intel-rapl/intel-rapl\:0/constraint_0_time_window_us
/sys/devices/virtual/powercap/intel-rapl/intel-rapl:0/constraint_0_time_window_us:1:9765
but from 'turbostat -d', the package is limited to 976us:
cpu0:
MSR_PKG_POWER_INFO:0x01200168(45 W TDP, RAPL 36 - 0 W, 0.000977 sec.)(Note, there appears to be a rounding issue in turbostat which needs to also be fixed. Looking at the values in the register it is clear the value is 1/1024 = 976us.)
After the patch we are limited by the maximum time window:
[root@intel-chiefriver-03 rhel7]# echo 10000 > /sys/devices/virtual/powercap/intel-rapl/intel-rapl\:0/constraint_0_time_window_us; egrep ^ /sys/devices/virtual/powercap/intel-rapl/intel-rapl\:0/constraint_0_time_window_us
-bash: echo: write error: Invalid argument /sys/devices/virtual/powercap/intel-rapl/intel-rapl:0/constraint_0_time_window_us:1:976
Cc "Rafael J. Wysocki" <[email protected]>Cc Prarit Bhargava <[email protected]>Cc Radivoje Jovanovic <[email protected]>Cc Seiichi Ikarashi <[email protected]>Cc Mathias Krause <[email protected]>Cc Ajay Thomas <[email protected]>Signed-off-by Prarit Bhargava <[email protected]>--- drivers/powercap/intel_rapl.c | 31 +++++++++++++++++++++++++++++++ drivers/powercap/powercap_sys.c | 6 ++++-- 2 files changed, 35 insertions(+), 2 deletions(-)drivers/powercap/intel_rapl.c +0
-0@@ -493,13 +493,42 @@ static int get_current_power_limit(struct powercap_zone *power_zone, int id,return ret; }
Quoted 4 lines
drivers/powercap/intel_rapl.c static int get_current_power_limit(struct powercap_zone *power_zone, int id, · line 497
Quoted 2 lines
drivers/powercap/intel_rapl.c +2
-0+static int get_max_time_window (struct powercap_zone *power_zone,int id, +u64 *data)The 2nd arg "id" is not necessary.
It is required by the powercap_sys code. If I remove this then I have to create separate sysfs store and show files and I think it isn't worth it to do so.
P.
Quoted 4 lines
--
Seiichi
PATCH 2/3 powercap, intel_rapl, implement check for minimum time window
In reply to Prarit BhargavaUsing an small value for the time window results in a bogus value for the time window. For example,
[root@intel-chiefriver-03 linux]# echo 950 > /sys/devices/virtual/powercap/intel-rapl/intel-rapl\:0/constraint_0_time_window_us; egrep ^ /sys/devices/virtual/powercap/intel-rapl/intel-rapl\:0/constraint_0_time_window_us -bash: echo: write error: Invalid argument /sys/devices/virtual/powercap/intel-rapl/intel-rapl:0/constraint_0_time_window_us:1:4501502475370496
The Intel ASDM doesn't explicitly define a minimum time window. The MSR_RAPL_POWER_UNIT register, read during initialization, does specify a minimum time window unit so that can be used as a lower bound for error checking.
After this change the minimum time window is properly clamped:
[root@intel-chiefriver-03 linux]# echo 950 > /sys/devices/virtual/powercap/intel-rapl/intel-rapl\:0/constraint_0_time_window_us; egrep ^ /sys/devices/virtual/powercap/intel-rapl/intel-rapl\:0/constraint_0_time_window_us -bash: echo: write error: Invalid argument /sys/devices/virtual/powercap/intel-rapl/intel-rapl:0/constraint_0_time_window_us:1:976
--- drivers/powercap/intel_rapl.c | 11 +++++++++-- 1 file changed, 9 insertions(+), 2 deletions(-)
drivers/powercap/intel_rapl.c +9 -2
@@ -516,6 +516,7 @@ static int set_time_window(struct powercap_zone *power_zone, int id, u64 window) {struct rapl_domain *rd; +struct rapl_package *rp;int ret =0 ;u64 max_window;
@@ -524,12 +525,18 @@ static int set_time_window(struct powercap_zone *power_zone, int id,if (ret <0 ) goto out; -if (window > max_window) { + rd =power_zone_to_rapl_domain (power_zone); + rp =find_package_by_id (rd->package_id); +/* +* The Intel ASDM doesn't explicitly define a minimum time window. +* The MSR_RAPL_POWER_UNIT register, read during initialization, +* does contain the smallest unit of time that can be measured. +*/ +if ((window > max_window) || (window < rp->time_unit)) { ret = -EINVAL ; goto out; } - rd =power_zone_to_rapl_domain (power_zone);switch (rd->rpl[id].prim_id) {case PL1_ENABLE :rapl_write_data_raw (rd,TIME_WINDOW1 , window);
Signature · 1 line
1.7.9.3
PATCH 3/3 powercap, intel_rapl, Add ignore_max_time_window_check module parameter for broken BIOSes
In reply to Prarit BhargavaSome systems erroneously set the maximum time window field of MSR_PKG_POWER_INFO register to 0. This results in a user not being able to set the time windows for the package. In some cases, however, RAPL will still continue to work with a small window (albeit through some trial and error). This patch adds a ignore_max_time_window_check module parameter to avoid the maximum time window check in set_time_window().
[v2]: change name to max_time_window_check, fix (val == 0) check
--- drivers/powercap/intel_rapl.c | 18 +++++++++++++++--- 1 file changed, 15 insertions(+), 3 deletions(-)
drivers/powercap/intel_rapl.c +15 -3
@@ -505,13 +505,24 @@ static int get_max_time_window(struct powercap_zone *power_zone, int id,if (rapl_read_data_raw (rd,MAX_TIME_WINDOW , true, &val)) ret = -EIO ; -else +else { *data = val; - +if (val ==0 ) +pr_warn_once (FW_BUG "intel_rapl: Maximum Time Window is zero. This is a BIOS bug that should be reported to your hardware or BIOS vendor. The value of zero may prevent Intel RAPL from functioning properly. Most bugs can be avoided by setting the ignore_max_window_check module parameter.\n" ); + }put_online_cpus ();return ret; } +/* Some BIOSes incorrectly program the maximum time window in the +* MSR_PKG_POWER_INFO register. Some of these systems still have functional +* RAPL registers, etc., so give the user the option of disabling the maximum +* time window check. +*/ +static int ignore_max_time_window_check; +module_param (ignore_max_time_window_check, int,0444 ); +MODULE_PARM_DESC (ignore_max_time_window_check,"Ignore maximum time window check. A bug should be reported to your hardware or BIOS vendor if this parameter is used." ); +static int set_time_window (struct powercap_zone *power_zone,int id,u64 window) {
@@ -532,7 +543,8 @@ static int set_time_window(struct powercap_zone *power_zone, int id,* The MSR_RAPL_POWER_UNIT register, read during initialization, * does contain the smallest unit of time that can be measured. */ -if ((window > max_window) || (window < rp->time_unit)) { +if ((!ignore_max_time_window_check && (window > max_window)) || + (window < rp->time_unit)) { ret = -EINVAL ; goto out; }
Signature · 1 line
1.7.9.3
Re: [PATCH 3/3] powercap, intel_rapl, Add ignore_max_time_window_check module parameter for broken BIOSes
Quoted 25 lines · Prarit Bhargava
Some systems erroneously set the maximum time window field of
MSR_PKG_POWER_INFOregister to 0. This results in a user not being able to set the time windows for the package. In some cases, however, RAPL will still continue to work with a small window (albeit through some trial and error). This patch adds aignore_max_time_window_checkmodule parameter to avoid the maximum time window check inset_time_window().[v2]: change name to
max_time_window_check, fix (val == 0) checkCc "Rafael J. Wysocki" <[email protected]>Cc Prarit Bhargava <[email protected]>Cc Radivoje Jovanovic <[email protected]>Cc Seiichi Ikarashi <[email protected]>Cc Mathias Krause <[email protected]>Cc Ajay Thomas <[email protected]>Signed-off-by Prarit Bhargava <[email protected]>--- drivers/powercap/intel_rapl.c | 18 +++++++++++++++--- 1 file changed, 15 insertions(+), 3 deletions(-)drivers/powercap/intel_rapl.c +0
-0
drivers/powercap/intel_rapl.c static int get_max_time_window(struct powercap_zone *power_zone, int id, · line 511
Quoted 9 lines
drivers/powercap/intel_rapl.c +3
-2@@ -505,13 +505,24 @@ static int get_max_time_window(struct powercap_zone *power_zone, int id,if (rapl_read_data_raw (rd,MAX_TIME_WINDOW , true, &val)) ret = -EIO ; -else +else { *data = val; - +if (val ==0 ) +pr_warn_once (FW_BUG "intel_rapl: Maximum Time Window is zero. This is a BIOS bug that should be reported to your hardware or BIOS vendor. The value of zero may prevent Intel RAPL from functioning properly. Most bugs can be avoided by setting the ignore_max_window_check module parameter.\n" );
The correct name is ignore_max_time_window_check here.
--
Seiichi