2 * Intel Running Average Power Limit (RAPL) Driver
3 * Copyright (c) 2013, Intel Corporation.
5 * This program is free software; you can redistribute it and/or modify it
6 * under the terms and conditions of the GNU General Public License,
7 * version 2, as published by the Free Software Foundation.
9 * This program is distributed in the hope it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
14 * You should have received a copy of the GNU General Public License along with
15 * this program; if not, write to the Free Software Foundation, Inc.
18 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
20 #include <linux/kernel.h>
21 #include <linux/module.h>
22 #include <linux/list.h>
23 #include <linux/types.h>
24 #include <linux/device.h>
25 #include <linux/slab.h>
26 #include <linux/log2.h>
27 #include <linux/bitmap.h>
28 #include <linux/delay.h>
29 #include <linux/sysfs.h>
30 #include <linux/cpu.h>
31 #include <linux/powercap.h>
32 #include <asm/iosf_mbi.h>
34 #include <asm/processor.h>
35 #include <asm/cpu_device_id.h>
37 /* bitmasks for RAPL MSRs, used by primitive access functions */
38 #define ENERGY_STATUS_MASK 0xffffffff
40 #define POWER_LIMIT1_MASK 0x7FFF
41 #define POWER_LIMIT1_ENABLE BIT(15)
42 #define POWER_LIMIT1_CLAMP BIT(16)
44 #define POWER_LIMIT2_MASK (0x7FFFULL<<32)
45 #define POWER_LIMIT2_ENABLE BIT_ULL(47)
46 #define POWER_LIMIT2_CLAMP BIT_ULL(48)
47 #define POWER_PACKAGE_LOCK BIT_ULL(63)
48 #define POWER_PP_LOCK BIT(31)
50 #define TIME_WINDOW1_MASK (0x7FULL<<17)
51 #define TIME_WINDOW2_MASK (0x7FULL<<49)
53 #define POWER_UNIT_OFFSET 0
54 #define POWER_UNIT_MASK 0x0F
56 #define ENERGY_UNIT_OFFSET 0x08
57 #define ENERGY_UNIT_MASK 0x1F00
59 #define TIME_UNIT_OFFSET 0x10
60 #define TIME_UNIT_MASK 0xF0000
62 #define POWER_INFO_MAX_MASK (0x7fffULL<<32)
63 #define POWER_INFO_MIN_MASK (0x7fffULL<<16)
64 #define POWER_INFO_MAX_TIME_WIN_MASK (0x3fULL<<48)
65 #define POWER_INFO_THERMAL_SPEC_MASK 0x7fff
67 #define PERF_STATUS_THROTTLE_TIME_MASK 0xffffffff
68 #define PP_POLICY_MASK 0x1F
70 /* Non HW constants */
71 #define RAPL_PRIMITIVE_DERIVED BIT(1) /* not from raw data */
72 #define RAPL_PRIMITIVE_DUMMY BIT(2)
74 #define TIME_WINDOW_MAX_MSEC 40000
75 #define TIME_WINDOW_MIN_MSEC 250
76 #define ENERGY_UNIT_SCALE 1000 /* scale from driver unit to powercap unit */
78 ARBITRARY_UNIT
, /* no translation */
84 enum rapl_domain_type
{
85 RAPL_DOMAIN_PACKAGE
, /* entire package/socket */
86 RAPL_DOMAIN_PP0
, /* core power plane */
87 RAPL_DOMAIN_PP1
, /* graphics uncore */
88 RAPL_DOMAIN_DRAM
,/* DRAM control_type */
92 enum rapl_domain_msr_id
{
93 RAPL_DOMAIN_MSR_LIMIT
,
94 RAPL_DOMAIN_MSR_STATUS
,
96 RAPL_DOMAIN_MSR_POLICY
,
101 /* per domain data, some are optional */
102 enum rapl_primitives
{
108 PL1_ENABLE
, /* power limit 1, aka long term */
109 PL1_CLAMP
, /* allow frequency to go below OS request */
110 PL2_ENABLE
, /* power limit 2, aka short term, instantaneous */
113 TIME_WINDOW1
, /* long term */
114 TIME_WINDOW2
, /* short term */
123 /* below are not raw primitive data */
128 #define NR_RAW_PRIMITIVES (NR_RAPL_PRIMITIVES - 2)
130 /* Can be expanded to include events, etc.*/
131 struct rapl_domain_data
{
132 u64 primitives
[NR_RAPL_PRIMITIVES
];
133 unsigned long timestamp
;
137 #define DOMAIN_STATE_INACTIVE BIT(0)
138 #define DOMAIN_STATE_POWER_LIMIT_SET BIT(1)
139 #define DOMAIN_STATE_BIOS_LOCKED BIT(2)
141 #define NR_POWER_LIMITS (2)
142 struct rapl_power_limit
{
143 struct powercap_zone_constraint
*constraint
;
144 int prim_id
; /* primitive ID used to enable */
145 struct rapl_domain
*domain
;
149 static const char pl1_name
[] = "long_term";
150 static const char pl2_name
[] = "short_term";
154 enum rapl_domain_type id
;
155 int msrs
[RAPL_DOMAIN_MSR_MAX
];
156 struct powercap_zone power_zone
;
157 struct rapl_domain_data rdd
;
158 struct rapl_power_limit rpl
[NR_POWER_LIMITS
];
159 u64 attr_map
; /* track capabilities */
161 unsigned int domain_energy_unit
;
164 #define power_zone_to_rapl_domain(_zone) \
165 container_of(_zone, struct rapl_domain, power_zone)
168 /* Each physical package contains multiple domains, these are the common
169 * data across RAPL domains within a package.
171 struct rapl_package
{
172 unsigned int id
; /* physical package/socket id */
173 unsigned int nr_domains
;
174 unsigned long domain_map
; /* bit map of active domains */
175 unsigned int power_unit
;
176 unsigned int energy_unit
;
177 unsigned int time_unit
;
178 struct rapl_domain
*domains
; /* array of domains, sized at runtime */
179 struct powercap_zone
*power_zone
; /* keep track of parent zone */
180 int nr_cpus
; /* active cpus on the package, topology info is lost during
181 * cpu hotplug. so we have to track ourselves.
183 unsigned long power_limit_irq
; /* keep track of package power limit
184 * notify interrupt enable status.
186 struct list_head plist
;
189 struct rapl_defaults
{
190 int (*check_unit
)(struct rapl_package
*rp
, int cpu
);
191 void (*set_floor_freq
)(struct rapl_domain
*rd
, bool mode
);
192 u64 (*compute_time_window
)(struct rapl_package
*rp
, u64 val
,
194 unsigned int dram_domain_energy_unit
;
196 static struct rapl_defaults
*rapl_defaults
;
198 /* Sideband MBI registers */
199 #define IOSF_CPU_POWER_BUDGET_CTL (0x2)
201 #define PACKAGE_PLN_INT_SAVED BIT(0)
202 #define MAX_PRIM_NAME (32)
204 /* per domain data. used to describe individual knobs such that access function
205 * can be consolidated into one instead of many inline functions.
207 struct rapl_primitive_info
{
211 enum rapl_domain_msr_id id
;
216 #define PRIMITIVE_INFO_INIT(p, m, s, i, u, f) { \
225 static void rapl_init_domains(struct rapl_package
*rp
);
226 static int rapl_read_data_raw(struct rapl_domain
*rd
,
227 enum rapl_primitives prim
,
228 bool xlate
, u64
*data
);
229 static int rapl_write_data_raw(struct rapl_domain
*rd
,
230 enum rapl_primitives prim
,
231 unsigned long long value
);
232 static u64
rapl_unit_xlate(struct rapl_domain
*rd
, int package
,
233 enum unit_type type
, u64 value
,
235 static void package_power_limit_irq_save(int package_id
);
237 static LIST_HEAD(rapl_packages
); /* guarded by CPU hotplug lock */
239 static const char * const rapl_domain_names
[] = {
246 static struct powercap_control_type
*control_type
; /* PowerCap Controller */
248 /* caller to ensure CPU hotplug lock is held */
249 static struct rapl_package
*find_package_by_id(int id
)
251 struct rapl_package
*rp
;
253 list_for_each_entry(rp
, &rapl_packages
, plist
) {
261 /* caller to ensure CPU hotplug lock is held */
262 static int find_active_cpu_on_package(int package_id
)
266 for_each_online_cpu(i
) {
267 if (topology_physical_package_id(i
) == package_id
)
270 /* all CPUs on this package are offline */
275 /* caller must hold cpu hotplug lock */
276 static void rapl_cleanup_data(void)
278 struct rapl_package
*p
, *tmp
;
280 list_for_each_entry_safe(p
, tmp
, &rapl_packages
, plist
) {
287 static int get_energy_counter(struct powercap_zone
*power_zone
, u64
*energy_raw
)
289 struct rapl_domain
*rd
;
292 /* prevent CPU hotplug, make sure the RAPL domain does not go
293 * away while reading the counter.
296 rd
= power_zone_to_rapl_domain(power_zone
);
298 if (!rapl_read_data_raw(rd
, ENERGY_COUNTER
, true, &energy_now
)) {
299 *energy_raw
= energy_now
;
309 static int get_max_energy_counter(struct powercap_zone
*pcd_dev
, u64
*energy
)
311 struct rapl_domain
*rd
= power_zone_to_rapl_domain(pcd_dev
);
313 *energy
= rapl_unit_xlate(rd
, 0, ENERGY_UNIT
, ENERGY_STATUS_MASK
, 0);
317 static int release_zone(struct powercap_zone
*power_zone
)
319 struct rapl_domain
*rd
= power_zone_to_rapl_domain(power_zone
);
320 struct rapl_package
*rp
;
322 /* package zone is the last zone of a package, we can free
323 * memory here since all children has been unregistered.
325 if (rd
->id
== RAPL_DOMAIN_PACKAGE
) {
326 rp
= find_package_by_id(rd
->package_id
);
328 dev_warn(&power_zone
->dev
, "no package id %s\n",
340 static int find_nr_power_limit(struct rapl_domain
*rd
)
344 for (i
= 0; i
< NR_POWER_LIMITS
; i
++) {
345 if (rd
->rpl
[i
].name
== NULL
)
352 static int set_domain_enable(struct powercap_zone
*power_zone
, bool mode
)
354 struct rapl_domain
*rd
= power_zone_to_rapl_domain(power_zone
);
356 if (rd
->state
& DOMAIN_STATE_BIOS_LOCKED
)
360 rapl_write_data_raw(rd
, PL1_ENABLE
, mode
);
361 rapl_defaults
->set_floor_freq(rd
, mode
);
367 static int get_domain_enable(struct powercap_zone
*power_zone
, bool *mode
)
369 struct rapl_domain
*rd
= power_zone_to_rapl_domain(power_zone
);
372 if (rd
->state
& DOMAIN_STATE_BIOS_LOCKED
) {
377 if (rapl_read_data_raw(rd
, PL1_ENABLE
, true, &val
)) {
387 /* per RAPL domain ops, in the order of rapl_domain_type */
388 static struct powercap_zone_ops zone_ops
[] = {
389 /* RAPL_DOMAIN_PACKAGE */
391 .get_energy_uj
= get_energy_counter
,
392 .get_max_energy_range_uj
= get_max_energy_counter
,
393 .release
= release_zone
,
394 .set_enable
= set_domain_enable
,
395 .get_enable
= get_domain_enable
,
397 /* RAPL_DOMAIN_PP0 */
399 .get_energy_uj
= get_energy_counter
,
400 .get_max_energy_range_uj
= get_max_energy_counter
,
401 .release
= release_zone
,
402 .set_enable
= set_domain_enable
,
403 .get_enable
= get_domain_enable
,
405 /* RAPL_DOMAIN_PP1 */
407 .get_energy_uj
= get_energy_counter
,
408 .get_max_energy_range_uj
= get_max_energy_counter
,
409 .release
= release_zone
,
410 .set_enable
= set_domain_enable
,
411 .get_enable
= get_domain_enable
,
413 /* RAPL_DOMAIN_DRAM */
415 .get_energy_uj
= get_energy_counter
,
416 .get_max_energy_range_uj
= get_max_energy_counter
,
417 .release
= release_zone
,
418 .set_enable
= set_domain_enable
,
419 .get_enable
= get_domain_enable
,
423 static int set_power_limit(struct powercap_zone
*power_zone
, int id
,
426 struct rapl_domain
*rd
;
427 struct rapl_package
*rp
;
431 rd
= power_zone_to_rapl_domain(power_zone
);
432 rp
= find_package_by_id(rd
->package_id
);
438 if (rd
->state
& DOMAIN_STATE_BIOS_LOCKED
) {
439 dev_warn(&power_zone
->dev
, "%s locked by BIOS, monitoring only\n",
445 switch (rd
->rpl
[id
].prim_id
) {
447 rapl_write_data_raw(rd
, POWER_LIMIT1
, power_limit
);
450 rapl_write_data_raw(rd
, POWER_LIMIT2
, power_limit
);
456 package_power_limit_irq_save(rd
->package_id
);
462 static int get_current_power_limit(struct powercap_zone
*power_zone
, int id
,
465 struct rapl_domain
*rd
;
471 rd
= power_zone_to_rapl_domain(power_zone
);
472 switch (rd
->rpl
[id
].prim_id
) {
483 if (rapl_read_data_raw(rd
, prim
, true, &val
))
493 static int set_time_window(struct powercap_zone
*power_zone
, int id
,
496 struct rapl_domain
*rd
;
500 rd
= power_zone_to_rapl_domain(power_zone
);
501 switch (rd
->rpl
[id
].prim_id
) {
503 rapl_write_data_raw(rd
, TIME_WINDOW1
, window
);
506 rapl_write_data_raw(rd
, TIME_WINDOW2
, window
);
515 static int get_time_window(struct powercap_zone
*power_zone
, int id
, u64
*data
)
517 struct rapl_domain
*rd
;
522 rd
= power_zone_to_rapl_domain(power_zone
);
523 switch (rd
->rpl
[id
].prim_id
) {
525 ret
= rapl_read_data_raw(rd
, TIME_WINDOW1
, true, &val
);
528 ret
= rapl_read_data_raw(rd
, TIME_WINDOW2
, true, &val
);
541 static const char *get_constraint_name(struct powercap_zone
*power_zone
, int id
)
543 struct rapl_power_limit
*rpl
;
544 struct rapl_domain
*rd
;
546 rd
= power_zone_to_rapl_domain(power_zone
);
547 rpl
= (struct rapl_power_limit
*) &rd
->rpl
[id
];
553 static int get_max_power(struct powercap_zone
*power_zone
, int id
,
556 struct rapl_domain
*rd
;
562 rd
= power_zone_to_rapl_domain(power_zone
);
563 switch (rd
->rpl
[id
].prim_id
) {
565 prim
= THERMAL_SPEC_POWER
;
574 if (rapl_read_data_raw(rd
, prim
, true, &val
))
584 static struct powercap_zone_constraint_ops constraint_ops
= {
585 .set_power_limit_uw
= set_power_limit
,
586 .get_power_limit_uw
= get_current_power_limit
,
587 .set_time_window_us
= set_time_window
,
588 .get_time_window_us
= get_time_window
,
589 .get_max_power_uw
= get_max_power
,
590 .get_name
= get_constraint_name
,
593 /* called after domain detection and package level data are set */
594 static void rapl_init_domains(struct rapl_package
*rp
)
597 struct rapl_domain
*rd
= rp
->domains
;
599 for (i
= 0; i
< RAPL_DOMAIN_MAX
; i
++) {
600 unsigned int mask
= rp
->domain_map
& (1 << i
);
602 case BIT(RAPL_DOMAIN_PACKAGE
):
603 rd
->name
= rapl_domain_names
[RAPL_DOMAIN_PACKAGE
];
604 rd
->id
= RAPL_DOMAIN_PACKAGE
;
605 rd
->msrs
[0] = MSR_PKG_POWER_LIMIT
;
606 rd
->msrs
[1] = MSR_PKG_ENERGY_STATUS
;
607 rd
->msrs
[2] = MSR_PKG_PERF_STATUS
;
609 rd
->msrs
[4] = MSR_PKG_POWER_INFO
;
610 rd
->rpl
[0].prim_id
= PL1_ENABLE
;
611 rd
->rpl
[0].name
= pl1_name
;
612 rd
->rpl
[1].prim_id
= PL2_ENABLE
;
613 rd
->rpl
[1].name
= pl2_name
;
615 case BIT(RAPL_DOMAIN_PP0
):
616 rd
->name
= rapl_domain_names
[RAPL_DOMAIN_PP0
];
617 rd
->id
= RAPL_DOMAIN_PP0
;
618 rd
->msrs
[0] = MSR_PP0_POWER_LIMIT
;
619 rd
->msrs
[1] = MSR_PP0_ENERGY_STATUS
;
621 rd
->msrs
[3] = MSR_PP0_POLICY
;
623 rd
->rpl
[0].prim_id
= PL1_ENABLE
;
624 rd
->rpl
[0].name
= pl1_name
;
626 case BIT(RAPL_DOMAIN_PP1
):
627 rd
->name
= rapl_domain_names
[RAPL_DOMAIN_PP1
];
628 rd
->id
= RAPL_DOMAIN_PP1
;
629 rd
->msrs
[0] = MSR_PP1_POWER_LIMIT
;
630 rd
->msrs
[1] = MSR_PP1_ENERGY_STATUS
;
632 rd
->msrs
[3] = MSR_PP1_POLICY
;
634 rd
->rpl
[0].prim_id
= PL1_ENABLE
;
635 rd
->rpl
[0].name
= pl1_name
;
637 case BIT(RAPL_DOMAIN_DRAM
):
638 rd
->name
= rapl_domain_names
[RAPL_DOMAIN_DRAM
];
639 rd
->id
= RAPL_DOMAIN_DRAM
;
640 rd
->msrs
[0] = MSR_DRAM_POWER_LIMIT
;
641 rd
->msrs
[1] = MSR_DRAM_ENERGY_STATUS
;
642 rd
->msrs
[2] = MSR_DRAM_PERF_STATUS
;
644 rd
->msrs
[4] = MSR_DRAM_POWER_INFO
;
645 rd
->rpl
[0].prim_id
= PL1_ENABLE
;
646 rd
->rpl
[0].name
= pl1_name
;
647 rd
->domain_energy_unit
=
648 rapl_defaults
->dram_domain_energy_unit
;
649 if (rd
->domain_energy_unit
)
650 pr_info("DRAM domain energy unit %dpj\n",
651 rd
->domain_energy_unit
);
655 rd
->package_id
= rp
->id
;
661 static u64
rapl_unit_xlate(struct rapl_domain
*rd
, int package
,
662 enum unit_type type
, u64 value
,
666 struct rapl_package
*rp
;
669 rp
= find_package_by_id(package
);
675 units
= rp
->power_unit
;
678 scale
= ENERGY_UNIT_SCALE
;
679 /* per domain unit takes precedence */
680 if (rd
&& rd
->domain_energy_unit
)
681 units
= rd
->domain_energy_unit
;
683 units
= rp
->energy_unit
;
686 return rapl_defaults
->compute_time_window(rp
, value
, to_raw
);
693 return div64_u64(value
, units
) * scale
;
697 return div64_u64(value
, scale
);
700 /* in the order of enum rapl_primitives */
701 static struct rapl_primitive_info rpi
[] = {
702 /* name, mask, shift, msr index, unit divisor */
703 PRIMITIVE_INFO_INIT(ENERGY_COUNTER
, ENERGY_STATUS_MASK
, 0,
704 RAPL_DOMAIN_MSR_STATUS
, ENERGY_UNIT
, 0),
705 PRIMITIVE_INFO_INIT(POWER_LIMIT1
, POWER_LIMIT1_MASK
, 0,
706 RAPL_DOMAIN_MSR_LIMIT
, POWER_UNIT
, 0),
707 PRIMITIVE_INFO_INIT(POWER_LIMIT2
, POWER_LIMIT2_MASK
, 32,
708 RAPL_DOMAIN_MSR_LIMIT
, POWER_UNIT
, 0),
709 PRIMITIVE_INFO_INIT(FW_LOCK
, POWER_PP_LOCK
, 31,
710 RAPL_DOMAIN_MSR_LIMIT
, ARBITRARY_UNIT
, 0),
711 PRIMITIVE_INFO_INIT(PL1_ENABLE
, POWER_LIMIT1_ENABLE
, 15,
712 RAPL_DOMAIN_MSR_LIMIT
, ARBITRARY_UNIT
, 0),
713 PRIMITIVE_INFO_INIT(PL1_CLAMP
, POWER_LIMIT1_CLAMP
, 16,
714 RAPL_DOMAIN_MSR_LIMIT
, ARBITRARY_UNIT
, 0),
715 PRIMITIVE_INFO_INIT(PL2_ENABLE
, POWER_LIMIT2_ENABLE
, 47,
716 RAPL_DOMAIN_MSR_LIMIT
, ARBITRARY_UNIT
, 0),
717 PRIMITIVE_INFO_INIT(PL2_CLAMP
, POWER_LIMIT2_CLAMP
, 48,
718 RAPL_DOMAIN_MSR_LIMIT
, ARBITRARY_UNIT
, 0),
719 PRIMITIVE_INFO_INIT(TIME_WINDOW1
, TIME_WINDOW1_MASK
, 17,
720 RAPL_DOMAIN_MSR_LIMIT
, TIME_UNIT
, 0),
721 PRIMITIVE_INFO_INIT(TIME_WINDOW2
, TIME_WINDOW2_MASK
, 49,
722 RAPL_DOMAIN_MSR_LIMIT
, TIME_UNIT
, 0),
723 PRIMITIVE_INFO_INIT(THERMAL_SPEC_POWER
, POWER_INFO_THERMAL_SPEC_MASK
,
724 0, RAPL_DOMAIN_MSR_INFO
, POWER_UNIT
, 0),
725 PRIMITIVE_INFO_INIT(MAX_POWER
, POWER_INFO_MAX_MASK
, 32,
726 RAPL_DOMAIN_MSR_INFO
, POWER_UNIT
, 0),
727 PRIMITIVE_INFO_INIT(MIN_POWER
, POWER_INFO_MIN_MASK
, 16,
728 RAPL_DOMAIN_MSR_INFO
, POWER_UNIT
, 0),
729 PRIMITIVE_INFO_INIT(MAX_TIME_WINDOW
, POWER_INFO_MAX_TIME_WIN_MASK
, 48,
730 RAPL_DOMAIN_MSR_INFO
, TIME_UNIT
, 0),
731 PRIMITIVE_INFO_INIT(THROTTLED_TIME
, PERF_STATUS_THROTTLE_TIME_MASK
, 0,
732 RAPL_DOMAIN_MSR_PERF
, TIME_UNIT
, 0),
733 PRIMITIVE_INFO_INIT(PRIORITY_LEVEL
, PP_POLICY_MASK
, 0,
734 RAPL_DOMAIN_MSR_POLICY
, ARBITRARY_UNIT
, 0),
736 PRIMITIVE_INFO_INIT(AVERAGE_POWER
, 0, 0, 0, POWER_UNIT
,
737 RAPL_PRIMITIVE_DERIVED
),
741 /* Read primitive data based on its related struct rapl_primitive_info.
742 * if xlate flag is set, return translated data based on data units, i.e.
743 * time, energy, and power.
744 * RAPL MSRs are non-architectual and are laid out not consistently across
745 * domains. Here we use primitive info to allow writing consolidated access
747 * For a given primitive, it is processed by MSR mask and shift. Unit conversion
748 * is pre-assigned based on RAPL unit MSRs read at init time.
749 * 63-------------------------- 31--------------------------- 0
751 * | |<- shift ----------------|
752 * 63-------------------------- 31--------------------------- 0
754 static int rapl_read_data_raw(struct rapl_domain
*rd
,
755 enum rapl_primitives prim
,
756 bool xlate
, u64
*data
)
760 struct rapl_primitive_info
*rp
= &rpi
[prim
];
763 if (!rp
->name
|| rp
->flag
& RAPL_PRIMITIVE_DUMMY
)
766 msr
= rd
->msrs
[rp
->id
];
769 /* use physical package id to look up active cpus */
770 cpu
= find_active_cpu_on_package(rd
->package_id
);
774 /* special-case package domain, which uses a different bit*/
775 if (prim
== FW_LOCK
&& rd
->id
== RAPL_DOMAIN_PACKAGE
) {
776 rp
->mask
= POWER_PACKAGE_LOCK
;
779 /* non-hardware data are collected by the polling thread */
780 if (rp
->flag
& RAPL_PRIMITIVE_DERIVED
) {
781 *data
= rd
->rdd
.primitives
[prim
];
785 if (rdmsrl_safe_on_cpu(cpu
, msr
, &value
)) {
786 pr_debug("failed to read msr 0x%x on cpu %d\n", msr
, cpu
);
790 final
= value
& rp
->mask
;
791 final
= final
>> rp
->shift
;
793 *data
= rapl_unit_xlate(rd
, rd
->package_id
, rp
->unit
, final
, 0);
800 /* Similar use of primitive info in the read counterpart */
801 static int rapl_write_data_raw(struct rapl_domain
*rd
,
802 enum rapl_primitives prim
,
803 unsigned long long value
)
807 struct rapl_primitive_info
*rp
= &rpi
[prim
];
810 cpu
= find_active_cpu_on_package(rd
->package_id
);
813 msr
= rd
->msrs
[rp
->id
];
814 if (rdmsrl_safe_on_cpu(cpu
, msr
, &msr_val
)) {
815 dev_dbg(&rd
->power_zone
.dev
,
816 "failed to read msr 0x%x on cpu %d\n", msr
, cpu
);
819 value
= rapl_unit_xlate(rd
, rd
->package_id
, rp
->unit
, value
, 1);
820 msr_val
&= ~rp
->mask
;
821 msr_val
|= value
<< rp
->shift
;
822 if (wrmsrl_safe_on_cpu(cpu
, msr
, msr_val
)) {
823 dev_dbg(&rd
->power_zone
.dev
,
824 "failed to write msr 0x%x on cpu %d\n", msr
, cpu
);
832 * Raw RAPL data stored in MSRs are in certain scales. We need to
833 * convert them into standard units based on the units reported in
834 * the RAPL unit MSRs. This is specific to CPUs as the method to
835 * calculate units differ on different CPUs.
836 * We convert the units to below format based on CPUs.
838 * energy unit: picoJoules : Represented in picoJoules by default
839 * power unit : microWatts : Represented in milliWatts by default
840 * time unit : microseconds: Represented in seconds by default
842 static int rapl_check_unit_core(struct rapl_package
*rp
, int cpu
)
847 if (rdmsrl_safe_on_cpu(cpu
, MSR_RAPL_POWER_UNIT
, &msr_val
)) {
848 pr_err("Failed to read power unit MSR 0x%x on CPU %d, exit.\n",
849 MSR_RAPL_POWER_UNIT
, cpu
);
853 value
= (msr_val
& ENERGY_UNIT_MASK
) >> ENERGY_UNIT_OFFSET
;
854 rp
->energy_unit
= ENERGY_UNIT_SCALE
* 1000000 / (1 << value
);
856 value
= (msr_val
& POWER_UNIT_MASK
) >> POWER_UNIT_OFFSET
;
857 rp
->power_unit
= 1000000 / (1 << value
);
859 value
= (msr_val
& TIME_UNIT_MASK
) >> TIME_UNIT_OFFSET
;
860 rp
->time_unit
= 1000000 / (1 << value
);
862 pr_debug("Core CPU package %d energy=%dpJ, time=%dus, power=%duW\n",
863 rp
->id
, rp
->energy_unit
, rp
->time_unit
, rp
->power_unit
);
868 static int rapl_check_unit_atom(struct rapl_package
*rp
, int cpu
)
873 if (rdmsrl_safe_on_cpu(cpu
, MSR_RAPL_POWER_UNIT
, &msr_val
)) {
874 pr_err("Failed to read power unit MSR 0x%x on CPU %d, exit.\n",
875 MSR_RAPL_POWER_UNIT
, cpu
);
878 value
= (msr_val
& ENERGY_UNIT_MASK
) >> ENERGY_UNIT_OFFSET
;
879 rp
->energy_unit
= ENERGY_UNIT_SCALE
* 1 << value
;
881 value
= (msr_val
& POWER_UNIT_MASK
) >> POWER_UNIT_OFFSET
;
882 rp
->power_unit
= (1 << value
) * 1000;
884 value
= (msr_val
& TIME_UNIT_MASK
) >> TIME_UNIT_OFFSET
;
885 rp
->time_unit
= 1000000 / (1 << value
);
887 pr_debug("Atom package %d energy=%dpJ, time=%dus, power=%duW\n",
888 rp
->id
, rp
->energy_unit
, rp
->time_unit
, rp
->power_unit
);
895 * When package power limit is set artificially low by RAPL, LVT
896 * thermal interrupt for package power limit should be ignored
897 * since we are not really exceeding the real limit. The intention
898 * is to avoid excessive interrupts while we are trying to save power.
899 * A useful feature might be routing the package_power_limit interrupt
900 * to userspace via eventfd. once we have a usecase, this is simple
901 * to do by adding an atomic notifier.
904 static void package_power_limit_irq_save(int package_id
)
908 struct rapl_package
*rp
;
910 rp
= find_package_by_id(package_id
);
914 if (!boot_cpu_has(X86_FEATURE_PTS
) || !boot_cpu_has(X86_FEATURE_PLN
))
917 cpu
= find_active_cpu_on_package(package_id
);
920 /* save the state of PLN irq mask bit before disabling it */
921 rdmsr_safe_on_cpu(cpu
, MSR_IA32_PACKAGE_THERM_INTERRUPT
, &l
, &h
);
922 if (!(rp
->power_limit_irq
& PACKAGE_PLN_INT_SAVED
)) {
923 rp
->power_limit_irq
= l
& PACKAGE_THERM_INT_PLN_ENABLE
;
924 rp
->power_limit_irq
|= PACKAGE_PLN_INT_SAVED
;
926 l
&= ~PACKAGE_THERM_INT_PLN_ENABLE
;
927 wrmsr_on_cpu(cpu
, MSR_IA32_PACKAGE_THERM_INTERRUPT
, l
, h
);
930 /* restore per package power limit interrupt enable state */
931 static void package_power_limit_irq_restore(int package_id
)
935 struct rapl_package
*rp
;
937 rp
= find_package_by_id(package_id
);
941 if (!boot_cpu_has(X86_FEATURE_PTS
) || !boot_cpu_has(X86_FEATURE_PLN
))
944 cpu
= find_active_cpu_on_package(package_id
);
948 /* irq enable state not saved, nothing to restore */
949 if (!(rp
->power_limit_irq
& PACKAGE_PLN_INT_SAVED
))
951 rdmsr_safe_on_cpu(cpu
, MSR_IA32_PACKAGE_THERM_INTERRUPT
, &l
, &h
);
953 if (rp
->power_limit_irq
& PACKAGE_THERM_INT_PLN_ENABLE
)
954 l
|= PACKAGE_THERM_INT_PLN_ENABLE
;
956 l
&= ~PACKAGE_THERM_INT_PLN_ENABLE
;
958 wrmsr_on_cpu(cpu
, MSR_IA32_PACKAGE_THERM_INTERRUPT
, l
, h
);
961 static void set_floor_freq_default(struct rapl_domain
*rd
, bool mode
)
963 int nr_powerlimit
= find_nr_power_limit(rd
);
965 /* always enable clamp such that p-state can go below OS requested
966 * range. power capping priority over guranteed frequency.
968 rapl_write_data_raw(rd
, PL1_CLAMP
, mode
);
970 /* some domains have pl2 */
971 if (nr_powerlimit
> 1) {
972 rapl_write_data_raw(rd
, PL2_ENABLE
, mode
);
973 rapl_write_data_raw(rd
, PL2_CLAMP
, mode
);
977 static void set_floor_freq_atom(struct rapl_domain
*rd
, bool enable
)
979 static u32 power_ctrl_orig_val
;
982 if (!power_ctrl_orig_val
)
983 iosf_mbi_read(BT_MBI_UNIT_PMC
, BT_MBI_PMC_READ
,
984 IOSF_CPU_POWER_BUDGET_CTL
, &power_ctrl_orig_val
);
985 mdata
= power_ctrl_orig_val
;
987 mdata
&= ~(0x7f << 8);
990 iosf_mbi_write(BT_MBI_UNIT_PMC
, BT_MBI_PMC_WRITE
,
991 IOSF_CPU_POWER_BUDGET_CTL
, mdata
);
994 static u64
rapl_compute_time_window_core(struct rapl_package
*rp
, u64 value
,
997 u64 f
, y
; /* fraction and exp. used for time unit */
1000 * Special processing based on 2^Y*(1+F/4), refer
1001 * to Intel Software Developer's manual Vol.3B: CH 14.9.3.
1004 f
= (value
& 0x60) >> 5;
1006 value
= (1 << y
) * (4 + f
) * rp
->time_unit
/ 4;
1008 do_div(value
, rp
->time_unit
);
1010 f
= div64_u64(4 * (value
- (1 << y
)), 1 << y
);
1011 value
= (y
& 0x1f) | ((f
& 0x3) << 5);
1016 static u64
rapl_compute_time_window_atom(struct rapl_package
*rp
, u64 value
,
1020 * Atom time unit encoding is straight forward val * time_unit,
1021 * where time_unit is default to 1 sec. Never 0.
1024 return (value
) ? value
*= rp
->time_unit
: rp
->time_unit
;
1026 value
= div64_u64(value
, rp
->time_unit
);
1031 static const struct rapl_defaults rapl_defaults_core
= {
1032 .check_unit
= rapl_check_unit_core
,
1033 .set_floor_freq
= set_floor_freq_default
,
1034 .compute_time_window
= rapl_compute_time_window_core
,
1037 static const struct rapl_defaults rapl_defaults_hsw_server
= {
1038 .check_unit
= rapl_check_unit_core
,
1039 .set_floor_freq
= set_floor_freq_default
,
1040 .compute_time_window
= rapl_compute_time_window_core
,
1041 .dram_domain_energy_unit
= 15300,
1044 static const struct rapl_defaults rapl_defaults_atom
= {
1045 .check_unit
= rapl_check_unit_atom
,
1046 .set_floor_freq
= set_floor_freq_atom
,
1047 .compute_time_window
= rapl_compute_time_window_atom
,
1050 #define RAPL_CPU(_model, _ops) { \
1051 .vendor = X86_VENDOR_INTEL, \
1054 .driver_data = (kernel_ulong_t)&_ops, \
1057 static const struct x86_cpu_id rapl_ids
[] = {
1058 RAPL_CPU(0x2a, rapl_defaults_core
),/* Sandy Bridge */
1059 RAPL_CPU(0x2d, rapl_defaults_core
),/* Sandy Bridge EP */
1060 RAPL_CPU(0x37, rapl_defaults_atom
),/* Valleyview */
1061 RAPL_CPU(0x3a, rapl_defaults_core
),/* Ivy Bridge */
1062 RAPL_CPU(0x3c, rapl_defaults_core
),/* Haswell */
1063 RAPL_CPU(0x3d, rapl_defaults_core
),/* Broadwell */
1064 RAPL_CPU(0x3f, rapl_defaults_hsw_server
),/* Haswell servers */
1065 RAPL_CPU(0x45, rapl_defaults_core
),/* Haswell ULT */
1066 RAPL_CPU(0x4C, rapl_defaults_atom
),/* Braswell */
1067 RAPL_CPU(0x4A, rapl_defaults_atom
),/* Tangier */
1068 RAPL_CPU(0x56, rapl_defaults_core
),/* Future Xeon */
1069 RAPL_CPU(0x5A, rapl_defaults_atom
),/* Annidale */
1072 MODULE_DEVICE_TABLE(x86cpu
, rapl_ids
);
1074 /* read once for all raw primitive data for all packages, domains */
1075 static void rapl_update_domain_data(void)
1079 struct rapl_package
*rp
;
1081 list_for_each_entry(rp
, &rapl_packages
, plist
) {
1082 for (dmn
= 0; dmn
< rp
->nr_domains
; dmn
++) {
1083 pr_debug("update package %d domain %s data\n", rp
->id
,
1084 rp
->domains
[dmn
].name
);
1085 /* exclude non-raw primitives */
1086 for (prim
= 0; prim
< NR_RAW_PRIMITIVES
; prim
++)
1087 if (!rapl_read_data_raw(&rp
->domains
[dmn
], prim
,
1090 rp
->domains
[dmn
].rdd
.primitives
[prim
] =
1097 static int rapl_unregister_powercap(void)
1099 struct rapl_package
*rp
;
1100 struct rapl_domain
*rd
, *rd_package
= NULL
;
1102 /* unregister all active rapl packages from the powercap layer,
1105 list_for_each_entry(rp
, &rapl_packages
, plist
) {
1106 package_power_limit_irq_restore(rp
->id
);
1108 for (rd
= rp
->domains
; rd
< rp
->domains
+ rp
->nr_domains
;
1110 pr_debug("remove package, undo power limit on %d: %s\n",
1112 rapl_write_data_raw(rd
, PL1_ENABLE
, 0);
1113 rapl_write_data_raw(rd
, PL2_ENABLE
, 0);
1114 rapl_write_data_raw(rd
, PL1_CLAMP
, 0);
1115 rapl_write_data_raw(rd
, PL2_CLAMP
, 0);
1116 if (rd
->id
== RAPL_DOMAIN_PACKAGE
) {
1120 powercap_unregister_zone(control_type
, &rd
->power_zone
);
1122 /* do the package zone last */
1124 powercap_unregister_zone(control_type
,
1125 &rd_package
->power_zone
);
1127 powercap_unregister_control_type(control_type
);
1132 static int rapl_package_register_powercap(struct rapl_package
*rp
)
1134 struct rapl_domain
*rd
;
1136 char dev_name
[17]; /* max domain name = 7 + 1 + 8 for int + 1 for null*/
1137 struct powercap_zone
*power_zone
= NULL
;
1140 /* first we register package domain as the parent zone*/
1141 for (rd
= rp
->domains
; rd
< rp
->domains
+ rp
->nr_domains
; rd
++) {
1142 if (rd
->id
== RAPL_DOMAIN_PACKAGE
) {
1143 nr_pl
= find_nr_power_limit(rd
);
1144 pr_debug("register socket %d package domain %s\n",
1146 memset(dev_name
, 0, sizeof(dev_name
));
1147 snprintf(dev_name
, sizeof(dev_name
), "%s-%d",
1149 power_zone
= powercap_register_zone(&rd
->power_zone
,
1155 if (IS_ERR(power_zone
)) {
1156 pr_debug("failed to register package, %d\n",
1158 ret
= PTR_ERR(power_zone
);
1161 /* track parent zone in per package/socket data */
1162 rp
->power_zone
= power_zone
;
1163 /* done, only one package domain per socket */
1168 pr_err("no package domain found, unknown topology!\n");
1172 /* now register domains as children of the socket/package*/
1173 for (rd
= rp
->domains
; rd
< rp
->domains
+ rp
->nr_domains
; rd
++) {
1174 if (rd
->id
== RAPL_DOMAIN_PACKAGE
)
1176 /* number of power limits per domain varies */
1177 nr_pl
= find_nr_power_limit(rd
);
1178 power_zone
= powercap_register_zone(&rd
->power_zone
,
1179 control_type
, rd
->name
,
1181 &zone_ops
[rd
->id
], nr_pl
,
1184 if (IS_ERR(power_zone
)) {
1185 pr_debug("failed to register power_zone, %d:%s:%s\n",
1186 rp
->id
, rd
->name
, dev_name
);
1187 ret
= PTR_ERR(power_zone
);
1195 /* clean up previously initialized domains within the package if we
1196 * failed after the first domain setup.
1198 while (--rd
>= rp
->domains
) {
1199 pr_debug("unregister package %d domain %s\n", rp
->id
, rd
->name
);
1200 powercap_unregister_zone(control_type
, &rd
->power_zone
);
1206 static int rapl_register_powercap(void)
1208 struct rapl_domain
*rd
;
1209 struct rapl_package
*rp
;
1212 control_type
= powercap_register_control_type(NULL
, "intel-rapl", NULL
);
1213 if (IS_ERR(control_type
)) {
1214 pr_debug("failed to register powercap control_type.\n");
1215 return PTR_ERR(control_type
);
1217 /* read the initial data */
1218 rapl_update_domain_data();
1219 list_for_each_entry(rp
, &rapl_packages
, plist
)
1220 if (rapl_package_register_powercap(rp
))
1221 goto err_cleanup_package
;
1224 err_cleanup_package
:
1225 /* clean up previously initialized packages */
1226 list_for_each_entry_continue_reverse(rp
, &rapl_packages
, plist
) {
1227 for (rd
= rp
->domains
; rd
< rp
->domains
+ rp
->nr_domains
;
1229 pr_debug("unregister zone/package %d, %s domain\n",
1231 powercap_unregister_zone(control_type
, &rd
->power_zone
);
1238 static int rapl_check_domain(int cpu
, int domain
)
1244 case RAPL_DOMAIN_PACKAGE
:
1245 msr
= MSR_PKG_ENERGY_STATUS
;
1247 case RAPL_DOMAIN_PP0
:
1248 msr
= MSR_PP0_ENERGY_STATUS
;
1250 case RAPL_DOMAIN_PP1
:
1251 msr
= MSR_PP1_ENERGY_STATUS
;
1253 case RAPL_DOMAIN_DRAM
:
1254 msr
= MSR_DRAM_ENERGY_STATUS
;
1257 pr_err("invalid domain id %d\n", domain
);
1260 /* make sure domain counters are available and contains non-zero
1261 * values, otherwise skip it.
1263 if (rdmsrl_safe_on_cpu(cpu
, msr
, &val
) || !val
)
1269 /* Detect active and valid domains for the given CPU, caller must
1270 * ensure the CPU belongs to the targeted package and CPU hotlug is disabled.
1272 static int rapl_detect_domains(struct rapl_package
*rp
, int cpu
)
1276 struct rapl_domain
*rd
;
1279 for (i
= 0; i
< RAPL_DOMAIN_MAX
; i
++) {
1280 /* use physical package id to read counters */
1281 if (!rapl_check_domain(cpu
, i
)) {
1282 rp
->domain_map
|= 1 << i
;
1283 pr_info("Found RAPL domain %s\n", rapl_domain_names
[i
]);
1286 rp
->nr_domains
= bitmap_weight(&rp
->domain_map
, RAPL_DOMAIN_MAX
);
1287 if (!rp
->nr_domains
) {
1288 pr_err("no valid rapl domains found in package %d\n", rp
->id
);
1292 pr_debug("found %d domains on package %d\n", rp
->nr_domains
, rp
->id
);
1294 rp
->domains
= kcalloc(rp
->nr_domains
+ 1, sizeof(struct rapl_domain
),
1300 rapl_init_domains(rp
);
1302 for (rd
= rp
->domains
; rd
< rp
->domains
+ rp
->nr_domains
; rd
++) {
1303 /* check if the domain is locked by BIOS */
1304 if (rapl_read_data_raw(rd
, FW_LOCK
, false, &locked
)) {
1305 pr_info("RAPL package %d domain %s locked by BIOS\n",
1307 rd
->state
|= DOMAIN_STATE_BIOS_LOCKED
;
1316 static bool is_package_new(int package
)
1318 struct rapl_package
*rp
;
1320 /* caller prevents cpu hotplug, there will be no new packages added
1321 * or deleted while traversing the package list, no need for locking.
1323 list_for_each_entry(rp
, &rapl_packages
, plist
)
1324 if (package
== rp
->id
)
1330 /* RAPL interface can be made of a two-level hierarchy: package level and domain
1331 * level. We first detect the number of packages then domains of each package.
1332 * We have to consider the possiblity of CPU online/offline due to hotplug and
1335 static int rapl_detect_topology(void)
1339 struct rapl_package
*new_package
, *rp
;
1341 for_each_online_cpu(i
) {
1342 phy_package_id
= topology_physical_package_id(i
);
1343 if (is_package_new(phy_package_id
)) {
1344 new_package
= kzalloc(sizeof(*rp
), GFP_KERNEL
);
1346 rapl_cleanup_data();
1349 /* add the new package to the list */
1350 new_package
->id
= phy_package_id
;
1351 new_package
->nr_cpus
= 1;
1353 /* check if the package contains valid domains */
1354 if (rapl_detect_domains(new_package
, i
) ||
1355 rapl_defaults
->check_unit(new_package
, i
)) {
1356 kfree(new_package
->domains
);
1358 /* free up the packages already initialized */
1359 rapl_cleanup_data();
1362 INIT_LIST_HEAD(&new_package
->plist
);
1363 list_add(&new_package
->plist
, &rapl_packages
);
1365 rp
= find_package_by_id(phy_package_id
);
1374 /* called from CPU hotplug notifier, hotplug lock held */
1375 static void rapl_remove_package(struct rapl_package
*rp
)
1377 struct rapl_domain
*rd
, *rd_package
= NULL
;
1379 for (rd
= rp
->domains
; rd
< rp
->domains
+ rp
->nr_domains
; rd
++) {
1380 if (rd
->id
== RAPL_DOMAIN_PACKAGE
) {
1384 pr_debug("remove package %d, %s domain\n", rp
->id
, rd
->name
);
1385 powercap_unregister_zone(control_type
, &rd
->power_zone
);
1387 /* do parent zone last */
1388 powercap_unregister_zone(control_type
, &rd_package
->power_zone
);
1389 list_del(&rp
->plist
);
1393 /* called from CPU hotplug notifier, hotplug lock held */
1394 static int rapl_add_package(int cpu
)
1398 struct rapl_package
*rp
;
1400 phy_package_id
= topology_physical_package_id(cpu
);
1401 rp
= kzalloc(sizeof(struct rapl_package
), GFP_KERNEL
);
1405 /* add the new package to the list */
1406 rp
->id
= phy_package_id
;
1408 /* check if the package contains valid domains */
1409 if (rapl_detect_domains(rp
, cpu
) ||
1410 rapl_defaults
->check_unit(rp
, cpu
)) {
1412 goto err_free_package
;
1414 if (!rapl_package_register_powercap(rp
)) {
1415 INIT_LIST_HEAD(&rp
->plist
);
1416 list_add(&rp
->plist
, &rapl_packages
);
1427 /* Handles CPU hotplug on multi-socket systems.
1428 * If a CPU goes online as the first CPU of the physical package
1429 * we add the RAPL package to the system. Similarly, when the last
1430 * CPU of the package is removed, we remove the RAPL package and its
1431 * associated domains. Cooling devices are handled accordingly at
1434 static int rapl_cpu_callback(struct notifier_block
*nfb
,
1435 unsigned long action
, void *hcpu
)
1437 unsigned long cpu
= (unsigned long)hcpu
;
1439 struct rapl_package
*rp
;
1441 phy_package_id
= topology_physical_package_id(cpu
);
1444 case CPU_ONLINE_FROZEN
:
1445 case CPU_DOWN_FAILED
:
1446 case CPU_DOWN_FAILED_FROZEN
:
1447 rp
= find_package_by_id(phy_package_id
);
1451 rapl_add_package(cpu
);
1453 case CPU_DOWN_PREPARE
:
1454 case CPU_DOWN_PREPARE_FROZEN
:
1455 rp
= find_package_by_id(phy_package_id
);
1458 if (--rp
->nr_cpus
== 0)
1459 rapl_remove_package(rp
);
1465 static struct notifier_block rapl_cpu_notifier
= {
1466 .notifier_call
= rapl_cpu_callback
,
1469 static int __init
rapl_init(void)
1472 const struct x86_cpu_id
*id
;
1474 id
= x86_match_cpu(rapl_ids
);
1476 pr_err("driver does not support CPU family %d model %d\n",
1477 boot_cpu_data
.x86
, boot_cpu_data
.x86_model
);
1482 rapl_defaults
= (struct rapl_defaults
*)id
->driver_data
;
1484 cpu_notifier_register_begin();
1486 /* prevent CPU hotplug during detection */
1488 ret
= rapl_detect_topology();
1492 if (rapl_register_powercap()) {
1493 rapl_cleanup_data();
1497 __register_hotcpu_notifier(&rapl_cpu_notifier
);
1500 cpu_notifier_register_done();
1505 static void __exit
rapl_exit(void)
1507 cpu_notifier_register_begin();
1509 __unregister_hotcpu_notifier(&rapl_cpu_notifier
);
1510 rapl_unregister_powercap();
1511 rapl_cleanup_data();
1513 cpu_notifier_register_done();
1516 module_init(rapl_init
);
1517 module_exit(rapl_exit
);
1519 MODULE_DESCRIPTION("Driver for Intel RAPL (Running Average Power Limit)");
1520 MODULE_AUTHOR("Jacob Pan <jacob.jun.pan@intel.com>");
1521 MODULE_LICENSE("GPL v2");