2 * turbostat -- show CPU frequency and C-state residency
3 * on modern Intel turbo-capable processors.
5 * Copyright (c) 2013 Intel Corporation.
6 * Len Brown <len.brown@intel.com>
8 * This program is free software; you can redistribute it and/or modify it
9 * under the terms and conditions of the GNU General Public License,
10 * version 2, as published by the Free Software Foundation.
12 * This program is distributed in the hope it will be useful, but WITHOUT
13 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
14 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
17 * You should have received a copy of the GNU General Public License along with
18 * this program; if not, write to the Free Software Foundation, Inc.,
19 * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
24 #include INTEL_FAMILY_HEADER
29 #include <sys/types.h>
32 #include <sys/select.h>
33 #include <sys/resource.h>
45 #include <linux/capability.h>
48 char *proc_stat = "/proc/stat";
51 struct timeval interval_tv = {5, 0};
52 struct timespec interval_ts = {5, 0};
53 struct timespec one_msec = {0, 1000000};
54 unsigned int num_iterations;
58 unsigned int sums_need_wide_columns;
59 unsigned int rapl_joules;
60 unsigned int summary_only;
61 unsigned int list_header_only;
62 unsigned int dump_only;
63 unsigned int do_snb_cstates;
64 unsigned int do_knl_cstates;
65 unsigned int do_slm_cstates;
66 unsigned int do_cnl_cstates;
67 unsigned int use_c1_residency_msr;
68 unsigned int has_aperf;
70 unsigned int do_irtl_snb;
71 unsigned int do_irtl_hsw;
72 unsigned int units = 1000000; /* MHz etc */
73 unsigned int genuine_intel;
74 unsigned int has_invariant_tsc;
75 unsigned int do_nhm_platform_info;
76 unsigned int no_MSR_MISC_PWR_MGMT;
77 unsigned int aperf_mperf_multiplier = 1;
80 unsigned int has_base_hz;
81 double tsc_tweak = 1.0;
82 unsigned int show_pkg_only;
83 unsigned int show_core_only;
84 char *output_buffer, *outp;
88 unsigned long long gfx_cur_rc6_ms;
89 unsigned long long cpuidle_cur_cpu_lpi_us;
90 unsigned long long cpuidle_cur_sys_lpi_us;
91 unsigned int gfx_cur_mhz;
92 unsigned int tcc_activation_temp;
93 unsigned int tcc_activation_temp_override;
94 double rapl_power_units, rapl_time_units;
95 double rapl_dram_energy_units, rapl_energy_units;
96 double rapl_joule_counter_range;
97 unsigned int do_core_perf_limit_reasons;
98 unsigned int has_automatic_cstate_conversion;
99 unsigned int do_gfx_perf_limit_reasons;
100 unsigned int do_ring_perf_limit_reasons;
101 unsigned int crystal_hz;
102 unsigned long long tsc_hz;
104 double discover_bclk(unsigned int family, unsigned int model);
105 unsigned int has_hwp; /* IA32_PM_ENABLE, IA32_HWP_CAPABILITIES */
106 /* IA32_HWP_REQUEST, IA32_HWP_STATUS */
107 unsigned int has_hwp_notify; /* IA32_HWP_INTERRUPT */
108 unsigned int has_hwp_activity_window; /* IA32_HWP_REQUEST[bits 41:32] */
109 unsigned int has_hwp_epp; /* IA32_HWP_REQUEST[bits 31:24] */
110 unsigned int has_hwp_pkg; /* IA32_HWP_REQUEST_PKG */
111 unsigned int has_misc_feature_control;
112 unsigned int first_counter_read = 1;
114 #define RAPL_PKG (1 << 0)
115 /* 0x610 MSR_PKG_POWER_LIMIT */
116 /* 0x611 MSR_PKG_ENERGY_STATUS */
117 #define RAPL_PKG_PERF_STATUS (1 << 1)
118 /* 0x613 MSR_PKG_PERF_STATUS */
119 #define RAPL_PKG_POWER_INFO (1 << 2)
120 /* 0x614 MSR_PKG_POWER_INFO */
122 #define RAPL_DRAM (1 << 3)
123 /* 0x618 MSR_DRAM_POWER_LIMIT */
124 /* 0x619 MSR_DRAM_ENERGY_STATUS */
125 #define RAPL_DRAM_PERF_STATUS (1 << 4)
126 /* 0x61b MSR_DRAM_PERF_STATUS */
127 #define RAPL_DRAM_POWER_INFO (1 << 5)
128 /* 0x61c MSR_DRAM_POWER_INFO */
130 #define RAPL_CORES_POWER_LIMIT (1 << 6)
131 /* 0x638 MSR_PP0_POWER_LIMIT */
132 #define RAPL_CORE_POLICY (1 << 7)
133 /* 0x63a MSR_PP0_POLICY */
135 #define RAPL_GFX (1 << 8)
136 /* 0x640 MSR_PP1_POWER_LIMIT */
137 /* 0x641 MSR_PP1_ENERGY_STATUS */
138 /* 0x642 MSR_PP1_POLICY */
140 #define RAPL_CORES_ENERGY_STATUS (1 << 9)
141 /* 0x639 MSR_PP0_ENERGY_STATUS */
142 #define RAPL_CORES (RAPL_CORES_ENERGY_STATUS | RAPL_CORES_POWER_LIMIT)
143 #define TJMAX_DEFAULT 100
145 #define MAX(a, b) ((a) > (b) ? (a) : (b))
148 * buffer size used by sscanf() for added column names
149 * Usually truncated to 7 characters, but also handles 18 columns for raw 64-bit counters
151 #define NAME_BYTES 20
152 #define PATH_BYTES 128
157 #define CPU_SUBSET_MAXCPUS 1024 /* need to use before probe... */
158 cpu_set_t *cpu_present_set, *cpu_affinity_set, *cpu_subset;
159 size_t cpu_present_setsize, cpu_affinity_setsize, cpu_subset_size;
160 #define MAX_ADDED_COUNTERS 8
161 #define MAX_ADDED_THREAD_COUNTERS 24
162 #define BITMASK_SIZE 32
165 struct timeval tv_begin;
166 struct timeval tv_end;
167 unsigned long long tsc;
168 unsigned long long aperf;
169 unsigned long long mperf;
170 unsigned long long c1;
171 unsigned long long irq_count;
172 unsigned int smi_count;
174 unsigned int apic_id;
175 unsigned int x2apic_id;
177 #define CPU_IS_FIRST_THREAD_IN_CORE 0x2
178 #define CPU_IS_FIRST_CORE_IN_PACKAGE 0x4
179 unsigned long long counter[MAX_ADDED_THREAD_COUNTERS];
180 } *thread_even, *thread_odd;
183 unsigned long long c3;
184 unsigned long long c6;
185 unsigned long long c7;
186 unsigned long long mc6_us; /* duplicate as per-core for now, even though per module */
187 unsigned int core_temp_c;
188 unsigned int core_id;
189 unsigned long long counter[MAX_ADDED_COUNTERS];
190 } *core_even, *core_odd;
193 unsigned long long pc2;
194 unsigned long long pc3;
195 unsigned long long pc6;
196 unsigned long long pc7;
197 unsigned long long pc8;
198 unsigned long long pc9;
199 unsigned long long pc10;
200 unsigned long long cpu_lpi;
201 unsigned long long sys_lpi;
202 unsigned long long pkg_wtd_core_c0;
203 unsigned long long pkg_any_core_c0;
204 unsigned long long pkg_any_gfxe_c0;
205 unsigned long long pkg_both_core_gfxe_c0;
206 long long gfx_rc6_ms;
207 unsigned int gfx_mhz;
208 unsigned int package_id;
209 unsigned int energy_pkg; /* MSR_PKG_ENERGY_STATUS */
210 unsigned int energy_dram; /* MSR_DRAM_ENERGY_STATUS */
211 unsigned int energy_cores; /* MSR_PP0_ENERGY_STATUS */
212 unsigned int energy_gfx; /* MSR_PP1_ENERGY_STATUS */
213 unsigned int rapl_pkg_perf_status; /* MSR_PKG_PERF_STATUS */
214 unsigned int rapl_dram_perf_status; /* MSR_DRAM_PERF_STATUS */
215 unsigned int pkg_temp_c;
216 unsigned long long counter[MAX_ADDED_COUNTERS];
217 } *package_even, *package_odd;
219 #define ODD_COUNTERS thread_odd, core_odd, package_odd
220 #define EVEN_COUNTERS thread_even, core_even, package_even
222 #define GET_THREAD(thread_base, thread_no, core_no, node_no, pkg_no) \
225 topo.nodes_per_pkg * topo.cores_per_node * topo.threads_per_core) + \
226 ((node_no) * topo.cores_per_node * topo.threads_per_core) + \
227 ((core_no) * topo.threads_per_core) + \
230 #define GET_CORE(core_base, core_no, node_no, pkg_no) \
232 ((pkg_no) * topo.nodes_per_pkg * topo.cores_per_node) + \
233 ((node_no) * topo.cores_per_node) + \
237 #define GET_PKG(pkg_base, pkg_no) (pkg_base + pkg_no)
239 enum counter_scope {SCOPE_CPU, SCOPE_CORE, SCOPE_PACKAGE};
240 enum counter_type {COUNTER_ITEMS, COUNTER_CYCLES, COUNTER_SECONDS, COUNTER_USEC};
241 enum counter_format {FORMAT_RAW, FORMAT_DELTA, FORMAT_PERCENT};
244 unsigned int msr_num;
245 char name[NAME_BYTES];
246 char path[PATH_BYTES];
248 enum counter_type type;
249 enum counter_format format;
250 struct msr_counter *next;
252 #define FLAGS_HIDE (1 << 0)
253 #define FLAGS_SHOW (1 << 1)
254 #define SYSFS_PERCPU (1 << 1)
257 struct sys_counters {
258 unsigned int added_thread_counters;
259 unsigned int added_core_counters;
260 unsigned int added_package_counters;
261 struct msr_counter *tp;
262 struct msr_counter *cp;
263 struct msr_counter *pp;
266 struct system_summary {
267 struct thread_data threads;
268 struct core_data cores;
269 struct pkg_data packages;
272 struct cpu_topology {
273 int physical_package_id;
275 int physical_node_id;
276 int logical_node_id; /* 0-based count within the package */
277 int physical_core_id;
279 cpu_set_t *put_ids; /* Processing Unit/Thread IDs */
290 int threads_per_core;
293 struct timeval tv_even, tv_odd, tv_delta;
295 int *irq_column_2_cpu; /* /proc/interrupts column numbers */
296 int *irqs_per_cpu; /* indexed by cpu_num */
298 void setup_all_buffers(void);
300 int cpu_is_not_present(int cpu)
302 return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set);
305 * run func(thread, core, package) in topology order
306 * skip non-present cpus
309 int for_all_cpus(int (func)(struct thread_data *, struct core_data *, struct pkg_data *),
310 struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base)
312 int retval, pkg_no, core_no, thread_no, node_no;
314 for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
315 for (core_no = 0; core_no < topo.cores_per_node; ++core_no) {
316 for (node_no = 0; node_no < topo.nodes_per_pkg;
318 for (thread_no = 0; thread_no <
319 topo.threads_per_core; ++thread_no) {
320 struct thread_data *t;
324 t = GET_THREAD(thread_base, thread_no,
328 if (cpu_is_not_present(t->cpu_id))
331 c = GET_CORE(core_base, core_no,
333 p = GET_PKG(pkg_base, pkg_no);
335 retval = func(t, c, p);
345 int cpu_migrate(int cpu)
347 CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
348 CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set);
349 if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1)
354 int get_msr_fd(int cpu)
364 sprintf(pathname, "/dev/cpu/%d/msr", cpu);
365 fd = open(pathname, O_RDONLY);
367 err(-1, "%s open failed, try chown or chmod +r /dev/cpu/*/msr, or run as root", pathname);
374 int get_msr(int cpu, off_t offset, unsigned long long *msr)
378 retval = pread(get_msr_fd(cpu), msr, sizeof(*msr), offset);
380 if (retval != sizeof *msr)
381 err(-1, "cpu%d: msr offset 0x%llx read failed", cpu, (unsigned long long)offset);
387 * This list matches the column headers, except
388 * 1. built-in only, the sysfs counters are not here -- we learn of those at run-time
389 * 2. Core and CPU are moved to the end, we can't have strings that contain them
390 * matching on them for --show and --hide.
392 struct msr_counter bic[] = {
394 { 0x0, "Time_Of_Day_Seconds" },
402 { 0x0, "SMI", "", 32, 0, FORMAT_DELTA, NULL},
445 #define MAX_BIC (sizeof(bic) / sizeof(struct msr_counter))
446 #define BIC_USEC (1ULL << 0)
447 #define BIC_TOD (1ULL << 1)
448 #define BIC_Package (1ULL << 2)
449 #define BIC_Node (1ULL << 3)
450 #define BIC_Avg_MHz (1ULL << 4)
451 #define BIC_Busy (1ULL << 5)
452 #define BIC_Bzy_MHz (1ULL << 6)
453 #define BIC_TSC_MHz (1ULL << 7)
454 #define BIC_IRQ (1ULL << 8)
455 #define BIC_SMI (1ULL << 9)
456 #define BIC_sysfs (1ULL << 10)
457 #define BIC_CPU_c1 (1ULL << 11)
458 #define BIC_CPU_c3 (1ULL << 12)
459 #define BIC_CPU_c6 (1ULL << 13)
460 #define BIC_CPU_c7 (1ULL << 14)
461 #define BIC_ThreadC (1ULL << 15)
462 #define BIC_CoreTmp (1ULL << 16)
463 #define BIC_CoreCnt (1ULL << 17)
464 #define BIC_PkgTmp (1ULL << 18)
465 #define BIC_GFX_rc6 (1ULL << 19)
466 #define BIC_GFXMHz (1ULL << 20)
467 #define BIC_Pkgpc2 (1ULL << 21)
468 #define BIC_Pkgpc3 (1ULL << 22)
469 #define BIC_Pkgpc6 (1ULL << 23)
470 #define BIC_Pkgpc7 (1ULL << 24)
471 #define BIC_Pkgpc8 (1ULL << 25)
472 #define BIC_Pkgpc9 (1ULL << 26)
473 #define BIC_Pkgpc10 (1ULL << 27)
474 #define BIC_CPU_LPI (1ULL << 28)
475 #define BIC_SYS_LPI (1ULL << 29)
476 #define BIC_PkgWatt (1ULL << 30)
477 #define BIC_CorWatt (1ULL << 31)
478 #define BIC_GFXWatt (1ULL << 32)
479 #define BIC_PkgCnt (1ULL << 33)
480 #define BIC_RAMWatt (1ULL << 34)
481 #define BIC_PKG__ (1ULL << 35)
482 #define BIC_RAM__ (1ULL << 36)
483 #define BIC_Pkg_J (1ULL << 37)
484 #define BIC_Cor_J (1ULL << 38)
485 #define BIC_GFX_J (1ULL << 39)
486 #define BIC_RAM_J (1ULL << 40)
487 #define BIC_Mod_c6 (1ULL << 41)
488 #define BIC_Totl_c0 (1ULL << 42)
489 #define BIC_Any_c0 (1ULL << 43)
490 #define BIC_GFX_c0 (1ULL << 44)
491 #define BIC_CPUGFX (1ULL << 45)
492 #define BIC_Core (1ULL << 46)
493 #define BIC_CPU (1ULL << 47)
494 #define BIC_APIC (1ULL << 48)
495 #define BIC_X2APIC (1ULL << 49)
497 #define BIC_DISABLED_BY_DEFAULT (BIC_USEC | BIC_TOD | BIC_APIC | BIC_X2APIC)
499 unsigned long long bic_enabled = (0xFFFFFFFFFFFFFFFFULL & ~BIC_DISABLED_BY_DEFAULT);
500 unsigned long long bic_present = BIC_USEC | BIC_TOD | BIC_sysfs | BIC_APIC | BIC_X2APIC;
502 #define DO_BIC(COUNTER_NAME) (bic_enabled & bic_present & COUNTER_NAME)
503 #define ENABLE_BIC(COUNTER_NAME) (bic_enabled |= COUNTER_NAME)
504 #define BIC_PRESENT(COUNTER_BIT) (bic_present |= COUNTER_BIT)
505 #define BIC_NOT_PRESENT(COUNTER_BIT) (bic_present &= ~COUNTER_BIT)
508 #define MAX_DEFERRED 16
509 char *deferred_skip_names[MAX_DEFERRED];
510 int deferred_skip_index;
513 * HIDE_LIST - hide this list of counters, show the rest [default]
514 * SHOW_LIST - show this list of counters, hide the rest
516 enum show_hide_mode { SHOW_LIST, HIDE_LIST } global_show_hide_mode = HIDE_LIST;
521 "Usage: turbostat [OPTIONS][(--interval seconds) | COMMAND ...]\n"
523 "Turbostat forks the specified COMMAND and prints statistics\n"
524 "when COMMAND completes.\n"
525 "If no COMMAND is specified, turbostat wakes every 5-seconds\n"
526 "to print statistics, until interrupted.\n"
527 "--add add a counter\n"
528 " eg. --add msr0x10,u64,cpu,delta,MY_TSC\n"
529 "--cpu cpu-set limit output to summary plus cpu-set:\n"
530 " {core | package | j,k,l..m,n-p }\n"
531 "--quiet skip decoding system configuration header\n"
532 "--interval sec.subsec Override default 5-second measurement interval\n"
533 "--help print this help message\n"
534 "--list list column headers only\n"
535 "--num_iterations num number of the measurement iterations\n"
536 "--out file create or truncate \"file\" for all output\n"
537 "--version print version information\n"
539 "For more help, run \"man turbostat\"\n");
544 * for all the strings in comma separate name_list,
545 * set the approprate bit in return value.
547 unsigned long long bic_lookup(char *name_list, enum show_hide_mode mode)
550 unsigned long long retval = 0;
555 comma = strchr(name_list, ',');
560 if (!strcmp(name_list, "all"))
563 for (i = 0; i < MAX_BIC; ++i) {
564 if (!strcmp(name_list, bic[i].name)) {
565 retval |= (1ULL << i);
570 if (mode == SHOW_LIST) {
571 fprintf(stderr, "Invalid counter name: %s\n", name_list);
574 deferred_skip_names[deferred_skip_index++] = name_list;
576 fprintf(stderr, "deferred \"%s\"\n", name_list);
577 if (deferred_skip_index >= MAX_DEFERRED) {
578 fprintf(stderr, "More than max %d un-recognized --skip options '%s'\n",
579 MAX_DEFERRED, name_list);
594 void print_header(char *delim)
596 struct msr_counter *mp;
599 if (DO_BIC(BIC_USEC))
600 outp += sprintf(outp, "%susec", (printed++ ? delim : ""));
602 outp += sprintf(outp, "%sTime_Of_Day_Seconds", (printed++ ? delim : ""));
603 if (DO_BIC(BIC_Package))
604 outp += sprintf(outp, "%sPackage", (printed++ ? delim : ""));
605 if (DO_BIC(BIC_Node))
606 outp += sprintf(outp, "%sNode", (printed++ ? delim : ""));
607 if (DO_BIC(BIC_Core))
608 outp += sprintf(outp, "%sCore", (printed++ ? delim : ""));
610 outp += sprintf(outp, "%sCPU", (printed++ ? delim : ""));
611 if (DO_BIC(BIC_APIC))
612 outp += sprintf(outp, "%sAPIC", (printed++ ? delim : ""));
613 if (DO_BIC(BIC_X2APIC))
614 outp += sprintf(outp, "%sX2APIC", (printed++ ? delim : ""));
615 if (DO_BIC(BIC_Avg_MHz))
616 outp += sprintf(outp, "%sAvg_MHz", (printed++ ? delim : ""));
617 if (DO_BIC(BIC_Busy))
618 outp += sprintf(outp, "%sBusy%%", (printed++ ? delim : ""));
619 if (DO_BIC(BIC_Bzy_MHz))
620 outp += sprintf(outp, "%sBzy_MHz", (printed++ ? delim : ""));
621 if (DO_BIC(BIC_TSC_MHz))
622 outp += sprintf(outp, "%sTSC_MHz", (printed++ ? delim : ""));
624 if (DO_BIC(BIC_IRQ)) {
625 if (sums_need_wide_columns)
626 outp += sprintf(outp, "%s IRQ", (printed++ ? delim : ""));
628 outp += sprintf(outp, "%sIRQ", (printed++ ? delim : ""));
632 outp += sprintf(outp, "%sSMI", (printed++ ? delim : ""));
634 for (mp = sys.tp; mp; mp = mp->next) {
636 if (mp->format == FORMAT_RAW) {
638 outp += sprintf(outp, "%s%18.18s", (printed++ ? delim : ""), mp->name);
640 outp += sprintf(outp, "%s%10.10s", (printed++ ? delim : ""), mp->name);
642 if ((mp->type == COUNTER_ITEMS) && sums_need_wide_columns)
643 outp += sprintf(outp, "%s%8s", (printed++ ? delim : ""), mp->name);
645 outp += sprintf(outp, "%s%s", (printed++ ? delim : ""), mp->name);
649 if (DO_BIC(BIC_CPU_c1))
650 outp += sprintf(outp, "%sCPU%%c1", (printed++ ? delim : ""));
651 if (DO_BIC(BIC_CPU_c3) && !do_slm_cstates && !do_knl_cstates && !do_cnl_cstates)
652 outp += sprintf(outp, "%sCPU%%c3", (printed++ ? delim : ""));
653 if (DO_BIC(BIC_CPU_c6))
654 outp += sprintf(outp, "%sCPU%%c6", (printed++ ? delim : ""));
655 if (DO_BIC(BIC_CPU_c7))
656 outp += sprintf(outp, "%sCPU%%c7", (printed++ ? delim : ""));
658 if (DO_BIC(BIC_Mod_c6))
659 outp += sprintf(outp, "%sMod%%c6", (printed++ ? delim : ""));
661 if (DO_BIC(BIC_CoreTmp))
662 outp += sprintf(outp, "%sCoreTmp", (printed++ ? delim : ""));
664 for (mp = sys.cp; mp; mp = mp->next) {
665 if (mp->format == FORMAT_RAW) {
667 outp += sprintf(outp, "%s%18.18s", delim, mp->name);
669 outp += sprintf(outp, "%s%10.10s", delim, mp->name);
671 if ((mp->type == COUNTER_ITEMS) && sums_need_wide_columns)
672 outp += sprintf(outp, "%s%8s", delim, mp->name);
674 outp += sprintf(outp, "%s%s", delim, mp->name);
678 if (DO_BIC(BIC_PkgTmp))
679 outp += sprintf(outp, "%sPkgTmp", (printed++ ? delim : ""));
681 if (DO_BIC(BIC_GFX_rc6))
682 outp += sprintf(outp, "%sGFX%%rc6", (printed++ ? delim : ""));
684 if (DO_BIC(BIC_GFXMHz))
685 outp += sprintf(outp, "%sGFXMHz", (printed++ ? delim : ""));
687 if (DO_BIC(BIC_Totl_c0))
688 outp += sprintf(outp, "%sTotl%%C0", (printed++ ? delim : ""));
689 if (DO_BIC(BIC_Any_c0))
690 outp += sprintf(outp, "%sAny%%C0", (printed++ ? delim : ""));
691 if (DO_BIC(BIC_GFX_c0))
692 outp += sprintf(outp, "%sGFX%%C0", (printed++ ? delim : ""));
693 if (DO_BIC(BIC_CPUGFX))
694 outp += sprintf(outp, "%sCPUGFX%%", (printed++ ? delim : ""));
696 if (DO_BIC(BIC_Pkgpc2))
697 outp += sprintf(outp, "%sPkg%%pc2", (printed++ ? delim : ""));
698 if (DO_BIC(BIC_Pkgpc3))
699 outp += sprintf(outp, "%sPkg%%pc3", (printed++ ? delim : ""));
700 if (DO_BIC(BIC_Pkgpc6))
701 outp += sprintf(outp, "%sPkg%%pc6", (printed++ ? delim : ""));
702 if (DO_BIC(BIC_Pkgpc7))
703 outp += sprintf(outp, "%sPkg%%pc7", (printed++ ? delim : ""));
704 if (DO_BIC(BIC_Pkgpc8))
705 outp += sprintf(outp, "%sPkg%%pc8", (printed++ ? delim : ""));
706 if (DO_BIC(BIC_Pkgpc9))
707 outp += sprintf(outp, "%sPkg%%pc9", (printed++ ? delim : ""));
708 if (DO_BIC(BIC_Pkgpc10))
709 outp += sprintf(outp, "%sPk%%pc10", (printed++ ? delim : ""));
710 if (DO_BIC(BIC_CPU_LPI))
711 outp += sprintf(outp, "%sCPU%%LPI", (printed++ ? delim : ""));
712 if (DO_BIC(BIC_SYS_LPI))
713 outp += sprintf(outp, "%sSYS%%LPI", (printed++ ? delim : ""));
715 if (do_rapl && !rapl_joules) {
716 if (DO_BIC(BIC_PkgWatt))
717 outp += sprintf(outp, "%sPkgWatt", (printed++ ? delim : ""));
718 if (DO_BIC(BIC_CorWatt))
719 outp += sprintf(outp, "%sCorWatt", (printed++ ? delim : ""));
720 if (DO_BIC(BIC_GFXWatt))
721 outp += sprintf(outp, "%sGFXWatt", (printed++ ? delim : ""));
722 if (DO_BIC(BIC_RAMWatt))
723 outp += sprintf(outp, "%sRAMWatt", (printed++ ? delim : ""));
724 if (DO_BIC(BIC_PKG__))
725 outp += sprintf(outp, "%sPKG_%%", (printed++ ? delim : ""));
726 if (DO_BIC(BIC_RAM__))
727 outp += sprintf(outp, "%sRAM_%%", (printed++ ? delim : ""));
728 } else if (do_rapl && rapl_joules) {
729 if (DO_BIC(BIC_Pkg_J))
730 outp += sprintf(outp, "%sPkg_J", (printed++ ? delim : ""));
731 if (DO_BIC(BIC_Cor_J))
732 outp += sprintf(outp, "%sCor_J", (printed++ ? delim : ""));
733 if (DO_BIC(BIC_GFX_J))
734 outp += sprintf(outp, "%sGFX_J", (printed++ ? delim : ""));
735 if (DO_BIC(BIC_RAM_J))
736 outp += sprintf(outp, "%sRAM_J", (printed++ ? delim : ""));
737 if (DO_BIC(BIC_PKG__))
738 outp += sprintf(outp, "%sPKG_%%", (printed++ ? delim : ""));
739 if (DO_BIC(BIC_RAM__))
740 outp += sprintf(outp, "%sRAM_%%", (printed++ ? delim : ""));
742 for (mp = sys.pp; mp; mp = mp->next) {
743 if (mp->format == FORMAT_RAW) {
745 outp += sprintf(outp, "%s%18.18s", delim, mp->name);
747 outp += sprintf(outp, "%s%10.10s", delim, mp->name);
749 if ((mp->type == COUNTER_ITEMS) && sums_need_wide_columns)
750 outp += sprintf(outp, "%s%8s", delim, mp->name);
752 outp += sprintf(outp, "%s%s", delim, mp->name);
756 outp += sprintf(outp, "\n");
759 int dump_counters(struct thread_data *t, struct core_data *c,
763 struct msr_counter *mp;
765 outp += sprintf(outp, "t %p, c %p, p %p\n", t, c, p);
768 outp += sprintf(outp, "CPU: %d flags 0x%x\n",
769 t->cpu_id, t->flags);
770 outp += sprintf(outp, "TSC: %016llX\n", t->tsc);
771 outp += sprintf(outp, "aperf: %016llX\n", t->aperf);
772 outp += sprintf(outp, "mperf: %016llX\n", t->mperf);
773 outp += sprintf(outp, "c1: %016llX\n", t->c1);
776 outp += sprintf(outp, "IRQ: %lld\n", t->irq_count);
778 outp += sprintf(outp, "SMI: %d\n", t->smi_count);
780 for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
781 outp += sprintf(outp, "tADDED [%d] msr0x%x: %08llX\n",
782 i, mp->msr_num, t->counter[i]);
787 outp += sprintf(outp, "core: %d\n", c->core_id);
788 outp += sprintf(outp, "c3: %016llX\n", c->c3);
789 outp += sprintf(outp, "c6: %016llX\n", c->c6);
790 outp += sprintf(outp, "c7: %016llX\n", c->c7);
791 outp += sprintf(outp, "DTS: %dC\n", c->core_temp_c);
793 for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
794 outp += sprintf(outp, "cADDED [%d] msr0x%x: %08llX\n",
795 i, mp->msr_num, c->counter[i]);
797 outp += sprintf(outp, "mc6_us: %016llX\n", c->mc6_us);
801 outp += sprintf(outp, "package: %d\n", p->package_id);
803 outp += sprintf(outp, "Weighted cores: %016llX\n", p->pkg_wtd_core_c0);
804 outp += sprintf(outp, "Any cores: %016llX\n", p->pkg_any_core_c0);
805 outp += sprintf(outp, "Any GFX: %016llX\n", p->pkg_any_gfxe_c0);
806 outp += sprintf(outp, "CPU + GFX: %016llX\n", p->pkg_both_core_gfxe_c0);
808 outp += sprintf(outp, "pc2: %016llX\n", p->pc2);
809 if (DO_BIC(BIC_Pkgpc3))
810 outp += sprintf(outp, "pc3: %016llX\n", p->pc3);
811 if (DO_BIC(BIC_Pkgpc6))
812 outp += sprintf(outp, "pc6: %016llX\n", p->pc6);
813 if (DO_BIC(BIC_Pkgpc7))
814 outp += sprintf(outp, "pc7: %016llX\n", p->pc7);
815 outp += sprintf(outp, "pc8: %016llX\n", p->pc8);
816 outp += sprintf(outp, "pc9: %016llX\n", p->pc9);
817 outp += sprintf(outp, "pc10: %016llX\n", p->pc10);
818 outp += sprintf(outp, "pc10: %016llX\n", p->pc10);
819 outp += sprintf(outp, "cpu_lpi: %016llX\n", p->cpu_lpi);
820 outp += sprintf(outp, "sys_lpi: %016llX\n", p->sys_lpi);
821 outp += sprintf(outp, "Joules PKG: %0X\n", p->energy_pkg);
822 outp += sprintf(outp, "Joules COR: %0X\n", p->energy_cores);
823 outp += sprintf(outp, "Joules GFX: %0X\n", p->energy_gfx);
824 outp += sprintf(outp, "Joules RAM: %0X\n", p->energy_dram);
825 outp += sprintf(outp, "Throttle PKG: %0X\n",
826 p->rapl_pkg_perf_status);
827 outp += sprintf(outp, "Throttle RAM: %0X\n",
828 p->rapl_dram_perf_status);
829 outp += sprintf(outp, "PTM: %dC\n", p->pkg_temp_c);
831 for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
832 outp += sprintf(outp, "pADDED [%d] msr0x%x: %08llX\n",
833 i, mp->msr_num, p->counter[i]);
837 outp += sprintf(outp, "\n");
843 * column formatting convention & formats
845 int format_counters(struct thread_data *t, struct core_data *c,
848 double interval_float, tsc;
851 struct msr_counter *mp;
855 /* if showing only 1st thread in core and this isn't one, bail out */
856 if (show_core_only && !(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
859 /* if showing only 1st thread in pkg and this isn't one, bail out */
860 if (show_pkg_only && !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
863 /*if not summary line and --cpu is used */
864 if ((t != &average.threads) &&
865 (cpu_subset && !CPU_ISSET_S(t->cpu_id, cpu_subset_size, cpu_subset)))
868 if (DO_BIC(BIC_USEC)) {
869 /* on each row, print how many usec each timestamp took to gather */
872 timersub(&t->tv_end, &t->tv_begin, &tv);
873 outp += sprintf(outp, "%5ld\t", tv.tv_sec * 1000000 + tv.tv_usec);
876 /* Time_Of_Day_Seconds: on each row, print sec.usec last timestamp taken */
878 outp += sprintf(outp, "%10ld.%06ld\t", t->tv_end.tv_sec, t->tv_end.tv_usec);
880 interval_float = tv_delta.tv_sec + tv_delta.tv_usec/1000000.0;
882 tsc = t->tsc * tsc_tweak;
884 /* topo columns, print blanks on 1st (average) line */
885 if (t == &average.threads) {
886 if (DO_BIC(BIC_Package))
887 outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
888 if (DO_BIC(BIC_Node))
889 outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
890 if (DO_BIC(BIC_Core))
891 outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
893 outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
894 if (DO_BIC(BIC_APIC))
895 outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
896 if (DO_BIC(BIC_X2APIC))
897 outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
899 if (DO_BIC(BIC_Package)) {
901 outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), p->package_id);
903 outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
905 if (DO_BIC(BIC_Node)) {
907 outp += sprintf(outp, "%s%d",
908 (printed++ ? delim : ""),
909 cpus[t->cpu_id].physical_node_id);
911 outp += sprintf(outp, "%s-",
912 (printed++ ? delim : ""));
914 if (DO_BIC(BIC_Core)) {
916 outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), c->core_id);
918 outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
921 outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), t->cpu_id);
922 if (DO_BIC(BIC_APIC))
923 outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), t->apic_id);
924 if (DO_BIC(BIC_X2APIC))
925 outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), t->x2apic_id);
928 if (DO_BIC(BIC_Avg_MHz))
929 outp += sprintf(outp, "%s%.0f", (printed++ ? delim : ""),
930 1.0 / units * t->aperf / interval_float);
932 if (DO_BIC(BIC_Busy))
933 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * t->mperf/tsc);
935 if (DO_BIC(BIC_Bzy_MHz)) {
937 outp += sprintf(outp, "%s%.0f", (printed++ ? delim : ""), base_hz / units * t->aperf / t->mperf);
939 outp += sprintf(outp, "%s%.0f", (printed++ ? delim : ""),
940 tsc / units * t->aperf / t->mperf / interval_float);
943 if (DO_BIC(BIC_TSC_MHz))
944 outp += sprintf(outp, "%s%.0f", (printed++ ? delim : ""), 1.0 * t->tsc/units/interval_float);
947 if (DO_BIC(BIC_IRQ)) {
948 if (sums_need_wide_columns)
949 outp += sprintf(outp, "%s%8lld", (printed++ ? delim : ""), t->irq_count);
951 outp += sprintf(outp, "%s%lld", (printed++ ? delim : ""), t->irq_count);
956 outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), t->smi_count);
959 for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
960 if (mp->format == FORMAT_RAW) {
962 outp += sprintf(outp, "%s0x%08x", (printed++ ? delim : ""), (unsigned int) t->counter[i]);
964 outp += sprintf(outp, "%s0x%016llx", (printed++ ? delim : ""), t->counter[i]);
965 } else if (mp->format == FORMAT_DELTA) {
966 if ((mp->type == COUNTER_ITEMS) && sums_need_wide_columns)
967 outp += sprintf(outp, "%s%8lld", (printed++ ? delim : ""), t->counter[i]);
969 outp += sprintf(outp, "%s%lld", (printed++ ? delim : ""), t->counter[i]);
970 } else if (mp->format == FORMAT_PERCENT) {
971 if (mp->type == COUNTER_USEC)
972 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), t->counter[i]/interval_float/10000);
974 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * t->counter[i]/tsc);
979 if (DO_BIC(BIC_CPU_c1))
980 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * t->c1/tsc);
983 /* print per-core data only for 1st thread in core */
984 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
987 if (DO_BIC(BIC_CPU_c3) && !do_slm_cstates && !do_knl_cstates && !do_cnl_cstates)
988 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * c->c3/tsc);
989 if (DO_BIC(BIC_CPU_c6))
990 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * c->c6/tsc);
991 if (DO_BIC(BIC_CPU_c7))
992 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * c->c7/tsc);
995 if (DO_BIC(BIC_Mod_c6))
996 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * c->mc6_us / tsc);
998 if (DO_BIC(BIC_CoreTmp))
999 outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), c->core_temp_c);
1001 for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
1002 if (mp->format == FORMAT_RAW) {
1003 if (mp->width == 32)
1004 outp += sprintf(outp, "%s0x%08x", (printed++ ? delim : ""), (unsigned int) c->counter[i]);
1006 outp += sprintf(outp, "%s0x%016llx", (printed++ ? delim : ""), c->counter[i]);
1007 } else if (mp->format == FORMAT_DELTA) {
1008 if ((mp->type == COUNTER_ITEMS) && sums_need_wide_columns)
1009 outp += sprintf(outp, "%s%8lld", (printed++ ? delim : ""), c->counter[i]);
1011 outp += sprintf(outp, "%s%lld", (printed++ ? delim : ""), c->counter[i]);
1012 } else if (mp->format == FORMAT_PERCENT) {
1013 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * c->counter[i]/tsc);
1017 /* print per-package data only for 1st core in package */
1018 if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
1022 if (DO_BIC(BIC_PkgTmp))
1023 outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), p->pkg_temp_c);
1026 if (DO_BIC(BIC_GFX_rc6)) {
1027 if (p->gfx_rc6_ms == -1) { /* detect GFX counter reset */
1028 outp += sprintf(outp, "%s**.**", (printed++ ? delim : ""));
1030 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""),
1031 p->gfx_rc6_ms / 10.0 / interval_float);
1036 if (DO_BIC(BIC_GFXMHz))
1037 outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), p->gfx_mhz);
1039 /* Totl%C0, Any%C0 GFX%C0 CPUGFX% */
1040 if (DO_BIC(BIC_Totl_c0))
1041 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pkg_wtd_core_c0/tsc);
1042 if (DO_BIC(BIC_Any_c0))
1043 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pkg_any_core_c0/tsc);
1044 if (DO_BIC(BIC_GFX_c0))
1045 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pkg_any_gfxe_c0/tsc);
1046 if (DO_BIC(BIC_CPUGFX))
1047 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pkg_both_core_gfxe_c0/tsc);
1049 if (DO_BIC(BIC_Pkgpc2))
1050 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pc2/tsc);
1051 if (DO_BIC(BIC_Pkgpc3))
1052 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pc3/tsc);
1053 if (DO_BIC(BIC_Pkgpc6))
1054 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pc6/tsc);
1055 if (DO_BIC(BIC_Pkgpc7))
1056 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pc7/tsc);
1057 if (DO_BIC(BIC_Pkgpc8))
1058 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pc8/tsc);
1059 if (DO_BIC(BIC_Pkgpc9))
1060 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pc9/tsc);
1061 if (DO_BIC(BIC_Pkgpc10))
1062 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pc10/tsc);
1064 if (DO_BIC(BIC_CPU_LPI))
1065 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->cpu_lpi / 1000000.0 / interval_float);
1066 if (DO_BIC(BIC_SYS_LPI))
1067 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->sys_lpi / 1000000.0 / interval_float);
1070 * If measurement interval exceeds minimum RAPL Joule Counter range,
1071 * indicate that results are suspect by printing "**" in fraction place.
1073 if (interval_float < rapl_joule_counter_range)
1078 if (DO_BIC(BIC_PkgWatt))
1079 outp += sprintf(outp, fmt8, (printed++ ? delim : ""), p->energy_pkg * rapl_energy_units / interval_float);
1080 if (DO_BIC(BIC_CorWatt))
1081 outp += sprintf(outp, fmt8, (printed++ ? delim : ""), p->energy_cores * rapl_energy_units / interval_float);
1082 if (DO_BIC(BIC_GFXWatt))
1083 outp += sprintf(outp, fmt8, (printed++ ? delim : ""), p->energy_gfx * rapl_energy_units / interval_float);
1084 if (DO_BIC(BIC_RAMWatt))
1085 outp += sprintf(outp, fmt8, (printed++ ? delim : ""), p->energy_dram * rapl_dram_energy_units / interval_float);
1086 if (DO_BIC(BIC_Pkg_J))
1087 outp += sprintf(outp, fmt8, (printed++ ? delim : ""), p->energy_pkg * rapl_energy_units);
1088 if (DO_BIC(BIC_Cor_J))
1089 outp += sprintf(outp, fmt8, (printed++ ? delim : ""), p->energy_cores * rapl_energy_units);
1090 if (DO_BIC(BIC_GFX_J))
1091 outp += sprintf(outp, fmt8, (printed++ ? delim : ""), p->energy_gfx * rapl_energy_units);
1092 if (DO_BIC(BIC_RAM_J))
1093 outp += sprintf(outp, fmt8, (printed++ ? delim : ""), p->energy_dram * rapl_dram_energy_units);
1094 if (DO_BIC(BIC_PKG__))
1095 outp += sprintf(outp, fmt8, (printed++ ? delim : ""), 100.0 * p->rapl_pkg_perf_status * rapl_time_units / interval_float);
1096 if (DO_BIC(BIC_RAM__))
1097 outp += sprintf(outp, fmt8, (printed++ ? delim : ""), 100.0 * p->rapl_dram_perf_status * rapl_time_units / interval_float);
1099 for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
1100 if (mp->format == FORMAT_RAW) {
1101 if (mp->width == 32)
1102 outp += sprintf(outp, "%s0x%08x", (printed++ ? delim : ""), (unsigned int) p->counter[i]);
1104 outp += sprintf(outp, "%s0x%016llx", (printed++ ? delim : ""), p->counter[i]);
1105 } else if (mp->format == FORMAT_DELTA) {
1106 if ((mp->type == COUNTER_ITEMS) && sums_need_wide_columns)
1107 outp += sprintf(outp, "%s%8lld", (printed++ ? delim : ""), p->counter[i]);
1109 outp += sprintf(outp, "%s%lld", (printed++ ? delim : ""), p->counter[i]);
1110 } else if (mp->format == FORMAT_PERCENT) {
1111 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->counter[i]/tsc);
1116 if (*(outp - 1) != '\n')
1117 outp += sprintf(outp, "\n");
1122 void flush_output_stdout(void)
1131 fputs(output_buffer, filep);
1134 outp = output_buffer;
1136 void flush_output_stderr(void)
1138 fputs(output_buffer, outf);
1140 outp = output_buffer;
1142 void format_all_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1146 if (!printed || !summary_only)
1149 if (topo.num_cpus > 1)
1150 format_counters(&average.threads, &average.cores,
1158 for_all_cpus(format_counters, t, c, p);
1161 #define DELTA_WRAP32(new, old) \
1165 old = 0x100000000 + new - old; \
1169 delta_package(struct pkg_data *new, struct pkg_data *old)
1172 struct msr_counter *mp;
1175 if (DO_BIC(BIC_Totl_c0))
1176 old->pkg_wtd_core_c0 = new->pkg_wtd_core_c0 - old->pkg_wtd_core_c0;
1177 if (DO_BIC(BIC_Any_c0))
1178 old->pkg_any_core_c0 = new->pkg_any_core_c0 - old->pkg_any_core_c0;
1179 if (DO_BIC(BIC_GFX_c0))
1180 old->pkg_any_gfxe_c0 = new->pkg_any_gfxe_c0 - old->pkg_any_gfxe_c0;
1181 if (DO_BIC(BIC_CPUGFX))
1182 old->pkg_both_core_gfxe_c0 = new->pkg_both_core_gfxe_c0 - old->pkg_both_core_gfxe_c0;
1184 old->pc2 = new->pc2 - old->pc2;
1185 if (DO_BIC(BIC_Pkgpc3))
1186 old->pc3 = new->pc3 - old->pc3;
1187 if (DO_BIC(BIC_Pkgpc6))
1188 old->pc6 = new->pc6 - old->pc6;
1189 if (DO_BIC(BIC_Pkgpc7))
1190 old->pc7 = new->pc7 - old->pc7;
1191 old->pc8 = new->pc8 - old->pc8;
1192 old->pc9 = new->pc9 - old->pc9;
1193 old->pc10 = new->pc10 - old->pc10;
1194 old->cpu_lpi = new->cpu_lpi - old->cpu_lpi;
1195 old->sys_lpi = new->sys_lpi - old->sys_lpi;
1196 old->pkg_temp_c = new->pkg_temp_c;
1198 /* flag an error when rc6 counter resets/wraps */
1199 if (old->gfx_rc6_ms > new->gfx_rc6_ms)
1200 old->gfx_rc6_ms = -1;
1202 old->gfx_rc6_ms = new->gfx_rc6_ms - old->gfx_rc6_ms;
1204 old->gfx_mhz = new->gfx_mhz;
1206 DELTA_WRAP32(new->energy_pkg, old->energy_pkg);
1207 DELTA_WRAP32(new->energy_cores, old->energy_cores);
1208 DELTA_WRAP32(new->energy_gfx, old->energy_gfx);
1209 DELTA_WRAP32(new->energy_dram, old->energy_dram);
1210 DELTA_WRAP32(new->rapl_pkg_perf_status, old->rapl_pkg_perf_status);
1211 DELTA_WRAP32(new->rapl_dram_perf_status, old->rapl_dram_perf_status);
1213 for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
1214 if (mp->format == FORMAT_RAW)
1215 old->counter[i] = new->counter[i];
1217 old->counter[i] = new->counter[i] - old->counter[i];
1224 delta_core(struct core_data *new, struct core_data *old)
1227 struct msr_counter *mp;
1229 old->c3 = new->c3 - old->c3;
1230 old->c6 = new->c6 - old->c6;
1231 old->c7 = new->c7 - old->c7;
1232 old->core_temp_c = new->core_temp_c;
1233 old->mc6_us = new->mc6_us - old->mc6_us;
1235 for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
1236 if (mp->format == FORMAT_RAW)
1237 old->counter[i] = new->counter[i];
1239 old->counter[i] = new->counter[i] - old->counter[i];
1247 delta_thread(struct thread_data *new, struct thread_data *old,
1248 struct core_data *core_delta)
1251 struct msr_counter *mp;
1253 /* we run cpuid just the 1st time, copy the results */
1254 if (DO_BIC(BIC_APIC))
1255 new->apic_id = old->apic_id;
1256 if (DO_BIC(BIC_X2APIC))
1257 new->x2apic_id = old->x2apic_id;
1260 * the timestamps from start of measurement interval are in "old"
1261 * the timestamp from end of measurement interval are in "new"
1262 * over-write old w/ new so we can print end of interval values
1265 old->tv_begin = new->tv_begin;
1266 old->tv_end = new->tv_end;
1268 old->tsc = new->tsc - old->tsc;
1270 /* check for TSC < 1 Mcycles over interval */
1271 if (old->tsc < (1000 * 1000))
1272 errx(-3, "Insanely slow TSC rate, TSC stops in idle?\n"
1273 "You can disable all c-states by booting with \"idle=poll\"\n"
1274 "or just the deep ones with \"processor.max_cstate=1\"");
1276 old->c1 = new->c1 - old->c1;
1278 if (DO_BIC(BIC_Avg_MHz) || DO_BIC(BIC_Busy) || DO_BIC(BIC_Bzy_MHz)) {
1279 if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
1280 old->aperf = new->aperf - old->aperf;
1281 old->mperf = new->mperf - old->mperf;
1288 if (use_c1_residency_msr) {
1290 * Some models have a dedicated C1 residency MSR,
1291 * which should be more accurate than the derivation below.
1295 * As counter collection is not atomic,
1296 * it is possible for mperf's non-halted cycles + idle states
1297 * to exceed TSC's all cycles: show c1 = 0% in that case.
1299 if ((old->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > (old->tsc * tsc_tweak))
1302 /* normal case, derive c1 */
1303 old->c1 = (old->tsc * tsc_tweak) - old->mperf - core_delta->c3
1304 - core_delta->c6 - core_delta->c7;
1308 if (old->mperf == 0) {
1310 fprintf(outf, "cpu%d MPERF 0!\n", old->cpu_id);
1311 old->mperf = 1; /* divide by 0 protection */
1314 if (DO_BIC(BIC_IRQ))
1315 old->irq_count = new->irq_count - old->irq_count;
1317 if (DO_BIC(BIC_SMI))
1318 old->smi_count = new->smi_count - old->smi_count;
1320 for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
1321 if (mp->format == FORMAT_RAW)
1322 old->counter[i] = new->counter[i];
1324 old->counter[i] = new->counter[i] - old->counter[i];
1329 int delta_cpu(struct thread_data *t, struct core_data *c,
1330 struct pkg_data *p, struct thread_data *t2,
1331 struct core_data *c2, struct pkg_data *p2)
1335 /* calculate core delta only for 1st thread in core */
1336 if (t->flags & CPU_IS_FIRST_THREAD_IN_CORE)
1339 /* always calculate thread delta */
1340 retval = delta_thread(t, t2, c2); /* c2 is core delta */
1344 /* calculate package delta only for 1st core in package */
1345 if (t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)
1346 retval = delta_package(p, p2);
1351 void clear_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1354 struct msr_counter *mp;
1356 t->tv_begin.tv_sec = 0;
1357 t->tv_begin.tv_usec = 0;
1358 t->tv_end.tv_sec = 0;
1359 t->tv_end.tv_usec = 0;
1369 /* tells format_counters to dump all fields from this set */
1370 t->flags = CPU_IS_FIRST_THREAD_IN_CORE | CPU_IS_FIRST_CORE_IN_PACKAGE;
1378 p->pkg_wtd_core_c0 = 0;
1379 p->pkg_any_core_c0 = 0;
1380 p->pkg_any_gfxe_c0 = 0;
1381 p->pkg_both_core_gfxe_c0 = 0;
1384 if (DO_BIC(BIC_Pkgpc3))
1386 if (DO_BIC(BIC_Pkgpc6))
1388 if (DO_BIC(BIC_Pkgpc7))
1398 p->energy_cores = 0;
1400 p->rapl_pkg_perf_status = 0;
1401 p->rapl_dram_perf_status = 0;
1406 for (i = 0, mp = sys.tp; mp; i++, mp = mp->next)
1409 for (i = 0, mp = sys.cp; mp; i++, mp = mp->next)
1412 for (i = 0, mp = sys.pp; mp; i++, mp = mp->next)
1415 int sum_counters(struct thread_data *t, struct core_data *c,
1419 struct msr_counter *mp;
1421 /* copy un-changing apic_id's */
1422 if (DO_BIC(BIC_APIC))
1423 average.threads.apic_id = t->apic_id;
1424 if (DO_BIC(BIC_X2APIC))
1425 average.threads.x2apic_id = t->x2apic_id;
1427 /* remember first tv_begin */
1428 if (average.threads.tv_begin.tv_sec == 0)
1429 average.threads.tv_begin = t->tv_begin;
1431 /* remember last tv_end */
1432 average.threads.tv_end = t->tv_end;
1434 average.threads.tsc += t->tsc;
1435 average.threads.aperf += t->aperf;
1436 average.threads.mperf += t->mperf;
1437 average.threads.c1 += t->c1;
1439 average.threads.irq_count += t->irq_count;
1440 average.threads.smi_count += t->smi_count;
1442 for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
1443 if (mp->format == FORMAT_RAW)
1445 average.threads.counter[i] += t->counter[i];
1448 /* sum per-core values only for 1st thread in core */
1449 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
1452 average.cores.c3 += c->c3;
1453 average.cores.c6 += c->c6;
1454 average.cores.c7 += c->c7;
1455 average.cores.mc6_us += c->mc6_us;
1457 average.cores.core_temp_c = MAX(average.cores.core_temp_c, c->core_temp_c);
1459 for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
1460 if (mp->format == FORMAT_RAW)
1462 average.cores.counter[i] += c->counter[i];
1465 /* sum per-pkg values only for 1st core in pkg */
1466 if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
1469 if (DO_BIC(BIC_Totl_c0))
1470 average.packages.pkg_wtd_core_c0 += p->pkg_wtd_core_c0;
1471 if (DO_BIC(BIC_Any_c0))
1472 average.packages.pkg_any_core_c0 += p->pkg_any_core_c0;
1473 if (DO_BIC(BIC_GFX_c0))
1474 average.packages.pkg_any_gfxe_c0 += p->pkg_any_gfxe_c0;
1475 if (DO_BIC(BIC_CPUGFX))
1476 average.packages.pkg_both_core_gfxe_c0 += p->pkg_both_core_gfxe_c0;
1478 average.packages.pc2 += p->pc2;
1479 if (DO_BIC(BIC_Pkgpc3))
1480 average.packages.pc3 += p->pc3;
1481 if (DO_BIC(BIC_Pkgpc6))
1482 average.packages.pc6 += p->pc6;
1483 if (DO_BIC(BIC_Pkgpc7))
1484 average.packages.pc7 += p->pc7;
1485 average.packages.pc8 += p->pc8;
1486 average.packages.pc9 += p->pc9;
1487 average.packages.pc10 += p->pc10;
1489 average.packages.cpu_lpi = p->cpu_lpi;
1490 average.packages.sys_lpi = p->sys_lpi;
1492 average.packages.energy_pkg += p->energy_pkg;
1493 average.packages.energy_dram += p->energy_dram;
1494 average.packages.energy_cores += p->energy_cores;
1495 average.packages.energy_gfx += p->energy_gfx;
1497 average.packages.gfx_rc6_ms = p->gfx_rc6_ms;
1498 average.packages.gfx_mhz = p->gfx_mhz;
1500 average.packages.pkg_temp_c = MAX(average.packages.pkg_temp_c, p->pkg_temp_c);
1502 average.packages.rapl_pkg_perf_status += p->rapl_pkg_perf_status;
1503 average.packages.rapl_dram_perf_status += p->rapl_dram_perf_status;
1505 for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
1506 if (mp->format == FORMAT_RAW)
1508 average.packages.counter[i] += p->counter[i];
1513 * sum the counters for all cpus in the system
1514 * compute the weighted average
1516 void compute_average(struct thread_data *t, struct core_data *c,
1520 struct msr_counter *mp;
1522 clear_counters(&average.threads, &average.cores, &average.packages);
1524 for_all_cpus(sum_counters, t, c, p);
1526 average.threads.tsc /= topo.num_cpus;
1527 average.threads.aperf /= topo.num_cpus;
1528 average.threads.mperf /= topo.num_cpus;
1529 average.threads.c1 /= topo.num_cpus;
1531 if (average.threads.irq_count > 9999999)
1532 sums_need_wide_columns = 1;
1534 average.cores.c3 /= topo.num_cores;
1535 average.cores.c6 /= topo.num_cores;
1536 average.cores.c7 /= topo.num_cores;
1537 average.cores.mc6_us /= topo.num_cores;
1539 if (DO_BIC(BIC_Totl_c0))
1540 average.packages.pkg_wtd_core_c0 /= topo.num_packages;
1541 if (DO_BIC(BIC_Any_c0))
1542 average.packages.pkg_any_core_c0 /= topo.num_packages;
1543 if (DO_BIC(BIC_GFX_c0))
1544 average.packages.pkg_any_gfxe_c0 /= topo.num_packages;
1545 if (DO_BIC(BIC_CPUGFX))
1546 average.packages.pkg_both_core_gfxe_c0 /= topo.num_packages;
1548 average.packages.pc2 /= topo.num_packages;
1549 if (DO_BIC(BIC_Pkgpc3))
1550 average.packages.pc3 /= topo.num_packages;
1551 if (DO_BIC(BIC_Pkgpc6))
1552 average.packages.pc6 /= topo.num_packages;
1553 if (DO_BIC(BIC_Pkgpc7))
1554 average.packages.pc7 /= topo.num_packages;
1556 average.packages.pc8 /= topo.num_packages;
1557 average.packages.pc9 /= topo.num_packages;
1558 average.packages.pc10 /= topo.num_packages;
1560 for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
1561 if (mp->format == FORMAT_RAW)
1563 if (mp->type == COUNTER_ITEMS) {
1564 if (average.threads.counter[i] > 9999999)
1565 sums_need_wide_columns = 1;
1568 average.threads.counter[i] /= topo.num_cpus;
1570 for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
1571 if (mp->format == FORMAT_RAW)
1573 if (mp->type == COUNTER_ITEMS) {
1574 if (average.cores.counter[i] > 9999999)
1575 sums_need_wide_columns = 1;
1577 average.cores.counter[i] /= topo.num_cores;
1579 for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
1580 if (mp->format == FORMAT_RAW)
1582 if (mp->type == COUNTER_ITEMS) {
1583 if (average.packages.counter[i] > 9999999)
1584 sums_need_wide_columns = 1;
1586 average.packages.counter[i] /= topo.num_packages;
1590 static unsigned long long rdtsc(void)
1592 unsigned int low, high;
1594 asm volatile("rdtsc" : "=a" (low), "=d" (high));
1596 return low | ((unsigned long long)high) << 32;
1600 * Open a file, and exit on failure
1602 FILE *fopen_or_die(const char *path, const char *mode)
1604 FILE *filep = fopen(path, mode);
1607 err(1, "%s: open failed", path);
1611 * snapshot_sysfs_counter()
1613 * return snapshot of given counter
1615 unsigned long long snapshot_sysfs_counter(char *path)
1619 unsigned long long counter;
1621 fp = fopen_or_die(path, "r");
1623 retval = fscanf(fp, "%lld", &counter);
1625 err(1, "snapshot_sysfs_counter(%s)", path);
1632 int get_mp(int cpu, struct msr_counter *mp, unsigned long long *counterp)
1634 if (mp->msr_num != 0) {
1635 if (get_msr(cpu, mp->msr_num, counterp))
1638 char path[128 + PATH_BYTES];
1640 if (mp->flags & SYSFS_PERCPU) {
1641 sprintf(path, "/sys/devices/system/cpu/cpu%d/%s",
1644 *counterp = snapshot_sysfs_counter(path);
1646 *counterp = snapshot_sysfs_counter(mp->path);
1653 void get_apic_id(struct thread_data *t)
1655 unsigned int eax, ebx, ecx, edx, max_level;
1657 eax = ebx = ecx = edx = 0;
1662 __cpuid(0, max_level, ebx, ecx, edx);
1664 __cpuid(1, eax, ebx, ecx, edx);
1665 t->apic_id = (ebx >> 24) & 0xf;
1667 if (max_level < 0xb)
1670 if (!DO_BIC(BIC_X2APIC))
1674 __cpuid(0xb, eax, ebx, ecx, edx);
1677 if (debug && (t->apic_id != t->x2apic_id))
1678 fprintf(stderr, "cpu%d: apic 0x%x x2apic 0x%x\n", t->cpu_id, t->apic_id, t->x2apic_id);
1684 * acquire and record local counters for that cpu
1686 int get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1688 int cpu = t->cpu_id;
1689 unsigned long long msr;
1690 int aperf_mperf_retry_count = 0;
1691 struct msr_counter *mp;
1694 gettimeofday(&t->tv_begin, (struct timezone *)NULL);
1696 if (cpu_migrate(cpu)) {
1697 fprintf(outf, "Could not migrate to CPU %d\n", cpu);
1701 if (first_counter_read)
1704 t->tsc = rdtsc(); /* we are running on local CPU of interest */
1706 if (DO_BIC(BIC_Avg_MHz) || DO_BIC(BIC_Busy) || DO_BIC(BIC_Bzy_MHz)) {
1707 unsigned long long tsc_before, tsc_between, tsc_after, aperf_time, mperf_time;
1710 * The TSC, APERF and MPERF must be read together for
1711 * APERF/MPERF and MPERF/TSC to give accurate results.
1713 * Unfortunately, APERF and MPERF are read by
1714 * individual system call, so delays may occur
1715 * between them. If the time to read them
1716 * varies by a large amount, we re-read them.
1720 * This initial dummy APERF read has been seen to
1721 * reduce jitter in the subsequent reads.
1724 if (get_msr(cpu, MSR_IA32_APERF, &t->aperf))
1727 t->tsc = rdtsc(); /* re-read close to APERF */
1729 tsc_before = t->tsc;
1731 if (get_msr(cpu, MSR_IA32_APERF, &t->aperf))
1734 tsc_between = rdtsc();
1736 if (get_msr(cpu, MSR_IA32_MPERF, &t->mperf))
1739 tsc_after = rdtsc();
1741 aperf_time = tsc_between - tsc_before;
1742 mperf_time = tsc_after - tsc_between;
1745 * If the system call latency to read APERF and MPERF
1746 * differ by more than 2x, then try again.
1748 if ((aperf_time > (2 * mperf_time)) || (mperf_time > (2 * aperf_time))) {
1749 aperf_mperf_retry_count++;
1750 if (aperf_mperf_retry_count < 5)
1753 warnx("cpu%d jitter %lld %lld",
1754 cpu, aperf_time, mperf_time);
1756 aperf_mperf_retry_count = 0;
1758 t->aperf = t->aperf * aperf_mperf_multiplier;
1759 t->mperf = t->mperf * aperf_mperf_multiplier;
1762 if (DO_BIC(BIC_IRQ))
1763 t->irq_count = irqs_per_cpu[cpu];
1764 if (DO_BIC(BIC_SMI)) {
1765 if (get_msr(cpu, MSR_SMI_COUNT, &msr))
1767 t->smi_count = msr & 0xFFFFFFFF;
1769 if (DO_BIC(BIC_CPU_c1) && use_c1_residency_msr) {
1770 if (get_msr(cpu, MSR_CORE_C1_RES, &t->c1))
1774 for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
1775 if (get_mp(cpu, mp, &t->counter[i]))
1779 /* collect core counters only for 1st thread in core */
1780 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
1783 if (DO_BIC(BIC_CPU_c3) && !do_slm_cstates && !do_knl_cstates && !do_cnl_cstates) {
1784 if (get_msr(cpu, MSR_CORE_C3_RESIDENCY, &c->c3))
1788 if (DO_BIC(BIC_CPU_c6) && !do_knl_cstates) {
1789 if (get_msr(cpu, MSR_CORE_C6_RESIDENCY, &c->c6))
1791 } else if (do_knl_cstates) {
1792 if (get_msr(cpu, MSR_KNL_CORE_C6_RESIDENCY, &c->c6))
1796 if (DO_BIC(BIC_CPU_c7))
1797 if (get_msr(cpu, MSR_CORE_C7_RESIDENCY, &c->c7))
1800 if (DO_BIC(BIC_Mod_c6))
1801 if (get_msr(cpu, MSR_MODULE_C6_RES_MS, &c->mc6_us))
1804 if (DO_BIC(BIC_CoreTmp)) {
1805 if (get_msr(cpu, MSR_IA32_THERM_STATUS, &msr))
1807 c->core_temp_c = tcc_activation_temp - ((msr >> 16) & 0x7F);
1810 for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
1811 if (get_mp(cpu, mp, &c->counter[i]))
1815 /* collect package counters only for 1st core in package */
1816 if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
1819 if (DO_BIC(BIC_Totl_c0)) {
1820 if (get_msr(cpu, MSR_PKG_WEIGHTED_CORE_C0_RES, &p->pkg_wtd_core_c0))
1823 if (DO_BIC(BIC_Any_c0)) {
1824 if (get_msr(cpu, MSR_PKG_ANY_CORE_C0_RES, &p->pkg_any_core_c0))
1827 if (DO_BIC(BIC_GFX_c0)) {
1828 if (get_msr(cpu, MSR_PKG_ANY_GFXE_C0_RES, &p->pkg_any_gfxe_c0))
1831 if (DO_BIC(BIC_CPUGFX)) {
1832 if (get_msr(cpu, MSR_PKG_BOTH_CORE_GFXE_C0_RES, &p->pkg_both_core_gfxe_c0))
1835 if (DO_BIC(BIC_Pkgpc3))
1836 if (get_msr(cpu, MSR_PKG_C3_RESIDENCY, &p->pc3))
1838 if (DO_BIC(BIC_Pkgpc6)) {
1839 if (do_slm_cstates) {
1840 if (get_msr(cpu, MSR_ATOM_PKG_C6_RESIDENCY, &p->pc6))
1843 if (get_msr(cpu, MSR_PKG_C6_RESIDENCY, &p->pc6))
1848 if (DO_BIC(BIC_Pkgpc2))
1849 if (get_msr(cpu, MSR_PKG_C2_RESIDENCY, &p->pc2))
1851 if (DO_BIC(BIC_Pkgpc7))
1852 if (get_msr(cpu, MSR_PKG_C7_RESIDENCY, &p->pc7))
1854 if (DO_BIC(BIC_Pkgpc8))
1855 if (get_msr(cpu, MSR_PKG_C8_RESIDENCY, &p->pc8))
1857 if (DO_BIC(BIC_Pkgpc9))
1858 if (get_msr(cpu, MSR_PKG_C9_RESIDENCY, &p->pc9))
1860 if (DO_BIC(BIC_Pkgpc10))
1861 if (get_msr(cpu, MSR_PKG_C10_RESIDENCY, &p->pc10))
1864 if (DO_BIC(BIC_CPU_LPI))
1865 p->cpu_lpi = cpuidle_cur_cpu_lpi_us;
1866 if (DO_BIC(BIC_SYS_LPI))
1867 p->sys_lpi = cpuidle_cur_sys_lpi_us;
1869 if (do_rapl & RAPL_PKG) {
1870 if (get_msr(cpu, MSR_PKG_ENERGY_STATUS, &msr))
1872 p->energy_pkg = msr & 0xFFFFFFFF;
1874 if (do_rapl & RAPL_CORES_ENERGY_STATUS) {
1875 if (get_msr(cpu, MSR_PP0_ENERGY_STATUS, &msr))
1877 p->energy_cores = msr & 0xFFFFFFFF;
1879 if (do_rapl & RAPL_DRAM) {
1880 if (get_msr(cpu, MSR_DRAM_ENERGY_STATUS, &msr))
1882 p->energy_dram = msr & 0xFFFFFFFF;
1884 if (do_rapl & RAPL_GFX) {
1885 if (get_msr(cpu, MSR_PP1_ENERGY_STATUS, &msr))
1887 p->energy_gfx = msr & 0xFFFFFFFF;
1889 if (do_rapl & RAPL_PKG_PERF_STATUS) {
1890 if (get_msr(cpu, MSR_PKG_PERF_STATUS, &msr))
1892 p->rapl_pkg_perf_status = msr & 0xFFFFFFFF;
1894 if (do_rapl & RAPL_DRAM_PERF_STATUS) {
1895 if (get_msr(cpu, MSR_DRAM_PERF_STATUS, &msr))
1897 p->rapl_dram_perf_status = msr & 0xFFFFFFFF;
1899 if (DO_BIC(BIC_PkgTmp)) {
1900 if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_STATUS, &msr))
1902 p->pkg_temp_c = tcc_activation_temp - ((msr >> 16) & 0x7F);
1905 if (DO_BIC(BIC_GFX_rc6))
1906 p->gfx_rc6_ms = gfx_cur_rc6_ms;
1908 if (DO_BIC(BIC_GFXMHz))
1909 p->gfx_mhz = gfx_cur_mhz;
1911 for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
1912 if (get_mp(cpu, mp, &p->counter[i]))
1916 gettimeofday(&t->tv_end, (struct timezone *)NULL);
1922 * MSR_PKG_CST_CONFIG_CONTROL decoding for pkg_cstate_limit:
1923 * If you change the values, note they are used both in comparisons
1924 * (>= PCL__7) and to index pkg_cstate_limit_strings[].
1927 #define PCLUKN 0 /* Unknown */
1928 #define PCLRSV 1 /* Reserved */
1929 #define PCL__0 2 /* PC0 */
1930 #define PCL__1 3 /* PC1 */
1931 #define PCL__2 4 /* PC2 */
1932 #define PCL__3 5 /* PC3 */
1933 #define PCL__4 6 /* PC4 */
1934 #define PCL__6 7 /* PC6 */
1935 #define PCL_6N 8 /* PC6 No Retention */
1936 #define PCL_6R 9 /* PC6 Retention */
1937 #define PCL__7 10 /* PC7 */
1938 #define PCL_7S 11 /* PC7 Shrink */
1939 #define PCL__8 12 /* PC8 */
1940 #define PCL__9 13 /* PC9 */
1941 #define PCLUNL 14 /* Unlimited */
1943 int pkg_cstate_limit = PCLUKN;
1944 char *pkg_cstate_limit_strings[] = { "reserved", "unknown", "pc0", "pc1", "pc2",
1945 "pc3", "pc4", "pc6", "pc6n", "pc6r", "pc7", "pc7s", "pc8", "pc9", "unlimited"};
1947 int nhm_pkg_cstate_limits[16] = {PCL__0, PCL__1, PCL__3, PCL__6, PCL__7, PCLRSV, PCLRSV, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV};
1948 int snb_pkg_cstate_limits[16] = {PCL__0, PCL__2, PCL_6N, PCL_6R, PCL__7, PCL_7S, PCLRSV, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV};
1949 int hsw_pkg_cstate_limits[16] = {PCL__0, PCL__2, PCL__3, PCL__6, PCL__7, PCL_7S, PCL__8, PCL__9, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV};
1950 int slv_pkg_cstate_limits[16] = {PCL__0, PCL__1, PCLRSV, PCLRSV, PCL__4, PCLRSV, PCL__6, PCL__7, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCL__6, PCL__7};
1951 int amt_pkg_cstate_limits[16] = {PCLUNL, PCL__1, PCL__2, PCLRSV, PCLRSV, PCLRSV, PCL__6, PCL__7, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV};
1952 int phi_pkg_cstate_limits[16] = {PCL__0, PCL__2, PCL_6N, PCL_6R, PCLRSV, PCLRSV, PCLRSV, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV};
1953 int bxt_pkg_cstate_limits[16] = {PCL__0, PCL__2, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV};
1954 int skx_pkg_cstate_limits[16] = {PCL__0, PCL__2, PCL_6N, PCL_6R, PCLRSV, PCLRSV, PCLRSV, PCLUNL, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV};
1958 calculate_tsc_tweak()
1960 tsc_tweak = base_hz / tsc_hz;
1964 dump_nhm_platform_info(void)
1966 unsigned long long msr;
1969 get_msr(base_cpu, MSR_PLATFORM_INFO, &msr);
1971 fprintf(outf, "cpu%d: MSR_PLATFORM_INFO: 0x%08llx\n", base_cpu, msr);
1973 ratio = (msr >> 40) & 0xFF;
1974 fprintf(outf, "%d * %.1f = %.1f MHz max efficiency frequency\n",
1975 ratio, bclk, ratio * bclk);
1977 ratio = (msr >> 8) & 0xFF;
1978 fprintf(outf, "%d * %.1f = %.1f MHz base frequency\n",
1979 ratio, bclk, ratio * bclk);
1981 get_msr(base_cpu, MSR_IA32_POWER_CTL, &msr);
1982 fprintf(outf, "cpu%d: MSR_IA32_POWER_CTL: 0x%08llx (C1E auto-promotion: %sabled)\n",
1983 base_cpu, msr, msr & 0x2 ? "EN" : "DIS");
1989 dump_hsw_turbo_ratio_limits(void)
1991 unsigned long long msr;
1994 get_msr(base_cpu, MSR_TURBO_RATIO_LIMIT2, &msr);
1996 fprintf(outf, "cpu%d: MSR_TURBO_RATIO_LIMIT2: 0x%08llx\n", base_cpu, msr);
1998 ratio = (msr >> 8) & 0xFF;
2000 fprintf(outf, "%d * %.1f = %.1f MHz max turbo 18 active cores\n",
2001 ratio, bclk, ratio * bclk);
2003 ratio = (msr >> 0) & 0xFF;
2005 fprintf(outf, "%d * %.1f = %.1f MHz max turbo 17 active cores\n",
2006 ratio, bclk, ratio * bclk);
2011 dump_ivt_turbo_ratio_limits(void)
2013 unsigned long long msr;
2016 get_msr(base_cpu, MSR_TURBO_RATIO_LIMIT1, &msr);
2018 fprintf(outf, "cpu%d: MSR_TURBO_RATIO_LIMIT1: 0x%08llx\n", base_cpu, msr);
2020 ratio = (msr >> 56) & 0xFF;
2022 fprintf(outf, "%d * %.1f = %.1f MHz max turbo 16 active cores\n",
2023 ratio, bclk, ratio * bclk);
2025 ratio = (msr >> 48) & 0xFF;
2027 fprintf(outf, "%d * %.1f = %.1f MHz max turbo 15 active cores\n",
2028 ratio, bclk, ratio * bclk);
2030 ratio = (msr >> 40) & 0xFF;
2032 fprintf(outf, "%d * %.1f = %.1f MHz max turbo 14 active cores\n",
2033 ratio, bclk, ratio * bclk);
2035 ratio = (msr >> 32) & 0xFF;
2037 fprintf(outf, "%d * %.1f = %.1f MHz max turbo 13 active cores\n",
2038 ratio, bclk, ratio * bclk);
2040 ratio = (msr >> 24) & 0xFF;
2042 fprintf(outf, "%d * %.1f = %.1f MHz max turbo 12 active cores\n",
2043 ratio, bclk, ratio * bclk);
2045 ratio = (msr >> 16) & 0xFF;
2047 fprintf(outf, "%d * %.1f = %.1f MHz max turbo 11 active cores\n",
2048 ratio, bclk, ratio * bclk);
2050 ratio = (msr >> 8) & 0xFF;
2052 fprintf(outf, "%d * %.1f = %.1f MHz max turbo 10 active cores\n",
2053 ratio, bclk, ratio * bclk);
2055 ratio = (msr >> 0) & 0xFF;
2057 fprintf(outf, "%d * %.1f = %.1f MHz max turbo 9 active cores\n",
2058 ratio, bclk, ratio * bclk);
2061 int has_turbo_ratio_group_limits(int family, int model)
2068 case INTEL_FAM6_ATOM_GOLDMONT:
2069 case INTEL_FAM6_SKYLAKE_X:
2070 case INTEL_FAM6_ATOM_DENVERTON:
2077 dump_turbo_ratio_limits(int family, int model)
2079 unsigned long long msr, core_counts;
2080 unsigned int ratio, group_size;
2082 get_msr(base_cpu, MSR_TURBO_RATIO_LIMIT, &msr);
2083 fprintf(outf, "cpu%d: MSR_TURBO_RATIO_LIMIT: 0x%08llx\n", base_cpu, msr);
2085 if (has_turbo_ratio_group_limits(family, model)) {
2086 get_msr(base_cpu, MSR_TURBO_RATIO_LIMIT1, &core_counts);
2087 fprintf(outf, "cpu%d: MSR_TURBO_RATIO_LIMIT1: 0x%08llx\n", base_cpu, core_counts);
2089 core_counts = 0x0807060504030201;
2092 ratio = (msr >> 56) & 0xFF;
2093 group_size = (core_counts >> 56) & 0xFF;
2095 fprintf(outf, "%d * %.1f = %.1f MHz max turbo %d active cores\n",
2096 ratio, bclk, ratio * bclk, group_size);
2098 ratio = (msr >> 48) & 0xFF;
2099 group_size = (core_counts >> 48) & 0xFF;
2101 fprintf(outf, "%d * %.1f = %.1f MHz max turbo %d active cores\n",
2102 ratio, bclk, ratio * bclk, group_size);
2104 ratio = (msr >> 40) & 0xFF;
2105 group_size = (core_counts >> 40) & 0xFF;
2107 fprintf(outf, "%d * %.1f = %.1f MHz max turbo %d active cores\n",
2108 ratio, bclk, ratio * bclk, group_size);
2110 ratio = (msr >> 32) & 0xFF;
2111 group_size = (core_counts >> 32) & 0xFF;
2113 fprintf(outf, "%d * %.1f = %.1f MHz max turbo %d active cores\n",
2114 ratio, bclk, ratio * bclk, group_size);
2116 ratio = (msr >> 24) & 0xFF;
2117 group_size = (core_counts >> 24) & 0xFF;
2119 fprintf(outf, "%d * %.1f = %.1f MHz max turbo %d active cores\n",
2120 ratio, bclk, ratio * bclk, group_size);
2122 ratio = (msr >> 16) & 0xFF;
2123 group_size = (core_counts >> 16) & 0xFF;
2125 fprintf(outf, "%d * %.1f = %.1f MHz max turbo %d active cores\n",
2126 ratio, bclk, ratio * bclk, group_size);
2128 ratio = (msr >> 8) & 0xFF;
2129 group_size = (core_counts >> 8) & 0xFF;
2131 fprintf(outf, "%d * %.1f = %.1f MHz max turbo %d active cores\n",
2132 ratio, bclk, ratio * bclk, group_size);
2134 ratio = (msr >> 0) & 0xFF;
2135 group_size = (core_counts >> 0) & 0xFF;
2137 fprintf(outf, "%d * %.1f = %.1f MHz max turbo %d active cores\n",
2138 ratio, bclk, ratio * bclk, group_size);
2143 dump_atom_turbo_ratio_limits(void)
2145 unsigned long long msr;
2148 get_msr(base_cpu, MSR_ATOM_CORE_RATIOS, &msr);
2149 fprintf(outf, "cpu%d: MSR_ATOM_CORE_RATIOS: 0x%08llx\n", base_cpu, msr & 0xFFFFFFFF);
2151 ratio = (msr >> 0) & 0x3F;
2153 fprintf(outf, "%d * %.1f = %.1f MHz minimum operating frequency\n",
2154 ratio, bclk, ratio * bclk);
2156 ratio = (msr >> 8) & 0x3F;
2158 fprintf(outf, "%d * %.1f = %.1f MHz low frequency mode (LFM)\n",
2159 ratio, bclk, ratio * bclk);
2161 ratio = (msr >> 16) & 0x3F;
2163 fprintf(outf, "%d * %.1f = %.1f MHz base frequency\n",
2164 ratio, bclk, ratio * bclk);
2166 get_msr(base_cpu, MSR_ATOM_CORE_TURBO_RATIOS, &msr);
2167 fprintf(outf, "cpu%d: MSR_ATOM_CORE_TURBO_RATIOS: 0x%08llx\n", base_cpu, msr & 0xFFFFFFFF);
2169 ratio = (msr >> 24) & 0x3F;
2171 fprintf(outf, "%d * %.1f = %.1f MHz max turbo 4 active cores\n",
2172 ratio, bclk, ratio * bclk);
2174 ratio = (msr >> 16) & 0x3F;
2176 fprintf(outf, "%d * %.1f = %.1f MHz max turbo 3 active cores\n",
2177 ratio, bclk, ratio * bclk);
2179 ratio = (msr >> 8) & 0x3F;
2181 fprintf(outf, "%d * %.1f = %.1f MHz max turbo 2 active cores\n",
2182 ratio, bclk, ratio * bclk);
2184 ratio = (msr >> 0) & 0x3F;
2186 fprintf(outf, "%d * %.1f = %.1f MHz max turbo 1 active core\n",
2187 ratio, bclk, ratio * bclk);
2191 dump_knl_turbo_ratio_limits(void)
2193 const unsigned int buckets_no = 7;
2195 unsigned long long msr;
2196 int delta_cores, delta_ratio;
2198 unsigned int cores[buckets_no];
2199 unsigned int ratio[buckets_no];
2201 get_msr(base_cpu, MSR_TURBO_RATIO_LIMIT, &msr);
2203 fprintf(outf, "cpu%d: MSR_TURBO_RATIO_LIMIT: 0x%08llx\n",
2207 * Turbo encoding in KNL is as follows:
2209 * [7:1] -- Base value of number of active cores of bucket 1.
2210 * [15:8] -- Base value of freq ratio of bucket 1.
2211 * [20:16] -- +ve delta of number of active cores of bucket 2.
2212 * i.e. active cores of bucket 2 =
2213 * active cores of bucket 1 + delta
2214 * [23:21] -- Negative delta of freq ratio of bucket 2.
2215 * i.e. freq ratio of bucket 2 =
2216 * freq ratio of bucket 1 - delta
2217 * [28:24]-- +ve delta of number of active cores of bucket 3.
2218 * [31:29]-- -ve delta of freq ratio of bucket 3.
2219 * [36:32]-- +ve delta of number of active cores of bucket 4.
2220 * [39:37]-- -ve delta of freq ratio of bucket 4.
2221 * [44:40]-- +ve delta of number of active cores of bucket 5.
2222 * [47:45]-- -ve delta of freq ratio of bucket 5.
2223 * [52:48]-- +ve delta of number of active cores of bucket 6.
2224 * [55:53]-- -ve delta of freq ratio of bucket 6.
2225 * [60:56]-- +ve delta of number of active cores of bucket 7.
2226 * [63:61]-- -ve delta of freq ratio of bucket 7.
2230 cores[b_nr] = (msr & 0xFF) >> 1;
2231 ratio[b_nr] = (msr >> 8) & 0xFF;
2233 for (i = 16; i < 64; i += 8) {
2234 delta_cores = (msr >> i) & 0x1F;
2235 delta_ratio = (msr >> (i + 5)) & 0x7;
2237 cores[b_nr + 1] = cores[b_nr] + delta_cores;
2238 ratio[b_nr + 1] = ratio[b_nr] - delta_ratio;
2242 for (i = buckets_no - 1; i >= 0; i--)
2243 if (i > 0 ? ratio[i] != ratio[i - 1] : 1)
2245 "%d * %.1f = %.1f MHz max turbo %d active cores\n",
2246 ratio[i], bclk, ratio[i] * bclk, cores[i]);
2250 dump_nhm_cst_cfg(void)
2252 unsigned long long msr;
2254 get_msr(base_cpu, MSR_PKG_CST_CONFIG_CONTROL, &msr);
2256 fprintf(outf, "cpu%d: MSR_PKG_CST_CONFIG_CONTROL: 0x%08llx", base_cpu, msr);
2258 fprintf(outf, " (%s%s%s%s%slocked, pkg-cstate-limit=%d (%s)",
2259 (msr & SNB_C3_AUTO_UNDEMOTE) ? "UNdemote-C3, " : "",
2260 (msr & SNB_C1_AUTO_UNDEMOTE) ? "UNdemote-C1, " : "",
2261 (msr & NHM_C3_AUTO_DEMOTE) ? "demote-C3, " : "",
2262 (msr & NHM_C1_AUTO_DEMOTE) ? "demote-C1, " : "",
2263 (msr & (1 << 15)) ? "" : "UN",
2264 (unsigned int)msr & 0xF,
2265 pkg_cstate_limit_strings[pkg_cstate_limit]);
2267 #define AUTOMATIC_CSTATE_CONVERSION (1UL << 16)
2268 if (has_automatic_cstate_conversion) {
2269 fprintf(outf, ", automatic c-state conversion=%s",
2270 (msr & AUTOMATIC_CSTATE_CONVERSION) ? "on" : "off");
2273 fprintf(outf, ")\n");
2279 dump_config_tdp(void)
2281 unsigned long long msr;
2283 get_msr(base_cpu, MSR_CONFIG_TDP_NOMINAL, &msr);
2284 fprintf(outf, "cpu%d: MSR_CONFIG_TDP_NOMINAL: 0x%08llx", base_cpu, msr);
2285 fprintf(outf, " (base_ratio=%d)\n", (unsigned int)msr & 0xFF);
2287 get_msr(base_cpu, MSR_CONFIG_TDP_LEVEL_1, &msr);
2288 fprintf(outf, "cpu%d: MSR_CONFIG_TDP_LEVEL_1: 0x%08llx (", base_cpu, msr);
2290 fprintf(outf, "PKG_MIN_PWR_LVL1=%d ", (unsigned int)(msr >> 48) & 0x7FFF);
2291 fprintf(outf, "PKG_MAX_PWR_LVL1=%d ", (unsigned int)(msr >> 32) & 0x7FFF);
2292 fprintf(outf, "LVL1_RATIO=%d ", (unsigned int)(msr >> 16) & 0xFF);
2293 fprintf(outf, "PKG_TDP_LVL1=%d", (unsigned int)(msr) & 0x7FFF);
2295 fprintf(outf, ")\n");
2297 get_msr(base_cpu, MSR_CONFIG_TDP_LEVEL_2, &msr);
2298 fprintf(outf, "cpu%d: MSR_CONFIG_TDP_LEVEL_2: 0x%08llx (", base_cpu, msr);
2300 fprintf(outf, "PKG_MIN_PWR_LVL2=%d ", (unsigned int)(msr >> 48) & 0x7FFF);
2301 fprintf(outf, "PKG_MAX_PWR_LVL2=%d ", (unsigned int)(msr >> 32) & 0x7FFF);
2302 fprintf(outf, "LVL2_RATIO=%d ", (unsigned int)(msr >> 16) & 0xFF);
2303 fprintf(outf, "PKG_TDP_LVL2=%d", (unsigned int)(msr) & 0x7FFF);
2305 fprintf(outf, ")\n");
2307 get_msr(base_cpu, MSR_CONFIG_TDP_CONTROL, &msr);
2308 fprintf(outf, "cpu%d: MSR_CONFIG_TDP_CONTROL: 0x%08llx (", base_cpu, msr);
2310 fprintf(outf, "TDP_LEVEL=%d ", (unsigned int)(msr) & 0x3);
2311 fprintf(outf, " lock=%d", (unsigned int)(msr >> 31) & 1);
2312 fprintf(outf, ")\n");
2314 get_msr(base_cpu, MSR_TURBO_ACTIVATION_RATIO, &msr);
2315 fprintf(outf, "cpu%d: MSR_TURBO_ACTIVATION_RATIO: 0x%08llx (", base_cpu, msr);
2316 fprintf(outf, "MAX_NON_TURBO_RATIO=%d", (unsigned int)(msr) & 0xFF);
2317 fprintf(outf, " lock=%d", (unsigned int)(msr >> 31) & 1);
2318 fprintf(outf, ")\n");
2321 unsigned int irtl_time_units[] = {1, 32, 1024, 32768, 1048576, 33554432, 0, 0 };
2323 void print_irtl(void)
2325 unsigned long long msr;
2327 get_msr(base_cpu, MSR_PKGC3_IRTL, &msr);
2328 fprintf(outf, "cpu%d: MSR_PKGC3_IRTL: 0x%08llx (", base_cpu, msr);
2329 fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
2330 (msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
2332 get_msr(base_cpu, MSR_PKGC6_IRTL, &msr);
2333 fprintf(outf, "cpu%d: MSR_PKGC6_IRTL: 0x%08llx (", base_cpu, msr);
2334 fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
2335 (msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
2337 get_msr(base_cpu, MSR_PKGC7_IRTL, &msr);
2338 fprintf(outf, "cpu%d: MSR_PKGC7_IRTL: 0x%08llx (", base_cpu, msr);
2339 fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
2340 (msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
2345 get_msr(base_cpu, MSR_PKGC8_IRTL, &msr);
2346 fprintf(outf, "cpu%d: MSR_PKGC8_IRTL: 0x%08llx (", base_cpu, msr);
2347 fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
2348 (msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
2350 get_msr(base_cpu, MSR_PKGC9_IRTL, &msr);
2351 fprintf(outf, "cpu%d: MSR_PKGC9_IRTL: 0x%08llx (", base_cpu, msr);
2352 fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
2353 (msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
2355 get_msr(base_cpu, MSR_PKGC10_IRTL, &msr);
2356 fprintf(outf, "cpu%d: MSR_PKGC10_IRTL: 0x%08llx (", base_cpu, msr);
2357 fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
2358 (msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
2361 void free_fd_percpu(void)
2365 for (i = 0; i < topo.max_cpu_num + 1; ++i) {
2366 if (fd_percpu[i] != 0)
2367 close(fd_percpu[i]);
2373 void free_all_buffers(void)
2377 CPU_FREE(cpu_present_set);
2378 cpu_present_set = NULL;
2379 cpu_present_setsize = 0;
2381 CPU_FREE(cpu_affinity_set);
2382 cpu_affinity_set = NULL;
2383 cpu_affinity_setsize = 0;
2391 package_even = NULL;
2401 free(output_buffer);
2402 output_buffer = NULL;
2407 free(irq_column_2_cpu);
2410 for (i = 0; i <= topo.max_cpu_num; ++i) {
2411 if (cpus[i].put_ids)
2412 CPU_FREE(cpus[i].put_ids);
2419 * Parse a file containing a single int.
2421 int parse_int_file(const char *fmt, ...)
2424 char path[PATH_MAX];
2428 va_start(args, fmt);
2429 vsnprintf(path, sizeof(path), fmt, args);
2431 filep = fopen_or_die(path, "r");
2432 if (fscanf(filep, "%d", &value) != 1)
2433 err(1, "%s: failed to parse number from file", path);
2439 * cpu_is_first_core_in_package(cpu)
2440 * return 1 if given CPU is 1st core in package
2442 int cpu_is_first_core_in_package(int cpu)
2444 return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", cpu);
2447 int get_physical_package_id(int cpu)
2449 return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", cpu);
2452 int get_core_id(int cpu)
2454 return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", cpu);
2457 void set_node_data(void)
2463 struct pkg_node_info {
2468 pni = calloc(topo.num_packages, sizeof(struct pkg_node_info));
2470 err(1, "calloc pkg_node_count");
2472 for (pkg = 0; pkg < topo.num_packages; pkg++)
2473 pni[pkg].min = topo.num_cpus;
2475 for (node = 0; node <= topo.max_node_num; node++) {
2476 /* find the "first" cpu in the node */
2477 sprintf(path, "/sys/bus/node/devices/node%d/cpulist", node);
2478 filep = fopen(path, "r");
2481 fscanf(filep, "%d", &cpu);
2484 pkg = cpus[cpu].physical_package_id;
2487 if (node < pni[pkg].min)
2488 pni[pkg].min = node;
2491 for (pkg = 0; pkg < topo.num_packages; pkg++)
2492 if (pni[pkg].count > topo.nodes_per_pkg)
2493 topo.nodes_per_pkg = pni[0].count;
2495 for (cpu = 0; cpu < topo.num_cpus; cpu++) {
2496 pkg = cpus[cpu].physical_package_id;
2497 node = cpus[cpu].physical_node_id;
2498 cpus[cpu].logical_node_id = node - pni[pkg].min;
2504 int get_physical_node_id(struct cpu_topology *thiscpu)
2509 int cpu = thiscpu->logical_cpu_id;
2511 for (i = 0; i <= topo.max_cpu_num; i++) {
2512 sprintf(path, "/sys/devices/system/cpu/cpu%d/node%i/cpulist",
2514 filep = fopen(path, "r");
2523 int get_thread_siblings(struct cpu_topology *thiscpu)
2525 char path[80], character;
2528 int so, shift, sib_core;
2529 int cpu = thiscpu->logical_cpu_id;
2530 int offset = topo.max_cpu_num + 1;
2534 thiscpu->put_ids = CPU_ALLOC((topo.max_cpu_num + 1));
2535 if (thiscpu->thread_id < 0)
2536 thiscpu->thread_id = thread_id++;
2537 if (!thiscpu->put_ids)
2540 size = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
2541 CPU_ZERO_S(size, thiscpu->put_ids);
2544 "/sys/devices/system/cpu/cpu%d/topology/thread_siblings", cpu);
2545 filep = fopen_or_die(path, "r");
2547 offset -= BITMASK_SIZE;
2548 fscanf(filep, "%lx%c", &map, &character);
2549 for (shift = 0; shift < BITMASK_SIZE; shift++) {
2550 if ((map >> shift) & 0x1) {
2551 so = shift + offset;
2552 sib_core = get_core_id(so);
2553 if (sib_core == thiscpu->physical_core_id) {
2554 CPU_SET_S(so, size, thiscpu->put_ids);
2556 (cpus[so].thread_id < 0))
2557 cpus[so].thread_id =
2562 } while (!strncmp(&character, ",", 1));
2565 return CPU_COUNT_S(size, thiscpu->put_ids);
2569 * run func(thread, core, package) in topology order
2570 * skip non-present cpus
2573 int for_all_cpus_2(int (func)(struct thread_data *, struct core_data *,
2574 struct pkg_data *, struct thread_data *, struct core_data *,
2575 struct pkg_data *), struct thread_data *thread_base,
2576 struct core_data *core_base, struct pkg_data *pkg_base,
2577 struct thread_data *thread_base2, struct core_data *core_base2,
2578 struct pkg_data *pkg_base2)
2580 int retval, pkg_no, node_no, core_no, thread_no;
2582 for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
2583 for (node_no = 0; node_no < topo.nodes_per_pkg; ++node_no) {
2584 for (core_no = 0; core_no < topo.cores_per_node;
2586 for (thread_no = 0; thread_no <
2587 topo.threads_per_core; ++thread_no) {
2588 struct thread_data *t, *t2;
2589 struct core_data *c, *c2;
2590 struct pkg_data *p, *p2;
2592 t = GET_THREAD(thread_base, thread_no,
2596 if (cpu_is_not_present(t->cpu_id))
2599 t2 = GET_THREAD(thread_base2, thread_no,
2603 c = GET_CORE(core_base, core_no,
2605 c2 = GET_CORE(core_base2, core_no,
2609 p = GET_PKG(pkg_base, pkg_no);
2610 p2 = GET_PKG(pkg_base2, pkg_no);
2612 retval = func(t, c, p, t2, c2, p2);
2623 * run func(cpu) on every cpu in /proc/stat
2624 * return max_cpu number
2626 int for_all_proc_cpus(int (func)(int))
2632 fp = fopen_or_die(proc_stat, "r");
2634 retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
2636 err(1, "%s: failed to parse format", proc_stat);
2639 retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
2643 retval = func(cpu_num);
2653 void re_initialize(void)
2656 setup_all_buffers();
2657 printf("turbostat: re-initialized with num_cpus %d\n", topo.num_cpus);
2660 void set_max_cpu_num(void)
2663 unsigned long dummy;
2665 topo.max_cpu_num = 0;
2666 filep = fopen_or_die(
2667 "/sys/devices/system/cpu/cpu0/topology/thread_siblings",
2669 while (fscanf(filep, "%lx,", &dummy) == 1)
2670 topo.max_cpu_num += BITMASK_SIZE;
2672 topo.max_cpu_num--; /* 0 based */
2677 * remember the last one seen, it will be the max
2679 int count_cpus(int cpu)
2684 int mark_cpu_present(int cpu)
2686 CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
2690 int init_thread_id(int cpu)
2692 cpus[cpu].thread_id = -1;
2697 * snapshot_proc_interrupts()
2699 * read and record summary of /proc/interrupts
2701 * return 1 if config change requires a restart, else return 0
2703 int snapshot_proc_interrupts(void)
2709 fp = fopen_or_die("/proc/interrupts", "r");
2713 /* read 1st line of /proc/interrupts to get cpu* name for each column */
2714 for (column = 0; column < topo.num_cpus; ++column) {
2717 retval = fscanf(fp, " CPU%d", &cpu_number);
2721 if (cpu_number > topo.max_cpu_num) {
2722 warn("/proc/interrupts: cpu%d: > %d", cpu_number, topo.max_cpu_num);
2726 irq_column_2_cpu[column] = cpu_number;
2727 irqs_per_cpu[cpu_number] = 0;
2730 /* read /proc/interrupt count lines and sum up irqs per cpu */
2735 retval = fscanf(fp, " %s:", buf); /* flush irq# "N:" */
2739 /* read the count per cpu */
2740 for (column = 0; column < topo.num_cpus; ++column) {
2742 int cpu_number, irq_count;
2744 retval = fscanf(fp, " %d", &irq_count);
2748 cpu_number = irq_column_2_cpu[column];
2749 irqs_per_cpu[cpu_number] += irq_count;
2753 while (getc(fp) != '\n')
2754 ; /* flush interrupt description */
2760 * snapshot_gfx_rc6_ms()
2762 * record snapshot of
2763 * /sys/class/drm/card0/power/rc6_residency_ms
2765 * return 1 if config change requires a restart, else return 0
2767 int snapshot_gfx_rc6_ms(void)
2772 fp = fopen_or_die("/sys/class/drm/card0/power/rc6_residency_ms", "r");
2774 retval = fscanf(fp, "%lld", &gfx_cur_rc6_ms);
2783 * snapshot_gfx_mhz()
2785 * record snapshot of
2786 * /sys/class/graphics/fb0/device/drm/card0/gt_cur_freq_mhz
2788 * return 1 if config change requires a restart, else return 0
2790 int snapshot_gfx_mhz(void)
2796 fp = fopen_or_die("/sys/class/graphics/fb0/device/drm/card0/gt_cur_freq_mhz", "r");
2802 retval = fscanf(fp, "%d", &gfx_cur_mhz);
2810 * snapshot_cpu_lpi()
2812 * record snapshot of
2813 * /sys/devices/system/cpu/cpuidle/low_power_idle_cpu_residency_us
2815 * return 1 if config change requires a restart, else return 0
2817 int snapshot_cpu_lpi_us(void)
2822 fp = fopen_or_die("/sys/devices/system/cpu/cpuidle/low_power_idle_cpu_residency_us", "r");
2824 retval = fscanf(fp, "%lld", &cpuidle_cur_cpu_lpi_us);
2833 * snapshot_sys_lpi()
2835 * record snapshot of
2836 * /sys/devices/system/cpu/cpuidle/low_power_idle_system_residency_us
2838 * return 1 if config change requires a restart, else return 0
2840 int snapshot_sys_lpi_us(void)
2845 fp = fopen_or_die("/sys/devices/system/cpu/cpuidle/low_power_idle_system_residency_us", "r");
2847 retval = fscanf(fp, "%lld", &cpuidle_cur_sys_lpi_us);
2856 * snapshot /proc and /sys files
2858 * return 1 if configuration restart needed, else return 0
2860 int snapshot_proc_sysfs_files(void)
2862 if (DO_BIC(BIC_IRQ))
2863 if (snapshot_proc_interrupts())
2866 if (DO_BIC(BIC_GFX_rc6))
2867 snapshot_gfx_rc6_ms();
2869 if (DO_BIC(BIC_GFXMHz))
2872 if (DO_BIC(BIC_CPU_LPI))
2873 snapshot_cpu_lpi_us();
2875 if (DO_BIC(BIC_SYS_LPI))
2876 snapshot_sys_lpi_us();
2883 static void signal_handler (int signal)
2889 fprintf(stderr, " SIGINT\n");
2893 fprintf(stderr, "SIGUSR1\n");
2896 /* make sure this manually-invoked interval is at least 1ms long */
2897 nanosleep(&one_msec, NULL);
2900 void setup_signal_handler(void)
2902 struct sigaction sa;
2904 memset(&sa, 0, sizeof(sa));
2906 sa.sa_handler = &signal_handler;
2908 if (sigaction(SIGINT, &sa, NULL) < 0)
2909 err(1, "sigaction SIGINT");
2910 if (sigaction(SIGUSR1, &sa, NULL) < 0)
2911 err(1, "sigaction SIGUSR1");
2916 struct timeval select_timeout;
2921 FD_SET(0, &readfds);
2923 if (!isatty(fileno(stdin))) {
2924 nanosleep(&interval_ts, NULL);
2928 select_timeout = interval_tv;
2929 retval = select(1, &readfds, NULL, NULL, &select_timeout);
2932 switch (getc(stdin)) {
2937 /* make sure this manually-invoked interval is at least 1ms long */
2938 nanosleep(&one_msec, NULL);
2943 void turbostat_loop()
2949 setup_signal_handler();
2954 snapshot_proc_sysfs_files();
2955 retval = for_all_cpus(get_counters, EVEN_COUNTERS);
2956 first_counter_read = 0;
2959 } else if (retval == -1) {
2960 if (restarted > 1) {
2968 gettimeofday(&tv_even, (struct timezone *)NULL);
2971 if (for_all_proc_cpus(cpu_is_not_present)) {
2976 if (snapshot_proc_sysfs_files())
2978 retval = for_all_cpus(get_counters, ODD_COUNTERS);
2981 } else if (retval == -1) {
2985 gettimeofday(&tv_odd, (struct timezone *)NULL);
2986 timersub(&tv_odd, &tv_even, &tv_delta);
2987 if (for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS)) {
2991 compute_average(EVEN_COUNTERS);
2992 format_all_counters(EVEN_COUNTERS);
2993 flush_output_stdout();
2996 if (num_iterations && ++done_iters >= num_iterations)
2999 if (snapshot_proc_sysfs_files())
3001 retval = for_all_cpus(get_counters, EVEN_COUNTERS);
3004 } else if (retval == -1) {
3008 gettimeofday(&tv_even, (struct timezone *)NULL);
3009 timersub(&tv_even, &tv_odd, &tv_delta);
3010 if (for_all_cpus_2(delta_cpu, EVEN_COUNTERS, ODD_COUNTERS)) {
3014 compute_average(ODD_COUNTERS);
3015 format_all_counters(ODD_COUNTERS);
3016 flush_output_stdout();
3019 if (num_iterations && ++done_iters >= num_iterations)
3024 void check_dev_msr()
3029 sprintf(pathname, "/dev/cpu/%d/msr", base_cpu);
3030 if (stat(pathname, &sb))
3031 if (system("/sbin/modprobe msr > /dev/null 2>&1"))
3032 err(-5, "no /dev/cpu/0/msr, Try \"# modprobe msr\" ");
3035 void check_permissions()
3037 struct __user_cap_header_struct cap_header_data;
3038 cap_user_header_t cap_header = &cap_header_data;
3039 struct __user_cap_data_struct cap_data_data;
3040 cap_user_data_t cap_data = &cap_data_data;
3041 extern int capget(cap_user_header_t hdrp, cap_user_data_t datap);
3045 /* check for CAP_SYS_RAWIO */
3046 cap_header->pid = getpid();
3047 cap_header->version = _LINUX_CAPABILITY_VERSION;
3048 if (capget(cap_header, cap_data) < 0)
3049 err(-6, "capget(2) failed");
3051 if ((cap_data->effective & (1 << CAP_SYS_RAWIO)) == 0) {
3053 warnx("capget(CAP_SYS_RAWIO) failed,"
3054 " try \"# setcap cap_sys_rawio=ep %s\"", progname);
3057 /* test file permissions */
3058 sprintf(pathname, "/dev/cpu/%d/msr", base_cpu);
3059 if (euidaccess(pathname, R_OK)) {
3061 warn("/dev/cpu/0/msr open failed, try chown or chmod +r /dev/cpu/*/msr");
3064 /* if all else fails, thell them to be root */
3067 warnx("... or simply run as root");
3074 * NHM adds support for additional MSRs:
3076 * MSR_SMI_COUNT 0x00000034
3078 * MSR_PLATFORM_INFO 0x000000ce
3079 * MSR_PKG_CST_CONFIG_CONTROL 0x000000e2
3081 * MSR_MISC_PWR_MGMT 0x000001aa
3083 * MSR_PKG_C3_RESIDENCY 0x000003f8
3084 * MSR_PKG_C6_RESIDENCY 0x000003f9
3085 * MSR_CORE_C3_RESIDENCY 0x000003fc
3086 * MSR_CORE_C6_RESIDENCY 0x000003fd
3089 * sets global pkg_cstate_limit to decode MSR_PKG_CST_CONFIG_CONTROL
3090 * sets has_misc_feature_control
3092 int probe_nhm_msrs(unsigned int family, unsigned int model)
3094 unsigned long long msr;
3095 unsigned int base_ratio;
3096 int *pkg_cstate_limits;
3104 bclk = discover_bclk(family, model);
3107 case INTEL_FAM6_NEHALEM_EP: /* Core i7, Xeon 5500 series - Bloomfield, Gainstown NHM-EP */
3108 case INTEL_FAM6_NEHALEM: /* Core i7 and i5 Processor - Clarksfield, Lynnfield, Jasper Forest */
3109 case 0x1F: /* Core i7 and i5 Processor - Nehalem */
3110 case INTEL_FAM6_WESTMERE: /* Westmere Client - Clarkdale, Arrandale */
3111 case INTEL_FAM6_WESTMERE_EP: /* Westmere EP - Gulftown */
3112 case INTEL_FAM6_NEHALEM_EX: /* Nehalem-EX Xeon - Beckton */
3113 case INTEL_FAM6_WESTMERE_EX: /* Westmere-EX Xeon - Eagleton */
3114 pkg_cstate_limits = nhm_pkg_cstate_limits;
3116 case INTEL_FAM6_SANDYBRIDGE: /* SNB */
3117 case INTEL_FAM6_SANDYBRIDGE_X: /* SNB Xeon */
3118 case INTEL_FAM6_IVYBRIDGE: /* IVB */
3119 case INTEL_FAM6_IVYBRIDGE_X: /* IVB Xeon */
3120 pkg_cstate_limits = snb_pkg_cstate_limits;
3121 has_misc_feature_control = 1;
3123 case INTEL_FAM6_HASWELL_CORE: /* HSW */
3124 case INTEL_FAM6_HASWELL_X: /* HSX */
3125 case INTEL_FAM6_HASWELL_ULT: /* HSW */
3126 case INTEL_FAM6_HASWELL_GT3E: /* HSW */
3127 case INTEL_FAM6_BROADWELL_CORE: /* BDW */
3128 case INTEL_FAM6_BROADWELL_GT3E: /* BDW */
3129 case INTEL_FAM6_BROADWELL_X: /* BDX */
3130 case INTEL_FAM6_BROADWELL_XEON_D: /* BDX-DE */
3131 case INTEL_FAM6_SKYLAKE_MOBILE: /* SKL */
3132 case INTEL_FAM6_SKYLAKE_DESKTOP: /* SKL */
3133 case INTEL_FAM6_KABYLAKE_MOBILE: /* KBL */
3134 case INTEL_FAM6_KABYLAKE_DESKTOP: /* KBL */
3135 case INTEL_FAM6_CANNONLAKE_MOBILE: /* CNL */
3136 pkg_cstate_limits = hsw_pkg_cstate_limits;
3137 has_misc_feature_control = 1;
3139 case INTEL_FAM6_SKYLAKE_X: /* SKX */
3140 pkg_cstate_limits = skx_pkg_cstate_limits;
3141 has_misc_feature_control = 1;
3143 case INTEL_FAM6_ATOM_SILVERMONT1: /* BYT */
3144 no_MSR_MISC_PWR_MGMT = 1;
3145 case INTEL_FAM6_ATOM_SILVERMONT2: /* AVN */
3146 pkg_cstate_limits = slv_pkg_cstate_limits;
3148 case INTEL_FAM6_ATOM_AIRMONT: /* AMT */
3149 pkg_cstate_limits = amt_pkg_cstate_limits;
3150 no_MSR_MISC_PWR_MGMT = 1;
3152 case INTEL_FAM6_XEON_PHI_KNL: /* PHI */
3153 case INTEL_FAM6_XEON_PHI_KNM:
3154 pkg_cstate_limits = phi_pkg_cstate_limits;
3156 case INTEL_FAM6_ATOM_GOLDMONT: /* BXT */
3157 case INTEL_FAM6_ATOM_GEMINI_LAKE:
3158 case INTEL_FAM6_ATOM_DENVERTON: /* DNV */
3159 pkg_cstate_limits = bxt_pkg_cstate_limits;
3164 get_msr(base_cpu, MSR_PKG_CST_CONFIG_CONTROL, &msr);
3165 pkg_cstate_limit = pkg_cstate_limits[msr & 0xF];
3167 get_msr(base_cpu, MSR_PLATFORM_INFO, &msr);
3168 base_ratio = (msr >> 8) & 0xFF;
3170 base_hz = base_ratio * bclk * 1000000;
3175 * SLV client has support for unique MSRs:
3177 * MSR_CC6_DEMOTION_POLICY_CONFIG
3178 * MSR_MC6_DEMOTION_POLICY_CONFIG
3181 int has_slv_msrs(unsigned int family, unsigned int model)
3187 case INTEL_FAM6_ATOM_SILVERMONT1:
3188 case INTEL_FAM6_ATOM_MERRIFIELD:
3189 case INTEL_FAM6_ATOM_MOOREFIELD:
3194 int is_dnv(unsigned int family, unsigned int model)
3201 case INTEL_FAM6_ATOM_DENVERTON:
3206 int is_bdx(unsigned int family, unsigned int model)
3213 case INTEL_FAM6_BROADWELL_X:
3214 case INTEL_FAM6_BROADWELL_XEON_D:
3219 int is_skx(unsigned int family, unsigned int model)
3226 case INTEL_FAM6_SKYLAKE_X:
3232 int has_turbo_ratio_limit(unsigned int family, unsigned int model)
3234 if (has_slv_msrs(family, model))
3238 /* Nehalem compatible, but do not include turbo-ratio limit support */
3239 case INTEL_FAM6_NEHALEM_EX: /* Nehalem-EX Xeon - Beckton */
3240 case INTEL_FAM6_WESTMERE_EX: /* Westmere-EX Xeon - Eagleton */
3241 case INTEL_FAM6_XEON_PHI_KNL: /* PHI - Knights Landing (different MSR definition) */
3242 case INTEL_FAM6_XEON_PHI_KNM:
3248 int has_atom_turbo_ratio_limit(unsigned int family, unsigned int model)
3250 if (has_slv_msrs(family, model))
3255 int has_ivt_turbo_ratio_limit(unsigned int family, unsigned int model)
3264 case INTEL_FAM6_IVYBRIDGE_X: /* IVB Xeon */
3265 case INTEL_FAM6_HASWELL_X: /* HSW Xeon */
3271 int has_hsw_turbo_ratio_limit(unsigned int family, unsigned int model)
3280 case INTEL_FAM6_HASWELL_X: /* HSW Xeon */
3287 int has_knl_turbo_ratio_limit(unsigned int family, unsigned int model)
3296 case INTEL_FAM6_XEON_PHI_KNL: /* Knights Landing */
3297 case INTEL_FAM6_XEON_PHI_KNM:
3303 int has_glm_turbo_ratio_limit(unsigned int family, unsigned int model)
3312 case INTEL_FAM6_ATOM_GOLDMONT:
3313 case INTEL_FAM6_SKYLAKE_X:
3319 int has_config_tdp(unsigned int family, unsigned int model)
3328 case INTEL_FAM6_IVYBRIDGE: /* IVB */
3329 case INTEL_FAM6_HASWELL_CORE: /* HSW */
3330 case INTEL_FAM6_HASWELL_X: /* HSX */
3331 case INTEL_FAM6_HASWELL_ULT: /* HSW */
3332 case INTEL_FAM6_HASWELL_GT3E: /* HSW */
3333 case INTEL_FAM6_BROADWELL_CORE: /* BDW */
3334 case INTEL_FAM6_BROADWELL_GT3E: /* BDW */
3335 case INTEL_FAM6_BROADWELL_X: /* BDX */
3336 case INTEL_FAM6_BROADWELL_XEON_D: /* BDX-DE */
3337 case INTEL_FAM6_SKYLAKE_MOBILE: /* SKL */
3338 case INTEL_FAM6_SKYLAKE_DESKTOP: /* SKL */
3339 case INTEL_FAM6_KABYLAKE_MOBILE: /* KBL */
3340 case INTEL_FAM6_KABYLAKE_DESKTOP: /* KBL */
3341 case INTEL_FAM6_CANNONLAKE_MOBILE: /* CNL */
3342 case INTEL_FAM6_SKYLAKE_X: /* SKX */
3344 case INTEL_FAM6_XEON_PHI_KNL: /* Knights Landing */
3345 case INTEL_FAM6_XEON_PHI_KNM:
3353 dump_cstate_pstate_config_info(unsigned int family, unsigned int model)
3355 if (!do_nhm_platform_info)
3358 dump_nhm_platform_info();
3360 if (has_hsw_turbo_ratio_limit(family, model))
3361 dump_hsw_turbo_ratio_limits();
3363 if (has_ivt_turbo_ratio_limit(family, model))
3364 dump_ivt_turbo_ratio_limits();
3366 if (has_turbo_ratio_limit(family, model))
3367 dump_turbo_ratio_limits(family, model);
3369 if (has_atom_turbo_ratio_limit(family, model))
3370 dump_atom_turbo_ratio_limits();
3372 if (has_knl_turbo_ratio_limit(family, model))
3373 dump_knl_turbo_ratio_limits();
3375 if (has_config_tdp(family, model))
3382 dump_sysfs_cstate_config(void)
3391 if (!DO_BIC(BIC_sysfs))
3394 for (state = 0; state < 10; ++state) {
3396 sprintf(path, "/sys/devices/system/cpu/cpu%d/cpuidle/state%d/name",
3398 input = fopen(path, "r");
3401 fgets(name_buf, sizeof(name_buf), input);
3403 /* truncate "C1-HSW\n" to "C1", or truncate "C1\n" to "C1" */
3404 sp = strchr(name_buf, '-');
3406 sp = strchrnul(name_buf, '\n');
3411 sprintf(path, "/sys/devices/system/cpu/cpu%d/cpuidle/state%d/desc",
3413 input = fopen(path, "r");
3416 fgets(desc, sizeof(desc), input);
3418 fprintf(outf, "cpu%d: %s: %s", base_cpu, name_buf, desc);
3423 dump_sysfs_pstate_config(void)
3426 char driver_buf[64];
3427 char governor_buf[64];
3431 sprintf(path, "/sys/devices/system/cpu/cpu%d/cpufreq/scaling_driver",
3433 input = fopen(path, "r");
3434 if (input == NULL) {
3435 fprintf(stderr, "NSFOD %s\n", path);
3438 fgets(driver_buf, sizeof(driver_buf), input);
3441 sprintf(path, "/sys/devices/system/cpu/cpu%d/cpufreq/scaling_governor",
3443 input = fopen(path, "r");
3444 if (input == NULL) {
3445 fprintf(stderr, "NSFOD %s\n", path);
3448 fgets(governor_buf, sizeof(governor_buf), input);
3451 fprintf(outf, "cpu%d: cpufreq driver: %s", base_cpu, driver_buf);
3452 fprintf(outf, "cpu%d: cpufreq governor: %s", base_cpu, governor_buf);
3454 sprintf(path, "/sys/devices/system/cpu/cpufreq/boost");
3455 input = fopen(path, "r");
3456 if (input != NULL) {
3457 fscanf(input, "%d", &turbo);
3458 fprintf(outf, "cpufreq boost: %d\n", turbo);
3462 sprintf(path, "/sys/devices/system/cpu/intel_pstate/no_turbo");
3463 input = fopen(path, "r");
3464 if (input != NULL) {
3465 fscanf(input, "%d", &turbo);
3466 fprintf(outf, "cpufreq intel_pstate no_turbo: %d\n", turbo);
3474 * Decode the ENERGY_PERF_BIAS MSR
3476 int print_epb(struct thread_data *t, struct core_data *c, struct pkg_data *p)
3478 unsigned long long msr;
3487 /* EPB is per-package */
3488 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
3491 if (cpu_migrate(cpu)) {
3492 fprintf(outf, "Could not migrate to CPU %d\n", cpu);
3496 if (get_msr(cpu, MSR_IA32_ENERGY_PERF_BIAS, &msr))
3499 switch (msr & 0xF) {
3500 case ENERGY_PERF_BIAS_PERFORMANCE:
3501 epb_string = "performance";
3503 case ENERGY_PERF_BIAS_NORMAL:
3504 epb_string = "balanced";
3506 case ENERGY_PERF_BIAS_POWERSAVE:
3507 epb_string = "powersave";
3510 epb_string = "custom";
3513 fprintf(outf, "cpu%d: MSR_IA32_ENERGY_PERF_BIAS: 0x%08llx (%s)\n", cpu, msr, epb_string);
3519 * Decode the MSR_HWP_CAPABILITIES
3521 int print_hwp(struct thread_data *t, struct core_data *c, struct pkg_data *p)
3523 unsigned long long msr;
3531 /* MSR_HWP_CAPABILITIES is per-package */
3532 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
3535 if (cpu_migrate(cpu)) {
3536 fprintf(outf, "Could not migrate to CPU %d\n", cpu);
3540 if (get_msr(cpu, MSR_PM_ENABLE, &msr))
3543 fprintf(outf, "cpu%d: MSR_PM_ENABLE: 0x%08llx (%sHWP)\n",
3544 cpu, msr, (msr & (1 << 0)) ? "" : "No-");
3546 /* MSR_PM_ENABLE[1] == 1 if HWP is enabled and MSRs visible */
3547 if ((msr & (1 << 0)) == 0)
3550 if (get_msr(cpu, MSR_HWP_CAPABILITIES, &msr))
3553 fprintf(outf, "cpu%d: MSR_HWP_CAPABILITIES: 0x%08llx "
3554 "(high %d guar %d eff %d low %d)\n",
3556 (unsigned int)HWP_HIGHEST_PERF(msr),
3557 (unsigned int)HWP_GUARANTEED_PERF(msr),
3558 (unsigned int)HWP_MOSTEFFICIENT_PERF(msr),
3559 (unsigned int)HWP_LOWEST_PERF(msr));
3561 if (get_msr(cpu, MSR_HWP_REQUEST, &msr))
3564 fprintf(outf, "cpu%d: MSR_HWP_REQUEST: 0x%08llx "
3565 "(min %d max %d des %d epp 0x%x window 0x%x pkg 0x%x)\n",
3567 (unsigned int)(((msr) >> 0) & 0xff),
3568 (unsigned int)(((msr) >> 8) & 0xff),
3569 (unsigned int)(((msr) >> 16) & 0xff),
3570 (unsigned int)(((msr) >> 24) & 0xff),
3571 (unsigned int)(((msr) >> 32) & 0xff3),
3572 (unsigned int)(((msr) >> 42) & 0x1));
3575 if (get_msr(cpu, MSR_HWP_REQUEST_PKG, &msr))
3578 fprintf(outf, "cpu%d: MSR_HWP_REQUEST_PKG: 0x%08llx "
3579 "(min %d max %d des %d epp 0x%x window 0x%x)\n",
3581 (unsigned int)(((msr) >> 0) & 0xff),
3582 (unsigned int)(((msr) >> 8) & 0xff),
3583 (unsigned int)(((msr) >> 16) & 0xff),
3584 (unsigned int)(((msr) >> 24) & 0xff),
3585 (unsigned int)(((msr) >> 32) & 0xff3));
3587 if (has_hwp_notify) {
3588 if (get_msr(cpu, MSR_HWP_INTERRUPT, &msr))
3591 fprintf(outf, "cpu%d: MSR_HWP_INTERRUPT: 0x%08llx "
3592 "(%s_Guaranteed_Perf_Change, %s_Excursion_Min)\n",
3594 ((msr) & 0x1) ? "EN" : "Dis",
3595 ((msr) & 0x2) ? "EN" : "Dis");
3597 if (get_msr(cpu, MSR_HWP_STATUS, &msr))
3600 fprintf(outf, "cpu%d: MSR_HWP_STATUS: 0x%08llx "
3601 "(%sGuaranteed_Perf_Change, %sExcursion_Min)\n",
3603 ((msr) & 0x1) ? "" : "No-",
3604 ((msr) & 0x2) ? "" : "No-");
3610 * print_perf_limit()
3612 int print_perf_limit(struct thread_data *t, struct core_data *c, struct pkg_data *p)
3614 unsigned long long msr;
3620 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
3623 if (cpu_migrate(cpu)) {
3624 fprintf(outf, "Could not migrate to CPU %d\n", cpu);
3628 if (do_core_perf_limit_reasons) {
3629 get_msr(cpu, MSR_CORE_PERF_LIMIT_REASONS, &msr);
3630 fprintf(outf, "cpu%d: MSR_CORE_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
3631 fprintf(outf, " (Active: %s%s%s%s%s%s%s%s%s%s%s%s%s%s)",
3632 (msr & 1 << 15) ? "bit15, " : "",
3633 (msr & 1 << 14) ? "bit14, " : "",
3634 (msr & 1 << 13) ? "Transitions, " : "",
3635 (msr & 1 << 12) ? "MultiCoreTurbo, " : "",
3636 (msr & 1 << 11) ? "PkgPwrL2, " : "",
3637 (msr & 1 << 10) ? "PkgPwrL1, " : "",
3638 (msr & 1 << 9) ? "CorePwr, " : "",
3639 (msr & 1 << 8) ? "Amps, " : "",
3640 (msr & 1 << 6) ? "VR-Therm, " : "",
3641 (msr & 1 << 5) ? "Auto-HWP, " : "",
3642 (msr & 1 << 4) ? "Graphics, " : "",
3643 (msr & 1 << 2) ? "bit2, " : "",
3644 (msr & 1 << 1) ? "ThermStatus, " : "",
3645 (msr & 1 << 0) ? "PROCHOT, " : "");
3646 fprintf(outf, " (Logged: %s%s%s%s%s%s%s%s%s%s%s%s%s%s)\n",
3647 (msr & 1 << 31) ? "bit31, " : "",
3648 (msr & 1 << 30) ? "bit30, " : "",
3649 (msr & 1 << 29) ? "Transitions, " : "",
3650 (msr & 1 << 28) ? "MultiCoreTurbo, " : "",
3651 (msr & 1 << 27) ? "PkgPwrL2, " : "",
3652 (msr & 1 << 26) ? "PkgPwrL1, " : "",
3653 (msr & 1 << 25) ? "CorePwr, " : "",
3654 (msr & 1 << 24) ? "Amps, " : "",
3655 (msr & 1 << 22) ? "VR-Therm, " : "",
3656 (msr & 1 << 21) ? "Auto-HWP, " : "",
3657 (msr & 1 << 20) ? "Graphics, " : "",
3658 (msr & 1 << 18) ? "bit18, " : "",
3659 (msr & 1 << 17) ? "ThermStatus, " : "",
3660 (msr & 1 << 16) ? "PROCHOT, " : "");
3663 if (do_gfx_perf_limit_reasons) {
3664 get_msr(cpu, MSR_GFX_PERF_LIMIT_REASONS, &msr);
3665 fprintf(outf, "cpu%d: MSR_GFX_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
3666 fprintf(outf, " (Active: %s%s%s%s%s%s%s%s)",
3667 (msr & 1 << 0) ? "PROCHOT, " : "",
3668 (msr & 1 << 1) ? "ThermStatus, " : "",
3669 (msr & 1 << 4) ? "Graphics, " : "",
3670 (msr & 1 << 6) ? "VR-Therm, " : "",
3671 (msr & 1 << 8) ? "Amps, " : "",
3672 (msr & 1 << 9) ? "GFXPwr, " : "",
3673 (msr & 1 << 10) ? "PkgPwrL1, " : "",
3674 (msr & 1 << 11) ? "PkgPwrL2, " : "");
3675 fprintf(outf, " (Logged: %s%s%s%s%s%s%s%s)\n",
3676 (msr & 1 << 16) ? "PROCHOT, " : "",
3677 (msr & 1 << 17) ? "ThermStatus, " : "",
3678 (msr & 1 << 20) ? "Graphics, " : "",
3679 (msr & 1 << 22) ? "VR-Therm, " : "",
3680 (msr & 1 << 24) ? "Amps, " : "",
3681 (msr & 1 << 25) ? "GFXPwr, " : "",
3682 (msr & 1 << 26) ? "PkgPwrL1, " : "",
3683 (msr & 1 << 27) ? "PkgPwrL2, " : "");
3685 if (do_ring_perf_limit_reasons) {
3686 get_msr(cpu, MSR_RING_PERF_LIMIT_REASONS, &msr);
3687 fprintf(outf, "cpu%d: MSR_RING_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
3688 fprintf(outf, " (Active: %s%s%s%s%s%s)",
3689 (msr & 1 << 0) ? "PROCHOT, " : "",
3690 (msr & 1 << 1) ? "ThermStatus, " : "",
3691 (msr & 1 << 6) ? "VR-Therm, " : "",
3692 (msr & 1 << 8) ? "Amps, " : "",
3693 (msr & 1 << 10) ? "PkgPwrL1, " : "",
3694 (msr & 1 << 11) ? "PkgPwrL2, " : "");
3695 fprintf(outf, " (Logged: %s%s%s%s%s%s)\n",
3696 (msr & 1 << 16) ? "PROCHOT, " : "",
3697 (msr & 1 << 17) ? "ThermStatus, " : "",
3698 (msr & 1 << 22) ? "VR-Therm, " : "",
3699 (msr & 1 << 24) ? "Amps, " : "",
3700 (msr & 1 << 26) ? "PkgPwrL1, " : "",
3701 (msr & 1 << 27) ? "PkgPwrL2, " : "");
3706 #define RAPL_POWER_GRANULARITY 0x7FFF /* 15 bit power granularity */
3707 #define RAPL_TIME_GRANULARITY 0x3F /* 6 bit time granularity */
3709 double get_tdp(unsigned int model)
3711 unsigned long long msr;
3713 if (do_rapl & RAPL_PKG_POWER_INFO)
3714 if (!get_msr(base_cpu, MSR_PKG_POWER_INFO, &msr))
3715 return ((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units;
3718 case INTEL_FAM6_ATOM_SILVERMONT1:
3719 case INTEL_FAM6_ATOM_SILVERMONT2:
3727 * rapl_dram_energy_units_probe()
3728 * Energy units are either hard-coded, or come from RAPL Energy Unit MSR.
3731 rapl_dram_energy_units_probe(int model, double rapl_energy_units)
3733 /* only called for genuine_intel, family 6 */
3736 case INTEL_FAM6_HASWELL_X: /* HSX */
3737 case INTEL_FAM6_BROADWELL_X: /* BDX */
3738 case INTEL_FAM6_BROADWELL_XEON_D: /* BDX-DE */
3739 case INTEL_FAM6_XEON_PHI_KNL: /* KNL */
3740 case INTEL_FAM6_XEON_PHI_KNM:
3741 return (rapl_dram_energy_units = 15.3 / 1000000);
3743 return (rapl_energy_units);
3751 * sets do_rapl, rapl_power_units, rapl_energy_units, rapl_time_units
3753 void rapl_probe(unsigned int family, unsigned int model)
3755 unsigned long long msr;
3756 unsigned int time_unit;
3766 case INTEL_FAM6_SANDYBRIDGE:
3767 case INTEL_FAM6_IVYBRIDGE:
3768 case INTEL_FAM6_HASWELL_CORE: /* HSW */
3769 case INTEL_FAM6_HASWELL_ULT: /* HSW */
3770 case INTEL_FAM6_HASWELL_GT3E: /* HSW */
3771 case INTEL_FAM6_BROADWELL_CORE: /* BDW */
3772 case INTEL_FAM6_BROADWELL_GT3E: /* BDW */
3773 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_GFX | RAPL_PKG_POWER_INFO;
3775 BIC_PRESENT(BIC_Pkg_J);
3776 BIC_PRESENT(BIC_Cor_J);
3777 BIC_PRESENT(BIC_GFX_J);
3779 BIC_PRESENT(BIC_PkgWatt);
3780 BIC_PRESENT(BIC_CorWatt);
3781 BIC_PRESENT(BIC_GFXWatt);
3784 case INTEL_FAM6_ATOM_GOLDMONT: /* BXT */
3785 case INTEL_FAM6_ATOM_GEMINI_LAKE:
3786 do_rapl = RAPL_PKG | RAPL_PKG_POWER_INFO;
3788 BIC_PRESENT(BIC_Pkg_J);
3790 BIC_PRESENT(BIC_PkgWatt);
3792 case INTEL_FAM6_SKYLAKE_MOBILE: /* SKL */
3793 case INTEL_FAM6_SKYLAKE_DESKTOP: /* SKL */
3794 case INTEL_FAM6_KABYLAKE_MOBILE: /* KBL */
3795 case INTEL_FAM6_KABYLAKE_DESKTOP: /* KBL */
3796 case INTEL_FAM6_CANNONLAKE_MOBILE: /* CNL */
3797 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_DRAM | RAPL_DRAM_PERF_STATUS | RAPL_PKG_PERF_STATUS | RAPL_GFX | RAPL_PKG_POWER_INFO;
3798 BIC_PRESENT(BIC_PKG__);
3799 BIC_PRESENT(BIC_RAM__);
3801 BIC_PRESENT(BIC_Pkg_J);
3802 BIC_PRESENT(BIC_Cor_J);
3803 BIC_PRESENT(BIC_RAM_J);
3804 BIC_PRESENT(BIC_GFX_J);
3806 BIC_PRESENT(BIC_PkgWatt);
3807 BIC_PRESENT(BIC_CorWatt);
3808 BIC_PRESENT(BIC_RAMWatt);
3809 BIC_PRESENT(BIC_GFXWatt);
3812 case INTEL_FAM6_HASWELL_X: /* HSX */
3813 case INTEL_FAM6_BROADWELL_X: /* BDX */
3814 case INTEL_FAM6_BROADWELL_XEON_D: /* BDX-DE */
3815 case INTEL_FAM6_SKYLAKE_X: /* SKX */
3816 case INTEL_FAM6_XEON_PHI_KNL: /* KNL */
3817 case INTEL_FAM6_XEON_PHI_KNM:
3818 do_rapl = RAPL_PKG | RAPL_DRAM | RAPL_DRAM_POWER_INFO | RAPL_DRAM_PERF_STATUS | RAPL_PKG_PERF_STATUS | RAPL_PKG_POWER_INFO;
3819 BIC_PRESENT(BIC_PKG__);
3820 BIC_PRESENT(BIC_RAM__);
3822 BIC_PRESENT(BIC_Pkg_J);
3823 BIC_PRESENT(BIC_RAM_J);
3825 BIC_PRESENT(BIC_PkgWatt);
3826 BIC_PRESENT(BIC_RAMWatt);
3829 case INTEL_FAM6_SANDYBRIDGE_X:
3830 case INTEL_FAM6_IVYBRIDGE_X:
3831 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_DRAM | RAPL_DRAM_POWER_INFO | RAPL_PKG_PERF_STATUS | RAPL_DRAM_PERF_STATUS | RAPL_PKG_POWER_INFO;
3832 BIC_PRESENT(BIC_PKG__);
3833 BIC_PRESENT(BIC_RAM__);
3835 BIC_PRESENT(BIC_Pkg_J);
3836 BIC_PRESENT(BIC_Cor_J);
3837 BIC_PRESENT(BIC_RAM_J);
3839 BIC_PRESENT(BIC_PkgWatt);
3840 BIC_PRESENT(BIC_CorWatt);
3841 BIC_PRESENT(BIC_RAMWatt);
3844 case INTEL_FAM6_ATOM_SILVERMONT1: /* BYT */
3845 case INTEL_FAM6_ATOM_SILVERMONT2: /* AVN */
3846 do_rapl = RAPL_PKG | RAPL_CORES;
3848 BIC_PRESENT(BIC_Pkg_J);
3849 BIC_PRESENT(BIC_Cor_J);
3851 BIC_PRESENT(BIC_PkgWatt);
3852 BIC_PRESENT(BIC_CorWatt);
3855 case INTEL_FAM6_ATOM_DENVERTON: /* DNV */
3856 do_rapl = RAPL_PKG | RAPL_DRAM | RAPL_DRAM_POWER_INFO | RAPL_DRAM_PERF_STATUS | RAPL_PKG_PERF_STATUS | RAPL_PKG_POWER_INFO | RAPL_CORES_ENERGY_STATUS;
3857 BIC_PRESENT(BIC_PKG__);
3858 BIC_PRESENT(BIC_RAM__);
3860 BIC_PRESENT(BIC_Pkg_J);
3861 BIC_PRESENT(BIC_Cor_J);
3862 BIC_PRESENT(BIC_RAM_J);
3864 BIC_PRESENT(BIC_PkgWatt);
3865 BIC_PRESENT(BIC_CorWatt);
3866 BIC_PRESENT(BIC_RAMWatt);
3873 /* units on package 0, verify later other packages match */
3874 if (get_msr(base_cpu, MSR_RAPL_POWER_UNIT, &msr))
3877 rapl_power_units = 1.0 / (1 << (msr & 0xF));
3878 if (model == INTEL_FAM6_ATOM_SILVERMONT1)
3879 rapl_energy_units = 1.0 * (1 << (msr >> 8 & 0x1F)) / 1000000;
3881 rapl_energy_units = 1.0 / (1 << (msr >> 8 & 0x1F));
3883 rapl_dram_energy_units = rapl_dram_energy_units_probe(model, rapl_energy_units);
3885 time_unit = msr >> 16 & 0xF;
3889 rapl_time_units = 1.0 / (1 << (time_unit));
3891 tdp = get_tdp(model);
3893 rapl_joule_counter_range = 0xFFFFFFFF * rapl_energy_units / tdp;
3895 fprintf(outf, "RAPL: %.0f sec. Joule Counter Range, at %.0f Watts\n", rapl_joule_counter_range, tdp);
3900 void perf_limit_reasons_probe(unsigned int family, unsigned int model)
3909 case INTEL_FAM6_HASWELL_CORE: /* HSW */
3910 case INTEL_FAM6_HASWELL_ULT: /* HSW */
3911 case INTEL_FAM6_HASWELL_GT3E: /* HSW */
3912 do_gfx_perf_limit_reasons = 1;
3913 case INTEL_FAM6_HASWELL_X: /* HSX */
3914 do_core_perf_limit_reasons = 1;
3915 do_ring_perf_limit_reasons = 1;
3921 void automatic_cstate_conversion_probe(unsigned int family, unsigned int model)
3923 if (is_skx(family, model) || is_bdx(family, model))
3924 has_automatic_cstate_conversion = 1;
3927 int print_thermal(struct thread_data *t, struct core_data *c, struct pkg_data *p)
3929 unsigned long long msr;
3930 unsigned int dts, dts2;
3933 if (!(do_dts || do_ptm))
3938 /* DTS is per-core, no need to print for each thread */
3939 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
3942 if (cpu_migrate(cpu)) {
3943 fprintf(outf, "Could not migrate to CPU %d\n", cpu);
3947 if (do_ptm && (t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) {
3948 if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_STATUS, &msr))
3951 dts = (msr >> 16) & 0x7F;
3952 fprintf(outf, "cpu%d: MSR_IA32_PACKAGE_THERM_STATUS: 0x%08llx (%d C)\n",
3953 cpu, msr, tcc_activation_temp - dts);
3955 if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_INTERRUPT, &msr))
3958 dts = (msr >> 16) & 0x7F;
3959 dts2 = (msr >> 8) & 0x7F;
3960 fprintf(outf, "cpu%d: MSR_IA32_PACKAGE_THERM_INTERRUPT: 0x%08llx (%d C, %d C)\n",
3961 cpu, msr, tcc_activation_temp - dts, tcc_activation_temp - dts2);
3965 if (do_dts && debug) {
3966 unsigned int resolution;
3968 if (get_msr(cpu, MSR_IA32_THERM_STATUS, &msr))
3971 dts = (msr >> 16) & 0x7F;
3972 resolution = (msr >> 27) & 0xF;
3973 fprintf(outf, "cpu%d: MSR_IA32_THERM_STATUS: 0x%08llx (%d C +/- %d)\n",
3974 cpu, msr, tcc_activation_temp - dts, resolution);
3976 if (get_msr(cpu, MSR_IA32_THERM_INTERRUPT, &msr))
3979 dts = (msr >> 16) & 0x7F;
3980 dts2 = (msr >> 8) & 0x7F;
3981 fprintf(outf, "cpu%d: MSR_IA32_THERM_INTERRUPT: 0x%08llx (%d C, %d C)\n",
3982 cpu, msr, tcc_activation_temp - dts, tcc_activation_temp - dts2);
3988 void print_power_limit_msr(int cpu, unsigned long long msr, char *label)
3990 fprintf(outf, "cpu%d: %s: %sabled (%f Watts, %f sec, clamp %sabled)\n",
3992 ((msr >> 15) & 1) ? "EN" : "DIS",
3993 ((msr >> 0) & 0x7FFF) * rapl_power_units,
3994 (1.0 + (((msr >> 22) & 0x3)/4.0)) * (1 << ((msr >> 17) & 0x1F)) * rapl_time_units,
3995 (((msr >> 16) & 1) ? "EN" : "DIS"));
4000 int print_rapl(struct thread_data *t, struct core_data *c, struct pkg_data *p)
4002 unsigned long long msr;
4008 /* RAPL counters are per package, so print only for 1st thread/package */
4009 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
4013 if (cpu_migrate(cpu)) {
4014 fprintf(outf, "Could not migrate to CPU %d\n", cpu);
4018 if (get_msr(cpu, MSR_RAPL_POWER_UNIT, &msr))
4021 fprintf(outf, "cpu%d: MSR_RAPL_POWER_UNIT: 0x%08llx (%f Watts, %f Joules, %f sec.)\n", cpu, msr,
4022 rapl_power_units, rapl_energy_units, rapl_time_units);
4024 if (do_rapl & RAPL_PKG_POWER_INFO) {
4026 if (get_msr(cpu, MSR_PKG_POWER_INFO, &msr))
4030 fprintf(outf, "cpu%d: MSR_PKG_POWER_INFO: 0x%08llx (%.0f W TDP, RAPL %.0f - %.0f W, %f sec.)\n",
4032 ((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units,
4033 ((msr >> 16) & RAPL_POWER_GRANULARITY) * rapl_power_units,
4034 ((msr >> 32) & RAPL_POWER_GRANULARITY) * rapl_power_units,
4035 ((msr >> 48) & RAPL_TIME_GRANULARITY) * rapl_time_units);
4038 if (do_rapl & RAPL_PKG) {
4040 if (get_msr(cpu, MSR_PKG_POWER_LIMIT, &msr))
4043 fprintf(outf, "cpu%d: MSR_PKG_POWER_LIMIT: 0x%08llx (%slocked)\n",
4044 cpu, msr, (msr >> 63) & 1 ? "" : "UN");
4046 print_power_limit_msr(cpu, msr, "PKG Limit #1");
4047 fprintf(outf, "cpu%d: PKG Limit #2: %sabled (%f Watts, %f* sec, clamp %sabled)\n",
4049 ((msr >> 47) & 1) ? "EN" : "DIS",
4050 ((msr >> 32) & 0x7FFF) * rapl_power_units,
4051 (1.0 + (((msr >> 54) & 0x3)/4.0)) * (1 << ((msr >> 49) & 0x1F)) * rapl_time_units,
4052 ((msr >> 48) & 1) ? "EN" : "DIS");
4055 if (do_rapl & RAPL_DRAM_POWER_INFO) {
4056 if (get_msr(cpu, MSR_DRAM_POWER_INFO, &msr))
4059 fprintf(outf, "cpu%d: MSR_DRAM_POWER_INFO,: 0x%08llx (%.0f W TDP, RAPL %.0f - %.0f W, %f sec.)\n",
4061 ((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units,
4062 ((msr >> 16) & RAPL_POWER_GRANULARITY) * rapl_power_units,
4063 ((msr >> 32) & RAPL_POWER_GRANULARITY) * rapl_power_units,
4064 ((msr >> 48) & RAPL_TIME_GRANULARITY) * rapl_time_units);
4066 if (do_rapl & RAPL_DRAM) {
4067 if (get_msr(cpu, MSR_DRAM_POWER_LIMIT, &msr))
4069 fprintf(outf, "cpu%d: MSR_DRAM_POWER_LIMIT: 0x%08llx (%slocked)\n",
4070 cpu, msr, (msr >> 31) & 1 ? "" : "UN");
4072 print_power_limit_msr(cpu, msr, "DRAM Limit");
4074 if (do_rapl & RAPL_CORE_POLICY) {
4075 if (get_msr(cpu, MSR_PP0_POLICY, &msr))
4078 fprintf(outf, "cpu%d: MSR_PP0_POLICY: %lld\n", cpu, msr & 0xF);
4080 if (do_rapl & RAPL_CORES_POWER_LIMIT) {
4081 if (get_msr(cpu, MSR_PP0_POWER_LIMIT, &msr))
4083 fprintf(outf, "cpu%d: MSR_PP0_POWER_LIMIT: 0x%08llx (%slocked)\n",
4084 cpu, msr, (msr >> 31) & 1 ? "" : "UN");
4085 print_power_limit_msr(cpu, msr, "Cores Limit");
4087 if (do_rapl & RAPL_GFX) {
4088 if (get_msr(cpu, MSR_PP1_POLICY, &msr))
4091 fprintf(outf, "cpu%d: MSR_PP1_POLICY: %lld\n", cpu, msr & 0xF);
4093 if (get_msr(cpu, MSR_PP1_POWER_LIMIT, &msr))
4095 fprintf(outf, "cpu%d: MSR_PP1_POWER_LIMIT: 0x%08llx (%slocked)\n",
4096 cpu, msr, (msr >> 31) & 1 ? "" : "UN");
4097 print_power_limit_msr(cpu, msr, "GFX Limit");
4103 * SNB adds support for additional MSRs:
4105 * MSR_PKG_C7_RESIDENCY 0x000003fa
4106 * MSR_CORE_C7_RESIDENCY 0x000003fe
4107 * MSR_PKG_C2_RESIDENCY 0x0000060d
4110 int has_snb_msrs(unsigned int family, unsigned int model)
4116 case INTEL_FAM6_SANDYBRIDGE:
4117 case INTEL_FAM6_SANDYBRIDGE_X:
4118 case INTEL_FAM6_IVYBRIDGE: /* IVB */
4119 case INTEL_FAM6_IVYBRIDGE_X: /* IVB Xeon */
4120 case INTEL_FAM6_HASWELL_CORE: /* HSW */
4121 case INTEL_FAM6_HASWELL_X: /* HSW */
4122 case INTEL_FAM6_HASWELL_ULT: /* HSW */
4123 case INTEL_FAM6_HASWELL_GT3E: /* HSW */
4124 case INTEL_FAM6_BROADWELL_CORE: /* BDW */
4125 case INTEL_FAM6_BROADWELL_GT3E: /* BDW */
4126 case INTEL_FAM6_BROADWELL_X: /* BDX */
4127 case INTEL_FAM6_BROADWELL_XEON_D: /* BDX-DE */
4128 case INTEL_FAM6_SKYLAKE_MOBILE: /* SKL */
4129 case INTEL_FAM6_SKYLAKE_DESKTOP: /* SKL */
4130 case INTEL_FAM6_KABYLAKE_MOBILE: /* KBL */
4131 case INTEL_FAM6_KABYLAKE_DESKTOP: /* KBL */
4132 case INTEL_FAM6_CANNONLAKE_MOBILE: /* CNL */
4133 case INTEL_FAM6_SKYLAKE_X: /* SKX */
4134 case INTEL_FAM6_ATOM_GOLDMONT: /* BXT */
4135 case INTEL_FAM6_ATOM_GEMINI_LAKE:
4136 case INTEL_FAM6_ATOM_DENVERTON: /* DNV */
4143 * HSW adds support for additional MSRs:
4145 * MSR_PKG_C8_RESIDENCY 0x00000630
4146 * MSR_PKG_C9_RESIDENCY 0x00000631
4147 * MSR_PKG_C10_RESIDENCY 0x00000632
4149 * MSR_PKGC8_IRTL 0x00000633
4150 * MSR_PKGC9_IRTL 0x00000634
4151 * MSR_PKGC10_IRTL 0x00000635
4154 int has_hsw_msrs(unsigned int family, unsigned int model)
4160 case INTEL_FAM6_HASWELL_ULT: /* HSW */
4161 case INTEL_FAM6_BROADWELL_CORE: /* BDW */
4162 case INTEL_FAM6_SKYLAKE_MOBILE: /* SKL */
4163 case INTEL_FAM6_SKYLAKE_DESKTOP: /* SKL */
4164 case INTEL_FAM6_KABYLAKE_MOBILE: /* KBL */
4165 case INTEL_FAM6_KABYLAKE_DESKTOP: /* KBL */
4166 case INTEL_FAM6_CANNONLAKE_MOBILE: /* CNL */
4167 case INTEL_FAM6_ATOM_GOLDMONT: /* BXT */
4168 case INTEL_FAM6_ATOM_GEMINI_LAKE:
4175 * SKL adds support for additional MSRS:
4177 * MSR_PKG_WEIGHTED_CORE_C0_RES 0x00000658
4178 * MSR_PKG_ANY_CORE_C0_RES 0x00000659
4179 * MSR_PKG_ANY_GFXE_C0_RES 0x0000065A
4180 * MSR_PKG_BOTH_CORE_GFXE_C0_RES 0x0000065B
4182 int has_skl_msrs(unsigned int family, unsigned int model)
4188 case INTEL_FAM6_SKYLAKE_MOBILE: /* SKL */
4189 case INTEL_FAM6_SKYLAKE_DESKTOP: /* SKL */
4190 case INTEL_FAM6_KABYLAKE_MOBILE: /* KBL */
4191 case INTEL_FAM6_KABYLAKE_DESKTOP: /* KBL */
4192 case INTEL_FAM6_CANNONLAKE_MOBILE: /* CNL */
4198 int is_slm(unsigned int family, unsigned int model)
4203 case INTEL_FAM6_ATOM_SILVERMONT1: /* BYT */
4204 case INTEL_FAM6_ATOM_SILVERMONT2: /* AVN */
4210 int is_knl(unsigned int family, unsigned int model)
4215 case INTEL_FAM6_XEON_PHI_KNL: /* KNL */
4216 case INTEL_FAM6_XEON_PHI_KNM:
4222 int is_cnl(unsigned int family, unsigned int model)
4228 case INTEL_FAM6_CANNONLAKE_MOBILE: /* CNL */
4235 unsigned int get_aperf_mperf_multiplier(unsigned int family, unsigned int model)
4237 if (is_knl(family, model))
4242 #define SLM_BCLK_FREQS 5
4243 double slm_freq_table[SLM_BCLK_FREQS] = { 83.3, 100.0, 133.3, 116.7, 80.0};
4245 double slm_bclk(void)
4247 unsigned long long msr = 3;
4251 if (get_msr(base_cpu, MSR_FSB_FREQ, &msr))
4252 fprintf(outf, "SLM BCLK: unknown\n");
4255 if (i >= SLM_BCLK_FREQS) {
4256 fprintf(outf, "SLM BCLK[%d] invalid\n", i);
4259 freq = slm_freq_table[i];
4262 fprintf(outf, "SLM BCLK: %.1f Mhz\n", freq);
4267 double discover_bclk(unsigned int family, unsigned int model)
4269 if (has_snb_msrs(family, model) || is_knl(family, model))
4271 else if (is_slm(family, model))
4278 * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where
4279 * the Thermal Control Circuit (TCC) activates.
4280 * This is usually equal to tjMax.
4282 * Older processors do not have this MSR, so there we guess,
4283 * but also allow cmdline over-ride with -T.
4285 * Several MSR temperature values are in units of degrees-C
4286 * below this value, including the Digital Thermal Sensor (DTS),
4287 * Package Thermal Management Sensor (PTM), and thermal event thresholds.
4289 int set_temperature_target(struct thread_data *t, struct core_data *c, struct pkg_data *p)
4291 unsigned long long msr;
4292 unsigned int target_c_local;
4295 /* tcc_activation_temp is used only for dts or ptm */
4296 if (!(do_dts || do_ptm))
4299 /* this is a per-package concept */
4300 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
4304 if (cpu_migrate(cpu)) {
4305 fprintf(outf, "Could not migrate to CPU %d\n", cpu);
4309 if (tcc_activation_temp_override != 0) {
4310 tcc_activation_temp = tcc_activation_temp_override;
4311 fprintf(outf, "cpu%d: Using cmdline TCC Target (%d C)\n",
4312 cpu, tcc_activation_temp);
4316 /* Temperature Target MSR is Nehalem and newer only */
4317 if (!do_nhm_platform_info)
4320 if (get_msr(base_cpu, MSR_IA32_TEMPERATURE_TARGET, &msr))
4323 target_c_local = (msr >> 16) & 0xFF;
4326 fprintf(outf, "cpu%d: MSR_IA32_TEMPERATURE_TARGET: 0x%08llx (%d C)\n",
4327 cpu, msr, target_c_local);
4329 if (!target_c_local)
4332 tcc_activation_temp = target_c_local;
4337 tcc_activation_temp = TJMAX_DEFAULT;
4338 fprintf(outf, "cpu%d: Guessing tjMax %d C, Please use -T to specify\n",
4339 cpu, tcc_activation_temp);
4344 void decode_feature_control_msr(void)
4346 unsigned long long msr;
4348 if (!get_msr(base_cpu, MSR_IA32_FEATURE_CONTROL, &msr))
4349 fprintf(outf, "cpu%d: MSR_IA32_FEATURE_CONTROL: 0x%08llx (%sLocked %s)\n",
4351 msr & FEATURE_CONTROL_LOCKED ? "" : "UN-",
4352 msr & (1 << 18) ? "SGX" : "");
4355 void decode_misc_enable_msr(void)
4357 unsigned long long msr;
4362 if (!get_msr(base_cpu, MSR_IA32_MISC_ENABLE, &msr))
4363 fprintf(outf, "cpu%d: MSR_IA32_MISC_ENABLE: 0x%08llx (%sTCC %sEIST %sMWAIT %sPREFETCH %sTURBO)\n",
4365 msr & MSR_IA32_MISC_ENABLE_TM1 ? "" : "No-",
4366 msr & MSR_IA32_MISC_ENABLE_ENHANCED_SPEEDSTEP ? "" : "No-",
4367 msr & MSR_IA32_MISC_ENABLE_MWAIT ? "" : "No-",
4368 msr & MSR_IA32_MISC_ENABLE_PREFETCH_DISABLE ? "No-" : "",
4369 msr & MSR_IA32_MISC_ENABLE_TURBO_DISABLE ? "No-" : "");
4372 void decode_misc_feature_control(void)
4374 unsigned long long msr;
4376 if (!has_misc_feature_control)
4379 if (!get_msr(base_cpu, MSR_MISC_FEATURE_CONTROL, &msr))
4380 fprintf(outf, "cpu%d: MSR_MISC_FEATURE_CONTROL: 0x%08llx (%sL2-Prefetch %sL2-Prefetch-pair %sL1-Prefetch %sL1-IP-Prefetch)\n",
4382 msr & (0 << 0) ? "No-" : "",
4383 msr & (1 << 0) ? "No-" : "",
4384 msr & (2 << 0) ? "No-" : "",
4385 msr & (3 << 0) ? "No-" : "");
4388 * Decode MSR_MISC_PWR_MGMT
4390 * Decode the bits according to the Nehalem documentation
4391 * bit[0] seems to continue to have same meaning going forward
4394 void decode_misc_pwr_mgmt_msr(void)
4396 unsigned long long msr;
4398 if (!do_nhm_platform_info)
4401 if (no_MSR_MISC_PWR_MGMT)
4404 if (!get_msr(base_cpu, MSR_MISC_PWR_MGMT, &msr))
4405 fprintf(outf, "cpu%d: MSR_MISC_PWR_MGMT: 0x%08llx (%sable-EIST_Coordination %sable-EPB %sable-OOB)\n",
4407 msr & (1 << 0) ? "DIS" : "EN",
4408 msr & (1 << 1) ? "EN" : "DIS",
4409 msr & (1 << 8) ? "EN" : "DIS");
4412 * Decode MSR_CC6_DEMOTION_POLICY_CONFIG, MSR_MC6_DEMOTION_POLICY_CONFIG
4414 * This MSRs are present on Silvermont processors,
4415 * Intel Atom processor E3000 series (Baytrail), and friends.
4417 void decode_c6_demotion_policy_msr(void)
4419 unsigned long long msr;
4421 if (!get_msr(base_cpu, MSR_CC6_DEMOTION_POLICY_CONFIG, &msr))
4422 fprintf(outf, "cpu%d: MSR_CC6_DEMOTION_POLICY_CONFIG: 0x%08llx (%sable-CC6-Demotion)\n",
4423 base_cpu, msr, msr & (1 << 0) ? "EN" : "DIS");
4425 if (!get_msr(base_cpu, MSR_MC6_DEMOTION_POLICY_CONFIG, &msr))
4426 fprintf(outf, "cpu%d: MSR_MC6_DEMOTION_POLICY_CONFIG: 0x%08llx (%sable-MC6-Demotion)\n",
4427 base_cpu, msr, msr & (1 << 0) ? "EN" : "DIS");
4430 void process_cpuid()
4432 unsigned int eax, ebx, ecx, edx, max_level, max_extended_level;
4433 unsigned int fms, family, model, stepping;
4434 unsigned int has_turbo;
4436 eax = ebx = ecx = edx = 0;
4438 __cpuid(0, max_level, ebx, ecx, edx);
4440 if (ebx == 0x756e6547 && edx == 0x49656e69 && ecx == 0x6c65746e)
4444 fprintf(outf, "CPUID(0): %.4s%.4s%.4s ",
4445 (char *)&ebx, (char *)&edx, (char *)&ecx);
4447 __cpuid(1, fms, ebx, ecx, edx);
4448 family = (fms >> 8) & 0xf;
4449 model = (fms >> 4) & 0xf;
4450 stepping = fms & 0xf;
4451 if (family == 6 || family == 0xf)
4452 model += ((fms >> 16) & 0xf) << 4;
4455 fprintf(outf, "%d CPUID levels; family:model:stepping 0x%x:%x:%x (%d:%d:%d)\n",
4456 max_level, family, model, stepping, family, model, stepping);
4457 fprintf(outf, "CPUID(1): %s %s %s %s %s %s %s %s %s %s\n",
4458 ecx & (1 << 0) ? "SSE3" : "-",
4459 ecx & (1 << 3) ? "MONITOR" : "-",
4460 ecx & (1 << 6) ? "SMX" : "-",
4461 ecx & (1 << 7) ? "EIST" : "-",
4462 ecx & (1 << 8) ? "TM2" : "-",
4463 edx & (1 << 4) ? "TSC" : "-",
4464 edx & (1 << 5) ? "MSR" : "-",
4465 edx & (1 << 22) ? "ACPI-TM" : "-",
4466 edx & (1 << 28) ? "HT" : "-",
4467 edx & (1 << 29) ? "TM" : "-");
4470 if (!(edx & (1 << 5)))
4471 errx(1, "CPUID: no MSR");
4474 * check max extended function levels of CPUID.
4475 * This is needed to check for invariant TSC.
4476 * This check is valid for both Intel and AMD.
4478 ebx = ecx = edx = 0;
4479 __cpuid(0x80000000, max_extended_level, ebx, ecx, edx);
4481 if (max_extended_level >= 0x80000007) {
4484 * Non-Stop TSC is advertised by CPUID.EAX=0x80000007: EDX.bit8
4485 * this check is valid for both Intel and AMD
4487 __cpuid(0x80000007, eax, ebx, ecx, edx);
4488 has_invariant_tsc = edx & (1 << 8);
4492 * APERF/MPERF is advertised by CPUID.EAX=0x6: ECX.bit0
4493 * this check is valid for both Intel and AMD
4496 __cpuid(0x6, eax, ebx, ecx, edx);
4497 has_aperf = ecx & (1 << 0);
4499 BIC_PRESENT(BIC_Avg_MHz);
4500 BIC_PRESENT(BIC_Busy);
4501 BIC_PRESENT(BIC_Bzy_MHz);
4503 do_dts = eax & (1 << 0);
4505 BIC_PRESENT(BIC_CoreTmp);
4506 has_turbo = eax & (1 << 1);
4507 do_ptm = eax & (1 << 6);
4509 BIC_PRESENT(BIC_PkgTmp);
4510 has_hwp = eax & (1 << 7);
4511 has_hwp_notify = eax & (1 << 8);
4512 has_hwp_activity_window = eax & (1 << 9);
4513 has_hwp_epp = eax & (1 << 10);
4514 has_hwp_pkg = eax & (1 << 11);
4515 has_epb = ecx & (1 << 3);
4518 fprintf(outf, "CPUID(6): %sAPERF, %sTURBO, %sDTS, %sPTM, %sHWP, "
4519 "%sHWPnotify, %sHWPwindow, %sHWPepp, %sHWPpkg, %sEPB\n",
4520 has_aperf ? "" : "No-",
4521 has_turbo ? "" : "No-",
4522 do_dts ? "" : "No-",
4523 do_ptm ? "" : "No-",
4524 has_hwp ? "" : "No-",
4525 has_hwp_notify ? "" : "No-",
4526 has_hwp_activity_window ? "" : "No-",
4527 has_hwp_epp ? "" : "No-",
4528 has_hwp_pkg ? "" : "No-",
4529 has_epb ? "" : "No-");
4532 decode_misc_enable_msr();
4535 if (max_level >= 0x7 && !quiet) {
4540 __cpuid_count(0x7, 0, eax, ebx, ecx, edx);
4542 has_sgx = ebx & (1 << 2);
4543 fprintf(outf, "CPUID(7): %sSGX\n", has_sgx ? "" : "No-");
4546 decode_feature_control_msr();
4549 if (max_level >= 0x15) {
4550 unsigned int eax_crystal;
4551 unsigned int ebx_tsc;
4554 * CPUID 15H TSC/Crystal ratio, possibly Crystal Hz
4556 eax_crystal = ebx_tsc = crystal_hz = edx = 0;
4557 __cpuid(0x15, eax_crystal, ebx_tsc, crystal_hz, edx);
4561 if (!quiet && (ebx != 0))
4562 fprintf(outf, "CPUID(0x15): eax_crystal: %d ebx_tsc: %d ecx_crystal_hz: %d\n",
4563 eax_crystal, ebx_tsc, crystal_hz);
4565 if (crystal_hz == 0)
4567 case INTEL_FAM6_SKYLAKE_MOBILE: /* SKL */
4568 case INTEL_FAM6_SKYLAKE_DESKTOP: /* SKL */
4569 case INTEL_FAM6_KABYLAKE_MOBILE: /* KBL */
4570 case INTEL_FAM6_KABYLAKE_DESKTOP: /* KBL */
4571 crystal_hz = 24000000; /* 24.0 MHz */
4573 case INTEL_FAM6_ATOM_DENVERTON: /* DNV */
4574 crystal_hz = 25000000; /* 25.0 MHz */
4576 case INTEL_FAM6_ATOM_GOLDMONT: /* BXT */
4577 case INTEL_FAM6_ATOM_GEMINI_LAKE:
4578 crystal_hz = 19200000; /* 19.2 MHz */
4585 tsc_hz = (unsigned long long) crystal_hz * ebx_tsc / eax_crystal;
4587 fprintf(outf, "TSC: %lld MHz (%d Hz * %d / %d / 1000000)\n",
4588 tsc_hz / 1000000, crystal_hz, ebx_tsc, eax_crystal);
4592 if (max_level >= 0x16) {
4593 unsigned int base_mhz, max_mhz, bus_mhz, edx;
4596 * CPUID 16H Base MHz, Max MHz, Bus MHz
4598 base_mhz = max_mhz = bus_mhz = edx = 0;
4600 __cpuid(0x16, base_mhz, max_mhz, bus_mhz, edx);
4602 fprintf(outf, "CPUID(0x16): base_mhz: %d max_mhz: %d bus_mhz: %d\n",
4603 base_mhz, max_mhz, bus_mhz);
4607 aperf_mperf_multiplier = get_aperf_mperf_multiplier(family, model);
4609 BIC_PRESENT(BIC_IRQ);
4610 BIC_PRESENT(BIC_TSC_MHz);
4612 if (probe_nhm_msrs(family, model)) {
4613 do_nhm_platform_info = 1;
4614 BIC_PRESENT(BIC_CPU_c1);
4615 BIC_PRESENT(BIC_CPU_c3);
4616 BIC_PRESENT(BIC_CPU_c6);
4617 BIC_PRESENT(BIC_SMI);
4619 do_snb_cstates = has_snb_msrs(family, model);
4622 BIC_PRESENT(BIC_CPU_c7);
4624 do_irtl_snb = has_snb_msrs(family, model);
4625 if (do_snb_cstates && (pkg_cstate_limit >= PCL__2))
4626 BIC_PRESENT(BIC_Pkgpc2);
4627 if (pkg_cstate_limit >= PCL__3)
4628 BIC_PRESENT(BIC_Pkgpc3);
4629 if (pkg_cstate_limit >= PCL__6)
4630 BIC_PRESENT(BIC_Pkgpc6);
4631 if (do_snb_cstates && (pkg_cstate_limit >= PCL__7))
4632 BIC_PRESENT(BIC_Pkgpc7);
4633 if (has_slv_msrs(family, model)) {
4634 BIC_NOT_PRESENT(BIC_Pkgpc2);
4635 BIC_NOT_PRESENT(BIC_Pkgpc3);
4636 BIC_PRESENT(BIC_Pkgpc6);
4637 BIC_NOT_PRESENT(BIC_Pkgpc7);
4638 BIC_PRESENT(BIC_Mod_c6);
4639 use_c1_residency_msr = 1;
4641 if (is_dnv(family, model)) {
4642 BIC_PRESENT(BIC_CPU_c1);
4643 BIC_NOT_PRESENT(BIC_CPU_c3);
4644 BIC_NOT_PRESENT(BIC_Pkgpc3);
4645 BIC_NOT_PRESENT(BIC_CPU_c7);
4646 BIC_NOT_PRESENT(BIC_Pkgpc7);
4647 use_c1_residency_msr = 1;
4649 if (is_skx(family, model)) {
4650 BIC_NOT_PRESENT(BIC_CPU_c3);
4651 BIC_NOT_PRESENT(BIC_Pkgpc3);
4652 BIC_NOT_PRESENT(BIC_CPU_c7);
4653 BIC_NOT_PRESENT(BIC_Pkgpc7);
4655 if (is_bdx(family, model)) {
4656 BIC_NOT_PRESENT(BIC_CPU_c7);
4657 BIC_NOT_PRESENT(BIC_Pkgpc7);
4659 if (has_hsw_msrs(family, model)) {
4660 BIC_PRESENT(BIC_Pkgpc8);
4661 BIC_PRESENT(BIC_Pkgpc9);
4662 BIC_PRESENT(BIC_Pkgpc10);
4664 do_irtl_hsw = has_hsw_msrs(family, model);
4665 if (has_skl_msrs(family, model)) {
4666 BIC_PRESENT(BIC_Totl_c0);
4667 BIC_PRESENT(BIC_Any_c0);
4668 BIC_PRESENT(BIC_GFX_c0);
4669 BIC_PRESENT(BIC_CPUGFX);
4671 do_slm_cstates = is_slm(family, model);
4672 do_knl_cstates = is_knl(family, model);
4673 do_cnl_cstates = is_cnl(family, model);
4676 decode_misc_pwr_mgmt_msr();
4678 if (!quiet && has_slv_msrs(family, model))
4679 decode_c6_demotion_policy_msr();
4681 rapl_probe(family, model);
4682 perf_limit_reasons_probe(family, model);
4683 automatic_cstate_conversion_probe(family, model);
4686 dump_cstate_pstate_config_info(family, model);
4689 dump_sysfs_cstate_config();
4691 dump_sysfs_pstate_config();
4693 if (has_skl_msrs(family, model))
4694 calculate_tsc_tweak();
4696 if (!access("/sys/class/drm/card0/power/rc6_residency_ms", R_OK))
4697 BIC_PRESENT(BIC_GFX_rc6);
4699 if (!access("/sys/class/graphics/fb0/device/drm/card0/gt_cur_freq_mhz", R_OK))
4700 BIC_PRESENT(BIC_GFXMHz);
4702 if (!access("/sys/devices/system/cpu/cpuidle/low_power_idle_cpu_residency_us", R_OK))
4703 BIC_PRESENT(BIC_CPU_LPI);
4705 BIC_NOT_PRESENT(BIC_CPU_LPI);
4707 if (!access("/sys/devices/system/cpu/cpuidle/low_power_idle_system_residency_us", R_OK))
4708 BIC_PRESENT(BIC_SYS_LPI);
4710 BIC_NOT_PRESENT(BIC_SYS_LPI);
4713 decode_misc_feature_control();
4719 * in /dev/cpu/ return success for names that are numbers
4720 * ie. filter out ".", "..", "microcode".
4722 int dir_filter(const struct dirent *dirp)
4724 if (isdigit(dirp->d_name[0]))
4730 int open_dev_cpu_msr(int dummy1)
4735 void topology_probe()
4738 int max_core_id = 0;
4739 int max_package_id = 0;
4740 int max_siblings = 0;
4742 /* Initialize num_cpus, max_cpu_num */
4745 for_all_proc_cpus(count_cpus);
4746 if (!summary_only && topo.num_cpus > 1)
4747 BIC_PRESENT(BIC_CPU);
4750 fprintf(outf, "num_cpus %d max_cpu_num %d\n", topo.num_cpus, topo.max_cpu_num);
4752 cpus = calloc(1, (topo.max_cpu_num + 1) * sizeof(struct cpu_topology));
4754 err(1, "calloc cpus");
4757 * Allocate and initialize cpu_present_set
4759 cpu_present_set = CPU_ALLOC((topo.max_cpu_num + 1));
4760 if (cpu_present_set == NULL)
4761 err(3, "CPU_ALLOC");
4762 cpu_present_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
4763 CPU_ZERO_S(cpu_present_setsize, cpu_present_set);
4764 for_all_proc_cpus(mark_cpu_present);
4767 * Validate that all cpus in cpu_subset are also in cpu_present_set
4769 for (i = 0; i < CPU_SUBSET_MAXCPUS; ++i) {
4770 if (CPU_ISSET_S(i, cpu_subset_size, cpu_subset))
4771 if (!CPU_ISSET_S(i, cpu_present_setsize, cpu_present_set))
4772 err(1, "cpu%d not present", i);
4776 * Allocate and initialize cpu_affinity_set
4778 cpu_affinity_set = CPU_ALLOC((topo.max_cpu_num + 1));
4779 if (cpu_affinity_set == NULL)
4780 err(3, "CPU_ALLOC");
4781 cpu_affinity_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
4782 CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
4784 for_all_proc_cpus(init_thread_id);
4788 * find max_core_id, max_package_id
4790 for (i = 0; i <= topo.max_cpu_num; ++i) {
4793 if (cpu_is_not_present(i)) {
4795 fprintf(outf, "cpu%d NOT PRESENT\n", i);
4799 cpus[i].logical_cpu_id = i;
4801 /* get package information */
4802 cpus[i].physical_package_id = get_physical_package_id(i);
4803 if (cpus[i].physical_package_id > max_package_id)
4804 max_package_id = cpus[i].physical_package_id;
4806 /* get numa node information */
4807 cpus[i].physical_node_id = get_physical_node_id(&cpus[i]);
4808 if (cpus[i].physical_node_id > topo.max_node_num)
4809 topo.max_node_num = cpus[i].physical_node_id;
4811 /* get core information */
4812 cpus[i].physical_core_id = get_core_id(i);
4813 if (cpus[i].physical_core_id > max_core_id)
4814 max_core_id = cpus[i].physical_core_id;
4816 /* get thread information */
4817 siblings = get_thread_siblings(&cpus[i]);
4818 if (siblings > max_siblings)
4819 max_siblings = siblings;
4820 if (cpus[i].thread_id != -1)
4825 "cpu %d pkg %d node %d core %d thread %d\n",
4826 i, cpus[i].physical_package_id,
4827 cpus[i].physical_node_id,
4828 cpus[i].physical_core_id,
4832 topo.cores_per_node = max_core_id + 1;
4834 fprintf(outf, "max_core_id %d, sizing for %d cores per package\n",
4835 max_core_id, topo.cores_per_node);
4836 if (!summary_only && topo.cores_per_node > 1)
4837 BIC_PRESENT(BIC_Core);
4839 topo.num_packages = max_package_id + 1;
4841 fprintf(outf, "max_package_id %d, sizing for %d packages\n",
4842 max_package_id, topo.num_packages);
4843 if (!summary_only && topo.num_packages > 1)
4844 BIC_PRESENT(BIC_Package);
4848 fprintf(outf, "nodes_per_pkg %d\n", topo.nodes_per_pkg);
4849 if (!summary_only && topo.nodes_per_pkg > 1)
4850 BIC_PRESENT(BIC_Node);
4852 topo.threads_per_core = max_siblings;
4854 fprintf(outf, "max_siblings %d\n", max_siblings);
4858 allocate_counters(struct thread_data **t, struct core_data **c,
4859 struct pkg_data **p)
4862 int num_cores = topo.cores_per_node * topo.nodes_per_pkg *
4864 int num_threads = topo.threads_per_core * num_cores;
4866 *t = calloc(num_threads, sizeof(struct thread_data));
4870 for (i = 0; i < num_threads; i++)
4871 (*t)[i].cpu_id = -1;
4873 *c = calloc(num_cores, sizeof(struct core_data));
4877 for (i = 0; i < num_cores; i++)
4878 (*c)[i].core_id = -1;
4880 *p = calloc(topo.num_packages, sizeof(struct pkg_data));
4884 for (i = 0; i < topo.num_packages; i++)
4885 (*p)[i].package_id = i;
4889 err(1, "calloc counters");
4894 * set FIRST_THREAD_IN_CORE and FIRST_CORE_IN_PACKAGE
4896 void init_counter(struct thread_data *thread_base, struct core_data *core_base,
4897 struct pkg_data *pkg_base, int cpu_id)
4899 int pkg_id = cpus[cpu_id].physical_package_id;
4900 int node_id = cpus[cpu_id].logical_node_id;
4901 int core_id = cpus[cpu_id].physical_core_id;
4902 int thread_id = cpus[cpu_id].thread_id;
4903 struct thread_data *t;
4904 struct core_data *c;
4907 t = GET_THREAD(thread_base, thread_id, core_id, node_id, pkg_id);
4908 c = GET_CORE(core_base, core_id, node_id, pkg_id);
4909 p = GET_PKG(pkg_base, pkg_id);
4912 if (thread_id == 0) {
4913 t->flags |= CPU_IS_FIRST_THREAD_IN_CORE;
4914 if (cpu_is_first_core_in_package(cpu_id))
4915 t->flags |= CPU_IS_FIRST_CORE_IN_PACKAGE;
4918 c->core_id = core_id;
4919 p->package_id = pkg_id;
4923 int initialize_counters(int cpu_id)
4925 init_counter(EVEN_COUNTERS, cpu_id);
4926 init_counter(ODD_COUNTERS, cpu_id);
4930 void allocate_output_buffer()
4932 output_buffer = calloc(1, (1 + topo.num_cpus) * 1024);
4933 outp = output_buffer;
4935 err(-1, "calloc output buffer");
4937 void allocate_fd_percpu(void)
4939 fd_percpu = calloc(topo.max_cpu_num + 1, sizeof(int));
4940 if (fd_percpu == NULL)
4941 err(-1, "calloc fd_percpu");
4943 void allocate_irq_buffers(void)
4945 irq_column_2_cpu = calloc(topo.num_cpus, sizeof(int));
4946 if (irq_column_2_cpu == NULL)
4947 err(-1, "calloc %d", topo.num_cpus);
4949 irqs_per_cpu = calloc(topo.max_cpu_num + 1, sizeof(int));
4950 if (irqs_per_cpu == NULL)
4951 err(-1, "calloc %d", topo.max_cpu_num + 1);
4953 void setup_all_buffers(void)
4956 allocate_irq_buffers();
4957 allocate_fd_percpu();
4958 allocate_counters(&thread_even, &core_even, &package_even);
4959 allocate_counters(&thread_odd, &core_odd, &package_odd);
4960 allocate_output_buffer();
4961 for_all_proc_cpus(initialize_counters);
4964 void set_base_cpu(void)
4966 base_cpu = sched_getcpu();
4968 err(-ENODEV, "No valid cpus found");
4971 fprintf(outf, "base_cpu = %d\n", base_cpu);
4974 void turbostat_init()
4976 setup_all_buffers();
4979 check_permissions();
4984 for_all_cpus(print_hwp, ODD_COUNTERS);
4987 for_all_cpus(print_epb, ODD_COUNTERS);
4990 for_all_cpus(print_perf_limit, ODD_COUNTERS);
4993 for_all_cpus(print_rapl, ODD_COUNTERS);
4995 for_all_cpus(set_temperature_target, ODD_COUNTERS);
4998 for_all_cpus(print_thermal, ODD_COUNTERS);
5000 if (!quiet && do_irtl_snb)
5004 int fork_it(char **argv)
5009 snapshot_proc_sysfs_files();
5010 status = for_all_cpus(get_counters, EVEN_COUNTERS);
5011 first_counter_read = 0;
5014 /* clear affinity side-effect of get_counters() */
5015 sched_setaffinity(0, cpu_present_setsize, cpu_present_set);
5016 gettimeofday(&tv_even, (struct timezone *)NULL);
5021 execvp(argv[0], argv);
5022 err(errno, "exec %s", argv[0]);
5026 if (child_pid == -1)
5029 signal(SIGINT, SIG_IGN);
5030 signal(SIGQUIT, SIG_IGN);
5031 if (waitpid(child_pid, &status, 0) == -1)
5032 err(status, "waitpid");
5035 * n.b. fork_it() does not check for errors from for_all_cpus()
5036 * because re-starting is problematic when forking
5038 snapshot_proc_sysfs_files();
5039 for_all_cpus(get_counters, ODD_COUNTERS);
5040 gettimeofday(&tv_odd, (struct timezone *)NULL);
5041 timersub(&tv_odd, &tv_even, &tv_delta);
5042 if (for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS))
5043 fprintf(outf, "%s: Counter reset detected\n", progname);
5045 compute_average(EVEN_COUNTERS);
5046 format_all_counters(EVEN_COUNTERS);
5049 fprintf(outf, "%.6f sec\n", tv_delta.tv_sec + tv_delta.tv_usec/1000000.0);
5051 flush_output_stderr();
5056 int get_and_dump_counters(void)
5060 snapshot_proc_sysfs_files();
5061 status = for_all_cpus(get_counters, ODD_COUNTERS);
5065 status = for_all_cpus(dump_counters, ODD_COUNTERS);
5069 flush_output_stdout();
5074 void print_version() {
5075 fprintf(outf, "turbostat version 18.06.01"
5076 " - Len Brown <lenb@kernel.org>\n");
5079 int add_counter(unsigned int msr_num, char *path, char *name,
5080 unsigned int width, enum counter_scope scope,
5081 enum counter_type type, enum counter_format format, int flags)
5083 struct msr_counter *msrp;
5085 msrp = calloc(1, sizeof(struct msr_counter));
5091 msrp->msr_num = msr_num;
5092 strncpy(msrp->name, name, NAME_BYTES);
5094 strncpy(msrp->path, path, PATH_BYTES);
5095 msrp->width = width;
5097 msrp->format = format;
5098 msrp->flags = flags;
5103 msrp->next = sys.tp;
5105 sys.added_thread_counters++;
5106 if (sys.added_thread_counters > MAX_ADDED_THREAD_COUNTERS) {
5107 fprintf(stderr, "exceeded max %d added thread counters\n",
5108 MAX_ADDED_COUNTERS);
5114 msrp->next = sys.cp;
5116 sys.added_core_counters++;
5117 if (sys.added_core_counters > MAX_ADDED_COUNTERS) {
5118 fprintf(stderr, "exceeded max %d added core counters\n",
5119 MAX_ADDED_COUNTERS);
5125 msrp->next = sys.pp;
5127 sys.added_package_counters++;
5128 if (sys.added_package_counters > MAX_ADDED_COUNTERS) {
5129 fprintf(stderr, "exceeded max %d added package counters\n",
5130 MAX_ADDED_COUNTERS);
5139 void parse_add_command(char *add_command)
5143 char name_buffer[NAME_BYTES] = "";
5146 enum counter_scope scope = SCOPE_CPU;
5147 enum counter_type type = COUNTER_CYCLES;
5148 enum counter_format format = FORMAT_DELTA;
5150 while (add_command) {
5152 if (sscanf(add_command, "msr0x%x", &msr_num) == 1)
5155 if (sscanf(add_command, "msr%d", &msr_num) == 1)
5158 if (*add_command == '/') {
5163 if (sscanf(add_command, "u%d", &width) == 1) {
5164 if ((width == 32) || (width == 64))
5168 if (!strncmp(add_command, "cpu", strlen("cpu"))) {
5172 if (!strncmp(add_command, "core", strlen("core"))) {
5176 if (!strncmp(add_command, "package", strlen("package"))) {
5177 scope = SCOPE_PACKAGE;
5180 if (!strncmp(add_command, "cycles", strlen("cycles"))) {
5181 type = COUNTER_CYCLES;
5184 if (!strncmp(add_command, "seconds", strlen("seconds"))) {
5185 type = COUNTER_SECONDS;
5188 if (!strncmp(add_command, "usec", strlen("usec"))) {
5189 type = COUNTER_USEC;
5192 if (!strncmp(add_command, "raw", strlen("raw"))) {
5193 format = FORMAT_RAW;
5196 if (!strncmp(add_command, "delta", strlen("delta"))) {
5197 format = FORMAT_DELTA;
5200 if (!strncmp(add_command, "percent", strlen("percent"))) {
5201 format = FORMAT_PERCENT;
5205 if (sscanf(add_command, "%18s,%*s", name_buffer) == 1) { /* 18 < NAME_BYTES */
5208 eos = strchr(name_buffer, ',');
5215 add_command = strchr(add_command, ',');
5217 *add_command = '\0';
5222 if ((msr_num == 0) && (path == NULL)) {
5223 fprintf(stderr, "--add: (msrDDD | msr0xXXX | /path_to_counter ) required\n");
5227 /* generate default column header */
5228 if (*name_buffer == '\0') {
5230 sprintf(name_buffer, "M0x%x%s", msr_num, format == FORMAT_PERCENT ? "%" : "");
5232 sprintf(name_buffer, "M0X%x%s", msr_num, format == FORMAT_PERCENT ? "%" : "");
5235 if (add_counter(msr_num, path, name_buffer, width, scope, type, format, 0))
5244 int is_deferred_skip(char *name)
5248 for (i = 0; i < deferred_skip_index; ++i)
5249 if (!strcmp(name, deferred_skip_names[i]))
5254 void probe_sysfs(void)
5262 if (!DO_BIC(BIC_sysfs))
5265 for (state = 10; state >= 0; --state) {
5267 sprintf(path, "/sys/devices/system/cpu/cpu%d/cpuidle/state%d/name",
5269 input = fopen(path, "r");
5272 fgets(name_buf, sizeof(name_buf), input);
5274 /* truncate "C1-HSW\n" to "C1", or truncate "C1\n" to "C1" */
5275 sp = strchr(name_buf, '-');
5277 sp = strchrnul(name_buf, '\n');
5283 sprintf(path, "cpuidle/state%d/time", state);
5285 if (is_deferred_skip(name_buf))
5288 add_counter(0, path, name_buf, 64, SCOPE_CPU, COUNTER_USEC,
5289 FORMAT_PERCENT, SYSFS_PERCPU);
5292 for (state = 10; state >= 0; --state) {
5294 sprintf(path, "/sys/devices/system/cpu/cpu%d/cpuidle/state%d/name",
5296 input = fopen(path, "r");
5299 fgets(name_buf, sizeof(name_buf), input);
5300 /* truncate "C1-HSW\n" to "C1", or truncate "C1\n" to "C1" */
5301 sp = strchr(name_buf, '-');
5303 sp = strchrnul(name_buf, '\n');
5307 sprintf(path, "cpuidle/state%d/usage", state);
5309 if (is_deferred_skip(name_buf))
5312 add_counter(0, path, name_buf, 64, SCOPE_CPU, COUNTER_ITEMS,
5313 FORMAT_DELTA, SYSFS_PERCPU);
5320 * parse cpuset with following syntax
5321 * 1,2,4..6,8-10 and set bits in cpu_subset
5323 void parse_cpu_command(char *optarg)
5325 unsigned int start, end;
5328 if (!strcmp(optarg, "core")) {
5334 if (!strcmp(optarg, "package")) {
5340 if (show_core_only || show_pkg_only)
5343 cpu_subset = CPU_ALLOC(CPU_SUBSET_MAXCPUS);
5344 if (cpu_subset == NULL)
5345 err(3, "CPU_ALLOC");
5346 cpu_subset_size = CPU_ALLOC_SIZE(CPU_SUBSET_MAXCPUS);
5348 CPU_ZERO_S(cpu_subset_size, cpu_subset);
5352 while (next && *next) {
5354 if (*next == '-') /* no negative cpu numbers */
5357 start = strtoul(next, &next, 10);
5359 if (start >= CPU_SUBSET_MAXCPUS)
5361 CPU_SET_S(start, cpu_subset_size, cpu_subset);
5372 next += 1; /* start range */
5373 } else if (*next == '.') {
5376 next += 1; /* start range */
5381 end = strtoul(next, &next, 10);
5385 while (++start <= end) {
5386 if (start >= CPU_SUBSET_MAXCPUS)
5388 CPU_SET_S(start, cpu_subset_size, cpu_subset);
5393 else if (*next != '\0')
5400 fprintf(stderr, "\"--cpu %s\" malformed\n", optarg);
5406 void cmdline(int argc, char **argv)
5409 int option_index = 0;
5410 static struct option long_options[] = {
5411 {"add", required_argument, 0, 'a'},
5412 {"cpu", required_argument, 0, 'c'},
5413 {"Dump", no_argument, 0, 'D'},
5414 {"debug", no_argument, 0, 'd'}, /* internal, not documented */
5415 {"enable", required_argument, 0, 'e'},
5416 {"interval", required_argument, 0, 'i'},
5417 {"num_iterations", required_argument, 0, 'n'},
5418 {"help", no_argument, 0, 'h'},
5419 {"hide", required_argument, 0, 'H'}, // meh, -h taken by --help
5420 {"Joules", no_argument, 0, 'J'},
5421 {"list", no_argument, 0, 'l'},
5422 {"out", required_argument, 0, 'o'},
5423 {"quiet", no_argument, 0, 'q'},
5424 {"show", required_argument, 0, 's'},
5425 {"Summary", no_argument, 0, 'S'},
5426 {"TCC", required_argument, 0, 'T'},
5427 {"version", no_argument, 0, 'v' },
5433 while ((opt = getopt_long_only(argc, argv, "+C:c:Dde:hi:Jn:o:qST:v",
5434 long_options, &option_index)) != -1) {
5437 parse_add_command(optarg);
5440 parse_cpu_command(optarg);
5446 /* --enable specified counter */
5447 bic_enabled = bic_enabled | bic_lookup(optarg, SHOW_LIST);
5451 ENABLE_BIC(BIC_DISABLED_BY_DEFAULT);
5455 * --hide: do not show those specified
5456 * multiple invocations simply clear more bits in enabled mask
5458 bic_enabled &= ~bic_lookup(optarg, HIDE_LIST);
5466 double interval = strtod(optarg, NULL);
5468 if (interval < 0.001) {
5469 fprintf(outf, "interval %f seconds is too small\n",
5474 interval_tv.tv_sec = interval_ts.tv_sec = interval;
5475 interval_tv.tv_usec = (interval - interval_tv.tv_sec) * 1000000;
5476 interval_ts.tv_nsec = (interval - interval_ts.tv_sec) * 1000000000;
5483 ENABLE_BIC(BIC_DISABLED_BY_DEFAULT);
5488 outf = fopen_or_die(optarg, "w");
5494 num_iterations = strtod(optarg, NULL);
5496 if (num_iterations <= 0) {
5497 fprintf(outf, "iterations %d should be positive number\n",
5504 * --show: show only those specified
5505 * The 1st invocation will clear and replace the enabled mask
5506 * subsequent invocations can add to it.
5509 bic_enabled = bic_lookup(optarg, SHOW_LIST);
5511 bic_enabled |= bic_lookup(optarg, SHOW_LIST);
5518 tcc_activation_temp_override = atoi(optarg);
5528 int main(int argc, char **argv)
5531 cmdline(argc, argv);
5540 /* dump counters and exit */
5542 return get_and_dump_counters();
5544 /* list header and exit */
5545 if (list_header_only) {
5547 flush_output_stdout();
5552 * if any params left, it must be a command to fork
5555 return fork_it(argv + optind);