1 // SPDX-License-Identifier: GPL-2.0-only
3 * turbostat -- show CPU frequency and C-state residency
4 * on modern Intel and AMD processors.
6 * Copyright (c) 2013 Intel Corporation.
7 * Len Brown <len.brown@intel.com>
12 #include INTEL_FAMILY_HEADER
17 #include <sys/types.h>
20 #include <sys/select.h>
21 #include <sys/resource.h>
33 #include <sys/capability.h>
36 #include <linux/perf_event.h>
37 #include <asm/unistd.h>
39 char *proc_stat = "/proc/stat";
42 int *fd_instr_count_percpu;
43 struct timeval interval_tv = {5, 0};
44 struct timespec interval_ts = {5, 0};
45 unsigned int num_iterations;
49 unsigned int sums_need_wide_columns;
50 unsigned int rapl_joules;
51 unsigned int summary_only;
52 unsigned int list_header_only;
53 unsigned int dump_only;
54 unsigned int do_snb_cstates;
55 unsigned int do_knl_cstates;
56 unsigned int do_slm_cstates;
57 unsigned int use_c1_residency_msr;
58 unsigned int has_aperf;
60 unsigned int do_irtl_snb;
61 unsigned int do_irtl_hsw;
62 unsigned int units = 1000000; /* MHz etc */
63 unsigned int genuine_intel;
64 unsigned int authentic_amd;
65 unsigned int hygon_genuine;
66 unsigned int max_level, max_extended_level;
67 unsigned int has_invariant_tsc;
68 unsigned int do_nhm_platform_info;
69 unsigned int no_MSR_MISC_PWR_MGMT;
70 unsigned int aperf_mperf_multiplier = 1;
73 unsigned int has_base_hz;
74 double tsc_tweak = 1.0;
75 unsigned int show_pkg_only;
76 unsigned int show_core_only;
77 char *output_buffer, *outp;
82 unsigned long long gfx_cur_rc6_ms;
83 unsigned long long cpuidle_cur_cpu_lpi_us;
84 unsigned long long cpuidle_cur_sys_lpi_us;
85 unsigned int gfx_cur_mhz;
86 unsigned int gfx_act_mhz;
87 unsigned int tcc_activation_temp;
88 unsigned int tcc_activation_temp_override;
89 double rapl_power_units, rapl_time_units;
90 double rapl_dram_energy_units, rapl_energy_units;
91 double rapl_joule_counter_range;
92 unsigned int do_core_perf_limit_reasons;
93 unsigned int has_automatic_cstate_conversion;
94 unsigned int do_gfx_perf_limit_reasons;
95 unsigned int do_ring_perf_limit_reasons;
96 unsigned int crystal_hz;
97 unsigned long long tsc_hz;
99 double discover_bclk(unsigned int family, unsigned int model);
100 unsigned int has_hwp; /* IA32_PM_ENABLE, IA32_HWP_CAPABILITIES */
101 /* IA32_HWP_REQUEST, IA32_HWP_STATUS */
102 unsigned int has_hwp_notify; /* IA32_HWP_INTERRUPT */
103 unsigned int has_hwp_activity_window; /* IA32_HWP_REQUEST[bits 41:32] */
104 unsigned int has_hwp_epp; /* IA32_HWP_REQUEST[bits 31:24] */
105 unsigned int has_hwp_pkg; /* IA32_HWP_REQUEST_PKG */
106 unsigned int has_misc_feature_control;
107 unsigned int first_counter_read = 1;
110 #define RAPL_PKG (1 << 0)
111 /* 0x610 MSR_PKG_POWER_LIMIT */
112 /* 0x611 MSR_PKG_ENERGY_STATUS */
113 #define RAPL_PKG_PERF_STATUS (1 << 1)
114 /* 0x613 MSR_PKG_PERF_STATUS */
115 #define RAPL_PKG_POWER_INFO (1 << 2)
116 /* 0x614 MSR_PKG_POWER_INFO */
118 #define RAPL_DRAM (1 << 3)
119 /* 0x618 MSR_DRAM_POWER_LIMIT */
120 /* 0x619 MSR_DRAM_ENERGY_STATUS */
121 #define RAPL_DRAM_PERF_STATUS (1 << 4)
122 /* 0x61b MSR_DRAM_PERF_STATUS */
123 #define RAPL_DRAM_POWER_INFO (1 << 5)
124 /* 0x61c MSR_DRAM_POWER_INFO */
126 #define RAPL_CORES_POWER_LIMIT (1 << 6)
127 /* 0x638 MSR_PP0_POWER_LIMIT */
128 #define RAPL_CORE_POLICY (1 << 7)
129 /* 0x63a MSR_PP0_POLICY */
131 #define RAPL_GFX (1 << 8)
132 /* 0x640 MSR_PP1_POWER_LIMIT */
133 /* 0x641 MSR_PP1_ENERGY_STATUS */
134 /* 0x642 MSR_PP1_POLICY */
136 #define RAPL_CORES_ENERGY_STATUS (1 << 9)
137 /* 0x639 MSR_PP0_ENERGY_STATUS */
138 #define RAPL_PER_CORE_ENERGY (1 << 10)
139 /* Indicates cores energy collection is per-core,
140 * not per-package. */
141 #define RAPL_AMD_F17H (1 << 11)
142 /* 0xc0010299 MSR_RAPL_PWR_UNIT */
143 /* 0xc001029a MSR_CORE_ENERGY_STAT */
144 /* 0xc001029b MSR_PKG_ENERGY_STAT */
145 #define RAPL_CORES (RAPL_CORES_ENERGY_STATUS | RAPL_CORES_POWER_LIMIT)
146 #define TJMAX_DEFAULT 100
148 /* MSRs that are not yet in the kernel-provided header. */
149 #define MSR_RAPL_PWR_UNIT 0xc0010299
150 #define MSR_CORE_ENERGY_STAT 0xc001029a
151 #define MSR_PKG_ENERGY_STAT 0xc001029b
153 #define MAX(a, b) ((a) > (b) ? (a) : (b))
156 * buffer size used by sscanf() for added column names
157 * Usually truncated to 7 characters, but also handles 18 columns for raw 64-bit counters
159 #define NAME_BYTES 20
160 #define PATH_BYTES 128
165 #define CPU_SUBSET_MAXCPUS 1024 /* need to use before probe... */
166 cpu_set_t *cpu_present_set, *cpu_affinity_set, *cpu_subset;
167 size_t cpu_present_setsize, cpu_affinity_setsize, cpu_subset_size;
168 #define MAX_ADDED_COUNTERS 8
169 #define MAX_ADDED_THREAD_COUNTERS 24
170 #define BITMASK_SIZE 32
173 struct timeval tv_begin;
174 struct timeval tv_end;
175 struct timeval tv_delta;
176 unsigned long long tsc;
177 unsigned long long aperf;
178 unsigned long long mperf;
179 unsigned long long c1;
180 unsigned long long instr_count;
181 unsigned long long irq_count;
182 unsigned int smi_count;
184 unsigned int apic_id;
185 unsigned int x2apic_id;
187 #define CPU_IS_FIRST_THREAD_IN_CORE 0x2
188 #define CPU_IS_FIRST_CORE_IN_PACKAGE 0x4
189 unsigned long long counter[MAX_ADDED_THREAD_COUNTERS];
190 } *thread_even, *thread_odd;
193 unsigned long long c3;
194 unsigned long long c6;
195 unsigned long long c7;
196 unsigned long long mc6_us; /* duplicate as per-core for now, even though per module */
197 unsigned int core_temp_c;
198 unsigned int core_energy; /* MSR_CORE_ENERGY_STAT */
199 unsigned int core_id;
200 unsigned long long counter[MAX_ADDED_COUNTERS];
201 } *core_even, *core_odd;
204 unsigned long long pc2;
205 unsigned long long pc3;
206 unsigned long long pc6;
207 unsigned long long pc7;
208 unsigned long long pc8;
209 unsigned long long pc9;
210 unsigned long long pc10;
211 unsigned long long cpu_lpi;
212 unsigned long long sys_lpi;
213 unsigned long long pkg_wtd_core_c0;
214 unsigned long long pkg_any_core_c0;
215 unsigned long long pkg_any_gfxe_c0;
216 unsigned long long pkg_both_core_gfxe_c0;
217 long long gfx_rc6_ms;
218 unsigned int gfx_mhz;
219 unsigned int gfx_act_mhz;
220 unsigned int package_id;
221 unsigned long long energy_pkg; /* MSR_PKG_ENERGY_STATUS */
222 unsigned long long energy_dram; /* MSR_DRAM_ENERGY_STATUS */
223 unsigned long long energy_cores; /* MSR_PP0_ENERGY_STATUS */
224 unsigned long long energy_gfx; /* MSR_PP1_ENERGY_STATUS */
225 unsigned long long rapl_pkg_perf_status; /* MSR_PKG_PERF_STATUS */
226 unsigned long long rapl_dram_perf_status; /* MSR_DRAM_PERF_STATUS */
227 unsigned int pkg_temp_c;
228 unsigned long long counter[MAX_ADDED_COUNTERS];
229 } *package_even, *package_odd;
231 #define ODD_COUNTERS thread_odd, core_odd, package_odd
232 #define EVEN_COUNTERS thread_even, core_even, package_even
234 #define GET_THREAD(thread_base, thread_no, core_no, node_no, pkg_no) \
237 topo.nodes_per_pkg * topo.cores_per_node * topo.threads_per_core) + \
238 ((node_no) * topo.cores_per_node * topo.threads_per_core) + \
239 ((core_no) * topo.threads_per_core) + \
242 #define GET_CORE(core_base, core_no, node_no, pkg_no) \
244 ((pkg_no) * topo.nodes_per_pkg * topo.cores_per_node) + \
245 ((node_no) * topo.cores_per_node) + \
249 #define GET_PKG(pkg_base, pkg_no) (pkg_base + pkg_no)
251 enum counter_scope {SCOPE_CPU, SCOPE_CORE, SCOPE_PACKAGE};
252 enum counter_type {COUNTER_ITEMS, COUNTER_CYCLES, COUNTER_SECONDS, COUNTER_USEC};
253 enum counter_format {FORMAT_RAW, FORMAT_DELTA, FORMAT_PERCENT};
256 unsigned int msr_num;
257 char name[NAME_BYTES];
258 char path[PATH_BYTES];
260 enum counter_type type;
261 enum counter_format format;
262 struct msr_counter *next;
264 #define FLAGS_HIDE (1 << 0)
265 #define FLAGS_SHOW (1 << 1)
266 #define SYSFS_PERCPU (1 << 1)
270 * The accumulated sum of MSR is defined as a monotonic
271 * increasing MSR, it will be accumulated periodically,
272 * despite its register's bit width.
284 int get_msr_sum(int cpu, off_t offset, unsigned long long *msr);
286 struct msr_sum_array {
287 /* get_msr_sum() = sum + (get_msr() - last) */
289 /*The accumulated MSR value is updated by the timer*/
290 unsigned long long sum;
291 /*The MSR footprint recorded in last timer*/
292 unsigned long long last;
293 } entries[IDX_COUNT];
296 /* The percpu MSR sum array.*/
297 struct msr_sum_array *per_cpu_msr_sum;
299 int idx_to_offset(int idx)
305 offset = MSR_PKG_ENERGY_STATUS;
307 case IDX_DRAM_ENERGY:
308 offset = MSR_DRAM_ENERGY_STATUS;
311 offset = MSR_PP0_ENERGY_STATUS;
314 offset = MSR_PP1_ENERGY_STATUS;
317 offset = MSR_PKG_PERF_STATUS;
320 offset = MSR_DRAM_PERF_STATUS;
328 int offset_to_idx(int offset)
333 case MSR_PKG_ENERGY_STATUS:
334 idx = IDX_PKG_ENERGY;
336 case MSR_DRAM_ENERGY_STATUS:
337 idx = IDX_DRAM_ENERGY;
339 case MSR_PP0_ENERGY_STATUS:
340 idx = IDX_PP0_ENERGY;
342 case MSR_PP1_ENERGY_STATUS:
343 idx = IDX_PP1_ENERGY;
345 case MSR_PKG_PERF_STATUS:
348 case MSR_DRAM_PERF_STATUS:
357 int idx_valid(int idx)
361 return do_rapl & RAPL_PKG;
362 case IDX_DRAM_ENERGY:
363 return do_rapl & RAPL_DRAM;
365 return do_rapl & RAPL_CORES_ENERGY_STATUS;
367 return do_rapl & RAPL_GFX;
369 return do_rapl & RAPL_PKG_PERF_STATUS;
371 return do_rapl & RAPL_DRAM_PERF_STATUS;
376 struct sys_counters {
377 unsigned int added_thread_counters;
378 unsigned int added_core_counters;
379 unsigned int added_package_counters;
380 struct msr_counter *tp;
381 struct msr_counter *cp;
382 struct msr_counter *pp;
385 struct system_summary {
386 struct thread_data threads;
387 struct core_data cores;
388 struct pkg_data packages;
391 struct cpu_topology {
392 int physical_package_id;
395 int physical_node_id;
396 int logical_node_id; /* 0-based count within the package */
397 int physical_core_id;
399 cpu_set_t *put_ids; /* Processing Unit/Thread IDs */
411 int threads_per_core;
414 struct timeval tv_even, tv_odd, tv_delta;
416 int *irq_column_2_cpu; /* /proc/interrupts column numbers */
417 int *irqs_per_cpu; /* indexed by cpu_num */
419 void setup_all_buffers(void);
422 char *sys_lpi_file_sysfs = "/sys/devices/system/cpu/cpuidle/low_power_idle_system_residency_us";
423 char *sys_lpi_file_debugfs = "/sys/kernel/debug/pmc_core/slp_s0_residency_usec";
425 int cpu_is_not_present(int cpu)
427 return !CPU_ISSET_S(cpu, cpu_present_setsize, cpu_present_set);
430 * run func(thread, core, package) in topology order
431 * skip non-present cpus
434 int for_all_cpus(int (func)(struct thread_data *, struct core_data *, struct pkg_data *),
435 struct thread_data *thread_base, struct core_data *core_base, struct pkg_data *pkg_base)
437 int retval, pkg_no, core_no, thread_no, node_no;
439 for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
440 for (node_no = 0; node_no < topo.nodes_per_pkg; node_no++) {
441 for (core_no = 0; core_no < topo.cores_per_node; ++core_no) {
442 for (thread_no = 0; thread_no <
443 topo.threads_per_core; ++thread_no) {
444 struct thread_data *t;
448 t = GET_THREAD(thread_base, thread_no,
452 if (cpu_is_not_present(t->cpu_id))
455 c = GET_CORE(core_base, core_no,
457 p = GET_PKG(pkg_base, pkg_no);
459 retval = func(t, c, p);
469 int cpu_migrate(int cpu)
471 CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
472 CPU_SET_S(cpu, cpu_affinity_setsize, cpu_affinity_set);
473 if (sched_setaffinity(0, cpu_affinity_setsize, cpu_affinity_set) == -1)
478 int get_msr_fd(int cpu)
488 sprintf(pathname, "/dev/cpu/%d/msr", cpu);
489 fd = open(pathname, O_RDONLY);
491 err(-1, "%s open failed, try chown or chmod +r /dev/cpu/*/msr, or run as root", pathname);
498 static long perf_event_open(struct perf_event_attr *hw_event, pid_t pid, int cpu, int group_fd, unsigned long flags)
500 return syscall(__NR_perf_event_open, hw_event, pid, cpu, group_fd, flags);
503 static int perf_instr_count_open(int cpu_num)
505 struct perf_event_attr pea;
508 memset(&pea, 0, sizeof(struct perf_event_attr));
509 pea.type = PERF_TYPE_HARDWARE;
510 pea.size = sizeof(struct perf_event_attr);
511 pea.config = PERF_COUNT_HW_INSTRUCTIONS;
513 /* counter for cpu_num, including user + kernel and all processes */
514 fd = perf_event_open(&pea, -1, cpu_num, -1, 0);
516 err(-1, "cpu%d: perf instruction counter\n", cpu_num);
521 int get_instr_count_fd(int cpu)
523 if (fd_instr_count_percpu[cpu])
524 return fd_instr_count_percpu[cpu];
526 fd_instr_count_percpu[cpu] = perf_instr_count_open(cpu);
528 return fd_instr_count_percpu[cpu];
531 int get_msr(int cpu, off_t offset, unsigned long long *msr)
535 retval = pread(get_msr_fd(cpu), msr, sizeof(*msr), offset);
537 if (retval != sizeof *msr)
538 err(-1, "cpu%d: msr offset 0x%llx read failed", cpu, (unsigned long long)offset);
544 * This list matches the column headers, except
545 * 1. built-in only, the sysfs counters are not here -- we learn of those at run-time
546 * 2. Core and CPU are moved to the end, we can't have strings that contain them
547 * matching on them for --show and --hide.
549 struct msr_counter bic[] = {
551 { 0x0, "Time_Of_Day_Seconds" },
559 { 0x0, "SMI", "", 32, 0, FORMAT_DELTA, NULL},
605 #define MAX_BIC (sizeof(bic) / sizeof(struct msr_counter))
606 #define BIC_USEC (1ULL << 0)
607 #define BIC_TOD (1ULL << 1)
608 #define BIC_Package (1ULL << 2)
609 #define BIC_Node (1ULL << 3)
610 #define BIC_Avg_MHz (1ULL << 4)
611 #define BIC_Busy (1ULL << 5)
612 #define BIC_Bzy_MHz (1ULL << 6)
613 #define BIC_TSC_MHz (1ULL << 7)
614 #define BIC_IRQ (1ULL << 8)
615 #define BIC_SMI (1ULL << 9)
616 #define BIC_sysfs (1ULL << 10)
617 #define BIC_CPU_c1 (1ULL << 11)
618 #define BIC_CPU_c3 (1ULL << 12)
619 #define BIC_CPU_c6 (1ULL << 13)
620 #define BIC_CPU_c7 (1ULL << 14)
621 #define BIC_ThreadC (1ULL << 15)
622 #define BIC_CoreTmp (1ULL << 16)
623 #define BIC_CoreCnt (1ULL << 17)
624 #define BIC_PkgTmp (1ULL << 18)
625 #define BIC_GFX_rc6 (1ULL << 19)
626 #define BIC_GFXMHz (1ULL << 20)
627 #define BIC_Pkgpc2 (1ULL << 21)
628 #define BIC_Pkgpc3 (1ULL << 22)
629 #define BIC_Pkgpc6 (1ULL << 23)
630 #define BIC_Pkgpc7 (1ULL << 24)
631 #define BIC_Pkgpc8 (1ULL << 25)
632 #define BIC_Pkgpc9 (1ULL << 26)
633 #define BIC_Pkgpc10 (1ULL << 27)
634 #define BIC_CPU_LPI (1ULL << 28)
635 #define BIC_SYS_LPI (1ULL << 29)
636 #define BIC_PkgWatt (1ULL << 30)
637 #define BIC_CorWatt (1ULL << 31)
638 #define BIC_GFXWatt (1ULL << 32)
639 #define BIC_PkgCnt (1ULL << 33)
640 #define BIC_RAMWatt (1ULL << 34)
641 #define BIC_PKG__ (1ULL << 35)
642 #define BIC_RAM__ (1ULL << 36)
643 #define BIC_Pkg_J (1ULL << 37)
644 #define BIC_Cor_J (1ULL << 38)
645 #define BIC_GFX_J (1ULL << 39)
646 #define BIC_RAM_J (1ULL << 40)
647 #define BIC_Mod_c6 (1ULL << 41)
648 #define BIC_Totl_c0 (1ULL << 42)
649 #define BIC_Any_c0 (1ULL << 43)
650 #define BIC_GFX_c0 (1ULL << 44)
651 #define BIC_CPUGFX (1ULL << 45)
652 #define BIC_Core (1ULL << 46)
653 #define BIC_CPU (1ULL << 47)
654 #define BIC_APIC (1ULL << 48)
655 #define BIC_X2APIC (1ULL << 49)
656 #define BIC_Die (1ULL << 50)
657 #define BIC_GFXACTMHz (1ULL << 51)
658 #define BIC_IPC (1ULL << 52)
660 #define BIC_DISABLED_BY_DEFAULT (BIC_USEC | BIC_TOD | BIC_APIC | BIC_X2APIC)
662 unsigned long long bic_enabled = (0xFFFFFFFFFFFFFFFFULL & ~BIC_DISABLED_BY_DEFAULT);
663 unsigned long long bic_present = BIC_USEC | BIC_TOD | BIC_sysfs | BIC_APIC | BIC_X2APIC;
665 #define DO_BIC(COUNTER_NAME) (bic_enabled & bic_present & COUNTER_NAME)
666 #define DO_BIC_READ(COUNTER_NAME) (bic_present & COUNTER_NAME)
667 #define ENABLE_BIC(COUNTER_NAME) (bic_enabled |= COUNTER_NAME)
668 #define BIC_PRESENT(COUNTER_BIT) (bic_present |= COUNTER_BIT)
669 #define BIC_NOT_PRESENT(COUNTER_BIT) (bic_present &= ~COUNTER_BIT)
670 #define BIC_IS_ENABLED(COUNTER_BIT) (bic_enabled & COUNTER_BIT)
673 #define MAX_DEFERRED 16
674 char *deferred_skip_names[MAX_DEFERRED];
675 int deferred_skip_index;
678 * HIDE_LIST - hide this list of counters, show the rest [default]
679 * SHOW_LIST - show this list of counters, hide the rest
681 enum show_hide_mode { SHOW_LIST, HIDE_LIST } global_show_hide_mode = HIDE_LIST;
686 "Usage: turbostat [OPTIONS][(--interval seconds) | COMMAND ...]\n"
688 "Turbostat forks the specified COMMAND and prints statistics\n"
689 "when COMMAND completes.\n"
690 "If no COMMAND is specified, turbostat wakes every 5-seconds\n"
691 "to print statistics, until interrupted.\n"
692 " -a, --add add a counter\n"
693 " eg. --add msr0x10,u64,cpu,delta,MY_TSC\n"
694 " -c, --cpu cpu-set limit output to summary plus cpu-set:\n"
695 " {core | package | j,k,l..m,n-p }\n"
696 " -d, --debug displays usec, Time_Of_Day_Seconds and more debugging\n"
697 " -D, --Dump displays the raw counter values\n"
698 " -e, --enable [all | column]\n"
699 " shows all or the specified disabled column\n"
700 " -H, --hide [column|column,column,...]\n"
701 " hide the specified column(s)\n"
702 " -i, --interval sec.subsec\n"
703 " Override default 5-second measurement interval\n"
704 " -J, --Joules displays energy in Joules instead of Watts\n"
705 " -l, --list list column headers only\n"
706 " -n, --num_iterations num\n"
707 " number of the measurement iterations\n"
709 " create or truncate \"file\" for all output\n"
710 " -q, --quiet skip decoding system configuration header\n"
711 " -s, --show [column|column,column,...]\n"
712 " show only the specified column(s)\n"
714 " limits output to 1-line system summary per interval\n"
715 " -T, --TCC temperature\n"
716 " sets the Thermal Control Circuit temperature in\n"
718 " -h, --help print this help message\n"
719 " -v, --version print version information\n"
721 "For more help, run \"man turbostat\"\n");
726 * for all the strings in comma separate name_list,
727 * set the approprate bit in return value.
729 unsigned long long bic_lookup(char *name_list, enum show_hide_mode mode)
732 unsigned long long retval = 0;
737 comma = strchr(name_list, ',');
742 if (!strcmp(name_list, "all"))
745 for (i = 0; i < MAX_BIC; ++i) {
746 if (!strcmp(name_list, bic[i].name)) {
747 retval |= (1ULL << i);
752 if (mode == SHOW_LIST) {
753 fprintf(stderr, "Invalid counter name: %s\n", name_list);
756 deferred_skip_names[deferred_skip_index++] = name_list;
758 fprintf(stderr, "deferred \"%s\"\n", name_list);
759 if (deferred_skip_index >= MAX_DEFERRED) {
760 fprintf(stderr, "More than max %d un-recognized --skip options '%s'\n",
761 MAX_DEFERRED, name_list);
776 void print_header(char *delim)
778 struct msr_counter *mp;
781 if (DO_BIC(BIC_USEC))
782 outp += sprintf(outp, "%susec", (printed++ ? delim : ""));
784 outp += sprintf(outp, "%sTime_Of_Day_Seconds", (printed++ ? delim : ""));
785 if (DO_BIC(BIC_Package))
786 outp += sprintf(outp, "%sPackage", (printed++ ? delim : ""));
788 outp += sprintf(outp, "%sDie", (printed++ ? delim : ""));
789 if (DO_BIC(BIC_Node))
790 outp += sprintf(outp, "%sNode", (printed++ ? delim : ""));
791 if (DO_BIC(BIC_Core))
792 outp += sprintf(outp, "%sCore", (printed++ ? delim : ""));
794 outp += sprintf(outp, "%sCPU", (printed++ ? delim : ""));
795 if (DO_BIC(BIC_APIC))
796 outp += sprintf(outp, "%sAPIC", (printed++ ? delim : ""));
797 if (DO_BIC(BIC_X2APIC))
798 outp += sprintf(outp, "%sX2APIC", (printed++ ? delim : ""));
799 if (DO_BIC(BIC_Avg_MHz))
800 outp += sprintf(outp, "%sAvg_MHz", (printed++ ? delim : ""));
801 if (DO_BIC(BIC_Busy))
802 outp += sprintf(outp, "%sBusy%%", (printed++ ? delim : ""));
803 if (DO_BIC(BIC_Bzy_MHz))
804 outp += sprintf(outp, "%sBzy_MHz", (printed++ ? delim : ""));
805 if (DO_BIC(BIC_TSC_MHz))
806 outp += sprintf(outp, "%sTSC_MHz", (printed++ ? delim : ""));
809 outp += sprintf(outp, "%sIPC", (printed++ ? delim : ""));
811 if (DO_BIC(BIC_IRQ)) {
812 if (sums_need_wide_columns)
813 outp += sprintf(outp, "%s IRQ", (printed++ ? delim : ""));
815 outp += sprintf(outp, "%sIRQ", (printed++ ? delim : ""));
819 outp += sprintf(outp, "%sSMI", (printed++ ? delim : ""));
821 for (mp = sys.tp; mp; mp = mp->next) {
823 if (mp->format == FORMAT_RAW) {
825 outp += sprintf(outp, "%s%18.18s", (printed++ ? delim : ""), mp->name);
827 outp += sprintf(outp, "%s%10.10s", (printed++ ? delim : ""), mp->name);
829 if ((mp->type == COUNTER_ITEMS) && sums_need_wide_columns)
830 outp += sprintf(outp, "%s%8s", (printed++ ? delim : ""), mp->name);
832 outp += sprintf(outp, "%s%s", (printed++ ? delim : ""), mp->name);
836 if (DO_BIC(BIC_CPU_c1))
837 outp += sprintf(outp, "%sCPU%%c1", (printed++ ? delim : ""));
838 if (DO_BIC(BIC_CPU_c3))
839 outp += sprintf(outp, "%sCPU%%c3", (printed++ ? delim : ""));
840 if (DO_BIC(BIC_CPU_c6))
841 outp += sprintf(outp, "%sCPU%%c6", (printed++ ? delim : ""));
842 if (DO_BIC(BIC_CPU_c7))
843 outp += sprintf(outp, "%sCPU%%c7", (printed++ ? delim : ""));
845 if (DO_BIC(BIC_Mod_c6))
846 outp += sprintf(outp, "%sMod%%c6", (printed++ ? delim : ""));
848 if (DO_BIC(BIC_CoreTmp))
849 outp += sprintf(outp, "%sCoreTmp", (printed++ ? delim : ""));
851 if (do_rapl && !rapl_joules) {
852 if (DO_BIC(BIC_CorWatt) && (do_rapl & RAPL_PER_CORE_ENERGY))
853 outp += sprintf(outp, "%sCorWatt", (printed++ ? delim : ""));
854 } else if (do_rapl && rapl_joules) {
855 if (DO_BIC(BIC_Cor_J) && (do_rapl & RAPL_PER_CORE_ENERGY))
856 outp += sprintf(outp, "%sCor_J", (printed++ ? delim : ""));
859 for (mp = sys.cp; mp; mp = mp->next) {
860 if (mp->format == FORMAT_RAW) {
862 outp += sprintf(outp, "%s%18.18s", delim, mp->name);
864 outp += sprintf(outp, "%s%10.10s", delim, mp->name);
866 if ((mp->type == COUNTER_ITEMS) && sums_need_wide_columns)
867 outp += sprintf(outp, "%s%8s", delim, mp->name);
869 outp += sprintf(outp, "%s%s", delim, mp->name);
873 if (DO_BIC(BIC_PkgTmp))
874 outp += sprintf(outp, "%sPkgTmp", (printed++ ? delim : ""));
876 if (DO_BIC(BIC_GFX_rc6))
877 outp += sprintf(outp, "%sGFX%%rc6", (printed++ ? delim : ""));
879 if (DO_BIC(BIC_GFXMHz))
880 outp += sprintf(outp, "%sGFXMHz", (printed++ ? delim : ""));
882 if (DO_BIC(BIC_GFXACTMHz))
883 outp += sprintf(outp, "%sGFXAMHz", (printed++ ? delim : ""));
885 if (DO_BIC(BIC_Totl_c0))
886 outp += sprintf(outp, "%sTotl%%C0", (printed++ ? delim : ""));
887 if (DO_BIC(BIC_Any_c0))
888 outp += sprintf(outp, "%sAny%%C0", (printed++ ? delim : ""));
889 if (DO_BIC(BIC_GFX_c0))
890 outp += sprintf(outp, "%sGFX%%C0", (printed++ ? delim : ""));
891 if (DO_BIC(BIC_CPUGFX))
892 outp += sprintf(outp, "%sCPUGFX%%", (printed++ ? delim : ""));
894 if (DO_BIC(BIC_Pkgpc2))
895 outp += sprintf(outp, "%sPkg%%pc2", (printed++ ? delim : ""));
896 if (DO_BIC(BIC_Pkgpc3))
897 outp += sprintf(outp, "%sPkg%%pc3", (printed++ ? delim : ""));
898 if (DO_BIC(BIC_Pkgpc6))
899 outp += sprintf(outp, "%sPkg%%pc6", (printed++ ? delim : ""));
900 if (DO_BIC(BIC_Pkgpc7))
901 outp += sprintf(outp, "%sPkg%%pc7", (printed++ ? delim : ""));
902 if (DO_BIC(BIC_Pkgpc8))
903 outp += sprintf(outp, "%sPkg%%pc8", (printed++ ? delim : ""));
904 if (DO_BIC(BIC_Pkgpc9))
905 outp += sprintf(outp, "%sPkg%%pc9", (printed++ ? delim : ""));
906 if (DO_BIC(BIC_Pkgpc10))
907 outp += sprintf(outp, "%sPk%%pc10", (printed++ ? delim : ""));
908 if (DO_BIC(BIC_CPU_LPI))
909 outp += sprintf(outp, "%sCPU%%LPI", (printed++ ? delim : ""));
910 if (DO_BIC(BIC_SYS_LPI))
911 outp += sprintf(outp, "%sSYS%%LPI", (printed++ ? delim : ""));
913 if (do_rapl && !rapl_joules) {
914 if (DO_BIC(BIC_PkgWatt))
915 outp += sprintf(outp, "%sPkgWatt", (printed++ ? delim : ""));
916 if (DO_BIC(BIC_CorWatt) && !(do_rapl & RAPL_PER_CORE_ENERGY))
917 outp += sprintf(outp, "%sCorWatt", (printed++ ? delim : ""));
918 if (DO_BIC(BIC_GFXWatt))
919 outp += sprintf(outp, "%sGFXWatt", (printed++ ? delim : ""));
920 if (DO_BIC(BIC_RAMWatt))
921 outp += sprintf(outp, "%sRAMWatt", (printed++ ? delim : ""));
922 if (DO_BIC(BIC_PKG__))
923 outp += sprintf(outp, "%sPKG_%%", (printed++ ? delim : ""));
924 if (DO_BIC(BIC_RAM__))
925 outp += sprintf(outp, "%sRAM_%%", (printed++ ? delim : ""));
926 } else if (do_rapl && rapl_joules) {
927 if (DO_BIC(BIC_Pkg_J))
928 outp += sprintf(outp, "%sPkg_J", (printed++ ? delim : ""));
929 if (DO_BIC(BIC_Cor_J) && !(do_rapl & RAPL_PER_CORE_ENERGY))
930 outp += sprintf(outp, "%sCor_J", (printed++ ? delim : ""));
931 if (DO_BIC(BIC_GFX_J))
932 outp += sprintf(outp, "%sGFX_J", (printed++ ? delim : ""));
933 if (DO_BIC(BIC_RAM_J))
934 outp += sprintf(outp, "%sRAM_J", (printed++ ? delim : ""));
935 if (DO_BIC(BIC_PKG__))
936 outp += sprintf(outp, "%sPKG_%%", (printed++ ? delim : ""));
937 if (DO_BIC(BIC_RAM__))
938 outp += sprintf(outp, "%sRAM_%%", (printed++ ? delim : ""));
940 for (mp = sys.pp; mp; mp = mp->next) {
941 if (mp->format == FORMAT_RAW) {
943 outp += sprintf(outp, "%s%18.18s", delim, mp->name);
945 outp += sprintf(outp, "%s%10.10s", delim, mp->name);
947 if ((mp->type == COUNTER_ITEMS) && sums_need_wide_columns)
948 outp += sprintf(outp, "%s%8s", delim, mp->name);
950 outp += sprintf(outp, "%s%s", delim, mp->name);
954 outp += sprintf(outp, "\n");
957 int dump_counters(struct thread_data *t, struct core_data *c,
961 struct msr_counter *mp;
963 outp += sprintf(outp, "t %p, c %p, p %p\n", t, c, p);
966 outp += sprintf(outp, "CPU: %d flags 0x%x\n",
967 t->cpu_id, t->flags);
968 outp += sprintf(outp, "TSC: %016llX\n", t->tsc);
969 outp += sprintf(outp, "aperf: %016llX\n", t->aperf);
970 outp += sprintf(outp, "mperf: %016llX\n", t->mperf);
971 outp += sprintf(outp, "c1: %016llX\n", t->c1);
974 outp += sprintf(outp, "IPC: %lld\n", t->instr_count);
977 outp += sprintf(outp, "IRQ: %lld\n", t->irq_count);
979 outp += sprintf(outp, "SMI: %d\n", t->smi_count);
981 for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
982 outp += sprintf(outp, "tADDED [%d] msr0x%x: %08llX\n",
983 i, mp->msr_num, t->counter[i]);
988 outp += sprintf(outp, "core: %d\n", c->core_id);
989 outp += sprintf(outp, "c3: %016llX\n", c->c3);
990 outp += sprintf(outp, "c6: %016llX\n", c->c6);
991 outp += sprintf(outp, "c7: %016llX\n", c->c7);
992 outp += sprintf(outp, "DTS: %dC\n", c->core_temp_c);
993 outp += sprintf(outp, "Joules: %0X\n", c->core_energy);
995 for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
996 outp += sprintf(outp, "cADDED [%d] msr0x%x: %08llX\n",
997 i, mp->msr_num, c->counter[i]);
999 outp += sprintf(outp, "mc6_us: %016llX\n", c->mc6_us);
1003 outp += sprintf(outp, "package: %d\n", p->package_id);
1005 outp += sprintf(outp, "Weighted cores: %016llX\n", p->pkg_wtd_core_c0);
1006 outp += sprintf(outp, "Any cores: %016llX\n", p->pkg_any_core_c0);
1007 outp += sprintf(outp, "Any GFX: %016llX\n", p->pkg_any_gfxe_c0);
1008 outp += sprintf(outp, "CPU + GFX: %016llX\n", p->pkg_both_core_gfxe_c0);
1010 outp += sprintf(outp, "pc2: %016llX\n", p->pc2);
1011 if (DO_BIC(BIC_Pkgpc3))
1012 outp += sprintf(outp, "pc3: %016llX\n", p->pc3);
1013 if (DO_BIC(BIC_Pkgpc6))
1014 outp += sprintf(outp, "pc6: %016llX\n", p->pc6);
1015 if (DO_BIC(BIC_Pkgpc7))
1016 outp += sprintf(outp, "pc7: %016llX\n", p->pc7);
1017 outp += sprintf(outp, "pc8: %016llX\n", p->pc8);
1018 outp += sprintf(outp, "pc9: %016llX\n", p->pc9);
1019 outp += sprintf(outp, "pc10: %016llX\n", p->pc10);
1020 outp += sprintf(outp, "cpu_lpi: %016llX\n", p->cpu_lpi);
1021 outp += sprintf(outp, "sys_lpi: %016llX\n", p->sys_lpi);
1022 outp += sprintf(outp, "Joules PKG: %0llX\n", p->energy_pkg);
1023 outp += sprintf(outp, "Joules COR: %0llX\n", p->energy_cores);
1024 outp += sprintf(outp, "Joules GFX: %0llX\n", p->energy_gfx);
1025 outp += sprintf(outp, "Joules RAM: %0llX\n", p->energy_dram);
1026 outp += sprintf(outp, "Throttle PKG: %0llX\n",
1027 p->rapl_pkg_perf_status);
1028 outp += sprintf(outp, "Throttle RAM: %0llX\n",
1029 p->rapl_dram_perf_status);
1030 outp += sprintf(outp, "PTM: %dC\n", p->pkg_temp_c);
1032 for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
1033 outp += sprintf(outp, "pADDED [%d] msr0x%x: %08llX\n",
1034 i, mp->msr_num, p->counter[i]);
1038 outp += sprintf(outp, "\n");
1044 * column formatting convention & formats
1046 int format_counters(struct thread_data *t, struct core_data *c,
1049 double interval_float, tsc;
1052 struct msr_counter *mp;
1056 /* if showing only 1st thread in core and this isn't one, bail out */
1057 if (show_core_only && !(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
1060 /* if showing only 1st thread in pkg and this isn't one, bail out */
1061 if (show_pkg_only && !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
1064 /*if not summary line and --cpu is used */
1065 if ((t != &average.threads) &&
1066 (cpu_subset && !CPU_ISSET_S(t->cpu_id, cpu_subset_size, cpu_subset)))
1069 if (DO_BIC(BIC_USEC)) {
1070 /* on each row, print how many usec each timestamp took to gather */
1073 timersub(&t->tv_end, &t->tv_begin, &tv);
1074 outp += sprintf(outp, "%5ld\t", tv.tv_sec * 1000000 + tv.tv_usec);
1077 /* Time_Of_Day_Seconds: on each row, print sec.usec last timestamp taken */
1078 if (DO_BIC(BIC_TOD))
1079 outp += sprintf(outp, "%10ld.%06ld\t", t->tv_end.tv_sec, t->tv_end.tv_usec);
1081 interval_float = t->tv_delta.tv_sec + t->tv_delta.tv_usec/1000000.0;
1083 tsc = t->tsc * tsc_tweak;
1085 /* topo columns, print blanks on 1st (average) line */
1086 if (t == &average.threads) {
1087 if (DO_BIC(BIC_Package))
1088 outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
1089 if (DO_BIC(BIC_Die))
1090 outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
1091 if (DO_BIC(BIC_Node))
1092 outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
1093 if (DO_BIC(BIC_Core))
1094 outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
1095 if (DO_BIC(BIC_CPU))
1096 outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
1097 if (DO_BIC(BIC_APIC))
1098 outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
1099 if (DO_BIC(BIC_X2APIC))
1100 outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
1102 if (DO_BIC(BIC_Package)) {
1104 outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), p->package_id);
1106 outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
1108 if (DO_BIC(BIC_Die)) {
1110 outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), cpus[t->cpu_id].die_id);
1112 outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
1114 if (DO_BIC(BIC_Node)) {
1116 outp += sprintf(outp, "%s%d",
1117 (printed++ ? delim : ""),
1118 cpus[t->cpu_id].physical_node_id);
1120 outp += sprintf(outp, "%s-",
1121 (printed++ ? delim : ""));
1123 if (DO_BIC(BIC_Core)) {
1125 outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), c->core_id);
1127 outp += sprintf(outp, "%s-", (printed++ ? delim : ""));
1129 if (DO_BIC(BIC_CPU))
1130 outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), t->cpu_id);
1131 if (DO_BIC(BIC_APIC))
1132 outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), t->apic_id);
1133 if (DO_BIC(BIC_X2APIC))
1134 outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), t->x2apic_id);
1137 if (DO_BIC(BIC_Avg_MHz))
1138 outp += sprintf(outp, "%s%.0f", (printed++ ? delim : ""),
1139 1.0 / units * t->aperf / interval_float);
1141 if (DO_BIC(BIC_Busy))
1142 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * t->mperf/tsc);
1144 if (DO_BIC(BIC_Bzy_MHz)) {
1146 outp += sprintf(outp, "%s%.0f", (printed++ ? delim : ""), base_hz / units * t->aperf / t->mperf);
1148 outp += sprintf(outp, "%s%.0f", (printed++ ? delim : ""),
1149 tsc / units * t->aperf / t->mperf / interval_float);
1152 if (DO_BIC(BIC_TSC_MHz))
1153 outp += sprintf(outp, "%s%.0f", (printed++ ? delim : ""), 1.0 * t->tsc/units/interval_float);
1155 if (DO_BIC(BIC_IPC))
1156 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 1.0 * t->instr_count / t->aperf);
1159 if (DO_BIC(BIC_IRQ)) {
1160 if (sums_need_wide_columns)
1161 outp += sprintf(outp, "%s%8lld", (printed++ ? delim : ""), t->irq_count);
1163 outp += sprintf(outp, "%s%lld", (printed++ ? delim : ""), t->irq_count);
1167 if (DO_BIC(BIC_SMI))
1168 outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), t->smi_count);
1170 /* Added counters */
1171 for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
1172 if (mp->format == FORMAT_RAW) {
1173 if (mp->width == 32)
1174 outp += sprintf(outp, "%s0x%08x", (printed++ ? delim : ""), (unsigned int) t->counter[i]);
1176 outp += sprintf(outp, "%s0x%016llx", (printed++ ? delim : ""), t->counter[i]);
1177 } else if (mp->format == FORMAT_DELTA) {
1178 if ((mp->type == COUNTER_ITEMS) && sums_need_wide_columns)
1179 outp += sprintf(outp, "%s%8lld", (printed++ ? delim : ""), t->counter[i]);
1181 outp += sprintf(outp, "%s%lld", (printed++ ? delim : ""), t->counter[i]);
1182 } else if (mp->format == FORMAT_PERCENT) {
1183 if (mp->type == COUNTER_USEC)
1184 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), t->counter[i]/interval_float/10000);
1186 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * t->counter[i]/tsc);
1191 if (DO_BIC(BIC_CPU_c1))
1192 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * t->c1/tsc);
1195 /* print per-core data only for 1st thread in core */
1196 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
1199 if (DO_BIC(BIC_CPU_c3))
1200 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * c->c3/tsc);
1201 if (DO_BIC(BIC_CPU_c6))
1202 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * c->c6/tsc);
1203 if (DO_BIC(BIC_CPU_c7))
1204 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * c->c7/tsc);
1207 if (DO_BIC(BIC_Mod_c6))
1208 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * c->mc6_us / tsc);
1210 if (DO_BIC(BIC_CoreTmp))
1211 outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), c->core_temp_c);
1213 for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
1214 if (mp->format == FORMAT_RAW) {
1215 if (mp->width == 32)
1216 outp += sprintf(outp, "%s0x%08x", (printed++ ? delim : ""), (unsigned int) c->counter[i]);
1218 outp += sprintf(outp, "%s0x%016llx", (printed++ ? delim : ""), c->counter[i]);
1219 } else if (mp->format == FORMAT_DELTA) {
1220 if ((mp->type == COUNTER_ITEMS) && sums_need_wide_columns)
1221 outp += sprintf(outp, "%s%8lld", (printed++ ? delim : ""), c->counter[i]);
1223 outp += sprintf(outp, "%s%lld", (printed++ ? delim : ""), c->counter[i]);
1224 } else if (mp->format == FORMAT_PERCENT) {
1225 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * c->counter[i]/tsc);
1231 if (DO_BIC(BIC_CorWatt) && (do_rapl & RAPL_PER_CORE_ENERGY))
1232 outp += sprintf(outp, fmt8, (printed++ ? delim : ""), c->core_energy * rapl_energy_units / interval_float);
1233 if (DO_BIC(BIC_Cor_J) && (do_rapl & RAPL_PER_CORE_ENERGY))
1234 outp += sprintf(outp, fmt8, (printed++ ? delim : ""), c->core_energy * rapl_energy_units);
1236 /* print per-package data only for 1st core in package */
1237 if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
1241 if (DO_BIC(BIC_PkgTmp))
1242 outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), p->pkg_temp_c);
1245 if (DO_BIC(BIC_GFX_rc6)) {
1246 if (p->gfx_rc6_ms == -1) { /* detect GFX counter reset */
1247 outp += sprintf(outp, "%s**.**", (printed++ ? delim : ""));
1249 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""),
1250 p->gfx_rc6_ms / 10.0 / interval_float);
1255 if (DO_BIC(BIC_GFXMHz))
1256 outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), p->gfx_mhz);
1259 if (DO_BIC(BIC_GFXACTMHz))
1260 outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), p->gfx_act_mhz);
1262 /* Totl%C0, Any%C0 GFX%C0 CPUGFX% */
1263 if (DO_BIC(BIC_Totl_c0))
1264 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pkg_wtd_core_c0/tsc);
1265 if (DO_BIC(BIC_Any_c0))
1266 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pkg_any_core_c0/tsc);
1267 if (DO_BIC(BIC_GFX_c0))
1268 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pkg_any_gfxe_c0/tsc);
1269 if (DO_BIC(BIC_CPUGFX))
1270 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pkg_both_core_gfxe_c0/tsc);
1272 if (DO_BIC(BIC_Pkgpc2))
1273 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pc2/tsc);
1274 if (DO_BIC(BIC_Pkgpc3))
1275 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pc3/tsc);
1276 if (DO_BIC(BIC_Pkgpc6))
1277 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pc6/tsc);
1278 if (DO_BIC(BIC_Pkgpc7))
1279 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pc7/tsc);
1280 if (DO_BIC(BIC_Pkgpc8))
1281 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pc8/tsc);
1282 if (DO_BIC(BIC_Pkgpc9))
1283 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pc9/tsc);
1284 if (DO_BIC(BIC_Pkgpc10))
1285 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->pc10/tsc);
1287 if (DO_BIC(BIC_CPU_LPI))
1288 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->cpu_lpi / 1000000.0 / interval_float);
1289 if (DO_BIC(BIC_SYS_LPI))
1290 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->sys_lpi / 1000000.0 / interval_float);
1292 if (DO_BIC(BIC_PkgWatt))
1293 outp += sprintf(outp, fmt8, (printed++ ? delim : ""), p->energy_pkg * rapl_energy_units / interval_float);
1294 if (DO_BIC(BIC_CorWatt) && !(do_rapl & RAPL_PER_CORE_ENERGY))
1295 outp += sprintf(outp, fmt8, (printed++ ? delim : ""), p->energy_cores * rapl_energy_units / interval_float);
1296 if (DO_BIC(BIC_GFXWatt))
1297 outp += sprintf(outp, fmt8, (printed++ ? delim : ""), p->energy_gfx * rapl_energy_units / interval_float);
1298 if (DO_BIC(BIC_RAMWatt))
1299 outp += sprintf(outp, fmt8, (printed++ ? delim : ""), p->energy_dram * rapl_dram_energy_units / interval_float);
1300 if (DO_BIC(BIC_Pkg_J))
1301 outp += sprintf(outp, fmt8, (printed++ ? delim : ""), p->energy_pkg * rapl_energy_units);
1302 if (DO_BIC(BIC_Cor_J) && !(do_rapl & RAPL_PER_CORE_ENERGY))
1303 outp += sprintf(outp, fmt8, (printed++ ? delim : ""), p->energy_cores * rapl_energy_units);
1304 if (DO_BIC(BIC_GFX_J))
1305 outp += sprintf(outp, fmt8, (printed++ ? delim : ""), p->energy_gfx * rapl_energy_units);
1306 if (DO_BIC(BIC_RAM_J))
1307 outp += sprintf(outp, fmt8, (printed++ ? delim : ""), p->energy_dram * rapl_dram_energy_units);
1308 if (DO_BIC(BIC_PKG__))
1309 outp += sprintf(outp, fmt8, (printed++ ? delim : ""), 100.0 * p->rapl_pkg_perf_status * rapl_time_units / interval_float);
1310 if (DO_BIC(BIC_RAM__))
1311 outp += sprintf(outp, fmt8, (printed++ ? delim : ""), 100.0 * p->rapl_dram_perf_status * rapl_time_units / interval_float);
1313 for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
1314 if (mp->format == FORMAT_RAW) {
1315 if (mp->width == 32)
1316 outp += sprintf(outp, "%s0x%08x", (printed++ ? delim : ""), (unsigned int) p->counter[i]);
1318 outp += sprintf(outp, "%s0x%016llx", (printed++ ? delim : ""), p->counter[i]);
1319 } else if (mp->format == FORMAT_DELTA) {
1320 if ((mp->type == COUNTER_ITEMS) && sums_need_wide_columns)
1321 outp += sprintf(outp, "%s%8lld", (printed++ ? delim : ""), p->counter[i]);
1323 outp += sprintf(outp, "%s%lld", (printed++ ? delim : ""), p->counter[i]);
1324 } else if (mp->format == FORMAT_PERCENT) {
1325 outp += sprintf(outp, "%s%.2f", (printed++ ? delim : ""), 100.0 * p->counter[i]/tsc);
1330 if (*(outp - 1) != '\n')
1331 outp += sprintf(outp, "\n");
1336 void flush_output_stdout(void)
1345 fputs(output_buffer, filep);
1348 outp = output_buffer;
1350 void flush_output_stderr(void)
1352 fputs(output_buffer, outf);
1354 outp = output_buffer;
1356 void format_all_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1360 if (!printed || !summary_only)
1363 format_counters(&average.threads, &average.cores, &average.packages);
1370 for_all_cpus(format_counters, t, c, p);
1373 #define DELTA_WRAP32(new, old) \
1374 old = ((((unsigned long long)new << 32) - ((unsigned long long)old << 32)) >> 32);
1377 delta_package(struct pkg_data *new, struct pkg_data *old)
1380 struct msr_counter *mp;
1383 if (DO_BIC(BIC_Totl_c0))
1384 old->pkg_wtd_core_c0 = new->pkg_wtd_core_c0 - old->pkg_wtd_core_c0;
1385 if (DO_BIC(BIC_Any_c0))
1386 old->pkg_any_core_c0 = new->pkg_any_core_c0 - old->pkg_any_core_c0;
1387 if (DO_BIC(BIC_GFX_c0))
1388 old->pkg_any_gfxe_c0 = new->pkg_any_gfxe_c0 - old->pkg_any_gfxe_c0;
1389 if (DO_BIC(BIC_CPUGFX))
1390 old->pkg_both_core_gfxe_c0 = new->pkg_both_core_gfxe_c0 - old->pkg_both_core_gfxe_c0;
1392 old->pc2 = new->pc2 - old->pc2;
1393 if (DO_BIC(BIC_Pkgpc3))
1394 old->pc3 = new->pc3 - old->pc3;
1395 if (DO_BIC(BIC_Pkgpc6))
1396 old->pc6 = new->pc6 - old->pc6;
1397 if (DO_BIC(BIC_Pkgpc7))
1398 old->pc7 = new->pc7 - old->pc7;
1399 old->pc8 = new->pc8 - old->pc8;
1400 old->pc9 = new->pc9 - old->pc9;
1401 old->pc10 = new->pc10 - old->pc10;
1402 old->cpu_lpi = new->cpu_lpi - old->cpu_lpi;
1403 old->sys_lpi = new->sys_lpi - old->sys_lpi;
1404 old->pkg_temp_c = new->pkg_temp_c;
1406 /* flag an error when rc6 counter resets/wraps */
1407 if (old->gfx_rc6_ms > new->gfx_rc6_ms)
1408 old->gfx_rc6_ms = -1;
1410 old->gfx_rc6_ms = new->gfx_rc6_ms - old->gfx_rc6_ms;
1412 old->gfx_mhz = new->gfx_mhz;
1413 old->gfx_act_mhz = new->gfx_act_mhz;
1415 old->energy_pkg = new->energy_pkg - old->energy_pkg;
1416 old->energy_cores = new->energy_cores - old->energy_cores;
1417 old->energy_gfx = new->energy_gfx - old->energy_gfx;
1418 old->energy_dram = new->energy_dram - old->energy_dram;
1419 old->rapl_pkg_perf_status = new->rapl_pkg_perf_status - old->rapl_pkg_perf_status;
1420 old->rapl_dram_perf_status = new->rapl_dram_perf_status - old->rapl_dram_perf_status;
1422 for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
1423 if (mp->format == FORMAT_RAW)
1424 old->counter[i] = new->counter[i];
1426 old->counter[i] = new->counter[i] - old->counter[i];
1433 delta_core(struct core_data *new, struct core_data *old)
1436 struct msr_counter *mp;
1438 old->c3 = new->c3 - old->c3;
1439 old->c6 = new->c6 - old->c6;
1440 old->c7 = new->c7 - old->c7;
1441 old->core_temp_c = new->core_temp_c;
1442 old->mc6_us = new->mc6_us - old->mc6_us;
1444 DELTA_WRAP32(new->core_energy, old->core_energy);
1446 for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
1447 if (mp->format == FORMAT_RAW)
1448 old->counter[i] = new->counter[i];
1450 old->counter[i] = new->counter[i] - old->counter[i];
1454 int soft_c1_residency_display(int bic)
1456 if (!DO_BIC(BIC_CPU_c1) || use_c1_residency_msr)
1459 return DO_BIC_READ(bic);
1466 delta_thread(struct thread_data *new, struct thread_data *old,
1467 struct core_data *core_delta)
1470 struct msr_counter *mp;
1472 /* we run cpuid just the 1st time, copy the results */
1473 if (DO_BIC(BIC_APIC))
1474 new->apic_id = old->apic_id;
1475 if (DO_BIC(BIC_X2APIC))
1476 new->x2apic_id = old->x2apic_id;
1479 * the timestamps from start of measurement interval are in "old"
1480 * the timestamp from end of measurement interval are in "new"
1481 * over-write old w/ new so we can print end of interval values
1484 timersub(&new->tv_begin, &old->tv_begin, &old->tv_delta);
1485 old->tv_begin = new->tv_begin;
1486 old->tv_end = new->tv_end;
1488 old->tsc = new->tsc - old->tsc;
1490 /* check for TSC < 1 Mcycles over interval */
1491 if (old->tsc < (1000 * 1000))
1492 errx(-3, "Insanely slow TSC rate, TSC stops in idle?\n"
1493 "You can disable all c-states by booting with \"idle=poll\"\n"
1494 "or just the deep ones with \"processor.max_cstate=1\"");
1496 old->c1 = new->c1 - old->c1;
1498 if (DO_BIC(BIC_Avg_MHz) || DO_BIC(BIC_Busy) || DO_BIC(BIC_Bzy_MHz) ||
1499 soft_c1_residency_display(BIC_Avg_MHz)) {
1500 if ((new->aperf > old->aperf) && (new->mperf > old->mperf)) {
1501 old->aperf = new->aperf - old->aperf;
1502 old->mperf = new->mperf - old->mperf;
1509 if (use_c1_residency_msr) {
1511 * Some models have a dedicated C1 residency MSR,
1512 * which should be more accurate than the derivation below.
1516 * As counter collection is not atomic,
1517 * it is possible for mperf's non-halted cycles + idle states
1518 * to exceed TSC's all cycles: show c1 = 0% in that case.
1520 if ((old->mperf + core_delta->c3 + core_delta->c6 + core_delta->c7) > (old->tsc * tsc_tweak))
1523 /* normal case, derive c1 */
1524 old->c1 = (old->tsc * tsc_tweak) - old->mperf - core_delta->c3
1525 - core_delta->c6 - core_delta->c7;
1529 if (old->mperf == 0) {
1531 fprintf(outf, "cpu%d MPERF 0!\n", old->cpu_id);
1532 old->mperf = 1; /* divide by 0 protection */
1535 if (DO_BIC(BIC_IPC))
1536 old->instr_count = new->instr_count - old->instr_count;
1538 if (DO_BIC(BIC_IRQ))
1539 old->irq_count = new->irq_count - old->irq_count;
1541 if (DO_BIC(BIC_SMI))
1542 old->smi_count = new->smi_count - old->smi_count;
1544 for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
1545 if (mp->format == FORMAT_RAW)
1546 old->counter[i] = new->counter[i];
1548 old->counter[i] = new->counter[i] - old->counter[i];
1553 int delta_cpu(struct thread_data *t, struct core_data *c,
1554 struct pkg_data *p, struct thread_data *t2,
1555 struct core_data *c2, struct pkg_data *p2)
1559 /* calculate core delta only for 1st thread in core */
1560 if (t->flags & CPU_IS_FIRST_THREAD_IN_CORE)
1563 /* always calculate thread delta */
1564 retval = delta_thread(t, t2, c2); /* c2 is core delta */
1568 /* calculate package delta only for 1st core in package */
1569 if (t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)
1570 retval = delta_package(p, p2);
1575 void clear_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1578 struct msr_counter *mp;
1580 t->tv_begin.tv_sec = 0;
1581 t->tv_begin.tv_usec = 0;
1582 t->tv_end.tv_sec = 0;
1583 t->tv_end.tv_usec = 0;
1584 t->tv_delta.tv_sec = 0;
1585 t->tv_delta.tv_usec = 0;
1597 /* tells format_counters to dump all fields from this set */
1598 t->flags = CPU_IS_FIRST_THREAD_IN_CORE | CPU_IS_FIRST_CORE_IN_PACKAGE;
1607 p->pkg_wtd_core_c0 = 0;
1608 p->pkg_any_core_c0 = 0;
1609 p->pkg_any_gfxe_c0 = 0;
1610 p->pkg_both_core_gfxe_c0 = 0;
1613 if (DO_BIC(BIC_Pkgpc3))
1615 if (DO_BIC(BIC_Pkgpc6))
1617 if (DO_BIC(BIC_Pkgpc7))
1627 p->energy_cores = 0;
1629 p->rapl_pkg_perf_status = 0;
1630 p->rapl_dram_perf_status = 0;
1636 for (i = 0, mp = sys.tp; mp; i++, mp = mp->next)
1639 for (i = 0, mp = sys.cp; mp; i++, mp = mp->next)
1642 for (i = 0, mp = sys.pp; mp; i++, mp = mp->next)
1645 int sum_counters(struct thread_data *t, struct core_data *c,
1649 struct msr_counter *mp;
1651 /* copy un-changing apic_id's */
1652 if (DO_BIC(BIC_APIC))
1653 average.threads.apic_id = t->apic_id;
1654 if (DO_BIC(BIC_X2APIC))
1655 average.threads.x2apic_id = t->x2apic_id;
1657 /* remember first tv_begin */
1658 if (average.threads.tv_begin.tv_sec == 0)
1659 average.threads.tv_begin = t->tv_begin;
1661 /* remember last tv_end */
1662 average.threads.tv_end = t->tv_end;
1664 average.threads.tsc += t->tsc;
1665 average.threads.aperf += t->aperf;
1666 average.threads.mperf += t->mperf;
1667 average.threads.c1 += t->c1;
1669 average.threads.instr_count += t->instr_count;
1671 average.threads.irq_count += t->irq_count;
1672 average.threads.smi_count += t->smi_count;
1674 for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
1675 if (mp->format == FORMAT_RAW)
1677 average.threads.counter[i] += t->counter[i];
1680 /* sum per-core values only for 1st thread in core */
1681 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
1684 average.cores.c3 += c->c3;
1685 average.cores.c6 += c->c6;
1686 average.cores.c7 += c->c7;
1687 average.cores.mc6_us += c->mc6_us;
1689 average.cores.core_temp_c = MAX(average.cores.core_temp_c, c->core_temp_c);
1691 average.cores.core_energy += c->core_energy;
1693 for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
1694 if (mp->format == FORMAT_RAW)
1696 average.cores.counter[i] += c->counter[i];
1699 /* sum per-pkg values only for 1st core in pkg */
1700 if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
1703 if (DO_BIC(BIC_Totl_c0))
1704 average.packages.pkg_wtd_core_c0 += p->pkg_wtd_core_c0;
1705 if (DO_BIC(BIC_Any_c0))
1706 average.packages.pkg_any_core_c0 += p->pkg_any_core_c0;
1707 if (DO_BIC(BIC_GFX_c0))
1708 average.packages.pkg_any_gfxe_c0 += p->pkg_any_gfxe_c0;
1709 if (DO_BIC(BIC_CPUGFX))
1710 average.packages.pkg_both_core_gfxe_c0 += p->pkg_both_core_gfxe_c0;
1712 average.packages.pc2 += p->pc2;
1713 if (DO_BIC(BIC_Pkgpc3))
1714 average.packages.pc3 += p->pc3;
1715 if (DO_BIC(BIC_Pkgpc6))
1716 average.packages.pc6 += p->pc6;
1717 if (DO_BIC(BIC_Pkgpc7))
1718 average.packages.pc7 += p->pc7;
1719 average.packages.pc8 += p->pc8;
1720 average.packages.pc9 += p->pc9;
1721 average.packages.pc10 += p->pc10;
1723 average.packages.cpu_lpi = p->cpu_lpi;
1724 average.packages.sys_lpi = p->sys_lpi;
1726 average.packages.energy_pkg += p->energy_pkg;
1727 average.packages.energy_dram += p->energy_dram;
1728 average.packages.energy_cores += p->energy_cores;
1729 average.packages.energy_gfx += p->energy_gfx;
1731 average.packages.gfx_rc6_ms = p->gfx_rc6_ms;
1732 average.packages.gfx_mhz = p->gfx_mhz;
1733 average.packages.gfx_act_mhz = p->gfx_act_mhz;
1735 average.packages.pkg_temp_c = MAX(average.packages.pkg_temp_c, p->pkg_temp_c);
1737 average.packages.rapl_pkg_perf_status += p->rapl_pkg_perf_status;
1738 average.packages.rapl_dram_perf_status += p->rapl_dram_perf_status;
1740 for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
1741 if (mp->format == FORMAT_RAW)
1743 average.packages.counter[i] += p->counter[i];
1748 * sum the counters for all cpus in the system
1749 * compute the weighted average
1751 void compute_average(struct thread_data *t, struct core_data *c,
1755 struct msr_counter *mp;
1757 clear_counters(&average.threads, &average.cores, &average.packages);
1759 for_all_cpus(sum_counters, t, c, p);
1761 /* Use the global time delta for the average. */
1762 average.threads.tv_delta = tv_delta;
1764 average.threads.tsc /= topo.num_cpus;
1765 average.threads.aperf /= topo.num_cpus;
1766 average.threads.mperf /= topo.num_cpus;
1767 average.threads.instr_count /= topo.num_cpus;
1768 average.threads.c1 /= topo.num_cpus;
1770 if (average.threads.irq_count > 9999999)
1771 sums_need_wide_columns = 1;
1773 average.cores.c3 /= topo.num_cores;
1774 average.cores.c6 /= topo.num_cores;
1775 average.cores.c7 /= topo.num_cores;
1776 average.cores.mc6_us /= topo.num_cores;
1778 if (DO_BIC(BIC_Totl_c0))
1779 average.packages.pkg_wtd_core_c0 /= topo.num_packages;
1780 if (DO_BIC(BIC_Any_c0))
1781 average.packages.pkg_any_core_c0 /= topo.num_packages;
1782 if (DO_BIC(BIC_GFX_c0))
1783 average.packages.pkg_any_gfxe_c0 /= topo.num_packages;
1784 if (DO_BIC(BIC_CPUGFX))
1785 average.packages.pkg_both_core_gfxe_c0 /= topo.num_packages;
1787 average.packages.pc2 /= topo.num_packages;
1788 if (DO_BIC(BIC_Pkgpc3))
1789 average.packages.pc3 /= topo.num_packages;
1790 if (DO_BIC(BIC_Pkgpc6))
1791 average.packages.pc6 /= topo.num_packages;
1792 if (DO_BIC(BIC_Pkgpc7))
1793 average.packages.pc7 /= topo.num_packages;
1795 average.packages.pc8 /= topo.num_packages;
1796 average.packages.pc9 /= topo.num_packages;
1797 average.packages.pc10 /= topo.num_packages;
1799 for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
1800 if (mp->format == FORMAT_RAW)
1802 if (mp->type == COUNTER_ITEMS) {
1803 if (average.threads.counter[i] > 9999999)
1804 sums_need_wide_columns = 1;
1807 average.threads.counter[i] /= topo.num_cpus;
1809 for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
1810 if (mp->format == FORMAT_RAW)
1812 if (mp->type == COUNTER_ITEMS) {
1813 if (average.cores.counter[i] > 9999999)
1814 sums_need_wide_columns = 1;
1816 average.cores.counter[i] /= topo.num_cores;
1818 for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
1819 if (mp->format == FORMAT_RAW)
1821 if (mp->type == COUNTER_ITEMS) {
1822 if (average.packages.counter[i] > 9999999)
1823 sums_need_wide_columns = 1;
1825 average.packages.counter[i] /= topo.num_packages;
1829 static unsigned long long rdtsc(void)
1831 unsigned int low, high;
1833 asm volatile("rdtsc" : "=a" (low), "=d" (high));
1835 return low | ((unsigned long long)high) << 32;
1839 * Open a file, and exit on failure
1841 FILE *fopen_or_die(const char *path, const char *mode)
1843 FILE *filep = fopen(path, mode);
1846 err(1, "%s: open failed", path);
1850 * snapshot_sysfs_counter()
1852 * return snapshot of given counter
1854 unsigned long long snapshot_sysfs_counter(char *path)
1858 unsigned long long counter;
1860 fp = fopen_or_die(path, "r");
1862 retval = fscanf(fp, "%lld", &counter);
1864 err(1, "snapshot_sysfs_counter(%s)", path);
1871 int get_mp(int cpu, struct msr_counter *mp, unsigned long long *counterp)
1873 if (mp->msr_num != 0) {
1874 if (get_msr(cpu, mp->msr_num, counterp))
1877 char path[128 + PATH_BYTES];
1879 if (mp->flags & SYSFS_PERCPU) {
1880 sprintf(path, "/sys/devices/system/cpu/cpu%d/%s",
1883 *counterp = snapshot_sysfs_counter(path);
1885 *counterp = snapshot_sysfs_counter(mp->path);
1892 int get_epb(int cpu)
1894 char path[128 + PATH_BYTES];
1895 unsigned long long msr;
1899 sprintf(path, "/sys/devices/system/cpu/cpu%d/power/energy_perf_bias", cpu);
1901 fp = fopen(path, "r");
1905 ret = fscanf(fp, "%d", &epb);
1907 err(1, "%s(%s)", __func__, path);
1914 get_msr(cpu, MSR_IA32_ENERGY_PERF_BIAS, &msr);
1919 void get_apic_id(struct thread_data *t)
1921 unsigned int eax, ebx, ecx, edx;
1923 if (DO_BIC(BIC_APIC)) {
1924 eax = ebx = ecx = edx = 0;
1925 __cpuid(1, eax, ebx, ecx, edx);
1927 t->apic_id = (ebx >> 24) & 0xff;
1930 if (!DO_BIC(BIC_X2APIC))
1933 if (authentic_amd || hygon_genuine) {
1934 unsigned int topology_extensions;
1936 if (max_extended_level < 0x8000001e)
1939 eax = ebx = ecx = edx = 0;
1940 __cpuid(0x80000001, eax, ebx, ecx, edx);
1941 topology_extensions = ecx & (1 << 22);
1943 if (topology_extensions == 0)
1946 eax = ebx = ecx = edx = 0;
1947 __cpuid(0x8000001e, eax, ebx, ecx, edx);
1956 if (max_level < 0xb)
1960 __cpuid(0xb, eax, ebx, ecx, edx);
1963 if (debug && (t->apic_id != (t->x2apic_id & 0xff)))
1964 fprintf(outf, "cpu%d: BIOS BUG: apic 0x%x x2apic 0x%x\n",
1965 t->cpu_id, t->apic_id, t->x2apic_id);
1971 * acquire and record local counters for that cpu
1973 int get_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
1975 int cpu = t->cpu_id;
1976 unsigned long long msr;
1977 int aperf_mperf_retry_count = 0;
1978 struct msr_counter *mp;
1981 if (cpu_migrate(cpu)) {
1982 fprintf(outf, "get_counters: Could not migrate to CPU %d\n", cpu);
1986 gettimeofday(&t->tv_begin, (struct timezone *)NULL);
1988 if (first_counter_read)
1991 t->tsc = rdtsc(); /* we are running on local CPU of interest */
1993 if (DO_BIC(BIC_Avg_MHz) || DO_BIC(BIC_Busy) || DO_BIC(BIC_Bzy_MHz) ||
1994 soft_c1_residency_display(BIC_Avg_MHz)) {
1995 unsigned long long tsc_before, tsc_between, tsc_after, aperf_time, mperf_time;
1998 * The TSC, APERF and MPERF must be read together for
1999 * APERF/MPERF and MPERF/TSC to give accurate results.
2001 * Unfortunately, APERF and MPERF are read by
2002 * individual system call, so delays may occur
2003 * between them. If the time to read them
2004 * varies by a large amount, we re-read them.
2008 * This initial dummy APERF read has been seen to
2009 * reduce jitter in the subsequent reads.
2012 if (get_msr(cpu, MSR_IA32_APERF, &t->aperf))
2015 t->tsc = rdtsc(); /* re-read close to APERF */
2017 tsc_before = t->tsc;
2019 if (get_msr(cpu, MSR_IA32_APERF, &t->aperf))
2022 tsc_between = rdtsc();
2024 if (get_msr(cpu, MSR_IA32_MPERF, &t->mperf))
2027 tsc_after = rdtsc();
2029 aperf_time = tsc_between - tsc_before;
2030 mperf_time = tsc_after - tsc_between;
2033 * If the system call latency to read APERF and MPERF
2034 * differ by more than 2x, then try again.
2036 if ((aperf_time > (2 * mperf_time)) || (mperf_time > (2 * aperf_time))) {
2037 aperf_mperf_retry_count++;
2038 if (aperf_mperf_retry_count < 5)
2041 warnx("cpu%d jitter %lld %lld",
2042 cpu, aperf_time, mperf_time);
2044 aperf_mperf_retry_count = 0;
2046 t->aperf = t->aperf * aperf_mperf_multiplier;
2047 t->mperf = t->mperf * aperf_mperf_multiplier;
2050 if (DO_BIC(BIC_IPC))
2051 if (read(get_instr_count_fd(cpu), &t->instr_count, sizeof(long long)) != sizeof(long long))
2054 if (DO_BIC(BIC_IRQ))
2055 t->irq_count = irqs_per_cpu[cpu];
2056 if (DO_BIC(BIC_SMI)) {
2057 if (get_msr(cpu, MSR_SMI_COUNT, &msr))
2059 t->smi_count = msr & 0xFFFFFFFF;
2061 if (DO_BIC(BIC_CPU_c1) && use_c1_residency_msr) {
2062 if (get_msr(cpu, MSR_CORE_C1_RES, &t->c1))
2066 for (i = 0, mp = sys.tp; mp; i++, mp = mp->next) {
2067 if (get_mp(cpu, mp, &t->counter[i]))
2071 /* collect core counters only for 1st thread in core */
2072 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
2075 if (DO_BIC(BIC_CPU_c3) || soft_c1_residency_display(BIC_CPU_c3)) {
2076 if (get_msr(cpu, MSR_CORE_C3_RESIDENCY, &c->c3))
2080 if ((DO_BIC(BIC_CPU_c6) || soft_c1_residency_display(BIC_CPU_c6)) && !do_knl_cstates) {
2081 if (get_msr(cpu, MSR_CORE_C6_RESIDENCY, &c->c6))
2083 } else if (do_knl_cstates || soft_c1_residency_display(BIC_CPU_c6)) {
2084 if (get_msr(cpu, MSR_KNL_CORE_C6_RESIDENCY, &c->c6))
2088 if (DO_BIC(BIC_CPU_c7) || soft_c1_residency_display(BIC_CPU_c7))
2089 if (get_msr(cpu, MSR_CORE_C7_RESIDENCY, &c->c7))
2092 if (DO_BIC(BIC_Mod_c6))
2093 if (get_msr(cpu, MSR_MODULE_C6_RES_MS, &c->mc6_us))
2096 if (DO_BIC(BIC_CoreTmp)) {
2097 if (get_msr(cpu, MSR_IA32_THERM_STATUS, &msr))
2099 c->core_temp_c = tcc_activation_temp - ((msr >> 16) & 0x7F);
2102 if (do_rapl & RAPL_AMD_F17H) {
2103 if (get_msr(cpu, MSR_CORE_ENERGY_STAT, &msr))
2105 c->core_energy = msr & 0xFFFFFFFF;
2108 for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
2109 if (get_mp(cpu, mp, &c->counter[i]))
2113 /* collect package counters only for 1st core in package */
2114 if (!(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
2117 if (DO_BIC(BIC_Totl_c0)) {
2118 if (get_msr(cpu, MSR_PKG_WEIGHTED_CORE_C0_RES, &p->pkg_wtd_core_c0))
2121 if (DO_BIC(BIC_Any_c0)) {
2122 if (get_msr(cpu, MSR_PKG_ANY_CORE_C0_RES, &p->pkg_any_core_c0))
2125 if (DO_BIC(BIC_GFX_c0)) {
2126 if (get_msr(cpu, MSR_PKG_ANY_GFXE_C0_RES, &p->pkg_any_gfxe_c0))
2129 if (DO_BIC(BIC_CPUGFX)) {
2130 if (get_msr(cpu, MSR_PKG_BOTH_CORE_GFXE_C0_RES, &p->pkg_both_core_gfxe_c0))
2133 if (DO_BIC(BIC_Pkgpc3))
2134 if (get_msr(cpu, MSR_PKG_C3_RESIDENCY, &p->pc3))
2136 if (DO_BIC(BIC_Pkgpc6)) {
2137 if (do_slm_cstates) {
2138 if (get_msr(cpu, MSR_ATOM_PKG_C6_RESIDENCY, &p->pc6))
2141 if (get_msr(cpu, MSR_PKG_C6_RESIDENCY, &p->pc6))
2146 if (DO_BIC(BIC_Pkgpc2))
2147 if (get_msr(cpu, MSR_PKG_C2_RESIDENCY, &p->pc2))
2149 if (DO_BIC(BIC_Pkgpc7))
2150 if (get_msr(cpu, MSR_PKG_C7_RESIDENCY, &p->pc7))
2152 if (DO_BIC(BIC_Pkgpc8))
2153 if (get_msr(cpu, MSR_PKG_C8_RESIDENCY, &p->pc8))
2155 if (DO_BIC(BIC_Pkgpc9))
2156 if (get_msr(cpu, MSR_PKG_C9_RESIDENCY, &p->pc9))
2158 if (DO_BIC(BIC_Pkgpc10))
2159 if (get_msr(cpu, MSR_PKG_C10_RESIDENCY, &p->pc10))
2162 if (DO_BIC(BIC_CPU_LPI))
2163 p->cpu_lpi = cpuidle_cur_cpu_lpi_us;
2164 if (DO_BIC(BIC_SYS_LPI))
2165 p->sys_lpi = cpuidle_cur_sys_lpi_us;
2167 if (do_rapl & RAPL_PKG) {
2168 if (get_msr_sum(cpu, MSR_PKG_ENERGY_STATUS, &msr))
2170 p->energy_pkg = msr;
2172 if (do_rapl & RAPL_CORES_ENERGY_STATUS) {
2173 if (get_msr_sum(cpu, MSR_PP0_ENERGY_STATUS, &msr))
2175 p->energy_cores = msr;
2177 if (do_rapl & RAPL_DRAM) {
2178 if (get_msr_sum(cpu, MSR_DRAM_ENERGY_STATUS, &msr))
2180 p->energy_dram = msr;
2182 if (do_rapl & RAPL_GFX) {
2183 if (get_msr_sum(cpu, MSR_PP1_ENERGY_STATUS, &msr))
2185 p->energy_gfx = msr;
2187 if (do_rapl & RAPL_PKG_PERF_STATUS) {
2188 if (get_msr_sum(cpu, MSR_PKG_PERF_STATUS, &msr))
2190 p->rapl_pkg_perf_status = msr;
2192 if (do_rapl & RAPL_DRAM_PERF_STATUS) {
2193 if (get_msr_sum(cpu, MSR_DRAM_PERF_STATUS, &msr))
2195 p->rapl_dram_perf_status = msr;
2197 if (do_rapl & RAPL_AMD_F17H) {
2198 if (get_msr_sum(cpu, MSR_PKG_ENERGY_STAT, &msr))
2200 p->energy_pkg = msr;
2202 if (DO_BIC(BIC_PkgTmp)) {
2203 if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_STATUS, &msr))
2205 p->pkg_temp_c = tcc_activation_temp - ((msr >> 16) & 0x7F);
2208 if (DO_BIC(BIC_GFX_rc6))
2209 p->gfx_rc6_ms = gfx_cur_rc6_ms;
2211 if (DO_BIC(BIC_GFXMHz))
2212 p->gfx_mhz = gfx_cur_mhz;
2214 if (DO_BIC(BIC_GFXACTMHz))
2215 p->gfx_act_mhz = gfx_act_mhz;
2217 for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
2218 if (get_mp(cpu, mp, &p->counter[i]))
2222 gettimeofday(&t->tv_end, (struct timezone *)NULL);
2228 * MSR_PKG_CST_CONFIG_CONTROL decoding for pkg_cstate_limit:
2229 * If you change the values, note they are used both in comparisons
2230 * (>= PCL__7) and to index pkg_cstate_limit_strings[].
2233 #define PCLUKN 0 /* Unknown */
2234 #define PCLRSV 1 /* Reserved */
2235 #define PCL__0 2 /* PC0 */
2236 #define PCL__1 3 /* PC1 */
2237 #define PCL__2 4 /* PC2 */
2238 #define PCL__3 5 /* PC3 */
2239 #define PCL__4 6 /* PC4 */
2240 #define PCL__6 7 /* PC6 */
2241 #define PCL_6N 8 /* PC6 No Retention */
2242 #define PCL_6R 9 /* PC6 Retention */
2243 #define PCL__7 10 /* PC7 */
2244 #define PCL_7S 11 /* PC7 Shrink */
2245 #define PCL__8 12 /* PC8 */
2246 #define PCL__9 13 /* PC9 */
2247 #define PCL_10 14 /* PC10 */
2248 #define PCLUNL 15 /* Unlimited */
2250 int pkg_cstate_limit = PCLUKN;
2251 char *pkg_cstate_limit_strings[] = { "reserved", "unknown", "pc0", "pc1", "pc2",
2252 "pc3", "pc4", "pc6", "pc6n", "pc6r", "pc7", "pc7s", "pc8", "pc9", "pc10", "unlimited"};
2254 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};
2255 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};
2256 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};
2257 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};
2258 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};
2259 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};
2260 int glm_pkg_cstate_limits[16] = {PCLUNL, PCL__1, PCL__3, PCL__6, PCL__7, PCL_7S, PCL__8, PCL__9, PCL_10, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV, PCLRSV};
2261 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};
2265 calculate_tsc_tweak()
2267 tsc_tweak = base_hz / tsc_hz;
2271 dump_nhm_platform_info(void)
2273 unsigned long long msr;
2276 get_msr(base_cpu, MSR_PLATFORM_INFO, &msr);
2278 fprintf(outf, "cpu%d: MSR_PLATFORM_INFO: 0x%08llx\n", base_cpu, msr);
2280 ratio = (msr >> 40) & 0xFF;
2281 fprintf(outf, "%d * %.1f = %.1f MHz max efficiency frequency\n",
2282 ratio, bclk, ratio * bclk);
2284 ratio = (msr >> 8) & 0xFF;
2285 fprintf(outf, "%d * %.1f = %.1f MHz base frequency\n",
2286 ratio, bclk, ratio * bclk);
2288 get_msr(base_cpu, MSR_IA32_POWER_CTL, &msr);
2289 fprintf(outf, "cpu%d: MSR_IA32_POWER_CTL: 0x%08llx (C1E auto-promotion: %sabled)\n",
2290 base_cpu, msr, msr & 0x2 ? "EN" : "DIS");
2296 dump_hsw_turbo_ratio_limits(void)
2298 unsigned long long msr;
2301 get_msr(base_cpu, MSR_TURBO_RATIO_LIMIT2, &msr);
2303 fprintf(outf, "cpu%d: MSR_TURBO_RATIO_LIMIT2: 0x%08llx\n", base_cpu, msr);
2305 ratio = (msr >> 8) & 0xFF;
2307 fprintf(outf, "%d * %.1f = %.1f MHz max turbo 18 active cores\n",
2308 ratio, bclk, ratio * bclk);
2310 ratio = (msr >> 0) & 0xFF;
2312 fprintf(outf, "%d * %.1f = %.1f MHz max turbo 17 active cores\n",
2313 ratio, bclk, ratio * bclk);
2318 dump_ivt_turbo_ratio_limits(void)
2320 unsigned long long msr;
2323 get_msr(base_cpu, MSR_TURBO_RATIO_LIMIT1, &msr);
2325 fprintf(outf, "cpu%d: MSR_TURBO_RATIO_LIMIT1: 0x%08llx\n", base_cpu, msr);
2327 ratio = (msr >> 56) & 0xFF;
2329 fprintf(outf, "%d * %.1f = %.1f MHz max turbo 16 active cores\n",
2330 ratio, bclk, ratio * bclk);
2332 ratio = (msr >> 48) & 0xFF;
2334 fprintf(outf, "%d * %.1f = %.1f MHz max turbo 15 active cores\n",
2335 ratio, bclk, ratio * bclk);
2337 ratio = (msr >> 40) & 0xFF;
2339 fprintf(outf, "%d * %.1f = %.1f MHz max turbo 14 active cores\n",
2340 ratio, bclk, ratio * bclk);
2342 ratio = (msr >> 32) & 0xFF;
2344 fprintf(outf, "%d * %.1f = %.1f MHz max turbo 13 active cores\n",
2345 ratio, bclk, ratio * bclk);
2347 ratio = (msr >> 24) & 0xFF;
2349 fprintf(outf, "%d * %.1f = %.1f MHz max turbo 12 active cores\n",
2350 ratio, bclk, ratio * bclk);
2352 ratio = (msr >> 16) & 0xFF;
2354 fprintf(outf, "%d * %.1f = %.1f MHz max turbo 11 active cores\n",
2355 ratio, bclk, ratio * bclk);
2357 ratio = (msr >> 8) & 0xFF;
2359 fprintf(outf, "%d * %.1f = %.1f MHz max turbo 10 active cores\n",
2360 ratio, bclk, ratio * bclk);
2362 ratio = (msr >> 0) & 0xFF;
2364 fprintf(outf, "%d * %.1f = %.1f MHz max turbo 9 active cores\n",
2365 ratio, bclk, ratio * bclk);
2368 int has_turbo_ratio_group_limits(int family, int model)
2375 case INTEL_FAM6_ATOM_GOLDMONT:
2376 case INTEL_FAM6_SKYLAKE_X:
2377 case INTEL_FAM6_ATOM_GOLDMONT_D:
2378 case INTEL_FAM6_ATOM_TREMONT_D:
2385 dump_turbo_ratio_limits(int family, int model)
2387 unsigned long long msr, core_counts;
2388 unsigned int ratio, group_size;
2390 get_msr(base_cpu, MSR_TURBO_RATIO_LIMIT, &msr);
2391 fprintf(outf, "cpu%d: MSR_TURBO_RATIO_LIMIT: 0x%08llx\n", base_cpu, msr);
2393 if (has_turbo_ratio_group_limits(family, model)) {
2394 get_msr(base_cpu, MSR_TURBO_RATIO_LIMIT1, &core_counts);
2395 fprintf(outf, "cpu%d: MSR_TURBO_RATIO_LIMIT1: 0x%08llx\n", base_cpu, core_counts);
2397 core_counts = 0x0807060504030201;
2400 ratio = (msr >> 56) & 0xFF;
2401 group_size = (core_counts >> 56) & 0xFF;
2403 fprintf(outf, "%d * %.1f = %.1f MHz max turbo %d active cores\n",
2404 ratio, bclk, ratio * bclk, group_size);
2406 ratio = (msr >> 48) & 0xFF;
2407 group_size = (core_counts >> 48) & 0xFF;
2409 fprintf(outf, "%d * %.1f = %.1f MHz max turbo %d active cores\n",
2410 ratio, bclk, ratio * bclk, group_size);
2412 ratio = (msr >> 40) & 0xFF;
2413 group_size = (core_counts >> 40) & 0xFF;
2415 fprintf(outf, "%d * %.1f = %.1f MHz max turbo %d active cores\n",
2416 ratio, bclk, ratio * bclk, group_size);
2418 ratio = (msr >> 32) & 0xFF;
2419 group_size = (core_counts >> 32) & 0xFF;
2421 fprintf(outf, "%d * %.1f = %.1f MHz max turbo %d active cores\n",
2422 ratio, bclk, ratio * bclk, group_size);
2424 ratio = (msr >> 24) & 0xFF;
2425 group_size = (core_counts >> 24) & 0xFF;
2427 fprintf(outf, "%d * %.1f = %.1f MHz max turbo %d active cores\n",
2428 ratio, bclk, ratio * bclk, group_size);
2430 ratio = (msr >> 16) & 0xFF;
2431 group_size = (core_counts >> 16) & 0xFF;
2433 fprintf(outf, "%d * %.1f = %.1f MHz max turbo %d active cores\n",
2434 ratio, bclk, ratio * bclk, group_size);
2436 ratio = (msr >> 8) & 0xFF;
2437 group_size = (core_counts >> 8) & 0xFF;
2439 fprintf(outf, "%d * %.1f = %.1f MHz max turbo %d active cores\n",
2440 ratio, bclk, ratio * bclk, group_size);
2442 ratio = (msr >> 0) & 0xFF;
2443 group_size = (core_counts >> 0) & 0xFF;
2445 fprintf(outf, "%d * %.1f = %.1f MHz max turbo %d active cores\n",
2446 ratio, bclk, ratio * bclk, group_size);
2451 dump_atom_turbo_ratio_limits(void)
2453 unsigned long long msr;
2456 get_msr(base_cpu, MSR_ATOM_CORE_RATIOS, &msr);
2457 fprintf(outf, "cpu%d: MSR_ATOM_CORE_RATIOS: 0x%08llx\n", base_cpu, msr & 0xFFFFFFFF);
2459 ratio = (msr >> 0) & 0x3F;
2461 fprintf(outf, "%d * %.1f = %.1f MHz minimum operating frequency\n",
2462 ratio, bclk, ratio * bclk);
2464 ratio = (msr >> 8) & 0x3F;
2466 fprintf(outf, "%d * %.1f = %.1f MHz low frequency mode (LFM)\n",
2467 ratio, bclk, ratio * bclk);
2469 ratio = (msr >> 16) & 0x3F;
2471 fprintf(outf, "%d * %.1f = %.1f MHz base frequency\n",
2472 ratio, bclk, ratio * bclk);
2474 get_msr(base_cpu, MSR_ATOM_CORE_TURBO_RATIOS, &msr);
2475 fprintf(outf, "cpu%d: MSR_ATOM_CORE_TURBO_RATIOS: 0x%08llx\n", base_cpu, msr & 0xFFFFFFFF);
2477 ratio = (msr >> 24) & 0x3F;
2479 fprintf(outf, "%d * %.1f = %.1f MHz max turbo 4 active cores\n",
2480 ratio, bclk, ratio * bclk);
2482 ratio = (msr >> 16) & 0x3F;
2484 fprintf(outf, "%d * %.1f = %.1f MHz max turbo 3 active cores\n",
2485 ratio, bclk, ratio * bclk);
2487 ratio = (msr >> 8) & 0x3F;
2489 fprintf(outf, "%d * %.1f = %.1f MHz max turbo 2 active cores\n",
2490 ratio, bclk, ratio * bclk);
2492 ratio = (msr >> 0) & 0x3F;
2494 fprintf(outf, "%d * %.1f = %.1f MHz max turbo 1 active core\n",
2495 ratio, bclk, ratio * bclk);
2499 dump_knl_turbo_ratio_limits(void)
2501 const unsigned int buckets_no = 7;
2503 unsigned long long msr;
2504 int delta_cores, delta_ratio;
2506 unsigned int cores[buckets_no];
2507 unsigned int ratio[buckets_no];
2509 get_msr(base_cpu, MSR_TURBO_RATIO_LIMIT, &msr);
2511 fprintf(outf, "cpu%d: MSR_TURBO_RATIO_LIMIT: 0x%08llx\n",
2515 * Turbo encoding in KNL is as follows:
2517 * [7:1] -- Base value of number of active cores of bucket 1.
2518 * [15:8] -- Base value of freq ratio of bucket 1.
2519 * [20:16] -- +ve delta of number of active cores of bucket 2.
2520 * i.e. active cores of bucket 2 =
2521 * active cores of bucket 1 + delta
2522 * [23:21] -- Negative delta of freq ratio of bucket 2.
2523 * i.e. freq ratio of bucket 2 =
2524 * freq ratio of bucket 1 - delta
2525 * [28:24]-- +ve delta of number of active cores of bucket 3.
2526 * [31:29]-- -ve delta of freq ratio of bucket 3.
2527 * [36:32]-- +ve delta of number of active cores of bucket 4.
2528 * [39:37]-- -ve delta of freq ratio of bucket 4.
2529 * [44:40]-- +ve delta of number of active cores of bucket 5.
2530 * [47:45]-- -ve delta of freq ratio of bucket 5.
2531 * [52:48]-- +ve delta of number of active cores of bucket 6.
2532 * [55:53]-- -ve delta of freq ratio of bucket 6.
2533 * [60:56]-- +ve delta of number of active cores of bucket 7.
2534 * [63:61]-- -ve delta of freq ratio of bucket 7.
2538 cores[b_nr] = (msr & 0xFF) >> 1;
2539 ratio[b_nr] = (msr >> 8) & 0xFF;
2541 for (i = 16; i < 64; i += 8) {
2542 delta_cores = (msr >> i) & 0x1F;
2543 delta_ratio = (msr >> (i + 5)) & 0x7;
2545 cores[b_nr + 1] = cores[b_nr] + delta_cores;
2546 ratio[b_nr + 1] = ratio[b_nr] - delta_ratio;
2550 for (i = buckets_no - 1; i >= 0; i--)
2551 if (i > 0 ? ratio[i] != ratio[i - 1] : 1)
2553 "%d * %.1f = %.1f MHz max turbo %d active cores\n",
2554 ratio[i], bclk, ratio[i] * bclk, cores[i]);
2558 dump_nhm_cst_cfg(void)
2560 unsigned long long msr;
2562 get_msr(base_cpu, MSR_PKG_CST_CONFIG_CONTROL, &msr);
2564 fprintf(outf, "cpu%d: MSR_PKG_CST_CONFIG_CONTROL: 0x%08llx", base_cpu, msr);
2566 fprintf(outf, " (%s%s%s%s%slocked, pkg-cstate-limit=%d (%s)",
2567 (msr & SNB_C3_AUTO_UNDEMOTE) ? "UNdemote-C3, " : "",
2568 (msr & SNB_C1_AUTO_UNDEMOTE) ? "UNdemote-C1, " : "",
2569 (msr & NHM_C3_AUTO_DEMOTE) ? "demote-C3, " : "",
2570 (msr & NHM_C1_AUTO_DEMOTE) ? "demote-C1, " : "",
2571 (msr & (1 << 15)) ? "" : "UN",
2572 (unsigned int)msr & 0xF,
2573 pkg_cstate_limit_strings[pkg_cstate_limit]);
2575 #define AUTOMATIC_CSTATE_CONVERSION (1UL << 16)
2576 if (has_automatic_cstate_conversion) {
2577 fprintf(outf, ", automatic c-state conversion=%s",
2578 (msr & AUTOMATIC_CSTATE_CONVERSION) ? "on" : "off");
2581 fprintf(outf, ")\n");
2587 dump_config_tdp(void)
2589 unsigned long long msr;
2591 get_msr(base_cpu, MSR_CONFIG_TDP_NOMINAL, &msr);
2592 fprintf(outf, "cpu%d: MSR_CONFIG_TDP_NOMINAL: 0x%08llx", base_cpu, msr);
2593 fprintf(outf, " (base_ratio=%d)\n", (unsigned int)msr & 0xFF);
2595 get_msr(base_cpu, MSR_CONFIG_TDP_LEVEL_1, &msr);
2596 fprintf(outf, "cpu%d: MSR_CONFIG_TDP_LEVEL_1: 0x%08llx (", base_cpu, msr);
2598 fprintf(outf, "PKG_MIN_PWR_LVL1=%d ", (unsigned int)(msr >> 48) & 0x7FFF);
2599 fprintf(outf, "PKG_MAX_PWR_LVL1=%d ", (unsigned int)(msr >> 32) & 0x7FFF);
2600 fprintf(outf, "LVL1_RATIO=%d ", (unsigned int)(msr >> 16) & 0xFF);
2601 fprintf(outf, "PKG_TDP_LVL1=%d", (unsigned int)(msr) & 0x7FFF);
2603 fprintf(outf, ")\n");
2605 get_msr(base_cpu, MSR_CONFIG_TDP_LEVEL_2, &msr);
2606 fprintf(outf, "cpu%d: MSR_CONFIG_TDP_LEVEL_2: 0x%08llx (", base_cpu, msr);
2608 fprintf(outf, "PKG_MIN_PWR_LVL2=%d ", (unsigned int)(msr >> 48) & 0x7FFF);
2609 fprintf(outf, "PKG_MAX_PWR_LVL2=%d ", (unsigned int)(msr >> 32) & 0x7FFF);
2610 fprintf(outf, "LVL2_RATIO=%d ", (unsigned int)(msr >> 16) & 0xFF);
2611 fprintf(outf, "PKG_TDP_LVL2=%d", (unsigned int)(msr) & 0x7FFF);
2613 fprintf(outf, ")\n");
2615 get_msr(base_cpu, MSR_CONFIG_TDP_CONTROL, &msr);
2616 fprintf(outf, "cpu%d: MSR_CONFIG_TDP_CONTROL: 0x%08llx (", base_cpu, msr);
2618 fprintf(outf, "TDP_LEVEL=%d ", (unsigned int)(msr) & 0x3);
2619 fprintf(outf, " lock=%d", (unsigned int)(msr >> 31) & 1);
2620 fprintf(outf, ")\n");
2622 get_msr(base_cpu, MSR_TURBO_ACTIVATION_RATIO, &msr);
2623 fprintf(outf, "cpu%d: MSR_TURBO_ACTIVATION_RATIO: 0x%08llx (", base_cpu, msr);
2624 fprintf(outf, "MAX_NON_TURBO_RATIO=%d", (unsigned int)(msr) & 0xFF);
2625 fprintf(outf, " lock=%d", (unsigned int)(msr >> 31) & 1);
2626 fprintf(outf, ")\n");
2629 unsigned int irtl_time_units[] = {1, 32, 1024, 32768, 1048576, 33554432, 0, 0 };
2631 void print_irtl(void)
2633 unsigned long long msr;
2635 get_msr(base_cpu, MSR_PKGC3_IRTL, &msr);
2636 fprintf(outf, "cpu%d: MSR_PKGC3_IRTL: 0x%08llx (", base_cpu, msr);
2637 fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
2638 (msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
2640 get_msr(base_cpu, MSR_PKGC6_IRTL, &msr);
2641 fprintf(outf, "cpu%d: MSR_PKGC6_IRTL: 0x%08llx (", base_cpu, msr);
2642 fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
2643 (msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
2645 get_msr(base_cpu, MSR_PKGC7_IRTL, &msr);
2646 fprintf(outf, "cpu%d: MSR_PKGC7_IRTL: 0x%08llx (", base_cpu, msr);
2647 fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
2648 (msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
2653 get_msr(base_cpu, MSR_PKGC8_IRTL, &msr);
2654 fprintf(outf, "cpu%d: MSR_PKGC8_IRTL: 0x%08llx (", base_cpu, msr);
2655 fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
2656 (msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
2658 get_msr(base_cpu, MSR_PKGC9_IRTL, &msr);
2659 fprintf(outf, "cpu%d: MSR_PKGC9_IRTL: 0x%08llx (", base_cpu, msr);
2660 fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
2661 (msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
2663 get_msr(base_cpu, MSR_PKGC10_IRTL, &msr);
2664 fprintf(outf, "cpu%d: MSR_PKGC10_IRTL: 0x%08llx (", base_cpu, msr);
2665 fprintf(outf, "%svalid, %lld ns)\n", msr & (1 << 15) ? "" : "NOT",
2666 (msr & 0x3FF) * irtl_time_units[(msr >> 10) & 0x3]);
2669 void free_fd_percpu(void)
2673 for (i = 0; i < topo.max_cpu_num + 1; ++i) {
2674 if (fd_percpu[i] != 0)
2675 close(fd_percpu[i]);
2681 void free_all_buffers(void)
2685 CPU_FREE(cpu_present_set);
2686 cpu_present_set = NULL;
2687 cpu_present_setsize = 0;
2689 CPU_FREE(cpu_affinity_set);
2690 cpu_affinity_set = NULL;
2691 cpu_affinity_setsize = 0;
2699 package_even = NULL;
2709 free(output_buffer);
2710 output_buffer = NULL;
2715 free(irq_column_2_cpu);
2718 for (i = 0; i <= topo.max_cpu_num; ++i) {
2719 if (cpus[i].put_ids)
2720 CPU_FREE(cpus[i].put_ids);
2727 * Parse a file containing a single int.
2728 * Return 0 if file can not be opened
2729 * Exit if file can be opened, but can not be parsed
2731 int parse_int_file(const char *fmt, ...)
2734 char path[PATH_MAX];
2738 va_start(args, fmt);
2739 vsnprintf(path, sizeof(path), fmt, args);
2741 filep = fopen(path, "r");
2744 if (fscanf(filep, "%d", &value) != 1)
2745 err(1, "%s: failed to parse number from file", path);
2751 * cpu_is_first_core_in_package(cpu)
2752 * return 1 if given CPU is 1st core in package
2754 int cpu_is_first_core_in_package(int cpu)
2756 return cpu == parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_siblings_list", cpu);
2759 int get_physical_package_id(int cpu)
2761 return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/physical_package_id", cpu);
2764 int get_die_id(int cpu)
2766 return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/die_id", cpu);
2769 int get_core_id(int cpu)
2771 return parse_int_file("/sys/devices/system/cpu/cpu%d/topology/core_id", cpu);
2774 void set_node_data(void)
2776 int pkg, node, lnode, cpu, cpux;
2779 /* initialize logical_node_id */
2780 for (cpu = 0; cpu <= topo.max_cpu_num; ++cpu)
2781 cpus[cpu].logical_node_id = -1;
2784 for (pkg = 0; pkg < topo.num_packages; pkg++) {
2786 for (cpu = 0; cpu <= topo.max_cpu_num; ++cpu) {
2787 if (cpus[cpu].physical_package_id != pkg)
2789 /* find a cpu with an unset logical_node_id */
2790 if (cpus[cpu].logical_node_id != -1)
2792 cpus[cpu].logical_node_id = lnode;
2793 node = cpus[cpu].physical_node_id;
2796 * find all matching cpus on this pkg and set
2797 * the logical_node_id
2799 for (cpux = cpu; cpux <= topo.max_cpu_num; cpux++) {
2800 if ((cpus[cpux].physical_package_id == pkg) &&
2801 (cpus[cpux].physical_node_id == node)) {
2802 cpus[cpux].logical_node_id = lnode;
2807 if (lnode > topo.nodes_per_pkg)
2808 topo.nodes_per_pkg = lnode;
2810 if (cpu_count >= topo.max_cpu_num)
2815 int get_physical_node_id(struct cpu_topology *thiscpu)
2820 int cpu = thiscpu->logical_cpu_id;
2822 for (i = 0; i <= topo.max_cpu_num; i++) {
2823 sprintf(path, "/sys/devices/system/cpu/cpu%d/node%i/cpulist",
2825 filep = fopen(path, "r");
2834 int get_thread_siblings(struct cpu_topology *thiscpu)
2836 char path[80], character;
2839 int so, shift, sib_core;
2840 int cpu = thiscpu->logical_cpu_id;
2841 int offset = topo.max_cpu_num + 1;
2845 thiscpu->put_ids = CPU_ALLOC((topo.max_cpu_num + 1));
2846 if (thiscpu->thread_id < 0)
2847 thiscpu->thread_id = thread_id++;
2848 if (!thiscpu->put_ids)
2851 size = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
2852 CPU_ZERO_S(size, thiscpu->put_ids);
2855 "/sys/devices/system/cpu/cpu%d/topology/thread_siblings", cpu);
2856 filep = fopen(path, "r");
2859 warnx("%s: open failed", path);
2863 offset -= BITMASK_SIZE;
2864 if (fscanf(filep, "%lx%c", &map, &character) != 2)
2865 err(1, "%s: failed to parse file", path);
2866 for (shift = 0; shift < BITMASK_SIZE; shift++) {
2867 if ((map >> shift) & 0x1) {
2868 so = shift + offset;
2869 sib_core = get_core_id(so);
2870 if (sib_core == thiscpu->physical_core_id) {
2871 CPU_SET_S(so, size, thiscpu->put_ids);
2873 (cpus[so].thread_id < 0))
2874 cpus[so].thread_id =
2879 } while (!strncmp(&character, ",", 1));
2882 return CPU_COUNT_S(size, thiscpu->put_ids);
2886 * run func(thread, core, package) in topology order
2887 * skip non-present cpus
2890 int for_all_cpus_2(int (func)(struct thread_data *, struct core_data *,
2891 struct pkg_data *, struct thread_data *, struct core_data *,
2892 struct pkg_data *), struct thread_data *thread_base,
2893 struct core_data *core_base, struct pkg_data *pkg_base,
2894 struct thread_data *thread_base2, struct core_data *core_base2,
2895 struct pkg_data *pkg_base2)
2897 int retval, pkg_no, node_no, core_no, thread_no;
2899 for (pkg_no = 0; pkg_no < topo.num_packages; ++pkg_no) {
2900 for (node_no = 0; node_no < topo.nodes_per_pkg; ++node_no) {
2901 for (core_no = 0; core_no < topo.cores_per_node;
2903 for (thread_no = 0; thread_no <
2904 topo.threads_per_core; ++thread_no) {
2905 struct thread_data *t, *t2;
2906 struct core_data *c, *c2;
2907 struct pkg_data *p, *p2;
2909 t = GET_THREAD(thread_base, thread_no,
2913 if (cpu_is_not_present(t->cpu_id))
2916 t2 = GET_THREAD(thread_base2, thread_no,
2920 c = GET_CORE(core_base, core_no,
2922 c2 = GET_CORE(core_base2, core_no,
2926 p = GET_PKG(pkg_base, pkg_no);
2927 p2 = GET_PKG(pkg_base2, pkg_no);
2929 retval = func(t, c, p, t2, c2, p2);
2940 * run func(cpu) on every cpu in /proc/stat
2941 * return max_cpu number
2943 int for_all_proc_cpus(int (func)(int))
2949 fp = fopen_or_die(proc_stat, "r");
2951 retval = fscanf(fp, "cpu %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n");
2953 err(1, "%s: failed to parse format", proc_stat);
2956 retval = fscanf(fp, "cpu%u %*d %*d %*d %*d %*d %*d %*d %*d %*d %*d\n", &cpu_num);
2960 retval = func(cpu_num);
2970 void re_initialize(void)
2973 setup_all_buffers();
2974 fprintf(outf, "turbostat: re-initialized with num_cpus %d\n", topo.num_cpus);
2977 void set_max_cpu_num(void)
2981 unsigned long dummy;
2984 base_cpu = sched_getcpu();
2986 err(1, "cannot find calling cpu ID");
2988 "/sys/devices/system/cpu/cpu%d/topology/thread_siblings",
2991 filep = fopen_or_die(pathname, "r");
2992 topo.max_cpu_num = 0;
2993 while (fscanf(filep, "%lx,", &dummy) == 1)
2994 topo.max_cpu_num += BITMASK_SIZE;
2996 topo.max_cpu_num--; /* 0 based */
3001 * remember the last one seen, it will be the max
3003 int count_cpus(int cpu)
3008 int mark_cpu_present(int cpu)
3010 CPU_SET_S(cpu, cpu_present_setsize, cpu_present_set);
3014 int init_thread_id(int cpu)
3016 cpus[cpu].thread_id = -1;
3021 * snapshot_proc_interrupts()
3023 * read and record summary of /proc/interrupts
3025 * return 1 if config change requires a restart, else return 0
3027 int snapshot_proc_interrupts(void)
3033 fp = fopen_or_die("/proc/interrupts", "r");
3037 /* read 1st line of /proc/interrupts to get cpu* name for each column */
3038 for (column = 0; column < topo.num_cpus; ++column) {
3041 retval = fscanf(fp, " CPU%d", &cpu_number);
3045 if (cpu_number > topo.max_cpu_num) {
3046 warn("/proc/interrupts: cpu%d: > %d", cpu_number, topo.max_cpu_num);
3050 irq_column_2_cpu[column] = cpu_number;
3051 irqs_per_cpu[cpu_number] = 0;
3054 /* read /proc/interrupt count lines and sum up irqs per cpu */
3059 retval = fscanf(fp, " %s:", buf); /* flush irq# "N:" */
3063 /* read the count per cpu */
3064 for (column = 0; column < topo.num_cpus; ++column) {
3066 int cpu_number, irq_count;
3068 retval = fscanf(fp, " %d", &irq_count);
3072 cpu_number = irq_column_2_cpu[column];
3073 irqs_per_cpu[cpu_number] += irq_count;
3077 while (getc(fp) != '\n')
3078 ; /* flush interrupt description */
3084 * snapshot_gfx_rc6_ms()
3086 * record snapshot of
3087 * /sys/class/drm/card0/power/rc6_residency_ms
3089 * return 1 if config change requires a restart, else return 0
3091 int snapshot_gfx_rc6_ms(void)
3096 fp = fopen_or_die("/sys/class/drm/card0/power/rc6_residency_ms", "r");
3098 retval = fscanf(fp, "%lld", &gfx_cur_rc6_ms);
3107 * snapshot_gfx_mhz()
3109 * record snapshot of
3110 * /sys/class/graphics/fb0/device/drm/card0/gt_cur_freq_mhz
3112 * return 1 if config change requires a restart, else return 0
3114 int snapshot_gfx_mhz(void)
3120 fp = fopen_or_die("/sys/class/graphics/fb0/device/drm/card0/gt_cur_freq_mhz", "r");
3126 retval = fscanf(fp, "%d", &gfx_cur_mhz);
3134 * snapshot_gfx_cur_mhz()
3136 * record snapshot of
3137 * /sys/class/graphics/fb0/device/drm/card0/gt_act_freq_mhz
3139 * return 1 if config change requires a restart, else return 0
3141 int snapshot_gfx_act_mhz(void)
3147 fp = fopen_or_die("/sys/class/graphics/fb0/device/drm/card0/gt_act_freq_mhz", "r");
3153 retval = fscanf(fp, "%d", &gfx_act_mhz);
3155 err(1, "GFX ACT MHz");
3161 * snapshot_cpu_lpi()
3163 * record snapshot of
3164 * /sys/devices/system/cpu/cpuidle/low_power_idle_cpu_residency_us
3166 int snapshot_cpu_lpi_us(void)
3171 fp = fopen_or_die("/sys/devices/system/cpu/cpuidle/low_power_idle_cpu_residency_us", "r");
3173 retval = fscanf(fp, "%lld", &cpuidle_cur_cpu_lpi_us);
3175 fprintf(stderr, "Disabling Low Power Idle CPU output\n");
3176 BIC_NOT_PRESENT(BIC_CPU_LPI);
3186 * snapshot_sys_lpi()
3188 * record snapshot of sys_lpi_file
3190 int snapshot_sys_lpi_us(void)
3195 fp = fopen_or_die(sys_lpi_file, "r");
3197 retval = fscanf(fp, "%lld", &cpuidle_cur_sys_lpi_us);
3199 fprintf(stderr, "Disabling Low Power Idle System output\n");
3200 BIC_NOT_PRESENT(BIC_SYS_LPI);
3209 * snapshot /proc and /sys files
3211 * return 1 if configuration restart needed, else return 0
3213 int snapshot_proc_sysfs_files(void)
3215 if (DO_BIC(BIC_IRQ))
3216 if (snapshot_proc_interrupts())
3219 if (DO_BIC(BIC_GFX_rc6))
3220 snapshot_gfx_rc6_ms();
3222 if (DO_BIC(BIC_GFXMHz))
3225 if (DO_BIC(BIC_GFXACTMHz))
3226 snapshot_gfx_act_mhz();
3228 if (DO_BIC(BIC_CPU_LPI))
3229 snapshot_cpu_lpi_us();
3231 if (DO_BIC(BIC_SYS_LPI))
3232 snapshot_sys_lpi_us();
3239 static void signal_handler (int signal)
3245 fprintf(stderr, " SIGINT\n");
3249 fprintf(stderr, "SIGUSR1\n");
3254 void setup_signal_handler(void)
3256 struct sigaction sa;
3258 memset(&sa, 0, sizeof(sa));
3260 sa.sa_handler = &signal_handler;
3262 if (sigaction(SIGINT, &sa, NULL) < 0)
3263 err(1, "sigaction SIGINT");
3264 if (sigaction(SIGUSR1, &sa, NULL) < 0)
3265 err(1, "sigaction SIGUSR1");
3270 struct timeval tout;
3271 struct timespec rest;
3276 FD_SET(0, &readfds);
3279 nanosleep(&interval_ts, NULL);
3284 retval = select(1, &readfds, NULL, NULL, &tout);
3287 switch (getc(stdin)) {
3293 * 'stdin' is a pipe closed on the other end. There
3294 * won't be any further input.
3297 /* Sleep the rest of the time */
3298 rest.tv_sec = (tout.tv_sec + tout.tv_usec / 1000000);
3299 rest.tv_nsec = (tout.tv_usec % 1000000) * 1000;
3300 nanosleep(&rest, NULL);
3305 int get_msr_sum(int cpu, off_t offset, unsigned long long *msr)
3308 unsigned long long msr_cur, msr_last;
3310 if (!per_cpu_msr_sum)
3313 idx = offset_to_idx(offset);
3316 /* get_msr_sum() = sum + (get_msr() - last) */
3317 ret = get_msr(cpu, offset, &msr_cur);
3320 msr_last = per_cpu_msr_sum[cpu].entries[idx].last;
3321 DELTA_WRAP32(msr_cur, msr_last);
3322 *msr = msr_last + per_cpu_msr_sum[cpu].entries[idx].sum;
3329 /* Timer callback, update the sum of MSRs periodically. */
3330 static int update_msr_sum(struct thread_data *t, struct core_data *c, struct pkg_data *p)
3333 int cpu = t->cpu_id;
3335 for (i = IDX_PKG_ENERGY; i < IDX_COUNT; i++) {
3336 unsigned long long msr_cur, msr_last;
3341 offset = idx_to_offset(i);
3344 ret = get_msr(cpu, offset, &msr_cur);
3346 fprintf(outf, "Can not update msr(0x%x)\n", offset);
3350 msr_last = per_cpu_msr_sum[cpu].entries[i].last;
3351 per_cpu_msr_sum[cpu].entries[i].last = msr_cur & 0xffffffff;
3353 DELTA_WRAP32(msr_cur, msr_last);
3354 per_cpu_msr_sum[cpu].entries[i].sum += msr_last;
3360 msr_record_handler(union sigval v)
3362 for_all_cpus(update_msr_sum, EVEN_COUNTERS);
3365 void msr_sum_record(void)
3367 struct itimerspec its;
3368 struct sigevent sev;
3370 per_cpu_msr_sum = calloc(topo.max_cpu_num + 1, sizeof(struct msr_sum_array));
3371 if (!per_cpu_msr_sum) {
3372 fprintf(outf, "Can not allocate memory for long time MSR.\n");
3376 * Signal handler might be restricted, so use thread notifier instead.
3378 memset(&sev, 0, sizeof(struct sigevent));
3379 sev.sigev_notify = SIGEV_THREAD;
3380 sev.sigev_notify_function = msr_record_handler;
3382 sev.sigev_value.sival_ptr = &timerid;
3383 if (timer_create(CLOCK_REALTIME, &sev, &timerid) == -1) {
3384 fprintf(outf, "Can not create timer.\n");
3388 its.it_value.tv_sec = 0;
3389 its.it_value.tv_nsec = 1;
3391 * A wraparound time has been calculated early.
3392 * Some sources state that the peak power for a
3393 * microprocessor is usually 1.5 times the TDP rating,
3394 * use 2 * TDP for safety.
3396 its.it_interval.tv_sec = rapl_joule_counter_range / 2;
3397 its.it_interval.tv_nsec = 0;
3399 if (timer_settime(timerid, 0, &its, NULL) == -1) {
3400 fprintf(outf, "Can not set timer.\n");
3406 timer_delete(timerid);
3408 free(per_cpu_msr_sum);
3411 void turbostat_loop()
3417 setup_signal_handler();
3422 snapshot_proc_sysfs_files();
3423 retval = for_all_cpus(get_counters, EVEN_COUNTERS);
3424 first_counter_read = 0;
3427 } else if (retval == -1) {
3428 if (restarted > 10) {
3436 gettimeofday(&tv_even, (struct timezone *)NULL);
3439 if (for_all_proc_cpus(cpu_is_not_present)) {
3444 if (snapshot_proc_sysfs_files())
3446 retval = for_all_cpus(get_counters, ODD_COUNTERS);
3449 } else if (retval == -1) {
3453 gettimeofday(&tv_odd, (struct timezone *)NULL);
3454 timersub(&tv_odd, &tv_even, &tv_delta);
3455 if (for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS)) {
3459 compute_average(EVEN_COUNTERS);
3460 format_all_counters(EVEN_COUNTERS);
3461 flush_output_stdout();
3464 if (num_iterations && ++done_iters >= num_iterations)
3467 if (snapshot_proc_sysfs_files())
3469 retval = for_all_cpus(get_counters, EVEN_COUNTERS);
3472 } else if (retval == -1) {
3476 gettimeofday(&tv_even, (struct timezone *)NULL);
3477 timersub(&tv_even, &tv_odd, &tv_delta);
3478 if (for_all_cpus_2(delta_cpu, EVEN_COUNTERS, ODD_COUNTERS)) {
3482 compute_average(ODD_COUNTERS);
3483 format_all_counters(ODD_COUNTERS);
3484 flush_output_stdout();
3487 if (num_iterations && ++done_iters >= num_iterations)
3492 void check_dev_msr()
3497 sprintf(pathname, "/dev/cpu/%d/msr", base_cpu);
3498 if (stat(pathname, &sb))
3499 if (system("/sbin/modprobe msr > /dev/null 2>&1"))
3500 err(-5, "no /dev/cpu/0/msr, Try \"# modprobe msr\" ");
3504 * check for CAP_SYS_RAWIO
3505 * return 0 on success
3508 int check_for_cap_sys_rawio(void)
3511 cap_flag_value_t cap_flag_value;
3513 caps = cap_get_proc();
3515 err(-6, "cap_get_proc\n");
3517 if (cap_get_flag(caps, CAP_SYS_RAWIO, CAP_EFFECTIVE, &cap_flag_value))
3518 err(-6, "cap_get\n");
3520 if (cap_flag_value != CAP_SET) {
3521 warnx("capget(CAP_SYS_RAWIO) failed,"
3522 " try \"# setcap cap_sys_rawio=ep %s\"", progname);
3526 if (cap_free(caps) == -1)
3527 err(-6, "cap_free\n");
3531 void check_permissions(void)
3536 /* check for CAP_SYS_RAWIO */
3537 do_exit += check_for_cap_sys_rawio();
3539 /* test file permissions */
3540 sprintf(pathname, "/dev/cpu/%d/msr", base_cpu);
3541 if (euidaccess(pathname, R_OK)) {
3543 warn("/dev/cpu/0/msr open failed, try chown or chmod +r /dev/cpu/*/msr");
3546 /* if all else fails, thell them to be root */
3549 warnx("... or simply run as root");
3556 * NHM adds support for additional MSRs:
3558 * MSR_SMI_COUNT 0x00000034
3560 * MSR_PLATFORM_INFO 0x000000ce
3561 * MSR_PKG_CST_CONFIG_CONTROL 0x000000e2
3563 * MSR_MISC_PWR_MGMT 0x000001aa
3565 * MSR_PKG_C3_RESIDENCY 0x000003f8
3566 * MSR_PKG_C6_RESIDENCY 0x000003f9
3567 * MSR_CORE_C3_RESIDENCY 0x000003fc
3568 * MSR_CORE_C6_RESIDENCY 0x000003fd
3571 * sets global pkg_cstate_limit to decode MSR_PKG_CST_CONFIG_CONTROL
3572 * sets has_misc_feature_control
3574 int probe_nhm_msrs(unsigned int family, unsigned int model)
3576 unsigned long long msr;
3577 unsigned int base_ratio;
3578 int *pkg_cstate_limits;
3586 bclk = discover_bclk(family, model);
3589 case INTEL_FAM6_NEHALEM: /* Core i7 and i5 Processor - Clarksfield, Lynnfield, Jasper Forest */
3590 case INTEL_FAM6_NEHALEM_EX: /* Nehalem-EX Xeon - Beckton */
3591 pkg_cstate_limits = nhm_pkg_cstate_limits;
3593 case INTEL_FAM6_SANDYBRIDGE: /* SNB */
3594 case INTEL_FAM6_SANDYBRIDGE_X: /* SNB Xeon */
3595 case INTEL_FAM6_IVYBRIDGE: /* IVB */
3596 case INTEL_FAM6_IVYBRIDGE_X: /* IVB Xeon */
3597 pkg_cstate_limits = snb_pkg_cstate_limits;
3598 has_misc_feature_control = 1;
3600 case INTEL_FAM6_HASWELL: /* HSW */
3601 case INTEL_FAM6_HASWELL_G: /* HSW */
3602 case INTEL_FAM6_HASWELL_X: /* HSX */
3603 case INTEL_FAM6_HASWELL_L: /* HSW */
3604 case INTEL_FAM6_BROADWELL: /* BDW */
3605 case INTEL_FAM6_BROADWELL_G: /* BDW */
3606 case INTEL_FAM6_BROADWELL_X: /* BDX */
3607 case INTEL_FAM6_SKYLAKE_L: /* SKL */
3608 case INTEL_FAM6_CANNONLAKE_L: /* CNL */
3609 pkg_cstate_limits = hsw_pkg_cstate_limits;
3610 has_misc_feature_control = 1;
3612 case INTEL_FAM6_SKYLAKE_X: /* SKX */
3613 pkg_cstate_limits = skx_pkg_cstate_limits;
3614 has_misc_feature_control = 1;
3616 case INTEL_FAM6_ATOM_SILVERMONT: /* BYT */
3617 no_MSR_MISC_PWR_MGMT = 1;
3618 case INTEL_FAM6_ATOM_SILVERMONT_D: /* AVN */
3619 pkg_cstate_limits = slv_pkg_cstate_limits;
3621 case INTEL_FAM6_ATOM_AIRMONT: /* AMT */
3622 pkg_cstate_limits = amt_pkg_cstate_limits;
3623 no_MSR_MISC_PWR_MGMT = 1;
3625 case INTEL_FAM6_XEON_PHI_KNL: /* PHI */
3626 pkg_cstate_limits = phi_pkg_cstate_limits;
3628 case INTEL_FAM6_ATOM_GOLDMONT: /* BXT */
3629 case INTEL_FAM6_ATOM_GOLDMONT_PLUS:
3630 case INTEL_FAM6_ATOM_GOLDMONT_D: /* DNV */
3631 case INTEL_FAM6_ATOM_TREMONT: /* EHL */
3632 case INTEL_FAM6_ATOM_TREMONT_D: /* JVL */
3633 pkg_cstate_limits = glm_pkg_cstate_limits;
3638 get_msr(base_cpu, MSR_PKG_CST_CONFIG_CONTROL, &msr);
3639 pkg_cstate_limit = pkg_cstate_limits[msr & 0xF];
3641 get_msr(base_cpu, MSR_PLATFORM_INFO, &msr);
3642 base_ratio = (msr >> 8) & 0xFF;
3644 base_hz = base_ratio * bclk * 1000000;
3649 * SLV client has support for unique MSRs:
3651 * MSR_CC6_DEMOTION_POLICY_CONFIG
3652 * MSR_MC6_DEMOTION_POLICY_CONFIG
3655 int has_slv_msrs(unsigned int family, unsigned int model)
3661 case INTEL_FAM6_ATOM_SILVERMONT:
3662 case INTEL_FAM6_ATOM_SILVERMONT_MID:
3663 case INTEL_FAM6_ATOM_AIRMONT_MID:
3668 int is_dnv(unsigned int family, unsigned int model)
3675 case INTEL_FAM6_ATOM_GOLDMONT_D:
3680 int is_bdx(unsigned int family, unsigned int model)
3687 case INTEL_FAM6_BROADWELL_X:
3692 int is_skx(unsigned int family, unsigned int model)
3699 case INTEL_FAM6_SKYLAKE_X:
3704 int is_ehl(unsigned int family, unsigned int model)
3710 case INTEL_FAM6_ATOM_TREMONT:
3715 int is_jvl(unsigned int family, unsigned int model)
3721 case INTEL_FAM6_ATOM_TREMONT_D:
3727 int has_turbo_ratio_limit(unsigned int family, unsigned int model)
3729 if (has_slv_msrs(family, model))
3733 /* Nehalem compatible, but do not include turbo-ratio limit support */
3734 case INTEL_FAM6_NEHALEM_EX: /* Nehalem-EX Xeon - Beckton */
3735 case INTEL_FAM6_XEON_PHI_KNL: /* PHI - Knights Landing (different MSR definition) */
3741 int has_atom_turbo_ratio_limit(unsigned int family, unsigned int model)
3743 if (has_slv_msrs(family, model))
3748 int has_ivt_turbo_ratio_limit(unsigned int family, unsigned int model)
3757 case INTEL_FAM6_IVYBRIDGE_X: /* IVB Xeon */
3758 case INTEL_FAM6_HASWELL_X: /* HSW Xeon */
3764 int has_hsw_turbo_ratio_limit(unsigned int family, unsigned int model)
3773 case INTEL_FAM6_HASWELL_X: /* HSW Xeon */
3780 int has_knl_turbo_ratio_limit(unsigned int family, unsigned int model)
3789 case INTEL_FAM6_XEON_PHI_KNL: /* Knights Landing */
3795 int has_glm_turbo_ratio_limit(unsigned int family, unsigned int model)
3804 case INTEL_FAM6_ATOM_GOLDMONT:
3805 case INTEL_FAM6_SKYLAKE_X:
3811 int has_config_tdp(unsigned int family, unsigned int model)
3820 case INTEL_FAM6_IVYBRIDGE: /* IVB */
3821 case INTEL_FAM6_HASWELL: /* HSW */
3822 case INTEL_FAM6_HASWELL_X: /* HSX */
3823 case INTEL_FAM6_HASWELL_L: /* HSW */
3824 case INTEL_FAM6_HASWELL_G: /* HSW */
3825 case INTEL_FAM6_BROADWELL: /* BDW */
3826 case INTEL_FAM6_BROADWELL_G: /* BDW */
3827 case INTEL_FAM6_BROADWELL_X: /* BDX */
3828 case INTEL_FAM6_SKYLAKE_L: /* SKL */
3829 case INTEL_FAM6_CANNONLAKE_L: /* CNL */
3830 case INTEL_FAM6_SKYLAKE_X: /* SKX */
3832 case INTEL_FAM6_XEON_PHI_KNL: /* Knights Landing */
3840 remove_underbar(char *s)
3854 dump_cstate_pstate_config_info(unsigned int family, unsigned int model)
3856 if (!do_nhm_platform_info)
3859 dump_nhm_platform_info();
3861 if (has_hsw_turbo_ratio_limit(family, model))
3862 dump_hsw_turbo_ratio_limits();
3864 if (has_ivt_turbo_ratio_limit(family, model))
3865 dump_ivt_turbo_ratio_limits();
3867 if (has_turbo_ratio_limit(family, model))
3868 dump_turbo_ratio_limits(family, model);
3870 if (has_atom_turbo_ratio_limit(family, model))
3871 dump_atom_turbo_ratio_limits();
3873 if (has_knl_turbo_ratio_limit(family, model))
3874 dump_knl_turbo_ratio_limits();
3876 if (has_config_tdp(family, model))
3882 static void dump_sysfs_file(char *path)
3885 char cpuidle_buf[64];
3887 input = fopen(path, "r");
3888 if (input == NULL) {
3890 fprintf(outf, "NSFOD %s\n", path);
3893 if (!fgets(cpuidle_buf, sizeof(cpuidle_buf), input))
3894 err(1, "%s: failed to read file", path);
3897 fprintf(outf, "%s: %s", strrchr(path, '/') + 1, cpuidle_buf);
3900 dump_sysfs_cstate_config(void)
3909 if (access("/sys/devices/system/cpu/cpuidle", R_OK)) {
3910 fprintf(outf, "cpuidle not loaded\n");
3914 dump_sysfs_file("/sys/devices/system/cpu/cpuidle/current_driver");
3915 dump_sysfs_file("/sys/devices/system/cpu/cpuidle/current_governor");
3916 dump_sysfs_file("/sys/devices/system/cpu/cpuidle/current_governor_ro");
3918 for (state = 0; state < 10; ++state) {
3920 sprintf(path, "/sys/devices/system/cpu/cpu%d/cpuidle/state%d/name",
3922 input = fopen(path, "r");
3925 if (!fgets(name_buf, sizeof(name_buf), input))
3926 err(1, "%s: failed to read file", path);
3928 /* truncate "C1-HSW\n" to "C1", or truncate "C1\n" to "C1" */
3929 sp = strchr(name_buf, '-');
3931 sp = strchrnul(name_buf, '\n');
3935 remove_underbar(name_buf);
3937 sprintf(path, "/sys/devices/system/cpu/cpu%d/cpuidle/state%d/desc",
3939 input = fopen(path, "r");
3942 if (!fgets(desc, sizeof(desc), input))
3943 err(1, "%s: failed to read file", path);
3945 fprintf(outf, "cpu%d: %s: %s", base_cpu, name_buf, desc);
3950 dump_sysfs_pstate_config(void)
3953 char driver_buf[64];
3954 char governor_buf[64];
3958 sprintf(path, "/sys/devices/system/cpu/cpu%d/cpufreq/scaling_driver",
3960 input = fopen(path, "r");
3961 if (input == NULL) {
3962 fprintf(outf, "NSFOD %s\n", path);
3965 if (!fgets(driver_buf, sizeof(driver_buf), input))
3966 err(1, "%s: failed to read file", path);
3969 sprintf(path, "/sys/devices/system/cpu/cpu%d/cpufreq/scaling_governor",
3971 input = fopen(path, "r");
3972 if (input == NULL) {
3973 fprintf(outf, "NSFOD %s\n", path);
3976 if (!fgets(governor_buf, sizeof(governor_buf), input))
3977 err(1, "%s: failed to read file", path);
3980 fprintf(outf, "cpu%d: cpufreq driver: %s", base_cpu, driver_buf);
3981 fprintf(outf, "cpu%d: cpufreq governor: %s", base_cpu, governor_buf);
3983 sprintf(path, "/sys/devices/system/cpu/cpufreq/boost");
3984 input = fopen(path, "r");
3985 if (input != NULL) {
3986 if (fscanf(input, "%d", &turbo) != 1)
3987 err(1, "%s: failed to parse number from file", path);
3988 fprintf(outf, "cpufreq boost: %d\n", turbo);
3992 sprintf(path, "/sys/devices/system/cpu/intel_pstate/no_turbo");
3993 input = fopen(path, "r");
3994 if (input != NULL) {
3995 if (fscanf(input, "%d", &turbo) != 1)
3996 err(1, "%s: failed to parse number from file", path);
3997 fprintf(outf, "cpufreq intel_pstate no_turbo: %d\n", turbo);
4005 * Decode the ENERGY_PERF_BIAS MSR
4007 int print_epb(struct thread_data *t, struct core_data *c, struct pkg_data *p)
4017 /* EPB is per-package */
4018 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
4021 if (cpu_migrate(cpu)) {
4022 fprintf(outf, "print_epb: Could not migrate to CPU %d\n", cpu);
4031 case ENERGY_PERF_BIAS_PERFORMANCE:
4032 epb_string = "performance";
4034 case ENERGY_PERF_BIAS_NORMAL:
4035 epb_string = "balanced";
4037 case ENERGY_PERF_BIAS_POWERSAVE:
4038 epb_string = "powersave";
4041 epb_string = "custom";
4044 fprintf(outf, "cpu%d: EPB: %d (%s)\n", cpu, epb, epb_string);
4050 * Decode the MSR_HWP_CAPABILITIES
4052 int print_hwp(struct thread_data *t, struct core_data *c, struct pkg_data *p)
4054 unsigned long long msr;
4062 /* MSR_HWP_CAPABILITIES is per-package */
4063 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
4066 if (cpu_migrate(cpu)) {
4067 fprintf(outf, "print_hwp: Could not migrate to CPU %d\n", cpu);
4071 if (get_msr(cpu, MSR_PM_ENABLE, &msr))
4074 fprintf(outf, "cpu%d: MSR_PM_ENABLE: 0x%08llx (%sHWP)\n",
4075 cpu, msr, (msr & (1 << 0)) ? "" : "No-");
4077 /* MSR_PM_ENABLE[1] == 1 if HWP is enabled and MSRs visible */
4078 if ((msr & (1 << 0)) == 0)
4081 if (get_msr(cpu, MSR_HWP_CAPABILITIES, &msr))
4084 fprintf(outf, "cpu%d: MSR_HWP_CAPABILITIES: 0x%08llx "
4085 "(high %d guar %d eff %d low %d)\n",
4087 (unsigned int)HWP_HIGHEST_PERF(msr),
4088 (unsigned int)HWP_GUARANTEED_PERF(msr),
4089 (unsigned int)HWP_MOSTEFFICIENT_PERF(msr),
4090 (unsigned int)HWP_LOWEST_PERF(msr));
4092 if (get_msr(cpu, MSR_HWP_REQUEST, &msr))
4095 fprintf(outf, "cpu%d: MSR_HWP_REQUEST: 0x%08llx "
4096 "(min %d max %d des %d epp 0x%x window 0x%x pkg 0x%x)\n",
4098 (unsigned int)(((msr) >> 0) & 0xff),
4099 (unsigned int)(((msr) >> 8) & 0xff),
4100 (unsigned int)(((msr) >> 16) & 0xff),
4101 (unsigned int)(((msr) >> 24) & 0xff),
4102 (unsigned int)(((msr) >> 32) & 0xff3),
4103 (unsigned int)(((msr) >> 42) & 0x1));
4106 if (get_msr(cpu, MSR_HWP_REQUEST_PKG, &msr))
4109 fprintf(outf, "cpu%d: MSR_HWP_REQUEST_PKG: 0x%08llx "
4110 "(min %d max %d des %d epp 0x%x window 0x%x)\n",
4112 (unsigned int)(((msr) >> 0) & 0xff),
4113 (unsigned int)(((msr) >> 8) & 0xff),
4114 (unsigned int)(((msr) >> 16) & 0xff),
4115 (unsigned int)(((msr) >> 24) & 0xff),
4116 (unsigned int)(((msr) >> 32) & 0xff3));
4118 if (has_hwp_notify) {
4119 if (get_msr(cpu, MSR_HWP_INTERRUPT, &msr))
4122 fprintf(outf, "cpu%d: MSR_HWP_INTERRUPT: 0x%08llx "
4123 "(%s_Guaranteed_Perf_Change, %s_Excursion_Min)\n",
4125 ((msr) & 0x1) ? "EN" : "Dis",
4126 ((msr) & 0x2) ? "EN" : "Dis");
4128 if (get_msr(cpu, MSR_HWP_STATUS, &msr))
4131 fprintf(outf, "cpu%d: MSR_HWP_STATUS: 0x%08llx "
4132 "(%sGuaranteed_Perf_Change, %sExcursion_Min)\n",
4134 ((msr) & 0x1) ? "" : "No-",
4135 ((msr) & 0x2) ? "" : "No-");
4141 * print_perf_limit()
4143 int print_perf_limit(struct thread_data *t, struct core_data *c, struct pkg_data *p)
4145 unsigned long long msr;
4151 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
4154 if (cpu_migrate(cpu)) {
4155 fprintf(outf, "print_perf_limit: Could not migrate to CPU %d\n", cpu);
4159 if (do_core_perf_limit_reasons) {
4160 get_msr(cpu, MSR_CORE_PERF_LIMIT_REASONS, &msr);
4161 fprintf(outf, "cpu%d: MSR_CORE_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
4162 fprintf(outf, " (Active: %s%s%s%s%s%s%s%s%s%s%s%s%s%s)",
4163 (msr & 1 << 15) ? "bit15, " : "",
4164 (msr & 1 << 14) ? "bit14, " : "",
4165 (msr & 1 << 13) ? "Transitions, " : "",
4166 (msr & 1 << 12) ? "MultiCoreTurbo, " : "",
4167 (msr & 1 << 11) ? "PkgPwrL2, " : "",
4168 (msr & 1 << 10) ? "PkgPwrL1, " : "",
4169 (msr & 1 << 9) ? "CorePwr, " : "",
4170 (msr & 1 << 8) ? "Amps, " : "",
4171 (msr & 1 << 6) ? "VR-Therm, " : "",
4172 (msr & 1 << 5) ? "Auto-HWP, " : "",
4173 (msr & 1 << 4) ? "Graphics, " : "",
4174 (msr & 1 << 2) ? "bit2, " : "",
4175 (msr & 1 << 1) ? "ThermStatus, " : "",
4176 (msr & 1 << 0) ? "PROCHOT, " : "");
4177 fprintf(outf, " (Logged: %s%s%s%s%s%s%s%s%s%s%s%s%s%s)\n",
4178 (msr & 1 << 31) ? "bit31, " : "",
4179 (msr & 1 << 30) ? "bit30, " : "",
4180 (msr & 1 << 29) ? "Transitions, " : "",
4181 (msr & 1 << 28) ? "MultiCoreTurbo, " : "",
4182 (msr & 1 << 27) ? "PkgPwrL2, " : "",
4183 (msr & 1 << 26) ? "PkgPwrL1, " : "",
4184 (msr & 1 << 25) ? "CorePwr, " : "",
4185 (msr & 1 << 24) ? "Amps, " : "",
4186 (msr & 1 << 22) ? "VR-Therm, " : "",
4187 (msr & 1 << 21) ? "Auto-HWP, " : "",
4188 (msr & 1 << 20) ? "Graphics, " : "",
4189 (msr & 1 << 18) ? "bit18, " : "",
4190 (msr & 1 << 17) ? "ThermStatus, " : "",
4191 (msr & 1 << 16) ? "PROCHOT, " : "");
4194 if (do_gfx_perf_limit_reasons) {
4195 get_msr(cpu, MSR_GFX_PERF_LIMIT_REASONS, &msr);
4196 fprintf(outf, "cpu%d: MSR_GFX_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
4197 fprintf(outf, " (Active: %s%s%s%s%s%s%s%s)",
4198 (msr & 1 << 0) ? "PROCHOT, " : "",
4199 (msr & 1 << 1) ? "ThermStatus, " : "",
4200 (msr & 1 << 4) ? "Graphics, " : "",
4201 (msr & 1 << 6) ? "VR-Therm, " : "",
4202 (msr & 1 << 8) ? "Amps, " : "",
4203 (msr & 1 << 9) ? "GFXPwr, " : "",
4204 (msr & 1 << 10) ? "PkgPwrL1, " : "",
4205 (msr & 1 << 11) ? "PkgPwrL2, " : "");
4206 fprintf(outf, " (Logged: %s%s%s%s%s%s%s%s)\n",
4207 (msr & 1 << 16) ? "PROCHOT, " : "",
4208 (msr & 1 << 17) ? "ThermStatus, " : "",
4209 (msr & 1 << 20) ? "Graphics, " : "",
4210 (msr & 1 << 22) ? "VR-Therm, " : "",
4211 (msr & 1 << 24) ? "Amps, " : "",
4212 (msr & 1 << 25) ? "GFXPwr, " : "",
4213 (msr & 1 << 26) ? "PkgPwrL1, " : "",
4214 (msr & 1 << 27) ? "PkgPwrL2, " : "");
4216 if (do_ring_perf_limit_reasons) {
4217 get_msr(cpu, MSR_RING_PERF_LIMIT_REASONS, &msr);
4218 fprintf(outf, "cpu%d: MSR_RING_PERF_LIMIT_REASONS, 0x%08llx", cpu, msr);
4219 fprintf(outf, " (Active: %s%s%s%s%s%s)",
4220 (msr & 1 << 0) ? "PROCHOT, " : "",
4221 (msr & 1 << 1) ? "ThermStatus, " : "",
4222 (msr & 1 << 6) ? "VR-Therm, " : "",
4223 (msr & 1 << 8) ? "Amps, " : "",
4224 (msr & 1 << 10) ? "PkgPwrL1, " : "",
4225 (msr & 1 << 11) ? "PkgPwrL2, " : "");
4226 fprintf(outf, " (Logged: %s%s%s%s%s%s)\n",
4227 (msr & 1 << 16) ? "PROCHOT, " : "",
4228 (msr & 1 << 17) ? "ThermStatus, " : "",
4229 (msr & 1 << 22) ? "VR-Therm, " : "",
4230 (msr & 1 << 24) ? "Amps, " : "",
4231 (msr & 1 << 26) ? "PkgPwrL1, " : "",
4232 (msr & 1 << 27) ? "PkgPwrL2, " : "");
4237 #define RAPL_POWER_GRANULARITY 0x7FFF /* 15 bit power granularity */
4238 #define RAPL_TIME_GRANULARITY 0x3F /* 6 bit time granularity */
4240 double get_tdp_intel(unsigned int model)
4242 unsigned long long msr;
4244 if (do_rapl & RAPL_PKG_POWER_INFO)
4245 if (!get_msr(base_cpu, MSR_PKG_POWER_INFO, &msr))
4246 return ((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units;
4249 case INTEL_FAM6_ATOM_SILVERMONT:
4250 case INTEL_FAM6_ATOM_SILVERMONT_D:
4257 double get_tdp_amd(unsigned int family)
4259 /* This is the max stock TDP of HEDT/Server Fam17h+ chips */
4264 * rapl_dram_energy_units_probe()
4265 * Energy units are either hard-coded, or come from RAPL Energy Unit MSR.
4268 rapl_dram_energy_units_probe(int model, double rapl_energy_units)
4270 /* only called for genuine_intel, family 6 */
4273 case INTEL_FAM6_HASWELL_X: /* HSX */
4274 case INTEL_FAM6_BROADWELL_X: /* BDX */
4275 case INTEL_FAM6_XEON_PHI_KNL: /* KNL */
4276 return (rapl_dram_energy_units = 15.3 / 1000000);
4278 return (rapl_energy_units);
4282 void rapl_probe_intel(unsigned int family, unsigned int model)
4284 unsigned long long msr;
4285 unsigned int time_unit;
4292 case INTEL_FAM6_SANDYBRIDGE:
4293 case INTEL_FAM6_IVYBRIDGE:
4294 case INTEL_FAM6_HASWELL: /* HSW */
4295 case INTEL_FAM6_HASWELL_L: /* HSW */
4296 case INTEL_FAM6_HASWELL_G: /* HSW */
4297 case INTEL_FAM6_BROADWELL: /* BDW */
4298 case INTEL_FAM6_BROADWELL_G: /* BDW */
4299 do_rapl = RAPL_PKG | RAPL_CORES | RAPL_CORE_POLICY | RAPL_GFX | RAPL_PKG_POWER_INFO;
4301 BIC_PRESENT(BIC_Pkg_J);
4302 BIC_PRESENT(BIC_Cor_J);
4303 BIC_PRESENT(BIC_GFX_J);
4305 BIC_PRESENT(BIC_PkgWatt);
4306 BIC_PRESENT(BIC_CorWatt);
4307 BIC_PRESENT(BIC_GFXWatt);
4310 case INTEL_FAM6_ATOM_GOLDMONT: /* BXT */
4311 case INTEL_FAM6_ATOM_GOLDMONT_PLUS:
4312 do_rapl = RAPL_PKG | RAPL_PKG_POWER_INFO;
4314 BIC_PRESENT(BIC_Pkg_J);
4316 BIC_PRESENT(BIC_PkgWatt);
4318 case INTEL_FAM6_ATOM_TREMONT: /* EHL */
4319 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;
4321 BIC_PRESENT(BIC_Pkg_J);
4322 BIC_PRESENT(BIC_Cor_J);
4323 BIC_PRESENT(BIC_RAM_J);
4324 BIC_PRESENT(BIC_GFX_J);
4326 BIC_PRESENT(BIC_PkgWatt);
4327 BIC_PRESENT(BIC_CorWatt);
4328 BIC_PRESENT(BIC_RAMWatt);
4329 BIC_PRESENT(BIC_GFXWatt);
4332 case INTEL_FAM6_ATOM_TREMONT_D: /* JVL */
4333 do_rapl = RAPL_PKG | RAPL_PKG_PERF_STATUS | RAPL_PKG_POWER_INFO;
4334 BIC_PRESENT(BIC_PKG__);
4336 BIC_PRESENT(BIC_Pkg_J);
4338 BIC_PRESENT(BIC_PkgWatt);
4340 case INTEL_FAM6_SKYLAKE_L: /* SKL */
4341 case INTEL_FAM6_CANNONLAKE_L: /* CNL */
4342 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;
4343 BIC_PRESENT(BIC_PKG__);
4344 BIC_PRESENT(BIC_RAM__);
4346 BIC_PRESENT(BIC_Pkg_J);
4347 BIC_PRESENT(BIC_Cor_J);
4348 BIC_PRESENT(BIC_RAM_J);
4349 BIC_PRESENT(BIC_GFX_J);
4351 BIC_PRESENT(BIC_PkgWatt);
4352 BIC_PRESENT(BIC_CorWatt);
4353 BIC_PRESENT(BIC_RAMWatt);
4354 BIC_PRESENT(BIC_GFXWatt);
4357 case INTEL_FAM6_HASWELL_X: /* HSX */
4358 case INTEL_FAM6_BROADWELL_X: /* BDX */
4359 case INTEL_FAM6_SKYLAKE_X: /* SKX */
4360 case INTEL_FAM6_XEON_PHI_KNL: /* KNL */
4361 do_rapl = RAPL_PKG | RAPL_DRAM | RAPL_DRAM_POWER_INFO | RAPL_DRAM_PERF_STATUS | RAPL_PKG_PERF_STATUS | RAPL_PKG_POWER_INFO;
4362 BIC_PRESENT(BIC_PKG__);
4363 BIC_PRESENT(BIC_RAM__);
4365 BIC_PRESENT(BIC_Pkg_J);
4366 BIC_PRESENT(BIC_RAM_J);
4368 BIC_PRESENT(BIC_PkgWatt);
4369 BIC_PRESENT(BIC_RAMWatt);
4372 case INTEL_FAM6_SANDYBRIDGE_X:
4373 case INTEL_FAM6_IVYBRIDGE_X:
4374 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;
4375 BIC_PRESENT(BIC_PKG__);
4376 BIC_PRESENT(BIC_RAM__);
4378 BIC_PRESENT(BIC_Pkg_J);
4379 BIC_PRESENT(BIC_Cor_J);
4380 BIC_PRESENT(BIC_RAM_J);
4382 BIC_PRESENT(BIC_PkgWatt);
4383 BIC_PRESENT(BIC_CorWatt);
4384 BIC_PRESENT(BIC_RAMWatt);
4387 case INTEL_FAM6_ATOM_SILVERMONT: /* BYT */
4388 case INTEL_FAM6_ATOM_SILVERMONT_D: /* AVN */
4389 do_rapl = RAPL_PKG | RAPL_CORES;
4391 BIC_PRESENT(BIC_Pkg_J);
4392 BIC_PRESENT(BIC_Cor_J);
4394 BIC_PRESENT(BIC_PkgWatt);
4395 BIC_PRESENT(BIC_CorWatt);
4398 case INTEL_FAM6_ATOM_GOLDMONT_D: /* DNV */
4399 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;
4400 BIC_PRESENT(BIC_PKG__);
4401 BIC_PRESENT(BIC_RAM__);
4403 BIC_PRESENT(BIC_Pkg_J);
4404 BIC_PRESENT(BIC_Cor_J);
4405 BIC_PRESENT(BIC_RAM_J);
4407 BIC_PRESENT(BIC_PkgWatt);
4408 BIC_PRESENT(BIC_CorWatt);
4409 BIC_PRESENT(BIC_RAMWatt);
4416 /* units on package 0, verify later other packages match */
4417 if (get_msr(base_cpu, MSR_RAPL_POWER_UNIT, &msr))
4420 rapl_power_units = 1.0 / (1 << (msr & 0xF));
4421 if (model == INTEL_FAM6_ATOM_SILVERMONT)
4422 rapl_energy_units = 1.0 * (1 << (msr >> 8 & 0x1F)) / 1000000;
4424 rapl_energy_units = 1.0 / (1 << (msr >> 8 & 0x1F));
4426 rapl_dram_energy_units = rapl_dram_energy_units_probe(model, rapl_energy_units);
4428 time_unit = msr >> 16 & 0xF;
4432 rapl_time_units = 1.0 / (1 << (time_unit));
4434 tdp = get_tdp_intel(model);
4436 rapl_joule_counter_range = 0xFFFFFFFF * rapl_energy_units / tdp;
4438 fprintf(outf, "RAPL: %.0f sec. Joule Counter Range, at %.0f Watts\n", rapl_joule_counter_range, tdp);
4441 void rapl_probe_amd(unsigned int family, unsigned int model)
4443 unsigned long long msr;
4444 unsigned int eax, ebx, ecx, edx;
4445 unsigned int has_rapl = 0;
4448 if (max_extended_level >= 0x80000007) {
4449 __cpuid(0x80000007, eax, ebx, ecx, edx);
4450 /* RAPL (Fam 17h+) */
4451 has_rapl = edx & (1 << 14);
4454 if (!has_rapl || family < 0x17)
4457 do_rapl = RAPL_AMD_F17H | RAPL_PER_CORE_ENERGY;
4459 BIC_PRESENT(BIC_Pkg_J);
4460 BIC_PRESENT(BIC_Cor_J);
4462 BIC_PRESENT(BIC_PkgWatt);
4463 BIC_PRESENT(BIC_CorWatt);
4466 if (get_msr(base_cpu, MSR_RAPL_PWR_UNIT, &msr))
4469 rapl_time_units = ldexp(1.0, -(msr >> 16 & 0xf));
4470 rapl_energy_units = ldexp(1.0, -(msr >> 8 & 0x1f));
4471 rapl_power_units = ldexp(1.0, -(msr & 0xf));
4473 tdp = get_tdp_amd(family);
4475 rapl_joule_counter_range = 0xFFFFFFFF * rapl_energy_units / tdp;
4477 fprintf(outf, "RAPL: %.0f sec. Joule Counter Range, at %.0f Watts\n", rapl_joule_counter_range, tdp);
4483 * sets do_rapl, rapl_power_units, rapl_energy_units, rapl_time_units
4485 void rapl_probe(unsigned int family, unsigned int model)
4488 rapl_probe_intel(family, model);
4489 if (authentic_amd || hygon_genuine)
4490 rapl_probe_amd(family, model);
4493 void perf_limit_reasons_probe(unsigned int family, unsigned int model)
4502 case INTEL_FAM6_HASWELL: /* HSW */
4503 case INTEL_FAM6_HASWELL_L: /* HSW */
4504 case INTEL_FAM6_HASWELL_G: /* HSW */
4505 do_gfx_perf_limit_reasons = 1;
4506 case INTEL_FAM6_HASWELL_X: /* HSX */
4507 do_core_perf_limit_reasons = 1;
4508 do_ring_perf_limit_reasons = 1;
4514 void automatic_cstate_conversion_probe(unsigned int family, unsigned int model)
4516 if (is_skx(family, model) || is_bdx(family, model))
4517 has_automatic_cstate_conversion = 1;
4520 int print_thermal(struct thread_data *t, struct core_data *c, struct pkg_data *p)
4522 unsigned long long msr;
4523 unsigned int dts, dts2;
4526 if (!(do_dts || do_ptm))
4531 /* DTS is per-core, no need to print for each thread */
4532 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE))
4535 if (cpu_migrate(cpu)) {
4536 fprintf(outf, "print_thermal: Could not migrate to CPU %d\n", cpu);
4540 if (do_ptm && (t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE)) {
4541 if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_STATUS, &msr))
4544 dts = (msr >> 16) & 0x7F;
4545 fprintf(outf, "cpu%d: MSR_IA32_PACKAGE_THERM_STATUS: 0x%08llx (%d C)\n",
4546 cpu, msr, tcc_activation_temp - dts);
4548 if (get_msr(cpu, MSR_IA32_PACKAGE_THERM_INTERRUPT, &msr))
4551 dts = (msr >> 16) & 0x7F;
4552 dts2 = (msr >> 8) & 0x7F;
4553 fprintf(outf, "cpu%d: MSR_IA32_PACKAGE_THERM_INTERRUPT: 0x%08llx (%d C, %d C)\n",
4554 cpu, msr, tcc_activation_temp - dts, tcc_activation_temp - dts2);
4558 if (do_dts && debug) {
4559 unsigned int resolution;
4561 if (get_msr(cpu, MSR_IA32_THERM_STATUS, &msr))
4564 dts = (msr >> 16) & 0x7F;
4565 resolution = (msr >> 27) & 0xF;
4566 fprintf(outf, "cpu%d: MSR_IA32_THERM_STATUS: 0x%08llx (%d C +/- %d)\n",
4567 cpu, msr, tcc_activation_temp - dts, resolution);
4569 if (get_msr(cpu, MSR_IA32_THERM_INTERRUPT, &msr))
4572 dts = (msr >> 16) & 0x7F;
4573 dts2 = (msr >> 8) & 0x7F;
4574 fprintf(outf, "cpu%d: MSR_IA32_THERM_INTERRUPT: 0x%08llx (%d C, %d C)\n",
4575 cpu, msr, tcc_activation_temp - dts, tcc_activation_temp - dts2);
4581 void print_power_limit_msr(int cpu, unsigned long long msr, char *label)
4583 fprintf(outf, "cpu%d: %s: %sabled (%f Watts, %f sec, clamp %sabled)\n",
4585 ((msr >> 15) & 1) ? "EN" : "DIS",
4586 ((msr >> 0) & 0x7FFF) * rapl_power_units,
4587 (1.0 + (((msr >> 22) & 0x3)/4.0)) * (1 << ((msr >> 17) & 0x1F)) * rapl_time_units,
4588 (((msr >> 16) & 1) ? "EN" : "DIS"));
4593 int print_rapl(struct thread_data *t, struct core_data *c, struct pkg_data *p)
4595 unsigned long long msr;
4596 const char *msr_name;
4602 /* RAPL counters are per package, so print only for 1st thread/package */
4603 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
4607 if (cpu_migrate(cpu)) {
4608 fprintf(outf, "print_rapl: Could not migrate to CPU %d\n", cpu);
4612 if (do_rapl & RAPL_AMD_F17H) {
4613 msr_name = "MSR_RAPL_PWR_UNIT";
4614 if (get_msr(cpu, MSR_RAPL_PWR_UNIT, &msr))
4617 msr_name = "MSR_RAPL_POWER_UNIT";
4618 if (get_msr(cpu, MSR_RAPL_POWER_UNIT, &msr))
4622 fprintf(outf, "cpu%d: %s: 0x%08llx (%f Watts, %f Joules, %f sec.)\n", cpu, msr_name, msr,
4623 rapl_power_units, rapl_energy_units, rapl_time_units);
4625 if (do_rapl & RAPL_PKG_POWER_INFO) {
4627 if (get_msr(cpu, MSR_PKG_POWER_INFO, &msr))
4631 fprintf(outf, "cpu%d: MSR_PKG_POWER_INFO: 0x%08llx (%.0f W TDP, RAPL %.0f - %.0f W, %f sec.)\n",
4633 ((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units,
4634 ((msr >> 16) & RAPL_POWER_GRANULARITY) * rapl_power_units,
4635 ((msr >> 32) & RAPL_POWER_GRANULARITY) * rapl_power_units,
4636 ((msr >> 48) & RAPL_TIME_GRANULARITY) * rapl_time_units);
4639 if (do_rapl & RAPL_PKG) {
4641 if (get_msr(cpu, MSR_PKG_POWER_LIMIT, &msr))
4644 fprintf(outf, "cpu%d: MSR_PKG_POWER_LIMIT: 0x%08llx (%slocked)\n",
4645 cpu, msr, (msr >> 63) & 1 ? "" : "UN");
4647 print_power_limit_msr(cpu, msr, "PKG Limit #1");
4648 fprintf(outf, "cpu%d: PKG Limit #2: %sabled (%f Watts, %f* sec, clamp %sabled)\n",
4650 ((msr >> 47) & 1) ? "EN" : "DIS",
4651 ((msr >> 32) & 0x7FFF) * rapl_power_units,
4652 (1.0 + (((msr >> 54) & 0x3)/4.0)) * (1 << ((msr >> 49) & 0x1F)) * rapl_time_units,
4653 ((msr >> 48) & 1) ? "EN" : "DIS");
4656 if (do_rapl & RAPL_DRAM_POWER_INFO) {
4657 if (get_msr(cpu, MSR_DRAM_POWER_INFO, &msr))
4660 fprintf(outf, "cpu%d: MSR_DRAM_POWER_INFO,: 0x%08llx (%.0f W TDP, RAPL %.0f - %.0f W, %f sec.)\n",
4662 ((msr >> 0) & RAPL_POWER_GRANULARITY) * rapl_power_units,
4663 ((msr >> 16) & RAPL_POWER_GRANULARITY) * rapl_power_units,
4664 ((msr >> 32) & RAPL_POWER_GRANULARITY) * rapl_power_units,
4665 ((msr >> 48) & RAPL_TIME_GRANULARITY) * rapl_time_units);
4667 if (do_rapl & RAPL_DRAM) {
4668 if (get_msr(cpu, MSR_DRAM_POWER_LIMIT, &msr))
4670 fprintf(outf, "cpu%d: MSR_DRAM_POWER_LIMIT: 0x%08llx (%slocked)\n",
4671 cpu, msr, (msr >> 31) & 1 ? "" : "UN");
4673 print_power_limit_msr(cpu, msr, "DRAM Limit");
4675 if (do_rapl & RAPL_CORE_POLICY) {
4676 if (get_msr(cpu, MSR_PP0_POLICY, &msr))
4679 fprintf(outf, "cpu%d: MSR_PP0_POLICY: %lld\n", cpu, msr & 0xF);
4681 if (do_rapl & RAPL_CORES_POWER_LIMIT) {
4682 if (get_msr(cpu, MSR_PP0_POWER_LIMIT, &msr))
4684 fprintf(outf, "cpu%d: MSR_PP0_POWER_LIMIT: 0x%08llx (%slocked)\n",
4685 cpu, msr, (msr >> 31) & 1 ? "" : "UN");
4686 print_power_limit_msr(cpu, msr, "Cores Limit");
4688 if (do_rapl & RAPL_GFX) {
4689 if (get_msr(cpu, MSR_PP1_POLICY, &msr))
4692 fprintf(outf, "cpu%d: MSR_PP1_POLICY: %lld\n", cpu, msr & 0xF);
4694 if (get_msr(cpu, MSR_PP1_POWER_LIMIT, &msr))
4696 fprintf(outf, "cpu%d: MSR_PP1_POWER_LIMIT: 0x%08llx (%slocked)\n",
4697 cpu, msr, (msr >> 31) & 1 ? "" : "UN");
4698 print_power_limit_msr(cpu, msr, "GFX Limit");
4704 * SNB adds support for additional MSRs:
4706 * MSR_PKG_C7_RESIDENCY 0x000003fa
4707 * MSR_CORE_C7_RESIDENCY 0x000003fe
4708 * MSR_PKG_C2_RESIDENCY 0x0000060d
4711 int has_snb_msrs(unsigned int family, unsigned int model)
4717 case INTEL_FAM6_SANDYBRIDGE:
4718 case INTEL_FAM6_SANDYBRIDGE_X:
4719 case INTEL_FAM6_IVYBRIDGE: /* IVB */
4720 case INTEL_FAM6_IVYBRIDGE_X: /* IVB Xeon */
4721 case INTEL_FAM6_HASWELL: /* HSW */
4722 case INTEL_FAM6_HASWELL_X: /* HSW */
4723 case INTEL_FAM6_HASWELL_L: /* HSW */
4724 case INTEL_FAM6_HASWELL_G: /* HSW */
4725 case INTEL_FAM6_BROADWELL: /* BDW */
4726 case INTEL_FAM6_BROADWELL_G: /* BDW */
4727 case INTEL_FAM6_BROADWELL_X: /* BDX */
4728 case INTEL_FAM6_SKYLAKE_L: /* SKL */
4729 case INTEL_FAM6_CANNONLAKE_L: /* CNL */
4730 case INTEL_FAM6_SKYLAKE_X: /* SKX */
4731 case INTEL_FAM6_ATOM_GOLDMONT: /* BXT */
4732 case INTEL_FAM6_ATOM_GOLDMONT_PLUS:
4733 case INTEL_FAM6_ATOM_GOLDMONT_D: /* DNV */
4734 case INTEL_FAM6_ATOM_TREMONT: /* EHL */
4735 case INTEL_FAM6_ATOM_TREMONT_D: /* JVL */
4742 * HSW ULT added support for C8/C9/C10 MSRs:
4744 * MSR_PKG_C8_RESIDENCY 0x00000630
4745 * MSR_PKG_C9_RESIDENCY 0x00000631
4746 * MSR_PKG_C10_RESIDENCY 0x00000632
4748 * MSR_PKGC8_IRTL 0x00000633
4749 * MSR_PKGC9_IRTL 0x00000634
4750 * MSR_PKGC10_IRTL 0x00000635
4753 int has_c8910_msrs(unsigned int family, unsigned int model)
4759 case INTEL_FAM6_HASWELL_L: /* HSW */
4760 case INTEL_FAM6_BROADWELL: /* BDW */
4761 case INTEL_FAM6_SKYLAKE_L: /* SKL */
4762 case INTEL_FAM6_CANNONLAKE_L: /* CNL */
4763 case INTEL_FAM6_ATOM_GOLDMONT: /* BXT */
4764 case INTEL_FAM6_ATOM_GOLDMONT_PLUS:
4765 case INTEL_FAM6_ATOM_TREMONT: /* EHL */
4772 * SKL adds support for additional MSRS:
4774 * MSR_PKG_WEIGHTED_CORE_C0_RES 0x00000658
4775 * MSR_PKG_ANY_CORE_C0_RES 0x00000659
4776 * MSR_PKG_ANY_GFXE_C0_RES 0x0000065A
4777 * MSR_PKG_BOTH_CORE_GFXE_C0_RES 0x0000065B
4779 int has_skl_msrs(unsigned int family, unsigned int model)
4785 case INTEL_FAM6_SKYLAKE_L: /* SKL */
4786 case INTEL_FAM6_CANNONLAKE_L: /* CNL */
4792 int is_slm(unsigned int family, unsigned int model)
4797 case INTEL_FAM6_ATOM_SILVERMONT: /* BYT */
4798 case INTEL_FAM6_ATOM_SILVERMONT_D: /* AVN */
4804 int is_knl(unsigned int family, unsigned int model)
4809 case INTEL_FAM6_XEON_PHI_KNL: /* KNL */
4815 int is_cnl(unsigned int family, unsigned int model)
4821 case INTEL_FAM6_CANNONLAKE_L: /* CNL */
4828 unsigned int get_aperf_mperf_multiplier(unsigned int family, unsigned int model)
4830 if (is_knl(family, model))
4835 #define SLM_BCLK_FREQS 5
4836 double slm_freq_table[SLM_BCLK_FREQS] = { 83.3, 100.0, 133.3, 116.7, 80.0};
4838 double slm_bclk(void)
4840 unsigned long long msr = 3;
4844 if (get_msr(base_cpu, MSR_FSB_FREQ, &msr))
4845 fprintf(outf, "SLM BCLK: unknown\n");
4848 if (i >= SLM_BCLK_FREQS) {
4849 fprintf(outf, "SLM BCLK[%d] invalid\n", i);
4852 freq = slm_freq_table[i];
4855 fprintf(outf, "SLM BCLK: %.1f Mhz\n", freq);
4860 double discover_bclk(unsigned int family, unsigned int model)
4862 if (has_snb_msrs(family, model) || is_knl(family, model))
4864 else if (is_slm(family, model))
4871 * MSR_IA32_TEMPERATURE_TARGET indicates the temperature where
4872 * the Thermal Control Circuit (TCC) activates.
4873 * This is usually equal to tjMax.
4875 * Older processors do not have this MSR, so there we guess,
4876 * but also allow cmdline over-ride with -T.
4878 * Several MSR temperature values are in units of degrees-C
4879 * below this value, including the Digital Thermal Sensor (DTS),
4880 * Package Thermal Management Sensor (PTM), and thermal event thresholds.
4882 int read_tcc_activation_temp()
4884 unsigned long long msr;
4885 unsigned int tcc, target_c, offset_c;
4887 /* Temperature Target MSR is Nehalem and newer only */
4888 if (!do_nhm_platform_info)
4891 if (get_msr(base_cpu, MSR_IA32_TEMPERATURE_TARGET, &msr))
4894 target_c = (msr >> 16) & 0xFF;
4896 offset_c = (msr >> 24) & 0xF;
4898 tcc = target_c - offset_c;
4901 fprintf(outf, "cpu%d: MSR_IA32_TEMPERATURE_TARGET: 0x%08llx (%d C) (%d default - %d offset)\n",
4902 base_cpu, msr, tcc, target_c, offset_c);
4907 int set_temperature_target(struct thread_data *t, struct core_data *c, struct pkg_data *p)
4909 /* tcc_activation_temp is used only for dts or ptm */
4910 if (!(do_dts || do_ptm))
4913 /* this is a per-package concept */
4914 if (!(t->flags & CPU_IS_FIRST_THREAD_IN_CORE) || !(t->flags & CPU_IS_FIRST_CORE_IN_PACKAGE))
4917 if (tcc_activation_temp_override != 0) {
4918 tcc_activation_temp = tcc_activation_temp_override;
4919 fprintf(outf, "Using cmdline TCC Target (%d C)\n", tcc_activation_temp);
4923 tcc_activation_temp = read_tcc_activation_temp();
4924 if (tcc_activation_temp)
4927 tcc_activation_temp = TJMAX_DEFAULT;
4928 fprintf(outf, "Guessing tjMax %d C, Please use -T to specify\n", tcc_activation_temp);
4933 void decode_feature_control_msr(void)
4935 unsigned long long msr;
4937 if (!get_msr(base_cpu, MSR_IA32_FEAT_CTL, &msr))
4938 fprintf(outf, "cpu%d: MSR_IA32_FEATURE_CONTROL: 0x%08llx (%sLocked %s)\n",
4940 msr & FEAT_CTL_LOCKED ? "" : "UN-",
4941 msr & (1 << 18) ? "SGX" : "");
4944 void decode_misc_enable_msr(void)
4946 unsigned long long msr;
4951 if (!get_msr(base_cpu, MSR_IA32_MISC_ENABLE, &msr))
4952 fprintf(outf, "cpu%d: MSR_IA32_MISC_ENABLE: 0x%08llx (%sTCC %sEIST %sMWAIT %sPREFETCH %sTURBO)\n",
4954 msr & MSR_IA32_MISC_ENABLE_TM1 ? "" : "No-",
4955 msr & MSR_IA32_MISC_ENABLE_ENHANCED_SPEEDSTEP ? "" : "No-",
4956 msr & MSR_IA32_MISC_ENABLE_MWAIT ? "" : "No-",
4957 msr & MSR_IA32_MISC_ENABLE_PREFETCH_DISABLE ? "No-" : "",
4958 msr & MSR_IA32_MISC_ENABLE_TURBO_DISABLE ? "No-" : "");
4961 void decode_misc_feature_control(void)
4963 unsigned long long msr;
4965 if (!has_misc_feature_control)
4968 if (!get_msr(base_cpu, MSR_MISC_FEATURE_CONTROL, &msr))
4969 fprintf(outf, "cpu%d: MSR_MISC_FEATURE_CONTROL: 0x%08llx (%sL2-Prefetch %sL2-Prefetch-pair %sL1-Prefetch %sL1-IP-Prefetch)\n",
4971 msr & (0 << 0) ? "No-" : "",
4972 msr & (1 << 0) ? "No-" : "",
4973 msr & (2 << 0) ? "No-" : "",
4974 msr & (3 << 0) ? "No-" : "");
4977 * Decode MSR_MISC_PWR_MGMT
4979 * Decode the bits according to the Nehalem documentation
4980 * bit[0] seems to continue to have same meaning going forward
4983 void decode_misc_pwr_mgmt_msr(void)
4985 unsigned long long msr;
4987 if (!do_nhm_platform_info)
4990 if (no_MSR_MISC_PWR_MGMT)
4993 if (!get_msr(base_cpu, MSR_MISC_PWR_MGMT, &msr))
4994 fprintf(outf, "cpu%d: MSR_MISC_PWR_MGMT: 0x%08llx (%sable-EIST_Coordination %sable-EPB %sable-OOB)\n",
4996 msr & (1 << 0) ? "DIS" : "EN",
4997 msr & (1 << 1) ? "EN" : "DIS",
4998 msr & (1 << 8) ? "EN" : "DIS");
5001 * Decode MSR_CC6_DEMOTION_POLICY_CONFIG, MSR_MC6_DEMOTION_POLICY_CONFIG
5003 * This MSRs are present on Silvermont processors,
5004 * Intel Atom processor E3000 series (Baytrail), and friends.
5006 void decode_c6_demotion_policy_msr(void)
5008 unsigned long long msr;
5010 if (!get_msr(base_cpu, MSR_CC6_DEMOTION_POLICY_CONFIG, &msr))
5011 fprintf(outf, "cpu%d: MSR_CC6_DEMOTION_POLICY_CONFIG: 0x%08llx (%sable-CC6-Demotion)\n",
5012 base_cpu, msr, msr & (1 << 0) ? "EN" : "DIS");
5014 if (!get_msr(base_cpu, MSR_MC6_DEMOTION_POLICY_CONFIG, &msr))
5015 fprintf(outf, "cpu%d: MSR_MC6_DEMOTION_POLICY_CONFIG: 0x%08llx (%sable-MC6-Demotion)\n",
5016 base_cpu, msr, msr & (1 << 0) ? "EN" : "DIS");
5020 * When models are the same, for the purpose of turbostat, reuse
5022 unsigned int intel_model_duplicates(unsigned int model)
5026 case INTEL_FAM6_NEHALEM_EP: /* Core i7, Xeon 5500 series - Bloomfield, Gainstown NHM-EP */
5027 case INTEL_FAM6_NEHALEM: /* Core i7 and i5 Processor - Clarksfield, Lynnfield, Jasper Forest */
5028 case 0x1F: /* Core i7 and i5 Processor - Nehalem */
5029 case INTEL_FAM6_WESTMERE: /* Westmere Client - Clarkdale, Arrandale */
5030 case INTEL_FAM6_WESTMERE_EP: /* Westmere EP - Gulftown */
5031 return INTEL_FAM6_NEHALEM;
5033 case INTEL_FAM6_NEHALEM_EX: /* Nehalem-EX Xeon - Beckton */
5034 case INTEL_FAM6_WESTMERE_EX: /* Westmere-EX Xeon - Eagleton */
5035 return INTEL_FAM6_NEHALEM_EX;
5037 case INTEL_FAM6_XEON_PHI_KNM:
5038 return INTEL_FAM6_XEON_PHI_KNL;
5040 case INTEL_FAM6_BROADWELL_X:
5041 case INTEL_FAM6_BROADWELL_D: /* BDX-DE */
5042 return INTEL_FAM6_BROADWELL_X;
5044 case INTEL_FAM6_SKYLAKE_L:
5045 case INTEL_FAM6_SKYLAKE:
5046 case INTEL_FAM6_KABYLAKE_L:
5047 case INTEL_FAM6_KABYLAKE:
5048 case INTEL_FAM6_COMETLAKE_L:
5049 case INTEL_FAM6_COMETLAKE:
5050 return INTEL_FAM6_SKYLAKE_L;
5052 case INTEL_FAM6_ICELAKE_L:
5053 case INTEL_FAM6_ICELAKE_NNPI:
5054 case INTEL_FAM6_TIGERLAKE_L:
5055 case INTEL_FAM6_TIGERLAKE:
5056 case INTEL_FAM6_ROCKETLAKE:
5057 case INTEL_FAM6_LAKEFIELD:
5058 case INTEL_FAM6_ALDERLAKE:
5059 return INTEL_FAM6_CANNONLAKE_L;
5061 case INTEL_FAM6_ATOM_TREMONT_L:
5062 return INTEL_FAM6_ATOM_TREMONT;
5064 case INTEL_FAM6_ICELAKE_X:
5065 case INTEL_FAM6_SAPPHIRERAPIDS_X:
5066 return INTEL_FAM6_SKYLAKE_X;
5071 void print_dev_latency(void)
5073 char *path = "/dev/cpu_dma_latency";
5078 fd = open(path, O_RDONLY);
5080 warn("fopen %s\n", path);
5084 retval = read(fd, (void *)&value, sizeof(int));
5085 if (retval != sizeof(int)) {
5086 warn("read %s\n", path);
5090 fprintf(outf, "/dev/cpu_dma_latency: %d usec (%s)\n",
5091 value, value == 2000000000 ? "default" : "constrained");
5098 * Linux-perf manages the the HW instructions-retired counter
5099 * by enabling when requested, and hiding rollover
5101 void linux_perf_init(void)
5103 if (!BIC_IS_ENABLED(BIC_IPC))
5106 if (access("/proc/sys/kernel/perf_event_paranoid", F_OK))
5109 fd_instr_count_percpu = calloc(topo.max_cpu_num + 1, sizeof(int));
5110 if (fd_instr_count_percpu == NULL)
5111 err(-1, "calloc fd_instr_count_percpu");
5113 BIC_PRESENT(BIC_IPC);
5116 void process_cpuid()
5118 unsigned int eax, ebx, ecx, edx;
5119 unsigned int fms, family, model, stepping, ecx_flags, edx_flags;
5120 unsigned int has_turbo;
5121 unsigned long long ucode_patch = 0;
5123 eax = ebx = ecx = edx = 0;
5125 __cpuid(0, max_level, ebx, ecx, edx);
5127 if (ebx == 0x756e6547 && ecx == 0x6c65746e && edx == 0x49656e69)
5129 else if (ebx == 0x68747541 && ecx == 0x444d4163 && edx == 0x69746e65)
5131 else if (ebx == 0x6f677948 && ecx == 0x656e6975 && edx == 0x6e65476e)
5135 fprintf(outf, "CPUID(0): %.4s%.4s%.4s 0x%x CPUID levels\n",
5136 (char *)&ebx, (char *)&edx, (char *)&ecx, max_level);
5138 __cpuid(1, fms, ebx, ecx, edx);
5139 family = (fms >> 8) & 0xf;
5140 model = (fms >> 4) & 0xf;
5141 stepping = fms & 0xf;
5143 family += (fms >> 20) & 0xff;
5145 model += ((fms >> 16) & 0xf) << 4;
5149 if (get_msr(sched_getcpu(), MSR_IA32_UCODE_REV, &ucode_patch))
5150 warnx("get_msr(UCODE)\n");
5153 * check max extended function levels of CPUID.
5154 * This is needed to check for invariant TSC.
5155 * This check is valid for both Intel and AMD.
5157 ebx = ecx = edx = 0;
5158 __cpuid(0x80000000, max_extended_level, ebx, ecx, edx);
5161 fprintf(outf, "CPUID(1): family:model:stepping 0x%x:%x:%x (%d:%d:%d) microcode 0x%x\n",
5162 family, model, stepping, family, model, stepping, (unsigned int)((ucode_patch >> 32) & 0xFFFFFFFF));
5163 fprintf(outf, "CPUID(0x80000000): max_extended_levels: 0x%x\n", max_extended_level);
5164 fprintf(outf, "CPUID(1): %s %s %s %s %s %s %s %s %s %s\n",
5165 ecx_flags & (1 << 0) ? "SSE3" : "-",
5166 ecx_flags & (1 << 3) ? "MONITOR" : "-",
5167 ecx_flags & (1 << 6) ? "SMX" : "-",
5168 ecx_flags & (1 << 7) ? "EIST" : "-",
5169 ecx_flags & (1 << 8) ? "TM2" : "-",
5170 edx_flags & (1 << 4) ? "TSC" : "-",
5171 edx_flags & (1 << 5) ? "MSR" : "-",
5172 edx_flags & (1 << 22) ? "ACPI-TM" : "-",
5173 edx_flags & (1 << 28) ? "HT" : "-",
5174 edx_flags & (1 << 29) ? "TM" : "-");
5177 model = intel_model_duplicates(model);
5179 if (!(edx_flags & (1 << 5)))
5180 errx(1, "CPUID: no MSR");
5182 if (max_extended_level >= 0x80000007) {
5185 * Non-Stop TSC is advertised by CPUID.EAX=0x80000007: EDX.bit8
5186 * this check is valid for both Intel and AMD
5188 __cpuid(0x80000007, eax, ebx, ecx, edx);
5189 has_invariant_tsc = edx & (1 << 8);
5193 * APERF/MPERF is advertised by CPUID.EAX=0x6: ECX.bit0
5194 * this check is valid for both Intel and AMD
5197 __cpuid(0x6, eax, ebx, ecx, edx);
5198 has_aperf = ecx & (1 << 0);
5200 BIC_PRESENT(BIC_Avg_MHz);
5201 BIC_PRESENT(BIC_Busy);
5202 BIC_PRESENT(BIC_Bzy_MHz);
5204 do_dts = eax & (1 << 0);
5206 BIC_PRESENT(BIC_CoreTmp);
5207 has_turbo = eax & (1 << 1);
5208 do_ptm = eax & (1 << 6);
5210 BIC_PRESENT(BIC_PkgTmp);
5211 has_hwp = eax & (1 << 7);
5212 has_hwp_notify = eax & (1 << 8);
5213 has_hwp_activity_window = eax & (1 << 9);
5214 has_hwp_epp = eax & (1 << 10);
5215 has_hwp_pkg = eax & (1 << 11);
5216 has_epb = ecx & (1 << 3);
5219 fprintf(outf, "CPUID(6): %sAPERF, %sTURBO, %sDTS, %sPTM, %sHWP, "
5220 "%sHWPnotify, %sHWPwindow, %sHWPepp, %sHWPpkg, %sEPB\n",
5221 has_aperf ? "" : "No-",
5222 has_turbo ? "" : "No-",
5223 do_dts ? "" : "No-",
5224 do_ptm ? "" : "No-",
5225 has_hwp ? "" : "No-",
5226 has_hwp_notify ? "" : "No-",
5227 has_hwp_activity_window ? "" : "No-",
5228 has_hwp_epp ? "" : "No-",
5229 has_hwp_pkg ? "" : "No-",
5230 has_epb ? "" : "No-");
5233 decode_misc_enable_msr();
5236 if (max_level >= 0x7 && !quiet) {
5241 __cpuid_count(0x7, 0, eax, ebx, ecx, edx);
5243 has_sgx = ebx & (1 << 2);
5244 fprintf(outf, "CPUID(7): %sSGX\n", has_sgx ? "" : "No-");
5247 decode_feature_control_msr();
5250 if (max_level >= 0x15) {
5251 unsigned int eax_crystal;
5252 unsigned int ebx_tsc;
5255 * CPUID 15H TSC/Crystal ratio, possibly Crystal Hz
5257 eax_crystal = ebx_tsc = crystal_hz = edx = 0;
5258 __cpuid(0x15, eax_crystal, ebx_tsc, crystal_hz, edx);
5262 if (!quiet && (ebx != 0))
5263 fprintf(outf, "CPUID(0x15): eax_crystal: %d ebx_tsc: %d ecx_crystal_hz: %d\n",
5264 eax_crystal, ebx_tsc, crystal_hz);
5266 if (crystal_hz == 0)
5268 case INTEL_FAM6_SKYLAKE_L: /* SKL */
5269 crystal_hz = 24000000; /* 24.0 MHz */
5271 case INTEL_FAM6_ATOM_GOLDMONT_D: /* DNV */
5272 crystal_hz = 25000000; /* 25.0 MHz */
5274 case INTEL_FAM6_ATOM_GOLDMONT: /* BXT */
5275 case INTEL_FAM6_ATOM_GOLDMONT_PLUS:
5276 crystal_hz = 19200000; /* 19.2 MHz */
5283 tsc_hz = (unsigned long long) crystal_hz * ebx_tsc / eax_crystal;
5285 fprintf(outf, "TSC: %lld MHz (%d Hz * %d / %d / 1000000)\n",
5286 tsc_hz / 1000000, crystal_hz, ebx_tsc, eax_crystal);
5290 if (max_level >= 0x16) {
5291 unsigned int base_mhz, max_mhz, bus_mhz, edx;
5294 * CPUID 16H Base MHz, Max MHz, Bus MHz
5296 base_mhz = max_mhz = bus_mhz = edx = 0;
5298 __cpuid(0x16, base_mhz, max_mhz, bus_mhz, edx);
5300 fprintf(outf, "CPUID(0x16): base_mhz: %d max_mhz: %d bus_mhz: %d\n",
5301 base_mhz, max_mhz, bus_mhz);
5305 aperf_mperf_multiplier = get_aperf_mperf_multiplier(family, model);
5307 BIC_PRESENT(BIC_IRQ);
5308 BIC_PRESENT(BIC_TSC_MHz);
5310 if (probe_nhm_msrs(family, model)) {
5311 do_nhm_platform_info = 1;
5312 BIC_PRESENT(BIC_CPU_c1);
5313 BIC_PRESENT(BIC_CPU_c3);
5314 BIC_PRESENT(BIC_CPU_c6);
5315 BIC_PRESENT(BIC_SMI);
5317 do_snb_cstates = has_snb_msrs(family, model);
5320 BIC_PRESENT(BIC_CPU_c7);
5322 do_irtl_snb = has_snb_msrs(family, model);
5323 if (do_snb_cstates && (pkg_cstate_limit >= PCL__2))
5324 BIC_PRESENT(BIC_Pkgpc2);
5325 if (pkg_cstate_limit >= PCL__3)
5326 BIC_PRESENT(BIC_Pkgpc3);
5327 if (pkg_cstate_limit >= PCL__6)
5328 BIC_PRESENT(BIC_Pkgpc6);
5329 if (do_snb_cstates && (pkg_cstate_limit >= PCL__7))
5330 BIC_PRESENT(BIC_Pkgpc7);
5331 if (has_slv_msrs(family, model)) {
5332 BIC_NOT_PRESENT(BIC_Pkgpc2);
5333 BIC_NOT_PRESENT(BIC_Pkgpc3);
5334 BIC_PRESENT(BIC_Pkgpc6);
5335 BIC_NOT_PRESENT(BIC_Pkgpc7);
5336 BIC_PRESENT(BIC_Mod_c6);
5337 use_c1_residency_msr = 1;
5339 if (is_jvl(family, model)) {
5340 BIC_NOT_PRESENT(BIC_CPU_c3);
5341 BIC_NOT_PRESENT(BIC_CPU_c7);
5342 BIC_NOT_PRESENT(BIC_Pkgpc2);
5343 BIC_NOT_PRESENT(BIC_Pkgpc3);
5344 BIC_NOT_PRESENT(BIC_Pkgpc6);
5345 BIC_NOT_PRESENT(BIC_Pkgpc7);
5347 if (is_dnv(family, model)) {
5348 BIC_PRESENT(BIC_CPU_c1);
5349 BIC_NOT_PRESENT(BIC_CPU_c3);
5350 BIC_NOT_PRESENT(BIC_Pkgpc3);
5351 BIC_NOT_PRESENT(BIC_CPU_c7);
5352 BIC_NOT_PRESENT(BIC_Pkgpc7);
5353 use_c1_residency_msr = 1;
5355 if (is_skx(family, model)) {
5356 BIC_NOT_PRESENT(BIC_CPU_c3);
5357 BIC_NOT_PRESENT(BIC_Pkgpc3);
5358 BIC_NOT_PRESENT(BIC_CPU_c7);
5359 BIC_NOT_PRESENT(BIC_Pkgpc7);
5361 if (is_bdx(family, model)) {
5362 BIC_NOT_PRESENT(BIC_CPU_c7);
5363 BIC_NOT_PRESENT(BIC_Pkgpc7);
5365 if (has_c8910_msrs(family, model)) {
5366 if (pkg_cstate_limit >= PCL__8)
5367 BIC_PRESENT(BIC_Pkgpc8);
5368 if (pkg_cstate_limit >= PCL__9)
5369 BIC_PRESENT(BIC_Pkgpc9);
5370 if (pkg_cstate_limit >= PCL_10)
5371 BIC_PRESENT(BIC_Pkgpc10);
5373 do_irtl_hsw = has_c8910_msrs(family, model);
5374 if (has_skl_msrs(family, model)) {
5375 BIC_PRESENT(BIC_Totl_c0);
5376 BIC_PRESENT(BIC_Any_c0);
5377 BIC_PRESENT(BIC_GFX_c0);
5378 BIC_PRESENT(BIC_CPUGFX);
5380 do_slm_cstates = is_slm(family, model);
5381 do_knl_cstates = is_knl(family, model);
5383 if (do_slm_cstates || do_knl_cstates || is_cnl(family, model) ||
5384 is_ehl(family, model))
5385 BIC_NOT_PRESENT(BIC_CPU_c3);
5388 decode_misc_pwr_mgmt_msr();
5390 if (!quiet && has_slv_msrs(family, model))
5391 decode_c6_demotion_policy_msr();
5393 rapl_probe(family, model);
5394 perf_limit_reasons_probe(family, model);
5395 automatic_cstate_conversion_probe(family, model);
5398 dump_cstate_pstate_config_info(family, model);
5401 print_dev_latency();
5403 dump_sysfs_cstate_config();
5405 dump_sysfs_pstate_config();
5407 if (has_skl_msrs(family, model))
5408 calculate_tsc_tweak();
5410 if (!access("/sys/class/drm/card0/power/rc6_residency_ms", R_OK))
5411 BIC_PRESENT(BIC_GFX_rc6);
5413 if (!access("/sys/class/graphics/fb0/device/drm/card0/gt_cur_freq_mhz", R_OK))
5414 BIC_PRESENT(BIC_GFXMHz);
5416 if (!access("/sys/class/graphics/fb0/device/drm/card0/gt_act_freq_mhz", R_OK))
5417 BIC_PRESENT(BIC_GFXACTMHz);
5419 if (!access("/sys/devices/system/cpu/cpuidle/low_power_idle_cpu_residency_us", R_OK))
5420 BIC_PRESENT(BIC_CPU_LPI);
5422 BIC_NOT_PRESENT(BIC_CPU_LPI);
5424 if (!access(sys_lpi_file_sysfs, R_OK)) {
5425 sys_lpi_file = sys_lpi_file_sysfs;
5426 BIC_PRESENT(BIC_SYS_LPI);
5427 } else if (!access(sys_lpi_file_debugfs, R_OK)) {
5428 sys_lpi_file = sys_lpi_file_debugfs;
5429 BIC_PRESENT(BIC_SYS_LPI);
5431 sys_lpi_file_sysfs = NULL;
5432 BIC_NOT_PRESENT(BIC_SYS_LPI);
5436 decode_misc_feature_control();
5442 * in /dev/cpu/ return success for names that are numbers
5443 * ie. filter out ".", "..", "microcode".
5445 int dir_filter(const struct dirent *dirp)
5447 if (isdigit(dirp->d_name[0]))
5453 int open_dev_cpu_msr(int dummy1)
5458 void topology_probe()
5461 int max_core_id = 0;
5462 int max_package_id = 0;
5464 int max_siblings = 0;
5466 /* Initialize num_cpus, max_cpu_num */
5469 for_all_proc_cpus(count_cpus);
5470 if (!summary_only && topo.num_cpus > 1)
5471 BIC_PRESENT(BIC_CPU);
5474 fprintf(outf, "num_cpus %d max_cpu_num %d\n", topo.num_cpus, topo.max_cpu_num);
5476 cpus = calloc(1, (topo.max_cpu_num + 1) * sizeof(struct cpu_topology));
5478 err(1, "calloc cpus");
5481 * Allocate and initialize cpu_present_set
5483 cpu_present_set = CPU_ALLOC((topo.max_cpu_num + 1));
5484 if (cpu_present_set == NULL)
5485 err(3, "CPU_ALLOC");
5486 cpu_present_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
5487 CPU_ZERO_S(cpu_present_setsize, cpu_present_set);
5488 for_all_proc_cpus(mark_cpu_present);
5491 * Validate that all cpus in cpu_subset are also in cpu_present_set
5493 for (i = 0; i < CPU_SUBSET_MAXCPUS; ++i) {
5494 if (CPU_ISSET_S(i, cpu_subset_size, cpu_subset))
5495 if (!CPU_ISSET_S(i, cpu_present_setsize, cpu_present_set))
5496 err(1, "cpu%d not present", i);
5500 * Allocate and initialize cpu_affinity_set
5502 cpu_affinity_set = CPU_ALLOC((topo.max_cpu_num + 1));
5503 if (cpu_affinity_set == NULL)
5504 err(3, "CPU_ALLOC");
5505 cpu_affinity_setsize = CPU_ALLOC_SIZE((topo.max_cpu_num + 1));
5506 CPU_ZERO_S(cpu_affinity_setsize, cpu_affinity_set);
5508 for_all_proc_cpus(init_thread_id);
5512 * find max_core_id, max_package_id
5514 for (i = 0; i <= topo.max_cpu_num; ++i) {
5517 if (cpu_is_not_present(i)) {
5519 fprintf(outf, "cpu%d NOT PRESENT\n", i);
5523 cpus[i].logical_cpu_id = i;
5525 /* get package information */
5526 cpus[i].physical_package_id = get_physical_package_id(i);
5527 if (cpus[i].physical_package_id > max_package_id)
5528 max_package_id = cpus[i].physical_package_id;
5530 /* get die information */
5531 cpus[i].die_id = get_die_id(i);
5532 if (cpus[i].die_id > max_die_id)
5533 max_die_id = cpus[i].die_id;
5535 /* get numa node information */
5536 cpus[i].physical_node_id = get_physical_node_id(&cpus[i]);
5537 if (cpus[i].physical_node_id > topo.max_node_num)
5538 topo.max_node_num = cpus[i].physical_node_id;
5540 /* get core information */
5541 cpus[i].physical_core_id = get_core_id(i);
5542 if (cpus[i].physical_core_id > max_core_id)
5543 max_core_id = cpus[i].physical_core_id;
5545 /* get thread information */
5546 siblings = get_thread_siblings(&cpus[i]);
5547 if (siblings > max_siblings)
5548 max_siblings = siblings;
5549 if (cpus[i].thread_id == 0)
5553 topo.cores_per_node = max_core_id + 1;
5555 fprintf(outf, "max_core_id %d, sizing for %d cores per package\n",
5556 max_core_id, topo.cores_per_node);
5557 if (!summary_only && topo.cores_per_node > 1)
5558 BIC_PRESENT(BIC_Core);
5560 topo.num_die = max_die_id + 1;
5562 fprintf(outf, "max_die_id %d, sizing for %d die\n",
5563 max_die_id, topo.num_die);
5564 if (!summary_only && topo.num_die > 1)
5565 BIC_PRESENT(BIC_Die);
5567 topo.num_packages = max_package_id + 1;
5569 fprintf(outf, "max_package_id %d, sizing for %d packages\n",
5570 max_package_id, topo.num_packages);
5571 if (!summary_only && topo.num_packages > 1)
5572 BIC_PRESENT(BIC_Package);
5576 fprintf(outf, "nodes_per_pkg %d\n", topo.nodes_per_pkg);
5577 if (!summary_only && topo.nodes_per_pkg > 1)
5578 BIC_PRESENT(BIC_Node);
5580 topo.threads_per_core = max_siblings;
5582 fprintf(outf, "max_siblings %d\n", max_siblings);
5587 for (i = 0; i <= topo.max_cpu_num; ++i) {
5588 if (cpu_is_not_present(i))
5591 "cpu %d pkg %d die %d node %d lnode %d core %d thread %d\n",
5592 i, cpus[i].physical_package_id, cpus[i].die_id,
5593 cpus[i].physical_node_id,
5594 cpus[i].logical_node_id,
5595 cpus[i].physical_core_id,
5602 allocate_counters(struct thread_data **t, struct core_data **c,
5603 struct pkg_data **p)
5606 int num_cores = topo.cores_per_node * topo.nodes_per_pkg *
5608 int num_threads = topo.threads_per_core * num_cores;
5610 *t = calloc(num_threads, sizeof(struct thread_data));
5614 for (i = 0; i < num_threads; i++)
5615 (*t)[i].cpu_id = -1;
5617 *c = calloc(num_cores, sizeof(struct core_data));
5621 for (i = 0; i < num_cores; i++)
5622 (*c)[i].core_id = -1;
5624 *p = calloc(topo.num_packages, sizeof(struct pkg_data));
5628 for (i = 0; i < topo.num_packages; i++)
5629 (*p)[i].package_id = i;
5633 err(1, "calloc counters");
5638 * set FIRST_THREAD_IN_CORE and FIRST_CORE_IN_PACKAGE
5640 void init_counter(struct thread_data *thread_base, struct core_data *core_base,
5641 struct pkg_data *pkg_base, int cpu_id)
5643 int pkg_id = cpus[cpu_id].physical_package_id;
5644 int node_id = cpus[cpu_id].logical_node_id;
5645 int core_id = cpus[cpu_id].physical_core_id;
5646 int thread_id = cpus[cpu_id].thread_id;
5647 struct thread_data *t;
5648 struct core_data *c;
5652 /* Workaround for systems where physical_node_id==-1
5653 * and logical_node_id==(-1 - topo.num_cpus)
5658 t = GET_THREAD(thread_base, thread_id, core_id, node_id, pkg_id);
5659 c = GET_CORE(core_base, core_id, node_id, pkg_id);
5660 p = GET_PKG(pkg_base, pkg_id);
5663 if (thread_id == 0) {
5664 t->flags |= CPU_IS_FIRST_THREAD_IN_CORE;
5665 if (cpu_is_first_core_in_package(cpu_id))
5666 t->flags |= CPU_IS_FIRST_CORE_IN_PACKAGE;
5669 c->core_id = core_id;
5670 p->package_id = pkg_id;
5674 int initialize_counters(int cpu_id)
5676 init_counter(EVEN_COUNTERS, cpu_id);
5677 init_counter(ODD_COUNTERS, cpu_id);
5681 void allocate_output_buffer()
5683 output_buffer = calloc(1, (1 + topo.num_cpus) * 2048);
5684 outp = output_buffer;
5686 err(-1, "calloc output buffer");
5688 void allocate_fd_percpu(void)
5690 fd_percpu = calloc(topo.max_cpu_num + 1, sizeof(int));
5691 if (fd_percpu == NULL)
5692 err(-1, "calloc fd_percpu");
5694 void allocate_irq_buffers(void)
5696 irq_column_2_cpu = calloc(topo.num_cpus, sizeof(int));
5697 if (irq_column_2_cpu == NULL)
5698 err(-1, "calloc %d", topo.num_cpus);
5700 irqs_per_cpu = calloc(topo.max_cpu_num + 1, sizeof(int));
5701 if (irqs_per_cpu == NULL)
5702 err(-1, "calloc %d", topo.max_cpu_num + 1);
5704 void setup_all_buffers(void)
5707 allocate_irq_buffers();
5708 allocate_fd_percpu();
5709 allocate_counters(&thread_even, &core_even, &package_even);
5710 allocate_counters(&thread_odd, &core_odd, &package_odd);
5711 allocate_output_buffer();
5712 for_all_proc_cpus(initialize_counters);
5715 void set_base_cpu(void)
5717 base_cpu = sched_getcpu();
5719 err(-ENODEV, "No valid cpus found");
5722 fprintf(outf, "base_cpu = %d\n", base_cpu);
5725 void turbostat_init()
5727 setup_all_buffers();
5730 check_permissions();
5736 for_all_cpus(print_hwp, ODD_COUNTERS);
5739 for_all_cpus(print_epb, ODD_COUNTERS);
5742 for_all_cpus(print_perf_limit, ODD_COUNTERS);
5745 for_all_cpus(print_rapl, ODD_COUNTERS);
5747 for_all_cpus(set_temperature_target, ODD_COUNTERS);
5750 for_all_cpus(print_thermal, ODD_COUNTERS);
5752 if (!quiet && do_irtl_snb)
5756 int fork_it(char **argv)
5761 snapshot_proc_sysfs_files();
5762 status = for_all_cpus(get_counters, EVEN_COUNTERS);
5763 first_counter_read = 0;
5766 /* clear affinity side-effect of get_counters() */
5767 sched_setaffinity(0, cpu_present_setsize, cpu_present_set);
5768 gettimeofday(&tv_even, (struct timezone *)NULL);
5773 execvp(argv[0], argv);
5774 err(errno, "exec %s", argv[0]);
5778 if (child_pid == -1)
5781 signal(SIGINT, SIG_IGN);
5782 signal(SIGQUIT, SIG_IGN);
5783 if (waitpid(child_pid, &status, 0) == -1)
5784 err(status, "waitpid");
5786 if (WIFEXITED(status))
5787 status = WEXITSTATUS(status);
5790 * n.b. fork_it() does not check for errors from for_all_cpus()
5791 * because re-starting is problematic when forking
5793 snapshot_proc_sysfs_files();
5794 for_all_cpus(get_counters, ODD_COUNTERS);
5795 gettimeofday(&tv_odd, (struct timezone *)NULL);
5796 timersub(&tv_odd, &tv_even, &tv_delta);
5797 if (for_all_cpus_2(delta_cpu, ODD_COUNTERS, EVEN_COUNTERS))
5798 fprintf(outf, "%s: Counter reset detected\n", progname);
5800 compute_average(EVEN_COUNTERS);
5801 format_all_counters(EVEN_COUNTERS);
5804 fprintf(outf, "%.6f sec\n", tv_delta.tv_sec + tv_delta.tv_usec/1000000.0);
5806 flush_output_stderr();
5811 int get_and_dump_counters(void)
5815 snapshot_proc_sysfs_files();
5816 status = for_all_cpus(get_counters, ODD_COUNTERS);
5820 status = for_all_cpus(dump_counters, ODD_COUNTERS);
5824 flush_output_stdout();
5829 void print_version() {
5830 fprintf(outf, "turbostat version 20.09.30"
5831 " - Len Brown <lenb@kernel.org>\n");
5834 int add_counter(unsigned int msr_num, char *path, char *name,
5835 unsigned int width, enum counter_scope scope,
5836 enum counter_type type, enum counter_format format, int flags)
5838 struct msr_counter *msrp;
5840 msrp = calloc(1, sizeof(struct msr_counter));
5846 msrp->msr_num = msr_num;
5847 strncpy(msrp->name, name, NAME_BYTES - 1);
5849 strncpy(msrp->path, path, PATH_BYTES - 1);
5850 msrp->width = width;
5852 msrp->format = format;
5853 msrp->flags = flags;
5858 msrp->next = sys.tp;
5860 sys.added_thread_counters++;
5861 if (sys.added_thread_counters > MAX_ADDED_THREAD_COUNTERS) {
5862 fprintf(stderr, "exceeded max %d added thread counters\n",
5863 MAX_ADDED_COUNTERS);
5869 msrp->next = sys.cp;
5871 sys.added_core_counters++;
5872 if (sys.added_core_counters > MAX_ADDED_COUNTERS) {
5873 fprintf(stderr, "exceeded max %d added core counters\n",
5874 MAX_ADDED_COUNTERS);
5880 msrp->next = sys.pp;
5882 sys.added_package_counters++;
5883 if (sys.added_package_counters > MAX_ADDED_COUNTERS) {
5884 fprintf(stderr, "exceeded max %d added package counters\n",
5885 MAX_ADDED_COUNTERS);
5894 void parse_add_command(char *add_command)
5898 char name_buffer[NAME_BYTES] = "";
5901 enum counter_scope scope = SCOPE_CPU;
5902 enum counter_type type = COUNTER_CYCLES;
5903 enum counter_format format = FORMAT_DELTA;
5905 while (add_command) {
5907 if (sscanf(add_command, "msr0x%x", &msr_num) == 1)
5910 if (sscanf(add_command, "msr%d", &msr_num) == 1)
5913 if (*add_command == '/') {
5918 if (sscanf(add_command, "u%d", &width) == 1) {
5919 if ((width == 32) || (width == 64))
5923 if (!strncmp(add_command, "cpu", strlen("cpu"))) {
5927 if (!strncmp(add_command, "core", strlen("core"))) {
5931 if (!strncmp(add_command, "package", strlen("package"))) {
5932 scope = SCOPE_PACKAGE;
5935 if (!strncmp(add_command, "cycles", strlen("cycles"))) {
5936 type = COUNTER_CYCLES;
5939 if (!strncmp(add_command, "seconds", strlen("seconds"))) {
5940 type = COUNTER_SECONDS;
5943 if (!strncmp(add_command, "usec", strlen("usec"))) {
5944 type = COUNTER_USEC;
5947 if (!strncmp(add_command, "raw", strlen("raw"))) {
5948 format = FORMAT_RAW;
5951 if (!strncmp(add_command, "delta", strlen("delta"))) {
5952 format = FORMAT_DELTA;
5955 if (!strncmp(add_command, "percent", strlen("percent"))) {
5956 format = FORMAT_PERCENT;
5960 if (sscanf(add_command, "%18s,%*s", name_buffer) == 1) { /* 18 < NAME_BYTES */
5963 eos = strchr(name_buffer, ',');
5970 add_command = strchr(add_command, ',');
5972 *add_command = '\0';
5977 if ((msr_num == 0) && (path == NULL)) {
5978 fprintf(stderr, "--add: (msrDDD | msr0xXXX | /path_to_counter ) required\n");
5982 /* generate default column header */
5983 if (*name_buffer == '\0') {
5985 sprintf(name_buffer, "M0x%x%s", msr_num, format == FORMAT_PERCENT ? "%" : "");
5987 sprintf(name_buffer, "M0X%x%s", msr_num, format == FORMAT_PERCENT ? "%" : "");
5990 if (add_counter(msr_num, path, name_buffer, width, scope, type, format, 0))
5999 int is_deferred_skip(char *name)
6003 for (i = 0; i < deferred_skip_index; ++i)
6004 if (!strcmp(name, deferred_skip_names[i]))
6009 void probe_sysfs(void)
6017 if (!DO_BIC(BIC_sysfs))
6020 for (state = 10; state >= 0; --state) {
6022 sprintf(path, "/sys/devices/system/cpu/cpu%d/cpuidle/state%d/name",
6024 input = fopen(path, "r");
6027 if (!fgets(name_buf, sizeof(name_buf), input))
6028 err(1, "%s: failed to read file", path);
6030 /* truncate "C1-HSW\n" to "C1", or truncate "C1\n" to "C1" */
6031 sp = strchr(name_buf, '-');
6033 sp = strchrnul(name_buf, '\n');
6037 remove_underbar(name_buf);
6041 sprintf(path, "cpuidle/state%d/time", state);
6043 if (is_deferred_skip(name_buf))
6046 add_counter(0, path, name_buf, 64, SCOPE_CPU, COUNTER_USEC,
6047 FORMAT_PERCENT, SYSFS_PERCPU);
6050 for (state = 10; state >= 0; --state) {
6052 sprintf(path, "/sys/devices/system/cpu/cpu%d/cpuidle/state%d/name",
6054 input = fopen(path, "r");
6057 if (!fgets(name_buf, sizeof(name_buf), input))
6058 err(1, "%s: failed to read file", path);
6059 /* truncate "C1-HSW\n" to "C1", or truncate "C1\n" to "C1" */
6060 sp = strchr(name_buf, '-');
6062 sp = strchrnul(name_buf, '\n');
6066 remove_underbar(name_buf);
6068 sprintf(path, "cpuidle/state%d/usage", state);
6070 if (is_deferred_skip(name_buf))
6073 add_counter(0, path, name_buf, 64, SCOPE_CPU, COUNTER_ITEMS,
6074 FORMAT_DELTA, SYSFS_PERCPU);
6081 * parse cpuset with following syntax
6082 * 1,2,4..6,8-10 and set bits in cpu_subset
6084 void parse_cpu_command(char *optarg)
6086 unsigned int start, end;
6089 if (!strcmp(optarg, "core")) {
6095 if (!strcmp(optarg, "package")) {
6101 if (show_core_only || show_pkg_only)
6104 cpu_subset = CPU_ALLOC(CPU_SUBSET_MAXCPUS);
6105 if (cpu_subset == NULL)
6106 err(3, "CPU_ALLOC");
6107 cpu_subset_size = CPU_ALLOC_SIZE(CPU_SUBSET_MAXCPUS);
6109 CPU_ZERO_S(cpu_subset_size, cpu_subset);
6113 while (next && *next) {
6115 if (*next == '-') /* no negative cpu numbers */
6118 start = strtoul(next, &next, 10);
6120 if (start >= CPU_SUBSET_MAXCPUS)
6122 CPU_SET_S(start, cpu_subset_size, cpu_subset);
6133 next += 1; /* start range */
6134 } else if (*next == '.') {
6137 next += 1; /* start range */
6142 end = strtoul(next, &next, 10);
6146 while (++start <= end) {
6147 if (start >= CPU_SUBSET_MAXCPUS)
6149 CPU_SET_S(start, cpu_subset_size, cpu_subset);
6154 else if (*next != '\0')
6161 fprintf(stderr, "\"--cpu %s\" malformed\n", optarg);
6167 void cmdline(int argc, char **argv)
6170 int option_index = 0;
6171 static struct option long_options[] = {
6172 {"add", required_argument, 0, 'a'},
6173 {"cpu", required_argument, 0, 'c'},
6174 {"Dump", no_argument, 0, 'D'},
6175 {"debug", no_argument, 0, 'd'}, /* internal, not documented */
6176 {"enable", required_argument, 0, 'e'},
6177 {"interval", required_argument, 0, 'i'},
6178 {"IPC", no_argument, 0, 'I'},
6179 {"num_iterations", required_argument, 0, 'n'},
6180 {"help", no_argument, 0, 'h'},
6181 {"hide", required_argument, 0, 'H'}, // meh, -h taken by --help
6182 {"Joules", no_argument, 0, 'J'},
6183 {"list", no_argument, 0, 'l'},
6184 {"out", required_argument, 0, 'o'},
6185 {"quiet", no_argument, 0, 'q'},
6186 {"show", required_argument, 0, 's'},
6187 {"Summary", no_argument, 0, 'S'},
6188 {"TCC", required_argument, 0, 'T'},
6189 {"version", no_argument, 0, 'v' },
6195 while ((opt = getopt_long_only(argc, argv, "+C:c:Dde:hi:Jn:o:qST:v",
6196 long_options, &option_index)) != -1) {
6199 parse_add_command(optarg);
6202 parse_cpu_command(optarg);
6208 /* --enable specified counter */
6209 bic_enabled = bic_enabled | bic_lookup(optarg, SHOW_LIST);
6213 ENABLE_BIC(BIC_DISABLED_BY_DEFAULT);
6217 * --hide: do not show those specified
6218 * multiple invocations simply clear more bits in enabled mask
6220 bic_enabled &= ~bic_lookup(optarg, HIDE_LIST);
6228 double interval = strtod(optarg, NULL);
6230 if (interval < 0.001) {
6231 fprintf(outf, "interval %f seconds is too small\n",
6236 interval_tv.tv_sec = interval_ts.tv_sec = interval;
6237 interval_tv.tv_usec = (interval - interval_tv.tv_sec) * 1000000;
6238 interval_ts.tv_nsec = (interval - interval_ts.tv_sec) * 1000000000;
6245 ENABLE_BIC(BIC_DISABLED_BY_DEFAULT);
6250 outf = fopen_or_die(optarg, "w");
6256 num_iterations = strtod(optarg, NULL);
6258 if (num_iterations <= 0) {
6259 fprintf(outf, "iterations %d should be positive number\n",
6266 * --show: show only those specified
6267 * The 1st invocation will clear and replace the enabled mask
6268 * subsequent invocations can add to it.
6271 bic_enabled = bic_lookup(optarg, SHOW_LIST);
6273 bic_enabled |= bic_lookup(optarg, SHOW_LIST);
6280 tcc_activation_temp_override = atoi(optarg);
6290 int main(int argc, char **argv)
6293 cmdline(argc, argv);
6302 /* dump counters and exit */
6304 return get_and_dump_counters();
6306 /* list header and exit */
6307 if (list_header_only) {
6309 flush_output_stdout();
6315 * if any params left, it must be a command to fork
6318 return fork_it(argv + optind);