Commit 2dfc818b35cbea59188cc86e86e0a0efce2b0dbe

Authored by Thomas Renninger
Committed by Dominik Brodowski
1 parent 75f25bd31d

cpupower: mperf monitor - Use TSC to calculate max frequency if possible

Which makes the implementation independent from cpufreq drivers.
Therefore this would also work on a Xen kernel where the hypervisor
is doing frequency switching and idle entering.

Signed-off-by: Thomas Renninger <trenn@suse.de>
Signed-off-by: Dominik Brodowski <linux@dominikbrodowski.net>

Showing 2 changed files with 131 additions and 48 deletions Side-by-side Diff

tools/power/cpupower/Makefile
... ... @@ -24,7 +24,7 @@
24 24  
25 25 # Set the following to `true' to make a unstripped, unoptimized
26 26 # binary. Leave this set to `false' for production use.
27   -DEBUG ?= false
  27 +DEBUG ?= true
28 28  
29 29 # make the build silent. Set this to something else to make it noisy again.
30 30 V ?= false
tools/power/cpupower/utils/idle_monitor/mperf_monitor.c
... ... @@ -22,12 +22,15 @@
22 22  
23 23 #define MSR_TSC 0x10
24 24  
  25 +#define MSR_AMD_HWCR 0xc0010015
  26 +
25 27 enum mperf_id { C0 = 0, Cx, AVG_FREQ, MPERF_CSTATE_COUNT };
26 28  
27 29 static int mperf_get_count_percent(unsigned int self_id, double *percent,
28 30 unsigned int cpu);
29 31 static int mperf_get_count_freq(unsigned int id, unsigned long long *count,
30 32 unsigned int cpu);
  33 +static struct timespec time_start, time_end;
31 34  
32 35 static cstate_t mperf_cstates[MPERF_CSTATE_COUNT] = {
33 36 {
34 37  
35 38  
36 39  
... ... @@ -54,19 +57,33 @@
54 57 },
55 58 };
56 59  
  60 +enum MAX_FREQ_MODE { MAX_FREQ_SYSFS, MAX_FREQ_TSC_REF };
  61 +static int max_freq_mode;
  62 +/*
  63 + * The max frequency mperf is ticking at (in C0), either retrieved via:
  64 + * 1) calculated after measurements if we know TSC ticks at mperf/P0 frequency
  65 + * 2) cpufreq /sys/devices/.../cpu0/cpufreq/cpuinfo_max_freq at init time
  66 + * 1. Is preferred as it also works without cpufreq subsystem (e.g. on Xen)
  67 + */
  68 +static unsigned long max_frequency;
  69 +
57 70 static unsigned long long tsc_at_measure_start;
58 71 static unsigned long long tsc_at_measure_end;
59   -static unsigned long max_frequency;
60 72 static unsigned long long *mperf_previous_count;
61 73 static unsigned long long *aperf_previous_count;
62 74 static unsigned long long *mperf_current_count;
63 75 static unsigned long long *aperf_current_count;
  76 +
64 77 /* valid flag for all CPUs. If a MSR read failed it will be zero */
65 78 static int *is_valid;
66 79  
67 80 static int mperf_get_tsc(unsigned long long *tsc)
68 81 {
69   - return read_msr(0, MSR_TSC, tsc);
  82 + int ret;
  83 + ret = read_msr(0, MSR_TSC, tsc);
  84 + if (ret)
  85 + dprint("Reading TSC MSR failed, returning %llu\n", *tsc);
  86 + return ret;
70 87 }
71 88  
72 89 static int mperf_init_stats(unsigned int cpu)
73 90  
... ... @@ -97,36 +114,11 @@
97 114 return 0;
98 115 }
99 116  
100   -/*
101   - * get_average_perf()
102   - *
103   - * Returns the average performance (also considers boosted frequencies)
104   - *
105   - * Input:
106   - * aperf_diff: Difference of the aperf register over a time period
107   - * mperf_diff: Difference of the mperf register over the same time period
108   - * max_freq: Maximum frequency (P0)
109   - *
110   - * Returns:
111   - * Average performance over the time period
112   - */
113   -static unsigned long get_average_perf(unsigned long long aperf_diff,
114   - unsigned long long mperf_diff)
115   -{
116   - unsigned int perf_percent = 0;
117   - if (((unsigned long)(-1) / 100) < aperf_diff) {
118   - int shift_count = 7;
119   - aperf_diff >>= shift_count;
120   - mperf_diff >>= shift_count;
121   - }
122   - perf_percent = (aperf_diff * 100) / mperf_diff;
123   - return (max_frequency * perf_percent) / 100;
124   -}
125   -
126 117 static int mperf_get_count_percent(unsigned int id, double *percent,
127 118 unsigned int cpu)
128 119 {
129 120 unsigned long long aperf_diff, mperf_diff, tsc_diff;
  121 + unsigned long long timediff;
130 122  
131 123 if (!is_valid[cpu])
132 124 return -1;
133 125  
... ... @@ -136,11 +128,19 @@
136 128  
137 129 mperf_diff = mperf_current_count[cpu] - mperf_previous_count[cpu];
138 130 aperf_diff = aperf_current_count[cpu] - aperf_previous_count[cpu];
139   - tsc_diff = tsc_at_measure_end - tsc_at_measure_start;
140 131  
141   - *percent = 100.0 * mperf_diff / tsc_diff;
142   - dprint("%s: mperf_diff: %llu, tsc_diff: %llu\n",
143   - mperf_cstates[id].name, mperf_diff, tsc_diff);
  132 + if (max_freq_mode == MAX_FREQ_TSC_REF) {
  133 + tsc_diff = tsc_at_measure_end - tsc_at_measure_start;
  134 + *percent = 100.0 * mperf_diff / tsc_diff;
  135 + dprint("%s: TSC Ref - mperf_diff: %llu, tsc_diff: %llu\n",
  136 + mperf_cstates[id].name, mperf_diff, tsc_diff);
  137 + } else if (max_freq_mode == MAX_FREQ_SYSFS) {
  138 + timediff = timespec_diff_us(time_start, time_end);
  139 + *percent = 100.0 * mperf_diff / timediff;
  140 + dprint("%s: MAXFREQ - mperf_diff: %llu, time_diff: %llu\n",
  141 + mperf_cstates[id].name, mperf_diff, timediff);
  142 + } else
  143 + return -1;
144 144  
145 145 if (id == Cx)
146 146 *percent = 100.0 - *percent;
... ... @@ -154,7 +154,7 @@
154 154 static int mperf_get_count_freq(unsigned int id, unsigned long long *count,
155 155 unsigned int cpu)
156 156 {
157   - unsigned long long aperf_diff, mperf_diff;
  157 + unsigned long long aperf_diff, mperf_diff, time_diff, tsc_diff;
158 158  
159 159 if (id != AVG_FREQ)
160 160 return 1;
161 161  
... ... @@ -165,11 +165,21 @@
165 165 mperf_diff = mperf_current_count[cpu] - mperf_previous_count[cpu];
166 166 aperf_diff = aperf_current_count[cpu] - aperf_previous_count[cpu];
167 167  
168   - /* Return MHz for now, might want to return KHz if column width is more
169   - generic */
170   - *count = get_average_perf(aperf_diff, mperf_diff) / 1000;
171   - dprint("%s: %llu\n", mperf_cstates[id].name, *count);
  168 + if (max_freq_mode == MAX_FREQ_TSC_REF) {
  169 + /* Calculate max_freq from TSC count */
  170 + tsc_diff = tsc_at_measure_end - tsc_at_measure_start;
  171 + time_diff = timespec_diff_us(time_start, time_end);
  172 + max_frequency = tsc_diff / time_diff;
  173 + }
172 174  
  175 + *count = max_frequency * ((double)aperf_diff / mperf_diff);
  176 + dprint("%s: Average freq based on %s maximum frequency:\n",
  177 + mperf_cstates[id].name,
  178 + (max_freq_mode == MAX_FREQ_TSC_REF) ? "TSC calculated" : "sysfs read");
  179 + dprint("%max_frequency: %lu", max_frequency);
  180 + dprint("aperf_diff: %llu\n", aperf_diff);
  181 + dprint("mperf_diff: %llu\n", mperf_diff);
  182 + dprint("avg freq: %llu\n", *count);
173 183 return 0;
174 184 }
175 185  
... ... @@ -178,6 +188,7 @@
178 188 int cpu;
179 189 unsigned long long dbg;
180 190  
  191 + clock_gettime(CLOCK_REALTIME, &time_start);
181 192 mperf_get_tsc(&tsc_at_measure_start);
182 193  
183 194 for (cpu = 0; cpu < cpu_count; cpu++)
184 195  
185 196  
186 197  
187 198  
188 199  
189 200  
190 201  
... ... @@ -193,32 +204,104 @@
193 204 unsigned long long dbg;
194 205 int cpu;
195 206  
196   - mperf_get_tsc(&tsc_at_measure_end);
197   -
198 207 for (cpu = 0; cpu < cpu_count; cpu++)
199 208 mperf_measure_stats(cpu);
200 209  
  210 + mperf_get_tsc(&tsc_at_measure_end);
  211 + clock_gettime(CLOCK_REALTIME, &time_end);
  212 +
201 213 mperf_get_tsc(&dbg);
202 214 dprint("TSC diff: %llu\n", dbg - tsc_at_measure_end);
203 215  
204 216 return 0;
205 217 }
206 218  
207   -struct cpuidle_monitor mperf_monitor;
208   -
209   -struct cpuidle_monitor *mperf_register(void)
  219 +/*
  220 + * Mperf register is defined to tick at P0 (maximum) frequency
  221 + *
  222 + * Instead of reading out P0 which can be tricky to read out from HW,
  223 + * we use TSC counter if it reliably ticks at P0/mperf frequency.
  224 + *
  225 + * Still try to fall back to:
  226 + * /sys/devices/system/cpu/cpu0/cpufreq/cpuinfo_max_freq
  227 + * on older Intel HW without invariant TSC feature.
  228 + * Or on AMD machines where TSC does not tick at P0 (do not exist yet, but
  229 + * it's still double checked (MSR_AMD_HWCR)).
  230 + *
  231 + * On these machines the user would still get useful mperf
  232 + * stats when acpi-cpufreq driver is loaded.
  233 + */
  234 +static int init_maxfreq_mode(void)
210 235 {
  236 + int ret;
  237 + unsigned long long hwcr;
211 238 unsigned long min;
212 239  
213   - if (!(cpupower_cpu_info.caps & CPUPOWER_CAP_APERF))
214   - return NULL;
  240 + if (!cpupower_cpu_info.caps & CPUPOWER_CAP_INV_TSC)
  241 + goto use_sysfs;
215 242  
216   - /* Assume min/max all the same on all cores */
  243 + if (cpupower_cpu_info.vendor == X86_VENDOR_AMD) {
  244 + /* MSR_AMD_HWCR tells us whether TSC runs at P0/mperf
  245 + * freq.
  246 + * A test whether hwcr is accessable/available would be:
  247 + * (cpupower_cpu_info.family > 0x10 ||
  248 + * cpupower_cpu_info.family == 0x10 &&
  249 + * cpupower_cpu_info.model >= 0x2))
  250 + * This should be the case for all aperf/mperf
  251 + * capable AMD machines and is therefore safe to test here.
  252 + * Compare with Linus kernel git commit: acf01734b1747b1ec4
  253 + */
  254 + ret = read_msr(0, MSR_AMD_HWCR, &hwcr);
  255 + /*
  256 + * If the MSR read failed, assume a Xen system that did
  257 + * not explicitly provide access to it and assume TSC works
  258 + */
  259 + if (ret != 0) {
  260 + dprint("TSC read 0x%x failed - assume TSC working\n",
  261 + MSR_AMD_HWCR);
  262 + return 0;
  263 + } else if (1 & (hwcr >> 24)) {
  264 + max_freq_mode = MAX_FREQ_TSC_REF;
  265 + return 0;
  266 + } else { /* Use sysfs max frequency if available */ }
  267 + } else if (cpupower_cpu_info.vendor == X86_VENDOR_INTEL) {
  268 + /*
  269 + * On Intel we assume mperf (in C0) is ticking at same
  270 + * rate than TSC
  271 + */
  272 + max_freq_mode = MAX_FREQ_TSC_REF;
  273 + return 0;
  274 + }
  275 +use_sysfs:
217 276 if (cpufreq_get_hardware_limits(0, &min, &max_frequency)) {
218 277 dprint("Cannot retrieve max freq from cpufreq kernel "
219 278 "subsystem\n");
220   - return NULL;
  279 + return -1;
221 280 }
  281 + max_freq_mode = MAX_FREQ_SYSFS;
  282 + return 0;
  283 +}
  284 +
  285 +/*
  286 + * This monitor provides:
  287 + *
  288 + * 1) Average frequency a CPU resided in
  289 + * This always works if the CPU has aperf/mperf capabilities
  290 + *
  291 + * 2) C0 and Cx (any sleep state) time a CPU resided in
  292 + * Works if mperf timer stops ticking in sleep states which
  293 + * seem to be the case on all current HW.
  294 + * Both is directly retrieved from HW registers and is independent
  295 + * from kernel statistics.
  296 + */
  297 +struct cpuidle_monitor mperf_monitor;
  298 +struct cpuidle_monitor *mperf_register(void)
  299 +{
  300 + if (!(cpupower_cpu_info.caps & CPUPOWER_CAP_APERF))
  301 + return NULL;
  302 +
  303 + if (init_maxfreq_mode())
  304 + return NULL;
222 305  
223 306 /* Free this at program termination */
224 307 is_valid = calloc(cpu_count, sizeof(int));