| 1 | // SPDX-License-Identifier: GPL-2.0-only |
| 2 | /* |
| 3 | * Performance events - AMD Processor Power Reporting Mechanism |
| 4 | * |
| 5 | * Copyright (C) 2016 Advanced Micro Devices, Inc. |
| 6 | * |
| 7 | * Author: Huang Rui <ray.huang@amd.com> |
| 8 | */ |
| 9 | |
| 10 | #include <linux/module.h> |
| 11 | #include <linux/slab.h> |
| 12 | #include <linux/perf_event.h> |
| 13 | #include <asm/cpu_device_id.h> |
| 14 | #include <asm/msr.h> |
| 15 | #include "../perf_event.h" |
| 16 | |
| 17 | /* Event code: LSB 8 bits, passed in attr->config any other bit is reserved. */ |
| 18 | #define AMD_POWER_EVENT_MASK 0xFFULL |
| 19 | |
| 20 | /* |
| 21 | * Accumulated power status counters. |
| 22 | */ |
| 23 | #define AMD_POWER_EVENTSEL_PKG 1 |
| 24 | |
| 25 | /* |
| 26 | * The ratio of compute unit power accumulator sample period to the |
| 27 | * PTSC period. |
| 28 | */ |
| 29 | static unsigned int cpu_pwr_sample_ratio; |
| 30 | |
| 31 | /* Maximum accumulated power of a compute unit. */ |
| 32 | static u64 max_cu_acc_power; |
| 33 | |
| 34 | static struct pmu pmu_class; |
| 35 | |
| 36 | /* |
| 37 | * Accumulated power represents the sum of each compute unit's (CU) power |
| 38 | * consumption. On any core of each CU we read the total accumulated power from |
| 39 | * MSR_F15H_CU_PWR_ACCUMULATOR. cpu_mask represents CPU bit map of all cores |
| 40 | * which are picked to measure the power for the CUs they belong to. |
| 41 | */ |
| 42 | static cpumask_t cpu_mask; |
| 43 | |
| 44 | static void event_update(struct perf_event *event) |
| 45 | { |
| 46 | struct hw_perf_event *hwc = &event->hw; |
| 47 | u64 prev_pwr_acc, new_pwr_acc, prev_ptsc, new_ptsc; |
| 48 | u64 delta, tdelta; |
| 49 | |
| 50 | prev_pwr_acc = hwc->pwr_acc; |
| 51 | prev_ptsc = hwc->ptsc; |
| 52 | rdmsrq(MSR_F15H_CU_PWR_ACCUMULATOR, new_pwr_acc); |
| 53 | rdmsrq(MSR_F15H_PTSC, new_ptsc); |
| 54 | |
| 55 | /* |
| 56 | * Calculate the CU power consumption over a time period, the unit of |
| 57 | * final value (delta) is micro-Watts. Then add it to the event count. |
| 58 | */ |
| 59 | if (new_pwr_acc < prev_pwr_acc) { |
| 60 | delta = max_cu_acc_power + new_pwr_acc; |
| 61 | delta -= prev_pwr_acc; |
| 62 | } else |
| 63 | delta = new_pwr_acc - prev_pwr_acc; |
| 64 | |
| 65 | delta *= cpu_pwr_sample_ratio * 1000; |
| 66 | tdelta = new_ptsc - prev_ptsc; |
| 67 | |
| 68 | do_div(delta, tdelta); |
| 69 | local64_add(delta, &event->count); |
| 70 | } |
| 71 | |
| 72 | static void __pmu_event_start(struct perf_event *event) |
| 73 | { |
| 74 | if (WARN_ON_ONCE(!(event->hw.state & PERF_HES_STOPPED))) |
| 75 | return; |
| 76 | |
| 77 | event->hw.state = 0; |
| 78 | |
| 79 | rdmsrq(MSR_F15H_PTSC, event->hw.ptsc); |
| 80 | rdmsrq(MSR_F15H_CU_PWR_ACCUMULATOR, event->hw.pwr_acc); |
| 81 | } |
| 82 | |
| 83 | static void pmu_event_start(struct perf_event *event, int mode) |
| 84 | { |
| 85 | __pmu_event_start(event); |
| 86 | } |
| 87 | |
| 88 | static void pmu_event_stop(struct perf_event *event, int mode) |
| 89 | { |
| 90 | struct hw_perf_event *hwc = &event->hw; |
| 91 | |
| 92 | /* Mark event as deactivated and stopped. */ |
| 93 | if (!(hwc->state & PERF_HES_STOPPED)) |
| 94 | hwc->state |= PERF_HES_STOPPED; |
| 95 | |
| 96 | /* Check if software counter update is necessary. */ |
| 97 | if ((mode & PERF_EF_UPDATE) && !(hwc->state & PERF_HES_UPTODATE)) { |
| 98 | /* |
| 99 | * Drain the remaining delta count out of an event |
| 100 | * that we are disabling: |
| 101 | */ |
| 102 | event_update(event); |
| 103 | hwc->state |= PERF_HES_UPTODATE; |
| 104 | } |
| 105 | } |
| 106 | |
| 107 | static int pmu_event_add(struct perf_event *event, int mode) |
| 108 | { |
| 109 | struct hw_perf_event *hwc = &event->hw; |
| 110 | |
| 111 | hwc->state = PERF_HES_UPTODATE | PERF_HES_STOPPED; |
| 112 | |
| 113 | if (mode & PERF_EF_START) |
| 114 | __pmu_event_start(event); |
| 115 | |
| 116 | return 0; |
| 117 | } |
| 118 | |
| 119 | static void pmu_event_del(struct perf_event *event, int flags) |
| 120 | { |
| 121 | pmu_event_stop(event, PERF_EF_UPDATE); |
| 122 | } |
| 123 | |
| 124 | static int pmu_event_init(struct perf_event *event) |
| 125 | { |
| 126 | u64 cfg = event->attr.config & AMD_POWER_EVENT_MASK; |
| 127 | |
| 128 | /* Only look at AMD power events. */ |
| 129 | if (event->attr.type != pmu_class.type) |
| 130 | return -ENOENT; |
| 131 | |
| 132 | /* Unsupported modes and filters. */ |
| 133 | if (event->attr.sample_period) |
| 134 | return -EINVAL; |
| 135 | |
| 136 | if (cfg != AMD_POWER_EVENTSEL_PKG) |
| 137 | return -EINVAL; |
| 138 | |
| 139 | return 0; |
| 140 | } |
| 141 | |
| 142 | static void pmu_event_read(struct perf_event *event) |
| 143 | { |
| 144 | event_update(event); |
| 145 | } |
| 146 | |
| 147 | static ssize_t |
| 148 | get_attr_cpumask(struct device *dev, struct device_attribute *attr, char *buf) |
| 149 | { |
| 150 | return cpumap_print_to_pagebuf(list: true, buf, mask: &cpu_mask); |
| 151 | } |
| 152 | |
| 153 | static DEVICE_ATTR(cpumask, S_IRUGO, get_attr_cpumask, NULL); |
| 154 | |
| 155 | static struct attribute *pmu_attrs[] = { |
| 156 | &dev_attr_cpumask.attr, |
| 157 | NULL, |
| 158 | }; |
| 159 | |
| 160 | static struct attribute_group pmu_attr_group = { |
| 161 | .attrs = pmu_attrs, |
| 162 | }; |
| 163 | |
| 164 | /* |
| 165 | * Currently it only supports to report the power of each |
| 166 | * processor/package. |
| 167 | */ |
| 168 | EVENT_ATTR_STR(power-pkg, power_pkg, "event=0x01" ); |
| 169 | |
| 170 | EVENT_ATTR_STR(power-pkg.unit, power_pkg_unit, "mWatts" ); |
| 171 | |
| 172 | /* Convert the count from micro-Watts to milli-Watts. */ |
| 173 | EVENT_ATTR_STR(power-pkg.scale, power_pkg_scale, "1.000000e-3" ); |
| 174 | |
| 175 | static struct attribute *events_attr[] = { |
| 176 | EVENT_PTR(power_pkg), |
| 177 | EVENT_PTR(power_pkg_unit), |
| 178 | EVENT_PTR(power_pkg_scale), |
| 179 | NULL, |
| 180 | }; |
| 181 | |
| 182 | static struct attribute_group pmu_events_group = { |
| 183 | .name = "events" , |
| 184 | .attrs = events_attr, |
| 185 | }; |
| 186 | |
| 187 | PMU_FORMAT_ATTR(event, "config:0-7" ); |
| 188 | |
| 189 | static struct attribute *formats_attr[] = { |
| 190 | &format_attr_event.attr, |
| 191 | NULL, |
| 192 | }; |
| 193 | |
| 194 | static struct attribute_group pmu_format_group = { |
| 195 | .name = "format" , |
| 196 | .attrs = formats_attr, |
| 197 | }; |
| 198 | |
| 199 | static const struct attribute_group *attr_groups[] = { |
| 200 | &pmu_attr_group, |
| 201 | &pmu_format_group, |
| 202 | &pmu_events_group, |
| 203 | NULL, |
| 204 | }; |
| 205 | |
| 206 | static struct pmu pmu_class = { |
| 207 | .attr_groups = attr_groups, |
| 208 | /* system-wide only */ |
| 209 | .task_ctx_nr = perf_invalid_context, |
| 210 | .event_init = pmu_event_init, |
| 211 | .add = pmu_event_add, |
| 212 | .del = pmu_event_del, |
| 213 | .start = pmu_event_start, |
| 214 | .stop = pmu_event_stop, |
| 215 | .read = pmu_event_read, |
| 216 | .capabilities = PERF_PMU_CAP_NO_EXCLUDE, |
| 217 | .module = THIS_MODULE, |
| 218 | }; |
| 219 | |
| 220 | static int power_cpu_exit(unsigned int cpu) |
| 221 | { |
| 222 | int target; |
| 223 | |
| 224 | if (!cpumask_test_and_clear_cpu(cpu, cpumask: &cpu_mask)) |
| 225 | return 0; |
| 226 | |
| 227 | /* |
| 228 | * Find a new CPU on the same compute unit, if was set in cpumask |
| 229 | * and still some CPUs on compute unit. Then migrate event and |
| 230 | * context to new CPU. |
| 231 | */ |
| 232 | target = cpumask_any_but(topology_sibling_cpumask(cpu), cpu); |
| 233 | if (target < nr_cpumask_bits) { |
| 234 | cpumask_set_cpu(cpu: target, dstp: &cpu_mask); |
| 235 | perf_pmu_migrate_context(pmu: &pmu_class, src_cpu: cpu, dst_cpu: target); |
| 236 | } |
| 237 | return 0; |
| 238 | } |
| 239 | |
| 240 | static int power_cpu_init(unsigned int cpu) |
| 241 | { |
| 242 | int target; |
| 243 | |
| 244 | /* |
| 245 | * 1) If any CPU is set at cpu_mask in the same compute unit, do |
| 246 | * nothing. |
| 247 | * 2) If no CPU is set at cpu_mask in the same compute unit, |
| 248 | * set current ONLINE CPU. |
| 249 | * |
| 250 | * Note: if there is a CPU aside of the new one already in the |
| 251 | * sibling mask, then it is also in cpu_mask. |
| 252 | */ |
| 253 | target = cpumask_any_but(topology_sibling_cpumask(cpu), cpu); |
| 254 | if (target >= nr_cpumask_bits) |
| 255 | cpumask_set_cpu(cpu, dstp: &cpu_mask); |
| 256 | return 0; |
| 257 | } |
| 258 | |
| 259 | static const struct x86_cpu_id cpu_match[] = { |
| 260 | X86_MATCH_VENDOR_FAM(AMD, 0x15, NULL), |
| 261 | {}, |
| 262 | }; |
| 263 | |
| 264 | static int __init amd_power_pmu_init(void) |
| 265 | { |
| 266 | int ret; |
| 267 | |
| 268 | if (!x86_match_cpu(match: cpu_match)) |
| 269 | return -ENODEV; |
| 270 | |
| 271 | if (!boot_cpu_has(X86_FEATURE_ACC_POWER)) |
| 272 | return -ENODEV; |
| 273 | |
| 274 | cpu_pwr_sample_ratio = cpuid_ecx(op: 0x80000007); |
| 275 | |
| 276 | if (rdmsrq_safe(MSR_F15H_CU_MAX_PWR_ACCUMULATOR, p: &max_cu_acc_power)) { |
| 277 | pr_err("Failed to read max compute unit power accumulator MSR\n" ); |
| 278 | return -ENODEV; |
| 279 | } |
| 280 | |
| 281 | |
| 282 | cpuhp_setup_state(state: CPUHP_AP_PERF_X86_AMD_POWER_ONLINE, |
| 283 | name: "perf/x86/amd/power:online" , |
| 284 | startup: power_cpu_init, teardown: power_cpu_exit); |
| 285 | |
| 286 | ret = perf_pmu_register(pmu: &pmu_class, name: "power" , type: -1); |
| 287 | if (WARN_ON(ret)) { |
| 288 | pr_warn("AMD Power PMU registration failed\n" ); |
| 289 | return ret; |
| 290 | } |
| 291 | |
| 292 | pr_info("AMD Power PMU detected\n" ); |
| 293 | return ret; |
| 294 | } |
| 295 | module_init(amd_power_pmu_init); |
| 296 | |
| 297 | static void __exit amd_power_pmu_exit(void) |
| 298 | { |
| 299 | cpuhp_remove_state_nocalls(state: CPUHP_AP_PERF_X86_AMD_POWER_ONLINE); |
| 300 | perf_pmu_unregister(pmu: &pmu_class); |
| 301 | } |
| 302 | module_exit(amd_power_pmu_exit); |
| 303 | |
| 304 | MODULE_AUTHOR("Huang Rui <ray.huang@amd.com>" ); |
| 305 | MODULE_DESCRIPTION("AMD Processor Power Reporting Mechanism" ); |
| 306 | MODULE_LICENSE("GPL v2" ); |
| 307 | |