Olav Haugan | 9306c80 | 2016-08-18 17:22:44 -0700 | [diff] [blame^] | 1 | /* Copyright (c) 2014-2016, The Linux Foundation. All rights reserved. |
| 2 | * |
| 3 | * This program is free software; you can redistribute it and/or modify |
| 4 | * it under the terms of the GNU General Public License version 2 and |
| 5 | * only version 2 as published by the Free Software Foundation. |
| 6 | * |
| 7 | * This program is distributed in the hope that it will be useful, |
| 8 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 9 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 10 | * GNU General Public License for more details. |
| 11 | */ |
| 12 | |
| 13 | #include <linux/init.h> |
| 14 | #include <linux/notifier.h> |
| 15 | #include <linux/cpu.h> |
| 16 | #include <linux/cpumask.h> |
| 17 | #include <linux/cpufreq.h> |
| 18 | #include <linux/timer.h> |
| 19 | #include <linux/kthread.h> |
| 20 | #include <linux/sched.h> |
| 21 | #include <linux/sched/rt.h> |
| 22 | |
| 23 | #include <trace/events/power.h> |
| 24 | |
| 25 | #define MAX_CPUS_PER_GROUP 4 |
| 26 | |
| 27 | struct cpu_data { |
| 28 | /* Per CPU data. */ |
| 29 | bool inited; |
| 30 | bool online; |
| 31 | bool rejected; |
| 32 | bool is_busy; |
| 33 | bool not_preferred; |
| 34 | unsigned int busy; |
| 35 | unsigned int cpu; |
| 36 | struct list_head sib; |
| 37 | unsigned int first_cpu; |
| 38 | |
| 39 | /* Per cluster data set only on first CPU */ |
| 40 | unsigned int min_cpus; |
| 41 | unsigned int max_cpus; |
| 42 | unsigned int offline_delay_ms; |
| 43 | unsigned int busy_up_thres[MAX_CPUS_PER_GROUP]; |
| 44 | unsigned int busy_down_thres[MAX_CPUS_PER_GROUP]; |
| 45 | unsigned int online_cpus; |
| 46 | unsigned int avail_cpus; |
| 47 | unsigned int num_cpus; |
| 48 | unsigned int need_cpus; |
| 49 | unsigned int task_thres; |
| 50 | s64 need_ts; |
| 51 | struct list_head lru; |
| 52 | bool pending; |
| 53 | spinlock_t pending_lock; |
| 54 | bool is_big_cluster; |
| 55 | int nrrun; |
| 56 | bool nrrun_changed; |
| 57 | struct timer_list timer; |
| 58 | struct task_struct *hotplug_thread; |
| 59 | struct kobject kobj; |
| 60 | }; |
| 61 | |
| 62 | static DEFINE_PER_CPU(struct cpu_data, cpu_state); |
| 63 | static DEFINE_SPINLOCK(state_lock); |
| 64 | static void apply_need(struct cpu_data *f); |
| 65 | static void wake_up_hotplug_thread(struct cpu_data *state); |
| 66 | |
| 67 | /* ========================= sysfs interface =========================== */ |
| 68 | |
| 69 | static ssize_t store_min_cpus(struct cpu_data *state, |
| 70 | const char *buf, size_t count) |
| 71 | { |
| 72 | unsigned int val; |
| 73 | |
| 74 | if (sscanf(buf, "%u\n", &val) != 1) |
| 75 | return -EINVAL; |
| 76 | |
| 77 | state->min_cpus = min(val, state->max_cpus); |
| 78 | wake_up_hotplug_thread(state); |
| 79 | |
| 80 | return count; |
| 81 | } |
| 82 | |
| 83 | static ssize_t show_min_cpus(struct cpu_data *state, char *buf) |
| 84 | { |
| 85 | return snprintf(buf, PAGE_SIZE, "%u\n", state->min_cpus); |
| 86 | } |
| 87 | |
| 88 | static ssize_t store_max_cpus(struct cpu_data *state, |
| 89 | const char *buf, size_t count) |
| 90 | { |
| 91 | unsigned int val; |
| 92 | |
| 93 | if (sscanf(buf, "%u\n", &val) != 1) |
| 94 | return -EINVAL; |
| 95 | |
| 96 | val = min(val, state->num_cpus); |
| 97 | state->max_cpus = val; |
| 98 | state->min_cpus = min(state->min_cpus, state->max_cpus); |
| 99 | wake_up_hotplug_thread(state); |
| 100 | |
| 101 | return count; |
| 102 | } |
| 103 | |
| 104 | static ssize_t show_max_cpus(struct cpu_data *state, char *buf) |
| 105 | { |
| 106 | return snprintf(buf, PAGE_SIZE, "%u\n", state->max_cpus); |
| 107 | } |
| 108 | |
| 109 | static ssize_t store_offline_delay_ms(struct cpu_data *state, |
| 110 | const char *buf, size_t count) |
| 111 | { |
| 112 | unsigned int val; |
| 113 | |
| 114 | if (sscanf(buf, "%u\n", &val) != 1) |
| 115 | return -EINVAL; |
| 116 | |
| 117 | state->offline_delay_ms = val; |
| 118 | apply_need(state); |
| 119 | |
| 120 | return count; |
| 121 | } |
| 122 | |
| 123 | static ssize_t show_task_thres(struct cpu_data *state, char *buf) |
| 124 | { |
| 125 | return snprintf(buf, PAGE_SIZE, "%u\n", state->task_thres); |
| 126 | } |
| 127 | |
| 128 | static ssize_t store_task_thres(struct cpu_data *state, |
| 129 | const char *buf, size_t count) |
| 130 | { |
| 131 | unsigned int val; |
| 132 | |
| 133 | if (sscanf(buf, "%u\n", &val) != 1) |
| 134 | return -EINVAL; |
| 135 | |
| 136 | if (val < state->num_cpus) |
| 137 | return -EINVAL; |
| 138 | |
| 139 | state->task_thres = val; |
| 140 | apply_need(state); |
| 141 | |
| 142 | return count; |
| 143 | } |
| 144 | |
| 145 | static ssize_t show_offline_delay_ms(struct cpu_data *state, char *buf) |
| 146 | { |
| 147 | return snprintf(buf, PAGE_SIZE, "%u\n", state->offline_delay_ms); |
| 148 | } |
| 149 | |
| 150 | static ssize_t store_busy_up_thres(struct cpu_data *state, |
| 151 | const char *buf, size_t count) |
| 152 | { |
| 153 | unsigned int val[MAX_CPUS_PER_GROUP]; |
| 154 | int ret, i; |
| 155 | |
| 156 | ret = sscanf(buf, "%u %u %u %u\n", &val[0], &val[1], &val[2], &val[3]); |
| 157 | if (ret != 1 && ret != state->num_cpus) |
| 158 | return -EINVAL; |
| 159 | |
| 160 | if (ret == 1) { |
| 161 | for (i = 0; i < state->num_cpus; i++) |
| 162 | state->busy_up_thres[i] = val[0]; |
| 163 | } else { |
| 164 | for (i = 0; i < state->num_cpus; i++) |
| 165 | state->busy_up_thres[i] = val[i]; |
| 166 | } |
| 167 | apply_need(state); |
| 168 | return count; |
| 169 | } |
| 170 | |
| 171 | static ssize_t show_busy_up_thres(struct cpu_data *state, char *buf) |
| 172 | { |
| 173 | int i, count = 0; |
| 174 | |
| 175 | for (i = 0; i < state->num_cpus; i++) |
| 176 | count += snprintf(buf + count, PAGE_SIZE - count, "%u ", |
| 177 | state->busy_up_thres[i]); |
| 178 | count += snprintf(buf + count, PAGE_SIZE - count, "\n"); |
| 179 | return count; |
| 180 | } |
| 181 | |
| 182 | static ssize_t store_busy_down_thres(struct cpu_data *state, |
| 183 | const char *buf, size_t count) |
| 184 | { |
| 185 | unsigned int val[MAX_CPUS_PER_GROUP]; |
| 186 | int ret, i; |
| 187 | |
| 188 | ret = sscanf(buf, "%u %u %u %u\n", &val[0], &val[1], &val[2], &val[3]); |
| 189 | if (ret != 1 && ret != state->num_cpus) |
| 190 | return -EINVAL; |
| 191 | |
| 192 | if (ret == 1) { |
| 193 | for (i = 0; i < state->num_cpus; i++) |
| 194 | state->busy_down_thres[i] = val[0]; |
| 195 | } else { |
| 196 | for (i = 0; i < state->num_cpus; i++) |
| 197 | state->busy_down_thres[i] = val[i]; |
| 198 | } |
| 199 | apply_need(state); |
| 200 | return count; |
| 201 | } |
| 202 | |
| 203 | static ssize_t show_busy_down_thres(struct cpu_data *state, char *buf) |
| 204 | { |
| 205 | int i, count = 0; |
| 206 | |
| 207 | for (i = 0; i < state->num_cpus; i++) |
| 208 | count += snprintf(buf + count, PAGE_SIZE - count, "%u ", |
| 209 | state->busy_down_thres[i]); |
| 210 | count += snprintf(buf + count, PAGE_SIZE - count, "\n"); |
| 211 | return count; |
| 212 | } |
| 213 | |
| 214 | static ssize_t store_is_big_cluster(struct cpu_data *state, |
| 215 | const char *buf, size_t count) |
| 216 | { |
| 217 | unsigned int val; |
| 218 | |
| 219 | if (sscanf(buf, "%u\n", &val) != 1) |
| 220 | return -EINVAL; |
| 221 | |
| 222 | state->is_big_cluster = val ? 1 : 0; |
| 223 | return count; |
| 224 | } |
| 225 | |
| 226 | static ssize_t show_is_big_cluster(struct cpu_data *state, char *buf) |
| 227 | { |
| 228 | return snprintf(buf, PAGE_SIZE, "%u\n", state->is_big_cluster); |
| 229 | } |
| 230 | |
| 231 | static ssize_t show_cpus(struct cpu_data *state, char *buf) |
| 232 | { |
| 233 | struct cpu_data *c; |
| 234 | ssize_t count = 0; |
| 235 | unsigned long flags; |
| 236 | |
| 237 | spin_lock_irqsave(&state_lock, flags); |
| 238 | list_for_each_entry(c, &state->lru, sib) { |
| 239 | count += snprintf(buf + count, PAGE_SIZE - count, |
| 240 | "CPU%u (%s)\n", c->cpu, |
| 241 | c->online ? "Online" : "Offline"); |
| 242 | } |
| 243 | spin_unlock_irqrestore(&state_lock, flags); |
| 244 | return count; |
| 245 | } |
| 246 | |
| 247 | static ssize_t show_need_cpus(struct cpu_data *state, char *buf) |
| 248 | { |
| 249 | return snprintf(buf, PAGE_SIZE, "%u\n", state->need_cpus); |
| 250 | } |
| 251 | |
| 252 | static ssize_t show_online_cpus(struct cpu_data *state, char *buf) |
| 253 | { |
| 254 | return snprintf(buf, PAGE_SIZE, "%u\n", state->online_cpus); |
| 255 | } |
| 256 | |
| 257 | static ssize_t show_global_state(struct cpu_data *state, char *buf) |
| 258 | { |
| 259 | struct cpu_data *c; |
| 260 | ssize_t count = 0; |
| 261 | unsigned int cpu; |
| 262 | |
| 263 | for_each_possible_cpu(cpu) { |
| 264 | count += snprintf(buf + count, PAGE_SIZE - count, |
| 265 | "CPU%u\n", cpu); |
| 266 | c = &per_cpu(cpu_state, cpu); |
| 267 | if (!c->inited) |
| 268 | continue; |
| 269 | count += snprintf(buf + count, PAGE_SIZE - count, |
| 270 | "\tCPU: %u\n", c->cpu); |
| 271 | count += snprintf(buf + count, PAGE_SIZE - count, |
| 272 | "\tOnline: %u\n", c->online); |
| 273 | count += snprintf(buf + count, PAGE_SIZE - count, |
| 274 | "\tRejected: %u\n", c->rejected); |
| 275 | count += snprintf(buf + count, PAGE_SIZE - count, |
| 276 | "\tFirst CPU: %u\n", c->first_cpu); |
| 277 | count += snprintf(buf + count, PAGE_SIZE - count, |
| 278 | "\tBusy%%: %u\n", c->busy); |
| 279 | count += snprintf(buf + count, PAGE_SIZE - count, |
| 280 | "\tIs busy: %u\n", c->is_busy); |
| 281 | if (c->cpu != c->first_cpu) |
| 282 | continue; |
| 283 | count += snprintf(buf + count, PAGE_SIZE - count, |
| 284 | "\tNr running: %u\n", c->nrrun); |
| 285 | count += snprintf(buf + count, PAGE_SIZE - count, |
| 286 | "\tAvail CPUs: %u\n", c->avail_cpus); |
| 287 | count += snprintf(buf + count, PAGE_SIZE - count, |
| 288 | "\tNeed CPUs: %u\n", c->need_cpus); |
| 289 | } |
| 290 | |
| 291 | return count; |
| 292 | } |
| 293 | |
| 294 | static ssize_t store_not_preferred(struct cpu_data *state, |
| 295 | const char *buf, size_t count) |
| 296 | { |
| 297 | struct cpu_data *c; |
| 298 | unsigned int i, first_cpu; |
| 299 | unsigned int val[MAX_CPUS_PER_GROUP]; |
| 300 | int ret; |
| 301 | |
| 302 | ret = sscanf(buf, "%u %u %u %u\n", &val[0], &val[1], &val[2], &val[3]); |
| 303 | if (ret != 1 && ret != state->num_cpus) |
| 304 | return -EINVAL; |
| 305 | |
| 306 | first_cpu = state->first_cpu; |
| 307 | |
| 308 | for (i = 0; i < state->num_cpus; i++) { |
| 309 | c = &per_cpu(cpu_state, first_cpu); |
| 310 | c->not_preferred = val[i]; |
| 311 | first_cpu++; |
| 312 | } |
| 313 | |
| 314 | return count; |
| 315 | } |
| 316 | |
| 317 | static ssize_t show_not_preferred(struct cpu_data *state, char *buf) |
| 318 | { |
| 319 | struct cpu_data *c; |
| 320 | ssize_t count = 0; |
| 321 | unsigned int i, first_cpu; |
| 322 | |
| 323 | first_cpu = state->first_cpu; |
| 324 | |
| 325 | for (i = 0; i < state->num_cpus; i++) { |
| 326 | c = &per_cpu(cpu_state, first_cpu); |
| 327 | count += snprintf(buf + count, PAGE_SIZE - count, |
| 328 | "\tCPU:%d %u\n", first_cpu, c->not_preferred); |
| 329 | first_cpu++; |
| 330 | } |
| 331 | |
| 332 | return count; |
| 333 | } |
| 334 | |
| 335 | struct core_ctl_attr { |
| 336 | struct attribute attr; |
| 337 | ssize_t (*show)(struct cpu_data *, char *); |
| 338 | ssize_t (*store)(struct cpu_data *, const char *, size_t count); |
| 339 | }; |
| 340 | |
| 341 | #define core_ctl_attr_ro(_name) \ |
| 342 | static struct core_ctl_attr _name = \ |
| 343 | __ATTR(_name, 0444, show_##_name, NULL) |
| 344 | |
| 345 | #define core_ctl_attr_rw(_name) \ |
| 346 | static struct core_ctl_attr _name = \ |
| 347 | __ATTR(_name, 0644, show_##_name, store_##_name) |
| 348 | |
| 349 | core_ctl_attr_rw(min_cpus); |
| 350 | core_ctl_attr_rw(max_cpus); |
| 351 | core_ctl_attr_rw(offline_delay_ms); |
| 352 | core_ctl_attr_rw(busy_up_thres); |
| 353 | core_ctl_attr_rw(busy_down_thres); |
| 354 | core_ctl_attr_rw(task_thres); |
| 355 | core_ctl_attr_rw(is_big_cluster); |
| 356 | core_ctl_attr_ro(cpus); |
| 357 | core_ctl_attr_ro(need_cpus); |
| 358 | core_ctl_attr_ro(online_cpus); |
| 359 | core_ctl_attr_ro(global_state); |
| 360 | core_ctl_attr_rw(not_preferred); |
| 361 | |
| 362 | static struct attribute *default_attrs[] = { |
| 363 | &min_cpus.attr, |
| 364 | &max_cpus.attr, |
| 365 | &offline_delay_ms.attr, |
| 366 | &busy_up_thres.attr, |
| 367 | &busy_down_thres.attr, |
| 368 | &task_thres.attr, |
| 369 | &is_big_cluster.attr, |
| 370 | &cpus.attr, |
| 371 | &need_cpus.attr, |
| 372 | &online_cpus.attr, |
| 373 | &global_state.attr, |
| 374 | ¬_preferred.attr, |
| 375 | NULL |
| 376 | }; |
| 377 | |
| 378 | #define to_cpu_data(k) container_of(k, struct cpu_data, kobj) |
| 379 | #define to_attr(a) container_of(a, struct core_ctl_attr, attr) |
| 380 | static ssize_t show(struct kobject *kobj, struct attribute *attr, char *buf) |
| 381 | { |
| 382 | struct cpu_data *data = to_cpu_data(kobj); |
| 383 | struct core_ctl_attr *cattr = to_attr(attr); |
| 384 | ssize_t ret = -EIO; |
| 385 | |
| 386 | if (cattr->show) |
| 387 | ret = cattr->show(data, buf); |
| 388 | |
| 389 | return ret; |
| 390 | } |
| 391 | |
| 392 | static ssize_t store(struct kobject *kobj, struct attribute *attr, |
| 393 | const char *buf, size_t count) |
| 394 | { |
| 395 | struct cpu_data *data = to_cpu_data(kobj); |
| 396 | struct core_ctl_attr *cattr = to_attr(attr); |
| 397 | ssize_t ret = -EIO; |
| 398 | |
| 399 | if (cattr->store) |
| 400 | ret = cattr->store(data, buf, count); |
| 401 | |
| 402 | return ret; |
| 403 | } |
| 404 | |
| 405 | static const struct sysfs_ops sysfs_ops = { |
| 406 | .show = show, |
| 407 | .store = store, |
| 408 | }; |
| 409 | |
| 410 | static struct kobj_type ktype_core_ctl = { |
| 411 | .sysfs_ops = &sysfs_ops, |
| 412 | .default_attrs = default_attrs, |
| 413 | }; |
| 414 | |
| 415 | /* ==================== runqueue based core count =================== */ |
| 416 | |
| 417 | #define RQ_AVG_TOLERANCE 2 |
| 418 | #define RQ_AVG_DEFAULT_MS 20 |
| 419 | #define NR_RUNNING_TOLERANCE 5 |
| 420 | static unsigned int rq_avg_period_ms = RQ_AVG_DEFAULT_MS; |
| 421 | |
| 422 | static s64 rq_avg_timestamp_ms; |
| 423 | static struct timer_list rq_avg_timer; |
| 424 | |
| 425 | static void update_running_avg(bool trigger_update) |
| 426 | { |
| 427 | int cpu; |
| 428 | struct cpu_data *pcpu; |
| 429 | int avg, iowait_avg, big_avg, old_nrrun; |
| 430 | s64 now; |
| 431 | unsigned long flags; |
| 432 | |
| 433 | spin_lock_irqsave(&state_lock, flags); |
| 434 | |
| 435 | now = ktime_to_ms(ktime_get()); |
| 436 | if (now - rq_avg_timestamp_ms < rq_avg_period_ms - RQ_AVG_TOLERANCE) { |
| 437 | spin_unlock_irqrestore(&state_lock, flags); |
| 438 | return; |
| 439 | } |
| 440 | rq_avg_timestamp_ms = now; |
| 441 | sched_get_nr_running_avg(&avg, &iowait_avg, &big_avg); |
| 442 | |
| 443 | spin_unlock_irqrestore(&state_lock, flags); |
| 444 | |
| 445 | /* |
| 446 | * Round up to the next integer if the average nr running tasks |
| 447 | * is within NR_RUNNING_TOLERANCE/100 of the next integer. |
| 448 | * If normal rounding up is used, it will allow a transient task |
| 449 | * to trigger online event. By the time core is onlined, the task |
| 450 | * has finished. |
| 451 | * Rounding to closest suffers same problem because scheduler |
| 452 | * might only provide running stats per jiffy, and a transient |
| 453 | * task could skew the number for one jiffy. If core control |
| 454 | * samples every 2 jiffies, it will observe 0.5 additional running |
| 455 | * average which rounds up to 1 task. |
| 456 | */ |
| 457 | avg = (avg + NR_RUNNING_TOLERANCE) / 100; |
| 458 | big_avg = (big_avg + NR_RUNNING_TOLERANCE) / 100; |
| 459 | |
| 460 | for_each_possible_cpu(cpu) { |
| 461 | pcpu = &per_cpu(cpu_state, cpu); |
| 462 | if (!pcpu->inited || pcpu->first_cpu != cpu) |
| 463 | continue; |
| 464 | old_nrrun = pcpu->nrrun; |
| 465 | /* |
| 466 | * Big cluster only need to take care of big tasks, but if |
| 467 | * there are not enough big cores, big tasks need to be run |
| 468 | * on little as well. Thus for little's runqueue stat, it |
| 469 | * has to use overall runqueue average, or derive what big |
| 470 | * tasks would have to be run on little. The latter approach |
| 471 | * is not easy to get given core control reacts much slower |
| 472 | * than scheduler, and can't predict scheduler's behavior. |
| 473 | */ |
| 474 | pcpu->nrrun = pcpu->is_big_cluster ? big_avg : avg; |
| 475 | if (pcpu->nrrun != old_nrrun) { |
| 476 | if (trigger_update) |
| 477 | apply_need(pcpu); |
| 478 | else |
| 479 | pcpu->nrrun_changed = true; |
| 480 | } |
| 481 | } |
| 482 | } |
| 483 | |
| 484 | /* adjust needed CPUs based on current runqueue information */ |
| 485 | static unsigned int apply_task_need(struct cpu_data *f, unsigned int new_need) |
| 486 | { |
| 487 | /* Online all cores if there are enough tasks */ |
| 488 | if (f->nrrun >= f->task_thres) |
| 489 | return f->num_cpus; |
| 490 | |
| 491 | /* only online more cores if there are tasks to run */ |
| 492 | if (f->nrrun > new_need) |
| 493 | return new_need + 1; |
| 494 | |
| 495 | return new_need; |
| 496 | } |
| 497 | |
| 498 | static u64 round_to_nw_start(void) |
| 499 | { |
| 500 | unsigned long step = msecs_to_jiffies(rq_avg_period_ms); |
| 501 | u64 jif = get_jiffies_64(); |
| 502 | |
| 503 | do_div(jif, step); |
| 504 | return (jif + 1) * step; |
| 505 | } |
| 506 | |
| 507 | static void rq_avg_timer_func(unsigned long not_used) |
| 508 | { |
| 509 | update_running_avg(true); |
| 510 | mod_timer(&rq_avg_timer, round_to_nw_start()); |
| 511 | } |
| 512 | |
| 513 | /* ======================= load based core count ====================== */ |
| 514 | |
| 515 | static unsigned int apply_limits(struct cpu_data *f, unsigned int need_cpus) |
| 516 | { |
| 517 | return min(max(f->min_cpus, need_cpus), f->max_cpus); |
| 518 | } |
| 519 | |
| 520 | static bool eval_need(struct cpu_data *f) |
| 521 | { |
| 522 | unsigned long flags; |
| 523 | struct cpu_data *c; |
| 524 | unsigned int need_cpus = 0, last_need, thres_idx; |
| 525 | int ret = 0; |
| 526 | bool need_flag = false; |
| 527 | s64 now; |
| 528 | |
| 529 | if (unlikely(!f->inited)) |
| 530 | return 0; |
| 531 | |
| 532 | spin_lock_irqsave(&state_lock, flags); |
| 533 | thres_idx = f->online_cpus ? f->online_cpus - 1 : 0; |
| 534 | list_for_each_entry(c, &f->lru, sib) { |
| 535 | if (c->busy >= f->busy_up_thres[thres_idx]) |
| 536 | c->is_busy = true; |
| 537 | else if (c->busy < f->busy_down_thres[thres_idx]) |
| 538 | c->is_busy = false; |
| 539 | need_cpus += c->is_busy; |
| 540 | } |
| 541 | need_cpus = apply_task_need(f, need_cpus); |
| 542 | need_flag = apply_limits(f, need_cpus) != apply_limits(f, f->need_cpus); |
| 543 | last_need = f->need_cpus; |
| 544 | |
| 545 | now = ktime_to_ms(ktime_get()); |
| 546 | |
| 547 | if (need_cpus == last_need) { |
| 548 | f->need_ts = now; |
| 549 | spin_unlock_irqrestore(&state_lock, flags); |
| 550 | return 0; |
| 551 | } |
| 552 | |
| 553 | if (need_cpus > last_need) { |
| 554 | ret = 1; |
| 555 | } else if (need_cpus < last_need) { |
| 556 | s64 elapsed = now - f->need_ts; |
| 557 | |
| 558 | if (elapsed >= f->offline_delay_ms) { |
| 559 | ret = 1; |
| 560 | } else { |
| 561 | mod_timer(&f->timer, jiffies + |
| 562 | msecs_to_jiffies(f->offline_delay_ms)); |
| 563 | } |
| 564 | } |
| 565 | |
| 566 | if (ret) { |
| 567 | f->need_ts = now; |
| 568 | f->need_cpus = need_cpus; |
| 569 | } |
| 570 | |
| 571 | trace_core_ctl_eval_need(f->cpu, last_need, need_cpus, |
| 572 | ret && need_flag); |
| 573 | spin_unlock_irqrestore(&state_lock, flags); |
| 574 | |
| 575 | return ret && need_flag; |
| 576 | } |
| 577 | |
| 578 | static void apply_need(struct cpu_data *f) |
| 579 | { |
| 580 | if (eval_need(f)) |
| 581 | wake_up_hotplug_thread(f); |
| 582 | } |
| 583 | |
| 584 | static int core_ctl_set_busy(unsigned int cpu, unsigned int busy) |
| 585 | { |
| 586 | struct cpu_data *c = &per_cpu(cpu_state, cpu); |
| 587 | struct cpu_data *f; |
| 588 | unsigned int old_is_busy = c->is_busy; |
| 589 | |
| 590 | if (!c->inited) |
| 591 | return 0; |
| 592 | f = &per_cpu(cpu_state, c->first_cpu); |
| 593 | |
| 594 | update_running_avg(false); |
| 595 | if (c->busy == busy && !f->nrrun_changed) |
| 596 | return 0; |
| 597 | c->busy = busy; |
| 598 | f->nrrun_changed = false; |
| 599 | |
| 600 | apply_need(f); |
| 601 | trace_core_ctl_set_busy(cpu, busy, old_is_busy, c->is_busy); |
| 602 | return 0; |
| 603 | } |
| 604 | |
| 605 | /* ========================= core count enforcement ==================== */ |
| 606 | |
| 607 | /* |
| 608 | * If current thread is hotplug thread, don't attempt to wake up |
| 609 | * itself or other hotplug threads because it will deadlock. Instead, |
| 610 | * schedule a timer to fire in next timer tick and wake up the thread. |
| 611 | */ |
| 612 | static void wake_up_hotplug_thread(struct cpu_data *state) |
| 613 | { |
| 614 | unsigned long flags; |
| 615 | int cpu; |
| 616 | struct cpu_data *pcpu; |
| 617 | bool no_wakeup = false; |
| 618 | |
| 619 | for_each_possible_cpu(cpu) { |
| 620 | pcpu = &per_cpu(cpu_state, cpu); |
| 621 | if (cpu != pcpu->first_cpu) |
| 622 | continue; |
| 623 | if (pcpu->hotplug_thread == current) { |
| 624 | no_wakeup = true; |
| 625 | break; |
| 626 | } |
| 627 | } |
| 628 | |
| 629 | spin_lock_irqsave(&state->pending_lock, flags); |
| 630 | state->pending = true; |
| 631 | spin_unlock_irqrestore(&state->pending_lock, flags); |
| 632 | |
| 633 | if (no_wakeup) { |
| 634 | spin_lock_irqsave(&state_lock, flags); |
| 635 | mod_timer(&state->timer, jiffies); |
| 636 | spin_unlock_irqrestore(&state_lock, flags); |
| 637 | } else { |
| 638 | wake_up_process(state->hotplug_thread); |
| 639 | } |
| 640 | } |
| 641 | |
| 642 | static void core_ctl_timer_func(unsigned long cpu) |
| 643 | { |
| 644 | struct cpu_data *state = &per_cpu(cpu_state, cpu); |
| 645 | unsigned long flags; |
| 646 | |
| 647 | if (eval_need(state)) { |
| 648 | spin_lock_irqsave(&state->pending_lock, flags); |
| 649 | state->pending = true; |
| 650 | spin_unlock_irqrestore(&state->pending_lock, flags); |
| 651 | wake_up_process(state->hotplug_thread); |
| 652 | } |
| 653 | |
| 654 | } |
| 655 | |
| 656 | static int core_ctl_online_core(unsigned int cpu) |
| 657 | { |
| 658 | int ret; |
| 659 | struct device *dev; |
| 660 | |
| 661 | lock_device_hotplug(); |
| 662 | dev = get_cpu_device(cpu); |
| 663 | if (!dev) { |
| 664 | pr_err("%s: failed to get cpu%d device\n", __func__, cpu); |
| 665 | ret = -ENODEV; |
| 666 | } else { |
| 667 | ret = device_online(dev); |
| 668 | } |
| 669 | unlock_device_hotplug(); |
| 670 | return ret; |
| 671 | } |
| 672 | |
| 673 | static int core_ctl_offline_core(unsigned int cpu) |
| 674 | { |
| 675 | int ret; |
| 676 | struct device *dev; |
| 677 | |
| 678 | lock_device_hotplug(); |
| 679 | dev = get_cpu_device(cpu); |
| 680 | if (!dev) { |
| 681 | pr_err("%s: failed to get cpu%d device\n", __func__, cpu); |
| 682 | ret = -ENODEV; |
| 683 | } else { |
| 684 | ret = device_offline(dev); |
| 685 | } |
| 686 | unlock_device_hotplug(); |
| 687 | return ret; |
| 688 | } |
| 689 | |
| 690 | static void __ref do_hotplug(struct cpu_data *f) |
| 691 | { |
| 692 | unsigned int need; |
| 693 | struct cpu_data *c, *tmp; |
| 694 | |
| 695 | need = apply_limits(f, f->need_cpus); |
| 696 | pr_debug("Trying to adjust group %u to %u\n", f->first_cpu, need); |
| 697 | |
| 698 | if (f->online_cpus > need) { |
| 699 | list_for_each_entry_safe(c, tmp, &f->lru, sib) { |
| 700 | if (!c->online) |
| 701 | continue; |
| 702 | |
| 703 | if (f->online_cpus == need) |
| 704 | break; |
| 705 | |
| 706 | /* Don't offline busy CPUs. */ |
| 707 | if (c->is_busy) |
| 708 | continue; |
| 709 | |
| 710 | pr_debug("Trying to Offline CPU%u\n", c->cpu); |
| 711 | if (core_ctl_offline_core(c->cpu)) |
| 712 | pr_debug("Unable to Offline CPU%u\n", c->cpu); |
| 713 | } |
| 714 | |
| 715 | /* |
| 716 | * If the number of online CPUs is within the limits, then |
| 717 | * don't force any busy CPUs offline. |
| 718 | */ |
| 719 | if (f->online_cpus <= f->max_cpus) |
| 720 | return; |
| 721 | |
| 722 | list_for_each_entry_safe(c, tmp, &f->lru, sib) { |
| 723 | if (!c->online) |
| 724 | continue; |
| 725 | |
| 726 | if (f->online_cpus <= f->max_cpus) |
| 727 | break; |
| 728 | |
| 729 | pr_debug("Trying to Offline CPU%u\n", c->cpu); |
| 730 | if (core_ctl_offline_core(c->cpu)) |
| 731 | pr_debug("Unable to Offline CPU%u\n", c->cpu); |
| 732 | } |
| 733 | } else if (f->online_cpus < need) { |
| 734 | list_for_each_entry_safe(c, tmp, &f->lru, sib) { |
| 735 | if (c->online || c->rejected || c->not_preferred) |
| 736 | continue; |
| 737 | if (f->online_cpus == need) |
| 738 | break; |
| 739 | |
| 740 | pr_debug("Trying to Online CPU%u\n", c->cpu); |
| 741 | if (core_ctl_online_core(c->cpu)) |
| 742 | pr_debug("Unable to Online CPU%u\n", c->cpu); |
| 743 | } |
| 744 | |
| 745 | if (f->online_cpus == need) |
| 746 | return; |
| 747 | |
| 748 | |
| 749 | list_for_each_entry_safe(c, tmp, &f->lru, sib) { |
| 750 | if (c->online || c->rejected || !c->not_preferred) |
| 751 | continue; |
| 752 | if (f->online_cpus == need) |
| 753 | break; |
| 754 | |
| 755 | pr_debug("Trying to Online CPU%u\n", c->cpu); |
| 756 | if (core_ctl_online_core(c->cpu)) |
| 757 | pr_debug("Unable to Online CPU%u\n", c->cpu); |
| 758 | } |
| 759 | |
| 760 | } |
| 761 | } |
| 762 | |
| 763 | static int __ref try_hotplug(void *data) |
| 764 | { |
| 765 | struct cpu_data *f = data; |
| 766 | unsigned long flags; |
| 767 | |
| 768 | while (1) { |
| 769 | set_current_state(TASK_INTERRUPTIBLE); |
| 770 | spin_lock_irqsave(&f->pending_lock, flags); |
| 771 | if (!f->pending) { |
| 772 | spin_unlock_irqrestore(&f->pending_lock, flags); |
| 773 | schedule(); |
| 774 | if (kthread_should_stop()) |
| 775 | break; |
| 776 | spin_lock_irqsave(&f->pending_lock, flags); |
| 777 | } |
| 778 | set_current_state(TASK_RUNNING); |
| 779 | f->pending = false; |
| 780 | spin_unlock_irqrestore(&f->pending_lock, flags); |
| 781 | |
| 782 | do_hotplug(f); |
| 783 | } |
| 784 | |
| 785 | return 0; |
| 786 | } |
| 787 | |
| 788 | static int __ref cpu_callback(struct notifier_block *nfb, |
| 789 | unsigned long action, void *hcpu) |
| 790 | { |
| 791 | uint32_t cpu = (uintptr_t)hcpu; |
| 792 | struct cpu_data *state = &per_cpu(cpu_state, cpu); |
| 793 | struct cpu_data *f; |
| 794 | int ret = NOTIFY_OK; |
| 795 | unsigned long flags; |
| 796 | |
| 797 | /* Don't affect suspend resume */ |
| 798 | if (action & CPU_TASKS_FROZEN) |
| 799 | return NOTIFY_OK; |
| 800 | |
| 801 | if (unlikely(!state->inited)) |
| 802 | return NOTIFY_OK; |
| 803 | |
| 804 | f = &per_cpu(cpu_state, state->first_cpu); |
| 805 | |
| 806 | switch (action) { |
| 807 | case CPU_UP_PREPARE: |
| 808 | |
| 809 | /* If online state of CPU somehow got out of sync, fix it. */ |
| 810 | if (state->online) { |
| 811 | f->online_cpus--; |
| 812 | state->online = false; |
| 813 | pr_warn("CPU%d offline when state is online\n", cpu); |
| 814 | } |
| 815 | |
| 816 | if (state->rejected) { |
| 817 | state->rejected = false; |
| 818 | f->avail_cpus++; |
| 819 | } |
| 820 | |
| 821 | /* |
| 822 | * If a CPU is in the process of coming up, mark it as online |
| 823 | * so that there's no race with hotplug thread bringing up more |
| 824 | * CPUs than necessary. |
| 825 | */ |
| 826 | if (apply_limits(f, f->need_cpus) <= f->online_cpus) { |
| 827 | pr_debug("Prevent CPU%d onlining\n", cpu); |
| 828 | ret = NOTIFY_BAD; |
| 829 | } else { |
| 830 | state->online = true; |
| 831 | f->online_cpus++; |
| 832 | } |
| 833 | break; |
| 834 | |
| 835 | case CPU_ONLINE: |
| 836 | /* |
| 837 | * Moving to the end of the list should only happen in |
| 838 | * CPU_ONLINE and not on CPU_UP_PREPARE to prevent an |
| 839 | * infinite list traversal when thermal (or other entities) |
| 840 | * reject trying to online CPUs. |
| 841 | */ |
| 842 | spin_lock_irqsave(&state_lock, flags); |
| 843 | list_del(&state->sib); |
| 844 | list_add_tail(&state->sib, &f->lru); |
| 845 | spin_unlock_irqrestore(&state_lock, flags); |
| 846 | break; |
| 847 | |
| 848 | case CPU_DEAD: |
| 849 | /* Move a CPU to the end of the LRU when it goes offline. */ |
| 850 | spin_lock_irqsave(&state_lock, flags); |
| 851 | list_del(&state->sib); |
| 852 | list_add_tail(&state->sib, &f->lru); |
| 853 | spin_unlock_irqrestore(&state_lock, flags); |
| 854 | |
| 855 | /* Fall through */ |
| 856 | |
| 857 | case CPU_UP_CANCELED: |
| 858 | |
| 859 | /* If online state of CPU somehow got out of sync, fix it. */ |
| 860 | if (!state->online) { |
| 861 | f->online_cpus++; |
| 862 | pr_warn("CPU%d online when state is offline\n", cpu); |
| 863 | } |
| 864 | |
| 865 | if (!state->rejected && action == CPU_UP_CANCELED) { |
| 866 | state->rejected = true; |
| 867 | f->avail_cpus--; |
| 868 | } |
| 869 | |
| 870 | state->online = false; |
| 871 | state->busy = 0; |
| 872 | f->online_cpus--; |
| 873 | break; |
| 874 | } |
| 875 | |
| 876 | if (f->online_cpus < apply_limits(f, f->need_cpus) |
| 877 | && f->online_cpus < f->avail_cpus |
| 878 | && action == CPU_DEAD) |
| 879 | wake_up_hotplug_thread(f); |
| 880 | |
| 881 | return ret; |
| 882 | } |
| 883 | |
| 884 | static struct notifier_block __refdata cpu_notifier = { |
| 885 | .notifier_call = cpu_callback, |
| 886 | }; |
| 887 | |
| 888 | /* ============================ init code ============================== */ |
| 889 | |
| 890 | static int group_init(struct cpumask *mask) |
| 891 | { |
| 892 | struct device *dev; |
| 893 | unsigned int first_cpu = cpumask_first(mask); |
| 894 | struct cpu_data *f = &per_cpu(cpu_state, first_cpu); |
| 895 | struct cpu_data *state; |
| 896 | unsigned int cpu; |
| 897 | struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 }; |
| 898 | |
| 899 | if (likely(f->inited)) |
| 900 | return 0; |
| 901 | |
| 902 | dev = get_cpu_device(first_cpu); |
| 903 | if (!dev) |
| 904 | return -ENODEV; |
| 905 | |
| 906 | pr_info("Creating CPU group %d\n", first_cpu); |
| 907 | |
| 908 | f->num_cpus = cpumask_weight(mask); |
| 909 | if (f->num_cpus > MAX_CPUS_PER_GROUP) { |
| 910 | pr_err("HW configuration not supported\n"); |
| 911 | return -EINVAL; |
| 912 | } |
| 913 | f->min_cpus = 1; |
| 914 | f->max_cpus = f->num_cpus; |
| 915 | f->need_cpus = f->num_cpus; |
| 916 | f->avail_cpus = f->num_cpus; |
| 917 | f->offline_delay_ms = 100; |
| 918 | f->task_thres = UINT_MAX; |
| 919 | f->nrrun = f->num_cpus; |
| 920 | INIT_LIST_HEAD(&f->lru); |
| 921 | init_timer(&f->timer); |
| 922 | spin_lock_init(&f->pending_lock); |
| 923 | f->timer.function = core_ctl_timer_func; |
| 924 | f->timer.data = first_cpu; |
| 925 | |
| 926 | for_each_cpu(cpu, mask) { |
| 927 | pr_info("Init CPU%u state\n", cpu); |
| 928 | |
| 929 | state = &per_cpu(cpu_state, cpu); |
| 930 | state->cpu = cpu; |
| 931 | state->first_cpu = first_cpu; |
| 932 | |
| 933 | if (cpu_online(cpu)) { |
| 934 | f->online_cpus++; |
| 935 | state->online = true; |
| 936 | } |
| 937 | |
| 938 | list_add_tail(&state->sib, &f->lru); |
| 939 | } |
| 940 | |
| 941 | f->hotplug_thread = kthread_run(try_hotplug, (void *) f, |
| 942 | "core_ctl/%d", first_cpu); |
| 943 | sched_setscheduler_nocheck(f->hotplug_thread, SCHED_FIFO, ¶m); |
| 944 | |
| 945 | for_each_cpu(cpu, mask) { |
| 946 | state = &per_cpu(cpu_state, cpu); |
| 947 | state->inited = true; |
| 948 | } |
| 949 | |
| 950 | kobject_init(&f->kobj, &ktype_core_ctl); |
| 951 | return kobject_add(&f->kobj, &dev->kobj, "core_ctl"); |
| 952 | } |
| 953 | |
| 954 | static int cpufreq_policy_cb(struct notifier_block *nb, unsigned long val, |
| 955 | void *data) |
| 956 | { |
| 957 | struct cpufreq_policy *policy = data; |
| 958 | |
| 959 | switch (val) { |
| 960 | case CPUFREQ_CREATE_POLICY: |
| 961 | group_init(policy->related_cpus); |
| 962 | break; |
| 963 | } |
| 964 | |
| 965 | return NOTIFY_OK; |
| 966 | } |
| 967 | |
| 968 | static struct notifier_block cpufreq_pol_nb = { |
| 969 | .notifier_call = cpufreq_policy_cb, |
| 970 | }; |
| 971 | |
| 972 | static int cpufreq_gov_cb(struct notifier_block *nb, unsigned long val, |
| 973 | void *data) |
| 974 | { |
| 975 | struct cpufreq_govinfo *info = data; |
| 976 | |
| 977 | switch (val) { |
| 978 | case CPUFREQ_LOAD_CHANGE: |
| 979 | core_ctl_set_busy(info->cpu, info->load); |
| 980 | break; |
| 981 | } |
| 982 | |
| 983 | return NOTIFY_OK; |
| 984 | } |
| 985 | |
| 986 | static struct notifier_block cpufreq_gov_nb = { |
| 987 | .notifier_call = cpufreq_gov_cb, |
| 988 | }; |
| 989 | |
| 990 | static int __init core_ctl_init(void) |
| 991 | { |
| 992 | struct cpufreq_policy *policy; |
| 993 | unsigned int cpu; |
| 994 | |
| 995 | register_cpu_notifier(&cpu_notifier); |
| 996 | cpufreq_register_notifier(&cpufreq_pol_nb, CPUFREQ_POLICY_NOTIFIER); |
| 997 | cpufreq_register_notifier(&cpufreq_gov_nb, CPUFREQ_GOVINFO_NOTIFIER); |
| 998 | init_timer_deferrable(&rq_avg_timer); |
| 999 | rq_avg_timer.function = rq_avg_timer_func; |
| 1000 | |
| 1001 | get_online_cpus(); |
| 1002 | for_each_online_cpu(cpu) { |
| 1003 | policy = cpufreq_cpu_get(cpu); |
| 1004 | if (policy) { |
| 1005 | group_init(policy->related_cpus); |
| 1006 | cpufreq_cpu_put(policy); |
| 1007 | } |
| 1008 | } |
| 1009 | put_online_cpus(); |
| 1010 | mod_timer(&rq_avg_timer, round_to_nw_start()); |
| 1011 | return 0; |
| 1012 | } |
| 1013 | |
| 1014 | late_initcall(core_ctl_init); |