| 1 | #include <asm.h> |
| 2 | #include <bootstage.h> |
| 3 | #include <crypto/prng.h> |
| 4 | #include <irq/idt.h> |
| 5 | #include <math/clamp.h> |
| 6 | #include <math/fixed.h> |
| 7 | #include <math/min_max.h> |
| 8 | #include <registry.h> |
| 9 | #include <sch/sched.h> |
| 10 | #include <smp/smp.h> |
| 11 | #include <stdatomic.h> |
| 12 | #include <stdint.h> |
| 13 | |
| 14 | #include "internal.h" |
| 15 | |
| 16 | /* TODO: #define things <--- do this */ |
| 17 | |
| 18 | static void derive_timeshare_prio_range(enum thread_activity_class cls, |
| 19 | uint32_t *min, uint32_t *max); |
| 20 | |
| 21 | #define THREAD_DELTA_UNIT (1ULL << 3) |
| 22 | |
| 23 | #define THREAD_BOOST_WAKE_SMALL \ |
| 24 | (4ULL * THREAD_DELTA_UNIT) /* provisional boost for wake */ |
| 25 | |
| 26 | #define THREAD_BOOST_WAKE_LARGE \ |
| 27 | (12ULL * THREAD_DELTA_UNIT) /* stronger boost if very interactive */ |
| 28 | |
| 29 | #define THREAD_PENALTY_CPU_RUN \ |
| 30 | (3ULL * THREAD_DELTA_UNIT) /* penalize heavy CPU usage per run period */ |
| 31 | |
| 32 | #define THREAD_DELTA_MAX (1 << 9) |
| 33 | |
| 34 | #define THREAD_REINSERT_THRESHOLD \ |
| 35 | (8ULL * THREAD_DELTA_UNIT) /* only reinsert if effective priority changes \ |
| 36 | >= this */ |
| 37 | |
| 38 | #define THREAD_HYSTERESIS_MS 250ULL |
| 39 | #define THREAD_PROV_BOOST_MS 50ULL /* provisional boost expiration */ |
| 40 | |
| 41 | #define THREAD_MUL_INTERACTIVE 3ULL /* Big boost */ |
| 42 | #define THREAD_MUL_IO_BOUND 2ULL /* Small boost */ |
| 43 | #define THREAD_MUL_CPU_BOUND 1ULL /* No boost */ |
| 44 | #define THREAD_MUL_SLEEPY 1ULL /* No boost */ |
| 45 | |
| 46 | /* Scheduling periods */ |
| 47 | #define MIN_PERIOD_MS 20ULL /* don’t go too short */ |
| 48 | #define MAX_PERIOD_MS 300ULL /* don’t go too long */ |
| 49 | #define BASE_PERIOD_MS 50ULL /* baseline for small loads */ |
| 50 | |
| 51 | /* Timeslices */ |
| 52 | #define MIN_SLICE_MS 2 /* smallest slice granularity */ |
| 53 | #define MAX_SLICE_MS 20 /* cap slice length to avoid hogs */ |
| 54 | |
| 55 | #define WAKE_FREQ_MAX 20 |
| 56 | #define WAKE_FREQ_SCALE 5 |
| 57 | |
| 58 | #define THREAD_SLICE_MIN 1 |
| 59 | #define THREAD_SLICE_MAX 16 |
| 60 | |
| 61 | #define THREAD_BASE_WEIGHT 1024 |
| 62 | #define THREAD_WEIGHT_SCALING 100 |
| 63 | |
| 64 | #define NICE_BASE_FP FX(1.022) |
| 65 | |
| 66 | #define SET_MUL(__multiplier) \ |
| 67 | class_mul = __multiplier; \ |
| 68 | break; |
| 69 | |
| 70 | static enum thread_activity_class |
| 71 | classify_activity(struct thread_activity_metrics m) { |
| 72 | if (m.run_ratio > 80 && m.block_ratio < 10) |
| 73 | return THREAD_ACTIVITY_CLASS_CPU_BOUND; |
| 74 | |
| 75 | if (m.block_ratio > 40 && m.wake_freq > 2) |
| 76 | return THREAD_ACTIVITY_CLASS_IO_BOUND; |
| 77 | |
| 78 | if (m.wake_freq > 5) |
| 79 | return THREAD_ACTIVITY_CLASS_INTERACTIVE; |
| 80 | |
| 81 | if (m.sleep_ratio > 50) |
| 82 | return THREAD_ACTIVITY_CLASS_SLEEPY; |
| 83 | |
| 84 | /* Let's not confuse the rest of the scheduler |
| 85 | * with an unknown classification */ |
| 86 | return THREAD_ACTIVITY_CLASS_CPU_BOUND; |
| 87 | } |
| 88 | |
| 89 | void thread_classify_activity(struct thread *t, uint64_t now_ms) { |
| 90 | /* Rate limit to prevent rapid bouncing */ |
| 91 | if (now_ms - t->last_class_change_ms < THREAD_HYSTERESIS_MS) |
| 92 | return; |
| 93 | |
| 94 | t->activity_class = classify_activity(m: t->activity_metrics); |
| 95 | t->last_class_change_ms = now_ms; |
| 96 | } |
| 97 | |
| 98 | static struct thread_activity_metrics calc_activity_metrics(struct thread *t) { |
| 99 | struct thread_activity_metrics m = {0}; |
| 100 | uint64_t total_duration = 0; |
| 101 | uint64_t total_block = 0; |
| 102 | uint64_t total_sleep = 0; |
| 103 | uint64_t total_run = 0; |
| 104 | uint64_t total_wake_count = 0; |
| 105 | |
| 106 | for (size_t i = 0; i < THREAD_ACTIVITY_BUCKET_COUNT; i++) { |
| 107 | struct thread_activity_bucket *b = &t->activity_stats->buckets[i]; |
| 108 | struct thread_runtime_bucket *rtb = &t->activity_stats->rt_buckets[i]; |
| 109 | |
| 110 | uint64_t run_time = rtb->run_time_ms; |
| 111 | uint64_t block_time = b->block_duration; |
| 112 | uint64_t sleep_time = b->sleep_duration; |
| 113 | |
| 114 | total_run += run_time; |
| 115 | total_block += block_time; |
| 116 | total_sleep += sleep_time; |
| 117 | total_wake_count += b->wake_count; |
| 118 | } |
| 119 | |
| 120 | total_duration = total_run + total_block + total_sleep; |
| 121 | if (total_duration == 0) |
| 122 | total_duration = 1; |
| 123 | |
| 124 | m.run_ratio = (total_run * 100) / total_duration; |
| 125 | m.block_ratio = (total_block * 100) / total_duration; |
| 126 | m.sleep_ratio = (total_sleep * 100) / total_duration; |
| 127 | |
| 128 | m.wake_freq = total_wake_count / THREAD_ACTIVITY_BUCKET_COUNT; |
| 129 | |
| 130 | return m; |
| 131 | } |
| 132 | |
| 133 | void thread_calculate_activity_data(struct thread *t) { |
| 134 | if (thread_get_state(t) == THREAD_STATE_IDLE_THREAD) |
| 135 | return; |
| 136 | |
| 137 | struct thread_activity_metrics mtcs = calc_activity_metrics(t); |
| 138 | t->activity_metrics = mtcs; |
| 139 | } |
| 140 | |
| 141 | static void derive_timeshare_prio_range(enum thread_activity_class cls, |
| 142 | uint32_t *min, uint32_t *max) { |
| 143 | switch (cls) { |
| 144 | case THREAD_ACTIVITY_CLASS_INTERACTIVE: |
| 145 | *min = THREAD_ACT_INTERACTIVE_MIN; |
| 146 | *max = THREAD_ACT_INTERACTIVE_MAX; |
| 147 | break; |
| 148 | |
| 149 | case THREAD_ACTIVITY_CLASS_IO_BOUND: |
| 150 | *min = THREAD_ACT_IO_BOUND_MIN; |
| 151 | *max = THREAD_ACT_IO_BOUND_MAX; |
| 152 | break; |
| 153 | |
| 154 | case THREAD_ACTIVITY_CLASS_CPU_BOUND: |
| 155 | *min = THREAD_ACT_CPU_BOUND_MIN; |
| 156 | *max = THREAD_ACT_CPU_BOUND_MAX; |
| 157 | break; |
| 158 | |
| 159 | case THREAD_ACTIVITY_CLASS_SLEEPY: |
| 160 | default: |
| 161 | *min = THREAD_ACT_SLEEPY_MIN; |
| 162 | *max = THREAD_ACT_SLEEPY_MAX; |
| 163 | break; |
| 164 | } |
| 165 | } |
| 166 | |
| 167 | static uint32_t compute_activity_score_pct(struct thread_activity_metrics *m) { |
| 168 | uint32_t wake_norm = m->wake_freq > WAKE_FREQ_MAX ? 100 : m->wake_freq * 5; |
| 169 | |
| 170 | uint32_t interactive_pct = (wake_norm * (100u - m->block_ratio)) / 100u; |
| 171 | |
| 172 | uint32_t cpu_factor = 100u - m->run_ratio; |
| 173 | |
| 174 | uint32_t score_pct = (interactive_pct * cpu_factor) / 100u; |
| 175 | |
| 176 | uint32_t bias = score_pct / 8; |
| 177 | |
| 178 | uint32_t final_pct = score_pct + bias; |
| 179 | |
| 180 | if (final_pct > 100) |
| 181 | final_pct = 100; |
| 182 | |
| 183 | return final_pct; |
| 184 | } |
| 185 | |
| 186 | static inline int32_t jitter_for_thread(void) { |
| 187 | uint32_t v = prng_next(); |
| 188 | uint32_t jit = THREAD_REINSERT_THRESHOLD >> 2; |
| 189 | int32_t j = (int32_t) (v % (2 * jit + 1)) - (int32_t) jit; |
| 190 | return j; |
| 191 | } |
| 192 | |
| 193 | static int get_class_multiplier(enum thread_activity_class class) { |
| 194 | int class_mul; |
| 195 | switch (class) { |
| 196 | case THREAD_ACTIVITY_CLASS_INTERACTIVE: SET_MUL(THREAD_MUL_INTERACTIVE); |
| 197 | case THREAD_ACTIVITY_CLASS_IO_BOUND: SET_MUL(THREAD_MUL_IO_BOUND); |
| 198 | case THREAD_ACTIVITY_CLASS_CPU_BOUND: SET_MUL(THREAD_MUL_CPU_BOUND); |
| 199 | case THREAD_ACTIVITY_CLASS_SLEEPY: |
| 200 | default: SET_MUL(0); /* Unclassified thread or sleepy thread */ |
| 201 | } |
| 202 | return class_mul; |
| 203 | } |
| 204 | |
| 205 | void thread_apply_wake_boost(struct thread *t) { |
| 206 | if (thread_is_rt(t)) |
| 207 | return; |
| 208 | |
| 209 | uint32_t score_pct = compute_activity_score_pct(m: &t->activity_metrics); |
| 210 | int32_t mul = get_class_multiplier(class: t->activity_class); |
| 211 | |
| 212 | int64_t delta_change64 = score_pct * THREAD_DELTA_UNIT * mul / 100; |
| 213 | |
| 214 | delta_change64 += jitter_for_thread(); |
| 215 | |
| 216 | int64_t new_delta = (int64_t) t->dynamic_delta + delta_change64; |
| 217 | CLAMP(new_delta, -THREAD_DELTA_MAX, THREAD_DELTA_MAX); |
| 218 | |
| 219 | t->dynamic_delta = (int32_t) new_delta; |
| 220 | |
| 221 | thread_update_effective_priority(t); |
| 222 | } |
| 223 | |
| 224 | static int32_t compute_cpu_penalty(struct thread *t, int32_t base_penalty) { |
| 225 | struct thread_activity_metrics *m = &t->activity_metrics; |
| 226 | uint32_t run_scale = m->run_ratio; |
| 227 | uint32_t wake_scale = m->wake_freq * 2; |
| 228 | if (wake_scale > 100) |
| 229 | wake_scale = 100; |
| 230 | |
| 231 | int32_t penalty = base_penalty * run_scale / 100; |
| 232 | penalty -= base_penalty * wake_scale / 200; |
| 233 | if (penalty < 1) |
| 234 | penalty = 1; |
| 235 | |
| 236 | return penalty; |
| 237 | } |
| 238 | |
| 239 | void thread_apply_cpu_penalty(struct thread *t) { |
| 240 | thread_calculate_activity_data(t); /* Give it another update */ |
| 241 | |
| 242 | /* Apply the penalty */ |
| 243 | if (t->activity_class == THREAD_ACTIVITY_CLASS_CPU_BOUND) { |
| 244 | /* Small penalty */ |
| 245 | int32_t penalty = compute_cpu_penalty(t, THREAD_PENALTY_CPU_RUN); |
| 246 | int64_t scaled_delta = penalty * t->weight / MAX(t->weight, 1ULL); |
| 247 | |
| 248 | CLAMP(scaled_delta, -THREAD_DELTA_MAX, THREAD_DELTA_MAX); |
| 249 | t->dynamic_delta -= scaled_delta; |
| 250 | } |
| 251 | |
| 252 | thread_update_effective_priority(t); |
| 253 | } |
| 254 | |
| 255 | static int64_t base_weight_of(struct thread *t) { |
| 256 | struct thread_activity_metrics *m = &t->activity_metrics; |
| 257 | |
| 258 | int64_t w = THREAD_BASE_WEIGHT; |
| 259 | |
| 260 | w += m->wake_freq * THREAD_WEIGHT_SCALING; |
| 261 | w += (THREAD_WEIGHT_SCALING - m->run_ratio) * (THREAD_WEIGHT_SCALING / 2); |
| 262 | w += t->dynamic_delta / THREAD_WEIGHT_SCALING; |
| 263 | |
| 264 | int32_t effective_boost = t->niceness + t->climb_state.effective_boost; |
| 265 | |
| 266 | if (effective_boost) { |
| 267 | int32_t nf = fx_pow_i32(NICE_BASE_FP, exp: effective_boost); |
| 268 | w = (w * nf) >> 16; |
| 269 | } |
| 270 | |
| 271 | if (w < 1) |
| 272 | w = 1; |
| 273 | |
| 274 | return w; |
| 275 | } |
| 276 | |
| 277 | void thread_update_effective_priority(struct thread *t) { |
| 278 | uint32_t min, max; |
| 279 | derive_timeshare_prio_range(cls: t->activity_class, min: &min, max: &max); |
| 280 | |
| 281 | int64_t avg = ((int64_t) min + (int64_t) max) / 2; |
| 282 | int64_t eff = avg + t->dynamic_delta; |
| 283 | CLAMP(eff, min, max); |
| 284 | |
| 285 | t->activity_score = (thread_prio_t) eff; |
| 286 | t->virtual_runtime_left = thread_virtual_runtime_left(t); |
| 287 | t->weight = base_weight_of(t); |
| 288 | } |
| 289 | |
| 290 | static uint64_t compute_period(struct scheduler *s) { |
| 291 | uint64_t load = s->total_thread_count; |
| 292 | uint64_t period = BASE_PERIOD_MS + (load * 2); /* Linear growth */ |
| 293 | CLAMP(period, MIN_PERIOD_MS, MAX_PERIOD_MS); |
| 294 | return period; |
| 295 | } |
| 296 | |
| 297 | static uint64_t map_activity_score(uint64_t score) { |
| 298 | const uint64_t min_score = THREAD_ACT_SLEEPY_MIN; |
| 299 | const uint64_t max_score = THREAD_ACT_INTERACTIVE_MAX; |
| 300 | const uint64_t slice_delta = THREAD_SLICE_MAX - THREAD_SLICE_MIN; |
| 301 | const uint64_t score_range = max_score - min_score; |
| 302 | |
| 303 | CLAMP(score, min_score, max_score); |
| 304 | |
| 305 | uint64_t score_delta = score - min_score; |
| 306 | |
| 307 | return 1 + (score_delta * slice_delta) / score_range; |
| 308 | } |
| 309 | |
| 310 | static uint64_t thread_derive_slice_count(struct thread *t) { |
| 311 | uint64_t base = map_activity_score(score: t->activity_score); |
| 312 | int64_t adjust = 1; |
| 313 | |
| 314 | if (t->activity_metrics.block_ratio > t->activity_metrics.run_ratio) |
| 315 | adjust += 1; |
| 316 | |
| 317 | if (t->activity_metrics.run_ratio > 70) |
| 318 | adjust -= 1; |
| 319 | |
| 320 | if (t->activity_metrics.sleep_ratio > 70) |
| 321 | adjust -= 1; |
| 322 | |
| 323 | int64_t result = (int64_t) base + adjust; |
| 324 | |
| 325 | CLAMP(result, THREAD_SLICE_MIN, THREAD_SLICE_MAX); |
| 326 | |
| 327 | return (uint64_t) result; |
| 328 | } |
| 329 | |
| 330 | static void allocate_slices(struct scheduler *s, uint64_t now_ms) { |
| 331 | s->period_ms = compute_period(s); |
| 332 | s->period_start_ms = now_ms; |
| 333 | |
| 334 | uint64_t total_weight = 0; |
| 335 | struct rbt_node *node; |
| 336 | rbt_for_each(node, &s->thread_rbt) { |
| 337 | struct thread *t = thread_from_rq_rbt_node(node); |
| 338 | total_weight += t->weight; |
| 339 | } |
| 340 | |
| 341 | if (total_weight == 0) |
| 342 | total_weight = 1; |
| 343 | |
| 344 | rbt_for_each(node, &s->thread_rbt) { |
| 345 | struct thread *t = thread_from_rq_rbt_node(node); |
| 346 | |
| 347 | uint64_t budget_ms = (s->period_ms * t->weight) / total_weight; |
| 348 | if (budget_ms < MIN_SLICE_MS) |
| 349 | budget_ms = MIN_SLICE_MS; |
| 350 | |
| 351 | t->period_runtime_raw_ms = 0; |
| 352 | t->budget_time_raw_ms = budget_ms; |
| 353 | |
| 354 | uint64_t slices = thread_derive_slice_count(t); |
| 355 | t->timeslice_length_raw_ms = budget_ms / slices; |
| 356 | |
| 357 | uint64_t mult = t->activity_score == 0 ? 1 : t->activity_score; |
| 358 | |
| 359 | t->virtual_period_runtime = 0; |
| 360 | t->virtual_budget = budget_ms * mult; |
| 361 | t->completed_period = s->current_period - 1; |
| 362 | } |
| 363 | } |
| 364 | |
| 365 | static void scheduler_update_thread_weights(struct scheduler *s) { |
| 366 | struct rbt_node *node; |
| 367 | rbt_for_each(node, &s->thread_rbt) { |
| 368 | struct thread *t = rbt_entry(node, struct thread, rq_tree_node); |
| 369 | |
| 370 | /* This will recalculate activity data |
| 371 | * and update the effective priority */ |
| 372 | thread_apply_cpu_penalty(t); |
| 373 | } |
| 374 | } |
| 375 | |
| 376 | void scheduler_period_start(struct scheduler *s, uint64_t now_ms) { |
| 377 | s->current_period++; |
| 378 | |
| 379 | scheduler_update_thread_weights(s); |
| 380 | |
| 381 | allocate_slices(s, now_ms); |
| 382 | |
| 383 | s->period_enabled = true; |
| 384 | } |
| 385 | |