| 1 | #include <kassert.h> |
| 2 | #include <sch/sched.h> |
| 3 | #include <smp/core.h> |
| 4 | #include <thread/apc.h> |
| 5 | #include <thread/thread.h> |
| 6 | |
| 7 | #include "sch/internal.h" |
| 8 | #include <mem/alloc.h> |
| 9 | |
| 10 | static inline bool safe_to_exec_apcs(void) { |
| 11 | if (irql_get() != IRQL_PASSIVE_LEVEL || !irq_not_in_interrupt()) |
| 12 | return false; |
| 13 | |
| 14 | struct thread *curr = thread_get_current(); |
| 15 | if (!curr || thread_get_state(t: curr) != THREAD_STATE_RUNNING) |
| 16 | return false; |
| 17 | |
| 18 | return true; |
| 19 | } |
| 20 | |
| 21 | static inline enum apc_state apc_state_load(struct apc *a) { |
| 22 | return atomic_load_explicit(&a->state, memory_order_acquire); |
| 23 | } |
| 24 | |
| 25 | static inline size_t apc_type_bit(enum apc_type t) { |
| 26 | return (size_t) 1ULL << (size_t) t; |
| 27 | } |
| 28 | |
| 29 | static inline bool apc_queue_empty(struct apc_queue *q) { |
| 30 | return q->head == NULL; |
| 31 | } |
| 32 | |
| 33 | static inline void apc_enqueue_tail(struct apc_queue *q, struct apc *a) { |
| 34 | a->next = NULL; |
| 35 | |
| 36 | if (!q->head) { |
| 37 | q->head = q->tail = a; |
| 38 | } else { |
| 39 | q->tail->next = a; |
| 40 | q->tail = a; |
| 41 | } |
| 42 | } |
| 43 | |
| 44 | static inline struct apc *apc_dequeue_head(struct apc_queue *q) { |
| 45 | struct apc *a = q->head; |
| 46 | if (!a) |
| 47 | return NULL; |
| 48 | |
| 49 | q->head = a->next; |
| 50 | if (!q->head) |
| 51 | q->tail = NULL; |
| 52 | |
| 53 | a->next = NULL; |
| 54 | return a; |
| 55 | } |
| 56 | |
| 57 | static inline void apc_add_tail(struct thread *t, struct apc *a, |
| 58 | enum apc_type type) { |
| 59 | apc_enqueue_tail(q: &t->apc_head[type], a); |
| 60 | } |
| 61 | |
| 62 | static inline bool apc_list_empty(struct thread *t, enum apc_type type) { |
| 63 | return apc_queue_empty(q: &t->apc_head[type]); |
| 64 | } |
| 65 | |
| 66 | static inline void apc_unset_bitmask(struct thread *t, enum apc_type type) { |
| 67 | atomic_fetch_and(&t->apc_pending_mask, ~apc_type_bit(type)); |
| 68 | } |
| 69 | |
| 70 | static inline void apc_set_bitmask(struct thread *t, enum apc_type type) { |
| 71 | atomic_fetch_or(&t->apc_pending_mask, apc_type_bit(type)); |
| 72 | } |
| 73 | |
| 74 | static inline bool thread_can_exec_special_apcs(struct thread *t) { |
| 75 | return t->special_apc_disable == 0 && |
| 76 | (atomic_load(&t->apc_pending_mask) & |
| 77 | apc_type_bit(t: APC_TYPE_SPECIAL_KERNEL)); |
| 78 | } |
| 79 | |
| 80 | static inline bool thread_can_exec_kernel_apcs(struct thread *t) { |
| 81 | return t->kernel_apc_disable == 0 && |
| 82 | (atomic_load(&t->apc_pending_mask) & apc_type_bit(t: APC_TYPE_KERNEL)); |
| 83 | } |
| 84 | |
| 85 | static inline bool thread_is_dying(struct thread *t) { |
| 86 | if (thread_get_flags(t) & THREAD_FLAG_DYING) |
| 87 | return true; |
| 88 | enum thread_state s = thread_get_state(t); |
| 89 | return s == THREAD_STATE_TERMINATED || s == THREAD_STATE_ZOMBIE; |
| 90 | } |
| 91 | |
| 92 | static bool thread_apc_sanity_check(struct thread *t) { |
| 93 | if (unlikely(thread_get_state(t) == THREAD_STATE_IDLE_THREAD)) |
| 94 | panic("Attempted to put an APC on the idle thread" ); |
| 95 | |
| 96 | if (unlikely(thread_is_dying(t))) |
| 97 | return false; |
| 98 | |
| 99 | return true; |
| 100 | } |
| 101 | |
| 102 | static void apc_execute(struct apc *a) { |
| 103 | kassert(irql_get() == IRQL_APC_LEVEL); |
| 104 | |
| 105 | struct thread *curr = thread_get_current(); |
| 106 | |
| 107 | thread_or_flags(t: curr, flags: THREAD_FLAG_EXECUTING_APC); |
| 108 | |
| 109 | a->func(a->ctx); |
| 110 | |
| 111 | thread_and_flags(t: curr, flags: ~THREAD_FLAG_EXECUTING_APC); |
| 112 | curr->total_apcs_ran++; |
| 113 | } |
| 114 | |
| 115 | static void deliver_apc_type(struct thread *t, enum apc_type type) { |
| 116 | while (true) { |
| 117 | bool ok; |
| 118 | enum irql irql = thread_acquire(t, success: &ok); |
| 119 | if (!ok) |
| 120 | return; |
| 121 | |
| 122 | struct apc *apc = apc_dequeue_head(q: &t->apc_head[type]); |
| 123 | |
| 124 | if (!apc) { |
| 125 | apc_unset_bitmask(t, type); |
| 126 | thread_release(t, irql); |
| 127 | return; |
| 128 | } |
| 129 | |
| 130 | kassert(apc->owner == t); |
| 131 | kassert(apc_state_load(apc) == APC_STATE_QUEUED); |
| 132 | apc->owner = NULL; |
| 133 | atomic_store_explicit(&apc->state, APC_STATE_EXECUTING, |
| 134 | memory_order_release); |
| 135 | |
| 136 | thread_release(t, irql); |
| 137 | |
| 138 | apc_execute(a: apc); |
| 139 | atomic_store_explicit(&apc->state, APC_STATE_IDLE, |
| 140 | memory_order_release); |
| 141 | apc_put(a: apc); |
| 142 | } |
| 143 | } |
| 144 | |
| 145 | static void add_apc_to_thread(struct thread *t, struct apc *a, |
| 146 | enum apc_type type) { |
| 147 | a->owner = t; |
| 148 | apc_add_tail(t, a, type); |
| 149 | apc_set_bitmask(t, type); |
| 150 | } |
| 151 | |
| 152 | static inline bool thread_is_active(struct thread *t) { |
| 153 | enum thread_state s = thread_get_state(t); |
| 154 | return s == THREAD_STATE_READY || s == THREAD_STATE_RUNNING; |
| 155 | } |
| 156 | |
| 157 | static void maybe_force_resched(struct thread *t) { |
| 158 | enum irql irql; |
| 159 | struct scheduler *sched = thread_get_scheduler(t, sirql_out: &irql); |
| 160 | |
| 161 | scheduler_force_resched(sched); |
| 162 | |
| 163 | spin_unlock(&sched->lock, irql); |
| 164 | } |
| 165 | |
| 166 | static void wake_if_waiting(struct thread *t) { |
| 167 | if (thread_is_active(t)) |
| 168 | maybe_force_resched(t); |
| 169 | |
| 170 | /* Get it running again */ |
| 171 | if (!thread_apc_sanity_check(t)) |
| 172 | return; |
| 173 | |
| 174 | /* set the wake_src as the thread that enqueued the APC */ |
| 175 | scheduler_wake_manual(t, /* wake_src = */ t); |
| 176 | } |
| 177 | |
| 178 | bool apc_enqueue(struct thread *t, struct apc *a, enum apc_type type) { |
| 179 | if (!t || !a || type >= APC_TYPE_COUNT || !thread_apc_sanity_check(t)) |
| 180 | return false; |
| 181 | |
| 182 | bool ok; |
| 183 | enum irql irql = thread_acquire(t, success: &ok); |
| 184 | if (!ok) |
| 185 | return false; |
| 186 | |
| 187 | if (!apc_get(a)) { |
| 188 | thread_release(t, irql); |
| 189 | return false; |
| 190 | } |
| 191 | |
| 192 | enum apc_state expected = APC_STATE_IDLE; |
| 193 | if (!atomic_compare_exchange_strong_explicit( |
| 194 | &a->state, &expected, APC_STATE_QUEUED, memory_order_acq_rel, |
| 195 | memory_order_acquire)) { |
| 196 | thread_release(t, irql); |
| 197 | apc_put(a); |
| 198 | return false; |
| 199 | } |
| 200 | |
| 201 | kassert(a->owner == NULL); |
| 202 | add_apc_to_thread(t, a, type); |
| 203 | thread_release(t, irql); |
| 204 | |
| 205 | /* Let's go and execute em */ |
| 206 | if (t == thread_get_current()) { |
| 207 | apc_check_and_deliver(t); |
| 208 | } else { |
| 209 | /* Not us, go wake up the other guy */ |
| 210 | wake_if_waiting(t); |
| 211 | } |
| 212 | return true; |
| 213 | } |
| 214 | |
| 215 | /* We can only enqueue and run from ourselves, no sync needed */ |
| 216 | bool apc_enqueue_event_apc(struct event_apc *a, struct apc_event_desc *desc) { |
| 217 | kassert(desc); |
| 218 | if (!a || !apc_get(a: &a->apc)) |
| 219 | return false; |
| 220 | |
| 221 | enum apc_state expected = APC_STATE_IDLE; |
| 222 | if (!atomic_compare_exchange_strong_explicit( |
| 223 | &a->apc.state, &expected, APC_STATE_QUEUED, memory_order_acq_rel, |
| 224 | memory_order_acquire)) { |
| 225 | apc_put(a: &a->apc); |
| 226 | return false; |
| 227 | } |
| 228 | |
| 229 | kassert(!a->apc.owner); |
| 230 | |
| 231 | a->desc = desc; |
| 232 | |
| 233 | struct thread *t = thread_get_current(); |
| 234 | if (!thread_apc_sanity_check(t)) { |
| 235 | atomic_store_explicit(&a->apc.state, APC_STATE_IDLE, |
| 236 | memory_order_release); |
| 237 | apc_put(a: &a->apc); |
| 238 | return false; |
| 239 | } |
| 240 | |
| 241 | apc_enqueue_tail(q: &t->event_apcs, a: &a->apc); |
| 242 | |
| 243 | a->apc.owner = t; |
| 244 | apc_set_bitmask(t, type: APC_TYPE_KERNEL); |
| 245 | return true; |
| 246 | } |
| 247 | |
| 248 | static bool try_cancel_from_queue(struct thread *t, struct apc *a, |
| 249 | enum apc_type type) { |
| 250 | struct apc_queue *q = &t->apc_head[type]; |
| 251 | |
| 252 | struct apc *prev = NULL; |
| 253 | struct apc *curr = q->head; |
| 254 | |
| 255 | while (curr) { |
| 256 | struct apc *next = curr->next; |
| 257 | |
| 258 | if (curr == a) { |
| 259 | kassert(curr->owner == t); |
| 260 | kassert(apc_state_load(curr) == APC_STATE_QUEUED); |
| 261 | if (prev) |
| 262 | prev->next = next; |
| 263 | else |
| 264 | q->head = next; |
| 265 | |
| 266 | if (q->tail == curr) |
| 267 | q->tail = prev; |
| 268 | |
| 269 | curr->next = NULL; |
| 270 | curr->owner = NULL; |
| 271 | |
| 272 | return true; |
| 273 | } |
| 274 | |
| 275 | prev = curr; |
| 276 | curr = next; |
| 277 | } |
| 278 | |
| 279 | return false; |
| 280 | } |
| 281 | |
| 282 | /* update pending mask if queue now empty */ |
| 283 | static inline void update_pending_mask(struct thread *t, enum apc_type type) { |
| 284 | if (apc_list_empty(t, type)) |
| 285 | atomic_fetch_and(&t->apc_pending_mask, ~apc_type_bit(type)); |
| 286 | } |
| 287 | |
| 288 | bool apc_cancel(struct thread *t, struct apc *a) { |
| 289 | if (!t || !a) |
| 290 | return false; |
| 291 | |
| 292 | bool removed = false; |
| 293 | bool ok; |
| 294 | enum irql irql = thread_acquire(t, success: &ok); |
| 295 | if (!ok) |
| 296 | return false; |
| 297 | |
| 298 | for (int type = 0; type < APC_TYPE_COUNT; type++) { |
| 299 | removed = try_cancel_from_queue(t, a, type); |
| 300 | |
| 301 | if (removed) { |
| 302 | update_pending_mask(t, type); |
| 303 | break; |
| 304 | } |
| 305 | } |
| 306 | |
| 307 | thread_release(t, irql); |
| 308 | if (removed) { |
| 309 | atomic_store_explicit(&a->state, APC_STATE_IDLE, memory_order_release); |
| 310 | apc_put(a); |
| 311 | } |
| 312 | return removed; |
| 313 | } |
| 314 | |
| 315 | struct apc *apc_create(void) { |
| 316 | return kmalloc(sizeof(struct apc)); |
| 317 | } |
| 318 | |
| 319 | struct event_apc *apc_event_apc_create(void) { |
| 320 | return kmalloc(sizeof(struct event_apc)); |
| 321 | } |
| 322 | |
| 323 | void apc_init(struct apc *a, apc_func_t fn, void *arg1, apc_destroy_t destroy) { |
| 324 | kassert(a); |
| 325 | kassert(fn); |
| 326 | a->func = fn; |
| 327 | a->ctx = arg1; |
| 328 | a->next = NULL; |
| 329 | a->owner = NULL; |
| 330 | refcount_init(rc: &a->refcount, val: 1); |
| 331 | atomic_store_explicit(&a->state, APC_STATE_IDLE, memory_order_relaxed); |
| 332 | a->destroy = destroy; |
| 333 | } |
| 334 | |
| 335 | void apc_event_apc_init(struct event_apc *a, apc_func_t fn, void *arg1, |
| 336 | apc_destroy_t destroy) { |
| 337 | apc_init(a: &a->apc, fn, arg1, destroy); |
| 338 | a->execute_times = 0; |
| 339 | } |
| 340 | |
| 341 | bool apc_get(struct apc *a) { |
| 342 | return a && refcount_inc_not_zero(rc: &a->refcount); |
| 343 | } |
| 344 | |
| 345 | void apc_put(struct apc *a) { |
| 346 | kassert(a); |
| 347 | if (!refcount_dec_and_test(rc: &a->refcount)) |
| 348 | return; |
| 349 | |
| 350 | kassert(apc_state_load(a) == APC_STATE_IDLE); |
| 351 | kassert(a->owner == NULL); |
| 352 | kassert(a->next == NULL); |
| 353 | if (a->destroy) |
| 354 | a->destroy(a); |
| 355 | } |
| 356 | |
| 357 | void apc_destroy_free(struct apc *a) { |
| 358 | kfree(a); |
| 359 | } |
| 360 | |
| 361 | static void apc_queue_splice(struct apc_queue *from, struct apc_queue *to) { |
| 362 | struct apc *a; |
| 363 | |
| 364 | while ((a = apc_dequeue_head(q: from))) |
| 365 | apc_enqueue_tail(q: to, a); |
| 366 | } |
| 367 | |
| 368 | void apc_rundown_thread(struct thread *t) { |
| 369 | struct apc_queue drained = {0}; |
| 370 | |
| 371 | /* Every other queue mutator will hold t->lock, |
| 372 | * and rundown has to do that too */ |
| 373 | enum irql irql = spin_lock_irq_disable(&t->lock); |
| 374 | |
| 375 | for (size_t type = 0; type < APC_TYPE_COUNT; type++) |
| 376 | apc_queue_splice(from: &t->apc_head[type], to: &drained); |
| 377 | |
| 378 | apc_queue_splice(from: &t->event_apcs, to: &drained); |
| 379 | apc_queue_splice(from: &t->to_exec_event_apcs, to: &drained); |
| 380 | atomic_store_explicit(&t->apc_pending_mask, 0, memory_order_release); |
| 381 | |
| 382 | spin_unlock(&t->lock, irql); |
| 383 | |
| 384 | /* We can't run the teardown from the critical section */ |
| 385 | struct apc *a; |
| 386 | while ((a = apc_dequeue_head(q: &drained))) { |
| 387 | a->owner = NULL; |
| 388 | atomic_store_explicit(&a->state, APC_STATE_IDLE, memory_order_release); |
| 389 | apc_put(a); |
| 390 | } |
| 391 | } |
| 392 | |
| 393 | static void bump_counters_on_queue(struct apc_queue *from, |
| 394 | struct apc_event_desc *desc, |
| 395 | struct apc_queue *to) { |
| 396 | struct apc *prev = NULL; |
| 397 | struct apc *curr = from->head; |
| 398 | |
| 399 | while (curr) { |
| 400 | struct apc *next = curr->next; |
| 401 | struct event_apc *eapc = container_of(curr, struct event_apc, apc); |
| 402 | |
| 403 | if (eapc->desc == desc) { |
| 404 | eapc->execute_times++; |
| 405 | |
| 406 | if (to) { |
| 407 | /* unlink */ |
| 408 | if (prev) |
| 409 | prev->next = next; |
| 410 | else |
| 411 | from->head = next; |
| 412 | |
| 413 | if (from->tail == curr) |
| 414 | from->tail = prev; |
| 415 | |
| 416 | /* enqueue into target */ |
| 417 | curr->next = NULL; |
| 418 | apc_enqueue_tail(q: to, a: curr); |
| 419 | |
| 420 | curr = next; |
| 421 | continue; |
| 422 | } |
| 423 | } |
| 424 | |
| 425 | prev = curr; |
| 426 | curr = next; |
| 427 | } |
| 428 | } |
| 429 | |
| 430 | static void apc_execute_event(struct apc *a) { |
| 431 | kassert(apc_state_load(a) == APC_STATE_QUEUED); |
| 432 | atomic_store_explicit(&a->state, APC_STATE_EXECUTING, memory_order_release); |
| 433 | apc_execute(a); |
| 434 | atomic_store_explicit(&a->state, APC_STATE_QUEUED, memory_order_release); |
| 435 | } |
| 436 | |
| 437 | void apc_event_signal(struct apc_event_desc *desc) { |
| 438 | /* here we want to do two things: first, we identify if it is safe to |
| 439 | * execute APCs. if it is, it must be guaranteed that the to_execute |
| 440 | * tree of event APCs is empty, because the irql_lower that should've |
| 441 | * happened would have executed anything on that tree. in this case, |
| 442 | * we check our event_apcs tree, and execute anything of relevance |
| 443 | * in there. if it is not safe to execute APCs, we will check |
| 444 | * the to_execute tree, increment counters for all relevant APCs, and then |
| 445 | * check the event_apcs tree, and move anything necessary over */ |
| 446 | struct thread *curr = thread_get_current(); |
| 447 | |
| 448 | if (safe_to_exec_apcs() && curr->kernel_apc_disable == 0) { |
| 449 | kassert(apc_queue_empty(&curr->to_exec_event_apcs)); |
| 450 | |
| 451 | enum irql irql = irql_raise(new_level: IRQL_APC_LEVEL); |
| 452 | |
| 453 | struct apc *a = curr->event_apcs.head; |
| 454 | |
| 455 | /* This will give us the "first node in a list" that matches our `desc` |
| 456 | * value. We can keep going this->right->right to find everyone else to |
| 457 | * execute */ |
| 458 | while (a) { |
| 459 | if (container_of(a, struct event_apc, apc)->desc == desc) |
| 460 | apc_execute_event(a); |
| 461 | |
| 462 | a = a->next; |
| 463 | } |
| 464 | |
| 465 | irql_lower(old_level: irql); |
| 466 | } else { |
| 467 | /* Cannot execute APCs right now. Search both trees, bump counters. */ |
| 468 | bump_counters_on_queue(from: &curr->to_exec_event_apcs, desc, NULL); |
| 469 | bump_counters_on_queue(from: &curr->event_apcs, desc, |
| 470 | to: &curr->to_exec_event_apcs); |
| 471 | |
| 472 | apc_set_bitmask(t: curr, type: APC_TYPE_KERNEL); |
| 473 | } |
| 474 | } |
| 475 | |
| 476 | void thread_exec_event_apcs(struct thread *t) { |
| 477 | struct apc *a; |
| 478 | |
| 479 | while ((a = apc_dequeue_head(q: &t->to_exec_event_apcs))) { |
| 480 | struct event_apc *eapc = container_of(a, struct event_apc, apc); |
| 481 | kassert(eapc->execute_times); |
| 482 | |
| 483 | for (size_t i = 0; i < eapc->execute_times; i++) |
| 484 | apc_execute_event(a); |
| 485 | |
| 486 | eapc->execute_times = 0; |
| 487 | |
| 488 | apc_enqueue_tail(q: &t->event_apcs, a); |
| 489 | } |
| 490 | |
| 491 | /* Just in case */ |
| 492 | apc_unset_bitmask(t, type: APC_TYPE_KERNEL); |
| 493 | } |
| 494 | |
| 495 | void apc_disable_special() { |
| 496 | thread_get_current()->special_apc_disable++; |
| 497 | } |
| 498 | |
| 499 | void apc_enable_special() { |
| 500 | struct thread *t = thread_get_current(); |
| 501 | kassert(t->special_apc_disable > 0); |
| 502 | |
| 503 | if (--t->special_apc_disable == 0) |
| 504 | apc_check_and_deliver(t); |
| 505 | } |
| 506 | |
| 507 | void apc_disable_kernel() { |
| 508 | thread_get_current()->kernel_apc_disable++; |
| 509 | } |
| 510 | |
| 511 | void apc_enable_kernel() { |
| 512 | struct thread *t = thread_get_current(); |
| 513 | kassert(t->kernel_apc_disable > 0); |
| 514 | |
| 515 | if (--t->kernel_apc_disable == 0) |
| 516 | apc_check_and_deliver(t); |
| 517 | } |
| 518 | |
| 519 | static inline bool thread_can_exec_any_apcs(struct thread *t) { |
| 520 | return thread_can_exec_special_apcs(t) || thread_can_exec_kernel_apcs(t); |
| 521 | } |
| 522 | |
| 523 | void thread_exec_apcs(struct thread *t) { |
| 524 | if (thread_can_exec_special_apcs(t)) |
| 525 | deliver_apc_type(t, type: APC_TYPE_SPECIAL_KERNEL); |
| 526 | |
| 527 | if (thread_can_exec_kernel_apcs(t)) { |
| 528 | deliver_apc_type(t, type: APC_TYPE_KERNEL); |
| 529 | thread_exec_event_apcs(t); |
| 530 | } |
| 531 | } |
| 532 | |
| 533 | void apc_check_and_deliver(struct thread *t) { |
| 534 | if (!t || !safe_to_exec_apcs() || !thread_can_exec_any_apcs(t)) |
| 535 | return; |
| 536 | |
| 537 | if (thread_get_flags(t) & (THREAD_FLAG_EXECUTING_APC | THREAD_FLAG_DYING)) |
| 538 | return; |
| 539 | |
| 540 | enum irql irql = irql_raise(new_level: IRQL_APC_LEVEL); |
| 541 | |
| 542 | thread_exec_apcs(t); |
| 543 | |
| 544 | irql_lower(old_level: irql); |
| 545 | } |
| 546 | |