| 1 | // SPDX-License-Identifier: LGPL-2.1-or-later |
| 2 | /* |
| 3 | * Copyright (C) 2019-2021 Paul Cercueil <paul@crapouillou.net> |
| 4 | */ |
| 5 | |
| 6 | #include "blockcache.h" |
| 7 | #include "debug.h" |
| 8 | #include "interpreter.h" |
| 9 | #include "lightrec-private.h" |
| 10 | #include "memmanager.h" |
| 11 | #include "reaper.h" |
| 12 | #include "slist.h" |
| 13 | |
| 14 | #include <errno.h> |
| 15 | #include <stdatomic.h> |
| 16 | #include <stdbool.h> |
| 17 | #include <stdlib.h> |
| 18 | #include <pthread.h> |
| 19 | #ifdef __linux__ |
| 20 | #include <unistd.h> |
| 21 | #endif |
| 22 | |
| 23 | struct block_rec { |
| 24 | struct block *block; |
| 25 | struct slist_elm slist; |
| 26 | unsigned int requests; |
| 27 | bool compiling; |
| 28 | }; |
| 29 | |
| 30 | struct recompiler_thd { |
| 31 | struct lightrec_cstate *cstate; |
| 32 | unsigned int tid; |
| 33 | pthread_t thd; |
| 34 | }; |
| 35 | |
| 36 | struct recompiler { |
| 37 | struct lightrec_state *state; |
| 38 | pthread_cond_t cond; |
| 39 | pthread_cond_t cond2; |
| 40 | pthread_mutex_t mutex; |
| 41 | bool stop, pause, must_flush; |
| 42 | struct slist_elm slist; |
| 43 | |
| 44 | pthread_mutex_t alloc_mutex; |
| 45 | |
| 46 | unsigned int nb_recs; |
| 47 | struct recompiler_thd thds[]; |
| 48 | }; |
| 49 | |
| 50 | static unsigned int get_processors_count(void) |
| 51 | { |
| 52 | int nb = 1; |
| 53 | |
| 54 | #if defined(PTW32_VERSION) |
| 55 | nb = pthread_num_processors_np(); |
| 56 | #elif defined(__APPLE__) || defined(__FreeBSD__) |
| 57 | int count; |
| 58 | size_t size = sizeof(count); |
| 59 | |
| 60 | nb = sysctlbyname("hw.ncpu", &count, &size, NULL, 0) ? 1 : count; |
| 61 | #elif defined(_SC_NPROCESSORS_ONLN) |
| 62 | nb = (int)sysconf(_SC_NPROCESSORS_ONLN); |
| 63 | #endif |
| 64 | |
| 65 | return nb < 1 ? 1 : nb; |
| 66 | } |
| 67 | |
| 68 | static struct block_rec * lightrec_get_best_elm(struct slist_elm *head) |
| 69 | { |
| 70 | struct block_rec *block_rec, *best = NULL; |
| 71 | struct slist_elm *elm; |
| 72 | |
| 73 | for (elm = slist_first(head); elm; elm = elm->next) { |
| 74 | block_rec = container_of(elm, struct block_rec, slist); |
| 75 | |
| 76 | if (!block_rec->compiling |
| 77 | && (!best || block_rec->requests > best->requests)) |
| 78 | best = block_rec; |
| 79 | } |
| 80 | |
| 81 | return best; |
| 82 | } |
| 83 | |
| 84 | static bool lightrec_cancel_block_rec(struct recompiler *rec, |
| 85 | struct block_rec *block_rec) |
| 86 | { |
| 87 | if (block_rec->compiling) { |
| 88 | /* Block is being recompiled - wait for |
| 89 | * completion */ |
| 90 | pthread_cond_wait(&rec->cond2, &rec->mutex); |
| 91 | |
| 92 | /* We can't guarantee the signal was for us. |
| 93 | * Since block_rec may have been removed while |
| 94 | * we were waiting on the condition, we cannot |
| 95 | * check block_rec->compiling again. The best |
| 96 | * thing is just to restart the function. */ |
| 97 | return false; |
| 98 | } |
| 99 | |
| 100 | /* Block is not yet being processed - remove it from the list */ |
| 101 | slist_remove(&rec->slist, &block_rec->slist); |
| 102 | lightrec_free(rec->state, MEM_FOR_LIGHTREC, |
| 103 | sizeof(*block_rec), block_rec); |
| 104 | |
| 105 | return true; |
| 106 | } |
| 107 | |
| 108 | static void lightrec_cancel_list(struct recompiler *rec) |
| 109 | { |
| 110 | struct block_rec *block_rec; |
| 111 | struct slist_elm *elm, *head = &rec->slist; |
| 112 | |
| 113 | for (elm = slist_first(head); elm; elm = slist_first(head)) { |
| 114 | block_rec = container_of(elm, struct block_rec, slist); |
| 115 | lightrec_cancel_block_rec(rec, block_rec); |
| 116 | } |
| 117 | } |
| 118 | |
| 119 | static void lightrec_flush_code_buffer(struct lightrec_state *state, void *d) |
| 120 | { |
| 121 | struct recompiler *rec = d; |
| 122 | |
| 123 | lightrec_remove_outdated_blocks(state->block_cache, NULL); |
| 124 | rec->must_flush = false; |
| 125 | } |
| 126 | |
| 127 | static void lightrec_compile_list(struct recompiler *rec, |
| 128 | struct recompiler_thd *thd) |
| 129 | { |
| 130 | struct block_rec *block_rec; |
| 131 | struct block *block; |
| 132 | int ret; |
| 133 | |
| 134 | while (!rec->pause && |
| 135 | !!(block_rec = lightrec_get_best_elm(&rec->slist))) { |
| 136 | block_rec->compiling = true; |
| 137 | block = block_rec->block; |
| 138 | |
| 139 | pthread_mutex_unlock(&rec->mutex); |
| 140 | |
| 141 | if (likely(!block_has_flag(block, BLOCK_IS_DEAD))) { |
| 142 | ret = lightrec_compile_block(thd->cstate, block); |
| 143 | if (ret == -ENOMEM) { |
| 144 | /* Code buffer is full. Request the reaper to |
| 145 | * flush it. */ |
| 146 | |
| 147 | pthread_mutex_lock(&rec->mutex); |
| 148 | block_rec->compiling = false; |
| 149 | pthread_cond_broadcast(&rec->cond2); |
| 150 | |
| 151 | if (!rec->must_flush) { |
| 152 | rec->must_flush = true; |
| 153 | lightrec_cancel_list(rec); |
| 154 | |
| 155 | lightrec_reaper_add(rec->state->reaper, |
| 156 | lightrec_flush_code_buffer, |
| 157 | rec); |
| 158 | } |
| 159 | return; |
| 160 | } |
| 161 | |
| 162 | if (ret) { |
| 163 | pr_err("Unable to compile block at "PC_FMT": %d\n", |
| 164 | block->pc, ret); |
| 165 | } |
| 166 | } |
| 167 | |
| 168 | pthread_mutex_lock(&rec->mutex); |
| 169 | |
| 170 | slist_remove(&rec->slist, &block_rec->slist); |
| 171 | lightrec_free(rec->state, MEM_FOR_LIGHTREC, |
| 172 | sizeof(*block_rec), block_rec); |
| 173 | pthread_cond_broadcast(&rec->cond2); |
| 174 | } |
| 175 | } |
| 176 | |
| 177 | static void * lightrec_recompiler_thd(void *d) |
| 178 | { |
| 179 | struct recompiler_thd *thd = d; |
| 180 | struct recompiler *rec = container_of(thd, struct recompiler, thds[thd->tid]); |
| 181 | |
| 182 | pthread_mutex_lock(&rec->mutex); |
| 183 | |
| 184 | while (!rec->stop) { |
| 185 | do { |
| 186 | pthread_cond_wait(&rec->cond, &rec->mutex); |
| 187 | |
| 188 | if (rec->stop) |
| 189 | goto out_unlock; |
| 190 | |
| 191 | } while (rec->pause || slist_empty(&rec->slist)); |
| 192 | |
| 193 | lightrec_compile_list(rec, thd); |
| 194 | } |
| 195 | |
| 196 | out_unlock: |
| 197 | pthread_mutex_unlock(&rec->mutex); |
| 198 | return NULL; |
| 199 | } |
| 200 | |
| 201 | struct recompiler *lightrec_recompiler_init(struct lightrec_state *state) |
| 202 | { |
| 203 | struct recompiler *rec; |
| 204 | unsigned int i, nb_recs, nb_cpus; |
| 205 | int ret; |
| 206 | |
| 207 | nb_cpus = get_processors_count(); |
| 208 | nb_recs = nb_cpus < 2 ? 1 : nb_cpus - 1; |
| 209 | |
| 210 | rec = lightrec_malloc(state, MEM_FOR_LIGHTREC, sizeof(*rec) |
| 211 | + nb_recs * sizeof(*rec->thds)); |
| 212 | if (!rec) { |
| 213 | pr_err("Cannot create recompiler: Out of memory\n"); |
| 214 | return NULL; |
| 215 | } |
| 216 | |
| 217 | for (i = 0; i < nb_recs; i++) { |
| 218 | rec->thds[i].tid = i; |
| 219 | rec->thds[i].cstate = NULL; |
| 220 | } |
| 221 | |
| 222 | for (i = 0; i < nb_recs; i++) { |
| 223 | rec->thds[i].cstate = lightrec_create_cstate(state); |
| 224 | if (!rec->thds[i].cstate) { |
| 225 | pr_err("Cannot create recompiler: Out of memory\n"); |
| 226 | goto err_free_cstates; |
| 227 | } |
| 228 | } |
| 229 | |
| 230 | rec->state = state; |
| 231 | rec->stop = false; |
| 232 | rec->pause = false; |
| 233 | rec->must_flush = false; |
| 234 | rec->nb_recs = nb_recs; |
| 235 | slist_init(&rec->slist); |
| 236 | |
| 237 | ret = pthread_cond_init(&rec->cond, NULL); |
| 238 | if (ret) { |
| 239 | pr_err("Cannot init cond variable: %d\n", ret); |
| 240 | goto err_free_cstates; |
| 241 | } |
| 242 | |
| 243 | ret = pthread_cond_init(&rec->cond2, NULL); |
| 244 | if (ret) { |
| 245 | pr_err("Cannot init cond variable: %d\n", ret); |
| 246 | goto err_cnd_destroy; |
| 247 | } |
| 248 | |
| 249 | ret = pthread_mutex_init(&rec->alloc_mutex, NULL); |
| 250 | if (ret) { |
| 251 | pr_err("Cannot init alloc mutex variable: %d\n", ret); |
| 252 | goto err_cnd2_destroy; |
| 253 | } |
| 254 | |
| 255 | ret = pthread_mutex_init(&rec->mutex, NULL); |
| 256 | if (ret) { |
| 257 | pr_err("Cannot init mutex variable: %d\n", ret); |
| 258 | goto err_alloc_mtx_destroy; |
| 259 | } |
| 260 | |
| 261 | for (i = 0; i < nb_recs; i++) { |
| 262 | ret = pthread_create(&rec->thds[i].thd, NULL, |
| 263 | lightrec_recompiler_thd, &rec->thds[i]); |
| 264 | if (ret) { |
| 265 | pr_err("Cannot create recompiler thread: %d\n", ret); |
| 266 | /* TODO: Handle cleanup properly */ |
| 267 | goto err_mtx_destroy; |
| 268 | } |
| 269 | } |
| 270 | |
| 271 | pr_info("Threaded recompiler started with %u workers.\n", nb_recs); |
| 272 | |
| 273 | return rec; |
| 274 | |
| 275 | err_mtx_destroy: |
| 276 | pthread_mutex_destroy(&rec->mutex); |
| 277 | err_alloc_mtx_destroy: |
| 278 | pthread_mutex_destroy(&rec->alloc_mutex); |
| 279 | err_cnd2_destroy: |
| 280 | pthread_cond_destroy(&rec->cond2); |
| 281 | err_cnd_destroy: |
| 282 | pthread_cond_destroy(&rec->cond); |
| 283 | err_free_cstates: |
| 284 | for (i = 0; i < nb_recs; i++) { |
| 285 | if (rec->thds[i].cstate) |
| 286 | lightrec_free_cstate(rec->thds[i].cstate); |
| 287 | } |
| 288 | lightrec_free(state, MEM_FOR_LIGHTREC, sizeof(*rec), rec); |
| 289 | return NULL; |
| 290 | } |
| 291 | |
| 292 | void lightrec_free_recompiler(struct recompiler *rec) |
| 293 | { |
| 294 | unsigned int i; |
| 295 | |
| 296 | rec->stop = true; |
| 297 | |
| 298 | /* Stop the thread */ |
| 299 | pthread_mutex_lock(&rec->mutex); |
| 300 | pthread_cond_broadcast(&rec->cond); |
| 301 | lightrec_cancel_list(rec); |
| 302 | pthread_mutex_unlock(&rec->mutex); |
| 303 | |
| 304 | for (i = 0; i < rec->nb_recs; i++) |
| 305 | pthread_join(rec->thds[i].thd, NULL); |
| 306 | |
| 307 | for (i = 0; i < rec->nb_recs; i++) |
| 308 | lightrec_free_cstate(rec->thds[i].cstate); |
| 309 | |
| 310 | pthread_mutex_destroy(&rec->mutex); |
| 311 | pthread_mutex_destroy(&rec->alloc_mutex); |
| 312 | pthread_cond_destroy(&rec->cond); |
| 313 | pthread_cond_destroy(&rec->cond2); |
| 314 | lightrec_free(rec->state, MEM_FOR_LIGHTREC, sizeof(*rec), rec); |
| 315 | } |
| 316 | |
| 317 | int lightrec_recompiler_add(struct recompiler *rec, struct block *block) |
| 318 | { |
| 319 | struct slist_elm *elm; |
| 320 | struct block_rec *block_rec; |
| 321 | u32 pc1, pc2; |
| 322 | int ret = 0; |
| 323 | |
| 324 | pthread_mutex_lock(&rec->mutex); |
| 325 | |
| 326 | /* If the recompiler must flush the code cache, we can't add the new |
| 327 | * job. It will be re-added next time the block's address is jumped to |
| 328 | * again. */ |
| 329 | if (rec->must_flush) |
| 330 | goto out_unlock; |
| 331 | |
| 332 | /* If the block is marked as dead, don't compile it, it will be removed |
| 333 | * as soon as it's safe. */ |
| 334 | if (block_has_flag(block, BLOCK_IS_DEAD)) |
| 335 | goto out_unlock; |
| 336 | |
| 337 | for (elm = slist_first(&rec->slist); elm; elm = elm->next) { |
| 338 | block_rec = container_of(elm, struct block_rec, slist); |
| 339 | |
| 340 | if (block_rec->block == block) { |
| 341 | /* The block to compile is already in the queue - |
| 342 | * increment its counter to increase its priority */ |
| 343 | block_rec->requests++; |
| 344 | goto out_unlock; |
| 345 | } |
| 346 | |
| 347 | pc1 = kunseg(block_rec->block->pc); |
| 348 | pc2 = kunseg(block->pc); |
| 349 | if (pc2 >= pc1 && pc2 < pc1 + block_rec->block->nb_ops * 4) { |
| 350 | /* The block we want to compile is already covered by |
| 351 | * another one in the queue - increment its counter to |
| 352 | * increase its priority */ |
| 353 | block_rec->requests++; |
| 354 | goto out_unlock; |
| 355 | } |
| 356 | } |
| 357 | |
| 358 | /* By the time this function was called, the block has been recompiled |
| 359 | * and ins't in the wait list anymore. Just return here. */ |
| 360 | if (block->function && !block_has_flag(block, BLOCK_SHOULD_RECOMPILE)) |
| 361 | goto out_unlock; |
| 362 | |
| 363 | block_rec = lightrec_malloc(rec->state, MEM_FOR_LIGHTREC, |
| 364 | sizeof(*block_rec)); |
| 365 | if (!block_rec) { |
| 366 | ret = -ENOMEM; |
| 367 | goto out_unlock; |
| 368 | } |
| 369 | |
| 370 | pr_debug("Adding block "PC_FMT" to recompiler\n", block->pc); |
| 371 | |
| 372 | block_rec->block = block; |
| 373 | block_rec->compiling = false; |
| 374 | block_rec->requests = 1; |
| 375 | |
| 376 | elm = &rec->slist; |
| 377 | |
| 378 | /* Push the new entry to the front of the queue */ |
| 379 | slist_append(elm, &block_rec->slist); |
| 380 | |
| 381 | /* Signal the thread */ |
| 382 | pthread_cond_signal(&rec->cond); |
| 383 | |
| 384 | out_unlock: |
| 385 | pthread_mutex_unlock(&rec->mutex); |
| 386 | |
| 387 | return ret; |
| 388 | } |
| 389 | |
| 390 | void lightrec_recompiler_remove(struct recompiler *rec, struct block *block) |
| 391 | { |
| 392 | struct block_rec *block_rec; |
| 393 | struct slist_elm *elm; |
| 394 | |
| 395 | pthread_mutex_lock(&rec->mutex); |
| 396 | |
| 397 | while (true) { |
| 398 | for (elm = slist_first(&rec->slist); elm; elm = elm->next) { |
| 399 | block_rec = container_of(elm, struct block_rec, slist); |
| 400 | |
| 401 | if (block_rec->block == block) { |
| 402 | if (lightrec_cancel_block_rec(rec, block_rec)) |
| 403 | goto out_unlock; |
| 404 | |
| 405 | break; |
| 406 | } |
| 407 | } |
| 408 | |
| 409 | if (!elm) |
| 410 | break; |
| 411 | } |
| 412 | |
| 413 | out_unlock: |
| 414 | pthread_mutex_unlock(&rec->mutex); |
| 415 | } |
| 416 | |
| 417 | void * lightrec_recompiler_run_first_pass(struct lightrec_state *state, |
| 418 | struct block *block, u32 *pc) |
| 419 | { |
| 420 | u8 old_flags; |
| 421 | |
| 422 | /* There's no point in running the first pass if the block will never |
| 423 | * be compiled. Let the main loop run the interpreter instead. */ |
| 424 | if (block_has_flag(block, BLOCK_NEVER_COMPILE)) |
| 425 | return NULL; |
| 426 | |
| 427 | /* The block is marked as dead, and will be removed the next time the |
| 428 | * reaper is run. In the meantime, the old function can still be |
| 429 | * executed. */ |
| 430 | if (block_has_flag(block, BLOCK_IS_DEAD)) |
| 431 | return block->function; |
| 432 | |
| 433 | /* If the block is already fully tagged, there is no point in running |
| 434 | * the first pass. Request a recompilation of the block, and maybe the |
| 435 | * interpreter will run the block in the meantime. */ |
| 436 | if (block_has_flag(block, BLOCK_FULLY_TAGGED)) |
| 437 | lightrec_recompiler_add(state->rec, block); |
| 438 | |
| 439 | if (likely(block->function)) { |
| 440 | if (block_has_flag(block, BLOCK_FULLY_TAGGED)) { |
| 441 | old_flags = block_set_flags(block, BLOCK_NO_OPCODE_LIST); |
| 442 | |
| 443 | if (!(old_flags & BLOCK_NO_OPCODE_LIST)) { |
| 444 | pr_debug("Block "PC_FMT" is fully tagged" |
| 445 | " - free opcode list\n", block->pc); |
| 446 | |
| 447 | /* The block was already compiled but the opcode list |
| 448 | * didn't get freed yet - do it now */ |
| 449 | lightrec_free_opcode_list(state, block->opcode_list); |
| 450 | } |
| 451 | } |
| 452 | |
| 453 | return block->function; |
| 454 | } |
| 455 | |
| 456 | /* Mark the opcode list as freed, so that the threaded compiler won't |
| 457 | * free it while we're using it in the interpreter. */ |
| 458 | old_flags = block_set_flags(block, BLOCK_NO_OPCODE_LIST); |
| 459 | |
| 460 | /* Block wasn't compiled yet - run the interpreter */ |
| 461 | *pc = lightrec_emulate_block(state, block, *pc); |
| 462 | |
| 463 | if (!(old_flags & BLOCK_NO_OPCODE_LIST)) |
| 464 | block_clear_flags(block, BLOCK_NO_OPCODE_LIST); |
| 465 | |
| 466 | /* The block got compiled while the interpreter was running. |
| 467 | * We can free the opcode list now. */ |
| 468 | if (block->function && block_has_flag(block, BLOCK_FULLY_TAGGED)) { |
| 469 | old_flags = block_set_flags(block, BLOCK_NO_OPCODE_LIST); |
| 470 | |
| 471 | if (!(old_flags & BLOCK_NO_OPCODE_LIST)) { |
| 472 | pr_debug("Block "PC_FMT" is fully tagged" |
| 473 | " - free opcode list\n", block->pc); |
| 474 | |
| 475 | lightrec_free_opcode_list(state, block->opcode_list); |
| 476 | } |
| 477 | } |
| 478 | |
| 479 | return NULL; |
| 480 | } |
| 481 | |
| 482 | void lightrec_code_alloc_lock(struct lightrec_state *state) |
| 483 | { |
| 484 | pthread_mutex_lock(&state->rec->alloc_mutex); |
| 485 | } |
| 486 | |
| 487 | void lightrec_code_alloc_unlock(struct lightrec_state *state) |
| 488 | { |
| 489 | pthread_mutex_unlock(&state->rec->alloc_mutex); |
| 490 | } |
| 491 | |
| 492 | void lightrec_recompiler_pause(struct recompiler *rec) |
| 493 | { |
| 494 | rec->pause = true; |
| 495 | |
| 496 | pthread_mutex_lock(&rec->mutex); |
| 497 | pthread_cond_broadcast(&rec->cond); |
| 498 | lightrec_cancel_list(rec); |
| 499 | pthread_mutex_unlock(&rec->mutex); |
| 500 | } |
| 501 | |
| 502 | void lightrec_recompiler_unpause(struct recompiler *rec) |
| 503 | { |
| 504 | rec->pause = false; |
| 505 | } |