| 1 | // SPDX-License-Identifier: LGPL-2.1-or-later |
| 2 | /* |
| 3 | * Copyright (C) 2019-2021 Paul Cercueil <paul@crapouillou.net> |
| 4 | */ |
| 5 | |
| 6 | #include "debug.h" |
| 7 | #include "interpreter.h" |
| 8 | #include "lightrec-private.h" |
| 9 | #include "memmanager.h" |
| 10 | #include "slist.h" |
| 11 | |
| 12 | #include <errno.h> |
| 13 | #include <stdatomic.h> |
| 14 | #include <stdbool.h> |
| 15 | #include <stdlib.h> |
| 16 | #include <pthread.h> |
| 17 | #ifdef __linux__ |
| 18 | #include <unistd.h> |
| 19 | #endif |
| 20 | |
| 21 | struct block_rec { |
| 22 | struct block *block; |
| 23 | struct slist_elm slist; |
| 24 | bool compiling; |
| 25 | }; |
| 26 | |
| 27 | struct recompiler_thd { |
| 28 | struct lightrec_cstate *cstate; |
| 29 | unsigned int tid; |
| 30 | pthread_t thd; |
| 31 | }; |
| 32 | |
| 33 | struct recompiler { |
| 34 | struct lightrec_state *state; |
| 35 | pthread_cond_t cond; |
| 36 | pthread_cond_t cond2; |
| 37 | pthread_mutex_t mutex; |
| 38 | bool stop; |
| 39 | struct slist_elm slist; |
| 40 | |
| 41 | unsigned int nb_recs; |
| 42 | struct recompiler_thd thds[]; |
| 43 | }; |
| 44 | |
| 45 | static unsigned int get_processors_count(void) |
| 46 | { |
| 47 | unsigned int nb; |
| 48 | |
| 49 | #if defined(PTW32_VERSION) |
| 50 | nb = pthread_num_processors_np(); |
| 51 | #elif defined(__APPLE__) || defined(__FreeBSD__) |
| 52 | int count; |
| 53 | size_t size = sizeof(count); |
| 54 | |
| 55 | nb = sysctlbyname("hw.ncpu", &count, &size, NULL, 0) ? 1 : count; |
| 56 | #elif defined(__linux__) |
| 57 | nb = sysconf(_SC_NPROCESSORS_ONLN); |
| 58 | #endif |
| 59 | |
| 60 | return nb < 1 ? 1 : nb; |
| 61 | } |
| 62 | |
| 63 | static struct slist_elm * lightrec_get_first_elm(struct slist_elm *head) |
| 64 | { |
| 65 | struct block_rec *block_rec; |
| 66 | struct slist_elm *elm; |
| 67 | |
| 68 | for (elm = slist_first(head); elm; elm = elm->next) { |
| 69 | block_rec = container_of(elm, struct block_rec, slist); |
| 70 | |
| 71 | if (!block_rec->compiling) |
| 72 | return elm; |
| 73 | } |
| 74 | |
| 75 | return NULL; |
| 76 | } |
| 77 | |
| 78 | static void lightrec_compile_list(struct recompiler *rec, |
| 79 | struct recompiler_thd *thd) |
| 80 | { |
| 81 | struct block_rec *block_rec; |
| 82 | struct slist_elm *next; |
| 83 | struct block *block; |
| 84 | int ret; |
| 85 | |
| 86 | while (!!(next = lightrec_get_first_elm(&rec->slist))) { |
| 87 | block_rec = container_of(next, struct block_rec, slist); |
| 88 | block_rec->compiling = true; |
| 89 | block = block_rec->block; |
| 90 | |
| 91 | pthread_mutex_unlock(&rec->mutex); |
| 92 | |
| 93 | if (likely(!(block->flags & BLOCK_IS_DEAD))) { |
| 94 | ret = lightrec_compile_block(thd->cstate, block); |
| 95 | if (ret) { |
| 96 | pr_err("Unable to compile block at PC 0x%x: %d\n", |
| 97 | block->pc, ret); |
| 98 | } |
| 99 | } |
| 100 | |
| 101 | pthread_mutex_lock(&rec->mutex); |
| 102 | |
| 103 | slist_remove(&rec->slist, next); |
| 104 | lightrec_free(rec->state, MEM_FOR_LIGHTREC, |
| 105 | sizeof(*block_rec), block_rec); |
| 106 | pthread_cond_signal(&rec->cond2); |
| 107 | } |
| 108 | } |
| 109 | |
| 110 | static void * lightrec_recompiler_thd(void *d) |
| 111 | { |
| 112 | struct recompiler_thd *thd = d; |
| 113 | struct recompiler *rec = container_of(thd, struct recompiler, thds[thd->tid]); |
| 114 | |
| 115 | pthread_mutex_lock(&rec->mutex); |
| 116 | |
| 117 | while (!rec->stop) { |
| 118 | do { |
| 119 | pthread_cond_wait(&rec->cond, &rec->mutex); |
| 120 | |
| 121 | if (rec->stop) |
| 122 | goto out_unlock; |
| 123 | |
| 124 | } while (slist_empty(&rec->slist)); |
| 125 | |
| 126 | lightrec_compile_list(rec, thd); |
| 127 | } |
| 128 | |
| 129 | out_unlock: |
| 130 | pthread_mutex_unlock(&rec->mutex); |
| 131 | return NULL; |
| 132 | } |
| 133 | |
| 134 | struct recompiler *lightrec_recompiler_init(struct lightrec_state *state) |
| 135 | { |
| 136 | struct recompiler *rec; |
| 137 | unsigned int i, nb_recs, nb_cpus; |
| 138 | int ret; |
| 139 | |
| 140 | nb_cpus = get_processors_count(); |
| 141 | nb_recs = nb_cpus < 2 ? 1 : nb_cpus - 1; |
| 142 | |
| 143 | rec = lightrec_malloc(state, MEM_FOR_LIGHTREC, sizeof(*rec) |
| 144 | + nb_recs * sizeof(*rec->thds)); |
| 145 | if (!rec) { |
| 146 | pr_err("Cannot create recompiler: Out of memory\n"); |
| 147 | return NULL; |
| 148 | } |
| 149 | |
| 150 | for (i = 0; i < nb_recs; i++) { |
| 151 | rec->thds[i].tid = i; |
| 152 | rec->thds[i].cstate = NULL; |
| 153 | } |
| 154 | |
| 155 | for (i = 0; i < nb_recs; i++) { |
| 156 | rec->thds[i].cstate = lightrec_create_cstate(state); |
| 157 | if (!rec->state) { |
| 158 | pr_err("Cannot create recompiler: Out of memory\n"); |
| 159 | goto err_free_cstates; |
| 160 | } |
| 161 | } |
| 162 | |
| 163 | rec->state = state; |
| 164 | rec->stop = false; |
| 165 | rec->nb_recs = nb_recs; |
| 166 | slist_init(&rec->slist); |
| 167 | |
| 168 | ret = pthread_cond_init(&rec->cond, NULL); |
| 169 | if (ret) { |
| 170 | pr_err("Cannot init cond variable: %d\n", ret); |
| 171 | goto err_free_cstates; |
| 172 | } |
| 173 | |
| 174 | ret = pthread_cond_init(&rec->cond2, NULL); |
| 175 | if (ret) { |
| 176 | pr_err("Cannot init cond variable: %d\n", ret); |
| 177 | goto err_cnd_destroy; |
| 178 | } |
| 179 | |
| 180 | ret = pthread_mutex_init(&rec->mutex, NULL); |
| 181 | if (ret) { |
| 182 | pr_err("Cannot init mutex variable: %d\n", ret); |
| 183 | goto err_cnd2_destroy; |
| 184 | } |
| 185 | |
| 186 | for (i = 0; i < nb_recs; i++) { |
| 187 | ret = pthread_create(&rec->thds[i].thd, NULL, |
| 188 | lightrec_recompiler_thd, &rec->thds[i]); |
| 189 | if (ret) { |
| 190 | pr_err("Cannot create recompiler thread: %d\n", ret); |
| 191 | /* TODO: Handle cleanup properly */ |
| 192 | goto err_mtx_destroy; |
| 193 | } |
| 194 | } |
| 195 | |
| 196 | pr_info("Threaded recompiler started with %u workers.\n", nb_recs); |
| 197 | |
| 198 | return rec; |
| 199 | |
| 200 | err_mtx_destroy: |
| 201 | pthread_mutex_destroy(&rec->mutex); |
| 202 | err_cnd2_destroy: |
| 203 | pthread_cond_destroy(&rec->cond2); |
| 204 | err_cnd_destroy: |
| 205 | pthread_cond_destroy(&rec->cond); |
| 206 | err_free_cstates: |
| 207 | for (i = 0; i < nb_recs; i++) { |
| 208 | if (rec->thds[i].cstate) |
| 209 | lightrec_free_cstate(rec->thds[i].cstate); |
| 210 | } |
| 211 | lightrec_free(state, MEM_FOR_LIGHTREC, sizeof(*rec), rec); |
| 212 | return NULL; |
| 213 | } |
| 214 | |
| 215 | void lightrec_free_recompiler(struct recompiler *rec) |
| 216 | { |
| 217 | unsigned int i; |
| 218 | |
| 219 | rec->stop = true; |
| 220 | |
| 221 | /* Stop the thread */ |
| 222 | pthread_mutex_lock(&rec->mutex); |
| 223 | pthread_cond_broadcast(&rec->cond); |
| 224 | pthread_mutex_unlock(&rec->mutex); |
| 225 | |
| 226 | for (i = 0; i < rec->nb_recs; i++) |
| 227 | pthread_join(rec->thds[i].thd, NULL); |
| 228 | |
| 229 | for (i = 0; i < rec->nb_recs; i++) |
| 230 | lightrec_free_cstate(rec->thds[i].cstate); |
| 231 | |
| 232 | pthread_mutex_destroy(&rec->mutex); |
| 233 | pthread_cond_destroy(&rec->cond); |
| 234 | pthread_cond_destroy(&rec->cond2); |
| 235 | lightrec_free(rec->state, MEM_FOR_LIGHTREC, sizeof(*rec), rec); |
| 236 | } |
| 237 | |
| 238 | int lightrec_recompiler_add(struct recompiler *rec, struct block *block) |
| 239 | { |
| 240 | struct slist_elm *elm, *prev; |
| 241 | struct block_rec *block_rec; |
| 242 | int ret = 0; |
| 243 | |
| 244 | pthread_mutex_lock(&rec->mutex); |
| 245 | |
| 246 | /* If the block is marked as dead, don't compile it, it will be removed |
| 247 | * as soon as it's safe. */ |
| 248 | if (block->flags & BLOCK_IS_DEAD) |
| 249 | goto out_unlock; |
| 250 | |
| 251 | for (elm = slist_first(&rec->slist), prev = NULL; elm; |
| 252 | prev = elm, elm = elm->next) { |
| 253 | block_rec = container_of(elm, struct block_rec, slist); |
| 254 | |
| 255 | if (block_rec->block == block) { |
| 256 | /* The block to compile is already in the queue - bump |
| 257 | * it to the top of the list, unless the block is being |
| 258 | * recompiled. */ |
| 259 | if (prev && !block_rec->compiling && |
| 260 | !(block->flags & BLOCK_SHOULD_RECOMPILE)) { |
| 261 | slist_remove_next(prev); |
| 262 | slist_append(&rec->slist, elm); |
| 263 | } |
| 264 | |
| 265 | goto out_unlock; |
| 266 | } |
| 267 | } |
| 268 | |
| 269 | /* By the time this function was called, the block has been recompiled |
| 270 | * and ins't in the wait list anymore. Just return here. */ |
| 271 | if (block->function && !(block->flags & BLOCK_SHOULD_RECOMPILE)) |
| 272 | goto out_unlock; |
| 273 | |
| 274 | block_rec = lightrec_malloc(rec->state, MEM_FOR_LIGHTREC, |
| 275 | sizeof(*block_rec)); |
| 276 | if (!block_rec) { |
| 277 | ret = -ENOMEM; |
| 278 | goto out_unlock; |
| 279 | } |
| 280 | |
| 281 | pr_debug("Adding block PC 0x%x to recompiler\n", block->pc); |
| 282 | |
| 283 | block_rec->block = block; |
| 284 | block_rec->compiling = false; |
| 285 | |
| 286 | elm = &rec->slist; |
| 287 | |
| 288 | /* If the block is being recompiled, push it to the end of the queue; |
| 289 | * otherwise push it to the front of the queue. */ |
| 290 | if (block->flags & BLOCK_SHOULD_RECOMPILE) |
| 291 | for (; elm->next; elm = elm->next); |
| 292 | |
| 293 | slist_append(elm, &block_rec->slist); |
| 294 | |
| 295 | /* Signal the thread */ |
| 296 | pthread_cond_signal(&rec->cond); |
| 297 | |
| 298 | out_unlock: |
| 299 | pthread_mutex_unlock(&rec->mutex); |
| 300 | |
| 301 | return ret; |
| 302 | } |
| 303 | |
| 304 | void lightrec_recompiler_remove(struct recompiler *rec, struct block *block) |
| 305 | { |
| 306 | struct block_rec *block_rec; |
| 307 | struct slist_elm *elm; |
| 308 | |
| 309 | pthread_mutex_lock(&rec->mutex); |
| 310 | |
| 311 | while (true) { |
| 312 | for (elm = slist_first(&rec->slist); elm; elm = elm->next) { |
| 313 | block_rec = container_of(elm, struct block_rec, slist); |
| 314 | |
| 315 | if (block_rec->block != block) |
| 316 | continue; |
| 317 | |
| 318 | if (block_rec->compiling) { |
| 319 | /* Block is being recompiled - wait for |
| 320 | * completion */ |
| 321 | pthread_cond_wait(&rec->cond2, &rec->mutex); |
| 322 | |
| 323 | /* We can't guarantee the signal was for us. |
| 324 | * Since block_rec may have been removed while |
| 325 | * we were waiting on the condition, we cannot |
| 326 | * check block_rec->compiling again. The best |
| 327 | * thing is just to restart the function. */ |
| 328 | break; |
| 329 | } else { |
| 330 | /* Block is not yet being processed - remove it |
| 331 | * from the list */ |
| 332 | slist_remove(&rec->slist, elm); |
| 333 | lightrec_free(rec->state, MEM_FOR_LIGHTREC, |
| 334 | sizeof(*block_rec), block_rec); |
| 335 | |
| 336 | goto out_unlock; |
| 337 | } |
| 338 | } |
| 339 | |
| 340 | if (!elm) |
| 341 | break; |
| 342 | } |
| 343 | |
| 344 | out_unlock: |
| 345 | pthread_mutex_unlock(&rec->mutex); |
| 346 | } |
| 347 | |
| 348 | void * lightrec_recompiler_run_first_pass(struct lightrec_state *state, |
| 349 | struct block *block, u32 *pc) |
| 350 | { |
| 351 | bool freed; |
| 352 | |
| 353 | /* There's no point in running the first pass if the block will never |
| 354 | * be compiled. Let the main loop run the interpreter instead. */ |
| 355 | if (block->flags & BLOCK_NEVER_COMPILE) |
| 356 | return NULL; |
| 357 | |
| 358 | /* If the block is already fully tagged, there is no point in running |
| 359 | * the first pass. Request a recompilation of the block, and maybe the |
| 360 | * interpreter will run the block in the meantime. */ |
| 361 | if (block->flags & BLOCK_FULLY_TAGGED) |
| 362 | lightrec_recompiler_add(state->rec, block); |
| 363 | |
| 364 | if (likely(block->function)) { |
| 365 | if (block->flags & BLOCK_FULLY_TAGGED) { |
| 366 | freed = atomic_flag_test_and_set(&block->op_list_freed); |
| 367 | |
| 368 | if (!freed) { |
| 369 | pr_debug("Block PC 0x%08x is fully tagged" |
| 370 | " - free opcode list\n", block->pc); |
| 371 | |
| 372 | /* The block was already compiled but the opcode list |
| 373 | * didn't get freed yet - do it now */ |
| 374 | lightrec_free_opcode_list(state, block); |
| 375 | block->opcode_list = NULL; |
| 376 | } |
| 377 | } |
| 378 | |
| 379 | return block->function; |
| 380 | } |
| 381 | |
| 382 | /* Mark the opcode list as freed, so that the threaded compiler won't |
| 383 | * free it while we're using it in the interpreter. */ |
| 384 | freed = atomic_flag_test_and_set(&block->op_list_freed); |
| 385 | |
| 386 | /* Block wasn't compiled yet - run the interpreter */ |
| 387 | *pc = lightrec_emulate_block(state, block, *pc); |
| 388 | |
| 389 | if (!freed) |
| 390 | atomic_flag_clear(&block->op_list_freed); |
| 391 | |
| 392 | /* The block got compiled while the interpreter was running. |
| 393 | * We can free the opcode list now. */ |
| 394 | if (block->function && (block->flags & BLOCK_FULLY_TAGGED) && |
| 395 | !atomic_flag_test_and_set(&block->op_list_freed)) { |
| 396 | pr_debug("Block PC 0x%08x is fully tagged" |
| 397 | " - free opcode list\n", block->pc); |
| 398 | |
| 399 | lightrec_free_opcode_list(state, block); |
| 400 | block->opcode_list = NULL; |
| 401 | } |
| 402 | |
| 403 | return NULL; |
| 404 | } |