| 1 | /* |
| 2 | * Memory region management for Tiny Code Generator for QEMU |
| 3 | * |
| 4 | * Copyright (c) 2008 Fabrice Bellard |
| 5 | * |
| 6 | * Permission is hereby granted, free of charge, to any person obtaining a copy |
| 7 | * of this software and associated documentation files (the "Software"), to deal |
| 8 | * in the Software without restriction, including without limitation the rights |
| 9 | * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell |
| 10 | * copies of the Software, and to permit persons to whom the Software is |
| 11 | * furnished to do so, subject to the following conditions: |
| 12 | * |
| 13 | * The above copyright notice and this permission notice shall be included in |
| 14 | * all copies or substantial portions of the Software. |
| 15 | * |
| 16 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 17 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 18 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL |
| 19 | * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 20 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 21 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN |
| 22 | * THE SOFTWARE. |
| 23 | */ |
| 24 | |
| 25 | #include "qemu/osdep.h" |
| 26 | #include "qemu/units.h" |
| 27 | #include "qemu/madvise.h" |
| 28 | #include "qemu/mprotect.h" |
| 29 | #include "qemu/memalign.h" |
| 30 | #include "qemu/cacheinfo.h" |
| 31 | #include "qemu/qtree.h" |
| 32 | #include "qapi/error.h" |
| 33 | #include "tcg/tcg.h" |
| 34 | #include "exec/translation-block.h" |
| 35 | #include "tcg-internal.h" |
| 36 | #include "host/cpuinfo.h" |
| 37 | |
| 38 | |
| 39 | /* |
| 40 | * Local source-level compatibility with Unix. |
| 41 | * Used by tcg_region_init below. |
| 42 | */ |
| 43 | #if defined(_WIN32) |
| 44 | #define PROT_READ 1 |
| 45 | #define PROT_WRITE 2 |
| 46 | #define PROT_EXEC 4 |
| 47 | #endif |
| 48 | |
| 49 | struct tcg_region_tree { |
| 50 | QemuMutex lock; |
| 51 | QTree *tree; |
| 52 | /* padding to avoid false sharing is computed at run-time */ |
| 53 | }; |
| 54 | |
| 55 | /* |
| 56 | * We divide code_gen_buffer into equally-sized "regions" that TCG threads |
| 57 | * dynamically allocate from as demand dictates. Given appropriate region |
| 58 | * sizing, this minimizes flushes even when some TCG threads generate a lot |
| 59 | * more code than others. |
| 60 | */ |
| 61 | struct tcg_region_state { |
| 62 | QemuMutex lock; |
| 63 | |
| 64 | /* fields set at init time */ |
| 65 | void *start_aligned; |
| 66 | void *after_prologue; |
| 67 | size_t n; |
| 68 | size_t size; /* size of one region */ |
| 69 | size_t stride; /* .size + guard size */ |
| 70 | size_t total_size; /* size of entire buffer, >= n * stride */ |
| 71 | |
| 72 | /* fields protected by the lock */ |
| 73 | size_t current; /* current region index */ |
| 74 | size_t agg_size_full; /* aggregate size of full regions */ |
| 75 | }; |
| 76 | |
| 77 | static struct tcg_region_state region; |
| 78 | |
| 79 | /* |
| 80 | * This is an array of struct tcg_region_tree's, with padding. |
| 81 | * We use void * to simplify the computation of region_trees[i]; each |
| 82 | * struct is found every tree_size bytes. |
| 83 | */ |
| 84 | static void *region_trees; |
| 85 | static size_t tree_size; |
| 86 | |
| 87 | bool in_code_gen_buffer(const void *p) |
| 88 | { |
| 89 | /* |
| 90 | * Much like it is valid to have a pointer to the byte past the |
| 91 | * end of an array (so long as you don't dereference it), allow |
| 92 | * a pointer to the byte past the end of the code gen buffer. |
| 93 | */ |
| 94 | return (size_t)(p - region.start_aligned) <= region.total_size; |
| 95 | } |
| 96 | |
| 97 | #ifndef CONFIG_TCG_INTERPRETER |
| 98 | static int host_prot_read_exec(void) |
| 99 | { |
| 100 | #if defined(CONFIG_LINUX) && defined(HOST_AARCH64) && defined(PROT_BTI) |
| 101 | if (cpuinfo & CPUINFO_BTI) { |
| 102 | return PROT_READ | PROT_EXEC | PROT_BTI; |
| 103 | } |
| 104 | #endif |
| 105 | return PROT_READ | PROT_EXEC; |
| 106 | } |
| 107 | #endif |
| 108 | |
| 109 | #ifdef CONFIG_DEBUG_TCG |
| 110 | const void *tcg_splitwx_to_rx(void *rw) |
| 111 | { |
| 112 | /* Pass NULL pointers unchanged. */ |
| 113 | if (rw) { |
| 114 | g_assert(in_code_gen_buffer(rw)); |
| 115 | rw += tcg_splitwx_diff; |
| 116 | } |
| 117 | return rw; |
| 118 | } |
| 119 | |
| 120 | void *tcg_splitwx_to_rw(const void *rx) |
| 121 | { |
| 122 | /* Pass NULL pointers unchanged. */ |
| 123 | if (rx) { |
| 124 | rx -= tcg_splitwx_diff; |
| 125 | /* Assert that we end with a pointer in the rw region. */ |
| 126 | g_assert(in_code_gen_buffer(rx)); |
| 127 | } |
| 128 | return (void *)rx; |
| 129 | } |
| 130 | #endif /* CONFIG_DEBUG_TCG */ |
| 131 | |
| 132 | /* compare a pointer @ptr and a tb_tc @s */ |
| 133 | static int ptr_cmp_tb_tc(const void *ptr, const struct tb_tc *s) |
| 134 | { |
| 135 | if (ptr >= s->ptr + s->size) { |
| 136 | return 1; |
| 137 | } else if (ptr < s->ptr) { |
| 138 | return -1; |
| 139 | } |
| 140 | return 0; |
| 141 | } |
| 142 | |
| 143 | static gint tb_tc_cmp(gconstpointer ap, gconstpointer bp, gpointer userdata) |
| 144 | { |
| 145 | const struct tb_tc *a = ap; |
| 146 | const struct tb_tc *b = bp; |
| 147 | |
| 148 | /* |
| 149 | * When both sizes are set, we know this isn't a lookup. |
| 150 | * This is the most likely case: every TB must be inserted; lookups |
| 151 | * are a lot less frequent. |
| 152 | */ |
| 153 | if (likely(a->size && b->size)) { |
| 154 | if (a->ptr > b->ptr) { |
| 155 | return 1; |
| 156 | } else if (a->ptr < b->ptr) { |
| 157 | return -1; |
| 158 | } |
| 159 | /* a->ptr == b->ptr should happen only on deletions */ |
| 160 | g_assert(a->size == b->size); |
| 161 | return 0; |
| 162 | } |
| 163 | /* |
| 164 | * All lookups have either .size field set to 0. |
| 165 | * From the glib sources we see that @ap is always the lookup key. However |
| 166 | * the docs provide no guarantee, so we just mark this case as likely. |
| 167 | */ |
| 168 | if (likely(a->size == 0)) { |
| 169 | return ptr_cmp_tb_tc(a->ptr, b); |
| 170 | } |
| 171 | return ptr_cmp_tb_tc(b->ptr, a); |
| 172 | } |
| 173 | |
| 174 | static void tb_destroy(gpointer value) |
| 175 | { |
| 176 | TranslationBlock *tb = value; |
| 177 | qemu_spin_destroy(&tb->jmp_lock); |
| 178 | } |
| 179 | |
| 180 | static void tcg_region_trees_init(void) |
| 181 | { |
| 182 | size_t i; |
| 183 | |
| 184 | tree_size = ROUND_UP(sizeof(struct tcg_region_tree), qemu_dcache_linesize); |
| 185 | region_trees = qemu_memalign(qemu_dcache_linesize, region.n * tree_size); |
| 186 | for (i = 0; i < region.n; i++) { |
| 187 | struct tcg_region_tree *rt = region_trees + i * tree_size; |
| 188 | |
| 189 | qemu_mutex_init(&rt->lock); |
| 190 | rt->tree = q_tree_new_full(tb_tc_cmp, NULL, NULL, tb_destroy); |
| 191 | } |
| 192 | } |
| 193 | |
| 194 | static struct tcg_region_tree *tc_ptr_to_region_tree(const void *p) |
| 195 | { |
| 196 | size_t region_idx; |
| 197 | |
| 198 | /* |
| 199 | * Like tcg_splitwx_to_rw, with no assert. The pc may come from |
| 200 | * a signal handler over which the caller has no control. |
| 201 | */ |
| 202 | if (!in_code_gen_buffer(p)) { |
| 203 | p -= tcg_splitwx_diff; |
| 204 | if (!in_code_gen_buffer(p)) { |
| 205 | return NULL; |
| 206 | } |
| 207 | } |
| 208 | |
| 209 | if (p < region.start_aligned) { |
| 210 | region_idx = 0; |
| 211 | } else { |
| 212 | ptrdiff_t offset = p - region.start_aligned; |
| 213 | |
| 214 | if (offset > region.stride * (region.n - 1)) { |
| 215 | region_idx = region.n - 1; |
| 216 | } else { |
| 217 | region_idx = offset / region.stride; |
| 218 | } |
| 219 | } |
| 220 | return region_trees + region_idx * tree_size; |
| 221 | } |
| 222 | |
| 223 | void tcg_tb_insert(TranslationBlock *tb) |
| 224 | { |
| 225 | struct tcg_region_tree *rt = tc_ptr_to_region_tree(tb->tc.ptr); |
| 226 | |
| 227 | g_assert(rt != NULL); |
| 228 | qemu_mutex_lock(&rt->lock); |
| 229 | q_tree_insert(rt->tree, &tb->tc, tb); |
| 230 | qemu_mutex_unlock(&rt->lock); |
| 231 | } |
| 232 | |
| 233 | void tcg_tb_remove(TranslationBlock *tb) |
| 234 | { |
| 235 | struct tcg_region_tree *rt = tc_ptr_to_region_tree(tb->tc.ptr); |
| 236 | |
| 237 | g_assert(rt != NULL); |
| 238 | qemu_mutex_lock(&rt->lock); |
| 239 | q_tree_remove(rt->tree, &tb->tc); |
| 240 | qemu_mutex_unlock(&rt->lock); |
| 241 | } |
| 242 | |
| 243 | /* |
| 244 | * Find the TB 'tb' such that |
| 245 | * tb->tc.ptr <= tc_ptr < tb->tc.ptr + tb->tc.size |
| 246 | * Return NULL if not found. |
| 247 | */ |
| 248 | TranslationBlock *tcg_tb_lookup(uintptr_t tc_ptr) |
| 249 | { |
| 250 | struct tcg_region_tree *rt = tc_ptr_to_region_tree((void *)tc_ptr); |
| 251 | TranslationBlock *tb; |
| 252 | struct tb_tc s = { .ptr = (void *)tc_ptr }; |
| 253 | |
| 254 | if (rt == NULL) { |
| 255 | return NULL; |
| 256 | } |
| 257 | |
| 258 | qemu_mutex_lock(&rt->lock); |
| 259 | tb = q_tree_lookup(rt->tree, &s); |
| 260 | qemu_mutex_unlock(&rt->lock); |
| 261 | return tb; |
| 262 | } |
| 263 | |
| 264 | static void tcg_region_tree_lock_all(void) |
| 265 | { |
| 266 | size_t i; |
| 267 | |
| 268 | for (i = 0; i < region.n; i++) { |
| 269 | struct tcg_region_tree *rt = region_trees + i * tree_size; |
| 270 | |
| 271 | qemu_mutex_lock(&rt->lock); |
| 272 | } |
| 273 | } |
| 274 | |
| 275 | static void tcg_region_tree_unlock_all(void) |
| 276 | { |
| 277 | size_t i; |
| 278 | |
| 279 | for (i = 0; i < region.n; i++) { |
| 280 | struct tcg_region_tree *rt = region_trees + i * tree_size; |
| 281 | |
| 282 | qemu_mutex_unlock(&rt->lock); |
| 283 | } |
| 284 | } |
| 285 | |
| 286 | void tcg_tb_foreach(GTraverseFunc func, gpointer user_data) |
| 287 | { |
| 288 | size_t i; |
| 289 | |
| 290 | tcg_region_tree_lock_all(); |
| 291 | for (i = 0; i < region.n; i++) { |
| 292 | struct tcg_region_tree *rt = region_trees + i * tree_size; |
| 293 | |
| 294 | q_tree_foreach(rt->tree, func, user_data); |
| 295 | } |
| 296 | tcg_region_tree_unlock_all(); |
| 297 | } |
| 298 | |
| 299 | size_t tcg_nb_tbs(void) |
| 300 | { |
| 301 | size_t nb_tbs = 0; |
| 302 | size_t i; |
| 303 | |
| 304 | tcg_region_tree_lock_all(); |
| 305 | for (i = 0; i < region.n; i++) { |
| 306 | struct tcg_region_tree *rt = region_trees + i * tree_size; |
| 307 | |
| 308 | nb_tbs += q_tree_nnodes(rt->tree); |
| 309 | } |
| 310 | tcg_region_tree_unlock_all(); |
| 311 | return nb_tbs; |
| 312 | } |
| 313 | |
| 314 | static void tcg_region_tree_reset_all(void) |
| 315 | { |
| 316 | size_t i; |
| 317 | |
| 318 | tcg_region_tree_lock_all(); |
| 319 | for (i = 0; i < region.n; i++) { |
| 320 | struct tcg_region_tree *rt = region_trees + i * tree_size; |
| 321 | |
| 322 | /* Increment the refcount first so that destroy acts as a reset */ |
| 323 | q_tree_ref(rt->tree); |
| 324 | q_tree_destroy(rt->tree); |
| 325 | } |
| 326 | tcg_region_tree_unlock_all(); |
| 327 | } |
| 328 | |
| 329 | static void tcg_region_bounds(size_t curr_region, void **pstart, void **pend) |
| 330 | { |
| 331 | void *start, *end; |
| 332 | |
| 333 | start = region.start_aligned + curr_region * region.stride; |
| 334 | end = start + region.size; |
| 335 | |
| 336 | if (curr_region == 0) { |
| 337 | start = region.after_prologue; |
| 338 | } |
| 339 | /* The final region may have a few extra pages due to earlier rounding. */ |
| 340 | if (curr_region == region.n - 1) { |
| 341 | end = region.start_aligned + region.total_size; |
| 342 | } |
| 343 | |
| 344 | *pstart = start; |
| 345 | *pend = end; |
| 346 | } |
| 347 | |
| 348 | static void tcg_region_assign(TCGContext *s, size_t curr_region) |
| 349 | { |
| 350 | void *start, *end; |
| 351 | |
| 352 | tcg_region_bounds(curr_region, &start, &end); |
| 353 | |
| 354 | s->code_gen_buffer = start; |
| 355 | s->code_gen_ptr = start; |
| 356 | s->code_gen_buffer_size = end - start; |
| 357 | s->code_gen_highwater = end - TCG_HIGHWATER; |
| 358 | } |
| 359 | |
| 360 | static bool tcg_region_alloc__locked(TCGContext *s) |
| 361 | { |
| 362 | if (region.current == region.n) { |
| 363 | return false; |
| 364 | } |
| 365 | tcg_region_assign(s, region.current); |
| 366 | region.current++; |
| 367 | return true; |
| 368 | } |
| 369 | |
| 370 | /* |
| 371 | * Request a new region once the one in use has filled up. |
| 372 | * Returns true on success. |
| 373 | */ |
| 374 | bool tcg_region_alloc(TCGContext *s) |
| 375 | { |
| 376 | bool ok; |
| 377 | /* read the region size now; alloc__locked will overwrite it on success */ |
| 378 | size_t size_full = s->code_gen_buffer_size; |
| 379 | |
| 380 | qemu_mutex_lock(®ion.lock); |
| 381 | ok = tcg_region_alloc__locked(s); |
| 382 | if (ok) { |
| 383 | region.agg_size_full += size_full - TCG_HIGHWATER; |
| 384 | } |
| 385 | qemu_mutex_unlock(®ion.lock); |
| 386 | return ok; |
| 387 | } |
| 388 | |
| 389 | /* |
| 390 | * Perform a context's first region allocation. |
| 391 | * This function does _not_ increment region.agg_size_full. |
| 392 | */ |
| 393 | static void tcg_region_initial_alloc__locked(TCGContext *s) |
| 394 | { |
| 395 | bool ok = tcg_region_alloc__locked(s); |
| 396 | g_assert(ok); |
| 397 | } |
| 398 | |
| 399 | void tcg_region_thread_initial_alloc(TCGContext *s) |
| 400 | { |
| 401 | bool ok; |
| 402 | |
| 403 | qemu_mutex_lock(®ion.lock); |
| 404 | ok = tcg_region_alloc__locked(s); |
| 405 | qemu_mutex_unlock(®ion.lock); |
| 406 | |
| 407 | /* |
| 408 | * A vCPU hotplug may happen at any time. When the new thread is |
| 409 | * started, the region pool may be exhausted. At this point in |
| 410 | * the new thread call stack, we are not in a position to fix this. |
| 411 | * Leave code_gen_ptr NULL, so that this thread's first call to |
| 412 | * tcg_tb_alloc() returns NULL, so that the translator performs |
| 413 | * a tb_flush() and retry. |
| 414 | * |
| 415 | * During the tb_flush(), tcg_region_reset_all() will assign a |
| 416 | * new region to all contexts, including this one. |
| 417 | */ |
| 418 | if (!ok) { |
| 419 | s->code_gen_buffer = NULL; |
| 420 | s->code_gen_ptr = NULL; |
| 421 | s->code_gen_buffer_size = 0; |
| 422 | s->code_gen_highwater = NULL; |
| 423 | } |
| 424 | } |
| 425 | |
| 426 | /* Call from a safe-work context */ |
| 427 | void tcg_region_reset_all(void) |
| 428 | { |
| 429 | unsigned int n_ctxs = qatomic_read(&tcg_cur_ctxs); |
| 430 | unsigned int i; |
| 431 | |
| 432 | qemu_mutex_lock(®ion.lock); |
| 433 | region.current = 0; |
| 434 | region.agg_size_full = 0; |
| 435 | |
| 436 | for (i = 0; i < n_ctxs; i++) { |
| 437 | TCGContext *s = qatomic_read(&tcg_ctxs[i]); |
| 438 | tcg_region_initial_alloc__locked(s); |
| 439 | } |
| 440 | qemu_mutex_unlock(®ion.lock); |
| 441 | |
| 442 | tcg_region_tree_reset_all(); |
| 443 | } |
| 444 | |
| 445 | static size_t tcg_n_regions(size_t tb_size, unsigned max_threads) |
| 446 | { |
| 447 | #ifdef CONFIG_USER_ONLY |
| 448 | return 1; |
| 449 | #else |
| 450 | size_t n_regions; |
| 451 | |
| 452 | /* |
| 453 | * It is likely that some vCPUs will translate more code than others, |
| 454 | * so we first try to set more regions than threads, with those regions |
| 455 | * being of reasonable size. If that's not possible we make do by evenly |
| 456 | * dividing the code_gen_buffer among the vCPUs. |
| 457 | * |
| 458 | * Use a single region if all we have is one vCPU thread. |
| 459 | */ |
| 460 | if (max_threads == 1) { |
| 461 | return 1; |
| 462 | } |
| 463 | |
| 464 | /* |
| 465 | * Try to have more regions than threads, with each region being >= 2 MB. |
| 466 | * If we can't, then just allocate one region per vCPU thread. |
| 467 | */ |
| 468 | n_regions = tb_size / (2 * MiB); |
| 469 | if (n_regions <= max_threads) { |
| 470 | return max_threads; |
| 471 | } |
| 472 | return MIN(n_regions, max_threads * 8); |
| 473 | #endif |
| 474 | } |
| 475 | |
| 476 | /* |
| 477 | * Minimum size of the code gen buffer. This number is randomly chosen, |
| 478 | * but not so small that we can't have a fair number of TB's live. |
| 479 | * |
| 480 | * Maximum size, MAX_CODE_GEN_BUFFER_SIZE, is defined in tcg-target.h. |
| 481 | * Unless otherwise indicated, this is constrained by the range of |
| 482 | * direct branches on the host cpu, as used by the TCG implementation |
| 483 | * of goto_tb. |
| 484 | */ |
| 485 | #define MIN_CODE_GEN_BUFFER_SIZE (1 * MiB) |
| 486 | |
| 487 | #ifdef CONFIG_USER_ONLY |
| 488 | /* |
| 489 | * As user-mode emulation typically means running multiple instances |
| 490 | * of the translator don't go too nuts with our default code gen |
| 491 | * buffer lest we make things too hard for the OS. |
| 492 | */ |
| 493 | #define DEFAULT_CODE_GEN_BUFFER_SIZE_1 (128 * MiB) |
| 494 | #else |
| 495 | /* |
| 496 | * We expect most system emulation to run one or two guests per host. |
| 497 | * Users running large scale system emulation may want to tweak their |
| 498 | * runtime setup via the tb-size control on the command line. |
| 499 | */ |
| 500 | #define DEFAULT_CODE_GEN_BUFFER_SIZE_1 (1 * GiB) |
| 501 | #endif |
| 502 | |
| 503 | #define DEFAULT_CODE_GEN_BUFFER_SIZE \ |
| 504 | (DEFAULT_CODE_GEN_BUFFER_SIZE_1 < MAX_CODE_GEN_BUFFER_SIZE \ |
| 505 | ? DEFAULT_CODE_GEN_BUFFER_SIZE_1 : MAX_CODE_GEN_BUFFER_SIZE) |
| 506 | |
| 507 | #ifdef USE_STATIC_CODE_GEN_BUFFER |
| 508 | static uint8_t static_code_gen_buffer[DEFAULT_CODE_GEN_BUFFER_SIZE] |
| 509 | __attribute__((aligned(CODE_GEN_ALIGN))); |
| 510 | |
| 511 | static int alloc_code_gen_buffer(size_t tb_size, int splitwx, Error **errp) |
| 512 | { |
| 513 | void *buf, *end; |
| 514 | size_t size; |
| 515 | |
| 516 | if (splitwx > 0) { |
| 517 | error_setg(errp, "jit split-wx not supported"); |
| 518 | return -1; |
| 519 | } |
| 520 | |
| 521 | /* page-align the beginning and end of the buffer */ |
| 522 | buf = static_code_gen_buffer; |
| 523 | end = static_code_gen_buffer + sizeof(static_code_gen_buffer); |
| 524 | buf = QEMU_ALIGN_PTR_UP(buf, qemu_real_host_page_size()); |
| 525 | end = QEMU_ALIGN_PTR_DOWN(end, qemu_real_host_page_size()); |
| 526 | |
| 527 | size = end - buf; |
| 528 | |
| 529 | /* Honor a command-line option limiting the size of the buffer. */ |
| 530 | if (size > tb_size) { |
| 531 | size = QEMU_ALIGN_DOWN(tb_size, qemu_real_host_page_size()); |
| 532 | } |
| 533 | |
| 534 | region.start_aligned = buf; |
| 535 | region.total_size = size; |
| 536 | |
| 537 | return PROT_READ | PROT_WRITE; |
| 538 | } |
| 539 | #elif defined(_WIN32) |
| 540 | static int alloc_code_gen_buffer(size_t size, int splitwx, Error **errp) |
| 541 | { |
| 542 | void *buf; |
| 543 | |
| 544 | if (splitwx > 0) { |
| 545 | error_setg(errp, "jit split-wx not supported"); |
| 546 | return -1; |
| 547 | } |
| 548 | |
| 549 | buf = VirtualAlloc(NULL, size, MEM_RESERVE | MEM_COMMIT, |
| 550 | PAGE_EXECUTE_READWRITE); |
| 551 | if (buf == NULL) { |
| 552 | error_setg_win32(errp, GetLastError(), |
| 553 | "allocate %zu bytes for jit buffer", size); |
| 554 | return false; |
| 555 | } |
| 556 | |
| 557 | region.start_aligned = buf; |
| 558 | region.total_size = size; |
| 559 | |
| 560 | return PROT_READ | PROT_WRITE | PROT_EXEC; |
| 561 | } |
| 562 | #else |
| 563 | static int alloc_code_gen_buffer_anon(size_t size, int prot, |
| 564 | int flags, Error **errp) |
| 565 | { |
| 566 | void *buf; |
| 567 | |
| 568 | buf = mmap(NULL, size, prot, flags, -1, 0); |
| 569 | if (buf == MAP_FAILED) { |
| 570 | error_setg_errno(errp, errno, |
| 571 | "allocate %zu bytes for jit buffer", size); |
| 572 | return -1; |
| 573 | } |
| 574 | |
| 575 | region.start_aligned = buf; |
| 576 | region.total_size = size; |
| 577 | return prot; |
| 578 | } |
| 579 | |
| 580 | #ifndef CONFIG_TCG_INTERPRETER |
| 581 | #ifdef CONFIG_POSIX |
| 582 | #include "qemu/memfd.h" |
| 583 | |
| 584 | static int alloc_code_gen_buffer_splitwx_memfd(size_t size, Error **errp) |
| 585 | { |
| 586 | void *buf_rw = NULL, *buf_rx = MAP_FAILED; |
| 587 | int fd = -1; |
| 588 | |
| 589 | buf_rw = qemu_memfd_alloc("tcg-jit", size, 0, &fd, errp); |
| 590 | if (buf_rw == NULL) { |
| 591 | goto fail; |
| 592 | } |
| 593 | |
| 594 | buf_rx = mmap(NULL, size, host_prot_read_exec(), MAP_SHARED, fd, 0); |
| 595 | if (buf_rx == MAP_FAILED) { |
| 596 | error_setg_errno(errp, errno, |
| 597 | "failed to map shared memory for execute"); |
| 598 | goto fail; |
| 599 | } |
| 600 | |
| 601 | close(fd); |
| 602 | region.start_aligned = buf_rw; |
| 603 | region.total_size = size; |
| 604 | tcg_splitwx_diff = buf_rx - buf_rw; |
| 605 | |
| 606 | return PROT_READ | PROT_WRITE; |
| 607 | |
| 608 | fail: |
| 609 | /* buf_rx is always equal to MAP_FAILED here and does not require cleanup */ |
| 610 | if (buf_rw) { |
| 611 | munmap(buf_rw, size); |
| 612 | } |
| 613 | if (fd >= 0) { |
| 614 | close(fd); |
| 615 | } |
| 616 | return -1; |
| 617 | } |
| 618 | #endif /* CONFIG_POSIX */ |
| 619 | |
| 620 | #ifdef CONFIG_DARWIN |
| 621 | #include <mach/mach.h> |
| 622 | |
| 623 | extern kern_return_t mach_vm_remap(vm_map_t target_task, |
| 624 | mach_vm_address_t *target_address, |
| 625 | mach_vm_size_t size, |
| 626 | mach_vm_offset_t mask, |
| 627 | int flags, |
| 628 | vm_map_t src_task, |
| 629 | mach_vm_address_t src_address, |
| 630 | boolean_t copy, |
| 631 | vm_prot_t *cur_protection, |
| 632 | vm_prot_t *max_protection, |
| 633 | vm_inherit_t inheritance); |
| 634 | |
| 635 | static int alloc_code_gen_buffer_splitwx_vmremap(size_t size, Error **errp) |
| 636 | { |
| 637 | kern_return_t ret; |
| 638 | mach_vm_address_t buf_rw, buf_rx; |
| 639 | vm_prot_t cur_prot, max_prot; |
| 640 | |
| 641 | /* Map the read-write portion via normal anon memory. */ |
| 642 | if (!alloc_code_gen_buffer_anon(size, PROT_READ | PROT_WRITE, |
| 643 | MAP_PRIVATE | MAP_ANONYMOUS, errp)) { |
| 644 | return -1; |
| 645 | } |
| 646 | |
| 647 | buf_rw = (mach_vm_address_t)region.start_aligned; |
| 648 | buf_rx = 0; |
| 649 | ret = mach_vm_remap(mach_task_self(), |
| 650 | &buf_rx, |
| 651 | size, |
| 652 | 0, |
| 653 | VM_FLAGS_ANYWHERE, |
| 654 | mach_task_self(), |
| 655 | buf_rw, |
| 656 | false, |
| 657 | &cur_prot, |
| 658 | &max_prot, |
| 659 | VM_INHERIT_NONE); |
| 660 | if (ret != KERN_SUCCESS) { |
| 661 | /* TODO: Convert "ret" to a human readable error message. */ |
| 662 | error_setg(errp, "vm_remap for jit splitwx failed"); |
| 663 | munmap((void *)buf_rw, size); |
| 664 | return -1; |
| 665 | } |
| 666 | |
| 667 | if (mprotect((void *)buf_rx, size, host_prot_read_exec()) != 0) { |
| 668 | error_setg_errno(errp, errno, "mprotect for jit splitwx"); |
| 669 | munmap((void *)buf_rx, size); |
| 670 | munmap((void *)buf_rw, size); |
| 671 | return -1; |
| 672 | } |
| 673 | |
| 674 | tcg_splitwx_diff = buf_rx - buf_rw; |
| 675 | return PROT_READ | PROT_WRITE; |
| 676 | } |
| 677 | #endif /* CONFIG_DARWIN */ |
| 678 | #endif /* CONFIG_TCG_INTERPRETER */ |
| 679 | |
| 680 | static int alloc_code_gen_buffer_splitwx(size_t size, Error **errp) |
| 681 | { |
| 682 | #ifndef CONFIG_TCG_INTERPRETER |
| 683 | # ifdef CONFIG_DARWIN |
| 684 | return alloc_code_gen_buffer_splitwx_vmremap(size, errp); |
| 685 | # endif |
| 686 | # ifdef CONFIG_POSIX |
| 687 | return alloc_code_gen_buffer_splitwx_memfd(size, errp); |
| 688 | # endif |
| 689 | #endif |
| 690 | error_setg(errp, "jit split-wx not supported"); |
| 691 | return -1; |
| 692 | } |
| 693 | |
| 694 | static int alloc_code_gen_buffer(size_t size, int splitwx, Error **errp) |
| 695 | { |
| 696 | ERRP_GUARD(); |
| 697 | int prot, flags; |
| 698 | |
| 699 | if (splitwx) { |
| 700 | prot = alloc_code_gen_buffer_splitwx(size, errp); |
| 701 | if (prot >= 0) { |
| 702 | return prot; |
| 703 | } |
| 704 | /* |
| 705 | * If splitwx force-on (1), fail; |
| 706 | * if splitwx default-on (-1), fall through to splitwx off. |
| 707 | */ |
| 708 | if (splitwx > 0) { |
| 709 | return -1; |
| 710 | } |
| 711 | error_free_or_abort(errp); |
| 712 | } |
| 713 | |
| 714 | /* |
| 715 | * macOS 11.2 has a bug (Apple Feedback FB8994773) in which mprotect |
| 716 | * rejects a permission change from RWX -> NONE when reserving the |
| 717 | * guard pages later. We can go the other way with the same number |
| 718 | * of syscalls, so always begin with PROT_NONE. |
| 719 | */ |
| 720 | prot = PROT_NONE; |
| 721 | flags = MAP_PRIVATE | MAP_ANONYMOUS; |
| 722 | #ifdef CONFIG_DARWIN |
| 723 | /* Applicable to both iOS and macOS (Apple Silicon). */ |
| 724 | if (!splitwx) { |
| 725 | flags |= MAP_JIT; |
| 726 | } |
| 727 | #endif |
| 728 | |
| 729 | return alloc_code_gen_buffer_anon(size, prot, flags, errp); |
| 730 | } |
| 731 | #endif /* USE_STATIC_CODE_GEN_BUFFER, WIN32, POSIX */ |
| 732 | |
| 733 | /* |
| 734 | * Initializes region partitioning. |
| 735 | * |
| 736 | * Called at init time from the parent thread (i.e. the one calling |
| 737 | * tcg_context_init), after the target's TCG globals have been set. |
| 738 | * |
| 739 | * Region partitioning works by splitting code_gen_buffer into separate regions, |
| 740 | * and then assigning regions to TCG threads so that the threads can translate |
| 741 | * code in parallel without synchronization. |
| 742 | * |
| 743 | * In system-mode the number of TCG threads is bounded by max_threads, |
| 744 | * |
| 745 | * In user-mode we use a single region. Having multiple regions in user-mode |
| 746 | * is not supported, because the number of vCPU threads (recall that each thread |
| 747 | * spawned by the guest corresponds to a vCPU thread) is only bounded by the |
| 748 | * OS, and usually this number is huge (tens of thousands is not uncommon). |
| 749 | * Thus, given this large bound on the number of vCPU threads and the fact |
| 750 | * that code_gen_buffer is allocated at compile-time, we cannot guarantee |
| 751 | * that the availability of at least one region per vCPU thread. |
| 752 | * |
| 753 | * However, this user-mode limitation is unlikely to be a significant problem |
| 754 | * in practice. Multi-threaded guests share most if not all of their translated |
| 755 | * code, which makes parallel code generation less appealing than in system-mode |
| 756 | */ |
| 757 | void tcg_region_init(size_t tb_size, int splitwx, unsigned max_threads) |
| 758 | { |
| 759 | const size_t page_size = qemu_real_host_page_size(); |
| 760 | size_t region_size; |
| 761 | int have_prot, need_prot; |
| 762 | |
| 763 | /* Size the buffer. */ |
| 764 | if (tb_size == 0) { |
| 765 | size_t phys_mem = qemu_get_host_physmem(); |
| 766 | if (phys_mem == 0) { |
| 767 | tb_size = DEFAULT_CODE_GEN_BUFFER_SIZE; |
| 768 | } else { |
| 769 | tb_size = QEMU_ALIGN_DOWN(phys_mem / 8, page_size); |
| 770 | tb_size = MIN(DEFAULT_CODE_GEN_BUFFER_SIZE, tb_size); |
| 771 | } |
| 772 | } |
| 773 | if (tb_size < MIN_CODE_GEN_BUFFER_SIZE) { |
| 774 | tb_size = MIN_CODE_GEN_BUFFER_SIZE; |
| 775 | } |
| 776 | if (tb_size > MAX_CODE_GEN_BUFFER_SIZE) { |
| 777 | tb_size = MAX_CODE_GEN_BUFFER_SIZE; |
| 778 | } |
| 779 | |
| 780 | have_prot = alloc_code_gen_buffer(tb_size, splitwx, &error_fatal); |
| 781 | assert(have_prot >= 0); |
| 782 | |
| 783 | /* Request large pages for the buffer and the splitwx. */ |
| 784 | qemu_madvise(region.start_aligned, region.total_size, QEMU_MADV_HUGEPAGE); |
| 785 | if (tcg_splitwx_diff) { |
| 786 | qemu_madvise(region.start_aligned + tcg_splitwx_diff, |
| 787 | region.total_size, QEMU_MADV_HUGEPAGE); |
| 788 | } |
| 789 | |
| 790 | /* |
| 791 | * Make region_size a multiple of page_size, using aligned as the start. |
| 792 | * As a result of this we might end up with a few extra pages at the end of |
| 793 | * the buffer; we will assign those to the last region. |
| 794 | */ |
| 795 | region.n = tcg_n_regions(tb_size, max_threads); |
| 796 | region_size = tb_size / region.n; |
| 797 | region_size = QEMU_ALIGN_DOWN(region_size, page_size); |
| 798 | |
| 799 | /* A region must have at least 2 pages; one code, one guard */ |
| 800 | g_assert(region_size >= 2 * page_size); |
| 801 | region.stride = region_size; |
| 802 | |
| 803 | /* Reserve space for guard pages. */ |
| 804 | region.size = region_size - page_size; |
| 805 | region.total_size -= page_size; |
| 806 | |
| 807 | /* |
| 808 | * The first region will be smaller than the others, via the prologue, |
| 809 | * which has yet to be allocated. For now, the first region begins at |
| 810 | * the page boundary. |
| 811 | */ |
| 812 | region.after_prologue = region.start_aligned; |
| 813 | |
| 814 | /* init the region struct */ |
| 815 | qemu_mutex_init(®ion.lock); |
| 816 | |
| 817 | /* |
| 818 | * Set guard pages in the rw buffer, as that's the one into which |
| 819 | * buffer overruns could occur. Do not set guard pages in the rx |
| 820 | * buffer -- let that one use hugepages throughout. |
| 821 | * Work with the page protections set up with the initial mapping. |
| 822 | */ |
| 823 | need_prot = PROT_READ | PROT_WRITE; |
| 824 | #ifndef CONFIG_TCG_INTERPRETER |
| 825 | if (tcg_splitwx_diff == 0) { |
| 826 | need_prot |= host_prot_read_exec(); |
| 827 | } |
| 828 | #endif |
| 829 | for (size_t i = 0, n = region.n; i < n; i++) { |
| 830 | void *start, *end; |
| 831 | |
| 832 | tcg_region_bounds(i, &start, &end); |
| 833 | if (have_prot != need_prot) { |
| 834 | int rc; |
| 835 | |
| 836 | if (need_prot == (PROT_READ | PROT_WRITE | PROT_EXEC)) { |
| 837 | rc = qemu_mprotect_rwx(start, end - start); |
| 838 | } else if (need_prot == (PROT_READ | PROT_WRITE)) { |
| 839 | rc = qemu_mprotect_rw(start, end - start); |
| 840 | } else { |
| 841 | #ifdef CONFIG_POSIX |
| 842 | rc = mprotect(start, end - start, need_prot); |
| 843 | if (rc) { |
| 844 | error_report("mprotect of jit buffer: %s", |
| 845 | strerror(errno)); |
| 846 | } |
| 847 | #else |
| 848 | g_assert_not_reached(); |
| 849 | #endif |
| 850 | } |
| 851 | if (rc) { |
| 852 | exit(1); |
| 853 | } |
| 854 | } |
| 855 | if (have_prot != 0) { |
| 856 | /* Guard pages are nice for bug detection but are not essential. */ |
| 857 | (void)qemu_mprotect_none(end, page_size); |
| 858 | } |
| 859 | } |
| 860 | |
| 861 | tcg_region_trees_init(); |
| 862 | |
| 863 | /* |
| 864 | * Leave the initial context initialized to the first region. |
| 865 | * This will be the context into which we generate the prologue. |
| 866 | * It is also the only context for CONFIG_USER_ONLY. |
| 867 | */ |
| 868 | tcg_region_initial_alloc__locked(&tcg_init_ctx); |
| 869 | } |
| 870 | |
| 871 | void tcg_region_prologue_set(TCGContext *s) |
| 872 | { |
| 873 | /* Deduct the prologue from the first region. */ |
| 874 | g_assert(region.start_aligned == s->code_gen_buffer); |
| 875 | region.after_prologue = s->code_ptr; |
| 876 | |
| 877 | /* Recompute boundaries of the first region. */ |
| 878 | tcg_region_assign(s, 0); |
| 879 | |
| 880 | /* Register the balance of the buffer with gdb. */ |
| 881 | tcg_register_jit(tcg_splitwx_to_rx(region.after_prologue), |
| 882 | region.start_aligned + region.total_size - |
| 883 | region.after_prologue); |
| 884 | } |
| 885 | |
| 886 | /* |
| 887 | * Returns the size (in bytes) of all translated code (i.e. from all regions) |
| 888 | * currently in the cache. |
| 889 | * See also: tcg_code_capacity() |
| 890 | * Do not confuse with tcg_current_code_size(); that one applies to a single |
| 891 | * TCG context. |
| 892 | */ |
| 893 | size_t tcg_code_size(void) |
| 894 | { |
| 895 | unsigned int n_ctxs = qatomic_read(&tcg_cur_ctxs); |
| 896 | unsigned int i; |
| 897 | size_t total; |
| 898 | |
| 899 | qemu_mutex_lock(®ion.lock); |
| 900 | total = region.agg_size_full; |
| 901 | for (i = 0; i < n_ctxs; i++) { |
| 902 | const TCGContext *s = qatomic_read(&tcg_ctxs[i]); |
| 903 | size_t size; |
| 904 | |
| 905 | size = qatomic_read(&s->code_gen_ptr) - s->code_gen_buffer; |
| 906 | g_assert(size <= s->code_gen_buffer_size); |
| 907 | total += size; |
| 908 | } |
| 909 | qemu_mutex_unlock(®ion.lock); |
| 910 | return total; |
| 911 | } |
| 912 | |
| 913 | /* |
| 914 | * Returns the code capacity (in bytes) of the entire cache, i.e. including all |
| 915 | * regions. |
| 916 | * See also: tcg_code_size() |
| 917 | */ |
| 918 | size_t tcg_code_capacity(void) |
| 919 | { |
| 920 | size_t guard_size, capacity; |
| 921 | |
| 922 | /* no need for synchronization; these variables are set at init time */ |
| 923 | guard_size = region.stride - region.size; |
| 924 | capacity = region.total_size; |
| 925 | capacity -= (region.n - 1) * guard_size; |
| 926 | capacity -= region.n * TCG_HIGHWATER; |
| 927 | |
| 928 | return capacity; |
| 929 | } |