| 1 | /* |
| 2 | * QEMU System Emulator |
| 3 | * |
| 4 | * Copyright (c) 2003-2008 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/cutils.h" |
| 27 | #include "migration/vmstate.h" |
| 28 | #include "qapi/error.h" |
| 29 | #include "qemu/error-report.h" |
| 30 | #include "system/cpus.h" |
| 31 | #include "system/qtest.h" |
| 32 | #include "qemu/main-loop.h" |
| 33 | #include "qemu/option.h" |
| 34 | #include "qemu/seqlock.h" |
| 35 | #include "system/replay.h" |
| 36 | #include "system/runstate.h" |
| 37 | #include "hw/core/cpu.h" |
| 38 | #include "exec/icount.h" |
| 39 | #include "system/cpu-timers-internal.h" |
| 40 | |
| 41 | /* |
| 42 | * ICOUNT: Instruction Counter |
| 43 | * |
| 44 | * this module is split off from cpu-timers because the icount part |
| 45 | * is TCG-specific, and does not need to be built for other accels. |
| 46 | */ |
| 47 | static bool icount_sleep = true; |
| 48 | /* Arbitrarily pick 1MIPS as the minimum allowable speed. */ |
| 49 | #define MAX_ICOUNT_SHIFT 10 |
| 50 | |
| 51 | bool icount_align_option; |
| 52 | |
| 53 | /* Do not count executed instructions */ |
| 54 | ICountMode use_icount = ICOUNT_DISABLED; |
| 55 | |
| 56 | static void icount_enable_precise(void) |
| 57 | { |
| 58 | /* Fixed conversion of insn to ns via "shift" option */ |
| 59 | use_icount = ICOUNT_PRECISE; |
| 60 | } |
| 61 | |
| 62 | static void icount_enable_adaptive(void) |
| 63 | { |
| 64 | /* Runtime adaptive algorithm to compute shift */ |
| 65 | use_icount = ICOUNT_ADAPTATIVE; |
| 66 | } |
| 67 | |
| 68 | /* |
| 69 | * The current number of executed instructions is based on what we |
| 70 | * originally budgeted minus the current state of the decrementing |
| 71 | * icount counters in extra/u16.low. |
| 72 | */ |
| 73 | static int64_t icount_get_executed(CPUState *cpu) |
| 74 | { |
| 75 | return (cpu->icount_budget - |
| 76 | (cpu->neg.icount_decr.u16.low + cpu->icount_extra)); |
| 77 | } |
| 78 | |
| 79 | /* |
| 80 | * Update the global shared timer_state.qemu_icount to take into |
| 81 | * account executed instructions. This is done by the TCG vCPU |
| 82 | * thread so the main-loop can see time has moved forward. |
| 83 | */ |
| 84 | static void icount_update_locked(CPUState *cpu) |
| 85 | { |
| 86 | int64_t executed = icount_get_executed(cpu); |
| 87 | cpu->icount_budget -= executed; |
| 88 | |
| 89 | qatomic_set(&timers_state.qemu_icount, |
| 90 | timers_state.qemu_icount + executed); |
| 91 | } |
| 92 | |
| 93 | /* |
| 94 | * Update the global shared timer_state.qemu_icount to take into |
| 95 | * account executed instructions. This is done by the TCG vCPU |
| 96 | * thread so the main-loop can see time has moved forward. |
| 97 | */ |
| 98 | void icount_update(CPUState *cpu) |
| 99 | { |
| 100 | seqlock_write_lock(&timers_state.vm_clock_seqlock, |
| 101 | &timers_state.vm_clock_lock); |
| 102 | icount_update_locked(cpu); |
| 103 | seqlock_write_unlock(&timers_state.vm_clock_seqlock, |
| 104 | &timers_state.vm_clock_lock); |
| 105 | } |
| 106 | |
| 107 | static int64_t icount_get_raw_locked(void) |
| 108 | { |
| 109 | CPUState *cpu = current_cpu; |
| 110 | |
| 111 | if (cpu && cpu->running) { |
| 112 | if (!cpu->neg.can_do_io) { |
| 113 | error_report("Bad icount read"); |
| 114 | exit(1); |
| 115 | } |
| 116 | /* Take into account what has run */ |
| 117 | icount_update_locked(cpu); |
| 118 | } |
| 119 | /* The read is protected by the seqlock, but needs atomic to avoid UB */ |
| 120 | return qatomic_read(&timers_state.qemu_icount); |
| 121 | } |
| 122 | |
| 123 | static int64_t icount_get_locked(void) |
| 124 | { |
| 125 | int64_t icount = icount_get_raw_locked(); |
| 126 | return qatomic_read(&timers_state.qemu_icount_bias) + icount_to_ns(icount); |
| 127 | } |
| 128 | |
| 129 | int64_t icount_get_raw(void) |
| 130 | { |
| 131 | int64_t icount; |
| 132 | unsigned start; |
| 133 | |
| 134 | do { |
| 135 | start = seqlock_read_begin(&timers_state.vm_clock_seqlock); |
| 136 | icount = icount_get_raw_locked(); |
| 137 | } while (seqlock_read_retry(&timers_state.vm_clock_seqlock, start)); |
| 138 | |
| 139 | return icount; |
| 140 | } |
| 141 | |
| 142 | /* Return the virtual CPU time, based on the instruction counter. */ |
| 143 | int64_t icount_get(void) |
| 144 | { |
| 145 | int64_t icount; |
| 146 | unsigned start; |
| 147 | |
| 148 | do { |
| 149 | start = seqlock_read_begin(&timers_state.vm_clock_seqlock); |
| 150 | icount = icount_get_locked(); |
| 151 | } while (seqlock_read_retry(&timers_state.vm_clock_seqlock, start)); |
| 152 | |
| 153 | return icount; |
| 154 | } |
| 155 | |
| 156 | int64_t icount_to_ns(int64_t icount) |
| 157 | { |
| 158 | return icount << qatomic_read(&timers_state.icount_time_shift); |
| 159 | } |
| 160 | |
| 161 | /* |
| 162 | * Correlation between real and virtual time is always going to be |
| 163 | * fairly approximate, so ignore small variation. |
| 164 | * When the guest is idle real and virtual time will be aligned in |
| 165 | * the IO wait loop. |
| 166 | */ |
| 167 | #define ICOUNT_WOBBLE (NANOSECONDS_PER_SECOND / 10) |
| 168 | |
| 169 | static void icount_adjust(void) |
| 170 | { |
| 171 | int64_t cur_time; |
| 172 | int64_t cur_icount; |
| 173 | int64_t delta; |
| 174 | |
| 175 | /* If the VM is not running, then do nothing. */ |
| 176 | if (!runstate_is_running()) { |
| 177 | return; |
| 178 | } |
| 179 | |
| 180 | seqlock_write_lock(&timers_state.vm_clock_seqlock, |
| 181 | &timers_state.vm_clock_lock); |
| 182 | cur_time = REPLAY_CLOCK_LOCKED(REPLAY_CLOCK_VIRTUAL_RT, |
| 183 | cpu_get_clock_locked()); |
| 184 | cur_icount = icount_get_locked(); |
| 185 | |
| 186 | delta = cur_icount - cur_time; |
| 187 | /* FIXME: This is a very crude algorithm, somewhat prone to oscillation. */ |
| 188 | if (delta > 0 |
| 189 | && timers_state.last_delta + ICOUNT_WOBBLE < delta * 2 |
| 190 | && timers_state.icount_time_shift > 0) { |
| 191 | /* The guest is getting too far ahead. Slow time down. */ |
| 192 | qatomic_set(&timers_state.icount_time_shift, |
| 193 | timers_state.icount_time_shift - 1); |
| 194 | } |
| 195 | if (delta < 0 |
| 196 | && timers_state.last_delta - ICOUNT_WOBBLE > delta * 2 |
| 197 | && timers_state.icount_time_shift < MAX_ICOUNT_SHIFT) { |
| 198 | /* The guest is getting too far behind. Speed time up. */ |
| 199 | qatomic_set(&timers_state.icount_time_shift, |
| 200 | timers_state.icount_time_shift + 1); |
| 201 | } |
| 202 | timers_state.last_delta = delta; |
| 203 | qatomic_set(&timers_state.qemu_icount_bias, |
| 204 | cur_icount - (timers_state.qemu_icount |
| 205 | << timers_state.icount_time_shift)); |
| 206 | seqlock_write_unlock(&timers_state.vm_clock_seqlock, |
| 207 | &timers_state.vm_clock_lock); |
| 208 | } |
| 209 | |
| 210 | static void icount_adjust_rt(void *opaque) |
| 211 | { |
| 212 | timer_mod(timers_state.icount_rt_timer, |
| 213 | qemu_clock_get_ms(QEMU_CLOCK_VIRTUAL_RT) + 1000); |
| 214 | icount_adjust(); |
| 215 | } |
| 216 | |
| 217 | static void icount_adjust_vm(void *opaque) |
| 218 | { |
| 219 | timer_mod(timers_state.icount_vm_timer, |
| 220 | qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) + |
| 221 | NANOSECONDS_PER_SECOND / 10); |
| 222 | icount_adjust(); |
| 223 | } |
| 224 | |
| 225 | int64_t icount_round(int64_t count) |
| 226 | { |
| 227 | int shift = qatomic_read(&timers_state.icount_time_shift); |
| 228 | return (count + (1 << shift) - 1) >> shift; |
| 229 | } |
| 230 | |
| 231 | static void icount_warp_rt(void) |
| 232 | { |
| 233 | unsigned seq; |
| 234 | int64_t warp_start; |
| 235 | |
| 236 | /* |
| 237 | * The icount_warp_timer is rescheduled soon after vm_clock_warp_start |
| 238 | * changes from -1 to another value, so the race here is okay. |
| 239 | */ |
| 240 | do { |
| 241 | seq = seqlock_read_begin(&timers_state.vm_clock_seqlock); |
| 242 | warp_start = timers_state.vm_clock_warp_start; |
| 243 | } while (seqlock_read_retry(&timers_state.vm_clock_seqlock, seq)); |
| 244 | |
| 245 | if (warp_start == -1) { |
| 246 | return; |
| 247 | } |
| 248 | |
| 249 | seqlock_write_lock(&timers_state.vm_clock_seqlock, |
| 250 | &timers_state.vm_clock_lock); |
| 251 | if (runstate_is_running()) { |
| 252 | int64_t clock = REPLAY_CLOCK_LOCKED(REPLAY_CLOCK_VIRTUAL_RT, |
| 253 | cpu_get_clock_locked()); |
| 254 | int64_t warp_delta; |
| 255 | |
| 256 | warp_delta = clock - timers_state.vm_clock_warp_start; |
| 257 | if (icount_enabled() == ICOUNT_ADAPTATIVE) { |
| 258 | /* |
| 259 | * In adaptive mode, do not let QEMU_CLOCK_VIRTUAL run too far |
| 260 | * ahead of real time (it might already be ahead so careful not |
| 261 | * to go backwards). |
| 262 | */ |
| 263 | int64_t cur_icount = icount_get_locked(); |
| 264 | int64_t delta = clock - cur_icount; |
| 265 | |
| 266 | if (delta < 0) { |
| 267 | delta = 0; |
| 268 | } |
| 269 | warp_delta = MIN(warp_delta, delta); |
| 270 | } |
| 271 | qatomic_set(&timers_state.qemu_icount_bias, |
| 272 | timers_state.qemu_icount_bias + warp_delta); |
| 273 | } |
| 274 | timers_state.vm_clock_warp_start = -1; |
| 275 | seqlock_write_unlock(&timers_state.vm_clock_seqlock, |
| 276 | &timers_state.vm_clock_lock); |
| 277 | |
| 278 | if (qemu_clock_expired(QEMU_CLOCK_VIRTUAL)) { |
| 279 | qemu_clock_notify(QEMU_CLOCK_VIRTUAL); |
| 280 | } |
| 281 | } |
| 282 | |
| 283 | static void icount_timer_cb(void *opaque) |
| 284 | { |
| 285 | /* |
| 286 | * No need for a checkpoint because the timer already synchronizes |
| 287 | * with CHECKPOINT_CLOCK_VIRTUAL_RT. |
| 288 | */ |
| 289 | icount_warp_rt(); |
| 290 | } |
| 291 | |
| 292 | void icount_start_warp_timer(void) |
| 293 | { |
| 294 | int64_t clock; |
| 295 | int64_t deadline; |
| 296 | |
| 297 | assert(icount_enabled()); |
| 298 | |
| 299 | /* |
| 300 | * Nothing to do if the VM is stopped: QEMU_CLOCK_VIRTUAL timers |
| 301 | * do not fire, so computing the deadline does not make sense. |
| 302 | */ |
| 303 | if (!runstate_is_running()) { |
| 304 | return; |
| 305 | } |
| 306 | |
| 307 | if (replay_mode != REPLAY_MODE_PLAY) { |
| 308 | if (!all_cpu_threads_idle()) { |
| 309 | return; |
| 310 | } |
| 311 | |
| 312 | if (qtest_enabled()) { |
| 313 | /* When testing, qtest commands advance icount. */ |
| 314 | return; |
| 315 | } |
| 316 | |
| 317 | replay_checkpoint(CHECKPOINT_CLOCK_WARP_START); |
| 318 | } else { |
| 319 | /* warp clock deterministically in record/replay mode */ |
| 320 | if (!replay_checkpoint(CHECKPOINT_CLOCK_WARP_START)) { |
| 321 | /* |
| 322 | * vCPU is sleeping and warp can't be started. |
| 323 | * It is probably a race condition: notification sent |
| 324 | * to vCPU was processed in advance and vCPU went to sleep. |
| 325 | * Therefore we have to wake it up for doing something. |
| 326 | */ |
| 327 | if (replay_has_event()) { |
| 328 | qemu_clock_notify(QEMU_CLOCK_VIRTUAL); |
| 329 | } |
| 330 | return; |
| 331 | } |
| 332 | } |
| 333 | |
| 334 | /* We want to use the earliest deadline from ALL vm_clocks */ |
| 335 | clock = qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL_RT); |
| 336 | deadline = qemu_clock_deadline_ns_all(QEMU_CLOCK_VIRTUAL, |
| 337 | ~QEMU_TIMER_ATTR_EXTERNAL); |
| 338 | if (deadline < 0) { |
| 339 | if (!icount_sleep) { |
| 340 | warn_report_once("icount sleep disabled and no active timers"); |
| 341 | } |
| 342 | return; |
| 343 | } |
| 344 | |
| 345 | if (deadline > 0) { |
| 346 | /* |
| 347 | * Ensure QEMU_CLOCK_VIRTUAL proceeds even when the virtual CPU goes to |
| 348 | * sleep. Otherwise, the CPU might be waiting for a future timer |
| 349 | * interrupt to wake it up, but the interrupt never comes because |
| 350 | * the vCPU isn't running any insns and thus doesn't advance the |
| 351 | * QEMU_CLOCK_VIRTUAL. |
| 352 | */ |
| 353 | if (!icount_sleep) { |
| 354 | /* |
| 355 | * We never let VCPUs sleep in no sleep icount mode. |
| 356 | * If there is a pending QEMU_CLOCK_VIRTUAL timer we just advance |
| 357 | * to the next QEMU_CLOCK_VIRTUAL event and notify it. |
| 358 | * It is useful when we want a deterministic execution time, |
| 359 | * isolated from host latencies. |
| 360 | */ |
| 361 | seqlock_write_lock(&timers_state.vm_clock_seqlock, |
| 362 | &timers_state.vm_clock_lock); |
| 363 | qatomic_set(&timers_state.qemu_icount_bias, |
| 364 | timers_state.qemu_icount_bias + deadline); |
| 365 | seqlock_write_unlock(&timers_state.vm_clock_seqlock, |
| 366 | &timers_state.vm_clock_lock); |
| 367 | qemu_clock_notify(QEMU_CLOCK_VIRTUAL); |
| 368 | } else { |
| 369 | /* |
| 370 | * We do stop VCPUs and only advance QEMU_CLOCK_VIRTUAL after some |
| 371 | * "real" time, (related to the time left until the next event) has |
| 372 | * passed. The QEMU_CLOCK_VIRTUAL_RT clock will do this. |
| 373 | * This avoids that the warps are visible externally; for example, |
| 374 | * you will not be sending network packets continuously instead of |
| 375 | * every 100ms. |
| 376 | */ |
| 377 | seqlock_write_lock(&timers_state.vm_clock_seqlock, |
| 378 | &timers_state.vm_clock_lock); |
| 379 | if (timers_state.vm_clock_warp_start == -1 |
| 380 | || timers_state.vm_clock_warp_start > clock) { |
| 381 | timers_state.vm_clock_warp_start = clock; |
| 382 | } |
| 383 | seqlock_write_unlock(&timers_state.vm_clock_seqlock, |
| 384 | &timers_state.vm_clock_lock); |
| 385 | timer_mod_anticipate(timers_state.icount_warp_timer, |
| 386 | clock + deadline); |
| 387 | } |
| 388 | } else if (deadline == 0) { |
| 389 | qemu_clock_notify(QEMU_CLOCK_VIRTUAL); |
| 390 | } |
| 391 | } |
| 392 | |
| 393 | void icount_account_warp_timer(void) |
| 394 | { |
| 395 | if (!icount_sleep) { |
| 396 | return; |
| 397 | } |
| 398 | |
| 399 | /* |
| 400 | * Nothing to do if the VM is stopped: QEMU_CLOCK_VIRTUAL timers |
| 401 | * do not fire, so computing the deadline does not make sense. |
| 402 | */ |
| 403 | if (!runstate_is_running()) { |
| 404 | return; |
| 405 | } |
| 406 | |
| 407 | replay_async_events(); |
| 408 | |
| 409 | /* warp clock deterministically in record/replay mode */ |
| 410 | if (!replay_checkpoint(CHECKPOINT_CLOCK_WARP_ACCOUNT)) { |
| 411 | return; |
| 412 | } |
| 413 | |
| 414 | timer_del(timers_state.icount_warp_timer); |
| 415 | icount_warp_rt(); |
| 416 | } |
| 417 | |
| 418 | bool icount_configure(QemuOpts *opts, Error **errp) |
| 419 | { |
| 420 | const char *option = qemu_opt_get(opts, "shift"); |
| 421 | bool sleep = qemu_opt_get_bool(opts, "sleep", true); |
| 422 | bool align = qemu_opt_get_bool(opts, "align", false); |
| 423 | long time_shift = -1; |
| 424 | |
| 425 | if (!option) { |
| 426 | if (qemu_opt_get(opts, "align") != NULL) { |
| 427 | error_setg(errp, "Please specify shift option when using align"); |
| 428 | return false; |
| 429 | } |
| 430 | return true; |
| 431 | } |
| 432 | |
| 433 | if (align && !sleep) { |
| 434 | error_setg(errp, "align=on and sleep=off are incompatible"); |
| 435 | return false; |
| 436 | } |
| 437 | |
| 438 | if (strcmp(option, "auto") != 0) { |
| 439 | if (qemu_strtol(option, NULL, 0, &time_shift) < 0 |
| 440 | || time_shift < 0 || time_shift > MAX_ICOUNT_SHIFT) { |
| 441 | error_setg(errp, "icount: Invalid shift value"); |
| 442 | return false; |
| 443 | } |
| 444 | } else if (icount_align_option) { |
| 445 | error_setg(errp, "shift=auto and align=on are incompatible"); |
| 446 | return false; |
| 447 | } else if (!icount_sleep) { |
| 448 | error_setg(errp, "shift=auto and sleep=off are incompatible"); |
| 449 | return false; |
| 450 | } |
| 451 | |
| 452 | icount_sleep = sleep; |
| 453 | if (icount_sleep) { |
| 454 | timers_state.icount_warp_timer = timer_new_ns(QEMU_CLOCK_VIRTUAL_RT, |
| 455 | icount_timer_cb, NULL); |
| 456 | } |
| 457 | |
| 458 | icount_align_option = align; |
| 459 | |
| 460 | if (time_shift >= 0) { |
| 461 | timers_state.icount_time_shift = time_shift; |
| 462 | icount_enable_precise(); |
| 463 | return true; |
| 464 | } |
| 465 | |
| 466 | icount_enable_adaptive(); |
| 467 | |
| 468 | /* |
| 469 | * 125MIPS seems a reasonable initial guess at the guest speed. |
| 470 | * It will be corrected fairly quickly anyway. |
| 471 | */ |
| 472 | timers_state.icount_time_shift = 3; |
| 473 | |
| 474 | /* |
| 475 | * Have both realtime and virtual time triggers for speed adjustment. |
| 476 | * The realtime trigger catches emulated time passing too slowly, |
| 477 | * the virtual time trigger catches emulated time passing too fast. |
| 478 | * Realtime triggers occur even when idle, so use them less frequently |
| 479 | * than VM triggers. |
| 480 | */ |
| 481 | timers_state.vm_clock_warp_start = -1; |
| 482 | timers_state.icount_rt_timer = timer_new_ms(QEMU_CLOCK_VIRTUAL_RT, |
| 483 | icount_adjust_rt, NULL); |
| 484 | timer_mod(timers_state.icount_rt_timer, |
| 485 | qemu_clock_get_ms(QEMU_CLOCK_VIRTUAL_RT) + 1000); |
| 486 | timers_state.icount_vm_timer = timer_new_ns(QEMU_CLOCK_VIRTUAL, |
| 487 | icount_adjust_vm, NULL); |
| 488 | timer_mod(timers_state.icount_vm_timer, |
| 489 | qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) + |
| 490 | NANOSECONDS_PER_SECOND / 10); |
| 491 | return true; |
| 492 | } |
| 493 | |
| 494 | void icount_notify_exit(void) |
| 495 | { |
| 496 | assert(icount_enabled()); |
| 497 | |
| 498 | if (current_cpu) { |
| 499 | qemu_cpu_kick(current_cpu); |
| 500 | qemu_clock_notify(QEMU_CLOCK_VIRTUAL); |
| 501 | } |
| 502 | } |