softmmu/icount.c

   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-common.h"
  27 #include "qemu/cutils.h"
  28 #include "migration/vmstate.h"
  29 #include "qapi/error.h"
  30 #include "qemu/error-report.h"
  31 #include "exec/exec-all.h"
  32 #include "sysemu/cpus.h"
  33 #include "sysemu/qtest.h"
  34 #include "qemu/main-loop.h"
  35 #include "qemu/option.h"
  36 #include "qemu/seqlock.h"
  37 #include "sysemu/replay.h"
  38 #include "sysemu/runstate.h"
  39 #include "hw/core/cpu.h"
  40 #include "sysemu/cpu-timers.h"
  41 #include "sysemu/cpu-throttle.h"
  42 #include "timers-state.h"
  43
  44 /*
  45  * ICOUNT: Instruction Counter
  46  *
  47  * this module is split off from cpu-timers because the icount part
  48  * is TCG-specific, and does not need to be built for other accels.
  49  */
  50 static bool icount_sleep = true;
  51 /* Arbitrarily pick 1MIPS as the minimum allowable speed.  */
  52 #define MAX_ICOUNT_SHIFT 10
  53
  54 /*
  55  * 0 = Do not count executed instructions.
  56  * 1 = Fixed conversion of insn to ns via "shift" option
  57  * 2 = Runtime adaptive algorithm to compute shift
  58  */
  59 int use_icount;
  60
  61 static void icount_enable_precise(void)
  62 {
  63     use_icount = 1;
  64 }
  65
  66 static void icount_enable_adaptive(void)
  67 {
  68     use_icount = 2;
  69 }
  70
  71 /*
  72  * The current number of executed instructions is based on what we
  73  * originally budgeted minus the current state of the decrementing
  74  * icount counters in extra/u16.low.
  75  */
  76 static int64_t icount_get_executed(CPUState *cpu)
  77 {
  78     return (cpu->icount_budget -
  79             (cpu_neg(cpu)->icount_decr.u16.low + cpu->icount_extra));
  80 }
  81
  82 /*
  83  * Update the global shared timer_state.qemu_icount to take into
  84  * account executed instructions. This is done by the TCG vCPU
  85  * thread so the main-loop can see time has moved forward.
  86  */
  87 static void icount_update_locked(CPUState *cpu)
  88 {
  89     int64_t executed = icount_get_executed(cpu);
  90     cpu->icount_budget -= executed;
  91
  92     qatomic_set_i64(&timers_state.qemu_icount,
  93                     timers_state.qemu_icount + executed);
  94 }
  95
  96 /*
  97  * Update the global shared timer_state.qemu_icount to take into
  98  * account executed instructions. This is done by the TCG vCPU
  99  * thread so the main-loop can see time has moved forward.
 100  */
 101 void icount_update(CPUState *cpu)
 102 {
 103     seqlock_write_lock(&timers_state.vm_clock_seqlock,
 104                        &timers_state.vm_clock_lock);
 105     icount_update_locked(cpu);
 106     seqlock_write_unlock(&timers_state.vm_clock_seqlock,
 107                          &timers_state.vm_clock_lock);
 108 }
 109
 110 static int64_t icount_get_raw_locked(void)
 111 {
 112     CPUState *cpu = current_cpu;
 113
 114     if (cpu && cpu->running) {
 115         if (!cpu->can_do_io) {
 116             error_report("Bad icount read");
 117             exit(1);
 118         }
 119         /* Take into account what has run */
 120         icount_update_locked(cpu);
 121     }
 122     /* The read is protected by the seqlock, but needs atomic64 to avoid UB */
 123     return qatomic_read_i64(&timers_state.qemu_icount);
 124 }
 125
 126 static int64_t icount_get_locked(void)
 127 {
 128     int64_t icount = icount_get_raw_locked();
 129     return qatomic_read_i64(&timers_state.qemu_icount_bias) +
 130         icount_to_ns(icount);
 131 }
 132
 133 int64_t icount_get_raw(void)
 134 {
 135     int64_t icount;
 136     unsigned start;
 137
 138     do {
 139         start = seqlock_read_begin(&timers_state.vm_clock_seqlock);
 140         icount = icount_get_raw_locked();
 141     } while (seqlock_read_retry(&timers_state.vm_clock_seqlock, start));
 142
 143     return icount;
 144 }
 145
 146 /* Return the virtual CPU time, based on the instruction counter.  */
 147 int64_t icount_get(void)
 148 {
 149     int64_t icount;
 150     unsigned start;
 151
 152     do {
 153         start = seqlock_read_begin(&timers_state.vm_clock_seqlock);
 154         icount = icount_get_locked();
 155     } while (seqlock_read_retry(&timers_state.vm_clock_seqlock, start));
 156
 157     return icount;
 158 }
 159
 160 int64_t icount_to_ns(int64_t icount)
 161 {
 162     return icount << qatomic_read(&timers_state.icount_time_shift);
 163 }
 164
 165 /*
 166  * Correlation between real and virtual time is always going to be
 167  * fairly approximate, so ignore small variation.
 168  * When the guest is idle real and virtual time will be aligned in
 169  * the IO wait loop.
 170  */
 171 #define ICOUNT_WOBBLE (NANOSECONDS_PER_SECOND / 10)
 172
 173 static void icount_adjust(void)
 174 {
 175     int64_t cur_time;
 176     int64_t cur_icount;
 177     int64_t delta;
 178
 179     /* Protected by TimersState mutex.  */
 180     static int64_t last_delta;
 181
 182     /* If the VM is not running, then do nothing.  */
 183     if (!runstate_is_running()) {
 184         return;
 185     }
 186
 187     seqlock_write_lock(&timers_state.vm_clock_seqlock,
 188                        &timers_state.vm_clock_lock);
 189     cur_time = REPLAY_CLOCK_LOCKED(REPLAY_CLOCK_VIRTUAL_RT,
 190                                    cpu_get_clock_locked());
 191     cur_icount = icount_get_locked();
 192
 193     delta = cur_icount - cur_time;
 194     /* FIXME: This is a very crude algorithm, somewhat prone to oscillation.  */
 195     if (delta > 0
 196         && last_delta + ICOUNT_WOBBLE < delta * 2
 197         && timers_state.icount_time_shift > 0) {
 198         /* The guest is getting too far ahead.  Slow time down.  */
 199         qatomic_set(&timers_state.icount_time_shift,
 200                     timers_state.icount_time_shift - 1);
 201     }
 202     if (delta < 0
 203         && last_delta - ICOUNT_WOBBLE > delta * 2
 204         && timers_state.icount_time_shift < MAX_ICOUNT_SHIFT) {
 205         /* The guest is getting too far behind.  Speed time up.  */
 206         qatomic_set(&timers_state.icount_time_shift,
 207                     timers_state.icount_time_shift + 1);
 208     }
 209     last_delta = delta;
 210     qatomic_set_i64(&timers_state.qemu_icount_bias,
 211                     cur_icount - (timers_state.qemu_icount
 212                                   << timers_state.icount_time_shift));
 213     seqlock_write_unlock(&timers_state.vm_clock_seqlock,
 214                          &timers_state.vm_clock_lock);
 215 }
 216
 217 static void icount_adjust_rt(void *opaque)
 218 {
 219     timer_mod(timers_state.icount_rt_timer,
 220               qemu_clock_get_ms(QEMU_CLOCK_VIRTUAL_RT) + 1000);
 221     icount_adjust();
 222 }
 223
 224 static void icount_adjust_vm(void *opaque)
 225 {
 226     timer_mod(timers_state.icount_vm_timer,
 227                    qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) +
 228                    NANOSECONDS_PER_SECOND / 10);
 229     icount_adjust();
 230 }
 231
 232 int64_t icount_round(int64_t count)
 233 {
 234     int shift = qatomic_read(&timers_state.icount_time_shift);
 235     return (count + (1 << shift) - 1) >> shift;
 236 }
 237
 238 static void icount_warp_rt(void)
 239 {
 240     unsigned seq;
 241     int64_t warp_start;
 242
 243     /*
 244      * The icount_warp_timer is rescheduled soon after vm_clock_warp_start
 245      * changes from -1 to another value, so the race here is okay.
 246      */
 247     do {
 248         seq = seqlock_read_begin(&timers_state.vm_clock_seqlock);
 249         warp_start = timers_state.vm_clock_warp_start;
 250     } while (seqlock_read_retry(&timers_state.vm_clock_seqlock, seq));
 251
 252     if (warp_start == -1) {
 253         return;
 254     }
 255
 256     seqlock_write_lock(&timers_state.vm_clock_seqlock,
 257                        &timers_state.vm_clock_lock);
 258     if (runstate_is_running()) {
 259         int64_t clock = REPLAY_CLOCK_LOCKED(REPLAY_CLOCK_VIRTUAL_RT,
 260                                             cpu_get_clock_locked());
 261         int64_t warp_delta;
 262
 263         warp_delta = clock - timers_state.vm_clock_warp_start;
 264         if (icount_enabled() == 2) {
 265             /*
 266              * In adaptive mode, do not let QEMU_CLOCK_VIRTUAL run too
 267              * far ahead of real time.
 268              */
 269             int64_t cur_icount = icount_get_locked();
 270             int64_t delta = clock - cur_icount;
 271             warp_delta = MIN(warp_delta, delta);
 272         }
 273         qatomic_set_i64(&timers_state.qemu_icount_bias,
 274                         timers_state.qemu_icount_bias + warp_delta);
 275     }
 276     timers_state.vm_clock_warp_start = -1;
 277     seqlock_write_unlock(&timers_state.vm_clock_seqlock,
 278                        &timers_state.vm_clock_lock);
 279
 280     if (qemu_clock_expired(QEMU_CLOCK_VIRTUAL)) {
 281         qemu_clock_notify(QEMU_CLOCK_VIRTUAL);
 282     }
 283 }
 284
 285 static void icount_timer_cb(void *opaque)
 286 {
 287     /*
 288      * No need for a checkpoint because the timer already synchronizes
 289      * with CHECKPOINT_CLOCK_VIRTUAL_RT.
 290      */
 291     icount_warp_rt();
 292 }
 293
 294 void icount_start_warp_timer(void)
 295 {
 296     int64_t clock;
 297     int64_t deadline;
 298
 299     assert(icount_enabled());
 300
 301     /*
 302      * Nothing to do if the VM is stopped: QEMU_CLOCK_VIRTUAL timers
 303      * do not fire, so computing the deadline does not make sense.
 304      */
 305     if (!runstate_is_running()) {
 306         return;
 307     }
 308
 309     if (replay_mode != REPLAY_MODE_PLAY) {
 310         if (!all_cpu_threads_idle()) {
 311             return;
 312         }
 313
 314         if (qtest_enabled()) {
 315             /* When testing, qtest commands advance icount.  */
 316             return;
 317         }
 318
 319         replay_checkpoint(CHECKPOINT_CLOCK_WARP_START);
 320     } else {
 321         /* warp clock deterministically in record/replay mode */
 322         if (!replay_checkpoint(CHECKPOINT_CLOCK_WARP_START)) {
 323             /*
 324              * vCPU is sleeping and warp can't be started.
 325              * It is probably a race condition: notification sent
 326              * to vCPU was processed in advance and vCPU went to sleep.
 327              * Therefore we have to wake it up for doing someting.
 328              */
 329             if (replay_has_checkpoint()) {
 330                 qemu_clock_notify(QEMU_CLOCK_VIRTUAL);
 331             }
 332             return;
 333         }
 334     }
 335
 336     /* We want to use the earliest deadline from ALL vm_clocks */
 337     clock = qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL_RT);
 338     deadline = qemu_clock_deadline_ns_all(QEMU_CLOCK_VIRTUAL,
 339                                           ~QEMU_TIMER_ATTR_EXTERNAL);
 340     if (deadline < 0) {
 341         static bool notified;
 342         if (!icount_sleep && !notified) {
 343             warn_report("icount sleep disabled and no active timers");
 344             notified = true;
 345         }
 346         return;
 347     }
 348
 349     if (deadline > 0) {
 350         /*
 351          * Ensure QEMU_CLOCK_VIRTUAL proceeds even when the virtual CPU goes to
 352          * sleep.  Otherwise, the CPU might be waiting for a future timer
 353          * interrupt to wake it up, but the interrupt never comes because
 354          * the vCPU isn't running any insns and thus doesn't advance the
 355          * QEMU_CLOCK_VIRTUAL.
 356          */
 357         if (!icount_sleep) {
 358             /*
 359              * We never let VCPUs sleep in no sleep icount mode.
 360              * If there is a pending QEMU_CLOCK_VIRTUAL timer we just advance
 361              * to the next QEMU_CLOCK_VIRTUAL event and notify it.
 362              * It is useful when we want a deterministic execution time,
 363              * isolated from host latencies.
 364              */
 365             seqlock_write_lock(&timers_state.vm_clock_seqlock,
 366                                &timers_state.vm_clock_lock);
 367             qatomic_set_i64(&timers_state.qemu_icount_bias,
 368                             timers_state.qemu_icount_bias + deadline);
 369             seqlock_write_unlock(&timers_state.vm_clock_seqlock,
 370                                  &timers_state.vm_clock_lock);
 371             qemu_clock_notify(QEMU_CLOCK_VIRTUAL);
 372         } else {
 373             /*
 374              * We do stop VCPUs and only advance QEMU_CLOCK_VIRTUAL after some
 375              * "real" time, (related to the time left until the next event) has
 376              * passed. The QEMU_CLOCK_VIRTUAL_RT clock will do this.
 377              * This avoids that the warps are visible externally; for example,
 378              * you will not be sending network packets continuously instead of
 379              * every 100ms.
 380              */
 381             seqlock_write_lock(&timers_state.vm_clock_seqlock,
 382                                &timers_state.vm_clock_lock);
 383             if (timers_state.vm_clock_warp_start == -1
 384                 || timers_state.vm_clock_warp_start > clock) {
 385                 timers_state.vm_clock_warp_start = clock;
 386             }
 387             seqlock_write_unlock(&timers_state.vm_clock_seqlock,
 388                                  &timers_state.vm_clock_lock);
 389             timer_mod_anticipate(timers_state.icount_warp_timer,
 390                                  clock + deadline);
 391         }
 392     } else if (deadline == 0) {
 393         qemu_clock_notify(QEMU_CLOCK_VIRTUAL);
 394     }
 395 }
 396
 397 void icount_account_warp_timer(void)
 398 {
 399     if (!icount_sleep) {
 400         return;
 401     }
 402
 403     /*
 404      * Nothing to do if the VM is stopped: QEMU_CLOCK_VIRTUAL timers
 405      * do not fire, so computing the deadline does not make sense.
 406      */
 407     if (!runstate_is_running()) {
 408         return;
 409     }
 410
 411     /* warp clock deterministically in record/replay mode */
 412     if (!replay_checkpoint(CHECKPOINT_CLOCK_WARP_ACCOUNT)) {
 413         return;
 414     }
 415
 416     timer_del(timers_state.icount_warp_timer);
 417     icount_warp_rt();
 418 }
 419
 420 void icount_configure(QemuOpts *opts, Error **errp)
 421 {
 422     const char *option = qemu_opt_get(opts, "shift");
 423     bool sleep = qemu_opt_get_bool(opts, "sleep", true);
 424     bool align = qemu_opt_get_bool(opts, "align", false);
 425     long time_shift = -1;
 426
 427     if (!option) {
 428         if (qemu_opt_get(opts, "align") != NULL) {
 429             error_setg(errp, "Please specify shift option when using align");
 430         }
 431         return;
 432     }
 433
 434     if (align && !sleep) {
 435         error_setg(errp, "align=on and sleep=off are incompatible");
 436         return;
 437     }
 438
 439     if (strcmp(option, "auto") != 0) {
 440         if (qemu_strtol(option, NULL, 0, &time_shift) < 0
 441             || time_shift < 0 || time_shift > MAX_ICOUNT_SHIFT) {
 442             error_setg(errp, "icount: Invalid shift value");
 443             return;
 444         }
 445     } else if (icount_align_option) {
 446         error_setg(errp, "shift=auto and align=on are incompatible");
 447         return;
 448     } else if (!icount_sleep) {
 449         error_setg(errp, "shift=auto and sleep=off are incompatible");
 450         return;
 451     }
 452
 453     icount_sleep = sleep;
 454     if (icount_sleep) {
 455         timers_state.icount_warp_timer = timer_new_ns(QEMU_CLOCK_VIRTUAL_RT,
 456                                          icount_timer_cb, NULL);
 457     }
 458
 459     icount_align_option = align;
 460
 461     if (time_shift >= 0) {
 462         timers_state.icount_time_shift = time_shift;
 463         icount_enable_precise();
 464         return;
 465     }
 466
 467     icount_enable_adaptive();
 468
 469     /*
 470      * 125MIPS seems a reasonable initial guess at the guest speed.
 471      * It will be corrected fairly quickly anyway.
 472      */
 473     timers_state.icount_time_shift = 3;
 474
 475     /*
 476      * Have both realtime and virtual time triggers for speed adjustment.
 477      * The realtime trigger catches emulated time passing too slowly,
 478      * the virtual time trigger catches emulated time passing too fast.
 479      * Realtime triggers occur even when idle, so use them less frequently
 480      * than VM triggers.
 481      */
 482     timers_state.vm_clock_warp_start = -1;
 483     timers_state.icount_rt_timer = timer_new_ms(QEMU_CLOCK_VIRTUAL_RT,
 484                                    icount_adjust_rt, NULL);
 485     timer_mod(timers_state.icount_rt_timer,
 486                    qemu_clock_get_ms(QEMU_CLOCK_VIRTUAL_RT) + 1000);
 487     timers_state.icount_vm_timer = timer_new_ns(QEMU_CLOCK_VIRTUAL,
 488                                         icount_adjust_vm, NULL);
 489     timer_mod(timers_state.icount_vm_timer,
 490                    qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) +
 491                    NANOSECONDS_PER_SECOND / 10);
 492 }