cpu-exec.c

   1 /*
   2  *  emulator main execution loop
   3  *
   4  *  Copyright (c) 2003-2005 Fabrice Bellard
   5  *
   6  * This library is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU Lesser General Public
   8  * License as published by the Free Software Foundation; either
   9  * version 2 of the License, or (at your option) any later version.
  10  *
  11  * This library is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14  * Lesser General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU Lesser General Public
  17  * License along with this library; if not, see <http://www.gnu.org/licenses/>.
  18  */
  19 #include "config.h"
  20 #include "cpu.h"
  21 #include "trace.h"
  22 #include "disas/disas.h"
  23 #include "tcg.h"
  24 #include "qemu/atomic.h"
  25 #include "sysemu/qtest.h"
  26 #include "qemu/timer.h"
  27 #include "exec/address-spaces.h"
  28 #include "qemu/rcu.h"
  29 #include "exec/tb-hash.h"
  30
  31 /* -icount align implementation. */
  32
  33 typedef struct SyncClocks {
  34     int64_t diff_clk;
  35     int64_t last_cpu_icount;
  36     int64_t realtime_clock;
  37 } SyncClocks;
  38
  39 #if !defined(CONFIG_USER_ONLY)
  40 /* Allow the guest to have a max 3ms advance.
  41  * The difference between the 2 clocks could therefore
  42  * oscillate around 0.
  43  */
  44 #define VM_CLOCK_ADVANCE 3000000
  45 #define THRESHOLD_REDUCE 1.5
  46 #define MAX_DELAY_PRINT_RATE 2000000000LL
  47 #define MAX_NB_PRINTS 100
  48
  49 static void align_clocks(SyncClocks *sc, const CPUState *cpu)
  50 {
  51     int64_t cpu_icount;
  52
  53     if (!icount_align_option) {
  54         return;
  55     }
  56
  57     cpu_icount = cpu->icount_extra + cpu->icount_decr.u16.low;
  58     sc->diff_clk += cpu_icount_to_ns(sc->last_cpu_icount - cpu_icount);
  59     sc->last_cpu_icount = cpu_icount;
  60
  61     if (sc->diff_clk > VM_CLOCK_ADVANCE) {
  62 #ifndef _WIN32
  63         struct timespec sleep_delay, rem_delay;
  64         sleep_delay.tv_sec = sc->diff_clk / 1000000000LL;
  65         sleep_delay.tv_nsec = sc->diff_clk % 1000000000LL;
  66         if (nanosleep(&sleep_delay, &rem_delay) < 0) {
  67             sc->diff_clk = rem_delay.tv_sec * 1000000000LL + rem_delay.tv_nsec;
  68         } else {
  69             sc->diff_clk = 0;
  70         }
  71 #else
  72         Sleep(sc->diff_clk / SCALE_MS);
  73         sc->diff_clk = 0;
  74 #endif
  75     }
  76 }
  77
  78 static void print_delay(const SyncClocks *sc)
  79 {
  80     static float threshold_delay;
  81     static int64_t last_realtime_clock;
  82     static int nb_prints;
  83
  84     if (icount_align_option &&
  85         sc->realtime_clock - last_realtime_clock >= MAX_DELAY_PRINT_RATE &&
  86         nb_prints < MAX_NB_PRINTS) {
  87         if ((-sc->diff_clk / (float)1000000000LL > threshold_delay) ||
  88             (-sc->diff_clk / (float)1000000000LL <
  89              (threshold_delay - THRESHOLD_REDUCE))) {
  90             threshold_delay = (-sc->diff_clk / 1000000000LL) + 1;
  91             printf("Warning: The guest is now late by %.1f to %.1f seconds\n",
  92                    threshold_delay - 1,
  93                    threshold_delay);
  94             nb_prints++;
  95             last_realtime_clock = sc->realtime_clock;
  96         }
  97     }
  98 }
  99
 100 static void init_delay_params(SyncClocks *sc,
 101                               const CPUState *cpu)
 102 {
 103     if (!icount_align_option) {
 104         return;
 105     }
 106     sc->realtime_clock = qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL_RT);
 107     sc->diff_clk = qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) - sc->realtime_clock;
 108     sc->last_cpu_icount = cpu->icount_extra + cpu->icount_decr.u16.low;
 109     if (sc->diff_clk < max_delay) {
 110         max_delay = sc->diff_clk;
 111     }
 112     if (sc->diff_clk > max_advance) {
 113         max_advance = sc->diff_clk;
 114     }
 115
 116     /* Print every 2s max if the guest is late. We limit the number
 117        of printed messages to NB_PRINT_MAX(currently 100) */
 118     print_delay(sc);
 119 }
 120 #else
 121 static void align_clocks(SyncClocks *sc, const CPUState *cpu)
 122 {
 123 }
 124
 125 static void init_delay_params(SyncClocks *sc, const CPUState *cpu)
 126 {
 127 }
 128 #endif /* CONFIG USER ONLY */
 129
 130 /* Execute a TB, and fix up the CPU state afterwards if necessary */
 131 static inline tcg_target_ulong cpu_tb_exec(CPUState *cpu, uint8_t *tb_ptr)
 132 {
 133     CPUArchState *env = cpu->env_ptr;
 134     uintptr_t next_tb;
 135
 136 #if defined(DEBUG_DISAS)
 137     if (qemu_loglevel_mask(CPU_LOG_TB_CPU)) {
 138 #if defined(TARGET_I386)
 139         log_cpu_state(cpu, CPU_DUMP_CCOP);
 140 #elif defined(TARGET_M68K)
 141         /* ??? Should not modify env state for dumping.  */
 142         cpu_m68k_flush_flags(env, env->cc_op);
 143         env->cc_op = CC_OP_FLAGS;
 144         env->sr = (env->sr & 0xffe0) | env->cc_dest | (env->cc_x << 4);
 145         log_cpu_state(cpu, 0);
 146 #else
 147         log_cpu_state(cpu, 0);
 148 #endif
 149     }
 150 #endif /* DEBUG_DISAS */
 151
 152     cpu->can_do_io = !use_icount;
 153     next_tb = tcg_qemu_tb_exec(env, tb_ptr);
 154     cpu->can_do_io = 1;
 155     trace_exec_tb_exit((void *) (next_tb & ~TB_EXIT_MASK),
 156                        next_tb & TB_EXIT_MASK);
 157
 158     if ((next_tb & TB_EXIT_MASK) > TB_EXIT_IDX1) {
 159         /* We didn't start executing this TB (eg because the instruction
 160          * counter hit zero); we must restore the guest PC to the address
 161          * of the start of the TB.
 162          */
 163         CPUClass *cc = CPU_GET_CLASS(cpu);
 164         TranslationBlock *tb = (TranslationBlock *)(next_tb & ~TB_EXIT_MASK);
 165         if (cc->synchronize_from_tb) {
 166             cc->synchronize_from_tb(cpu, tb);
 167         } else {
 168             assert(cc->set_pc);
 169             cc->set_pc(cpu, tb->pc);
 170         }
 171     }
 172     if ((next_tb & TB_EXIT_MASK) == TB_EXIT_REQUESTED) {
 173         /* We were asked to stop executing TBs (probably a pending
 174          * interrupt. We've now stopped, so clear the flag.
 175          */
 176         cpu->tcg_exit_req = 0;
 177     }
 178     return next_tb;
 179 }
 180
 181 /* Execute the code without caching the generated code. An interpreter
 182    could be used if available. */
 183 static void cpu_exec_nocache(CPUState *cpu, int max_cycles,
 184                              TranslationBlock *orig_tb)
 185 {
 186     TranslationBlock *tb;
 187
 188     /* Should never happen.
 189        We only end up here when an existing TB is too long.  */
 190     if (max_cycles > CF_COUNT_MASK)
 191         max_cycles = CF_COUNT_MASK;
 192
 193     tb = tb_gen_code(cpu, orig_tb->pc, orig_tb->cs_base, orig_tb->flags,
 194                      max_cycles | CF_NOCACHE);
 195     tb->orig_tb = tcg_ctx.tb_ctx.tb_invalidated_flag ? NULL : orig_tb;
 196     cpu->current_tb = tb;
 197     /* execute the generated code */
 198     trace_exec_tb_nocache(tb, tb->pc);
 199     cpu_tb_exec(cpu, tb->tc_ptr);
 200     cpu->current_tb = NULL;
 201     tb_phys_invalidate(tb, -1);
 202     tb_free(tb);
 203 }
 204
 205 static TranslationBlock *tb_find_physical(CPUState *cpu,
 206                                           target_ulong pc,
 207                                           target_ulong cs_base,
 208                                           uint64_t flags)
 209 {
 210     CPUArchState *env = (CPUArchState *)cpu->env_ptr;
 211     TranslationBlock *tb, **ptb1;
 212     unsigned int h;
 213     tb_page_addr_t phys_pc, phys_page1;
 214     target_ulong virt_page2;
 215
 216     tcg_ctx.tb_ctx.tb_invalidated_flag = 0;
 217
 218     /* find translated block using physical mappings */
 219     phys_pc = get_page_addr_code(env, pc);
 220     phys_page1 = phys_pc & TARGET_PAGE_MASK;
 221     h = tb_phys_hash_func(phys_pc);
 222     ptb1 = &tcg_ctx.tb_ctx.tb_phys_hash[h];
 223     for(;;) {
 224         tb = *ptb1;
 225         if (!tb) {
 226             return NULL;
 227         }
 228         if (tb->pc == pc &&
 229             tb->page_addr[0] == phys_page1 &&
 230             tb->cs_base == cs_base &&
 231             tb->flags == flags) {
 232             /* check next page if needed */
 233             if (tb->page_addr[1] != -1) {
 234                 tb_page_addr_t phys_page2;
 235
 236                 virt_page2 = (pc & TARGET_PAGE_MASK) +
 237                     TARGET_PAGE_SIZE;
 238                 phys_page2 = get_page_addr_code(env, virt_page2);
 239                 if (tb->page_addr[1] == phys_page2) {
 240                     break;
 241                 }
 242             } else {
 243                 break;
 244             }
 245         }
 246         ptb1 = &tb->phys_hash_next;
 247     }
 248
 249     /* Move the TB to the head of the list */
 250     *ptb1 = tb->phys_hash_next;
 251     tb->phys_hash_next = tcg_ctx.tb_ctx.tb_phys_hash[h];
 252     tcg_ctx.tb_ctx.tb_phys_hash[h] = tb;
 253     return tb;
 254 }
 255
 256 static TranslationBlock *tb_find_slow(CPUState *cpu,
 257                                       target_ulong pc,
 258                                       target_ulong cs_base,
 259                                       uint64_t flags)
 260 {
 261     TranslationBlock *tb;
 262
 263     tb = tb_find_physical(cpu, pc, cs_base, flags);
 264     if (tb) {
 265         goto found;
 266     }
 267
 268 #ifdef CONFIG_USER_ONLY
 269     /* mmap_lock is needed by tb_gen_code, and mmap_lock must be
 270      * taken outside tb_lock.  Since we're momentarily dropping
 271      * tb_lock, there's a chance that our desired tb has been
 272      * translated.
 273      */
 274     tb_unlock();
 275     mmap_lock();
 276     tb_lock();
 277     tb = tb_find_physical(cpu, pc, cs_base, flags);
 278     if (tb) {
 279         mmap_unlock();
 280         goto found;
 281     }
 282 #endif
 283
 284     /* if no translated code available, then translate it now */
 285     tb = tb_gen_code(cpu, pc, cs_base, flags, 0);
 286
 287 #ifdef CONFIG_USER_ONLY
 288     mmap_unlock();
 289 #endif
 290
 291 found:
 292     /* we add the TB in the virtual pc hash table */
 293     cpu->tb_jmp_cache[tb_jmp_cache_hash_func(pc)] = tb;
 294     return tb;
 295 }
 296
 297 static inline TranslationBlock *tb_find_fast(CPUState *cpu)
 298 {
 299     CPUArchState *env = (CPUArchState *)cpu->env_ptr;
 300     TranslationBlock *tb;
 301     target_ulong cs_base, pc;
 302     int flags;
 303
 304     /* we record a subset of the CPU state. It will
 305        always be the same before a given translated block
 306        is executed. */
 307     cpu_get_tb_cpu_state(env, &pc, &cs_base, &flags);
 308     tb = cpu->tb_jmp_cache[tb_jmp_cache_hash_func(pc)];
 309     if (unlikely(!tb || tb->pc != pc || tb->cs_base != cs_base ||
 310                  tb->flags != flags)) {
 311         tb = tb_find_slow(cpu, pc, cs_base, flags);
 312     }
 313     return tb;
 314 }
 315
 316 static void cpu_handle_debug_exception(CPUState *cpu)
 317 {
 318     CPUClass *cc = CPU_GET_CLASS(cpu);
 319     CPUWatchpoint *wp;
 320
 321     if (!cpu->watchpoint_hit) {
 322         QTAILQ_FOREACH(wp, &cpu->watchpoints, entry) {
 323             wp->flags &= ~BP_WATCHPOINT_HIT;
 324         }
 325     }
 326
 327     cc->debug_excp_handler(cpu);
 328 }
 329
 330 /* main execution loop */
 331
 332 int cpu_exec(CPUState *cpu)
 333 {
 334     CPUClass *cc = CPU_GET_CLASS(cpu);
 335 #ifdef TARGET_I386
 336     X86CPU *x86_cpu = X86_CPU(cpu);
 337     CPUArchState *env = &x86_cpu->env;
 338 #endif
 339     int ret, interrupt_request;
 340     TranslationBlock *tb;
 341     uint8_t *tc_ptr;
 342     uintptr_t next_tb;
 343     SyncClocks sc;
 344
 345     if (cpu->halted) {
 346         if (!cpu_has_work(cpu)) {
 347             return EXCP_HALTED;
 348         }
 349
 350         cpu->halted = 0;
 351     }
 352
 353     current_cpu = cpu;
 354     atomic_mb_set(&tcg_current_cpu, cpu);
 355     rcu_read_lock();
 356
 357     if (unlikely(atomic_mb_read(&exit_request))) {
 358         cpu->exit_request = 1;
 359     }
 360
 361     cc->cpu_exec_enter(cpu);
 362
 363     /* Calculate difference between guest clock and host clock.
 364      * This delay includes the delay of the last cycle, so
 365      * what we have to do is sleep until it is 0. As for the
 366      * advance/delay we gain here, we try to fix it next time.
 367      */
 368     init_delay_params(&sc, cpu);
 369
 370     /* prepare setjmp context for exception handling */
 371     for(;;) {
 372         if (sigsetjmp(cpu->jmp_env, 0) == 0) {
 373             /* if an exception is pending, we execute it here */
 374             if (cpu->exception_index >= 0) {
 375                 if (cpu->exception_index >= EXCP_INTERRUPT) {
 376                     /* exit request from the cpu execution loop */
 377                     ret = cpu->exception_index;
 378                     if (ret == EXCP_DEBUG) {
 379                         cpu_handle_debug_exception(cpu);
 380                     }
 381                     cpu->exception_index = -1;
 382                     break;
 383                 } else {
 384 #if defined(CONFIG_USER_ONLY)
 385                     /* if user mode only, we simulate a fake exception
 386                        which will be handled outside the cpu execution
 387                        loop */
 388 #if defined(TARGET_I386)
 389                     cc->do_interrupt(cpu);
 390 #endif
 391                     ret = cpu->exception_index;
 392                     cpu->exception_index = -1;
 393                     break;
 394 #else
 395                     cc->do_interrupt(cpu);
 396                     cpu->exception_index = -1;
 397 #endif
 398                 }
 399             }
 400
 401             next_tb = 0; /* force lookup of first TB */
 402             for(;;) {
 403                 interrupt_request = cpu->interrupt_request;
 404                 if (unlikely(interrupt_request)) {
 405                     if (unlikely(cpu->singlestep_enabled & SSTEP_NOIRQ)) {
 406                         /* Mask out external interrupts for this step. */
 407                         interrupt_request &= ~CPU_INTERRUPT_SSTEP_MASK;
 408                     }
 409                     if (interrupt_request & CPU_INTERRUPT_DEBUG) {
 410                         cpu->interrupt_request &= ~CPU_INTERRUPT_DEBUG;
 411                         cpu->exception_index = EXCP_DEBUG;
 412                         cpu_loop_exit(cpu);
 413                     }
 414                     if (interrupt_request & CPU_INTERRUPT_HALT) {
 415                         cpu->interrupt_request &= ~CPU_INTERRUPT_HALT;
 416                         cpu->halted = 1;
 417                         cpu->exception_index = EXCP_HLT;
 418                         cpu_loop_exit(cpu);
 419                     }
 420 #if defined(TARGET_I386)
 421                     if (interrupt_request & CPU_INTERRUPT_INIT) {
 422                         cpu_svm_check_intercept_param(env, SVM_EXIT_INIT, 0);
 423                         do_cpu_init(x86_cpu);
 424                         cpu->exception_index = EXCP_HALTED;
 425                         cpu_loop_exit(cpu);
 426                     }
 427 #else
 428                     if (interrupt_request & CPU_INTERRUPT_RESET) {
 429                         cpu_reset(cpu);
 430                     }
 431 #endif
 432                     /* The target hook has 3 exit conditions:
 433                        False when the interrupt isn't processed,
 434                        True when it is, and we should restart on a new TB,
 435                        and via longjmp via cpu_loop_exit.  */
 436                     if (cc->cpu_exec_interrupt(cpu, interrupt_request)) {
 437                         next_tb = 0;
 438                     }
 439                     /* Don't use the cached interrupt_request value,
 440                        do_interrupt may have updated the EXITTB flag. */
 441                     if (cpu->interrupt_request & CPU_INTERRUPT_EXITTB) {
 442                         cpu->interrupt_request &= ~CPU_INTERRUPT_EXITTB;
 443                         /* ensure that no TB jump will be modified as
 444                            the program flow was changed */
 445                         next_tb = 0;
 446                     }
 447                 }
 448                 if (unlikely(cpu->exit_request)) {
 449                     cpu->exit_request = 0;
 450                     cpu->exception_index = EXCP_INTERRUPT;
 451                     cpu_loop_exit(cpu);
 452                 }
 453                 tb_lock();
 454                 tb = tb_find_fast(cpu);
 455                 /* Note: we do it here to avoid a gcc bug on Mac OS X when
 456                    doing it in tb_find_slow */
 457                 if (tcg_ctx.tb_ctx.tb_invalidated_flag) {
 458                     /* as some TB could have been invalidated because
 459                        of memory exceptions while generating the code, we
 460                        must recompute the hash index here */
 461                     next_tb = 0;
 462                     tcg_ctx.tb_ctx.tb_invalidated_flag = 0;
 463                 }
 464                 if (qemu_loglevel_mask(CPU_LOG_EXEC)) {
 465                     qemu_log("Trace %p [" TARGET_FMT_lx "] %s\n",
 466                              tb->tc_ptr, tb->pc, lookup_symbol(tb->pc));
 467                 }
 468                 /* see if we can patch the calling TB. When the TB
 469                    spans two pages, we cannot safely do a direct
 470                    jump. */
 471                 if (next_tb != 0 && tb->page_addr[1] == -1) {
 472                     tb_add_jump((TranslationBlock *)(next_tb & ~TB_EXIT_MASK),
 473                                 next_tb & TB_EXIT_MASK, tb);
 474                 }
 475                 tb_unlock();
 476                 if (likely(!cpu->exit_request)) {
 477                     trace_exec_tb(tb, tb->pc);
 478                     tc_ptr = tb->tc_ptr;
 479                     /* execute the generated code */
 480                     cpu->current_tb = tb;
 481                     next_tb = cpu_tb_exec(cpu, tc_ptr);
 482                     cpu->current_tb = NULL;
 483                     switch (next_tb & TB_EXIT_MASK) {
 484                     case TB_EXIT_REQUESTED:
 485                         /* Something asked us to stop executing
 486                          * chained TBs; just continue round the main
 487                          * loop. Whatever requested the exit will also
 488                          * have set something else (eg exit_request or
 489                          * interrupt_request) which we will handle
 490                          * next time around the loop.  But we need to
 491                          * ensure the tcg_exit_req read in generated code
 492                          * comes before the next read of cpu->exit_request
 493                          * or cpu->interrupt_request.
 494                          */
 495                         smp_rmb();
 496                         next_tb = 0;
 497                         break;
 498                     case TB_EXIT_ICOUNT_EXPIRED:
 499                     {
 500                         /* Instruction counter expired.  */
 501                         int insns_left = cpu->icount_decr.u32;
 502                         if (cpu->icount_extra && insns_left >= 0) {
 503                             /* Refill decrementer and continue execution.  */
 504                             cpu->icount_extra += insns_left;
 505                             insns_left = MIN(0xffff, cpu->icount_extra);
 506                             cpu->icount_extra -= insns_left;
 507                             cpu->icount_decr.u16.low = insns_left;
 508                         } else {
 509                             if (insns_left > 0) {
 510                                 /* Execute remaining instructions.  */
 511                                 tb = (TranslationBlock *)(next_tb & ~TB_EXIT_MASK);
 512                                 cpu_exec_nocache(cpu, insns_left, tb);
 513                                 align_clocks(&sc, cpu);
 514                             }
 515                             cpu->exception_index = EXCP_INTERRUPT;
 516                             next_tb = 0;
 517                             cpu_loop_exit(cpu);
 518                         }
 519                         break;
 520                     }
 521                     default:
 522                         break;
 523                     }
 524                 }
 525                 /* Try to align the host and virtual clocks
 526                    if the guest is in advance */
 527                 align_clocks(&sc, cpu);
 528                 /* reset soft MMU for next block (it can currently
 529                    only be set by a memory fault) */
 530             } /* for(;;) */
 531         } else {
 532             /* Reload env after longjmp - the compiler may have smashed all
 533              * local variables as longjmp is marked 'noreturn'. */
 534             cpu = current_cpu;
 535             cc = CPU_GET_CLASS(cpu);
 536             cpu->can_do_io = 1;
 537 #ifdef TARGET_I386
 538             x86_cpu = X86_CPU(cpu);
 539             env = &x86_cpu->env;
 540 #endif
 541             tb_lock_reset();
 542         }
 543     } /* for(;;) */
 544
 545     cc->cpu_exec_exit(cpu);
 546     rcu_read_unlock();
 547
 548     /* fail safe : never use current_cpu outside cpu_exec() */
 549     current_cpu = NULL;
 550
 551     /* Does not need atomic_mb_set because a spurious wakeup is okay.  */
 552     atomic_set(&tcg_current_cpu, NULL);
 553     return ret;
 554 }