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 "exec/memory-internal.h"
  29 #include "qemu/rcu.h"
  30 #include "exec/tb-hash.h"
  31
  32 /* -icount align implementation. */
  33
  34 typedef struct SyncClocks {
  35     int64_t diff_clk;
  36     int64_t last_cpu_icount;
  37     int64_t realtime_clock;
  38 } SyncClocks;
  39
  40 #if !defined(CONFIG_USER_ONLY)
  41 /* Allow the guest to have a max 3ms advance.
  42  * The difference between the 2 clocks could therefore
  43  * oscillate around 0.
  44  */
  45 #define VM_CLOCK_ADVANCE 3000000
  46 #define THRESHOLD_REDUCE 1.5
  47 #define MAX_DELAY_PRINT_RATE 2000000000LL
  48 #define MAX_NB_PRINTS 100
  49
  50 static void align_clocks(SyncClocks *sc, const CPUState *cpu)
  51 {
  52     int64_t cpu_icount;
  53
  54     if (!icount_align_option) {
  55         return;
  56     }
  57
  58     cpu_icount = cpu->icount_extra + cpu->icount_decr.u16.low;
  59     sc->diff_clk += cpu_icount_to_ns(sc->last_cpu_icount - cpu_icount);
  60     sc->last_cpu_icount = cpu_icount;
  61
  62     if (sc->diff_clk > VM_CLOCK_ADVANCE) {
  63 #ifndef _WIN32
  64         struct timespec sleep_delay, rem_delay;
  65         sleep_delay.tv_sec = sc->diff_clk / 1000000000LL;
  66         sleep_delay.tv_nsec = sc->diff_clk % 1000000000LL;
  67         if (nanosleep(&sleep_delay, &rem_delay) < 0) {
  68             sc->diff_clk = rem_delay.tv_sec * 1000000000LL + rem_delay.tv_nsec;
  69         } else {
  70             sc->diff_clk = 0;
  71         }
  72 #else
  73         Sleep(sc->diff_clk / SCALE_MS);
  74         sc->diff_clk = 0;
  75 #endif
  76     }
  77 }
  78
  79 static void print_delay(const SyncClocks *sc)
  80 {
  81     static float threshold_delay;
  82     static int64_t last_realtime_clock;
  83     static int nb_prints;
  84
  85     if (icount_align_option &&
  86         sc->realtime_clock - last_realtime_clock >= MAX_DELAY_PRINT_RATE &&
  87         nb_prints < MAX_NB_PRINTS) {
  88         if ((-sc->diff_clk / (float)1000000000LL > threshold_delay) ||
  89             (-sc->diff_clk / (float)1000000000LL <
  90              (threshold_delay - THRESHOLD_REDUCE))) {
  91             threshold_delay = (-sc->diff_clk / 1000000000LL) + 1;
  92             printf("Warning: The guest is now late by %.1f to %.1f seconds\n",
  93                    threshold_delay - 1,
  94                    threshold_delay);
  95             nb_prints++;
  96             last_realtime_clock = sc->realtime_clock;
  97         }
  98     }
  99 }
 100
 101 static void init_delay_params(SyncClocks *sc,
 102                               const CPUState *cpu)
 103 {
 104     if (!icount_align_option) {
 105         return;
 106     }
 107     sc->realtime_clock = qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL_RT);
 108     sc->diff_clk = qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) - sc->realtime_clock;
 109     sc->last_cpu_icount = cpu->icount_extra + cpu->icount_decr.u16.low;
 110     if (sc->diff_clk < max_delay) {
 111         max_delay = sc->diff_clk;
 112     }
 113     if (sc->diff_clk > max_advance) {
 114         max_advance = sc->diff_clk;
 115     }
 116
 117     /* Print every 2s max if the guest is late. We limit the number
 118        of printed messages to NB_PRINT_MAX(currently 100) */
 119     print_delay(sc);
 120 }
 121 #else
 122 static void align_clocks(SyncClocks *sc, const CPUState *cpu)
 123 {
 124 }
 125
 126 static void init_delay_params(SyncClocks *sc, const CPUState *cpu)
 127 {
 128 }
 129 #endif /* CONFIG USER ONLY */
 130
 131 void cpu_loop_exit(CPUState *cpu)
 132 {
 133     cpu->current_tb = NULL;
 134     siglongjmp(cpu->jmp_env, 1);
 135 }
 136
 137 /* exit the current TB from a signal handler. The host registers are
 138    restored in a state compatible with the CPU emulator
 139  */
 140 #if defined(CONFIG_SOFTMMU)
 141 void cpu_resume_from_signal(CPUState *cpu, void *puc)
 142 {
 143     /* XXX: restore cpu registers saved in host registers */
 144
 145     cpu->exception_index = -1;
 146     siglongjmp(cpu->jmp_env, 1);
 147 }
 148
 149 void cpu_reload_memory_map(CPUState *cpu)
 150 {
 151     AddressSpaceDispatch *d;
 152
 153     if (qemu_in_vcpu_thread()) {
 154         /* Do not let the guest prolong the critical section as much as it
 155          * as it desires.
 156          *
 157          * Currently, this is prevented by the I/O thread's periodinc kicking
 158          * of the VCPU thread (iothread_requesting_mutex, qemu_cpu_kick_thread)
 159          * but this will go away once TCG's execution moves out of the global
 160          * mutex.
 161          *
 162          * This pair matches cpu_exec's rcu_read_lock()/rcu_read_unlock(), which
 163          * only protects cpu->as->dispatch.  Since we reload it below, we can
 164          * split the critical section.
 165          */
 166         rcu_read_unlock();
 167         rcu_read_lock();
 168     }
 169
 170     /* The CPU and TLB are protected by the iothread lock.  */
 171     d = atomic_rcu_read(&cpu->as->dispatch);
 172     cpu->memory_dispatch = d;
 173     tlb_flush(cpu, 1);
 174 }
 175 #endif
 176
 177 /* Execute a TB, and fix up the CPU state afterwards if necessary */
 178 static inline tcg_target_ulong cpu_tb_exec(CPUState *cpu, uint8_t *tb_ptr)
 179 {
 180     CPUArchState *env = cpu->env_ptr;
 181     uintptr_t next_tb;
 182
 183 #if defined(DEBUG_DISAS)
 184     if (qemu_loglevel_mask(CPU_LOG_TB_CPU)) {
 185 #if defined(TARGET_I386)
 186         log_cpu_state(cpu, CPU_DUMP_CCOP);
 187 #elif defined(TARGET_M68K)
 188         /* ??? Should not modify env state for dumping.  */
 189         cpu_m68k_flush_flags(env, env->cc_op);
 190         env->cc_op = CC_OP_FLAGS;
 191         env->sr = (env->sr & 0xffe0) | env->cc_dest | (env->cc_x << 4);
 192         log_cpu_state(cpu, 0);
 193 #else
 194         log_cpu_state(cpu, 0);
 195 #endif
 196     }
 197 #endif /* DEBUG_DISAS */
 198
 199     cpu->can_do_io = !use_icount;
 200     next_tb = tcg_qemu_tb_exec(env, tb_ptr);
 201     cpu->can_do_io = 1;
 202     trace_exec_tb_exit((void *) (next_tb & ~TB_EXIT_MASK),
 203                        next_tb & TB_EXIT_MASK);
 204
 205     if ((next_tb & TB_EXIT_MASK) > TB_EXIT_IDX1) {
 206         /* We didn't start executing this TB (eg because the instruction
 207          * counter hit zero); we must restore the guest PC to the address
 208          * of the start of the TB.
 209          */
 210         CPUClass *cc = CPU_GET_CLASS(cpu);
 211         TranslationBlock *tb = (TranslationBlock *)(next_tb & ~TB_EXIT_MASK);
 212         if (cc->synchronize_from_tb) {
 213             cc->synchronize_from_tb(cpu, tb);
 214         } else {
 215             assert(cc->set_pc);
 216             cc->set_pc(cpu, tb->pc);
 217         }
 218     }
 219     if ((next_tb & TB_EXIT_MASK) == TB_EXIT_REQUESTED) {
 220         /* We were asked to stop executing TBs (probably a pending
 221          * interrupt. We've now stopped, so clear the flag.
 222          */
 223         cpu->tcg_exit_req = 0;
 224     }
 225     return next_tb;
 226 }
 227
 228 /* Execute the code without caching the generated code. An interpreter
 229    could be used if available. */
 230 static void cpu_exec_nocache(CPUState *cpu, int max_cycles,
 231                              TranslationBlock *orig_tb)
 232 {
 233     TranslationBlock *tb;
 234
 235     /* Should never happen.
 236        We only end up here when an existing TB is too long.  */
 237     if (max_cycles > CF_COUNT_MASK)
 238         max_cycles = CF_COUNT_MASK;
 239
 240     tb = tb_gen_code(cpu, orig_tb->pc, orig_tb->cs_base, orig_tb->flags,
 241                      max_cycles | CF_NOCACHE);
 242     tb->orig_tb = tcg_ctx.tb_ctx.tb_invalidated_flag ? NULL : orig_tb;
 243     cpu->current_tb = tb;
 244     /* execute the generated code */
 245     trace_exec_tb_nocache(tb, tb->pc);
 246     cpu_tb_exec(cpu, tb->tc_ptr);
 247     cpu->current_tb = NULL;
 248     tb_phys_invalidate(tb, -1);
 249     tb_free(tb);
 250 }
 251
 252 static TranslationBlock *tb_find_slow(CPUState *cpu,
 253                                       target_ulong pc,
 254                                       target_ulong cs_base,
 255                                       uint64_t flags)
 256 {
 257     CPUArchState *env = (CPUArchState *)cpu->env_ptr;
 258     TranslationBlock *tb, **ptb1;
 259     unsigned int h;
 260     tb_page_addr_t phys_pc, phys_page1;
 261     target_ulong virt_page2;
 262
 263     tcg_ctx.tb_ctx.tb_invalidated_flag = 0;
 264
 265     /* find translated block using physical mappings */
 266     phys_pc = get_page_addr_code(env, pc);
 267     phys_page1 = phys_pc & TARGET_PAGE_MASK;
 268     h = tb_phys_hash_func(phys_pc);
 269     ptb1 = &tcg_ctx.tb_ctx.tb_phys_hash[h];
 270     for(;;) {
 271         tb = *ptb1;
 272         if (!tb)
 273             goto not_found;
 274         if (tb->pc == pc &&
 275             tb->page_addr[0] == phys_page1 &&
 276             tb->cs_base == cs_base &&
 277             tb->flags == flags) {
 278             /* check next page if needed */
 279             if (tb->page_addr[1] != -1) {
 280                 tb_page_addr_t phys_page2;
 281
 282                 virt_page2 = (pc & TARGET_PAGE_MASK) +
 283                     TARGET_PAGE_SIZE;
 284                 phys_page2 = get_page_addr_code(env, virt_page2);
 285                 if (tb->page_addr[1] == phys_page2)
 286                     goto found;
 287             } else {
 288                 goto found;
 289             }
 290         }
 291         ptb1 = &tb->phys_hash_next;
 292     }
 293  not_found:
 294    /* if no translated code available, then translate it now */
 295     tb = tb_gen_code(cpu, pc, cs_base, flags, 0);
 296
 297  found:
 298     /* Move the last found TB to the head of the list */
 299     if (likely(*ptb1)) {
 300         *ptb1 = tb->phys_hash_next;
 301         tb->phys_hash_next = tcg_ctx.tb_ctx.tb_phys_hash[h];
 302         tcg_ctx.tb_ctx.tb_phys_hash[h] = tb;
 303     }
 304     /* we add the TB in the virtual pc hash table */
 305     cpu->tb_jmp_cache[tb_jmp_cache_hash_func(pc)] = tb;
 306     return tb;
 307 }
 308
 309 static inline TranslationBlock *tb_find_fast(CPUState *cpu)
 310 {
 311     CPUArchState *env = (CPUArchState *)cpu->env_ptr;
 312     TranslationBlock *tb;
 313     target_ulong cs_base, pc;
 314     int flags;
 315
 316     /* we record a subset of the CPU state. It will
 317        always be the same before a given translated block
 318        is executed. */
 319     cpu_get_tb_cpu_state(env, &pc, &cs_base, &flags);
 320     tb = cpu->tb_jmp_cache[tb_jmp_cache_hash_func(pc)];
 321     if (unlikely(!tb || tb->pc != pc || tb->cs_base != cs_base ||
 322                  tb->flags != flags)) {
 323         tb = tb_find_slow(cpu, pc, cs_base, flags);
 324     }
 325     return tb;
 326 }
 327
 328 static void cpu_handle_debug_exception(CPUState *cpu)
 329 {
 330     CPUClass *cc = CPU_GET_CLASS(cpu);
 331     CPUWatchpoint *wp;
 332
 333     if (!cpu->watchpoint_hit) {
 334         QTAILQ_FOREACH(wp, &cpu->watchpoints, entry) {
 335             wp->flags &= ~BP_WATCHPOINT_HIT;
 336         }
 337     }
 338
 339     cc->debug_excp_handler(cpu);
 340 }
 341
 342 /* main execution loop */
 343
 344 volatile sig_atomic_t exit_request;
 345 CPUState *tcg_current_cpu;
 346
 347 int cpu_exec(CPUState *cpu)
 348 {
 349     CPUClass *cc = CPU_GET_CLASS(cpu);
 350 #ifdef TARGET_I386
 351     X86CPU *x86_cpu = X86_CPU(cpu);
 352     CPUArchState *env = &x86_cpu->env;
 353 #endif
 354     int ret, interrupt_request;
 355     TranslationBlock *tb;
 356     uint8_t *tc_ptr;
 357     uintptr_t next_tb;
 358     SyncClocks sc;
 359
 360     /* This must be volatile so it is not trashed by longjmp() */
 361     volatile bool have_tb_lock = false;
 362
 363     if (cpu->halted) {
 364         if (!cpu_has_work(cpu)) {
 365             return EXCP_HALTED;
 366         }
 367
 368         cpu->halted = 0;
 369     }
 370
 371     current_cpu = cpu;
 372     atomic_mb_set(&tcg_current_cpu, cpu);
 373     rcu_read_lock();
 374
 375     if (unlikely(exit_request)) {
 376         cpu->exit_request = 1;
 377     }
 378
 379     cc->cpu_exec_enter(cpu);
 380
 381     /* Calculate difference between guest clock and host clock.
 382      * This delay includes the delay of the last cycle, so
 383      * what we have to do is sleep until it is 0. As for the
 384      * advance/delay we gain here, we try to fix it next time.
 385      */
 386     init_delay_params(&sc, cpu);
 387
 388     /* prepare setjmp context for exception handling */
 389     for(;;) {
 390         if (sigsetjmp(cpu->jmp_env, 0) == 0) {
 391             /* if an exception is pending, we execute it here */
 392             if (cpu->exception_index >= 0) {
 393                 if (cpu->exception_index >= EXCP_INTERRUPT) {
 394                     /* exit request from the cpu execution loop */
 395                     ret = cpu->exception_index;
 396                     if (ret == EXCP_DEBUG) {
 397                         cpu_handle_debug_exception(cpu);
 398                     }
 399                     cpu->exception_index = -1;
 400                     break;
 401                 } else {
 402 #if defined(CONFIG_USER_ONLY)
 403                     /* if user mode only, we simulate a fake exception
 404                        which will be handled outside the cpu execution
 405                        loop */
 406 #if defined(TARGET_I386)
 407                     cc->do_interrupt(cpu);
 408 #endif
 409                     ret = cpu->exception_index;
 410                     cpu->exception_index = -1;
 411                     break;
 412 #else
 413                     cc->do_interrupt(cpu);
 414                     cpu->exception_index = -1;
 415 #endif
 416                 }
 417             }
 418
 419             next_tb = 0; /* force lookup of first TB */
 420             for(;;) {
 421                 interrupt_request = cpu->interrupt_request;
 422                 if (unlikely(interrupt_request)) {
 423                     if (unlikely(cpu->singlestep_enabled & SSTEP_NOIRQ)) {
 424                         /* Mask out external interrupts for this step. */
 425                         interrupt_request &= ~CPU_INTERRUPT_SSTEP_MASK;
 426                     }
 427                     if (interrupt_request & CPU_INTERRUPT_DEBUG) {
 428                         cpu->interrupt_request &= ~CPU_INTERRUPT_DEBUG;
 429                         cpu->exception_index = EXCP_DEBUG;
 430                         cpu_loop_exit(cpu);
 431                     }
 432                     if (interrupt_request & CPU_INTERRUPT_HALT) {
 433                         cpu->interrupt_request &= ~CPU_INTERRUPT_HALT;
 434                         cpu->halted = 1;
 435                         cpu->exception_index = EXCP_HLT;
 436                         cpu_loop_exit(cpu);
 437                     }
 438 #if defined(TARGET_I386)
 439                     if (interrupt_request & CPU_INTERRUPT_INIT) {
 440                         cpu_svm_check_intercept_param(env, SVM_EXIT_INIT, 0);
 441                         do_cpu_init(x86_cpu);
 442                         cpu->exception_index = EXCP_HALTED;
 443                         cpu_loop_exit(cpu);
 444                     }
 445 #else
 446                     if (interrupt_request & CPU_INTERRUPT_RESET) {
 447                         cpu_reset(cpu);
 448                     }
 449 #endif
 450                     /* The target hook has 3 exit conditions:
 451                        False when the interrupt isn't processed,
 452                        True when it is, and we should restart on a new TB,
 453                        and via longjmp via cpu_loop_exit.  */
 454                     if (cc->cpu_exec_interrupt(cpu, interrupt_request)) {
 455                         next_tb = 0;
 456                     }
 457                     /* Don't use the cached interrupt_request value,
 458                        do_interrupt may have updated the EXITTB flag. */
 459                     if (cpu->interrupt_request & CPU_INTERRUPT_EXITTB) {
 460                         cpu->interrupt_request &= ~CPU_INTERRUPT_EXITTB;
 461                         /* ensure that no TB jump will be modified as
 462                            the program flow was changed */
 463                         next_tb = 0;
 464                     }
 465                 }
 466                 if (unlikely(cpu->exit_request)) {
 467                     cpu->exit_request = 0;
 468                     cpu->exception_index = EXCP_INTERRUPT;
 469                     cpu_loop_exit(cpu);
 470                 }
 471                 spin_lock(&tcg_ctx.tb_ctx.tb_lock);
 472                 have_tb_lock = true;
 473                 tb = tb_find_fast(cpu);
 474                 /* Note: we do it here to avoid a gcc bug on Mac OS X when
 475                    doing it in tb_find_slow */
 476                 if (tcg_ctx.tb_ctx.tb_invalidated_flag) {
 477                     /* as some TB could have been invalidated because
 478                        of memory exceptions while generating the code, we
 479                        must recompute the hash index here */
 480                     next_tb = 0;
 481                     tcg_ctx.tb_ctx.tb_invalidated_flag = 0;
 482                 }
 483                 if (qemu_loglevel_mask(CPU_LOG_EXEC)) {
 484                     qemu_log("Trace %p [" TARGET_FMT_lx "] %s\n",
 485                              tb->tc_ptr, tb->pc, lookup_symbol(tb->pc));
 486                 }
 487                 /* see if we can patch the calling TB. When the TB
 488                    spans two pages, we cannot safely do a direct
 489                    jump. */
 490                 if (next_tb != 0 && tb->page_addr[1] == -1) {
 491                     tb_add_jump((TranslationBlock *)(next_tb & ~TB_EXIT_MASK),
 492                                 next_tb & TB_EXIT_MASK, tb);
 493                 }
 494                 have_tb_lock = false;
 495                 spin_unlock(&tcg_ctx.tb_ctx.tb_lock);
 496                 if (likely(!cpu->exit_request)) {
 497                     trace_exec_tb(tb, tb->pc);
 498                     tc_ptr = tb->tc_ptr;
 499                     /* execute the generated code */
 500                     cpu->current_tb = tb;
 501                     next_tb = cpu_tb_exec(cpu, tc_ptr);
 502                     cpu->current_tb = NULL;
 503                     switch (next_tb & TB_EXIT_MASK) {
 504                     case TB_EXIT_REQUESTED:
 505                         /* Something asked us to stop executing
 506                          * chained TBs; just continue round the main
 507                          * loop. Whatever requested the exit will also
 508                          * have set something else (eg exit_request or
 509                          * interrupt_request) which we will handle
 510                          * next time around the loop.  But we need to
 511                          * ensure the tcg_exit_req read in generated code
 512                          * comes before the next read of cpu->exit_request
 513                          * or cpu->interrupt_request.
 514                          */
 515                         smp_rmb();
 516                         next_tb = 0;
 517                         break;
 518                     case TB_EXIT_ICOUNT_EXPIRED:
 519                     {
 520                         /* Instruction counter expired.  */
 521                         int insns_left = cpu->icount_decr.u32;
 522                         if (cpu->icount_extra && insns_left >= 0) {
 523                             /* Refill decrementer and continue execution.  */
 524                             cpu->icount_extra += insns_left;
 525                             insns_left = MIN(0xffff, cpu->icount_extra);
 526                             cpu->icount_extra -= insns_left;
 527                             cpu->icount_decr.u16.low = insns_left;
 528                         } else {
 529                             if (insns_left > 0) {
 530                                 /* Execute remaining instructions.  */
 531                                 tb = (TranslationBlock *)(next_tb & ~TB_EXIT_MASK);
 532                                 cpu_exec_nocache(cpu, insns_left, tb);
 533                                 align_clocks(&sc, cpu);
 534                             }
 535                             cpu->exception_index = EXCP_INTERRUPT;
 536                             next_tb = 0;
 537                             cpu_loop_exit(cpu);
 538                         }
 539                         break;
 540                     }
 541                     default:
 542                         break;
 543                     }
 544                 }
 545                 /* Try to align the host and virtual clocks
 546                    if the guest is in advance */
 547                 align_clocks(&sc, cpu);
 548                 /* reset soft MMU for next block (it can currently
 549                    only be set by a memory fault) */
 550             } /* for(;;) */
 551         } else {
 552             /* Reload env after longjmp - the compiler may have smashed all
 553              * local variables as longjmp is marked 'noreturn'. */
 554             cpu = current_cpu;
 555             cc = CPU_GET_CLASS(cpu);
 556             cpu->can_do_io = 1;
 557 #ifdef TARGET_I386
 558             x86_cpu = X86_CPU(cpu);
 559             env = &x86_cpu->env;
 560 #endif
 561             if (have_tb_lock) {
 562                 spin_unlock(&tcg_ctx.tb_ctx.tb_lock);
 563                 have_tb_lock = false;
 564             }
 565         }
 566     } /* for(;;) */
 567
 568     cc->cpu_exec_exit(cpu);
 569     rcu_read_unlock();
 570
 571     /* fail safe : never use current_cpu outside cpu_exec() */
 572     current_cpu = NULL;
 573
 574     /* Does not need atomic_mb_set because a spurious wakeup is okay.  */
 575     atomic_set(&tcg_current_cpu, NULL);
 576     return ret;
 577 }