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 void cpu_loop_exit_restore(CPUState *cpu, uintptr_t pc)
 138 {
 139     if (pc) {
 140         cpu_restore_state(cpu, pc);
 141     }
 142     cpu->current_tb = NULL;
 143     siglongjmp(cpu->jmp_env, 1);
 144 }
 145
 146 /* exit the current TB from a signal handler. The host registers are
 147    restored in a state compatible with the CPU emulator
 148  */
 149 #if defined(CONFIG_SOFTMMU)
 150 void cpu_resume_from_signal(CPUState *cpu, void *puc)
 151 {
 152     /* XXX: restore cpu registers saved in host registers */
 153
 154     cpu->exception_index = -1;
 155     siglongjmp(cpu->jmp_env, 1);
 156 }
 157
 158 void cpu_reload_memory_map(CPUState *cpu)
 159 {
 160     AddressSpaceDispatch *d;
 161
 162     if (qemu_in_vcpu_thread()) {
 163         /* Do not let the guest prolong the critical section as much as it
 164          * as it desires.
 165          *
 166          * Currently, this is prevented by the I/O thread's periodinc kicking
 167          * of the VCPU thread (iothread_requesting_mutex, qemu_cpu_kick_thread)
 168          * but this will go away once TCG's execution moves out of the global
 169          * mutex.
 170          *
 171          * This pair matches cpu_exec's rcu_read_lock()/rcu_read_unlock(), which
 172          * only protects cpu->as->dispatch.  Since we reload it below, we can
 173          * split the critical section.
 174          */
 175         rcu_read_unlock();
 176         rcu_read_lock();
 177     }
 178
 179     /* The CPU and TLB are protected by the iothread lock.  */
 180     d = atomic_rcu_read(&cpu->as->dispatch);
 181     cpu->memory_dispatch = d;
 182     tlb_flush(cpu, 1);
 183 }
 184 #endif
 185
 186 /* Execute a TB, and fix up the CPU state afterwards if necessary */
 187 static inline tcg_target_ulong cpu_tb_exec(CPUState *cpu, uint8_t *tb_ptr)
 188 {
 189     CPUArchState *env = cpu->env_ptr;
 190     uintptr_t next_tb;
 191
 192 #if defined(DEBUG_DISAS)
 193     if (qemu_loglevel_mask(CPU_LOG_TB_CPU)) {
 194 #if defined(TARGET_I386)
 195         log_cpu_state(cpu, CPU_DUMP_CCOP);
 196 #elif defined(TARGET_M68K)
 197         /* ??? Should not modify env state for dumping.  */
 198         cpu_m68k_flush_flags(env, env->cc_op);
 199         env->cc_op = CC_OP_FLAGS;
 200         env->sr = (env->sr & 0xffe0) | env->cc_dest | (env->cc_x << 4);
 201         log_cpu_state(cpu, 0);
 202 #else
 203         log_cpu_state(cpu, 0);
 204 #endif
 205     }
 206 #endif /* DEBUG_DISAS */
 207
 208     cpu->can_do_io = !use_icount;
 209     next_tb = tcg_qemu_tb_exec(env, tb_ptr);
 210     cpu->can_do_io = 1;
 211     trace_exec_tb_exit((void *) (next_tb & ~TB_EXIT_MASK),
 212                        next_tb & TB_EXIT_MASK);
 213
 214     if ((next_tb & TB_EXIT_MASK) > TB_EXIT_IDX1) {
 215         /* We didn't start executing this TB (eg because the instruction
 216          * counter hit zero); we must restore the guest PC to the address
 217          * of the start of the TB.
 218          */
 219         CPUClass *cc = CPU_GET_CLASS(cpu);
 220         TranslationBlock *tb = (TranslationBlock *)(next_tb & ~TB_EXIT_MASK);
 221         if (cc->synchronize_from_tb) {
 222             cc->synchronize_from_tb(cpu, tb);
 223         } else {
 224             assert(cc->set_pc);
 225             cc->set_pc(cpu, tb->pc);
 226         }
 227     }
 228     if ((next_tb & TB_EXIT_MASK) == TB_EXIT_REQUESTED) {
 229         /* We were asked to stop executing TBs (probably a pending
 230          * interrupt. We've now stopped, so clear the flag.
 231          */
 232         cpu->tcg_exit_req = 0;
 233     }
 234     return next_tb;
 235 }
 236
 237 /* Execute the code without caching the generated code. An interpreter
 238    could be used if available. */
 239 static void cpu_exec_nocache(CPUState *cpu, int max_cycles,
 240                              TranslationBlock *orig_tb)
 241 {
 242     TranslationBlock *tb;
 243
 244     /* Should never happen.
 245        We only end up here when an existing TB is too long.  */
 246     if (max_cycles > CF_COUNT_MASK)
 247         max_cycles = CF_COUNT_MASK;
 248
 249     tb = tb_gen_code(cpu, orig_tb->pc, orig_tb->cs_base, orig_tb->flags,
 250                      max_cycles | CF_NOCACHE);
 251     tb->orig_tb = tcg_ctx.tb_ctx.tb_invalidated_flag ? NULL : orig_tb;
 252     cpu->current_tb = tb;
 253     /* execute the generated code */
 254     trace_exec_tb_nocache(tb, tb->pc);
 255     cpu_tb_exec(cpu, tb->tc_ptr);
 256     cpu->current_tb = NULL;
 257     tb_phys_invalidate(tb, -1);
 258     tb_free(tb);
 259 }
 260
 261 static TranslationBlock *tb_find_physical(CPUState *cpu,
 262                                           target_ulong pc,
 263                                           target_ulong cs_base,
 264                                           uint64_t flags)
 265 {
 266     CPUArchState *env = (CPUArchState *)cpu->env_ptr;
 267     TranslationBlock *tb, **ptb1;
 268     unsigned int h;
 269     tb_page_addr_t phys_pc, phys_page1;
 270     target_ulong virt_page2;
 271
 272     tcg_ctx.tb_ctx.tb_invalidated_flag = 0;
 273
 274     /* find translated block using physical mappings */
 275     phys_pc = get_page_addr_code(env, pc);
 276     phys_page1 = phys_pc & TARGET_PAGE_MASK;
 277     h = tb_phys_hash_func(phys_pc);
 278     ptb1 = &tcg_ctx.tb_ctx.tb_phys_hash[h];
 279     for(;;) {
 280         tb = *ptb1;
 281         if (!tb) {
 282             return NULL;
 283         }
 284         if (tb->pc == pc &&
 285             tb->page_addr[0] == phys_page1 &&
 286             tb->cs_base == cs_base &&
 287             tb->flags == flags) {
 288             /* check next page if needed */
 289             if (tb->page_addr[1] != -1) {
 290                 tb_page_addr_t phys_page2;
 291
 292                 virt_page2 = (pc & TARGET_PAGE_MASK) +
 293                     TARGET_PAGE_SIZE;
 294                 phys_page2 = get_page_addr_code(env, virt_page2);
 295                 if (tb->page_addr[1] == phys_page2) {
 296                     break;
 297                 }
 298             } else {
 299                 break;
 300             }
 301         }
 302         ptb1 = &tb->phys_hash_next;
 303     }
 304
 305     /* Move the TB to the head of the list */
 306     *ptb1 = tb->phys_hash_next;
 307     tb->phys_hash_next = tcg_ctx.tb_ctx.tb_phys_hash[h];
 308     tcg_ctx.tb_ctx.tb_phys_hash[h] = tb;
 309     return tb;
 310 }
 311
 312 static TranslationBlock *tb_find_slow(CPUState *cpu,
 313                                       target_ulong pc,
 314                                       target_ulong cs_base,
 315                                       uint64_t flags)
 316 {
 317     TranslationBlock *tb;
 318
 319     tb = tb_find_physical(cpu, pc, cs_base, flags);
 320     if (tb) {
 321         goto found;
 322     }
 323
 324 #ifdef CONFIG_USER_ONLY
 325     /* mmap_lock is needed by tb_gen_code, and mmap_lock must be
 326      * taken outside tb_lock.  Since we're momentarily dropping
 327      * tb_lock, there's a chance that our desired tb has been
 328      * translated.
 329      */
 330     tb_unlock();
 331     mmap_lock();
 332     tb_lock();
 333     tb = tb_find_physical(cpu, pc, cs_base, flags);
 334     if (tb) {
 335         mmap_unlock();
 336         goto found;
 337     }
 338 #endif
 339
 340     /* if no translated code available, then translate it now */
 341     tb = tb_gen_code(cpu, pc, cs_base, flags, 0);
 342
 343 #ifdef CONFIG_USER_ONLY
 344     mmap_unlock();
 345 #endif
 346
 347 found:
 348     /* we add the TB in the virtual pc hash table */
 349     cpu->tb_jmp_cache[tb_jmp_cache_hash_func(pc)] = tb;
 350     return tb;
 351 }
 352
 353 static inline TranslationBlock *tb_find_fast(CPUState *cpu)
 354 {
 355     CPUArchState *env = (CPUArchState *)cpu->env_ptr;
 356     TranslationBlock *tb;
 357     target_ulong cs_base, pc;
 358     int flags;
 359
 360     /* we record a subset of the CPU state. It will
 361        always be the same before a given translated block
 362        is executed. */
 363     cpu_get_tb_cpu_state(env, &pc, &cs_base, &flags);
 364     tb = cpu->tb_jmp_cache[tb_jmp_cache_hash_func(pc)];
 365     if (unlikely(!tb || tb->pc != pc || tb->cs_base != cs_base ||
 366                  tb->flags != flags)) {
 367         tb = tb_find_slow(cpu, pc, cs_base, flags);
 368     }
 369     return tb;
 370 }
 371
 372 static void cpu_handle_debug_exception(CPUState *cpu)
 373 {
 374     CPUClass *cc = CPU_GET_CLASS(cpu);
 375     CPUWatchpoint *wp;
 376
 377     if (!cpu->watchpoint_hit) {
 378         QTAILQ_FOREACH(wp, &cpu->watchpoints, entry) {
 379             wp->flags &= ~BP_WATCHPOINT_HIT;
 380         }
 381     }
 382
 383     cc->debug_excp_handler(cpu);
 384 }
 385
 386 /* main execution loop */
 387
 388 bool exit_request;
 389 CPUState *tcg_current_cpu;
 390
 391 int cpu_exec(CPUState *cpu)
 392 {
 393     CPUClass *cc = CPU_GET_CLASS(cpu);
 394 #ifdef TARGET_I386
 395     X86CPU *x86_cpu = X86_CPU(cpu);
 396     CPUArchState *env = &x86_cpu->env;
 397 #endif
 398     int ret, interrupt_request;
 399     TranslationBlock *tb;
 400     uint8_t *tc_ptr;
 401     uintptr_t next_tb;
 402     SyncClocks sc;
 403
 404     if (cpu->halted) {
 405         if (!cpu_has_work(cpu)) {
 406             return EXCP_HALTED;
 407         }
 408
 409         cpu->halted = 0;
 410     }
 411
 412     current_cpu = cpu;
 413     atomic_mb_set(&tcg_current_cpu, cpu);
 414     rcu_read_lock();
 415
 416     if (unlikely(atomic_mb_read(&exit_request))) {
 417         cpu->exit_request = 1;
 418     }
 419
 420     cc->cpu_exec_enter(cpu);
 421
 422     /* Calculate difference between guest clock and host clock.
 423      * This delay includes the delay of the last cycle, so
 424      * what we have to do is sleep until it is 0. As for the
 425      * advance/delay we gain here, we try to fix it next time.
 426      */
 427     init_delay_params(&sc, cpu);
 428
 429     /* prepare setjmp context for exception handling */
 430     for(;;) {
 431         if (sigsetjmp(cpu->jmp_env, 0) == 0) {
 432             /* if an exception is pending, we execute it here */
 433             if (cpu->exception_index >= 0) {
 434                 if (cpu->exception_index >= EXCP_INTERRUPT) {
 435                     /* exit request from the cpu execution loop */
 436                     ret = cpu->exception_index;
 437                     if (ret == EXCP_DEBUG) {
 438                         cpu_handle_debug_exception(cpu);
 439                     }
 440                     cpu->exception_index = -1;
 441                     break;
 442                 } else {
 443 #if defined(CONFIG_USER_ONLY)
 444                     /* if user mode only, we simulate a fake exception
 445                        which will be handled outside the cpu execution
 446                        loop */
 447 #if defined(TARGET_I386)
 448                     cc->do_interrupt(cpu);
 449 #endif
 450                     ret = cpu->exception_index;
 451                     cpu->exception_index = -1;
 452                     break;
 453 #else
 454                     cc->do_interrupt(cpu);
 455                     cpu->exception_index = -1;
 456 #endif
 457                 }
 458             }
 459
 460             next_tb = 0; /* force lookup of first TB */
 461             for(;;) {
 462                 interrupt_request = cpu->interrupt_request;
 463                 if (unlikely(interrupt_request)) {
 464                     if (unlikely(cpu->singlestep_enabled & SSTEP_NOIRQ)) {
 465                         /* Mask out external interrupts for this step. */
 466                         interrupt_request &= ~CPU_INTERRUPT_SSTEP_MASK;
 467                     }
 468                     if (interrupt_request & CPU_INTERRUPT_DEBUG) {
 469                         cpu->interrupt_request &= ~CPU_INTERRUPT_DEBUG;
 470                         cpu->exception_index = EXCP_DEBUG;
 471                         cpu_loop_exit(cpu);
 472                     }
 473                     if (interrupt_request & CPU_INTERRUPT_HALT) {
 474                         cpu->interrupt_request &= ~CPU_INTERRUPT_HALT;
 475                         cpu->halted = 1;
 476                         cpu->exception_index = EXCP_HLT;
 477                         cpu_loop_exit(cpu);
 478                     }
 479 #if defined(TARGET_I386)
 480                     if (interrupt_request & CPU_INTERRUPT_INIT) {
 481                         cpu_svm_check_intercept_param(env, SVM_EXIT_INIT, 0);
 482                         do_cpu_init(x86_cpu);
 483                         cpu->exception_index = EXCP_HALTED;
 484                         cpu_loop_exit(cpu);
 485                     }
 486 #else
 487                     if (interrupt_request & CPU_INTERRUPT_RESET) {
 488                         cpu_reset(cpu);
 489                     }
 490 #endif
 491                     /* The target hook has 3 exit conditions:
 492                        False when the interrupt isn't processed,
 493                        True when it is, and we should restart on a new TB,
 494                        and via longjmp via cpu_loop_exit.  */
 495                     if (cc->cpu_exec_interrupt(cpu, interrupt_request)) {
 496                         next_tb = 0;
 497                     }
 498                     /* Don't use the cached interrupt_request value,
 499                        do_interrupt may have updated the EXITTB flag. */
 500                     if (cpu->interrupt_request & CPU_INTERRUPT_EXITTB) {
 501                         cpu->interrupt_request &= ~CPU_INTERRUPT_EXITTB;
 502                         /* ensure that no TB jump will be modified as
 503                            the program flow was changed */
 504                         next_tb = 0;
 505                     }
 506                 }
 507                 if (unlikely(cpu->exit_request)) {
 508                     cpu->exit_request = 0;
 509                     cpu->exception_index = EXCP_INTERRUPT;
 510                     cpu_loop_exit(cpu);
 511                 }
 512                 tb_lock();
 513                 tb = tb_find_fast(cpu);
 514                 /* Note: we do it here to avoid a gcc bug on Mac OS X when
 515                    doing it in tb_find_slow */
 516                 if (tcg_ctx.tb_ctx.tb_invalidated_flag) {
 517                     /* as some TB could have been invalidated because
 518                        of memory exceptions while generating the code, we
 519                        must recompute the hash index here */
 520                     next_tb = 0;
 521                     tcg_ctx.tb_ctx.tb_invalidated_flag = 0;
 522                 }
 523                 if (qemu_loglevel_mask(CPU_LOG_EXEC)) {
 524                     qemu_log("Trace %p [" TARGET_FMT_lx "] %s\n",
 525                              tb->tc_ptr, tb->pc, lookup_symbol(tb->pc));
 526                 }
 527                 /* see if we can patch the calling TB. When the TB
 528                    spans two pages, we cannot safely do a direct
 529                    jump. */
 530                 if (next_tb != 0 && tb->page_addr[1] == -1) {
 531                     tb_add_jump((TranslationBlock *)(next_tb & ~TB_EXIT_MASK),
 532                                 next_tb & TB_EXIT_MASK, tb);
 533                 }
 534                 tb_unlock();
 535                 if (likely(!cpu->exit_request)) {
 536                     trace_exec_tb(tb, tb->pc);
 537                     tc_ptr = tb->tc_ptr;
 538                     /* execute the generated code */
 539                     cpu->current_tb = tb;
 540                     next_tb = cpu_tb_exec(cpu, tc_ptr);
 541                     cpu->current_tb = NULL;
 542                     switch (next_tb & TB_EXIT_MASK) {
 543                     case TB_EXIT_REQUESTED:
 544                         /* Something asked us to stop executing
 545                          * chained TBs; just continue round the main
 546                          * loop. Whatever requested the exit will also
 547                          * have set something else (eg exit_request or
 548                          * interrupt_request) which we will handle
 549                          * next time around the loop.  But we need to
 550                          * ensure the tcg_exit_req read in generated code
 551                          * comes before the next read of cpu->exit_request
 552                          * or cpu->interrupt_request.
 553                          */
 554                         smp_rmb();
 555                         next_tb = 0;
 556                         break;
 557                     case TB_EXIT_ICOUNT_EXPIRED:
 558                     {
 559                         /* Instruction counter expired.  */
 560                         int insns_left = cpu->icount_decr.u32;
 561                         if (cpu->icount_extra && insns_left >= 0) {
 562                             /* Refill decrementer and continue execution.  */
 563                             cpu->icount_extra += insns_left;
 564                             insns_left = MIN(0xffff, cpu->icount_extra);
 565                             cpu->icount_extra -= insns_left;
 566                             cpu->icount_decr.u16.low = insns_left;
 567                         } else {
 568                             if (insns_left > 0) {
 569                                 /* Execute remaining instructions.  */
 570                                 tb = (TranslationBlock *)(next_tb & ~TB_EXIT_MASK);
 571                                 cpu_exec_nocache(cpu, insns_left, tb);
 572                                 align_clocks(&sc, cpu);
 573                             }
 574                             cpu->exception_index = EXCP_INTERRUPT;
 575                             next_tb = 0;
 576                             cpu_loop_exit(cpu);
 577                         }
 578                         break;
 579                     }
 580                     default:
 581                         break;
 582                     }
 583                 }
 584                 /* Try to align the host and virtual clocks
 585                    if the guest is in advance */
 586                 align_clocks(&sc, cpu);
 587                 /* reset soft MMU for next block (it can currently
 588                    only be set by a memory fault) */
 589             } /* for(;;) */
 590         } else {
 591             /* Reload env after longjmp - the compiler may have smashed all
 592              * local variables as longjmp is marked 'noreturn'. */
 593             cpu = current_cpu;
 594             cc = CPU_GET_CLASS(cpu);
 595             cpu->can_do_io = 1;
 596 #ifdef TARGET_I386
 597             x86_cpu = X86_CPU(cpu);
 598             env = &x86_cpu->env;
 599 #endif
 600             tb_lock_reset();
 601         }
 602     } /* for(;;) */
 603
 604     cc->cpu_exec_exit(cpu);
 605     rcu_read_unlock();
 606
 607     /* fail safe : never use current_cpu outside cpu_exec() */
 608     current_cpu = NULL;
 609
 610     /* Does not need atomic_mb_set because a spurious wakeup is okay.  */
 611     atomic_set(&tcg_current_cpu, NULL);
 612     return ret;
 613 }