cpu-exec: Move TB chaining into tb_find_fast()
[sdk/emulator/qemu.git] / 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 "qemu/osdep.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 #include "exec/log.h"
31 #if defined(TARGET_I386) && !defined(CONFIG_USER_ONLY)
32 #include "hw/i386/apic.h"
33 #endif
34 #include "sysemu/replay.h"
35
36 /* -icount align implementation. */
37
38 typedef struct SyncClocks {
39     int64_t diff_clk;
40     int64_t last_cpu_icount;
41     int64_t realtime_clock;
42 } SyncClocks;
43
44 #if !defined(CONFIG_USER_ONLY)
45 /* Allow the guest to have a max 3ms advance.
46  * The difference between the 2 clocks could therefore
47  * oscillate around 0.
48  */
49 #define VM_CLOCK_ADVANCE 3000000
50 #define THRESHOLD_REDUCE 1.5
51 #define MAX_DELAY_PRINT_RATE 2000000000LL
52 #define MAX_NB_PRINTS 100
53
54 static void align_clocks(SyncClocks *sc, const CPUState *cpu)
55 {
56     int64_t cpu_icount;
57
58     if (!icount_align_option) {
59         return;
60     }
61
62     cpu_icount = cpu->icount_extra + cpu->icount_decr.u16.low;
63     sc->diff_clk += cpu_icount_to_ns(sc->last_cpu_icount - cpu_icount);
64     sc->last_cpu_icount = cpu_icount;
65
66     if (sc->diff_clk > VM_CLOCK_ADVANCE) {
67 #ifndef _WIN32
68         struct timespec sleep_delay, rem_delay;
69         sleep_delay.tv_sec = sc->diff_clk / 1000000000LL;
70         sleep_delay.tv_nsec = sc->diff_clk % 1000000000LL;
71         if (nanosleep(&sleep_delay, &rem_delay) < 0) {
72             sc->diff_clk = rem_delay.tv_sec * 1000000000LL + rem_delay.tv_nsec;
73         } else {
74             sc->diff_clk = 0;
75         }
76 #else
77         Sleep(sc->diff_clk / SCALE_MS);
78         sc->diff_clk = 0;
79 #endif
80     }
81 }
82
83 static void print_delay(const SyncClocks *sc)
84 {
85     static float threshold_delay;
86     static int64_t last_realtime_clock;
87     static int nb_prints;
88
89     if (icount_align_option &&
90         sc->realtime_clock - last_realtime_clock >= MAX_DELAY_PRINT_RATE &&
91         nb_prints < MAX_NB_PRINTS) {
92         if ((-sc->diff_clk / (float)1000000000LL > threshold_delay) ||
93             (-sc->diff_clk / (float)1000000000LL <
94              (threshold_delay - THRESHOLD_REDUCE))) {
95             threshold_delay = (-sc->diff_clk / 1000000000LL) + 1;
96             printf("Warning: The guest is now late by %.1f to %.1f seconds\n",
97                    threshold_delay - 1,
98                    threshold_delay);
99             nb_prints++;
100             last_realtime_clock = sc->realtime_clock;
101         }
102     }
103 }
104
105 static void init_delay_params(SyncClocks *sc,
106                               const CPUState *cpu)
107 {
108     if (!icount_align_option) {
109         return;
110     }
111     sc->realtime_clock = qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL_RT);
112     sc->diff_clk = qemu_clock_get_ns(QEMU_CLOCK_VIRTUAL) - sc->realtime_clock;
113     sc->last_cpu_icount = cpu->icount_extra + cpu->icount_decr.u16.low;
114     if (sc->diff_clk < max_delay) {
115         max_delay = sc->diff_clk;
116     }
117     if (sc->diff_clk > max_advance) {
118         max_advance = sc->diff_clk;
119     }
120
121     /* Print every 2s max if the guest is late. We limit the number
122        of printed messages to NB_PRINT_MAX(currently 100) */
123     print_delay(sc);
124 }
125 #else
126 static void align_clocks(SyncClocks *sc, const CPUState *cpu)
127 {
128 }
129
130 static void init_delay_params(SyncClocks *sc, const CPUState *cpu)
131 {
132 }
133 #endif /* CONFIG USER ONLY */
134
135 /* Execute a TB, and fix up the CPU state afterwards if necessary */
136 static inline tcg_target_ulong cpu_tb_exec(CPUState *cpu, TranslationBlock *itb)
137 {
138     CPUArchState *env = cpu->env_ptr;
139     uintptr_t ret;
140     TranslationBlock *last_tb;
141     int tb_exit;
142     uint8_t *tb_ptr = itb->tc_ptr;
143
144     qemu_log_mask_and_addr(CPU_LOG_EXEC, itb->pc,
145                            "Trace %p [" TARGET_FMT_lx "] %s\n",
146                            itb->tc_ptr, itb->pc, lookup_symbol(itb->pc));
147
148 #if defined(DEBUG_DISAS)
149     if (qemu_loglevel_mask(CPU_LOG_TB_CPU)) {
150 #if defined(TARGET_I386)
151         log_cpu_state(cpu, CPU_DUMP_CCOP);
152 #elif defined(TARGET_M68K)
153         /* ??? Should not modify env state for dumping.  */
154         cpu_m68k_flush_flags(env, env->cc_op);
155         env->cc_op = CC_OP_FLAGS;
156         env->sr = (env->sr & 0xffe0) | env->cc_dest | (env->cc_x << 4);
157         log_cpu_state(cpu, 0);
158 #else
159         log_cpu_state(cpu, 0);
160 #endif
161     }
162 #endif /* DEBUG_DISAS */
163
164     cpu->can_do_io = !use_icount;
165     ret = tcg_qemu_tb_exec(env, tb_ptr);
166     cpu->can_do_io = 1;
167     last_tb = (TranslationBlock *)(ret & ~TB_EXIT_MASK);
168     tb_exit = ret & TB_EXIT_MASK;
169     trace_exec_tb_exit(last_tb, tb_exit);
170
171     if (tb_exit > TB_EXIT_IDX1) {
172         /* We didn't start executing this TB (eg because the instruction
173          * counter hit zero); we must restore the guest PC to the address
174          * of the start of the TB.
175          */
176         CPUClass *cc = CPU_GET_CLASS(cpu);
177         qemu_log_mask_and_addr(CPU_LOG_EXEC, last_tb->pc,
178                                "Stopped execution of TB chain before %p ["
179                                TARGET_FMT_lx "] %s\n",
180                                last_tb->tc_ptr, last_tb->pc,
181                                lookup_symbol(last_tb->pc));
182         if (cc->synchronize_from_tb) {
183             cc->synchronize_from_tb(cpu, last_tb);
184         } else {
185             assert(cc->set_pc);
186             cc->set_pc(cpu, last_tb->pc);
187         }
188     }
189     if (tb_exit == TB_EXIT_REQUESTED) {
190         /* We were asked to stop executing TBs (probably a pending
191          * interrupt. We've now stopped, so clear the flag.
192          */
193         cpu->tcg_exit_req = 0;
194     }
195     return ret;
196 }
197
198 #ifndef CONFIG_USER_ONLY
199 /* Execute the code without caching the generated code. An interpreter
200    could be used if available. */
201 static void cpu_exec_nocache(CPUState *cpu, int max_cycles,
202                              TranslationBlock *orig_tb, bool ignore_icount)
203 {
204     TranslationBlock *tb;
205     bool old_tb_flushed;
206
207     /* Should never happen.
208        We only end up here when an existing TB is too long.  */
209     if (max_cycles > CF_COUNT_MASK)
210         max_cycles = CF_COUNT_MASK;
211
212     old_tb_flushed = cpu->tb_flushed;
213     cpu->tb_flushed = false;
214     tb = tb_gen_code(cpu, orig_tb->pc, orig_tb->cs_base, orig_tb->flags,
215                      max_cycles | CF_NOCACHE
216                          | (ignore_icount ? CF_IGNORE_ICOUNT : 0));
217     tb->orig_tb = cpu->tb_flushed ? NULL : orig_tb;
218     cpu->tb_flushed |= old_tb_flushed;
219     cpu->current_tb = tb;
220     /* execute the generated code */
221     trace_exec_tb_nocache(tb, tb->pc);
222     cpu_tb_exec(cpu, tb);
223     cpu->current_tb = NULL;
224     tb_phys_invalidate(tb, -1);
225     tb_free(tb);
226 }
227 #endif
228
229 static TranslationBlock *tb_find_physical(CPUState *cpu,
230                                           target_ulong pc,
231                                           target_ulong cs_base,
232                                           uint32_t flags)
233 {
234     CPUArchState *env = (CPUArchState *)cpu->env_ptr;
235     TranslationBlock *tb, **tb_hash_head, **ptb1;
236     unsigned int h;
237     tb_page_addr_t phys_pc, phys_page1;
238
239     /* find translated block using physical mappings */
240     phys_pc = get_page_addr_code(env, pc);
241     phys_page1 = phys_pc & TARGET_PAGE_MASK;
242     h = tb_phys_hash_func(phys_pc);
243
244     /* Start at head of the hash entry */
245     ptb1 = tb_hash_head = &tcg_ctx.tb_ctx.tb_phys_hash[h];
246     tb = *ptb1;
247
248     while (tb) {
249         if (tb->pc == pc &&
250             tb->page_addr[0] == phys_page1 &&
251             tb->cs_base == cs_base &&
252             tb->flags == flags) {
253
254             if (tb->page_addr[1] == -1) {
255                 /* done, we have a match */
256                 break;
257             } else {
258                 /* check next page if needed */
259                 target_ulong virt_page2 = (pc & TARGET_PAGE_MASK) +
260                                           TARGET_PAGE_SIZE;
261                 tb_page_addr_t phys_page2 = get_page_addr_code(env, virt_page2);
262
263                 if (tb->page_addr[1] == phys_page2) {
264                     break;
265                 }
266             }
267         }
268
269         ptb1 = &tb->phys_hash_next;
270         tb = *ptb1;
271     }
272
273     if (tb) {
274         /* Move the TB to the head of the list */
275         *ptb1 = tb->phys_hash_next;
276         tb->phys_hash_next = *tb_hash_head;
277         *tb_hash_head = tb;
278     }
279     return tb;
280 }
281
282 static TranslationBlock *tb_find_slow(CPUState *cpu,
283                                       target_ulong pc,
284                                       target_ulong cs_base,
285                                       uint32_t flags)
286 {
287     TranslationBlock *tb;
288
289     tb = tb_find_physical(cpu, pc, cs_base, flags);
290     if (tb) {
291         goto found;
292     }
293
294 #ifdef CONFIG_USER_ONLY
295     /* mmap_lock is needed by tb_gen_code, and mmap_lock must be
296      * taken outside tb_lock.  Since we're momentarily dropping
297      * tb_lock, there's a chance that our desired tb has been
298      * translated.
299      */
300     tb_unlock();
301     mmap_lock();
302     tb_lock();
303     tb = tb_find_physical(cpu, pc, cs_base, flags);
304     if (tb) {
305         mmap_unlock();
306         goto found;
307     }
308 #endif
309
310     /* if no translated code available, then translate it now */
311     tb = tb_gen_code(cpu, pc, cs_base, flags, 0);
312
313 #ifdef CONFIG_USER_ONLY
314     mmap_unlock();
315 #endif
316
317 found:
318     /* we add the TB in the virtual pc hash table */
319     cpu->tb_jmp_cache[tb_jmp_cache_hash_func(pc)] = tb;
320     return tb;
321 }
322
323 static inline TranslationBlock *tb_find_fast(CPUState *cpu,
324                                              TranslationBlock **last_tb,
325                                              int tb_exit)
326 {
327     CPUArchState *env = (CPUArchState *)cpu->env_ptr;
328     TranslationBlock *tb;
329     target_ulong cs_base, pc;
330     uint32_t flags;
331
332     /* we record a subset of the CPU state. It will
333        always be the same before a given translated block
334        is executed. */
335     cpu_get_tb_cpu_state(env, &pc, &cs_base, &flags);
336     tb_lock();
337     tb = cpu->tb_jmp_cache[tb_jmp_cache_hash_func(pc)];
338     if (unlikely(!tb || tb->pc != pc || tb->cs_base != cs_base ||
339                  tb->flags != flags)) {
340         tb = tb_find_slow(cpu, pc, cs_base, flags);
341     }
342     if (cpu->tb_flushed) {
343         /* Ensure that no TB jump will be modified as the
344          * translation buffer has been flushed.
345          */
346         *last_tb = NULL;
347         cpu->tb_flushed = false;
348     }
349     /* See if we can patch the calling TB. */
350     if (*last_tb && !qemu_loglevel_mask(CPU_LOG_TB_NOCHAIN)) {
351         tb_add_jump(*last_tb, tb_exit, tb);
352     }
353     tb_unlock();
354     return tb;
355 }
356
357 static void cpu_handle_debug_exception(CPUState *cpu)
358 {
359     CPUClass *cc = CPU_GET_CLASS(cpu);
360     CPUWatchpoint *wp;
361
362     if (!cpu->watchpoint_hit) {
363         QTAILQ_FOREACH(wp, &cpu->watchpoints, entry) {
364             wp->flags &= ~BP_WATCHPOINT_HIT;
365         }
366     }
367
368     cc->debug_excp_handler(cpu);
369 }
370
371 /* main execution loop */
372
373 int cpu_exec(CPUState *cpu)
374 {
375     CPUClass *cc = CPU_GET_CLASS(cpu);
376 #ifdef TARGET_I386
377     X86CPU *x86_cpu = X86_CPU(cpu);
378     CPUArchState *env = &x86_cpu->env;
379 #endif
380     int ret, interrupt_request;
381     TranslationBlock *tb, *last_tb;
382     int tb_exit = 0;
383     SyncClocks sc;
384
385     /* replay_interrupt may need current_cpu */
386     current_cpu = cpu;
387
388     if (cpu->halted) {
389 #if defined(TARGET_I386) && !defined(CONFIG_USER_ONLY)
390         if ((cpu->interrupt_request & CPU_INTERRUPT_POLL)
391             && replay_interrupt()) {
392             apic_poll_irq(x86_cpu->apic_state);
393             cpu_reset_interrupt(cpu, CPU_INTERRUPT_POLL);
394         }
395 #endif
396         if (!cpu_has_work(cpu)) {
397             current_cpu = NULL;
398             return EXCP_HALTED;
399         }
400
401         cpu->halted = 0;
402     }
403
404     atomic_mb_set(&tcg_current_cpu, cpu);
405     rcu_read_lock();
406
407     if (unlikely(atomic_mb_read(&exit_request))) {
408         cpu->exit_request = 1;
409     }
410
411     cc->cpu_exec_enter(cpu);
412
413     /* Calculate difference between guest clock and host clock.
414      * This delay includes the delay of the last cycle, so
415      * what we have to do is sleep until it is 0. As for the
416      * advance/delay we gain here, we try to fix it next time.
417      */
418     init_delay_params(&sc, cpu);
419
420     /* prepare setjmp context for exception handling */
421     for(;;) {
422         if (sigsetjmp(cpu->jmp_env, 0) == 0) {
423             /* if an exception is pending, we execute it here */
424             if (cpu->exception_index >= 0) {
425                 if (cpu->exception_index >= EXCP_INTERRUPT) {
426                     /* exit request from the cpu execution loop */
427                     ret = cpu->exception_index;
428                     if (ret == EXCP_DEBUG) {
429                         cpu_handle_debug_exception(cpu);
430                     }
431                     cpu->exception_index = -1;
432                     break;
433                 } else {
434 #if defined(CONFIG_USER_ONLY)
435                     /* if user mode only, we simulate a fake exception
436                        which will be handled outside the cpu execution
437                        loop */
438 #if defined(TARGET_I386)
439                     cc->do_interrupt(cpu);
440 #endif
441                     ret = cpu->exception_index;
442                     cpu->exception_index = -1;
443                     break;
444 #else
445                     if (replay_exception()) {
446                         cc->do_interrupt(cpu);
447                         cpu->exception_index = -1;
448                     } else if (!replay_has_interrupt()) {
449                         /* give a chance to iothread in replay mode */
450                         ret = EXCP_INTERRUPT;
451                         break;
452                     }
453 #endif
454                 }
455 #ifndef CONFIG_USER_ONLY
456             } else if (replay_has_exception()
457                        && cpu->icount_decr.u16.low + cpu->icount_extra == 0) {
458                 /* try to cause an exception pending in the log */
459                 last_tb = NULL; /* Avoid chaining TBs */
460                 cpu_exec_nocache(cpu, 1, tb_find_fast(cpu, &last_tb, 0), true);
461                 ret = -1;
462                 break;
463 #endif
464             }
465
466             last_tb = NULL; /* forget the last executed TB after exception */
467             cpu->tb_flushed = false; /* reset before first TB lookup */
468             for(;;) {
469                 interrupt_request = cpu->interrupt_request;
470                 if (unlikely(interrupt_request)) {
471                     if (unlikely(cpu->singlestep_enabled & SSTEP_NOIRQ)) {
472                         /* Mask out external interrupts for this step. */
473                         interrupt_request &= ~CPU_INTERRUPT_SSTEP_MASK;
474                     }
475                     if (interrupt_request & CPU_INTERRUPT_DEBUG) {
476                         cpu->interrupt_request &= ~CPU_INTERRUPT_DEBUG;
477                         cpu->exception_index = EXCP_DEBUG;
478                         cpu_loop_exit(cpu);
479                     }
480                     if (replay_mode == REPLAY_MODE_PLAY
481                         && !replay_has_interrupt()) {
482                         /* Do nothing */
483                     } else if (interrupt_request & CPU_INTERRUPT_HALT) {
484                         replay_interrupt();
485                         cpu->interrupt_request &= ~CPU_INTERRUPT_HALT;
486                         cpu->halted = 1;
487                         cpu->exception_index = EXCP_HLT;
488                         cpu_loop_exit(cpu);
489                     }
490 #if defined(TARGET_I386)
491                     else if (interrupt_request & CPU_INTERRUPT_INIT) {
492                         replay_interrupt();
493                         cpu_svm_check_intercept_param(env, SVM_EXIT_INIT, 0);
494                         do_cpu_init(x86_cpu);
495                         cpu->exception_index = EXCP_HALTED;
496                         cpu_loop_exit(cpu);
497                     }
498 #else
499                     else if (interrupt_request & CPU_INTERRUPT_RESET) {
500                         replay_interrupt();
501                         cpu_reset(cpu);
502                         cpu_loop_exit(cpu);
503                     }
504 #endif
505                     /* The target hook has 3 exit conditions:
506                        False when the interrupt isn't processed,
507                        True when it is, and we should restart on a new TB,
508                        and via longjmp via cpu_loop_exit.  */
509                     else {
510                         replay_interrupt();
511                         if (cc->cpu_exec_interrupt(cpu, interrupt_request)) {
512                             last_tb = NULL;
513                         }
514                     }
515                     /* Don't use the cached interrupt_request value,
516                        do_interrupt may have updated the EXITTB flag. */
517                     if (cpu->interrupt_request & CPU_INTERRUPT_EXITTB) {
518                         cpu->interrupt_request &= ~CPU_INTERRUPT_EXITTB;
519                         /* ensure that no TB jump will be modified as
520                            the program flow was changed */
521                         last_tb = NULL;
522                     }
523                 }
524                 if (unlikely(cpu->exit_request
525                              || replay_has_interrupt())) {
526                     cpu->exit_request = 0;
527                     cpu->exception_index = EXCP_INTERRUPT;
528                     cpu_loop_exit(cpu);
529                 }
530                 tb = tb_find_fast(cpu, &last_tb, tb_exit);
531                 if (likely(!cpu->exit_request)) {
532                     uintptr_t ret;
533                     trace_exec_tb(tb, tb->pc);
534                     /* execute the generated code */
535                     cpu->current_tb = tb;
536                     ret = cpu_tb_exec(cpu, tb);
537                     cpu->current_tb = NULL;
538                     last_tb = (TranslationBlock *)(ret & ~TB_EXIT_MASK);
539                     tb_exit = ret & TB_EXIT_MASK;
540                     switch (tb_exit) {
541                     case TB_EXIT_REQUESTED:
542                         /* Something asked us to stop executing
543                          * chained TBs; just continue round the main
544                          * loop. Whatever requested the exit will also
545                          * have set something else (eg exit_request or
546                          * interrupt_request) which we will handle
547                          * next time around the loop.  But we need to
548                          * ensure the tcg_exit_req read in generated code
549                          * comes before the next read of cpu->exit_request
550                          * or cpu->interrupt_request.
551                          */
552                         smp_rmb();
553                         last_tb = NULL;
554                         break;
555                     case TB_EXIT_ICOUNT_EXPIRED:
556                     {
557                         /* Instruction counter expired.  */
558 #ifdef CONFIG_USER_ONLY
559                         abort();
560 #else
561                         int insns_left = cpu->icount_decr.u32;
562                         if (cpu->icount_extra && insns_left >= 0) {
563                             /* Refill decrementer and continue execution.  */
564                             cpu->icount_extra += insns_left;
565                             insns_left = MIN(0xffff, cpu->icount_extra);
566                             cpu->icount_extra -= insns_left;
567                             cpu->icount_decr.u16.low = insns_left;
568                         } else {
569                             if (insns_left > 0) {
570                                 /* Execute remaining instructions.  */
571                                 cpu_exec_nocache(cpu, insns_left,
572                                                  last_tb, false);
573                                 align_clocks(&sc, cpu);
574                             }
575                             cpu->exception_index = EXCP_INTERRUPT;
576                             last_tb = NULL;
577                             cpu_loop_exit(cpu);
578                         }
579                         break;
580 #endif
581                     }
582                     default:
583                         break;
584                     }
585                 }
586                 /* Try to align the host and virtual clocks
587                    if the guest is in advance */
588                 align_clocks(&sc, cpu);
589                 /* reset soft MMU for next block (it can currently
590                    only be set by a memory fault) */
591             } /* for(;;) */
592         } else {
593 #if defined(__clang__) || !QEMU_GNUC_PREREQ(4, 6)
594             /* Some compilers wrongly smash all local variables after
595              * siglongjmp. There were bug reports for gcc 4.5.0 and clang.
596              * Reload essential local variables here for those compilers.
597              * Newer versions of gcc would complain about this code (-Wclobbered). */
598             cpu = current_cpu;
599             cc = CPU_GET_CLASS(cpu);
600 #ifdef TARGET_I386
601             x86_cpu = X86_CPU(cpu);
602             env = &x86_cpu->env;
603 #endif
604 #else /* buggy compiler */
605             /* Assert that the compiler does not smash local variables. */
606             g_assert(cpu == current_cpu);
607             g_assert(cc == CPU_GET_CLASS(cpu));
608 #ifdef TARGET_I386
609             g_assert(x86_cpu == X86_CPU(cpu));
610             g_assert(env == &x86_cpu->env);
611 #endif
612 #endif /* buggy compiler */
613             cpu->can_do_io = 1;
614             tb_lock_reset();
615         }
616     } /* for(;;) */
617
618     cc->cpu_exec_exit(cpu);
619     rcu_read_unlock();
620
621     /* fail safe : never use current_cpu outside cpu_exec() */
622     current_cpu = NULL;
623
624     /* Does not need atomic_mb_set because a spurious wakeup is okay.  */
625     atomic_set(&tcg_current_cpu, NULL);
626     return ret;
627 }