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