xref: /qemu/accel/tcg/translate-all.c (revision cc1f4b34d011e908dcaf24721f1d5808e02ab0bd)
1 /*
2  *  Host code generation
3  *
4  *  Copyright (c) 2003 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 2.1 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 
20 #include "qemu/osdep.h"
21 
22 #include "trace.h"
23 #include "disas/disas.h"
24 #include "tcg/tcg.h"
25 #include "exec/mmap-lock.h"
26 #include "tb-internal.h"
27 #include "tlb-bounds.h"
28 #include "exec/tb-flush.h"
29 #include "qemu/cacheinfo.h"
30 #include "qemu/target-info.h"
31 #include "exec/log.h"
32 #include "exec/icount.h"
33 #include "accel/tcg/cpu-ops.h"
34 #include "tb-jmp-cache.h"
35 #include "tb-hash.h"
36 #include "tb-context.h"
37 #include "tb-internal.h"
38 #include "internal-common.h"
39 #include "tcg/perf.h"
40 #include "tcg/insn-start-words.h"
41 
42 TBContext tb_ctx;
43 
44 /*
45  * Encode VAL as a signed leb128 sequence at P.
46  * Return P incremented past the encoded value.
47  */
48 static uint8_t *encode_sleb128(uint8_t *p, int64_t val)
49 {
50     int more, byte;
51 
52     do {
53         byte = val & 0x7f;
54         val >>= 7;
55         more = !((val == 0 && (byte & 0x40) == 0)
56                  || (val == -1 && (byte & 0x40) != 0));
57         if (more) {
58             byte |= 0x80;
59         }
60         *p++ = byte;
61     } while (more);
62 
63     return p;
64 }
65 
66 /*
67  * Decode a signed leb128 sequence at *PP; increment *PP past the
68  * decoded value.  Return the decoded value.
69  */
70 static int64_t decode_sleb128(const uint8_t **pp)
71 {
72     const uint8_t *p = *pp;
73     int64_t val = 0;
74     int byte, shift = 0;
75 
76     do {
77         byte = *p++;
78         val |= (int64_t)(byte & 0x7f) << shift;
79         shift += 7;
80     } while (byte & 0x80);
81     if (shift < 64 && (byte & 0x40)) {
82         val |= -(int64_t)1 << shift;
83     }
84 
85     *pp = p;
86     return val;
87 }
88 
89 /* Encode the data collected about the instructions while compiling TB.
90    Place the data at BLOCK, and return the number of bytes consumed.
91 
92    The logical table consists of INSN_START_WORDS uint64_t's,
93    which come from the target's insn_start data, followed by a uintptr_t
94    which comes from the host pc of the end of the code implementing the insn.
95 
96    Each line of the table is encoded as sleb128 deltas from the previous
97    line.  The seed for the first line is { tb->pc, 0..., tb->tc.ptr }.
98    That is, the first column is seeded with the guest pc, the last column
99    with the host pc, and the middle columns with zeros.  */
100 
101 static int encode_search(TranslationBlock *tb, uint8_t *block)
102 {
103     uint8_t *highwater = tcg_ctx->code_gen_highwater;
104     uint64_t *insn_data = tcg_ctx->gen_insn_data;
105     uint16_t *insn_end_off = tcg_ctx->gen_insn_end_off;
106     uint8_t *p = block;
107     int i, j, n;
108 
109     for (i = 0, n = tb->icount; i < n; ++i) {
110         uint64_t prev, curr;
111 
112         for (j = 0; j < INSN_START_WORDS; ++j) {
113             if (i == 0) {
114                 prev = (!(tb_cflags(tb) & CF_PCREL) && j == 0 ? tb->pc : 0);
115             } else {
116                 prev = insn_data[(i - 1) * INSN_START_WORDS + j];
117             }
118             curr = insn_data[i * INSN_START_WORDS + j];
119             p = encode_sleb128(p, curr - prev);
120         }
121         prev = (i == 0 ? 0 : insn_end_off[i - 1]);
122         curr = insn_end_off[i];
123         p = encode_sleb128(p, curr - prev);
124 
125         /* Test for (pending) buffer overflow.  The assumption is that any
126            one row beginning below the high water mark cannot overrun
127            the buffer completely.  Thus we can test for overflow after
128            encoding a row without having to check during encoding.  */
129         if (unlikely(p > highwater)) {
130             return -1;
131         }
132     }
133 
134     return p - block;
135 }
136 
137 static int cpu_unwind_data_from_tb(TranslationBlock *tb, uintptr_t host_pc,
138                                    uint64_t *data)
139 {
140     uintptr_t iter_pc = (uintptr_t)tb->tc.ptr;
141     const uint8_t *p = tb->tc.ptr + tb->tc.size;
142     int i, j, num_insns = tb->icount;
143 
144     host_pc -= GETPC_ADJ;
145 
146     if (host_pc < iter_pc) {
147         return -1;
148     }
149 
150     memset(data, 0, sizeof(uint64_t) * INSN_START_WORDS);
151     if (!(tb_cflags(tb) & CF_PCREL)) {
152         data[0] = tb->pc;
153     }
154 
155     /*
156      * Reconstruct the stored insn data while looking for the point
157      * at which the end of the insn exceeds host_pc.
158      */
159     for (i = 0; i < num_insns; ++i) {
160         for (j = 0; j < INSN_START_WORDS; ++j) {
161             data[j] += decode_sleb128(&p);
162         }
163         iter_pc += decode_sleb128(&p);
164         if (iter_pc > host_pc) {
165             return num_insns - i;
166         }
167     }
168     return -1;
169 }
170 
171 /*
172  * The cpu state corresponding to 'host_pc' is restored in
173  * preparation for exiting the TB.
174  */
175 void cpu_restore_state_from_tb(CPUState *cpu, TranslationBlock *tb,
176                                uintptr_t host_pc)
177 {
178     uint64_t data[INSN_START_WORDS];
179     int insns_left = cpu_unwind_data_from_tb(tb, host_pc, data);
180 
181     if (insns_left < 0) {
182         return;
183     }
184 
185     if (tb_cflags(tb) & CF_USE_ICOUNT) {
186         assert(icount_enabled());
187         /*
188          * Reset the cycle counter to the start of the block and
189          * shift if to the number of actually executed instructions.
190          */
191         cpu->neg.icount_decr.u16.low += insns_left;
192     }
193 
194     cpu->cc->tcg_ops->restore_state_to_opc(cpu, tb, data);
195 }
196 
197 bool cpu_restore_state(CPUState *cpu, uintptr_t host_pc)
198 {
199     /*
200      * The host_pc has to be in the rx region of the code buffer.
201      * If it is not we will not be able to resolve it here.
202      * The two cases where host_pc will not be correct are:
203      *
204      *  - fault during translation (instruction fetch)
205      *  - fault from helper (not using GETPC() macro)
206      *
207      * Either way we need return early as we can't resolve it here.
208      */
209     if (in_code_gen_buffer((const void *)(host_pc - tcg_splitwx_diff))) {
210         TranslationBlock *tb = tcg_tb_lookup(host_pc);
211         if (tb) {
212             cpu_restore_state_from_tb(cpu, tb, host_pc);
213             return true;
214         }
215     }
216     return false;
217 }
218 
219 bool cpu_unwind_state_data(CPUState *cpu, uintptr_t host_pc, uint64_t *data)
220 {
221     if (in_code_gen_buffer((const void *)(host_pc - tcg_splitwx_diff))) {
222         TranslationBlock *tb = tcg_tb_lookup(host_pc);
223         if (tb) {
224             return cpu_unwind_data_from_tb(tb, host_pc, data) >= 0;
225         }
226     }
227     return false;
228 }
229 
230 void page_init(void)
231 {
232     page_table_config_init();
233 }
234 
235 /*
236  * Isolate the portion of code gen which can setjmp/longjmp.
237  * Return the size of the generated code, or negative on error.
238  */
239 static int setjmp_gen_code(CPUArchState *env, TranslationBlock *tb,
240                            vaddr pc, void *host_pc,
241                            int *max_insns, int64_t *ti)
242 {
243     int ret = sigsetjmp(tcg_ctx->jmp_trans, 0);
244     if (unlikely(ret != 0)) {
245         return ret;
246     }
247 
248     tcg_func_start(tcg_ctx);
249 
250     CPUState *cs = env_cpu(env);
251     tcg_ctx->cpu = cs;
252     cs->cc->tcg_ops->translate_code(cs, tb, max_insns, pc, host_pc);
253 
254     assert(tb->size != 0);
255     tcg_ctx->cpu = NULL;
256     *max_insns = tb->icount;
257 
258     return tcg_gen_code(tcg_ctx, tb, pc);
259 }
260 
261 /* Called with mmap_lock held for user mode emulation.  */
262 TranslationBlock *tb_gen_code(CPUState *cpu, TCGTBCPUState s)
263 {
264     CPUArchState *env = cpu_env(cpu);
265     TranslationBlock *tb, *existing_tb;
266     tb_page_addr_t phys_pc, phys_p2;
267     tcg_insn_unit *gen_code_buf;
268     int gen_code_size, search_size, max_insns;
269     int64_t ti;
270     void *host_pc;
271 
272     assert_memory_lock();
273     qemu_thread_jit_write();
274 
275     phys_pc = get_page_addr_code_hostp(env, s.pc, &host_pc);
276 
277     if (phys_pc == -1) {
278         /* Generate a one-shot TB with 1 insn in it */
279         s.cflags = (s.cflags & ~CF_COUNT_MASK) | 1;
280     }
281 
282     max_insns = s.cflags & CF_COUNT_MASK;
283     if (max_insns == 0) {
284         max_insns = TCG_MAX_INSNS;
285     }
286     QEMU_BUILD_BUG_ON(CF_COUNT_MASK + 1 != TCG_MAX_INSNS);
287 
288  buffer_overflow:
289     assert_no_pages_locked();
290     tb = tcg_tb_alloc(tcg_ctx);
291     if (unlikely(!tb)) {
292         /* flush must be done */
293         tb_flush(cpu);
294         mmap_unlock();
295         /* Make the execution loop process the flush as soon as possible.  */
296         cpu->exception_index = EXCP_INTERRUPT;
297         cpu_loop_exit(cpu);
298     }
299 
300     gen_code_buf = tcg_ctx->code_gen_ptr;
301     tb->tc.ptr = tcg_splitwx_to_rx(gen_code_buf);
302     if (!(s.cflags & CF_PCREL)) {
303         tb->pc = s.pc;
304     }
305     tb->cs_base = s.cs_base;
306     tb->flags = s.flags;
307     tb->cflags = s.cflags;
308     tb_set_page_addr0(tb, phys_pc);
309     tb_set_page_addr1(tb, -1);
310     if (phys_pc != -1) {
311         tb_lock_page0(phys_pc);
312     }
313 
314     tcg_ctx->gen_tb = tb;
315     tcg_ctx->addr_type = target_long_bits() == 32 ? TCG_TYPE_I32 : TCG_TYPE_I64;
316 #ifdef CONFIG_SOFTMMU
317     tcg_ctx->page_bits = TARGET_PAGE_BITS;
318     tcg_ctx->page_mask = TARGET_PAGE_MASK;
319     tcg_ctx->tlb_dyn_max_bits = CPU_TLB_DYN_MAX_BITS;
320 #endif
321     tcg_ctx->guest_mo = cpu->cc->tcg_ops->guest_default_memory_order;
322 
323  restart_translate:
324     trace_translate_block(tb, s.pc, tb->tc.ptr);
325 
326     gen_code_size = setjmp_gen_code(env, tb, s.pc, host_pc, &max_insns, &ti);
327     if (unlikely(gen_code_size < 0)) {
328         switch (gen_code_size) {
329         case -1:
330             /*
331              * Overflow of code_gen_buffer, or the current slice of it.
332              *
333              * TODO: We don't need to re-do tcg_ops->translate_code, nor
334              * should we re-do the tcg optimization currently hidden
335              * inside tcg_gen_code.  All that should be required is to
336              * flush the TBs, allocate a new TB, re-initialize it per
337              * above, and re-do the actual code generation.
338              */
339             qemu_log_mask(CPU_LOG_TB_OP | CPU_LOG_TB_OP_OPT,
340                           "Restarting code generation for "
341                           "code_gen_buffer overflow\n");
342             tb_unlock_pages(tb);
343             tcg_ctx->gen_tb = NULL;
344             goto buffer_overflow;
345 
346         case -2:
347             /*
348              * The code generated for the TranslationBlock is too large.
349              * The maximum size allowed by the unwind info is 64k.
350              * There may be stricter constraints from relocations
351              * in the tcg backend.
352              *
353              * Try again with half as many insns as we attempted this time.
354              * If a single insn overflows, there's a bug somewhere...
355              */
356             assert(max_insns > 1);
357             max_insns /= 2;
358             qemu_log_mask(CPU_LOG_TB_OP | CPU_LOG_TB_OP_OPT,
359                           "Restarting code generation with "
360                           "smaller translation block (max %d insns)\n",
361                           max_insns);
362 
363             /*
364              * The half-sized TB may not cross pages.
365              * TODO: Fix all targets that cross pages except with
366              * the first insn, at which point this can't be reached.
367              */
368             phys_p2 = tb_page_addr1(tb);
369             if (unlikely(phys_p2 != -1)) {
370                 tb_unlock_page1(phys_pc, phys_p2);
371                 tb_set_page_addr1(tb, -1);
372             }
373             goto restart_translate;
374 
375         case -3:
376             /*
377              * We had a page lock ordering problem.  In order to avoid
378              * deadlock we had to drop the lock on page0, which means
379              * that everything we translated so far is compromised.
380              * Restart with locks held on both pages.
381              */
382             qemu_log_mask(CPU_LOG_TB_OP | CPU_LOG_TB_OP_OPT,
383                           "Restarting code generation with re-locked pages");
384             goto restart_translate;
385 
386         default:
387             g_assert_not_reached();
388         }
389     }
390     tcg_ctx->gen_tb = NULL;
391 
392     search_size = encode_search(tb, (void *)gen_code_buf + gen_code_size);
393     if (unlikely(search_size < 0)) {
394         tb_unlock_pages(tb);
395         goto buffer_overflow;
396     }
397     tb->tc.size = gen_code_size;
398 
399     /*
400      * For CF_PCREL, attribute all executions of the generated code
401      * to its first mapping.
402      */
403     perf_report_code(s.pc, tb, tcg_splitwx_to_rx(gen_code_buf));
404 
405     if (qemu_loglevel_mask(CPU_LOG_TB_OUT_ASM) &&
406         qemu_log_in_addr_range(s.pc)) {
407         FILE *logfile = qemu_log_trylock();
408         if (logfile) {
409             int code_size, data_size;
410             const tcg_target_ulong *rx_data_gen_ptr;
411             size_t chunk_start;
412             int insn = 0;
413 
414             if (tcg_ctx->data_gen_ptr) {
415                 rx_data_gen_ptr = tcg_splitwx_to_rx(tcg_ctx->data_gen_ptr);
416                 code_size = (const void *)rx_data_gen_ptr - tb->tc.ptr;
417                 data_size = gen_code_size - code_size;
418             } else {
419                 rx_data_gen_ptr = 0;
420                 code_size = gen_code_size;
421                 data_size = 0;
422             }
423 
424             /* Dump header and the first instruction */
425             fprintf(logfile, "OUT: [size=%d]\n", gen_code_size);
426             fprintf(logfile,
427                     "  -- guest addr 0x%016" PRIx64 " + tb prologue\n",
428                     tcg_ctx->gen_insn_data[insn * INSN_START_WORDS]);
429             chunk_start = tcg_ctx->gen_insn_end_off[insn];
430             disas(logfile, tb->tc.ptr, chunk_start);
431 
432             /*
433              * Dump each instruction chunk, wrapping up empty chunks into
434              * the next instruction. The whole array is offset so the
435              * first entry is the beginning of the 2nd instruction.
436              */
437             while (insn < tb->icount) {
438                 size_t chunk_end = tcg_ctx->gen_insn_end_off[insn];
439                 if (chunk_end > chunk_start) {
440                     fprintf(logfile, "  -- guest addr 0x%016" PRIx64 "\n",
441                             tcg_ctx->gen_insn_data[insn * INSN_START_WORDS]);
442                     disas(logfile, tb->tc.ptr + chunk_start,
443                           chunk_end - chunk_start);
444                     chunk_start = chunk_end;
445                 }
446                 insn++;
447             }
448 
449             if (chunk_start < code_size) {
450                 fprintf(logfile, "  -- tb slow paths + alignment\n");
451                 disas(logfile, tb->tc.ptr + chunk_start,
452                       code_size - chunk_start);
453             }
454 
455             /* Finally dump any data we may have after the block */
456             if (data_size) {
457                 int i;
458                 fprintf(logfile, "  data: [size=%d]\n", data_size);
459                 for (i = 0; i < data_size / sizeof(tcg_target_ulong); i++) {
460                     if (sizeof(tcg_target_ulong) == 8) {
461                         fprintf(logfile,
462                                 "0x%08" PRIxPTR ":  .quad  0x%016" TCG_PRIlx "\n",
463                                 (uintptr_t)&rx_data_gen_ptr[i], rx_data_gen_ptr[i]);
464                     } else if (sizeof(tcg_target_ulong) == 4) {
465                         fprintf(logfile,
466                                 "0x%08" PRIxPTR ":  .long  0x%08" TCG_PRIlx "\n",
467                                 (uintptr_t)&rx_data_gen_ptr[i], rx_data_gen_ptr[i]);
468                     } else {
469                         qemu_build_not_reached();
470                     }
471                 }
472             }
473             fprintf(logfile, "\n");
474             qemu_log_unlock(logfile);
475         }
476     }
477 
478     qatomic_set(&tcg_ctx->code_gen_ptr, (void *)
479         ROUND_UP((uintptr_t)gen_code_buf + gen_code_size + search_size,
480                  CODE_GEN_ALIGN));
481 
482     /* init jump list */
483     qemu_spin_init(&tb->jmp_lock);
484     tb->jmp_list_head = (uintptr_t)NULL;
485     tb->jmp_list_next[0] = (uintptr_t)NULL;
486     tb->jmp_list_next[1] = (uintptr_t)NULL;
487     tb->jmp_dest[0] = (uintptr_t)NULL;
488     tb->jmp_dest[1] = (uintptr_t)NULL;
489 
490     /* init original jump addresses which have been set during tcg_gen_code() */
491     if (tb->jmp_reset_offset[0] != TB_JMP_OFFSET_INVALID) {
492         tb_reset_jump(tb, 0);
493     }
494     if (tb->jmp_reset_offset[1] != TB_JMP_OFFSET_INVALID) {
495         tb_reset_jump(tb, 1);
496     }
497 
498     /*
499      * Insert TB into the corresponding region tree before publishing it
500      * through QHT. Otherwise rewinding happened in the TB might fail to
501      * lookup itself using host PC.
502      */
503     tcg_tb_insert(tb);
504 
505     /*
506      * If the TB is not associated with a physical RAM page then it must be
507      * a temporary one-insn TB.
508      *
509      * Such TBs must be added to region trees in order to make sure that
510      * restore_state_to_opc() - which on some architectures is not limited to
511      * rewinding, but also affects exception handling! - is called when such a
512      * TB causes an exception.
513      *
514      * At the same time, temporary one-insn TBs must be executed at most once,
515      * because subsequent reads from, e.g., I/O memory may return different
516      * values. So return early before attempting to link to other TBs or add
517      * to the QHT.
518      */
519     if (tb_page_addr0(tb) == -1) {
520         assert_no_pages_locked();
521         return tb;
522     }
523 
524     /*
525      * No explicit memory barrier is required -- tb_link_page() makes the
526      * TB visible in a consistent state.
527      */
528     existing_tb = tb_link_page(tb);
529     assert_no_pages_locked();
530 
531     /* if the TB already exists, discard what we just translated */
532     if (unlikely(existing_tb != tb)) {
533         uintptr_t orig_aligned = (uintptr_t)gen_code_buf;
534 
535         orig_aligned -= ROUND_UP(sizeof(*tb), qemu_icache_linesize);
536         qatomic_set(&tcg_ctx->code_gen_ptr, (void *)orig_aligned);
537         tcg_tb_remove(tb);
538         return existing_tb;
539     }
540     return tb;
541 }
542 
543 /* user-mode: call with mmap_lock held */
544 void tb_check_watchpoint(CPUState *cpu, uintptr_t retaddr)
545 {
546     TranslationBlock *tb;
547 
548     assert_memory_lock();
549 
550     tb = tcg_tb_lookup(retaddr);
551     if (tb) {
552         /* We can use retranslation to find the PC.  */
553         cpu_restore_state_from_tb(cpu, tb, retaddr);
554         tb_phys_invalidate(tb, -1);
555     } else {
556         /* The exception probably happened in a helper.  The CPU state should
557            have been saved before calling it. Fetch the PC from there.  */
558         CPUArchState *env = cpu_env(cpu);
559         TCGTBCPUState s = cpu->cc->tcg_ops->get_tb_cpu_state(cpu);
560         tb_page_addr_t addr = get_page_addr_code(env, s.pc);
561 
562         if (addr != -1) {
563             tb_invalidate_phys_range(cpu, addr, addr);
564         }
565     }
566 }
567 
568 #ifndef CONFIG_USER_ONLY
569 /*
570  * In deterministic execution mode, instructions doing device I/Os
571  * must be at the end of the TB.
572  *
573  * Called by softmmu_template.h, with iothread mutex not held.
574  */
575 void cpu_io_recompile(CPUState *cpu, uintptr_t retaddr)
576 {
577     TranslationBlock *tb;
578     CPUClass *cc;
579     uint32_t n;
580 
581     tb = tcg_tb_lookup(retaddr);
582     if (!tb) {
583         cpu_abort(cpu, "cpu_io_recompile: could not find TB for pc=%p",
584                   (void *)retaddr);
585     }
586     cpu_restore_state_from_tb(cpu, tb, retaddr);
587 
588     /*
589      * Some guests must re-execute the branch when re-executing a delay
590      * slot instruction.  When this is the case, adjust icount and N
591      * to account for the re-execution of the branch.
592      */
593     n = 1;
594     cc = cpu->cc;
595     if (cc->tcg_ops->io_recompile_replay_branch &&
596         cc->tcg_ops->io_recompile_replay_branch(cpu, tb)) {
597         cpu->neg.icount_decr.u16.low++;
598         n = 2;
599     }
600 
601     /*
602      * Exit the loop and potentially generate a new TB executing the
603      * just the I/O insns. We also limit instrumentation to memory
604      * operations only (which execute after completion) so we don't
605      * double instrument the instruction. Also don't let an IRQ sneak
606      * in before we execute it.
607      */
608     cpu->cflags_next_tb = curr_cflags(cpu) | CF_MEMI_ONLY | CF_NOIRQ | n;
609 
610     if (qemu_loglevel_mask(CPU_LOG_EXEC)) {
611         vaddr pc = cpu->cc->get_pc(cpu);
612         if (qemu_log_in_addr_range(pc)) {
613             qemu_log("cpu_io_recompile: rewound execution of TB to %016"
614                      VADDR_PRIx "\n", pc);
615         }
616     }
617 
618     cpu_loop_exit_noexc(cpu);
619 }
620 
621 #endif /* CONFIG_USER_ONLY */
622 
623 /*
624  * Called by generic code at e.g. cpu reset after cpu creation,
625  * therefore we must be prepared to allocate the jump cache.
626  */
627 void tcg_flush_jmp_cache(CPUState *cpu)
628 {
629     CPUJumpCache *jc = cpu->tb_jmp_cache;
630 
631     /* During early initialization, the cache may not yet be allocated. */
632     if (unlikely(jc == NULL)) {
633         return;
634     }
635 
636     for (int i = 0; i < TB_JMP_CACHE_SIZE; i++) {
637         qatomic_set(&jc->array[i].tb, NULL);
638     }
639 }
640