xref: /openbmc/qemu/accel/tcg/translate-all.c (revision 8f968b6a)
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 #include "qemu-common.h"
22 
23 #define NO_CPU_IO_DEFS
24 #include "trace.h"
25 #include "disas/disas.h"
26 #include "exec/exec-all.h"
27 #include "tcg/tcg.h"
28 #if defined(CONFIG_USER_ONLY)
29 #include "qemu.h"
30 #if defined(__FreeBSD__) || defined(__FreeBSD_kernel__)
31 #include <sys/param.h>
32 #if __FreeBSD_version >= 700104
33 #define HAVE_KINFO_GETVMMAP
34 #define sigqueue sigqueue_freebsd  /* avoid redefinition */
35 #include <sys/proc.h>
36 #include <machine/profile.h>
37 #define _KERNEL
38 #include <sys/user.h>
39 #undef _KERNEL
40 #undef sigqueue
41 #include <libutil.h>
42 #endif
43 #endif
44 #else
45 #include "exec/ram_addr.h"
46 #endif
47 
48 #include "exec/cputlb.h"
49 #include "exec/translate-all.h"
50 #include "qemu/bitmap.h"
51 #include "qemu/qemu-print.h"
52 #include "qemu/timer.h"
53 #include "qemu/main-loop.h"
54 #include "exec/log.h"
55 #include "sysemu/cpus.h"
56 #include "sysemu/cpu-timers.h"
57 #include "sysemu/tcg.h"
58 #include "qapi/error.h"
59 #include "hw/core/tcg-cpu-ops.h"
60 #include "tb-hash.h"
61 #include "tb-context.h"
62 #include "internal.h"
63 
64 /* #define DEBUG_TB_INVALIDATE */
65 /* #define DEBUG_TB_FLUSH */
66 /* make various TB consistency checks */
67 /* #define DEBUG_TB_CHECK */
68 
69 #ifdef DEBUG_TB_INVALIDATE
70 #define DEBUG_TB_INVALIDATE_GATE 1
71 #else
72 #define DEBUG_TB_INVALIDATE_GATE 0
73 #endif
74 
75 #ifdef DEBUG_TB_FLUSH
76 #define DEBUG_TB_FLUSH_GATE 1
77 #else
78 #define DEBUG_TB_FLUSH_GATE 0
79 #endif
80 
81 #if !defined(CONFIG_USER_ONLY)
82 /* TB consistency checks only implemented for usermode emulation.  */
83 #undef DEBUG_TB_CHECK
84 #endif
85 
86 #ifdef DEBUG_TB_CHECK
87 #define DEBUG_TB_CHECK_GATE 1
88 #else
89 #define DEBUG_TB_CHECK_GATE 0
90 #endif
91 
92 /* Access to the various translations structures need to be serialised via locks
93  * for consistency.
94  * In user-mode emulation access to the memory related structures are protected
95  * with mmap_lock.
96  * In !user-mode we use per-page locks.
97  */
98 #ifdef CONFIG_SOFTMMU
99 #define assert_memory_lock()
100 #else
101 #define assert_memory_lock() tcg_debug_assert(have_mmap_lock())
102 #endif
103 
104 #define SMC_BITMAP_USE_THRESHOLD 10
105 
106 typedef struct PageDesc {
107     /* list of TBs intersecting this ram page */
108     uintptr_t first_tb;
109 #ifdef CONFIG_SOFTMMU
110     /* in order to optimize self modifying code, we count the number
111        of lookups we do to a given page to use a bitmap */
112     unsigned long *code_bitmap;
113     unsigned int code_write_count;
114 #else
115     unsigned long flags;
116     void *target_data;
117 #endif
118 #ifndef CONFIG_USER_ONLY
119     QemuSpin lock;
120 #endif
121 } PageDesc;
122 
123 /**
124  * struct page_entry - page descriptor entry
125  * @pd:     pointer to the &struct PageDesc of the page this entry represents
126  * @index:  page index of the page
127  * @locked: whether the page is locked
128  *
129  * This struct helps us keep track of the locked state of a page, without
130  * bloating &struct PageDesc.
131  *
132  * A page lock protects accesses to all fields of &struct PageDesc.
133  *
134  * See also: &struct page_collection.
135  */
136 struct page_entry {
137     PageDesc *pd;
138     tb_page_addr_t index;
139     bool locked;
140 };
141 
142 /**
143  * struct page_collection - tracks a set of pages (i.e. &struct page_entry's)
144  * @tree:   Binary search tree (BST) of the pages, with key == page index
145  * @max:    Pointer to the page in @tree with the highest page index
146  *
147  * To avoid deadlock we lock pages in ascending order of page index.
148  * When operating on a set of pages, we need to keep track of them so that
149  * we can lock them in order and also unlock them later. For this we collect
150  * pages (i.e. &struct page_entry's) in a binary search @tree. Given that the
151  * @tree implementation we use does not provide an O(1) operation to obtain the
152  * highest-ranked element, we use @max to keep track of the inserted page
153  * with the highest index. This is valuable because if a page is not in
154  * the tree and its index is higher than @max's, then we can lock it
155  * without breaking the locking order rule.
156  *
157  * Note on naming: 'struct page_set' would be shorter, but we already have a few
158  * page_set_*() helpers, so page_collection is used instead to avoid confusion.
159  *
160  * See also: page_collection_lock().
161  */
162 struct page_collection {
163     GTree *tree;
164     struct page_entry *max;
165 };
166 
167 /* list iterators for lists of tagged pointers in TranslationBlock */
168 #define TB_FOR_EACH_TAGGED(head, tb, n, field)                          \
169     for (n = (head) & 1, tb = (TranslationBlock *)((head) & ~1);        \
170          tb; tb = (TranslationBlock *)tb->field[n], n = (uintptr_t)tb & 1, \
171              tb = (TranslationBlock *)((uintptr_t)tb & ~1))
172 
173 #define PAGE_FOR_EACH_TB(pagedesc, tb, n)                       \
174     TB_FOR_EACH_TAGGED((pagedesc)->first_tb, tb, n, page_next)
175 
176 #define TB_FOR_EACH_JMP(head_tb, tb, n)                                 \
177     TB_FOR_EACH_TAGGED((head_tb)->jmp_list_head, tb, n, jmp_list_next)
178 
179 /*
180  * In system mode we want L1_MAP to be based on ram offsets,
181  * while in user mode we want it to be based on virtual addresses.
182  *
183  * TODO: For user mode, see the caveat re host vs guest virtual
184  * address spaces near GUEST_ADDR_MAX.
185  */
186 #if !defined(CONFIG_USER_ONLY)
187 #if HOST_LONG_BITS < TARGET_PHYS_ADDR_SPACE_BITS
188 # define L1_MAP_ADDR_SPACE_BITS  HOST_LONG_BITS
189 #else
190 # define L1_MAP_ADDR_SPACE_BITS  TARGET_PHYS_ADDR_SPACE_BITS
191 #endif
192 #else
193 # define L1_MAP_ADDR_SPACE_BITS  MIN(HOST_LONG_BITS, TARGET_ABI_BITS)
194 #endif
195 
196 /* Size of the L2 (and L3, etc) page tables.  */
197 #define V_L2_BITS 10
198 #define V_L2_SIZE (1 << V_L2_BITS)
199 
200 /* Make sure all possible CPU event bits fit in tb->trace_vcpu_dstate */
201 QEMU_BUILD_BUG_ON(CPU_TRACE_DSTATE_MAX_EVENTS >
202                   sizeof_field(TranslationBlock, trace_vcpu_dstate)
203                   * BITS_PER_BYTE);
204 
205 /*
206  * L1 Mapping properties
207  */
208 static int v_l1_size;
209 static int v_l1_shift;
210 static int v_l2_levels;
211 
212 /* The bottom level has pointers to PageDesc, and is indexed by
213  * anything from 4 to (V_L2_BITS + 3) bits, depending on target page size.
214  */
215 #define V_L1_MIN_BITS 4
216 #define V_L1_MAX_BITS (V_L2_BITS + 3)
217 #define V_L1_MAX_SIZE (1 << V_L1_MAX_BITS)
218 
219 static void *l1_map[V_L1_MAX_SIZE];
220 
221 TBContext tb_ctx;
222 
223 static void page_table_config_init(void)
224 {
225     uint32_t v_l1_bits;
226 
227     assert(TARGET_PAGE_BITS);
228     /* The bits remaining after N lower levels of page tables.  */
229     v_l1_bits = (L1_MAP_ADDR_SPACE_BITS - TARGET_PAGE_BITS) % V_L2_BITS;
230     if (v_l1_bits < V_L1_MIN_BITS) {
231         v_l1_bits += V_L2_BITS;
232     }
233 
234     v_l1_size = 1 << v_l1_bits;
235     v_l1_shift = L1_MAP_ADDR_SPACE_BITS - TARGET_PAGE_BITS - v_l1_bits;
236     v_l2_levels = v_l1_shift / V_L2_BITS - 1;
237 
238     assert(v_l1_bits <= V_L1_MAX_BITS);
239     assert(v_l1_shift % V_L2_BITS == 0);
240     assert(v_l2_levels >= 0);
241 }
242 
243 /* Encode VAL as a signed leb128 sequence at P.
244    Return P incremented past the encoded value.  */
245 static uint8_t *encode_sleb128(uint8_t *p, target_long val)
246 {
247     int more, byte;
248 
249     do {
250         byte = val & 0x7f;
251         val >>= 7;
252         more = !((val == 0 && (byte & 0x40) == 0)
253                  || (val == -1 && (byte & 0x40) != 0));
254         if (more) {
255             byte |= 0x80;
256         }
257         *p++ = byte;
258     } while (more);
259 
260     return p;
261 }
262 
263 /* Decode a signed leb128 sequence at *PP; increment *PP past the
264    decoded value.  Return the decoded value.  */
265 static target_long decode_sleb128(const uint8_t **pp)
266 {
267     const uint8_t *p = *pp;
268     target_long val = 0;
269     int byte, shift = 0;
270 
271     do {
272         byte = *p++;
273         val |= (target_ulong)(byte & 0x7f) << shift;
274         shift += 7;
275     } while (byte & 0x80);
276     if (shift < TARGET_LONG_BITS && (byte & 0x40)) {
277         val |= -(target_ulong)1 << shift;
278     }
279 
280     *pp = p;
281     return val;
282 }
283 
284 /* Encode the data collected about the instructions while compiling TB.
285    Place the data at BLOCK, and return the number of bytes consumed.
286 
287    The logical table consists of TARGET_INSN_START_WORDS target_ulong's,
288    which come from the target's insn_start data, followed by a uintptr_t
289    which comes from the host pc of the end of the code implementing the insn.
290 
291    Each line of the table is encoded as sleb128 deltas from the previous
292    line.  The seed for the first line is { tb->pc, 0..., tb->tc.ptr }.
293    That is, the first column is seeded with the guest pc, the last column
294    with the host pc, and the middle columns with zeros.  */
295 
296 static int encode_search(TranslationBlock *tb, uint8_t *block)
297 {
298     uint8_t *highwater = tcg_ctx->code_gen_highwater;
299     uint8_t *p = block;
300     int i, j, n;
301 
302     for (i = 0, n = tb->icount; i < n; ++i) {
303         target_ulong prev;
304 
305         for (j = 0; j < TARGET_INSN_START_WORDS; ++j) {
306             if (i == 0) {
307                 prev = (j == 0 ? tb->pc : 0);
308             } else {
309                 prev = tcg_ctx->gen_insn_data[i - 1][j];
310             }
311             p = encode_sleb128(p, tcg_ctx->gen_insn_data[i][j] - prev);
312         }
313         prev = (i == 0 ? 0 : tcg_ctx->gen_insn_end_off[i - 1]);
314         p = encode_sleb128(p, tcg_ctx->gen_insn_end_off[i] - prev);
315 
316         /* Test for (pending) buffer overflow.  The assumption is that any
317            one row beginning below the high water mark cannot overrun
318            the buffer completely.  Thus we can test for overflow after
319            encoding a row without having to check during encoding.  */
320         if (unlikely(p > highwater)) {
321             return -1;
322         }
323     }
324 
325     return p - block;
326 }
327 
328 /* The cpu state corresponding to 'searched_pc' is restored.
329  * When reset_icount is true, current TB will be interrupted and
330  * icount should be recalculated.
331  */
332 static int cpu_restore_state_from_tb(CPUState *cpu, TranslationBlock *tb,
333                                      uintptr_t searched_pc, bool reset_icount)
334 {
335     target_ulong data[TARGET_INSN_START_WORDS] = { tb->pc };
336     uintptr_t host_pc = (uintptr_t)tb->tc.ptr;
337     CPUArchState *env = cpu->env_ptr;
338     const uint8_t *p = tb->tc.ptr + tb->tc.size;
339     int i, j, num_insns = tb->icount;
340 #ifdef CONFIG_PROFILER
341     TCGProfile *prof = &tcg_ctx->prof;
342     int64_t ti = profile_getclock();
343 #endif
344 
345     searched_pc -= GETPC_ADJ;
346 
347     if (searched_pc < host_pc) {
348         return -1;
349     }
350 
351     /* Reconstruct the stored insn data while looking for the point at
352        which the end of the insn exceeds the searched_pc.  */
353     for (i = 0; i < num_insns; ++i) {
354         for (j = 0; j < TARGET_INSN_START_WORDS; ++j) {
355             data[j] += decode_sleb128(&p);
356         }
357         host_pc += decode_sleb128(&p);
358         if (host_pc > searched_pc) {
359             goto found;
360         }
361     }
362     return -1;
363 
364  found:
365     if (reset_icount && (tb_cflags(tb) & CF_USE_ICOUNT)) {
366         assert(icount_enabled());
367         /* Reset the cycle counter to the start of the block
368            and shift if to the number of actually executed instructions */
369         cpu_neg(cpu)->icount_decr.u16.low += num_insns - i;
370     }
371     restore_state_to_opc(env, tb, data);
372 
373 #ifdef CONFIG_PROFILER
374     qatomic_set(&prof->restore_time,
375                 prof->restore_time + profile_getclock() - ti);
376     qatomic_set(&prof->restore_count, prof->restore_count + 1);
377 #endif
378     return 0;
379 }
380 
381 bool cpu_restore_state(CPUState *cpu, uintptr_t host_pc, bool will_exit)
382 {
383     /*
384      * The host_pc has to be in the rx region of the code buffer.
385      * If it is not we will not be able to resolve it here.
386      * The two cases where host_pc will not be correct are:
387      *
388      *  - fault during translation (instruction fetch)
389      *  - fault from helper (not using GETPC() macro)
390      *
391      * Either way we need return early as we can't resolve it here.
392      */
393     if (in_code_gen_buffer((const void *)(host_pc - tcg_splitwx_diff))) {
394         TranslationBlock *tb = tcg_tb_lookup(host_pc);
395         if (tb) {
396             cpu_restore_state_from_tb(cpu, tb, host_pc, will_exit);
397             return true;
398         }
399     }
400     return false;
401 }
402 
403 void page_init(void)
404 {
405     page_size_init();
406     page_table_config_init();
407 
408 #if defined(CONFIG_BSD) && defined(CONFIG_USER_ONLY)
409     {
410 #ifdef HAVE_KINFO_GETVMMAP
411         struct kinfo_vmentry *freep;
412         int i, cnt;
413 
414         freep = kinfo_getvmmap(getpid(), &cnt);
415         if (freep) {
416             mmap_lock();
417             for (i = 0; i < cnt; i++) {
418                 unsigned long startaddr, endaddr;
419 
420                 startaddr = freep[i].kve_start;
421                 endaddr = freep[i].kve_end;
422                 if (h2g_valid(startaddr)) {
423                     startaddr = h2g(startaddr) & TARGET_PAGE_MASK;
424 
425                     if (h2g_valid(endaddr)) {
426                         endaddr = h2g(endaddr);
427                         page_set_flags(startaddr, endaddr, PAGE_RESERVED);
428                     } else {
429 #if TARGET_ABI_BITS <= L1_MAP_ADDR_SPACE_BITS
430                         endaddr = ~0ul;
431                         page_set_flags(startaddr, endaddr, PAGE_RESERVED);
432 #endif
433                     }
434                 }
435             }
436             free(freep);
437             mmap_unlock();
438         }
439 #else
440         FILE *f;
441 
442         last_brk = (unsigned long)sbrk(0);
443 
444         f = fopen("/compat/linux/proc/self/maps", "r");
445         if (f) {
446             mmap_lock();
447 
448             do {
449                 unsigned long startaddr, endaddr;
450                 int n;
451 
452                 n = fscanf(f, "%lx-%lx %*[^\n]\n", &startaddr, &endaddr);
453 
454                 if (n == 2 && h2g_valid(startaddr)) {
455                     startaddr = h2g(startaddr) & TARGET_PAGE_MASK;
456 
457                     if (h2g_valid(endaddr)) {
458                         endaddr = h2g(endaddr);
459                     } else {
460                         endaddr = ~0ul;
461                     }
462                     page_set_flags(startaddr, endaddr, PAGE_RESERVED);
463                 }
464             } while (!feof(f));
465 
466             fclose(f);
467             mmap_unlock();
468         }
469 #endif
470     }
471 #endif
472 }
473 
474 static PageDesc *page_find_alloc(tb_page_addr_t index, int alloc)
475 {
476     PageDesc *pd;
477     void **lp;
478     int i;
479 
480     /* Level 1.  Always allocated.  */
481     lp = l1_map + ((index >> v_l1_shift) & (v_l1_size - 1));
482 
483     /* Level 2..N-1.  */
484     for (i = v_l2_levels; i > 0; i--) {
485         void **p = qatomic_rcu_read(lp);
486 
487         if (p == NULL) {
488             void *existing;
489 
490             if (!alloc) {
491                 return NULL;
492             }
493             p = g_new0(void *, V_L2_SIZE);
494             existing = qatomic_cmpxchg(lp, NULL, p);
495             if (unlikely(existing)) {
496                 g_free(p);
497                 p = existing;
498             }
499         }
500 
501         lp = p + ((index >> (i * V_L2_BITS)) & (V_L2_SIZE - 1));
502     }
503 
504     pd = qatomic_rcu_read(lp);
505     if (pd == NULL) {
506         void *existing;
507 
508         if (!alloc) {
509             return NULL;
510         }
511         pd = g_new0(PageDesc, V_L2_SIZE);
512 #ifndef CONFIG_USER_ONLY
513         {
514             int i;
515 
516             for (i = 0; i < V_L2_SIZE; i++) {
517                 qemu_spin_init(&pd[i].lock);
518             }
519         }
520 #endif
521         existing = qatomic_cmpxchg(lp, NULL, pd);
522         if (unlikely(existing)) {
523 #ifndef CONFIG_USER_ONLY
524             {
525                 int i;
526 
527                 for (i = 0; i < V_L2_SIZE; i++) {
528                     qemu_spin_destroy(&pd[i].lock);
529                 }
530             }
531 #endif
532             g_free(pd);
533             pd = existing;
534         }
535     }
536 
537     return pd + (index & (V_L2_SIZE - 1));
538 }
539 
540 static inline PageDesc *page_find(tb_page_addr_t index)
541 {
542     return page_find_alloc(index, 0);
543 }
544 
545 static void page_lock_pair(PageDesc **ret_p1, tb_page_addr_t phys1,
546                            PageDesc **ret_p2, tb_page_addr_t phys2, int alloc);
547 
548 /* In user-mode page locks aren't used; mmap_lock is enough */
549 #ifdef CONFIG_USER_ONLY
550 
551 #define assert_page_locked(pd) tcg_debug_assert(have_mmap_lock())
552 
553 static inline void page_lock(PageDesc *pd)
554 { }
555 
556 static inline void page_unlock(PageDesc *pd)
557 { }
558 
559 static inline void page_lock_tb(const TranslationBlock *tb)
560 { }
561 
562 static inline void page_unlock_tb(const TranslationBlock *tb)
563 { }
564 
565 struct page_collection *
566 page_collection_lock(tb_page_addr_t start, tb_page_addr_t end)
567 {
568     return NULL;
569 }
570 
571 void page_collection_unlock(struct page_collection *set)
572 { }
573 #else /* !CONFIG_USER_ONLY */
574 
575 #ifdef CONFIG_DEBUG_TCG
576 
577 static __thread GHashTable *ht_pages_locked_debug;
578 
579 static void ht_pages_locked_debug_init(void)
580 {
581     if (ht_pages_locked_debug) {
582         return;
583     }
584     ht_pages_locked_debug = g_hash_table_new(NULL, NULL);
585 }
586 
587 static bool page_is_locked(const PageDesc *pd)
588 {
589     PageDesc *found;
590 
591     ht_pages_locked_debug_init();
592     found = g_hash_table_lookup(ht_pages_locked_debug, pd);
593     return !!found;
594 }
595 
596 static void page_lock__debug(PageDesc *pd)
597 {
598     ht_pages_locked_debug_init();
599     g_assert(!page_is_locked(pd));
600     g_hash_table_insert(ht_pages_locked_debug, pd, pd);
601 }
602 
603 static void page_unlock__debug(const PageDesc *pd)
604 {
605     bool removed;
606 
607     ht_pages_locked_debug_init();
608     g_assert(page_is_locked(pd));
609     removed = g_hash_table_remove(ht_pages_locked_debug, pd);
610     g_assert(removed);
611 }
612 
613 static void
614 do_assert_page_locked(const PageDesc *pd, const char *file, int line)
615 {
616     if (unlikely(!page_is_locked(pd))) {
617         error_report("assert_page_lock: PageDesc %p not locked @ %s:%d",
618                      pd, file, line);
619         abort();
620     }
621 }
622 
623 #define assert_page_locked(pd) do_assert_page_locked(pd, __FILE__, __LINE__)
624 
625 void assert_no_pages_locked(void)
626 {
627     ht_pages_locked_debug_init();
628     g_assert(g_hash_table_size(ht_pages_locked_debug) == 0);
629 }
630 
631 #else /* !CONFIG_DEBUG_TCG */
632 
633 #define assert_page_locked(pd)
634 
635 static inline void page_lock__debug(const PageDesc *pd)
636 {
637 }
638 
639 static inline void page_unlock__debug(const PageDesc *pd)
640 {
641 }
642 
643 #endif /* CONFIG_DEBUG_TCG */
644 
645 static inline void page_lock(PageDesc *pd)
646 {
647     page_lock__debug(pd);
648     qemu_spin_lock(&pd->lock);
649 }
650 
651 static inline void page_unlock(PageDesc *pd)
652 {
653     qemu_spin_unlock(&pd->lock);
654     page_unlock__debug(pd);
655 }
656 
657 /* lock the page(s) of a TB in the correct acquisition order */
658 static inline void page_lock_tb(const TranslationBlock *tb)
659 {
660     page_lock_pair(NULL, tb->page_addr[0], NULL, tb->page_addr[1], 0);
661 }
662 
663 static inline void page_unlock_tb(const TranslationBlock *tb)
664 {
665     PageDesc *p1 = page_find(tb->page_addr[0] >> TARGET_PAGE_BITS);
666 
667     page_unlock(p1);
668     if (unlikely(tb->page_addr[1] != -1)) {
669         PageDesc *p2 = page_find(tb->page_addr[1] >> TARGET_PAGE_BITS);
670 
671         if (p2 != p1) {
672             page_unlock(p2);
673         }
674     }
675 }
676 
677 static inline struct page_entry *
678 page_entry_new(PageDesc *pd, tb_page_addr_t index)
679 {
680     struct page_entry *pe = g_malloc(sizeof(*pe));
681 
682     pe->index = index;
683     pe->pd = pd;
684     pe->locked = false;
685     return pe;
686 }
687 
688 static void page_entry_destroy(gpointer p)
689 {
690     struct page_entry *pe = p;
691 
692     g_assert(pe->locked);
693     page_unlock(pe->pd);
694     g_free(pe);
695 }
696 
697 /* returns false on success */
698 static bool page_entry_trylock(struct page_entry *pe)
699 {
700     bool busy;
701 
702     busy = qemu_spin_trylock(&pe->pd->lock);
703     if (!busy) {
704         g_assert(!pe->locked);
705         pe->locked = true;
706         page_lock__debug(pe->pd);
707     }
708     return busy;
709 }
710 
711 static void do_page_entry_lock(struct page_entry *pe)
712 {
713     page_lock(pe->pd);
714     g_assert(!pe->locked);
715     pe->locked = true;
716 }
717 
718 static gboolean page_entry_lock(gpointer key, gpointer value, gpointer data)
719 {
720     struct page_entry *pe = value;
721 
722     do_page_entry_lock(pe);
723     return FALSE;
724 }
725 
726 static gboolean page_entry_unlock(gpointer key, gpointer value, gpointer data)
727 {
728     struct page_entry *pe = value;
729 
730     if (pe->locked) {
731         pe->locked = false;
732         page_unlock(pe->pd);
733     }
734     return FALSE;
735 }
736 
737 /*
738  * Trylock a page, and if successful, add the page to a collection.
739  * Returns true ("busy") if the page could not be locked; false otherwise.
740  */
741 static bool page_trylock_add(struct page_collection *set, tb_page_addr_t addr)
742 {
743     tb_page_addr_t index = addr >> TARGET_PAGE_BITS;
744     struct page_entry *pe;
745     PageDesc *pd;
746 
747     pe = g_tree_lookup(set->tree, &index);
748     if (pe) {
749         return false;
750     }
751 
752     pd = page_find(index);
753     if (pd == NULL) {
754         return false;
755     }
756 
757     pe = page_entry_new(pd, index);
758     g_tree_insert(set->tree, &pe->index, pe);
759 
760     /*
761      * If this is either (1) the first insertion or (2) a page whose index
762      * is higher than any other so far, just lock the page and move on.
763      */
764     if (set->max == NULL || pe->index > set->max->index) {
765         set->max = pe;
766         do_page_entry_lock(pe);
767         return false;
768     }
769     /*
770      * Try to acquire out-of-order lock; if busy, return busy so that we acquire
771      * locks in order.
772      */
773     return page_entry_trylock(pe);
774 }
775 
776 static gint tb_page_addr_cmp(gconstpointer ap, gconstpointer bp, gpointer udata)
777 {
778     tb_page_addr_t a = *(const tb_page_addr_t *)ap;
779     tb_page_addr_t b = *(const tb_page_addr_t *)bp;
780 
781     if (a == b) {
782         return 0;
783     } else if (a < b) {
784         return -1;
785     }
786     return 1;
787 }
788 
789 /*
790  * Lock a range of pages ([@start,@end[) as well as the pages of all
791  * intersecting TBs.
792  * Locking order: acquire locks in ascending order of page index.
793  */
794 struct page_collection *
795 page_collection_lock(tb_page_addr_t start, tb_page_addr_t end)
796 {
797     struct page_collection *set = g_malloc(sizeof(*set));
798     tb_page_addr_t index;
799     PageDesc *pd;
800 
801     start >>= TARGET_PAGE_BITS;
802     end   >>= TARGET_PAGE_BITS;
803     g_assert(start <= end);
804 
805     set->tree = g_tree_new_full(tb_page_addr_cmp, NULL, NULL,
806                                 page_entry_destroy);
807     set->max = NULL;
808     assert_no_pages_locked();
809 
810  retry:
811     g_tree_foreach(set->tree, page_entry_lock, NULL);
812 
813     for (index = start; index <= end; index++) {
814         TranslationBlock *tb;
815         int n;
816 
817         pd = page_find(index);
818         if (pd == NULL) {
819             continue;
820         }
821         if (page_trylock_add(set, index << TARGET_PAGE_BITS)) {
822             g_tree_foreach(set->tree, page_entry_unlock, NULL);
823             goto retry;
824         }
825         assert_page_locked(pd);
826         PAGE_FOR_EACH_TB(pd, tb, n) {
827             if (page_trylock_add(set, tb->page_addr[0]) ||
828                 (tb->page_addr[1] != -1 &&
829                  page_trylock_add(set, tb->page_addr[1]))) {
830                 /* drop all locks, and reacquire in order */
831                 g_tree_foreach(set->tree, page_entry_unlock, NULL);
832                 goto retry;
833             }
834         }
835     }
836     return set;
837 }
838 
839 void page_collection_unlock(struct page_collection *set)
840 {
841     /* entries are unlocked and freed via page_entry_destroy */
842     g_tree_destroy(set->tree);
843     g_free(set);
844 }
845 
846 #endif /* !CONFIG_USER_ONLY */
847 
848 static void page_lock_pair(PageDesc **ret_p1, tb_page_addr_t phys1,
849                            PageDesc **ret_p2, tb_page_addr_t phys2, int alloc)
850 {
851     PageDesc *p1, *p2;
852     tb_page_addr_t page1;
853     tb_page_addr_t page2;
854 
855     assert_memory_lock();
856     g_assert(phys1 != -1);
857 
858     page1 = phys1 >> TARGET_PAGE_BITS;
859     page2 = phys2 >> TARGET_PAGE_BITS;
860 
861     p1 = page_find_alloc(page1, alloc);
862     if (ret_p1) {
863         *ret_p1 = p1;
864     }
865     if (likely(phys2 == -1)) {
866         page_lock(p1);
867         return;
868     } else if (page1 == page2) {
869         page_lock(p1);
870         if (ret_p2) {
871             *ret_p2 = p1;
872         }
873         return;
874     }
875     p2 = page_find_alloc(page2, alloc);
876     if (ret_p2) {
877         *ret_p2 = p2;
878     }
879     if (page1 < page2) {
880         page_lock(p1);
881         page_lock(p2);
882     } else {
883         page_lock(p2);
884         page_lock(p1);
885     }
886 }
887 
888 static bool tb_cmp(const void *ap, const void *bp)
889 {
890     const TranslationBlock *a = ap;
891     const TranslationBlock *b = bp;
892 
893     return a->pc == b->pc &&
894         a->cs_base == b->cs_base &&
895         a->flags == b->flags &&
896         (tb_cflags(a) & ~CF_INVALID) == (tb_cflags(b) & ~CF_INVALID) &&
897         a->trace_vcpu_dstate == b->trace_vcpu_dstate &&
898         a->page_addr[0] == b->page_addr[0] &&
899         a->page_addr[1] == b->page_addr[1];
900 }
901 
902 void tb_htable_init(void)
903 {
904     unsigned int mode = QHT_MODE_AUTO_RESIZE;
905 
906     qht_init(&tb_ctx.htable, tb_cmp, CODE_GEN_HTABLE_SIZE, mode);
907 }
908 
909 /* call with @p->lock held */
910 static inline void invalidate_page_bitmap(PageDesc *p)
911 {
912     assert_page_locked(p);
913 #ifdef CONFIG_SOFTMMU
914     g_free(p->code_bitmap);
915     p->code_bitmap = NULL;
916     p->code_write_count = 0;
917 #endif
918 }
919 
920 /* Set to NULL all the 'first_tb' fields in all PageDescs. */
921 static void page_flush_tb_1(int level, void **lp)
922 {
923     int i;
924 
925     if (*lp == NULL) {
926         return;
927     }
928     if (level == 0) {
929         PageDesc *pd = *lp;
930 
931         for (i = 0; i < V_L2_SIZE; ++i) {
932             page_lock(&pd[i]);
933             pd[i].first_tb = (uintptr_t)NULL;
934             invalidate_page_bitmap(pd + i);
935             page_unlock(&pd[i]);
936         }
937     } else {
938         void **pp = *lp;
939 
940         for (i = 0; i < V_L2_SIZE; ++i) {
941             page_flush_tb_1(level - 1, pp + i);
942         }
943     }
944 }
945 
946 static void page_flush_tb(void)
947 {
948     int i, l1_sz = v_l1_size;
949 
950     for (i = 0; i < l1_sz; i++) {
951         page_flush_tb_1(v_l2_levels, l1_map + i);
952     }
953 }
954 
955 static gboolean tb_host_size_iter(gpointer key, gpointer value, gpointer data)
956 {
957     const TranslationBlock *tb = value;
958     size_t *size = data;
959 
960     *size += tb->tc.size;
961     return false;
962 }
963 
964 /* flush all the translation blocks */
965 static void do_tb_flush(CPUState *cpu, run_on_cpu_data tb_flush_count)
966 {
967     bool did_flush = false;
968 
969     mmap_lock();
970     /* If it is already been done on request of another CPU,
971      * just retry.
972      */
973     if (tb_ctx.tb_flush_count != tb_flush_count.host_int) {
974         goto done;
975     }
976     did_flush = true;
977 
978     if (DEBUG_TB_FLUSH_GATE) {
979         size_t nb_tbs = tcg_nb_tbs();
980         size_t host_size = 0;
981 
982         tcg_tb_foreach(tb_host_size_iter, &host_size);
983         printf("qemu: flush code_size=%zu nb_tbs=%zu avg_tb_size=%zu\n",
984                tcg_code_size(), nb_tbs, nb_tbs > 0 ? host_size / nb_tbs : 0);
985     }
986 
987     CPU_FOREACH(cpu) {
988         cpu_tb_jmp_cache_clear(cpu);
989     }
990 
991     qht_reset_size(&tb_ctx.htable, CODE_GEN_HTABLE_SIZE);
992     page_flush_tb();
993 
994     tcg_region_reset_all();
995     /* XXX: flush processor icache at this point if cache flush is
996        expensive */
997     qatomic_mb_set(&tb_ctx.tb_flush_count, tb_ctx.tb_flush_count + 1);
998 
999 done:
1000     mmap_unlock();
1001     if (did_flush) {
1002         qemu_plugin_flush_cb();
1003     }
1004 }
1005 
1006 void tb_flush(CPUState *cpu)
1007 {
1008     if (tcg_enabled()) {
1009         unsigned tb_flush_count = qatomic_mb_read(&tb_ctx.tb_flush_count);
1010 
1011         if (cpu_in_exclusive_context(cpu)) {
1012             do_tb_flush(cpu, RUN_ON_CPU_HOST_INT(tb_flush_count));
1013         } else {
1014             async_safe_run_on_cpu(cpu, do_tb_flush,
1015                                   RUN_ON_CPU_HOST_INT(tb_flush_count));
1016         }
1017     }
1018 }
1019 
1020 /*
1021  * Formerly ifdef DEBUG_TB_CHECK. These debug functions are user-mode-only,
1022  * so in order to prevent bit rot we compile them unconditionally in user-mode,
1023  * and let the optimizer get rid of them by wrapping their user-only callers
1024  * with if (DEBUG_TB_CHECK_GATE).
1025  */
1026 #ifdef CONFIG_USER_ONLY
1027 
1028 static void do_tb_invalidate_check(void *p, uint32_t hash, void *userp)
1029 {
1030     TranslationBlock *tb = p;
1031     target_ulong addr = *(target_ulong *)userp;
1032 
1033     if (!(addr + TARGET_PAGE_SIZE <= tb->pc || addr >= tb->pc + tb->size)) {
1034         printf("ERROR invalidate: address=" TARGET_FMT_lx
1035                " PC=%08lx size=%04x\n", addr, (long)tb->pc, tb->size);
1036     }
1037 }
1038 
1039 /* verify that all the pages have correct rights for code
1040  *
1041  * Called with mmap_lock held.
1042  */
1043 static void tb_invalidate_check(target_ulong address)
1044 {
1045     address &= TARGET_PAGE_MASK;
1046     qht_iter(&tb_ctx.htable, do_tb_invalidate_check, &address);
1047 }
1048 
1049 static void do_tb_page_check(void *p, uint32_t hash, void *userp)
1050 {
1051     TranslationBlock *tb = p;
1052     int flags1, flags2;
1053 
1054     flags1 = page_get_flags(tb->pc);
1055     flags2 = page_get_flags(tb->pc + tb->size - 1);
1056     if ((flags1 & PAGE_WRITE) || (flags2 & PAGE_WRITE)) {
1057         printf("ERROR page flags: PC=%08lx size=%04x f1=%x f2=%x\n",
1058                (long)tb->pc, tb->size, flags1, flags2);
1059     }
1060 }
1061 
1062 /* verify that all the pages have correct rights for code */
1063 static void tb_page_check(void)
1064 {
1065     qht_iter(&tb_ctx.htable, do_tb_page_check, NULL);
1066 }
1067 
1068 #endif /* CONFIG_USER_ONLY */
1069 
1070 /*
1071  * user-mode: call with mmap_lock held
1072  * !user-mode: call with @pd->lock held
1073  */
1074 static inline void tb_page_remove(PageDesc *pd, TranslationBlock *tb)
1075 {
1076     TranslationBlock *tb1;
1077     uintptr_t *pprev;
1078     unsigned int n1;
1079 
1080     assert_page_locked(pd);
1081     pprev = &pd->first_tb;
1082     PAGE_FOR_EACH_TB(pd, tb1, n1) {
1083         if (tb1 == tb) {
1084             *pprev = tb1->page_next[n1];
1085             return;
1086         }
1087         pprev = &tb1->page_next[n1];
1088     }
1089     g_assert_not_reached();
1090 }
1091 
1092 /* remove @orig from its @n_orig-th jump list */
1093 static inline void tb_remove_from_jmp_list(TranslationBlock *orig, int n_orig)
1094 {
1095     uintptr_t ptr, ptr_locked;
1096     TranslationBlock *dest;
1097     TranslationBlock *tb;
1098     uintptr_t *pprev;
1099     int n;
1100 
1101     /* mark the LSB of jmp_dest[] so that no further jumps can be inserted */
1102     ptr = qatomic_or_fetch(&orig->jmp_dest[n_orig], 1);
1103     dest = (TranslationBlock *)(ptr & ~1);
1104     if (dest == NULL) {
1105         return;
1106     }
1107 
1108     qemu_spin_lock(&dest->jmp_lock);
1109     /*
1110      * While acquiring the lock, the jump might have been removed if the
1111      * destination TB was invalidated; check again.
1112      */
1113     ptr_locked = qatomic_read(&orig->jmp_dest[n_orig]);
1114     if (ptr_locked != ptr) {
1115         qemu_spin_unlock(&dest->jmp_lock);
1116         /*
1117          * The only possibility is that the jump was unlinked via
1118          * tb_jump_unlink(dest). Seeing here another destination would be a bug,
1119          * because we set the LSB above.
1120          */
1121         g_assert(ptr_locked == 1 && dest->cflags & CF_INVALID);
1122         return;
1123     }
1124     /*
1125      * We first acquired the lock, and since the destination pointer matches,
1126      * we know for sure that @orig is in the jmp list.
1127      */
1128     pprev = &dest->jmp_list_head;
1129     TB_FOR_EACH_JMP(dest, tb, n) {
1130         if (tb == orig && n == n_orig) {
1131             *pprev = tb->jmp_list_next[n];
1132             /* no need to set orig->jmp_dest[n]; setting the LSB was enough */
1133             qemu_spin_unlock(&dest->jmp_lock);
1134             return;
1135         }
1136         pprev = &tb->jmp_list_next[n];
1137     }
1138     g_assert_not_reached();
1139 }
1140 
1141 /* reset the jump entry 'n' of a TB so that it is not chained to
1142    another TB */
1143 static inline void tb_reset_jump(TranslationBlock *tb, int n)
1144 {
1145     uintptr_t addr = (uintptr_t)(tb->tc.ptr + tb->jmp_reset_offset[n]);
1146     tb_set_jmp_target(tb, n, addr);
1147 }
1148 
1149 /* remove any jumps to the TB */
1150 static inline void tb_jmp_unlink(TranslationBlock *dest)
1151 {
1152     TranslationBlock *tb;
1153     int n;
1154 
1155     qemu_spin_lock(&dest->jmp_lock);
1156 
1157     TB_FOR_EACH_JMP(dest, tb, n) {
1158         tb_reset_jump(tb, n);
1159         qatomic_and(&tb->jmp_dest[n], (uintptr_t)NULL | 1);
1160         /* No need to clear the list entry; setting the dest ptr is enough */
1161     }
1162     dest->jmp_list_head = (uintptr_t)NULL;
1163 
1164     qemu_spin_unlock(&dest->jmp_lock);
1165 }
1166 
1167 /*
1168  * In user-mode, call with mmap_lock held.
1169  * In !user-mode, if @rm_from_page_list is set, call with the TB's pages'
1170  * locks held.
1171  */
1172 static void do_tb_phys_invalidate(TranslationBlock *tb, bool rm_from_page_list)
1173 {
1174     CPUState *cpu;
1175     PageDesc *p;
1176     uint32_t h;
1177     tb_page_addr_t phys_pc;
1178     uint32_t orig_cflags = tb_cflags(tb);
1179 
1180     assert_memory_lock();
1181 
1182     /* make sure no further incoming jumps will be chained to this TB */
1183     qemu_spin_lock(&tb->jmp_lock);
1184     qatomic_set(&tb->cflags, tb->cflags | CF_INVALID);
1185     qemu_spin_unlock(&tb->jmp_lock);
1186 
1187     /* remove the TB from the hash list */
1188     phys_pc = tb->page_addr[0] + (tb->pc & ~TARGET_PAGE_MASK);
1189     h = tb_hash_func(phys_pc, tb->pc, tb->flags, orig_cflags,
1190                      tb->trace_vcpu_dstate);
1191     if (!qht_remove(&tb_ctx.htable, tb, h)) {
1192         return;
1193     }
1194 
1195     /* remove the TB from the page list */
1196     if (rm_from_page_list) {
1197         p = page_find(tb->page_addr[0] >> TARGET_PAGE_BITS);
1198         tb_page_remove(p, tb);
1199         invalidate_page_bitmap(p);
1200         if (tb->page_addr[1] != -1) {
1201             p = page_find(tb->page_addr[1] >> TARGET_PAGE_BITS);
1202             tb_page_remove(p, tb);
1203             invalidate_page_bitmap(p);
1204         }
1205     }
1206 
1207     /* remove the TB from the hash list */
1208     h = tb_jmp_cache_hash_func(tb->pc);
1209     CPU_FOREACH(cpu) {
1210         if (qatomic_read(&cpu->tb_jmp_cache[h]) == tb) {
1211             qatomic_set(&cpu->tb_jmp_cache[h], NULL);
1212         }
1213     }
1214 
1215     /* suppress this TB from the two jump lists */
1216     tb_remove_from_jmp_list(tb, 0);
1217     tb_remove_from_jmp_list(tb, 1);
1218 
1219     /* suppress any remaining jumps to this TB */
1220     tb_jmp_unlink(tb);
1221 
1222     qatomic_set(&tb_ctx.tb_phys_invalidate_count,
1223                 tb_ctx.tb_phys_invalidate_count + 1);
1224 }
1225 
1226 static void tb_phys_invalidate__locked(TranslationBlock *tb)
1227 {
1228     qemu_thread_jit_write();
1229     do_tb_phys_invalidate(tb, true);
1230     qemu_thread_jit_execute();
1231 }
1232 
1233 /* invalidate one TB
1234  *
1235  * Called with mmap_lock held in user-mode.
1236  */
1237 void tb_phys_invalidate(TranslationBlock *tb, tb_page_addr_t page_addr)
1238 {
1239     if (page_addr == -1 && tb->page_addr[0] != -1) {
1240         page_lock_tb(tb);
1241         do_tb_phys_invalidate(tb, true);
1242         page_unlock_tb(tb);
1243     } else {
1244         do_tb_phys_invalidate(tb, false);
1245     }
1246 }
1247 
1248 #ifdef CONFIG_SOFTMMU
1249 /* call with @p->lock held */
1250 static void build_page_bitmap(PageDesc *p)
1251 {
1252     int n, tb_start, tb_end;
1253     TranslationBlock *tb;
1254 
1255     assert_page_locked(p);
1256     p->code_bitmap = bitmap_new(TARGET_PAGE_SIZE);
1257 
1258     PAGE_FOR_EACH_TB(p, tb, n) {
1259         /* NOTE: this is subtle as a TB may span two physical pages */
1260         if (n == 0) {
1261             /* NOTE: tb_end may be after the end of the page, but
1262                it is not a problem */
1263             tb_start = tb->pc & ~TARGET_PAGE_MASK;
1264             tb_end = tb_start + tb->size;
1265             if (tb_end > TARGET_PAGE_SIZE) {
1266                 tb_end = TARGET_PAGE_SIZE;
1267              }
1268         } else {
1269             tb_start = 0;
1270             tb_end = ((tb->pc + tb->size) & ~TARGET_PAGE_MASK);
1271         }
1272         bitmap_set(p->code_bitmap, tb_start, tb_end - tb_start);
1273     }
1274 }
1275 #endif
1276 
1277 /* add the tb in the target page and protect it if necessary
1278  *
1279  * Called with mmap_lock held for user-mode emulation.
1280  * Called with @p->lock held in !user-mode.
1281  */
1282 static inline void tb_page_add(PageDesc *p, TranslationBlock *tb,
1283                                unsigned int n, tb_page_addr_t page_addr)
1284 {
1285 #ifndef CONFIG_USER_ONLY
1286     bool page_already_protected;
1287 #endif
1288 
1289     assert_page_locked(p);
1290 
1291     tb->page_addr[n] = page_addr;
1292     tb->page_next[n] = p->first_tb;
1293 #ifndef CONFIG_USER_ONLY
1294     page_already_protected = p->first_tb != (uintptr_t)NULL;
1295 #endif
1296     p->first_tb = (uintptr_t)tb | n;
1297     invalidate_page_bitmap(p);
1298 
1299 #if defined(CONFIG_USER_ONLY)
1300     if (p->flags & PAGE_WRITE) {
1301         target_ulong addr;
1302         PageDesc *p2;
1303         int prot;
1304 
1305         /* force the host page as non writable (writes will have a
1306            page fault + mprotect overhead) */
1307         page_addr &= qemu_host_page_mask;
1308         prot = 0;
1309         for (addr = page_addr; addr < page_addr + qemu_host_page_size;
1310             addr += TARGET_PAGE_SIZE) {
1311 
1312             p2 = page_find(addr >> TARGET_PAGE_BITS);
1313             if (!p2) {
1314                 continue;
1315             }
1316             prot |= p2->flags;
1317             p2->flags &= ~PAGE_WRITE;
1318           }
1319         mprotect(g2h_untagged(page_addr), qemu_host_page_size,
1320                  (prot & PAGE_BITS) & ~PAGE_WRITE);
1321         if (DEBUG_TB_INVALIDATE_GATE) {
1322             printf("protecting code page: 0x" TB_PAGE_ADDR_FMT "\n", page_addr);
1323         }
1324     }
1325 #else
1326     /* if some code is already present, then the pages are already
1327        protected. So we handle the case where only the first TB is
1328        allocated in a physical page */
1329     if (!page_already_protected) {
1330         tlb_protect_code(page_addr);
1331     }
1332 #endif
1333 }
1334 
1335 /*
1336  * Add a new TB and link it to the physical page tables. phys_page2 is
1337  * (-1) to indicate that only one page contains the TB.
1338  *
1339  * Called with mmap_lock held for user-mode emulation.
1340  *
1341  * Returns a pointer @tb, or a pointer to an existing TB that matches @tb.
1342  * Note that in !user-mode, another thread might have already added a TB
1343  * for the same block of guest code that @tb corresponds to. In that case,
1344  * the caller should discard the original @tb, and use instead the returned TB.
1345  */
1346 static TranslationBlock *
1347 tb_link_page(TranslationBlock *tb, tb_page_addr_t phys_pc,
1348              tb_page_addr_t phys_page2)
1349 {
1350     PageDesc *p;
1351     PageDesc *p2 = NULL;
1352     void *existing_tb = NULL;
1353     uint32_t h;
1354 
1355     assert_memory_lock();
1356     tcg_debug_assert(!(tb->cflags & CF_INVALID));
1357 
1358     /*
1359      * Add the TB to the page list, acquiring first the pages's locks.
1360      * We keep the locks held until after inserting the TB in the hash table,
1361      * so that if the insertion fails we know for sure that the TBs are still
1362      * in the page descriptors.
1363      * Note that inserting into the hash table first isn't an option, since
1364      * we can only insert TBs that are fully initialized.
1365      */
1366     page_lock_pair(&p, phys_pc, &p2, phys_page2, 1);
1367     tb_page_add(p, tb, 0, phys_pc & TARGET_PAGE_MASK);
1368     if (p2) {
1369         tb_page_add(p2, tb, 1, phys_page2);
1370     } else {
1371         tb->page_addr[1] = -1;
1372     }
1373 
1374     /* add in the hash table */
1375     h = tb_hash_func(phys_pc, tb->pc, tb->flags, tb->cflags,
1376                      tb->trace_vcpu_dstate);
1377     qht_insert(&tb_ctx.htable, tb, h, &existing_tb);
1378 
1379     /* remove TB from the page(s) if we couldn't insert it */
1380     if (unlikely(existing_tb)) {
1381         tb_page_remove(p, tb);
1382         invalidate_page_bitmap(p);
1383         if (p2) {
1384             tb_page_remove(p2, tb);
1385             invalidate_page_bitmap(p2);
1386         }
1387         tb = existing_tb;
1388     }
1389 
1390     if (p2 && p2 != p) {
1391         page_unlock(p2);
1392     }
1393     page_unlock(p);
1394 
1395 #ifdef CONFIG_USER_ONLY
1396     if (DEBUG_TB_CHECK_GATE) {
1397         tb_page_check();
1398     }
1399 #endif
1400     return tb;
1401 }
1402 
1403 /* Called with mmap_lock held for user mode emulation.  */
1404 TranslationBlock *tb_gen_code(CPUState *cpu,
1405                               target_ulong pc, target_ulong cs_base,
1406                               uint32_t flags, int cflags)
1407 {
1408     CPUArchState *env = cpu->env_ptr;
1409     TranslationBlock *tb, *existing_tb;
1410     tb_page_addr_t phys_pc, phys_page2;
1411     target_ulong virt_page2;
1412     tcg_insn_unit *gen_code_buf;
1413     int gen_code_size, search_size, max_insns;
1414 #ifdef CONFIG_PROFILER
1415     TCGProfile *prof = &tcg_ctx->prof;
1416     int64_t ti;
1417 #endif
1418 
1419     assert_memory_lock();
1420     qemu_thread_jit_write();
1421 
1422     phys_pc = get_page_addr_code(env, pc);
1423 
1424     if (phys_pc == -1) {
1425         /* Generate a one-shot TB with 1 insn in it */
1426         cflags = (cflags & ~CF_COUNT_MASK) | CF_LAST_IO | 1;
1427     }
1428 
1429     max_insns = cflags & CF_COUNT_MASK;
1430     if (max_insns == 0) {
1431         max_insns = CF_COUNT_MASK;
1432     }
1433     if (max_insns > TCG_MAX_INSNS) {
1434         max_insns = TCG_MAX_INSNS;
1435     }
1436     if (cpu->singlestep_enabled || singlestep) {
1437         max_insns = 1;
1438     }
1439 
1440  buffer_overflow:
1441     tb = tcg_tb_alloc(tcg_ctx);
1442     if (unlikely(!tb)) {
1443         /* flush must be done */
1444         tb_flush(cpu);
1445         mmap_unlock();
1446         /* Make the execution loop process the flush as soon as possible.  */
1447         cpu->exception_index = EXCP_INTERRUPT;
1448         cpu_loop_exit(cpu);
1449     }
1450 
1451     gen_code_buf = tcg_ctx->code_gen_ptr;
1452     tb->tc.ptr = tcg_splitwx_to_rx(gen_code_buf);
1453     tb->pc = pc;
1454     tb->cs_base = cs_base;
1455     tb->flags = flags;
1456     tb->cflags = cflags;
1457     tb->trace_vcpu_dstate = *cpu->trace_dstate;
1458     tcg_ctx->tb_cflags = cflags;
1459  tb_overflow:
1460 
1461 #ifdef CONFIG_PROFILER
1462     /* includes aborted translations because of exceptions */
1463     qatomic_set(&prof->tb_count1, prof->tb_count1 + 1);
1464     ti = profile_getclock();
1465 #endif
1466 
1467     gen_code_size = sigsetjmp(tcg_ctx->jmp_trans, 0);
1468     if (unlikely(gen_code_size != 0)) {
1469         goto error_return;
1470     }
1471 
1472     tcg_func_start(tcg_ctx);
1473 
1474     tcg_ctx->cpu = env_cpu(env);
1475     gen_intermediate_code(cpu, tb, max_insns);
1476     assert(tb->size != 0);
1477     tcg_ctx->cpu = NULL;
1478     max_insns = tb->icount;
1479 
1480     trace_translate_block(tb, tb->pc, tb->tc.ptr);
1481 
1482     /* generate machine code */
1483     tb->jmp_reset_offset[0] = TB_JMP_RESET_OFFSET_INVALID;
1484     tb->jmp_reset_offset[1] = TB_JMP_RESET_OFFSET_INVALID;
1485     tcg_ctx->tb_jmp_reset_offset = tb->jmp_reset_offset;
1486     if (TCG_TARGET_HAS_direct_jump) {
1487         tcg_ctx->tb_jmp_insn_offset = tb->jmp_target_arg;
1488         tcg_ctx->tb_jmp_target_addr = NULL;
1489     } else {
1490         tcg_ctx->tb_jmp_insn_offset = NULL;
1491         tcg_ctx->tb_jmp_target_addr = tb->jmp_target_arg;
1492     }
1493 
1494 #ifdef CONFIG_PROFILER
1495     qatomic_set(&prof->tb_count, prof->tb_count + 1);
1496     qatomic_set(&prof->interm_time,
1497                 prof->interm_time + profile_getclock() - ti);
1498     ti = profile_getclock();
1499 #endif
1500 
1501     gen_code_size = tcg_gen_code(tcg_ctx, tb);
1502     if (unlikely(gen_code_size < 0)) {
1503  error_return:
1504         switch (gen_code_size) {
1505         case -1:
1506             /*
1507              * Overflow of code_gen_buffer, or the current slice of it.
1508              *
1509              * TODO: We don't need to re-do gen_intermediate_code, nor
1510              * should we re-do the tcg optimization currently hidden
1511              * inside tcg_gen_code.  All that should be required is to
1512              * flush the TBs, allocate a new TB, re-initialize it per
1513              * above, and re-do the actual code generation.
1514              */
1515             qemu_log_mask(CPU_LOG_TB_OP | CPU_LOG_TB_OP_OPT,
1516                           "Restarting code generation for "
1517                           "code_gen_buffer overflow\n");
1518             goto buffer_overflow;
1519 
1520         case -2:
1521             /*
1522              * The code generated for the TranslationBlock is too large.
1523              * The maximum size allowed by the unwind info is 64k.
1524              * There may be stricter constraints from relocations
1525              * in the tcg backend.
1526              *
1527              * Try again with half as many insns as we attempted this time.
1528              * If a single insn overflows, there's a bug somewhere...
1529              */
1530             assert(max_insns > 1);
1531             max_insns /= 2;
1532             qemu_log_mask(CPU_LOG_TB_OP | CPU_LOG_TB_OP_OPT,
1533                           "Restarting code generation with "
1534                           "smaller translation block (max %d insns)\n",
1535                           max_insns);
1536             goto tb_overflow;
1537 
1538         default:
1539             g_assert_not_reached();
1540         }
1541     }
1542     search_size = encode_search(tb, (void *)gen_code_buf + gen_code_size);
1543     if (unlikely(search_size < 0)) {
1544         goto buffer_overflow;
1545     }
1546     tb->tc.size = gen_code_size;
1547 
1548 #ifdef CONFIG_PROFILER
1549     qatomic_set(&prof->code_time, prof->code_time + profile_getclock() - ti);
1550     qatomic_set(&prof->code_in_len, prof->code_in_len + tb->size);
1551     qatomic_set(&prof->code_out_len, prof->code_out_len + gen_code_size);
1552     qatomic_set(&prof->search_out_len, prof->search_out_len + search_size);
1553 #endif
1554 
1555 #ifdef DEBUG_DISAS
1556     if (qemu_loglevel_mask(CPU_LOG_TB_OUT_ASM) &&
1557         qemu_log_in_addr_range(tb->pc)) {
1558         FILE *logfile = qemu_log_lock();
1559         int code_size, data_size;
1560         const tcg_target_ulong *rx_data_gen_ptr;
1561         size_t chunk_start;
1562         int insn = 0;
1563 
1564         if (tcg_ctx->data_gen_ptr) {
1565             rx_data_gen_ptr = tcg_splitwx_to_rx(tcg_ctx->data_gen_ptr);
1566             code_size = (const void *)rx_data_gen_ptr - tb->tc.ptr;
1567             data_size = gen_code_size - code_size;
1568         } else {
1569             rx_data_gen_ptr = 0;
1570             code_size = gen_code_size;
1571             data_size = 0;
1572         }
1573 
1574         /* Dump header and the first instruction */
1575         qemu_log("OUT: [size=%d]\n", gen_code_size);
1576         qemu_log("  -- guest addr 0x" TARGET_FMT_lx " + tb prologue\n",
1577                  tcg_ctx->gen_insn_data[insn][0]);
1578         chunk_start = tcg_ctx->gen_insn_end_off[insn];
1579         log_disas(tb->tc.ptr, chunk_start);
1580 
1581         /*
1582          * Dump each instruction chunk, wrapping up empty chunks into
1583          * the next instruction. The whole array is offset so the
1584          * first entry is the beginning of the 2nd instruction.
1585          */
1586         while (insn < tb->icount) {
1587             size_t chunk_end = tcg_ctx->gen_insn_end_off[insn];
1588             if (chunk_end > chunk_start) {
1589                 qemu_log("  -- guest addr 0x" TARGET_FMT_lx "\n",
1590                          tcg_ctx->gen_insn_data[insn][0]);
1591                 log_disas(tb->tc.ptr + chunk_start, chunk_end - chunk_start);
1592                 chunk_start = chunk_end;
1593             }
1594             insn++;
1595         }
1596 
1597         if (chunk_start < code_size) {
1598             qemu_log("  -- tb slow paths + alignment\n");
1599             log_disas(tb->tc.ptr + chunk_start, code_size - chunk_start);
1600         }
1601 
1602         /* Finally dump any data we may have after the block */
1603         if (data_size) {
1604             int i;
1605             qemu_log("  data: [size=%d]\n", data_size);
1606             for (i = 0; i < data_size / sizeof(tcg_target_ulong); i++) {
1607                 if (sizeof(tcg_target_ulong) == 8) {
1608                     qemu_log("0x%08" PRIxPTR ":  .quad  0x%016" TCG_PRIlx "\n",
1609                              (uintptr_t)&rx_data_gen_ptr[i], rx_data_gen_ptr[i]);
1610                 } else if (sizeof(tcg_target_ulong) == 4) {
1611                     qemu_log("0x%08" PRIxPTR ":  .long  0x%08" TCG_PRIlx "\n",
1612                              (uintptr_t)&rx_data_gen_ptr[i], rx_data_gen_ptr[i]);
1613                 } else {
1614                     qemu_build_not_reached();
1615                 }
1616             }
1617         }
1618         qemu_log("\n");
1619         qemu_log_flush();
1620         qemu_log_unlock(logfile);
1621     }
1622 #endif
1623 
1624     qatomic_set(&tcg_ctx->code_gen_ptr, (void *)
1625         ROUND_UP((uintptr_t)gen_code_buf + gen_code_size + search_size,
1626                  CODE_GEN_ALIGN));
1627 
1628     /* init jump list */
1629     qemu_spin_init(&tb->jmp_lock);
1630     tb->jmp_list_head = (uintptr_t)NULL;
1631     tb->jmp_list_next[0] = (uintptr_t)NULL;
1632     tb->jmp_list_next[1] = (uintptr_t)NULL;
1633     tb->jmp_dest[0] = (uintptr_t)NULL;
1634     tb->jmp_dest[1] = (uintptr_t)NULL;
1635 
1636     /* init original jump addresses which have been set during tcg_gen_code() */
1637     if (tb->jmp_reset_offset[0] != TB_JMP_RESET_OFFSET_INVALID) {
1638         tb_reset_jump(tb, 0);
1639     }
1640     if (tb->jmp_reset_offset[1] != TB_JMP_RESET_OFFSET_INVALID) {
1641         tb_reset_jump(tb, 1);
1642     }
1643 
1644     /*
1645      * If the TB is not associated with a physical RAM page then
1646      * it must be a temporary one-insn TB, and we have nothing to do
1647      * except fill in the page_addr[] fields. Return early before
1648      * attempting to link to other TBs or add to the lookup table.
1649      */
1650     if (phys_pc == -1) {
1651         tb->page_addr[0] = tb->page_addr[1] = -1;
1652         return tb;
1653     }
1654 
1655     /*
1656      * Insert TB into the corresponding region tree before publishing it
1657      * through QHT. Otherwise rewinding happened in the TB might fail to
1658      * lookup itself using host PC.
1659      */
1660     tcg_tb_insert(tb);
1661 
1662     /* check next page if needed */
1663     virt_page2 = (pc + tb->size - 1) & TARGET_PAGE_MASK;
1664     phys_page2 = -1;
1665     if ((pc & TARGET_PAGE_MASK) != virt_page2) {
1666         phys_page2 = get_page_addr_code(env, virt_page2);
1667     }
1668     /*
1669      * No explicit memory barrier is required -- tb_link_page() makes the
1670      * TB visible in a consistent state.
1671      */
1672     existing_tb = tb_link_page(tb, phys_pc, phys_page2);
1673     /* if the TB already exists, discard what we just translated */
1674     if (unlikely(existing_tb != tb)) {
1675         uintptr_t orig_aligned = (uintptr_t)gen_code_buf;
1676 
1677         orig_aligned -= ROUND_UP(sizeof(*tb), qemu_icache_linesize);
1678         qatomic_set(&tcg_ctx->code_gen_ptr, (void *)orig_aligned);
1679         tcg_tb_remove(tb);
1680         return existing_tb;
1681     }
1682     return tb;
1683 }
1684 
1685 /*
1686  * @p must be non-NULL.
1687  * user-mode: call with mmap_lock held.
1688  * !user-mode: call with all @pages locked.
1689  */
1690 static void
1691 tb_invalidate_phys_page_range__locked(struct page_collection *pages,
1692                                       PageDesc *p, tb_page_addr_t start,
1693                                       tb_page_addr_t end,
1694                                       uintptr_t retaddr)
1695 {
1696     TranslationBlock *tb;
1697     tb_page_addr_t tb_start, tb_end;
1698     int n;
1699 #ifdef TARGET_HAS_PRECISE_SMC
1700     CPUState *cpu = current_cpu;
1701     CPUArchState *env = NULL;
1702     bool current_tb_not_found = retaddr != 0;
1703     bool current_tb_modified = false;
1704     TranslationBlock *current_tb = NULL;
1705     target_ulong current_pc = 0;
1706     target_ulong current_cs_base = 0;
1707     uint32_t current_flags = 0;
1708 #endif /* TARGET_HAS_PRECISE_SMC */
1709 
1710     assert_page_locked(p);
1711 
1712 #if defined(TARGET_HAS_PRECISE_SMC)
1713     if (cpu != NULL) {
1714         env = cpu->env_ptr;
1715     }
1716 #endif
1717 
1718     /* we remove all the TBs in the range [start, end[ */
1719     /* XXX: see if in some cases it could be faster to invalidate all
1720        the code */
1721     PAGE_FOR_EACH_TB(p, tb, n) {
1722         assert_page_locked(p);
1723         /* NOTE: this is subtle as a TB may span two physical pages */
1724         if (n == 0) {
1725             /* NOTE: tb_end may be after the end of the page, but
1726                it is not a problem */
1727             tb_start = tb->page_addr[0] + (tb->pc & ~TARGET_PAGE_MASK);
1728             tb_end = tb_start + tb->size;
1729         } else {
1730             tb_start = tb->page_addr[1];
1731             tb_end = tb_start + ((tb->pc + tb->size) & ~TARGET_PAGE_MASK);
1732         }
1733         if (!(tb_end <= start || tb_start >= end)) {
1734 #ifdef TARGET_HAS_PRECISE_SMC
1735             if (current_tb_not_found) {
1736                 current_tb_not_found = false;
1737                 /* now we have a real cpu fault */
1738                 current_tb = tcg_tb_lookup(retaddr);
1739             }
1740             if (current_tb == tb &&
1741                 (tb_cflags(current_tb) & CF_COUNT_MASK) != 1) {
1742                 /*
1743                  * If we are modifying the current TB, we must stop
1744                  * its execution. We could be more precise by checking
1745                  * that the modification is after the current PC, but it
1746                  * would require a specialized function to partially
1747                  * restore the CPU state.
1748                  */
1749                 current_tb_modified = true;
1750                 cpu_restore_state_from_tb(cpu, current_tb, retaddr, true);
1751                 cpu_get_tb_cpu_state(env, &current_pc, &current_cs_base,
1752                                      &current_flags);
1753             }
1754 #endif /* TARGET_HAS_PRECISE_SMC */
1755             tb_phys_invalidate__locked(tb);
1756         }
1757     }
1758 #if !defined(CONFIG_USER_ONLY)
1759     /* if no code remaining, no need to continue to use slow writes */
1760     if (!p->first_tb) {
1761         invalidate_page_bitmap(p);
1762         tlb_unprotect_code(start);
1763     }
1764 #endif
1765 #ifdef TARGET_HAS_PRECISE_SMC
1766     if (current_tb_modified) {
1767         page_collection_unlock(pages);
1768         /* Force execution of one insn next time.  */
1769         cpu->cflags_next_tb = 1 | curr_cflags(cpu);
1770         mmap_unlock();
1771         cpu_loop_exit_noexc(cpu);
1772     }
1773 #endif
1774 }
1775 
1776 /*
1777  * Invalidate all TBs which intersect with the target physical address range
1778  * [start;end[. NOTE: start and end must refer to the *same* physical page.
1779  * 'is_cpu_write_access' should be true if called from a real cpu write
1780  * access: the virtual CPU will exit the current TB if code is modified inside
1781  * this TB.
1782  *
1783  * Called with mmap_lock held for user-mode emulation
1784  */
1785 void tb_invalidate_phys_page_range(tb_page_addr_t start, tb_page_addr_t end)
1786 {
1787     struct page_collection *pages;
1788     PageDesc *p;
1789 
1790     assert_memory_lock();
1791 
1792     p = page_find(start >> TARGET_PAGE_BITS);
1793     if (p == NULL) {
1794         return;
1795     }
1796     pages = page_collection_lock(start, end);
1797     tb_invalidate_phys_page_range__locked(pages, p, start, end, 0);
1798     page_collection_unlock(pages);
1799 }
1800 
1801 /*
1802  * Invalidate all TBs which intersect with the target physical address range
1803  * [start;end[. NOTE: start and end may refer to *different* physical pages.
1804  * 'is_cpu_write_access' should be true if called from a real cpu write
1805  * access: the virtual CPU will exit the current TB if code is modified inside
1806  * this TB.
1807  *
1808  * Called with mmap_lock held for user-mode emulation.
1809  */
1810 #ifdef CONFIG_SOFTMMU
1811 void tb_invalidate_phys_range(ram_addr_t start, ram_addr_t end)
1812 #else
1813 void tb_invalidate_phys_range(target_ulong start, target_ulong end)
1814 #endif
1815 {
1816     struct page_collection *pages;
1817     tb_page_addr_t next;
1818 
1819     assert_memory_lock();
1820 
1821     pages = page_collection_lock(start, end);
1822     for (next = (start & TARGET_PAGE_MASK) + TARGET_PAGE_SIZE;
1823          start < end;
1824          start = next, next += TARGET_PAGE_SIZE) {
1825         PageDesc *pd = page_find(start >> TARGET_PAGE_BITS);
1826         tb_page_addr_t bound = MIN(next, end);
1827 
1828         if (pd == NULL) {
1829             continue;
1830         }
1831         tb_invalidate_phys_page_range__locked(pages, pd, start, bound, 0);
1832     }
1833     page_collection_unlock(pages);
1834 }
1835 
1836 #ifdef CONFIG_SOFTMMU
1837 /* len must be <= 8 and start must be a multiple of len.
1838  * Called via softmmu_template.h when code areas are written to with
1839  * iothread mutex not held.
1840  *
1841  * Call with all @pages in the range [@start, @start + len[ locked.
1842  */
1843 void tb_invalidate_phys_page_fast(struct page_collection *pages,
1844                                   tb_page_addr_t start, int len,
1845                                   uintptr_t retaddr)
1846 {
1847     PageDesc *p;
1848 
1849     assert_memory_lock();
1850 
1851     p = page_find(start >> TARGET_PAGE_BITS);
1852     if (!p) {
1853         return;
1854     }
1855 
1856     assert_page_locked(p);
1857     if (!p->code_bitmap &&
1858         ++p->code_write_count >= SMC_BITMAP_USE_THRESHOLD) {
1859         build_page_bitmap(p);
1860     }
1861     if (p->code_bitmap) {
1862         unsigned int nr;
1863         unsigned long b;
1864 
1865         nr = start & ~TARGET_PAGE_MASK;
1866         b = p->code_bitmap[BIT_WORD(nr)] >> (nr & (BITS_PER_LONG - 1));
1867         if (b & ((1 << len) - 1)) {
1868             goto do_invalidate;
1869         }
1870     } else {
1871     do_invalidate:
1872         tb_invalidate_phys_page_range__locked(pages, p, start, start + len,
1873                                               retaddr);
1874     }
1875 }
1876 #else
1877 /* Called with mmap_lock held. If pc is not 0 then it indicates the
1878  * host PC of the faulting store instruction that caused this invalidate.
1879  * Returns true if the caller needs to abort execution of the current
1880  * TB (because it was modified by this store and the guest CPU has
1881  * precise-SMC semantics).
1882  */
1883 static bool tb_invalidate_phys_page(tb_page_addr_t addr, uintptr_t pc)
1884 {
1885     TranslationBlock *tb;
1886     PageDesc *p;
1887     int n;
1888 #ifdef TARGET_HAS_PRECISE_SMC
1889     TranslationBlock *current_tb = NULL;
1890     CPUState *cpu = current_cpu;
1891     CPUArchState *env = NULL;
1892     int current_tb_modified = 0;
1893     target_ulong current_pc = 0;
1894     target_ulong current_cs_base = 0;
1895     uint32_t current_flags = 0;
1896 #endif
1897 
1898     assert_memory_lock();
1899 
1900     addr &= TARGET_PAGE_MASK;
1901     p = page_find(addr >> TARGET_PAGE_BITS);
1902     if (!p) {
1903         return false;
1904     }
1905 
1906 #ifdef TARGET_HAS_PRECISE_SMC
1907     if (p->first_tb && pc != 0) {
1908         current_tb = tcg_tb_lookup(pc);
1909     }
1910     if (cpu != NULL) {
1911         env = cpu->env_ptr;
1912     }
1913 #endif
1914     assert_page_locked(p);
1915     PAGE_FOR_EACH_TB(p, tb, n) {
1916 #ifdef TARGET_HAS_PRECISE_SMC
1917         if (current_tb == tb &&
1918             (tb_cflags(current_tb) & CF_COUNT_MASK) != 1) {
1919                 /* If we are modifying the current TB, we must stop
1920                    its execution. We could be more precise by checking
1921                    that the modification is after the current PC, but it
1922                    would require a specialized function to partially
1923                    restore the CPU state */
1924 
1925             current_tb_modified = 1;
1926             cpu_restore_state_from_tb(cpu, current_tb, pc, true);
1927             cpu_get_tb_cpu_state(env, &current_pc, &current_cs_base,
1928                                  &current_flags);
1929         }
1930 #endif /* TARGET_HAS_PRECISE_SMC */
1931         tb_phys_invalidate(tb, addr);
1932     }
1933     p->first_tb = (uintptr_t)NULL;
1934 #ifdef TARGET_HAS_PRECISE_SMC
1935     if (current_tb_modified) {
1936         /* Force execution of one insn next time.  */
1937         cpu->cflags_next_tb = 1 | curr_cflags(cpu);
1938         return true;
1939     }
1940 #endif
1941 
1942     return false;
1943 }
1944 #endif
1945 
1946 /* user-mode: call with mmap_lock held */
1947 void tb_check_watchpoint(CPUState *cpu, uintptr_t retaddr)
1948 {
1949     TranslationBlock *tb;
1950 
1951     assert_memory_lock();
1952 
1953     tb = tcg_tb_lookup(retaddr);
1954     if (tb) {
1955         /* We can use retranslation to find the PC.  */
1956         cpu_restore_state_from_tb(cpu, tb, retaddr, true);
1957         tb_phys_invalidate(tb, -1);
1958     } else {
1959         /* The exception probably happened in a helper.  The CPU state should
1960            have been saved before calling it. Fetch the PC from there.  */
1961         CPUArchState *env = cpu->env_ptr;
1962         target_ulong pc, cs_base;
1963         tb_page_addr_t addr;
1964         uint32_t flags;
1965 
1966         cpu_get_tb_cpu_state(env, &pc, &cs_base, &flags);
1967         addr = get_page_addr_code(env, pc);
1968         if (addr != -1) {
1969             tb_invalidate_phys_range(addr, addr + 1);
1970         }
1971     }
1972 }
1973 
1974 #ifndef CONFIG_USER_ONLY
1975 /*
1976  * In deterministic execution mode, instructions doing device I/Os
1977  * must be at the end of the TB.
1978  *
1979  * Called by softmmu_template.h, with iothread mutex not held.
1980  */
1981 void cpu_io_recompile(CPUState *cpu, uintptr_t retaddr)
1982 {
1983     TranslationBlock *tb;
1984     CPUClass *cc;
1985     uint32_t n;
1986 
1987     tb = tcg_tb_lookup(retaddr);
1988     if (!tb) {
1989         cpu_abort(cpu, "cpu_io_recompile: could not find TB for pc=%p",
1990                   (void *)retaddr);
1991     }
1992     cpu_restore_state_from_tb(cpu, tb, retaddr, true);
1993 
1994     /*
1995      * Some guests must re-execute the branch when re-executing a delay
1996      * slot instruction.  When this is the case, adjust icount and N
1997      * to account for the re-execution of the branch.
1998      */
1999     n = 1;
2000     cc = CPU_GET_CLASS(cpu);
2001     if (cc->tcg_ops->io_recompile_replay_branch &&
2002         cc->tcg_ops->io_recompile_replay_branch(cpu, tb)) {
2003         cpu_neg(cpu)->icount_decr.u16.low++;
2004         n = 2;
2005     }
2006 
2007     /*
2008      * Exit the loop and potentially generate a new TB executing the
2009      * just the I/O insns. We also limit instrumentation to memory
2010      * operations only (which execute after completion) so we don't
2011      * double instrument the instruction.
2012      */
2013     cpu->cflags_next_tb = curr_cflags(cpu) | CF_MEMI_ONLY | CF_LAST_IO | n;
2014 
2015     qemu_log_mask_and_addr(CPU_LOG_EXEC, tb->pc,
2016                            "cpu_io_recompile: rewound execution of TB to "
2017                            TARGET_FMT_lx "\n", tb->pc);
2018 
2019     cpu_loop_exit_noexc(cpu);
2020 }
2021 
2022 static void print_qht_statistics(struct qht_stats hst)
2023 {
2024     uint32_t hgram_opts;
2025     size_t hgram_bins;
2026     char *hgram;
2027 
2028     if (!hst.head_buckets) {
2029         return;
2030     }
2031     qemu_printf("TB hash buckets     %zu/%zu (%0.2f%% head buckets used)\n",
2032                 hst.used_head_buckets, hst.head_buckets,
2033                 (double)hst.used_head_buckets / hst.head_buckets * 100);
2034 
2035     hgram_opts =  QDIST_PR_BORDER | QDIST_PR_LABELS;
2036     hgram_opts |= QDIST_PR_100X   | QDIST_PR_PERCENT;
2037     if (qdist_xmax(&hst.occupancy) - qdist_xmin(&hst.occupancy) == 1) {
2038         hgram_opts |= QDIST_PR_NODECIMAL;
2039     }
2040     hgram = qdist_pr(&hst.occupancy, 10, hgram_opts);
2041     qemu_printf("TB hash occupancy   %0.2f%% avg chain occ. Histogram: %s\n",
2042                 qdist_avg(&hst.occupancy) * 100, hgram);
2043     g_free(hgram);
2044 
2045     hgram_opts = QDIST_PR_BORDER | QDIST_PR_LABELS;
2046     hgram_bins = qdist_xmax(&hst.chain) - qdist_xmin(&hst.chain);
2047     if (hgram_bins > 10) {
2048         hgram_bins = 10;
2049     } else {
2050         hgram_bins = 0;
2051         hgram_opts |= QDIST_PR_NODECIMAL | QDIST_PR_NOBINRANGE;
2052     }
2053     hgram = qdist_pr(&hst.chain, hgram_bins, hgram_opts);
2054     qemu_printf("TB hash avg chain   %0.3f buckets. Histogram: %s\n",
2055                 qdist_avg(&hst.chain), hgram);
2056     g_free(hgram);
2057 }
2058 
2059 struct tb_tree_stats {
2060     size_t nb_tbs;
2061     size_t host_size;
2062     size_t target_size;
2063     size_t max_target_size;
2064     size_t direct_jmp_count;
2065     size_t direct_jmp2_count;
2066     size_t cross_page;
2067 };
2068 
2069 static gboolean tb_tree_stats_iter(gpointer key, gpointer value, gpointer data)
2070 {
2071     const TranslationBlock *tb = value;
2072     struct tb_tree_stats *tst = data;
2073 
2074     tst->nb_tbs++;
2075     tst->host_size += tb->tc.size;
2076     tst->target_size += tb->size;
2077     if (tb->size > tst->max_target_size) {
2078         tst->max_target_size = tb->size;
2079     }
2080     if (tb->page_addr[1] != -1) {
2081         tst->cross_page++;
2082     }
2083     if (tb->jmp_reset_offset[0] != TB_JMP_RESET_OFFSET_INVALID) {
2084         tst->direct_jmp_count++;
2085         if (tb->jmp_reset_offset[1] != TB_JMP_RESET_OFFSET_INVALID) {
2086             tst->direct_jmp2_count++;
2087         }
2088     }
2089     return false;
2090 }
2091 
2092 void dump_exec_info(void)
2093 {
2094     struct tb_tree_stats tst = {};
2095     struct qht_stats hst;
2096     size_t nb_tbs, flush_full, flush_part, flush_elide;
2097 
2098     tcg_tb_foreach(tb_tree_stats_iter, &tst);
2099     nb_tbs = tst.nb_tbs;
2100     /* XXX: avoid using doubles ? */
2101     qemu_printf("Translation buffer state:\n");
2102     /*
2103      * Report total code size including the padding and TB structs;
2104      * otherwise users might think "-accel tcg,tb-size" is not honoured.
2105      * For avg host size we use the precise numbers from tb_tree_stats though.
2106      */
2107     qemu_printf("gen code size       %zu/%zu\n",
2108                 tcg_code_size(), tcg_code_capacity());
2109     qemu_printf("TB count            %zu\n", nb_tbs);
2110     qemu_printf("TB avg target size  %zu max=%zu bytes\n",
2111                 nb_tbs ? tst.target_size / nb_tbs : 0,
2112                 tst.max_target_size);
2113     qemu_printf("TB avg host size    %zu bytes (expansion ratio: %0.1f)\n",
2114                 nb_tbs ? tst.host_size / nb_tbs : 0,
2115                 tst.target_size ? (double)tst.host_size / tst.target_size : 0);
2116     qemu_printf("cross page TB count %zu (%zu%%)\n", tst.cross_page,
2117                 nb_tbs ? (tst.cross_page * 100) / nb_tbs : 0);
2118     qemu_printf("direct jump count   %zu (%zu%%) (2 jumps=%zu %zu%%)\n",
2119                 tst.direct_jmp_count,
2120                 nb_tbs ? (tst.direct_jmp_count * 100) / nb_tbs : 0,
2121                 tst.direct_jmp2_count,
2122                 nb_tbs ? (tst.direct_jmp2_count * 100) / nb_tbs : 0);
2123 
2124     qht_statistics_init(&tb_ctx.htable, &hst);
2125     print_qht_statistics(hst);
2126     qht_statistics_destroy(&hst);
2127 
2128     qemu_printf("\nStatistics:\n");
2129     qemu_printf("TB flush count      %u\n",
2130                 qatomic_read(&tb_ctx.tb_flush_count));
2131     qemu_printf("TB invalidate count %u\n",
2132                 qatomic_read(&tb_ctx.tb_phys_invalidate_count));
2133 
2134     tlb_flush_counts(&flush_full, &flush_part, &flush_elide);
2135     qemu_printf("TLB full flushes    %zu\n", flush_full);
2136     qemu_printf("TLB partial flushes %zu\n", flush_part);
2137     qemu_printf("TLB elided flushes  %zu\n", flush_elide);
2138     tcg_dump_info();
2139 }
2140 
2141 void dump_opcount_info(void)
2142 {
2143     tcg_dump_op_count();
2144 }
2145 
2146 #else /* CONFIG_USER_ONLY */
2147 
2148 void cpu_interrupt(CPUState *cpu, int mask)
2149 {
2150     g_assert(qemu_mutex_iothread_locked());
2151     cpu->interrupt_request |= mask;
2152     qatomic_set(&cpu_neg(cpu)->icount_decr.u16.high, -1);
2153 }
2154 
2155 /*
2156  * Walks guest process memory "regions" one by one
2157  * and calls callback function 'fn' for each region.
2158  */
2159 struct walk_memory_regions_data {
2160     walk_memory_regions_fn fn;
2161     void *priv;
2162     target_ulong start;
2163     int prot;
2164 };
2165 
2166 static int walk_memory_regions_end(struct walk_memory_regions_data *data,
2167                                    target_ulong end, int new_prot)
2168 {
2169     if (data->start != -1u) {
2170         int rc = data->fn(data->priv, data->start, end, data->prot);
2171         if (rc != 0) {
2172             return rc;
2173         }
2174     }
2175 
2176     data->start = (new_prot ? end : -1u);
2177     data->prot = new_prot;
2178 
2179     return 0;
2180 }
2181 
2182 static int walk_memory_regions_1(struct walk_memory_regions_data *data,
2183                                  target_ulong base, int level, void **lp)
2184 {
2185     target_ulong pa;
2186     int i, rc;
2187 
2188     if (*lp == NULL) {
2189         return walk_memory_regions_end(data, base, 0);
2190     }
2191 
2192     if (level == 0) {
2193         PageDesc *pd = *lp;
2194 
2195         for (i = 0; i < V_L2_SIZE; ++i) {
2196             int prot = pd[i].flags;
2197 
2198             pa = base | (i << TARGET_PAGE_BITS);
2199             if (prot != data->prot) {
2200                 rc = walk_memory_regions_end(data, pa, prot);
2201                 if (rc != 0) {
2202                     return rc;
2203                 }
2204             }
2205         }
2206     } else {
2207         void **pp = *lp;
2208 
2209         for (i = 0; i < V_L2_SIZE; ++i) {
2210             pa = base | ((target_ulong)i <<
2211                 (TARGET_PAGE_BITS + V_L2_BITS * level));
2212             rc = walk_memory_regions_1(data, pa, level - 1, pp + i);
2213             if (rc != 0) {
2214                 return rc;
2215             }
2216         }
2217     }
2218 
2219     return 0;
2220 }
2221 
2222 int walk_memory_regions(void *priv, walk_memory_regions_fn fn)
2223 {
2224     struct walk_memory_regions_data data;
2225     uintptr_t i, l1_sz = v_l1_size;
2226 
2227     data.fn = fn;
2228     data.priv = priv;
2229     data.start = -1u;
2230     data.prot = 0;
2231 
2232     for (i = 0; i < l1_sz; i++) {
2233         target_ulong base = i << (v_l1_shift + TARGET_PAGE_BITS);
2234         int rc = walk_memory_regions_1(&data, base, v_l2_levels, l1_map + i);
2235         if (rc != 0) {
2236             return rc;
2237         }
2238     }
2239 
2240     return walk_memory_regions_end(&data, 0, 0);
2241 }
2242 
2243 static int dump_region(void *priv, target_ulong start,
2244     target_ulong end, unsigned long prot)
2245 {
2246     FILE *f = (FILE *)priv;
2247 
2248     (void) fprintf(f, TARGET_FMT_lx"-"TARGET_FMT_lx
2249         " "TARGET_FMT_lx" %c%c%c\n",
2250         start, end, end - start,
2251         ((prot & PAGE_READ) ? 'r' : '-'),
2252         ((prot & PAGE_WRITE) ? 'w' : '-'),
2253         ((prot & PAGE_EXEC) ? 'x' : '-'));
2254 
2255     return 0;
2256 }
2257 
2258 /* dump memory mappings */
2259 void page_dump(FILE *f)
2260 {
2261     const int length = sizeof(target_ulong) * 2;
2262     (void) fprintf(f, "%-*s %-*s %-*s %s\n",
2263             length, "start", length, "end", length, "size", "prot");
2264     walk_memory_regions(f, dump_region);
2265 }
2266 
2267 int page_get_flags(target_ulong address)
2268 {
2269     PageDesc *p;
2270 
2271     p = page_find(address >> TARGET_PAGE_BITS);
2272     if (!p) {
2273         return 0;
2274     }
2275     return p->flags;
2276 }
2277 
2278 /* Modify the flags of a page and invalidate the code if necessary.
2279    The flag PAGE_WRITE_ORG is positioned automatically depending
2280    on PAGE_WRITE.  The mmap_lock should already be held.  */
2281 void page_set_flags(target_ulong start, target_ulong end, int flags)
2282 {
2283     target_ulong addr, len;
2284     bool reset_target_data;
2285 
2286     /* This function should never be called with addresses outside the
2287        guest address space.  If this assert fires, it probably indicates
2288        a missing call to h2g_valid.  */
2289     assert(end - 1 <= GUEST_ADDR_MAX);
2290     assert(start < end);
2291     /* Only set PAGE_ANON with new mappings. */
2292     assert(!(flags & PAGE_ANON) || (flags & PAGE_RESET));
2293     assert_memory_lock();
2294 
2295     start = start & TARGET_PAGE_MASK;
2296     end = TARGET_PAGE_ALIGN(end);
2297 
2298     if (flags & PAGE_WRITE) {
2299         flags |= PAGE_WRITE_ORG;
2300     }
2301     reset_target_data = !(flags & PAGE_VALID) || (flags & PAGE_RESET);
2302     flags &= ~PAGE_RESET;
2303 
2304     for (addr = start, len = end - start;
2305          len != 0;
2306          len -= TARGET_PAGE_SIZE, addr += TARGET_PAGE_SIZE) {
2307         PageDesc *p = page_find_alloc(addr >> TARGET_PAGE_BITS, 1);
2308 
2309         /* If the write protection bit is set, then we invalidate
2310            the code inside.  */
2311         if (!(p->flags & PAGE_WRITE) &&
2312             (flags & PAGE_WRITE) &&
2313             p->first_tb) {
2314             tb_invalidate_phys_page(addr, 0);
2315         }
2316         if (reset_target_data) {
2317             g_free(p->target_data);
2318             p->target_data = NULL;
2319             p->flags = flags;
2320         } else {
2321             /* Using mprotect on a page does not change MAP_ANON. */
2322             p->flags = (p->flags & PAGE_ANON) | flags;
2323         }
2324     }
2325 }
2326 
2327 void *page_get_target_data(target_ulong address)
2328 {
2329     PageDesc *p = page_find(address >> TARGET_PAGE_BITS);
2330     return p ? p->target_data : NULL;
2331 }
2332 
2333 void *page_alloc_target_data(target_ulong address, size_t size)
2334 {
2335     PageDesc *p = page_find(address >> TARGET_PAGE_BITS);
2336     void *ret = NULL;
2337 
2338     if (p->flags & PAGE_VALID) {
2339         ret = p->target_data;
2340         if (!ret) {
2341             p->target_data = ret = g_malloc0(size);
2342         }
2343     }
2344     return ret;
2345 }
2346 
2347 int page_check_range(target_ulong start, target_ulong len, int flags)
2348 {
2349     PageDesc *p;
2350     target_ulong end;
2351     target_ulong addr;
2352 
2353     /* This function should never be called with addresses outside the
2354        guest address space.  If this assert fires, it probably indicates
2355        a missing call to h2g_valid.  */
2356     if (TARGET_ABI_BITS > L1_MAP_ADDR_SPACE_BITS) {
2357         assert(start < ((target_ulong)1 << L1_MAP_ADDR_SPACE_BITS));
2358     }
2359 
2360     if (len == 0) {
2361         return 0;
2362     }
2363     if (start + len - 1 < start) {
2364         /* We've wrapped around.  */
2365         return -1;
2366     }
2367 
2368     /* must do before we loose bits in the next step */
2369     end = TARGET_PAGE_ALIGN(start + len);
2370     start = start & TARGET_PAGE_MASK;
2371 
2372     for (addr = start, len = end - start;
2373          len != 0;
2374          len -= TARGET_PAGE_SIZE, addr += TARGET_PAGE_SIZE) {
2375         p = page_find(addr >> TARGET_PAGE_BITS);
2376         if (!p) {
2377             return -1;
2378         }
2379         if (!(p->flags & PAGE_VALID)) {
2380             return -1;
2381         }
2382 
2383         if ((flags & PAGE_READ) && !(p->flags & PAGE_READ)) {
2384             return -1;
2385         }
2386         if (flags & PAGE_WRITE) {
2387             if (!(p->flags & PAGE_WRITE_ORG)) {
2388                 return -1;
2389             }
2390             /* unprotect the page if it was put read-only because it
2391                contains translated code */
2392             if (!(p->flags & PAGE_WRITE)) {
2393                 if (!page_unprotect(addr, 0)) {
2394                     return -1;
2395                 }
2396             }
2397         }
2398     }
2399     return 0;
2400 }
2401 
2402 /* called from signal handler: invalidate the code and unprotect the
2403  * page. Return 0 if the fault was not handled, 1 if it was handled,
2404  * and 2 if it was handled but the caller must cause the TB to be
2405  * immediately exited. (We can only return 2 if the 'pc' argument is
2406  * non-zero.)
2407  */
2408 int page_unprotect(target_ulong address, uintptr_t pc)
2409 {
2410     unsigned int prot;
2411     bool current_tb_invalidated;
2412     PageDesc *p;
2413     target_ulong host_start, host_end, addr;
2414 
2415     /* Technically this isn't safe inside a signal handler.  However we
2416        know this only ever happens in a synchronous SEGV handler, so in
2417        practice it seems to be ok.  */
2418     mmap_lock();
2419 
2420     p = page_find(address >> TARGET_PAGE_BITS);
2421     if (!p) {
2422         mmap_unlock();
2423         return 0;
2424     }
2425 
2426     /* if the page was really writable, then we change its
2427        protection back to writable */
2428     if (p->flags & PAGE_WRITE_ORG) {
2429         current_tb_invalidated = false;
2430         if (p->flags & PAGE_WRITE) {
2431             /* If the page is actually marked WRITE then assume this is because
2432              * this thread raced with another one which got here first and
2433              * set the page to PAGE_WRITE and did the TB invalidate for us.
2434              */
2435 #ifdef TARGET_HAS_PRECISE_SMC
2436             TranslationBlock *current_tb = tcg_tb_lookup(pc);
2437             if (current_tb) {
2438                 current_tb_invalidated = tb_cflags(current_tb) & CF_INVALID;
2439             }
2440 #endif
2441         } else {
2442             host_start = address & qemu_host_page_mask;
2443             host_end = host_start + qemu_host_page_size;
2444 
2445             prot = 0;
2446             for (addr = host_start; addr < host_end; addr += TARGET_PAGE_SIZE) {
2447                 p = page_find(addr >> TARGET_PAGE_BITS);
2448                 p->flags |= PAGE_WRITE;
2449                 prot |= p->flags;
2450 
2451                 /* and since the content will be modified, we must invalidate
2452                    the corresponding translated code. */
2453                 current_tb_invalidated |= tb_invalidate_phys_page(addr, pc);
2454 #ifdef CONFIG_USER_ONLY
2455                 if (DEBUG_TB_CHECK_GATE) {
2456                     tb_invalidate_check(addr);
2457                 }
2458 #endif
2459             }
2460             mprotect((void *)g2h_untagged(host_start), qemu_host_page_size,
2461                      prot & PAGE_BITS);
2462         }
2463         mmap_unlock();
2464         /* If current TB was invalidated return to main loop */
2465         return current_tb_invalidated ? 2 : 1;
2466     }
2467     mmap_unlock();
2468     return 0;
2469 }
2470 #endif /* CONFIG_USER_ONLY */
2471 
2472 /* This is a wrapper for common code that can not use CONFIG_SOFTMMU */
2473 void tcg_flush_softmmu_tlb(CPUState *cs)
2474 {
2475 #ifdef CONFIG_SOFTMMU
2476     tlb_flush(cs);
2477 #endif
2478 }
2479