master
c 1,225 lines 34.3 KB
Raw
1 /*
2 * Translation Block Maintenance
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/interval-tree.h"
22 #include "qemu/qtree.h"
23 #include "exec/cputlb.h"
24 #include "exec/log.h"
25 #include "exec/page-protection.h"
26 #include "exec/mmap-lock.h"
27 #include "exec/tb-flush.h"
28 #include "exec/target_page.h"
29 #include "accel/tcg/cpu-loop.h"
30 #include "accel/tcg/cpu-ops.h"
31 #include "tb-internal.h"
32 #include "system/tcg.h"
33 #include "tcg/tcg.h"
34 #include "tb-hash.h"
35 #include "tb-context.h"
36 #include "internal-common.h"
37 #ifdef CONFIG_USER_ONLY
38 #include "user/page-protection.h"
39 #define runstate_is_running() true
40 #else
41 #include "system-page-protection.h"
42 #include "system/runstate.h"
43 #endif
44 #include "trace.h"
45
46 /* List iterators for lists of tagged pointers in TranslationBlock. */
47 #define TB_FOR_EACH_TAGGED(head, tb, n, field) \
48 for (n = (head) & 1, tb = (TranslationBlock *)((head) & ~1); \
49 tb; tb = (TranslationBlock *)tb->field[n], n = (uintptr_t)tb & 1, \
50 tb = (TranslationBlock *)((uintptr_t)tb & ~1))
51
52 #define TB_FOR_EACH_JMP(head_tb, tb, n) \
53 TB_FOR_EACH_TAGGED((head_tb)->jmp_list_head, tb, n, jmp_list_next)
54
55 static bool tb_cmp(const void *ap, const void *bp)
56 {
57 const TranslationBlock *a = ap;
58 const TranslationBlock *b = bp;
59
60 return ((tb_cflags(a) & CF_PCREL || a->pc == b->pc) &&
61 a->cs_base == b->cs_base &&
62 a->flags == b->flags &&
63 (tb_cflags(a) & ~CF_INVALID) == (tb_cflags(b) & ~CF_INVALID) &&
64 tb_page_addr0(a) == tb_page_addr0(b) &&
65 tb_page_addr1(a) == tb_page_addr1(b));
66 }
67
68 void tb_htable_init(void)
69 {
70 unsigned int mode = QHT_MODE_AUTO_RESIZE;
71
72 qht_init(&tb_ctx.htable, tb_cmp, CODE_GEN_HTABLE_SIZE, mode);
73 }
74
75 typedef struct PageDesc PageDesc;
76
77 #ifdef CONFIG_USER_ONLY
78
79 /*
80 * In user-mode page locks aren't used; mmap_lock is enough.
81 */
82 #define assert_page_locked(pd) tcg_debug_assert(have_mmap_lock())
83
84 static inline void tb_lock_pages(const TranslationBlock *tb) { }
85
86 /*
87 * For user-only, since we are protecting all of memory with a single lock,
88 * and because the two pages of a TranslationBlock are always contiguous,
89 * use a single data structure to record all TranslationBlocks.
90 */
91 static IntervalTreeRoot tb_root;
92
93 static void tb_remove_all(void)
94 {
95 /*
96 * Only called from tb_flush__exclusive_or_serial, where we have already
97 * asserted that we're in an exclusive state.
98 */
99 memset(&tb_root, 0, sizeof(tb_root));
100 }
101
102 /* Call with mmap_lock held. */
103 static void tb_record(TranslationBlock *tb)
104 {
105 vaddr addr;
106 int flags;
107
108 assert_memory_lock();
109 tb->itree.last = tb->itree.start + tb->size - 1;
110
111 /* translator_loop() must have made all TB pages non-writable */
112 addr = tb_page_addr0(tb);
113 flags = page_get_flags(addr);
114 assert(!(flags & PAGE_WRITE));
115
116 addr = tb_page_addr1(tb);
117 if (addr != -1) {
118 flags = page_get_flags(addr);
119 assert(!(flags & PAGE_WRITE));
120 }
121
122 interval_tree_insert(&tb->itree, &tb_root);
123 }
124
125 /* Call with mmap_lock held. */
126 static void tb_remove(TranslationBlock *tb)
127 {
128 assert_memory_lock();
129 interval_tree_remove(&tb->itree, &tb_root);
130 }
131
132 /* TODO: For now, still shared with translate-all.c for system mode. */
133 #define PAGE_FOR_EACH_TB(start, last, pagedesc, T, N) \
134 for (T = foreach_tb_first(start, last), \
135 N = foreach_tb_next(T, start, last); \
136 T != NULL; \
137 T = N, N = foreach_tb_next(N, start, last))
138
139 typedef TranslationBlock *PageForEachNext;
140
141 static PageForEachNext foreach_tb_first(tb_page_addr_t start,
142 tb_page_addr_t last)
143 {
144 IntervalTreeNode *n = interval_tree_iter_first(&tb_root, start, last);
145 return n ? container_of(n, TranslationBlock, itree) : NULL;
146 }
147
148 static PageForEachNext foreach_tb_next(PageForEachNext tb,
149 tb_page_addr_t start,
150 tb_page_addr_t last)
151 {
152 IntervalTreeNode *n;
153
154 if (tb) {
155 n = interval_tree_iter_next(&tb->itree, start, last);
156 if (n) {
157 return container_of(n, TranslationBlock, itree);
158 }
159 }
160 return NULL;
161 }
162
163 #else
164 /*
165 * In system mode we want L1_MAP to be based on ram offsets.
166 */
167 #define L1_MAP_ADDR_SPACE_BITS HOST_LONG_BITS
168
169 /* Size of the L2 (and L3, etc) page tables. */
170 #define V_L2_BITS 10
171 #define V_L2_SIZE (1 << V_L2_BITS)
172
173 /*
174 * L1 Mapping properties
175 */
176 static int v_l1_size;
177 static int v_l1_shift;
178 static int v_l2_levels;
179
180 /*
181 * The bottom level has pointers to PageDesc, and is indexed by
182 * anything from 4 to (V_L2_BITS + 3) bits, depending on target page size.
183 */
184 #define V_L1_MIN_BITS 4
185 #define V_L1_MAX_BITS (V_L2_BITS + 3)
186 #define V_L1_MAX_SIZE (1 << V_L1_MAX_BITS)
187
188 static void *l1_map[V_L1_MAX_SIZE];
189
190 struct PageDesc {
191 QemuSpin lock;
192 /* list of TBs intersecting this ram page */
193 uintptr_t first_tb;
194 };
195
196 void page_table_config_init(void)
197 {
198 uint32_t v_l1_bits;
199
200 assert(TARGET_PAGE_BITS);
201 /* The bits remaining after N lower levels of page tables. */
202 v_l1_bits = (L1_MAP_ADDR_SPACE_BITS - TARGET_PAGE_BITS) % V_L2_BITS;
203 if (v_l1_bits < V_L1_MIN_BITS) {
204 v_l1_bits += V_L2_BITS;
205 }
206
207 v_l1_size = 1 << v_l1_bits;
208 v_l1_shift = L1_MAP_ADDR_SPACE_BITS - TARGET_PAGE_BITS - v_l1_bits;
209 v_l2_levels = v_l1_shift / V_L2_BITS - 1;
210
211 assert(v_l1_bits <= V_L1_MAX_BITS);
212 assert(v_l1_shift % V_L2_BITS == 0);
213 assert(v_l2_levels >= 0);
214 }
215
216 static PageDesc *page_find_alloc(tb_page_addr_t index, bool alloc)
217 {
218 PageDesc *pd;
219 void **lp;
220
221 /* Level 1. Always allocated. */
222 lp = l1_map + ((index >> v_l1_shift) & (v_l1_size - 1));
223
224 /* Level 2..N-1. */
225 for (int i = v_l2_levels; i > 0; i--) {
226 void **p = qatomic_rcu_read(lp);
227
228 if (p == NULL) {
229 void *existing;
230
231 if (!alloc) {
232 return NULL;
233 }
234 p = g_new0(void *, V_L2_SIZE);
235 existing = qatomic_cmpxchg(lp, NULL, p);
236 if (unlikely(existing)) {
237 g_free(p);
238 p = existing;
239 }
240 }
241
242 lp = p + ((index >> (i * V_L2_BITS)) & (V_L2_SIZE - 1));
243 }
244
245 pd = qatomic_rcu_read(lp);
246 if (pd == NULL) {
247 void *existing;
248
249 if (!alloc) {
250 return NULL;
251 }
252
253 pd = g_new0(PageDesc, V_L2_SIZE);
254 for (int i = 0; i < V_L2_SIZE; i++) {
255 qemu_spin_init(&pd[i].lock);
256 }
257
258 existing = qatomic_cmpxchg(lp, NULL, pd);
259 if (unlikely(existing)) {
260 for (int i = 0; i < V_L2_SIZE; i++) {
261 qemu_spin_destroy(&pd[i].lock);
262 }
263 g_free(pd);
264 pd = existing;
265 }
266 }
267
268 return pd + (index & (V_L2_SIZE - 1));
269 }
270
271 static inline PageDesc *page_find(tb_page_addr_t index)
272 {
273 return page_find_alloc(index, false);
274 }
275
276 /**
277 * struct page_entry - page descriptor entry
278 * @pd: pointer to the &struct PageDesc of the page this entry represents
279 * @index: page index of the page
280 * @locked: whether the page is locked
281 *
282 * This struct helps us keep track of the locked state of a page, without
283 * bloating &struct PageDesc.
284 *
285 * A page lock protects accesses to all fields of &struct PageDesc.
286 *
287 * See also: &struct page_collection.
288 */
289 struct page_entry {
290 PageDesc *pd;
291 tb_page_addr_t index;
292 bool locked;
293 };
294
295 /**
296 * struct page_collection - tracks a set of pages (i.e. &struct page_entry's)
297 * @tree: Binary search tree (BST) of the pages, with key == page index
298 * @max: Pointer to the page in @tree with the highest page index
299 *
300 * To avoid deadlock we lock pages in ascending order of page index.
301 * When operating on a set of pages, we need to keep track of them so that
302 * we can lock them in order and also unlock them later. For this we collect
303 * pages (i.e. &struct page_entry's) in a binary search @tree. Given that the
304 * @tree implementation we use does not provide an O(1) operation to obtain the
305 * highest-ranked element, we use @max to keep track of the inserted page
306 * with the highest index. This is valuable because if a page is not in
307 * the tree and its index is higher than @max's, then we can lock it
308 * without breaking the locking order rule.
309 *
310 * Note on naming: 'struct page_set' would be shorter, but we already have a few
311 * page_set_*() helpers, so page_collection is used instead to avoid confusion.
312 *
313 * See also: page_collection_lock().
314 */
315 struct page_collection {
316 QTree *tree;
317 struct page_entry *max;
318 };
319
320 typedef int PageForEachNext;
321 #define PAGE_FOR_EACH_TB(start, last, pagedesc, tb, n) \
322 TB_FOR_EACH_TAGGED((pagedesc)->first_tb, tb, n, page_next)
323
324 #ifdef CONFIG_DEBUG_TCG
325
326 static __thread GHashTable *ht_pages_locked_debug;
327
328 static void ht_pages_locked_debug_init(void)
329 {
330 if (ht_pages_locked_debug) {
331 return;
332 }
333 ht_pages_locked_debug = g_hash_table_new(NULL, NULL);
334 }
335
336 static bool page_is_locked(const PageDesc *pd)
337 {
338 PageDesc *found;
339
340 ht_pages_locked_debug_init();
341 found = g_hash_table_lookup(ht_pages_locked_debug, pd);
342 return !!found;
343 }
344
345 static void page_lock__debug(PageDesc *pd)
346 {
347 ht_pages_locked_debug_init();
348 g_assert(!page_is_locked(pd));
349 g_hash_table_insert(ht_pages_locked_debug, pd, pd);
350 }
351
352 static void page_unlock__debug(const PageDesc *pd)
353 {
354 bool removed;
355
356 ht_pages_locked_debug_init();
357 g_assert(page_is_locked(pd));
358 removed = g_hash_table_remove(ht_pages_locked_debug, pd);
359 g_assert(removed);
360 }
361
362 static void do_assert_page_locked(const PageDesc *pd,
363 const char *file, int line)
364 {
365 if (unlikely(!page_is_locked(pd))) {
366 error_report("assert_page_lock: PageDesc %p not locked @ %s:%d",
367 pd, file, line);
368 abort();
369 }
370 }
371 #define assert_page_locked(pd) do_assert_page_locked(pd, __FILE__, __LINE__)
372
373 void assert_no_pages_locked(void)
374 {
375 ht_pages_locked_debug_init();
376 g_assert(g_hash_table_size(ht_pages_locked_debug) == 0);
377 }
378
379 #else /* !CONFIG_DEBUG_TCG */
380
381 static inline void page_lock__debug(const PageDesc *pd) { }
382 static inline void page_unlock__debug(const PageDesc *pd) { }
383 static inline void assert_page_locked(const PageDesc *pd) { }
384
385 #endif /* CONFIG_DEBUG_TCG */
386
387 static void page_lock(PageDesc *pd)
388 {
389 page_lock__debug(pd);
390 qemu_spin_lock(&pd->lock);
391 }
392
393 /* Like qemu_spin_trylock, returns false on success */
394 static bool page_trylock(PageDesc *pd)
395 {
396 bool busy = qemu_spin_trylock(&pd->lock);
397 if (!busy) {
398 page_lock__debug(pd);
399 }
400 return busy;
401 }
402
403 static void page_unlock(PageDesc *pd)
404 {
405 qemu_spin_unlock(&pd->lock);
406 page_unlock__debug(pd);
407 }
408
409 void tb_lock_page0(tb_page_addr_t paddr)
410 {
411 page_lock(page_find_alloc(paddr >> TARGET_PAGE_BITS, true));
412 }
413
414 void tb_lock_page1(tb_page_addr_t paddr0, tb_page_addr_t paddr1)
415 {
416 tb_page_addr_t pindex0 = paddr0 >> TARGET_PAGE_BITS;
417 tb_page_addr_t pindex1 = paddr1 >> TARGET_PAGE_BITS;
418 PageDesc *pd0, *pd1;
419
420 if (pindex0 == pindex1) {
421 /* Identical pages, and the first page is already locked. */
422 return;
423 }
424
425 pd1 = page_find_alloc(pindex1, true);
426 if (pindex0 < pindex1) {
427 /* Correct locking order, we may block. */
428 page_lock(pd1);
429 return;
430 }
431
432 /* Incorrect locking order, we cannot block lest we deadlock. */
433 if (!page_trylock(pd1)) {
434 return;
435 }
436
437 /*
438 * Drop the lock on page0 and get both page locks in the right order.
439 * Restart translation via longjmp.
440 */
441 pd0 = page_find_alloc(pindex0, false);
442 page_unlock(pd0);
443 page_lock(pd1);
444 page_lock(pd0);
445 siglongjmp(tcg_ctx->jmp_trans, -3);
446 }
447
448 void tb_unlock_page1(tb_page_addr_t paddr0, tb_page_addr_t paddr1)
449 {
450 tb_page_addr_t pindex0 = paddr0 >> TARGET_PAGE_BITS;
451 tb_page_addr_t pindex1 = paddr1 >> TARGET_PAGE_BITS;
452
453 if (pindex0 != pindex1) {
454 page_unlock(page_find_alloc(pindex1, false));
455 }
456 }
457
458 static void tb_lock_pages(TranslationBlock *tb)
459 {
460 tb_page_addr_t paddr0 = tb_page_addr0(tb);
461 tb_page_addr_t paddr1 = tb_page_addr1(tb);
462 tb_page_addr_t pindex0 = paddr0 >> TARGET_PAGE_BITS;
463 tb_page_addr_t pindex1 = paddr1 >> TARGET_PAGE_BITS;
464
465 if (unlikely(paddr0 == -1)) {
466 return;
467 }
468 if (unlikely(paddr1 != -1) && pindex0 != pindex1) {
469 if (pindex0 < pindex1) {
470 page_lock(page_find_alloc(pindex0, true));
471 page_lock(page_find_alloc(pindex1, true));
472 return;
473 }
474 page_lock(page_find_alloc(pindex1, true));
475 }
476 page_lock(page_find_alloc(pindex0, true));
477 }
478
479 void tb_unlock_pages(TranslationBlock *tb)
480 {
481 tb_page_addr_t paddr0 = tb_page_addr0(tb);
482 tb_page_addr_t paddr1 = tb_page_addr1(tb);
483 tb_page_addr_t pindex0 = paddr0 >> TARGET_PAGE_BITS;
484 tb_page_addr_t pindex1 = paddr1 >> TARGET_PAGE_BITS;
485
486 if (unlikely(paddr0 == -1)) {
487 return;
488 }
489 if (unlikely(paddr1 != -1) && pindex0 != pindex1) {
490 page_unlock(page_find_alloc(pindex1, false));
491 }
492 page_unlock(page_find_alloc(pindex0, false));
493 }
494
495 static inline struct page_entry *
496 page_entry_new(PageDesc *pd, tb_page_addr_t index)
497 {
498 struct page_entry *pe = g_malloc(sizeof(*pe));
499
500 pe->index = index;
501 pe->pd = pd;
502 pe->locked = false;
503 return pe;
504 }
505
506 static void page_entry_destroy(gpointer p)
507 {
508 struct page_entry *pe = p;
509
510 g_assert(pe->locked);
511 page_unlock(pe->pd);
512 g_free(pe);
513 }
514
515 /* returns false on success */
516 static bool page_entry_trylock(struct page_entry *pe)
517 {
518 bool busy = page_trylock(pe->pd);
519 if (!busy) {
520 g_assert(!pe->locked);
521 pe->locked = true;
522 }
523 return busy;
524 }
525
526 static void do_page_entry_lock(struct page_entry *pe)
527 {
528 page_lock(pe->pd);
529 g_assert(!pe->locked);
530 pe->locked = true;
531 }
532
533 static gboolean page_entry_lock(gpointer key, gpointer value, gpointer data)
534 {
535 struct page_entry *pe = value;
536
537 do_page_entry_lock(pe);
538 return FALSE;
539 }
540
541 static gboolean page_entry_unlock(gpointer key, gpointer value, gpointer data)
542 {
543 struct page_entry *pe = value;
544
545 if (pe->locked) {
546 pe->locked = false;
547 page_unlock(pe->pd);
548 }
549 return FALSE;
550 }
551
552 /*
553 * Trylock a page, and if successful, add the page to a collection.
554 * Returns true ("busy") if the page could not be locked; false otherwise.
555 */
556 static bool page_trylock_add(struct page_collection *set, tb_page_addr_t addr)
557 {
558 tb_page_addr_t index = addr >> TARGET_PAGE_BITS;
559 struct page_entry *pe;
560 PageDesc *pd;
561
562 pe = q_tree_lookup(set->tree, &index);
563 if (pe) {
564 return false;
565 }
566
567 pd = page_find(index);
568 if (pd == NULL) {
569 return false;
570 }
571
572 pe = page_entry_new(pd, index);
573 q_tree_insert(set->tree, &pe->index, pe);
574
575 /*
576 * If this is either (1) the first insertion or (2) a page whose index
577 * is higher than any other so far, just lock the page and move on.
578 */
579 if (set->max == NULL || pe->index > set->max->index) {
580 set->max = pe;
581 do_page_entry_lock(pe);
582 return false;
583 }
584 /*
585 * Try to acquire out-of-order lock; if busy, return busy so that we acquire
586 * locks in order.
587 */
588 return page_entry_trylock(pe);
589 }
590
591 static gint tb_page_addr_cmp(gconstpointer ap, gconstpointer bp, gpointer udata)
592 {
593 tb_page_addr_t a = *(const tb_page_addr_t *)ap;
594 tb_page_addr_t b = *(const tb_page_addr_t *)bp;
595
596 if (a == b) {
597 return 0;
598 } else if (a < b) {
599 return -1;
600 }
601 return 1;
602 }
603
604 /*
605 * Lock a range of pages ([@start,@last]) as well as the pages of all
606 * intersecting TBs.
607 * Locking order: acquire locks in ascending order of page index.
608 */
609 static struct page_collection *page_collection_lock(tb_page_addr_t start,
610 tb_page_addr_t last)
611 {
612 struct page_collection *set = g_malloc(sizeof(*set));
613 tb_page_addr_t index;
614 PageDesc *pd;
615
616 start >>= TARGET_PAGE_BITS;
617 last >>= TARGET_PAGE_BITS;
618 g_assert(start <= last);
619
620 set->tree = q_tree_new_full(tb_page_addr_cmp, NULL, NULL,
621 page_entry_destroy);
622 set->max = NULL;
623 assert_no_pages_locked();
624
625 retry:
626 q_tree_foreach(set->tree, page_entry_lock, NULL);
627
628 for (index = start; index <= last; index++) {
629 TranslationBlock *tb;
630 PageForEachNext n;
631
632 pd = page_find(index);
633 if (pd == NULL) {
634 continue;
635 }
636 if (page_trylock_add(set, index << TARGET_PAGE_BITS)) {
637 q_tree_foreach(set->tree, page_entry_unlock, NULL);
638 goto retry;
639 }
640 assert_page_locked(pd);
641 PAGE_FOR_EACH_TB(unused, unused, pd, tb, n) {
642 if (page_trylock_add(set, tb_page_addr0(tb)) ||
643 (tb_page_addr1(tb) != -1 &&
644 page_trylock_add(set, tb_page_addr1(tb)))) {
645 /* drop all locks, and reacquire in order */
646 q_tree_foreach(set->tree, page_entry_unlock, NULL);
647 goto retry;
648 }
649 }
650 }
651 return set;
652 }
653
654 static void page_collection_unlock(struct page_collection *set)
655 {
656 /* entries are unlocked and freed via page_entry_destroy */
657 q_tree_destroy(set->tree);
658 g_free(set);
659 }
660
661 /* Set to NULL all the 'first_tb' fields in all PageDescs. */
662 static void tb_remove_all_1(int level, void **lp)
663 {
664 int i;
665
666 if (*lp == NULL) {
667 return;
668 }
669 if (level == 0) {
670 PageDesc *pd = *lp;
671
672 for (i = 0; i < V_L2_SIZE; ++i) {
673 page_lock(&pd[i]);
674 pd[i].first_tb = (uintptr_t)NULL;
675 page_unlock(&pd[i]);
676 }
677 } else {
678 void **pp = *lp;
679
680 for (i = 0; i < V_L2_SIZE; ++i) {
681 tb_remove_all_1(level - 1, pp + i);
682 }
683 }
684 }
685
686 static void tb_remove_all(void)
687 {
688 int i, l1_sz = v_l1_size;
689
690 for (i = 0; i < l1_sz; i++) {
691 tb_remove_all_1(v_l2_levels, l1_map + i);
692 }
693 }
694
695 /*
696 * Add the tb in the target page and protect it if necessary.
697 * Called with @p->lock held.
698 */
699 static void tb_page_add(PageDesc *p, TranslationBlock *tb, unsigned int n)
700 {
701 bool page_already_protected;
702
703 assert_page_locked(p);
704
705 tb->page_next[n] = p->first_tb;
706 page_already_protected = p->first_tb != 0;
707 p->first_tb = (uintptr_t)tb | n;
708
709 /*
710 * If some code is already present, then the pages are already
711 * protected. So we handle the case where only the first TB is
712 * allocated in a physical page.
713 */
714 if (!page_already_protected) {
715 tlb_protect_code(tb->page_addr[n] & TARGET_PAGE_MASK);
716 }
717 }
718
719 static void tb_record(TranslationBlock *tb)
720 {
721 tb_page_addr_t paddr0 = tb_page_addr0(tb);
722 tb_page_addr_t paddr1 = tb_page_addr1(tb);
723 tb_page_addr_t pindex0 = paddr0 >> TARGET_PAGE_BITS;
724 tb_page_addr_t pindex1 = paddr1 >> TARGET_PAGE_BITS;
725
726 assert(paddr0 != -1);
727 if (unlikely(paddr1 != -1) && pindex0 != pindex1) {
728 tb_page_add(page_find_alloc(pindex1, false), tb, 1);
729 }
730 tb_page_add(page_find_alloc(pindex0, false), tb, 0);
731 }
732
733 static void tb_page_remove(PageDesc *pd, TranslationBlock *tb)
734 {
735 TranslationBlock *tb1;
736 uintptr_t *pprev;
737 PageForEachNext n1;
738
739 assert_page_locked(pd);
740 pprev = &pd->first_tb;
741 PAGE_FOR_EACH_TB(unused, unused, pd, tb1, n1) {
742 if (tb1 == tb) {
743 *pprev = tb1->page_next[n1];
744 return;
745 }
746 pprev = &tb1->page_next[n1];
747 }
748 g_assert_not_reached();
749 }
750
751 static void tb_remove(TranslationBlock *tb)
752 {
753 tb_page_addr_t paddr0 = tb_page_addr0(tb);
754 tb_page_addr_t paddr1 = tb_page_addr1(tb);
755 tb_page_addr_t pindex0 = paddr0 >> TARGET_PAGE_BITS;
756 tb_page_addr_t pindex1 = paddr1 >> TARGET_PAGE_BITS;
757
758 assert(paddr0 != -1);
759 if (unlikely(paddr1 != -1) && pindex0 != pindex1) {
760 tb_page_remove(page_find_alloc(pindex1, false), tb);
761 }
762 tb_page_remove(page_find_alloc(pindex0, false), tb);
763 }
764 #endif /* CONFIG_USER_ONLY */
765
766 /*
767 * Flush all the translation blocks.
768 * Must be called from a context in which no cpus are running,
769 * e.g. start_exclusive() or vm_stop().
770 */
771 void tb_flush__exclusive_or_serial(void)
772 {
773 CPUState *cpu;
774
775 trace_tb_flush();
776 assert(tcg_enabled());
777 /* Note that cpu_in_serial_context checks cpu_in_exclusive_context. */
778 assert(!runstate_is_running() ||
779 (current_cpu && cpu_in_serial_context(current_cpu)));
780
781 CPU_FOREACH(cpu) {
782 tcg_flush_jmp_cache(cpu);
783 }
784
785 qht_reset_size(&tb_ctx.htable, CODE_GEN_HTABLE_SIZE);
786 tb_remove_all();
787
788 tcg_region_reset_all();
789 /* XXX: flush processor icache at this point if cache flush is expensive */
790 qatomic_inc(&tb_ctx.tb_flush_count);
791 qemu_plugin_flush_cb();
792 }
793
794 static void do_tb_flush(CPUState *cpu, run_on_cpu_data tb_flush_count)
795 {
796 /* If it is already been done on request of another CPU, just retry. */
797 if (tb_ctx.tb_flush_count == tb_flush_count.host_int) {
798 tb_flush__exclusive_or_serial();
799 }
800 }
801
802 void queue_tb_flush(CPUState *cs)
803 {
804 if (tcg_enabled()) {
805 unsigned tb_flush_count = qatomic_read(&tb_ctx.tb_flush_count);
806 async_safe_run_on_cpu(cs, do_tb_flush,
807 RUN_ON_CPU_HOST_INT(tb_flush_count));
808 }
809 }
810
811 /* remove @orig from its @n_orig-th jump list */
812 static inline void tb_remove_from_jmp_list(TranslationBlock *orig, int n_orig)
813 {
814 uintptr_t ptr, ptr_locked;
815 TranslationBlock *dest;
816 TranslationBlock *tb;
817 uintptr_t *pprev;
818 int n;
819
820 /* mark the LSB of jmp_dest[] so that no further jumps can be inserted */
821 ptr = qatomic_or_fetch(&orig->jmp_dest[n_orig], 1);
822 dest = (TranslationBlock *)(ptr & ~1);
823 if (dest == NULL) {
824 return;
825 }
826
827 qemu_spin_lock(&dest->jmp_lock);
828 /*
829 * While acquiring the lock, the jump might have been removed if the
830 * destination TB was invalidated; check again.
831 */
832 ptr_locked = qatomic_read(&orig->jmp_dest[n_orig]);
833 if (ptr_locked != ptr) {
834 qemu_spin_unlock(&dest->jmp_lock);
835 /*
836 * The only possibility is that the jump was unlinked via
837 * tb_jump_unlink(dest). Seeing here another destination would be a bug,
838 * because we set the LSB above.
839 */
840 g_assert(ptr_locked == 1 && dest->cflags & CF_INVALID);
841 return;
842 }
843 /*
844 * We first acquired the lock, and since the destination pointer matches,
845 * we know for sure that @orig is in the jmp list.
846 */
847 if (dest == orig) {
848 /*
849 * In the case of a TB that links to itself, removing the entry
850 * from the list means that it won't be present later during
851 * tb_jmp_unlink -- unlink now.
852 */
853 tb_reset_jump(orig, n_orig);
854 }
855 pprev = &dest->jmp_list_head;
856 TB_FOR_EACH_JMP(dest, tb, n) {
857 if (tb == orig && n == n_orig) {
858 *pprev = tb->jmp_list_next[n];
859 /* no need to set orig->jmp_dest[n]; setting the LSB was enough */
860 qemu_spin_unlock(&dest->jmp_lock);
861 return;
862 }
863 pprev = &tb->jmp_list_next[n];
864 }
865 g_assert_not_reached();
866 }
867
868 /*
869 * Reset the jump entry 'n' of a TB so that it is not chained to another TB.
870 */
871 void tb_reset_jump(TranslationBlock *tb, int n)
872 {
873 uintptr_t addr = (uintptr_t)(tb->tc.ptr + tb->jmp_reset_offset[n]);
874 tb_set_jmp_target(tb, n, addr);
875 }
876
877 /* remove any jumps to the TB */
878 static inline void tb_jmp_unlink(TranslationBlock *dest)
879 {
880 TranslationBlock *tb;
881 int n;
882
883 qemu_spin_lock(&dest->jmp_lock);
884
885 TB_FOR_EACH_JMP(dest, tb, n) {
886 tb_reset_jump(tb, n);
887 qatomic_and(&tb->jmp_dest[n], (uintptr_t)NULL | 1);
888 /* No need to clear the list entry; setting the dest ptr is enough */
889 }
890 dest->jmp_list_head = (uintptr_t)NULL;
891
892 qemu_spin_unlock(&dest->jmp_lock);
893 }
894
895 static void tb_jmp_cache_inval_tb(TranslationBlock *tb)
896 {
897 CPUState *cpu;
898
899 if (tb_cflags(tb) & CF_PCREL) {
900 /* A TB may be at any virtual address */
901 CPU_FOREACH(cpu) {
902 tcg_flush_jmp_cache(cpu);
903 }
904 } else {
905 uint32_t h = tb_jmp_cache_hash_func(tb->pc);
906
907 CPU_FOREACH(cpu) {
908 CPUJumpCache *jc = cpu->tb_jmp_cache;
909
910 if (qatomic_read(&jc->array[h].tb) == tb) {
911 qatomic_set(&jc->array[h].tb, NULL);
912 }
913 }
914 }
915 }
916
917 /*
918 * In user-mode, call with mmap_lock held.
919 * In !user-mode, if @rm_from_page_list is set, call with the TB's pages'
920 * locks held.
921 */
922 static void do_tb_phys_invalidate(TranslationBlock *tb, bool rm_from_page_list)
923 {
924 uint32_t h;
925 tb_page_addr_t phys_pc;
926 uint32_t orig_cflags = tb_cflags(tb);
927
928 assert_memory_lock();
929 qemu_thread_jit_write();
930
931 /* make sure no further incoming jumps will be chained to this TB */
932 qemu_spin_lock(&tb->jmp_lock);
933 qatomic_set(&tb->cflags, tb->cflags | CF_INVALID);
934 qemu_spin_unlock(&tb->jmp_lock);
935
936 /* remove the TB from the hash list */
937 phys_pc = tb_page_addr0(tb);
938 h = tb_hash_func(phys_pc, (orig_cflags & CF_PCREL ? 0 : tb->pc),
939 tb->flags, tb->cs_base, orig_cflags);
940 if (qht_remove(&tb_ctx.htable, tb, h)) {
941
942 /* remove the TB from the page list */
943 if (rm_from_page_list) {
944 tb_remove(tb);
945 }
946
947 /* remove the TB from the hash list */
948 tb_jmp_cache_inval_tb(tb);
949
950 /* suppress this TB from the two jump lists */
951 tb_remove_from_jmp_list(tb, 0);
952 tb_remove_from_jmp_list(tb, 1);
953
954 /* suppress any remaining jumps to this TB */
955 tb_jmp_unlink(tb);
956
957 qatomic_set(&tb_ctx.tb_phys_invalidate_count,
958 tb_ctx.tb_phys_invalidate_count + 1);
959 }
960
961 qemu_thread_jit_execute();
962 }
963
964 /*
965 * Invalidate one TB.
966 * Called with mmap_lock held in user-mode.
967 */
968 void tb_phys_invalidate(TranslationBlock *tb, tb_page_addr_t page_addr)
969 {
970 if (page_addr == -1 && tb_page_addr0(tb) != -1) {
971 tb_lock_pages(tb);
972 do_tb_phys_invalidate(tb, true);
973 tb_unlock_pages(tb);
974 } else {
975 do_tb_phys_invalidate(tb, false);
976 }
977 }
978
979 /*
980 * Add a new TB and link it to the physical page tables.
981 * Called with mmap_lock held for user-mode emulation.
982 *
983 * Returns a pointer @tb, or a pointer to an existing TB that matches @tb.
984 * Note that in !user-mode, another thread might have already added a TB
985 * for the same block of guest code that @tb corresponds to. In that case,
986 * the caller should discard the original @tb, and use instead the returned TB.
987 */
988 TranslationBlock *tb_link_page(TranslationBlock *tb)
989 {
990 void *existing_tb = NULL;
991 uint32_t h;
992
993 assert_memory_lock();
994 tcg_debug_assert(!(tb->cflags & CF_INVALID));
995
996 tb_record(tb);
997
998 /* add in the hash table */
999 h = tb_hash_func(tb_page_addr0(tb), (tb->cflags & CF_PCREL ? 0 : tb->pc),
1000 tb->flags, tb->cs_base, tb->cflags);
1001 qht_insert(&tb_ctx.htable, tb, h, &existing_tb);
1002
1003 /* remove TB from the page(s) if we couldn't insert it */
1004 if (unlikely(existing_tb)) {
1005 tb_remove(tb);
1006 tb_unlock_pages(tb);
1007 return existing_tb;
1008 }
1009
1010 tb_unlock_pages(tb);
1011 return tb;
1012 }
1013
1014 #ifdef CONFIG_USER_ONLY
1015 /*
1016 * Invalidate all TBs which intersect with the target address range.
1017 * Called with mmap_lock held for user-mode emulation.
1018 * NOTE: this function must not be called while a TB is running.
1019 */
1020 void tb_invalidate_phys_range(CPUState *cpu, tb_page_addr_t start,
1021 tb_page_addr_t last)
1022 {
1023 TranslationBlock *tb;
1024 PageForEachNext n;
1025
1026 assert_memory_lock();
1027
1028 PAGE_FOR_EACH_TB(start, last, unused, tb, n) {
1029 do_tb_phys_invalidate(tb, true);
1030 }
1031 }
1032
1033 /*
1034 * Invalidate all TBs which intersect with the target address page @addr.
1035 * Called with mmap_lock held for user-mode emulation
1036 * NOTE: this function must not be called while a TB is running.
1037 */
1038 static void tb_invalidate_phys_page(tb_page_addr_t addr)
1039 {
1040 tb_page_addr_t start, last;
1041
1042 start = addr & TARGET_PAGE_MASK;
1043 last = addr | ~TARGET_PAGE_MASK;
1044 tb_invalidate_phys_range(NULL, start, last);
1045 }
1046
1047 /*
1048 * Called with mmap_lock held. If pc is not 0 then it indicates the
1049 * host PC of the faulting store instruction that caused this invalidate.
1050 * Returns true if the caller needs to abort execution of the current TB.
1051 */
1052 bool tb_invalidate_phys_page_unwind(CPUState *cpu, tb_page_addr_t addr,
1053 uintptr_t pc)
1054 {
1055 TranslationBlock *current_tb;
1056 bool current_tb_modified;
1057 TranslationBlock *tb;
1058 PageForEachNext n;
1059 tb_page_addr_t last;
1060
1061 /*
1062 * Without precise smc semantics, or when outside of a TB,
1063 * we can skip to invalidate.
1064 */
1065 if (!pc || !cpu || !cpu->cc->tcg_ops->precise_smc) {
1066 tb_invalidate_phys_page(addr);
1067 return false;
1068 }
1069
1070 assert_memory_lock();
1071 current_tb = tcg_tb_lookup(pc);
1072
1073 last = addr | ~TARGET_PAGE_MASK;
1074 addr &= TARGET_PAGE_MASK;
1075 current_tb_modified = false;
1076
1077 PAGE_FOR_EACH_TB(addr, last, unused, tb, n) {
1078 if (current_tb == tb &&
1079 (tb_cflags(current_tb) & CF_COUNT_MASK) != 1) {
1080 /*
1081 * If we are modifying the current TB, we must stop its
1082 * execution. We could be more precise by checking that
1083 * the modification is after the current PC, but it would
1084 * require a specialized function to partially restore
1085 * the CPU state.
1086 */
1087 current_tb_modified = true;
1088 cpu_restore_state_from_tb(cpu, current_tb, pc);
1089 }
1090 do_tb_phys_invalidate(tb, true);
1091 }
1092
1093 if (current_tb_modified) {
1094 /* Force execution of one insn next time. */
1095 cpu->cflags_next_tb = 1 | CF_NOIRQ | curr_cflags(cpu);
1096 return true;
1097 }
1098 return false;
1099 }
1100 #else
1101 /*
1102 * @p must be non-NULL.
1103 * Call with all @pages locked.
1104 * (@cpu, @retaddr) may be (NULL, 0) outside of a cpu context,
1105 * in which case precise_smc need not be detected.
1106 */
1107 static void
1108 tb_invalidate_phys_page_range__locked(CPUState *cpu,
1109 struct page_collection *pages,
1110 PageDesc *p, tb_page_addr_t start,
1111 tb_page_addr_t last,
1112 uintptr_t retaddr)
1113 {
1114 TranslationBlock *tb;
1115 PageForEachNext n;
1116 bool current_tb_modified = false;
1117 TranslationBlock *current_tb = NULL;
1118
1119 /* Range may not cross a page. */
1120 tcg_debug_assert(((start ^ last) & TARGET_PAGE_MASK) == 0);
1121
1122 if (retaddr && cpu && cpu->cc->tcg_ops->precise_smc) {
1123 current_tb = tcg_tb_lookup(retaddr);
1124 }
1125
1126 /*
1127 * We remove all the TBs in the range [start, last].
1128 * XXX: see if in some cases it could be faster to invalidate all the code
1129 */
1130 PAGE_FOR_EACH_TB(start, last, p, tb, n) {
1131 tb_page_addr_t tb_start, tb_last;
1132
1133 /* NOTE: this is subtle as a TB may span two physical pages */
1134 tb_start = tb_page_addr0(tb);
1135 tb_last = tb_start + tb->size - 1;
1136 if (n == 0) {
1137 tb_last = MIN(tb_last, tb_start | ~TARGET_PAGE_MASK);
1138 } else {
1139 tb_start = tb_page_addr1(tb);
1140 tb_last = tb_start + (tb_last & ~TARGET_PAGE_MASK);
1141 }
1142 if (!(tb_last < start || tb_start > last)) {
1143 if (unlikely(current_tb == tb) &&
1144 (tb_cflags(current_tb) & CF_COUNT_MASK) != 1) {
1145 /*
1146 * If we are modifying the current TB, we must stop
1147 * its execution. We could be more precise by checking
1148 * that the modification is after the current PC, but it
1149 * would require a specialized function to partially
1150 * restore the CPU state.
1151 */
1152 current_tb_modified = true;
1153 cpu_restore_state_from_tb(cpu, current_tb, retaddr);
1154 }
1155 do_tb_phys_invalidate(tb, true);
1156 }
1157 }
1158
1159 /* if no code remaining, no need to continue to use slow writes */
1160 if (!p->first_tb) {
1161 tlb_unprotect_code(start);
1162 }
1163
1164 if (unlikely(current_tb_modified)) {
1165 page_collection_unlock(pages);
1166 /* Force execution of one insn next time. */
1167 cpu->cflags_next_tb = 1 | CF_NOIRQ | curr_cflags(cpu);
1168 cpu_loop_exit_noexc(cpu);
1169 }
1170 }
1171
1172 /*
1173 * Invalidate all TBs which intersect with the target physical address range
1174 * [start;last]. NOTE: start and end may refer to *different* physical pages.
1175 * 'is_cpu_write_access' should be true if called from a real cpu write
1176 * access: the virtual CPU will exit the current TB if code is modified inside
1177 * this TB.
1178 */
1179 void tb_invalidate_phys_range(CPUState *cpu, tb_page_addr_t start,
1180 tb_page_addr_t last)
1181 {
1182 struct page_collection *pages;
1183 tb_page_addr_t index, index_last;
1184
1185 pages = page_collection_lock(start, last);
1186
1187 index_last = last >> TARGET_PAGE_BITS;
1188 for (index = start >> TARGET_PAGE_BITS; index <= index_last; index++) {
1189 PageDesc *pd = page_find(index);
1190 tb_page_addr_t page_start, page_last;
1191
1192 if (pd == NULL) {
1193 continue;
1194 }
1195 assert_page_locked(pd);
1196 page_start = index << TARGET_PAGE_BITS;
1197 page_last = page_start | ~TARGET_PAGE_MASK;
1198 page_last = MIN(page_last, last);
1199 tb_invalidate_phys_page_range__locked(cpu, pages, pd,
1200 page_start, page_last, 0);
1201 }
1202 page_collection_unlock(pages);
1203 }
1204
1205 /*
1206 * len must be <= 8 and start must be a multiple of len.
1207 * Called via softmmu_template.h when code areas are written to with
1208 * iothread mutex not held.
1209 */
1210 void tb_invalidate_phys_range_fast(CPUState *cpu, ram_addr_t start,
1211 unsigned len, uintptr_t ra)
1212 {
1213 PageDesc *p = page_find(start >> TARGET_PAGE_BITS);
1214
1215 if (p) {
1216 ram_addr_t last = start + len - 1;
1217 struct page_collection *pages = page_collection_lock(start, last);
1218
1219 tb_invalidate_phys_page_range__locked(cpu, pages, p,
1220 start, last, ra);
1221 page_collection_unlock(pages);
1222 }
1223 }
1224
1225 #endif /* CONFIG_USER_ONLY */