xref: /openbmc/qemu/accel/tcg/cputlb.c (revision 21063bce)
1 /*
2  *  Common CPU TLB handling
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/main-loop.h"
22 #include "hw/core/tcg-cpu-ops.h"
23 #include "exec/exec-all.h"
24 #include "exec/memory.h"
25 #include "exec/cpu_ldst.h"
26 #include "exec/cputlb.h"
27 #include "exec/memory-internal.h"
28 #include "exec/ram_addr.h"
29 #include "tcg/tcg.h"
30 #include "qemu/error-report.h"
31 #include "exec/log.h"
32 #include "exec/helper-proto.h"
33 #include "qemu/atomic.h"
34 #include "qemu/atomic128.h"
35 #include "exec/translate-all.h"
36 #include "trace.h"
37 #include "tb-hash.h"
38 #include "internal.h"
39 #ifdef CONFIG_PLUGIN
40 #include "qemu/plugin-memory.h"
41 #endif
42 #include "tcg/tcg-ldst.h"
43 
44 /* DEBUG defines, enable DEBUG_TLB_LOG to log to the CPU_LOG_MMU target */
45 /* #define DEBUG_TLB */
46 /* #define DEBUG_TLB_LOG */
47 
48 #ifdef DEBUG_TLB
49 # define DEBUG_TLB_GATE 1
50 # ifdef DEBUG_TLB_LOG
51 #  define DEBUG_TLB_LOG_GATE 1
52 # else
53 #  define DEBUG_TLB_LOG_GATE 0
54 # endif
55 #else
56 # define DEBUG_TLB_GATE 0
57 # define DEBUG_TLB_LOG_GATE 0
58 #endif
59 
60 #define tlb_debug(fmt, ...) do { \
61     if (DEBUG_TLB_LOG_GATE) { \
62         qemu_log_mask(CPU_LOG_MMU, "%s: " fmt, __func__, \
63                       ## __VA_ARGS__); \
64     } else if (DEBUG_TLB_GATE) { \
65         fprintf(stderr, "%s: " fmt, __func__, ## __VA_ARGS__); \
66     } \
67 } while (0)
68 
69 #define assert_cpu_is_self(cpu) do {                              \
70         if (DEBUG_TLB_GATE) {                                     \
71             g_assert(!(cpu)->created || qemu_cpu_is_self(cpu));   \
72         }                                                         \
73     } while (0)
74 
75 /* run_on_cpu_data.target_ptr should always be big enough for a
76  * target_ulong even on 32 bit builds */
77 QEMU_BUILD_BUG_ON(sizeof(target_ulong) > sizeof(run_on_cpu_data));
78 
79 /* We currently can't handle more than 16 bits in the MMUIDX bitmask.
80  */
81 QEMU_BUILD_BUG_ON(NB_MMU_MODES > 16);
82 #define ALL_MMUIDX_BITS ((1 << NB_MMU_MODES) - 1)
83 
84 static inline size_t tlb_n_entries(CPUTLBDescFast *fast)
85 {
86     return (fast->mask >> CPU_TLB_ENTRY_BITS) + 1;
87 }
88 
89 static inline size_t sizeof_tlb(CPUTLBDescFast *fast)
90 {
91     return fast->mask + (1 << CPU_TLB_ENTRY_BITS);
92 }
93 
94 static void tlb_window_reset(CPUTLBDesc *desc, int64_t ns,
95                              size_t max_entries)
96 {
97     desc->window_begin_ns = ns;
98     desc->window_max_entries = max_entries;
99 }
100 
101 static void tb_jmp_cache_clear_page(CPUState *cpu, target_ulong page_addr)
102 {
103     CPUJumpCache *jc = cpu->tb_jmp_cache;
104     int i, i0;
105 
106     if (unlikely(!jc)) {
107         return;
108     }
109 
110     i0 = tb_jmp_cache_hash_page(page_addr);
111     for (i = 0; i < TB_JMP_PAGE_SIZE; i++) {
112         qatomic_set(&jc->array[i0 + i].tb, NULL);
113     }
114 }
115 
116 /**
117  * tlb_mmu_resize_locked() - perform TLB resize bookkeeping; resize if necessary
118  * @desc: The CPUTLBDesc portion of the TLB
119  * @fast: The CPUTLBDescFast portion of the same TLB
120  *
121  * Called with tlb_lock_held.
122  *
123  * We have two main constraints when resizing a TLB: (1) we only resize it
124  * on a TLB flush (otherwise we'd have to take a perf hit by either rehashing
125  * the array or unnecessarily flushing it), which means we do not control how
126  * frequently the resizing can occur; (2) we don't have access to the guest's
127  * future scheduling decisions, and therefore have to decide the magnitude of
128  * the resize based on past observations.
129  *
130  * In general, a memory-hungry process can benefit greatly from an appropriately
131  * sized TLB, since a guest TLB miss is very expensive. This doesn't mean that
132  * we just have to make the TLB as large as possible; while an oversized TLB
133  * results in minimal TLB miss rates, it also takes longer to be flushed
134  * (flushes can be _very_ frequent), and the reduced locality can also hurt
135  * performance.
136  *
137  * To achieve near-optimal performance for all kinds of workloads, we:
138  *
139  * 1. Aggressively increase the size of the TLB when the use rate of the
140  * TLB being flushed is high, since it is likely that in the near future this
141  * memory-hungry process will execute again, and its memory hungriness will
142  * probably be similar.
143  *
144  * 2. Slowly reduce the size of the TLB as the use rate declines over a
145  * reasonably large time window. The rationale is that if in such a time window
146  * we have not observed a high TLB use rate, it is likely that we won't observe
147  * it in the near future. In that case, once a time window expires we downsize
148  * the TLB to match the maximum use rate observed in the window.
149  *
150  * 3. Try to keep the maximum use rate in a time window in the 30-70% range,
151  * since in that range performance is likely near-optimal. Recall that the TLB
152  * is direct mapped, so we want the use rate to be low (or at least not too
153  * high), since otherwise we are likely to have a significant amount of
154  * conflict misses.
155  */
156 static void tlb_mmu_resize_locked(CPUTLBDesc *desc, CPUTLBDescFast *fast,
157                                   int64_t now)
158 {
159     size_t old_size = tlb_n_entries(fast);
160     size_t rate;
161     size_t new_size = old_size;
162     int64_t window_len_ms = 100;
163     int64_t window_len_ns = window_len_ms * 1000 * 1000;
164     bool window_expired = now > desc->window_begin_ns + window_len_ns;
165 
166     if (desc->n_used_entries > desc->window_max_entries) {
167         desc->window_max_entries = desc->n_used_entries;
168     }
169     rate = desc->window_max_entries * 100 / old_size;
170 
171     if (rate > 70) {
172         new_size = MIN(old_size << 1, 1 << CPU_TLB_DYN_MAX_BITS);
173     } else if (rate < 30 && window_expired) {
174         size_t ceil = pow2ceil(desc->window_max_entries);
175         size_t expected_rate = desc->window_max_entries * 100 / ceil;
176 
177         /*
178          * Avoid undersizing when the max number of entries seen is just below
179          * a pow2. For instance, if max_entries == 1025, the expected use rate
180          * would be 1025/2048==50%. However, if max_entries == 1023, we'd get
181          * 1023/1024==99.9% use rate, so we'd likely end up doubling the size
182          * later. Thus, make sure that the expected use rate remains below 70%.
183          * (and since we double the size, that means the lowest rate we'd
184          * expect to get is 35%, which is still in the 30-70% range where
185          * we consider that the size is appropriate.)
186          */
187         if (expected_rate > 70) {
188             ceil *= 2;
189         }
190         new_size = MAX(ceil, 1 << CPU_TLB_DYN_MIN_BITS);
191     }
192 
193     if (new_size == old_size) {
194         if (window_expired) {
195             tlb_window_reset(desc, now, desc->n_used_entries);
196         }
197         return;
198     }
199 
200     g_free(fast->table);
201     g_free(desc->fulltlb);
202 
203     tlb_window_reset(desc, now, 0);
204     /* desc->n_used_entries is cleared by the caller */
205     fast->mask = (new_size - 1) << CPU_TLB_ENTRY_BITS;
206     fast->table = g_try_new(CPUTLBEntry, new_size);
207     desc->fulltlb = g_try_new(CPUTLBEntryFull, new_size);
208 
209     /*
210      * If the allocations fail, try smaller sizes. We just freed some
211      * memory, so going back to half of new_size has a good chance of working.
212      * Increased memory pressure elsewhere in the system might cause the
213      * allocations to fail though, so we progressively reduce the allocation
214      * size, aborting if we cannot even allocate the smallest TLB we support.
215      */
216     while (fast->table == NULL || desc->fulltlb == NULL) {
217         if (new_size == (1 << CPU_TLB_DYN_MIN_BITS)) {
218             error_report("%s: %s", __func__, strerror(errno));
219             abort();
220         }
221         new_size = MAX(new_size >> 1, 1 << CPU_TLB_DYN_MIN_BITS);
222         fast->mask = (new_size - 1) << CPU_TLB_ENTRY_BITS;
223 
224         g_free(fast->table);
225         g_free(desc->fulltlb);
226         fast->table = g_try_new(CPUTLBEntry, new_size);
227         desc->fulltlb = g_try_new(CPUTLBEntryFull, new_size);
228     }
229 }
230 
231 static void tlb_mmu_flush_locked(CPUTLBDesc *desc, CPUTLBDescFast *fast)
232 {
233     desc->n_used_entries = 0;
234     desc->large_page_addr = -1;
235     desc->large_page_mask = -1;
236     desc->vindex = 0;
237     memset(fast->table, -1, sizeof_tlb(fast));
238     memset(desc->vtable, -1, sizeof(desc->vtable));
239 }
240 
241 static void tlb_flush_one_mmuidx_locked(CPUArchState *env, int mmu_idx,
242                                         int64_t now)
243 {
244     CPUTLBDesc *desc = &env_tlb(env)->d[mmu_idx];
245     CPUTLBDescFast *fast = &env_tlb(env)->f[mmu_idx];
246 
247     tlb_mmu_resize_locked(desc, fast, now);
248     tlb_mmu_flush_locked(desc, fast);
249 }
250 
251 static void tlb_mmu_init(CPUTLBDesc *desc, CPUTLBDescFast *fast, int64_t now)
252 {
253     size_t n_entries = 1 << CPU_TLB_DYN_DEFAULT_BITS;
254 
255     tlb_window_reset(desc, now, 0);
256     desc->n_used_entries = 0;
257     fast->mask = (n_entries - 1) << CPU_TLB_ENTRY_BITS;
258     fast->table = g_new(CPUTLBEntry, n_entries);
259     desc->fulltlb = g_new(CPUTLBEntryFull, n_entries);
260     tlb_mmu_flush_locked(desc, fast);
261 }
262 
263 static inline void tlb_n_used_entries_inc(CPUArchState *env, uintptr_t mmu_idx)
264 {
265     env_tlb(env)->d[mmu_idx].n_used_entries++;
266 }
267 
268 static inline void tlb_n_used_entries_dec(CPUArchState *env, uintptr_t mmu_idx)
269 {
270     env_tlb(env)->d[mmu_idx].n_used_entries--;
271 }
272 
273 void tlb_init(CPUState *cpu)
274 {
275     CPUArchState *env = cpu->env_ptr;
276     int64_t now = get_clock_realtime();
277     int i;
278 
279     qemu_spin_init(&env_tlb(env)->c.lock);
280 
281     /* All tlbs are initialized flushed. */
282     env_tlb(env)->c.dirty = 0;
283 
284     for (i = 0; i < NB_MMU_MODES; i++) {
285         tlb_mmu_init(&env_tlb(env)->d[i], &env_tlb(env)->f[i], now);
286     }
287 }
288 
289 void tlb_destroy(CPUState *cpu)
290 {
291     CPUArchState *env = cpu->env_ptr;
292     int i;
293 
294     qemu_spin_destroy(&env_tlb(env)->c.lock);
295     for (i = 0; i < NB_MMU_MODES; i++) {
296         CPUTLBDesc *desc = &env_tlb(env)->d[i];
297         CPUTLBDescFast *fast = &env_tlb(env)->f[i];
298 
299         g_free(fast->table);
300         g_free(desc->fulltlb);
301     }
302 }
303 
304 /* flush_all_helper: run fn across all cpus
305  *
306  * If the wait flag is set then the src cpu's helper will be queued as
307  * "safe" work and the loop exited creating a synchronisation point
308  * where all queued work will be finished before execution starts
309  * again.
310  */
311 static void flush_all_helper(CPUState *src, run_on_cpu_func fn,
312                              run_on_cpu_data d)
313 {
314     CPUState *cpu;
315 
316     CPU_FOREACH(cpu) {
317         if (cpu != src) {
318             async_run_on_cpu(cpu, fn, d);
319         }
320     }
321 }
322 
323 void tlb_flush_counts(size_t *pfull, size_t *ppart, size_t *pelide)
324 {
325     CPUState *cpu;
326     size_t full = 0, part = 0, elide = 0;
327 
328     CPU_FOREACH(cpu) {
329         CPUArchState *env = cpu->env_ptr;
330 
331         full += qatomic_read(&env_tlb(env)->c.full_flush_count);
332         part += qatomic_read(&env_tlb(env)->c.part_flush_count);
333         elide += qatomic_read(&env_tlb(env)->c.elide_flush_count);
334     }
335     *pfull = full;
336     *ppart = part;
337     *pelide = elide;
338 }
339 
340 static void tlb_flush_by_mmuidx_async_work(CPUState *cpu, run_on_cpu_data data)
341 {
342     CPUArchState *env = cpu->env_ptr;
343     uint16_t asked = data.host_int;
344     uint16_t all_dirty, work, to_clean;
345     int64_t now = get_clock_realtime();
346 
347     assert_cpu_is_self(cpu);
348 
349     tlb_debug("mmu_idx:0x%04" PRIx16 "\n", asked);
350 
351     qemu_spin_lock(&env_tlb(env)->c.lock);
352 
353     all_dirty = env_tlb(env)->c.dirty;
354     to_clean = asked & all_dirty;
355     all_dirty &= ~to_clean;
356     env_tlb(env)->c.dirty = all_dirty;
357 
358     for (work = to_clean; work != 0; work &= work - 1) {
359         int mmu_idx = ctz32(work);
360         tlb_flush_one_mmuidx_locked(env, mmu_idx, now);
361     }
362 
363     qemu_spin_unlock(&env_tlb(env)->c.lock);
364 
365     tcg_flush_jmp_cache(cpu);
366 
367     if (to_clean == ALL_MMUIDX_BITS) {
368         qatomic_set(&env_tlb(env)->c.full_flush_count,
369                    env_tlb(env)->c.full_flush_count + 1);
370     } else {
371         qatomic_set(&env_tlb(env)->c.part_flush_count,
372                    env_tlb(env)->c.part_flush_count + ctpop16(to_clean));
373         if (to_clean != asked) {
374             qatomic_set(&env_tlb(env)->c.elide_flush_count,
375                        env_tlb(env)->c.elide_flush_count +
376                        ctpop16(asked & ~to_clean));
377         }
378     }
379 }
380 
381 void tlb_flush_by_mmuidx(CPUState *cpu, uint16_t idxmap)
382 {
383     tlb_debug("mmu_idx: 0x%" PRIx16 "\n", idxmap);
384 
385     if (cpu->created && !qemu_cpu_is_self(cpu)) {
386         async_run_on_cpu(cpu, tlb_flush_by_mmuidx_async_work,
387                          RUN_ON_CPU_HOST_INT(idxmap));
388     } else {
389         tlb_flush_by_mmuidx_async_work(cpu, RUN_ON_CPU_HOST_INT(idxmap));
390     }
391 }
392 
393 void tlb_flush(CPUState *cpu)
394 {
395     tlb_flush_by_mmuidx(cpu, ALL_MMUIDX_BITS);
396 }
397 
398 void tlb_flush_by_mmuidx_all_cpus(CPUState *src_cpu, uint16_t idxmap)
399 {
400     const run_on_cpu_func fn = tlb_flush_by_mmuidx_async_work;
401 
402     tlb_debug("mmu_idx: 0x%"PRIx16"\n", idxmap);
403 
404     flush_all_helper(src_cpu, fn, RUN_ON_CPU_HOST_INT(idxmap));
405     fn(src_cpu, RUN_ON_CPU_HOST_INT(idxmap));
406 }
407 
408 void tlb_flush_all_cpus(CPUState *src_cpu)
409 {
410     tlb_flush_by_mmuidx_all_cpus(src_cpu, ALL_MMUIDX_BITS);
411 }
412 
413 void tlb_flush_by_mmuidx_all_cpus_synced(CPUState *src_cpu, uint16_t idxmap)
414 {
415     const run_on_cpu_func fn = tlb_flush_by_mmuidx_async_work;
416 
417     tlb_debug("mmu_idx: 0x%"PRIx16"\n", idxmap);
418 
419     flush_all_helper(src_cpu, fn, RUN_ON_CPU_HOST_INT(idxmap));
420     async_safe_run_on_cpu(src_cpu, fn, RUN_ON_CPU_HOST_INT(idxmap));
421 }
422 
423 void tlb_flush_all_cpus_synced(CPUState *src_cpu)
424 {
425     tlb_flush_by_mmuidx_all_cpus_synced(src_cpu, ALL_MMUIDX_BITS);
426 }
427 
428 static bool tlb_hit_page_mask_anyprot(CPUTLBEntry *tlb_entry,
429                                       target_ulong page, target_ulong mask)
430 {
431     page &= mask;
432     mask &= TARGET_PAGE_MASK | TLB_INVALID_MASK;
433 
434     return (page == (tlb_entry->addr_read & mask) ||
435             page == (tlb_addr_write(tlb_entry) & mask) ||
436             page == (tlb_entry->addr_code & mask));
437 }
438 
439 static inline bool tlb_hit_page_anyprot(CPUTLBEntry *tlb_entry,
440                                         target_ulong page)
441 {
442     return tlb_hit_page_mask_anyprot(tlb_entry, page, -1);
443 }
444 
445 /**
446  * tlb_entry_is_empty - return true if the entry is not in use
447  * @te: pointer to CPUTLBEntry
448  */
449 static inline bool tlb_entry_is_empty(const CPUTLBEntry *te)
450 {
451     return te->addr_read == -1 && te->addr_write == -1 && te->addr_code == -1;
452 }
453 
454 /* Called with tlb_c.lock held */
455 static bool tlb_flush_entry_mask_locked(CPUTLBEntry *tlb_entry,
456                                         target_ulong page,
457                                         target_ulong mask)
458 {
459     if (tlb_hit_page_mask_anyprot(tlb_entry, page, mask)) {
460         memset(tlb_entry, -1, sizeof(*tlb_entry));
461         return true;
462     }
463     return false;
464 }
465 
466 static inline bool tlb_flush_entry_locked(CPUTLBEntry *tlb_entry,
467                                           target_ulong page)
468 {
469     return tlb_flush_entry_mask_locked(tlb_entry, page, -1);
470 }
471 
472 /* Called with tlb_c.lock held */
473 static void tlb_flush_vtlb_page_mask_locked(CPUArchState *env, int mmu_idx,
474                                             target_ulong page,
475                                             target_ulong mask)
476 {
477     CPUTLBDesc *d = &env_tlb(env)->d[mmu_idx];
478     int k;
479 
480     assert_cpu_is_self(env_cpu(env));
481     for (k = 0; k < CPU_VTLB_SIZE; k++) {
482         if (tlb_flush_entry_mask_locked(&d->vtable[k], page, mask)) {
483             tlb_n_used_entries_dec(env, mmu_idx);
484         }
485     }
486 }
487 
488 static inline void tlb_flush_vtlb_page_locked(CPUArchState *env, int mmu_idx,
489                                               target_ulong page)
490 {
491     tlb_flush_vtlb_page_mask_locked(env, mmu_idx, page, -1);
492 }
493 
494 static void tlb_flush_page_locked(CPUArchState *env, int midx,
495                                   target_ulong page)
496 {
497     target_ulong lp_addr = env_tlb(env)->d[midx].large_page_addr;
498     target_ulong lp_mask = env_tlb(env)->d[midx].large_page_mask;
499 
500     /* Check if we need to flush due to large pages.  */
501     if ((page & lp_mask) == lp_addr) {
502         tlb_debug("forcing full flush midx %d ("
503                   TARGET_FMT_lx "/" TARGET_FMT_lx ")\n",
504                   midx, lp_addr, lp_mask);
505         tlb_flush_one_mmuidx_locked(env, midx, get_clock_realtime());
506     } else {
507         if (tlb_flush_entry_locked(tlb_entry(env, midx, page), page)) {
508             tlb_n_used_entries_dec(env, midx);
509         }
510         tlb_flush_vtlb_page_locked(env, midx, page);
511     }
512 }
513 
514 /**
515  * tlb_flush_page_by_mmuidx_async_0:
516  * @cpu: cpu on which to flush
517  * @addr: page of virtual address to flush
518  * @idxmap: set of mmu_idx to flush
519  *
520  * Helper for tlb_flush_page_by_mmuidx and friends, flush one page
521  * at @addr from the tlbs indicated by @idxmap from @cpu.
522  */
523 static void tlb_flush_page_by_mmuidx_async_0(CPUState *cpu,
524                                              target_ulong addr,
525                                              uint16_t idxmap)
526 {
527     CPUArchState *env = cpu->env_ptr;
528     int mmu_idx;
529 
530     assert_cpu_is_self(cpu);
531 
532     tlb_debug("page addr:" TARGET_FMT_lx " mmu_map:0x%x\n", addr, idxmap);
533 
534     qemu_spin_lock(&env_tlb(env)->c.lock);
535     for (mmu_idx = 0; mmu_idx < NB_MMU_MODES; mmu_idx++) {
536         if ((idxmap >> mmu_idx) & 1) {
537             tlb_flush_page_locked(env, mmu_idx, addr);
538         }
539     }
540     qemu_spin_unlock(&env_tlb(env)->c.lock);
541 
542     /*
543      * Discard jump cache entries for any tb which might potentially
544      * overlap the flushed page, which includes the previous.
545      */
546     tb_jmp_cache_clear_page(cpu, addr - TARGET_PAGE_SIZE);
547     tb_jmp_cache_clear_page(cpu, addr);
548 }
549 
550 /**
551  * tlb_flush_page_by_mmuidx_async_1:
552  * @cpu: cpu on which to flush
553  * @data: encoded addr + idxmap
554  *
555  * Helper for tlb_flush_page_by_mmuidx and friends, called through
556  * async_run_on_cpu.  The idxmap parameter is encoded in the page
557  * offset of the target_ptr field.  This limits the set of mmu_idx
558  * that can be passed via this method.
559  */
560 static void tlb_flush_page_by_mmuidx_async_1(CPUState *cpu,
561                                              run_on_cpu_data data)
562 {
563     target_ulong addr_and_idxmap = (target_ulong) data.target_ptr;
564     target_ulong addr = addr_and_idxmap & TARGET_PAGE_MASK;
565     uint16_t idxmap = addr_and_idxmap & ~TARGET_PAGE_MASK;
566 
567     tlb_flush_page_by_mmuidx_async_0(cpu, addr, idxmap);
568 }
569 
570 typedef struct {
571     target_ulong addr;
572     uint16_t idxmap;
573 } TLBFlushPageByMMUIdxData;
574 
575 /**
576  * tlb_flush_page_by_mmuidx_async_2:
577  * @cpu: cpu on which to flush
578  * @data: allocated addr + idxmap
579  *
580  * Helper for tlb_flush_page_by_mmuidx and friends, called through
581  * async_run_on_cpu.  The addr+idxmap parameters are stored in a
582  * TLBFlushPageByMMUIdxData structure that has been allocated
583  * specifically for this helper.  Free the structure when done.
584  */
585 static void tlb_flush_page_by_mmuidx_async_2(CPUState *cpu,
586                                              run_on_cpu_data data)
587 {
588     TLBFlushPageByMMUIdxData *d = data.host_ptr;
589 
590     tlb_flush_page_by_mmuidx_async_0(cpu, d->addr, d->idxmap);
591     g_free(d);
592 }
593 
594 void tlb_flush_page_by_mmuidx(CPUState *cpu, target_ulong addr, uint16_t idxmap)
595 {
596     tlb_debug("addr: "TARGET_FMT_lx" mmu_idx:%" PRIx16 "\n", addr, idxmap);
597 
598     /* This should already be page aligned */
599     addr &= TARGET_PAGE_MASK;
600 
601     if (qemu_cpu_is_self(cpu)) {
602         tlb_flush_page_by_mmuidx_async_0(cpu, addr, idxmap);
603     } else if (idxmap < TARGET_PAGE_SIZE) {
604         /*
605          * Most targets have only a few mmu_idx.  In the case where
606          * we can stuff idxmap into the low TARGET_PAGE_BITS, avoid
607          * allocating memory for this operation.
608          */
609         async_run_on_cpu(cpu, tlb_flush_page_by_mmuidx_async_1,
610                          RUN_ON_CPU_TARGET_PTR(addr | idxmap));
611     } else {
612         TLBFlushPageByMMUIdxData *d = g_new(TLBFlushPageByMMUIdxData, 1);
613 
614         /* Otherwise allocate a structure, freed by the worker.  */
615         d->addr = addr;
616         d->idxmap = idxmap;
617         async_run_on_cpu(cpu, tlb_flush_page_by_mmuidx_async_2,
618                          RUN_ON_CPU_HOST_PTR(d));
619     }
620 }
621 
622 void tlb_flush_page(CPUState *cpu, target_ulong addr)
623 {
624     tlb_flush_page_by_mmuidx(cpu, addr, ALL_MMUIDX_BITS);
625 }
626 
627 void tlb_flush_page_by_mmuidx_all_cpus(CPUState *src_cpu, target_ulong addr,
628                                        uint16_t idxmap)
629 {
630     tlb_debug("addr: "TARGET_FMT_lx" mmu_idx:%"PRIx16"\n", addr, idxmap);
631 
632     /* This should already be page aligned */
633     addr &= TARGET_PAGE_MASK;
634 
635     /*
636      * Allocate memory to hold addr+idxmap only when needed.
637      * See tlb_flush_page_by_mmuidx for details.
638      */
639     if (idxmap < TARGET_PAGE_SIZE) {
640         flush_all_helper(src_cpu, tlb_flush_page_by_mmuidx_async_1,
641                          RUN_ON_CPU_TARGET_PTR(addr | idxmap));
642     } else {
643         CPUState *dst_cpu;
644 
645         /* Allocate a separate data block for each destination cpu.  */
646         CPU_FOREACH(dst_cpu) {
647             if (dst_cpu != src_cpu) {
648                 TLBFlushPageByMMUIdxData *d
649                     = g_new(TLBFlushPageByMMUIdxData, 1);
650 
651                 d->addr = addr;
652                 d->idxmap = idxmap;
653                 async_run_on_cpu(dst_cpu, tlb_flush_page_by_mmuidx_async_2,
654                                  RUN_ON_CPU_HOST_PTR(d));
655             }
656         }
657     }
658 
659     tlb_flush_page_by_mmuidx_async_0(src_cpu, addr, idxmap);
660 }
661 
662 void tlb_flush_page_all_cpus(CPUState *src, target_ulong addr)
663 {
664     tlb_flush_page_by_mmuidx_all_cpus(src, addr, ALL_MMUIDX_BITS);
665 }
666 
667 void tlb_flush_page_by_mmuidx_all_cpus_synced(CPUState *src_cpu,
668                                               target_ulong addr,
669                                               uint16_t idxmap)
670 {
671     tlb_debug("addr: "TARGET_FMT_lx" mmu_idx:%"PRIx16"\n", addr, idxmap);
672 
673     /* This should already be page aligned */
674     addr &= TARGET_PAGE_MASK;
675 
676     /*
677      * Allocate memory to hold addr+idxmap only when needed.
678      * See tlb_flush_page_by_mmuidx for details.
679      */
680     if (idxmap < TARGET_PAGE_SIZE) {
681         flush_all_helper(src_cpu, tlb_flush_page_by_mmuidx_async_1,
682                          RUN_ON_CPU_TARGET_PTR(addr | idxmap));
683         async_safe_run_on_cpu(src_cpu, tlb_flush_page_by_mmuidx_async_1,
684                               RUN_ON_CPU_TARGET_PTR(addr | idxmap));
685     } else {
686         CPUState *dst_cpu;
687         TLBFlushPageByMMUIdxData *d;
688 
689         /* Allocate a separate data block for each destination cpu.  */
690         CPU_FOREACH(dst_cpu) {
691             if (dst_cpu != src_cpu) {
692                 d = g_new(TLBFlushPageByMMUIdxData, 1);
693                 d->addr = addr;
694                 d->idxmap = idxmap;
695                 async_run_on_cpu(dst_cpu, tlb_flush_page_by_mmuidx_async_2,
696                                  RUN_ON_CPU_HOST_PTR(d));
697             }
698         }
699 
700         d = g_new(TLBFlushPageByMMUIdxData, 1);
701         d->addr = addr;
702         d->idxmap = idxmap;
703         async_safe_run_on_cpu(src_cpu, tlb_flush_page_by_mmuidx_async_2,
704                               RUN_ON_CPU_HOST_PTR(d));
705     }
706 }
707 
708 void tlb_flush_page_all_cpus_synced(CPUState *src, target_ulong addr)
709 {
710     tlb_flush_page_by_mmuidx_all_cpus_synced(src, addr, ALL_MMUIDX_BITS);
711 }
712 
713 static void tlb_flush_range_locked(CPUArchState *env, int midx,
714                                    target_ulong addr, target_ulong len,
715                                    unsigned bits)
716 {
717     CPUTLBDesc *d = &env_tlb(env)->d[midx];
718     CPUTLBDescFast *f = &env_tlb(env)->f[midx];
719     target_ulong mask = MAKE_64BIT_MASK(0, bits);
720 
721     /*
722      * If @bits is smaller than the tlb size, there may be multiple entries
723      * within the TLB; otherwise all addresses that match under @mask hit
724      * the same TLB entry.
725      * TODO: Perhaps allow bits to be a few bits less than the size.
726      * For now, just flush the entire TLB.
727      *
728      * If @len is larger than the tlb size, then it will take longer to
729      * test all of the entries in the TLB than it will to flush it all.
730      */
731     if (mask < f->mask || len > f->mask) {
732         tlb_debug("forcing full flush midx %d ("
733                   TARGET_FMT_lx "/" TARGET_FMT_lx "+" TARGET_FMT_lx ")\n",
734                   midx, addr, mask, len);
735         tlb_flush_one_mmuidx_locked(env, midx, get_clock_realtime());
736         return;
737     }
738 
739     /*
740      * Check if we need to flush due to large pages.
741      * Because large_page_mask contains all 1's from the msb,
742      * we only need to test the end of the range.
743      */
744     if (((addr + len - 1) & d->large_page_mask) == d->large_page_addr) {
745         tlb_debug("forcing full flush midx %d ("
746                   TARGET_FMT_lx "/" TARGET_FMT_lx ")\n",
747                   midx, d->large_page_addr, d->large_page_mask);
748         tlb_flush_one_mmuidx_locked(env, midx, get_clock_realtime());
749         return;
750     }
751 
752     for (target_ulong i = 0; i < len; i += TARGET_PAGE_SIZE) {
753         target_ulong page = addr + i;
754         CPUTLBEntry *entry = tlb_entry(env, midx, page);
755 
756         if (tlb_flush_entry_mask_locked(entry, page, mask)) {
757             tlb_n_used_entries_dec(env, midx);
758         }
759         tlb_flush_vtlb_page_mask_locked(env, midx, page, mask);
760     }
761 }
762 
763 typedef struct {
764     target_ulong addr;
765     target_ulong len;
766     uint16_t idxmap;
767     uint16_t bits;
768 } TLBFlushRangeData;
769 
770 static void tlb_flush_range_by_mmuidx_async_0(CPUState *cpu,
771                                               TLBFlushRangeData d)
772 {
773     CPUArchState *env = cpu->env_ptr;
774     int mmu_idx;
775 
776     assert_cpu_is_self(cpu);
777 
778     tlb_debug("range:" TARGET_FMT_lx "/%u+" TARGET_FMT_lx " mmu_map:0x%x\n",
779               d.addr, d.bits, d.len, d.idxmap);
780 
781     qemu_spin_lock(&env_tlb(env)->c.lock);
782     for (mmu_idx = 0; mmu_idx < NB_MMU_MODES; mmu_idx++) {
783         if ((d.idxmap >> mmu_idx) & 1) {
784             tlb_flush_range_locked(env, mmu_idx, d.addr, d.len, d.bits);
785         }
786     }
787     qemu_spin_unlock(&env_tlb(env)->c.lock);
788 
789     /*
790      * If the length is larger than the jump cache size, then it will take
791      * longer to clear each entry individually than it will to clear it all.
792      */
793     if (d.len >= (TARGET_PAGE_SIZE * TB_JMP_CACHE_SIZE)) {
794         tcg_flush_jmp_cache(cpu);
795         return;
796     }
797 
798     /*
799      * Discard jump cache entries for any tb which might potentially
800      * overlap the flushed pages, which includes the previous.
801      */
802     d.addr -= TARGET_PAGE_SIZE;
803     for (target_ulong i = 0, n = d.len / TARGET_PAGE_SIZE + 1; i < n; i++) {
804         tb_jmp_cache_clear_page(cpu, d.addr);
805         d.addr += TARGET_PAGE_SIZE;
806     }
807 }
808 
809 static void tlb_flush_range_by_mmuidx_async_1(CPUState *cpu,
810                                               run_on_cpu_data data)
811 {
812     TLBFlushRangeData *d = data.host_ptr;
813     tlb_flush_range_by_mmuidx_async_0(cpu, *d);
814     g_free(d);
815 }
816 
817 void tlb_flush_range_by_mmuidx(CPUState *cpu, target_ulong addr,
818                                target_ulong len, uint16_t idxmap,
819                                unsigned bits)
820 {
821     TLBFlushRangeData d;
822 
823     /*
824      * If all bits are significant, and len is small,
825      * this devolves to tlb_flush_page.
826      */
827     if (bits >= TARGET_LONG_BITS && len <= TARGET_PAGE_SIZE) {
828         tlb_flush_page_by_mmuidx(cpu, addr, idxmap);
829         return;
830     }
831     /* If no page bits are significant, this devolves to tlb_flush. */
832     if (bits < TARGET_PAGE_BITS) {
833         tlb_flush_by_mmuidx(cpu, idxmap);
834         return;
835     }
836 
837     /* This should already be page aligned */
838     d.addr = addr & TARGET_PAGE_MASK;
839     d.len = len;
840     d.idxmap = idxmap;
841     d.bits = bits;
842 
843     if (qemu_cpu_is_self(cpu)) {
844         tlb_flush_range_by_mmuidx_async_0(cpu, d);
845     } else {
846         /* Otherwise allocate a structure, freed by the worker.  */
847         TLBFlushRangeData *p = g_memdup(&d, sizeof(d));
848         async_run_on_cpu(cpu, tlb_flush_range_by_mmuidx_async_1,
849                          RUN_ON_CPU_HOST_PTR(p));
850     }
851 }
852 
853 void tlb_flush_page_bits_by_mmuidx(CPUState *cpu, target_ulong addr,
854                                    uint16_t idxmap, unsigned bits)
855 {
856     tlb_flush_range_by_mmuidx(cpu, addr, TARGET_PAGE_SIZE, idxmap, bits);
857 }
858 
859 void tlb_flush_range_by_mmuidx_all_cpus(CPUState *src_cpu,
860                                         target_ulong addr, target_ulong len,
861                                         uint16_t idxmap, unsigned bits)
862 {
863     TLBFlushRangeData d;
864     CPUState *dst_cpu;
865 
866     /*
867      * If all bits are significant, and len is small,
868      * this devolves to tlb_flush_page.
869      */
870     if (bits >= TARGET_LONG_BITS && len <= TARGET_PAGE_SIZE) {
871         tlb_flush_page_by_mmuidx_all_cpus(src_cpu, addr, idxmap);
872         return;
873     }
874     /* If no page bits are significant, this devolves to tlb_flush. */
875     if (bits < TARGET_PAGE_BITS) {
876         tlb_flush_by_mmuidx_all_cpus(src_cpu, idxmap);
877         return;
878     }
879 
880     /* This should already be page aligned */
881     d.addr = addr & TARGET_PAGE_MASK;
882     d.len = len;
883     d.idxmap = idxmap;
884     d.bits = bits;
885 
886     /* Allocate a separate data block for each destination cpu.  */
887     CPU_FOREACH(dst_cpu) {
888         if (dst_cpu != src_cpu) {
889             TLBFlushRangeData *p = g_memdup(&d, sizeof(d));
890             async_run_on_cpu(dst_cpu,
891                              tlb_flush_range_by_mmuidx_async_1,
892                              RUN_ON_CPU_HOST_PTR(p));
893         }
894     }
895 
896     tlb_flush_range_by_mmuidx_async_0(src_cpu, d);
897 }
898 
899 void tlb_flush_page_bits_by_mmuidx_all_cpus(CPUState *src_cpu,
900                                             target_ulong addr,
901                                             uint16_t idxmap, unsigned bits)
902 {
903     tlb_flush_range_by_mmuidx_all_cpus(src_cpu, addr, TARGET_PAGE_SIZE,
904                                        idxmap, bits);
905 }
906 
907 void tlb_flush_range_by_mmuidx_all_cpus_synced(CPUState *src_cpu,
908                                                target_ulong addr,
909                                                target_ulong len,
910                                                uint16_t idxmap,
911                                                unsigned bits)
912 {
913     TLBFlushRangeData d, *p;
914     CPUState *dst_cpu;
915 
916     /*
917      * If all bits are significant, and len is small,
918      * this devolves to tlb_flush_page.
919      */
920     if (bits >= TARGET_LONG_BITS && len <= TARGET_PAGE_SIZE) {
921         tlb_flush_page_by_mmuidx_all_cpus_synced(src_cpu, addr, idxmap);
922         return;
923     }
924     /* If no page bits are significant, this devolves to tlb_flush. */
925     if (bits < TARGET_PAGE_BITS) {
926         tlb_flush_by_mmuidx_all_cpus_synced(src_cpu, idxmap);
927         return;
928     }
929 
930     /* This should already be page aligned */
931     d.addr = addr & TARGET_PAGE_MASK;
932     d.len = len;
933     d.idxmap = idxmap;
934     d.bits = bits;
935 
936     /* Allocate a separate data block for each destination cpu.  */
937     CPU_FOREACH(dst_cpu) {
938         if (dst_cpu != src_cpu) {
939             p = g_memdup(&d, sizeof(d));
940             async_run_on_cpu(dst_cpu, tlb_flush_range_by_mmuidx_async_1,
941                              RUN_ON_CPU_HOST_PTR(p));
942         }
943     }
944 
945     p = g_memdup(&d, sizeof(d));
946     async_safe_run_on_cpu(src_cpu, tlb_flush_range_by_mmuidx_async_1,
947                           RUN_ON_CPU_HOST_PTR(p));
948 }
949 
950 void tlb_flush_page_bits_by_mmuidx_all_cpus_synced(CPUState *src_cpu,
951                                                    target_ulong addr,
952                                                    uint16_t idxmap,
953                                                    unsigned bits)
954 {
955     tlb_flush_range_by_mmuidx_all_cpus_synced(src_cpu, addr, TARGET_PAGE_SIZE,
956                                               idxmap, bits);
957 }
958 
959 /* update the TLBs so that writes to code in the virtual page 'addr'
960    can be detected */
961 void tlb_protect_code(ram_addr_t ram_addr)
962 {
963     cpu_physical_memory_test_and_clear_dirty(ram_addr & TARGET_PAGE_MASK,
964                                              TARGET_PAGE_SIZE,
965                                              DIRTY_MEMORY_CODE);
966 }
967 
968 /* update the TLB so that writes in physical page 'phys_addr' are no longer
969    tested for self modifying code */
970 void tlb_unprotect_code(ram_addr_t ram_addr)
971 {
972     cpu_physical_memory_set_dirty_flag(ram_addr, DIRTY_MEMORY_CODE);
973 }
974 
975 
976 /*
977  * Dirty write flag handling
978  *
979  * When the TCG code writes to a location it looks up the address in
980  * the TLB and uses that data to compute the final address. If any of
981  * the lower bits of the address are set then the slow path is forced.
982  * There are a number of reasons to do this but for normal RAM the
983  * most usual is detecting writes to code regions which may invalidate
984  * generated code.
985  *
986  * Other vCPUs might be reading their TLBs during guest execution, so we update
987  * te->addr_write with qatomic_set. We don't need to worry about this for
988  * oversized guests as MTTCG is disabled for them.
989  *
990  * Called with tlb_c.lock held.
991  */
992 static void tlb_reset_dirty_range_locked(CPUTLBEntry *tlb_entry,
993                                          uintptr_t start, uintptr_t length)
994 {
995     uintptr_t addr = tlb_entry->addr_write;
996 
997     if ((addr & (TLB_INVALID_MASK | TLB_MMIO |
998                  TLB_DISCARD_WRITE | TLB_NOTDIRTY)) == 0) {
999         addr &= TARGET_PAGE_MASK;
1000         addr += tlb_entry->addend;
1001         if ((addr - start) < length) {
1002 #if TCG_OVERSIZED_GUEST
1003             tlb_entry->addr_write |= TLB_NOTDIRTY;
1004 #else
1005             qatomic_set(&tlb_entry->addr_write,
1006                        tlb_entry->addr_write | TLB_NOTDIRTY);
1007 #endif
1008         }
1009     }
1010 }
1011 
1012 /*
1013  * Called with tlb_c.lock held.
1014  * Called only from the vCPU context, i.e. the TLB's owner thread.
1015  */
1016 static inline void copy_tlb_helper_locked(CPUTLBEntry *d, const CPUTLBEntry *s)
1017 {
1018     *d = *s;
1019 }
1020 
1021 /* This is a cross vCPU call (i.e. another vCPU resetting the flags of
1022  * the target vCPU).
1023  * We must take tlb_c.lock to avoid racing with another vCPU update. The only
1024  * thing actually updated is the target TLB entry ->addr_write flags.
1025  */
1026 void tlb_reset_dirty(CPUState *cpu, ram_addr_t start1, ram_addr_t length)
1027 {
1028     CPUArchState *env;
1029 
1030     int mmu_idx;
1031 
1032     env = cpu->env_ptr;
1033     qemu_spin_lock(&env_tlb(env)->c.lock);
1034     for (mmu_idx = 0; mmu_idx < NB_MMU_MODES; mmu_idx++) {
1035         unsigned int i;
1036         unsigned int n = tlb_n_entries(&env_tlb(env)->f[mmu_idx]);
1037 
1038         for (i = 0; i < n; i++) {
1039             tlb_reset_dirty_range_locked(&env_tlb(env)->f[mmu_idx].table[i],
1040                                          start1, length);
1041         }
1042 
1043         for (i = 0; i < CPU_VTLB_SIZE; i++) {
1044             tlb_reset_dirty_range_locked(&env_tlb(env)->d[mmu_idx].vtable[i],
1045                                          start1, length);
1046         }
1047     }
1048     qemu_spin_unlock(&env_tlb(env)->c.lock);
1049 }
1050 
1051 /* Called with tlb_c.lock held */
1052 static inline void tlb_set_dirty1_locked(CPUTLBEntry *tlb_entry,
1053                                          target_ulong vaddr)
1054 {
1055     if (tlb_entry->addr_write == (vaddr | TLB_NOTDIRTY)) {
1056         tlb_entry->addr_write = vaddr;
1057     }
1058 }
1059 
1060 /* update the TLB corresponding to virtual page vaddr
1061    so that it is no longer dirty */
1062 void tlb_set_dirty(CPUState *cpu, target_ulong vaddr)
1063 {
1064     CPUArchState *env = cpu->env_ptr;
1065     int mmu_idx;
1066 
1067     assert_cpu_is_self(cpu);
1068 
1069     vaddr &= TARGET_PAGE_MASK;
1070     qemu_spin_lock(&env_tlb(env)->c.lock);
1071     for (mmu_idx = 0; mmu_idx < NB_MMU_MODES; mmu_idx++) {
1072         tlb_set_dirty1_locked(tlb_entry(env, mmu_idx, vaddr), vaddr);
1073     }
1074 
1075     for (mmu_idx = 0; mmu_idx < NB_MMU_MODES; mmu_idx++) {
1076         int k;
1077         for (k = 0; k < CPU_VTLB_SIZE; k++) {
1078             tlb_set_dirty1_locked(&env_tlb(env)->d[mmu_idx].vtable[k], vaddr);
1079         }
1080     }
1081     qemu_spin_unlock(&env_tlb(env)->c.lock);
1082 }
1083 
1084 /* Our TLB does not support large pages, so remember the area covered by
1085    large pages and trigger a full TLB flush if these are invalidated.  */
1086 static void tlb_add_large_page(CPUArchState *env, int mmu_idx,
1087                                target_ulong vaddr, target_ulong size)
1088 {
1089     target_ulong lp_addr = env_tlb(env)->d[mmu_idx].large_page_addr;
1090     target_ulong lp_mask = ~(size - 1);
1091 
1092     if (lp_addr == (target_ulong)-1) {
1093         /* No previous large page.  */
1094         lp_addr = vaddr;
1095     } else {
1096         /* Extend the existing region to include the new page.
1097            This is a compromise between unnecessary flushes and
1098            the cost of maintaining a full variable size TLB.  */
1099         lp_mask &= env_tlb(env)->d[mmu_idx].large_page_mask;
1100         while (((lp_addr ^ vaddr) & lp_mask) != 0) {
1101             lp_mask <<= 1;
1102         }
1103     }
1104     env_tlb(env)->d[mmu_idx].large_page_addr = lp_addr & lp_mask;
1105     env_tlb(env)->d[mmu_idx].large_page_mask = lp_mask;
1106 }
1107 
1108 /*
1109  * Add a new TLB entry. At most one entry for a given virtual address
1110  * is permitted. Only a single TARGET_PAGE_SIZE region is mapped, the
1111  * supplied size is only used by tlb_flush_page.
1112  *
1113  * Called from TCG-generated code, which is under an RCU read-side
1114  * critical section.
1115  */
1116 void tlb_set_page_full(CPUState *cpu, int mmu_idx,
1117                        target_ulong vaddr, CPUTLBEntryFull *full)
1118 {
1119     CPUArchState *env = cpu->env_ptr;
1120     CPUTLB *tlb = env_tlb(env);
1121     CPUTLBDesc *desc = &tlb->d[mmu_idx];
1122     MemoryRegionSection *section;
1123     unsigned int index;
1124     target_ulong address;
1125     target_ulong write_address;
1126     uintptr_t addend;
1127     CPUTLBEntry *te, tn;
1128     hwaddr iotlb, xlat, sz, paddr_page;
1129     target_ulong vaddr_page;
1130     int asidx, wp_flags, prot;
1131     bool is_ram, is_romd;
1132 
1133     assert_cpu_is_self(cpu);
1134 
1135     if (full->lg_page_size <= TARGET_PAGE_BITS) {
1136         sz = TARGET_PAGE_SIZE;
1137     } else {
1138         sz = (hwaddr)1 << full->lg_page_size;
1139         tlb_add_large_page(env, mmu_idx, vaddr, sz);
1140     }
1141     vaddr_page = vaddr & TARGET_PAGE_MASK;
1142     paddr_page = full->phys_addr & TARGET_PAGE_MASK;
1143 
1144     prot = full->prot;
1145     asidx = cpu_asidx_from_attrs(cpu, full->attrs);
1146     section = address_space_translate_for_iotlb(cpu, asidx, paddr_page,
1147                                                 &xlat, &sz, full->attrs, &prot);
1148     assert(sz >= TARGET_PAGE_SIZE);
1149 
1150     tlb_debug("vaddr=" TARGET_FMT_lx " paddr=0x" HWADDR_FMT_plx
1151               " prot=%x idx=%d\n",
1152               vaddr, full->phys_addr, prot, mmu_idx);
1153 
1154     address = vaddr_page;
1155     if (full->lg_page_size < TARGET_PAGE_BITS) {
1156         /* Repeat the MMU check and TLB fill on every access.  */
1157         address |= TLB_INVALID_MASK;
1158     }
1159     if (full->attrs.byte_swap) {
1160         address |= TLB_BSWAP;
1161     }
1162 
1163     is_ram = memory_region_is_ram(section->mr);
1164     is_romd = memory_region_is_romd(section->mr);
1165 
1166     if (is_ram || is_romd) {
1167         /* RAM and ROMD both have associated host memory. */
1168         addend = (uintptr_t)memory_region_get_ram_ptr(section->mr) + xlat;
1169     } else {
1170         /* I/O does not; force the host address to NULL. */
1171         addend = 0;
1172     }
1173 
1174     write_address = address;
1175     if (is_ram) {
1176         iotlb = memory_region_get_ram_addr(section->mr) + xlat;
1177         /*
1178          * Computing is_clean is expensive; avoid all that unless
1179          * the page is actually writable.
1180          */
1181         if (prot & PAGE_WRITE) {
1182             if (section->readonly) {
1183                 write_address |= TLB_DISCARD_WRITE;
1184             } else if (cpu_physical_memory_is_clean(iotlb)) {
1185                 write_address |= TLB_NOTDIRTY;
1186             }
1187         }
1188     } else {
1189         /* I/O or ROMD */
1190         iotlb = memory_region_section_get_iotlb(cpu, section) + xlat;
1191         /*
1192          * Writes to romd devices must go through MMIO to enable write.
1193          * Reads to romd devices go through the ram_ptr found above,
1194          * but of course reads to I/O must go through MMIO.
1195          */
1196         write_address |= TLB_MMIO;
1197         if (!is_romd) {
1198             address = write_address;
1199         }
1200     }
1201 
1202     wp_flags = cpu_watchpoint_address_matches(cpu, vaddr_page,
1203                                               TARGET_PAGE_SIZE);
1204 
1205     index = tlb_index(env, mmu_idx, vaddr_page);
1206     te = tlb_entry(env, mmu_idx, vaddr_page);
1207 
1208     /*
1209      * Hold the TLB lock for the rest of the function. We could acquire/release
1210      * the lock several times in the function, but it is faster to amortize the
1211      * acquisition cost by acquiring it just once. Note that this leads to
1212      * a longer critical section, but this is not a concern since the TLB lock
1213      * is unlikely to be contended.
1214      */
1215     qemu_spin_lock(&tlb->c.lock);
1216 
1217     /* Note that the tlb is no longer clean.  */
1218     tlb->c.dirty |= 1 << mmu_idx;
1219 
1220     /* Make sure there's no cached translation for the new page.  */
1221     tlb_flush_vtlb_page_locked(env, mmu_idx, vaddr_page);
1222 
1223     /*
1224      * Only evict the old entry to the victim tlb if it's for a
1225      * different page; otherwise just overwrite the stale data.
1226      */
1227     if (!tlb_hit_page_anyprot(te, vaddr_page) && !tlb_entry_is_empty(te)) {
1228         unsigned vidx = desc->vindex++ % CPU_VTLB_SIZE;
1229         CPUTLBEntry *tv = &desc->vtable[vidx];
1230 
1231         /* Evict the old entry into the victim tlb.  */
1232         copy_tlb_helper_locked(tv, te);
1233         desc->vfulltlb[vidx] = desc->fulltlb[index];
1234         tlb_n_used_entries_dec(env, mmu_idx);
1235     }
1236 
1237     /* refill the tlb */
1238     /*
1239      * At this point iotlb contains a physical section number in the lower
1240      * TARGET_PAGE_BITS, and either
1241      *  + the ram_addr_t of the page base of the target RAM (RAM)
1242      *  + the offset within section->mr of the page base (I/O, ROMD)
1243      * We subtract the vaddr_page (which is page aligned and thus won't
1244      * disturb the low bits) to give an offset which can be added to the
1245      * (non-page-aligned) vaddr of the eventual memory access to get
1246      * the MemoryRegion offset for the access. Note that the vaddr we
1247      * subtract here is that of the page base, and not the same as the
1248      * vaddr we add back in io_readx()/io_writex()/get_page_addr_code().
1249      */
1250     desc->fulltlb[index] = *full;
1251     desc->fulltlb[index].xlat_section = iotlb - vaddr_page;
1252     desc->fulltlb[index].phys_addr = paddr_page;
1253     desc->fulltlb[index].prot = prot;
1254 
1255     /* Now calculate the new entry */
1256     tn.addend = addend - vaddr_page;
1257     if (prot & PAGE_READ) {
1258         tn.addr_read = address;
1259         if (wp_flags & BP_MEM_READ) {
1260             tn.addr_read |= TLB_WATCHPOINT;
1261         }
1262     } else {
1263         tn.addr_read = -1;
1264     }
1265 
1266     if (prot & PAGE_EXEC) {
1267         tn.addr_code = address;
1268     } else {
1269         tn.addr_code = -1;
1270     }
1271 
1272     tn.addr_write = -1;
1273     if (prot & PAGE_WRITE) {
1274         tn.addr_write = write_address;
1275         if (prot & PAGE_WRITE_INV) {
1276             tn.addr_write |= TLB_INVALID_MASK;
1277         }
1278         if (wp_flags & BP_MEM_WRITE) {
1279             tn.addr_write |= TLB_WATCHPOINT;
1280         }
1281     }
1282 
1283     copy_tlb_helper_locked(te, &tn);
1284     tlb_n_used_entries_inc(env, mmu_idx);
1285     qemu_spin_unlock(&tlb->c.lock);
1286 }
1287 
1288 void tlb_set_page_with_attrs(CPUState *cpu, target_ulong vaddr,
1289                              hwaddr paddr, MemTxAttrs attrs, int prot,
1290                              int mmu_idx, target_ulong size)
1291 {
1292     CPUTLBEntryFull full = {
1293         .phys_addr = paddr,
1294         .attrs = attrs,
1295         .prot = prot,
1296         .lg_page_size = ctz64(size)
1297     };
1298 
1299     assert(is_power_of_2(size));
1300     tlb_set_page_full(cpu, mmu_idx, vaddr, &full);
1301 }
1302 
1303 void tlb_set_page(CPUState *cpu, target_ulong vaddr,
1304                   hwaddr paddr, int prot,
1305                   int mmu_idx, target_ulong size)
1306 {
1307     tlb_set_page_with_attrs(cpu, vaddr, paddr, MEMTXATTRS_UNSPECIFIED,
1308                             prot, mmu_idx, size);
1309 }
1310 
1311 /*
1312  * Note: tlb_fill() can trigger a resize of the TLB. This means that all of the
1313  * caller's prior references to the TLB table (e.g. CPUTLBEntry pointers) must
1314  * be discarded and looked up again (e.g. via tlb_entry()).
1315  */
1316 static void tlb_fill(CPUState *cpu, target_ulong addr, int size,
1317                      MMUAccessType access_type, int mmu_idx, uintptr_t retaddr)
1318 {
1319     bool ok;
1320 
1321     /*
1322      * This is not a probe, so only valid return is success; failure
1323      * should result in exception + longjmp to the cpu loop.
1324      */
1325     ok = cpu->cc->tcg_ops->tlb_fill(cpu, addr, size,
1326                                     access_type, mmu_idx, false, retaddr);
1327     assert(ok);
1328 }
1329 
1330 static inline void cpu_unaligned_access(CPUState *cpu, vaddr addr,
1331                                         MMUAccessType access_type,
1332                                         int mmu_idx, uintptr_t retaddr)
1333 {
1334     cpu->cc->tcg_ops->do_unaligned_access(cpu, addr, access_type,
1335                                           mmu_idx, retaddr);
1336 }
1337 
1338 static inline void cpu_transaction_failed(CPUState *cpu, hwaddr physaddr,
1339                                           vaddr addr, unsigned size,
1340                                           MMUAccessType access_type,
1341                                           int mmu_idx, MemTxAttrs attrs,
1342                                           MemTxResult response,
1343                                           uintptr_t retaddr)
1344 {
1345     CPUClass *cc = CPU_GET_CLASS(cpu);
1346 
1347     if (!cpu->ignore_memory_transaction_failures &&
1348         cc->tcg_ops->do_transaction_failed) {
1349         cc->tcg_ops->do_transaction_failed(cpu, physaddr, addr, size,
1350                                            access_type, mmu_idx, attrs,
1351                                            response, retaddr);
1352     }
1353 }
1354 
1355 static uint64_t io_readx(CPUArchState *env, CPUTLBEntryFull *full,
1356                          int mmu_idx, target_ulong addr, uintptr_t retaddr,
1357                          MMUAccessType access_type, MemOp op)
1358 {
1359     CPUState *cpu = env_cpu(env);
1360     hwaddr mr_offset;
1361     MemoryRegionSection *section;
1362     MemoryRegion *mr;
1363     uint64_t val;
1364     MemTxResult r;
1365 
1366     section = iotlb_to_section(cpu, full->xlat_section, full->attrs);
1367     mr = section->mr;
1368     mr_offset = (full->xlat_section & TARGET_PAGE_MASK) + addr;
1369     cpu->mem_io_pc = retaddr;
1370     if (!cpu->can_do_io) {
1371         cpu_io_recompile(cpu, retaddr);
1372     }
1373 
1374     {
1375         QEMU_IOTHREAD_LOCK_GUARD();
1376         r = memory_region_dispatch_read(mr, mr_offset, &val, op, full->attrs);
1377     }
1378 
1379     if (r != MEMTX_OK) {
1380         hwaddr physaddr = mr_offset +
1381             section->offset_within_address_space -
1382             section->offset_within_region;
1383 
1384         cpu_transaction_failed(cpu, physaddr, addr, memop_size(op), access_type,
1385                                mmu_idx, full->attrs, r, retaddr);
1386     }
1387     return val;
1388 }
1389 
1390 /*
1391  * Save a potentially trashed CPUTLBEntryFull for later lookup by plugin.
1392  * This is read by tlb_plugin_lookup if the fulltlb entry doesn't match
1393  * because of the side effect of io_writex changing memory layout.
1394  */
1395 static void save_iotlb_data(CPUState *cs, MemoryRegionSection *section,
1396                             hwaddr mr_offset)
1397 {
1398 #ifdef CONFIG_PLUGIN
1399     SavedIOTLB *saved = &cs->saved_iotlb;
1400     saved->section = section;
1401     saved->mr_offset = mr_offset;
1402 #endif
1403 }
1404 
1405 static void io_writex(CPUArchState *env, CPUTLBEntryFull *full,
1406                       int mmu_idx, uint64_t val, target_ulong addr,
1407                       uintptr_t retaddr, MemOp op)
1408 {
1409     CPUState *cpu = env_cpu(env);
1410     hwaddr mr_offset;
1411     MemoryRegionSection *section;
1412     MemoryRegion *mr;
1413     MemTxResult r;
1414 
1415     section = iotlb_to_section(cpu, full->xlat_section, full->attrs);
1416     mr = section->mr;
1417     mr_offset = (full->xlat_section & TARGET_PAGE_MASK) + addr;
1418     if (!cpu->can_do_io) {
1419         cpu_io_recompile(cpu, retaddr);
1420     }
1421     cpu->mem_io_pc = retaddr;
1422 
1423     /*
1424      * The memory_region_dispatch may trigger a flush/resize
1425      * so for plugins we save the iotlb_data just in case.
1426      */
1427     save_iotlb_data(cpu, section, mr_offset);
1428 
1429     {
1430         QEMU_IOTHREAD_LOCK_GUARD();
1431         r = memory_region_dispatch_write(mr, mr_offset, val, op, full->attrs);
1432     }
1433 
1434     if (r != MEMTX_OK) {
1435         hwaddr physaddr = mr_offset +
1436             section->offset_within_address_space -
1437             section->offset_within_region;
1438 
1439         cpu_transaction_failed(cpu, physaddr, addr, memop_size(op),
1440                                MMU_DATA_STORE, mmu_idx, full->attrs, r,
1441                                retaddr);
1442     }
1443 }
1444 
1445 static inline target_ulong tlb_read_ofs(CPUTLBEntry *entry, size_t ofs)
1446 {
1447 #if TCG_OVERSIZED_GUEST
1448     return *(target_ulong *)((uintptr_t)entry + ofs);
1449 #else
1450     /* ofs might correspond to .addr_write, so use qatomic_read */
1451     return qatomic_read((target_ulong *)((uintptr_t)entry + ofs));
1452 #endif
1453 }
1454 
1455 /* Return true if ADDR is present in the victim tlb, and has been copied
1456    back to the main tlb.  */
1457 static bool victim_tlb_hit(CPUArchState *env, size_t mmu_idx, size_t index,
1458                            size_t elt_ofs, target_ulong page)
1459 {
1460     size_t vidx;
1461 
1462     assert_cpu_is_self(env_cpu(env));
1463     for (vidx = 0; vidx < CPU_VTLB_SIZE; ++vidx) {
1464         CPUTLBEntry *vtlb = &env_tlb(env)->d[mmu_idx].vtable[vidx];
1465         target_ulong cmp;
1466 
1467         /* elt_ofs might correspond to .addr_write, so use qatomic_read */
1468 #if TCG_OVERSIZED_GUEST
1469         cmp = *(target_ulong *)((uintptr_t)vtlb + elt_ofs);
1470 #else
1471         cmp = qatomic_read((target_ulong *)((uintptr_t)vtlb + elt_ofs));
1472 #endif
1473 
1474         if (cmp == page) {
1475             /* Found entry in victim tlb, swap tlb and iotlb.  */
1476             CPUTLBEntry tmptlb, *tlb = &env_tlb(env)->f[mmu_idx].table[index];
1477 
1478             qemu_spin_lock(&env_tlb(env)->c.lock);
1479             copy_tlb_helper_locked(&tmptlb, tlb);
1480             copy_tlb_helper_locked(tlb, vtlb);
1481             copy_tlb_helper_locked(vtlb, &tmptlb);
1482             qemu_spin_unlock(&env_tlb(env)->c.lock);
1483 
1484             CPUTLBEntryFull *f1 = &env_tlb(env)->d[mmu_idx].fulltlb[index];
1485             CPUTLBEntryFull *f2 = &env_tlb(env)->d[mmu_idx].vfulltlb[vidx];
1486             CPUTLBEntryFull tmpf;
1487             tmpf = *f1; *f1 = *f2; *f2 = tmpf;
1488             return true;
1489         }
1490     }
1491     return false;
1492 }
1493 
1494 /* Macro to call the above, with local variables from the use context.  */
1495 #define VICTIM_TLB_HIT(TY, ADDR) \
1496   victim_tlb_hit(env, mmu_idx, index, offsetof(CPUTLBEntry, TY), \
1497                  (ADDR) & TARGET_PAGE_MASK)
1498 
1499 static void notdirty_write(CPUState *cpu, vaddr mem_vaddr, unsigned size,
1500                            CPUTLBEntryFull *full, uintptr_t retaddr)
1501 {
1502     ram_addr_t ram_addr = mem_vaddr + full->xlat_section;
1503 
1504     trace_memory_notdirty_write_access(mem_vaddr, ram_addr, size);
1505 
1506     if (!cpu_physical_memory_get_dirty_flag(ram_addr, DIRTY_MEMORY_CODE)) {
1507         tb_invalidate_phys_range_fast(ram_addr, size, retaddr);
1508     }
1509 
1510     /*
1511      * Set both VGA and migration bits for simplicity and to remove
1512      * the notdirty callback faster.
1513      */
1514     cpu_physical_memory_set_dirty_range(ram_addr, size, DIRTY_CLIENTS_NOCODE);
1515 
1516     /* We remove the notdirty callback only if the code has been flushed. */
1517     if (!cpu_physical_memory_is_clean(ram_addr)) {
1518         trace_memory_notdirty_set_dirty(mem_vaddr);
1519         tlb_set_dirty(cpu, mem_vaddr);
1520     }
1521 }
1522 
1523 static int probe_access_internal(CPUArchState *env, target_ulong addr,
1524                                  int fault_size, MMUAccessType access_type,
1525                                  int mmu_idx, bool nonfault,
1526                                  void **phost, CPUTLBEntryFull **pfull,
1527                                  uintptr_t retaddr)
1528 {
1529     uintptr_t index = tlb_index(env, mmu_idx, addr);
1530     CPUTLBEntry *entry = tlb_entry(env, mmu_idx, addr);
1531     target_ulong tlb_addr, page_addr;
1532     size_t elt_ofs;
1533     int flags;
1534 
1535     switch (access_type) {
1536     case MMU_DATA_LOAD:
1537         elt_ofs = offsetof(CPUTLBEntry, addr_read);
1538         break;
1539     case MMU_DATA_STORE:
1540         elt_ofs = offsetof(CPUTLBEntry, addr_write);
1541         break;
1542     case MMU_INST_FETCH:
1543         elt_ofs = offsetof(CPUTLBEntry, addr_code);
1544         break;
1545     default:
1546         g_assert_not_reached();
1547     }
1548     tlb_addr = tlb_read_ofs(entry, elt_ofs);
1549 
1550     flags = TLB_FLAGS_MASK;
1551     page_addr = addr & TARGET_PAGE_MASK;
1552     if (!tlb_hit_page(tlb_addr, page_addr)) {
1553         if (!victim_tlb_hit(env, mmu_idx, index, elt_ofs, page_addr)) {
1554             CPUState *cs = env_cpu(env);
1555 
1556             if (!cs->cc->tcg_ops->tlb_fill(cs, addr, fault_size, access_type,
1557                                            mmu_idx, nonfault, retaddr)) {
1558                 /* Non-faulting page table read failed.  */
1559                 *phost = NULL;
1560                 *pfull = NULL;
1561                 return TLB_INVALID_MASK;
1562             }
1563 
1564             /* TLB resize via tlb_fill may have moved the entry.  */
1565             index = tlb_index(env, mmu_idx, addr);
1566             entry = tlb_entry(env, mmu_idx, addr);
1567 
1568             /*
1569              * With PAGE_WRITE_INV, we set TLB_INVALID_MASK immediately,
1570              * to force the next access through tlb_fill.  We've just
1571              * called tlb_fill, so we know that this entry *is* valid.
1572              */
1573             flags &= ~TLB_INVALID_MASK;
1574         }
1575         tlb_addr = tlb_read_ofs(entry, elt_ofs);
1576     }
1577     flags &= tlb_addr;
1578 
1579     *pfull = &env_tlb(env)->d[mmu_idx].fulltlb[index];
1580 
1581     /* Fold all "mmio-like" bits into TLB_MMIO.  This is not RAM.  */
1582     if (unlikely(flags & ~(TLB_WATCHPOINT | TLB_NOTDIRTY))) {
1583         *phost = NULL;
1584         return TLB_MMIO;
1585     }
1586 
1587     /* Everything else is RAM. */
1588     *phost = (void *)((uintptr_t)addr + entry->addend);
1589     return flags;
1590 }
1591 
1592 int probe_access_full(CPUArchState *env, target_ulong addr,
1593                       MMUAccessType access_type, int mmu_idx,
1594                       bool nonfault, void **phost, CPUTLBEntryFull **pfull,
1595                       uintptr_t retaddr)
1596 {
1597     int flags = probe_access_internal(env, addr, 0, access_type, mmu_idx,
1598                                       nonfault, phost, pfull, retaddr);
1599 
1600     /* Handle clean RAM pages.  */
1601     if (unlikely(flags & TLB_NOTDIRTY)) {
1602         notdirty_write(env_cpu(env), addr, 1, *pfull, retaddr);
1603         flags &= ~TLB_NOTDIRTY;
1604     }
1605 
1606     return flags;
1607 }
1608 
1609 int probe_access_flags(CPUArchState *env, target_ulong addr,
1610                        MMUAccessType access_type, int mmu_idx,
1611                        bool nonfault, void **phost, uintptr_t retaddr)
1612 {
1613     CPUTLBEntryFull *full;
1614 
1615     return probe_access_full(env, addr, access_type, mmu_idx,
1616                              nonfault, phost, &full, retaddr);
1617 }
1618 
1619 void *probe_access(CPUArchState *env, target_ulong addr, int size,
1620                    MMUAccessType access_type, int mmu_idx, uintptr_t retaddr)
1621 {
1622     CPUTLBEntryFull *full;
1623     void *host;
1624     int flags;
1625 
1626     g_assert(-(addr | TARGET_PAGE_MASK) >= size);
1627 
1628     flags = probe_access_internal(env, addr, size, access_type, mmu_idx,
1629                                   false, &host, &full, retaddr);
1630 
1631     /* Per the interface, size == 0 merely faults the access. */
1632     if (size == 0) {
1633         return NULL;
1634     }
1635 
1636     if (unlikely(flags & (TLB_NOTDIRTY | TLB_WATCHPOINT))) {
1637         /* Handle watchpoints.  */
1638         if (flags & TLB_WATCHPOINT) {
1639             int wp_access = (access_type == MMU_DATA_STORE
1640                              ? BP_MEM_WRITE : BP_MEM_READ);
1641             cpu_check_watchpoint(env_cpu(env), addr, size,
1642                                  full->attrs, wp_access, retaddr);
1643         }
1644 
1645         /* Handle clean RAM pages.  */
1646         if (flags & TLB_NOTDIRTY) {
1647             notdirty_write(env_cpu(env), addr, 1, full, retaddr);
1648         }
1649     }
1650 
1651     return host;
1652 }
1653 
1654 void *tlb_vaddr_to_host(CPUArchState *env, abi_ptr addr,
1655                         MMUAccessType access_type, int mmu_idx)
1656 {
1657     CPUTLBEntryFull *full;
1658     void *host;
1659     int flags;
1660 
1661     flags = probe_access_internal(env, addr, 0, access_type,
1662                                   mmu_idx, true, &host, &full, 0);
1663 
1664     /* No combination of flags are expected by the caller. */
1665     return flags ? NULL : host;
1666 }
1667 
1668 /*
1669  * Return a ram_addr_t for the virtual address for execution.
1670  *
1671  * Return -1 if we can't translate and execute from an entire page
1672  * of RAM.  This will force us to execute by loading and translating
1673  * one insn at a time, without caching.
1674  *
1675  * NOTE: This function will trigger an exception if the page is
1676  * not executable.
1677  */
1678 tb_page_addr_t get_page_addr_code_hostp(CPUArchState *env, target_ulong addr,
1679                                         void **hostp)
1680 {
1681     CPUTLBEntryFull *full;
1682     void *p;
1683 
1684     (void)probe_access_internal(env, addr, 1, MMU_INST_FETCH,
1685                                 cpu_mmu_index(env, true), false, &p, &full, 0);
1686     if (p == NULL) {
1687         return -1;
1688     }
1689     if (hostp) {
1690         *hostp = p;
1691     }
1692     return qemu_ram_addr_from_host_nofail(p);
1693 }
1694 
1695 #ifdef CONFIG_PLUGIN
1696 /*
1697  * Perform a TLB lookup and populate the qemu_plugin_hwaddr structure.
1698  * This should be a hot path as we will have just looked this path up
1699  * in the softmmu lookup code (or helper). We don't handle re-fills or
1700  * checking the victim table. This is purely informational.
1701  *
1702  * This almost never fails as the memory access being instrumented
1703  * should have just filled the TLB. The one corner case is io_writex
1704  * which can cause TLB flushes and potential resizing of the TLBs
1705  * losing the information we need. In those cases we need to recover
1706  * data from a copy of the CPUTLBEntryFull. As long as this always occurs
1707  * from the same thread (which a mem callback will be) this is safe.
1708  */
1709 
1710 bool tlb_plugin_lookup(CPUState *cpu, target_ulong addr, int mmu_idx,
1711                        bool is_store, struct qemu_plugin_hwaddr *data)
1712 {
1713     CPUArchState *env = cpu->env_ptr;
1714     CPUTLBEntry *tlbe = tlb_entry(env, mmu_idx, addr);
1715     uintptr_t index = tlb_index(env, mmu_idx, addr);
1716     target_ulong tlb_addr = is_store ? tlb_addr_write(tlbe) : tlbe->addr_read;
1717 
1718     if (likely(tlb_hit(tlb_addr, addr))) {
1719         /* We must have an iotlb entry for MMIO */
1720         if (tlb_addr & TLB_MMIO) {
1721             CPUTLBEntryFull *full;
1722             full = &env_tlb(env)->d[mmu_idx].fulltlb[index];
1723             data->is_io = true;
1724             data->v.io.section =
1725                 iotlb_to_section(cpu, full->xlat_section, full->attrs);
1726             data->v.io.offset = (full->xlat_section & TARGET_PAGE_MASK) + addr;
1727         } else {
1728             data->is_io = false;
1729             data->v.ram.hostaddr = (void *)((uintptr_t)addr + tlbe->addend);
1730         }
1731         return true;
1732     } else {
1733         SavedIOTLB *saved = &cpu->saved_iotlb;
1734         data->is_io = true;
1735         data->v.io.section = saved->section;
1736         data->v.io.offset = saved->mr_offset;
1737         return true;
1738     }
1739 }
1740 
1741 #endif
1742 
1743 /*
1744  * Probe for an atomic operation.  Do not allow unaligned operations,
1745  * or io operations to proceed.  Return the host address.
1746  *
1747  * @prot may be PAGE_READ, PAGE_WRITE, or PAGE_READ|PAGE_WRITE.
1748  */
1749 static void *atomic_mmu_lookup(CPUArchState *env, target_ulong addr,
1750                                MemOpIdx oi, int size, int prot,
1751                                uintptr_t retaddr)
1752 {
1753     uintptr_t mmu_idx = get_mmuidx(oi);
1754     MemOp mop = get_memop(oi);
1755     int a_bits = get_alignment_bits(mop);
1756     uintptr_t index;
1757     CPUTLBEntry *tlbe;
1758     target_ulong tlb_addr;
1759     void *hostaddr;
1760 
1761     tcg_debug_assert(mmu_idx < NB_MMU_MODES);
1762 
1763     /* Adjust the given return address.  */
1764     retaddr -= GETPC_ADJ;
1765 
1766     /* Enforce guest required alignment.  */
1767     if (unlikely(a_bits > 0 && (addr & ((1 << a_bits) - 1)))) {
1768         /* ??? Maybe indicate atomic op to cpu_unaligned_access */
1769         cpu_unaligned_access(env_cpu(env), addr, MMU_DATA_STORE,
1770                              mmu_idx, retaddr);
1771     }
1772 
1773     /* Enforce qemu required alignment.  */
1774     if (unlikely(addr & (size - 1))) {
1775         /* We get here if guest alignment was not requested,
1776            or was not enforced by cpu_unaligned_access above.
1777            We might widen the access and emulate, but for now
1778            mark an exception and exit the cpu loop.  */
1779         goto stop_the_world;
1780     }
1781 
1782     index = tlb_index(env, mmu_idx, addr);
1783     tlbe = tlb_entry(env, mmu_idx, addr);
1784 
1785     /* Check TLB entry and enforce page permissions.  */
1786     if (prot & PAGE_WRITE) {
1787         tlb_addr = tlb_addr_write(tlbe);
1788         if (!tlb_hit(tlb_addr, addr)) {
1789             if (!VICTIM_TLB_HIT(addr_write, addr)) {
1790                 tlb_fill(env_cpu(env), addr, size,
1791                          MMU_DATA_STORE, mmu_idx, retaddr);
1792                 index = tlb_index(env, mmu_idx, addr);
1793                 tlbe = tlb_entry(env, mmu_idx, addr);
1794             }
1795             tlb_addr = tlb_addr_write(tlbe) & ~TLB_INVALID_MASK;
1796         }
1797 
1798         /* Let the guest notice RMW on a write-only page.  */
1799         if ((prot & PAGE_READ) &&
1800             unlikely(tlbe->addr_read != (tlb_addr & ~TLB_NOTDIRTY))) {
1801             tlb_fill(env_cpu(env), addr, size,
1802                      MMU_DATA_LOAD, mmu_idx, retaddr);
1803             /*
1804              * Since we don't support reads and writes to different addresses,
1805              * and we do have the proper page loaded for write, this shouldn't
1806              * ever return.  But just in case, handle via stop-the-world.
1807              */
1808             goto stop_the_world;
1809         }
1810     } else /* if (prot & PAGE_READ) */ {
1811         tlb_addr = tlbe->addr_read;
1812         if (!tlb_hit(tlb_addr, addr)) {
1813             if (!VICTIM_TLB_HIT(addr_write, addr)) {
1814                 tlb_fill(env_cpu(env), addr, size,
1815                          MMU_DATA_LOAD, mmu_idx, retaddr);
1816                 index = tlb_index(env, mmu_idx, addr);
1817                 tlbe = tlb_entry(env, mmu_idx, addr);
1818             }
1819             tlb_addr = tlbe->addr_read & ~TLB_INVALID_MASK;
1820         }
1821     }
1822 
1823     /* Notice an IO access or a needs-MMU-lookup access */
1824     if (unlikely(tlb_addr & TLB_MMIO)) {
1825         /* There's really nothing that can be done to
1826            support this apart from stop-the-world.  */
1827         goto stop_the_world;
1828     }
1829 
1830     hostaddr = (void *)((uintptr_t)addr + tlbe->addend);
1831 
1832     if (unlikely(tlb_addr & TLB_NOTDIRTY)) {
1833         notdirty_write(env_cpu(env), addr, size,
1834                        &env_tlb(env)->d[mmu_idx].fulltlb[index], retaddr);
1835     }
1836 
1837     return hostaddr;
1838 
1839  stop_the_world:
1840     cpu_loop_exit_atomic(env_cpu(env), retaddr);
1841 }
1842 
1843 /*
1844  * Verify that we have passed the correct MemOp to the correct function.
1845  *
1846  * In the case of the helper_*_mmu functions, we will have done this by
1847  * using the MemOp to look up the helper during code generation.
1848  *
1849  * In the case of the cpu_*_mmu functions, this is up to the caller.
1850  * We could present one function to target code, and dispatch based on
1851  * the MemOp, but so far we have worked hard to avoid an indirect function
1852  * call along the memory path.
1853  */
1854 static void validate_memop(MemOpIdx oi, MemOp expected)
1855 {
1856 #ifdef CONFIG_DEBUG_TCG
1857     MemOp have = get_memop(oi) & (MO_SIZE | MO_BSWAP);
1858     assert(have == expected);
1859 #endif
1860 }
1861 
1862 /*
1863  * Load Helpers
1864  *
1865  * We support two different access types. SOFTMMU_CODE_ACCESS is
1866  * specifically for reading instructions from system memory. It is
1867  * called by the translation loop and in some helpers where the code
1868  * is disassembled. It shouldn't be called directly by guest code.
1869  */
1870 
1871 typedef uint64_t FullLoadHelper(CPUArchState *env, target_ulong addr,
1872                                 MemOpIdx oi, uintptr_t retaddr);
1873 
1874 static inline uint64_t QEMU_ALWAYS_INLINE
1875 load_memop(const void *haddr, MemOp op)
1876 {
1877     switch (op) {
1878     case MO_UB:
1879         return ldub_p(haddr);
1880     case MO_BEUW:
1881         return lduw_be_p(haddr);
1882     case MO_LEUW:
1883         return lduw_le_p(haddr);
1884     case MO_BEUL:
1885         return (uint32_t)ldl_be_p(haddr);
1886     case MO_LEUL:
1887         return (uint32_t)ldl_le_p(haddr);
1888     case MO_BEUQ:
1889         return ldq_be_p(haddr);
1890     case MO_LEUQ:
1891         return ldq_le_p(haddr);
1892     default:
1893         qemu_build_not_reached();
1894     }
1895 }
1896 
1897 static inline uint64_t QEMU_ALWAYS_INLINE
1898 load_helper(CPUArchState *env, target_ulong addr, MemOpIdx oi,
1899             uintptr_t retaddr, MemOp op, bool code_read,
1900             FullLoadHelper *full_load)
1901 {
1902     const size_t tlb_off = code_read ?
1903         offsetof(CPUTLBEntry, addr_code) : offsetof(CPUTLBEntry, addr_read);
1904     const MMUAccessType access_type =
1905         code_read ? MMU_INST_FETCH : MMU_DATA_LOAD;
1906     const unsigned a_bits = get_alignment_bits(get_memop(oi));
1907     const size_t size = memop_size(op);
1908     uintptr_t mmu_idx = get_mmuidx(oi);
1909     uintptr_t index;
1910     CPUTLBEntry *entry;
1911     target_ulong tlb_addr;
1912     void *haddr;
1913     uint64_t res;
1914 
1915     tcg_debug_assert(mmu_idx < NB_MMU_MODES);
1916 
1917     /* Handle CPU specific unaligned behaviour */
1918     if (addr & ((1 << a_bits) - 1)) {
1919         cpu_unaligned_access(env_cpu(env), addr, access_type,
1920                              mmu_idx, retaddr);
1921     }
1922 
1923     index = tlb_index(env, mmu_idx, addr);
1924     entry = tlb_entry(env, mmu_idx, addr);
1925     tlb_addr = code_read ? entry->addr_code : entry->addr_read;
1926 
1927     /* If the TLB entry is for a different page, reload and try again.  */
1928     if (!tlb_hit(tlb_addr, addr)) {
1929         if (!victim_tlb_hit(env, mmu_idx, index, tlb_off,
1930                             addr & TARGET_PAGE_MASK)) {
1931             tlb_fill(env_cpu(env), addr, size,
1932                      access_type, mmu_idx, retaddr);
1933             index = tlb_index(env, mmu_idx, addr);
1934             entry = tlb_entry(env, mmu_idx, addr);
1935         }
1936         tlb_addr = code_read ? entry->addr_code : entry->addr_read;
1937         tlb_addr &= ~TLB_INVALID_MASK;
1938     }
1939 
1940     /* Handle anything that isn't just a straight memory access.  */
1941     if (unlikely(tlb_addr & ~TARGET_PAGE_MASK)) {
1942         CPUTLBEntryFull *full;
1943         bool need_swap;
1944 
1945         /* For anything that is unaligned, recurse through full_load.  */
1946         if ((addr & (size - 1)) != 0) {
1947             goto do_unaligned_access;
1948         }
1949 
1950         full = &env_tlb(env)->d[mmu_idx].fulltlb[index];
1951 
1952         /* Handle watchpoints.  */
1953         if (unlikely(tlb_addr & TLB_WATCHPOINT)) {
1954             /* On watchpoint hit, this will longjmp out.  */
1955             cpu_check_watchpoint(env_cpu(env), addr, size,
1956                                  full->attrs, BP_MEM_READ, retaddr);
1957         }
1958 
1959         need_swap = size > 1 && (tlb_addr & TLB_BSWAP);
1960 
1961         /* Handle I/O access.  */
1962         if (likely(tlb_addr & TLB_MMIO)) {
1963             return io_readx(env, full, mmu_idx, addr, retaddr,
1964                             access_type, op ^ (need_swap * MO_BSWAP));
1965         }
1966 
1967         haddr = (void *)((uintptr_t)addr + entry->addend);
1968 
1969         /*
1970          * Keep these two load_memop separate to ensure that the compiler
1971          * is able to fold the entire function to a single instruction.
1972          * There is a build-time assert inside to remind you of this.  ;-)
1973          */
1974         if (unlikely(need_swap)) {
1975             return load_memop(haddr, op ^ MO_BSWAP);
1976         }
1977         return load_memop(haddr, op);
1978     }
1979 
1980     /* Handle slow unaligned access (it spans two pages or IO).  */
1981     if (size > 1
1982         && unlikely((addr & ~TARGET_PAGE_MASK) + size - 1
1983                     >= TARGET_PAGE_SIZE)) {
1984         target_ulong addr1, addr2;
1985         uint64_t r1, r2;
1986         unsigned shift;
1987     do_unaligned_access:
1988         addr1 = addr & ~((target_ulong)size - 1);
1989         addr2 = addr1 + size;
1990         r1 = full_load(env, addr1, oi, retaddr);
1991         r2 = full_load(env, addr2, oi, retaddr);
1992         shift = (addr & (size - 1)) * 8;
1993 
1994         if (memop_big_endian(op)) {
1995             /* Big-endian combine.  */
1996             res = (r1 << shift) | (r2 >> ((size * 8) - shift));
1997         } else {
1998             /* Little-endian combine.  */
1999             res = (r1 >> shift) | (r2 << ((size * 8) - shift));
2000         }
2001         return res & MAKE_64BIT_MASK(0, size * 8);
2002     }
2003 
2004     haddr = (void *)((uintptr_t)addr + entry->addend);
2005     return load_memop(haddr, op);
2006 }
2007 
2008 /*
2009  * For the benefit of TCG generated code, we want to avoid the
2010  * complication of ABI-specific return type promotion and always
2011  * return a value extended to the register size of the host. This is
2012  * tcg_target_long, except in the case of a 32-bit host and 64-bit
2013  * data, and for that we always have uint64_t.
2014  *
2015  * We don't bother with this widened value for SOFTMMU_CODE_ACCESS.
2016  */
2017 
2018 static uint64_t full_ldub_mmu(CPUArchState *env, target_ulong addr,
2019                               MemOpIdx oi, uintptr_t retaddr)
2020 {
2021     validate_memop(oi, MO_UB);
2022     return load_helper(env, addr, oi, retaddr, MO_UB, false, full_ldub_mmu);
2023 }
2024 
2025 tcg_target_ulong helper_ret_ldub_mmu(CPUArchState *env, target_ulong addr,
2026                                      MemOpIdx oi, uintptr_t retaddr)
2027 {
2028     return full_ldub_mmu(env, addr, oi, retaddr);
2029 }
2030 
2031 static uint64_t full_le_lduw_mmu(CPUArchState *env, target_ulong addr,
2032                                  MemOpIdx oi, uintptr_t retaddr)
2033 {
2034     validate_memop(oi, MO_LEUW);
2035     return load_helper(env, addr, oi, retaddr, MO_LEUW, false,
2036                        full_le_lduw_mmu);
2037 }
2038 
2039 tcg_target_ulong helper_le_lduw_mmu(CPUArchState *env, target_ulong addr,
2040                                     MemOpIdx oi, uintptr_t retaddr)
2041 {
2042     return full_le_lduw_mmu(env, addr, oi, retaddr);
2043 }
2044 
2045 static uint64_t full_be_lduw_mmu(CPUArchState *env, target_ulong addr,
2046                                  MemOpIdx oi, uintptr_t retaddr)
2047 {
2048     validate_memop(oi, MO_BEUW);
2049     return load_helper(env, addr, oi, retaddr, MO_BEUW, false,
2050                        full_be_lduw_mmu);
2051 }
2052 
2053 tcg_target_ulong helper_be_lduw_mmu(CPUArchState *env, target_ulong addr,
2054                                     MemOpIdx oi, uintptr_t retaddr)
2055 {
2056     return full_be_lduw_mmu(env, addr, oi, retaddr);
2057 }
2058 
2059 static uint64_t full_le_ldul_mmu(CPUArchState *env, target_ulong addr,
2060                                  MemOpIdx oi, uintptr_t retaddr)
2061 {
2062     validate_memop(oi, MO_LEUL);
2063     return load_helper(env, addr, oi, retaddr, MO_LEUL, false,
2064                        full_le_ldul_mmu);
2065 }
2066 
2067 tcg_target_ulong helper_le_ldul_mmu(CPUArchState *env, target_ulong addr,
2068                                     MemOpIdx oi, uintptr_t retaddr)
2069 {
2070     return full_le_ldul_mmu(env, addr, oi, retaddr);
2071 }
2072 
2073 static uint64_t full_be_ldul_mmu(CPUArchState *env, target_ulong addr,
2074                                  MemOpIdx oi, uintptr_t retaddr)
2075 {
2076     validate_memop(oi, MO_BEUL);
2077     return load_helper(env, addr, oi, retaddr, MO_BEUL, false,
2078                        full_be_ldul_mmu);
2079 }
2080 
2081 tcg_target_ulong helper_be_ldul_mmu(CPUArchState *env, target_ulong addr,
2082                                     MemOpIdx oi, uintptr_t retaddr)
2083 {
2084     return full_be_ldul_mmu(env, addr, oi, retaddr);
2085 }
2086 
2087 uint64_t helper_le_ldq_mmu(CPUArchState *env, target_ulong addr,
2088                            MemOpIdx oi, uintptr_t retaddr)
2089 {
2090     validate_memop(oi, MO_LEUQ);
2091     return load_helper(env, addr, oi, retaddr, MO_LEUQ, false,
2092                        helper_le_ldq_mmu);
2093 }
2094 
2095 uint64_t helper_be_ldq_mmu(CPUArchState *env, target_ulong addr,
2096                            MemOpIdx oi, uintptr_t retaddr)
2097 {
2098     validate_memop(oi, MO_BEUQ);
2099     return load_helper(env, addr, oi, retaddr, MO_BEUQ, false,
2100                        helper_be_ldq_mmu);
2101 }
2102 
2103 /*
2104  * Provide signed versions of the load routines as well.  We can of course
2105  * avoid this for 64-bit data, or for 32-bit data on 32-bit host.
2106  */
2107 
2108 
2109 tcg_target_ulong helper_ret_ldsb_mmu(CPUArchState *env, target_ulong addr,
2110                                      MemOpIdx oi, uintptr_t retaddr)
2111 {
2112     return (int8_t)helper_ret_ldub_mmu(env, addr, oi, retaddr);
2113 }
2114 
2115 tcg_target_ulong helper_le_ldsw_mmu(CPUArchState *env, target_ulong addr,
2116                                     MemOpIdx oi, uintptr_t retaddr)
2117 {
2118     return (int16_t)helper_le_lduw_mmu(env, addr, oi, retaddr);
2119 }
2120 
2121 tcg_target_ulong helper_be_ldsw_mmu(CPUArchState *env, target_ulong addr,
2122                                     MemOpIdx oi, uintptr_t retaddr)
2123 {
2124     return (int16_t)helper_be_lduw_mmu(env, addr, oi, retaddr);
2125 }
2126 
2127 tcg_target_ulong helper_le_ldsl_mmu(CPUArchState *env, target_ulong addr,
2128                                     MemOpIdx oi, uintptr_t retaddr)
2129 {
2130     return (int32_t)helper_le_ldul_mmu(env, addr, oi, retaddr);
2131 }
2132 
2133 tcg_target_ulong helper_be_ldsl_mmu(CPUArchState *env, target_ulong addr,
2134                                     MemOpIdx oi, uintptr_t retaddr)
2135 {
2136     return (int32_t)helper_be_ldul_mmu(env, addr, oi, retaddr);
2137 }
2138 
2139 /*
2140  * Load helpers for cpu_ldst.h.
2141  */
2142 
2143 static inline uint64_t cpu_load_helper(CPUArchState *env, abi_ptr addr,
2144                                        MemOpIdx oi, uintptr_t retaddr,
2145                                        FullLoadHelper *full_load)
2146 {
2147     uint64_t ret;
2148 
2149     ret = full_load(env, addr, oi, retaddr);
2150     qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_R);
2151     return ret;
2152 }
2153 
2154 uint8_t cpu_ldb_mmu(CPUArchState *env, abi_ptr addr, MemOpIdx oi, uintptr_t ra)
2155 {
2156     return cpu_load_helper(env, addr, oi, ra, full_ldub_mmu);
2157 }
2158 
2159 uint16_t cpu_ldw_be_mmu(CPUArchState *env, abi_ptr addr,
2160                         MemOpIdx oi, uintptr_t ra)
2161 {
2162     return cpu_load_helper(env, addr, oi, ra, full_be_lduw_mmu);
2163 }
2164 
2165 uint32_t cpu_ldl_be_mmu(CPUArchState *env, abi_ptr addr,
2166                         MemOpIdx oi, uintptr_t ra)
2167 {
2168     return cpu_load_helper(env, addr, oi, ra, full_be_ldul_mmu);
2169 }
2170 
2171 uint64_t cpu_ldq_be_mmu(CPUArchState *env, abi_ptr addr,
2172                         MemOpIdx oi, uintptr_t ra)
2173 {
2174     return cpu_load_helper(env, addr, oi, ra, helper_be_ldq_mmu);
2175 }
2176 
2177 uint16_t cpu_ldw_le_mmu(CPUArchState *env, abi_ptr addr,
2178                         MemOpIdx oi, uintptr_t ra)
2179 {
2180     return cpu_load_helper(env, addr, oi, ra, full_le_lduw_mmu);
2181 }
2182 
2183 uint32_t cpu_ldl_le_mmu(CPUArchState *env, abi_ptr addr,
2184                         MemOpIdx oi, uintptr_t ra)
2185 {
2186     return cpu_load_helper(env, addr, oi, ra, full_le_ldul_mmu);
2187 }
2188 
2189 uint64_t cpu_ldq_le_mmu(CPUArchState *env, abi_ptr addr,
2190                         MemOpIdx oi, uintptr_t ra)
2191 {
2192     return cpu_load_helper(env, addr, oi, ra, helper_le_ldq_mmu);
2193 }
2194 
2195 Int128 cpu_ld16_be_mmu(CPUArchState *env, abi_ptr addr,
2196                        MemOpIdx oi, uintptr_t ra)
2197 {
2198     MemOp mop = get_memop(oi);
2199     int mmu_idx = get_mmuidx(oi);
2200     MemOpIdx new_oi;
2201     unsigned a_bits;
2202     uint64_t h, l;
2203 
2204     tcg_debug_assert((mop & (MO_BSWAP|MO_SSIZE)) == (MO_BE|MO_128));
2205     a_bits = get_alignment_bits(mop);
2206 
2207     /* Handle CPU specific unaligned behaviour */
2208     if (addr & ((1 << a_bits) - 1)) {
2209         cpu_unaligned_access(env_cpu(env), addr, MMU_DATA_LOAD,
2210                              mmu_idx, ra);
2211     }
2212 
2213     /* Construct an unaligned 64-bit replacement MemOpIdx. */
2214     mop = (mop & ~(MO_SIZE | MO_AMASK)) | MO_64 | MO_UNALN;
2215     new_oi = make_memop_idx(mop, mmu_idx);
2216 
2217     h = helper_be_ldq_mmu(env, addr, new_oi, ra);
2218     l = helper_be_ldq_mmu(env, addr + 8, new_oi, ra);
2219 
2220     qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_R);
2221     return int128_make128(l, h);
2222 }
2223 
2224 Int128 cpu_ld16_le_mmu(CPUArchState *env, abi_ptr addr,
2225                        MemOpIdx oi, uintptr_t ra)
2226 {
2227     MemOp mop = get_memop(oi);
2228     int mmu_idx = get_mmuidx(oi);
2229     MemOpIdx new_oi;
2230     unsigned a_bits;
2231     uint64_t h, l;
2232 
2233     tcg_debug_assert((mop & (MO_BSWAP|MO_SSIZE)) == (MO_LE|MO_128));
2234     a_bits = get_alignment_bits(mop);
2235 
2236     /* Handle CPU specific unaligned behaviour */
2237     if (addr & ((1 << a_bits) - 1)) {
2238         cpu_unaligned_access(env_cpu(env), addr, MMU_DATA_LOAD,
2239                              mmu_idx, ra);
2240     }
2241 
2242     /* Construct an unaligned 64-bit replacement MemOpIdx. */
2243     mop = (mop & ~(MO_SIZE | MO_AMASK)) | MO_64 | MO_UNALN;
2244     new_oi = make_memop_idx(mop, mmu_idx);
2245 
2246     l = helper_le_ldq_mmu(env, addr, new_oi, ra);
2247     h = helper_le_ldq_mmu(env, addr + 8, new_oi, ra);
2248 
2249     qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_R);
2250     return int128_make128(l, h);
2251 }
2252 
2253 /*
2254  * Store Helpers
2255  */
2256 
2257 static inline void QEMU_ALWAYS_INLINE
2258 store_memop(void *haddr, uint64_t val, MemOp op)
2259 {
2260     switch (op) {
2261     case MO_UB:
2262         stb_p(haddr, val);
2263         break;
2264     case MO_BEUW:
2265         stw_be_p(haddr, val);
2266         break;
2267     case MO_LEUW:
2268         stw_le_p(haddr, val);
2269         break;
2270     case MO_BEUL:
2271         stl_be_p(haddr, val);
2272         break;
2273     case MO_LEUL:
2274         stl_le_p(haddr, val);
2275         break;
2276     case MO_BEUQ:
2277         stq_be_p(haddr, val);
2278         break;
2279     case MO_LEUQ:
2280         stq_le_p(haddr, val);
2281         break;
2282     default:
2283         qemu_build_not_reached();
2284     }
2285 }
2286 
2287 static void full_stb_mmu(CPUArchState *env, target_ulong addr, uint64_t val,
2288                          MemOpIdx oi, uintptr_t retaddr);
2289 
2290 static void __attribute__((noinline))
2291 store_helper_unaligned(CPUArchState *env, target_ulong addr, uint64_t val,
2292                        uintptr_t retaddr, size_t size, uintptr_t mmu_idx,
2293                        bool big_endian)
2294 {
2295     const size_t tlb_off = offsetof(CPUTLBEntry, addr_write);
2296     uintptr_t index, index2;
2297     CPUTLBEntry *entry, *entry2;
2298     target_ulong page1, page2, tlb_addr, tlb_addr2;
2299     MemOpIdx oi;
2300     size_t size2;
2301     int i;
2302 
2303     /*
2304      * Ensure the second page is in the TLB.  Note that the first page
2305      * is already guaranteed to be filled, and that the second page
2306      * cannot evict the first.  An exception to this rule is PAGE_WRITE_INV
2307      * handling: the first page could have evicted itself.
2308      */
2309     page1 = addr & TARGET_PAGE_MASK;
2310     page2 = (addr + size) & TARGET_PAGE_MASK;
2311     size2 = (addr + size) & ~TARGET_PAGE_MASK;
2312     index2 = tlb_index(env, mmu_idx, page2);
2313     entry2 = tlb_entry(env, mmu_idx, page2);
2314 
2315     tlb_addr2 = tlb_addr_write(entry2);
2316     if (page1 != page2 && !tlb_hit_page(tlb_addr2, page2)) {
2317         if (!victim_tlb_hit(env, mmu_idx, index2, tlb_off, page2)) {
2318             tlb_fill(env_cpu(env), page2, size2, MMU_DATA_STORE,
2319                      mmu_idx, retaddr);
2320             index2 = tlb_index(env, mmu_idx, page2);
2321             entry2 = tlb_entry(env, mmu_idx, page2);
2322         }
2323         tlb_addr2 = tlb_addr_write(entry2);
2324     }
2325 
2326     index = tlb_index(env, mmu_idx, addr);
2327     entry = tlb_entry(env, mmu_idx, addr);
2328     tlb_addr = tlb_addr_write(entry);
2329 
2330     /*
2331      * Handle watchpoints.  Since this may trap, all checks
2332      * must happen before any store.
2333      */
2334     if (unlikely(tlb_addr & TLB_WATCHPOINT)) {
2335         cpu_check_watchpoint(env_cpu(env), addr, size - size2,
2336                              env_tlb(env)->d[mmu_idx].fulltlb[index].attrs,
2337                              BP_MEM_WRITE, retaddr);
2338     }
2339     if (unlikely(tlb_addr2 & TLB_WATCHPOINT)) {
2340         cpu_check_watchpoint(env_cpu(env), page2, size2,
2341                              env_tlb(env)->d[mmu_idx].fulltlb[index2].attrs,
2342                              BP_MEM_WRITE, retaddr);
2343     }
2344 
2345     /*
2346      * XXX: not efficient, but simple.
2347      * This loop must go in the forward direction to avoid issues
2348      * with self-modifying code in Windows 64-bit.
2349      */
2350     oi = make_memop_idx(MO_UB, mmu_idx);
2351     if (big_endian) {
2352         for (i = 0; i < size; ++i) {
2353             /* Big-endian extract.  */
2354             uint8_t val8 = val >> (((size - 1) * 8) - (i * 8));
2355             full_stb_mmu(env, addr + i, val8, oi, retaddr);
2356         }
2357     } else {
2358         for (i = 0; i < size; ++i) {
2359             /* Little-endian extract.  */
2360             uint8_t val8 = val >> (i * 8);
2361             full_stb_mmu(env, addr + i, val8, oi, retaddr);
2362         }
2363     }
2364 }
2365 
2366 static inline void QEMU_ALWAYS_INLINE
2367 store_helper(CPUArchState *env, target_ulong addr, uint64_t val,
2368              MemOpIdx oi, uintptr_t retaddr, MemOp op)
2369 {
2370     const size_t tlb_off = offsetof(CPUTLBEntry, addr_write);
2371     const unsigned a_bits = get_alignment_bits(get_memop(oi));
2372     const size_t size = memop_size(op);
2373     uintptr_t mmu_idx = get_mmuidx(oi);
2374     uintptr_t index;
2375     CPUTLBEntry *entry;
2376     target_ulong tlb_addr;
2377     void *haddr;
2378 
2379     tcg_debug_assert(mmu_idx < NB_MMU_MODES);
2380 
2381     /* Handle CPU specific unaligned behaviour */
2382     if (addr & ((1 << a_bits) - 1)) {
2383         cpu_unaligned_access(env_cpu(env), addr, MMU_DATA_STORE,
2384                              mmu_idx, retaddr);
2385     }
2386 
2387     index = tlb_index(env, mmu_idx, addr);
2388     entry = tlb_entry(env, mmu_idx, addr);
2389     tlb_addr = tlb_addr_write(entry);
2390 
2391     /* If the TLB entry is for a different page, reload and try again.  */
2392     if (!tlb_hit(tlb_addr, addr)) {
2393         if (!victim_tlb_hit(env, mmu_idx, index, tlb_off,
2394             addr & TARGET_PAGE_MASK)) {
2395             tlb_fill(env_cpu(env), addr, size, MMU_DATA_STORE,
2396                      mmu_idx, retaddr);
2397             index = tlb_index(env, mmu_idx, addr);
2398             entry = tlb_entry(env, mmu_idx, addr);
2399         }
2400         tlb_addr = tlb_addr_write(entry) & ~TLB_INVALID_MASK;
2401     }
2402 
2403     /* Handle anything that isn't just a straight memory access.  */
2404     if (unlikely(tlb_addr & ~TARGET_PAGE_MASK)) {
2405         CPUTLBEntryFull *full;
2406         bool need_swap;
2407 
2408         /* For anything that is unaligned, recurse through byte stores.  */
2409         if ((addr & (size - 1)) != 0) {
2410             goto do_unaligned_access;
2411         }
2412 
2413         full = &env_tlb(env)->d[mmu_idx].fulltlb[index];
2414 
2415         /* Handle watchpoints.  */
2416         if (unlikely(tlb_addr & TLB_WATCHPOINT)) {
2417             /* On watchpoint hit, this will longjmp out.  */
2418             cpu_check_watchpoint(env_cpu(env), addr, size,
2419                                  full->attrs, BP_MEM_WRITE, retaddr);
2420         }
2421 
2422         need_swap = size > 1 && (tlb_addr & TLB_BSWAP);
2423 
2424         /* Handle I/O access.  */
2425         if (tlb_addr & TLB_MMIO) {
2426             io_writex(env, full, mmu_idx, val, addr, retaddr,
2427                       op ^ (need_swap * MO_BSWAP));
2428             return;
2429         }
2430 
2431         /* Ignore writes to ROM.  */
2432         if (unlikely(tlb_addr & TLB_DISCARD_WRITE)) {
2433             return;
2434         }
2435 
2436         /* Handle clean RAM pages.  */
2437         if (tlb_addr & TLB_NOTDIRTY) {
2438             notdirty_write(env_cpu(env), addr, size, full, retaddr);
2439         }
2440 
2441         haddr = (void *)((uintptr_t)addr + entry->addend);
2442 
2443         /*
2444          * Keep these two store_memop separate to ensure that the compiler
2445          * is able to fold the entire function to a single instruction.
2446          * There is a build-time assert inside to remind you of this.  ;-)
2447          */
2448         if (unlikely(need_swap)) {
2449             store_memop(haddr, val, op ^ MO_BSWAP);
2450         } else {
2451             store_memop(haddr, val, op);
2452         }
2453         return;
2454     }
2455 
2456     /* Handle slow unaligned access (it spans two pages or IO).  */
2457     if (size > 1
2458         && unlikely((addr & ~TARGET_PAGE_MASK) + size - 1
2459                      >= TARGET_PAGE_SIZE)) {
2460     do_unaligned_access:
2461         store_helper_unaligned(env, addr, val, retaddr, size,
2462                                mmu_idx, memop_big_endian(op));
2463         return;
2464     }
2465 
2466     haddr = (void *)((uintptr_t)addr + entry->addend);
2467     store_memop(haddr, val, op);
2468 }
2469 
2470 static void __attribute__((noinline))
2471 full_stb_mmu(CPUArchState *env, target_ulong addr, uint64_t val,
2472              MemOpIdx oi, uintptr_t retaddr)
2473 {
2474     validate_memop(oi, MO_UB);
2475     store_helper(env, addr, val, oi, retaddr, MO_UB);
2476 }
2477 
2478 void helper_ret_stb_mmu(CPUArchState *env, target_ulong addr, uint8_t val,
2479                         MemOpIdx oi, uintptr_t retaddr)
2480 {
2481     full_stb_mmu(env, addr, val, oi, retaddr);
2482 }
2483 
2484 static void full_le_stw_mmu(CPUArchState *env, target_ulong addr, uint64_t val,
2485                             MemOpIdx oi, uintptr_t retaddr)
2486 {
2487     validate_memop(oi, MO_LEUW);
2488     store_helper(env, addr, val, oi, retaddr, MO_LEUW);
2489 }
2490 
2491 void helper_le_stw_mmu(CPUArchState *env, target_ulong addr, uint16_t val,
2492                        MemOpIdx oi, uintptr_t retaddr)
2493 {
2494     full_le_stw_mmu(env, addr, val, oi, retaddr);
2495 }
2496 
2497 static void full_be_stw_mmu(CPUArchState *env, target_ulong addr, uint64_t val,
2498                             MemOpIdx oi, uintptr_t retaddr)
2499 {
2500     validate_memop(oi, MO_BEUW);
2501     store_helper(env, addr, val, oi, retaddr, MO_BEUW);
2502 }
2503 
2504 void helper_be_stw_mmu(CPUArchState *env, target_ulong addr, uint16_t val,
2505                        MemOpIdx oi, uintptr_t retaddr)
2506 {
2507     full_be_stw_mmu(env, addr, val, oi, retaddr);
2508 }
2509 
2510 static void full_le_stl_mmu(CPUArchState *env, target_ulong addr, uint64_t val,
2511                             MemOpIdx oi, uintptr_t retaddr)
2512 {
2513     validate_memop(oi, MO_LEUL);
2514     store_helper(env, addr, val, oi, retaddr, MO_LEUL);
2515 }
2516 
2517 void helper_le_stl_mmu(CPUArchState *env, target_ulong addr, uint32_t val,
2518                        MemOpIdx oi, uintptr_t retaddr)
2519 {
2520     full_le_stl_mmu(env, addr, val, oi, retaddr);
2521 }
2522 
2523 static void full_be_stl_mmu(CPUArchState *env, target_ulong addr, uint64_t val,
2524                             MemOpIdx oi, uintptr_t retaddr)
2525 {
2526     validate_memop(oi, MO_BEUL);
2527     store_helper(env, addr, val, oi, retaddr, MO_BEUL);
2528 }
2529 
2530 void helper_be_stl_mmu(CPUArchState *env, target_ulong addr, uint32_t val,
2531                        MemOpIdx oi, uintptr_t retaddr)
2532 {
2533     full_be_stl_mmu(env, addr, val, oi, retaddr);
2534 }
2535 
2536 void helper_le_stq_mmu(CPUArchState *env, target_ulong addr, uint64_t val,
2537                        MemOpIdx oi, uintptr_t retaddr)
2538 {
2539     validate_memop(oi, MO_LEUQ);
2540     store_helper(env, addr, val, oi, retaddr, MO_LEUQ);
2541 }
2542 
2543 void helper_be_stq_mmu(CPUArchState *env, target_ulong addr, uint64_t val,
2544                        MemOpIdx oi, uintptr_t retaddr)
2545 {
2546     validate_memop(oi, MO_BEUQ);
2547     store_helper(env, addr, val, oi, retaddr, MO_BEUQ);
2548 }
2549 
2550 /*
2551  * Store Helpers for cpu_ldst.h
2552  */
2553 
2554 typedef void FullStoreHelper(CPUArchState *env, target_ulong addr,
2555                              uint64_t val, MemOpIdx oi, uintptr_t retaddr);
2556 
2557 static inline void cpu_store_helper(CPUArchState *env, target_ulong addr,
2558                                     uint64_t val, MemOpIdx oi, uintptr_t ra,
2559                                     FullStoreHelper *full_store)
2560 {
2561     full_store(env, addr, val, oi, ra);
2562     qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_W);
2563 }
2564 
2565 void cpu_stb_mmu(CPUArchState *env, target_ulong addr, uint8_t val,
2566                  MemOpIdx oi, uintptr_t retaddr)
2567 {
2568     cpu_store_helper(env, addr, val, oi, retaddr, full_stb_mmu);
2569 }
2570 
2571 void cpu_stw_be_mmu(CPUArchState *env, target_ulong addr, uint16_t val,
2572                     MemOpIdx oi, uintptr_t retaddr)
2573 {
2574     cpu_store_helper(env, addr, val, oi, retaddr, full_be_stw_mmu);
2575 }
2576 
2577 void cpu_stl_be_mmu(CPUArchState *env, target_ulong addr, uint32_t val,
2578                     MemOpIdx oi, uintptr_t retaddr)
2579 {
2580     cpu_store_helper(env, addr, val, oi, retaddr, full_be_stl_mmu);
2581 }
2582 
2583 void cpu_stq_be_mmu(CPUArchState *env, target_ulong addr, uint64_t val,
2584                     MemOpIdx oi, uintptr_t retaddr)
2585 {
2586     cpu_store_helper(env, addr, val, oi, retaddr, helper_be_stq_mmu);
2587 }
2588 
2589 void cpu_stw_le_mmu(CPUArchState *env, target_ulong addr, uint16_t val,
2590                     MemOpIdx oi, uintptr_t retaddr)
2591 {
2592     cpu_store_helper(env, addr, val, oi, retaddr, full_le_stw_mmu);
2593 }
2594 
2595 void cpu_stl_le_mmu(CPUArchState *env, target_ulong addr, uint32_t val,
2596                     MemOpIdx oi, uintptr_t retaddr)
2597 {
2598     cpu_store_helper(env, addr, val, oi, retaddr, full_le_stl_mmu);
2599 }
2600 
2601 void cpu_stq_le_mmu(CPUArchState *env, target_ulong addr, uint64_t val,
2602                     MemOpIdx oi, uintptr_t retaddr)
2603 {
2604     cpu_store_helper(env, addr, val, oi, retaddr, helper_le_stq_mmu);
2605 }
2606 
2607 void cpu_st16_be_mmu(CPUArchState *env, abi_ptr addr, Int128 val,
2608                      MemOpIdx oi, uintptr_t ra)
2609 {
2610     MemOp mop = get_memop(oi);
2611     int mmu_idx = get_mmuidx(oi);
2612     MemOpIdx new_oi;
2613     unsigned a_bits;
2614 
2615     tcg_debug_assert((mop & (MO_BSWAP|MO_SSIZE)) == (MO_BE|MO_128));
2616     a_bits = get_alignment_bits(mop);
2617 
2618     /* Handle CPU specific unaligned behaviour */
2619     if (addr & ((1 << a_bits) - 1)) {
2620         cpu_unaligned_access(env_cpu(env), addr, MMU_DATA_STORE,
2621                              mmu_idx, ra);
2622     }
2623 
2624     /* Construct an unaligned 64-bit replacement MemOpIdx. */
2625     mop = (mop & ~(MO_SIZE | MO_AMASK)) | MO_64 | MO_UNALN;
2626     new_oi = make_memop_idx(mop, mmu_idx);
2627 
2628     helper_be_stq_mmu(env, addr, int128_gethi(val), new_oi, ra);
2629     helper_be_stq_mmu(env, addr + 8, int128_getlo(val), new_oi, ra);
2630 
2631     qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_W);
2632 }
2633 
2634 void cpu_st16_le_mmu(CPUArchState *env, abi_ptr addr, Int128 val,
2635                      MemOpIdx oi, uintptr_t ra)
2636 {
2637     MemOp mop = get_memop(oi);
2638     int mmu_idx = get_mmuidx(oi);
2639     MemOpIdx new_oi;
2640     unsigned a_bits;
2641 
2642     tcg_debug_assert((mop & (MO_BSWAP|MO_SSIZE)) == (MO_LE|MO_128));
2643     a_bits = get_alignment_bits(mop);
2644 
2645     /* Handle CPU specific unaligned behaviour */
2646     if (addr & ((1 << a_bits) - 1)) {
2647         cpu_unaligned_access(env_cpu(env), addr, MMU_DATA_STORE,
2648                              mmu_idx, ra);
2649     }
2650 
2651     /* Construct an unaligned 64-bit replacement MemOpIdx. */
2652     mop = (mop & ~(MO_SIZE | MO_AMASK)) | MO_64 | MO_UNALN;
2653     new_oi = make_memop_idx(mop, mmu_idx);
2654 
2655     helper_le_stq_mmu(env, addr, int128_getlo(val), new_oi, ra);
2656     helper_le_stq_mmu(env, addr + 8, int128_gethi(val), new_oi, ra);
2657 
2658     qemu_plugin_vcpu_mem_cb(env_cpu(env), addr, oi, QEMU_PLUGIN_MEM_W);
2659 }
2660 
2661 #include "ldst_common.c.inc"
2662 
2663 /*
2664  * First set of functions passes in OI and RETADDR.
2665  * This makes them callable from other helpers.
2666  */
2667 
2668 #define ATOMIC_NAME(X) \
2669     glue(glue(glue(cpu_atomic_ ## X, SUFFIX), END), _mmu)
2670 
2671 #define ATOMIC_MMU_CLEANUP
2672 
2673 #include "atomic_common.c.inc"
2674 
2675 #define DATA_SIZE 1
2676 #include "atomic_template.h"
2677 
2678 #define DATA_SIZE 2
2679 #include "atomic_template.h"
2680 
2681 #define DATA_SIZE 4
2682 #include "atomic_template.h"
2683 
2684 #ifdef CONFIG_ATOMIC64
2685 #define DATA_SIZE 8
2686 #include "atomic_template.h"
2687 #endif
2688 
2689 #if HAVE_CMPXCHG128 || HAVE_ATOMIC128
2690 #define DATA_SIZE 16
2691 #include "atomic_template.h"
2692 #endif
2693 
2694 /* Code access functions.  */
2695 
2696 static uint64_t full_ldub_code(CPUArchState *env, target_ulong addr,
2697                                MemOpIdx oi, uintptr_t retaddr)
2698 {
2699     return load_helper(env, addr, oi, retaddr, MO_8, true, full_ldub_code);
2700 }
2701 
2702 uint32_t cpu_ldub_code(CPUArchState *env, abi_ptr addr)
2703 {
2704     MemOpIdx oi = make_memop_idx(MO_UB, cpu_mmu_index(env, true));
2705     return full_ldub_code(env, addr, oi, 0);
2706 }
2707 
2708 static uint64_t full_lduw_code(CPUArchState *env, target_ulong addr,
2709                                MemOpIdx oi, uintptr_t retaddr)
2710 {
2711     return load_helper(env, addr, oi, retaddr, MO_TEUW, true, full_lduw_code);
2712 }
2713 
2714 uint32_t cpu_lduw_code(CPUArchState *env, abi_ptr addr)
2715 {
2716     MemOpIdx oi = make_memop_idx(MO_TEUW, cpu_mmu_index(env, true));
2717     return full_lduw_code(env, addr, oi, 0);
2718 }
2719 
2720 static uint64_t full_ldl_code(CPUArchState *env, target_ulong addr,
2721                               MemOpIdx oi, uintptr_t retaddr)
2722 {
2723     return load_helper(env, addr, oi, retaddr, MO_TEUL, true, full_ldl_code);
2724 }
2725 
2726 uint32_t cpu_ldl_code(CPUArchState *env, abi_ptr addr)
2727 {
2728     MemOpIdx oi = make_memop_idx(MO_TEUL, cpu_mmu_index(env, true));
2729     return full_ldl_code(env, addr, oi, 0);
2730 }
2731 
2732 static uint64_t full_ldq_code(CPUArchState *env, target_ulong addr,
2733                               MemOpIdx oi, uintptr_t retaddr)
2734 {
2735     return load_helper(env, addr, oi, retaddr, MO_TEUQ, true, full_ldq_code);
2736 }
2737 
2738 uint64_t cpu_ldq_code(CPUArchState *env, abi_ptr addr)
2739 {
2740     MemOpIdx oi = make_memop_idx(MO_TEUQ, cpu_mmu_index(env, true));
2741     return full_ldq_code(env, addr, oi, 0);
2742 }
2743