Merge remote-tracking branch 'remotes/cohuck-gitlab/tags/s390x-20210121' into staging
[qemu.git] / softmmu / physmem.c
1 /*
2 * RAM allocation and memory access
3 *
4 * Copyright (c) 2003 Fabrice Bellard
5 *
6 * This library is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU Lesser General Public
8 * License as published by the Free Software Foundation; either
9 * version 2.1 of the License, or (at your option) any later version.
10 *
11 * This library is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 * Lesser General Public License for more details.
15 *
16 * You should have received a copy of the GNU Lesser General Public
17 * License along with this library; if not, see <http://www.gnu.org/licenses/>.
18 */
19
20 #include "qemu/osdep.h"
21 #include "qemu-common.h"
22 #include "qapi/error.h"
23
24 #include "qemu/cutils.h"
25 #include "qemu/cacheflush.h"
26 #include "cpu.h"
27 #include "exec/exec-all.h"
28 #include "exec/target_page.h"
29 #include "hw/qdev-core.h"
30 #include "hw/qdev-properties.h"
31 #include "hw/boards.h"
32 #include "hw/xen/xen.h"
33 #include "sysemu/kvm.h"
34 #include "sysemu/sysemu.h"
35 #include "sysemu/tcg.h"
36 #include "sysemu/qtest.h"
37 #include "qemu/timer.h"
38 #include "qemu/config-file.h"
39 #include "qemu/error-report.h"
40 #include "qemu/qemu-print.h"
41 #include "exec/memory.h"
42 #include "exec/ioport.h"
43 #include "sysemu/dma.h"
44 #include "sysemu/hostmem.h"
45 #include "sysemu/hw_accel.h"
46 #include "exec/address-spaces.h"
47 #include "sysemu/xen-mapcache.h"
48 #include "trace/trace-root.h"
49
50 #ifdef CONFIG_FALLOCATE_PUNCH_HOLE
51 #include <linux/falloc.h>
52 #endif
53
54 #include "qemu/rcu_queue.h"
55 #include "qemu/main-loop.h"
56 #include "exec/translate-all.h"
57 #include "sysemu/replay.h"
58
59 #include "exec/memory-internal.h"
60 #include "exec/ram_addr.h"
61 #include "exec/log.h"
62
63 #include "qemu/pmem.h"
64
65 #include "migration/vmstate.h"
66
67 #include "qemu/range.h"
68 #ifndef _WIN32
69 #include "qemu/mmap-alloc.h"
70 #endif
71
72 #include "monitor/monitor.h"
73
74 #ifdef CONFIG_LIBDAXCTL
75 #include <daxctl/libdaxctl.h>
76 #endif
77
78 //#define DEBUG_SUBPAGE
79
80 /* ram_list is read under rcu_read_lock()/rcu_read_unlock(). Writes
81 * are protected by the ramlist lock.
82 */
83 RAMList ram_list = { .blocks = QLIST_HEAD_INITIALIZER(ram_list.blocks) };
84
85 static MemoryRegion *system_memory;
86 static MemoryRegion *system_io;
87
88 AddressSpace address_space_io;
89 AddressSpace address_space_memory;
90
91 static MemoryRegion io_mem_unassigned;
92
93 typedef struct PhysPageEntry PhysPageEntry;
94
95 struct PhysPageEntry {
96 /* How many bits skip to next level (in units of L2_SIZE). 0 for a leaf. */
97 uint32_t skip : 6;
98 /* index into phys_sections (!skip) or phys_map_nodes (skip) */
99 uint32_t ptr : 26;
100 };
101
102 #define PHYS_MAP_NODE_NIL (((uint32_t)~0) >> 6)
103
104 /* Size of the L2 (and L3, etc) page tables. */
105 #define ADDR_SPACE_BITS 64
106
107 #define P_L2_BITS 9
108 #define P_L2_SIZE (1 << P_L2_BITS)
109
110 #define P_L2_LEVELS (((ADDR_SPACE_BITS - TARGET_PAGE_BITS - 1) / P_L2_BITS) + 1)
111
112 typedef PhysPageEntry Node[P_L2_SIZE];
113
114 typedef struct PhysPageMap {
115 struct rcu_head rcu;
116
117 unsigned sections_nb;
118 unsigned sections_nb_alloc;
119 unsigned nodes_nb;
120 unsigned nodes_nb_alloc;
121 Node *nodes;
122 MemoryRegionSection *sections;
123 } PhysPageMap;
124
125 struct AddressSpaceDispatch {
126 MemoryRegionSection *mru_section;
127 /* This is a multi-level map on the physical address space.
128 * The bottom level has pointers to MemoryRegionSections.
129 */
130 PhysPageEntry phys_map;
131 PhysPageMap map;
132 };
133
134 #define SUBPAGE_IDX(addr) ((addr) & ~TARGET_PAGE_MASK)
135 typedef struct subpage_t {
136 MemoryRegion iomem;
137 FlatView *fv;
138 hwaddr base;
139 uint16_t sub_section[];
140 } subpage_t;
141
142 #define PHYS_SECTION_UNASSIGNED 0
143
144 static void io_mem_init(void);
145 static void memory_map_init(void);
146 static void tcg_log_global_after_sync(MemoryListener *listener);
147 static void tcg_commit(MemoryListener *listener);
148
149 /**
150 * CPUAddressSpace: all the information a CPU needs about an AddressSpace
151 * @cpu: the CPU whose AddressSpace this is
152 * @as: the AddressSpace itself
153 * @memory_dispatch: its dispatch pointer (cached, RCU protected)
154 * @tcg_as_listener: listener for tracking changes to the AddressSpace
155 */
156 struct CPUAddressSpace {
157 CPUState *cpu;
158 AddressSpace *as;
159 struct AddressSpaceDispatch *memory_dispatch;
160 MemoryListener tcg_as_listener;
161 };
162
163 struct DirtyBitmapSnapshot {
164 ram_addr_t start;
165 ram_addr_t end;
166 unsigned long dirty[];
167 };
168
169 static void phys_map_node_reserve(PhysPageMap *map, unsigned nodes)
170 {
171 static unsigned alloc_hint = 16;
172 if (map->nodes_nb + nodes > map->nodes_nb_alloc) {
173 map->nodes_nb_alloc = MAX(alloc_hint, map->nodes_nb + nodes);
174 map->nodes = g_renew(Node, map->nodes, map->nodes_nb_alloc);
175 alloc_hint = map->nodes_nb_alloc;
176 }
177 }
178
179 static uint32_t phys_map_node_alloc(PhysPageMap *map, bool leaf)
180 {
181 unsigned i;
182 uint32_t ret;
183 PhysPageEntry e;
184 PhysPageEntry *p;
185
186 ret = map->nodes_nb++;
187 p = map->nodes[ret];
188 assert(ret != PHYS_MAP_NODE_NIL);
189 assert(ret != map->nodes_nb_alloc);
190
191 e.skip = leaf ? 0 : 1;
192 e.ptr = leaf ? PHYS_SECTION_UNASSIGNED : PHYS_MAP_NODE_NIL;
193 for (i = 0; i < P_L2_SIZE; ++i) {
194 memcpy(&p[i], &e, sizeof(e));
195 }
196 return ret;
197 }
198
199 static void phys_page_set_level(PhysPageMap *map, PhysPageEntry *lp,
200 hwaddr *index, uint64_t *nb, uint16_t leaf,
201 int level)
202 {
203 PhysPageEntry *p;
204 hwaddr step = (hwaddr)1 << (level * P_L2_BITS);
205
206 if (lp->skip && lp->ptr == PHYS_MAP_NODE_NIL) {
207 lp->ptr = phys_map_node_alloc(map, level == 0);
208 }
209 p = map->nodes[lp->ptr];
210 lp = &p[(*index >> (level * P_L2_BITS)) & (P_L2_SIZE - 1)];
211
212 while (*nb && lp < &p[P_L2_SIZE]) {
213 if ((*index & (step - 1)) == 0 && *nb >= step) {
214 lp->skip = 0;
215 lp->ptr = leaf;
216 *index += step;
217 *nb -= step;
218 } else {
219 phys_page_set_level(map, lp, index, nb, leaf, level - 1);
220 }
221 ++lp;
222 }
223 }
224
225 static void phys_page_set(AddressSpaceDispatch *d,
226 hwaddr index, uint64_t nb,
227 uint16_t leaf)
228 {
229 /* Wildly overreserve - it doesn't matter much. */
230 phys_map_node_reserve(&d->map, 3 * P_L2_LEVELS);
231
232 phys_page_set_level(&d->map, &d->phys_map, &index, &nb, leaf, P_L2_LEVELS - 1);
233 }
234
235 /* Compact a non leaf page entry. Simply detect that the entry has a single child,
236 * and update our entry so we can skip it and go directly to the destination.
237 */
238 static void phys_page_compact(PhysPageEntry *lp, Node *nodes)
239 {
240 unsigned valid_ptr = P_L2_SIZE;
241 int valid = 0;
242 PhysPageEntry *p;
243 int i;
244
245 if (lp->ptr == PHYS_MAP_NODE_NIL) {
246 return;
247 }
248
249 p = nodes[lp->ptr];
250 for (i = 0; i < P_L2_SIZE; i++) {
251 if (p[i].ptr == PHYS_MAP_NODE_NIL) {
252 continue;
253 }
254
255 valid_ptr = i;
256 valid++;
257 if (p[i].skip) {
258 phys_page_compact(&p[i], nodes);
259 }
260 }
261
262 /* We can only compress if there's only one child. */
263 if (valid != 1) {
264 return;
265 }
266
267 assert(valid_ptr < P_L2_SIZE);
268
269 /* Don't compress if it won't fit in the # of bits we have. */
270 if (P_L2_LEVELS >= (1 << 6) &&
271 lp->skip + p[valid_ptr].skip >= (1 << 6)) {
272 return;
273 }
274
275 lp->ptr = p[valid_ptr].ptr;
276 if (!p[valid_ptr].skip) {
277 /* If our only child is a leaf, make this a leaf. */
278 /* By design, we should have made this node a leaf to begin with so we
279 * should never reach here.
280 * But since it's so simple to handle this, let's do it just in case we
281 * change this rule.
282 */
283 lp->skip = 0;
284 } else {
285 lp->skip += p[valid_ptr].skip;
286 }
287 }
288
289 void address_space_dispatch_compact(AddressSpaceDispatch *d)
290 {
291 if (d->phys_map.skip) {
292 phys_page_compact(&d->phys_map, d->map.nodes);
293 }
294 }
295
296 static inline bool section_covers_addr(const MemoryRegionSection *section,
297 hwaddr addr)
298 {
299 /* Memory topology clips a memory region to [0, 2^64); size.hi > 0 means
300 * the section must cover the entire address space.
301 */
302 return int128_gethi(section->size) ||
303 range_covers_byte(section->offset_within_address_space,
304 int128_getlo(section->size), addr);
305 }
306
307 static MemoryRegionSection *phys_page_find(AddressSpaceDispatch *d, hwaddr addr)
308 {
309 PhysPageEntry lp = d->phys_map, *p;
310 Node *nodes = d->map.nodes;
311 MemoryRegionSection *sections = d->map.sections;
312 hwaddr index = addr >> TARGET_PAGE_BITS;
313 int i;
314
315 for (i = P_L2_LEVELS; lp.skip && (i -= lp.skip) >= 0;) {
316 if (lp.ptr == PHYS_MAP_NODE_NIL) {
317 return &sections[PHYS_SECTION_UNASSIGNED];
318 }
319 p = nodes[lp.ptr];
320 lp = p[(index >> (i * P_L2_BITS)) & (P_L2_SIZE - 1)];
321 }
322
323 if (section_covers_addr(&sections[lp.ptr], addr)) {
324 return &sections[lp.ptr];
325 } else {
326 return &sections[PHYS_SECTION_UNASSIGNED];
327 }
328 }
329
330 /* Called from RCU critical section */
331 static MemoryRegionSection *address_space_lookup_region(AddressSpaceDispatch *d,
332 hwaddr addr,
333 bool resolve_subpage)
334 {
335 MemoryRegionSection *section = qatomic_read(&d->mru_section);
336 subpage_t *subpage;
337
338 if (!section || section == &d->map.sections[PHYS_SECTION_UNASSIGNED] ||
339 !section_covers_addr(section, addr)) {
340 section = phys_page_find(d, addr);
341 qatomic_set(&d->mru_section, section);
342 }
343 if (resolve_subpage && section->mr->subpage) {
344 subpage = container_of(section->mr, subpage_t, iomem);
345 section = &d->map.sections[subpage->sub_section[SUBPAGE_IDX(addr)]];
346 }
347 return section;
348 }
349
350 /* Called from RCU critical section */
351 static MemoryRegionSection *
352 address_space_translate_internal(AddressSpaceDispatch *d, hwaddr addr, hwaddr *xlat,
353 hwaddr *plen, bool resolve_subpage)
354 {
355 MemoryRegionSection *section;
356 MemoryRegion *mr;
357 Int128 diff;
358
359 section = address_space_lookup_region(d, addr, resolve_subpage);
360 /* Compute offset within MemoryRegionSection */
361 addr -= section->offset_within_address_space;
362
363 /* Compute offset within MemoryRegion */
364 *xlat = addr + section->offset_within_region;
365
366 mr = section->mr;
367
368 /* MMIO registers can be expected to perform full-width accesses based only
369 * on their address, without considering adjacent registers that could
370 * decode to completely different MemoryRegions. When such registers
371 * exist (e.g. I/O ports 0xcf8 and 0xcf9 on most PC chipsets), MMIO
372 * regions overlap wildly. For this reason we cannot clamp the accesses
373 * here.
374 *
375 * If the length is small (as is the case for address_space_ldl/stl),
376 * everything works fine. If the incoming length is large, however,
377 * the caller really has to do the clamping through memory_access_size.
378 */
379 if (memory_region_is_ram(mr)) {
380 diff = int128_sub(section->size, int128_make64(addr));
381 *plen = int128_get64(int128_min(diff, int128_make64(*plen)));
382 }
383 return section;
384 }
385
386 /**
387 * address_space_translate_iommu - translate an address through an IOMMU
388 * memory region and then through the target address space.
389 *
390 * @iommu_mr: the IOMMU memory region that we start the translation from
391 * @addr: the address to be translated through the MMU
392 * @xlat: the translated address offset within the destination memory region.
393 * It cannot be %NULL.
394 * @plen_out: valid read/write length of the translated address. It
395 * cannot be %NULL.
396 * @page_mask_out: page mask for the translated address. This
397 * should only be meaningful for IOMMU translated
398 * addresses, since there may be huge pages that this bit
399 * would tell. It can be %NULL if we don't care about it.
400 * @is_write: whether the translation operation is for write
401 * @is_mmio: whether this can be MMIO, set true if it can
402 * @target_as: the address space targeted by the IOMMU
403 * @attrs: transaction attributes
404 *
405 * This function is called from RCU critical section. It is the common
406 * part of flatview_do_translate and address_space_translate_cached.
407 */
408 static MemoryRegionSection address_space_translate_iommu(IOMMUMemoryRegion *iommu_mr,
409 hwaddr *xlat,
410 hwaddr *plen_out,
411 hwaddr *page_mask_out,
412 bool is_write,
413 bool is_mmio,
414 AddressSpace **target_as,
415 MemTxAttrs attrs)
416 {
417 MemoryRegionSection *section;
418 hwaddr page_mask = (hwaddr)-1;
419
420 do {
421 hwaddr addr = *xlat;
422 IOMMUMemoryRegionClass *imrc = memory_region_get_iommu_class_nocheck(iommu_mr);
423 int iommu_idx = 0;
424 IOMMUTLBEntry iotlb;
425
426 if (imrc->attrs_to_index) {
427 iommu_idx = imrc->attrs_to_index(iommu_mr, attrs);
428 }
429
430 iotlb = imrc->translate(iommu_mr, addr, is_write ?
431 IOMMU_WO : IOMMU_RO, iommu_idx);
432
433 if (!(iotlb.perm & (1 << is_write))) {
434 goto unassigned;
435 }
436
437 addr = ((iotlb.translated_addr & ~iotlb.addr_mask)
438 | (addr & iotlb.addr_mask));
439 page_mask &= iotlb.addr_mask;
440 *plen_out = MIN(*plen_out, (addr | iotlb.addr_mask) - addr + 1);
441 *target_as = iotlb.target_as;
442
443 section = address_space_translate_internal(
444 address_space_to_dispatch(iotlb.target_as), addr, xlat,
445 plen_out, is_mmio);
446
447 iommu_mr = memory_region_get_iommu(section->mr);
448 } while (unlikely(iommu_mr));
449
450 if (page_mask_out) {
451 *page_mask_out = page_mask;
452 }
453 return *section;
454
455 unassigned:
456 return (MemoryRegionSection) { .mr = &io_mem_unassigned };
457 }
458
459 /**
460 * flatview_do_translate - translate an address in FlatView
461 *
462 * @fv: the flat view that we want to translate on
463 * @addr: the address to be translated in above address space
464 * @xlat: the translated address offset within memory region. It
465 * cannot be @NULL.
466 * @plen_out: valid read/write length of the translated address. It
467 * can be @NULL when we don't care about it.
468 * @page_mask_out: page mask for the translated address. This
469 * should only be meaningful for IOMMU translated
470 * addresses, since there may be huge pages that this bit
471 * would tell. It can be @NULL if we don't care about it.
472 * @is_write: whether the translation operation is for write
473 * @is_mmio: whether this can be MMIO, set true if it can
474 * @target_as: the address space targeted by the IOMMU
475 * @attrs: memory transaction attributes
476 *
477 * This function is called from RCU critical section
478 */
479 static MemoryRegionSection flatview_do_translate(FlatView *fv,
480 hwaddr addr,
481 hwaddr *xlat,
482 hwaddr *plen_out,
483 hwaddr *page_mask_out,
484 bool is_write,
485 bool is_mmio,
486 AddressSpace **target_as,
487 MemTxAttrs attrs)
488 {
489 MemoryRegionSection *section;
490 IOMMUMemoryRegion *iommu_mr;
491 hwaddr plen = (hwaddr)(-1);
492
493 if (!plen_out) {
494 plen_out = &plen;
495 }
496
497 section = address_space_translate_internal(
498 flatview_to_dispatch(fv), addr, xlat,
499 plen_out, is_mmio);
500
501 iommu_mr = memory_region_get_iommu(section->mr);
502 if (unlikely(iommu_mr)) {
503 return address_space_translate_iommu(iommu_mr, xlat,
504 plen_out, page_mask_out,
505 is_write, is_mmio,
506 target_as, attrs);
507 }
508 if (page_mask_out) {
509 /* Not behind an IOMMU, use default page size. */
510 *page_mask_out = ~TARGET_PAGE_MASK;
511 }
512
513 return *section;
514 }
515
516 /* Called from RCU critical section */
517 IOMMUTLBEntry address_space_get_iotlb_entry(AddressSpace *as, hwaddr addr,
518 bool is_write, MemTxAttrs attrs)
519 {
520 MemoryRegionSection section;
521 hwaddr xlat, page_mask;
522
523 /*
524 * This can never be MMIO, and we don't really care about plen,
525 * but page mask.
526 */
527 section = flatview_do_translate(address_space_to_flatview(as), addr, &xlat,
528 NULL, &page_mask, is_write, false, &as,
529 attrs);
530
531 /* Illegal translation */
532 if (section.mr == &io_mem_unassigned) {
533 goto iotlb_fail;
534 }
535
536 /* Convert memory region offset into address space offset */
537 xlat += section.offset_within_address_space -
538 section.offset_within_region;
539
540 return (IOMMUTLBEntry) {
541 .target_as = as,
542 .iova = addr & ~page_mask,
543 .translated_addr = xlat & ~page_mask,
544 .addr_mask = page_mask,
545 /* IOTLBs are for DMAs, and DMA only allows on RAMs. */
546 .perm = IOMMU_RW,
547 };
548
549 iotlb_fail:
550 return (IOMMUTLBEntry) {0};
551 }
552
553 /* Called from RCU critical section */
554 MemoryRegion *flatview_translate(FlatView *fv, hwaddr addr, hwaddr *xlat,
555 hwaddr *plen, bool is_write,
556 MemTxAttrs attrs)
557 {
558 MemoryRegion *mr;
559 MemoryRegionSection section;
560 AddressSpace *as = NULL;
561
562 /* This can be MMIO, so setup MMIO bit. */
563 section = flatview_do_translate(fv, addr, xlat, plen, NULL,
564 is_write, true, &as, attrs);
565 mr = section.mr;
566
567 if (xen_enabled() && memory_access_is_direct(mr, is_write)) {
568 hwaddr page = ((addr & TARGET_PAGE_MASK) + TARGET_PAGE_SIZE) - addr;
569 *plen = MIN(page, *plen);
570 }
571
572 return mr;
573 }
574
575 typedef struct TCGIOMMUNotifier {
576 IOMMUNotifier n;
577 MemoryRegion *mr;
578 CPUState *cpu;
579 int iommu_idx;
580 bool active;
581 } TCGIOMMUNotifier;
582
583 static void tcg_iommu_unmap_notify(IOMMUNotifier *n, IOMMUTLBEntry *iotlb)
584 {
585 TCGIOMMUNotifier *notifier = container_of(n, TCGIOMMUNotifier, n);
586
587 if (!notifier->active) {
588 return;
589 }
590 tlb_flush(notifier->cpu);
591 notifier->active = false;
592 /* We leave the notifier struct on the list to avoid reallocating it later.
593 * Generally the number of IOMMUs a CPU deals with will be small.
594 * In any case we can't unregister the iommu notifier from a notify
595 * callback.
596 */
597 }
598
599 static void tcg_register_iommu_notifier(CPUState *cpu,
600 IOMMUMemoryRegion *iommu_mr,
601 int iommu_idx)
602 {
603 /* Make sure this CPU has an IOMMU notifier registered for this
604 * IOMMU/IOMMU index combination, so that we can flush its TLB
605 * when the IOMMU tells us the mappings we've cached have changed.
606 */
607 MemoryRegion *mr = MEMORY_REGION(iommu_mr);
608 TCGIOMMUNotifier *notifier;
609 int i;
610
611 for (i = 0; i < cpu->iommu_notifiers->len; i++) {
612 notifier = g_array_index(cpu->iommu_notifiers, TCGIOMMUNotifier *, i);
613 if (notifier->mr == mr && notifier->iommu_idx == iommu_idx) {
614 break;
615 }
616 }
617 if (i == cpu->iommu_notifiers->len) {
618 /* Not found, add a new entry at the end of the array */
619 cpu->iommu_notifiers = g_array_set_size(cpu->iommu_notifiers, i + 1);
620 notifier = g_new0(TCGIOMMUNotifier, 1);
621 g_array_index(cpu->iommu_notifiers, TCGIOMMUNotifier *, i) = notifier;
622
623 notifier->mr = mr;
624 notifier->iommu_idx = iommu_idx;
625 notifier->cpu = cpu;
626 /* Rather than trying to register interest in the specific part
627 * of the iommu's address space that we've accessed and then
628 * expand it later as subsequent accesses touch more of it, we
629 * just register interest in the whole thing, on the assumption
630 * that iommu reconfiguration will be rare.
631 */
632 iommu_notifier_init(&notifier->n,
633 tcg_iommu_unmap_notify,
634 IOMMU_NOTIFIER_UNMAP,
635 0,
636 HWADDR_MAX,
637 iommu_idx);
638 memory_region_register_iommu_notifier(notifier->mr, &notifier->n,
639 &error_fatal);
640 }
641
642 if (!notifier->active) {
643 notifier->active = true;
644 }
645 }
646
647 void tcg_iommu_free_notifier_list(CPUState *cpu)
648 {
649 /* Destroy the CPU's notifier list */
650 int i;
651 TCGIOMMUNotifier *notifier;
652
653 for (i = 0; i < cpu->iommu_notifiers->len; i++) {
654 notifier = g_array_index(cpu->iommu_notifiers, TCGIOMMUNotifier *, i);
655 memory_region_unregister_iommu_notifier(notifier->mr, &notifier->n);
656 g_free(notifier);
657 }
658 g_array_free(cpu->iommu_notifiers, true);
659 }
660
661 void tcg_iommu_init_notifier_list(CPUState *cpu)
662 {
663 cpu->iommu_notifiers = g_array_new(false, true, sizeof(TCGIOMMUNotifier *));
664 }
665
666 /* Called from RCU critical section */
667 MemoryRegionSection *
668 address_space_translate_for_iotlb(CPUState *cpu, int asidx, hwaddr addr,
669 hwaddr *xlat, hwaddr *plen,
670 MemTxAttrs attrs, int *prot)
671 {
672 MemoryRegionSection *section;
673 IOMMUMemoryRegion *iommu_mr;
674 IOMMUMemoryRegionClass *imrc;
675 IOMMUTLBEntry iotlb;
676 int iommu_idx;
677 AddressSpaceDispatch *d =
678 qatomic_rcu_read(&cpu->cpu_ases[asidx].memory_dispatch);
679
680 for (;;) {
681 section = address_space_translate_internal(d, addr, &addr, plen, false);
682
683 iommu_mr = memory_region_get_iommu(section->mr);
684 if (!iommu_mr) {
685 break;
686 }
687
688 imrc = memory_region_get_iommu_class_nocheck(iommu_mr);
689
690 iommu_idx = imrc->attrs_to_index(iommu_mr, attrs);
691 tcg_register_iommu_notifier(cpu, iommu_mr, iommu_idx);
692 /* We need all the permissions, so pass IOMMU_NONE so the IOMMU
693 * doesn't short-cut its translation table walk.
694 */
695 iotlb = imrc->translate(iommu_mr, addr, IOMMU_NONE, iommu_idx);
696 addr = ((iotlb.translated_addr & ~iotlb.addr_mask)
697 | (addr & iotlb.addr_mask));
698 /* Update the caller's prot bits to remove permissions the IOMMU
699 * is giving us a failure response for. If we get down to no
700 * permissions left at all we can give up now.
701 */
702 if (!(iotlb.perm & IOMMU_RO)) {
703 *prot &= ~(PAGE_READ | PAGE_EXEC);
704 }
705 if (!(iotlb.perm & IOMMU_WO)) {
706 *prot &= ~PAGE_WRITE;
707 }
708
709 if (!*prot) {
710 goto translate_fail;
711 }
712
713 d = flatview_to_dispatch(address_space_to_flatview(iotlb.target_as));
714 }
715
716 assert(!memory_region_is_iommu(section->mr));
717 *xlat = addr;
718 return section;
719
720 translate_fail:
721 return &d->map.sections[PHYS_SECTION_UNASSIGNED];
722 }
723
724 void cpu_address_space_init(CPUState *cpu, int asidx,
725 const char *prefix, MemoryRegion *mr)
726 {
727 CPUAddressSpace *newas;
728 AddressSpace *as = g_new0(AddressSpace, 1);
729 char *as_name;
730
731 assert(mr);
732 as_name = g_strdup_printf("%s-%d", prefix, cpu->cpu_index);
733 address_space_init(as, mr, as_name);
734 g_free(as_name);
735
736 /* Target code should have set num_ases before calling us */
737 assert(asidx < cpu->num_ases);
738
739 if (asidx == 0) {
740 /* address space 0 gets the convenience alias */
741 cpu->as = as;
742 }
743
744 /* KVM cannot currently support multiple address spaces. */
745 assert(asidx == 0 || !kvm_enabled());
746
747 if (!cpu->cpu_ases) {
748 cpu->cpu_ases = g_new0(CPUAddressSpace, cpu->num_ases);
749 }
750
751 newas = &cpu->cpu_ases[asidx];
752 newas->cpu = cpu;
753 newas->as = as;
754 if (tcg_enabled()) {
755 newas->tcg_as_listener.log_global_after_sync = tcg_log_global_after_sync;
756 newas->tcg_as_listener.commit = tcg_commit;
757 memory_listener_register(&newas->tcg_as_listener, as);
758 }
759 }
760
761 AddressSpace *cpu_get_address_space(CPUState *cpu, int asidx)
762 {
763 /* Return the AddressSpace corresponding to the specified index */
764 return cpu->cpu_ases[asidx].as;
765 }
766
767 /* Add a watchpoint. */
768 int cpu_watchpoint_insert(CPUState *cpu, vaddr addr, vaddr len,
769 int flags, CPUWatchpoint **watchpoint)
770 {
771 CPUWatchpoint *wp;
772 vaddr in_page;
773
774 /* forbid ranges which are empty or run off the end of the address space */
775 if (len == 0 || (addr + len - 1) < addr) {
776 error_report("tried to set invalid watchpoint at %"
777 VADDR_PRIx ", len=%" VADDR_PRIu, addr, len);
778 return -EINVAL;
779 }
780 wp = g_malloc(sizeof(*wp));
781
782 wp->vaddr = addr;
783 wp->len = len;
784 wp->flags = flags;
785
786 /* keep all GDB-injected watchpoints in front */
787 if (flags & BP_GDB) {
788 QTAILQ_INSERT_HEAD(&cpu->watchpoints, wp, entry);
789 } else {
790 QTAILQ_INSERT_TAIL(&cpu->watchpoints, wp, entry);
791 }
792
793 in_page = -(addr | TARGET_PAGE_MASK);
794 if (len <= in_page) {
795 tlb_flush_page(cpu, addr);
796 } else {
797 tlb_flush(cpu);
798 }
799
800 if (watchpoint)
801 *watchpoint = wp;
802 return 0;
803 }
804
805 /* Remove a specific watchpoint. */
806 int cpu_watchpoint_remove(CPUState *cpu, vaddr addr, vaddr len,
807 int flags)
808 {
809 CPUWatchpoint *wp;
810
811 QTAILQ_FOREACH(wp, &cpu->watchpoints, entry) {
812 if (addr == wp->vaddr && len == wp->len
813 && flags == (wp->flags & ~BP_WATCHPOINT_HIT)) {
814 cpu_watchpoint_remove_by_ref(cpu, wp);
815 return 0;
816 }
817 }
818 return -ENOENT;
819 }
820
821 /* Remove a specific watchpoint by reference. */
822 void cpu_watchpoint_remove_by_ref(CPUState *cpu, CPUWatchpoint *watchpoint)
823 {
824 QTAILQ_REMOVE(&cpu->watchpoints, watchpoint, entry);
825
826 tlb_flush_page(cpu, watchpoint->vaddr);
827
828 g_free(watchpoint);
829 }
830
831 /* Remove all matching watchpoints. */
832 void cpu_watchpoint_remove_all(CPUState *cpu, int mask)
833 {
834 CPUWatchpoint *wp, *next;
835
836 QTAILQ_FOREACH_SAFE(wp, &cpu->watchpoints, entry, next) {
837 if (wp->flags & mask) {
838 cpu_watchpoint_remove_by_ref(cpu, wp);
839 }
840 }
841 }
842
843 /* Return true if this watchpoint address matches the specified
844 * access (ie the address range covered by the watchpoint overlaps
845 * partially or completely with the address range covered by the
846 * access).
847 */
848 static inline bool watchpoint_address_matches(CPUWatchpoint *wp,
849 vaddr addr, vaddr len)
850 {
851 /* We know the lengths are non-zero, but a little caution is
852 * required to avoid errors in the case where the range ends
853 * exactly at the top of the address space and so addr + len
854 * wraps round to zero.
855 */
856 vaddr wpend = wp->vaddr + wp->len - 1;
857 vaddr addrend = addr + len - 1;
858
859 return !(addr > wpend || wp->vaddr > addrend);
860 }
861
862 /* Return flags for watchpoints that match addr + prot. */
863 int cpu_watchpoint_address_matches(CPUState *cpu, vaddr addr, vaddr len)
864 {
865 CPUWatchpoint *wp;
866 int ret = 0;
867
868 QTAILQ_FOREACH(wp, &cpu->watchpoints, entry) {
869 if (watchpoint_address_matches(wp, addr, len)) {
870 ret |= wp->flags;
871 }
872 }
873 return ret;
874 }
875
876 /* Called from RCU critical section */
877 static RAMBlock *qemu_get_ram_block(ram_addr_t addr)
878 {
879 RAMBlock *block;
880
881 block = qatomic_rcu_read(&ram_list.mru_block);
882 if (block && addr - block->offset < block->max_length) {
883 return block;
884 }
885 RAMBLOCK_FOREACH(block) {
886 if (addr - block->offset < block->max_length) {
887 goto found;
888 }
889 }
890
891 fprintf(stderr, "Bad ram offset %" PRIx64 "\n", (uint64_t)addr);
892 abort();
893
894 found:
895 /* It is safe to write mru_block outside the iothread lock. This
896 * is what happens:
897 *
898 * mru_block = xxx
899 * rcu_read_unlock()
900 * xxx removed from list
901 * rcu_read_lock()
902 * read mru_block
903 * mru_block = NULL;
904 * call_rcu(reclaim_ramblock, xxx);
905 * rcu_read_unlock()
906 *
907 * qatomic_rcu_set is not needed here. The block was already published
908 * when it was placed into the list. Here we're just making an extra
909 * copy of the pointer.
910 */
911 ram_list.mru_block = block;
912 return block;
913 }
914
915 static void tlb_reset_dirty_range_all(ram_addr_t start, ram_addr_t length)
916 {
917 CPUState *cpu;
918 ram_addr_t start1;
919 RAMBlock *block;
920 ram_addr_t end;
921
922 assert(tcg_enabled());
923 end = TARGET_PAGE_ALIGN(start + length);
924 start &= TARGET_PAGE_MASK;
925
926 RCU_READ_LOCK_GUARD();
927 block = qemu_get_ram_block(start);
928 assert(block == qemu_get_ram_block(end - 1));
929 start1 = (uintptr_t)ramblock_ptr(block, start - block->offset);
930 CPU_FOREACH(cpu) {
931 tlb_reset_dirty(cpu, start1, length);
932 }
933 }
934
935 /* Note: start and end must be within the same ram block. */
936 bool cpu_physical_memory_test_and_clear_dirty(ram_addr_t start,
937 ram_addr_t length,
938 unsigned client)
939 {
940 DirtyMemoryBlocks *blocks;
941 unsigned long end, page, start_page;
942 bool dirty = false;
943 RAMBlock *ramblock;
944 uint64_t mr_offset, mr_size;
945
946 if (length == 0) {
947 return false;
948 }
949
950 end = TARGET_PAGE_ALIGN(start + length) >> TARGET_PAGE_BITS;
951 start_page = start >> TARGET_PAGE_BITS;
952 page = start_page;
953
954 WITH_RCU_READ_LOCK_GUARD() {
955 blocks = qatomic_rcu_read(&ram_list.dirty_memory[client]);
956 ramblock = qemu_get_ram_block(start);
957 /* Range sanity check on the ramblock */
958 assert(start >= ramblock->offset &&
959 start + length <= ramblock->offset + ramblock->used_length);
960
961 while (page < end) {
962 unsigned long idx = page / DIRTY_MEMORY_BLOCK_SIZE;
963 unsigned long offset = page % DIRTY_MEMORY_BLOCK_SIZE;
964 unsigned long num = MIN(end - page,
965 DIRTY_MEMORY_BLOCK_SIZE - offset);
966
967 dirty |= bitmap_test_and_clear_atomic(blocks->blocks[idx],
968 offset, num);
969 page += num;
970 }
971
972 mr_offset = (ram_addr_t)(start_page << TARGET_PAGE_BITS) - ramblock->offset;
973 mr_size = (end - start_page) << TARGET_PAGE_BITS;
974 memory_region_clear_dirty_bitmap(ramblock->mr, mr_offset, mr_size);
975 }
976
977 if (dirty && tcg_enabled()) {
978 tlb_reset_dirty_range_all(start, length);
979 }
980
981 return dirty;
982 }
983
984 DirtyBitmapSnapshot *cpu_physical_memory_snapshot_and_clear_dirty
985 (MemoryRegion *mr, hwaddr offset, hwaddr length, unsigned client)
986 {
987 DirtyMemoryBlocks *blocks;
988 ram_addr_t start = memory_region_get_ram_addr(mr) + offset;
989 unsigned long align = 1UL << (TARGET_PAGE_BITS + BITS_PER_LEVEL);
990 ram_addr_t first = QEMU_ALIGN_DOWN(start, align);
991 ram_addr_t last = QEMU_ALIGN_UP(start + length, align);
992 DirtyBitmapSnapshot *snap;
993 unsigned long page, end, dest;
994
995 snap = g_malloc0(sizeof(*snap) +
996 ((last - first) >> (TARGET_PAGE_BITS + 3)));
997 snap->start = first;
998 snap->end = last;
999
1000 page = first >> TARGET_PAGE_BITS;
1001 end = last >> TARGET_PAGE_BITS;
1002 dest = 0;
1003
1004 WITH_RCU_READ_LOCK_GUARD() {
1005 blocks = qatomic_rcu_read(&ram_list.dirty_memory[client]);
1006
1007 while (page < end) {
1008 unsigned long idx = page / DIRTY_MEMORY_BLOCK_SIZE;
1009 unsigned long offset = page % DIRTY_MEMORY_BLOCK_SIZE;
1010 unsigned long num = MIN(end - page,
1011 DIRTY_MEMORY_BLOCK_SIZE - offset);
1012
1013 assert(QEMU_IS_ALIGNED(offset, (1 << BITS_PER_LEVEL)));
1014 assert(QEMU_IS_ALIGNED(num, (1 << BITS_PER_LEVEL)));
1015 offset >>= BITS_PER_LEVEL;
1016
1017 bitmap_copy_and_clear_atomic(snap->dirty + dest,
1018 blocks->blocks[idx] + offset,
1019 num);
1020 page += num;
1021 dest += num >> BITS_PER_LEVEL;
1022 }
1023 }
1024
1025 if (tcg_enabled()) {
1026 tlb_reset_dirty_range_all(start, length);
1027 }
1028
1029 memory_region_clear_dirty_bitmap(mr, offset, length);
1030
1031 return snap;
1032 }
1033
1034 bool cpu_physical_memory_snapshot_get_dirty(DirtyBitmapSnapshot *snap,
1035 ram_addr_t start,
1036 ram_addr_t length)
1037 {
1038 unsigned long page, end;
1039
1040 assert(start >= snap->start);
1041 assert(start + length <= snap->end);
1042
1043 end = TARGET_PAGE_ALIGN(start + length - snap->start) >> TARGET_PAGE_BITS;
1044 page = (start - snap->start) >> TARGET_PAGE_BITS;
1045
1046 while (page < end) {
1047 if (test_bit(page, snap->dirty)) {
1048 return true;
1049 }
1050 page++;
1051 }
1052 return false;
1053 }
1054
1055 /* Called from RCU critical section */
1056 hwaddr memory_region_section_get_iotlb(CPUState *cpu,
1057 MemoryRegionSection *section)
1058 {
1059 AddressSpaceDispatch *d = flatview_to_dispatch(section->fv);
1060 return section - d->map.sections;
1061 }
1062
1063 static int subpage_register(subpage_t *mmio, uint32_t start, uint32_t end,
1064 uint16_t section);
1065 static subpage_t *subpage_init(FlatView *fv, hwaddr base);
1066
1067 static void *(*phys_mem_alloc)(size_t size, uint64_t *align, bool shared) =
1068 qemu_anon_ram_alloc;
1069
1070 /*
1071 * Set a custom physical guest memory alloator.
1072 * Accelerators with unusual needs may need this. Hopefully, we can
1073 * get rid of it eventually.
1074 */
1075 void phys_mem_set_alloc(void *(*alloc)(size_t, uint64_t *align, bool shared))
1076 {
1077 phys_mem_alloc = alloc;
1078 }
1079
1080 static uint16_t phys_section_add(PhysPageMap *map,
1081 MemoryRegionSection *section)
1082 {
1083 /* The physical section number is ORed with a page-aligned
1084 * pointer to produce the iotlb entries. Thus it should
1085 * never overflow into the page-aligned value.
1086 */
1087 assert(map->sections_nb < TARGET_PAGE_SIZE);
1088
1089 if (map->sections_nb == map->sections_nb_alloc) {
1090 map->sections_nb_alloc = MAX(map->sections_nb_alloc * 2, 16);
1091 map->sections = g_renew(MemoryRegionSection, map->sections,
1092 map->sections_nb_alloc);
1093 }
1094 map->sections[map->sections_nb] = *section;
1095 memory_region_ref(section->mr);
1096 return map->sections_nb++;
1097 }
1098
1099 static void phys_section_destroy(MemoryRegion *mr)
1100 {
1101 bool have_sub_page = mr->subpage;
1102
1103 memory_region_unref(mr);
1104
1105 if (have_sub_page) {
1106 subpage_t *subpage = container_of(mr, subpage_t, iomem);
1107 object_unref(OBJECT(&subpage->iomem));
1108 g_free(subpage);
1109 }
1110 }
1111
1112 static void phys_sections_free(PhysPageMap *map)
1113 {
1114 while (map->sections_nb > 0) {
1115 MemoryRegionSection *section = &map->sections[--map->sections_nb];
1116 phys_section_destroy(section->mr);
1117 }
1118 g_free(map->sections);
1119 g_free(map->nodes);
1120 }
1121
1122 static void register_subpage(FlatView *fv, MemoryRegionSection *section)
1123 {
1124 AddressSpaceDispatch *d = flatview_to_dispatch(fv);
1125 subpage_t *subpage;
1126 hwaddr base = section->offset_within_address_space
1127 & TARGET_PAGE_MASK;
1128 MemoryRegionSection *existing = phys_page_find(d, base);
1129 MemoryRegionSection subsection = {
1130 .offset_within_address_space = base,
1131 .size = int128_make64(TARGET_PAGE_SIZE),
1132 };
1133 hwaddr start, end;
1134
1135 assert(existing->mr->subpage || existing->mr == &io_mem_unassigned);
1136
1137 if (!(existing->mr->subpage)) {
1138 subpage = subpage_init(fv, base);
1139 subsection.fv = fv;
1140 subsection.mr = &subpage->iomem;
1141 phys_page_set(d, base >> TARGET_PAGE_BITS, 1,
1142 phys_section_add(&d->map, &subsection));
1143 } else {
1144 subpage = container_of(existing->mr, subpage_t, iomem);
1145 }
1146 start = section->offset_within_address_space & ~TARGET_PAGE_MASK;
1147 end = start + int128_get64(section->size) - 1;
1148 subpage_register(subpage, start, end,
1149 phys_section_add(&d->map, section));
1150 }
1151
1152
1153 static void register_multipage(FlatView *fv,
1154 MemoryRegionSection *section)
1155 {
1156 AddressSpaceDispatch *d = flatview_to_dispatch(fv);
1157 hwaddr start_addr = section->offset_within_address_space;
1158 uint16_t section_index = phys_section_add(&d->map, section);
1159 uint64_t num_pages = int128_get64(int128_rshift(section->size,
1160 TARGET_PAGE_BITS));
1161
1162 assert(num_pages);
1163 phys_page_set(d, start_addr >> TARGET_PAGE_BITS, num_pages, section_index);
1164 }
1165
1166 /*
1167 * The range in *section* may look like this:
1168 *
1169 * |s|PPPPPPP|s|
1170 *
1171 * where s stands for subpage and P for page.
1172 */
1173 void flatview_add_to_dispatch(FlatView *fv, MemoryRegionSection *section)
1174 {
1175 MemoryRegionSection remain = *section;
1176 Int128 page_size = int128_make64(TARGET_PAGE_SIZE);
1177
1178 /* register first subpage */
1179 if (remain.offset_within_address_space & ~TARGET_PAGE_MASK) {
1180 uint64_t left = TARGET_PAGE_ALIGN(remain.offset_within_address_space)
1181 - remain.offset_within_address_space;
1182
1183 MemoryRegionSection now = remain;
1184 now.size = int128_min(int128_make64(left), now.size);
1185 register_subpage(fv, &now);
1186 if (int128_eq(remain.size, now.size)) {
1187 return;
1188 }
1189 remain.size = int128_sub(remain.size, now.size);
1190 remain.offset_within_address_space += int128_get64(now.size);
1191 remain.offset_within_region += int128_get64(now.size);
1192 }
1193
1194 /* register whole pages */
1195 if (int128_ge(remain.size, page_size)) {
1196 MemoryRegionSection now = remain;
1197 now.size = int128_and(now.size, int128_neg(page_size));
1198 register_multipage(fv, &now);
1199 if (int128_eq(remain.size, now.size)) {
1200 return;
1201 }
1202 remain.size = int128_sub(remain.size, now.size);
1203 remain.offset_within_address_space += int128_get64(now.size);
1204 remain.offset_within_region += int128_get64(now.size);
1205 }
1206
1207 /* register last subpage */
1208 register_subpage(fv, &remain);
1209 }
1210
1211 void qemu_flush_coalesced_mmio_buffer(void)
1212 {
1213 if (kvm_enabled())
1214 kvm_flush_coalesced_mmio_buffer();
1215 }
1216
1217 void qemu_mutex_lock_ramlist(void)
1218 {
1219 qemu_mutex_lock(&ram_list.mutex);
1220 }
1221
1222 void qemu_mutex_unlock_ramlist(void)
1223 {
1224 qemu_mutex_unlock(&ram_list.mutex);
1225 }
1226
1227 void ram_block_dump(Monitor *mon)
1228 {
1229 RAMBlock *block;
1230 char *psize;
1231
1232 RCU_READ_LOCK_GUARD();
1233 monitor_printf(mon, "%24s %8s %18s %18s %18s\n",
1234 "Block Name", "PSize", "Offset", "Used", "Total");
1235 RAMBLOCK_FOREACH(block) {
1236 psize = size_to_str(block->page_size);
1237 monitor_printf(mon, "%24s %8s 0x%016" PRIx64 " 0x%016" PRIx64
1238 " 0x%016" PRIx64 "\n", block->idstr, psize,
1239 (uint64_t)block->offset,
1240 (uint64_t)block->used_length,
1241 (uint64_t)block->max_length);
1242 g_free(psize);
1243 }
1244 }
1245
1246 #ifdef __linux__
1247 /*
1248 * FIXME TOCTTOU: this iterates over memory backends' mem-path, which
1249 * may or may not name the same files / on the same filesystem now as
1250 * when we actually open and map them. Iterate over the file
1251 * descriptors instead, and use qemu_fd_getpagesize().
1252 */
1253 static int find_min_backend_pagesize(Object *obj, void *opaque)
1254 {
1255 long *hpsize_min = opaque;
1256
1257 if (object_dynamic_cast(obj, TYPE_MEMORY_BACKEND)) {
1258 HostMemoryBackend *backend = MEMORY_BACKEND(obj);
1259 long hpsize = host_memory_backend_pagesize(backend);
1260
1261 if (host_memory_backend_is_mapped(backend) && (hpsize < *hpsize_min)) {
1262 *hpsize_min = hpsize;
1263 }
1264 }
1265
1266 return 0;
1267 }
1268
1269 static int find_max_backend_pagesize(Object *obj, void *opaque)
1270 {
1271 long *hpsize_max = opaque;
1272
1273 if (object_dynamic_cast(obj, TYPE_MEMORY_BACKEND)) {
1274 HostMemoryBackend *backend = MEMORY_BACKEND(obj);
1275 long hpsize = host_memory_backend_pagesize(backend);
1276
1277 if (host_memory_backend_is_mapped(backend) && (hpsize > *hpsize_max)) {
1278 *hpsize_max = hpsize;
1279 }
1280 }
1281
1282 return 0;
1283 }
1284
1285 /*
1286 * TODO: We assume right now that all mapped host memory backends are
1287 * used as RAM, however some might be used for different purposes.
1288 */
1289 long qemu_minrampagesize(void)
1290 {
1291 long hpsize = LONG_MAX;
1292 Object *memdev_root = object_resolve_path("/objects", NULL);
1293
1294 object_child_foreach(memdev_root, find_min_backend_pagesize, &hpsize);
1295 return hpsize;
1296 }
1297
1298 long qemu_maxrampagesize(void)
1299 {
1300 long pagesize = 0;
1301 Object *memdev_root = object_resolve_path("/objects", NULL);
1302
1303 object_child_foreach(memdev_root, find_max_backend_pagesize, &pagesize);
1304 return pagesize;
1305 }
1306 #else
1307 long qemu_minrampagesize(void)
1308 {
1309 return qemu_real_host_page_size;
1310 }
1311 long qemu_maxrampagesize(void)
1312 {
1313 return qemu_real_host_page_size;
1314 }
1315 #endif
1316
1317 #ifdef CONFIG_POSIX
1318 static int64_t get_file_size(int fd)
1319 {
1320 int64_t size;
1321 #if defined(__linux__)
1322 struct stat st;
1323
1324 if (fstat(fd, &st) < 0) {
1325 return -errno;
1326 }
1327
1328 /* Special handling for devdax character devices */
1329 if (S_ISCHR(st.st_mode)) {
1330 g_autofree char *subsystem_path = NULL;
1331 g_autofree char *subsystem = NULL;
1332
1333 subsystem_path = g_strdup_printf("/sys/dev/char/%d:%d/subsystem",
1334 major(st.st_rdev), minor(st.st_rdev));
1335 subsystem = g_file_read_link(subsystem_path, NULL);
1336
1337 if (subsystem && g_str_has_suffix(subsystem, "/dax")) {
1338 g_autofree char *size_path = NULL;
1339 g_autofree char *size_str = NULL;
1340
1341 size_path = g_strdup_printf("/sys/dev/char/%d:%d/size",
1342 major(st.st_rdev), minor(st.st_rdev));
1343
1344 if (g_file_get_contents(size_path, &size_str, NULL, NULL)) {
1345 return g_ascii_strtoll(size_str, NULL, 0);
1346 }
1347 }
1348 }
1349 #endif /* defined(__linux__) */
1350
1351 /* st.st_size may be zero for special files yet lseek(2) works */
1352 size = lseek(fd, 0, SEEK_END);
1353 if (size < 0) {
1354 return -errno;
1355 }
1356 return size;
1357 }
1358
1359 static int64_t get_file_align(int fd)
1360 {
1361 int64_t align = -1;
1362 #if defined(__linux__) && defined(CONFIG_LIBDAXCTL)
1363 struct stat st;
1364
1365 if (fstat(fd, &st) < 0) {
1366 return -errno;
1367 }
1368
1369 /* Special handling for devdax character devices */
1370 if (S_ISCHR(st.st_mode)) {
1371 g_autofree char *path = NULL;
1372 g_autofree char *rpath = NULL;
1373 struct daxctl_ctx *ctx;
1374 struct daxctl_region *region;
1375 int rc = 0;
1376
1377 path = g_strdup_printf("/sys/dev/char/%d:%d",
1378 major(st.st_rdev), minor(st.st_rdev));
1379 rpath = realpath(path, NULL);
1380
1381 rc = daxctl_new(&ctx);
1382 if (rc) {
1383 return -1;
1384 }
1385
1386 daxctl_region_foreach(ctx, region) {
1387 if (strstr(rpath, daxctl_region_get_path(region))) {
1388 align = daxctl_region_get_align(region);
1389 break;
1390 }
1391 }
1392 daxctl_unref(ctx);
1393 }
1394 #endif /* defined(__linux__) && defined(CONFIG_LIBDAXCTL) */
1395
1396 return align;
1397 }
1398
1399 static int file_ram_open(const char *path,
1400 const char *region_name,
1401 bool *created,
1402 Error **errp)
1403 {
1404 char *filename;
1405 char *sanitized_name;
1406 char *c;
1407 int fd = -1;
1408
1409 *created = false;
1410 for (;;) {
1411 fd = open(path, O_RDWR);
1412 if (fd >= 0) {
1413 /* @path names an existing file, use it */
1414 break;
1415 }
1416 if (errno == ENOENT) {
1417 /* @path names a file that doesn't exist, create it */
1418 fd = open(path, O_RDWR | O_CREAT | O_EXCL, 0644);
1419 if (fd >= 0) {
1420 *created = true;
1421 break;
1422 }
1423 } else if (errno == EISDIR) {
1424 /* @path names a directory, create a file there */
1425 /* Make name safe to use with mkstemp by replacing '/' with '_'. */
1426 sanitized_name = g_strdup(region_name);
1427 for (c = sanitized_name; *c != '\0'; c++) {
1428 if (*c == '/') {
1429 *c = '_';
1430 }
1431 }
1432
1433 filename = g_strdup_printf("%s/qemu_back_mem.%s.XXXXXX", path,
1434 sanitized_name);
1435 g_free(sanitized_name);
1436
1437 fd = mkstemp(filename);
1438 if (fd >= 0) {
1439 unlink(filename);
1440 g_free(filename);
1441 break;
1442 }
1443 g_free(filename);
1444 }
1445 if (errno != EEXIST && errno != EINTR) {
1446 error_setg_errno(errp, errno,
1447 "can't open backing store %s for guest RAM",
1448 path);
1449 return -1;
1450 }
1451 /*
1452 * Try again on EINTR and EEXIST. The latter happens when
1453 * something else creates the file between our two open().
1454 */
1455 }
1456
1457 return fd;
1458 }
1459
1460 static void *file_ram_alloc(RAMBlock *block,
1461 ram_addr_t memory,
1462 int fd,
1463 bool truncate,
1464 Error **errp)
1465 {
1466 void *area;
1467
1468 block->page_size = qemu_fd_getpagesize(fd);
1469 if (block->mr->align % block->page_size) {
1470 error_setg(errp, "alignment 0x%" PRIx64
1471 " must be multiples of page size 0x%zx",
1472 block->mr->align, block->page_size);
1473 return NULL;
1474 } else if (block->mr->align && !is_power_of_2(block->mr->align)) {
1475 error_setg(errp, "alignment 0x%" PRIx64
1476 " must be a power of two", block->mr->align);
1477 return NULL;
1478 }
1479 block->mr->align = MAX(block->page_size, block->mr->align);
1480 #if defined(__s390x__)
1481 if (kvm_enabled()) {
1482 block->mr->align = MAX(block->mr->align, QEMU_VMALLOC_ALIGN);
1483 }
1484 #endif
1485
1486 if (memory < block->page_size) {
1487 error_setg(errp, "memory size 0x" RAM_ADDR_FMT " must be equal to "
1488 "or larger than page size 0x%zx",
1489 memory, block->page_size);
1490 return NULL;
1491 }
1492
1493 memory = ROUND_UP(memory, block->page_size);
1494
1495 /*
1496 * ftruncate is not supported by hugetlbfs in older
1497 * hosts, so don't bother bailing out on errors.
1498 * If anything goes wrong with it under other filesystems,
1499 * mmap will fail.
1500 *
1501 * Do not truncate the non-empty backend file to avoid corrupting
1502 * the existing data in the file. Disabling shrinking is not
1503 * enough. For example, the current vNVDIMM implementation stores
1504 * the guest NVDIMM labels at the end of the backend file. If the
1505 * backend file is later extended, QEMU will not be able to find
1506 * those labels. Therefore, extending the non-empty backend file
1507 * is disabled as well.
1508 */
1509 if (truncate && ftruncate(fd, memory)) {
1510 perror("ftruncate");
1511 }
1512
1513 area = qemu_ram_mmap(fd, memory, block->mr->align,
1514 block->flags & RAM_SHARED, block->flags & RAM_PMEM);
1515 if (area == MAP_FAILED) {
1516 error_setg_errno(errp, errno,
1517 "unable to map backing store for guest RAM");
1518 return NULL;
1519 }
1520
1521 block->fd = fd;
1522 return area;
1523 }
1524 #endif
1525
1526 /* Allocate space within the ram_addr_t space that governs the
1527 * dirty bitmaps.
1528 * Called with the ramlist lock held.
1529 */
1530 static ram_addr_t find_ram_offset(ram_addr_t size)
1531 {
1532 RAMBlock *block, *next_block;
1533 ram_addr_t offset = RAM_ADDR_MAX, mingap = RAM_ADDR_MAX;
1534
1535 assert(size != 0); /* it would hand out same offset multiple times */
1536
1537 if (QLIST_EMPTY_RCU(&ram_list.blocks)) {
1538 return 0;
1539 }
1540
1541 RAMBLOCK_FOREACH(block) {
1542 ram_addr_t candidate, next = RAM_ADDR_MAX;
1543
1544 /* Align blocks to start on a 'long' in the bitmap
1545 * which makes the bitmap sync'ing take the fast path.
1546 */
1547 candidate = block->offset + block->max_length;
1548 candidate = ROUND_UP(candidate, BITS_PER_LONG << TARGET_PAGE_BITS);
1549
1550 /* Search for the closest following block
1551 * and find the gap.
1552 */
1553 RAMBLOCK_FOREACH(next_block) {
1554 if (next_block->offset >= candidate) {
1555 next = MIN(next, next_block->offset);
1556 }
1557 }
1558
1559 /* If it fits remember our place and remember the size
1560 * of gap, but keep going so that we might find a smaller
1561 * gap to fill so avoiding fragmentation.
1562 */
1563 if (next - candidate >= size && next - candidate < mingap) {
1564 offset = candidate;
1565 mingap = next - candidate;
1566 }
1567
1568 trace_find_ram_offset_loop(size, candidate, offset, next, mingap);
1569 }
1570
1571 if (offset == RAM_ADDR_MAX) {
1572 fprintf(stderr, "Failed to find gap of requested size: %" PRIu64 "\n",
1573 (uint64_t)size);
1574 abort();
1575 }
1576
1577 trace_find_ram_offset(size, offset);
1578
1579 return offset;
1580 }
1581
1582 static unsigned long last_ram_page(void)
1583 {
1584 RAMBlock *block;
1585 ram_addr_t last = 0;
1586
1587 RCU_READ_LOCK_GUARD();
1588 RAMBLOCK_FOREACH(block) {
1589 last = MAX(last, block->offset + block->max_length);
1590 }
1591 return last >> TARGET_PAGE_BITS;
1592 }
1593
1594 static void qemu_ram_setup_dump(void *addr, ram_addr_t size)
1595 {
1596 int ret;
1597
1598 /* Use MADV_DONTDUMP, if user doesn't want the guest memory in the core */
1599 if (!machine_dump_guest_core(current_machine)) {
1600 ret = qemu_madvise(addr, size, QEMU_MADV_DONTDUMP);
1601 if (ret) {
1602 perror("qemu_madvise");
1603 fprintf(stderr, "madvise doesn't support MADV_DONTDUMP, "
1604 "but dump_guest_core=off specified\n");
1605 }
1606 }
1607 }
1608
1609 const char *qemu_ram_get_idstr(RAMBlock *rb)
1610 {
1611 return rb->idstr;
1612 }
1613
1614 void *qemu_ram_get_host_addr(RAMBlock *rb)
1615 {
1616 return rb->host;
1617 }
1618
1619 ram_addr_t qemu_ram_get_offset(RAMBlock *rb)
1620 {
1621 return rb->offset;
1622 }
1623
1624 ram_addr_t qemu_ram_get_used_length(RAMBlock *rb)
1625 {
1626 return rb->used_length;
1627 }
1628
1629 bool qemu_ram_is_shared(RAMBlock *rb)
1630 {
1631 return rb->flags & RAM_SHARED;
1632 }
1633
1634 /* Note: Only set at the start of postcopy */
1635 bool qemu_ram_is_uf_zeroable(RAMBlock *rb)
1636 {
1637 return rb->flags & RAM_UF_ZEROPAGE;
1638 }
1639
1640 void qemu_ram_set_uf_zeroable(RAMBlock *rb)
1641 {
1642 rb->flags |= RAM_UF_ZEROPAGE;
1643 }
1644
1645 bool qemu_ram_is_migratable(RAMBlock *rb)
1646 {
1647 return rb->flags & RAM_MIGRATABLE;
1648 }
1649
1650 void qemu_ram_set_migratable(RAMBlock *rb)
1651 {
1652 rb->flags |= RAM_MIGRATABLE;
1653 }
1654
1655 void qemu_ram_unset_migratable(RAMBlock *rb)
1656 {
1657 rb->flags &= ~RAM_MIGRATABLE;
1658 }
1659
1660 /* Called with iothread lock held. */
1661 void qemu_ram_set_idstr(RAMBlock *new_block, const char *name, DeviceState *dev)
1662 {
1663 RAMBlock *block;
1664
1665 assert(new_block);
1666 assert(!new_block->idstr[0]);
1667
1668 if (dev) {
1669 char *id = qdev_get_dev_path(dev);
1670 if (id) {
1671 snprintf(new_block->idstr, sizeof(new_block->idstr), "%s/", id);
1672 g_free(id);
1673 }
1674 }
1675 pstrcat(new_block->idstr, sizeof(new_block->idstr), name);
1676
1677 RCU_READ_LOCK_GUARD();
1678 RAMBLOCK_FOREACH(block) {
1679 if (block != new_block &&
1680 !strcmp(block->idstr, new_block->idstr)) {
1681 fprintf(stderr, "RAMBlock \"%s\" already registered, abort!\n",
1682 new_block->idstr);
1683 abort();
1684 }
1685 }
1686 }
1687
1688 /* Called with iothread lock held. */
1689 void qemu_ram_unset_idstr(RAMBlock *block)
1690 {
1691 /* FIXME: arch_init.c assumes that this is not called throughout
1692 * migration. Ignore the problem since hot-unplug during migration
1693 * does not work anyway.
1694 */
1695 if (block) {
1696 memset(block->idstr, 0, sizeof(block->idstr));
1697 }
1698 }
1699
1700 size_t qemu_ram_pagesize(RAMBlock *rb)
1701 {
1702 return rb->page_size;
1703 }
1704
1705 /* Returns the largest size of page in use */
1706 size_t qemu_ram_pagesize_largest(void)
1707 {
1708 RAMBlock *block;
1709 size_t largest = 0;
1710
1711 RAMBLOCK_FOREACH(block) {
1712 largest = MAX(largest, qemu_ram_pagesize(block));
1713 }
1714
1715 return largest;
1716 }
1717
1718 static int memory_try_enable_merging(void *addr, size_t len)
1719 {
1720 if (!machine_mem_merge(current_machine)) {
1721 /* disabled by the user */
1722 return 0;
1723 }
1724
1725 return qemu_madvise(addr, len, QEMU_MADV_MERGEABLE);
1726 }
1727
1728 /* Only legal before guest might have detected the memory size: e.g. on
1729 * incoming migration, or right after reset.
1730 *
1731 * As memory core doesn't know how is memory accessed, it is up to
1732 * resize callback to update device state and/or add assertions to detect
1733 * misuse, if necessary.
1734 */
1735 int qemu_ram_resize(RAMBlock *block, ram_addr_t newsize, Error **errp)
1736 {
1737 const ram_addr_t unaligned_size = newsize;
1738
1739 assert(block);
1740
1741 newsize = HOST_PAGE_ALIGN(newsize);
1742
1743 if (block->used_length == newsize) {
1744 /*
1745 * We don't have to resize the ram block (which only knows aligned
1746 * sizes), however, we have to notify if the unaligned size changed.
1747 */
1748 if (unaligned_size != memory_region_size(block->mr)) {
1749 memory_region_set_size(block->mr, unaligned_size);
1750 if (block->resized) {
1751 block->resized(block->idstr, unaligned_size, block->host);
1752 }
1753 }
1754 return 0;
1755 }
1756
1757 if (!(block->flags & RAM_RESIZEABLE)) {
1758 error_setg_errno(errp, EINVAL,
1759 "Size mismatch: %s: 0x" RAM_ADDR_FMT
1760 " != 0x" RAM_ADDR_FMT, block->idstr,
1761 newsize, block->used_length);
1762 return -EINVAL;
1763 }
1764
1765 if (block->max_length < newsize) {
1766 error_setg_errno(errp, EINVAL,
1767 "Size too large: %s: 0x" RAM_ADDR_FMT
1768 " > 0x" RAM_ADDR_FMT, block->idstr,
1769 newsize, block->max_length);
1770 return -EINVAL;
1771 }
1772
1773 cpu_physical_memory_clear_dirty_range(block->offset, block->used_length);
1774 block->used_length = newsize;
1775 cpu_physical_memory_set_dirty_range(block->offset, block->used_length,
1776 DIRTY_CLIENTS_ALL);
1777 memory_region_set_size(block->mr, unaligned_size);
1778 if (block->resized) {
1779 block->resized(block->idstr, unaligned_size, block->host);
1780 }
1781 return 0;
1782 }
1783
1784 /*
1785 * Trigger sync on the given ram block for range [start, start + length]
1786 * with the backing store if one is available.
1787 * Otherwise no-op.
1788 * @Note: this is supposed to be a synchronous op.
1789 */
1790 void qemu_ram_msync(RAMBlock *block, ram_addr_t start, ram_addr_t length)
1791 {
1792 /* The requested range should fit in within the block range */
1793 g_assert((start + length) <= block->used_length);
1794
1795 #ifdef CONFIG_LIBPMEM
1796 /* The lack of support for pmem should not block the sync */
1797 if (ramblock_is_pmem(block)) {
1798 void *addr = ramblock_ptr(block, start);
1799 pmem_persist(addr, length);
1800 return;
1801 }
1802 #endif
1803 if (block->fd >= 0) {
1804 /**
1805 * Case there is no support for PMEM or the memory has not been
1806 * specified as persistent (or is not one) - use the msync.
1807 * Less optimal but still achieves the same goal
1808 */
1809 void *addr = ramblock_ptr(block, start);
1810 if (qemu_msync(addr, length, block->fd)) {
1811 warn_report("%s: failed to sync memory range: start: "
1812 RAM_ADDR_FMT " length: " RAM_ADDR_FMT,
1813 __func__, start, length);
1814 }
1815 }
1816 }
1817
1818 /* Called with ram_list.mutex held */
1819 static void dirty_memory_extend(ram_addr_t old_ram_size,
1820 ram_addr_t new_ram_size)
1821 {
1822 ram_addr_t old_num_blocks = DIV_ROUND_UP(old_ram_size,
1823 DIRTY_MEMORY_BLOCK_SIZE);
1824 ram_addr_t new_num_blocks = DIV_ROUND_UP(new_ram_size,
1825 DIRTY_MEMORY_BLOCK_SIZE);
1826 int i;
1827
1828 /* Only need to extend if block count increased */
1829 if (new_num_blocks <= old_num_blocks) {
1830 return;
1831 }
1832
1833 for (i = 0; i < DIRTY_MEMORY_NUM; i++) {
1834 DirtyMemoryBlocks *old_blocks;
1835 DirtyMemoryBlocks *new_blocks;
1836 int j;
1837
1838 old_blocks = qatomic_rcu_read(&ram_list.dirty_memory[i]);
1839 new_blocks = g_malloc(sizeof(*new_blocks) +
1840 sizeof(new_blocks->blocks[0]) * new_num_blocks);
1841
1842 if (old_num_blocks) {
1843 memcpy(new_blocks->blocks, old_blocks->blocks,
1844 old_num_blocks * sizeof(old_blocks->blocks[0]));
1845 }
1846
1847 for (j = old_num_blocks; j < new_num_blocks; j++) {
1848 new_blocks->blocks[j] = bitmap_new(DIRTY_MEMORY_BLOCK_SIZE);
1849 }
1850
1851 qatomic_rcu_set(&ram_list.dirty_memory[i], new_blocks);
1852
1853 if (old_blocks) {
1854 g_free_rcu(old_blocks, rcu);
1855 }
1856 }
1857 }
1858
1859 static void ram_block_add(RAMBlock *new_block, Error **errp, bool shared)
1860 {
1861 RAMBlock *block;
1862 RAMBlock *last_block = NULL;
1863 ram_addr_t old_ram_size, new_ram_size;
1864 Error *err = NULL;
1865
1866 old_ram_size = last_ram_page();
1867
1868 qemu_mutex_lock_ramlist();
1869 new_block->offset = find_ram_offset(new_block->max_length);
1870
1871 if (!new_block->host) {
1872 if (xen_enabled()) {
1873 xen_ram_alloc(new_block->offset, new_block->max_length,
1874 new_block->mr, &err);
1875 if (err) {
1876 error_propagate(errp, err);
1877 qemu_mutex_unlock_ramlist();
1878 return;
1879 }
1880 } else {
1881 new_block->host = phys_mem_alloc(new_block->max_length,
1882 &new_block->mr->align, shared);
1883 if (!new_block->host) {
1884 error_setg_errno(errp, errno,
1885 "cannot set up guest memory '%s'",
1886 memory_region_name(new_block->mr));
1887 qemu_mutex_unlock_ramlist();
1888 return;
1889 }
1890 memory_try_enable_merging(new_block->host, new_block->max_length);
1891 }
1892 }
1893
1894 new_ram_size = MAX(old_ram_size,
1895 (new_block->offset + new_block->max_length) >> TARGET_PAGE_BITS);
1896 if (new_ram_size > old_ram_size) {
1897 dirty_memory_extend(old_ram_size, new_ram_size);
1898 }
1899 /* Keep the list sorted from biggest to smallest block. Unlike QTAILQ,
1900 * QLIST (which has an RCU-friendly variant) does not have insertion at
1901 * tail, so save the last element in last_block.
1902 */
1903 RAMBLOCK_FOREACH(block) {
1904 last_block = block;
1905 if (block->max_length < new_block->max_length) {
1906 break;
1907 }
1908 }
1909 if (block) {
1910 QLIST_INSERT_BEFORE_RCU(block, new_block, next);
1911 } else if (last_block) {
1912 QLIST_INSERT_AFTER_RCU(last_block, new_block, next);
1913 } else { /* list is empty */
1914 QLIST_INSERT_HEAD_RCU(&ram_list.blocks, new_block, next);
1915 }
1916 ram_list.mru_block = NULL;
1917
1918 /* Write list before version */
1919 smp_wmb();
1920 ram_list.version++;
1921 qemu_mutex_unlock_ramlist();
1922
1923 cpu_physical_memory_set_dirty_range(new_block->offset,
1924 new_block->used_length,
1925 DIRTY_CLIENTS_ALL);
1926
1927 if (new_block->host) {
1928 qemu_ram_setup_dump(new_block->host, new_block->max_length);
1929 qemu_madvise(new_block->host, new_block->max_length, QEMU_MADV_HUGEPAGE);
1930 /*
1931 * MADV_DONTFORK is also needed by KVM in absence of synchronous MMU
1932 * Configure it unless the machine is a qtest server, in which case
1933 * KVM is not used and it may be forked (eg for fuzzing purposes).
1934 */
1935 if (!qtest_enabled()) {
1936 qemu_madvise(new_block->host, new_block->max_length,
1937 QEMU_MADV_DONTFORK);
1938 }
1939 ram_block_notify_add(new_block->host, new_block->max_length);
1940 }
1941 }
1942
1943 #ifdef CONFIG_POSIX
1944 RAMBlock *qemu_ram_alloc_from_fd(ram_addr_t size, MemoryRegion *mr,
1945 uint32_t ram_flags, int fd,
1946 Error **errp)
1947 {
1948 RAMBlock *new_block;
1949 Error *local_err = NULL;
1950 int64_t file_size, file_align;
1951
1952 /* Just support these ram flags by now. */
1953 assert((ram_flags & ~(RAM_SHARED | RAM_PMEM)) == 0);
1954
1955 if (xen_enabled()) {
1956 error_setg(errp, "-mem-path not supported with Xen");
1957 return NULL;
1958 }
1959
1960 if (kvm_enabled() && !kvm_has_sync_mmu()) {
1961 error_setg(errp,
1962 "host lacks kvm mmu notifiers, -mem-path unsupported");
1963 return NULL;
1964 }
1965
1966 if (phys_mem_alloc != qemu_anon_ram_alloc) {
1967 /*
1968 * file_ram_alloc() needs to allocate just like
1969 * phys_mem_alloc, but we haven't bothered to provide
1970 * a hook there.
1971 */
1972 error_setg(errp,
1973 "-mem-path not supported with this accelerator");
1974 return NULL;
1975 }
1976
1977 size = HOST_PAGE_ALIGN(size);
1978 file_size = get_file_size(fd);
1979 if (file_size > 0 && file_size < size) {
1980 error_setg(errp, "backing store size 0x%" PRIx64
1981 " does not match 'size' option 0x" RAM_ADDR_FMT,
1982 file_size, size);
1983 return NULL;
1984 }
1985
1986 file_align = get_file_align(fd);
1987 if (file_align > 0 && mr && file_align > mr->align) {
1988 error_setg(errp, "backing store align 0x%" PRIx64
1989 " is larger than 'align' option 0x%" PRIx64,
1990 file_align, mr->align);
1991 return NULL;
1992 }
1993
1994 new_block = g_malloc0(sizeof(*new_block));
1995 new_block->mr = mr;
1996 new_block->used_length = size;
1997 new_block->max_length = size;
1998 new_block->flags = ram_flags;
1999 new_block->host = file_ram_alloc(new_block, size, fd, !file_size, errp);
2000 if (!new_block->host) {
2001 g_free(new_block);
2002 return NULL;
2003 }
2004
2005 ram_block_add(new_block, &local_err, ram_flags & RAM_SHARED);
2006 if (local_err) {
2007 g_free(new_block);
2008 error_propagate(errp, local_err);
2009 return NULL;
2010 }
2011 return new_block;
2012
2013 }
2014
2015
2016 RAMBlock *qemu_ram_alloc_from_file(ram_addr_t size, MemoryRegion *mr,
2017 uint32_t ram_flags, const char *mem_path,
2018 Error **errp)
2019 {
2020 int fd;
2021 bool created;
2022 RAMBlock *block;
2023
2024 fd = file_ram_open(mem_path, memory_region_name(mr), &created, errp);
2025 if (fd < 0) {
2026 return NULL;
2027 }
2028
2029 block = qemu_ram_alloc_from_fd(size, mr, ram_flags, fd, errp);
2030 if (!block) {
2031 if (created) {
2032 unlink(mem_path);
2033 }
2034 close(fd);
2035 return NULL;
2036 }
2037
2038 return block;
2039 }
2040 #endif
2041
2042 static
2043 RAMBlock *qemu_ram_alloc_internal(ram_addr_t size, ram_addr_t max_size,
2044 void (*resized)(const char*,
2045 uint64_t length,
2046 void *host),
2047 void *host, bool resizeable, bool share,
2048 MemoryRegion *mr, Error **errp)
2049 {
2050 RAMBlock *new_block;
2051 Error *local_err = NULL;
2052
2053 size = HOST_PAGE_ALIGN(size);
2054 max_size = HOST_PAGE_ALIGN(max_size);
2055 new_block = g_malloc0(sizeof(*new_block));
2056 new_block->mr = mr;
2057 new_block->resized = resized;
2058 new_block->used_length = size;
2059 new_block->max_length = max_size;
2060 assert(max_size >= size);
2061 new_block->fd = -1;
2062 new_block->page_size = qemu_real_host_page_size;
2063 new_block->host = host;
2064 if (host) {
2065 new_block->flags |= RAM_PREALLOC;
2066 }
2067 if (resizeable) {
2068 new_block->flags |= RAM_RESIZEABLE;
2069 }
2070 ram_block_add(new_block, &local_err, share);
2071 if (local_err) {
2072 g_free(new_block);
2073 error_propagate(errp, local_err);
2074 return NULL;
2075 }
2076 return new_block;
2077 }
2078
2079 RAMBlock *qemu_ram_alloc_from_ptr(ram_addr_t size, void *host,
2080 MemoryRegion *mr, Error **errp)
2081 {
2082 return qemu_ram_alloc_internal(size, size, NULL, host, false,
2083 false, mr, errp);
2084 }
2085
2086 RAMBlock *qemu_ram_alloc(ram_addr_t size, bool share,
2087 MemoryRegion *mr, Error **errp)
2088 {
2089 return qemu_ram_alloc_internal(size, size, NULL, NULL, false,
2090 share, mr, errp);
2091 }
2092
2093 RAMBlock *qemu_ram_alloc_resizeable(ram_addr_t size, ram_addr_t maxsz,
2094 void (*resized)(const char*,
2095 uint64_t length,
2096 void *host),
2097 MemoryRegion *mr, Error **errp)
2098 {
2099 return qemu_ram_alloc_internal(size, maxsz, resized, NULL, true,
2100 false, mr, errp);
2101 }
2102
2103 static void reclaim_ramblock(RAMBlock *block)
2104 {
2105 if (block->flags & RAM_PREALLOC) {
2106 ;
2107 } else if (xen_enabled()) {
2108 xen_invalidate_map_cache_entry(block->host);
2109 #ifndef _WIN32
2110 } else if (block->fd >= 0) {
2111 qemu_ram_munmap(block->fd, block->host, block->max_length);
2112 close(block->fd);
2113 #endif
2114 } else {
2115 qemu_anon_ram_free(block->host, block->max_length);
2116 }
2117 g_free(block);
2118 }
2119
2120 void qemu_ram_free(RAMBlock *block)
2121 {
2122 if (!block) {
2123 return;
2124 }
2125
2126 if (block->host) {
2127 ram_block_notify_remove(block->host, block->max_length);
2128 }
2129
2130 qemu_mutex_lock_ramlist();
2131 QLIST_REMOVE_RCU(block, next);
2132 ram_list.mru_block = NULL;
2133 /* Write list before version */
2134 smp_wmb();
2135 ram_list.version++;
2136 call_rcu(block, reclaim_ramblock, rcu);
2137 qemu_mutex_unlock_ramlist();
2138 }
2139
2140 #ifndef _WIN32
2141 void qemu_ram_remap(ram_addr_t addr, ram_addr_t length)
2142 {
2143 RAMBlock *block;
2144 ram_addr_t offset;
2145 int flags;
2146 void *area, *vaddr;
2147
2148 RAMBLOCK_FOREACH(block) {
2149 offset = addr - block->offset;
2150 if (offset < block->max_length) {
2151 vaddr = ramblock_ptr(block, offset);
2152 if (block->flags & RAM_PREALLOC) {
2153 ;
2154 } else if (xen_enabled()) {
2155 abort();
2156 } else {
2157 flags = MAP_FIXED;
2158 if (block->fd >= 0) {
2159 flags |= (block->flags & RAM_SHARED ?
2160 MAP_SHARED : MAP_PRIVATE);
2161 area = mmap(vaddr, length, PROT_READ | PROT_WRITE,
2162 flags, block->fd, offset);
2163 } else {
2164 /*
2165 * Remap needs to match alloc. Accelerators that
2166 * set phys_mem_alloc never remap. If they did,
2167 * we'd need a remap hook here.
2168 */
2169 assert(phys_mem_alloc == qemu_anon_ram_alloc);
2170
2171 flags |= MAP_PRIVATE | MAP_ANONYMOUS;
2172 area = mmap(vaddr, length, PROT_READ | PROT_WRITE,
2173 flags, -1, 0);
2174 }
2175 if (area != vaddr) {
2176 error_report("Could not remap addr: "
2177 RAM_ADDR_FMT "@" RAM_ADDR_FMT "",
2178 length, addr);
2179 exit(1);
2180 }
2181 memory_try_enable_merging(vaddr, length);
2182 qemu_ram_setup_dump(vaddr, length);
2183 }
2184 }
2185 }
2186 }
2187 #endif /* !_WIN32 */
2188
2189 /* Return a host pointer to ram allocated with qemu_ram_alloc.
2190 * This should not be used for general purpose DMA. Use address_space_map
2191 * or address_space_rw instead. For local memory (e.g. video ram) that the
2192 * device owns, use memory_region_get_ram_ptr.
2193 *
2194 * Called within RCU critical section.
2195 */
2196 void *qemu_map_ram_ptr(RAMBlock *ram_block, ram_addr_t addr)
2197 {
2198 RAMBlock *block = ram_block;
2199
2200 if (block == NULL) {
2201 block = qemu_get_ram_block(addr);
2202 addr -= block->offset;
2203 }
2204
2205 if (xen_enabled() && block->host == NULL) {
2206 /* We need to check if the requested address is in the RAM
2207 * because we don't want to map the entire memory in QEMU.
2208 * In that case just map until the end of the page.
2209 */
2210 if (block->offset == 0) {
2211 return xen_map_cache(addr, 0, 0, false);
2212 }
2213
2214 block->host = xen_map_cache(block->offset, block->max_length, 1, false);
2215 }
2216 return ramblock_ptr(block, addr);
2217 }
2218
2219 /* Return a host pointer to guest's ram. Similar to qemu_map_ram_ptr
2220 * but takes a size argument.
2221 *
2222 * Called within RCU critical section.
2223 */
2224 static void *qemu_ram_ptr_length(RAMBlock *ram_block, ram_addr_t addr,
2225 hwaddr *size, bool lock)
2226 {
2227 RAMBlock *block = ram_block;
2228 if (*size == 0) {
2229 return NULL;
2230 }
2231
2232 if (block == NULL) {
2233 block = qemu_get_ram_block(addr);
2234 addr -= block->offset;
2235 }
2236 *size = MIN(*size, block->max_length - addr);
2237
2238 if (xen_enabled() && block->host == NULL) {
2239 /* We need to check if the requested address is in the RAM
2240 * because we don't want to map the entire memory in QEMU.
2241 * In that case just map the requested area.
2242 */
2243 if (block->offset == 0) {
2244 return xen_map_cache(addr, *size, lock, lock);
2245 }
2246
2247 block->host = xen_map_cache(block->offset, block->max_length, 1, lock);
2248 }
2249
2250 return ramblock_ptr(block, addr);
2251 }
2252
2253 /* Return the offset of a hostpointer within a ramblock */
2254 ram_addr_t qemu_ram_block_host_offset(RAMBlock *rb, void *host)
2255 {
2256 ram_addr_t res = (uint8_t *)host - (uint8_t *)rb->host;
2257 assert((uintptr_t)host >= (uintptr_t)rb->host);
2258 assert(res < rb->max_length);
2259
2260 return res;
2261 }
2262
2263 /*
2264 * Translates a host ptr back to a RAMBlock, a ram_addr and an offset
2265 * in that RAMBlock.
2266 *
2267 * ptr: Host pointer to look up
2268 * round_offset: If true round the result offset down to a page boundary
2269 * *ram_addr: set to result ram_addr
2270 * *offset: set to result offset within the RAMBlock
2271 *
2272 * Returns: RAMBlock (or NULL if not found)
2273 *
2274 * By the time this function returns, the returned pointer is not protected
2275 * by RCU anymore. If the caller is not within an RCU critical section and
2276 * does not hold the iothread lock, it must have other means of protecting the
2277 * pointer, such as a reference to the region that includes the incoming
2278 * ram_addr_t.
2279 */
2280 RAMBlock *qemu_ram_block_from_host(void *ptr, bool round_offset,
2281 ram_addr_t *offset)
2282 {
2283 RAMBlock *block;
2284 uint8_t *host = ptr;
2285
2286 if (xen_enabled()) {
2287 ram_addr_t ram_addr;
2288 RCU_READ_LOCK_GUARD();
2289 ram_addr = xen_ram_addr_from_mapcache(ptr);
2290 block = qemu_get_ram_block(ram_addr);
2291 if (block) {
2292 *offset = ram_addr - block->offset;
2293 }
2294 return block;
2295 }
2296
2297 RCU_READ_LOCK_GUARD();
2298 block = qatomic_rcu_read(&ram_list.mru_block);
2299 if (block && block->host && host - block->host < block->max_length) {
2300 goto found;
2301 }
2302
2303 RAMBLOCK_FOREACH(block) {
2304 /* This case append when the block is not mapped. */
2305 if (block->host == NULL) {
2306 continue;
2307 }
2308 if (host - block->host < block->max_length) {
2309 goto found;
2310 }
2311 }
2312
2313 return NULL;
2314
2315 found:
2316 *offset = (host - block->host);
2317 if (round_offset) {
2318 *offset &= TARGET_PAGE_MASK;
2319 }
2320 return block;
2321 }
2322
2323 /*
2324 * Finds the named RAMBlock
2325 *
2326 * name: The name of RAMBlock to find
2327 *
2328 * Returns: RAMBlock (or NULL if not found)
2329 */
2330 RAMBlock *qemu_ram_block_by_name(const char *name)
2331 {
2332 RAMBlock *block;
2333
2334 RAMBLOCK_FOREACH(block) {
2335 if (!strcmp(name, block->idstr)) {
2336 return block;
2337 }
2338 }
2339
2340 return NULL;
2341 }
2342
2343 /* Some of the softmmu routines need to translate from a host pointer
2344 (typically a TLB entry) back to a ram offset. */
2345 ram_addr_t qemu_ram_addr_from_host(void *ptr)
2346 {
2347 RAMBlock *block;
2348 ram_addr_t offset;
2349
2350 block = qemu_ram_block_from_host(ptr, false, &offset);
2351 if (!block) {
2352 return RAM_ADDR_INVALID;
2353 }
2354
2355 return block->offset + offset;
2356 }
2357
2358 /* Generate a debug exception if a watchpoint has been hit. */
2359 void cpu_check_watchpoint(CPUState *cpu, vaddr addr, vaddr len,
2360 MemTxAttrs attrs, int flags, uintptr_t ra)
2361 {
2362 CPUClass *cc = CPU_GET_CLASS(cpu);
2363 CPUWatchpoint *wp;
2364
2365 assert(tcg_enabled());
2366 if (cpu->watchpoint_hit) {
2367 /*
2368 * We re-entered the check after replacing the TB.
2369 * Now raise the debug interrupt so that it will
2370 * trigger after the current instruction.
2371 */
2372 qemu_mutex_lock_iothread();
2373 cpu_interrupt(cpu, CPU_INTERRUPT_DEBUG);
2374 qemu_mutex_unlock_iothread();
2375 return;
2376 }
2377
2378 addr = cc->adjust_watchpoint_address(cpu, addr, len);
2379 QTAILQ_FOREACH(wp, &cpu->watchpoints, entry) {
2380 if (watchpoint_address_matches(wp, addr, len)
2381 && (wp->flags & flags)) {
2382 if (replay_running_debug()) {
2383 /*
2384 * Don't process the watchpoints when we are
2385 * in a reverse debugging operation.
2386 */
2387 replay_breakpoint();
2388 return;
2389 }
2390 if (flags == BP_MEM_READ) {
2391 wp->flags |= BP_WATCHPOINT_HIT_READ;
2392 } else {
2393 wp->flags |= BP_WATCHPOINT_HIT_WRITE;
2394 }
2395 wp->hitaddr = MAX(addr, wp->vaddr);
2396 wp->hitattrs = attrs;
2397 if (!cpu->watchpoint_hit) {
2398 if (wp->flags & BP_CPU &&
2399 !cc->debug_check_watchpoint(cpu, wp)) {
2400 wp->flags &= ~BP_WATCHPOINT_HIT;
2401 continue;
2402 }
2403 cpu->watchpoint_hit = wp;
2404
2405 mmap_lock();
2406 tb_check_watchpoint(cpu, ra);
2407 if (wp->flags & BP_STOP_BEFORE_ACCESS) {
2408 cpu->exception_index = EXCP_DEBUG;
2409 mmap_unlock();
2410 cpu_loop_exit_restore(cpu, ra);
2411 } else {
2412 /* Force execution of one insn next time. */
2413 cpu->cflags_next_tb = 1 | curr_cflags();
2414 mmap_unlock();
2415 if (ra) {
2416 cpu_restore_state(cpu, ra, true);
2417 }
2418 cpu_loop_exit_noexc(cpu);
2419 }
2420 }
2421 } else {
2422 wp->flags &= ~BP_WATCHPOINT_HIT;
2423 }
2424 }
2425 }
2426
2427 static MemTxResult flatview_read(FlatView *fv, hwaddr addr,
2428 MemTxAttrs attrs, void *buf, hwaddr len);
2429 static MemTxResult flatview_write(FlatView *fv, hwaddr addr, MemTxAttrs attrs,
2430 const void *buf, hwaddr len);
2431 static bool flatview_access_valid(FlatView *fv, hwaddr addr, hwaddr len,
2432 bool is_write, MemTxAttrs attrs);
2433
2434 static MemTxResult subpage_read(void *opaque, hwaddr addr, uint64_t *data,
2435 unsigned len, MemTxAttrs attrs)
2436 {
2437 subpage_t *subpage = opaque;
2438 uint8_t buf[8];
2439 MemTxResult res;
2440
2441 #if defined(DEBUG_SUBPAGE)
2442 printf("%s: subpage %p len %u addr " TARGET_FMT_plx "\n", __func__,
2443 subpage, len, addr);
2444 #endif
2445 res = flatview_read(subpage->fv, addr + subpage->base, attrs, buf, len);
2446 if (res) {
2447 return res;
2448 }
2449 *data = ldn_p(buf, len);
2450 return MEMTX_OK;
2451 }
2452
2453 static MemTxResult subpage_write(void *opaque, hwaddr addr,
2454 uint64_t value, unsigned len, MemTxAttrs attrs)
2455 {
2456 subpage_t *subpage = opaque;
2457 uint8_t buf[8];
2458
2459 #if defined(DEBUG_SUBPAGE)
2460 printf("%s: subpage %p len %u addr " TARGET_FMT_plx
2461 " value %"PRIx64"\n",
2462 __func__, subpage, len, addr, value);
2463 #endif
2464 stn_p(buf, len, value);
2465 return flatview_write(subpage->fv, addr + subpage->base, attrs, buf, len);
2466 }
2467
2468 static bool subpage_accepts(void *opaque, hwaddr addr,
2469 unsigned len, bool is_write,
2470 MemTxAttrs attrs)
2471 {
2472 subpage_t *subpage = opaque;
2473 #if defined(DEBUG_SUBPAGE)
2474 printf("%s: subpage %p %c len %u addr " TARGET_FMT_plx "\n",
2475 __func__, subpage, is_write ? 'w' : 'r', len, addr);
2476 #endif
2477
2478 return flatview_access_valid(subpage->fv, addr + subpage->base,
2479 len, is_write, attrs);
2480 }
2481
2482 static const MemoryRegionOps subpage_ops = {
2483 .read_with_attrs = subpage_read,
2484 .write_with_attrs = subpage_write,
2485 .impl.min_access_size = 1,
2486 .impl.max_access_size = 8,
2487 .valid.min_access_size = 1,
2488 .valid.max_access_size = 8,
2489 .valid.accepts = subpage_accepts,
2490 .endianness = DEVICE_NATIVE_ENDIAN,
2491 };
2492
2493 static int subpage_register(subpage_t *mmio, uint32_t start, uint32_t end,
2494 uint16_t section)
2495 {
2496 int idx, eidx;
2497
2498 if (start >= TARGET_PAGE_SIZE || end >= TARGET_PAGE_SIZE)
2499 return -1;
2500 idx = SUBPAGE_IDX(start);
2501 eidx = SUBPAGE_IDX(end);
2502 #if defined(DEBUG_SUBPAGE)
2503 printf("%s: %p start %08x end %08x idx %08x eidx %08x section %d\n",
2504 __func__, mmio, start, end, idx, eidx, section);
2505 #endif
2506 for (; idx <= eidx; idx++) {
2507 mmio->sub_section[idx] = section;
2508 }
2509
2510 return 0;
2511 }
2512
2513 static subpage_t *subpage_init(FlatView *fv, hwaddr base)
2514 {
2515 subpage_t *mmio;
2516
2517 /* mmio->sub_section is set to PHYS_SECTION_UNASSIGNED with g_malloc0 */
2518 mmio = g_malloc0(sizeof(subpage_t) + TARGET_PAGE_SIZE * sizeof(uint16_t));
2519 mmio->fv = fv;
2520 mmio->base = base;
2521 memory_region_init_io(&mmio->iomem, NULL, &subpage_ops, mmio,
2522 NULL, TARGET_PAGE_SIZE);
2523 mmio->iomem.subpage = true;
2524 #if defined(DEBUG_SUBPAGE)
2525 printf("%s: %p base " TARGET_FMT_plx " len %08x\n", __func__,
2526 mmio, base, TARGET_PAGE_SIZE);
2527 #endif
2528
2529 return mmio;
2530 }
2531
2532 static uint16_t dummy_section(PhysPageMap *map, FlatView *fv, MemoryRegion *mr)
2533 {
2534 assert(fv);
2535 MemoryRegionSection section = {
2536 .fv = fv,
2537 .mr = mr,
2538 .offset_within_address_space = 0,
2539 .offset_within_region = 0,
2540 .size = int128_2_64(),
2541 };
2542
2543 return phys_section_add(map, &section);
2544 }
2545
2546 MemoryRegionSection *iotlb_to_section(CPUState *cpu,
2547 hwaddr index, MemTxAttrs attrs)
2548 {
2549 int asidx = cpu_asidx_from_attrs(cpu, attrs);
2550 CPUAddressSpace *cpuas = &cpu->cpu_ases[asidx];
2551 AddressSpaceDispatch *d = qatomic_rcu_read(&cpuas->memory_dispatch);
2552 MemoryRegionSection *sections = d->map.sections;
2553
2554 return &sections[index & ~TARGET_PAGE_MASK];
2555 }
2556
2557 static void io_mem_init(void)
2558 {
2559 memory_region_init_io(&io_mem_unassigned, NULL, &unassigned_mem_ops, NULL,
2560 NULL, UINT64_MAX);
2561 }
2562
2563 AddressSpaceDispatch *address_space_dispatch_new(FlatView *fv)
2564 {
2565 AddressSpaceDispatch *d = g_new0(AddressSpaceDispatch, 1);
2566 uint16_t n;
2567
2568 n = dummy_section(&d->map, fv, &io_mem_unassigned);
2569 assert(n == PHYS_SECTION_UNASSIGNED);
2570
2571 d->phys_map = (PhysPageEntry) { .ptr = PHYS_MAP_NODE_NIL, .skip = 1 };
2572
2573 return d;
2574 }
2575
2576 void address_space_dispatch_free(AddressSpaceDispatch *d)
2577 {
2578 phys_sections_free(&d->map);
2579 g_free(d);
2580 }
2581
2582 static void do_nothing(CPUState *cpu, run_on_cpu_data d)
2583 {
2584 }
2585
2586 static void tcg_log_global_after_sync(MemoryListener *listener)
2587 {
2588 CPUAddressSpace *cpuas;
2589
2590 /* Wait for the CPU to end the current TB. This avoids the following
2591 * incorrect race:
2592 *
2593 * vCPU migration
2594 * ---------------------- -------------------------
2595 * TLB check -> slow path
2596 * notdirty_mem_write
2597 * write to RAM
2598 * mark dirty
2599 * clear dirty flag
2600 * TLB check -> fast path
2601 * read memory
2602 * write to RAM
2603 *
2604 * by pushing the migration thread's memory read after the vCPU thread has
2605 * written the memory.
2606 */
2607 if (replay_mode == REPLAY_MODE_NONE) {
2608 /*
2609 * VGA can make calls to this function while updating the screen.
2610 * In record/replay mode this causes a deadlock, because
2611 * run_on_cpu waits for rr mutex. Therefore no races are possible
2612 * in this case and no need for making run_on_cpu when
2613 * record/replay is not enabled.
2614 */
2615 cpuas = container_of(listener, CPUAddressSpace, tcg_as_listener);
2616 run_on_cpu(cpuas->cpu, do_nothing, RUN_ON_CPU_NULL);
2617 }
2618 }
2619
2620 static void tcg_commit(MemoryListener *listener)
2621 {
2622 CPUAddressSpace *cpuas;
2623 AddressSpaceDispatch *d;
2624
2625 assert(tcg_enabled());
2626 /* since each CPU stores ram addresses in its TLB cache, we must
2627 reset the modified entries */
2628 cpuas = container_of(listener, CPUAddressSpace, tcg_as_listener);
2629 cpu_reloading_memory_map();
2630 /* The CPU and TLB are protected by the iothread lock.
2631 * We reload the dispatch pointer now because cpu_reloading_memory_map()
2632 * may have split the RCU critical section.
2633 */
2634 d = address_space_to_dispatch(cpuas->as);
2635 qatomic_rcu_set(&cpuas->memory_dispatch, d);
2636 tlb_flush(cpuas->cpu);
2637 }
2638
2639 static void memory_map_init(void)
2640 {
2641 system_memory = g_malloc(sizeof(*system_memory));
2642
2643 memory_region_init(system_memory, NULL, "system", UINT64_MAX);
2644 address_space_init(&address_space_memory, system_memory, "memory");
2645
2646 system_io = g_malloc(sizeof(*system_io));
2647 memory_region_init_io(system_io, NULL, &unassigned_io_ops, NULL, "io",
2648 65536);
2649 address_space_init(&address_space_io, system_io, "I/O");
2650 }
2651
2652 MemoryRegion *get_system_memory(void)
2653 {
2654 return system_memory;
2655 }
2656
2657 MemoryRegion *get_system_io(void)
2658 {
2659 return system_io;
2660 }
2661
2662 static void invalidate_and_set_dirty(MemoryRegion *mr, hwaddr addr,
2663 hwaddr length)
2664 {
2665 uint8_t dirty_log_mask = memory_region_get_dirty_log_mask(mr);
2666 addr += memory_region_get_ram_addr(mr);
2667
2668 /* No early return if dirty_log_mask is or becomes 0, because
2669 * cpu_physical_memory_set_dirty_range will still call
2670 * xen_modified_memory.
2671 */
2672 if (dirty_log_mask) {
2673 dirty_log_mask =
2674 cpu_physical_memory_range_includes_clean(addr, length, dirty_log_mask);
2675 }
2676 if (dirty_log_mask & (1 << DIRTY_MEMORY_CODE)) {
2677 assert(tcg_enabled());
2678 tb_invalidate_phys_range(addr, addr + length);
2679 dirty_log_mask &= ~(1 << DIRTY_MEMORY_CODE);
2680 }
2681 cpu_physical_memory_set_dirty_range(addr, length, dirty_log_mask);
2682 }
2683
2684 void memory_region_flush_rom_device(MemoryRegion *mr, hwaddr addr, hwaddr size)
2685 {
2686 /*
2687 * In principle this function would work on other memory region types too,
2688 * but the ROM device use case is the only one where this operation is
2689 * necessary. Other memory regions should use the
2690 * address_space_read/write() APIs.
2691 */
2692 assert(memory_region_is_romd(mr));
2693
2694 invalidate_and_set_dirty(mr, addr, size);
2695 }
2696
2697 static int memory_access_size(MemoryRegion *mr, unsigned l, hwaddr addr)
2698 {
2699 unsigned access_size_max = mr->ops->valid.max_access_size;
2700
2701 /* Regions are assumed to support 1-4 byte accesses unless
2702 otherwise specified. */
2703 if (access_size_max == 0) {
2704 access_size_max = 4;
2705 }
2706
2707 /* Bound the maximum access by the alignment of the address. */
2708 if (!mr->ops->impl.unaligned) {
2709 unsigned align_size_max = addr & -addr;
2710 if (align_size_max != 0 && align_size_max < access_size_max) {
2711 access_size_max = align_size_max;
2712 }
2713 }
2714
2715 /* Don't attempt accesses larger than the maximum. */
2716 if (l > access_size_max) {
2717 l = access_size_max;
2718 }
2719 l = pow2floor(l);
2720
2721 return l;
2722 }
2723
2724 static bool prepare_mmio_access(MemoryRegion *mr)
2725 {
2726 bool release_lock = false;
2727
2728 if (!qemu_mutex_iothread_locked()) {
2729 qemu_mutex_lock_iothread();
2730 release_lock = true;
2731 }
2732 if (mr->flush_coalesced_mmio) {
2733 qemu_flush_coalesced_mmio_buffer();
2734 }
2735
2736 return release_lock;
2737 }
2738
2739 /* Called within RCU critical section. */
2740 static MemTxResult flatview_write_continue(FlatView *fv, hwaddr addr,
2741 MemTxAttrs attrs,
2742 const void *ptr,
2743 hwaddr len, hwaddr addr1,
2744 hwaddr l, MemoryRegion *mr)
2745 {
2746 uint8_t *ram_ptr;
2747 uint64_t val;
2748 MemTxResult result = MEMTX_OK;
2749 bool release_lock = false;
2750 const uint8_t *buf = ptr;
2751
2752 for (;;) {
2753 if (!memory_access_is_direct(mr, true)) {
2754 release_lock |= prepare_mmio_access(mr);
2755 l = memory_access_size(mr, l, addr1);
2756 /* XXX: could force current_cpu to NULL to avoid
2757 potential bugs */
2758 val = ldn_he_p(buf, l);
2759 result |= memory_region_dispatch_write(mr, addr1, val,
2760 size_memop(l), attrs);
2761 } else {
2762 /* RAM case */
2763 ram_ptr = qemu_ram_ptr_length(mr->ram_block, addr1, &l, false);
2764 memcpy(ram_ptr, buf, l);
2765 invalidate_and_set_dirty(mr, addr1, l);
2766 }
2767
2768 if (release_lock) {
2769 qemu_mutex_unlock_iothread();
2770 release_lock = false;
2771 }
2772
2773 len -= l;
2774 buf += l;
2775 addr += l;
2776
2777 if (!len) {
2778 break;
2779 }
2780
2781 l = len;
2782 mr = flatview_translate(fv, addr, &addr1, &l, true, attrs);
2783 }
2784
2785 return result;
2786 }
2787
2788 /* Called from RCU critical section. */
2789 static MemTxResult flatview_write(FlatView *fv, hwaddr addr, MemTxAttrs attrs,
2790 const void *buf, hwaddr len)
2791 {
2792 hwaddr l;
2793 hwaddr addr1;
2794 MemoryRegion *mr;
2795 MemTxResult result = MEMTX_OK;
2796
2797 l = len;
2798 mr = flatview_translate(fv, addr, &addr1, &l, true, attrs);
2799 result = flatview_write_continue(fv, addr, attrs, buf, len,
2800 addr1, l, mr);
2801
2802 return result;
2803 }
2804
2805 /* Called within RCU critical section. */
2806 MemTxResult flatview_read_continue(FlatView *fv, hwaddr addr,
2807 MemTxAttrs attrs, void *ptr,
2808 hwaddr len, hwaddr addr1, hwaddr l,
2809 MemoryRegion *mr)
2810 {
2811 uint8_t *ram_ptr;
2812 uint64_t val;
2813 MemTxResult result = MEMTX_OK;
2814 bool release_lock = false;
2815 uint8_t *buf = ptr;
2816
2817 for (;;) {
2818 if (!memory_access_is_direct(mr, false)) {
2819 /* I/O case */
2820 release_lock |= prepare_mmio_access(mr);
2821 l = memory_access_size(mr, l, addr1);
2822 result |= memory_region_dispatch_read(mr, addr1, &val,
2823 size_memop(l), attrs);
2824 stn_he_p(buf, l, val);
2825 } else {
2826 /* RAM case */
2827 fuzz_dma_read_cb(addr, len, mr, false);
2828 ram_ptr = qemu_ram_ptr_length(mr->ram_block, addr1, &l, false);
2829 memcpy(buf, ram_ptr, l);
2830 }
2831
2832 if (release_lock) {
2833 qemu_mutex_unlock_iothread();
2834 release_lock = false;
2835 }
2836
2837 len -= l;
2838 buf += l;
2839 addr += l;
2840
2841 if (!len) {
2842 break;
2843 }
2844
2845 l = len;
2846 mr = flatview_translate(fv, addr, &addr1, &l, false, attrs);
2847 }
2848
2849 return result;
2850 }
2851
2852 /* Called from RCU critical section. */
2853 static MemTxResult flatview_read(FlatView *fv, hwaddr addr,
2854 MemTxAttrs attrs, void *buf, hwaddr len)
2855 {
2856 hwaddr l;
2857 hwaddr addr1;
2858 MemoryRegion *mr;
2859
2860 l = len;
2861 mr = flatview_translate(fv, addr, &addr1, &l, false, attrs);
2862 return flatview_read_continue(fv, addr, attrs, buf, len,
2863 addr1, l, mr);
2864 }
2865
2866 MemTxResult address_space_read_full(AddressSpace *as, hwaddr addr,
2867 MemTxAttrs attrs, void *buf, hwaddr len)
2868 {
2869 MemTxResult result = MEMTX_OK;
2870 FlatView *fv;
2871
2872 if (len > 0) {
2873 RCU_READ_LOCK_GUARD();
2874 fv = address_space_to_flatview(as);
2875 result = flatview_read(fv, addr, attrs, buf, len);
2876 }
2877
2878 return result;
2879 }
2880
2881 MemTxResult address_space_write(AddressSpace *as, hwaddr addr,
2882 MemTxAttrs attrs,
2883 const void *buf, hwaddr len)
2884 {
2885 MemTxResult result = MEMTX_OK;
2886 FlatView *fv;
2887
2888 if (len > 0) {
2889 RCU_READ_LOCK_GUARD();
2890 fv = address_space_to_flatview(as);
2891 result = flatview_write(fv, addr, attrs, buf, len);
2892 }
2893
2894 return result;
2895 }
2896
2897 MemTxResult address_space_rw(AddressSpace *as, hwaddr addr, MemTxAttrs attrs,
2898 void *buf, hwaddr len, bool is_write)
2899 {
2900 if (is_write) {
2901 return address_space_write(as, addr, attrs, buf, len);
2902 } else {
2903 return address_space_read_full(as, addr, attrs, buf, len);
2904 }
2905 }
2906
2907 void cpu_physical_memory_rw(hwaddr addr, void *buf,
2908 hwaddr len, bool is_write)
2909 {
2910 address_space_rw(&address_space_memory, addr, MEMTXATTRS_UNSPECIFIED,
2911 buf, len, is_write);
2912 }
2913
2914 enum write_rom_type {
2915 WRITE_DATA,
2916 FLUSH_CACHE,
2917 };
2918
2919 static inline MemTxResult address_space_write_rom_internal(AddressSpace *as,
2920 hwaddr addr,
2921 MemTxAttrs attrs,
2922 const void *ptr,
2923 hwaddr len,
2924 enum write_rom_type type)
2925 {
2926 hwaddr l;
2927 uint8_t *ram_ptr;
2928 hwaddr addr1;
2929 MemoryRegion *mr;
2930 const uint8_t *buf = ptr;
2931
2932 RCU_READ_LOCK_GUARD();
2933 while (len > 0) {
2934 l = len;
2935 mr = address_space_translate(as, addr, &addr1, &l, true, attrs);
2936
2937 if (!(memory_region_is_ram(mr) ||
2938 memory_region_is_romd(mr))) {
2939 l = memory_access_size(mr, l, addr1);
2940 } else {
2941 /* ROM/RAM case */
2942 ram_ptr = qemu_map_ram_ptr(mr->ram_block, addr1);
2943 switch (type) {
2944 case WRITE_DATA:
2945 memcpy(ram_ptr, buf, l);
2946 invalidate_and_set_dirty(mr, addr1, l);
2947 break;
2948 case FLUSH_CACHE:
2949 flush_idcache_range((uintptr_t)ram_ptr, (uintptr_t)ram_ptr, l);
2950 break;
2951 }
2952 }
2953 len -= l;
2954 buf += l;
2955 addr += l;
2956 }
2957 return MEMTX_OK;
2958 }
2959
2960 /* used for ROM loading : can write in RAM and ROM */
2961 MemTxResult address_space_write_rom(AddressSpace *as, hwaddr addr,
2962 MemTxAttrs attrs,
2963 const void *buf, hwaddr len)
2964 {
2965 return address_space_write_rom_internal(as, addr, attrs,
2966 buf, len, WRITE_DATA);
2967 }
2968
2969 void cpu_flush_icache_range(hwaddr start, hwaddr len)
2970 {
2971 /*
2972 * This function should do the same thing as an icache flush that was
2973 * triggered from within the guest. For TCG we are always cache coherent,
2974 * so there is no need to flush anything. For KVM / Xen we need to flush
2975 * the host's instruction cache at least.
2976 */
2977 if (tcg_enabled()) {
2978 return;
2979 }
2980
2981 address_space_write_rom_internal(&address_space_memory,
2982 start, MEMTXATTRS_UNSPECIFIED,
2983 NULL, len, FLUSH_CACHE);
2984 }
2985
2986 typedef struct {
2987 MemoryRegion *mr;
2988 void *buffer;
2989 hwaddr addr;
2990 hwaddr len;
2991 bool in_use;
2992 } BounceBuffer;
2993
2994 static BounceBuffer bounce;
2995
2996 typedef struct MapClient {
2997 QEMUBH *bh;
2998 QLIST_ENTRY(MapClient) link;
2999 } MapClient;
3000
3001 QemuMutex map_client_list_lock;
3002 static QLIST_HEAD(, MapClient) map_client_list
3003 = QLIST_HEAD_INITIALIZER(map_client_list);
3004
3005 static void cpu_unregister_map_client_do(MapClient *client)
3006 {
3007 QLIST_REMOVE(client, link);
3008 g_free(client);
3009 }
3010
3011 static void cpu_notify_map_clients_locked(void)
3012 {
3013 MapClient *client;
3014
3015 while (!QLIST_EMPTY(&map_client_list)) {
3016 client = QLIST_FIRST(&map_client_list);
3017 qemu_bh_schedule(client->bh);
3018 cpu_unregister_map_client_do(client);
3019 }
3020 }
3021
3022 void cpu_register_map_client(QEMUBH *bh)
3023 {
3024 MapClient *client = g_malloc(sizeof(*client));
3025
3026 qemu_mutex_lock(&map_client_list_lock);
3027 client->bh = bh;
3028 QLIST_INSERT_HEAD(&map_client_list, client, link);
3029 if (!qatomic_read(&bounce.in_use)) {
3030 cpu_notify_map_clients_locked();
3031 }
3032 qemu_mutex_unlock(&map_client_list_lock);
3033 }
3034
3035 void cpu_exec_init_all(void)
3036 {
3037 qemu_mutex_init(&ram_list.mutex);
3038 /* The data structures we set up here depend on knowing the page size,
3039 * so no more changes can be made after this point.
3040 * In an ideal world, nothing we did before we had finished the
3041 * machine setup would care about the target page size, and we could
3042 * do this much later, rather than requiring board models to state
3043 * up front what their requirements are.
3044 */
3045 finalize_target_page_bits();
3046 io_mem_init();
3047 memory_map_init();
3048 qemu_mutex_init(&map_client_list_lock);
3049 }
3050
3051 void cpu_unregister_map_client(QEMUBH *bh)
3052 {
3053 MapClient *client;
3054
3055 qemu_mutex_lock(&map_client_list_lock);
3056 QLIST_FOREACH(client, &map_client_list, link) {
3057 if (client->bh == bh) {
3058 cpu_unregister_map_client_do(client);
3059 break;
3060 }
3061 }
3062 qemu_mutex_unlock(&map_client_list_lock);
3063 }
3064
3065 static void cpu_notify_map_clients(void)
3066 {
3067 qemu_mutex_lock(&map_client_list_lock);
3068 cpu_notify_map_clients_locked();
3069 qemu_mutex_unlock(&map_client_list_lock);
3070 }
3071
3072 static bool flatview_access_valid(FlatView *fv, hwaddr addr, hwaddr len,
3073 bool is_write, MemTxAttrs attrs)
3074 {
3075 MemoryRegion *mr;
3076 hwaddr l, xlat;
3077
3078 while (len > 0) {
3079 l = len;
3080 mr = flatview_translate(fv, addr, &xlat, &l, is_write, attrs);
3081 if (!memory_access_is_direct(mr, is_write)) {
3082 l = memory_access_size(mr, l, addr);
3083 if (!memory_region_access_valid(mr, xlat, l, is_write, attrs)) {
3084 return false;
3085 }
3086 }
3087
3088 len -= l;
3089 addr += l;
3090 }
3091 return true;
3092 }
3093
3094 bool address_space_access_valid(AddressSpace *as, hwaddr addr,
3095 hwaddr len, bool is_write,
3096 MemTxAttrs attrs)
3097 {
3098 FlatView *fv;
3099 bool result;
3100
3101 RCU_READ_LOCK_GUARD();
3102 fv = address_space_to_flatview(as);
3103 result = flatview_access_valid(fv, addr, len, is_write, attrs);
3104 return result;
3105 }
3106
3107 static hwaddr
3108 flatview_extend_translation(FlatView *fv, hwaddr addr,
3109 hwaddr target_len,
3110 MemoryRegion *mr, hwaddr base, hwaddr len,
3111 bool is_write, MemTxAttrs attrs)
3112 {
3113 hwaddr done = 0;
3114 hwaddr xlat;
3115 MemoryRegion *this_mr;
3116
3117 for (;;) {
3118 target_len -= len;
3119 addr += len;
3120 done += len;
3121 if (target_len == 0) {
3122 return done;
3123 }
3124
3125 len = target_len;
3126 this_mr = flatview_translate(fv, addr, &xlat,
3127 &len, is_write, attrs);
3128 if (this_mr != mr || xlat != base + done) {
3129 return done;
3130 }
3131 }
3132 }
3133
3134 /* Map a physical memory region into a host virtual address.
3135 * May map a subset of the requested range, given by and returned in *plen.
3136 * May return NULL if resources needed to perform the mapping are exhausted.
3137 * Use only for reads OR writes - not for read-modify-write operations.
3138 * Use cpu_register_map_client() to know when retrying the map operation is
3139 * likely to succeed.
3140 */
3141 void *address_space_map(AddressSpace *as,
3142 hwaddr addr,
3143 hwaddr *plen,
3144 bool is_write,
3145 MemTxAttrs attrs)
3146 {
3147 hwaddr len = *plen;
3148 hwaddr l, xlat;
3149 MemoryRegion *mr;
3150 void *ptr;
3151 FlatView *fv;
3152
3153 if (len == 0) {
3154 return NULL;
3155 }
3156
3157 l = len;
3158 RCU_READ_LOCK_GUARD();
3159 fv = address_space_to_flatview(as);
3160 mr = flatview_translate(fv, addr, &xlat, &l, is_write, attrs);
3161
3162 if (!memory_access_is_direct(mr, is_write)) {
3163 if (qatomic_xchg(&bounce.in_use, true)) {
3164 *plen = 0;
3165 return NULL;
3166 }
3167 /* Avoid unbounded allocations */
3168 l = MIN(l, TARGET_PAGE_SIZE);
3169 bounce.buffer = qemu_memalign(TARGET_PAGE_SIZE, l);
3170 bounce.addr = addr;
3171 bounce.len = l;
3172
3173 memory_region_ref(mr);
3174 bounce.mr = mr;
3175 if (!is_write) {
3176 flatview_read(fv, addr, MEMTXATTRS_UNSPECIFIED,
3177 bounce.buffer, l);
3178 }
3179
3180 *plen = l;
3181 return bounce.buffer;
3182 }
3183
3184
3185 memory_region_ref(mr);
3186 *plen = flatview_extend_translation(fv, addr, len, mr, xlat,
3187 l, is_write, attrs);
3188 fuzz_dma_read_cb(addr, *plen, mr, is_write);
3189 ptr = qemu_ram_ptr_length(mr->ram_block, xlat, plen, true);
3190
3191 return ptr;
3192 }
3193
3194 /* Unmaps a memory region previously mapped by address_space_map().
3195 * Will also mark the memory as dirty if is_write is true. access_len gives
3196 * the amount of memory that was actually read or written by the caller.
3197 */
3198 void address_space_unmap(AddressSpace *as, void *buffer, hwaddr len,
3199 bool is_write, hwaddr access_len)
3200 {
3201 if (buffer != bounce.buffer) {
3202 MemoryRegion *mr;
3203 ram_addr_t addr1;
3204
3205 mr = memory_region_from_host(buffer, &addr1);
3206 assert(mr != NULL);
3207 if (is_write) {
3208 invalidate_and_set_dirty(mr, addr1, access_len);
3209 }
3210 if (xen_enabled()) {
3211 xen_invalidate_map_cache_entry(buffer);
3212 }
3213 memory_region_unref(mr);
3214 return;
3215 }
3216 if (is_write) {
3217 address_space_write(as, bounce.addr, MEMTXATTRS_UNSPECIFIED,
3218 bounce.buffer, access_len);
3219 }
3220 qemu_vfree(bounce.buffer);
3221 bounce.buffer = NULL;
3222 memory_region_unref(bounce.mr);
3223 qatomic_mb_set(&bounce.in_use, false);
3224 cpu_notify_map_clients();
3225 }
3226
3227 void *cpu_physical_memory_map(hwaddr addr,
3228 hwaddr *plen,
3229 bool is_write)
3230 {
3231 return address_space_map(&address_space_memory, addr, plen, is_write,
3232 MEMTXATTRS_UNSPECIFIED);
3233 }
3234
3235 void cpu_physical_memory_unmap(void *buffer, hwaddr len,
3236 bool is_write, hwaddr access_len)
3237 {
3238 return address_space_unmap(&address_space_memory, buffer, len, is_write, access_len);
3239 }
3240
3241 #define ARG1_DECL AddressSpace *as
3242 #define ARG1 as
3243 #define SUFFIX
3244 #define TRANSLATE(...) address_space_translate(as, __VA_ARGS__)
3245 #define RCU_READ_LOCK(...) rcu_read_lock()
3246 #define RCU_READ_UNLOCK(...) rcu_read_unlock()
3247 #include "memory_ldst.c.inc"
3248
3249 int64_t address_space_cache_init(MemoryRegionCache *cache,
3250 AddressSpace *as,
3251 hwaddr addr,
3252 hwaddr len,
3253 bool is_write)
3254 {
3255 AddressSpaceDispatch *d;
3256 hwaddr l;
3257 MemoryRegion *mr;
3258 Int128 diff;
3259
3260 assert(len > 0);
3261
3262 l = len;
3263 cache->fv = address_space_get_flatview(as);
3264 d = flatview_to_dispatch(cache->fv);
3265 cache->mrs = *address_space_translate_internal(d, addr, &cache->xlat, &l, true);
3266
3267 /*
3268 * cache->xlat is now relative to cache->mrs.mr, not to the section itself.
3269 * Take that into account to compute how many bytes are there between
3270 * cache->xlat and the end of the section.
3271 */
3272 diff = int128_sub(cache->mrs.size,
3273 int128_make64(cache->xlat - cache->mrs.offset_within_region));
3274 l = int128_get64(int128_min(diff, int128_make64(l)));
3275
3276 mr = cache->mrs.mr;
3277 memory_region_ref(mr);
3278 if (memory_access_is_direct(mr, is_write)) {
3279 /* We don't care about the memory attributes here as we're only
3280 * doing this if we found actual RAM, which behaves the same
3281 * regardless of attributes; so UNSPECIFIED is fine.
3282 */
3283 l = flatview_extend_translation(cache->fv, addr, len, mr,
3284 cache->xlat, l, is_write,
3285 MEMTXATTRS_UNSPECIFIED);
3286 cache->ptr = qemu_ram_ptr_length(mr->ram_block, cache->xlat, &l, true);
3287 } else {
3288 cache->ptr = NULL;
3289 }
3290
3291 cache->len = l;
3292 cache->is_write = is_write;
3293 return l;
3294 }
3295
3296 void address_space_cache_invalidate(MemoryRegionCache *cache,
3297 hwaddr addr,
3298 hwaddr access_len)
3299 {
3300 assert(cache->is_write);
3301 if (likely(cache->ptr)) {
3302 invalidate_and_set_dirty(cache->mrs.mr, addr + cache->xlat, access_len);
3303 }
3304 }
3305
3306 void address_space_cache_destroy(MemoryRegionCache *cache)
3307 {
3308 if (!cache->mrs.mr) {
3309 return;
3310 }
3311
3312 if (xen_enabled()) {
3313 xen_invalidate_map_cache_entry(cache->ptr);
3314 }
3315 memory_region_unref(cache->mrs.mr);
3316 flatview_unref(cache->fv);
3317 cache->mrs.mr = NULL;
3318 cache->fv = NULL;
3319 }
3320
3321 /* Called from RCU critical section. This function has the same
3322 * semantics as address_space_translate, but it only works on a
3323 * predefined range of a MemoryRegion that was mapped with
3324 * address_space_cache_init.
3325 */
3326 static inline MemoryRegion *address_space_translate_cached(
3327 MemoryRegionCache *cache, hwaddr addr, hwaddr *xlat,
3328 hwaddr *plen, bool is_write, MemTxAttrs attrs)
3329 {
3330 MemoryRegionSection section;
3331 MemoryRegion *mr;
3332 IOMMUMemoryRegion *iommu_mr;
3333 AddressSpace *target_as;
3334
3335 assert(!cache->ptr);
3336 *xlat = addr + cache->xlat;
3337
3338 mr = cache->mrs.mr;
3339 iommu_mr = memory_region_get_iommu(mr);
3340 if (!iommu_mr) {
3341 /* MMIO region. */
3342 return mr;
3343 }
3344
3345 section = address_space_translate_iommu(iommu_mr, xlat, plen,
3346 NULL, is_write, true,
3347 &target_as, attrs);
3348 return section.mr;
3349 }
3350
3351 /* Called from RCU critical section. address_space_read_cached uses this
3352 * out of line function when the target is an MMIO or IOMMU region.
3353 */
3354 MemTxResult
3355 address_space_read_cached_slow(MemoryRegionCache *cache, hwaddr addr,
3356 void *buf, hwaddr len)
3357 {
3358 hwaddr addr1, l;
3359 MemoryRegion *mr;
3360
3361 l = len;
3362 mr = address_space_translate_cached(cache, addr, &addr1, &l, false,
3363 MEMTXATTRS_UNSPECIFIED);
3364 return flatview_read_continue(cache->fv,
3365 addr, MEMTXATTRS_UNSPECIFIED, buf, len,
3366 addr1, l, mr);
3367 }
3368
3369 /* Called from RCU critical section. address_space_write_cached uses this
3370 * out of line function when the target is an MMIO or IOMMU region.
3371 */
3372 MemTxResult
3373 address_space_write_cached_slow(MemoryRegionCache *cache, hwaddr addr,
3374 const void *buf, hwaddr len)
3375 {
3376 hwaddr addr1, l;
3377 MemoryRegion *mr;
3378
3379 l = len;
3380 mr = address_space_translate_cached(cache, addr, &addr1, &l, true,
3381 MEMTXATTRS_UNSPECIFIED);
3382 return flatview_write_continue(cache->fv,
3383 addr, MEMTXATTRS_UNSPECIFIED, buf, len,
3384 addr1, l, mr);
3385 }
3386
3387 #define ARG1_DECL MemoryRegionCache *cache
3388 #define ARG1 cache
3389 #define SUFFIX _cached_slow
3390 #define TRANSLATE(...) address_space_translate_cached(cache, __VA_ARGS__)
3391 #define RCU_READ_LOCK() ((void)0)
3392 #define RCU_READ_UNLOCK() ((void)0)
3393 #include "memory_ldst.c.inc"
3394
3395 /* virtual memory access for debug (includes writing to ROM) */
3396 int cpu_memory_rw_debug(CPUState *cpu, target_ulong addr,
3397 void *ptr, target_ulong len, bool is_write)
3398 {
3399 hwaddr phys_addr;
3400 target_ulong l, page;
3401 uint8_t *buf = ptr;
3402
3403 cpu_synchronize_state(cpu);
3404 while (len > 0) {
3405 int asidx;
3406 MemTxAttrs attrs;
3407 MemTxResult res;
3408
3409 page = addr & TARGET_PAGE_MASK;
3410 phys_addr = cpu_get_phys_page_attrs_debug(cpu, page, &attrs);
3411 asidx = cpu_asidx_from_attrs(cpu, attrs);
3412 /* if no physical page mapped, return an error */
3413 if (phys_addr == -1)
3414 return -1;
3415 l = (page + TARGET_PAGE_SIZE) - addr;
3416 if (l > len)
3417 l = len;
3418 phys_addr += (addr & ~TARGET_PAGE_MASK);
3419 if (is_write) {
3420 res = address_space_write_rom(cpu->cpu_ases[asidx].as, phys_addr,
3421 attrs, buf, l);
3422 } else {
3423 res = address_space_read(cpu->cpu_ases[asidx].as, phys_addr,
3424 attrs, buf, l);
3425 }
3426 if (res != MEMTX_OK) {
3427 return -1;
3428 }
3429 len -= l;
3430 buf += l;
3431 addr += l;
3432 }
3433 return 0;
3434 }
3435
3436 /*
3437 * Allows code that needs to deal with migration bitmaps etc to still be built
3438 * target independent.
3439