memory.c

   1 /*
   2  * Physical memory management
   3  *
   4  * Copyright 2011 Red Hat, Inc. and/or its affiliates
   5  *
   6  * Authors:
   7  *  Avi Kivity <[email protected]>
   8  *
   9  * This work is licensed under the terms of the GNU GPL, version 2.  See
  10  * the COPYING file in the top-level directory.
  11  *
  12  * Contributions after 2012-01-13 are licensed under the terms of the
  13  * GNU GPL, version 2 or (at your option) any later version.
  14  */
  15
  16 #include "exec/memory.h"
  17 #include "exec/address-spaces.h"
  18 #include "exec/ioport.h"
  19 #include "qemu/bitops.h"
  20 #include "qom/object.h"
  21 #include "sysemu/kvm.h"
  22 #include "trace.h"
  23 #include <assert.h>
  24
  25 #include "exec/memory-internal.h"
  26
  27 //#define DEBUG_UNASSIGNED
  28
  29 static unsigned memory_region_transaction_depth;
  30 static bool memory_region_update_pending;
  31 static bool global_dirty_log = false;
  32
  33 /* flat_view_mutex is taken around reading as->current_map; the critical
  34  * section is extremely short, so I'm using a single mutex for every AS.
  35  * We could also RCU for the read-side.
  36  *
  37  * The BQL is taken around transaction commits, hence both locks are taken
  38  * while writing to as->current_map (with the BQL taken outside).
  39  */
  40 static QemuMutex flat_view_mutex;
  41
  42 static QTAILQ_HEAD(memory_listeners, MemoryListener) memory_listeners
  43     = QTAILQ_HEAD_INITIALIZER(memory_listeners);
  44
  45 static QTAILQ_HEAD(, AddressSpace) address_spaces
  46     = QTAILQ_HEAD_INITIALIZER(address_spaces);
  47
  48 static void memory_init(void)
  49 {
  50     qemu_mutex_init(&flat_view_mutex);
  51 }
  52
  53 typedef struct AddrRange AddrRange;
  54
  55 /*
  56  * Note using signed integers limits us to physical addresses at most
  57  * 63 bits wide.  They are needed for negative offsetting in aliases
  58  * (large MemoryRegion::alias_offset).
  59  */
  60 struct AddrRange {
  61     Int128 start;
  62     Int128 size;
  63 };
  64
  65 static AddrRange addrrange_make(Int128 start, Int128 size)
  66 {
  67     return (AddrRange) { start, size };
  68 }
  69
  70 static bool addrrange_equal(AddrRange r1, AddrRange r2)
  71 {
  72     return int128_eq(r1.start, r2.start) && int128_eq(r1.size, r2.size);
  73 }
  74
  75 static Int128 addrrange_end(AddrRange r)
  76 {
  77     return int128_add(r.start, r.size);
  78 }
  79
  80 static AddrRange addrrange_shift(AddrRange range, Int128 delta)
  81 {
  82     int128_addto(&range.start, delta);
  83     return range;
  84 }
  85
  86 static bool addrrange_contains(AddrRange range, Int128 addr)
  87 {
  88     return int128_ge(addr, range.start)
  89         && int128_lt(addr, addrrange_end(range));
  90 }
  91
  92 static bool addrrange_intersects(AddrRange r1, AddrRange r2)
  93 {
  94     return addrrange_contains(r1, r2.start)
  95         || addrrange_contains(r2, r1.start);
  96 }
  97
  98 static AddrRange addrrange_intersection(AddrRange r1, AddrRange r2)
  99 {
 100     Int128 start = int128_max(r1.start, r2.start);
 101     Int128 end = int128_min(addrrange_end(r1), addrrange_end(r2));
 102     return addrrange_make(start, int128_sub(end, start));
 103 }
 104
 105 enum ListenerDirection { Forward, Reverse };
 106
 107 static bool memory_listener_match(MemoryListener *listener,
 108                                   MemoryRegionSection *section)
 109 {
 110     return !listener->address_space_filter
 111         || listener->address_space_filter == section->address_space;
 112 }
 113
 114 #define MEMORY_LISTENER_CALL_GLOBAL(_callback, _direction, _args...)    \
 115     do {                                                                \
 116         MemoryListener *_listener;                                      \
 117                                                                         \
 118         switch (_direction) {                                           \
 119         case Forward:                                                   \
 120             QTAILQ_FOREACH(_listener, &memory_listeners, link) {        \
 121                 if (_listener->_callback) {                             \
 122                     _listener->_callback(_listener, ##_args);           \
 123                 }                                                       \
 124             }                                                           \
 125             break;                                                      \
 126         case Reverse:                                                   \
 127             QTAILQ_FOREACH_REVERSE(_listener, &memory_listeners,        \
 128                                    memory_listeners, link) {            \
 129                 if (_listener->_callback) {                             \
 130                     _listener->_callback(_listener, ##_args);           \
 131                 }                                                       \
 132             }                                                           \
 133             break;                                                      \
 134         default:                                                        \
 135             abort();                                                    \
 136         }                                                               \
 137     } while (0)
 138
 139 #define MEMORY_LISTENER_CALL(_callback, _direction, _section, _args...) \
 140     do {                                                                \
 141         MemoryListener *_listener;                                      \
 142                                                                         \
 143         switch (_direction) {                                           \
 144         case Forward:                                                   \
 145             QTAILQ_FOREACH(_listener, &memory_listeners, link) {        \
 146                 if (_listener->_callback                                \
 147                     && memory_listener_match(_listener, _section)) {    \
 148                     _listener->_callback(_listener, _section, ##_args); \
 149                 }                                                       \
 150             }                                                           \
 151             break;                                                      \
 152         case Reverse:                                                   \
 153             QTAILQ_FOREACH_REVERSE(_listener, &memory_listeners,        \
 154                                    memory_listeners, link) {            \
 155                 if (_listener->_callback                                \
 156                     && memory_listener_match(_listener, _section)) {    \
 157                     _listener->_callback(_listener, _section, ##_args); \
 158                 }                                                       \
 159             }                                                           \
 160             break;                                                      \
 161         default:                                                        \
 162             abort();                                                    \
 163         }                                                               \
 164     } while (0)
 165
 166 /* No need to ref/unref .mr, the FlatRange keeps it alive.  */
 167 #define MEMORY_LISTENER_UPDATE_REGION(fr, as, dir, callback)            \
 168     MEMORY_LISTENER_CALL(callback, dir, (&(MemoryRegionSection) {       \
 169         .mr = (fr)->mr,                                                 \
 170         .address_space = (as),                                          \
 171         .offset_within_region = (fr)->offset_in_region,                 \
 172         .size = (fr)->addr.size,                                        \
 173         .offset_within_address_space = int128_get64((fr)->addr.start),  \
 174         .readonly = (fr)->readonly,                                     \
 175               }))
 176
 177 struct CoalescedMemoryRange {
 178     AddrRange addr;
 179     QTAILQ_ENTRY(CoalescedMemoryRange) link;
 180 };
 181
 182 struct MemoryRegionIoeventfd {
 183     AddrRange addr;
 184     bool match_data;
 185     uint64_t data;
 186     EventNotifier *e;
 187 };
 188
 189 static bool memory_region_ioeventfd_before(MemoryRegionIoeventfd a,
 190                                            MemoryRegionIoeventfd b)
 191 {
 192     if (int128_lt(a.addr.start, b.addr.start)) {
 193         return true;
 194     } else if (int128_gt(a.addr.start, b.addr.start)) {
 195         return false;
 196     } else if (int128_lt(a.addr.size, b.addr.size)) {
 197         return true;
 198     } else if (int128_gt(a.addr.size, b.addr.size)) {
 199         return false;
 200     } else if (a.match_data < b.match_data) {
 201         return true;
 202     } else  if (a.match_data > b.match_data) {
 203         return false;
 204     } else if (a.match_data) {
 205         if (a.data < b.data) {
 206             return true;
 207         } else if (a.data > b.data) {
 208             return false;
 209         }
 210     }
 211     if (a.e < b.e) {
 212         return true;
 213     } else if (a.e > b.e) {
 214         return false;
 215     }
 216     return false;
 217 }
 218
 219 static bool memory_region_ioeventfd_equal(MemoryRegionIoeventfd a,
 220                                           MemoryRegionIoeventfd b)
 221 {
 222     return !memory_region_ioeventfd_before(a, b)
 223         && !memory_region_ioeventfd_before(b, a);
 224 }
 225
 226 typedef struct FlatRange FlatRange;
 227 typedef struct FlatView FlatView;
 228
 229 /* Range of memory in the global map.  Addresses are absolute. */
 230 struct FlatRange {
 231     MemoryRegion *mr;
 232     hwaddr offset_in_region;
 233     AddrRange addr;
 234     uint8_t dirty_log_mask;
 235     bool romd_mode;
 236     bool readonly;
 237 };
 238
 239 /* Flattened global view of current active memory hierarchy.  Kept in sorted
 240  * order.
 241  */
 242 struct FlatView {
 243     unsigned ref;
 244     FlatRange *ranges;
 245     unsigned nr;
 246     unsigned nr_allocated;
 247 };
 248
 249 typedef struct AddressSpaceOps AddressSpaceOps;
 250
 251 #define FOR_EACH_FLAT_RANGE(var, view)          \
 252     for (var = (view)->ranges; var < (view)->ranges + (view)->nr; ++var)
 253
 254 static bool flatrange_equal(FlatRange *a, FlatRange *b)
 255 {
 256     return a->mr == b->mr
 257         && addrrange_equal(a->addr, b->addr)
 258         && a->offset_in_region == b->offset_in_region
 259         && a->romd_mode == b->romd_mode
 260         && a->readonly == b->readonly;
 261 }
 262
 263 static void flatview_init(FlatView *view)
 264 {
 265     view->ref = 1;
 266     view->ranges = NULL;
 267     view->nr = 0;
 268     view->nr_allocated = 0;
 269 }
 270
 271 /* Insert a range into a given position.  Caller is responsible for maintaining
 272  * sorting order.
 273  */
 274 static void flatview_insert(FlatView *view, unsigned pos, FlatRange *range)
 275 {
 276     if (view->nr == view->nr_allocated) {
 277         view->nr_allocated = MAX(2 * view->nr, 10);
 278         view->ranges = g_realloc(view->ranges,
 279                                     view->nr_allocated * sizeof(*view->ranges));
 280     }
 281     memmove(view->ranges + pos + 1, view->ranges + pos,
 282             (view->nr - pos) * sizeof(FlatRange));
 283     view->ranges[pos] = *range;
 284     memory_region_ref(range->mr);
 285     ++view->nr;
 286 }
 287
 288 static void flatview_destroy(FlatView *view)
 289 {
 290     int i;
 291
 292     for (i = 0; i < view->nr; i++) {
 293         memory_region_unref(view->ranges[i].mr);
 294     }
 295     g_free(view->ranges);
 296     g_free(view);
 297 }
 298
 299 static void flatview_ref(FlatView *view)
 300 {
 301     atomic_inc(&view->ref);
 302 }
 303
 304 static void flatview_unref(FlatView *view)
 305 {
 306     if (atomic_fetch_dec(&view->ref) == 1) {
 307         flatview_destroy(view);
 308     }
 309 }
 310
 311 static bool can_merge(FlatRange *r1, FlatRange *r2)
 312 {
 313     return int128_eq(addrrange_end(r1->addr), r2->addr.start)
 314         && r1->mr == r2->mr
 315         && int128_eq(int128_add(int128_make64(r1->offset_in_region),
 316                                 r1->addr.size),
 317                      int128_make64(r2->offset_in_region))
 318         && r1->dirty_log_mask == r2->dirty_log_mask
 319         && r1->romd_mode == r2->romd_mode
 320         && r1->readonly == r2->readonly;
 321 }
 322
 323 /* Attempt to simplify a view by merging adjacent ranges */
 324 static void flatview_simplify(FlatView *view)
 325 {
 326     unsigned i, j;
 327
 328     i = 0;
 329     while (i < view->nr) {
 330         j = i + 1;
 331         while (j < view->nr
 332                && can_merge(&view->ranges[j-1], &view->ranges[j])) {
 333             int128_addto(&view->ranges[i].addr.size, view->ranges[j].addr.size);
 334             ++j;
 335         }
 336         ++i;
 337         memmove(&view->ranges[i], &view->ranges[j],
 338                 (view->nr - j) * sizeof(view->ranges[j]));
 339         view->nr -= j - i;
 340     }
 341 }
 342
 343 static bool memory_region_big_endian(MemoryRegion *mr)
 344 {
 345 #ifdef TARGET_WORDS_BIGENDIAN
 346     return mr->ops->endianness != DEVICE_LITTLE_ENDIAN;
 347 #else
 348     return mr->ops->endianness == DEVICE_BIG_ENDIAN;
 349 #endif
 350 }
 351
 352 static bool memory_region_wrong_endianness(MemoryRegion *mr)
 353 {
 354 #ifdef TARGET_WORDS_BIGENDIAN
 355     return mr->ops->endianness == DEVICE_LITTLE_ENDIAN;
 356 #else
 357     return mr->ops->endianness == DEVICE_BIG_ENDIAN;
 358 #endif
 359 }
 360
 361 static void adjust_endianness(MemoryRegion *mr, uint64_t *data, unsigned size)
 362 {
 363     if (memory_region_wrong_endianness(mr)) {
 364         switch (size) {
 365         case 1:
 366             break;
 367         case 2:
 368             *data = bswap16(*data);
 369             break;
 370         case 4:
 371             *data = bswap32(*data);
 372             break;
 373         case 8:
 374             *data = bswap64(*data);
 375             break;
 376         default:
 377             abort();
 378         }
 379     }
 380 }
 381
 382 static void memory_region_oldmmio_read_accessor(MemoryRegion *mr,
 383                                                 hwaddr addr,
 384                                                 uint64_t *value,
 385                                                 unsigned size,
 386                                                 unsigned shift,
 387                                                 uint64_t mask)
 388 {
 389     uint64_t tmp;
 390
 391     tmp = mr->ops->old_mmio.read[ctz32(size)](mr->opaque, addr);
 392     trace_memory_region_ops_read(mr, addr, tmp, size);
 393     *value |= (tmp & mask) << shift;
 394 }
 395
 396 static void memory_region_read_accessor(MemoryRegion *mr,
 397                                         hwaddr addr,
 398                                         uint64_t *value,
 399                                         unsigned size,
 400                                         unsigned shift,
 401                                         uint64_t mask)
 402 {
 403     uint64_t tmp;
 404
 405     if (mr->flush_coalesced_mmio) {
 406         qemu_flush_coalesced_mmio_buffer();
 407     }
 408     tmp = mr->ops->read(mr->opaque, addr, size);
 409     trace_memory_region_ops_read(mr, addr, tmp, size);
 410     *value |= (tmp & mask) << shift;
 411 }
 412
 413 static void memory_region_oldmmio_write_accessor(MemoryRegion *mr,
 414                                                  hwaddr addr,
 415                                                  uint64_t *value,
 416                                                  unsigned size,
 417                                                  unsigned shift,
 418                                                  uint64_t mask)
 419 {
 420     uint64_t tmp;
 421
 422     tmp = (*value >> shift) & mask;
 423     trace_memory_region_ops_write(mr, addr, tmp, size);
 424     mr->ops->old_mmio.write[ctz32(size)](mr->opaque, addr, tmp);
 425 }
 426
 427 static void memory_region_write_accessor(MemoryRegion *mr,
 428                                          hwaddr addr,
 429                                          uint64_t *value,
 430                                          unsigned size,
 431                                          unsigned shift,
 432                                          uint64_t mask)
 433 {
 434     uint64_t tmp;
 435
 436     if (mr->flush_coalesced_mmio) {
 437         qemu_flush_coalesced_mmio_buffer();
 438     }
 439     tmp = (*value >> shift) & mask;
 440     trace_memory_region_ops_write(mr, addr, tmp, size);
 441     mr->ops->write(mr->opaque, addr, tmp, size);
 442 }
 443
 444 static void access_with_adjusted_size(hwaddr addr,
 445                                       uint64_t *value,
 446                                       unsigned size,
 447                                       unsigned access_size_min,
 448                                       unsigned access_size_max,
 449                                       void (*access)(MemoryRegion *mr,
 450                                                      hwaddr addr,
 451                                                      uint64_t *value,
 452                                                      unsigned size,
 453                                                      unsigned shift,
 454                                                      uint64_t mask),
 455                                       MemoryRegion *mr)
 456 {
 457     uint64_t access_mask;
 458     unsigned access_size;
 459     unsigned i;
 460
 461     if (!access_size_min) {
 462         access_size_min = 1;
 463     }
 464     if (!access_size_max) {
 465         access_size_max = 4;
 466     }
 467
 468     /* FIXME: support unaligned access? */
 469     access_size = MAX(MIN(size, access_size_max), access_size_min);
 470     access_mask = -1ULL >> (64 - access_size * 8);
 471     if (memory_region_big_endian(mr)) {
 472         for (i = 0; i < size; i += access_size) {
 473             access(mr, addr + i, value, access_size,
 474                    (size - access_size - i) * 8, access_mask);
 475         }
 476     } else {
 477         for (i = 0; i < size; i += access_size) {
 478             access(mr, addr + i, value, access_size, i * 8, access_mask);
 479         }
 480     }
 481 }
 482
 483 static AddressSpace *memory_region_to_address_space(MemoryRegion *mr)
 484 {
 485     AddressSpace *as;
 486
 487     while (mr->parent) {
 488         mr = mr->parent;
 489     }
 490     QTAILQ_FOREACH(as, &address_spaces, address_spaces_link) {
 491         if (mr == as->root) {
 492             return as;
 493         }
 494     }
 495     abort();
 496 }
 497
 498 /* Render a memory region into the global view.  Ranges in @view obscure
 499  * ranges in @mr.
 500  */
 501 static void render_memory_region(FlatView *view,
 502                                  MemoryRegion *mr,
 503                                  Int128 base,
 504                                  AddrRange clip,
 505                                  bool readonly)
 506 {
 507     MemoryRegion *subregion;
 508     unsigned i;
 509     hwaddr offset_in_region;
 510     Int128 remain;
 511     Int128 now;
 512     FlatRange fr;
 513     AddrRange tmp;
 514
 515     if (!mr->enabled) {
 516         return;
 517     }
 518
 519     int128_addto(&base, int128_make64(mr->addr));
 520     readonly |= mr->readonly;
 521
 522     tmp = addrrange_make(base, mr->size);
 523
 524     if (!addrrange_intersects(tmp, clip)) {
 525         return;
 526     }
 527
 528     clip = addrrange_intersection(tmp, clip);
 529
 530     if (mr->alias) {
 531         int128_subfrom(&base, int128_make64(mr->alias->addr));
 532         int128_subfrom(&base, int128_make64(mr->alias_offset));
 533         render_memory_region(view, mr->alias, base, clip, readonly);
 534         return;
 535     }
 536
 537     /* Render subregions in priority order. */
 538     QTAILQ_FOREACH(subregion, &mr->subregions, subregions_link) {
 539         render_memory_region(view, subregion, base, clip, readonly);
 540     }
 541
 542     if (!mr->terminates) {
 543         return;
 544     }
 545
 546     offset_in_region = int128_get64(int128_sub(clip.start, base));
 547     base = clip.start;
 548     remain = clip.size;
 549
 550     fr.mr = mr;
 551     fr.dirty_log_mask = mr->dirty_log_mask;
 552     fr.romd_mode = mr->romd_mode;
 553     fr.readonly = readonly;
 554
 555     /* Render the region itself into any gaps left by the current view. */
 556     for (i = 0; i < view->nr && int128_nz(remain); ++i) {
 557         if (int128_ge(base, addrrange_end(view->ranges[i].addr))) {
 558             continue;
 559         }
 560         if (int128_lt(base, view->ranges[i].addr.start)) {
 561             now = int128_min(remain,
 562                              int128_sub(view->ranges[i].addr.start, base));
 563             fr.offset_in_region = offset_in_region;
 564             fr.addr = addrrange_make(base, now);
 565             flatview_insert(view, i, &fr);
 566             ++i;
 567             int128_addto(&base, now);
 568             offset_in_region += int128_get64(now);
 569             int128_subfrom(&remain, now);
 570         }
 571         now = int128_sub(int128_min(int128_add(base, remain),
 572                                     addrrange_end(view->ranges[i].addr)),
 573                          base);
 574         int128_addto(&base, now);
 575         offset_in_region += int128_get64(now);
 576         int128_subfrom(&remain, now);
 577     }
 578     if (int128_nz(remain)) {
 579         fr.offset_in_region = offset_in_region;
 580         fr.addr = addrrange_make(base, remain);
 581         flatview_insert(view, i, &fr);
 582     }
 583 }
 584
 585 /* Render a memory topology into a list of disjoint absolute ranges. */
 586 static FlatView *generate_memory_topology(MemoryRegion *mr)
 587 {
 588     FlatView *view;
 589
 590     view = g_new(FlatView, 1);
 591     flatview_init(view);
 592
 593     if (mr) {
 594         render_memory_region(view, mr, int128_zero(),
 595                              addrrange_make(int128_zero(), int128_2_64()), false);
 596     }
 597     flatview_simplify(view);
 598
 599     return view;
 600 }
 601
 602 static void address_space_add_del_ioeventfds(AddressSpace *as,
 603                                              MemoryRegionIoeventfd *fds_new,
 604                                              unsigned fds_new_nb,
 605                                              MemoryRegionIoeventfd *fds_old,
 606                                              unsigned fds_old_nb)
 607 {
 608     unsigned iold, inew;
 609     MemoryRegionIoeventfd *fd;
 610     MemoryRegionSection section;
 611
 612     /* Generate a symmetric difference of the old and new fd sets, adding
 613      * and deleting as necessary.
 614      */
 615
 616     iold = inew = 0;
 617     while (iold < fds_old_nb || inew < fds_new_nb) {
 618         if (iold < fds_old_nb
 619             && (inew == fds_new_nb
 620                 || memory_region_ioeventfd_before(fds_old[iold],
 621                                                   fds_new[inew]))) {
 622             fd = &fds_old[iold];
 623             section = (MemoryRegionSection) {
 624                 .address_space = as,
 625                 .offset_within_address_space = int128_get64(fd->addr.start),
 626                 .size = fd->addr.size,
 627             };
 628             MEMORY_LISTENER_CALL(eventfd_del, Forward, &section,
 629                                  fd->match_data, fd->data, fd->e);
 630             ++iold;
 631         } else if (inew < fds_new_nb
 632                    && (iold == fds_old_nb
 633                        || memory_region_ioeventfd_before(fds_new[inew],
 634                                                          fds_old[iold]))) {
 635             fd = &fds_new[inew];
 636             section = (MemoryRegionSection) {
 637                 .address_space = as,
 638                 .offset_within_address_space = int128_get64(fd->addr.start),
 639                 .size = fd->addr.size,
 640             };
 641             MEMORY_LISTENER_CALL(eventfd_add, Reverse, &section,
 642                                  fd->match_data, fd->data, fd->e);
 643             ++inew;
 644         } else {
 645             ++iold;
 646             ++inew;
 647         }
 648     }
 649 }
 650
 651 static FlatView *address_space_get_flatview(AddressSpace *as)
 652 {
 653     FlatView *view;
 654
 655     qemu_mutex_lock(&flat_view_mutex);
 656     view = as->current_map;
 657     flatview_ref(view);
 658     qemu_mutex_unlock(&flat_view_mutex);
 659     return view;
 660 }
 661
 662 static void address_space_update_ioeventfds(AddressSpace *as)
 663 {
 664     FlatView *view;
 665     FlatRange *fr;
 666     unsigned ioeventfd_nb = 0;
 667     MemoryRegionIoeventfd *ioeventfds = NULL;
 668     AddrRange tmp;
 669     unsigned i;
 670
 671     view = address_space_get_flatview(as);
 672     FOR_EACH_FLAT_RANGE(fr, view) {
 673         for (i = 0; i < fr->mr->ioeventfd_nb; ++i) {
 674             tmp = addrrange_shift(fr->mr->ioeventfds[i].addr,
 675                                   int128_sub(fr->addr.start,
 676                                              int128_make64(fr->offset_in_region)));
 677             if (addrrange_intersects(fr->addr, tmp)) {
 678                 ++ioeventfd_nb;
 679                 ioeventfds = g_realloc(ioeventfds,
 680                                           ioeventfd_nb * sizeof(*ioeventfds));
 681                 ioeventfds[ioeventfd_nb-1] = fr->mr->ioeventfds[i];
 682                 ioeventfds[ioeventfd_nb-1].addr = tmp;
 683             }
 684         }
 685     }
 686
 687     address_space_add_del_ioeventfds(as, ioeventfds, ioeventfd_nb,
 688                                      as->ioeventfds, as->ioeventfd_nb);
 689
 690     g_free(as->ioeventfds);
 691     as->ioeventfds = ioeventfds;
 692     as->ioeventfd_nb = ioeventfd_nb;
 693     flatview_unref(view);
 694 }
 695
 696 static void address_space_update_topology_pass(AddressSpace *as,
 697                                                const FlatView *old_view,
 698                                                const FlatView *new_view,
 699                                                bool adding)
 700 {
 701     unsigned iold, inew;
 702     FlatRange *frold, *frnew;
 703
 704     /* Generate a symmetric difference of the old and new memory maps.
 705      * Kill ranges in the old map, and instantiate ranges in the new map.
 706      */
 707     iold = inew = 0;
 708     while (iold < old_view->nr || inew < new_view->nr) {
 709         if (iold < old_view->nr) {
 710             frold = &old_view->ranges[iold];
 711         } else {
 712             frold = NULL;
 713         }
 714         if (inew < new_view->nr) {
 715             frnew = &new_view->ranges[inew];
 716         } else {
 717             frnew = NULL;
 718         }
 719
 720         if (frold
 721             && (!frnew
 722                 || int128_lt(frold->addr.start, frnew->addr.start)
 723                 || (int128_eq(frold->addr.start, frnew->addr.start)
 724                     && !flatrange_equal(frold, frnew)))) {
 725             /* In old but not in new, or in both but attributes changed. */
 726
 727             if (!adding) {
 728                 MEMORY_LISTENER_UPDATE_REGION(frold, as, Reverse, region_del);
 729             }
 730
 731             ++iold;
 732         } else if (frold && frnew && flatrange_equal(frold, frnew)) {
 733             /* In both and unchanged (except logging may have changed) */
 734
 735             if (adding) {
 736                 MEMORY_LISTENER_UPDATE_REGION(frnew, as, Forward, region_nop);
 737                 if (frold->dirty_log_mask && !frnew->dirty_log_mask) {
 738                     MEMORY_LISTENER_UPDATE_REGION(frnew, as, Reverse, log_stop);
 739                 } else if (frnew->dirty_log_mask && !frold->dirty_log_mask) {
 740                     MEMORY_LISTENER_UPDATE_REGION(frnew, as, Forward, log_start);
 741                 }
 742             }
 743
 744             ++iold;
 745             ++inew;
 746         } else {
 747             /* In new */
 748
 749             if (adding) {
 750                 MEMORY_LISTENER_UPDATE_REGION(frnew, as, Forward, region_add);
 751             }
 752
 753             ++inew;
 754         }
 755     }
 756 }
 757
 758
 759 static void address_space_update_topology(AddressSpace *as)
 760 {
 761     FlatView *old_view = address_space_get_flatview(as);
 762     FlatView *new_view = generate_memory_topology(as->root);
 763
 764     address_space_update_topology_pass(as, old_view, new_view, false);
 765     address_space_update_topology_pass(as, old_view, new_view, true);
 766
 767     qemu_mutex_lock(&flat_view_mutex);
 768     flatview_unref(as->current_map);
 769     as->current_map = new_view;
 770     qemu_mutex_unlock(&flat_view_mutex);
 771
 772     /* Note that all the old MemoryRegions are still alive up to this
 773      * point.  This relieves most MemoryListeners from the need to
 774      * ref/unref the MemoryRegions they get---unless they use them
 775      * outside the iothread mutex, in which case precise reference
 776      * counting is necessary.
 777      */
 778     flatview_unref(old_view);
 779
 780     address_space_update_ioeventfds(as);
 781 }
 782
 783 void memory_region_transaction_begin(void)
 784 {
 785     qemu_flush_coalesced_mmio_buffer();
 786     ++memory_region_transaction_depth;
 787 }
 788
 789 void memory_region_transaction_commit(void)
 790 {
 791     AddressSpace *as;
 792
 793     assert(memory_region_transaction_depth);
 794     --memory_region_transaction_depth;
 795     if (!memory_region_transaction_depth && memory_region_update_pending) {
 796         memory_region_update_pending = false;
 797         MEMORY_LISTENER_CALL_GLOBAL(begin, Forward);
 798
 799         QTAILQ_FOREACH(as, &address_spaces, address_spaces_link) {
 800             address_space_update_topology(as);
 801         }
 802
 803         MEMORY_LISTENER_CALL_GLOBAL(commit, Forward);
 804     }
 805 }
 806
 807 static void memory_region_destructor_none(MemoryRegion *mr)
 808 {
 809 }
 810
 811 static void memory_region_destructor_ram(MemoryRegion *mr)
 812 {
 813     qemu_ram_free(mr->ram_addr);
 814 }
 815
 816 static void memory_region_destructor_alias(MemoryRegion *mr)
 817 {
 818     memory_region_unref(mr->alias);
 819 }
 820
 821 static void memory_region_destructor_ram_from_ptr(MemoryRegion *mr)
 822 {
 823     qemu_ram_free_from_ptr(mr->ram_addr);
 824 }
 825
 826 static void memory_region_destructor_rom_device(MemoryRegion *mr)
 827 {
 828     qemu_ram_free(mr->ram_addr & TARGET_PAGE_MASK);
 829 }
 830
 831 void memory_region_init(MemoryRegion *mr,
 832                         Object *owner,
 833                         const char *name,
 834                         uint64_t size)
 835 {
 836     mr->ops = &unassigned_mem_ops;
 837     mr->opaque = NULL;
 838     mr->owner = owner;
 839     mr->iommu_ops = NULL;
 840     mr->parent = NULL;
 841     mr->size = int128_make64(size);
 842     if (size == UINT64_MAX) {
 843         mr->size = int128_2_64();
 844     }
 845     mr->addr = 0;
 846     mr->subpage = false;
 847     mr->enabled = true;
 848     mr->terminates = false;
 849     mr->ram = false;
 850     mr->romd_mode = true;
 851     mr->readonly = false;
 852     mr->rom_device = false;
 853     mr->destructor = memory_region_destructor_none;
 854     mr->priority = 0;
 855     mr->may_overlap = false;
 856     mr->alias = NULL;
 857     QTAILQ_INIT(&mr->subregions);
 858     memset(&mr->subregions_link, 0, sizeof mr->subregions_link);
 859     QTAILQ_INIT(&mr->coalesced);
 860     mr->name = g_strdup(name);
 861     mr->dirty_log_mask = 0;
 862     mr->ioeventfd_nb = 0;
 863     mr->ioeventfds = NULL;
 864     mr->flush_coalesced_mmio = false;
 865 }
 866
 867 static uint64_t unassigned_mem_read(void *opaque, hwaddr addr,
 868                                     unsigned size)
 869 {
 870 #ifdef DEBUG_UNASSIGNED
 871     printf("Unassigned mem read " TARGET_FMT_plx "\n", addr);
 872 #endif
 873     if (current_cpu != NULL) {
 874         cpu_unassigned_access(current_cpu, addr, false, false, 0, size);
 875     }
 876     return -1ULL;
 877 }
 878
 879 static void unassigned_mem_write(void *opaque, hwaddr addr,
 880                                  uint64_t val, unsigned size)
 881 {
 882 #ifdef DEBUG_UNASSIGNED
 883     printf("Unassigned mem write " TARGET_FMT_plx " = 0x%"PRIx64"\n", addr, val);
 884 #endif
 885     if (current_cpu != NULL) {
 886         cpu_unassigned_access(current_cpu, addr, true, false, 0, size);
 887     }
 888 }
 889
 890 static bool unassigned_mem_accepts(void *opaque, hwaddr addr,
 891                                    unsigned size, bool is_write)
 892 {
 893     return false;
 894 }
 895
 896 const MemoryRegionOps unassigned_mem_ops = {
 897     .valid.accepts = unassigned_mem_accepts,
 898     .endianness = DEVICE_NATIVE_ENDIAN,
 899 };
 900
 901 bool memory_region_access_valid(MemoryRegion *mr,
 902                                 hwaddr addr,
 903                                 unsigned size,
 904                                 bool is_write)
 905 {
 906     int access_size_min, access_size_max;
 907     int access_size, i;
 908
 909     if (!mr->ops->valid.unaligned && (addr & (size - 1))) {
 910         return false;
 911     }
 912
 913     if (!mr->ops->valid.accepts) {
 914         return true;
 915     }
 916
 917     access_size_min = mr->ops->valid.min_access_size;
 918     if (!mr->ops->valid.min_access_size) {
 919         access_size_min = 1;
 920     }
 921
 922     access_size_max = mr->ops->valid.max_access_size;
 923     if (!mr->ops->valid.max_access_size) {
 924         access_size_max = 4;
 925     }
 926
 927     access_size = MAX(MIN(size, access_size_max), access_size_min);
 928     for (i = 0; i < size; i += access_size) {
 929         if (!mr->ops->valid.accepts(mr->opaque, addr + i, access_size,
 930                                     is_write)) {
 931             return false;
 932         }
 933     }
 934
 935     return true;
 936 }
 937
 938 static uint64_t memory_region_dispatch_read1(MemoryRegion *mr,
 939                                              hwaddr addr,
 940                                              unsigned size)
 941 {
 942     uint64_t data = 0;
 943
 944     if (mr->ops->read) {
 945         access_with_adjusted_size(addr, &data, size,
 946                                   mr->ops->impl.min_access_size,
 947                                   mr->ops->impl.max_access_size,
 948                                   memory_region_read_accessor, mr);
 949     } else {
 950         access_with_adjusted_size(addr, &data, size, 1, 4,
 951                                   memory_region_oldmmio_read_accessor, mr);
 952     }
 953
 954     return data;
 955 }
 956
 957 static bool memory_region_dispatch_read(MemoryRegion *mr,
 958                                         hwaddr addr,
 959                                         uint64_t *pval,
 960                                         unsigned size)
 961 {
 962     if (!memory_region_access_valid(mr, addr, size, false)) {
 963         *pval = unassigned_mem_read(mr, addr, size);
 964         return true;
 965     }
 966
 967     *pval = memory_region_dispatch_read1(mr, addr, size);
 968     adjust_endianness(mr, pval, size);
 969     return false;
 970 }
 971
 972 static bool memory_region_dispatch_write(MemoryRegion *mr,
 973                                          hwaddr addr,
 974                                          uint64_t data,
 975                                          unsigned size)
 976 {
 977     if (!memory_region_access_valid(mr, addr, size, true)) {
 978         unassigned_mem_write(mr, addr, data, size);
 979         return true;
 980     }
 981
 982     adjust_endianness(mr, &data, size);
 983
 984     if (mr->ops->write) {
 985         access_with_adjusted_size(addr, &data, size,
 986                                   mr->ops->impl.min_access_size,
 987                                   mr->ops->impl.max_access_size,
 988                                   memory_region_write_accessor, mr);
 989     } else {
 990         access_with_adjusted_size(addr, &data, size, 1, 4,
 991                                   memory_region_oldmmio_write_accessor, mr);
 992     }
 993     return false;
 994 }
 995
 996 void memory_region_init_io(MemoryRegion *mr,
 997                            Object *owner,
 998                            const MemoryRegionOps *ops,
 999                            void *opaque,
1000                            const char *name,
1001                            uint64_t size)
1002 {
1003     memory_region_init(mr, owner, name, size);
1004     mr->ops = ops;
1005     mr->opaque = opaque;
1006     mr->terminates = true;
1007     mr->ram_addr = ~(ram_addr_t)0;
1008 }
1009
1010 void memory_region_init_ram(MemoryRegion *mr,
1011                             Object *owner,
1012                             const char *name,
1013                             uint64_t size)
1014 {
1015     memory_region_init(mr, owner, name, size);
1016     mr->ram = true;
1017     mr->terminates = true;
1018     mr->destructor = memory_region_destructor_ram;
1019     mr->ram_addr = qemu_ram_alloc(size, mr);
1020 }
1021
1022 void memory_region_init_ram_ptr(MemoryRegion *mr,
1023                                 Object *owner,
1024                                 const char *name,
1025                                 uint64_t size,
1026                                 void *ptr)
1027 {
1028     memory_region_init(mr, owner, name, size);
1029     mr->ram = true;
1030     mr->terminates = true;
1031     mr->destructor = memory_region_destructor_ram_from_ptr;
1032     mr->ram_addr = qemu_ram_alloc_from_ptr(size, ptr, mr);
1033 }
1034
1035 void memory_region_init_alias(MemoryRegion *mr,
1036                               Object *owner,
1037                               const char *name,
1038                               MemoryRegion *orig,
1039                               hwaddr offset,
1040                               uint64_t size)
1041 {
1042     memory_region_init(mr, owner, name, size);
1043     memory_region_ref(orig);
1044     mr->destructor = memory_region_destructor_alias;
1045     mr->alias = orig;
1046     mr->alias_offset = offset;
1047 }
1048
1049 void memory_region_init_rom_device(MemoryRegion *mr,
1050                                    Object *owner,
1051                                    const MemoryRegionOps *ops,
1052                                    void *opaque,
1053                                    const char *name,
1054                                    uint64_t size)
1055 {
1056     memory_region_init(mr, owner, name, size);
1057     mr->ops = ops;
1058     mr->opaque = opaque;
1059     mr->terminates = true;
1060     mr->rom_device = true;
1061     mr->destructor = memory_region_destructor_rom_device;
1062     mr->ram_addr = qemu_ram_alloc(size, mr);
1063 }
1064
1065 void memory_region_init_iommu(MemoryRegion *mr,
1066                               Object *owner,
1067                               const MemoryRegionIOMMUOps *ops,
1068                               const char *name,
1069                               uint64_t size)
1070 {
1071     memory_region_init(mr, owner, name, size);
1072     mr->iommu_ops = ops,
1073     mr->terminates = true;  /* then re-forwards */
1074     notifier_list_init(&mr->iommu_notify);
1075 }
1076
1077 void memory_region_init_reservation(MemoryRegion *mr,
1078                                     Object *owner,
1079                                     const char *name,
1080                                     uint64_t size)
1081 {
1082     memory_region_init_io(mr, owner, &unassigned_mem_ops, mr, name, size);
1083 }
1084
1085 void memory_region_destroy(MemoryRegion *mr)
1086 {
1087     assert(QTAILQ_EMPTY(&mr->subregions));
1088     assert(memory_region_transaction_depth == 0);
1089     mr->destructor(mr);
1090     memory_region_clear_coalescing(mr);
1091     g_free((char *)mr->name);
1092     g_free(mr->ioeventfds);
1093 }
1094
1095 Object *memory_region_owner(MemoryRegion *mr)
1096 {
1097     return mr->owner;
1098 }
1099
1100 void memory_region_ref(MemoryRegion *mr)
1101 {
1102     if (mr && mr->owner) {
1103         object_ref(mr->owner);
1104     }
1105 }
1106
1107 void memory_region_unref(MemoryRegion *mr)
1108 {
1109     if (mr && mr->owner) {
1110         object_unref(mr->owner);
1111     }
1112 }
1113
1114 uint64_t memory_region_size(MemoryRegion *mr)
1115 {
1116     if (int128_eq(mr->size, int128_2_64())) {
1117         return UINT64_MAX;
1118     }
1119     return int128_get64(mr->size);
1120 }
1121
1122 const char *memory_region_name(MemoryRegion *mr)
1123 {
1124     return mr->name;
1125 }
1126
1127 bool memory_region_is_ram(MemoryRegion *mr)
1128 {
1129     return mr->ram;
1130 }
1131
1132 bool memory_region_is_logging(MemoryRegion *mr)
1133 {
1134     return mr->dirty_log_mask;
1135 }
1136
1137 bool memory_region_is_rom(MemoryRegion *mr)
1138 {
1139     return mr->ram && mr->readonly;
1140 }
1141
1142 bool memory_region_is_iommu(MemoryRegion *mr)
1143 {
1144     return mr->iommu_ops;
1145 }
1146
1147 void memory_region_register_iommu_notifier(MemoryRegion *mr, Notifier *n)
1148 {
1149     notifier_list_add(&mr->iommu_notify, n);
1150 }
1151
1152 void memory_region_unregister_iommu_notifier(Notifier *n)
1153 {
1154     notifier_remove(n);
1155 }
1156
1157 void memory_region_notify_iommu(MemoryRegion *mr,
1158                                 IOMMUTLBEntry entry)
1159 {
1160     assert(memory_region_is_iommu(mr));
1161     notifier_list_notify(&mr->iommu_notify, &entry);
1162 }
1163
1164 void memory_region_set_log(MemoryRegion *mr, bool log, unsigned client)
1165 {
1166     uint8_t mask = 1 << client;
1167
1168     memory_region_transaction_begin();
1169     mr->dirty_log_mask = (mr->dirty_log_mask & ~mask) | (log * mask);
1170     memory_region_update_pending |= mr->enabled;
1171     memory_region_transaction_commit();
1172 }
1173
1174 bool memory_region_get_dirty(MemoryRegion *mr, hwaddr addr,
1175                              hwaddr size, unsigned client)
1176 {
1177     assert(mr->terminates);
1178     return cpu_physical_memory_get_dirty(mr->ram_addr + addr, size,
1179                                          1 << client);
1180 }
1181
1182 void memory_region_set_dirty(MemoryRegion *mr, hwaddr addr,
1183                              hwaddr size)
1184 {
1185     assert(mr->terminates);
1186     return cpu_physical_memory_set_dirty_range(mr->ram_addr + addr, size, -1);
1187 }
1188
1189 bool memory_region_test_and_clear_dirty(MemoryRegion *mr, hwaddr addr,
1190                                         hwaddr size, unsigned client)
1191 {
1192     bool ret;
1193     assert(mr->terminates);
1194     ret = cpu_physical_memory_get_dirty(mr->ram_addr + addr, size,
1195                                         1 << client);
1196     if (ret) {
1197         cpu_physical_memory_reset_dirty(mr->ram_addr + addr,
1198                                         mr->ram_addr + addr + size,
1199                                         1 << client);
1200     }
1201     return ret;
1202 }
1203
1204
1205 void memory_region_sync_dirty_bitmap(MemoryRegion *mr)
1206 {
1207     AddressSpace *as;
1208     FlatRange *fr;
1209
1210     QTAILQ_FOREACH(as, &address_spaces, address_spaces_link) {
1211         FlatView *view = address_space_get_flatview(as);
1212         FOR_EACH_FLAT_RANGE(fr, view) {
1213             if (fr->mr == mr) {
1214                 MEMORY_LISTENER_UPDATE_REGION(fr, as, Forward, log_sync);
1215             }
1216         }
1217         flatview_unref(view);
1218     }
1219 }
1220
1221 void memory_region_set_readonly(MemoryRegion *mr, bool readonly)
1222 {
1223     if (mr->readonly != readonly) {
1224         memory_region_transaction_begin();
1225         mr->readonly = readonly;
1226         memory_region_update_pending |= mr->enabled;
1227         memory_region_transaction_commit();
1228     }
1229 }
1230
1231 void memory_region_rom_device_set_romd(MemoryRegion *mr, bool romd_mode)
1232 {
1233     if (mr->romd_mode != romd_mode) {
1234         memory_region_transaction_begin();
1235         mr->romd_mode = romd_mode;
1236         memory_region_update_pending |= mr->enabled;
1237         memory_region_transaction_commit();
1238     }
1239 }
1240
1241 void memory_region_reset_dirty(MemoryRegion *mr, hwaddr addr,
1242                                hwaddr size, unsigned client)
1243 {
1244     assert(mr->terminates);
1245     cpu_physical_memory_reset_dirty(mr->ram_addr + addr,
1246                                     mr->ram_addr + addr + size,
1247                                     1 << client);
1248 }
1249
1250 void *memory_region_get_ram_ptr(MemoryRegion *mr)
1251 {
1252     if (mr->alias) {
1253         return memory_region_get_ram_ptr(mr->alias) + mr->alias_offset;
1254     }
1255
1256     assert(mr->terminates);
1257
1258     return qemu_get_ram_ptr(mr->ram_addr & TARGET_PAGE_MASK);
1259 }
1260
1261 static void memory_region_update_coalesced_range_as(MemoryRegion *mr, AddressSpace *as)
1262 {
1263     FlatView *view;
1264     FlatRange *fr;
1265     CoalescedMemoryRange *cmr;
1266     AddrRange tmp;
1267     MemoryRegionSection section;
1268
1269     view = address_space_get_flatview(as);
1270     FOR_EACH_FLAT_RANGE(fr, view) {
1271         if (fr->mr == mr) {
1272             section = (MemoryRegionSection) {
1273                 .address_space = as,
1274                 .offset_within_address_space = int128_get64(fr->addr.start),
1275                 .size = fr->addr.size,
1276             };
1277
1278             MEMORY_LISTENER_CALL(coalesced_mmio_del, Reverse, &section,
1279                                  int128_get64(fr->addr.start),
1280                                  int128_get64(fr->addr.size));
1281             QTAILQ_FOREACH(cmr, &mr->coalesced, link) {
1282                 tmp = addrrange_shift(cmr->addr,
1283                                       int128_sub(fr->addr.start,
1284                                                  int128_make64(fr->offset_in_region)));
1285                 if (!addrrange_intersects(tmp, fr->addr)) {
1286                     continue;
1287                 }
1288                 tmp = addrrange_intersection(tmp, fr->addr);
1289                 MEMORY_LISTENER_CALL(coalesced_mmio_add, Forward, &section,
1290                                      int128_get64(tmp.start),
1291                                      int128_get64(tmp.size));
1292             }
1293         }
1294     }
1295     flatview_unref(view);
1296 }
1297
1298 static void memory_region_update_coalesced_range(MemoryRegion *mr)
1299 {
1300     AddressSpace *as;
1301
1302     QTAILQ_FOREACH(as, &address_spaces, address_spaces_link) {
1303         memory_region_update_coalesced_range_as(mr, as);
1304     }
1305 }
1306
1307 void memory_region_set_coalescing(MemoryRegion *mr)
1308 {
1309     memory_region_clear_coalescing(mr);
1310     memory_region_add_coalescing(mr, 0, int128_get64(mr->size));
1311 }
1312
1313 void memory_region_add_coalescing(MemoryRegion *mr,
1314                                   hwaddr offset,
1315                                   uint64_t size)
1316 {
1317     CoalescedMemoryRange *cmr = g_malloc(sizeof(*cmr));
1318
1319     cmr->addr = addrrange_make(int128_make64(offset), int128_make64(size));
1320     QTAILQ_INSERT_TAIL(&mr->coalesced, cmr, link);
1321     memory_region_update_coalesced_range(mr);
1322     memory_region_set_flush_coalesced(mr);
1323 }
1324
1325 void memory_region_clear_coalescing(MemoryRegion *mr)
1326 {
1327     CoalescedMemoryRange *cmr;
1328
1329     qemu_flush_coalesced_mmio_buffer();
1330     mr->flush_coalesced_mmio = false;
1331
1332     while (!QTAILQ_EMPTY(&mr->coalesced)) {
1333         cmr = QTAILQ_FIRST(&mr->coalesced);
1334         QTAILQ_REMOVE(&mr->coalesced, cmr, link);
1335         g_free(cmr);
1336     }
1337     memory_region_update_coalesced_range(mr);
1338 }
1339
1340 void memory_region_set_flush_coalesced(MemoryRegion *mr)
1341 {
1342     mr->flush_coalesced_mmio = true;
1343 }
1344
1345 void memory_region_clear_flush_coalesced(MemoryRegion *mr)
1346 {
1347     qemu_flush_coalesced_mmio_buffer();
1348     if (QTAILQ_EMPTY(&mr->coalesced)) {
1349         mr->flush_coalesced_mmio = false;
1350     }
1351 }
1352
1353 void memory_region_add_eventfd(MemoryRegion *mr,
1354                                hwaddr addr,
1355                                unsigned size,
1356                                bool match_data,
1357                                uint64_t data,
1358                                EventNotifier *e)
1359 {
1360     MemoryRegionIoeventfd mrfd = {
1361         .addr.start = int128_make64(addr),
1362         .addr.size = int128_make64(size),
1363         .match_data = match_data,
1364         .data = data,
1365         .e = e,
1366     };
1367     unsigned i;
1368
1369     adjust_endianness(mr, &mrfd.data, size);
1370     memory_region_transaction_begin();
1371     for (i = 0; i < mr->ioeventfd_nb; ++i) {
1372         if (memory_region_ioeventfd_before(mrfd, mr->ioeventfds[i])) {
1373             break;
1374         }
1375     }
1376     ++mr->ioeventfd_nb;
1377     mr->ioeventfds = g_realloc(mr->ioeventfds,
1378                                   sizeof(*mr->ioeventfds) * mr->ioeventfd_nb);
1379     memmove(&mr->ioeventfds[i+1], &mr->ioeventfds[i],
1380             sizeof(*mr->ioeventfds) * (mr->ioeventfd_nb-1 - i));
1381     mr->ioeventfds[i] = mrfd;
1382     memory_region_update_pending |= mr->enabled;
1383     memory_region_transaction_commit();
1384 }
1385
1386 void memory_region_del_eventfd(MemoryRegion *mr,
1387                                hwaddr addr,
1388                                unsigned size,
1389                                bool match_data,
1390                                uint64_t data,
1391                                EventNotifier *e)
1392 {
1393     MemoryRegionIoeventfd mrfd = {
1394         .addr.start = int128_make64(addr),
1395         .addr.size = int128_make64(size),
1396         .match_data = match_data,
1397         .data = data,
1398         .e = e,
1399     };
1400     unsigned i;
1401
1402     adjust_endianness(mr, &mrfd.data, size);
1403     memory_region_transaction_begin();
1404     for (i = 0; i < mr->ioeventfd_nb; ++i) {
1405         if (memory_region_ioeventfd_equal(mrfd, mr->ioeventfds[i])) {
1406             break;
1407         }
1408     }
1409     assert(i != mr->ioeventfd_nb);
1410     memmove(&mr->ioeventfds[i], &mr->ioeventfds[i+1],
1411             sizeof(*mr->ioeventfds) * (mr->ioeventfd_nb - (i+1)));
1412     --mr->ioeventfd_nb;
1413     mr->ioeventfds = g_realloc(mr->ioeventfds,
1414                                   sizeof(*mr->ioeventfds)*mr->ioeventfd_nb + 1);
1415     memory_region_update_pending |= mr->enabled;
1416     memory_region_transaction_commit();
1417 }
1418
1419 static void memory_region_add_subregion_common(MemoryRegion *mr,
1420                                                hwaddr offset,
1421                                                MemoryRegion *subregion)
1422 {
1423     MemoryRegion *other;
1424
1425     memory_region_transaction_begin();
1426
1427     assert(!subregion->parent);
1428     memory_region_ref(subregion);
1429     subregion->parent = mr;
1430     subregion->addr = offset;
1431     QTAILQ_FOREACH(other, &mr->subregions, subregions_link) {
1432         if (subregion->may_overlap || other->may_overlap) {
1433             continue;
1434         }
1435         if (int128_ge(int128_make64(offset),
1436                       int128_add(int128_make64(other->addr), other->size))
1437             || int128_le(int128_add(int128_make64(offset), subregion->size),
1438                          int128_make64(other->addr))) {
1439             continue;
1440         }
1441 #if 0
1442         printf("warning: subregion collision %llx/%llx (%s) "
1443                "vs %llx/%llx (%s)\n",
1444                (unsigned long long)offset,
1445                (unsigned long long)int128_get64(subregion->size),
1446                subregion->name,
1447                (unsigned long long)other->addr,
1448                (unsigned long long)int128_get64(other->size),
1449                other->name);
1450 #endif
1451     }
1452     QTAILQ_FOREACH(other, &mr->subregions, subregions_link) {
1453         if (subregion->priority >= other->priority) {
1454             QTAILQ_INSERT_BEFORE(other, subregion, subregions_link);
1455             goto done;
1456         }
1457     }
1458     QTAILQ_INSERT_TAIL(&mr->subregions, subregion, subregions_link);
1459 done:
1460     memory_region_update_pending |= mr->enabled && subregion->enabled;
1461     memory_region_transaction_commit();
1462 }
1463
1464
1465 void memory_region_add_subregion(MemoryRegion *mr,
1466                                  hwaddr offset,
1467                                  MemoryRegion *subregion)
1468 {
1469     subregion->may_overlap = false;
1470     subregion->priority = 0;
1471     memory_region_add_subregion_common(mr, offset, subregion);
1472 }
1473
1474 void memory_region_add_subregion_overlap(MemoryRegion *mr,
1475                                          hwaddr offset,
1476                                          MemoryRegion *subregion,
1477                                          unsigned priority)
1478 {
1479     subregion->may_overlap = true;
1480     subregion->priority = priority;
1481     memory_region_add_subregion_common(mr, offset, subregion);
1482 }
1483
1484 void memory_region_del_subregion(MemoryRegion *mr,
1485                                  MemoryRegion *subregion)
1486 {
1487     memory_region_transaction_begin();
1488     assert(subregion->parent == mr);
1489     subregion->parent = NULL;
1490     QTAILQ_REMOVE(&mr->subregions, subregion, subregions_link);
1491     memory_region_unref(subregion);
1492     memory_region_update_pending |= mr->enabled && subregion->enabled;
1493     memory_region_transaction_commit();
1494 }
1495
1496 void memory_region_set_enabled(MemoryRegion *mr, bool enabled)
1497 {
1498     if (enabled == mr->enabled) {
1499         return;
1500     }
1501     memory_region_transaction_begin();
1502     mr->enabled = enabled;
1503     memory_region_update_pending = true;
1504     memory_region_transaction_commit();
1505 }
1506
1507 void memory_region_set_address(MemoryRegion *mr, hwaddr addr)
1508 {
1509     MemoryRegion *parent = mr->parent;
1510     unsigned priority = mr->priority;
1511     bool may_overlap = mr->may_overlap;
1512
1513     if (addr == mr->addr || !parent) {
1514         mr->addr = addr;
1515         return;
1516     }
1517
1518     memory_region_transaction_begin();
1519     memory_region_ref(mr);
1520     memory_region_del_subregion(parent, mr);
1521     if (may_overlap) {
1522         memory_region_add_subregion_overlap(parent, addr, mr, priority);
1523     } else {
1524         memory_region_add_subregion(parent, addr, mr);
1525     }
1526     memory_region_unref(mr);
1527     memory_region_transaction_commit();
1528 }
1529
1530 void memory_region_set_alias_offset(MemoryRegion *mr, hwaddr offset)
1531 {
1532     assert(mr->alias);
1533
1534     if (offset == mr->alias_offset) {
1535         return;
1536     }
1537
1538     memory_region_transaction_begin();
1539     mr->alias_offset = offset;
1540     memory_region_update_pending |= mr->enabled;
1541     memory_region_transaction_commit();
1542 }
1543
1544 ram_addr_t memory_region_get_ram_addr(MemoryRegion *mr)
1545 {
1546     return mr->ram_addr;
1547 }
1548
1549 static int cmp_flatrange_addr(const void *addr_, const void *fr_)
1550 {
1551     const AddrRange *addr = addr_;
1552     const FlatRange *fr = fr_;
1553
1554     if (int128_le(addrrange_end(*addr), fr->addr.start)) {
1555         return -1;
1556     } else if (int128_ge(addr->start, addrrange_end(fr->addr))) {
1557         return 1;
1558     }
1559     return 0;
1560 }
1561
1562 static FlatRange *flatview_lookup(FlatView *view, AddrRange addr)
1563 {
1564     return bsearch(&addr, view->ranges, view->nr,
1565                    sizeof(FlatRange), cmp_flatrange_addr);
1566 }
1567
1568 bool memory_region_present(MemoryRegion *parent, hwaddr addr)
1569 {
1570     MemoryRegion *mr = memory_region_find(parent, addr, 1).mr;
1571     if (!mr) {
1572         return false;
1573     }
1574     memory_region_unref(mr);
1575     return true;
1576 }
1577
1578 MemoryRegionSection memory_region_find(MemoryRegion *mr,
1579                                        hwaddr addr, uint64_t size)
1580 {
1581     MemoryRegionSection ret = { .mr = NULL };
1582     MemoryRegion *root;
1583     AddressSpace *as;
1584     AddrRange range;
1585     FlatView *view;
1586     FlatRange *fr;
1587
1588     addr += mr->addr;
1589     for (root = mr; root->parent; ) {
1590         root = root->parent;
1591         addr += root->addr;
1592     }
1593
1594     as = memory_region_to_address_space(root);
1595     range = addrrange_make(int128_make64(addr), int128_make64(size));
1596
1597     view = address_space_get_flatview(as);
1598     fr = flatview_lookup(view, range);
1599     if (!fr) {
1600         return ret;
1601     }
1602
1603     while (fr > view->ranges && addrrange_intersects(fr[-1].addr, range)) {
1604         --fr;
1605     }
1606
1607     ret.mr = fr->mr;
1608     ret.address_space = as;
1609     range = addrrange_intersection(range, fr->addr);
1610     ret.offset_within_region = fr->offset_in_region;
1611     ret.offset_within_region += int128_get64(int128_sub(range.start,
1612                                                         fr->addr.start));
1613     ret.size = range.size;
1614     ret.offset_within_address_space = int128_get64(range.start);
1615     ret.readonly = fr->readonly;
1616     memory_region_ref(ret.mr);
1617
1618     flatview_unref(view);
1619     return ret;
1620 }
1621
1622 void address_space_sync_dirty_bitmap(AddressSpace *as)
1623 {
1624     FlatView *view;
1625     FlatRange *fr;
1626
1627     view = address_space_get_flatview(as);
1628     FOR_EACH_FLAT_RANGE(fr, view) {
1629         MEMORY_LISTENER_UPDATE_REGION(fr, as, Forward, log_sync);
1630     }
1631     flatview_unref(view);
1632 }
1633
1634 void memory_global_dirty_log_start(void)
1635 {
1636     global_dirty_log = true;
1637     MEMORY_LISTENER_CALL_GLOBAL(log_global_start, Forward);
1638 }
1639
1640 void memory_global_dirty_log_stop(void)
1641 {
1642     global_dirty_log = false;
1643     MEMORY_LISTENER_CALL_GLOBAL(log_global_stop, Reverse);
1644 }
1645
1646 static void listener_add_address_space(MemoryListener *listener,
1647                                        AddressSpace *as)
1648 {
1649     FlatView *view;
1650     FlatRange *fr;
1651
1652     if (listener->address_space_filter
1653         && listener->address_space_filter != as) {
1654         return;
1655     }
1656
1657     if (global_dirty_log) {
1658         if (listener->log_global_start) {
1659             listener->log_global_start(listener);
1660         }
1661     }
1662
1663     view = address_space_get_flatview(as);
1664     FOR_EACH_FLAT_RANGE(fr, view) {
1665         MemoryRegionSection section = {
1666             .mr = fr->mr,
1667             .address_space = as,
1668             .offset_within_region = fr->offset_in_region,
1669             .size = fr->addr.size,
1670             .offset_within_address_space = int128_get64(fr->addr.start),
1671             .readonly = fr->readonly,
1672         };
1673         if (listener->region_add) {
1674             listener->region_add(listener, &section);
1675         }
1676     }
1677     flatview_unref(view);
1678 }
1679
1680 void memory_listener_register(MemoryListener *listener, AddressSpace *filter)
1681 {
1682     MemoryListener *other = NULL;
1683     AddressSpace *as;
1684
1685     listener->address_space_filter = filter;
1686     if (QTAILQ_EMPTY(&memory_listeners)
1687         || listener->priority >= QTAILQ_LAST(&memory_listeners,
1688                                              memory_listeners)->priority) {
1689         QTAILQ_INSERT_TAIL(&memory_listeners, listener, link);
1690     } else {
1691         QTAILQ_FOREACH(other, &memory_listeners, link) {
1692             if (listener->priority < other->priority) {
1693                 break;
1694             }
1695         }
1696         QTAILQ_INSERT_BEFORE(other, listener, link);
1697     }
1698
1699     QTAILQ_FOREACH(as, &address_spaces, address_spaces_link) {
1700         listener_add_address_space(listener, as);
1701     }
1702 }
1703
1704 void memory_listener_unregister(MemoryListener *listener)
1705 {
1706     QTAILQ_REMOVE(&memory_listeners, listener, link);
1707 }
1708
1709 void address_space_init(AddressSpace *as, MemoryRegion *root, const char *name)
1710 {
1711     if (QTAILQ_EMPTY(&address_spaces)) {
1712         memory_init();
1713     }
1714
1715     memory_region_transaction_begin();
1716     as->root = root;
1717     as->current_map = g_new(FlatView, 1);
1718     flatview_init(as->current_map);
1719     as->ioeventfd_nb = 0;
1720     as->ioeventfds = NULL;
1721     QTAILQ_INSERT_TAIL(&address_spaces, as, address_spaces_link);
1722     as->name = g_strdup(name ? name : "anonymous");
1723     address_space_init_dispatch(as);
1724     memory_region_update_pending |= root->enabled;
1725     memory_region_transaction_commit();
1726 }
1727
1728 void address_space_destroy(AddressSpace *as)
1729 {
1730     /* Flush out anything from MemoryListeners listening in on this */
1731     memory_region_transaction_begin();
1732     as->root = NULL;
1733     memory_region_transaction_commit();
1734     QTAILQ_REMOVE(&address_spaces, as, address_spaces_link);
1735     address_space_destroy_dispatch(as);
1736     flatview_unref(as->current_map);
1737     g_free(as->name);
1738     g_free(as->ioeventfds);
1739 }
1740
1741 bool io_mem_read(MemoryRegion *mr, hwaddr addr, uint64_t *pval, unsigned size)
1742 {
1743     return memory_region_dispatch_read(mr, addr, pval, size);
1744 }
1745
1746 bool io_mem_write(MemoryRegion *mr, hwaddr addr,
1747                   uint64_t val, unsigned size)
1748 {
1749     return memory_region_dispatch_write(mr, addr, val, size);
1750 }
1751
1752 typedef struct MemoryRegionList MemoryRegionList;
1753
1754 struct MemoryRegionList {
1755     const MemoryRegion *mr;
1756     bool printed;
1757     QTAILQ_ENTRY(MemoryRegionList) queue;
1758 };
1759
1760 typedef QTAILQ_HEAD(queue, MemoryRegionList) MemoryRegionListHead;
1761
1762 static void mtree_print_mr(fprintf_function mon_printf, void *f,
1763                            const MemoryRegion *mr, unsigned int level,
1764                            hwaddr base,
1765                            MemoryRegionListHead *alias_print_queue)
1766 {
1767     MemoryRegionList *new_ml, *ml, *next_ml;
1768     MemoryRegionListHead submr_print_queue;
1769     const MemoryRegion *submr;
1770     unsigned int i;
1771
1772     if (!mr || !mr->enabled) {
1773         return;
1774     }
1775
1776     for (i = 0; i < level; i++) {
1777         mon_printf(f, "  ");
1778     }
1779
1780     if (mr->alias) {
1781         MemoryRegionList *ml;
1782         bool found = false;
1783
1784         /* check if the alias is already in the queue */
1785         QTAILQ_FOREACH(ml, alias_print_queue, queue) {
1786             if (ml->mr == mr->alias && !ml->printed) {
1787                 found = true;
1788             }
1789         }
1790
1791         if (!found) {
1792             ml = g_new(MemoryRegionList, 1);
1793             ml->mr = mr->alias;
1794             ml->printed = false;
1795             QTAILQ_INSERT_TAIL(alias_print_queue, ml, queue);
1796         }
1797         mon_printf(f, TARGET_FMT_plx "-" TARGET_FMT_plx
1798                    " (prio %d, %c%c): alias %s @%s " TARGET_FMT_plx
1799                    "-" TARGET_FMT_plx "\n",
1800                    base + mr->addr,
1801                    base + mr->addr
1802                    + (int128_nz(mr->size) ?
1803                       (hwaddr)int128_get64(int128_sub(mr->size,
1804                                                       int128_one())) : 0),
1805                    mr->priority,
1806                    mr->romd_mode ? 'R' : '-',
1807                    !mr->readonly && !(mr->rom_device && mr->romd_mode) ? 'W'
1808                                                                        : '-',
1809                    mr->name,
1810                    mr->alias->name,
1811                    mr->alias_offset,
1812                    mr->alias_offset
1813                    + (hwaddr)int128_get64(mr->size) - 1);
1814     } else {
1815         mon_printf(f,
1816                    TARGET_FMT_plx "-" TARGET_FMT_plx " (prio %d, %c%c): %s\n",
1817                    base + mr->addr,
1818                    base + mr->addr
1819                    + (int128_nz(mr->size) ?
1820                       (hwaddr)int128_get64(int128_sub(mr->size,
1821                                                       int128_one())) : 0),
1822                    mr->priority,
1823                    mr->romd_mode ? 'R' : '-',
1824                    !mr->readonly && !(mr->rom_device && mr->romd_mode) ? 'W'
1825                                                                        : '-',
1826                    mr->name);
1827     }
1828
1829     QTAILQ_INIT(&submr_print_queue);
1830
1831     QTAILQ_FOREACH(submr, &mr->subregions, subregions_link) {
1832         new_ml = g_new(MemoryRegionList, 1);
1833         new_ml->mr = submr;
1834         QTAILQ_FOREACH(ml, &submr_print_queue, queue) {
1835             if (new_ml->mr->addr < ml->mr->addr ||
1836                 (new_ml->mr->addr == ml->mr->addr &&
1837                  new_ml->mr->priority > ml->mr->priority)) {
1838                 QTAILQ_INSERT_BEFORE(ml, new_ml, queue);
1839                 new_ml = NULL;
1840                 break;
1841             }
1842         }
1843         if (new_ml) {
1844             QTAILQ_INSERT_TAIL(&submr_print_queue, new_ml, queue);
1845         }
1846     }
1847
1848     QTAILQ_FOREACH(ml, &submr_print_queue, queue) {
1849         mtree_print_mr(mon_printf, f, ml->mr, level + 1, base + mr->addr,
1850                        alias_print_queue);
1851     }
1852
1853     QTAILQ_FOREACH_SAFE(ml, &submr_print_queue, queue, next_ml) {
1854         g_free(ml);
1855     }
1856 }
1857
1858 void mtree_info(fprintf_function mon_printf, void *f)
1859 {
1860     MemoryRegionListHead ml_head;
1861     MemoryRegionList *ml, *ml2;
1862     AddressSpace *as;
1863
1864     QTAILQ_INIT(&ml_head);
1865
1866     QTAILQ_FOREACH(as, &address_spaces, address_spaces_link) {
1867         mon_printf(f, "%s\n", as->name);
1868         mtree_print_mr(mon_printf, f, as->root, 0, 0, &ml_head);
1869     }
1870
1871     mon_printf(f, "aliases\n");
1872     /* print aliased regions */
1873     QTAILQ_FOREACH(ml, &ml_head, queue) {
1874         if (!ml->printed) {
1875             mon_printf(f, "%s\n", ml->mr->name);
1876             mtree_print_mr(mon_printf, f, ml->mr, 0, 0, &ml_head);
1877         }
1878     }
1879
1880     QTAILQ_FOREACH_SAFE(ml, &ml_head, queue, ml2) {
1881         g_free(ml);
1882     }
1883 }