memory.c

   1 /*
   2  * Physical memory management
   3  *
   4  * Copyright 2011 Red Hat, Inc. and/or its affiliates
   5  *
   6  * Authors:
   7  *  Avi Kivity <avi@redhat.com>
   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 "trace.h"
  22 #include <assert.h>
  23
  24 #include "exec/memory-internal.h"
  25 #include "exec/ram_addr.h"
  26
  27 //#define DEBUG_UNASSIGNED
  28
  29 static unsigned memory_region_transaction_depth;
  30 static bool memory_region_update_pending;
  31 static bool ioeventfd_update_pending;
  32 static bool global_dirty_log = false;
  33
  34 /* flat_view_mutex is taken around reading as->current_map; the critical
  35  * section is extremely short, so I'm using a single mutex for every AS.
  36  * We could also RCU for the read-side.
  37  *
  38  * The BQL is taken around transaction commits, hence both locks are taken
  39  * while writing to as->current_map (with the BQL taken outside).
  40  */
  41 static QemuMutex flat_view_mutex;
  42
  43 static QTAILQ_HEAD(memory_listeners, MemoryListener) memory_listeners
  44     = QTAILQ_HEAD_INITIALIZER(memory_listeners);
  45
  46 static QTAILQ_HEAD(, AddressSpace) address_spaces
  47     = QTAILQ_HEAD_INITIALIZER(address_spaces);
  48
  49 static void memory_init(void)
  50 {
  51     qemu_mutex_init(&flat_view_mutex);
  52 }
  53
  54 typedef struct AddrRange AddrRange;
  55
  56 /*
  57  * Note using signed integers limits us to physical addresses at most
  58  * 63 bits wide.  They are needed for negative offsetting in aliases
  59  * (large MemoryRegion::alias_offset).
  60  */
  61 struct AddrRange {
  62     Int128 start;
  63     Int128 size;
  64 };
  65
  66 static AddrRange addrrange_make(Int128 start, Int128 size)
  67 {
  68     return (AddrRange) { start, size };
  69 }
  70
  71 static bool addrrange_equal(AddrRange r1, AddrRange r2)
  72 {
  73     return int128_eq(r1.start, r2.start) && int128_eq(r1.size, r2.size);
  74 }
  75
  76 static Int128 addrrange_end(AddrRange r)
  77 {
  78     return int128_add(r.start, r.size);
  79 }
  80
  81 static AddrRange addrrange_shift(AddrRange range, Int128 delta)
  82 {
  83     int128_addto(&range.start, delta);
  84     return range;
  85 }
  86
  87 static bool addrrange_contains(AddrRange range, Int128 addr)
  88 {
  89     return int128_ge(addr, range.start)
  90         && int128_lt(addr, addrrange_end(range));
  91 }
  92
  93 static bool addrrange_intersects(AddrRange r1, AddrRange r2)
  94 {
  95     return addrrange_contains(r1, r2.start)
  96         || addrrange_contains(r2, r1.start);
  97 }
  98
  99 static AddrRange addrrange_intersection(AddrRange r1, AddrRange r2)
 100 {
 101     Int128 start = int128_max(r1.start, r2.start);
 102     Int128 end = int128_min(addrrange_end(r1), addrrange_end(r2));
 103     return addrrange_make(start, int128_sub(end, start));
 104 }
 105
 106 enum ListenerDirection { Forward, Reverse };
 107
 108 static bool memory_listener_match(MemoryListener *listener,
 109                                   MemoryRegionSection *section)
 110 {
 111     return !listener->address_space_filter
 112         || listener->address_space_filter == section->address_space;
 113 }
 114
 115 #define MEMORY_LISTENER_CALL_GLOBAL(_callback, _direction, _args...)    \
 116     do {                                                                \
 117         MemoryListener *_listener;                                      \
 118                                                                         \
 119         switch (_direction) {                                           \
 120         case Forward:                                                   \
 121             QTAILQ_FOREACH(_listener, &memory_listeners, link) {        \
 122                 if (_listener->_callback) {                             \
 123                     _listener->_callback(_listener, ##_args);           \
 124                 }                                                       \
 125             }                                                           \
 126             break;                                                      \
 127         case Reverse:                                                   \
 128             QTAILQ_FOREACH_REVERSE(_listener, &memory_listeners,        \
 129                                    memory_listeners, link) {            \
 130                 if (_listener->_callback) {                             \
 131                     _listener->_callback(_listener, ##_args);           \
 132                 }                                                       \
 133             }                                                           \
 134             break;                                                      \
 135         default:                                                        \
 136             abort();                                                    \
 137         }                                                               \
 138     } while (0)
 139
 140 #define MEMORY_LISTENER_CALL(_callback, _direction, _section, _args...) \
 141     do {                                                                \
 142         MemoryListener *_listener;                                      \
 143                                                                         \
 144         switch (_direction) {                                           \
 145         case Forward:                                                   \
 146             QTAILQ_FOREACH(_listener, &memory_listeners, link) {        \
 147                 if (_listener->_callback                                \
 148                     && memory_listener_match(_listener, _section)) {    \
 149                     _listener->_callback(_listener, _section, ##_args); \
 150                 }                                                       \
 151             }                                                           \
 152             break;                                                      \
 153         case Reverse:                                                   \
 154             QTAILQ_FOREACH_REVERSE(_listener, &memory_listeners,        \
 155                                    memory_listeners, link) {            \
 156                 if (_listener->_callback                                \
 157                     && memory_listener_match(_listener, _section)) {    \
 158                     _listener->_callback(_listener, _section, ##_args); \
 159                 }                                                       \
 160             }                                                           \
 161             break;                                                      \
 162         default:                                                        \
 163             abort();                                                    \
 164         }                                                               \
 165     } while (0)
 166
 167 /* No need to ref/unref .mr, the FlatRange keeps it alive.  */
 168 #define MEMORY_LISTENER_UPDATE_REGION(fr, as, dir, callback)            \
 169     MEMORY_LISTENER_CALL(callback, dir, (&(MemoryRegionSection) {       \
 170         .mr = (fr)->mr,                                                 \
 171         .address_space = (as),                                          \
 172         .offset_within_region = (fr)->offset_in_region,                 \
 173         .size = (fr)->addr.size,                                        \
 174         .offset_within_address_space = int128_get64((fr)->addr.start),  \
 175         .readonly = (fr)->readonly,                                     \
 176               }))
 177
 178 struct CoalescedMemoryRange {
 179     AddrRange addr;
 180     QTAILQ_ENTRY(CoalescedMemoryRange) link;
 181 };
 182
 183 struct MemoryRegionIoeventfd {
 184     AddrRange addr;
 185     bool match_data;
 186     uint64_t data;
 187     EventNotifier *e;
 188 };
 189
 190 static bool memory_region_ioeventfd_before(MemoryRegionIoeventfd a,
 191                                            MemoryRegionIoeventfd b)
 192 {
 193     if (int128_lt(a.addr.start, b.addr.start)) {
 194         return true;
 195     } else if (int128_gt(a.addr.start, b.addr.start)) {
 196         return false;
 197     } else if (int128_lt(a.addr.size, b.addr.size)) {
 198         return true;
 199     } else if (int128_gt(a.addr.size, b.addr.size)) {
 200         return false;
 201     } else if (a.match_data < b.match_data) {
 202         return true;
 203     } else  if (a.match_data > b.match_data) {
 204         return false;
 205     } else if (a.match_data) {
 206         if (a.data < b.data) {
 207             return true;
 208         } else if (a.data > b.data) {
 209             return false;
 210         }
 211     }
 212     if (a.e < b.e) {
 213         return true;
 214     } else if (a.e > b.e) {
 215         return false;
 216     }
 217     return false;
 218 }
 219
 220 static bool memory_region_ioeventfd_equal(MemoryRegionIoeventfd a,
 221                                           MemoryRegionIoeventfd b)
 222 {
 223     return !memory_region_ioeventfd_before(a, b)
 224         && !memory_region_ioeventfd_before(b, a);
 225 }
 226
 227 typedef struct FlatRange FlatRange;
 228 typedef struct FlatView FlatView;
 229
 230 /* Range of memory in the global map.  Addresses are absolute. */
 231 struct FlatRange {
 232     MemoryRegion *mr;
 233     hwaddr offset_in_region;
 234     AddrRange addr;
 235     uint8_t dirty_log_mask;
 236     bool romd_mode;
 237     bool readonly;
 238 };
 239
 240 /* Flattened global view of current active memory hierarchy.  Kept in sorted
 241  * order.
 242  */
 243 struct FlatView {
 244     unsigned ref;
 245     FlatRange *ranges;
 246     unsigned nr;
 247     unsigned nr_allocated;
 248 };
 249
 250 typedef struct AddressSpaceOps AddressSpaceOps;
 251
 252 #define FOR_EACH_FLAT_RANGE(var, view)          \
 253     for (var = (view)->ranges; var < (view)->ranges + (view)->nr; ++var)
 254
 255 static bool flatrange_equal(FlatRange *a, FlatRange *b)
 256 {
 257     return a->mr == b->mr
 258         && addrrange_equal(a->addr, b->addr)
 259         && a->offset_in_region == b->offset_in_region
 260         && a->romd_mode == b->romd_mode
 261         && a->readonly == b->readonly;
 262 }
 263
 264 static void flatview_init(FlatView *view)
 265 {
 266     view->ref = 1;
 267     view->ranges = NULL;
 268     view->nr = 0;
 269     view->nr_allocated = 0;
 270 }
 271
 272 /* Insert a range into a given position.  Caller is responsible for maintaining
 273  * sorting order.
 274  */
 275 static void flatview_insert(FlatView *view, unsigned pos, FlatRange *range)
 276 {
 277     if (view->nr == view->nr_allocated) {
 278         view->nr_allocated = MAX(2 * view->nr, 10);
 279         view->ranges = g_realloc(view->ranges,
 280                                     view->nr_allocated * sizeof(*view->ranges));
 281     }
 282     memmove(view->ranges + pos + 1, view->ranges + pos,
 283             (view->nr - pos) * sizeof(FlatRange));
 284     view->ranges[pos] = *range;
 285     memory_region_ref(range->mr);
 286     ++view->nr;
 287 }
 288
 289 static void flatview_destroy(FlatView *view)
 290 {
 291     int i;
 292
 293     for (i = 0; i < view->nr; i++) {
 294         memory_region_unref(view->ranges[i].mr);
 295     }
 296     g_free(view->ranges);
 297     g_free(view);
 298 }
 299
 300 static void flatview_ref(FlatView *view)
 301 {
 302     atomic_inc(&view->ref);
 303 }
 304
 305 static void flatview_unref(FlatView *view)
 306 {
 307     if (atomic_fetch_dec(&view->ref) == 1) {
 308         flatview_destroy(view);
 309     }
 310 }
 311
 312 static bool can_merge(FlatRange *r1, FlatRange *r2)
 313 {
 314     return int128_eq(addrrange_end(r1->addr), r2->addr.start)
 315         && r1->mr == r2->mr
 316         && int128_eq(int128_add(int128_make64(r1->offset_in_region),
 317                                 r1->addr.size),
 318                      int128_make64(r2->offset_in_region))
 319         && r1->dirty_log_mask == r2->dirty_log_mask
 320         && r1->romd_mode == r2->romd_mode
 321         && r1->readonly == r2->readonly;
 322 }
 323
 324 /* Attempt to simplify a view by merging adjacent ranges */
 325 static void flatview_simplify(FlatView *view)
 326 {
 327     unsigned i, j;
 328
 329     i = 0;
 330     while (i < view->nr) {
 331         j = i + 1;
 332         while (j < view->nr
 333                && can_merge(&view->ranges[j-1], &view->ranges[j])) {
 334             int128_addto(&view->ranges[i].addr.size, view->ranges[j].addr.size);
 335             ++j;
 336         }
 337         ++i;
 338         memmove(&view->ranges[i], &view->ranges[j],
 339                 (view->nr - j) * sizeof(view->ranges[j]));
 340         view->nr -= j - i;
 341     }
 342 }
 343
 344 static bool memory_region_big_endian(MemoryRegion *mr)
 345 {
 346 #ifdef TARGET_WORDS_BIGENDIAN
 347     return mr->ops->endianness != DEVICE_LITTLE_ENDIAN;
 348 #else
 349     return mr->ops->endianness == DEVICE_BIG_ENDIAN;
 350 #endif
 351 }
 352
 353 static bool memory_region_wrong_endianness(MemoryRegion *mr)
 354 {
 355 #ifdef TARGET_WORDS_BIGENDIAN
 356     return mr->ops->endianness == DEVICE_LITTLE_ENDIAN;
 357 #else
 358     return mr->ops->endianness == DEVICE_BIG_ENDIAN;
 359 #endif
 360 }
 361
 362 static void adjust_endianness(MemoryRegion *mr, uint64_t *data, unsigned size)
 363 {
 364     if (memory_region_wrong_endianness(mr)) {
 365         switch (size) {
 366         case 1:
 367             break;
 368         case 2:
 369             *data = bswap16(*data);
 370             break;
 371         case 4:
 372             *data = bswap32(*data);
 373             break;
 374         case 8:
 375             *data = bswap64(*data);
 376             break;
 377         default:
 378             abort();
 379         }
 380     }
 381 }
 382
 383 static void memory_region_oldmmio_read_accessor(MemoryRegion *mr,
 384                                                 hwaddr addr,
 385                                                 uint64_t *value,
 386                                                 unsigned size,
 387                                                 unsigned shift,
 388                                                 uint64_t mask)
 389 {
 390     uint64_t tmp;
 391
 392     tmp = mr->ops->old_mmio.read[ctz32(size)](mr->opaque, addr);
 393     trace_memory_region_ops_read(mr, addr, tmp, size);
 394     *value |= (tmp & mask) << shift;
 395 }
 396
 397 static void memory_region_read_accessor(MemoryRegion *mr,
 398                                         hwaddr addr,
 399                                         uint64_t *value,
 400                                         unsigned size,
 401                                         unsigned shift,
 402                                         uint64_t mask)
 403 {
 404     uint64_t tmp;
 405
 406     if (mr->flush_coalesced_mmio) {
 407         qemu_flush_coalesced_mmio_buffer();
 408     }
 409     tmp = mr->ops->read(mr->opaque, addr, size);
 410     trace_memory_region_ops_read(mr, addr, tmp, size);
 411     *value |= (tmp & mask) << shift;
 412 }
 413
 414 static void memory_region_oldmmio_write_accessor(MemoryRegion *mr,
 415                                                  hwaddr addr,
 416                                                  uint64_t *value,
 417                                                  unsigned size,
 418                                                  unsigned shift,
 419                                                  uint64_t mask)
 420 {
 421     uint64_t tmp;
 422
 423     tmp = (*value >> shift) & mask;
 424     trace_memory_region_ops_write(mr, addr, tmp, size);
 425     mr->ops->old_mmio.write[ctz32(size)](mr->opaque, addr, tmp);
 426 }
 427
 428 static void memory_region_write_accessor(MemoryRegion *mr,
 429                                          hwaddr addr,
 430                                          uint64_t *value,
 431                                          unsigned size,
 432                                          unsigned shift,
 433                                          uint64_t mask)
 434 {
 435     uint64_t tmp;
 436
 437     if (mr->flush_coalesced_mmio) {
 438         qemu_flush_coalesced_mmio_buffer();
 439     }
 440     tmp = (*value >> shift) & mask;
 441     trace_memory_region_ops_write(mr, addr, tmp, size);
 442     mr->ops->write(mr->opaque, addr, tmp, size);
 443 }
 444
 445 static void access_with_adjusted_size(hwaddr addr,
 446                                       uint64_t *value,
 447                                       unsigned size,
 448                                       unsigned access_size_min,
 449                                       unsigned access_size_max,
 450                                       void (*access)(MemoryRegion *mr,
 451                                                      hwaddr addr,
 452                                                      uint64_t *value,
 453                                                      unsigned size,
 454                                                      unsigned shift,
 455                                                      uint64_t mask),
 456                                       MemoryRegion *mr)
 457 {
 458     uint64_t access_mask;
 459     unsigned access_size;
 460     unsigned i;
 461
 462     if (!access_size_min) {
 463         access_size_min = 1;
 464     }
 465     if (!access_size_max) {
 466         access_size_max = 4;
 467     }
 468
 469     /* FIXME: support unaligned access? */
 470     access_size = MAX(MIN(size, access_size_max), access_size_min);
 471     access_mask = -1ULL >> (64 - access_size * 8);
 472     if (memory_region_big_endian(mr)) {
 473         for (i = 0; i < size; i += access_size) {
 474             access(mr, addr + i, value, access_size,
 475                    (size - access_size - i) * 8, access_mask);
 476         }
 477     } else {
 478         for (i = 0; i < size; i += access_size) {
 479             access(mr, addr + i, value, access_size, i * 8, access_mask);
 480         }
 481     }
 482 }
 483
 484 static AddressSpace *memory_region_to_address_space(MemoryRegion *mr)
 485 {
 486     AddressSpace *as;
 487
 488     while (mr->container) {
 489         mr = mr->container;
 490     }
 491     QTAILQ_FOREACH(as, &address_spaces, address_spaces_link) {
 492         if (mr == as->root) {
 493             return as;
 494         }
 495     }
 496     abort();
 497 }
 498
 499 /* Render a memory region into the global view.  Ranges in @view obscure
 500  * ranges in @mr.
 501  */
 502 static void render_memory_region(FlatView *view,
 503                                  MemoryRegion *mr,
 504                                  Int128 base,
 505                                  AddrRange clip,
 506                                  bool readonly)
 507 {
 508     MemoryRegion *subregion;
 509     unsigned i;
 510     hwaddr offset_in_region;
 511     Int128 remain;
 512     Int128 now;
 513     FlatRange fr;
 514     AddrRange tmp;
 515
 516     if (!mr->enabled) {
 517         return;
 518     }
 519
 520     int128_addto(&base, int128_make64(mr->addr));
 521     readonly |= mr->readonly;
 522
 523     tmp = addrrange_make(base, mr->size);
 524
 525     if (!addrrange_intersects(tmp, clip)) {
 526         return;
 527     }
 528
 529     clip = addrrange_intersection(tmp, clip);
 530
 531     if (mr->alias) {
 532         int128_subfrom(&base, int128_make64(mr->alias->addr));
 533         int128_subfrom(&base, int128_make64(mr->alias_offset));
 534         render_memory_region(view, mr->alias, base, clip, readonly);
 535         return;
 536     }
 537
 538     /* Render subregions in priority order. */
 539     QTAILQ_FOREACH(subregion, &mr->subregions, subregions_link) {
 540         render_memory_region(view, subregion, base, clip, readonly);
 541     }
 542
 543     if (!mr->terminates) {
 544         return;
 545     }
 546
 547     offset_in_region = int128_get64(int128_sub(clip.start, base));
 548     base = clip.start;
 549     remain = clip.size;
 550
 551     fr.mr = mr;
 552     fr.dirty_log_mask = mr->dirty_log_mask;
 553     fr.romd_mode = mr->romd_mode;
 554     fr.readonly = readonly;
 555
 556     /* Render the region itself into any gaps left by the current view. */
 557     for (i = 0; i < view->nr && int128_nz(remain); ++i) {
 558         if (int128_ge(base, addrrange_end(view->ranges[i].addr))) {
 559             continue;
 560         }
 561         if (int128_lt(base, view->ranges[i].addr.start)) {
 562             now = int128_min(remain,
 563                              int128_sub(view->ranges[i].addr.start, base));
 564             fr.offset_in_region = offset_in_region;
 565             fr.addr = addrrange_make(base, now);
 566             flatview_insert(view, i, &fr);
 567             ++i;
 568             int128_addto(&base, now);
 569             offset_in_region += int128_get64(now);
 570             int128_subfrom(&remain, now);
 571         }
 572         now = int128_sub(int128_min(int128_add(base, remain),
 573                                     addrrange_end(view->ranges[i].addr)),
 574                          base);
 575         int128_addto(&base, now);
 576         offset_in_region += int128_get64(now);
 577         int128_subfrom(&remain, now);
 578     }
 579     if (int128_nz(remain)) {
 580         fr.offset_in_region = offset_in_region;
 581         fr.addr = addrrange_make(base, remain);
 582         flatview_insert(view, i, &fr);
 583     }
 584 }
 585
 586 /* Render a memory topology into a list of disjoint absolute ranges. */
 587 static FlatView *generate_memory_topology(MemoryRegion *mr)
 588 {
 589     FlatView *view;
 590
 591     view = g_new(FlatView, 1);
 592     flatview_init(view);
 593
 594     if (mr) {
 595         render_memory_region(view, mr, int128_zero(),
 596                              addrrange_make(int128_zero(), int128_2_64()), false);
 597     }
 598     flatview_simplify(view);
 599
 600     return view;
 601 }
 602
 603 static void address_space_add_del_ioeventfds(AddressSpace *as,
 604                                              MemoryRegionIoeventfd *fds_new,
 605                                              unsigned fds_new_nb,
 606                                              MemoryRegionIoeventfd *fds_old,
 607                                              unsigned fds_old_nb)
 608 {
 609     unsigned iold, inew;
 610     MemoryRegionIoeventfd *fd;
 611     MemoryRegionSection section;
 612
 613     /* Generate a symmetric difference of the old and new fd sets, adding
 614      * and deleting as necessary.
 615      */
 616
 617     iold = inew = 0;
 618     while (iold < fds_old_nb || inew < fds_new_nb) {
 619         if (iold < fds_old_nb
 620             && (inew == fds_new_nb
 621                 || memory_region_ioeventfd_before(fds_old[iold],
 622                                                   fds_new[inew]))) {
 623             fd = &fds_old[iold];
 624             section = (MemoryRegionSection) {
 625                 .address_space = as,
 626                 .offset_within_address_space = int128_get64(fd->addr.start),
 627                 .size = fd->addr.size,
 628             };
 629             MEMORY_LISTENER_CALL(eventfd_del, Forward, &section,
 630                                  fd->match_data, fd->data, fd->e);
 631             ++iold;
 632         } else if (inew < fds_new_nb
 633                    && (iold == fds_old_nb
 634                        || memory_region_ioeventfd_before(fds_new[inew],
 635                                                          fds_old[iold]))) {
 636             fd = &fds_new[inew];
 637             section = (MemoryRegionSection) {
 638                 .address_space = as,
 639                 .offset_within_address_space = int128_get64(fd->addr.start),
 640                 .size = fd->addr.size,
 641             };
 642             MEMORY_LISTENER_CALL(eventfd_add, Reverse, &section,
 643                                  fd->match_data, fd->data, fd->e);
 644             ++inew;
 645         } else {
 646             ++iold;
 647             ++inew;
 648         }
 649     }
 650 }
 651
 652 static FlatView *address_space_get_flatview(AddressSpace *as)
 653 {
 654     FlatView *view;
 655
 656     qemu_mutex_lock(&flat_view_mutex);
 657     view = as->current_map;
 658     flatview_ref(view);
 659     qemu_mutex_unlock(&flat_view_mutex);
 660     return view;
 661 }
 662
 663 static void address_space_update_ioeventfds(AddressSpace *as)
 664 {
 665     FlatView *view;
 666     FlatRange *fr;
 667     unsigned ioeventfd_nb = 0;
 668     MemoryRegionIoeventfd *ioeventfds = NULL;
 669     AddrRange tmp;
 670     unsigned i;
 671
 672     view = address_space_get_flatview(as);
 673     FOR_EACH_FLAT_RANGE(fr, view) {
 674         for (i = 0; i < fr->mr->ioeventfd_nb; ++i) {
 675             tmp = addrrange_shift(fr->mr->ioeventfds[i].addr,
 676                                   int128_sub(fr->addr.start,
 677                                              int128_make64(fr->offset_in_region)));
 678             if (addrrange_intersects(fr->addr, tmp)) {
 679                 ++ioeventfd_nb;
 680                 ioeventfds = g_realloc(ioeventfds,
 681                                           ioeventfd_nb * sizeof(*ioeventfds));
 682                 ioeventfds[ioeventfd_nb-1] = fr->mr->ioeventfds[i];
 683                 ioeventfds[ioeventfd_nb-1].addr = tmp;
 684             }
 685         }
 686     }
 687
 688     address_space_add_del_ioeventfds(as, ioeventfds, ioeventfd_nb,
 689                                      as->ioeventfds, as->ioeventfd_nb);
 690
 691     g_free(as->ioeventfds);
 692     as->ioeventfds = ioeventfds;
 693     as->ioeventfd_nb = ioeventfd_nb;
 694     flatview_unref(view);
 695 }
 696
 697 static void address_space_update_topology_pass(AddressSpace *as,
 698                                                const FlatView *old_view,
 699                                                const FlatView *new_view,
 700                                                bool adding)
 701 {
 702     unsigned iold, inew;
 703     FlatRange *frold, *frnew;
 704
 705     /* Generate a symmetric difference of the old and new memory maps.
 706      * Kill ranges in the old map, and instantiate ranges in the new map.
 707      */
 708     iold = inew = 0;
 709     while (iold < old_view->nr || inew < new_view->nr) {
 710         if (iold < old_view->nr) {
 711             frold = &old_view->ranges[iold];
 712         } else {
 713             frold = NULL;
 714         }
 715         if (inew < new_view->nr) {
 716             frnew = &new_view->ranges[inew];
 717         } else {
 718             frnew = NULL;
 719         }
 720
 721         if (frold
 722             && (!frnew
 723                 || int128_lt(frold->addr.start, frnew->addr.start)
 724                 || (int128_eq(frold->addr.start, frnew->addr.start)
 725                     && !flatrange_equal(frold, frnew)))) {
 726             /* In old but not in new, or in both but attributes changed. */
 727
 728             if (!adding) {
 729                 MEMORY_LISTENER_UPDATE_REGION(frold, as, Reverse, region_del);
 730             }
 731
 732             ++iold;
 733         } else if (frold && frnew && flatrange_equal(frold, frnew)) {
 734             /* In both and unchanged (except logging may have changed) */
 735
 736             if (adding) {
 737                 MEMORY_LISTENER_UPDATE_REGION(frnew, as, Forward, region_nop);
 738                 if (frold->dirty_log_mask && !frnew->dirty_log_mask) {
 739                     MEMORY_LISTENER_UPDATE_REGION(frnew, as, Reverse, log_stop);
 740                 } else if (frnew->dirty_log_mask && !frold->dirty_log_mask) {
 741                     MEMORY_LISTENER_UPDATE_REGION(frnew, as, Forward, log_start);
 742                 }
 743             }
 744
 745             ++iold;
 746             ++inew;
 747         } else {
 748             /* In new */
 749
 750             if (adding) {
 751                 MEMORY_LISTENER_UPDATE_REGION(frnew, as, Forward, region_add);
 752             }
 753
 754             ++inew;
 755         }
 756     }
 757 }
 758
 759
 760 static void address_space_update_topology(AddressSpace *as)
 761 {
 762     FlatView *old_view = address_space_get_flatview(as);
 763     FlatView *new_view = generate_memory_topology(as->root);
 764
 765     address_space_update_topology_pass(as, old_view, new_view, false);
 766     address_space_update_topology_pass(as, old_view, new_view, true);
 767
 768     qemu_mutex_lock(&flat_view_mutex);
 769     flatview_unref(as->current_map);
 770     as->current_map = new_view;
 771     qemu_mutex_unlock(&flat_view_mutex);
 772
 773     /* Note that all the old MemoryRegions are still alive up to this
 774      * point.  This relieves most MemoryListeners from the need to
 775      * ref/unref the MemoryRegions they get---unless they use them
 776      * outside the iothread mutex, in which case precise reference
 777      * counting is necessary.
 778      */
 779     flatview_unref(old_view);
 780
 781     address_space_update_ioeventfds(as);
 782 }
 783
 784 void memory_region_transaction_begin(void)
 785 {
 786     qemu_flush_coalesced_mmio_buffer();
 787     ++memory_region_transaction_depth;
 788 }
 789
 790 static void memory_region_clear_pending(void)
 791 {
 792     memory_region_update_pending = false;
 793     ioeventfd_update_pending = false;
 794 }
 795
 796 void memory_region_transaction_commit(void)
 797 {
 798     AddressSpace *as;
 799
 800     assert(memory_region_transaction_depth);
 801     --memory_region_transaction_depth;
 802     if (!memory_region_transaction_depth) {
 803         if (memory_region_update_pending) {
 804             MEMORY_LISTENER_CALL_GLOBAL(begin, Forward);
 805
 806             QTAILQ_FOREACH(as, &address_spaces, address_spaces_link) {
 807                 address_space_update_topology(as);
 808             }
 809
 810             MEMORY_LISTENER_CALL_GLOBAL(commit, Forward);
 811         } else if (ioeventfd_update_pending) {
 812             QTAILQ_FOREACH(as, &address_spaces, address_spaces_link) {
 813                 address_space_update_ioeventfds(as);
 814             }
 815         }
 816         memory_region_clear_pending();
 817    }
 818 }
 819
 820 static void memory_region_destructor_none(MemoryRegion *mr)
 821 {
 822 }
 823
 824 static void memory_region_destructor_ram(MemoryRegion *mr)
 825 {
 826     qemu_ram_free(mr->ram_addr);
 827 }
 828
 829 static void memory_region_destructor_alias(MemoryRegion *mr)
 830 {
 831     memory_region_unref(mr->alias);
 832 }
 833
 834 static void memory_region_destructor_ram_from_ptr(MemoryRegion *mr)
 835 {
 836     qemu_ram_free_from_ptr(mr->ram_addr);
 837 }
 838
 839 static void memory_region_destructor_rom_device(MemoryRegion *mr)
 840 {
 841     qemu_ram_free(mr->ram_addr & TARGET_PAGE_MASK);
 842 }
 843
 844 void memory_region_init(MemoryRegion *mr,
 845                         Object *owner,
 846                         const char *name,
 847                         uint64_t size)
 848 {
 849     mr->ops = &unassigned_mem_ops;
 850     mr->opaque = NULL;
 851     mr->owner = owner;
 852     mr->iommu_ops = NULL;
 853     mr->container = NULL;
 854     mr->size = int128_make64(size);
 855     if (size == UINT64_MAX) {
 856         mr->size = int128_2_64();
 857     }
 858     mr->addr = 0;
 859     mr->subpage = false;
 860     mr->enabled = true;
 861     mr->terminates = false;
 862     mr->ram = false;
 863     mr->romd_mode = true;
 864     mr->readonly = false;
 865     mr->rom_device = false;
 866     mr->destructor = memory_region_destructor_none;
 867     mr->priority = 0;
 868     mr->may_overlap = false;
 869     mr->alias = NULL;
 870     QTAILQ_INIT(&mr->subregions);
 871     memset(&mr->subregions_link, 0, sizeof mr->subregions_link);
 872     QTAILQ_INIT(&mr->coalesced);
 873     mr->name = g_strdup(name);
 874     mr->dirty_log_mask = 0;
 875     mr->ioeventfd_nb = 0;
 876     mr->ioeventfds = NULL;
 877     mr->flush_coalesced_mmio = false;
 878 }
 879
 880 static uint64_t unassigned_mem_read(void *opaque, hwaddr addr,
 881                                     unsigned size)
 882 {
 883 #ifdef DEBUG_UNASSIGNED
 884     printf("Unassigned mem read " TARGET_FMT_plx "\n", addr);
 885 #endif
 886     if (current_cpu != NULL) {
 887         cpu_unassigned_access(current_cpu, addr, false, false, 0, size);
 888     }
 889     return 0;
 890 }
 891
 892 static void unassigned_mem_write(void *opaque, hwaddr addr,
 893                                  uint64_t val, unsigned size)
 894 {
 895 #ifdef DEBUG_UNASSIGNED
 896     printf("Unassigned mem write " TARGET_FMT_plx " = 0x%"PRIx64"\n", addr, val);
 897 #endif
 898     if (current_cpu != NULL) {
 899         cpu_unassigned_access(current_cpu, addr, true, false, 0, size);
 900     }
 901 }
 902
 903 static bool unassigned_mem_accepts(void *opaque, hwaddr addr,
 904                                    unsigned size, bool is_write)
 905 {
 906     return false;
 907 }
 908
 909 const MemoryRegionOps unassigned_mem_ops = {
 910     .valid.accepts = unassigned_mem_accepts,
 911     .endianness = DEVICE_NATIVE_ENDIAN,
 912 };
 913
 914 bool memory_region_access_valid(MemoryRegion *mr,
 915                                 hwaddr addr,
 916                                 unsigned size,
 917                                 bool is_write)
 918 {
 919     int access_size_min, access_size_max;
 920     int access_size, i;
 921
 922     if (!mr->ops->valid.unaligned && (addr & (size - 1))) {
 923         return false;
 924     }
 925
 926     if (!mr->ops->valid.accepts) {
 927         return true;
 928     }
 929
 930     access_size_min = mr->ops->valid.min_access_size;
 931     if (!mr->ops->valid.min_access_size) {
 932         access_size_min = 1;
 933     }
 934
 935     access_size_max = mr->ops->valid.max_access_size;
 936     if (!mr->ops->valid.max_access_size) {
 937         access_size_max = 4;
 938     }
 939
 940     access_size = MAX(MIN(size, access_size_max), access_size_min);
 941     for (i = 0; i < size; i += access_size) {
 942         if (!mr->ops->valid.accepts(mr->opaque, addr + i, access_size,
 943                                     is_write)) {
 944             return false;
 945         }
 946     }
 947
 948     return true;
 949 }
 950
 951 static uint64_t memory_region_dispatch_read1(MemoryRegion *mr,
 952                                              hwaddr addr,
 953                                              unsigned size)
 954 {
 955     uint64_t data = 0;
 956
 957     if (mr->ops->read) {
 958         access_with_adjusted_size(addr, &data, size,
 959                                   mr->ops->impl.min_access_size,
 960                                   mr->ops->impl.max_access_size,
 961                                   memory_region_read_accessor, mr);
 962     } else {
 963         access_with_adjusted_size(addr, &data, size, 1, 4,
 964                                   memory_region_oldmmio_read_accessor, mr);
 965     }
 966
 967     return data;
 968 }
 969
 970 static bool memory_region_dispatch_read(MemoryRegion *mr,
 971                                         hwaddr addr,
 972                                         uint64_t *pval,
 973                                         unsigned size)
 974 {
 975     if (!memory_region_access_valid(mr, addr, size, false)) {
 976         *pval = unassigned_mem_read(mr, addr, size);
 977         return true;
 978     }
 979
 980     *pval = memory_region_dispatch_read1(mr, addr, size);
 981     adjust_endianness(mr, pval, size);
 982     return false;
 983 }
 984
 985 static bool memory_region_dispatch_write(MemoryRegion *mr,
 986                                          hwaddr addr,
 987                                          uint64_t data,
 988                                          unsigned size)
 989 {
 990     if (!memory_region_access_valid(mr, addr, size, true)) {
 991         unassigned_mem_write(mr, addr, data, size);
 992         return true;
 993     }
 994
 995     adjust_endianness(mr, &data, size);
 996
 997     if (mr->ops->write) {
 998         access_with_adjusted_size(addr, &data, size,
 999                                   mr->ops->impl.min_access_size,
1000                                   mr->ops->impl.max_access_size,
1001                                   memory_region_write_accessor, mr);
1002     } else {
1003         access_with_adjusted_size(addr, &data, size, 1, 4,
1004                                   memory_region_oldmmio_write_accessor, mr);
1005     }
1006     return false;
1007 }
1008
1009 void memory_region_init_io(MemoryRegion *mr,
1010                            Object *owner,
1011                            const MemoryRegionOps *ops,
1012                            void *opaque,
1013                            const char *name,
1014                            uint64_t size)
1015 {
1016     memory_region_init(mr, owner, name, size);
1017     mr->ops = ops;
1018     mr->opaque = opaque;
1019     mr->terminates = true;
1020     mr->ram_addr = ~(ram_addr_t)0;
1021 }
1022
1023 void memory_region_init_ram(MemoryRegion *mr,
1024                             Object *owner,
1025                             const char *name,
1026                             uint64_t size)
1027 {
1028     memory_region_init(mr, owner, name, size);
1029     mr->ram = true;
1030     mr->terminates = true;
1031     mr->destructor = memory_region_destructor_ram;
1032     mr->ram_addr = qemu_ram_alloc(size, mr);
1033 }
1034
1035 void memory_region_init_ram_ptr(MemoryRegion *mr,
1036                                 Object *owner,
1037                                 const char *name,
1038                                 uint64_t size,
1039                                 void *ptr)
1040 {
1041     memory_region_init(mr, owner, name, size);
1042     mr->ram = true;
1043     mr->terminates = true;
1044     mr->destructor = memory_region_destructor_ram_from_ptr;
1045     mr->ram_addr = qemu_ram_alloc_from_ptr(size, ptr, mr);
1046 }
1047
1048 void memory_region_init_alias(MemoryRegion *mr,
1049                               Object *owner,
1050                               const char *name,
1051                               MemoryRegion *orig,
1052                               hwaddr offset,
1053                               uint64_t size)
1054 {
1055     memory_region_init(mr, owner, name, size);
1056     memory_region_ref(orig);
1057     mr->destructor = memory_region_destructor_alias;
1058     mr->alias = orig;
1059     mr->alias_offset = offset;
1060 }
1061
1062 void memory_region_init_rom_device(MemoryRegion *mr,
1063                                    Object *owner,
1064                                    const MemoryRegionOps *ops,
1065                                    void *opaque,
1066                                    const char *name,
1067                                    uint64_t size)
1068 {
1069     memory_region_init(mr, owner, name, size);
1070     mr->ops = ops;
1071     mr->opaque = opaque;
1072     mr->terminates = true;
1073     mr->rom_device = true;
1074     mr->destructor = memory_region_destructor_rom_device;
1075     mr->ram_addr = qemu_ram_alloc(size, mr);
1076 }
1077
1078 void memory_region_init_iommu(MemoryRegion *mr,
1079                               Object *owner,
1080                               const MemoryRegionIOMMUOps *ops,
1081                               const char *name,
1082                               uint64_t size)
1083 {
1084     memory_region_init(mr, owner, name, size);
1085     mr->iommu_ops = ops,
1086     mr->terminates = true;  /* then re-forwards */
1087     notifier_list_init(&mr->iommu_notify);
1088 }
1089
1090 void memory_region_init_reservation(MemoryRegion *mr,
1091                                     Object *owner,
1092                                     const char *name,
1093                                     uint64_t size)
1094 {
1095     memory_region_init_io(mr, owner, &unassigned_mem_ops, mr, name, size);
1096 }
1097
1098 void memory_region_destroy(MemoryRegion *mr)
1099 {
1100     assert(QTAILQ_EMPTY(&mr->subregions));
1101     assert(memory_region_transaction_depth == 0);
1102     mr->destructor(mr);
1103     memory_region_clear_coalescing(mr);
1104     g_free((char *)mr->name);
1105     g_free(mr->ioeventfds);
1106 }
1107
1108 Object *memory_region_owner(MemoryRegion *mr)
1109 {
1110     return mr->owner;
1111 }
1112
1113 void memory_region_ref(MemoryRegion *mr)
1114 {
1115     if (mr && mr->owner) {
1116         object_ref(mr->owner);
1117     }
1118 }
1119
1120 void memory_region_unref(MemoryRegion *mr)
1121 {
1122     if (mr && mr->owner) {
1123         object_unref(mr->owner);
1124     }
1125 }
1126
1127 uint64_t memory_region_size(MemoryRegion *mr)
1128 {
1129     if (int128_eq(mr->size, int128_2_64())) {
1130         return UINT64_MAX;
1131     }
1132     return int128_get64(mr->size);
1133 }
1134
1135 const char *memory_region_name(MemoryRegion *mr)
1136 {
1137     return mr->name;
1138 }
1139
1140 bool memory_region_is_ram(MemoryRegion *mr)
1141 {
1142     return mr->ram;
1143 }
1144
1145 bool memory_region_is_logging(MemoryRegion *mr)
1146 {
1147     return mr->dirty_log_mask;
1148 }
1149
1150 bool memory_region_is_rom(MemoryRegion *mr)
1151 {
1152     return mr->ram && mr->readonly;
1153 }
1154
1155 bool memory_region_is_iommu(MemoryRegion *mr)
1156 {
1157     return mr->iommu_ops;
1158 }
1159
1160 void memory_region_register_iommu_notifier(MemoryRegion *mr, Notifier *n)
1161 {
1162     notifier_list_add(&mr->iommu_notify, n);
1163 }
1164
1165 void memory_region_unregister_iommu_notifier(Notifier *n)
1166 {
1167     notifier_remove(n);
1168 }
1169
1170 void memory_region_notify_iommu(MemoryRegion *mr,
1171                                 IOMMUTLBEntry entry)
1172 {
1173     assert(memory_region_is_iommu(mr));
1174     notifier_list_notify(&mr->iommu_notify, &entry);
1175 }
1176
1177 void memory_region_set_log(MemoryRegion *mr, bool log, unsigned client)
1178 {
1179     uint8_t mask = 1 << client;
1180
1181     memory_region_transaction_begin();
1182     mr->dirty_log_mask = (mr->dirty_log_mask & ~mask) | (log * mask);
1183     memory_region_update_pending |= mr->enabled;
1184     memory_region_transaction_commit();
1185 }
1186
1187 bool memory_region_get_dirty(MemoryRegion *mr, hwaddr addr,
1188                              hwaddr size, unsigned client)
1189 {
1190     assert(mr->terminates);
1191     return cpu_physical_memory_get_dirty(mr->ram_addr + addr, size, client);
1192 }
1193
1194 void memory_region_set_dirty(MemoryRegion *mr, hwaddr addr,
1195                              hwaddr size)
1196 {
1197     assert(mr->terminates);
1198     cpu_physical_memory_set_dirty_range(mr->ram_addr + addr, size);
1199 }
1200
1201 bool memory_region_test_and_clear_dirty(MemoryRegion *mr, hwaddr addr,
1202                                         hwaddr size, unsigned client)
1203 {
1204     bool ret;
1205     assert(mr->terminates);
1206     ret = cpu_physical_memory_get_dirty(mr->ram_addr + addr, size, client);
1207     if (ret) {
1208         cpu_physical_memory_reset_dirty(mr->ram_addr + addr, size, client);
1209     }
1210     return ret;
1211 }
1212
1213
1214 void memory_region_sync_dirty_bitmap(MemoryRegion *mr)
1215 {
1216     AddressSpace *as;
1217     FlatRange *fr;
1218
1219     QTAILQ_FOREACH(as, &address_spaces, address_spaces_link) {
1220         FlatView *view = address_space_get_flatview(as);
1221         FOR_EACH_FLAT_RANGE(fr, view) {
1222             if (fr->mr == mr) {
1223                 MEMORY_LISTENER_UPDATE_REGION(fr, as, Forward, log_sync);
1224             }
1225         }
1226         flatview_unref(view);
1227     }
1228 }
1229
1230 void memory_region_set_readonly(MemoryRegion *mr, bool readonly)
1231 {
1232     if (mr->readonly != readonly) {
1233         memory_region_transaction_begin();
1234         mr->readonly = readonly;
1235         memory_region_update_pending |= mr->enabled;
1236         memory_region_transaction_commit();
1237     }
1238 }
1239
1240 void memory_region_rom_device_set_romd(MemoryRegion *mr, bool romd_mode)
1241 {
1242     if (mr->romd_mode != romd_mode) {
1243         memory_region_transaction_begin();
1244         mr->romd_mode = romd_mode;
1245         memory_region_update_pending |= mr->enabled;
1246         memory_region_transaction_commit();
1247     }
1248 }
1249
1250 void memory_region_reset_dirty(MemoryRegion *mr, hwaddr addr,
1251                                hwaddr size, unsigned client)
1252 {
1253     assert(mr->terminates);
1254     cpu_physical_memory_reset_dirty(mr->ram_addr + addr, size, client);
1255 }
1256
1257 void *memory_region_get_ram_ptr(MemoryRegion *mr)
1258 {
1259     if (mr->alias) {
1260         return memory_region_get_ram_ptr(mr->alias) + mr->alias_offset;
1261     }
1262
1263     assert(mr->terminates);
1264
1265     return qemu_get_ram_ptr(mr->ram_addr & TARGET_PAGE_MASK);
1266 }
1267
1268 static void memory_region_update_coalesced_range_as(MemoryRegion *mr, AddressSpace *as)
1269 {
1270     FlatView *view;
1271     FlatRange *fr;
1272     CoalescedMemoryRange *cmr;
1273     AddrRange tmp;
1274     MemoryRegionSection section;
1275
1276     view = address_space_get_flatview(as);
1277     FOR_EACH_FLAT_RANGE(fr, view) {
1278         if (fr->mr == mr) {
1279             section = (MemoryRegionSection) {
1280                 .address_space = as,
1281                 .offset_within_address_space = int128_get64(fr->addr.start),
1282                 .size = fr->addr.size,
1283             };
1284
1285             MEMORY_LISTENER_CALL(coalesced_mmio_del, Reverse, &section,
1286                                  int128_get64(fr->addr.start),
1287                                  int128_get64(fr->addr.size));
1288             QTAILQ_FOREACH(cmr, &mr->coalesced, link) {
1289                 tmp = addrrange_shift(cmr->addr,
1290                                       int128_sub(fr->addr.start,
1291                                                  int128_make64(fr->offset_in_region)));
1292                 if (!addrrange_intersects(tmp, fr->addr)) {
1293                     continue;
1294                 }
1295                 tmp = addrrange_intersection(tmp, fr->addr);
1296                 MEMORY_LISTENER_CALL(coalesced_mmio_add, Forward, &section,
1297                                      int128_get64(tmp.start),
1298                                      int128_get64(tmp.size));
1299             }
1300         }
1301     }
1302     flatview_unref(view);
1303 }
1304
1305 static void memory_region_update_coalesced_range(MemoryRegion *mr)
1306 {
1307     AddressSpace *as;
1308
1309     QTAILQ_FOREACH(as, &address_spaces, address_spaces_link) {
1310         memory_region_update_coalesced_range_as(mr, as);
1311     }
1312 }
1313
1314 void memory_region_set_coalescing(MemoryRegion *mr)
1315 {
1316     memory_region_clear_coalescing(mr);
1317     memory_region_add_coalescing(mr, 0, int128_get64(mr->size));
1318 }
1319
1320 void memory_region_add_coalescing(MemoryRegion *mr,
1321                                   hwaddr offset,
1322                                   uint64_t size)
1323 {
1324     CoalescedMemoryRange *cmr = g_malloc(sizeof(*cmr));
1325
1326     cmr->addr = addrrange_make(int128_make64(offset), int128_make64(size));
1327     QTAILQ_INSERT_TAIL(&mr->coalesced, cmr, link);
1328     memory_region_update_coalesced_range(mr);
1329     memory_region_set_flush_coalesced(mr);
1330 }
1331
1332 void memory_region_clear_coalescing(MemoryRegion *mr)
1333 {
1334     CoalescedMemoryRange *cmr;
1335     bool updated = false;
1336
1337     qemu_flush_coalesced_mmio_buffer();
1338     mr->flush_coalesced_mmio = false;
1339
1340     while (!QTAILQ_EMPTY(&mr->coalesced)) {
1341         cmr = QTAILQ_FIRST(&mr->coalesced);
1342         QTAILQ_REMOVE(&mr->coalesced, cmr, link);
1343         g_free(cmr);
1344         updated = true;
1345     }
1346
1347     if (updated) {
1348         memory_region_update_coalesced_range(mr);
1349     }
1350 }
1351
1352 void memory_region_set_flush_coalesced(MemoryRegion *mr)
1353 {
1354     mr->flush_coalesced_mmio = true;
1355 }
1356
1357 void memory_region_clear_flush_coalesced(MemoryRegion *mr)
1358 {
1359     qemu_flush_coalesced_mmio_buffer();
1360     if (QTAILQ_EMPTY(&mr->coalesced)) {
1361         mr->flush_coalesced_mmio = false;
1362     }
1363 }
1364
1365 void memory_region_add_eventfd(MemoryRegion *mr,
1366                                hwaddr addr,
1367                                unsigned size,
1368                                bool match_data,
1369                                uint64_t data,
1370                                EventNotifier *e)
1371 {
1372     MemoryRegionIoeventfd mrfd = {
1373         .addr.start = int128_make64(addr),
1374         .addr.size = int128_make64(size),
1375         .match_data = match_data,
1376         .data = data,
1377         .e = e,
1378     };
1379     unsigned i;
1380
1381     adjust_endianness(mr, &mrfd.data, size);
1382     memory_region_transaction_begin();
1383     for (i = 0; i < mr->ioeventfd_nb; ++i) {
1384         if (memory_region_ioeventfd_before(mrfd, mr->ioeventfds[i])) {
1385             break;
1386         }
1387     }
1388     ++mr->ioeventfd_nb;
1389     mr->ioeventfds = g_realloc(mr->ioeventfds,
1390                                   sizeof(*mr->ioeventfds) * mr->ioeventfd_nb);
1391     memmove(&mr->ioeventfds[i+1], &mr->ioeventfds[i],
1392             sizeof(*mr->ioeventfds) * (mr->ioeventfd_nb-1 - i));
1393     mr->ioeventfds[i] = mrfd;
1394     ioeventfd_update_pending |= mr->enabled;
1395     memory_region_transaction_commit();
1396 }
1397
1398 void memory_region_del_eventfd(MemoryRegion *mr,
1399                                hwaddr addr,
1400                                unsigned size,
1401                                bool match_data,
1402                                uint64_t data,
1403                                EventNotifier *e)
1404 {
1405     MemoryRegionIoeventfd mrfd = {
1406         .addr.start = int128_make64(addr),
1407         .addr.size = int128_make64(size),
1408         .match_data = match_data,
1409         .data = data,
1410         .e = e,
1411     };
1412     unsigned i;
1413
1414     adjust_endianness(mr, &mrfd.data, size);
1415     memory_region_transaction_begin();
1416     for (i = 0; i < mr->ioeventfd_nb; ++i) {
1417         if (memory_region_ioeventfd_equal(mrfd, mr->ioeventfds[i])) {
1418             break;
1419         }
1420     }
1421     assert(i != mr->ioeventfd_nb);
1422     memmove(&mr->ioeventfds[i], &mr->ioeventfds[i+1],
1423             sizeof(*mr->ioeventfds) * (mr->ioeventfd_nb - (i+1)));
1424     --mr->ioeventfd_nb;
1425     mr->ioeventfds = g_realloc(mr->ioeventfds,
1426                                   sizeof(*mr->ioeventfds)*mr->ioeventfd_nb + 1);
1427     ioeventfd_update_pending |= mr->enabled;
1428     memory_region_transaction_commit();
1429 }
1430
1431 static void memory_region_update_container_subregions(MemoryRegion *subregion)
1432 {
1433     hwaddr offset = subregion->addr;
1434     MemoryRegion *mr = subregion->container;
1435     MemoryRegion *other;
1436
1437     memory_region_transaction_begin();
1438
1439     memory_region_ref(subregion);
1440     QTAILQ_FOREACH(other, &mr->subregions, subregions_link) {
1441         if (subregion->may_overlap || other->may_overlap) {
1442             continue;
1443         }
1444         if (int128_ge(int128_make64(offset),
1445                       int128_add(int128_make64(other->addr), other->size))
1446             || int128_le(int128_add(int128_make64(offset), subregion->size),
1447                          int128_make64(other->addr))) {
1448             continue;
1449         }
1450 #if 0
1451         printf("warning: subregion collision %llx/%llx (%s) "
1452                "vs %llx/%llx (%s)\n",
1453                (unsigned long long)offset,
1454                (unsigned long long)int128_get64(subregion->size),
1455                subregion->name,
1456                (unsigned long long)other->addr,
1457                (unsigned long long)int128_get64(other->size),
1458                other->name);
1459 #endif
1460     }
1461     QTAILQ_FOREACH(other, &mr->subregions, subregions_link) {
1462         if (subregion->priority >= other->priority) {
1463             QTAILQ_INSERT_BEFORE(other, subregion, subregions_link);
1464             goto done;
1465         }
1466     }
1467     QTAILQ_INSERT_TAIL(&mr->subregions, subregion, subregions_link);
1468 done:
1469     memory_region_update_pending |= mr->enabled && subregion->enabled;
1470     memory_region_transaction_commit();
1471 }
1472
1473 static void memory_region_add_subregion_common(MemoryRegion *mr,
1474                                                hwaddr offset,
1475                                                MemoryRegion *subregion)
1476 {
1477     assert(!subregion->container);
1478     subregion->container = mr;
1479     subregion->addr = offset;
1480     memory_region_update_container_subregions(subregion);
1481 }
1482
1483 void memory_region_add_subregion(MemoryRegion *mr,
1484                                  hwaddr offset,
1485                                  MemoryRegion *subregion)
1486 {
1487     subregion->may_overlap = false;
1488     subregion->priority = 0;
1489     memory_region_add_subregion_common(mr, offset, subregion);
1490 }
1491
1492 void memory_region_add_subregion_overlap(MemoryRegion *mr,
1493                                          hwaddr offset,
1494                                          MemoryRegion *subregion,
1495                                          int priority)
1496 {
1497     subregion->may_overlap = true;
1498     subregion->priority = priority;
1499     memory_region_add_subregion_common(mr, offset, subregion);
1500 }
1501
1502 void memory_region_del_subregion(MemoryRegion *mr,
1503                                  MemoryRegion *subregion)
1504 {
1505     memory_region_transaction_begin();
1506     assert(subregion->container == mr);
1507     subregion->container = NULL;
1508     QTAILQ_REMOVE(&mr->subregions, subregion, subregions_link);
1509     memory_region_unref(subregion);
1510     memory_region_update_pending |= mr->enabled && subregion->enabled;
1511     memory_region_transaction_commit();
1512 }
1513
1514 void memory_region_set_enabled(MemoryRegion *mr, bool enabled)
1515 {
1516     if (enabled == mr->enabled) {
1517         return;
1518     }
1519     memory_region_transaction_begin();
1520     mr->enabled = enabled;
1521     memory_region_update_pending = true;
1522     memory_region_transaction_commit();
1523 }
1524
1525 static void memory_region_readd_subregion(MemoryRegion *mr)
1526 {
1527     MemoryRegion *container = mr->container;
1528
1529     if (container) {
1530         memory_region_transaction_begin();
1531         memory_region_ref(mr);
1532         memory_region_del_subregion(container, mr);
1533         mr->container = container;
1534         memory_region_update_container_subregions(mr);
1535         memory_region_unref(mr);
1536         memory_region_transaction_commit();
1537     }
1538 }
1539
1540 void memory_region_set_address(MemoryRegion *mr, hwaddr addr)
1541 {
1542     if (addr != mr->addr) {
1543         mr->addr = addr;
1544         memory_region_readd_subregion(mr);
1545     }
1546 }
1547
1548 void memory_region_set_alias_offset(MemoryRegion *mr, hwaddr offset)
1549 {
1550     assert(mr->alias);
1551
1552     if (offset == mr->alias_offset) {
1553         return;
1554     }
1555
1556     memory_region_transaction_begin();
1557     mr->alias_offset = offset;
1558     memory_region_update_pending |= mr->enabled;
1559     memory_region_transaction_commit();
1560 }
1561
1562 ram_addr_t memory_region_get_ram_addr(MemoryRegion *mr)
1563 {
1564     return mr->ram_addr;
1565 }
1566
1567 static int cmp_flatrange_addr(const void *addr_, const void *fr_)
1568 {
1569     const AddrRange *addr = addr_;
1570     const FlatRange *fr = fr_;
1571
1572     if (int128_le(addrrange_end(*addr), fr->addr.start)) {
1573         return -1;
1574     } else if (int128_ge(addr->start, addrrange_end(fr->addr))) {
1575         return 1;
1576     }
1577     return 0;
1578 }
1579
1580 static FlatRange *flatview_lookup(FlatView *view, AddrRange addr)
1581 {
1582     return bsearch(&addr, view->ranges, view->nr,
1583                    sizeof(FlatRange), cmp_flatrange_addr);
1584 }
1585
1586 bool memory_region_present(MemoryRegion *container, hwaddr addr)
1587 {
1588     MemoryRegion *mr = memory_region_find(container, addr, 1).mr;
1589     if (!mr || (mr == container)) {
1590         return false;
1591     }
1592     memory_region_unref(mr);
1593     return true;
1594 }
1595
1596 MemoryRegionSection memory_region_find(MemoryRegion *mr,
1597                                        hwaddr addr, uint64_t size)
1598 {
1599     MemoryRegionSection ret = { .mr = NULL };
1600     MemoryRegion *root;
1601     AddressSpace *as;
1602     AddrRange range;
1603     FlatView *view;
1604     FlatRange *fr;
1605
1606     addr += mr->addr;
1607     for (root = mr; root->container; ) {
1608         root = root->container;
1609         addr += root->addr;
1610     }
1611
1612     as = memory_region_to_address_space(root);
1613     range = addrrange_make(int128_make64(addr), int128_make64(size));
1614
1615     view = address_space_get_flatview(as);
1616     fr = flatview_lookup(view, range);
1617     if (!fr) {
1618         flatview_unref(view);
1619         return ret;
1620     }
1621
1622     while (fr > view->ranges && addrrange_intersects(fr[-1].addr, range)) {
1623         --fr;
1624     }
1625
1626     ret.mr = fr->mr;
1627     ret.address_space = as;
1628     range = addrrange_intersection(range, fr->addr);
1629     ret.offset_within_region = fr->offset_in_region;
1630     ret.offset_within_region += int128_get64(int128_sub(range.start,
1631                                                         fr->addr.start));
1632     ret.size = range.size;
1633     ret.offset_within_address_space = int128_get64(range.start);
1634     ret.readonly = fr->readonly;
1635     memory_region_ref(ret.mr);
1636
1637     flatview_unref(view);
1638     return ret;
1639 }
1640
1641 void address_space_sync_dirty_bitmap(AddressSpace *as)
1642 {
1643     FlatView *view;
1644     FlatRange *fr;
1645
1646     view = address_space_get_flatview(as);
1647     FOR_EACH_FLAT_RANGE(fr, view) {
1648         MEMORY_LISTENER_UPDATE_REGION(fr, as, Forward, log_sync);
1649     }
1650     flatview_unref(view);
1651 }
1652
1653 void memory_global_dirty_log_start(void)
1654 {
1655     global_dirty_log = true;
1656     MEMORY_LISTENER_CALL_GLOBAL(log_global_start, Forward);
1657 }
1658
1659 void memory_global_dirty_log_stop(void)
1660 {
1661     global_dirty_log = false;
1662     MEMORY_LISTENER_CALL_GLOBAL(log_global_stop, Reverse);
1663 }
1664
1665 static void listener_add_address_space(MemoryListener *listener,
1666                                        AddressSpace *as)
1667 {
1668     FlatView *view;
1669     FlatRange *fr;
1670
1671     if (listener->address_space_filter
1672         && listener->address_space_filter != as) {
1673         return;
1674     }
1675
1676     if (global_dirty_log) {
1677         if (listener->log_global_start) {
1678             listener->log_global_start(listener);
1679         }
1680     }
1681
1682     view = address_space_get_flatview(as);
1683     FOR_EACH_FLAT_RANGE(fr, view) {
1684         MemoryRegionSection section = {
1685             .mr = fr->mr,
1686             .address_space = as,
1687             .offset_within_region = fr->offset_in_region,
1688             .size = fr->addr.size,
1689             .offset_within_address_space = int128_get64(fr->addr.start),
1690             .readonly = fr->readonly,
1691         };
1692         if (listener->region_add) {
1693             listener->region_add(listener, &section);
1694         }
1695     }
1696     flatview_unref(view);
1697 }
1698
1699 void memory_listener_register(MemoryListener *listener, AddressSpace *filter)
1700 {
1701     MemoryListener *other = NULL;
1702     AddressSpace *as;
1703
1704     listener->address_space_filter = filter;
1705     if (QTAILQ_EMPTY(&memory_listeners)
1706         || listener->priority >= QTAILQ_LAST(&memory_listeners,
1707                                              memory_listeners)->priority) {
1708         QTAILQ_INSERT_TAIL(&memory_listeners, listener, link);
1709     } else {
1710         QTAILQ_FOREACH(other, &memory_listeners, link) {
1711             if (listener->priority < other->priority) {
1712                 break;
1713             }
1714         }
1715         QTAILQ_INSERT_BEFORE(other, listener, link);
1716     }
1717
1718     QTAILQ_FOREACH(as, &address_spaces, address_spaces_link) {
1719         listener_add_address_space(listener, as);
1720     }
1721 }
1722
1723 void memory_listener_unregister(MemoryListener *listener)
1724 {
1725     QTAILQ_REMOVE(&memory_listeners, listener, link);
1726 }
1727
1728 void address_space_init(AddressSpace *as, MemoryRegion *root, const char *name)
1729 {
1730     if (QTAILQ_EMPTY(&address_spaces)) {
1731         memory_init();
1732     }
1733
1734     memory_region_transaction_begin();
1735     as->root = root;
1736     as->current_map = g_new(FlatView, 1);
1737     flatview_init(as->current_map);
1738     as->ioeventfd_nb = 0;
1739     as->ioeventfds = NULL;
1740     QTAILQ_INSERT_TAIL(&address_spaces, as, address_spaces_link);
1741     as->name = g_strdup(name ? name : "anonymous");
1742     address_space_init_dispatch(as);
1743     memory_region_update_pending |= root->enabled;
1744     memory_region_transaction_commit();
1745 }
1746
1747 void address_space_destroy(AddressSpace *as)
1748 {
1749     MemoryListener *listener;
1750
1751     /* Flush out anything from MemoryListeners listening in on this */
1752     memory_region_transaction_begin();
1753     as->root = NULL;
1754     memory_region_transaction_commit();
1755     QTAILQ_REMOVE(&address_spaces, as, address_spaces_link);
1756     address_space_destroy_dispatch(as);
1757
1758     QTAILQ_FOREACH(listener, &memory_listeners, link) {
1759         assert(listener->address_space_filter != as);
1760     }
1761
1762     flatview_unref(as->current_map);
1763     g_free(as->name);
1764     g_free(as->ioeventfds);
1765 }
1766
1767 bool io_mem_read(MemoryRegion *mr, hwaddr addr, uint64_t *pval, unsigned size)
1768 {
1769     return memory_region_dispatch_read(mr, addr, pval, size);
1770 }
1771
1772 bool io_mem_write(MemoryRegion *mr, hwaddr addr,
1773                   uint64_t val, unsigned size)
1774 {
1775     return memory_region_dispatch_write(mr, addr, val, size);
1776 }
1777
1778 typedef struct MemoryRegionList MemoryRegionList;
1779
1780 struct MemoryRegionList {
1781     const MemoryRegion *mr;
1782     bool printed;
1783     QTAILQ_ENTRY(MemoryRegionList) queue;
1784 };
1785
1786 typedef QTAILQ_HEAD(queue, MemoryRegionList) MemoryRegionListHead;
1787
1788 static void mtree_print_mr(fprintf_function mon_printf, void *f,
1789                            const MemoryRegion *mr, unsigned int level,
1790                            hwaddr base,
1791                            MemoryRegionListHead *alias_print_queue)
1792 {
1793     MemoryRegionList *new_ml, *ml, *next_ml;
1794     MemoryRegionListHead submr_print_queue;
1795     const MemoryRegion *submr;
1796     unsigned int i;
1797
1798     if (!mr || !mr->enabled) {
1799         return;
1800     }
1801
1802     for (i = 0; i < level; i++) {
1803         mon_printf(f, "  ");
1804     }
1805
1806     if (mr->alias) {
1807         MemoryRegionList *ml;
1808         bool found = false;
1809
1810         /* check if the alias is already in the queue */
1811         QTAILQ_FOREACH(ml, alias_print_queue, queue) {
1812             if (ml->mr == mr->alias && !ml->printed) {
1813                 found = true;
1814             }
1815         }
1816
1817         if (!found) {
1818             ml = g_new(MemoryRegionList, 1);
1819             ml->mr = mr->alias;
1820             ml->printed = false;
1821             QTAILQ_INSERT_TAIL(alias_print_queue, ml, queue);
1822         }
1823         mon_printf(f, TARGET_FMT_plx "-" TARGET_FMT_plx
1824                    " (prio %d, %c%c): alias %s @%s " TARGET_FMT_plx
1825                    "-" TARGET_FMT_plx "\n",
1826                    base + mr->addr,
1827                    base + mr->addr
1828                    + (int128_nz(mr->size) ?
1829                       (hwaddr)int128_get64(int128_sub(mr->size,
1830                                                       int128_one())) : 0),
1831                    mr->priority,
1832                    mr->romd_mode ? 'R' : '-',
1833                    !mr->readonly && !(mr->rom_device && mr->romd_mode) ? 'W'
1834                                                                        : '-',
1835                    mr->name,
1836                    mr->alias->name,
1837                    mr->alias_offset,
1838                    mr->alias_offset
1839                    + (int128_nz(mr->size) ?
1840                       (hwaddr)int128_get64(int128_sub(mr->size,
1841                                                       int128_one())) : 0));
1842     } else {
1843         mon_printf(f,
1844                    TARGET_FMT_plx "-" TARGET_FMT_plx " (prio %d, %c%c): %s\n",
1845                    base + mr->addr,
1846                    base + mr->addr
1847                    + (int128_nz(mr->size) ?
1848                       (hwaddr)int128_get64(int128_sub(mr->size,
1849                                                       int128_one())) : 0),
1850                    mr->priority,
1851                    mr->romd_mode ? 'R' : '-',
1852                    !mr->readonly && !(mr->rom_device && mr->romd_mode) ? 'W'
1853                                                                        : '-',
1854                    mr->name);
1855     }
1856
1857     QTAILQ_INIT(&submr_print_queue);
1858
1859     QTAILQ_FOREACH(submr, &mr->subregions, subregions_link) {
1860         new_ml = g_new(MemoryRegionList, 1);
1861         new_ml->mr = submr;
1862         QTAILQ_FOREACH(ml, &submr_print_queue, queue) {
1863             if (new_ml->mr->addr < ml->mr->addr ||
1864                 (new_ml->mr->addr == ml->mr->addr &&
1865                  new_ml->mr->priority > ml->mr->priority)) {
1866                 QTAILQ_INSERT_BEFORE(ml, new_ml, queue);
1867                 new_ml = NULL;
1868                 break;
1869             }
1870         }
1871         if (new_ml) {
1872             QTAILQ_INSERT_TAIL(&submr_print_queue, new_ml, queue);
1873         }
1874     }
1875
1876     QTAILQ_FOREACH(ml, &submr_print_queue, queue) {
1877         mtree_print_mr(mon_printf, f, ml->mr, level + 1, base + mr->addr,
1878                        alias_print_queue);
1879     }
1880
1881     QTAILQ_FOREACH_SAFE(ml, &submr_print_queue, queue, next_ml) {
1882         g_free(ml);
1883     }
1884 }
1885
1886 void mtree_info(fprintf_function mon_printf, void *f)
1887 {
1888     MemoryRegionListHead ml_head;
1889     MemoryRegionList *ml, *ml2;
1890     AddressSpace *as;
1891
1892     QTAILQ_INIT(&ml_head);
1893
1894     QTAILQ_FOREACH(as, &address_spaces, address_spaces_link) {
1895         mon_printf(f, "%s\n", as->name);
1896         mtree_print_mr(mon_printf, f, as->root, 0, 0, &ml_head);
1897     }
1898
1899     mon_printf(f, "aliases\n");
1900     /* print aliased regions */
1901     QTAILQ_FOREACH(ml, &ml_head, queue) {
1902         if (!ml->printed) {
1903             mon_printf(f, "%s\n", ml->mr->name);
1904             mtree_print_mr(mon_printf, f, ml->mr, 0, 0, &ml_head);
1905         }
1906     }
1907
1908     QTAILQ_FOREACH_SAFE(ml, &ml_head, queue, ml2) {
1909         g_free(ml);
1910     }
1911 }