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 #include "sysemu/sysemu.h"
  27
  28 //#define DEBUG_UNASSIGNED
  29
  30 static unsigned memory_region_transaction_depth;
  31 static bool memory_region_update_pending;
  32 static bool ioeventfd_update_pending;
  33 static bool global_dirty_log = false;
  34
  35 /* flat_view_mutex is taken around reading as->current_map; the critical
  36  * section is extremely short, so I'm using a single mutex for every AS.
  37  * We could also RCU for the read-side.
  38  *
  39  * The BQL is taken around transaction commits, hence both locks are taken
  40  * while writing to as->current_map (with the BQL taken outside).
  41  */
  42 static QemuMutex flat_view_mutex;
  43
  44 static QTAILQ_HEAD(memory_listeners, MemoryListener) memory_listeners
  45     = QTAILQ_HEAD_INITIALIZER(memory_listeners);
  46
  47 static QTAILQ_HEAD(, AddressSpace) address_spaces
  48     = QTAILQ_HEAD_INITIALIZER(address_spaces);
  49
  50 static void memory_init(void)
  51 {
  52     qemu_mutex_init(&flat_view_mutex);
  53 }
  54
  55 typedef struct AddrRange AddrRange;
  56
  57 /*
  58  * Note using signed integers limits us to physical addresses at most
  59  * 63 bits wide.  They are needed for negative offsetting in aliases
  60  * (large MemoryRegion::alias_offset).
  61  */
  62 struct AddrRange {
  63     Int128 start;
  64     Int128 size;
  65 };
  66
  67 static AddrRange addrrange_make(Int128 start, Int128 size)
  68 {
  69     return (AddrRange) { start, size };
  70 }
  71
  72 static bool addrrange_equal(AddrRange r1, AddrRange r2)
  73 {
  74     return int128_eq(r1.start, r2.start) && int128_eq(r1.size, r2.size);
  75 }
  76
  77 static Int128 addrrange_end(AddrRange r)
  78 {
  79     return int128_add(r.start, r.size);
  80 }
  81
  82 static AddrRange addrrange_shift(AddrRange range, Int128 delta)
  83 {
  84     int128_addto(&range.start, delta);
  85     return range;
  86 }
  87
  88 static bool addrrange_contains(AddrRange range, Int128 addr)
  89 {
  90     return int128_ge(addr, range.start)
  91         && int128_lt(addr, addrrange_end(range));
  92 }
  93
  94 static bool addrrange_intersects(AddrRange r1, AddrRange r2)
  95 {
  96     return addrrange_contains(r1, r2.start)
  97         || addrrange_contains(r2, r1.start);
  98 }
  99
 100 static AddrRange addrrange_intersection(AddrRange r1, AddrRange r2)
 101 {
 102     Int128 start = int128_max(r1.start, r2.start);
 103     Int128 end = int128_min(addrrange_end(r1), addrrange_end(r2));
 104     return addrrange_make(start, int128_sub(end, start));
 105 }
 106
 107 enum ListenerDirection { Forward, Reverse };
 108
 109 static bool memory_listener_match(MemoryListener *listener,
 110                                   MemoryRegionSection *section)
 111 {
 112     return !listener->address_space_filter
 113         || listener->address_space_filter == section->address_space;
 114 }
 115
 116 #define MEMORY_LISTENER_CALL_GLOBAL(_callback, _direction, _args...)    \
 117     do {                                                                \
 118         MemoryListener *_listener;                                      \
 119                                                                         \
 120         switch (_direction) {                                           \
 121         case Forward:                                                   \
 122             QTAILQ_FOREACH(_listener, &memory_listeners, link) {        \
 123                 if (_listener->_callback) {                             \
 124                     _listener->_callback(_listener, ##_args);           \
 125                 }                                                       \
 126             }                                                           \
 127             break;                                                      \
 128         case Reverse:                                                   \
 129             QTAILQ_FOREACH_REVERSE(_listener, &memory_listeners,        \
 130                                    memory_listeners, link) {            \
 131                 if (_listener->_callback) {                             \
 132                     _listener->_callback(_listener, ##_args);           \
 133                 }                                                       \
 134             }                                                           \
 135             break;                                                      \
 136         default:                                                        \
 137             abort();                                                    \
 138         }                                                               \
 139     } while (0)
 140
 141 #define MEMORY_LISTENER_CALL(_callback, _direction, _section, _args...) \
 142     do {                                                                \
 143         MemoryListener *_listener;                                      \
 144                                                                         \
 145         switch (_direction) {                                           \
 146         case Forward:                                                   \
 147             QTAILQ_FOREACH(_listener, &memory_listeners, link) {        \
 148                 if (_listener->_callback                                \
 149                     && memory_listener_match(_listener, _section)) {    \
 150                     _listener->_callback(_listener, _section, ##_args); \
 151                 }                                                       \
 152             }                                                           \
 153             break;                                                      \
 154         case Reverse:                                                   \
 155             QTAILQ_FOREACH_REVERSE(_listener, &memory_listeners,        \
 156                                    memory_listeners, link) {            \
 157                 if (_listener->_callback                                \
 158                     && memory_listener_match(_listener, _section)) {    \
 159                     _listener->_callback(_listener, _section, ##_args); \
 160                 }                                                       \
 161             }                                                           \
 162             break;                                                      \
 163         default:                                                        \
 164             abort();                                                    \
 165         }                                                               \
 166     } while (0)
 167
 168 /* No need to ref/unref .mr, the FlatRange keeps it alive.  */
 169 #define MEMORY_LISTENER_UPDATE_REGION(fr, as, dir, callback)            \
 170     MEMORY_LISTENER_CALL(callback, dir, (&(MemoryRegionSection) {       \
 171         .mr = (fr)->mr,                                                 \
 172         .address_space = (as),                                          \
 173         .offset_within_region = (fr)->offset_in_region,                 \
 174         .size = (fr)->addr.size,                                        \
 175         .offset_within_address_space = int128_get64((fr)->addr.start),  \
 176         .readonly = (fr)->readonly,                                     \
 177               }))
 178
 179 struct CoalescedMemoryRange {
 180     AddrRange addr;
 181     QTAILQ_ENTRY(CoalescedMemoryRange) link;
 182 };
 183
 184 struct MemoryRegionIoeventfd {
 185     AddrRange addr;
 186     bool match_data;
 187     uint64_t data;
 188     EventNotifier *e;
 189 };
 190
 191 static bool memory_region_ioeventfd_before(MemoryRegionIoeventfd a,
 192                                            MemoryRegionIoeventfd b)
 193 {
 194     if (int128_lt(a.addr.start, b.addr.start)) {
 195         return true;
 196     } else if (int128_gt(a.addr.start, b.addr.start)) {
 197         return false;
 198     } else if (int128_lt(a.addr.size, b.addr.size)) {
 199         return true;
 200     } else if (int128_gt(a.addr.size, b.addr.size)) {
 201         return false;
 202     } else if (a.match_data < b.match_data) {
 203         return true;
 204     } else  if (a.match_data > b.match_data) {
 205         return false;
 206     } else if (a.match_data) {
 207         if (a.data < b.data) {
 208             return true;
 209         } else if (a.data > b.data) {
 210             return false;
 211         }
 212     }
 213     if (a.e < b.e) {
 214         return true;
 215     } else if (a.e > b.e) {
 216         return false;
 217     }
 218     return false;
 219 }
 220
 221 static bool memory_region_ioeventfd_equal(MemoryRegionIoeventfd a,
 222                                           MemoryRegionIoeventfd b)
 223 {
 224     return !memory_region_ioeventfd_before(a, b)
 225         && !memory_region_ioeventfd_before(b, a);
 226 }
 227
 228 typedef struct FlatRange FlatRange;
 229 typedef struct FlatView FlatView;
 230
 231 /* Range of memory in the global map.  Addresses are absolute. */
 232 struct FlatRange {
 233     MemoryRegion *mr;
 234     hwaddr offset_in_region;
 235     AddrRange addr;
 236     uint8_t dirty_log_mask;
 237     bool romd_mode;
 238     bool readonly;
 239 };
 240
 241 /* Flattened global view of current active memory hierarchy.  Kept in sorted
 242  * order.
 243  */
 244 struct FlatView {
 245     unsigned ref;
 246     FlatRange *ranges;
 247     unsigned nr;
 248     unsigned nr_allocated;
 249 };
 250
 251 typedef struct AddressSpaceOps AddressSpaceOps;
 252
 253 #define FOR_EACH_FLAT_RANGE(var, view)          \
 254     for (var = (view)->ranges; var < (view)->ranges + (view)->nr; ++var)
 255
 256 static bool flatrange_equal(FlatRange *a, FlatRange *b)
 257 {
 258     return a->mr == b->mr
 259         && addrrange_equal(a->addr, b->addr)
 260         && a->offset_in_region == b->offset_in_region
 261         && a->romd_mode == b->romd_mode
 262         && a->readonly == b->readonly;
 263 }
 264
 265 static void flatview_init(FlatView *view)
 266 {
 267     view->ref = 1;
 268     view->ranges = NULL;
 269     view->nr = 0;
 270     view->nr_allocated = 0;
 271 }
 272
 273 /* Insert a range into a given position.  Caller is responsible for maintaining
 274  * sorting order.
 275  */
 276 static void flatview_insert(FlatView *view, unsigned pos, FlatRange *range)
 277 {
 278     if (view->nr == view->nr_allocated) {
 279         view->nr_allocated = MAX(2 * view->nr, 10);
 280         view->ranges = g_realloc(view->ranges,
 281                                     view->nr_allocated * sizeof(*view->ranges));
 282     }
 283     memmove(view->ranges + pos + 1, view->ranges + pos,
 284             (view->nr - pos) * sizeof(FlatRange));
 285     view->ranges[pos] = *range;
 286     memory_region_ref(range->mr);
 287     ++view->nr;
 288 }
 289
 290 static void flatview_destroy(FlatView *view)
 291 {
 292     int i;
 293
 294     for (i = 0; i < view->nr; i++) {
 295         memory_region_unref(view->ranges[i].mr);
 296     }
 297     g_free(view->ranges);
 298     g_free(view);
 299 }
 300
 301 static void flatview_ref(FlatView *view)
 302 {
 303     atomic_inc(&view->ref);
 304 }
 305
 306 static void flatview_unref(FlatView *view)
 307 {
 308     if (atomic_fetch_dec(&view->ref) == 1) {
 309         flatview_destroy(view);
 310     }
 311 }
 312
 313 static bool can_merge(FlatRange *r1, FlatRange *r2)
 314 {
 315     return int128_eq(addrrange_end(r1->addr), r2->addr.start)
 316         && r1->mr == r2->mr
 317         && int128_eq(int128_add(int128_make64(r1->offset_in_region),
 318                                 r1->addr.size),
 319                      int128_make64(r2->offset_in_region))
 320         && r1->dirty_log_mask == r2->dirty_log_mask
 321         && r1->romd_mode == r2->romd_mode
 322         && r1->readonly == r2->readonly;
 323 }
 324
 325 /* Attempt to simplify a view by merging adjacent ranges */
 326 static void flatview_simplify(FlatView *view)
 327 {
 328     unsigned i, j;
 329
 330     i = 0;
 331     while (i < view->nr) {
 332         j = i + 1;
 333         while (j < view->nr
 334                && can_merge(&view->ranges[j-1], &view->ranges[j])) {
 335             int128_addto(&view->ranges[i].addr.size, view->ranges[j].addr.size);
 336             ++j;
 337         }
 338         ++i;
 339         memmove(&view->ranges[i], &view->ranges[j],
 340                 (view->nr - j) * sizeof(view->ranges[j]));
 341         view->nr -= j - i;
 342     }
 343 }
 344
 345 static bool memory_region_big_endian(MemoryRegion *mr)
 346 {
 347 #ifdef TARGET_WORDS_BIGENDIAN
 348     return mr->ops->endianness != DEVICE_LITTLE_ENDIAN;
 349 #else
 350     return mr->ops->endianness == DEVICE_BIG_ENDIAN;
 351 #endif
 352 }
 353
 354 static bool memory_region_wrong_endianness(MemoryRegion *mr)
 355 {
 356 #ifdef TARGET_WORDS_BIGENDIAN
 357     return mr->ops->endianness == DEVICE_LITTLE_ENDIAN;
 358 #else
 359     return mr->ops->endianness == DEVICE_BIG_ENDIAN;
 360 #endif
 361 }
 362
 363 static void adjust_endianness(MemoryRegion *mr, uint64_t *data, unsigned size)
 364 {
 365     if (memory_region_wrong_endianness(mr)) {
 366         switch (size) {
 367         case 1:
 368             break;
 369         case 2:
 370             *data = bswap16(*data);
 371             break;
 372         case 4:
 373             *data = bswap32(*data);
 374             break;
 375         case 8:
 376             *data = bswap64(*data);
 377             break;
 378         default:
 379             abort();
 380         }
 381     }
 382 }
 383
 384 static void memory_region_oldmmio_read_accessor(MemoryRegion *mr,
 385                                                 hwaddr addr,
 386                                                 uint64_t *value,
 387                                                 unsigned size,
 388                                                 unsigned shift,
 389                                                 uint64_t mask)
 390 {
 391     uint64_t tmp;
 392
 393     tmp = mr->ops->old_mmio.read[ctz32(size)](mr->opaque, addr);
 394     trace_memory_region_ops_read(mr, addr, tmp, size);
 395     *value |= (tmp & mask) << shift;
 396 }
 397
 398 static void memory_region_read_accessor(MemoryRegion *mr,
 399                                         hwaddr addr,
 400                                         uint64_t *value,
 401                                         unsigned size,
 402                                         unsigned shift,
 403                                         uint64_t mask)
 404 {
 405     uint64_t tmp;
 406
 407     if (mr->flush_coalesced_mmio) {
 408         qemu_flush_coalesced_mmio_buffer();
 409     }
 410     tmp = mr->ops->read(mr->opaque, addr, size);
 411     trace_memory_region_ops_read(mr, addr, tmp, size);
 412     *value |= (tmp & mask) << shift;
 413 }
 414
 415 static void memory_region_oldmmio_write_accessor(MemoryRegion *mr,
 416                                                  hwaddr addr,
 417                                                  uint64_t *value,
 418                                                  unsigned size,
 419                                                  unsigned shift,
 420                                                  uint64_t mask)
 421 {
 422     uint64_t tmp;
 423
 424     tmp = (*value >> shift) & mask;
 425     trace_memory_region_ops_write(mr, addr, tmp, size);
 426     mr->ops->old_mmio.write[ctz32(size)](mr->opaque, addr, tmp);
 427 }
 428
 429 static void memory_region_write_accessor(MemoryRegion *mr,
 430                                          hwaddr addr,
 431                                          uint64_t *value,
 432                                          unsigned size,
 433                                          unsigned shift,
 434                                          uint64_t mask)
 435 {
 436     uint64_t tmp;
 437
 438     if (mr->flush_coalesced_mmio) {
 439         qemu_flush_coalesced_mmio_buffer();
 440     }
 441     tmp = (*value >> shift) & mask;
 442     trace_memory_region_ops_write(mr, addr, tmp, size);
 443     mr->ops->write(mr->opaque, addr, tmp, size);
 444 }
 445
 446 static void access_with_adjusted_size(hwaddr addr,
 447                                       uint64_t *value,
 448                                       unsigned size,
 449                                       unsigned access_size_min,
 450                                       unsigned access_size_max,
 451                                       void (*access)(MemoryRegion *mr,
 452                                                      hwaddr addr,
 453                                                      uint64_t *value,
 454                                                      unsigned size,
 455                                                      unsigned shift,
 456                                                      uint64_t mask),
 457                                       MemoryRegion *mr)
 458 {
 459     uint64_t access_mask;
 460     unsigned access_size;
 461     unsigned i;
 462
 463     if (!access_size_min) {
 464         access_size_min = 1;
 465     }
 466     if (!access_size_max) {
 467         access_size_max = 4;
 468     }
 469
 470     /* FIXME: support unaligned access? */
 471     access_size = MAX(MIN(size, access_size_max), access_size_min);
 472     access_mask = -1ULL >> (64 - access_size * 8);
 473     if (memory_region_big_endian(mr)) {
 474         for (i = 0; i < size; i += access_size) {
 475             access(mr, addr + i, value, access_size,
 476                    (size - access_size - i) * 8, access_mask);
 477         }
 478     } else {
 479         for (i = 0; i < size; i += access_size) {
 480             access(mr, addr + i, value, access_size, i * 8, access_mask);
 481         }
 482     }
 483 }
 484
 485 static AddressSpace *memory_region_to_address_space(MemoryRegion *mr)
 486 {
 487     AddressSpace *as;
 488
 489     while (mr->container) {
 490         mr = mr->container;
 491     }
 492     QTAILQ_FOREACH(as, &address_spaces, address_spaces_link) {
 493         if (mr == as->root) {
 494             return as;
 495         }
 496     }
 497     return NULL;
 498 }
 499
 500 /* Render a memory region into the global view.  Ranges in @view obscure
 501  * ranges in @mr.
 502  */
 503 static void render_memory_region(FlatView *view,
 504                                  MemoryRegion *mr,
 505                                  Int128 base,
 506                                  AddrRange clip,
 507                                  bool readonly)
 508 {
 509     MemoryRegion *subregion;
 510     unsigned i;
 511     hwaddr offset_in_region;
 512     Int128 remain;
 513     Int128 now;
 514     FlatRange fr;
 515     AddrRange tmp;
 516
 517     if (!mr->enabled) {
 518         return;
 519     }
 520
 521     int128_addto(&base, int128_make64(mr->addr));
 522     readonly |= mr->readonly;
 523
 524     tmp = addrrange_make(base, mr->size);
 525
 526     if (!addrrange_intersects(tmp, clip)) {
 527         return;
 528     }
 529
 530     clip = addrrange_intersection(tmp, clip);
 531
 532     if (mr->alias) {
 533         int128_subfrom(&base, int128_make64(mr->alias->addr));
 534         int128_subfrom(&base, int128_make64(mr->alias_offset));
 535         render_memory_region(view, mr->alias, base, clip, readonly);
 536         return;
 537     }
 538
 539     /* Render subregions in priority order. */
 540     QTAILQ_FOREACH(subregion, &mr->subregions, subregions_link) {
 541         render_memory_region(view, subregion, base, clip, readonly);
 542     }
 543
 544     if (!mr->terminates) {
 545         return;
 546     }
 547
 548     offset_in_region = int128_get64(int128_sub(clip.start, base));
 549     base = clip.start;
 550     remain = clip.size;
 551
 552     fr.mr = mr;
 553     fr.dirty_log_mask = mr->dirty_log_mask;
 554     fr.romd_mode = mr->romd_mode;
 555     fr.readonly = readonly;
 556
 557     /* Render the region itself into any gaps left by the current view. */
 558     for (i = 0; i < view->nr && int128_nz(remain); ++i) {
 559         if (int128_ge(base, addrrange_end(view->ranges[i].addr))) {
 560             continue;
 561         }
 562         if (int128_lt(base, view->ranges[i].addr.start)) {
 563             now = int128_min(remain,
 564                              int128_sub(view->ranges[i].addr.start, base));
 565             fr.offset_in_region = offset_in_region;
 566             fr.addr = addrrange_make(base, now);
 567             flatview_insert(view, i, &fr);
 568             ++i;
 569             int128_addto(&base, now);
 570             offset_in_region += int128_get64(now);
 571             int128_subfrom(&remain, now);
 572         }
 573         now = int128_sub(int128_min(int128_add(base, remain),
 574                                     addrrange_end(view->ranges[i].addr)),
 575                          base);
 576         int128_addto(&base, now);
 577         offset_in_region += int128_get64(now);
 578         int128_subfrom(&remain, now);
 579     }
 580     if (int128_nz(remain)) {
 581         fr.offset_in_region = offset_in_region;
 582         fr.addr = addrrange_make(base, remain);
 583         flatview_insert(view, i, &fr);
 584     }
 585 }
 586
 587 /* Render a memory topology into a list of disjoint absolute ranges. */
 588 static FlatView *generate_memory_topology(MemoryRegion *mr)
 589 {
 590     FlatView *view;
 591
 592     view = g_new(FlatView, 1);
 593     flatview_init(view);
 594
 595     if (mr) {
 596         render_memory_region(view, mr, int128_zero(),
 597                              addrrange_make(int128_zero(), int128_2_64()), false);
 598     }
 599     flatview_simplify(view);
 600
 601     return view;
 602 }
 603
 604 static void address_space_add_del_ioeventfds(AddressSpace *as,
 605                                              MemoryRegionIoeventfd *fds_new,
 606                                              unsigned fds_new_nb,
 607                                              MemoryRegionIoeventfd *fds_old,
 608                                              unsigned fds_old_nb)
 609 {
 610     unsigned iold, inew;
 611     MemoryRegionIoeventfd *fd;
 612     MemoryRegionSection section;
 613
 614     /* Generate a symmetric difference of the old and new fd sets, adding
 615      * and deleting as necessary.
 616      */
 617
 618     iold = inew = 0;
 619     while (iold < fds_old_nb || inew < fds_new_nb) {
 620         if (iold < fds_old_nb
 621             && (inew == fds_new_nb
 622                 || memory_region_ioeventfd_before(fds_old[iold],
 623                                                   fds_new[inew]))) {
 624             fd = &fds_old[iold];
 625             section = (MemoryRegionSection) {
 626                 .address_space = as,
 627                 .offset_within_address_space = int128_get64(fd->addr.start),
 628                 .size = fd->addr.size,
 629             };
 630             MEMORY_LISTENER_CALL(eventfd_del, Forward, &section,
 631                                  fd->match_data, fd->data, fd->e);
 632             ++iold;
 633         } else if (inew < fds_new_nb
 634                    && (iold == fds_old_nb
 635                        || memory_region_ioeventfd_before(fds_new[inew],
 636                                                          fds_old[iold]))) {
 637             fd = &fds_new[inew];
 638             section = (MemoryRegionSection) {
 639                 .address_space = as,
 640                 .offset_within_address_space = int128_get64(fd->addr.start),
 641                 .size = fd->addr.size,
 642             };
 643             MEMORY_LISTENER_CALL(eventfd_add, Reverse, &section,
 644                                  fd->match_data, fd->data, fd->e);
 645             ++inew;
 646         } else {
 647             ++iold;
 648             ++inew;
 649         }
 650     }
 651 }
 652
 653 static FlatView *address_space_get_flatview(AddressSpace *as)
 654 {
 655     FlatView *view;
 656
 657     qemu_mutex_lock(&flat_view_mutex);
 658     view = as->current_map;
 659     flatview_ref(view);
 660     qemu_mutex_unlock(&flat_view_mutex);
 661     return view;
 662 }
 663
 664 static void address_space_update_ioeventfds(AddressSpace *as)
 665 {
 666     FlatView *view;
 667     FlatRange *fr;
 668     unsigned ioeventfd_nb = 0;
 669     MemoryRegionIoeventfd *ioeventfds = NULL;
 670     AddrRange tmp;
 671     unsigned i;
 672
 673     view = address_space_get_flatview(as);
 674     FOR_EACH_FLAT_RANGE(fr, view) {
 675         for (i = 0; i < fr->mr->ioeventfd_nb; ++i) {
 676             tmp = addrrange_shift(fr->mr->ioeventfds[i].addr,
 677                                   int128_sub(fr->addr.start,
 678                                              int128_make64(fr->offset_in_region)));
 679             if (addrrange_intersects(fr->addr, tmp)) {
 680                 ++ioeventfd_nb;
 681                 ioeventfds = g_realloc(ioeventfds,
 682                                           ioeventfd_nb * sizeof(*ioeventfds));
 683                 ioeventfds[ioeventfd_nb-1] = fr->mr->ioeventfds[i];
 684                 ioeventfds[ioeventfd_nb-1].addr = tmp;
 685             }
 686         }
 687     }
 688
 689     address_space_add_del_ioeventfds(as, ioeventfds, ioeventfd_nb,
 690                                      as->ioeventfds, as->ioeventfd_nb);
 691
 692     g_free(as->ioeventfds);
 693     as->ioeventfds = ioeventfds;
 694     as->ioeventfd_nb = ioeventfd_nb;
 695     flatview_unref(view);
 696 }
 697
 698 static void address_space_update_topology_pass(AddressSpace *as,
 699                                                const FlatView *old_view,
 700                                                const FlatView *new_view,
 701                                                bool adding)
 702 {
 703     unsigned iold, inew;
 704     FlatRange *frold, *frnew;
 705
 706     /* Generate a symmetric difference of the old and new memory maps.
 707      * Kill ranges in the old map, and instantiate ranges in the new map.
 708      */
 709     iold = inew = 0;
 710     while (iold < old_view->nr || inew < new_view->nr) {
 711         if (iold < old_view->nr) {
 712             frold = &old_view->ranges[iold];
 713         } else {
 714             frold = NULL;
 715         }
 716         if (inew < new_view->nr) {
 717             frnew = &new_view->ranges[inew];
 718         } else {
 719             frnew = NULL;
 720         }
 721
 722         if (frold
 723             && (!frnew
 724                 || int128_lt(frold->addr.start, frnew->addr.start)
 725                 || (int128_eq(frold->addr.start, frnew->addr.start)
 726                     && !flatrange_equal(frold, frnew)))) {
 727             /* In old but not in new, or in both but attributes changed. */
 728
 729             if (!adding) {
 730                 MEMORY_LISTENER_UPDATE_REGION(frold, as, Reverse, region_del);
 731             }
 732
 733             ++iold;
 734         } else if (frold && frnew && flatrange_equal(frold, frnew)) {
 735             /* In both and unchanged (except logging may have changed) */
 736
 737             if (adding) {
 738                 MEMORY_LISTENER_UPDATE_REGION(frnew, as, Forward, region_nop);
 739                 if (frold->dirty_log_mask && !frnew->dirty_log_mask) {
 740                     MEMORY_LISTENER_UPDATE_REGION(frnew, as, Reverse, log_stop);
 741                 } else if (frnew->dirty_log_mask && !frold->dirty_log_mask) {
 742                     MEMORY_LISTENER_UPDATE_REGION(frnew, as, Forward, log_start);
 743                 }
 744             }
 745
 746             ++iold;
 747             ++inew;
 748         } else {
 749             /* In new */
 750
 751             if (adding) {
 752                 MEMORY_LISTENER_UPDATE_REGION(frnew, as, Forward, region_add);
 753             }
 754
 755             ++inew;
 756         }
 757     }
 758 }
 759
 760
 761 static void address_space_update_topology(AddressSpace *as)
 762 {
 763     FlatView *old_view = address_space_get_flatview(as);
 764     FlatView *new_view = generate_memory_topology(as->root);
 765
 766     address_space_update_topology_pass(as, old_view, new_view, false);
 767     address_space_update_topology_pass(as, old_view, new_view, true);
 768
 769     qemu_mutex_lock(&flat_view_mutex);
 770     flatview_unref(as->current_map);
 771     as->current_map = new_view;
 772     qemu_mutex_unlock(&flat_view_mutex);
 773
 774     /* Note that all the old MemoryRegions are still alive up to this
 775      * point.  This relieves most MemoryListeners from the need to
 776      * ref/unref the MemoryRegions they get---unless they use them
 777      * outside the iothread mutex, in which case precise reference
 778      * counting is necessary.
 779      */
 780     flatview_unref(old_view);
 781
 782     address_space_update_ioeventfds(as);
 783 }
 784
 785 void memory_region_transaction_begin(void)
 786 {
 787     qemu_flush_coalesced_mmio_buffer();
 788     ++memory_region_transaction_depth;
 789 }
 790
 791 static void memory_region_clear_pending(void)
 792 {
 793     memory_region_update_pending = false;
 794     ioeventfd_update_pending = false;
 795 }
 796
 797 void memory_region_transaction_commit(void)
 798 {
 799     AddressSpace *as;
 800
 801     assert(memory_region_transaction_depth);
 802     --memory_region_transaction_depth;
 803     if (!memory_region_transaction_depth) {
 804         if (memory_region_update_pending) {
 805             MEMORY_LISTENER_CALL_GLOBAL(begin, Forward);
 806
 807             QTAILQ_FOREACH(as, &address_spaces, address_spaces_link) {
 808                 address_space_update_topology(as);
 809             }
 810
 811             MEMORY_LISTENER_CALL_GLOBAL(commit, Forward);
 812         } else if (ioeventfd_update_pending) {
 813             QTAILQ_FOREACH(as, &address_spaces, address_spaces_link) {
 814                 address_space_update_ioeventfds(as);
 815             }
 816         }
 817         memory_region_clear_pending();
 818    }
 819 }
 820
 821 static void memory_region_destructor_none(MemoryRegion *mr)
 822 {
 823 }
 824
 825 static void memory_region_destructor_ram(MemoryRegion *mr)
 826 {
 827     qemu_ram_free(mr->ram_addr);
 828 }
 829
 830 static void memory_region_destructor_alias(MemoryRegion *mr)
 831 {
 832     memory_region_unref(mr->alias);
 833 }
 834
 835 static void memory_region_destructor_ram_from_ptr(MemoryRegion *mr)
 836 {
 837     qemu_ram_free_from_ptr(mr->ram_addr);
 838 }
 839
 840 static void memory_region_destructor_rom_device(MemoryRegion *mr)
 841 {
 842     qemu_ram_free(mr->ram_addr & TARGET_PAGE_MASK);
 843 }
 844
 845 void memory_region_init(MemoryRegion *mr,
 846                         Object *owner,
 847                         const char *name,
 848                         uint64_t size)
 849 {
 850     mr->ops = &unassigned_mem_ops;
 851     mr->opaque = NULL;
 852     mr->owner = owner;
 853     mr->iommu_ops = NULL;
 854     mr->container = NULL;
 855     mr->size = int128_make64(size);
 856     if (size == UINT64_MAX) {
 857         mr->size = int128_2_64();
 858     }
 859     mr->addr = 0;
 860     mr->subpage = false;
 861     mr->enabled = true;
 862     mr->terminates = false;
 863     mr->ram = false;
 864     mr->romd_mode = true;
 865     mr->readonly = false;
 866     mr->rom_device = false;
 867     mr->destructor = memory_region_destructor_none;
 868     mr->priority = 0;
 869     mr->may_overlap = false;
 870     mr->alias = NULL;
 871     QTAILQ_INIT(&mr->subregions);
 872     memset(&mr->subregions_link, 0, sizeof mr->subregions_link);
 873     QTAILQ_INIT(&mr->coalesced);
 874     mr->name = g_strdup(name);
 875     mr->dirty_log_mask = 0;
 876     mr->ioeventfd_nb = 0;
 877     mr->ioeventfds = NULL;
 878     mr->flush_coalesced_mmio = false;
 879 }
 880
 881 static uint64_t unassigned_mem_read(void *opaque, hwaddr addr,
 882                                     unsigned size)
 883 {
 884 #ifdef DEBUG_UNASSIGNED
 885     printf("Unassigned mem read " TARGET_FMT_plx "\n", addr);
 886 #endif
 887     if (current_cpu != NULL) {
 888         cpu_unassigned_access(current_cpu, addr, false, false, 0, size);
 889     }
 890     return 0;
 891 }
 892
 893 static void unassigned_mem_write(void *opaque, hwaddr addr,
 894                                  uint64_t val, unsigned size)
 895 {
 896 #ifdef DEBUG_UNASSIGNED
 897     printf("Unassigned mem write " TARGET_FMT_plx " = 0x%"PRIx64"\n", addr, val);
 898 #endif
 899     if (current_cpu != NULL) {
 900         cpu_unassigned_access(current_cpu, addr, true, false, 0, size);
 901     }
 902 }
 903
 904 static bool unassigned_mem_accepts(void *opaque, hwaddr addr,
 905                                    unsigned size, bool is_write)
 906 {
 907     return false;
 908 }
 909
 910 const MemoryRegionOps unassigned_mem_ops = {
 911     .valid.accepts = unassigned_mem_accepts,
 912     .endianness = DEVICE_NATIVE_ENDIAN,
 913 };
 914
 915 bool memory_region_access_valid(MemoryRegion *mr,
 916                                 hwaddr addr,
 917                                 unsigned size,
 918                                 bool is_write)
 919 {
 920     int access_size_min, access_size_max;
 921     int access_size, i;
 922
 923     if (!mr->ops->valid.unaligned && (addr & (size - 1))) {
 924         return false;
 925     }
 926
 927     if (!mr->ops->valid.accepts) {
 928         return true;
 929     }
 930
 931     access_size_min = mr->ops->valid.min_access_size;
 932     if (!mr->ops->valid.min_access_size) {
 933         access_size_min = 1;
 934     }
 935
 936     access_size_max = mr->ops->valid.max_access_size;
 937     if (!mr->ops->valid.max_access_size) {
 938         access_size_max = 4;
 939     }
 940
 941     access_size = MAX(MIN(size, access_size_max), access_size_min);
 942     for (i = 0; i < size; i += access_size) {
 943         if (!mr->ops->valid.accepts(mr->opaque, addr + i, access_size,
 944                                     is_write)) {
 945             return false;
 946         }
 947     }
 948
 949     return true;
 950 }
 951
 952 static uint64_t memory_region_dispatch_read1(MemoryRegion *mr,
 953                                              hwaddr addr,
 954                                              unsigned size)
 955 {
 956     uint64_t data = 0;
 957
 958     if (mr->ops->read) {
 959         access_with_adjusted_size(addr, &data, size,
 960                                   mr->ops->impl.min_access_size,
 961                                   mr->ops->impl.max_access_size,
 962                                   memory_region_read_accessor, mr);
 963     } else {
 964         access_with_adjusted_size(addr, &data, size, 1, 4,
 965                                   memory_region_oldmmio_read_accessor, mr);
 966     }
 967
 968     return data;
 969 }
 970
 971 static bool memory_region_dispatch_read(MemoryRegion *mr,
 972                                         hwaddr addr,
 973                                         uint64_t *pval,
 974                                         unsigned size)
 975 {
 976     if (!memory_region_access_valid(mr, addr, size, false)) {
 977         *pval = unassigned_mem_read(mr, addr, size);
 978         return true;
 979     }
 980
 981     *pval = memory_region_dispatch_read1(mr, addr, size);
 982     adjust_endianness(mr, pval, size);
 983     return false;
 984 }
 985
 986 static bool memory_region_dispatch_write(MemoryRegion *mr,
 987                                          hwaddr addr,
 988                                          uint64_t data,
 989                                          unsigned size)
 990 {
 991     if (!memory_region_access_valid(mr, addr, size, true)) {
 992         unassigned_mem_write(mr, addr, data, size);
 993         return true;
 994     }
 995
 996     adjust_endianness(mr, &data, size);
 997
 998     if (mr->ops->write) {
 999         access_with_adjusted_size(addr, &data, size,
1000                                   mr->ops->impl.min_access_size,
1001                                   mr->ops->impl.max_access_size,
1002                                   memory_region_write_accessor, mr);
1003     } else {
1004         access_with_adjusted_size(addr, &data, size, 1, 4,
1005                                   memory_region_oldmmio_write_accessor, mr);
1006     }
1007     return false;
1008 }
1009
1010 void memory_region_init_io(MemoryRegion *mr,
1011                            Object *owner,
1012                            const MemoryRegionOps *ops,
1013                            void *opaque,
1014                            const char *name,
1015                            uint64_t size)
1016 {
1017     memory_region_init(mr, owner, name, size);
1018     mr->ops = ops;
1019     mr->opaque = opaque;
1020     mr->terminates = true;
1021     mr->ram_addr = ~(ram_addr_t)0;
1022 }
1023
1024 void memory_region_init_ram(MemoryRegion *mr,
1025                             Object *owner,
1026                             const char *name,
1027                             uint64_t size)
1028 {
1029     memory_region_init(mr, owner, name, size);
1030     mr->ram = true;
1031     mr->terminates = true;
1032     mr->destructor = memory_region_destructor_ram;
1033     mr->ram_addr = qemu_ram_alloc(size, mr);
1034 }
1035
1036 #ifdef __linux__
1037 void memory_region_init_ram_from_file(MemoryRegion *mr,
1038                                       struct Object *owner,
1039                                       const char *name,
1040                                       uint64_t size,
1041                                       bool share,
1042                                       const char *path,
1043                                       Error **errp)
1044 {
1045     memory_region_init(mr, owner, name, size);
1046     mr->ram = true;
1047     mr->terminates = true;
1048     mr->destructor = memory_region_destructor_ram;
1049     mr->ram_addr = qemu_ram_alloc_from_file(size, mr, share, path, errp);
1050 }
1051 #endif
1052
1053 void memory_region_init_ram_ptr(MemoryRegion *mr,
1054                                 Object *owner,
1055                                 const char *name,
1056                                 uint64_t size,
1057                                 void *ptr)
1058 {
1059     memory_region_init(mr, owner, name, size);
1060     mr->ram = true;
1061     mr->terminates = true;
1062     mr->destructor = memory_region_destructor_ram_from_ptr;
1063     mr->ram_addr = qemu_ram_alloc_from_ptr(size, ptr, mr);
1064 }
1065
1066 void memory_region_init_alias(MemoryRegion *mr,
1067                               Object *owner,
1068                               const char *name,
1069                               MemoryRegion *orig,
1070                               hwaddr offset,
1071                               uint64_t size)
1072 {
1073     memory_region_init(mr, owner, name, size);
1074     memory_region_ref(orig);
1075     mr->destructor = memory_region_destructor_alias;
1076     mr->alias = orig;
1077     mr->alias_offset = offset;
1078 }
1079
1080 void memory_region_init_rom_device(MemoryRegion *mr,
1081                                    Object *owner,
1082                                    const MemoryRegionOps *ops,
1083                                    void *opaque,
1084                                    const char *name,
1085                                    uint64_t size)
1086 {
1087     memory_region_init(mr, owner, name, size);
1088     mr->ops = ops;
1089     mr->opaque = opaque;
1090     mr->terminates = true;
1091     mr->rom_device = true;
1092     mr->destructor = memory_region_destructor_rom_device;
1093     mr->ram_addr = qemu_ram_alloc(size, mr);
1094 }
1095
1096 void memory_region_init_iommu(MemoryRegion *mr,
1097                               Object *owner,
1098                               const MemoryRegionIOMMUOps *ops,
1099                               const char *name,
1100                               uint64_t size)
1101 {
1102     memory_region_init(mr, owner, name, size);
1103     mr->iommu_ops = ops,
1104     mr->terminates = true;  /* then re-forwards */
1105     notifier_list_init(&mr->iommu_notify);
1106 }
1107
1108 void memory_region_init_reservation(MemoryRegion *mr,
1109                                     Object *owner,
1110                                     const char *name,
1111                                     uint64_t size)
1112 {
1113     memory_region_init_io(mr, owner, &unassigned_mem_ops, mr, name, size);
1114 }
1115
1116 void memory_region_destroy(MemoryRegion *mr)
1117 {
1118     assert(QTAILQ_EMPTY(&mr->subregions));
1119     assert(memory_region_transaction_depth == 0);
1120     mr->destructor(mr);
1121     memory_region_clear_coalescing(mr);
1122     g_free((char *)mr->name);
1123     g_free(mr->ioeventfds);
1124 }
1125
1126 Object *memory_region_owner(MemoryRegion *mr)
1127 {
1128     return mr->owner;
1129 }
1130
1131 void memory_region_ref(MemoryRegion *mr)
1132 {
1133     if (mr && mr->owner) {
1134         object_ref(mr->owner);
1135     }
1136 }
1137
1138 void memory_region_unref(MemoryRegion *mr)
1139 {
1140     if (mr && mr->owner) {
1141         object_unref(mr->owner);
1142     }
1143 }
1144
1145 uint64_t memory_region_size(MemoryRegion *mr)
1146 {
1147     if (int128_eq(mr->size, int128_2_64())) {
1148         return UINT64_MAX;
1149     }
1150     return int128_get64(mr->size);
1151 }
1152
1153 const char *memory_region_name(MemoryRegion *mr)
1154 {
1155     return mr->name;
1156 }
1157
1158 bool memory_region_is_ram(MemoryRegion *mr)
1159 {
1160     return mr->ram;
1161 }
1162
1163 bool memory_region_is_logging(MemoryRegion *mr)
1164 {
1165     return mr->dirty_log_mask;
1166 }
1167
1168 bool memory_region_is_rom(MemoryRegion *mr)
1169 {
1170     return mr->ram && mr->readonly;
1171 }
1172
1173 bool memory_region_is_iommu(MemoryRegion *mr)
1174 {
1175     return mr->iommu_ops;
1176 }
1177
1178 void memory_region_register_iommu_notifier(MemoryRegion *mr, Notifier *n)
1179 {
1180     notifier_list_add(&mr->iommu_notify, n);
1181 }
1182
1183 void memory_region_unregister_iommu_notifier(Notifier *n)
1184 {
1185     notifier_remove(n);
1186 }
1187
1188 void memory_region_notify_iommu(MemoryRegion *mr,
1189                                 IOMMUTLBEntry entry)
1190 {
1191     assert(memory_region_is_iommu(mr));
1192     notifier_list_notify(&mr->iommu_notify, &entry);
1193 }
1194
1195 void memory_region_set_log(MemoryRegion *mr, bool log, unsigned client)
1196 {
1197     uint8_t mask = 1 << client;
1198
1199     memory_region_transaction_begin();
1200     mr->dirty_log_mask = (mr->dirty_log_mask & ~mask) | (log * mask);
1201     memory_region_update_pending |= mr->enabled;
1202     memory_region_transaction_commit();
1203 }
1204
1205 bool memory_region_get_dirty(MemoryRegion *mr, hwaddr addr,
1206                              hwaddr size, unsigned client)
1207 {
1208     assert(mr->terminates);
1209     return cpu_physical_memory_get_dirty(mr->ram_addr + addr, size, client);
1210 }
1211
1212 void memory_region_set_dirty(MemoryRegion *mr, hwaddr addr,
1213                              hwaddr size)
1214 {
1215     assert(mr->terminates);
1216     cpu_physical_memory_set_dirty_range(mr->ram_addr + addr, size);
1217 }
1218
1219 bool memory_region_test_and_clear_dirty(MemoryRegion *mr, hwaddr addr,
1220                                         hwaddr size, unsigned client)
1221 {
1222     bool ret;
1223     assert(mr->terminates);
1224     ret = cpu_physical_memory_get_dirty(mr->ram_addr + addr, size, client);
1225     if (ret) {
1226         cpu_physical_memory_reset_dirty(mr->ram_addr + addr, size, client);
1227     }
1228     return ret;
1229 }
1230
1231
1232 void memory_region_sync_dirty_bitmap(MemoryRegion *mr)
1233 {
1234     AddressSpace *as;
1235     FlatRange *fr;
1236
1237     QTAILQ_FOREACH(as, &address_spaces, address_spaces_link) {
1238         FlatView *view = address_space_get_flatview(as);
1239         FOR_EACH_FLAT_RANGE(fr, view) {
1240             if (fr->mr == mr) {
1241                 MEMORY_LISTENER_UPDATE_REGION(fr, as, Forward, log_sync);
1242             }
1243         }
1244         flatview_unref(view);
1245     }
1246 }
1247
1248 void memory_region_set_readonly(MemoryRegion *mr, bool readonly)
1249 {
1250     if (mr->readonly != readonly) {
1251         memory_region_transaction_begin();
1252         mr->readonly = readonly;
1253         memory_region_update_pending |= mr->enabled;
1254         memory_region_transaction_commit();
1255     }
1256 }
1257
1258 void memory_region_rom_device_set_romd(MemoryRegion *mr, bool romd_mode)
1259 {
1260     if (mr->romd_mode != romd_mode) {
1261         memory_region_transaction_begin();
1262         mr->romd_mode = romd_mode;
1263         memory_region_update_pending |= mr->enabled;
1264         memory_region_transaction_commit();
1265     }
1266 }
1267
1268 void memory_region_reset_dirty(MemoryRegion *mr, hwaddr addr,
1269                                hwaddr size, unsigned client)
1270 {
1271     assert(mr->terminates);
1272     cpu_physical_memory_reset_dirty(mr->ram_addr + addr, size, client);
1273 }
1274
1275 int memory_region_get_fd(MemoryRegion *mr)
1276 {
1277     if (mr->alias) {
1278         return memory_region_get_fd(mr->alias);
1279     }
1280
1281     assert(mr->terminates);
1282
1283     return qemu_get_ram_fd(mr->ram_addr & TARGET_PAGE_MASK);
1284 }
1285
1286 void *memory_region_get_ram_ptr(MemoryRegion *mr)
1287 {
1288     if (mr->alias) {
1289         return memory_region_get_ram_ptr(mr->alias) + mr->alias_offset;
1290     }
1291
1292     assert(mr->terminates);
1293
1294     return qemu_get_ram_ptr(mr->ram_addr & TARGET_PAGE_MASK);
1295 }
1296
1297 static void memory_region_update_coalesced_range_as(MemoryRegion *mr, AddressSpace *as)
1298 {
1299     FlatView *view;
1300     FlatRange *fr;
1301     CoalescedMemoryRange *cmr;
1302     AddrRange tmp;
1303     MemoryRegionSection section;
1304
1305     view = address_space_get_flatview(as);
1306     FOR_EACH_FLAT_RANGE(fr, view) {
1307         if (fr->mr == mr) {
1308             section = (MemoryRegionSection) {
1309                 .address_space = as,
1310                 .offset_within_address_space = int128_get64(fr->addr.start),
1311                 .size = fr->addr.size,
1312             };
1313
1314             MEMORY_LISTENER_CALL(coalesced_mmio_del, Reverse, &section,
1315                                  int128_get64(fr->addr.start),
1316                                  int128_get64(fr->addr.size));
1317             QTAILQ_FOREACH(cmr, &mr->coalesced, link) {
1318                 tmp = addrrange_shift(cmr->addr,
1319                                       int128_sub(fr->addr.start,
1320                                                  int128_make64(fr->offset_in_region)));
1321                 if (!addrrange_intersects(tmp, fr->addr)) {
1322                     continue;
1323                 }
1324                 tmp = addrrange_intersection(tmp, fr->addr);
1325                 MEMORY_LISTENER_CALL(coalesced_mmio_add, Forward, &section,
1326                                      int128_get64(tmp.start),
1327                                      int128_get64(tmp.size));
1328             }
1329         }
1330     }
1331     flatview_unref(view);
1332 }
1333
1334 static void memory_region_update_coalesced_range(MemoryRegion *mr)
1335 {
1336     AddressSpace *as;
1337
1338     QTAILQ_FOREACH(as, &address_spaces, address_spaces_link) {
1339         memory_region_update_coalesced_range_as(mr, as);
1340     }
1341 }
1342
1343 void memory_region_set_coalescing(MemoryRegion *mr)
1344 {
1345     memory_region_clear_coalescing(mr);
1346     memory_region_add_coalescing(mr, 0, int128_get64(mr->size));
1347 }
1348
1349 void memory_region_add_coalescing(MemoryRegion *mr,
1350                                   hwaddr offset,
1351                                   uint64_t size)
1352 {
1353     CoalescedMemoryRange *cmr = g_malloc(sizeof(*cmr));
1354
1355     cmr->addr = addrrange_make(int128_make64(offset), int128_make64(size));
1356     QTAILQ_INSERT_TAIL(&mr->coalesced, cmr, link);
1357     memory_region_update_coalesced_range(mr);
1358     memory_region_set_flush_coalesced(mr);
1359 }
1360
1361 void memory_region_clear_coalescing(MemoryRegion *mr)
1362 {
1363     CoalescedMemoryRange *cmr;
1364     bool updated = false;
1365
1366     qemu_flush_coalesced_mmio_buffer();
1367     mr->flush_coalesced_mmio = false;
1368
1369     while (!QTAILQ_EMPTY(&mr->coalesced)) {
1370         cmr = QTAILQ_FIRST(&mr->coalesced);
1371         QTAILQ_REMOVE(&mr->coalesced, cmr, link);
1372         g_free(cmr);
1373         updated = true;
1374     }
1375
1376     if (updated) {
1377         memory_region_update_coalesced_range(mr);
1378     }
1379 }
1380
1381 void memory_region_set_flush_coalesced(MemoryRegion *mr)
1382 {
1383     mr->flush_coalesced_mmio = true;
1384 }
1385
1386 void memory_region_clear_flush_coalesced(MemoryRegion *mr)
1387 {
1388     qemu_flush_coalesced_mmio_buffer();
1389     if (QTAILQ_EMPTY(&mr->coalesced)) {
1390         mr->flush_coalesced_mmio = false;
1391     }
1392 }
1393
1394 void memory_region_add_eventfd(MemoryRegion *mr,
1395                                hwaddr addr,
1396                                unsigned size,
1397                                bool match_data,
1398                                uint64_t data,
1399                                EventNotifier *e)
1400 {
1401     MemoryRegionIoeventfd mrfd = {
1402         .addr.start = int128_make64(addr),
1403         .addr.size = int128_make64(size),
1404         .match_data = match_data,
1405         .data = data,
1406         .e = e,
1407     };
1408     unsigned i;
1409
1410     adjust_endianness(mr, &mrfd.data, size);
1411     memory_region_transaction_begin();
1412     for (i = 0; i < mr->ioeventfd_nb; ++i) {
1413         if (memory_region_ioeventfd_before(mrfd, mr->ioeventfds[i])) {
1414             break;
1415         }
1416     }
1417     ++mr->ioeventfd_nb;
1418     mr->ioeventfds = g_realloc(mr->ioeventfds,
1419                                   sizeof(*mr->ioeventfds) * mr->ioeventfd_nb);
1420     memmove(&mr->ioeventfds[i+1], &mr->ioeventfds[i],
1421             sizeof(*mr->ioeventfds) * (mr->ioeventfd_nb-1 - i));
1422     mr->ioeventfds[i] = mrfd;
1423     ioeventfd_update_pending |= mr->enabled;
1424     memory_region_transaction_commit();
1425 }
1426
1427 void memory_region_del_eventfd(MemoryRegion *mr,
1428                                hwaddr addr,
1429                                unsigned size,
1430                                bool match_data,
1431                                uint64_t data,
1432                                EventNotifier *e)
1433 {
1434     MemoryRegionIoeventfd mrfd = {
1435         .addr.start = int128_make64(addr),
1436         .addr.size = int128_make64(size),
1437         .match_data = match_data,
1438         .data = data,
1439         .e = e,
1440     };
1441     unsigned i;
1442
1443     adjust_endianness(mr, &mrfd.data, size);
1444     memory_region_transaction_begin();
1445     for (i = 0; i < mr->ioeventfd_nb; ++i) {
1446         if (memory_region_ioeventfd_equal(mrfd, mr->ioeventfds[i])) {
1447             break;
1448         }
1449     }
1450     assert(i != mr->ioeventfd_nb);
1451     memmove(&mr->ioeventfds[i], &mr->ioeventfds[i+1],
1452             sizeof(*mr->ioeventfds) * (mr->ioeventfd_nb - (i+1)));
1453     --mr->ioeventfd_nb;
1454     mr->ioeventfds = g_realloc(mr->ioeventfds,
1455                                   sizeof(*mr->ioeventfds)*mr->ioeventfd_nb + 1);
1456     ioeventfd_update_pending |= mr->enabled;
1457     memory_region_transaction_commit();
1458 }
1459
1460 static void memory_region_update_container_subregions(MemoryRegion *subregion)
1461 {
1462     hwaddr offset = subregion->addr;
1463     MemoryRegion *mr = subregion->container;
1464     MemoryRegion *other;
1465
1466     memory_region_transaction_begin();
1467
1468     memory_region_ref(subregion);
1469     QTAILQ_FOREACH(other, &mr->subregions, subregions_link) {
1470         if (subregion->may_overlap || other->may_overlap) {
1471             continue;
1472         }
1473         if (int128_ge(int128_make64(offset),
1474                       int128_add(int128_make64(other->addr), other->size))
1475             || int128_le(int128_add(int128_make64(offset), subregion->size),
1476                          int128_make64(other->addr))) {
1477             continue;
1478         }
1479 #if 0
1480         printf("warning: subregion collision %llx/%llx (%s) "
1481                "vs %llx/%llx (%s)\n",
1482                (unsigned long long)offset,
1483                (unsigned long long)int128_get64(subregion->size),
1484                subregion->name,
1485                (unsigned long long)other->addr,
1486                (unsigned long long)int128_get64(other->size),
1487                other->name);
1488 #endif
1489     }
1490     QTAILQ_FOREACH(other, &mr->subregions, subregions_link) {
1491         if (subregion->priority >= other->priority) {
1492             QTAILQ_INSERT_BEFORE(other, subregion, subregions_link);
1493             goto done;
1494         }
1495     }
1496     QTAILQ_INSERT_TAIL(&mr->subregions, subregion, subregions_link);
1497 done:
1498     memory_region_update_pending |= mr->enabled && subregion->enabled;
1499     memory_region_transaction_commit();
1500 }
1501
1502 static void memory_region_add_subregion_common(MemoryRegion *mr,
1503                                                hwaddr offset,
1504                                                MemoryRegion *subregion)
1505 {
1506     assert(!subregion->container);
1507     subregion->container = mr;
1508     subregion->addr = offset;
1509     memory_region_update_container_subregions(subregion);
1510 }
1511
1512 void memory_region_add_subregion(MemoryRegion *mr,
1513                                  hwaddr offset,
1514                                  MemoryRegion *subregion)
1515 {
1516     subregion->may_overlap = false;
1517     subregion->priority = 0;
1518     memory_region_add_subregion_common(mr, offset, subregion);
1519 }
1520
1521 void memory_region_add_subregion_overlap(MemoryRegion *mr,
1522                                          hwaddr offset,
1523                                          MemoryRegion *subregion,
1524                                          int priority)
1525 {
1526     subregion->may_overlap = true;
1527     subregion->priority = priority;
1528     memory_region_add_subregion_common(mr, offset, subregion);
1529 }
1530
1531 void memory_region_del_subregion(MemoryRegion *mr,
1532                                  MemoryRegion *subregion)
1533 {
1534     memory_region_transaction_begin();
1535     assert(subregion->container == mr);
1536     subregion->container = NULL;
1537     QTAILQ_REMOVE(&mr->subregions, subregion, subregions_link);
1538     memory_region_unref(subregion);
1539     memory_region_update_pending |= mr->enabled && subregion->enabled;
1540     memory_region_transaction_commit();
1541 }
1542
1543 void memory_region_set_enabled(MemoryRegion *mr, bool enabled)
1544 {
1545     if (enabled == mr->enabled) {
1546         return;
1547     }
1548     memory_region_transaction_begin();
1549     mr->enabled = enabled;
1550     memory_region_update_pending = true;
1551     memory_region_transaction_commit();
1552 }
1553
1554 static void memory_region_readd_subregion(MemoryRegion *mr)
1555 {
1556     MemoryRegion *container = mr->container;
1557
1558     if (container) {
1559         memory_region_transaction_begin();
1560         memory_region_ref(mr);
1561         memory_region_del_subregion(container, mr);
1562         mr->container = container;
1563         memory_region_update_container_subregions(mr);
1564         memory_region_unref(mr);
1565         memory_region_transaction_commit();
1566     }
1567 }
1568
1569 void memory_region_set_address(MemoryRegion *mr, hwaddr addr)
1570 {
1571     if (addr != mr->addr) {
1572         mr->addr = addr;
1573         memory_region_readd_subregion(mr);
1574     }
1575 }
1576
1577 void memory_region_set_alias_offset(MemoryRegion *mr, hwaddr offset)
1578 {
1579     assert(mr->alias);
1580
1581     if (offset == mr->alias_offset) {
1582         return;
1583     }
1584
1585     memory_region_transaction_begin();
1586     mr->alias_offset = offset;
1587     memory_region_update_pending |= mr->enabled;
1588     memory_region_transaction_commit();
1589 }
1590
1591 ram_addr_t memory_region_get_ram_addr(MemoryRegion *mr)
1592 {
1593     return mr->ram_addr;
1594 }
1595
1596 static int cmp_flatrange_addr(const void *addr_, const void *fr_)
1597 {
1598     const AddrRange *addr = addr_;
1599     const FlatRange *fr = fr_;
1600
1601     if (int128_le(addrrange_end(*addr), fr->addr.start)) {
1602         return -1;
1603     } else if (int128_ge(addr->start, addrrange_end(fr->addr))) {
1604         return 1;
1605     }
1606     return 0;
1607 }
1608
1609 static FlatRange *flatview_lookup(FlatView *view, AddrRange addr)
1610 {
1611     return bsearch(&addr, view->ranges, view->nr,
1612                    sizeof(FlatRange), cmp_flatrange_addr);
1613 }
1614
1615 bool memory_region_present(MemoryRegion *container, hwaddr addr)
1616 {
1617     MemoryRegion *mr = memory_region_find(container, addr, 1).mr;
1618     if (!mr || (mr == container)) {
1619         return false;
1620     }
1621     memory_region_unref(mr);
1622     return true;
1623 }
1624
1625 bool memory_region_is_mapped(MemoryRegion *mr)
1626 {
1627     return mr->container ? true : false;
1628 }
1629
1630 MemoryRegionSection memory_region_find(MemoryRegion *mr,
1631                                        hwaddr addr, uint64_t size)
1632 {
1633     MemoryRegionSection ret = { .mr = NULL };
1634     MemoryRegion *root;
1635     AddressSpace *as;
1636     AddrRange range;
1637     FlatView *view;
1638     FlatRange *fr;
1639
1640     addr += mr->addr;
1641     for (root = mr; root->container; ) {
1642         root = root->container;
1643         addr += root->addr;
1644     }
1645
1646     as = memory_region_to_address_space(root);
1647     if (!as) {
1648         return ret;
1649     }
1650     range = addrrange_make(int128_make64(addr), int128_make64(size));
1651
1652     view = address_space_get_flatview(as);
1653     fr = flatview_lookup(view, range);
1654     if (!fr) {
1655         flatview_unref(view);
1656         return ret;
1657     }
1658
1659     while (fr > view->ranges && addrrange_intersects(fr[-1].addr, range)) {
1660         --fr;
1661     }
1662
1663     ret.mr = fr->mr;
1664     ret.address_space = as;
1665     range = addrrange_intersection(range, fr->addr);
1666     ret.offset_within_region = fr->offset_in_region;
1667     ret.offset_within_region += int128_get64(int128_sub(range.start,
1668                                                         fr->addr.start));
1669     ret.size = range.size;
1670     ret.offset_within_address_space = int128_get64(range.start);
1671     ret.readonly = fr->readonly;
1672     memory_region_ref(ret.mr);
1673
1674     flatview_unref(view);
1675     return ret;
1676 }
1677
1678 void address_space_sync_dirty_bitmap(AddressSpace *as)
1679 {
1680     FlatView *view;
1681     FlatRange *fr;
1682
1683     view = address_space_get_flatview(as);
1684     FOR_EACH_FLAT_RANGE(fr, view) {
1685         MEMORY_LISTENER_UPDATE_REGION(fr, as, Forward, log_sync);
1686     }
1687     flatview_unref(view);
1688 }
1689
1690 void memory_global_dirty_log_start(void)
1691 {
1692     global_dirty_log = true;
1693     MEMORY_LISTENER_CALL_GLOBAL(log_global_start, Forward);
1694 }
1695
1696 void memory_global_dirty_log_stop(void)
1697 {
1698     global_dirty_log = false;
1699     MEMORY_LISTENER_CALL_GLOBAL(log_global_stop, Reverse);
1700 }
1701
1702 static void listener_add_address_space(MemoryListener *listener,
1703                                        AddressSpace *as)
1704 {
1705     FlatView *view;
1706     FlatRange *fr;
1707
1708     if (listener->address_space_filter
1709         && listener->address_space_filter != as) {
1710         return;
1711     }
1712
1713     if (global_dirty_log) {
1714         if (listener->log_global_start) {
1715             listener->log_global_start(listener);
1716         }
1717     }
1718
1719     view = address_space_get_flatview(as);
1720     FOR_EACH_FLAT_RANGE(fr, view) {
1721         MemoryRegionSection section = {
1722             .mr = fr->mr,
1723             .address_space = as,
1724             .offset_within_region = fr->offset_in_region,
1725             .size = fr->addr.size,
1726             .offset_within_address_space = int128_get64(fr->addr.start),
1727             .readonly = fr->readonly,
1728         };
1729         if (listener->region_add) {
1730             listener->region_add(listener, &section);
1731         }
1732     }
1733     flatview_unref(view);
1734 }
1735
1736 void memory_listener_register(MemoryListener *listener, AddressSpace *filter)
1737 {
1738     MemoryListener *other = NULL;
1739     AddressSpace *as;
1740
1741     listener->address_space_filter = filter;
1742     if (QTAILQ_EMPTY(&memory_listeners)
1743         || listener->priority >= QTAILQ_LAST(&memory_listeners,
1744                                              memory_listeners)->priority) {
1745         QTAILQ_INSERT_TAIL(&memory_listeners, listener, link);
1746     } else {
1747         QTAILQ_FOREACH(other, &memory_listeners, link) {
1748             if (listener->priority < other->priority) {
1749                 break;
1750             }
1751         }
1752         QTAILQ_INSERT_BEFORE(other, listener, link);
1753     }
1754
1755     QTAILQ_FOREACH(as, &address_spaces, address_spaces_link) {
1756         listener_add_address_space(listener, as);
1757     }
1758 }
1759
1760 void memory_listener_unregister(MemoryListener *listener)
1761 {
1762     QTAILQ_REMOVE(&memory_listeners, listener, link);
1763 }
1764
1765 void address_space_init(AddressSpace *as, MemoryRegion *root, const char *name)
1766 {
1767     if (QTAILQ_EMPTY(&address_spaces)) {
1768         memory_init();
1769     }
1770
1771     memory_region_transaction_begin();
1772     as->root = root;
1773     as->current_map = g_new(FlatView, 1);
1774     flatview_init(as->current_map);
1775     as->ioeventfd_nb = 0;
1776     as->ioeventfds = NULL;
1777     QTAILQ_INSERT_TAIL(&address_spaces, as, address_spaces_link);
1778     as->name = g_strdup(name ? name : "anonymous");
1779     address_space_init_dispatch(as);
1780     memory_region_update_pending |= root->enabled;
1781     memory_region_transaction_commit();
1782 }
1783
1784 void address_space_destroy(AddressSpace *as)
1785 {
1786     MemoryListener *listener;
1787
1788     /* Flush out anything from MemoryListeners listening in on this */
1789     memory_region_transaction_begin();
1790     as->root = NULL;
1791     memory_region_transaction_commit();
1792     QTAILQ_REMOVE(&address_spaces, as, address_spaces_link);
1793     address_space_destroy_dispatch(as);
1794
1795     QTAILQ_FOREACH(listener, &memory_listeners, link) {
1796         assert(listener->address_space_filter != as);
1797     }
1798
1799     flatview_unref(as->current_map);
1800     g_free(as->name);
1801     g_free(as->ioeventfds);
1802 }
1803
1804 bool io_mem_read(MemoryRegion *mr, hwaddr addr, uint64_t *pval, unsigned size)
1805 {
1806     return memory_region_dispatch_read(mr, addr, pval, size);
1807 }
1808
1809 bool io_mem_write(MemoryRegion *mr, hwaddr addr,
1810                   uint64_t val, unsigned size)
1811 {
1812     return memory_region_dispatch_write(mr, addr, val, size);
1813 }
1814
1815 typedef struct MemoryRegionList MemoryRegionList;
1816
1817 struct MemoryRegionList {
1818     const MemoryRegion *mr;
1819     bool printed;
1820     QTAILQ_ENTRY(MemoryRegionList) queue;
1821 };
1822
1823 typedef QTAILQ_HEAD(queue, MemoryRegionList) MemoryRegionListHead;
1824
1825 static void mtree_print_mr(fprintf_function mon_printf, void *f,
1826                            const MemoryRegion *mr, unsigned int level,
1827                            hwaddr base,
1828                            MemoryRegionListHead *alias_print_queue)
1829 {
1830     MemoryRegionList *new_ml, *ml, *next_ml;
1831     MemoryRegionListHead submr_print_queue;
1832     const MemoryRegion *submr;
1833     unsigned int i;
1834
1835     if (!mr || !mr->enabled) {
1836         return;
1837     }
1838
1839     for (i = 0; i < level; i++) {
1840         mon_printf(f, "  ");
1841     }
1842
1843     if (mr->alias) {
1844         MemoryRegionList *ml;
1845         bool found = false;
1846
1847         /* check if the alias is already in the queue */
1848         QTAILQ_FOREACH(ml, alias_print_queue, queue) {
1849             if (ml->mr == mr->alias && !ml->printed) {
1850                 found = true;
1851             }
1852         }
1853
1854         if (!found) {
1855             ml = g_new(MemoryRegionList, 1);
1856             ml->mr = mr->alias;
1857             ml->printed = false;
1858             QTAILQ_INSERT_TAIL(alias_print_queue, ml, queue);
1859         }
1860         mon_printf(f, TARGET_FMT_plx "-" TARGET_FMT_plx
1861                    " (prio %d, %c%c): alias %s @%s " TARGET_FMT_plx
1862                    "-" TARGET_FMT_plx "\n",
1863                    base + mr->addr,
1864                    base + mr->addr
1865                    + (int128_nz(mr->size) ?
1866                       (hwaddr)int128_get64(int128_sub(mr->size,
1867                                                       int128_one())) : 0),
1868                    mr->priority,
1869                    mr->romd_mode ? 'R' : '-',
1870                    !mr->readonly && !(mr->rom_device && mr->romd_mode) ? 'W'
1871                                                                        : '-',
1872                    mr->name,
1873                    mr->alias->name,
1874                    mr->alias_offset,
1875                    mr->alias_offset
1876                    + (int128_nz(mr->size) ?
1877                       (hwaddr)int128_get64(int128_sub(mr->size,
1878                                                       int128_one())) : 0));
1879     } else {
1880         mon_printf(f,
1881                    TARGET_FMT_plx "-" TARGET_FMT_plx " (prio %d, %c%c): %s\n",
1882                    base + mr->addr,
1883                    base + mr->addr
1884                    + (int128_nz(mr->size) ?
1885                       (hwaddr)int128_get64(int128_sub(mr->size,
1886                                                       int128_one())) : 0),
1887                    mr->priority,
1888                    mr->romd_mode ? 'R' : '-',
1889                    !mr->readonly && !(mr->rom_device && mr->romd_mode) ? 'W'
1890                                                                        : '-',
1891                    mr->name);
1892     }
1893
1894     QTAILQ_INIT(&submr_print_queue);
1895
1896     QTAILQ_FOREACH(submr, &mr->subregions, subregions_link) {
1897         new_ml = g_new(MemoryRegionList, 1);
1898         new_ml->mr = submr;
1899         QTAILQ_FOREACH(ml, &submr_print_queue, queue) {
1900             if (new_ml->mr->addr < ml->mr->addr ||
1901                 (new_ml->mr->addr == ml->mr->addr &&
1902                  new_ml->mr->priority > ml->mr->priority)) {
1903                 QTAILQ_INSERT_BEFORE(ml, new_ml, queue);
1904                 new_ml = NULL;
1905                 break;
1906             }
1907         }
1908         if (new_ml) {
1909             QTAILQ_INSERT_TAIL(&submr_print_queue, new_ml, queue);
1910         }
1911     }
1912
1913     QTAILQ_FOREACH(ml, &submr_print_queue, queue) {
1914         mtree_print_mr(mon_printf, f, ml->mr, level + 1, base + mr->addr,
1915                        alias_print_queue);
1916     }
1917
1918     QTAILQ_FOREACH_SAFE(ml, &submr_print_queue, queue, next_ml) {
1919         g_free(ml);
1920     }
1921 }
1922
1923 void mtree_info(fprintf_function mon_printf, void *f)
1924 {
1925     MemoryRegionListHead ml_head;
1926     MemoryRegionList *ml, *ml2;
1927     AddressSpace *as;
1928
1929     QTAILQ_INIT(&ml_head);
1930
1931     QTAILQ_FOREACH(as, &address_spaces, address_spaces_link) {
1932         mon_printf(f, "%s\n", as->name);
1933         mtree_print_mr(mon_printf, f, as->root, 0, 0, &ml_head);
1934     }
1935
1936     mon_printf(f, "aliases\n");
1937     /* print aliased regions */
1938     QTAILQ_FOREACH(ml, &ml_head, queue) {
1939         if (!ml->printed) {
1940             mon_printf(f, "%s\n", ml->mr->name);
1941             mtree_print_mr(mon_printf, f, ml->mr, 0, 0, &ml_head);
1942         }
1943     }
1944
1945     QTAILQ_FOREACH_SAFE(ml, &ml_head, queue, ml2) {
1946         g_free(ml);
1947     }
1948 }