memory.c

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