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 = int128_get64((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 ajacent 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     /* Render the region itself into any gaps left by the current view. */
 560     for (i = 0; i < view->nr && int128_nz(remain); ++i) {
 561         if (int128_ge(base, addrrange_end(view->ranges[i].addr))) {
 562             continue;
 563         }
 564         if (int128_lt(base, view->ranges[i].addr.start)) {
 565             now = int128_min(remain,
 566                              int128_sub(view->ranges[i].addr.start, base));
 567             fr.mr = mr;
 568             fr.offset_in_region = offset_in_region;
 569             fr.addr = addrrange_make(base, now);
 570             fr.dirty_log_mask = mr->dirty_log_mask;
 571             fr.romd_mode = mr->romd_mode;
 572             fr.readonly = readonly;
 573             flatview_insert(view, i, &fr);
 574             ++i;
 575             int128_addto(&base, now);
 576             offset_in_region += int128_get64(now);
 577             int128_subfrom(&remain, now);
 578         }
 579         now = int128_sub(int128_min(int128_add(base, remain),
 580                                     addrrange_end(view->ranges[i].addr)),
 581                          base);
 582         int128_addto(&base, now);
 583         offset_in_region += int128_get64(now);
 584         int128_subfrom(&remain, now);
 585     }
 586     if (int128_nz(remain)) {
 587         fr.mr = mr;
 588         fr.offset_in_region = offset_in_region;
 589         fr.addr = addrrange_make(base, remain);
 590         fr.dirty_log_mask = mr->dirty_log_mask;
 591         fr.romd_mode = mr->romd_mode;
 592         fr.readonly = readonly;
 593         flatview_insert(view, i, &fr);
 594     }
 595 }
 596
 597 /* Render a memory topology into a list of disjoint absolute ranges. */
 598 static FlatView generate_memory_topology(MemoryRegion *mr)
 599 {
 600     FlatView view;
 601
 602     flatview_init(&view);
 603
 604     if (mr) {
 605         render_memory_region(&view, mr, int128_zero(),
 606                              addrrange_make(int128_zero(), int128_2_64()), false);
 607     }
 608     flatview_simplify(&view);
 609
 610     return view;
 611 }
 612
 613 static void address_space_add_del_ioeventfds(AddressSpace *as,
 614                                              MemoryRegionIoeventfd *fds_new,
 615                                              unsigned fds_new_nb,
 616                                              MemoryRegionIoeventfd *fds_old,
 617                                              unsigned fds_old_nb)
 618 {
 619     unsigned iold, inew;
 620     MemoryRegionIoeventfd *fd;
 621     MemoryRegionSection section;
 622
 623     /* Generate a symmetric difference of the old and new fd sets, adding
 624      * and deleting as necessary.
 625      */
 626
 627     iold = inew = 0;
 628     while (iold < fds_old_nb || inew < fds_new_nb) {
 629         if (iold < fds_old_nb
 630             && (inew == fds_new_nb
 631                 || memory_region_ioeventfd_before(fds_old[iold],
 632                                                   fds_new[inew]))) {
 633             fd = &fds_old[iold];
 634             section = (MemoryRegionSection) {
 635                 .address_space = as,
 636                 .offset_within_address_space = int128_get64(fd->addr.start),
 637                 .size = int128_get64(fd->addr.size),
 638             };
 639             MEMORY_LISTENER_CALL(eventfd_del, Forward, &section,
 640                                  fd->match_data, fd->data, fd->e);
 641             ++iold;
 642         } else if (inew < fds_new_nb
 643                    && (iold == fds_old_nb
 644                        || memory_region_ioeventfd_before(fds_new[inew],
 645                                                          fds_old[iold]))) {
 646             fd = &fds_new[inew];
 647             section = (MemoryRegionSection) {
 648                 .address_space = as,
 649                 .offset_within_address_space = int128_get64(fd->addr.start),
 650                 .size = int128_get64(fd->addr.size),
 651             };
 652             MEMORY_LISTENER_CALL(eventfd_add, Reverse, &section,
 653                                  fd->match_data, fd->data, fd->e);
 654             ++inew;
 655         } else {
 656             ++iold;
 657             ++inew;
 658         }
 659     }
 660 }
 661
 662 static void address_space_update_ioeventfds(AddressSpace *as)
 663 {
 664     FlatRange *fr;
 665     unsigned ioeventfd_nb = 0;
 666     MemoryRegionIoeventfd *ioeventfds = NULL;
 667     AddrRange tmp;
 668     unsigned i;
 669
 670     FOR_EACH_FLAT_RANGE(fr, as->current_map) {
 671         for (i = 0; i < fr->mr->ioeventfd_nb; ++i) {
 672             tmp = addrrange_shift(fr->mr->ioeventfds[i].addr,
 673                                   int128_sub(fr->addr.start,
 674                                              int128_make64(fr->offset_in_region)));
 675             if (addrrange_intersects(fr->addr, tmp)) {
 676                 ++ioeventfd_nb;
 677                 ioeventfds = g_realloc(ioeventfds,
 678                                           ioeventfd_nb * sizeof(*ioeventfds));
 679                 ioeventfds[ioeventfd_nb-1] = fr->mr->ioeventfds[i];
 680                 ioeventfds[ioeventfd_nb-1].addr = tmp;
 681             }
 682         }
 683     }
 684
 685     address_space_add_del_ioeventfds(as, ioeventfds, ioeventfd_nb,
 686                                      as->ioeventfds, as->ioeventfd_nb);
 687
 688     g_free(as->ioeventfds);
 689     as->ioeventfds = ioeventfds;
 690     as->ioeventfd_nb = ioeventfd_nb;
 691 }
 692
 693 static void address_space_update_topology_pass(AddressSpace *as,
 694                                                FlatView old_view,
 695                                                FlatView new_view,
 696                                                bool adding)
 697 {
 698     unsigned iold, inew;
 699     FlatRange *frold, *frnew;
 700
 701     /* Generate a symmetric difference of the old and new memory maps.
 702      * Kill ranges in the old map, and instantiate ranges in the new map.
 703      */
 704     iold = inew = 0;
 705     while (iold < old_view.nr || inew < new_view.nr) {
 706         if (iold < old_view.nr) {
 707             frold = &old_view.ranges[iold];
 708         } else {
 709             frold = NULL;
 710         }
 711         if (inew < new_view.nr) {
 712             frnew = &new_view.ranges[inew];
 713         } else {
 714             frnew = NULL;
 715         }
 716
 717         if (frold
 718             && (!frnew
 719                 || int128_lt(frold->addr.start, frnew->addr.start)
 720                 || (int128_eq(frold->addr.start, frnew->addr.start)
 721                     && !flatrange_equal(frold, frnew)))) {
 722             /* In old, but (not in new, or in new but attributes changed). */
 723
 724             if (!adding) {
 725                 MEMORY_LISTENER_UPDATE_REGION(frold, as, Reverse, region_del);
 726             }
 727
 728             ++iold;
 729         } else if (frold && frnew && flatrange_equal(frold, frnew)) {
 730             /* In both (logging may have changed) */
 731
 732             if (adding) {
 733                 MEMORY_LISTENER_UPDATE_REGION(frnew, as, Forward, region_nop);
 734                 if (frold->dirty_log_mask && !frnew->dirty_log_mask) {
 735                     MEMORY_LISTENER_UPDATE_REGION(frnew, as, Reverse, log_stop);
 736                 } else if (frnew->dirty_log_mask && !frold->dirty_log_mask) {
 737                     MEMORY_LISTENER_UPDATE_REGION(frnew, as, Forward, log_start);
 738                 }
 739             }
 740
 741             ++iold;
 742             ++inew;
 743         } else {
 744             /* In new */
 745
 746             if (adding) {
 747                 MEMORY_LISTENER_UPDATE_REGION(frnew, as, Forward, region_add);
 748             }
 749
 750             ++inew;
 751         }
 752     }
 753 }
 754
 755
 756 static void address_space_update_topology(AddressSpace *as)
 757 {
 758     FlatView old_view = *as->current_map;
 759     FlatView new_view = generate_memory_topology(as->root);
 760
 761     address_space_update_topology_pass(as, old_view, new_view, false);
 762     address_space_update_topology_pass(as, old_view, new_view, true);
 763
 764     *as->current_map = new_view;
 765     flatview_destroy(&old_view);
 766     address_space_update_ioeventfds(as);
 767 }
 768
 769 void memory_region_transaction_begin(void)
 770 {
 771     qemu_flush_coalesced_mmio_buffer();
 772     ++memory_region_transaction_depth;
 773 }
 774
 775 void memory_region_transaction_commit(void)
 776 {
 777     AddressSpace *as;
 778
 779     assert(memory_region_transaction_depth);
 780     --memory_region_transaction_depth;
 781     if (!memory_region_transaction_depth && memory_region_update_pending) {
 782         memory_region_update_pending = false;
 783         MEMORY_LISTENER_CALL_GLOBAL(begin, Forward);
 784
 785         QTAILQ_FOREACH(as, &address_spaces, address_spaces_link) {
 786             address_space_update_topology(as);
 787         }
 788
 789         MEMORY_LISTENER_CALL_GLOBAL(commit, Forward);
 790     }
 791 }
 792
 793 static void memory_region_destructor_none(MemoryRegion *mr)
 794 {
 795 }
 796
 797 static void memory_region_destructor_ram(MemoryRegion *mr)
 798 {
 799     qemu_ram_free(mr->ram_addr);
 800 }
 801
 802 static void memory_region_destructor_ram_from_ptr(MemoryRegion *mr)
 803 {
 804     qemu_ram_free_from_ptr(mr->ram_addr);
 805 }
 806
 807 static void memory_region_destructor_rom_device(MemoryRegion *mr)
 808 {
 809     qemu_ram_free(mr->ram_addr & TARGET_PAGE_MASK);
 810 }
 811
 812 static bool memory_region_wrong_endianness(MemoryRegion *mr)
 813 {
 814 #ifdef TARGET_WORDS_BIGENDIAN
 815     return mr->ops->endianness == DEVICE_LITTLE_ENDIAN;
 816 #else
 817     return mr->ops->endianness == DEVICE_BIG_ENDIAN;
 818 #endif
 819 }
 820
 821 void memory_region_init(MemoryRegion *mr,
 822                         const char *name,
 823                         uint64_t size)
 824 {
 825     mr->ops = &unassigned_mem_ops;
 826     mr->opaque = NULL;
 827     mr->parent = NULL;
 828     mr->size = int128_make64(size);
 829     if (size == UINT64_MAX) {
 830         mr->size = int128_2_64();
 831     }
 832     mr->addr = 0;
 833     mr->subpage = false;
 834     mr->enabled = true;
 835     mr->terminates = false;
 836     mr->ram = false;
 837     mr->romd_mode = true;
 838     mr->readonly = false;
 839     mr->rom_device = false;
 840     mr->destructor = memory_region_destructor_none;
 841     mr->priority = 0;
 842     mr->may_overlap = false;
 843     mr->alias = NULL;
 844     QTAILQ_INIT(&mr->subregions);
 845     memset(&mr->subregions_link, 0, sizeof mr->subregions_link);
 846     QTAILQ_INIT(&mr->coalesced);
 847     mr->name = g_strdup(name);
 848     mr->dirty_log_mask = 0;
 849     mr->ioeventfd_nb = 0;
 850     mr->ioeventfds = NULL;
 851     mr->flush_coalesced_mmio = false;
 852 }
 853
 854 static uint64_t unassigned_mem_read(void *opaque, hwaddr addr,
 855                                     unsigned size)
 856 {
 857 #ifdef DEBUG_UNASSIGNED
 858     printf("Unassigned mem read " TARGET_FMT_plx "\n", addr);
 859 #endif
 860 #if defined(TARGET_ALPHA) || defined(TARGET_SPARC) || defined(TARGET_MICROBLAZE)
 861     cpu_unassigned_access(cpu_single_env, addr, 0, 0, 0, size);
 862 #endif
 863     return 0;
 864 }
 865
 866 static void unassigned_mem_write(void *opaque, hwaddr addr,
 867                                  uint64_t val, unsigned size)
 868 {
 869 #ifdef DEBUG_UNASSIGNED
 870     printf("Unassigned mem write " TARGET_FMT_plx " = 0x%"PRIx64"\n", addr, val);
 871 #endif
 872 #if defined(TARGET_ALPHA) || defined(TARGET_SPARC) || defined(TARGET_MICROBLAZE)
 873     cpu_unassigned_access(cpu_single_env, addr, 1, 0, 0, size);
 874 #endif
 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_reservation(MemoryRegion *mr,
1067                                     const char *name,
1068                                     uint64_t size)
1069 {
1070     memory_region_init_io(mr, &unassigned_mem_ops, mr, name, size);
1071 }
1072
1073 void memory_region_destroy(MemoryRegion *mr)
1074 {
1075     assert(QTAILQ_EMPTY(&mr->subregions));
1076     assert(memory_region_transaction_depth == 0);
1077     mr->destructor(mr);
1078     memory_region_clear_coalescing(mr);
1079     g_free((char *)mr->name);
1080     g_free(mr->ioeventfds);
1081 }
1082
1083 uint64_t memory_region_size(MemoryRegion *mr)
1084 {
1085     if (int128_eq(mr->size, int128_2_64())) {
1086         return UINT64_MAX;
1087     }
1088     return int128_get64(mr->size);
1089 }
1090
1091 const char *memory_region_name(MemoryRegion *mr)
1092 {
1093     return mr->name;
1094 }
1095
1096 bool memory_region_is_ram(MemoryRegion *mr)
1097 {
1098     return mr->ram;
1099 }
1100
1101 bool memory_region_is_logging(MemoryRegion *mr)
1102 {
1103     return mr->dirty_log_mask;
1104 }
1105
1106 bool memory_region_is_rom(MemoryRegion *mr)
1107 {
1108     return mr->ram && mr->readonly;
1109 }
1110
1111 void memory_region_set_log(MemoryRegion *mr, bool log, unsigned client)
1112 {
1113     uint8_t mask = 1 << client;
1114
1115     memory_region_transaction_begin();
1116     mr->dirty_log_mask = (mr->dirty_log_mask & ~mask) | (log * mask);
1117     memory_region_update_pending |= mr->enabled;
1118     memory_region_transaction_commit();
1119 }
1120
1121 bool memory_region_get_dirty(MemoryRegion *mr, hwaddr addr,
1122                              hwaddr size, unsigned client)
1123 {
1124     assert(mr->terminates);
1125     return cpu_physical_memory_get_dirty(mr->ram_addr + addr, size,
1126                                          1 << client);
1127 }
1128
1129 void memory_region_set_dirty(MemoryRegion *mr, hwaddr addr,
1130                              hwaddr size)
1131 {
1132     assert(mr->terminates);
1133     return cpu_physical_memory_set_dirty_range(mr->ram_addr + addr, size, -1);
1134 }
1135
1136 bool memory_region_test_and_clear_dirty(MemoryRegion *mr, hwaddr addr,
1137                                         hwaddr size, unsigned client)
1138 {
1139     bool ret;
1140     assert(mr->terminates);
1141     ret = cpu_physical_memory_get_dirty(mr->ram_addr + addr, size,
1142                                         1 << client);
1143     if (ret) {
1144         cpu_physical_memory_reset_dirty(mr->ram_addr + addr,
1145                                         mr->ram_addr + addr + size,
1146                                         1 << client);
1147     }
1148     return ret;
1149 }
1150
1151
1152 void memory_region_sync_dirty_bitmap(MemoryRegion *mr)
1153 {
1154     AddressSpace *as;
1155     FlatRange *fr;
1156
1157     QTAILQ_FOREACH(as, &address_spaces, address_spaces_link) {
1158         FOR_EACH_FLAT_RANGE(fr, as->current_map) {
1159             if (fr->mr == mr) {
1160                 MEMORY_LISTENER_UPDATE_REGION(fr, as, Forward, log_sync);
1161             }
1162         }
1163     }
1164 }
1165
1166 void memory_region_set_readonly(MemoryRegion *mr, bool readonly)
1167 {
1168     if (mr->readonly != readonly) {
1169         memory_region_transaction_begin();
1170         mr->readonly = readonly;
1171         memory_region_update_pending |= mr->enabled;
1172         memory_region_transaction_commit();
1173     }
1174 }
1175
1176 void memory_region_rom_device_set_romd(MemoryRegion *mr, bool romd_mode)
1177 {
1178     if (mr->romd_mode != romd_mode) {
1179         memory_region_transaction_begin();
1180         mr->romd_mode = romd_mode;
1181         memory_region_update_pending |= mr->enabled;
1182         memory_region_transaction_commit();
1183     }
1184 }
1185
1186 void memory_region_reset_dirty(MemoryRegion *mr, hwaddr addr,
1187                                hwaddr size, unsigned client)
1188 {
1189     assert(mr->terminates);
1190     cpu_physical_memory_reset_dirty(mr->ram_addr + addr,
1191                                     mr->ram_addr + addr + size,
1192                                     1 << client);
1193 }
1194
1195 void *memory_region_get_ram_ptr(MemoryRegion *mr)
1196 {
1197     if (mr->alias) {
1198         return memory_region_get_ram_ptr(mr->alias) + mr->alias_offset;
1199     }
1200
1201     assert(mr->terminates);
1202
1203     return qemu_get_ram_ptr(mr->ram_addr & TARGET_PAGE_MASK);
1204 }
1205
1206 static void memory_region_update_coalesced_range_as(MemoryRegion *mr, AddressSpace *as)
1207 {
1208     FlatRange *fr;
1209     CoalescedMemoryRange *cmr;
1210     AddrRange tmp;
1211     MemoryRegionSection section;
1212
1213     FOR_EACH_FLAT_RANGE(fr, as->current_map) {
1214         if (fr->mr == mr) {
1215             section = (MemoryRegionSection) {
1216                 .address_space = as,
1217                 .offset_within_address_space = int128_get64(fr->addr.start),
1218                 .size = int128_get64(fr->addr.size),
1219             };
1220
1221             MEMORY_LISTENER_CALL(coalesced_mmio_del, Reverse, &section,
1222                                  int128_get64(fr->addr.start),
1223                                  int128_get64(fr->addr.size));
1224             QTAILQ_FOREACH(cmr, &mr->coalesced, link) {
1225                 tmp = addrrange_shift(cmr->addr,
1226                                       int128_sub(fr->addr.start,
1227                                                  int128_make64(fr->offset_in_region)));
1228                 if (!addrrange_intersects(tmp, fr->addr)) {
1229                     continue;
1230                 }
1231                 tmp = addrrange_intersection(tmp, fr->addr);
1232                 MEMORY_LISTENER_CALL(coalesced_mmio_add, Forward, &section,
1233                                      int128_get64(tmp.start),
1234                                      int128_get64(tmp.size));
1235             }
1236         }
1237     }
1238 }
1239
1240 static void memory_region_update_coalesced_range(MemoryRegion *mr)
1241 {
1242     AddressSpace *as;
1243
1244     QTAILQ_FOREACH(as, &address_spaces, address_spaces_link) {
1245         memory_region_update_coalesced_range_as(mr, as);
1246     }
1247 }
1248
1249 void memory_region_set_coalescing(MemoryRegion *mr)
1250 {
1251     memory_region_clear_coalescing(mr);
1252     memory_region_add_coalescing(mr, 0, int128_get64(mr->size));
1253 }
1254
1255 void memory_region_add_coalescing(MemoryRegion *mr,
1256                                   hwaddr offset,
1257                                   uint64_t size)
1258 {
1259     CoalescedMemoryRange *cmr = g_malloc(sizeof(*cmr));
1260
1261     cmr->addr = addrrange_make(int128_make64(offset), int128_make64(size));
1262     QTAILQ_INSERT_TAIL(&mr->coalesced, cmr, link);
1263     memory_region_update_coalesced_range(mr);
1264     memory_region_set_flush_coalesced(mr);
1265 }
1266
1267 void memory_region_clear_coalescing(MemoryRegion *mr)
1268 {
1269     CoalescedMemoryRange *cmr;
1270
1271     qemu_flush_coalesced_mmio_buffer();
1272     mr->flush_coalesced_mmio = false;
1273
1274     while (!QTAILQ_EMPTY(&mr->coalesced)) {
1275         cmr = QTAILQ_FIRST(&mr->coalesced);
1276         QTAILQ_REMOVE(&mr->coalesced, cmr, link);
1277         g_free(cmr);
1278     }
1279     memory_region_update_coalesced_range(mr);
1280 }
1281
1282 void memory_region_set_flush_coalesced(MemoryRegion *mr)
1283 {
1284     mr->flush_coalesced_mmio = true;
1285 }
1286
1287 void memory_region_clear_flush_coalesced(MemoryRegion *mr)
1288 {
1289     qemu_flush_coalesced_mmio_buffer();
1290     if (QTAILQ_EMPTY(&mr->coalesced)) {
1291         mr->flush_coalesced_mmio = false;
1292     }
1293 }
1294
1295 void memory_region_add_eventfd(MemoryRegion *mr,
1296                                hwaddr addr,
1297                                unsigned size,
1298                                bool match_data,
1299                                uint64_t data,
1300                                EventNotifier *e)
1301 {
1302     MemoryRegionIoeventfd mrfd = {
1303         .addr.start = int128_make64(addr),
1304         .addr.size = int128_make64(size),
1305         .match_data = match_data,
1306         .data = data,
1307         .e = e,
1308     };
1309     unsigned i;
1310
1311     adjust_endianness(mr, &mrfd.data, size);
1312     memory_region_transaction_begin();
1313     for (i = 0; i < mr->ioeventfd_nb; ++i) {
1314         if (memory_region_ioeventfd_before(mrfd, mr->ioeventfds[i])) {
1315             break;
1316         }
1317     }
1318     ++mr->ioeventfd_nb;
1319     mr->ioeventfds = g_realloc(mr->ioeventfds,
1320                                   sizeof(*mr->ioeventfds) * mr->ioeventfd_nb);
1321     memmove(&mr->ioeventfds[i+1], &mr->ioeventfds[i],
1322             sizeof(*mr->ioeventfds) * (mr->ioeventfd_nb-1 - i));
1323     mr->ioeventfds[i] = mrfd;
1324     memory_region_update_pending |= mr->enabled;
1325     memory_region_transaction_commit();
1326 }
1327
1328 void memory_region_del_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_equal(mrfd, mr->ioeventfds[i])) {
1348             break;
1349         }
1350     }
1351     assert(i != mr->ioeventfd_nb);
1352     memmove(&mr->ioeventfds[i], &mr->ioeventfds[i+1],
1353             sizeof(*mr->ioeventfds) * (mr->ioeventfd_nb - (i+1)));
1354     --mr->ioeventfd_nb;
1355     mr->ioeventfds = g_realloc(mr->ioeventfds,
1356                                   sizeof(*mr->ioeventfds)*mr->ioeventfd_nb + 1);
1357     memory_region_update_pending |= mr->enabled;
1358     memory_region_transaction_commit();
1359 }
1360
1361 static void memory_region_add_subregion_common(MemoryRegion *mr,
1362                                                hwaddr offset,
1363                                                MemoryRegion *subregion)
1364 {
1365     MemoryRegion *other;
1366
1367     memory_region_transaction_begin();
1368
1369     assert(!subregion->parent);
1370     subregion->parent = mr;
1371     subregion->addr = offset;
1372     QTAILQ_FOREACH(other, &mr->subregions, subregions_link) {
1373         if (subregion->may_overlap || other->may_overlap) {
1374             continue;
1375         }
1376         if (int128_ge(int128_make64(offset),
1377                       int128_add(int128_make64(other->addr), other->size))
1378             || int128_le(int128_add(int128_make64(offset), subregion->size),
1379                          int128_make64(other->addr))) {
1380             continue;
1381         }
1382 #if 0
1383         printf("warning: subregion collision %llx/%llx (%s) "
1384                "vs %llx/%llx (%s)\n",
1385                (unsigned long long)offset,
1386                (unsigned long long)int128_get64(subregion->size),
1387                subregion->name,
1388                (unsigned long long)other->addr,
1389                (unsigned long long)int128_get64(other->size),
1390                other->name);
1391 #endif
1392     }
1393     QTAILQ_FOREACH(other, &mr->subregions, subregions_link) {
1394         if (subregion->priority >= other->priority) {
1395             QTAILQ_INSERT_BEFORE(other, subregion, subregions_link);
1396             goto done;
1397         }
1398     }
1399     QTAILQ_INSERT_TAIL(&mr->subregions, subregion, subregions_link);
1400 done:
1401     memory_region_update_pending |= mr->enabled && subregion->enabled;
1402     memory_region_transaction_commit();
1403 }
1404
1405
1406 void memory_region_add_subregion(MemoryRegion *mr,
1407                                  hwaddr offset,
1408                                  MemoryRegion *subregion)
1409 {
1410     subregion->may_overlap = false;
1411     subregion->priority = 0;
1412     memory_region_add_subregion_common(mr, offset, subregion);
1413 }
1414
1415 void memory_region_add_subregion_overlap(MemoryRegion *mr,
1416                                          hwaddr offset,
1417                                          MemoryRegion *subregion,
1418                                          unsigned priority)
1419 {
1420     subregion->may_overlap = true;
1421     subregion->priority = priority;
1422     memory_region_add_subregion_common(mr, offset, subregion);
1423 }
1424
1425 void memory_region_del_subregion(MemoryRegion *mr,
1426                                  MemoryRegion *subregion)
1427 {
1428     memory_region_transaction_begin();
1429     assert(subregion->parent == mr);
1430     subregion->parent = NULL;
1431     QTAILQ_REMOVE(&mr->subregions, subregion, subregions_link);
1432     memory_region_update_pending |= mr->enabled && subregion->enabled;
1433     memory_region_transaction_commit();
1434 }
1435
1436 void memory_region_set_enabled(MemoryRegion *mr, bool enabled)
1437 {
1438     if (enabled == mr->enabled) {
1439         return;
1440     }
1441     memory_region_transaction_begin();
1442     mr->enabled = enabled;
1443     memory_region_update_pending = true;
1444     memory_region_transaction_commit();
1445 }
1446
1447 void memory_region_set_address(MemoryRegion *mr, hwaddr addr)
1448 {
1449     MemoryRegion *parent = mr->parent;
1450     unsigned priority = mr->priority;
1451     bool may_overlap = mr->may_overlap;
1452
1453     if (addr == mr->addr || !parent) {
1454         mr->addr = addr;
1455         return;
1456     }
1457
1458     memory_region_transaction_begin();
1459     memory_region_del_subregion(parent, mr);
1460     if (may_overlap) {
1461         memory_region_add_subregion_overlap(parent, addr, mr, priority);
1462     } else {
1463         memory_region_add_subregion(parent, addr, mr);
1464     }
1465     memory_region_transaction_commit();
1466 }
1467
1468 void memory_region_set_alias_offset(MemoryRegion *mr, hwaddr offset)
1469 {
1470     assert(mr->alias);
1471
1472     if (offset == mr->alias_offset) {
1473         return;
1474     }
1475
1476     memory_region_transaction_begin();
1477     mr->alias_offset = offset;
1478     memory_region_update_pending |= mr->enabled;
1479     memory_region_transaction_commit();
1480 }
1481
1482 ram_addr_t memory_region_get_ram_addr(MemoryRegion *mr)
1483 {
1484     return mr->ram_addr;
1485 }
1486
1487 static int cmp_flatrange_addr(const void *addr_, const void *fr_)
1488 {
1489     const AddrRange *addr = addr_;
1490     const FlatRange *fr = fr_;
1491
1492     if (int128_le(addrrange_end(*addr), fr->addr.start)) {
1493         return -1;
1494     } else if (int128_ge(addr->start, addrrange_end(fr->addr))) {
1495         return 1;
1496     }
1497     return 0;
1498 }
1499
1500 static FlatRange *address_space_lookup(AddressSpace *as, AddrRange addr)
1501 {
1502     return bsearch(&addr, as->current_map->ranges, as->current_map->nr,
1503                    sizeof(FlatRange), cmp_flatrange_addr);
1504 }
1505
1506 MemoryRegionSection memory_region_find(MemoryRegion *mr,
1507                                        hwaddr addr, uint64_t size)
1508 {
1509     MemoryRegionSection ret = { .mr = NULL, .size = 0 };
1510     MemoryRegion *root;
1511     AddressSpace *as;
1512     AddrRange range;
1513     FlatRange *fr;
1514
1515     addr += mr->addr;
1516     for (root = mr; root->parent; ) {
1517         root = root->parent;
1518         addr += root->addr;
1519     }
1520
1521     as = memory_region_to_address_space(root);
1522     range = addrrange_make(int128_make64(addr), int128_make64(size));
1523     fr = address_space_lookup(as, range);
1524     if (!fr) {
1525         return ret;
1526     }
1527
1528     while (fr > as->current_map->ranges
1529            && addrrange_intersects(fr[-1].addr, range)) {
1530         --fr;
1531     }
1532
1533     ret.mr = fr->mr;
1534     ret.address_space = as;
1535     range = addrrange_intersection(range, fr->addr);
1536     ret.offset_within_region = fr->offset_in_region;
1537     ret.offset_within_region += int128_get64(int128_sub(range.start,
1538                                                         fr->addr.start));
1539     ret.size = int128_get64(range.size);
1540     ret.offset_within_address_space = int128_get64(range.start);
1541     ret.readonly = fr->readonly;
1542     return ret;
1543 }
1544
1545 void address_space_sync_dirty_bitmap(AddressSpace *as)
1546 {
1547     FlatRange *fr;
1548
1549     FOR_EACH_FLAT_RANGE(fr, as->current_map) {
1550         MEMORY_LISTENER_UPDATE_REGION(fr, as, Forward, log_sync);
1551     }
1552 }
1553
1554 void memory_global_dirty_log_start(void)
1555 {
1556     global_dirty_log = true;
1557     MEMORY_LISTENER_CALL_GLOBAL(log_global_start, Forward);
1558 }
1559
1560 void memory_global_dirty_log_stop(void)
1561 {
1562     global_dirty_log = false;
1563     MEMORY_LISTENER_CALL_GLOBAL(log_global_stop, Reverse);
1564 }
1565
1566 static void listener_add_address_space(MemoryListener *listener,
1567                                        AddressSpace *as)
1568 {
1569     FlatRange *fr;
1570
1571     if (listener->address_space_filter
1572         && listener->address_space_filter != as) {
1573         return;
1574     }
1575
1576     if (global_dirty_log) {
1577         if (listener->log_global_start) {
1578             listener->log_global_start(listener);
1579         }
1580     }
1581
1582     FOR_EACH_FLAT_RANGE(fr, as->current_map) {
1583         MemoryRegionSection section = {
1584             .mr = fr->mr,
1585             .address_space = as,
1586             .offset_within_region = fr->offset_in_region,
1587             .size = int128_get64(fr->addr.size),
1588             .offset_within_address_space = int128_get64(fr->addr.start),
1589             .readonly = fr->readonly,
1590         };
1591         if (listener->region_add) {
1592             listener->region_add(listener, &section);
1593         }
1594     }
1595 }
1596
1597 void memory_listener_register(MemoryListener *listener, AddressSpace *filter)
1598 {
1599     MemoryListener *other = NULL;
1600     AddressSpace *as;
1601
1602     listener->address_space_filter = filter;
1603     if (QTAILQ_EMPTY(&memory_listeners)
1604         || listener->priority >= QTAILQ_LAST(&memory_listeners,
1605                                              memory_listeners)->priority) {
1606         QTAILQ_INSERT_TAIL(&memory_listeners, listener, link);
1607     } else {
1608         QTAILQ_FOREACH(other, &memory_listeners, link) {
1609             if (listener->priority < other->priority) {
1610                 break;
1611             }
1612         }
1613         QTAILQ_INSERT_BEFORE(other, listener, link);
1614     }
1615
1616     QTAILQ_FOREACH(as, &address_spaces, address_spaces_link) {
1617         listener_add_address_space(listener, as);
1618     }
1619 }
1620
1621 void memory_listener_unregister(MemoryListener *listener)
1622 {
1623     QTAILQ_REMOVE(&memory_listeners, listener, link);
1624 }
1625
1626 void address_space_init(AddressSpace *as, MemoryRegion *root)
1627 {
1628     memory_region_transaction_begin();
1629     as->root = root;
1630     as->current_map = g_new(FlatView, 1);
1631     flatview_init(as->current_map);
1632     as->ioeventfd_nb = 0;
1633     as->ioeventfds = NULL;
1634     QTAILQ_INSERT_TAIL(&address_spaces, as, address_spaces_link);
1635     as->name = NULL;
1636     address_space_init_dispatch(as);
1637     memory_region_update_pending |= root->enabled;
1638     memory_region_transaction_commit();
1639 }
1640
1641 void address_space_destroy(AddressSpace *as)
1642 {
1643     /* Flush out anything from MemoryListeners listening in on this */
1644     memory_region_transaction_begin();
1645     as->root = NULL;
1646     memory_region_transaction_commit();
1647     QTAILQ_REMOVE(&address_spaces, as, address_spaces_link);
1648     address_space_destroy_dispatch(as);
1649     flatview_destroy(as->current_map);
1650     g_free(as->current_map);
1651     g_free(as->ioeventfds);
1652 }
1653
1654 bool io_mem_read(MemoryRegion *mr, hwaddr addr, uint64_t *pval, unsigned size)
1655 {
1656     return memory_region_dispatch_read(mr, addr, pval, size);
1657 }
1658
1659 bool io_mem_write(MemoryRegion *mr, hwaddr addr,
1660                   uint64_t val, unsigned size)
1661 {
1662     return memory_region_dispatch_write(mr, addr, val, size);
1663 }
1664
1665 typedef struct MemoryRegionList MemoryRegionList;
1666
1667 struct MemoryRegionList {
1668     const MemoryRegion *mr;
1669     bool printed;
1670     QTAILQ_ENTRY(MemoryRegionList) queue;
1671 };
1672
1673 typedef QTAILQ_HEAD(queue, MemoryRegionList) MemoryRegionListHead;
1674
1675 static void mtree_print_mr(fprintf_function mon_printf, void *f,
1676                            const MemoryRegion *mr, unsigned int level,
1677                            hwaddr base,
1678                            MemoryRegionListHead *alias_print_queue)
1679 {
1680     MemoryRegionList *new_ml, *ml, *next_ml;
1681     MemoryRegionListHead submr_print_queue;
1682     const MemoryRegion *submr;
1683     unsigned int i;
1684
1685     if (!mr || !mr->enabled) {
1686         return;
1687     }
1688
1689     for (i = 0; i < level; i++) {
1690         mon_printf(f, "  ");
1691     }
1692
1693     if (mr->alias) {
1694         MemoryRegionList *ml;
1695         bool found = false;
1696
1697         /* check if the alias is already in the queue */
1698         QTAILQ_FOREACH(ml, alias_print_queue, queue) {
1699             if (ml->mr == mr->alias && !ml->printed) {
1700                 found = true;
1701             }
1702         }
1703
1704         if (!found) {
1705             ml = g_new(MemoryRegionList, 1);
1706             ml->mr = mr->alias;
1707             ml->printed = false;
1708             QTAILQ_INSERT_TAIL(alias_print_queue, ml, queue);
1709         }
1710         mon_printf(f, TARGET_FMT_plx "-" TARGET_FMT_plx
1711                    " (prio %d, %c%c): alias %s @%s " TARGET_FMT_plx
1712                    "-" TARGET_FMT_plx "\n",
1713                    base + mr->addr,
1714                    base + mr->addr
1715                    + (hwaddr)int128_get64(mr->size) - 1,
1716                    mr->priority,
1717                    mr->romd_mode ? 'R' : '-',
1718                    !mr->readonly && !(mr->rom_device && mr->romd_mode) ? 'W'
1719                                                                        : '-',
1720                    mr->name,
1721                    mr->alias->name,
1722                    mr->alias_offset,
1723                    mr->alias_offset
1724                    + (hwaddr)int128_get64(mr->size) - 1);
1725     } else {
1726         mon_printf(f,
1727                    TARGET_FMT_plx "-" TARGET_FMT_plx " (prio %d, %c%c): %s\n",
1728                    base + mr->addr,
1729                    base + mr->addr
1730                    + (hwaddr)int128_get64(mr->size) - 1,
1731                    mr->priority,
1732                    mr->romd_mode ? 'R' : '-',
1733                    !mr->readonly && !(mr->rom_device && mr->romd_mode) ? 'W'
1734                                                                        : '-',
1735                    mr->name);
1736     }
1737
1738     QTAILQ_INIT(&submr_print_queue);
1739
1740     QTAILQ_FOREACH(submr, &mr->subregions, subregions_link) {
1741         new_ml = g_new(MemoryRegionList, 1);
1742         new_ml->mr = submr;
1743         QTAILQ_FOREACH(ml, &submr_print_queue, queue) {
1744             if (new_ml->mr->addr < ml->mr->addr ||
1745                 (new_ml->mr->addr == ml->mr->addr &&
1746                  new_ml->mr->priority > ml->mr->priority)) {
1747                 QTAILQ_INSERT_BEFORE(ml, new_ml, queue);
1748                 new_ml = NULL;
1749                 break;
1750             }
1751         }
1752         if (new_ml) {
1753             QTAILQ_INSERT_TAIL(&submr_print_queue, new_ml, queue);
1754         }
1755     }
1756
1757     QTAILQ_FOREACH(ml, &submr_print_queue, queue) {
1758         mtree_print_mr(mon_printf, f, ml->mr, level + 1, base + mr->addr,
1759                        alias_print_queue);
1760     }
1761
1762     QTAILQ_FOREACH_SAFE(ml, &submr_print_queue, queue, next_ml) {
1763         g_free(ml);
1764     }
1765 }
1766
1767 void mtree_info(fprintf_function mon_printf, void *f)
1768 {
1769     MemoryRegionListHead ml_head;
1770     MemoryRegionList *ml, *ml2;
1771     AddressSpace *as;
1772
1773     QTAILQ_INIT(&ml_head);
1774
1775     QTAILQ_FOREACH(as, &address_spaces, address_spaces_link) {
1776         if (!as->name) {
1777             continue;
1778         }
1779         mon_printf(f, "%s\n", as->name);
1780         mtree_print_mr(mon_printf, f, as->root, 0, 0, &ml_head);
1781     }
1782
1783     mon_printf(f, "aliases\n");
1784     /* print aliased regions */
1785     QTAILQ_FOREACH(ml, &ml_head, queue) {
1786         if (!ml->printed) {
1787             mon_printf(f, "%s\n", ml->mr->name);
1788             mtree_print_mr(mon_printf, f, ml->mr, 0, 0, &ml_head);
1789         }
1790     }
1791
1792     QTAILQ_FOREACH_SAFE(ml, &ml_head, queue, ml2) {
1793         g_free(ml);
1794     }
1795 }