memory.c

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