memory.c

   1 /*
   2  * Physical memory management
   3  *
   4  * Copyright 2011 Red Hat, Inc. and/or its affiliates
   5  *
   6  * Authors:
   7  *  Avi Kivity <avi@redhat.com>
   8  *
   9  * This work is licensed under the terms of the GNU GPL, version 2.  See
  10  * the COPYING file in the top-level directory.
  11  *
  12  */
  13
  14 #include "memory.h"
  15 #include "exec-memory.h"
  16 #include "ioport.h"
  17 #include "bitops.h"
  18 #include "kvm.h"
  19 #include <assert.h>
  20
  21 unsigned memory_region_transaction_depth = 0;
  22
  23 typedef struct AddrRange AddrRange;
  24
  25 struct AddrRange {
  26     uint64_t start;
  27     uint64_t size;
  28 };
  29
  30 static AddrRange addrrange_make(uint64_t start, uint64_t size)
  31 {
  32     return (AddrRange) { start, size };
  33 }
  34
  35 static bool addrrange_equal(AddrRange r1, AddrRange r2)
  36 {
  37     return r1.start == r2.start && r1.size == r2.size;
  38 }
  39
  40 static uint64_t addrrange_end(AddrRange r)
  41 {
  42     return r.start + r.size;
  43 }
  44
  45 static AddrRange addrrange_shift(AddrRange range, int64_t delta)
  46 {
  47     range.start += delta;
  48     return range;
  49 }
  50
  51 static bool addrrange_intersects(AddrRange r1, AddrRange r2)
  52 {
  53     return (r1.start >= r2.start && r1.start < r2.start + r2.size)
  54         || (r2.start >= r1.start && r2.start < r1.start + r1.size);
  55 }
  56
  57 static AddrRange addrrange_intersection(AddrRange r1, AddrRange r2)
  58 {
  59     uint64_t start = MAX(r1.start, r2.start);
  60     /* off-by-one arithmetic to prevent overflow */
  61     uint64_t end = MIN(addrrange_end(r1) - 1, addrrange_end(r2) - 1);
  62     return addrrange_make(start, end - start + 1);
  63 }
  64
  65 struct CoalescedMemoryRange {
  66     AddrRange addr;
  67     QTAILQ_ENTRY(CoalescedMemoryRange) link;
  68 };
  69
  70 struct MemoryRegionIoeventfd {
  71     AddrRange addr;
  72     bool match_data;
  73     uint64_t data;
  74     int fd;
  75 };
  76
  77 static bool memory_region_ioeventfd_before(MemoryRegionIoeventfd a,
  78                                            MemoryRegionIoeventfd b)
  79 {
  80     if (a.addr.start < b.addr.start) {
  81         return true;
  82     } else if (a.addr.start > b.addr.start) {
  83         return false;
  84     } else if (a.addr.size < b.addr.size) {
  85         return true;
  86     } else if (a.addr.size > b.addr.size) {
  87         return false;
  88     } else if (a.match_data < b.match_data) {
  89         return true;
  90     } else  if (a.match_data > b.match_data) {
  91         return false;
  92     } else if (a.match_data) {
  93         if (a.data < b.data) {
  94             return true;
  95         } else if (a.data > b.data) {
  96             return false;
  97         }
  98     }
  99     if (a.fd < b.fd) {
 100         return true;
 101     } else if (a.fd > b.fd) {
 102         return false;
 103     }
 104     return false;
 105 }
 106
 107 static bool memory_region_ioeventfd_equal(MemoryRegionIoeventfd a,
 108                                           MemoryRegionIoeventfd b)
 109 {
 110     return !memory_region_ioeventfd_before(a, b)
 111         && !memory_region_ioeventfd_before(b, a);
 112 }
 113
 114 typedef struct FlatRange FlatRange;
 115 typedef struct FlatView FlatView;
 116
 117 /* Range of memory in the global map.  Addresses are absolute. */
 118 struct FlatRange {
 119     MemoryRegion *mr;
 120     target_phys_addr_t offset_in_region;
 121     AddrRange addr;
 122     uint8_t dirty_log_mask;
 123 };
 124
 125 /* Flattened global view of current active memory hierarchy.  Kept in sorted
 126  * order.
 127  */
 128 struct FlatView {
 129     FlatRange *ranges;
 130     unsigned nr;
 131     unsigned nr_allocated;
 132 };
 133
 134 typedef struct AddressSpace AddressSpace;
 135 typedef struct AddressSpaceOps AddressSpaceOps;
 136
 137 /* A system address space - I/O, memory, etc. */
 138 struct AddressSpace {
 139     const AddressSpaceOps *ops;
 140     MemoryRegion *root;
 141     FlatView current_map;
 142     int ioeventfd_nb;
 143     MemoryRegionIoeventfd *ioeventfds;
 144 };
 145
 146 struct AddressSpaceOps {
 147     void (*range_add)(AddressSpace *as, FlatRange *fr);
 148     void (*range_del)(AddressSpace *as, FlatRange *fr);
 149     void (*log_start)(AddressSpace *as, FlatRange *fr);
 150     void (*log_stop)(AddressSpace *as, FlatRange *fr);
 151     void (*ioeventfd_add)(AddressSpace *as, MemoryRegionIoeventfd *fd);
 152     void (*ioeventfd_del)(AddressSpace *as, MemoryRegionIoeventfd *fd);
 153 };
 154
 155 #define FOR_EACH_FLAT_RANGE(var, view)          \
 156     for (var = (view)->ranges; var < (view)->ranges + (view)->nr; ++var)
 157
 158 static bool flatrange_equal(FlatRange *a, FlatRange *b)
 159 {
 160     return a->mr == b->mr
 161         && addrrange_equal(a->addr, b->addr)
 162         && a->offset_in_region == b->offset_in_region;
 163 }
 164
 165 static void flatview_init(FlatView *view)
 166 {
 167     view->ranges = NULL;
 168     view->nr = 0;
 169     view->nr_allocated = 0;
 170 }
 171
 172 /* Insert a range into a given position.  Caller is responsible for maintaining
 173  * sorting order.
 174  */
 175 static void flatview_insert(FlatView *view, unsigned pos, FlatRange *range)
 176 {
 177     if (view->nr == view->nr_allocated) {
 178         view->nr_allocated = MAX(2 * view->nr, 10);
 179         view->ranges = qemu_realloc(view->ranges,
 180                                     view->nr_allocated * sizeof(*view->ranges));
 181     }
 182     memmove(view->ranges + pos + 1, view->ranges + pos,
 183             (view->nr - pos) * sizeof(FlatRange));
 184     view->ranges[pos] = *range;
 185     ++view->nr;
 186 }
 187
 188 static void flatview_destroy(FlatView *view)
 189 {
 190     qemu_free(view->ranges);
 191 }
 192
 193 static bool can_merge(FlatRange *r1, FlatRange *r2)
 194 {
 195     return addrrange_end(r1->addr) == r2->addr.start
 196         && r1->mr == r2->mr
 197         && r1->offset_in_region + r1->addr.size == r2->offset_in_region
 198         && r1->dirty_log_mask == r2->dirty_log_mask;
 199 }
 200
 201 /* Attempt to simplify a view by merging ajacent ranges */
 202 static void flatview_simplify(FlatView *view)
 203 {
 204     unsigned i, j;
 205
 206     i = 0;
 207     while (i < view->nr) {
 208         j = i + 1;
 209         while (j < view->nr
 210                && can_merge(&view->ranges[j-1], &view->ranges[j])) {
 211             view->ranges[i].addr.size += view->ranges[j].addr.size;
 212             ++j;
 213         }
 214         ++i;
 215         memmove(&view->ranges[i], &view->ranges[j],
 216                 (view->nr - j) * sizeof(view->ranges[j]));
 217         view->nr -= j - i;
 218     }
 219 }
 220
 221 static void memory_region_prepare_ram_addr(MemoryRegion *mr);
 222
 223 static void as_memory_range_add(AddressSpace *as, FlatRange *fr)
 224 {
 225     ram_addr_t phys_offset, region_offset;
 226
 227     memory_region_prepare_ram_addr(fr->mr);
 228
 229     phys_offset = fr->mr->ram_addr;
 230     region_offset = fr->offset_in_region;
 231     /* cpu_register_physical_memory_log() wants region_offset for
 232      * mmio, but prefers offseting phys_offset for RAM.  Humour it.
 233      */
 234     if ((phys_offset & ~TARGET_PAGE_MASK) <= IO_MEM_ROM) {
 235         phys_offset += region_offset;
 236         region_offset = 0;
 237     }
 238
 239     cpu_register_physical_memory_log(fr->addr.start,
 240                                      fr->addr.size,
 241                                      phys_offset,
 242                                      region_offset,
 243                                      fr->dirty_log_mask);
 244 }
 245
 246 static void as_memory_range_del(AddressSpace *as, FlatRange *fr)
 247 {
 248     cpu_register_physical_memory(fr->addr.start, fr->addr.size,
 249                                  IO_MEM_UNASSIGNED);
 250 }
 251
 252 static void as_memory_log_start(AddressSpace *as, FlatRange *fr)
 253 {
 254     cpu_physical_log_start(fr->addr.start, fr->addr.size);
 255 }
 256
 257 static void as_memory_log_stop(AddressSpace *as, FlatRange *fr)
 258 {
 259     cpu_physical_log_stop(fr->addr.start, fr->addr.size);
 260 }
 261
 262 static void as_memory_ioeventfd_add(AddressSpace *as, MemoryRegionIoeventfd *fd)
 263 {
 264     int r;
 265
 266     assert(fd->match_data && fd->addr.size == 4);
 267
 268     r = kvm_set_ioeventfd_mmio_long(fd->fd, fd->addr.start, fd->data, true);
 269     if (r < 0) {
 270         abort();
 271     }
 272 }
 273
 274 static void as_memory_ioeventfd_del(AddressSpace *as, MemoryRegionIoeventfd *fd)
 275 {
 276     int r;
 277
 278     r = kvm_set_ioeventfd_mmio_long(fd->fd, fd->addr.start, fd->data, false);
 279     if (r < 0) {
 280         abort();
 281     }
 282 }
 283
 284 static const AddressSpaceOps address_space_ops_memory = {
 285     .range_add = as_memory_range_add,
 286     .range_del = as_memory_range_del,
 287     .log_start = as_memory_log_start,
 288     .log_stop = as_memory_log_stop,
 289     .ioeventfd_add = as_memory_ioeventfd_add,
 290     .ioeventfd_del = as_memory_ioeventfd_del,
 291 };
 292
 293 static AddressSpace address_space_memory = {
 294     .ops = &address_space_ops_memory,
 295 };
 296
 297 static const MemoryRegionPortio *find_portio(MemoryRegion *mr, uint64_t offset,
 298                                              unsigned width, bool write)
 299 {
 300     const MemoryRegionPortio *mrp;
 301
 302     for (mrp = mr->ops->old_portio; mrp->size; ++mrp) {
 303         if (offset >= mrp->offset && offset < mrp->offset + mrp->len
 304             && width == mrp->size
 305             && (write ? (bool)mrp->write : (bool)mrp->read)) {
 306             return mrp;
 307         }
 308     }
 309     return NULL;
 310 }
 311
 312 static void memory_region_iorange_read(IORange *iorange,
 313                                        uint64_t offset,
 314                                        unsigned width,
 315                                        uint64_t *data)
 316 {
 317     MemoryRegion *mr = container_of(iorange, MemoryRegion, iorange);
 318
 319     if (mr->ops->old_portio) {
 320         const MemoryRegionPortio *mrp = find_portio(mr, offset, width, false);
 321
 322         *data = ((uint64_t)1 << (width * 8)) - 1;
 323         if (mrp) {
 324             *data = mrp->read(mr->opaque, offset - mrp->offset);
 325         }
 326         return;
 327     }
 328     *data = mr->ops->read(mr->opaque, offset, width);
 329 }
 330
 331 static void memory_region_iorange_write(IORange *iorange,
 332                                         uint64_t offset,
 333                                         unsigned width,
 334                                         uint64_t data)
 335 {
 336     MemoryRegion *mr = container_of(iorange, MemoryRegion, iorange);
 337
 338     if (mr->ops->old_portio) {
 339         const MemoryRegionPortio *mrp = find_portio(mr, offset, width, true);
 340
 341         if (mrp) {
 342             mrp->write(mr->opaque, offset - mrp->offset, data);
 343         }
 344         return;
 345     }
 346     mr->ops->write(mr->opaque, offset, data, width);
 347 }
 348
 349 static const IORangeOps memory_region_iorange_ops = {
 350     .read = memory_region_iorange_read,
 351     .write = memory_region_iorange_write,
 352 };
 353
 354 static void as_io_range_add(AddressSpace *as, FlatRange *fr)
 355 {
 356     iorange_init(&fr->mr->iorange, &memory_region_iorange_ops,
 357                  fr->addr.start,fr->addr.size);
 358     ioport_register(&fr->mr->iorange);
 359 }
 360
 361 static void as_io_range_del(AddressSpace *as, FlatRange *fr)
 362 {
 363     isa_unassign_ioport(fr->addr.start, fr->addr.size);
 364 }
 365
 366 static void as_io_ioeventfd_add(AddressSpace *as, MemoryRegionIoeventfd *fd)
 367 {
 368     int r;
 369
 370     assert(fd->match_data && fd->addr.size == 2);
 371
 372     r = kvm_set_ioeventfd_pio_word(fd->fd, fd->addr.start, fd->data, true);
 373     if (r < 0) {
 374         abort();
 375     }
 376 }
 377
 378 static void as_io_ioeventfd_del(AddressSpace *as, MemoryRegionIoeventfd *fd)
 379 {
 380     int r;
 381
 382     r = kvm_set_ioeventfd_pio_word(fd->fd, fd->addr.start, fd->data, false);
 383     if (r < 0) {
 384         abort();
 385     }
 386 }
 387
 388 static const AddressSpaceOps address_space_ops_io = {
 389     .range_add = as_io_range_add,
 390     .range_del = as_io_range_del,
 391     .ioeventfd_add = as_io_ioeventfd_add,
 392     .ioeventfd_del = as_io_ioeventfd_del,
 393 };
 394
 395 static AddressSpace address_space_io = {
 396     .ops = &address_space_ops_io,
 397 };
 398
 399 /* Render a memory region into the global view.  Ranges in @view obscure
 400  * ranges in @mr.
 401  */
 402 static void render_memory_region(FlatView *view,
 403                                  MemoryRegion *mr,
 404                                  target_phys_addr_t base,
 405                                  AddrRange clip)
 406 {
 407     MemoryRegion *subregion;
 408     unsigned i;
 409     target_phys_addr_t offset_in_region;
 410     uint64_t remain;
 411     uint64_t now;
 412     FlatRange fr;
 413     AddrRange tmp;
 414
 415     base += mr->addr;
 416
 417     tmp = addrrange_make(base, mr->size);
 418
 419     if (!addrrange_intersects(tmp, clip)) {
 420         return;
 421     }
 422
 423     clip = addrrange_intersection(tmp, clip);
 424
 425     if (mr->alias) {
 426         base -= mr->alias->addr;
 427         base -= mr->alias_offset;
 428         render_memory_region(view, mr->alias, base, clip);
 429         return;
 430     }
 431
 432     /* Render subregions in priority order. */
 433     QTAILQ_FOREACH(subregion, &mr->subregions, subregions_link) {
 434         render_memory_region(view, subregion, base, clip);
 435     }
 436
 437     if (!mr->terminates) {
 438         return;
 439     }
 440
 441     offset_in_region = clip.start - base;
 442     base = clip.start;
 443     remain = clip.size;
 444
 445     /* Render the region itself into any gaps left by the current view. */
 446     for (i = 0; i < view->nr && remain; ++i) {
 447         if (base >= addrrange_end(view->ranges[i].addr)) {
 448             continue;
 449         }
 450         if (base < view->ranges[i].addr.start) {
 451             now = MIN(remain, view->ranges[i].addr.start - base);
 452             fr.mr = mr;
 453             fr.offset_in_region = offset_in_region;
 454             fr.addr = addrrange_make(base, now);
 455             fr.dirty_log_mask = mr->dirty_log_mask;
 456             flatview_insert(view, i, &fr);
 457             ++i;
 458             base += now;
 459             offset_in_region += now;
 460             remain -= now;
 461         }
 462         if (base == view->ranges[i].addr.start) {
 463             now = MIN(remain, view->ranges[i].addr.size);
 464             base += now;
 465             offset_in_region += now;
 466             remain -= now;
 467         }
 468     }
 469     if (remain) {
 470         fr.mr = mr;
 471         fr.offset_in_region = offset_in_region;
 472         fr.addr = addrrange_make(base, remain);
 473         fr.dirty_log_mask = mr->dirty_log_mask;
 474         flatview_insert(view, i, &fr);
 475     }
 476 }
 477
 478 /* Render a memory topology into a list of disjoint absolute ranges. */
 479 static FlatView generate_memory_topology(MemoryRegion *mr)
 480 {
 481     FlatView view;
 482
 483     flatview_init(&view);
 484
 485     render_memory_region(&view, mr, 0, addrrange_make(0, UINT64_MAX));
 486     flatview_simplify(&view);
 487
 488     return view;
 489 }
 490
 491 static void address_space_add_del_ioeventfds(AddressSpace *as,
 492                                              MemoryRegionIoeventfd *fds_new,
 493                                              unsigned fds_new_nb,
 494                                              MemoryRegionIoeventfd *fds_old,
 495                                              unsigned fds_old_nb)
 496 {
 497     unsigned iold, inew;
 498
 499     /* Generate a symmetric difference of the old and new fd sets, adding
 500      * and deleting as necessary.
 501      */
 502
 503     iold = inew = 0;
 504     while (iold < fds_old_nb || inew < fds_new_nb) {
 505         if (iold < fds_old_nb
 506             && (inew == fds_new_nb
 507                 || memory_region_ioeventfd_before(fds_old[iold],
 508                                                   fds_new[inew]))) {
 509             as->ops->ioeventfd_del(as, &fds_old[iold]);
 510             ++iold;
 511         } else if (inew < fds_new_nb
 512                    && (iold == fds_old_nb
 513                        || memory_region_ioeventfd_before(fds_new[inew],
 514                                                          fds_old[iold]))) {
 515             as->ops->ioeventfd_add(as, &fds_new[inew]);
 516             ++inew;
 517         } else {
 518             ++iold;
 519             ++inew;
 520         }
 521     }
 522 }
 523
 524 static void address_space_update_ioeventfds(AddressSpace *as)
 525 {
 526     FlatRange *fr;
 527     unsigned ioeventfd_nb = 0;
 528     MemoryRegionIoeventfd *ioeventfds = NULL;
 529     AddrRange tmp;
 530     unsigned i;
 531
 532     FOR_EACH_FLAT_RANGE(fr, &as->current_map) {
 533         for (i = 0; i < fr->mr->ioeventfd_nb; ++i) {
 534             tmp = addrrange_shift(fr->mr->ioeventfds[i].addr,
 535                                   fr->addr.start - fr->offset_in_region);
 536             if (addrrange_intersects(fr->addr, tmp)) {
 537                 ++ioeventfd_nb;
 538                 ioeventfds = qemu_realloc(ioeventfds,
 539                                           ioeventfd_nb * sizeof(*ioeventfds));
 540                 ioeventfds[ioeventfd_nb-1] = fr->mr->ioeventfds[i];
 541                 ioeventfds[ioeventfd_nb-1].addr = tmp;
 542             }
 543         }
 544     }
 545
 546     address_space_add_del_ioeventfds(as, ioeventfds, ioeventfd_nb,
 547                                      as->ioeventfds, as->ioeventfd_nb);
 548
 549     qemu_free(as->ioeventfds);
 550     as->ioeventfds = ioeventfds;
 551     as->ioeventfd_nb = ioeventfd_nb;
 552 }
 553
 554 static void address_space_update_topology_pass(AddressSpace *as,
 555                                                FlatView old_view,
 556                                                FlatView new_view,
 557                                                bool adding)
 558 {
 559     unsigned iold, inew;
 560     FlatRange *frold, *frnew;
 561
 562     /* Generate a symmetric difference of the old and new memory maps.
 563      * Kill ranges in the old map, and instantiate ranges in the new map.
 564      */
 565     iold = inew = 0;
 566     while (iold < old_view.nr || inew < new_view.nr) {
 567         if (iold < old_view.nr) {
 568             frold = &old_view.ranges[iold];
 569         } else {
 570             frold = NULL;
 571         }
 572         if (inew < new_view.nr) {
 573             frnew = &new_view.ranges[inew];
 574         } else {
 575             frnew = NULL;
 576         }
 577
 578         if (frold
 579             && (!frnew
 580                 || frold->addr.start < frnew->addr.start
 581                 || (frold->addr.start == frnew->addr.start
 582                     && !flatrange_equal(frold, frnew)))) {
 583             /* In old, but (not in new, or in new but attributes changed). */
 584
 585             if (!adding) {
 586                 as->ops->range_del(as, frold);
 587             }
 588
 589             ++iold;
 590         } else if (frold && frnew && flatrange_equal(frold, frnew)) {
 591             /* In both (logging may have changed) */
 592
 593             if (adding) {
 594                 if (frold->dirty_log_mask && !frnew->dirty_log_mask) {
 595                     as->ops->log_stop(as, frnew);
 596                 } else if (frnew->dirty_log_mask && !frold->dirty_log_mask) {
 597                     as->ops->log_start(as, frnew);
 598                 }
 599             }
 600
 601             ++iold;
 602             ++inew;
 603         } else {
 604             /* In new */
 605
 606             if (adding) {
 607                 as->ops->range_add(as, frnew);
 608             }
 609
 610             ++inew;
 611         }
 612     }
 613 }
 614
 615
 616 static void address_space_update_topology(AddressSpace *as)
 617 {
 618     FlatView old_view = as->current_map;
 619     FlatView new_view = generate_memory_topology(as->root);
 620
 621     address_space_update_topology_pass(as, old_view, new_view, false);
 622     address_space_update_topology_pass(as, old_view, new_view, true);
 623
 624     as->current_map = new_view;
 625     flatview_destroy(&old_view);
 626     address_space_update_ioeventfds(as);
 627 }
 628
 629 static void memory_region_update_topology(void)
 630 {
 631     if (memory_region_transaction_depth) {
 632         return;
 633     }
 634
 635     if (address_space_memory.root) {
 636         address_space_update_topology(&address_space_memory);
 637     }
 638     if (address_space_io.root) {
 639         address_space_update_topology(&address_space_io);
 640     }
 641 }
 642
 643 void memory_region_transaction_begin(void)
 644 {
 645     ++memory_region_transaction_depth;
 646 }
 647
 648 void memory_region_transaction_commit(void)
 649 {
 650     assert(memory_region_transaction_depth);
 651     --memory_region_transaction_depth;
 652     memory_region_update_topology();
 653 }
 654
 655 void memory_region_init(MemoryRegion *mr,
 656                         const char *name,
 657                         uint64_t size)
 658 {
 659     mr->ops = NULL;
 660     mr->parent = NULL;
 661     mr->size = size;
 662     mr->addr = 0;
 663     mr->offset = 0;
 664     mr->terminates = false;
 665     mr->priority = 0;
 666     mr->may_overlap = false;
 667     mr->alias = NULL;
 668     QTAILQ_INIT(&mr->subregions);
 669     memset(&mr->subregions_link, 0, sizeof mr->subregions_link);
 670     QTAILQ_INIT(&mr->coalesced);
 671     mr->name = qemu_strdup(name);
 672     mr->dirty_log_mask = 0;
 673     mr->ioeventfd_nb = 0;
 674     mr->ioeventfds = NULL;
 675 }
 676
 677 static bool memory_region_access_valid(MemoryRegion *mr,
 678                                        target_phys_addr_t addr,
 679                                        unsigned size)
 680 {
 681     if (!mr->ops->valid.unaligned && (addr & (size - 1))) {
 682         return false;
 683     }
 684
 685     /* Treat zero as compatibility all valid */
 686     if (!mr->ops->valid.max_access_size) {
 687         return true;
 688     }
 689
 690     if (size > mr->ops->valid.max_access_size
 691         || size < mr->ops->valid.min_access_size) {
 692         return false;
 693     }
 694     return true;
 695 }
 696
 697 static uint32_t memory_region_read_thunk_n(void *_mr,
 698                                            target_phys_addr_t addr,
 699                                            unsigned size)
 700 {
 701     MemoryRegion *mr = _mr;
 702     unsigned access_size, access_size_min, access_size_max;
 703     uint64_t access_mask;
 704     uint32_t data = 0, tmp;
 705     unsigned i;
 706
 707     if (!memory_region_access_valid(mr, addr, size)) {
 708         return -1U; /* FIXME: better signalling */
 709     }
 710
 711     if (!mr->ops->read) {
 712         return mr->ops->old_mmio.read[bitops_ffsl(size)](mr->opaque, addr);
 713     }
 714
 715     /* FIXME: support unaligned access */
 716
 717     access_size_min = mr->ops->impl.min_access_size;
 718     if (!access_size_min) {
 719         access_size_min = 1;
 720     }
 721     access_size_max = mr->ops->impl.max_access_size;
 722     if (!access_size_max) {
 723         access_size_max = 4;
 724     }
 725     access_size = MAX(MIN(size, access_size_max), access_size_min);
 726     access_mask = -1ULL >> (64 - access_size * 8);
 727     addr += mr->offset;
 728     for (i = 0; i < size; i += access_size) {
 729         /* FIXME: big-endian support */
 730         tmp = mr->ops->read(mr->opaque, addr + i, access_size);
 731         data |= (tmp & access_mask) << (i * 8);
 732     }
 733
 734     return data;
 735 }
 736
 737 static void memory_region_write_thunk_n(void *_mr,
 738                                         target_phys_addr_t addr,
 739                                         unsigned size,
 740                                         uint64_t data)
 741 {
 742     MemoryRegion *mr = _mr;
 743     unsigned access_size, access_size_min, access_size_max;
 744     uint64_t access_mask;
 745     unsigned i;
 746
 747     if (!memory_region_access_valid(mr, addr, size)) {
 748         return; /* FIXME: better signalling */
 749     }
 750
 751     if (!mr->ops->write) {
 752         mr->ops->old_mmio.write[bitops_ffsl(size)](mr->opaque, addr, data);
 753         return;
 754     }
 755
 756     /* FIXME: support unaligned access */
 757
 758     access_size_min = mr->ops->impl.min_access_size;
 759     if (!access_size_min) {
 760         access_size_min = 1;
 761     }
 762     access_size_max = mr->ops->impl.max_access_size;
 763     if (!access_size_max) {
 764         access_size_max = 4;
 765     }
 766     access_size = MAX(MIN(size, access_size_max), access_size_min);
 767     access_mask = -1ULL >> (64 - access_size * 8);
 768     addr += mr->offset;
 769     for (i = 0; i < size; i += access_size) {
 770         /* FIXME: big-endian support */
 771         mr->ops->write(mr->opaque, addr + i, (data >> (i * 8)) & access_mask,
 772                        access_size);
 773     }
 774 }
 775
 776 static uint32_t memory_region_read_thunk_b(void *mr, target_phys_addr_t addr)
 777 {
 778     return memory_region_read_thunk_n(mr, addr, 1);
 779 }
 780
 781 static uint32_t memory_region_read_thunk_w(void *mr, target_phys_addr_t addr)
 782 {
 783     return memory_region_read_thunk_n(mr, addr, 2);
 784 }
 785
 786 static uint32_t memory_region_read_thunk_l(void *mr, target_phys_addr_t addr)
 787 {
 788     return memory_region_read_thunk_n(mr, addr, 4);
 789 }
 790
 791 static void memory_region_write_thunk_b(void *mr, target_phys_addr_t addr,
 792                                         uint32_t data)
 793 {
 794     memory_region_write_thunk_n(mr, addr, 1, data);
 795 }
 796
 797 static void memory_region_write_thunk_w(void *mr, target_phys_addr_t addr,
 798                                         uint32_t data)
 799 {
 800     memory_region_write_thunk_n(mr, addr, 2, data);
 801 }
 802
 803 static void memory_region_write_thunk_l(void *mr, target_phys_addr_t addr,
 804                                         uint32_t data)
 805 {
 806     memory_region_write_thunk_n(mr, addr, 4, data);
 807 }
 808
 809 static CPUReadMemoryFunc * const memory_region_read_thunk[] = {
 810     memory_region_read_thunk_b,
 811     memory_region_read_thunk_w,
 812     memory_region_read_thunk_l,
 813 };
 814
 815 static CPUWriteMemoryFunc * const memory_region_write_thunk[] = {
 816     memory_region_write_thunk_b,
 817     memory_region_write_thunk_w,
 818     memory_region_write_thunk_l,
 819 };
 820
 821 static void memory_region_prepare_ram_addr(MemoryRegion *mr)
 822 {
 823     if (mr->backend_registered) {
 824         return;
 825     }
 826
 827     mr->ram_addr = cpu_register_io_memory(memory_region_read_thunk,
 828                                           memory_region_write_thunk,
 829                                           mr,
 830                                           mr->ops->endianness);
 831     mr->backend_registered = true;
 832 }
 833
 834 void memory_region_init_io(MemoryRegion *mr,
 835                            const MemoryRegionOps *ops,
 836                            void *opaque,
 837                            const char *name,
 838                            uint64_t size)
 839 {
 840     memory_region_init(mr, name, size);
 841     mr->ops = ops;
 842     mr->opaque = opaque;
 843     mr->terminates = true;
 844     mr->backend_registered = false;
 845 }
 846
 847 void memory_region_init_ram(MemoryRegion *mr,
 848                             DeviceState *dev,
 849                             const char *name,
 850                             uint64_t size)
 851 {
 852     memory_region_init(mr, name, size);
 853     mr->terminates = true;
 854     mr->ram_addr = qemu_ram_alloc(dev, name, size);
 855     mr->backend_registered = true;
 856 }
 857
 858 void memory_region_init_ram_ptr(MemoryRegion *mr,
 859                                 DeviceState *dev,
 860                                 const char *name,
 861                                 uint64_t size,
 862                                 void *ptr)
 863 {
 864     memory_region_init(mr, name, size);
 865     mr->terminates = true;
 866     mr->ram_addr = qemu_ram_alloc_from_ptr(dev, name, size, ptr);
 867     mr->backend_registered = true;
 868 }
 869
 870 void memory_region_init_alias(MemoryRegion *mr,
 871                               const char *name,
 872                               MemoryRegion *orig,
 873                               target_phys_addr_t offset,
 874                               uint64_t size)
 875 {
 876     memory_region_init(mr, name, size);
 877     mr->alias = orig;
 878     mr->alias_offset = offset;
 879 }
 880
 881 void memory_region_destroy(MemoryRegion *mr)
 882 {
 883     assert(QTAILQ_EMPTY(&mr->subregions));
 884     memory_region_clear_coalescing(mr);
 885     qemu_free((char *)mr->name);
 886     qemu_free(mr->ioeventfds);
 887 }
 888
 889 uint64_t memory_region_size(MemoryRegion *mr)
 890 {
 891     return mr->size;
 892 }
 893
 894 void memory_region_set_offset(MemoryRegion *mr, target_phys_addr_t offset)
 895 {
 896     mr->offset = offset;
 897 }
 898
 899 void memory_region_set_log(MemoryRegion *mr, bool log, unsigned client)
 900 {
 901     uint8_t mask = 1 << client;
 902
 903     mr->dirty_log_mask = (mr->dirty_log_mask & ~mask) | (log * mask);
 904     memory_region_update_topology();
 905 }
 906
 907 bool memory_region_get_dirty(MemoryRegion *mr, target_phys_addr_t addr,
 908                              unsigned client)
 909 {
 910     assert(mr->terminates);
 911     return cpu_physical_memory_get_dirty(mr->ram_addr + addr, 1 << client);
 912 }
 913
 914 void memory_region_set_dirty(MemoryRegion *mr, target_phys_addr_t addr)
 915 {
 916     assert(mr->terminates);
 917     return cpu_physical_memory_set_dirty(mr->ram_addr + addr);
 918 }
 919
 920 void memory_region_sync_dirty_bitmap(MemoryRegion *mr)
 921 {
 922     FlatRange *fr;
 923
 924     FOR_EACH_FLAT_RANGE(fr, &address_space_memory.current_map) {
 925         if (fr->mr == mr) {
 926             cpu_physical_sync_dirty_bitmap(fr->addr.start,
 927                                            fr->addr.start + fr->addr.size);
 928         }
 929     }
 930 }
 931
 932 void memory_region_set_readonly(MemoryRegion *mr, bool readonly)
 933 {
 934     /* FIXME */
 935 }
 936
 937 void memory_region_reset_dirty(MemoryRegion *mr, target_phys_addr_t addr,
 938                                target_phys_addr_t size, unsigned client)
 939 {
 940     assert(mr->terminates);
 941     cpu_physical_memory_reset_dirty(mr->ram_addr + addr,
 942                                     mr->ram_addr + addr + size,
 943                                     1 << client);
 944 }
 945
 946 void *memory_region_get_ram_ptr(MemoryRegion *mr)
 947 {
 948     if (mr->alias) {
 949         return memory_region_get_ram_ptr(mr->alias) + mr->alias_offset;
 950     }
 951
 952     assert(mr->terminates);
 953
 954     return qemu_get_ram_ptr(mr->ram_addr);
 955 }
 956
 957 static void memory_region_update_coalesced_range(MemoryRegion *mr)
 958 {
 959     FlatRange *fr;
 960     CoalescedMemoryRange *cmr;
 961     AddrRange tmp;
 962
 963     FOR_EACH_FLAT_RANGE(fr, &address_space_memory.current_map) {
 964         if (fr->mr == mr) {
 965             qemu_unregister_coalesced_mmio(fr->addr.start, fr->addr.size);
 966             QTAILQ_FOREACH(cmr, &mr->coalesced, link) {
 967                 tmp = addrrange_shift(cmr->addr,
 968                                       fr->addr.start - fr->offset_in_region);
 969                 if (!addrrange_intersects(tmp, fr->addr)) {
 970                     continue;
 971                 }
 972                 tmp = addrrange_intersection(tmp, fr->addr);
 973                 qemu_register_coalesced_mmio(tmp.start, tmp.size);
 974             }
 975         }
 976     }
 977 }
 978
 979 void memory_region_set_coalescing(MemoryRegion *mr)
 980 {
 981     memory_region_clear_coalescing(mr);
 982     memory_region_add_coalescing(mr, 0, mr->size);
 983 }
 984
 985 void memory_region_add_coalescing(MemoryRegion *mr,
 986                                   target_phys_addr_t offset,
 987                                   uint64_t size)
 988 {
 989     CoalescedMemoryRange *cmr = qemu_malloc(sizeof(*cmr));
 990
 991     cmr->addr = addrrange_make(offset, size);
 992     QTAILQ_INSERT_TAIL(&mr->coalesced, cmr, link);
 993     memory_region_update_coalesced_range(mr);
 994 }
 995
 996 void memory_region_clear_coalescing(MemoryRegion *mr)
 997 {
 998     CoalescedMemoryRange *cmr;
 999
1000     while (!QTAILQ_EMPTY(&mr->coalesced)) {
1001         cmr = QTAILQ_FIRST(&mr->coalesced);
1002         QTAILQ_REMOVE(&mr->coalesced, cmr, link);
1003         qemu_free(cmr);
1004     }
1005     memory_region_update_coalesced_range(mr);
1006 }
1007
1008 void memory_region_add_eventfd(MemoryRegion *mr,
1009                                target_phys_addr_t addr,
1010                                unsigned size,
1011                                bool match_data,
1012                                uint64_t data,
1013                                int fd)
1014 {
1015     MemoryRegionIoeventfd mrfd = {
1016         .addr.start = addr,
1017         .addr.size = size,
1018         .match_data = match_data,
1019         .data = data,
1020         .fd = fd,
1021     };
1022     unsigned i;
1023
1024     for (i = 0; i < mr->ioeventfd_nb; ++i) {
1025         if (memory_region_ioeventfd_before(mrfd, mr->ioeventfds[i])) {
1026             break;
1027         }
1028     }
1029     ++mr->ioeventfd_nb;
1030     mr->ioeventfds = qemu_realloc(mr->ioeventfds,
1031                                   sizeof(*mr->ioeventfds) * mr->ioeventfd_nb);
1032     memmove(&mr->ioeventfds[i+1], &mr->ioeventfds[i],
1033             sizeof(*mr->ioeventfds) * (mr->ioeventfd_nb-1 - i));
1034     mr->ioeventfds[i] = mrfd;
1035     memory_region_update_topology();
1036 }
1037
1038 void memory_region_del_eventfd(MemoryRegion *mr,
1039                                target_phys_addr_t addr,
1040                                unsigned size,
1041                                bool match_data,
1042                                uint64_t data,
1043                                int fd)
1044 {
1045     MemoryRegionIoeventfd mrfd = {
1046         .addr.start = addr,
1047         .addr.size = size,
1048         .match_data = match_data,
1049         .data = data,
1050         .fd = fd,
1051     };
1052     unsigned i;
1053
1054     for (i = 0; i < mr->ioeventfd_nb; ++i) {
1055         if (memory_region_ioeventfd_equal(mrfd, mr->ioeventfds[i])) {
1056             break;
1057         }
1058     }
1059     assert(i != mr->ioeventfd_nb);
1060     memmove(&mr->ioeventfds[i], &mr->ioeventfds[i+1],
1061             sizeof(*mr->ioeventfds) * (mr->ioeventfd_nb - (i+1)));
1062     --mr->ioeventfd_nb;
1063     mr->ioeventfds = qemu_realloc(mr->ioeventfds,
1064                                   sizeof(*mr->ioeventfds)*mr->ioeventfd_nb + 1);
1065     memory_region_update_topology();
1066 }
1067
1068 static void memory_region_add_subregion_common(MemoryRegion *mr,
1069                                                target_phys_addr_t offset,
1070                                                MemoryRegion *subregion)
1071 {
1072     MemoryRegion *other;
1073
1074     assert(!subregion->parent);
1075     subregion->parent = mr;
1076     subregion->addr = offset;
1077     QTAILQ_FOREACH(other, &mr->subregions, subregions_link) {
1078         if (subregion->may_overlap || other->may_overlap) {
1079             continue;
1080         }
1081         if (offset >= other->offset + other->size
1082             || offset + subregion->size <= other->offset) {
1083             continue;
1084         }
1085         printf("warning: subregion collision %llx/%llx vs %llx/%llx\n",
1086                (unsigned long long)offset,
1087                (unsigned long long)subregion->size,
1088                (unsigned long long)other->offset,
1089                (unsigned long long)other->size);
1090     }
1091     QTAILQ_FOREACH(other, &mr->subregions, subregions_link) {
1092         if (subregion->priority >= other->priority) {
1093             QTAILQ_INSERT_BEFORE(other, subregion, subregions_link);
1094             goto done;
1095         }
1096     }
1097     QTAILQ_INSERT_TAIL(&mr->subregions, subregion, subregions_link);
1098 done:
1099     memory_region_update_topology();
1100 }
1101
1102
1103 void memory_region_add_subregion(MemoryRegion *mr,
1104                                  target_phys_addr_t offset,
1105                                  MemoryRegion *subregion)
1106 {
1107     subregion->may_overlap = false;
1108     subregion->priority = 0;
1109     memory_region_add_subregion_common(mr, offset, subregion);
1110 }
1111
1112 void memory_region_add_subregion_overlap(MemoryRegion *mr,
1113                                          target_phys_addr_t offset,
1114                                          MemoryRegion *subregion,
1115                                          unsigned priority)
1116 {
1117     subregion->may_overlap = true;
1118     subregion->priority = priority;
1119     memory_region_add_subregion_common(mr, offset, subregion);
1120 }
1121
1122 void memory_region_del_subregion(MemoryRegion *mr,
1123                                  MemoryRegion *subregion)
1124 {
1125     assert(subregion->parent == mr);
1126     subregion->parent = NULL;
1127     QTAILQ_REMOVE(&mr->subregions, subregion, subregions_link);
1128     memory_region_update_topology();
1129 }
1130
1131 void set_system_memory_map(MemoryRegion *mr)
1132 {
1133     address_space_memory.root = mr;
1134     memory_region_update_topology();
1135 }
1136
1137 void set_system_io_map(MemoryRegion *mr)
1138 {
1139     address_space_io.root = mr;
1140     memory_region_update_topology();
1141 }