migration/postcopy-ram.c

   1 /*
   2  * Postcopy migration for RAM
   3  *
   4  * Copyright 2013-2015 Red Hat, Inc. and/or its affiliates
   5  *
   6  * Authors:
   7  *  Dave Gilbert  <dgilbert@redhat.com>
   8  *
   9  * This work is licensed under the terms of the GNU GPL, version 2 or later.
  10  * See the COPYING file in the top-level directory.
  11  *
  12  */
  13
  14 /*
  15  * Postcopy is a migration technique where the execution flips from the
  16  * source to the destination before all the data has been copied.
  17  */
  18
  19 #include "qemu/osdep.h"
  20 #include <glib.h>
  21
  22 #include "qemu-common.h"
  23 #include "migration/migration.h"
  24 #include "migration/postcopy-ram.h"
  25 #include "sysemu/sysemu.h"
  26 #include "sysemu/balloon.h"
  27 #include "qemu/error-report.h"
  28 #include "trace.h"
  29
  30 /* Arbitrary limit on size of each discard command,
  31  * keeps them around ~200 bytes
  32  */
  33 #define MAX_DISCARDS_PER_COMMAND 12
  34
  35 struct PostcopyDiscardState {
  36     const char *ramblock_name;
  37     uint64_t offset; /* Bitmap entry for the 1st bit of this RAMBlock */
  38     uint16_t cur_entry;
  39     /*
  40      * Start and length of a discard range (bytes)
  41      */
  42     uint64_t start_list[MAX_DISCARDS_PER_COMMAND];
  43     uint64_t length_list[MAX_DISCARDS_PER_COMMAND];
  44     unsigned int nsentwords;
  45     unsigned int nsentcmds;
  46 };
  47
  48 /* Postcopy needs to detect accesses to pages that haven't yet been copied
  49  * across, and efficiently map new pages in, the techniques for doing this
  50  * are target OS specific.
  51  */
  52 #if defined(__linux__)
  53
  54 #include <poll.h>
  55 #include <sys/mman.h>
  56 #include <sys/ioctl.h>
  57 #include <sys/syscall.h>
  58 #include <asm/types.h> /* for __u64 */
  59 #endif
  60
  61 #if defined(__linux__) && defined(__NR_userfaultfd) && defined(CONFIG_EVENTFD)
  62 #include <sys/eventfd.h>
  63 #include <linux/userfaultfd.h>
  64
  65 static bool ufd_version_check(int ufd)
  66 {
  67     struct uffdio_api api_struct;
  68     uint64_t ioctl_mask;
  69
  70     api_struct.api = UFFD_API;
  71     api_struct.features = 0;
  72     if (ioctl(ufd, UFFDIO_API, &api_struct)) {
  73         error_report("postcopy_ram_supported_by_host: UFFDIO_API failed: %s",
  74                      strerror(errno));
  75         return false;
  76     }
  77
  78     ioctl_mask = (__u64)1 << _UFFDIO_REGISTER |
  79                  (__u64)1 << _UFFDIO_UNREGISTER;
  80     if ((api_struct.ioctls & ioctl_mask) != ioctl_mask) {
  81         error_report("Missing userfault features: %" PRIx64,
  82                      (uint64_t)(~api_struct.ioctls & ioctl_mask));
  83         return false;
  84     }
  85
  86     return true;
  87 }
  88
  89 /*
  90  * Note: This has the side effect of munlock'ing all of RAM, that's
  91  * normally fine since if the postcopy succeeds it gets turned back on at the
  92  * end.
  93  */
  94 bool postcopy_ram_supported_by_host(void)
  95 {
  96     long pagesize = getpagesize();
  97     int ufd = -1;
  98     bool ret = false; /* Error unless we change it */
  99     void *testarea = NULL;
 100     struct uffdio_register reg_struct;
 101     struct uffdio_range range_struct;
 102     uint64_t feature_mask;
 103
 104     if ((1ul << qemu_target_page_bits()) > pagesize) {
 105         error_report("Target page size bigger than host page size");
 106         goto out;
 107     }
 108
 109     ufd = syscall(__NR_userfaultfd, O_CLOEXEC);
 110     if (ufd == -1) {
 111         error_report("%s: userfaultfd not available: %s", __func__,
 112                      strerror(errno));
 113         goto out;
 114     }
 115
 116     /* Version and features check */
 117     if (!ufd_version_check(ufd)) {
 118         goto out;
 119     }
 120
 121     /*
 122      * userfault and mlock don't go together; we'll put it back later if
 123      * it was enabled.
 124      */
 125     if (munlockall()) {
 126         error_report("%s: munlockall: %s", __func__,  strerror(errno));
 127         return -1;
 128     }
 129
 130     /*
 131      *  We need to check that the ops we need are supported on anon memory
 132      *  To do that we need to register a chunk and see the flags that
 133      *  are returned.
 134      */
 135     testarea = mmap(NULL, pagesize, PROT_READ | PROT_WRITE, MAP_PRIVATE |
 136                                     MAP_ANONYMOUS, -1, 0);
 137     if (testarea == MAP_FAILED) {
 138         error_report("%s: Failed to map test area: %s", __func__,
 139                      strerror(errno));
 140         goto out;
 141     }
 142     g_assert(((size_t)testarea & (pagesize-1)) == 0);
 143
 144     reg_struct.range.start = (uintptr_t)testarea;
 145     reg_struct.range.len = pagesize;
 146     reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING;
 147
 148     if (ioctl(ufd, UFFDIO_REGISTER, &reg_struct)) {
 149         error_report("%s userfault register: %s", __func__, strerror(errno));
 150         goto out;
 151     }
 152
 153     range_struct.start = (uintptr_t)testarea;
 154     range_struct.len = pagesize;
 155     if (ioctl(ufd, UFFDIO_UNREGISTER, &range_struct)) {
 156         error_report("%s userfault unregister: %s", __func__, strerror(errno));
 157         goto out;
 158     }
 159
 160     feature_mask = (__u64)1 << _UFFDIO_WAKE |
 161                    (__u64)1 << _UFFDIO_COPY |
 162                    (__u64)1 << _UFFDIO_ZEROPAGE;
 163     if ((reg_struct.ioctls & feature_mask) != feature_mask) {
 164         error_report("Missing userfault map features: %" PRIx64,
 165                      (uint64_t)(~reg_struct.ioctls & feature_mask));
 166         goto out;
 167     }
 168
 169     /* Success! */
 170     ret = true;
 171 out:
 172     if (testarea) {
 173         munmap(testarea, pagesize);
 174     }
 175     if (ufd != -1) {
 176         close(ufd);
 177     }
 178     return ret;
 179 }
 180
 181 /**
 182  * postcopy_ram_discard_range: Discard a range of memory.
 183  * We can assume that if we've been called postcopy_ram_hosttest returned true.
 184  *
 185  * @mis: Current incoming migration state.
 186  * @start, @length: range of memory to discard.
 187  *
 188  * returns: 0 on success.
 189  */
 190 int postcopy_ram_discard_range(MigrationIncomingState *mis, uint8_t *start,
 191                                size_t length)
 192 {
 193     trace_postcopy_ram_discard_range(start, length);
 194     if (madvise(start, length, MADV_DONTNEED)) {
 195         error_report("%s MADV_DONTNEED: %s", __func__, strerror(errno));
 196         return -1;
 197     }
 198
 199     return 0;
 200 }
 201
 202 /*
 203  * Setup an area of RAM so that it *can* be used for postcopy later; this
 204  * must be done right at the start prior to pre-copy.
 205  * opaque should be the MIS.
 206  */
 207 static int init_range(const char *block_name, void *host_addr,
 208                       ram_addr_t offset, ram_addr_t length, void *opaque)
 209 {
 210     MigrationIncomingState *mis = opaque;
 211
 212     trace_postcopy_init_range(block_name, host_addr, offset, length);
 213
 214     /*
 215      * We need the whole of RAM to be truly empty for postcopy, so things
 216      * like ROMs and any data tables built during init must be zero'd
 217      * - we're going to get the copy from the source anyway.
 218      * (Precopy will just overwrite this data, so doesn't need the discard)
 219      */
 220     if (postcopy_ram_discard_range(mis, host_addr, length)) {
 221         return -1;
 222     }
 223
 224     return 0;
 225 }
 226
 227 /*
 228  * At the end of migration, undo the effects of init_range
 229  * opaque should be the MIS.
 230  */
 231 static int cleanup_range(const char *block_name, void *host_addr,
 232                         ram_addr_t offset, ram_addr_t length, void *opaque)
 233 {
 234     MigrationIncomingState *mis = opaque;
 235     struct uffdio_range range_struct;
 236     trace_postcopy_cleanup_range(block_name, host_addr, offset, length);
 237
 238     /*
 239      * We turned off hugepage for the precopy stage with postcopy enabled
 240      * we can turn it back on now.
 241      */
 242     qemu_madvise(host_addr, length, QEMU_MADV_HUGEPAGE);
 243
 244     /*
 245      * We can also turn off userfault now since we should have all the
 246      * pages.   It can be useful to leave it on to debug postcopy
 247      * if you're not sure it's always getting every page.
 248      */
 249     range_struct.start = (uintptr_t)host_addr;
 250     range_struct.len = length;
 251
 252     if (ioctl(mis->userfault_fd, UFFDIO_UNREGISTER, &range_struct)) {
 253         error_report("%s: userfault unregister %s", __func__, strerror(errno));
 254
 255         return -1;
 256     }
 257
 258     return 0;
 259 }
 260
 261 /*
 262  * Initialise postcopy-ram, setting the RAM to a state where we can go into
 263  * postcopy later; must be called prior to any precopy.
 264  * called from arch_init's similarly named ram_postcopy_incoming_init
 265  */
 266 int postcopy_ram_incoming_init(MigrationIncomingState *mis, size_t ram_pages)
 267 {
 268     if (qemu_ram_foreach_block(init_range, mis)) {
 269         return -1;
 270     }
 271
 272     return 0;
 273 }
 274
 275 /*
 276  * At the end of a migration where postcopy_ram_incoming_init was called.
 277  */
 278 int postcopy_ram_incoming_cleanup(MigrationIncomingState *mis)
 279 {
 280     trace_postcopy_ram_incoming_cleanup_entry();
 281
 282     if (mis->have_fault_thread) {
 283         uint64_t tmp64;
 284
 285         if (qemu_ram_foreach_block(cleanup_range, mis)) {
 286             return -1;
 287         }
 288         /*
 289          * Tell the fault_thread to exit, it's an eventfd that should
 290          * currently be at 0, we're going to increment it to 1
 291          */
 292         tmp64 = 1;
 293         if (write(mis->userfault_quit_fd, &tmp64, 8) == 8) {
 294             trace_postcopy_ram_incoming_cleanup_join();
 295             qemu_thread_join(&mis->fault_thread);
 296         } else {
 297             /* Not much we can do here, but may as well report it */
 298             error_report("%s: incrementing userfault_quit_fd: %s", __func__,
 299                          strerror(errno));
 300         }
 301         trace_postcopy_ram_incoming_cleanup_closeuf();
 302         close(mis->userfault_fd);
 303         close(mis->userfault_quit_fd);
 304         mis->have_fault_thread = false;
 305     }
 306
 307     qemu_balloon_inhibit(false);
 308
 309     if (enable_mlock) {
 310         if (os_mlock() < 0) {
 311             error_report("mlock: %s", strerror(errno));
 312             /*
 313              * It doesn't feel right to fail at this point, we have a valid
 314              * VM state.
 315              */
 316         }
 317     }
 318
 319     postcopy_state_set(POSTCOPY_INCOMING_END);
 320     migrate_send_rp_shut(mis, qemu_file_get_error(mis->from_src_file) != 0);
 321
 322     if (mis->postcopy_tmp_page) {
 323         munmap(mis->postcopy_tmp_page, getpagesize());
 324         mis->postcopy_tmp_page = NULL;
 325     }
 326     trace_postcopy_ram_incoming_cleanup_exit();
 327     return 0;
 328 }
 329
 330 /*
 331  * Disable huge pages on an area
 332  */
 333 static int nhp_range(const char *block_name, void *host_addr,
 334                     ram_addr_t offset, ram_addr_t length, void *opaque)
 335 {
 336     trace_postcopy_nhp_range(block_name, host_addr, offset, length);
 337
 338     /*
 339      * Before we do discards we need to ensure those discards really
 340      * do delete areas of the page, even if THP thinks a hugepage would
 341      * be a good idea, so force hugepages off.
 342      */
 343     qemu_madvise(host_addr, length, QEMU_MADV_NOHUGEPAGE);
 344
 345     return 0;
 346 }
 347
 348 /*
 349  * Userfault requires us to mark RAM as NOHUGEPAGE prior to discard
 350  * however leaving it until after precopy means that most of the precopy
 351  * data is still THPd
 352  */
 353 int postcopy_ram_prepare_discard(MigrationIncomingState *mis)
 354 {
 355     if (qemu_ram_foreach_block(nhp_range, mis)) {
 356         return -1;
 357     }
 358
 359     postcopy_state_set(POSTCOPY_INCOMING_DISCARD);
 360
 361     return 0;
 362 }
 363
 364 /*
 365  * Mark the given area of RAM as requiring notification to unwritten areas
 366  * Used as a  callback on qemu_ram_foreach_block.
 367  *   host_addr: Base of area to mark
 368  *   offset: Offset in the whole ram arena
 369  *   length: Length of the section
 370  *   opaque: MigrationIncomingState pointer
 371  * Returns 0 on success
 372  */
 373 static int ram_block_enable_notify(const char *block_name, void *host_addr,
 374                                    ram_addr_t offset, ram_addr_t length,
 375                                    void *opaque)
 376 {
 377     MigrationIncomingState *mis = opaque;
 378     struct uffdio_register reg_struct;
 379
 380     reg_struct.range.start = (uintptr_t)host_addr;
 381     reg_struct.range.len = length;
 382     reg_struct.mode = UFFDIO_REGISTER_MODE_MISSING;
 383
 384     /* Now tell our userfault_fd that it's responsible for this area */
 385     if (ioctl(mis->userfault_fd, UFFDIO_REGISTER, &reg_struct)) {
 386         error_report("%s userfault register: %s", __func__, strerror(errno));
 387         return -1;
 388     }
 389
 390     return 0;
 391 }
 392
 393 /*
 394  * Handle faults detected by the USERFAULT markings
 395  */
 396 static void *postcopy_ram_fault_thread(void *opaque)
 397 {
 398     MigrationIncomingState *mis = opaque;
 399     struct uffd_msg msg;
 400     int ret;
 401     size_t hostpagesize = getpagesize();
 402     RAMBlock *rb = NULL;
 403     RAMBlock *last_rb = NULL; /* last RAMBlock we sent part of */
 404
 405     trace_postcopy_ram_fault_thread_entry();
 406     qemu_sem_post(&mis->fault_thread_sem);
 407
 408     while (true) {
 409         ram_addr_t rb_offset;
 410         struct pollfd pfd[2];
 411
 412         /*
 413          * We're mainly waiting for the kernel to give us a faulting HVA,
 414          * however we can be told to quit via userfault_quit_fd which is
 415          * an eventfd
 416          */
 417         pfd[0].fd = mis->userfault_fd;
 418         pfd[0].events = POLLIN;
 419         pfd[0].revents = 0;
 420         pfd[1].fd = mis->userfault_quit_fd;
 421         pfd[1].events = POLLIN; /* Waiting for eventfd to go positive */
 422         pfd[1].revents = 0;
 423
 424         if (poll(pfd, 2, -1 /* Wait forever */) == -1) {
 425             error_report("%s: userfault poll: %s", __func__, strerror(errno));
 426             break;
 427         }
 428
 429         if (pfd[1].revents) {
 430             trace_postcopy_ram_fault_thread_quit();
 431             break;
 432         }
 433
 434         ret = read(mis->userfault_fd, &msg, sizeof(msg));
 435         if (ret != sizeof(msg)) {
 436             if (errno == EAGAIN) {
 437                 /*
 438                  * if a wake up happens on the other thread just after
 439                  * the poll, there is nothing to read.
 440                  */
 441                 continue;
 442             }
 443             if (ret < 0) {
 444                 error_report("%s: Failed to read full userfault message: %s",
 445                              __func__, strerror(errno));
 446                 break;
 447             } else {
 448                 error_report("%s: Read %d bytes from userfaultfd expected %zd",
 449                              __func__, ret, sizeof(msg));
 450                 break; /* Lost alignment, don't know what we'd read next */
 451             }
 452         }
 453         if (msg.event != UFFD_EVENT_PAGEFAULT) {
 454             error_report("%s: Read unexpected event %ud from userfaultfd",
 455                          __func__, msg.event);
 456             continue; /* It's not a page fault, shouldn't happen */
 457         }
 458
 459         rb = qemu_ram_block_from_host(
 460                  (void *)(uintptr_t)msg.arg.pagefault.address,
 461                  true, &rb_offset);
 462         if (!rb) {
 463             error_report("postcopy_ram_fault_thread: Fault outside guest: %"
 464                          PRIx64, (uint64_t)msg.arg.pagefault.address);
 465             break;
 466         }
 467
 468         rb_offset &= ~(hostpagesize - 1);
 469         trace_postcopy_ram_fault_thread_request(msg.arg.pagefault.address,
 470                                                 qemu_ram_get_idstr(rb),
 471                                                 rb_offset);
 472
 473         /*
 474          * Send the request to the source - we want to request one
 475          * of our host page sizes (which is >= TPS)
 476          */
 477         if (rb != last_rb) {
 478             last_rb = rb;
 479             migrate_send_rp_req_pages(mis, qemu_ram_get_idstr(rb),
 480                                      rb_offset, hostpagesize);
 481         } else {
 482             /* Save some space */
 483             migrate_send_rp_req_pages(mis, NULL,
 484                                      rb_offset, hostpagesize);
 485         }
 486     }
 487     trace_postcopy_ram_fault_thread_exit();
 488     return NULL;
 489 }
 490
 491 int postcopy_ram_enable_notify(MigrationIncomingState *mis)
 492 {
 493     /* Open the fd for the kernel to give us userfaults */
 494     mis->userfault_fd = syscall(__NR_userfaultfd, O_CLOEXEC | O_NONBLOCK);
 495     if (mis->userfault_fd == -1) {
 496         error_report("%s: Failed to open userfault fd: %s", __func__,
 497                      strerror(errno));
 498         return -1;
 499     }
 500
 501     /*
 502      * Although the host check already tested the API, we need to
 503      * do the check again as an ABI handshake on the new fd.
 504      */
 505     if (!ufd_version_check(mis->userfault_fd)) {
 506         return -1;
 507     }
 508
 509     /* Now an eventfd we use to tell the fault-thread to quit */
 510     mis->userfault_quit_fd = eventfd(0, EFD_CLOEXEC);
 511     if (mis->userfault_quit_fd == -1) {
 512         error_report("%s: Opening userfault_quit_fd: %s", __func__,
 513                      strerror(errno));
 514         close(mis->userfault_fd);
 515         return -1;
 516     }
 517
 518     qemu_sem_init(&mis->fault_thread_sem, 0);
 519     qemu_thread_create(&mis->fault_thread, "postcopy/fault",
 520                        postcopy_ram_fault_thread, mis, QEMU_THREAD_JOINABLE);
 521     qemu_sem_wait(&mis->fault_thread_sem);
 522     qemu_sem_destroy(&mis->fault_thread_sem);
 523     mis->have_fault_thread = true;
 524
 525     /* Mark so that we get notified of accesses to unwritten areas */
 526     if (qemu_ram_foreach_block(ram_block_enable_notify, mis)) {
 527         return -1;
 528     }
 529
 530     /*
 531      * Ballooning can mark pages as absent while we're postcopying
 532      * that would cause false userfaults.
 533      */
 534     qemu_balloon_inhibit(true);
 535
 536     trace_postcopy_ram_enable_notify();
 537
 538     return 0;
 539 }
 540
 541 /*
 542  * Place a host page (from) at (host) atomically
 543  * returns 0 on success
 544  */
 545 int postcopy_place_page(MigrationIncomingState *mis, void *host, void *from)
 546 {
 547     struct uffdio_copy copy_struct;
 548
 549     copy_struct.dst = (uint64_t)(uintptr_t)host;
 550     copy_struct.src = (uint64_t)(uintptr_t)from;
 551     copy_struct.len = getpagesize();
 552     copy_struct.mode = 0;
 553
 554     /* copy also acks to the kernel waking the stalled thread up
 555      * TODO: We can inhibit that ack and only do it if it was requested
 556      * which would be slightly cheaper, but we'd have to be careful
 557      * of the order of updating our page state.
 558      */
 559     if (ioctl(mis->userfault_fd, UFFDIO_COPY, &copy_struct)) {
 560         int e = errno;
 561         error_report("%s: %s copy host: %p from: %p",
 562                      __func__, strerror(e), host, from);
 563
 564         return -e;
 565     }
 566
 567     trace_postcopy_place_page(host);
 568     return 0;
 569 }
 570
 571 /*
 572  * Place a zero page at (host) atomically
 573  * returns 0 on success
 574  */
 575 int postcopy_place_page_zero(MigrationIncomingState *mis, void *host)
 576 {
 577     struct uffdio_zeropage zero_struct;
 578
 579     zero_struct.range.start = (uint64_t)(uintptr_t)host;
 580     zero_struct.range.len = getpagesize();
 581     zero_struct.mode = 0;
 582
 583     if (ioctl(mis->userfault_fd, UFFDIO_ZEROPAGE, &zero_struct)) {
 584         int e = errno;
 585         error_report("%s: %s zero host: %p",
 586                      __func__, strerror(e), host);
 587
 588         return -e;
 589     }
 590
 591     trace_postcopy_place_page_zero(host);
 592     return 0;
 593 }
 594
 595 /*
 596  * Returns a target page of memory that can be mapped at a later point in time
 597  * using postcopy_place_page
 598  * The same address is used repeatedly, postcopy_place_page just takes the
 599  * backing page away.
 600  * Returns: Pointer to allocated page
 601  *
 602  */
 603 void *postcopy_get_tmp_page(MigrationIncomingState *mis)
 604 {
 605     if (!mis->postcopy_tmp_page) {
 606         mis->postcopy_tmp_page = mmap(NULL, getpagesize(),
 607                              PROT_READ | PROT_WRITE, MAP_PRIVATE |
 608                              MAP_ANONYMOUS, -1, 0);
 609         if (!mis->postcopy_tmp_page) {
 610             error_report("%s: %s", __func__, strerror(errno));
 611             return NULL;
 612         }
 613     }
 614
 615     return mis->postcopy_tmp_page;
 616 }
 617
 618 #else
 619 /* No target OS support, stubs just fail */
 620 bool postcopy_ram_supported_by_host(void)
 621 {
 622     error_report("%s: No OS support", __func__);
 623     return false;
 624 }
 625
 626 int postcopy_ram_incoming_init(MigrationIncomingState *mis, size_t ram_pages)
 627 {
 628     error_report("postcopy_ram_incoming_init: No OS support");
 629     return -1;
 630 }
 631
 632 int postcopy_ram_incoming_cleanup(MigrationIncomingState *mis)
 633 {
 634     assert(0);
 635     return -1;
 636 }
 637
 638 int postcopy_ram_discard_range(MigrationIncomingState *mis, uint8_t *start,
 639                                size_t length)
 640 {
 641     assert(0);
 642     return -1;
 643 }
 644
 645 int postcopy_ram_prepare_discard(MigrationIncomingState *mis)
 646 {
 647     assert(0);
 648     return -1;
 649 }
 650
 651 int postcopy_ram_enable_notify(MigrationIncomingState *mis)
 652 {
 653     assert(0);
 654     return -1;
 655 }
 656
 657 int postcopy_place_page(MigrationIncomingState *mis, void *host, void *from)
 658 {
 659     assert(0);
 660     return -1;
 661 }
 662
 663 int postcopy_place_page_zero(MigrationIncomingState *mis, void *host)
 664 {
 665     assert(0);
 666     return -1;
 667 }
 668
 669 void *postcopy_get_tmp_page(MigrationIncomingState *mis)
 670 {
 671     assert(0);
 672     return NULL;
 673 }
 674
 675 #endif
 676
 677 /* ------------------------------------------------------------------------- */
 678
 679 /**
 680  * postcopy_discard_send_init: Called at the start of each RAMBlock before
 681  *   asking to discard individual ranges.
 682  *
 683  * @ms: The current migration state.
 684  * @offset: the bitmap offset of the named RAMBlock in the migration
 685  *   bitmap.
 686  * @name: RAMBlock that discards will operate on.
 687  *
 688  * returns: a new PDS.
 689  */
 690 PostcopyDiscardState *postcopy_discard_send_init(MigrationState *ms,
 691                                                  unsigned long offset,
 692                                                  const char *name)
 693 {
 694     PostcopyDiscardState *res = g_malloc0(sizeof(PostcopyDiscardState));
 695
 696     if (res) {
 697         res->ramblock_name = name;
 698         res->offset = offset;
 699     }
 700
 701     return res;
 702 }
 703
 704 /**
 705  * postcopy_discard_send_range: Called by the bitmap code for each chunk to
 706  *   discard. May send a discard message, may just leave it queued to
 707  *   be sent later.
 708  *
 709  * @ms: Current migration state.
 710  * @pds: Structure initialised by postcopy_discard_send_init().
 711  * @start,@length: a range of pages in the migration bitmap in the
 712  *   RAM block passed to postcopy_discard_send_init() (length=1 is one page)
 713  */
 714 void postcopy_discard_send_range(MigrationState *ms, PostcopyDiscardState *pds,
 715                                 unsigned long start, unsigned long length)
 716 {
 717     size_t tp_bits = qemu_target_page_bits();
 718     /* Convert to byte offsets within the RAM block */
 719     pds->start_list[pds->cur_entry] = (start - pds->offset) << tp_bits;
 720     pds->length_list[pds->cur_entry] = length << tp_bits;
 721     trace_postcopy_discard_send_range(pds->ramblock_name, start, length);
 722     pds->cur_entry++;
 723     pds->nsentwords++;
 724
 725     if (pds->cur_entry == MAX_DISCARDS_PER_COMMAND) {
 726         /* Full set, ship it! */
 727         qemu_savevm_send_postcopy_ram_discard(ms->to_dst_file,
 728                                               pds->ramblock_name,
 729                                               pds->cur_entry,
 730                                               pds->start_list,
 731                                               pds->length_list);
 732         pds->nsentcmds++;
 733         pds->cur_entry = 0;
 734     }
 735 }
 736
 737 /**
 738  * postcopy_discard_send_finish: Called at the end of each RAMBlock by the
 739  * bitmap code. Sends any outstanding discard messages, frees the PDS
 740  *
 741  * @ms: Current migration state.
 742  * @pds: Structure initialised by postcopy_discard_send_init().
 743  */
 744 void postcopy_discard_send_finish(MigrationState *ms, PostcopyDiscardState *pds)
 745 {
 746     /* Anything unsent? */
 747     if (pds->cur_entry) {
 748         qemu_savevm_send_postcopy_ram_discard(ms->to_dst_file,
 749                                               pds->ramblock_name,
 750                                               pds->cur_entry,
 751                                               pds->start_list,
 752                                               pds->length_list);
 753         pds->nsentcmds++;
 754     }
 755
 756     trace_postcopy_discard_send_finish(pds->ramblock_name, pds->nsentwords,
 757                                        pds->nsentcmds);
 758
 759     g_free(pds);
 760 }