block-migration.c

   1 /*
   2  * QEMU live block migration
   3  *
   4  * Copyright IBM, Corp. 2009
   5  *
   6  * Authors:
   7  *  Liran Schour   <lirans@il.ibm.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  * 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 "qemu-common.h"
  17 #include "block/block_int.h"
  18 #include "hw/hw.h"
  19 #include "qemu/queue.h"
  20 #include "qemu/timer.h"
  21 #include "migration/block.h"
  22 #include "migration/migration.h"
  23 #include "sysemu/blockdev.h"
  24 #include <assert.h>
  25
  26 #define BLOCK_SIZE                       (1 << 20)
  27 #define BDRV_SECTORS_PER_DIRTY_CHUNK     (BLOCK_SIZE >> BDRV_SECTOR_BITS)
  28
  29 #define BLK_MIG_FLAG_DEVICE_BLOCK       0x01
  30 #define BLK_MIG_FLAG_EOS                0x02
  31 #define BLK_MIG_FLAG_PROGRESS           0x04
  32 #define BLK_MIG_FLAG_ZERO_BLOCK         0x08
  33
  34 #define MAX_IS_ALLOCATED_SEARCH 65536
  35
  36 //#define DEBUG_BLK_MIGRATION
  37
  38 #ifdef DEBUG_BLK_MIGRATION
  39 #define DPRINTF(fmt, ...) \
  40     do { printf("blk_migration: " fmt, ## __VA_ARGS__); } while (0)
  41 #else
  42 #define DPRINTF(fmt, ...) \
  43     do { } while (0)
  44 #endif
  45
  46 typedef struct BlkMigDevState {
  47     /* Written during setup phase.  Can be read without a lock.  */
  48     BlockDriverState *bs;
  49     int shared_base;
  50     int64_t total_sectors;
  51     QSIMPLEQ_ENTRY(BlkMigDevState) entry;
  52
  53     /* Only used by migration thread.  Does not need a lock.  */
  54     int bulk_completed;
  55     int64_t cur_sector;
  56     int64_t cur_dirty;
  57
  58     /* Protected by block migration lock.  */
  59     unsigned long *aio_bitmap;
  60     int64_t completed_sectors;
  61     BdrvDirtyBitmap *dirty_bitmap;
  62 } BlkMigDevState;
  63
  64 typedef struct BlkMigBlock {
  65     /* Only used by migration thread.  */
  66     uint8_t *buf;
  67     BlkMigDevState *bmds;
  68     int64_t sector;
  69     int nr_sectors;
  70     struct iovec iov;
  71     QEMUIOVector qiov;
  72     BlockDriverAIOCB *aiocb;
  73
  74     /* Protected by block migration lock.  */
  75     int ret;
  76     QSIMPLEQ_ENTRY(BlkMigBlock) entry;
  77 } BlkMigBlock;
  78
  79 typedef struct BlkMigState {
  80     /* Written during setup phase.  Can be read without a lock.  */
  81     int blk_enable;
  82     int shared_base;
  83     QSIMPLEQ_HEAD(bmds_list, BlkMigDevState) bmds_list;
  84     int64_t total_sector_sum;
  85     bool zero_blocks;
  86
  87     /* Protected by lock.  */
  88     QSIMPLEQ_HEAD(blk_list, BlkMigBlock) blk_list;
  89     int submitted;
  90     int read_done;
  91
  92     /* Only used by migration thread.  Does not need a lock.  */
  93     int transferred;
  94     int prev_progress;
  95     int bulk_completed;
  96
  97     /* Lock must be taken _inside_ the iothread lock.  */
  98     QemuMutex lock;
  99 } BlkMigState;
 100
 101 static BlkMigState block_mig_state;
 102
 103 static void blk_mig_lock(void)
 104 {
 105     qemu_mutex_lock(&block_mig_state.lock);
 106 }
 107
 108 static void blk_mig_unlock(void)
 109 {
 110     qemu_mutex_unlock(&block_mig_state.lock);
 111 }
 112
 113 /* Must run outside of the iothread lock during the bulk phase,
 114  * or the VM will stall.
 115  */
 116
 117 static void blk_send(QEMUFile *f, BlkMigBlock * blk)
 118 {
 119     int len;
 120     uint64_t flags = BLK_MIG_FLAG_DEVICE_BLOCK;
 121
 122     if (block_mig_state.zero_blocks &&
 123         buffer_is_zero(blk->buf, BLOCK_SIZE)) {
 124         flags |= BLK_MIG_FLAG_ZERO_BLOCK;
 125     }
 126
 127     /* sector number and flags */
 128     qemu_put_be64(f, (blk->sector << BDRV_SECTOR_BITS)
 129                      | flags);
 130
 131     /* device name */
 132     len = strlen(blk->bmds->bs->device_name);
 133     qemu_put_byte(f, len);
 134     qemu_put_buffer(f, (uint8_t *)blk->bmds->bs->device_name, len);
 135
 136     /* if a block is zero we need to flush here since the network
 137      * bandwidth is now a lot higher than the storage device bandwidth.
 138      * thus if we queue zero blocks we slow down the migration */
 139     if (flags & BLK_MIG_FLAG_ZERO_BLOCK) {
 140         qemu_fflush(f);
 141         return;
 142     }
 143
 144     qemu_put_buffer(f, blk->buf, BLOCK_SIZE);
 145 }
 146
 147 int blk_mig_active(void)
 148 {
 149     return !QSIMPLEQ_EMPTY(&block_mig_state.bmds_list);
 150 }
 151
 152 uint64_t blk_mig_bytes_transferred(void)
 153 {
 154     BlkMigDevState *bmds;
 155     uint64_t sum = 0;
 156
 157     blk_mig_lock();
 158     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 159         sum += bmds->completed_sectors;
 160     }
 161     blk_mig_unlock();
 162     return sum << BDRV_SECTOR_BITS;
 163 }
 164
 165 uint64_t blk_mig_bytes_remaining(void)
 166 {
 167     return blk_mig_bytes_total() - blk_mig_bytes_transferred();
 168 }
 169
 170 uint64_t blk_mig_bytes_total(void)
 171 {
 172     BlkMigDevState *bmds;
 173     uint64_t sum = 0;
 174
 175     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 176         sum += bmds->total_sectors;
 177     }
 178     return sum << BDRV_SECTOR_BITS;
 179 }
 180
 181
 182 /* Called with migration lock held.  */
 183
 184 static int bmds_aio_inflight(BlkMigDevState *bmds, int64_t sector)
 185 {
 186     int64_t chunk = sector / (int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK;
 187
 188     if ((sector << BDRV_SECTOR_BITS) < bdrv_getlength(bmds->bs)) {
 189         return !!(bmds->aio_bitmap[chunk / (sizeof(unsigned long) * 8)] &
 190             (1UL << (chunk % (sizeof(unsigned long) * 8))));
 191     } else {
 192         return 0;
 193     }
 194 }
 195
 196 /* Called with migration lock held.  */
 197
 198 static void bmds_set_aio_inflight(BlkMigDevState *bmds, int64_t sector_num,
 199                              int nb_sectors, int set)
 200 {
 201     int64_t start, end;
 202     unsigned long val, idx, bit;
 203
 204     start = sector_num / BDRV_SECTORS_PER_DIRTY_CHUNK;
 205     end = (sector_num + nb_sectors - 1) / BDRV_SECTORS_PER_DIRTY_CHUNK;
 206
 207     for (; start <= end; start++) {
 208         idx = start / (sizeof(unsigned long) * 8);
 209         bit = start % (sizeof(unsigned long) * 8);
 210         val = bmds->aio_bitmap[idx];
 211         if (set) {
 212             val |= 1UL << bit;
 213         } else {
 214             val &= ~(1UL << bit);
 215         }
 216         bmds->aio_bitmap[idx] = val;
 217     }
 218 }
 219
 220 static void alloc_aio_bitmap(BlkMigDevState *bmds)
 221 {
 222     BlockDriverState *bs = bmds->bs;
 223     int64_t bitmap_size;
 224
 225     bitmap_size = (bdrv_getlength(bs) >> BDRV_SECTOR_BITS) +
 226             BDRV_SECTORS_PER_DIRTY_CHUNK * 8 - 1;
 227     bitmap_size /= BDRV_SECTORS_PER_DIRTY_CHUNK * 8;
 228
 229     bmds->aio_bitmap = g_malloc0(bitmap_size);
 230 }
 231
 232 /* Never hold migration lock when yielding to the main loop!  */
 233
 234 static void blk_mig_read_cb(void *opaque, int ret)
 235 {
 236     BlkMigBlock *blk = opaque;
 237
 238     blk_mig_lock();
 239     blk->ret = ret;
 240
 241     QSIMPLEQ_INSERT_TAIL(&block_mig_state.blk_list, blk, entry);
 242     bmds_set_aio_inflight(blk->bmds, blk->sector, blk->nr_sectors, 0);
 243
 244     block_mig_state.submitted--;
 245     block_mig_state.read_done++;
 246     assert(block_mig_state.submitted >= 0);
 247     blk_mig_unlock();
 248 }
 249
 250 /* Called with no lock taken.  */
 251
 252 static int mig_save_device_bulk(QEMUFile *f, BlkMigDevState *bmds)
 253 {
 254     int64_t total_sectors = bmds->total_sectors;
 255     int64_t cur_sector = bmds->cur_sector;
 256     BlockDriverState *bs = bmds->bs;
 257     BlkMigBlock *blk;
 258     int nr_sectors;
 259
 260     if (bmds->shared_base) {
 261         qemu_mutex_lock_iothread();
 262         while (cur_sector < total_sectors &&
 263                !bdrv_is_allocated(bs, cur_sector, MAX_IS_ALLOCATED_SEARCH,
 264                                   &nr_sectors)) {
 265             cur_sector += nr_sectors;
 266         }
 267         qemu_mutex_unlock_iothread();
 268     }
 269
 270     if (cur_sector >= total_sectors) {
 271         bmds->cur_sector = bmds->completed_sectors = total_sectors;
 272         return 1;
 273     }
 274
 275     bmds->completed_sectors = cur_sector;
 276
 277     cur_sector &= ~((int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK - 1);
 278
 279     /* we are going to transfer a full block even if it is not allocated */
 280     nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 281
 282     if (total_sectors - cur_sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 283         nr_sectors = total_sectors - cur_sector;
 284     }
 285
 286     blk = g_malloc(sizeof(BlkMigBlock));
 287     blk->buf = g_malloc(BLOCK_SIZE);
 288     blk->bmds = bmds;
 289     blk->sector = cur_sector;
 290     blk->nr_sectors = nr_sectors;
 291
 292     blk->iov.iov_base = blk->buf;
 293     blk->iov.iov_len = nr_sectors * BDRV_SECTOR_SIZE;
 294     qemu_iovec_init_external(&blk->qiov, &blk->iov, 1);
 295
 296     blk_mig_lock();
 297     block_mig_state.submitted++;
 298     blk_mig_unlock();
 299
 300     qemu_mutex_lock_iothread();
 301     blk->aiocb = bdrv_aio_readv(bs, cur_sector, &blk->qiov,
 302                                 nr_sectors, blk_mig_read_cb, blk);
 303
 304     bdrv_reset_dirty(bs, cur_sector, nr_sectors);
 305     qemu_mutex_unlock_iothread();
 306
 307     bmds->cur_sector = cur_sector + nr_sectors;
 308     return (bmds->cur_sector >= total_sectors);
 309 }
 310
 311 /* Called with iothread lock taken.  */
 312
 313 static int set_dirty_tracking(void)
 314 {
 315     BlkMigDevState *bmds;
 316     int ret;
 317
 318     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 319         bmds->dirty_bitmap = bdrv_create_dirty_bitmap(bmds->bs, BLOCK_SIZE,
 320                                                       NULL);
 321         if (!bmds->dirty_bitmap) {
 322             ret = -errno;
 323             goto fail;
 324         }
 325     }
 326     return 0;
 327
 328 fail:
 329     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 330         if (bmds->dirty_bitmap) {
 331             bdrv_release_dirty_bitmap(bmds->bs, bmds->dirty_bitmap);
 332         }
 333     }
 334     return ret;
 335 }
 336
 337 static void unset_dirty_tracking(void)
 338 {
 339     BlkMigDevState *bmds;
 340
 341     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 342         bdrv_release_dirty_bitmap(bmds->bs, bmds->dirty_bitmap);
 343     }
 344 }
 345
 346 static void init_blk_migration_it(void *opaque, BlockDriverState *bs)
 347 {
 348     BlkMigDevState *bmds;
 349     int64_t sectors;
 350
 351     if (!bdrv_is_read_only(bs)) {
 352         sectors = bdrv_getlength(bs) >> BDRV_SECTOR_BITS;
 353         if (sectors <= 0) {
 354             return;
 355         }
 356
 357         bmds = g_malloc0(sizeof(BlkMigDevState));
 358         bmds->bs = bs;
 359         bmds->bulk_completed = 0;
 360         bmds->total_sectors = sectors;
 361         bmds->completed_sectors = 0;
 362         bmds->shared_base = block_mig_state.shared_base;
 363         alloc_aio_bitmap(bmds);
 364         bdrv_set_in_use(bs, 1);
 365         bdrv_ref(bs);
 366
 367         block_mig_state.total_sector_sum += sectors;
 368
 369         if (bmds->shared_base) {
 370             DPRINTF("Start migration for %s with shared base image\n",
 371                     bs->device_name);
 372         } else {
 373             DPRINTF("Start full migration for %s\n", bs->device_name);
 374         }
 375
 376         QSIMPLEQ_INSERT_TAIL(&block_mig_state.bmds_list, bmds, entry);
 377     }
 378 }
 379
 380 static void init_blk_migration(QEMUFile *f)
 381 {
 382     block_mig_state.submitted = 0;
 383     block_mig_state.read_done = 0;
 384     block_mig_state.transferred = 0;
 385     block_mig_state.total_sector_sum = 0;
 386     block_mig_state.prev_progress = -1;
 387     block_mig_state.bulk_completed = 0;
 388     block_mig_state.zero_blocks = migrate_zero_blocks();
 389
 390     bdrv_iterate(init_blk_migration_it, NULL);
 391 }
 392
 393 /* Called with no lock taken.  */
 394
 395 static int blk_mig_save_bulked_block(QEMUFile *f)
 396 {
 397     int64_t completed_sector_sum = 0;
 398     BlkMigDevState *bmds;
 399     int progress;
 400     int ret = 0;
 401
 402     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 403         if (bmds->bulk_completed == 0) {
 404             if (mig_save_device_bulk(f, bmds) == 1) {
 405                 /* completed bulk section for this device */
 406                 bmds->bulk_completed = 1;
 407             }
 408             completed_sector_sum += bmds->completed_sectors;
 409             ret = 1;
 410             break;
 411         } else {
 412             completed_sector_sum += bmds->completed_sectors;
 413         }
 414     }
 415
 416     if (block_mig_state.total_sector_sum != 0) {
 417         progress = completed_sector_sum * 100 /
 418                    block_mig_state.total_sector_sum;
 419     } else {
 420         progress = 100;
 421     }
 422     if (progress != block_mig_state.prev_progress) {
 423         block_mig_state.prev_progress = progress;
 424         qemu_put_be64(f, (progress << BDRV_SECTOR_BITS)
 425                          | BLK_MIG_FLAG_PROGRESS);
 426         DPRINTF("Completed %d %%\r", progress);
 427     }
 428
 429     return ret;
 430 }
 431
 432 static void blk_mig_reset_dirty_cursor(void)
 433 {
 434     BlkMigDevState *bmds;
 435
 436     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 437         bmds->cur_dirty = 0;
 438     }
 439 }
 440
 441 /* Called with iothread lock taken.  */
 442
 443 static int mig_save_device_dirty(QEMUFile *f, BlkMigDevState *bmds,
 444                                  int is_async)
 445 {
 446     BlkMigBlock *blk;
 447     int64_t total_sectors = bmds->total_sectors;
 448     int64_t sector;
 449     int nr_sectors;
 450     int ret = -EIO;
 451
 452     for (sector = bmds->cur_dirty; sector < bmds->total_sectors;) {
 453         blk_mig_lock();
 454         if (bmds_aio_inflight(bmds, sector)) {
 455             blk_mig_unlock();
 456             bdrv_drain_all();
 457         } else {
 458             blk_mig_unlock();
 459         }
 460         if (bdrv_get_dirty(bmds->bs, bmds->dirty_bitmap, sector)) {
 461
 462             if (total_sectors - sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 463                 nr_sectors = total_sectors - sector;
 464             } else {
 465                 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 466             }
 467             blk = g_malloc(sizeof(BlkMigBlock));
 468             blk->buf = g_malloc(BLOCK_SIZE);
 469             blk->bmds = bmds;
 470             blk->sector = sector;
 471             blk->nr_sectors = nr_sectors;
 472
 473             if (is_async) {
 474                 blk->iov.iov_base = blk->buf;
 475                 blk->iov.iov_len = nr_sectors * BDRV_SECTOR_SIZE;
 476                 qemu_iovec_init_external(&blk->qiov, &blk->iov, 1);
 477
 478                 blk->aiocb = bdrv_aio_readv(bmds->bs, sector, &blk->qiov,
 479                                             nr_sectors, blk_mig_read_cb, blk);
 480
 481                 blk_mig_lock();
 482                 block_mig_state.submitted++;
 483                 bmds_set_aio_inflight(bmds, sector, nr_sectors, 1);
 484                 blk_mig_unlock();
 485             } else {
 486                 ret = bdrv_read(bmds->bs, sector, blk->buf, nr_sectors);
 487                 if (ret < 0) {
 488                     goto error;
 489                 }
 490                 blk_send(f, blk);
 491
 492                 g_free(blk->buf);
 493                 g_free(blk);
 494             }
 495
 496             bdrv_reset_dirty(bmds->bs, sector, nr_sectors);
 497             break;
 498         }
 499         sector += BDRV_SECTORS_PER_DIRTY_CHUNK;
 500         bmds->cur_dirty = sector;
 501     }
 502
 503     return (bmds->cur_dirty >= bmds->total_sectors);
 504
 505 error:
 506     DPRINTF("Error reading sector %" PRId64 "\n", sector);
 507     g_free(blk->buf);
 508     g_free(blk);
 509     return ret;
 510 }
 511
 512 /* Called with iothread lock taken.
 513  *
 514  * return value:
 515  * 0: too much data for max_downtime
 516  * 1: few enough data for max_downtime
 517 */
 518 static int blk_mig_save_dirty_block(QEMUFile *f, int is_async)
 519 {
 520     BlkMigDevState *bmds;
 521     int ret = 1;
 522
 523     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 524         ret = mig_save_device_dirty(f, bmds, is_async);
 525         if (ret <= 0) {
 526             break;
 527         }
 528     }
 529
 530     return ret;
 531 }
 532
 533 /* Called with no locks taken.  */
 534
 535 static int flush_blks(QEMUFile *f)
 536 {
 537     BlkMigBlock *blk;
 538     int ret = 0;
 539
 540     DPRINTF("%s Enter submitted %d read_done %d transferred %d\n",
 541             __FUNCTION__, block_mig_state.submitted, block_mig_state.read_done,
 542             block_mig_state.transferred);
 543
 544     blk_mig_lock();
 545     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 546         if (qemu_file_rate_limit(f)) {
 547             break;
 548         }
 549         if (blk->ret < 0) {
 550             ret = blk->ret;
 551             break;
 552         }
 553
 554         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 555         blk_mig_unlock();
 556         blk_send(f, blk);
 557         blk_mig_lock();
 558
 559         g_free(blk->buf);
 560         g_free(blk);
 561
 562         block_mig_state.read_done--;
 563         block_mig_state.transferred++;
 564         assert(block_mig_state.read_done >= 0);
 565     }
 566     blk_mig_unlock();
 567
 568     DPRINTF("%s Exit submitted %d read_done %d transferred %d\n", __FUNCTION__,
 569             block_mig_state.submitted, block_mig_state.read_done,
 570             block_mig_state.transferred);
 571     return ret;
 572 }
 573
 574 /* Called with iothread lock taken.  */
 575
 576 static int64_t get_remaining_dirty(void)
 577 {
 578     BlkMigDevState *bmds;
 579     int64_t dirty = 0;
 580
 581     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 582         dirty += bdrv_get_dirty_count(bmds->bs, bmds->dirty_bitmap);
 583     }
 584
 585     return dirty << BDRV_SECTOR_BITS;
 586 }
 587
 588 /* Called with iothread lock taken.  */
 589
 590 static void blk_mig_cleanup(void)
 591 {
 592     BlkMigDevState *bmds;
 593     BlkMigBlock *blk;
 594
 595     bdrv_drain_all();
 596
 597     unset_dirty_tracking();
 598
 599     blk_mig_lock();
 600     while ((bmds = QSIMPLEQ_FIRST(&block_mig_state.bmds_list)) != NULL) {
 601         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.bmds_list, entry);
 602         bdrv_set_in_use(bmds->bs, 0);
 603         bdrv_unref(bmds->bs);
 604         g_free(bmds->aio_bitmap);
 605         g_free(bmds);
 606     }
 607
 608     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 609         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 610         g_free(blk->buf);
 611         g_free(blk);
 612     }
 613     blk_mig_unlock();
 614 }
 615
 616 static void block_migration_cancel(void *opaque)
 617 {
 618     blk_mig_cleanup();
 619 }
 620
 621 static int block_save_setup(QEMUFile *f, void *opaque)
 622 {
 623     int ret;
 624
 625     DPRINTF("Enter save live setup submitted %d transferred %d\n",
 626             block_mig_state.submitted, block_mig_state.transferred);
 627
 628     qemu_mutex_lock_iothread();
 629
 630     /* start track dirty blocks */
 631     ret = set_dirty_tracking();
 632
 633     if (ret) {
 634         qemu_mutex_unlock_iothread();
 635         return ret;
 636     }
 637
 638     init_blk_migration(f);
 639
 640     qemu_mutex_unlock_iothread();
 641
 642     ret = flush_blks(f);
 643     blk_mig_reset_dirty_cursor();
 644     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 645
 646     return ret;
 647 }
 648
 649 static int block_save_iterate(QEMUFile *f, void *opaque)
 650 {
 651     int ret;
 652     int64_t last_ftell = qemu_ftell(f);
 653
 654     DPRINTF("Enter save live iterate submitted %d transferred %d\n",
 655             block_mig_state.submitted, block_mig_state.transferred);
 656
 657     ret = flush_blks(f);
 658     if (ret) {
 659         return ret;
 660     }
 661
 662     blk_mig_reset_dirty_cursor();
 663
 664     /* control the rate of transfer */
 665     blk_mig_lock();
 666     while ((block_mig_state.submitted +
 667             block_mig_state.read_done) * BLOCK_SIZE <
 668            qemu_file_get_rate_limit(f)) {
 669         blk_mig_unlock();
 670         if (block_mig_state.bulk_completed == 0) {
 671             /* first finish the bulk phase */
 672             if (blk_mig_save_bulked_block(f) == 0) {
 673                 /* finished saving bulk on all devices */
 674                 block_mig_state.bulk_completed = 1;
 675             }
 676             ret = 0;
 677         } else {
 678             /* Always called with iothread lock taken for
 679              * simplicity, block_save_complete also calls it.
 680              */
 681             qemu_mutex_lock_iothread();
 682             ret = blk_mig_save_dirty_block(f, 1);
 683             qemu_mutex_unlock_iothread();
 684         }
 685         if (ret < 0) {
 686             return ret;
 687         }
 688         blk_mig_lock();
 689         if (ret != 0) {
 690             /* no more dirty blocks */
 691             break;
 692         }
 693     }
 694     blk_mig_unlock();
 695
 696     ret = flush_blks(f);
 697     if (ret) {
 698         return ret;
 699     }
 700
 701     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 702     return qemu_ftell(f) - last_ftell;
 703 }
 704
 705 /* Called with iothread lock taken.  */
 706
 707 static int block_save_complete(QEMUFile *f, void *opaque)
 708 {
 709     int ret;
 710
 711     DPRINTF("Enter save live complete submitted %d transferred %d\n",
 712             block_mig_state.submitted, block_mig_state.transferred);
 713
 714     ret = flush_blks(f);
 715     if (ret) {
 716         return ret;
 717     }
 718
 719     blk_mig_reset_dirty_cursor();
 720
 721     /* we know for sure that save bulk is completed and
 722        all async read completed */
 723     blk_mig_lock();
 724     assert(block_mig_state.submitted == 0);
 725     blk_mig_unlock();
 726
 727     do {
 728         ret = blk_mig_save_dirty_block(f, 0);
 729         if (ret < 0) {
 730             return ret;
 731         }
 732     } while (ret == 0);
 733
 734     /* report completion */
 735     qemu_put_be64(f, (100 << BDRV_SECTOR_BITS) | BLK_MIG_FLAG_PROGRESS);
 736
 737     DPRINTF("Block migration completed\n");
 738
 739     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 740
 741     blk_mig_cleanup();
 742     return 0;
 743 }
 744
 745 static uint64_t block_save_pending(QEMUFile *f, void *opaque, uint64_t max_size)
 746 {
 747     /* Estimate pending number of bytes to send */
 748     uint64_t pending;
 749
 750     qemu_mutex_lock_iothread();
 751     blk_mig_lock();
 752     pending = get_remaining_dirty() +
 753                        block_mig_state.submitted * BLOCK_SIZE +
 754                        block_mig_state.read_done * BLOCK_SIZE;
 755
 756     /* Report at least one block pending during bulk phase */
 757     if (pending == 0 && !block_mig_state.bulk_completed) {
 758         pending = BLOCK_SIZE;
 759     }
 760     blk_mig_unlock();
 761     qemu_mutex_unlock_iothread();
 762
 763     DPRINTF("Enter save live pending  %" PRIu64 "\n", pending);
 764     return pending;
 765 }
 766
 767 static int block_load(QEMUFile *f, void *opaque, int version_id)
 768 {
 769     static int banner_printed;
 770     int len, flags;
 771     char device_name[256];
 772     int64_t addr;
 773     BlockDriverState *bs, *bs_prev = NULL;
 774     uint8_t *buf;
 775     int64_t total_sectors = 0;
 776     int nr_sectors;
 777     int ret;
 778
 779     do {
 780         addr = qemu_get_be64(f);
 781
 782         flags = addr & ~BDRV_SECTOR_MASK;
 783         addr >>= BDRV_SECTOR_BITS;
 784
 785         if (flags & BLK_MIG_FLAG_DEVICE_BLOCK) {
 786             /* get device name */
 787             len = qemu_get_byte(f);
 788             qemu_get_buffer(f, (uint8_t *)device_name, len);
 789             device_name[len] = '\0';
 790
 791             bs = bdrv_find(device_name);
 792             if (!bs) {
 793                 fprintf(stderr, "Error unknown block device %s\n",
 794                         device_name);
 795                 return -EINVAL;
 796             }
 797
 798             if (bs != bs_prev) {
 799                 bs_prev = bs;
 800                 total_sectors = bdrv_getlength(bs) >> BDRV_SECTOR_BITS;
 801                 if (total_sectors <= 0) {
 802                     error_report("Error getting length of block device %s",
 803                                  device_name);
 804                     return -EINVAL;
 805                 }
 806             }
 807
 808             if (total_sectors - addr < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 809                 nr_sectors = total_sectors - addr;
 810             } else {
 811                 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 812             }
 813
 814             if (flags & BLK_MIG_FLAG_ZERO_BLOCK) {
 815                 ret = bdrv_write_zeroes(bs, addr, nr_sectors,
 816                                         BDRV_REQ_MAY_UNMAP);
 817             } else {
 818                 buf = g_malloc(BLOCK_SIZE);
 819                 qemu_get_buffer(f, buf, BLOCK_SIZE);
 820                 ret = bdrv_write(bs, addr, buf, nr_sectors);
 821                 g_free(buf);
 822             }
 823
 824             if (ret < 0) {
 825                 return ret;
 826             }
 827         } else if (flags & BLK_MIG_FLAG_PROGRESS) {
 828             if (!banner_printed) {
 829                 printf("Receiving block device images\n");
 830                 banner_printed = 1;
 831             }
 832             printf("Completed %d %%%c", (int)addr,
 833                    (addr == 100) ? '\n' : '\r');
 834             fflush(stdout);
 835         } else if (!(flags & BLK_MIG_FLAG_EOS)) {
 836             fprintf(stderr, "Unknown block migration flags: %#x\n", flags);
 837             return -EINVAL;
 838         }
 839         ret = qemu_file_get_error(f);
 840         if (ret != 0) {
 841             return ret;
 842         }
 843     } while (!(flags & BLK_MIG_FLAG_EOS));
 844
 845     return 0;
 846 }
 847
 848 static void block_set_params(const MigrationParams *params, void *opaque)
 849 {
 850     block_mig_state.blk_enable = params->blk;
 851     block_mig_state.shared_base = params->shared;
 852
 853     /* shared base means that blk_enable = 1 */
 854     block_mig_state.blk_enable |= params->shared;
 855 }
 856
 857 static bool block_is_active(void *opaque)
 858 {
 859     return block_mig_state.blk_enable == 1;
 860 }
 861
 862 SaveVMHandlers savevm_block_handlers = {
 863     .set_params = block_set_params,
 864     .save_live_setup = block_save_setup,
 865     .save_live_iterate = block_save_iterate,
 866     .save_live_complete = block_save_complete,
 867     .save_live_pending = block_save_pending,
 868     .load_state = block_load,
 869     .cancel = block_migration_cancel,
 870     .is_active = block_is_active,
 871 };
 872
 873 void blk_mig_init(void)
 874 {
 875     QSIMPLEQ_INIT(&block_mig_state.bmds_list);
 876     QSIMPLEQ_INIT(&block_mig_state.blk_list);
 877     qemu_mutex_init(&block_mig_state.lock);
 878
 879     register_savevm_live(NULL, "block", 0, 1, &savevm_block_handlers,
 880                          &block_mig_state);
 881 }