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     Error *blocker;
  63 } BlkMigDevState;
  64
  65 typedef struct BlkMigBlock {
  66     /* Only used by migration thread.  */
  67     uint8_t *buf;
  68     BlkMigDevState *bmds;
  69     int64_t sector;
  70     int nr_sectors;
  71     struct iovec iov;
  72     QEMUIOVector qiov;
  73     BlockDriverAIOCB *aiocb;
  74
  75     /* Protected by block migration lock.  */
  76     int ret;
  77     QSIMPLEQ_ENTRY(BlkMigBlock) entry;
  78 } BlkMigBlock;
  79
  80 typedef struct BlkMigState {
  81     /* Written during setup phase.  Can be read without a lock.  */
  82     int blk_enable;
  83     int shared_base;
  84     QSIMPLEQ_HEAD(bmds_list, BlkMigDevState) bmds_list;
  85     int64_t total_sector_sum;
  86     bool zero_blocks;
  87
  88     /* Protected by lock.  */
  89     QSIMPLEQ_HEAD(blk_list, BlkMigBlock) blk_list;
  90     int submitted;
  91     int read_done;
  92
  93     /* Only used by migration thread.  Does not need a lock.  */
  94     int transferred;
  95     int prev_progress;
  96     int bulk_completed;
  97
  98     /* Lock must be taken _inside_ the iothread lock.  */
  99     QemuMutex lock;
 100 } BlkMigState;
 101
 102 static BlkMigState block_mig_state;
 103
 104 static void blk_mig_lock(void)
 105 {
 106     qemu_mutex_lock(&block_mig_state.lock);
 107 }
 108
 109 static void blk_mig_unlock(void)
 110 {
 111     qemu_mutex_unlock(&block_mig_state.lock);
 112 }
 113
 114 /* Must run outside of the iothread lock during the bulk phase,
 115  * or the VM will stall.
 116  */
 117
 118 static void blk_send(QEMUFile *f, BlkMigBlock * blk)
 119 {
 120     int len;
 121     uint64_t flags = BLK_MIG_FLAG_DEVICE_BLOCK;
 122
 123     if (block_mig_state.zero_blocks &&
 124         buffer_is_zero(blk->buf, BLOCK_SIZE)) {
 125         flags |= BLK_MIG_FLAG_ZERO_BLOCK;
 126     }
 127
 128     /* sector number and flags */
 129     qemu_put_be64(f, (blk->sector << BDRV_SECTOR_BITS)
 130                      | flags);
 131
 132     /* device name */
 133     len = strlen(blk->bmds->bs->device_name);
 134     qemu_put_byte(f, len);
 135     qemu_put_buffer(f, (uint8_t *)blk->bmds->bs->device_name, len);
 136
 137     /* if a block is zero we need to flush here since the network
 138      * bandwidth is now a lot higher than the storage device bandwidth.
 139      * thus if we queue zero blocks we slow down the migration */
 140     if (flags & BLK_MIG_FLAG_ZERO_BLOCK) {
 141         qemu_fflush(f);
 142         return;
 143     }
 144
 145     qemu_put_buffer(f, blk->buf, BLOCK_SIZE);
 146 }
 147
 148 int blk_mig_active(void)
 149 {
 150     return !QSIMPLEQ_EMPTY(&block_mig_state.bmds_list);
 151 }
 152
 153 uint64_t blk_mig_bytes_transferred(void)
 154 {
 155     BlkMigDevState *bmds;
 156     uint64_t sum = 0;
 157
 158     blk_mig_lock();
 159     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 160         sum += bmds->completed_sectors;
 161     }
 162     blk_mig_unlock();
 163     return sum << BDRV_SECTOR_BITS;
 164 }
 165
 166 uint64_t blk_mig_bytes_remaining(void)
 167 {
 168     return blk_mig_bytes_total() - blk_mig_bytes_transferred();
 169 }
 170
 171 uint64_t blk_mig_bytes_total(void)
 172 {
 173     BlkMigDevState *bmds;
 174     uint64_t sum = 0;
 175
 176     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 177         sum += bmds->total_sectors;
 178     }
 179     return sum << BDRV_SECTOR_BITS;
 180 }
 181
 182
 183 /* Called with migration lock held.  */
 184
 185 static int bmds_aio_inflight(BlkMigDevState *bmds, int64_t sector)
 186 {
 187     int64_t chunk = sector / (int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK;
 188
 189     if ((sector << BDRV_SECTOR_BITS) < bdrv_getlength(bmds->bs)) {
 190         return !!(bmds->aio_bitmap[chunk / (sizeof(unsigned long) * 8)] &
 191             (1UL << (chunk % (sizeof(unsigned long) * 8))));
 192     } else {
 193         return 0;
 194     }
 195 }
 196
 197 /* Called with migration lock held.  */
 198
 199 static void bmds_set_aio_inflight(BlkMigDevState *bmds, int64_t sector_num,
 200                              int nb_sectors, int set)
 201 {
 202     int64_t start, end;
 203     unsigned long val, idx, bit;
 204
 205     start = sector_num / BDRV_SECTORS_PER_DIRTY_CHUNK;
 206     end = (sector_num + nb_sectors - 1) / BDRV_SECTORS_PER_DIRTY_CHUNK;
 207
 208     for (; start <= end; start++) {
 209         idx = start / (sizeof(unsigned long) * 8);
 210         bit = start % (sizeof(unsigned long) * 8);
 211         val = bmds->aio_bitmap[idx];
 212         if (set) {
 213             val |= 1UL << bit;
 214         } else {
 215             val &= ~(1UL << bit);
 216         }
 217         bmds->aio_bitmap[idx] = val;
 218     }
 219 }
 220
 221 static void alloc_aio_bitmap(BlkMigDevState *bmds)
 222 {
 223     BlockDriverState *bs = bmds->bs;
 224     int64_t bitmap_size;
 225
 226     bitmap_size = (bdrv_getlength(bs) >> BDRV_SECTOR_BITS) +
 227             BDRV_SECTORS_PER_DIRTY_CHUNK * 8 - 1;
 228     bitmap_size /= BDRV_SECTORS_PER_DIRTY_CHUNK * 8;
 229
 230     bmds->aio_bitmap = g_malloc0(bitmap_size);
 231 }
 232
 233 /* Never hold migration lock when yielding to the main loop!  */
 234
 235 static void blk_mig_read_cb(void *opaque, int ret)
 236 {
 237     BlkMigBlock *blk = opaque;
 238
 239     blk_mig_lock();
 240     blk->ret = ret;
 241
 242     QSIMPLEQ_INSERT_TAIL(&block_mig_state.blk_list, blk, entry);
 243     bmds_set_aio_inflight(blk->bmds, blk->sector, blk->nr_sectors, 0);
 244
 245     block_mig_state.submitted--;
 246     block_mig_state.read_done++;
 247     assert(block_mig_state.submitted >= 0);
 248     blk_mig_unlock();
 249 }
 250
 251 /* Called with no lock taken.  */
 252
 253 static int mig_save_device_bulk(QEMUFile *f, BlkMigDevState *bmds)
 254 {
 255     int64_t total_sectors = bmds->total_sectors;
 256     int64_t cur_sector = bmds->cur_sector;
 257     BlockDriverState *bs = bmds->bs;
 258     BlkMigBlock *blk;
 259     int nr_sectors;
 260
 261     if (bmds->shared_base) {
 262         qemu_mutex_lock_iothread();
 263         while (cur_sector < total_sectors &&
 264                !bdrv_is_allocated(bs, cur_sector, MAX_IS_ALLOCATED_SEARCH,
 265                                   &nr_sectors)) {
 266             cur_sector += nr_sectors;
 267         }
 268         qemu_mutex_unlock_iothread();
 269     }
 270
 271     if (cur_sector >= total_sectors) {
 272         bmds->cur_sector = bmds->completed_sectors = total_sectors;
 273         return 1;
 274     }
 275
 276     bmds->completed_sectors = cur_sector;
 277
 278     cur_sector &= ~((int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK - 1);
 279
 280     /* we are going to transfer a full block even if it is not allocated */
 281     nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 282
 283     if (total_sectors - cur_sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 284         nr_sectors = total_sectors - cur_sector;
 285     }
 286
 287     blk = g_malloc(sizeof(BlkMigBlock));
 288     blk->buf = g_malloc(BLOCK_SIZE);
 289     blk->bmds = bmds;
 290     blk->sector = cur_sector;
 291     blk->nr_sectors = nr_sectors;
 292
 293     blk->iov.iov_base = blk->buf;
 294     blk->iov.iov_len = nr_sectors * BDRV_SECTOR_SIZE;
 295     qemu_iovec_init_external(&blk->qiov, &blk->iov, 1);
 296
 297     blk_mig_lock();
 298     block_mig_state.submitted++;
 299     blk_mig_unlock();
 300
 301     qemu_mutex_lock_iothread();
 302     blk->aiocb = bdrv_aio_readv(bs, cur_sector, &blk->qiov,
 303                                 nr_sectors, blk_mig_read_cb, blk);
 304
 305     bdrv_reset_dirty(bs, cur_sector, nr_sectors);
 306     qemu_mutex_unlock_iothread();
 307
 308     bmds->cur_sector = cur_sector + nr_sectors;
 309     return (bmds->cur_sector >= total_sectors);
 310 }
 311
 312 /* Called with iothread lock taken.  */
 313
 314 static int set_dirty_tracking(void)
 315 {
 316     BlkMigDevState *bmds;
 317     int ret;
 318
 319     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 320         bmds->dirty_bitmap = bdrv_create_dirty_bitmap(bmds->bs, BLOCK_SIZE,
 321                                                       NULL);
 322         if (!bmds->dirty_bitmap) {
 323             ret = -errno;
 324             goto fail;
 325         }
 326     }
 327     return 0;
 328
 329 fail:
 330     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 331         if (bmds->dirty_bitmap) {
 332             bdrv_release_dirty_bitmap(bmds->bs, bmds->dirty_bitmap);
 333         }
 334     }
 335     return ret;
 336 }
 337
 338 static void unset_dirty_tracking(void)
 339 {
 340     BlkMigDevState *bmds;
 341
 342     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 343         bdrv_release_dirty_bitmap(bmds->bs, bmds->dirty_bitmap);
 344     }
 345 }
 346
 347 static void init_blk_migration_it(void *opaque, BlockDriverState *bs)
 348 {
 349     BlkMigDevState *bmds;
 350     int64_t sectors;
 351
 352     if (!bdrv_is_read_only(bs)) {
 353         sectors = bdrv_getlength(bs) >> BDRV_SECTOR_BITS;
 354         if (sectors <= 0) {
 355             return;
 356         }
 357
 358         bmds = g_malloc0(sizeof(BlkMigDevState));
 359         bmds->bs = bs;
 360         bmds->bulk_completed = 0;
 361         bmds->total_sectors = sectors;
 362         bmds->completed_sectors = 0;
 363         bmds->shared_base = block_mig_state.shared_base;
 364         alloc_aio_bitmap(bmds);
 365         error_setg(&bmds->blocker, "block device is in use by migration");
 366         bdrv_op_block_all(bs, bmds->blocker);
 367         bdrv_ref(bs);
 368
 369         block_mig_state.total_sector_sum += sectors;
 370
 371         if (bmds->shared_base) {
 372             DPRINTF("Start migration for %s with shared base image\n",
 373                     bs->device_name);
 374         } else {
 375             DPRINTF("Start full migration for %s\n", bs->device_name);
 376         }
 377
 378         QSIMPLEQ_INSERT_TAIL(&block_mig_state.bmds_list, bmds, entry);
 379     }
 380 }
 381
 382 static void init_blk_migration(QEMUFile *f)
 383 {
 384     block_mig_state.submitted = 0;
 385     block_mig_state.read_done = 0;
 386     block_mig_state.transferred = 0;
 387     block_mig_state.total_sector_sum = 0;
 388     block_mig_state.prev_progress = -1;
 389     block_mig_state.bulk_completed = 0;
 390     block_mig_state.zero_blocks = migrate_zero_blocks();
 391
 392     bdrv_iterate(init_blk_migration_it, NULL);
 393 }
 394
 395 /* Called with no lock taken.  */
 396
 397 static int blk_mig_save_bulked_block(QEMUFile *f)
 398 {
 399     int64_t completed_sector_sum = 0;
 400     BlkMigDevState *bmds;
 401     int progress;
 402     int ret = 0;
 403
 404     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 405         if (bmds->bulk_completed == 0) {
 406             if (mig_save_device_bulk(f, bmds) == 1) {
 407                 /* completed bulk section for this device */
 408                 bmds->bulk_completed = 1;
 409             }
 410             completed_sector_sum += bmds->completed_sectors;
 411             ret = 1;
 412             break;
 413         } else {
 414             completed_sector_sum += bmds->completed_sectors;
 415         }
 416     }
 417
 418     if (block_mig_state.total_sector_sum != 0) {
 419         progress = completed_sector_sum * 100 /
 420                    block_mig_state.total_sector_sum;
 421     } else {
 422         progress = 100;
 423     }
 424     if (progress != block_mig_state.prev_progress) {
 425         block_mig_state.prev_progress = progress;
 426         qemu_put_be64(f, (progress << BDRV_SECTOR_BITS)
 427                          | BLK_MIG_FLAG_PROGRESS);
 428         DPRINTF("Completed %d %%\r", progress);
 429     }
 430
 431     return ret;
 432 }
 433
 434 static void blk_mig_reset_dirty_cursor(void)
 435 {
 436     BlkMigDevState *bmds;
 437
 438     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 439         bmds->cur_dirty = 0;
 440     }
 441 }
 442
 443 /* Called with iothread lock taken.  */
 444
 445 static int mig_save_device_dirty(QEMUFile *f, BlkMigDevState *bmds,
 446                                  int is_async)
 447 {
 448     BlkMigBlock *blk;
 449     int64_t total_sectors = bmds->total_sectors;
 450     int64_t sector;
 451     int nr_sectors;
 452     int ret = -EIO;
 453
 454     for (sector = bmds->cur_dirty; sector < bmds->total_sectors;) {
 455         blk_mig_lock();
 456         if (bmds_aio_inflight(bmds, sector)) {
 457             blk_mig_unlock();
 458             bdrv_drain_all();
 459         } else {
 460             blk_mig_unlock();
 461         }
 462         if (bdrv_get_dirty(bmds->bs, bmds->dirty_bitmap, sector)) {
 463
 464             if (total_sectors - sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 465                 nr_sectors = total_sectors - sector;
 466             } else {
 467                 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 468             }
 469             blk = g_malloc(sizeof(BlkMigBlock));
 470             blk->buf = g_malloc(BLOCK_SIZE);
 471             blk->bmds = bmds;
 472             blk->sector = sector;
 473             blk->nr_sectors = nr_sectors;
 474
 475             if (is_async) {
 476                 blk->iov.iov_base = blk->buf;
 477                 blk->iov.iov_len = nr_sectors * BDRV_SECTOR_SIZE;
 478                 qemu_iovec_init_external(&blk->qiov, &blk->iov, 1);
 479
 480                 blk->aiocb = bdrv_aio_readv(bmds->bs, sector, &blk->qiov,
 481                                             nr_sectors, blk_mig_read_cb, blk);
 482
 483                 blk_mig_lock();
 484                 block_mig_state.submitted++;
 485                 bmds_set_aio_inflight(bmds, sector, nr_sectors, 1);
 486                 blk_mig_unlock();
 487             } else {
 488                 ret = bdrv_read(bmds->bs, sector, blk->buf, nr_sectors);
 489                 if (ret < 0) {
 490                     goto error;
 491                 }
 492                 blk_send(f, blk);
 493
 494                 g_free(blk->buf);
 495                 g_free(blk);
 496             }
 497
 498             bdrv_reset_dirty(bmds->bs, sector, nr_sectors);
 499             break;
 500         }
 501         sector += BDRV_SECTORS_PER_DIRTY_CHUNK;
 502         bmds->cur_dirty = sector;
 503     }
 504
 505     return (bmds->cur_dirty >= bmds->total_sectors);
 506
 507 error:
 508     DPRINTF("Error reading sector %" PRId64 "\n", sector);
 509     g_free(blk->buf);
 510     g_free(blk);
 511     return ret;
 512 }
 513
 514 /* Called with iothread lock taken.
 515  *
 516  * return value:
 517  * 0: too much data for max_downtime
 518  * 1: few enough data for max_downtime
 519 */
 520 static int blk_mig_save_dirty_block(QEMUFile *f, int is_async)
 521 {
 522     BlkMigDevState *bmds;
 523     int ret = 1;
 524
 525     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 526         ret = mig_save_device_dirty(f, bmds, is_async);
 527         if (ret <= 0) {
 528             break;
 529         }
 530     }
 531
 532     return ret;
 533 }
 534
 535 /* Called with no locks taken.  */
 536
 537 static int flush_blks(QEMUFile *f)
 538 {
 539     BlkMigBlock *blk;
 540     int ret = 0;
 541
 542     DPRINTF("%s Enter submitted %d read_done %d transferred %d\n",
 543             __FUNCTION__, block_mig_state.submitted, block_mig_state.read_done,
 544             block_mig_state.transferred);
 545
 546     blk_mig_lock();
 547     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 548         if (qemu_file_rate_limit(f)) {
 549             break;
 550         }
 551         if (blk->ret < 0) {
 552             ret = blk->ret;
 553             break;
 554         }
 555
 556         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 557         blk_mig_unlock();
 558         blk_send(f, blk);
 559         blk_mig_lock();
 560
 561         g_free(blk->buf);
 562         g_free(blk);
 563
 564         block_mig_state.read_done--;
 565         block_mig_state.transferred++;
 566         assert(block_mig_state.read_done >= 0);
 567     }
 568     blk_mig_unlock();
 569
 570     DPRINTF("%s Exit submitted %d read_done %d transferred %d\n", __FUNCTION__,
 571             block_mig_state.submitted, block_mig_state.read_done,
 572             block_mig_state.transferred);
 573     return ret;
 574 }
 575
 576 /* Called with iothread lock taken.  */
 577
 578 static int64_t get_remaining_dirty(void)
 579 {
 580     BlkMigDevState *bmds;
 581     int64_t dirty = 0;
 582
 583     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 584         dirty += bdrv_get_dirty_count(bmds->bs, bmds->dirty_bitmap);
 585     }
 586
 587     return dirty << BDRV_SECTOR_BITS;
 588 }
 589
 590 /* Called with iothread lock taken.  */
 591
 592 static void blk_mig_cleanup(void)
 593 {
 594     BlkMigDevState *bmds;
 595     BlkMigBlock *blk;
 596
 597     bdrv_drain_all();
 598
 599     unset_dirty_tracking();
 600
 601     blk_mig_lock();
 602     while ((bmds = QSIMPLEQ_FIRST(&block_mig_state.bmds_list)) != NULL) {
 603         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.bmds_list, entry);
 604         bdrv_op_unblock_all(bmds->bs, bmds->blocker);
 605         error_free(bmds->blocker);
 606         bdrv_unref(bmds->bs);
 607         g_free(bmds->aio_bitmap);
 608         g_free(bmds);
 609     }
 610
 611     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 612         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 613         g_free(blk->buf);
 614         g_free(blk);
 615     }
 616     blk_mig_unlock();
 617 }
 618
 619 static void block_migration_cancel(void *opaque)
 620 {
 621     blk_mig_cleanup();
 622 }
 623
 624 static int block_save_setup(QEMUFile *f, void *opaque)
 625 {
 626     int ret;
 627
 628     DPRINTF("Enter save live setup submitted %d transferred %d\n",
 629             block_mig_state.submitted, block_mig_state.transferred);
 630
 631     qemu_mutex_lock_iothread();
 632     init_blk_migration(f);
 633
 634     /* start track dirty blocks */
 635     ret = set_dirty_tracking();
 636
 637     if (ret) {
 638         qemu_mutex_unlock_iothread();
 639         return ret;
 640     }
 641
 642     qemu_mutex_unlock_iothread();
 643
 644     ret = flush_blks(f);
 645     blk_mig_reset_dirty_cursor();
 646     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 647
 648     return ret;
 649 }
 650
 651 static int block_save_iterate(QEMUFile *f, void *opaque)
 652 {
 653     int ret;
 654     int64_t last_ftell = qemu_ftell(f);
 655
 656     DPRINTF("Enter save live iterate submitted %d transferred %d\n",
 657             block_mig_state.submitted, block_mig_state.transferred);
 658
 659     ret = flush_blks(f);
 660     if (ret) {
 661         return ret;
 662     }
 663
 664     blk_mig_reset_dirty_cursor();
 665
 666     /* control the rate of transfer */
 667     blk_mig_lock();
 668     while ((block_mig_state.submitted +
 669             block_mig_state.read_done) * BLOCK_SIZE <
 670            qemu_file_get_rate_limit(f)) {
 671         blk_mig_unlock();
 672         if (block_mig_state.bulk_completed == 0) {
 673             /* first finish the bulk phase */
 674             if (blk_mig_save_bulked_block(f) == 0) {
 675                 /* finished saving bulk on all devices */
 676                 block_mig_state.bulk_completed = 1;
 677             }
 678             ret = 0;
 679         } else {
 680             /* Always called with iothread lock taken for
 681              * simplicity, block_save_complete also calls it.
 682              */
 683             qemu_mutex_lock_iothread();
 684             ret = blk_mig_save_dirty_block(f, 1);
 685             qemu_mutex_unlock_iothread();
 686         }
 687         if (ret < 0) {
 688             return ret;
 689         }
 690         blk_mig_lock();
 691         if (ret != 0) {
 692             /* no more dirty blocks */
 693             break;
 694         }
 695     }
 696     blk_mig_unlock();
 697
 698     ret = flush_blks(f);
 699     if (ret) {
 700         return ret;
 701     }
 702
 703     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 704     return qemu_ftell(f) - last_ftell;
 705 }
 706
 707 /* Called with iothread lock taken.  */
 708
 709 static int block_save_complete(QEMUFile *f, void *opaque)
 710 {
 711     int ret;
 712
 713     DPRINTF("Enter save live complete submitted %d transferred %d\n",
 714             block_mig_state.submitted, block_mig_state.transferred);
 715
 716     ret = flush_blks(f);
 717     if (ret) {
 718         return ret;
 719     }
 720
 721     blk_mig_reset_dirty_cursor();
 722
 723     /* we know for sure that save bulk is completed and
 724        all async read completed */
 725     blk_mig_lock();
 726     assert(block_mig_state.submitted == 0);
 727     blk_mig_unlock();
 728
 729     do {
 730         ret = blk_mig_save_dirty_block(f, 0);
 731         if (ret < 0) {
 732             return ret;
 733         }
 734     } while (ret == 0);
 735
 736     /* report completion */
 737     qemu_put_be64(f, (100 << BDRV_SECTOR_BITS) | BLK_MIG_FLAG_PROGRESS);
 738
 739     DPRINTF("Block migration completed\n");
 740
 741     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 742
 743     blk_mig_cleanup();
 744     return 0;
 745 }
 746
 747 static uint64_t block_save_pending(QEMUFile *f, void *opaque, uint64_t max_size)
 748 {
 749     /* Estimate pending number of bytes to send */
 750     uint64_t pending;
 751
 752     qemu_mutex_lock_iothread();
 753     blk_mig_lock();
 754     pending = get_remaining_dirty() +
 755                        block_mig_state.submitted * BLOCK_SIZE +
 756                        block_mig_state.read_done * BLOCK_SIZE;
 757
 758     /* Report at least one block pending during bulk phase */
 759     if (pending == 0 && !block_mig_state.bulk_completed) {
 760         pending = BLOCK_SIZE;
 761     }
 762     blk_mig_unlock();
 763     qemu_mutex_unlock_iothread();
 764
 765     DPRINTF("Enter save live pending  %" PRIu64 "\n", pending);
 766     return pending;
 767 }
 768
 769 static int block_load(QEMUFile *f, void *opaque, int version_id)
 770 {
 771     static int banner_printed;
 772     int len, flags;
 773     char device_name[256];
 774     int64_t addr;
 775     BlockDriverState *bs, *bs_prev = NULL;
 776     uint8_t *buf;
 777     int64_t total_sectors = 0;
 778     int nr_sectors;
 779     int ret;
 780
 781     do {
 782         addr = qemu_get_be64(f);
 783
 784         flags = addr & ~BDRV_SECTOR_MASK;
 785         addr >>= BDRV_SECTOR_BITS;
 786
 787         if (flags & BLK_MIG_FLAG_DEVICE_BLOCK) {
 788             /* get device name */
 789             len = qemu_get_byte(f);
 790             qemu_get_buffer(f, (uint8_t *)device_name, len);
 791             device_name[len] = '\0';
 792
 793             bs = bdrv_find(device_name);
 794             if (!bs) {
 795                 fprintf(stderr, "Error unknown block device %s\n",
 796                         device_name);
 797                 return -EINVAL;
 798             }
 799
 800             if (bs != bs_prev) {
 801                 bs_prev = bs;
 802                 total_sectors = bdrv_getlength(bs) >> BDRV_SECTOR_BITS;
 803                 if (total_sectors <= 0) {
 804                     error_report("Error getting length of block device %s",
 805                                  device_name);
 806                     return -EINVAL;
 807                 }
 808             }
 809
 810             if (total_sectors - addr < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 811                 nr_sectors = total_sectors - addr;
 812             } else {
 813                 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 814             }
 815
 816             if (flags & BLK_MIG_FLAG_ZERO_BLOCK) {
 817                 ret = bdrv_write_zeroes(bs, addr, nr_sectors,
 818                                         BDRV_REQ_MAY_UNMAP);
 819             } else {
 820                 buf = g_malloc(BLOCK_SIZE);
 821                 qemu_get_buffer(f, buf, BLOCK_SIZE);
 822                 ret = bdrv_write(bs, addr, buf, nr_sectors);
 823                 g_free(buf);
 824             }
 825
 826             if (ret < 0) {
 827                 return ret;
 828             }
 829         } else if (flags & BLK_MIG_FLAG_PROGRESS) {
 830             if (!banner_printed) {
 831                 printf("Receiving block device images\n");
 832                 banner_printed = 1;
 833             }
 834             printf("Completed %d %%%c", (int)addr,
 835                    (addr == 100) ? '\n' : '\r');
 836             fflush(stdout);
 837         } else if (!(flags & BLK_MIG_FLAG_EOS)) {
 838             fprintf(stderr, "Unknown block migration flags: %#x\n", flags);
 839             return -EINVAL;
 840         }
 841         ret = qemu_file_get_error(f);
 842         if (ret != 0) {
 843             return ret;
 844         }
 845     } while (!(flags & BLK_MIG_FLAG_EOS));
 846
 847     return 0;
 848 }
 849
 850 static void block_set_params(const MigrationParams *params, void *opaque)
 851 {
 852     block_mig_state.blk_enable = params->blk;
 853     block_mig_state.shared_base = params->shared;
 854
 855     /* shared base means that blk_enable = 1 */
 856     block_mig_state.blk_enable |= params->shared;
 857 }
 858
 859 static bool block_is_active(void *opaque)
 860 {
 861     return block_mig_state.blk_enable == 1;
 862 }
 863
 864 static SaveVMHandlers savevm_block_handlers = {
 865     .set_params = block_set_params,
 866     .save_live_setup = block_save_setup,
 867     .save_live_iterate = block_save_iterate,
 868     .save_live_complete = block_save_complete,
 869     .save_live_pending = block_save_pending,
 870     .load_state = block_load,
 871     .cancel = block_migration_cancel,
 872     .is_active = block_is_active,
 873 };
 874
 875 void blk_mig_init(void)
 876 {
 877     QSIMPLEQ_INIT(&block_mig_state.bmds_list);
 878     QSIMPLEQ_INIT(&block_mig_state.blk_list);
 879     qemu_mutex_init(&block_mig_state.lock);
 880
 881     register_savevm_live(NULL, "block", 0, 1, &savevm_block_handlers,
 882                          &block_mig_state);
 883 }