migration/block.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/osdep.h"
  17 #include "qapi/error.h"
  18 #include "qemu/error-report.h"
  19 #include "qemu/main-loop.h"
  20 #include "qemu/cutils.h"
  21 #include "qemu/queue.h"
  22 #include "block.h"
  23 #include "block/dirty-bitmap.h"
  24 #include "migration/misc.h"
  25 #include "migration.h"
  26 #include "migration-stats.h"
  27 #include "migration/register.h"
  28 #include "qemu-file.h"
  29 #include "migration/vmstate.h"
  30 #include "sysemu/block-backend.h"
  31 #include "trace.h"
  32 #include "options.h"
  33
  34 #define BLK_MIG_BLOCK_SIZE           (1ULL << 20)
  35 #define BDRV_SECTORS_PER_DIRTY_CHUNK (BLK_MIG_BLOCK_SIZE >> BDRV_SECTOR_BITS)
  36
  37 #define BLK_MIG_FLAG_DEVICE_BLOCK       0x01
  38 #define BLK_MIG_FLAG_EOS                0x02
  39 #define BLK_MIG_FLAG_PROGRESS           0x04
  40 #define BLK_MIG_FLAG_ZERO_BLOCK         0x08
  41
  42 #define MAX_IS_ALLOCATED_SEARCH (65536 * BDRV_SECTOR_SIZE)
  43
  44 #define MAX_IO_BUFFERS 512
  45 #define MAX_PARALLEL_IO 16
  46
  47 typedef struct BlkMigDevState {
  48     /* Written during setup phase.  Can be read without a lock.  */
  49     BlockBackend *blk;
  50     char *blk_name;
  51     int shared_base;
  52     int64_t total_sectors;
  53     QSIMPLEQ_ENTRY(BlkMigDevState) entry;
  54     Error *blocker;
  55
  56     /* Only used by migration thread.  Does not need a lock.  */
  57     int bulk_completed;
  58     int64_t cur_sector;
  59     int64_t cur_dirty;
  60
  61     /* Data in the aio_bitmap is protected by block migration lock.
  62      * Allocation and free happen during setup and cleanup respectively.
  63      */
  64     unsigned long *aio_bitmap;
  65
  66     /* Protected by block migration lock.  */
  67     int64_t completed_sectors;
  68
  69     /* During migration this is protected by iothread lock / AioContext.
  70      * Allocation and free happen during setup and cleanup respectively.
  71      */
  72     BdrvDirtyBitmap *dirty_bitmap;
  73 } BlkMigDevState;
  74
  75 typedef struct BlkMigBlock {
  76     /* Only used by migration thread.  */
  77     uint8_t *buf;
  78     BlkMigDevState *bmds;
  79     int64_t sector;
  80     int nr_sectors;
  81     QEMUIOVector qiov;
  82     BlockAIOCB *aiocb;
  83
  84     /* Protected by block migration lock.  */
  85     int ret;
  86     QSIMPLEQ_ENTRY(BlkMigBlock) entry;
  87 } BlkMigBlock;
  88
  89 typedef struct BlkMigState {
  90     QSIMPLEQ_HEAD(, BlkMigDevState) bmds_list;
  91     int64_t total_sector_sum;
  92     bool zero_blocks;
  93
  94     /* Protected by lock.  */
  95     QSIMPLEQ_HEAD(, BlkMigBlock) blk_list;
  96     int submitted;
  97     int read_done;
  98
  99     /* Only used by migration thread.  Does not need a lock.  */
 100     int transferred;
 101     int prev_progress;
 102     int bulk_completed;
 103
 104     /* Lock must be taken _inside_ the iothread lock and any AioContexts.  */
 105     QemuMutex lock;
 106 } BlkMigState;
 107
 108 static BlkMigState block_mig_state;
 109
 110 static void blk_mig_lock(void)
 111 {
 112     qemu_mutex_lock(&block_mig_state.lock);
 113 }
 114
 115 static void blk_mig_unlock(void)
 116 {
 117     qemu_mutex_unlock(&block_mig_state.lock);
 118 }
 119
 120 /* Must run outside of the iothread lock during the bulk phase,
 121  * or the VM will stall.
 122  */
 123
 124 static void blk_send(QEMUFile *f, BlkMigBlock * blk)
 125 {
 126     int len;
 127     uint64_t flags = BLK_MIG_FLAG_DEVICE_BLOCK;
 128
 129     if (block_mig_state.zero_blocks &&
 130         buffer_is_zero(blk->buf, BLK_MIG_BLOCK_SIZE)) {
 131         flags |= BLK_MIG_FLAG_ZERO_BLOCK;
 132     }
 133
 134     /* sector number and flags */
 135     qemu_put_be64(f, (blk->sector << BDRV_SECTOR_BITS)
 136                      | flags);
 137
 138     /* device name */
 139     len = strlen(blk->bmds->blk_name);
 140     qemu_put_byte(f, len);
 141     qemu_put_buffer(f, (uint8_t *) blk->bmds->blk_name, len);
 142
 143     /* if a block is zero we need to flush here since the network
 144      * bandwidth is now a lot higher than the storage device bandwidth.
 145      * thus if we queue zero blocks we slow down the migration */
 146     if (flags & BLK_MIG_FLAG_ZERO_BLOCK) {
 147         qemu_fflush(f);
 148         return;
 149     }
 150
 151     qemu_put_buffer(f, blk->buf, BLK_MIG_BLOCK_SIZE);
 152 }
 153
 154 int blk_mig_active(void)
 155 {
 156     return !QSIMPLEQ_EMPTY(&block_mig_state.bmds_list);
 157 }
 158
 159 int blk_mig_bulk_active(void)
 160 {
 161     return blk_mig_active() && !block_mig_state.bulk_completed;
 162 }
 163
 164 uint64_t blk_mig_bytes_transferred(void)
 165 {
 166     BlkMigDevState *bmds;
 167     uint64_t sum = 0;
 168
 169     blk_mig_lock();
 170     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 171         sum += bmds->completed_sectors;
 172     }
 173     blk_mig_unlock();
 174     return sum << BDRV_SECTOR_BITS;
 175 }
 176
 177 uint64_t blk_mig_bytes_remaining(void)
 178 {
 179     return blk_mig_bytes_total() - blk_mig_bytes_transferred();
 180 }
 181
 182 uint64_t blk_mig_bytes_total(void)
 183 {
 184     BlkMigDevState *bmds;
 185     uint64_t sum = 0;
 186
 187     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 188         sum += bmds->total_sectors;
 189     }
 190     return sum << BDRV_SECTOR_BITS;
 191 }
 192
 193
 194 /* Called with migration lock held.  */
 195
 196 static int bmds_aio_inflight(BlkMigDevState *bmds, int64_t sector)
 197 {
 198     int64_t chunk = sector / (int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK;
 199
 200     if (sector < bmds->total_sectors) {
 201         return !!(bmds->aio_bitmap[chunk / (sizeof(unsigned long) * 8)] &
 202             (1UL << (chunk % (sizeof(unsigned long) * 8))));
 203     } else {
 204         return 0;
 205     }
 206 }
 207
 208 /* Called with migration lock held.  */
 209
 210 static void bmds_set_aio_inflight(BlkMigDevState *bmds, int64_t sector_num,
 211                              int nb_sectors, int set)
 212 {
 213     int64_t start, end;
 214     unsigned long val, idx, bit;
 215
 216     start = sector_num / BDRV_SECTORS_PER_DIRTY_CHUNK;
 217     end = (sector_num + nb_sectors - 1) / BDRV_SECTORS_PER_DIRTY_CHUNK;
 218
 219     for (; start <= end; start++) {
 220         idx = start / (sizeof(unsigned long) * 8);
 221         bit = start % (sizeof(unsigned long) * 8);
 222         val = bmds->aio_bitmap[idx];
 223         if (set) {
 224             val |= 1UL << bit;
 225         } else {
 226             val &= ~(1UL << bit);
 227         }
 228         bmds->aio_bitmap[idx] = val;
 229     }
 230 }
 231
 232 static void alloc_aio_bitmap(BlkMigDevState *bmds)
 233 {
 234     int64_t bitmap_size;
 235
 236     bitmap_size = bmds->total_sectors + BDRV_SECTORS_PER_DIRTY_CHUNK * 8 - 1;
 237     bitmap_size /= BDRV_SECTORS_PER_DIRTY_CHUNK * 8;
 238
 239     bmds->aio_bitmap = g_malloc0(bitmap_size);
 240 }
 241
 242 /* Never hold migration lock when yielding to the main loop!  */
 243
 244 static void blk_mig_read_cb(void *opaque, int ret)
 245 {
 246     BlkMigBlock *blk = opaque;
 247
 248     blk_mig_lock();
 249     blk->ret = ret;
 250
 251     QSIMPLEQ_INSERT_TAIL(&block_mig_state.blk_list, blk, entry);
 252     bmds_set_aio_inflight(blk->bmds, blk->sector, blk->nr_sectors, 0);
 253
 254     block_mig_state.submitted--;
 255     block_mig_state.read_done++;
 256     assert(block_mig_state.submitted >= 0);
 257     blk_mig_unlock();
 258 }
 259
 260 /* Called with no lock taken.  */
 261
 262 static int mig_save_device_bulk(QEMUFile *f, BlkMigDevState *bmds)
 263 {
 264     int64_t total_sectors = bmds->total_sectors;
 265     int64_t cur_sector = bmds->cur_sector;
 266     BlockBackend *bb = bmds->blk;
 267     BlkMigBlock *blk;
 268     int nr_sectors;
 269     int64_t count;
 270
 271     if (bmds->shared_base) {
 272         qemu_mutex_lock_iothread();
 273         aio_context_acquire(blk_get_aio_context(bb));
 274         /* Skip unallocated sectors; intentionally treats failure or
 275          * partial sector as an allocated sector */
 276         while (cur_sector < total_sectors &&
 277                !bdrv_is_allocated(blk_bs(bb), cur_sector * BDRV_SECTOR_SIZE,
 278                                   MAX_IS_ALLOCATED_SEARCH, &count)) {
 279             if (count < BDRV_SECTOR_SIZE) {
 280                 break;
 281             }
 282             cur_sector += count >> BDRV_SECTOR_BITS;
 283         }
 284         aio_context_release(blk_get_aio_context(bb));
 285         qemu_mutex_unlock_iothread();
 286     }
 287
 288     if (cur_sector >= total_sectors) {
 289         bmds->cur_sector = bmds->completed_sectors = total_sectors;
 290         return 1;
 291     }
 292
 293     bmds->completed_sectors = cur_sector;
 294
 295     cur_sector &= ~((int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK - 1);
 296
 297     /* we are going to transfer a full block even if it is not allocated */
 298     nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 299
 300     if (total_sectors - cur_sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 301         nr_sectors = total_sectors - cur_sector;
 302     }
 303
 304     blk = g_new(BlkMigBlock, 1);
 305     blk->buf = g_malloc(BLK_MIG_BLOCK_SIZE);
 306     blk->bmds = bmds;
 307     blk->sector = cur_sector;
 308     blk->nr_sectors = nr_sectors;
 309
 310     qemu_iovec_init_buf(&blk->qiov, blk->buf, nr_sectors * BDRV_SECTOR_SIZE);
 311
 312     blk_mig_lock();
 313     block_mig_state.submitted++;
 314     blk_mig_unlock();
 315
 316     /* We do not know if bs is under the main thread (and thus does
 317      * not acquire the AioContext when doing AIO) or rather under
 318      * dataplane.  Thus acquire both the iothread mutex and the
 319      * AioContext.
 320      *
 321      * This is ugly and will disappear when we make bdrv_* thread-safe,
 322      * without the need to acquire the AioContext.
 323      */
 324     qemu_mutex_lock_iothread();
 325     aio_context_acquire(blk_get_aio_context(bmds->blk));
 326     bdrv_reset_dirty_bitmap(bmds->dirty_bitmap, cur_sector * BDRV_SECTOR_SIZE,
 327                             nr_sectors * BDRV_SECTOR_SIZE);
 328     blk->aiocb = blk_aio_preadv(bb, cur_sector * BDRV_SECTOR_SIZE, &blk->qiov,
 329                                 0, blk_mig_read_cb, blk);
 330     aio_context_release(blk_get_aio_context(bmds->blk));
 331     qemu_mutex_unlock_iothread();
 332
 333     bmds->cur_sector = cur_sector + nr_sectors;
 334     return (bmds->cur_sector >= total_sectors);
 335 }
 336
 337 /* Called with iothread lock taken.  */
 338
 339 static int set_dirty_tracking(void)
 340 {
 341     BlkMigDevState *bmds;
 342     int ret;
 343
 344     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 345         bmds->dirty_bitmap = bdrv_create_dirty_bitmap(blk_bs(bmds->blk),
 346                                                       BLK_MIG_BLOCK_SIZE,
 347                                                       NULL, NULL);
 348         if (!bmds->dirty_bitmap) {
 349             ret = -errno;
 350             goto fail;
 351         }
 352     }
 353     return 0;
 354
 355 fail:
 356     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 357         if (bmds->dirty_bitmap) {
 358             bdrv_release_dirty_bitmap(bmds->dirty_bitmap);
 359         }
 360     }
 361     return ret;
 362 }
 363
 364 /* Called with iothread lock taken.  */
 365
 366 static void unset_dirty_tracking(void)
 367 {
 368     BlkMigDevState *bmds;
 369
 370     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 371         if (bmds->dirty_bitmap) {
 372             bdrv_release_dirty_bitmap(bmds->dirty_bitmap);
 373         }
 374     }
 375 }
 376
 377 static int init_blk_migration(QEMUFile *f)
 378 {
 379     BlockDriverState *bs;
 380     BlkMigDevState *bmds;
 381     int64_t sectors;
 382     BdrvNextIterator it;
 383     int i, num_bs = 0;
 384     struct {
 385         BlkMigDevState *bmds;
 386         BlockDriverState *bs;
 387     } *bmds_bs;
 388     Error *local_err = NULL;
 389     int ret;
 390
 391     block_mig_state.submitted = 0;
 392     block_mig_state.read_done = 0;
 393     block_mig_state.transferred = 0;
 394     block_mig_state.total_sector_sum = 0;
 395     block_mig_state.prev_progress = -1;
 396     block_mig_state.bulk_completed = 0;
 397     block_mig_state.zero_blocks = migrate_zero_blocks();
 398
 399     for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) {
 400         num_bs++;
 401     }
 402     bmds_bs = g_malloc0(num_bs * sizeof(*bmds_bs));
 403
 404     for (i = 0, bs = bdrv_first(&it); bs; bs = bdrv_next(&it), i++) {
 405         if (bdrv_is_read_only(bs)) {
 406             continue;
 407         }
 408
 409         sectors = bdrv_nb_sectors(bs);
 410         if (sectors <= 0) {
 411             ret = sectors;
 412             bdrv_next_cleanup(&it);
 413             goto out;
 414         }
 415
 416         bmds = g_new0(BlkMigDevState, 1);
 417         bmds->blk = blk_new(qemu_get_aio_context(),
 418                             BLK_PERM_CONSISTENT_READ, BLK_PERM_ALL);
 419         bmds->blk_name = g_strdup(bdrv_get_device_name(bs));
 420         bmds->bulk_completed = 0;
 421         bmds->total_sectors = sectors;
 422         bmds->completed_sectors = 0;
 423         bmds->shared_base = migrate_block_incremental();
 424
 425         assert(i < num_bs);
 426         bmds_bs[i].bmds = bmds;
 427         bmds_bs[i].bs = bs;
 428
 429         block_mig_state.total_sector_sum += sectors;
 430
 431         if (bmds->shared_base) {
 432             trace_migration_block_init_shared(bdrv_get_device_name(bs));
 433         } else {
 434             trace_migration_block_init_full(bdrv_get_device_name(bs));
 435         }
 436
 437         QSIMPLEQ_INSERT_TAIL(&block_mig_state.bmds_list, bmds, entry);
 438     }
 439
 440     /* Can only insert new BDSes now because doing so while iterating block
 441      * devices may end up in a deadlock (iterating the new BDSes, too). */
 442     for (i = 0; i < num_bs; i++) {
 443         bmds = bmds_bs[i].bmds;
 444         bs = bmds_bs[i].bs;
 445
 446         if (bmds) {
 447             ret = blk_insert_bs(bmds->blk, bs, &local_err);
 448             if (ret < 0) {
 449                 error_report_err(local_err);
 450                 goto out;
 451             }
 452
 453             alloc_aio_bitmap(bmds);
 454             error_setg(&bmds->blocker, "block device is in use by migration");
 455             bdrv_op_block_all(bs, bmds->blocker);
 456         }
 457     }
 458
 459     ret = 0;
 460 out:
 461     g_free(bmds_bs);
 462     return ret;
 463 }
 464
 465 /* Called with no lock taken.  */
 466
 467 static int blk_mig_save_bulked_block(QEMUFile *f)
 468 {
 469     int64_t completed_sector_sum = 0;
 470     BlkMigDevState *bmds;
 471     int progress;
 472     int ret = 0;
 473
 474     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 475         if (bmds->bulk_completed == 0) {
 476             if (mig_save_device_bulk(f, bmds) == 1) {
 477                 /* completed bulk section for this device */
 478                 bmds->bulk_completed = 1;
 479             }
 480             completed_sector_sum += bmds->completed_sectors;
 481             ret = 1;
 482             break;
 483         } else {
 484             completed_sector_sum += bmds->completed_sectors;
 485         }
 486     }
 487
 488     if (block_mig_state.total_sector_sum != 0) {
 489         progress = completed_sector_sum * 100 /
 490                    block_mig_state.total_sector_sum;
 491     } else {
 492         progress = 100;
 493     }
 494     if (progress != block_mig_state.prev_progress) {
 495         block_mig_state.prev_progress = progress;
 496         qemu_put_be64(f, (progress << BDRV_SECTOR_BITS)
 497                          | BLK_MIG_FLAG_PROGRESS);
 498         trace_migration_block_progression(progress);
 499     }
 500
 501     return ret;
 502 }
 503
 504 static void blk_mig_reset_dirty_cursor(void)
 505 {
 506     BlkMigDevState *bmds;
 507
 508     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 509         bmds->cur_dirty = 0;
 510     }
 511 }
 512
 513 /* Called with iothread lock and AioContext taken.  */
 514
 515 static int mig_save_device_dirty(QEMUFile *f, BlkMigDevState *bmds,
 516                                  int is_async)
 517 {
 518     BlkMigBlock *blk;
 519     int64_t total_sectors = bmds->total_sectors;
 520     int64_t sector;
 521     int nr_sectors;
 522     int ret = -EIO;
 523
 524     for (sector = bmds->cur_dirty; sector < bmds->total_sectors;) {
 525         blk_mig_lock();
 526         if (bmds_aio_inflight(bmds, sector)) {
 527             blk_mig_unlock();
 528             blk_drain(bmds->blk);
 529         } else {
 530             blk_mig_unlock();
 531         }
 532         bdrv_dirty_bitmap_lock(bmds->dirty_bitmap);
 533         if (bdrv_dirty_bitmap_get_locked(bmds->dirty_bitmap,
 534                                          sector * BDRV_SECTOR_SIZE)) {
 535             if (total_sectors - sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 536                 nr_sectors = total_sectors - sector;
 537             } else {
 538                 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 539             }
 540             bdrv_reset_dirty_bitmap_locked(bmds->dirty_bitmap,
 541                                            sector * BDRV_SECTOR_SIZE,
 542                                            nr_sectors * BDRV_SECTOR_SIZE);
 543             bdrv_dirty_bitmap_unlock(bmds->dirty_bitmap);
 544
 545             blk = g_new(BlkMigBlock, 1);
 546             blk->buf = g_malloc(BLK_MIG_BLOCK_SIZE);
 547             blk->bmds = bmds;
 548             blk->sector = sector;
 549             blk->nr_sectors = nr_sectors;
 550
 551             if (is_async) {
 552                 qemu_iovec_init_buf(&blk->qiov, blk->buf,
 553                                     nr_sectors * BDRV_SECTOR_SIZE);
 554
 555                 blk->aiocb = blk_aio_preadv(bmds->blk,
 556                                             sector * BDRV_SECTOR_SIZE,
 557                                             &blk->qiov, 0, blk_mig_read_cb,
 558                                             blk);
 559
 560                 blk_mig_lock();
 561                 block_mig_state.submitted++;
 562                 bmds_set_aio_inflight(bmds, sector, nr_sectors, 1);
 563                 blk_mig_unlock();
 564             } else {
 565                 ret = blk_pread(bmds->blk, sector * BDRV_SECTOR_SIZE,
 566                                 nr_sectors * BDRV_SECTOR_SIZE, blk->buf, 0);
 567                 if (ret < 0) {
 568                     goto error;
 569                 }
 570                 blk_send(f, blk);
 571
 572                 g_free(blk->buf);
 573                 g_free(blk);
 574             }
 575
 576             sector += nr_sectors;
 577             bmds->cur_dirty = sector;
 578             break;
 579         }
 580
 581         bdrv_dirty_bitmap_unlock(bmds->dirty_bitmap);
 582         sector += BDRV_SECTORS_PER_DIRTY_CHUNK;
 583         bmds->cur_dirty = sector;
 584     }
 585
 586     return (bmds->cur_dirty >= bmds->total_sectors);
 587
 588 error:
 589     trace_migration_block_save_device_dirty(sector);
 590     g_free(blk->buf);
 591     g_free(blk);
 592     return ret;
 593 }
 594
 595 /* Called with iothread lock taken.
 596  *
 597  * return value:
 598  * 0: too much data for max_downtime
 599  * 1: few enough data for max_downtime
 600 */
 601 static int blk_mig_save_dirty_block(QEMUFile *f, int is_async)
 602 {
 603     BlkMigDevState *bmds;
 604     int ret = 1;
 605
 606     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 607         aio_context_acquire(blk_get_aio_context(bmds->blk));
 608         ret = mig_save_device_dirty(f, bmds, is_async);
 609         aio_context_release(blk_get_aio_context(bmds->blk));
 610         if (ret <= 0) {
 611             break;
 612         }
 613     }
 614
 615     return ret;
 616 }
 617
 618 /* Called with no locks taken.  */
 619
 620 static int flush_blks(QEMUFile *f)
 621 {
 622     BlkMigBlock *blk;
 623     int ret = 0;
 624
 625     trace_migration_block_flush_blks("Enter", block_mig_state.submitted,
 626                                      block_mig_state.read_done,
 627                                      block_mig_state.transferred);
 628
 629     blk_mig_lock();
 630     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 631         if (migration_rate_exceeded(f)) {
 632             break;
 633         }
 634         if (blk->ret < 0) {
 635             ret = blk->ret;
 636             break;
 637         }
 638
 639         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 640         blk_mig_unlock();
 641         blk_send(f, blk);
 642         blk_mig_lock();
 643
 644         g_free(blk->buf);
 645         g_free(blk);
 646
 647         block_mig_state.read_done--;
 648         block_mig_state.transferred++;
 649         assert(block_mig_state.read_done >= 0);
 650     }
 651     blk_mig_unlock();
 652
 653     trace_migration_block_flush_blks("Exit", block_mig_state.submitted,
 654                                      block_mig_state.read_done,
 655                                      block_mig_state.transferred);
 656     return ret;
 657 }
 658
 659 /* Called with iothread lock taken.  */
 660
 661 static int64_t get_remaining_dirty(void)
 662 {
 663     BlkMigDevState *bmds;
 664     int64_t dirty = 0;
 665
 666     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 667         aio_context_acquire(blk_get_aio_context(bmds->blk));
 668         dirty += bdrv_get_dirty_count(bmds->dirty_bitmap);
 669         aio_context_release(blk_get_aio_context(bmds->blk));
 670     }
 671
 672     return dirty;
 673 }
 674
 675
 676
 677 /* Called with iothread lock taken.  */
 678 static void block_migration_cleanup_bmds(void)
 679 {
 680     BlkMigDevState *bmds;
 681     BlockDriverState *bs;
 682     AioContext *ctx;
 683
 684     unset_dirty_tracking();
 685
 686     while ((bmds = QSIMPLEQ_FIRST(&block_mig_state.bmds_list)) != NULL) {
 687         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.bmds_list, entry);
 688
 689         bs = blk_bs(bmds->blk);
 690         if (bs) {
 691             bdrv_op_unblock_all(bs, bmds->blocker);
 692         }
 693         error_free(bmds->blocker);
 694
 695         /* Save ctx, because bmds->blk can disappear during blk_unref.  */
 696         ctx = blk_get_aio_context(bmds->blk);
 697         aio_context_acquire(ctx);
 698         blk_unref(bmds->blk);
 699         aio_context_release(ctx);
 700
 701         g_free(bmds->blk_name);
 702         g_free(bmds->aio_bitmap);
 703         g_free(bmds);
 704     }
 705 }
 706
 707 /* Called with iothread lock taken.  */
 708 static void block_migration_cleanup(void *opaque)
 709 {
 710     BlkMigBlock *blk;
 711
 712     bdrv_drain_all();
 713
 714     block_migration_cleanup_bmds();
 715
 716     blk_mig_lock();
 717     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 718         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 719         g_free(blk->buf);
 720         g_free(blk);
 721     }
 722     blk_mig_unlock();
 723 }
 724
 725 static int block_save_setup(QEMUFile *f, void *opaque)
 726 {
 727     int ret;
 728
 729     trace_migration_block_save("setup", block_mig_state.submitted,
 730                                block_mig_state.transferred);
 731
 732     qemu_mutex_lock_iothread();
 733     ret = init_blk_migration(f);
 734     if (ret < 0) {
 735         qemu_mutex_unlock_iothread();
 736         return ret;
 737     }
 738
 739     /* start track dirty blocks */
 740     ret = set_dirty_tracking();
 741
 742     qemu_mutex_unlock_iothread();
 743
 744     if (ret) {
 745         return ret;
 746     }
 747
 748     ret = flush_blks(f);
 749     blk_mig_reset_dirty_cursor();
 750     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 751
 752     return ret;
 753 }
 754
 755 static int block_save_iterate(QEMUFile *f, void *opaque)
 756 {
 757     int ret;
 758     uint64_t last_bytes = qemu_file_transferred_noflush(f);
 759
 760     trace_migration_block_save("iterate", block_mig_state.submitted,
 761                                block_mig_state.transferred);
 762
 763     ret = flush_blks(f);
 764     if (ret) {
 765         return ret;
 766     }
 767
 768     blk_mig_reset_dirty_cursor();
 769
 770     /* control the rate of transfer */
 771     blk_mig_lock();
 772     while (block_mig_state.read_done * BLK_MIG_BLOCK_SIZE <
 773            migration_rate_get() &&
 774            block_mig_state.submitted < MAX_PARALLEL_IO &&
 775            (block_mig_state.submitted + block_mig_state.read_done) <
 776            MAX_IO_BUFFERS) {
 777         blk_mig_unlock();
 778         if (block_mig_state.bulk_completed == 0) {
 779             /* first finish the bulk phase */
 780             if (blk_mig_save_bulked_block(f) == 0) {
 781                 /* finished saving bulk on all devices */
 782                 block_mig_state.bulk_completed = 1;
 783             }
 784             ret = 0;
 785         } else {
 786             /* Always called with iothread lock taken for
 787              * simplicity, block_save_complete also calls it.
 788              */
 789             qemu_mutex_lock_iothread();
 790             ret = blk_mig_save_dirty_block(f, 1);
 791             qemu_mutex_unlock_iothread();
 792         }
 793         if (ret < 0) {
 794             return ret;
 795         }
 796         blk_mig_lock();
 797         if (ret != 0) {
 798             /* no more dirty blocks */
 799             break;
 800         }
 801     }
 802     blk_mig_unlock();
 803
 804     ret = flush_blks(f);
 805     if (ret) {
 806         return ret;
 807     }
 808
 809     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 810     uint64_t delta_bytes = qemu_file_transferred_noflush(f) - last_bytes;
 811     return (delta_bytes > 0);
 812 }
 813
 814 /* Called with iothread lock taken.  */
 815
 816 static int block_save_complete(QEMUFile *f, void *opaque)
 817 {
 818     int ret;
 819
 820     trace_migration_block_save("complete", block_mig_state.submitted,
 821                                block_mig_state.transferred);
 822
 823     ret = flush_blks(f);
 824     if (ret) {
 825         return ret;
 826     }
 827
 828     blk_mig_reset_dirty_cursor();
 829
 830     /* we know for sure that save bulk is completed and
 831        all async read completed */
 832     blk_mig_lock();
 833     assert(block_mig_state.submitted == 0);
 834     blk_mig_unlock();
 835
 836     do {
 837         ret = blk_mig_save_dirty_block(f, 0);
 838         if (ret < 0) {
 839             return ret;
 840         }
 841     } while (ret == 0);
 842
 843     /* report completion */
 844     qemu_put_be64(f, (100 << BDRV_SECTOR_BITS) | BLK_MIG_FLAG_PROGRESS);
 845
 846     trace_migration_block_save_complete();
 847
 848     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 849
 850     /* Make sure that our BlockBackends are gone, so that the block driver
 851      * nodes can be inactivated. */
 852     block_migration_cleanup_bmds();
 853
 854     return 0;
 855 }
 856
 857 static void block_state_pending(void *opaque, uint64_t *must_precopy,
 858                                 uint64_t *can_postcopy)
 859 {
 860     /* Estimate pending number of bytes to send */
 861     uint64_t pending;
 862
 863     qemu_mutex_lock_iothread();
 864     pending = get_remaining_dirty();
 865     qemu_mutex_unlock_iothread();
 866
 867     blk_mig_lock();
 868     pending += block_mig_state.submitted * BLK_MIG_BLOCK_SIZE +
 869                block_mig_state.read_done * BLK_MIG_BLOCK_SIZE;
 870     blk_mig_unlock();
 871
 872     /* Report at least one block pending during bulk phase */
 873     if (!pending && !block_mig_state.bulk_completed) {
 874         pending = BLK_MIG_BLOCK_SIZE;
 875     }
 876
 877     trace_migration_block_state_pending(pending);
 878     /* We don't do postcopy */
 879     *must_precopy += pending;
 880 }
 881
 882 static int block_load(QEMUFile *f, void *opaque, int version_id)
 883 {
 884     static int banner_printed;
 885     int len, flags;
 886     char device_name[256];
 887     int64_t addr;
 888     BlockBackend *blk, *blk_prev = NULL;
 889     Error *local_err = NULL;
 890     uint8_t *buf;
 891     int64_t total_sectors = 0;
 892     int nr_sectors;
 893     int ret;
 894     BlockDriverInfo bdi;
 895     int cluster_size = BLK_MIG_BLOCK_SIZE;
 896
 897     do {
 898         addr = qemu_get_be64(f);
 899
 900         flags = addr & (BDRV_SECTOR_SIZE - 1);
 901         addr >>= BDRV_SECTOR_BITS;
 902
 903         if (flags & BLK_MIG_FLAG_DEVICE_BLOCK) {
 904             /* get device name */
 905             len = qemu_get_byte(f);
 906             qemu_get_buffer(f, (uint8_t *)device_name, len);
 907             device_name[len] = '\0';
 908
 909             blk = blk_by_name(device_name);
 910             if (!blk) {
 911                 fprintf(stderr, "Error unknown block device %s\n",
 912                         device_name);
 913                 return -EINVAL;
 914             }
 915
 916             if (blk != blk_prev) {
 917                 blk_prev = blk;
 918                 total_sectors = blk_nb_sectors(blk);
 919                 if (total_sectors <= 0) {
 920                     error_report("Error getting length of block device %s",
 921                                  device_name);
 922                     return -EINVAL;
 923                 }
 924
 925                 blk_activate(blk, &local_err);
 926                 if (local_err) {
 927                     error_report_err(local_err);
 928                     return -EINVAL;
 929                 }
 930
 931                 ret = bdrv_get_info(blk_bs(blk), &bdi);
 932                 if (ret == 0 && bdi.cluster_size > 0 &&
 933                     bdi.cluster_size <= BLK_MIG_BLOCK_SIZE &&
 934                     BLK_MIG_BLOCK_SIZE % bdi.cluster_size == 0) {
 935                     cluster_size = bdi.cluster_size;
 936                 } else {
 937                     cluster_size = BLK_MIG_BLOCK_SIZE;
 938                 }
 939             }
 940
 941             if (total_sectors - addr < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 942                 nr_sectors = total_sectors - addr;
 943             } else {
 944                 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 945             }
 946
 947             if (flags & BLK_MIG_FLAG_ZERO_BLOCK) {
 948                 ret = blk_pwrite_zeroes(blk, addr * BDRV_SECTOR_SIZE,
 949                                         nr_sectors * BDRV_SECTOR_SIZE,
 950                                         BDRV_REQ_MAY_UNMAP);
 951             } else {
 952                 int i;
 953                 int64_t cur_addr;
 954                 uint8_t *cur_buf;
 955
 956                 buf = g_malloc(BLK_MIG_BLOCK_SIZE);
 957                 qemu_get_buffer(f, buf, BLK_MIG_BLOCK_SIZE);
 958                 for (i = 0; i < BLK_MIG_BLOCK_SIZE / cluster_size; i++) {
 959                     cur_addr = addr * BDRV_SECTOR_SIZE + i * cluster_size;
 960                     cur_buf = buf + i * cluster_size;
 961
 962                     if ((!block_mig_state.zero_blocks ||
 963                         cluster_size < BLK_MIG_BLOCK_SIZE) &&
 964                         buffer_is_zero(cur_buf, cluster_size)) {
 965                         ret = blk_pwrite_zeroes(blk, cur_addr,
 966                                                 cluster_size,
 967                                                 BDRV_REQ_MAY_UNMAP);
 968                     } else {
 969                         ret = blk_pwrite(blk, cur_addr, cluster_size, cur_buf,
 970                                          0);
 971                     }
 972                     if (ret < 0) {
 973                         break;
 974                     }
 975                 }
 976                 g_free(buf);
 977             }
 978
 979             if (ret < 0) {
 980                 return ret;
 981             }
 982         } else if (flags & BLK_MIG_FLAG_PROGRESS) {
 983             if (!banner_printed) {
 984                 printf("Receiving block device images\n");
 985                 banner_printed = 1;
 986             }
 987             printf("Completed %d %%%c", (int)addr,
 988                    (addr == 100) ? '\n' : '\r');
 989             fflush(stdout);
 990         } else if (!(flags & BLK_MIG_FLAG_EOS)) {
 991             fprintf(stderr, "Unknown block migration flags: 0x%x\n", flags);
 992             return -EINVAL;
 993         }
 994         ret = qemu_file_get_error(f);
 995         if (ret != 0) {
 996             return ret;
 997         }
 998     } while (!(flags & BLK_MIG_FLAG_EOS));
 999
1000     return 0;
1001 }
1002
1003 static bool block_is_active(void *opaque)
1004 {
1005     return migrate_block();
1006 }
1007
1008 static SaveVMHandlers savevm_block_handlers = {
1009     .save_setup = block_save_setup,
1010     .save_live_iterate = block_save_iterate,
1011     .save_live_complete_precopy = block_save_complete,
1012     .state_pending_exact = block_state_pending,
1013     .state_pending_estimate = block_state_pending,
1014     .load_state = block_load,
1015     .save_cleanup = block_migration_cleanup,
1016     .is_active = block_is_active,
1017 };
1018
1019 void blk_mig_init(void)
1020 {
1021     QSIMPLEQ_INIT(&block_mig_state.bmds_list);
1022     QSIMPLEQ_INIT(&block_mig_state.blk_list);
1023     qemu_mutex_init(&block_mig_state.lock);
1024
1025     register_savevm_live("block", 0, 1, &savevm_block_handlers,
1026                          &block_mig_state);
1027 }