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 bdrv_dirty_bitmap_lock().
  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 BQL.  */
 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 BQL 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         bql_lock();
 273         /* Skip unallocated sectors; intentionally treats failure or
 274          * partial sector as an allocated sector */
 275         while (cur_sector < total_sectors &&
 276                !bdrv_is_allocated(blk_bs(bb), cur_sector * BDRV_SECTOR_SIZE,
 277                                   MAX_IS_ALLOCATED_SEARCH, &count)) {
 278             if (count < BDRV_SECTOR_SIZE) {
 279                 break;
 280             }
 281             cur_sector += count >> BDRV_SECTOR_BITS;
 282         }
 283         bql_unlock();
 284     }
 285
 286     if (cur_sector >= total_sectors) {
 287         bmds->cur_sector = bmds->completed_sectors = total_sectors;
 288         return 1;
 289     }
 290
 291     bmds->completed_sectors = cur_sector;
 292
 293     cur_sector &= ~((int64_t)BDRV_SECTORS_PER_DIRTY_CHUNK - 1);
 294
 295     /* we are going to transfer a full block even if it is not allocated */
 296     nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 297
 298     if (total_sectors - cur_sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 299         nr_sectors = total_sectors - cur_sector;
 300     }
 301
 302     blk = g_new(BlkMigBlock, 1);
 303     blk->buf = g_malloc(BLK_MIG_BLOCK_SIZE);
 304     blk->bmds = bmds;
 305     blk->sector = cur_sector;
 306     blk->nr_sectors = nr_sectors;
 307
 308     qemu_iovec_init_buf(&blk->qiov, blk->buf, nr_sectors * BDRV_SECTOR_SIZE);
 309
 310     blk_mig_lock();
 311     block_mig_state.submitted++;
 312     blk_mig_unlock();
 313
 314     /*
 315      * The migration thread does not have an AioContext. Lock the BQL so that
 316      * I/O runs in the main loop AioContext (see
 317      * qemu_get_current_aio_context()).
 318      */
 319     bql_lock();
 320     bdrv_reset_dirty_bitmap(bmds->dirty_bitmap, cur_sector * BDRV_SECTOR_SIZE,
 321                             nr_sectors * BDRV_SECTOR_SIZE);
 322     blk->aiocb = blk_aio_preadv(bb, cur_sector * BDRV_SECTOR_SIZE, &blk->qiov,
 323                                 0, blk_mig_read_cb, blk);
 324     bql_unlock();
 325
 326     bmds->cur_sector = cur_sector + nr_sectors;
 327     return (bmds->cur_sector >= total_sectors);
 328 }
 329
 330 /* Called with the BQL taken.  */
 331
 332 static int set_dirty_tracking(void)
 333 {
 334     BlkMigDevState *bmds;
 335     int ret;
 336
 337     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 338         bmds->dirty_bitmap = bdrv_create_dirty_bitmap(blk_bs(bmds->blk),
 339                                                       BLK_MIG_BLOCK_SIZE,
 340                                                       NULL, NULL);
 341         if (!bmds->dirty_bitmap) {
 342             ret = -errno;
 343             goto fail;
 344         }
 345     }
 346     return 0;
 347
 348 fail:
 349     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 350         if (bmds->dirty_bitmap) {
 351             bdrv_release_dirty_bitmap(bmds->dirty_bitmap);
 352         }
 353     }
 354     return ret;
 355 }
 356
 357 /* Called with the BQL taken.  */
 358
 359 static void unset_dirty_tracking(void)
 360 {
 361     BlkMigDevState *bmds;
 362
 363     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 364         if (bmds->dirty_bitmap) {
 365             bdrv_release_dirty_bitmap(bmds->dirty_bitmap);
 366         }
 367     }
 368 }
 369
 370 static int init_blk_migration(QEMUFile *f)
 371 {
 372     BlockDriverState *bs;
 373     BlkMigDevState *bmds;
 374     int64_t sectors;
 375     BdrvNextIterator it;
 376     int i, num_bs = 0;
 377     struct {
 378         BlkMigDevState *bmds;
 379         BlockDriverState *bs;
 380     } *bmds_bs;
 381     Error *local_err = NULL;
 382     int ret;
 383
 384     GRAPH_RDLOCK_GUARD_MAINLOOP();
 385
 386     block_mig_state.submitted = 0;
 387     block_mig_state.read_done = 0;
 388     block_mig_state.transferred = 0;
 389     block_mig_state.total_sector_sum = 0;
 390     block_mig_state.prev_progress = -1;
 391     block_mig_state.bulk_completed = 0;
 392     block_mig_state.zero_blocks = migrate_zero_blocks();
 393
 394     for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) {
 395         num_bs++;
 396     }
 397     bmds_bs = g_malloc0(num_bs * sizeof(*bmds_bs));
 398
 399     for (i = 0, bs = bdrv_first(&it); bs; bs = bdrv_next(&it), i++) {
 400         if (bdrv_is_read_only(bs)) {
 401             continue;
 402         }
 403
 404         sectors = bdrv_nb_sectors(bs);
 405         if (sectors == 0) {
 406             continue;
 407         }
 408         if (sectors < 0) {
 409             ret = sectors;
 410             bdrv_next_cleanup(&it);
 411             goto out;
 412         }
 413
 414         bmds = g_new0(BlkMigDevState, 1);
 415         bmds->blk = blk_new(qemu_get_aio_context(),
 416                             BLK_PERM_CONSISTENT_READ, BLK_PERM_ALL);
 417         bmds->blk_name = g_strdup(bdrv_get_device_name(bs));
 418         bmds->bulk_completed = 0;
 419         bmds->total_sectors = sectors;
 420         bmds->completed_sectors = 0;
 421         bmds->shared_base = migrate_block_incremental();
 422
 423         assert(i < num_bs);
 424         bmds_bs[i].bmds = bmds;
 425         bmds_bs[i].bs = bs;
 426
 427         block_mig_state.total_sector_sum += sectors;
 428
 429         if (bmds->shared_base) {
 430             trace_migration_block_init_shared(bdrv_get_device_name(bs));
 431         } else {
 432             trace_migration_block_init_full(bdrv_get_device_name(bs));
 433         }
 434
 435         QSIMPLEQ_INSERT_TAIL(&block_mig_state.bmds_list, bmds, entry);
 436     }
 437
 438     /* Can only insert new BDSes now because doing so while iterating block
 439      * devices may end up in a deadlock (iterating the new BDSes, too). */
 440     for (i = 0; i < num_bs; i++) {
 441         bmds = bmds_bs[i].bmds;
 442         bs = bmds_bs[i].bs;
 443
 444         if (bmds) {
 445             ret = blk_insert_bs(bmds->blk, bs, &local_err);
 446             if (ret < 0) {
 447                 error_report_err(local_err);
 448                 goto out;
 449             }
 450
 451             alloc_aio_bitmap(bmds);
 452             error_setg(&bmds->blocker, "block device is in use by migration");
 453             bdrv_op_block_all(bs, bmds->blocker);
 454         }
 455     }
 456
 457     ret = 0;
 458 out:
 459     g_free(bmds_bs);
 460     return ret;
 461 }
 462
 463 /* Called with no lock taken.  */
 464
 465 static int blk_mig_save_bulked_block(QEMUFile *f)
 466 {
 467     int64_t completed_sector_sum = 0;
 468     BlkMigDevState *bmds;
 469     int progress;
 470     int ret = 0;
 471
 472     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 473         if (bmds->bulk_completed == 0) {
 474             if (mig_save_device_bulk(f, bmds) == 1) {
 475                 /* completed bulk section for this device */
 476                 bmds->bulk_completed = 1;
 477             }
 478             completed_sector_sum += bmds->completed_sectors;
 479             ret = 1;
 480             break;
 481         } else {
 482             completed_sector_sum += bmds->completed_sectors;
 483         }
 484     }
 485
 486     if (block_mig_state.total_sector_sum != 0) {
 487         progress = completed_sector_sum * 100 /
 488                    block_mig_state.total_sector_sum;
 489     } else {
 490         progress = 100;
 491     }
 492     if (progress != block_mig_state.prev_progress) {
 493         block_mig_state.prev_progress = progress;
 494         qemu_put_be64(f, (progress << BDRV_SECTOR_BITS)
 495                          | BLK_MIG_FLAG_PROGRESS);
 496         trace_migration_block_progression(progress);
 497     }
 498
 499     return ret;
 500 }
 501
 502 static void blk_mig_reset_dirty_cursor(void)
 503 {
 504     BlkMigDevState *bmds;
 505
 506     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 507         bmds->cur_dirty = 0;
 508     }
 509 }
 510
 511 /* Called with the BQL taken.  */
 512
 513 static int mig_save_device_dirty(QEMUFile *f, BlkMigDevState *bmds,
 514                                  int is_async)
 515 {
 516     BlkMigBlock *blk;
 517     int64_t total_sectors = bmds->total_sectors;
 518     int64_t sector;
 519     int nr_sectors;
 520     int ret = -EIO;
 521
 522     for (sector = bmds->cur_dirty; sector < bmds->total_sectors;) {
 523         blk_mig_lock();
 524         if (bmds_aio_inflight(bmds, sector)) {
 525             blk_mig_unlock();
 526             blk_drain(bmds->blk);
 527         } else {
 528             blk_mig_unlock();
 529         }
 530         bdrv_dirty_bitmap_lock(bmds->dirty_bitmap);
 531         if (bdrv_dirty_bitmap_get_locked(bmds->dirty_bitmap,
 532                                          sector * BDRV_SECTOR_SIZE)) {
 533             if (total_sectors - sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 534                 nr_sectors = total_sectors - sector;
 535             } else {
 536                 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 537             }
 538             bdrv_reset_dirty_bitmap_locked(bmds->dirty_bitmap,
 539                                            sector * BDRV_SECTOR_SIZE,
 540                                            nr_sectors * BDRV_SECTOR_SIZE);
 541             bdrv_dirty_bitmap_unlock(bmds->dirty_bitmap);
 542
 543             blk = g_new(BlkMigBlock, 1);
 544             blk->buf = g_malloc(BLK_MIG_BLOCK_SIZE);
 545             blk->bmds = bmds;
 546             blk->sector = sector;
 547             blk->nr_sectors = nr_sectors;
 548
 549             if (is_async) {
 550                 qemu_iovec_init_buf(&blk->qiov, blk->buf,
 551                                     nr_sectors * BDRV_SECTOR_SIZE);
 552
 553                 blk->aiocb = blk_aio_preadv(bmds->blk,
 554                                             sector * BDRV_SECTOR_SIZE,
 555                                             &blk->qiov, 0, blk_mig_read_cb,
 556                                             blk);
 557
 558                 blk_mig_lock();
 559                 block_mig_state.submitted++;
 560                 bmds_set_aio_inflight(bmds, sector, nr_sectors, 1);
 561                 blk_mig_unlock();
 562             } else {
 563                 ret = blk_pread(bmds->blk, sector * BDRV_SECTOR_SIZE,
 564                                 nr_sectors * BDRV_SECTOR_SIZE, blk->buf, 0);
 565                 if (ret < 0) {
 566                     goto error;
 567                 }
 568                 blk_send(f, blk);
 569
 570                 g_free(blk->buf);
 571                 g_free(blk);
 572             }
 573
 574             sector += nr_sectors;
 575             bmds->cur_dirty = sector;
 576             break;
 577         }
 578
 579         bdrv_dirty_bitmap_unlock(bmds->dirty_bitmap);
 580         sector += BDRV_SECTORS_PER_DIRTY_CHUNK;
 581         bmds->cur_dirty = sector;
 582     }
 583
 584     return (bmds->cur_dirty >= bmds->total_sectors);
 585
 586 error:
 587     trace_migration_block_save_device_dirty(sector);
 588     g_free(blk->buf);
 589     g_free(blk);
 590     return ret;
 591 }
 592
 593 /* Called with the BQL taken.
 594  *
 595  * return value:
 596  * 0: too much data for max_downtime
 597  * 1: few enough data for max_downtime
 598 */
 599 static int blk_mig_save_dirty_block(QEMUFile *f, int is_async)
 600 {
 601     BlkMigDevState *bmds;
 602     int ret = 1;
 603
 604     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 605         ret = mig_save_device_dirty(f, bmds, is_async);
 606         if (ret <= 0) {
 607             break;
 608         }
 609     }
 610
 611     return ret;
 612 }
 613
 614 /* Called with no locks taken.  */
 615
 616 static int flush_blks(QEMUFile *f)
 617 {
 618     BlkMigBlock *blk;
 619     int ret = 0;
 620
 621     trace_migration_block_flush_blks("Enter", block_mig_state.submitted,
 622                                      block_mig_state.read_done,
 623                                      block_mig_state.transferred);
 624
 625     blk_mig_lock();
 626     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 627         if (migration_rate_exceeded(f)) {
 628             break;
 629         }
 630         if (blk->ret < 0) {
 631             ret = blk->ret;
 632             break;
 633         }
 634
 635         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 636         blk_mig_unlock();
 637         blk_send(f, blk);
 638         blk_mig_lock();
 639
 640         g_free(blk->buf);
 641         g_free(blk);
 642
 643         block_mig_state.read_done--;
 644         block_mig_state.transferred++;
 645         assert(block_mig_state.read_done >= 0);
 646     }
 647     blk_mig_unlock();
 648
 649     trace_migration_block_flush_blks("Exit", block_mig_state.submitted,
 650                                      block_mig_state.read_done,
 651                                      block_mig_state.transferred);
 652     return ret;
 653 }
 654
 655 /* Called with the BQL taken.  */
 656
 657 static int64_t get_remaining_dirty(void)
 658 {
 659     BlkMigDevState *bmds;
 660     int64_t dirty = 0;
 661
 662     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 663         bdrv_dirty_bitmap_lock(bmds->dirty_bitmap);
 664         dirty += bdrv_get_dirty_count(bmds->dirty_bitmap);
 665         bdrv_dirty_bitmap_unlock(bmds->dirty_bitmap);
 666     }
 667
 668     return dirty;
 669 }
 670
 671
 672
 673 /* Called with the BQL taken.  */
 674 static void block_migration_cleanup_bmds(void)
 675 {
 676     BlkMigDevState *bmds;
 677     BlockDriverState *bs;
 678
 679     unset_dirty_tracking();
 680
 681     while ((bmds = QSIMPLEQ_FIRST(&block_mig_state.bmds_list)) != NULL) {
 682         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.bmds_list, entry);
 683
 684         bs = blk_bs(bmds->blk);
 685         if (bs) {
 686             bdrv_op_unblock_all(bs, bmds->blocker);
 687         }
 688         error_free(bmds->blocker);
 689         blk_unref(bmds->blk);
 690         g_free(bmds->blk_name);
 691         g_free(bmds->aio_bitmap);
 692         g_free(bmds);
 693     }
 694 }
 695
 696 /* Called with the BQL taken.  */
 697 static void block_migration_cleanup(void *opaque)
 698 {
 699     BlkMigBlock *blk;
 700
 701     bdrv_drain_all();
 702
 703     block_migration_cleanup_bmds();
 704
 705     blk_mig_lock();
 706     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 707         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 708         g_free(blk->buf);
 709         g_free(blk);
 710     }
 711     blk_mig_unlock();
 712 }
 713
 714 static int block_save_setup(QEMUFile *f, void *opaque)
 715 {
 716     int ret;
 717
 718     trace_migration_block_save("setup", block_mig_state.submitted,
 719                                block_mig_state.transferred);
 720
 721     warn_report("block migration is deprecated;"
 722                 " use blockdev-mirror with NBD instead");
 723
 724     ret = init_blk_migration(f);
 725     if (ret < 0) {
 726         return ret;
 727     }
 728
 729     /* start track dirty blocks */
 730     ret = set_dirty_tracking();
 731     if (ret) {
 732         return ret;
 733     }
 734
 735     ret = flush_blks(f);
 736     blk_mig_reset_dirty_cursor();
 737     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 738
 739     return ret;
 740 }
 741
 742 static int block_save_iterate(QEMUFile *f, void *opaque)
 743 {
 744     int ret;
 745     uint64_t last_bytes = qemu_file_transferred(f);
 746
 747     trace_migration_block_save("iterate", block_mig_state.submitted,
 748                                block_mig_state.transferred);
 749
 750     ret = flush_blks(f);
 751     if (ret) {
 752         return ret;
 753     }
 754
 755     blk_mig_reset_dirty_cursor();
 756
 757     /* control the rate of transfer */
 758     blk_mig_lock();
 759     while (block_mig_state.read_done * BLK_MIG_BLOCK_SIZE <
 760            migration_rate_get() &&
 761            block_mig_state.submitted < MAX_PARALLEL_IO &&
 762            (block_mig_state.submitted + block_mig_state.read_done) <
 763            MAX_IO_BUFFERS) {
 764         blk_mig_unlock();
 765         if (block_mig_state.bulk_completed == 0) {
 766             /* first finish the bulk phase */
 767             if (blk_mig_save_bulked_block(f) == 0) {
 768                 /* finished saving bulk on all devices */
 769                 block_mig_state.bulk_completed = 1;
 770             }
 771             ret = 0;
 772         } else {
 773             /* Always called with the BQL taken for
 774              * simplicity, block_save_complete also calls it.
 775              */
 776             bql_lock();
 777             ret = blk_mig_save_dirty_block(f, 1);
 778             bql_unlock();
 779         }
 780         if (ret < 0) {
 781             return ret;
 782         }
 783         blk_mig_lock();
 784         if (ret != 0) {
 785             /* no more dirty blocks */
 786             break;
 787         }
 788     }
 789     blk_mig_unlock();
 790
 791     ret = flush_blks(f);
 792     if (ret) {
 793         return ret;
 794     }
 795
 796     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 797     uint64_t delta_bytes = qemu_file_transferred(f) - last_bytes;
 798     return (delta_bytes > 0);
 799 }
 800
 801 /* Called with the BQL taken.  */
 802
 803 static int block_save_complete(QEMUFile *f, void *opaque)
 804 {
 805     int ret;
 806
 807     trace_migration_block_save("complete", block_mig_state.submitted,
 808                                block_mig_state.transferred);
 809
 810     ret = flush_blks(f);
 811     if (ret) {
 812         return ret;
 813     }
 814
 815     blk_mig_reset_dirty_cursor();
 816
 817     /* we know for sure that save bulk is completed and
 818        all async read completed */
 819     blk_mig_lock();
 820     assert(block_mig_state.submitted == 0);
 821     blk_mig_unlock();
 822
 823     do {
 824         ret = blk_mig_save_dirty_block(f, 0);
 825         if (ret < 0) {
 826             return ret;
 827         }
 828     } while (ret == 0);
 829
 830     /* report completion */
 831     qemu_put_be64(f, (100 << BDRV_SECTOR_BITS) | BLK_MIG_FLAG_PROGRESS);
 832
 833     trace_migration_block_save_complete();
 834
 835     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 836
 837     /* Make sure that our BlockBackends are gone, so that the block driver
 838      * nodes can be inactivated. */
 839     block_migration_cleanup_bmds();
 840
 841     return 0;
 842 }
 843
 844 static void block_state_pending(void *opaque, uint64_t *must_precopy,
 845                                 uint64_t *can_postcopy)
 846 {
 847     /* Estimate pending number of bytes to send */
 848     uint64_t pending;
 849
 850     bql_lock();
 851     pending = get_remaining_dirty();
 852     bql_unlock();
 853
 854     blk_mig_lock();
 855     pending += block_mig_state.submitted * BLK_MIG_BLOCK_SIZE +
 856                block_mig_state.read_done * BLK_MIG_BLOCK_SIZE;
 857     blk_mig_unlock();
 858
 859     /* Report at least one block pending during bulk phase */
 860     if (!pending && !block_mig_state.bulk_completed) {
 861         pending = BLK_MIG_BLOCK_SIZE;
 862     }
 863
 864     trace_migration_block_state_pending(pending);
 865     /* We don't do postcopy */
 866     *must_precopy += pending;
 867 }
 868
 869 static int block_load(QEMUFile *f, void *opaque, int version_id)
 870 {
 871     static int banner_printed;
 872     int len, flags;
 873     char device_name[256];
 874     int64_t addr;
 875     BlockBackend *blk, *blk_prev = NULL;
 876     Error *local_err = NULL;
 877     uint8_t *buf;
 878     int64_t total_sectors = 0;
 879     int nr_sectors;
 880     int ret;
 881     BlockDriverInfo bdi;
 882     int cluster_size = BLK_MIG_BLOCK_SIZE;
 883
 884     do {
 885         addr = qemu_get_be64(f);
 886
 887         flags = addr & (BDRV_SECTOR_SIZE - 1);
 888         addr >>= BDRV_SECTOR_BITS;
 889
 890         if (flags & BLK_MIG_FLAG_DEVICE_BLOCK) {
 891             /* get device name */
 892             len = qemu_get_byte(f);
 893             qemu_get_buffer(f, (uint8_t *)device_name, len);
 894             device_name[len] = '\0';
 895
 896             blk = blk_by_name(device_name);
 897             if (!blk) {
 898                 fprintf(stderr, "Error unknown block device %s\n",
 899                         device_name);
 900                 return -EINVAL;
 901             }
 902
 903             if (blk != blk_prev) {
 904                 blk_prev = blk;
 905                 total_sectors = blk_nb_sectors(blk);
 906                 if (total_sectors <= 0) {
 907                     error_report("Error getting length of block device %s",
 908                                  device_name);
 909                     return -EINVAL;
 910                 }
 911
 912                 blk_activate(blk, &local_err);
 913                 if (local_err) {
 914                     error_report_err(local_err);
 915                     return -EINVAL;
 916                 }
 917
 918                 ret = bdrv_get_info(blk_bs(blk), &bdi);
 919                 if (ret == 0 && bdi.cluster_size > 0 &&
 920                     bdi.cluster_size <= BLK_MIG_BLOCK_SIZE &&
 921                     BLK_MIG_BLOCK_SIZE % bdi.cluster_size == 0) {
 922                     cluster_size = bdi.cluster_size;
 923                 } else {
 924                     cluster_size = BLK_MIG_BLOCK_SIZE;
 925                 }
 926             }
 927
 928             if (total_sectors - addr < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 929                 nr_sectors = total_sectors - addr;
 930             } else {
 931                 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 932             }
 933
 934             if (flags & BLK_MIG_FLAG_ZERO_BLOCK) {
 935                 ret = blk_pwrite_zeroes(blk, addr * BDRV_SECTOR_SIZE,
 936                                         nr_sectors * BDRV_SECTOR_SIZE,
 937                                         BDRV_REQ_MAY_UNMAP);
 938             } else {
 939                 int i;
 940                 int64_t cur_addr;
 941                 uint8_t *cur_buf;
 942
 943                 buf = g_malloc(BLK_MIG_BLOCK_SIZE);
 944                 qemu_get_buffer(f, buf, BLK_MIG_BLOCK_SIZE);
 945                 for (i = 0; i < BLK_MIG_BLOCK_SIZE / cluster_size; i++) {
 946                     cur_addr = addr * BDRV_SECTOR_SIZE + i * cluster_size;
 947                     cur_buf = buf + i * cluster_size;
 948
 949                     if ((!block_mig_state.zero_blocks ||
 950                         cluster_size < BLK_MIG_BLOCK_SIZE) &&
 951                         buffer_is_zero(cur_buf, cluster_size)) {
 952                         ret = blk_pwrite_zeroes(blk, cur_addr,
 953                                                 cluster_size,
 954                                                 BDRV_REQ_MAY_UNMAP);
 955                     } else {
 956                         ret = blk_pwrite(blk, cur_addr, cluster_size, cur_buf,
 957                                          0);
 958                     }
 959                     if (ret < 0) {
 960                         break;
 961                     }
 962                 }
 963                 g_free(buf);
 964             }
 965
 966             if (ret < 0) {
 967                 return ret;
 968             }
 969         } else if (flags & BLK_MIG_FLAG_PROGRESS) {
 970             if (!banner_printed) {
 971                 printf("Receiving block device images\n");
 972                 banner_printed = 1;
 973             }
 974             printf("Completed %d %%%c", (int)addr,
 975                    (addr == 100) ? '\n' : '\r');
 976             fflush(stdout);
 977         } else if (!(flags & BLK_MIG_FLAG_EOS)) {
 978             fprintf(stderr, "Unknown block migration flags: 0x%x\n", flags);
 979             return -EINVAL;
 980         }
 981         ret = qemu_file_get_error(f);
 982         if (ret != 0) {
 983             return ret;
 984         }
 985     } while (!(flags & BLK_MIG_FLAG_EOS));
 986
 987     return 0;
 988 }
 989
 990 static bool block_is_active(void *opaque)
 991 {
 992     return migrate_block();
 993 }
 994
 995 static SaveVMHandlers savevm_block_handlers = {
 996     .save_setup = block_save_setup,
 997     .save_live_iterate = block_save_iterate,
 998     .save_live_complete_precopy = block_save_complete,
 999     .state_pending_exact = block_state_pending,
1000     .state_pending_estimate = block_state_pending,
1001     .load_state = block_load,
1002     .save_cleanup = block_migration_cleanup,
1003     .is_active = block_is_active,
1004 };
1005
1006 void blk_mig_init(void)
1007 {
1008     QSIMPLEQ_INIT(&block_mig_state.bmds_list);
1009     QSIMPLEQ_INIT(&block_mig_state.blk_list);
1010     qemu_mutex_init(&block_mig_state.lock);
1011
1012     register_savevm_live("block", 0, 1, &savevm_block_handlers,
1013                          &block_mig_state);
1014 }