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     GRAPH_RDLOCK_GUARD_MAINLOOP();
 392
 393     block_mig_state.submitted = 0;
 394     block_mig_state.read_done = 0;
 395     block_mig_state.transferred = 0;
 396     block_mig_state.total_sector_sum = 0;
 397     block_mig_state.prev_progress = -1;
 398     block_mig_state.bulk_completed = 0;
 399     block_mig_state.zero_blocks = migrate_zero_blocks();
 400
 401     for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) {
 402         num_bs++;
 403     }
 404     bmds_bs = g_malloc0(num_bs * sizeof(*bmds_bs));
 405
 406     for (i = 0, bs = bdrv_first(&it); bs; bs = bdrv_next(&it), i++) {
 407         if (bdrv_is_read_only(bs)) {
 408             continue;
 409         }
 410
 411         sectors = bdrv_nb_sectors(bs);
 412         if (sectors <= 0) {
 413             ret = sectors;
 414             bdrv_next_cleanup(&it);
 415             goto out;
 416         }
 417
 418         bmds = g_new0(BlkMigDevState, 1);
 419         bmds->blk = blk_new(qemu_get_aio_context(),
 420                             BLK_PERM_CONSISTENT_READ, BLK_PERM_ALL);
 421         bmds->blk_name = g_strdup(bdrv_get_device_name(bs));
 422         bmds->bulk_completed = 0;
 423         bmds->total_sectors = sectors;
 424         bmds->completed_sectors = 0;
 425         bmds->shared_base = migrate_block_incremental();
 426
 427         assert(i < num_bs);
 428         bmds_bs[i].bmds = bmds;
 429         bmds_bs[i].bs = bs;
 430
 431         block_mig_state.total_sector_sum += sectors;
 432
 433         if (bmds->shared_base) {
 434             trace_migration_block_init_shared(bdrv_get_device_name(bs));
 435         } else {
 436             trace_migration_block_init_full(bdrv_get_device_name(bs));
 437         }
 438
 439         QSIMPLEQ_INSERT_TAIL(&block_mig_state.bmds_list, bmds, entry);
 440     }
 441
 442     /* Can only insert new BDSes now because doing so while iterating block
 443      * devices may end up in a deadlock (iterating the new BDSes, too). */
 444     for (i = 0; i < num_bs; i++) {
 445         bmds = bmds_bs[i].bmds;
 446         bs = bmds_bs[i].bs;
 447
 448         if (bmds) {
 449             ret = blk_insert_bs(bmds->blk, bs, &local_err);
 450             if (ret < 0) {
 451                 error_report_err(local_err);
 452                 goto out;
 453             }
 454
 455             alloc_aio_bitmap(bmds);
 456             error_setg(&bmds->blocker, "block device is in use by migration");
 457             bdrv_op_block_all(bs, bmds->blocker);
 458         }
 459     }
 460
 461     ret = 0;
 462 out:
 463     g_free(bmds_bs);
 464     return ret;
 465 }
 466
 467 /* Called with no lock taken.  */
 468
 469 static int blk_mig_save_bulked_block(QEMUFile *f)
 470 {
 471     int64_t completed_sector_sum = 0;
 472     BlkMigDevState *bmds;
 473     int progress;
 474     int ret = 0;
 475
 476     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 477         if (bmds->bulk_completed == 0) {
 478             if (mig_save_device_bulk(f, bmds) == 1) {
 479                 /* completed bulk section for this device */
 480                 bmds->bulk_completed = 1;
 481             }
 482             completed_sector_sum += bmds->completed_sectors;
 483             ret = 1;
 484             break;
 485         } else {
 486             completed_sector_sum += bmds->completed_sectors;
 487         }
 488     }
 489
 490     if (block_mig_state.total_sector_sum != 0) {
 491         progress = completed_sector_sum * 100 /
 492                    block_mig_state.total_sector_sum;
 493     } else {
 494         progress = 100;
 495     }
 496     if (progress != block_mig_state.prev_progress) {
 497         block_mig_state.prev_progress = progress;
 498         qemu_put_be64(f, (progress << BDRV_SECTOR_BITS)
 499                          | BLK_MIG_FLAG_PROGRESS);
 500         trace_migration_block_progression(progress);
 501     }
 502
 503     return ret;
 504 }
 505
 506 static void blk_mig_reset_dirty_cursor(void)
 507 {
 508     BlkMigDevState *bmds;
 509
 510     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 511         bmds->cur_dirty = 0;
 512     }
 513 }
 514
 515 /* Called with iothread lock and AioContext taken.  */
 516
 517 static int mig_save_device_dirty(QEMUFile *f, BlkMigDevState *bmds,
 518                                  int is_async)
 519 {
 520     BlkMigBlock *blk;
 521     int64_t total_sectors = bmds->total_sectors;
 522     int64_t sector;
 523     int nr_sectors;
 524     int ret = -EIO;
 525
 526     for (sector = bmds->cur_dirty; sector < bmds->total_sectors;) {
 527         blk_mig_lock();
 528         if (bmds_aio_inflight(bmds, sector)) {
 529             blk_mig_unlock();
 530             blk_drain(bmds->blk);
 531         } else {
 532             blk_mig_unlock();
 533         }
 534         bdrv_dirty_bitmap_lock(bmds->dirty_bitmap);
 535         if (bdrv_dirty_bitmap_get_locked(bmds->dirty_bitmap,
 536                                          sector * BDRV_SECTOR_SIZE)) {
 537             if (total_sectors - sector < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 538                 nr_sectors = total_sectors - sector;
 539             } else {
 540                 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 541             }
 542             bdrv_reset_dirty_bitmap_locked(bmds->dirty_bitmap,
 543                                            sector * BDRV_SECTOR_SIZE,
 544                                            nr_sectors * BDRV_SECTOR_SIZE);
 545             bdrv_dirty_bitmap_unlock(bmds->dirty_bitmap);
 546
 547             blk = g_new(BlkMigBlock, 1);
 548             blk->buf = g_malloc(BLK_MIG_BLOCK_SIZE);
 549             blk->bmds = bmds;
 550             blk->sector = sector;
 551             blk->nr_sectors = nr_sectors;
 552
 553             if (is_async) {
 554                 qemu_iovec_init_buf(&blk->qiov, blk->buf,
 555                                     nr_sectors * BDRV_SECTOR_SIZE);
 556
 557                 blk->aiocb = blk_aio_preadv(bmds->blk,
 558                                             sector * BDRV_SECTOR_SIZE,
 559                                             &blk->qiov, 0, blk_mig_read_cb,
 560                                             blk);
 561
 562                 blk_mig_lock();
 563                 block_mig_state.submitted++;
 564                 bmds_set_aio_inflight(bmds, sector, nr_sectors, 1);
 565                 blk_mig_unlock();
 566             } else {
 567                 ret = blk_pread(bmds->blk, sector * BDRV_SECTOR_SIZE,
 568                                 nr_sectors * BDRV_SECTOR_SIZE, blk->buf, 0);
 569                 if (ret < 0) {
 570                     goto error;
 571                 }
 572                 blk_send(f, blk);
 573
 574                 g_free(blk->buf);
 575                 g_free(blk);
 576             }
 577
 578             sector += nr_sectors;
 579             bmds->cur_dirty = sector;
 580             break;
 581         }
 582
 583         bdrv_dirty_bitmap_unlock(bmds->dirty_bitmap);
 584         sector += BDRV_SECTORS_PER_DIRTY_CHUNK;
 585         bmds->cur_dirty = sector;
 586     }
 587
 588     return (bmds->cur_dirty >= bmds->total_sectors);
 589
 590 error:
 591     trace_migration_block_save_device_dirty(sector);
 592     g_free(blk->buf);
 593     g_free(blk);
 594     return ret;
 595 }
 596
 597 /* Called with iothread lock taken.
 598  *
 599  * return value:
 600  * 0: too much data for max_downtime
 601  * 1: few enough data for max_downtime
 602 */
 603 static int blk_mig_save_dirty_block(QEMUFile *f, int is_async)
 604 {
 605     BlkMigDevState *bmds;
 606     int ret = 1;
 607
 608     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 609         aio_context_acquire(blk_get_aio_context(bmds->blk));
 610         ret = mig_save_device_dirty(f, bmds, is_async);
 611         aio_context_release(blk_get_aio_context(bmds->blk));
 612         if (ret <= 0) {
 613             break;
 614         }
 615     }
 616
 617     return ret;
 618 }
 619
 620 /* Called with no locks taken.  */
 621
 622 static int flush_blks(QEMUFile *f)
 623 {
 624     BlkMigBlock *blk;
 625     int ret = 0;
 626
 627     trace_migration_block_flush_blks("Enter", block_mig_state.submitted,
 628                                      block_mig_state.read_done,
 629                                      block_mig_state.transferred);
 630
 631     blk_mig_lock();
 632     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 633         if (migration_rate_exceeded(f)) {
 634             break;
 635         }
 636         if (blk->ret < 0) {
 637             ret = blk->ret;
 638             break;
 639         }
 640
 641         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 642         blk_mig_unlock();
 643         blk_send(f, blk);
 644         blk_mig_lock();
 645
 646         g_free(blk->buf);
 647         g_free(blk);
 648
 649         block_mig_state.read_done--;
 650         block_mig_state.transferred++;
 651         assert(block_mig_state.read_done >= 0);
 652     }
 653     blk_mig_unlock();
 654
 655     trace_migration_block_flush_blks("Exit", block_mig_state.submitted,
 656                                      block_mig_state.read_done,
 657                                      block_mig_state.transferred);
 658     return ret;
 659 }
 660
 661 /* Called with iothread lock taken.  */
 662
 663 static int64_t get_remaining_dirty(void)
 664 {
 665     BlkMigDevState *bmds;
 666     int64_t dirty = 0;
 667
 668     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 669         aio_context_acquire(blk_get_aio_context(bmds->blk));
 670         dirty += bdrv_get_dirty_count(bmds->dirty_bitmap);
 671         aio_context_release(blk_get_aio_context(bmds->blk));
 672     }
 673
 674     return dirty;
 675 }
 676
 677
 678
 679 /* Called with iothread lock taken.  */
 680 static void block_migration_cleanup_bmds(void)
 681 {
 682     BlkMigDevState *bmds;
 683     BlockDriverState *bs;
 684     AioContext *ctx;
 685
 686     unset_dirty_tracking();
 687
 688     while ((bmds = QSIMPLEQ_FIRST(&block_mig_state.bmds_list)) != NULL) {
 689         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.bmds_list, entry);
 690
 691         bs = blk_bs(bmds->blk);
 692         if (bs) {
 693             bdrv_op_unblock_all(bs, bmds->blocker);
 694         }
 695         error_free(bmds->blocker);
 696
 697         /* Save ctx, because bmds->blk can disappear during blk_unref.  */
 698         ctx = blk_get_aio_context(bmds->blk);
 699         aio_context_acquire(ctx);
 700         blk_unref(bmds->blk);
 701         aio_context_release(ctx);
 702
 703         g_free(bmds->blk_name);
 704         g_free(bmds->aio_bitmap);
 705         g_free(bmds);
 706     }
 707 }
 708
 709 /* Called with iothread lock taken.  */
 710 static void block_migration_cleanup(void *opaque)
 711 {
 712     BlkMigBlock *blk;
 713
 714     bdrv_drain_all();
 715
 716     block_migration_cleanup_bmds();
 717
 718     blk_mig_lock();
 719     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 720         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 721         g_free(blk->buf);
 722         g_free(blk);
 723     }
 724     blk_mig_unlock();
 725 }
 726
 727 static int block_save_setup(QEMUFile *f, void *opaque)
 728 {
 729     int ret;
 730
 731     trace_migration_block_save("setup", block_mig_state.submitted,
 732                                block_mig_state.transferred);
 733
 734     ret = init_blk_migration(f);
 735     if (ret < 0) {
 736         return ret;
 737     }
 738
 739     /* start track dirty blocks */
 740     ret = set_dirty_tracking();
 741     if (ret) {
 742         return ret;
 743     }
 744
 745     ret = flush_blks(f);
 746     blk_mig_reset_dirty_cursor();
 747     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 748
 749     return ret;
 750 }
 751
 752 static int block_save_iterate(QEMUFile *f, void *opaque)
 753 {
 754     int ret;
 755     uint64_t last_bytes = qemu_file_transferred_noflush(f);
 756
 757     trace_migration_block_save("iterate", block_mig_state.submitted,
 758                                block_mig_state.transferred);
 759
 760     ret = flush_blks(f);
 761     if (ret) {
 762         return ret;
 763     }
 764
 765     blk_mig_reset_dirty_cursor();
 766
 767     /* control the rate of transfer */
 768     blk_mig_lock();
 769     while (block_mig_state.read_done * BLK_MIG_BLOCK_SIZE <
 770            migration_rate_get() &&
 771            block_mig_state.submitted < MAX_PARALLEL_IO &&
 772            (block_mig_state.submitted + block_mig_state.read_done) <
 773            MAX_IO_BUFFERS) {
 774         blk_mig_unlock();
 775         if (block_mig_state.bulk_completed == 0) {
 776             /* first finish the bulk phase */
 777             if (blk_mig_save_bulked_block(f) == 0) {
 778                 /* finished saving bulk on all devices */
 779                 block_mig_state.bulk_completed = 1;
 780             }
 781             ret = 0;
 782         } else {
 783             /* Always called with iothread lock taken for
 784              * simplicity, block_save_complete also calls it.
 785              */
 786             qemu_mutex_lock_iothread();
 787             ret = blk_mig_save_dirty_block(f, 1);
 788             qemu_mutex_unlock_iothread();
 789         }
 790         if (ret < 0) {
 791             return ret;
 792         }
 793         blk_mig_lock();
 794         if (ret != 0) {
 795             /* no more dirty blocks */
 796             break;
 797         }
 798     }
 799     blk_mig_unlock();
 800
 801     ret = flush_blks(f);
 802     if (ret) {
 803         return ret;
 804     }
 805
 806     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 807     uint64_t delta_bytes = qemu_file_transferred_noflush(f) - last_bytes;
 808     return (delta_bytes > 0);
 809 }
 810
 811 /* Called with iothread lock taken.  */
 812
 813 static int block_save_complete(QEMUFile *f, void *opaque)
 814 {
 815     int ret;
 816
 817     trace_migration_block_save("complete", block_mig_state.submitted,
 818                                block_mig_state.transferred);
 819
 820     ret = flush_blks(f);
 821     if (ret) {
 822         return ret;
 823     }
 824
 825     blk_mig_reset_dirty_cursor();
 826
 827     /* we know for sure that save bulk is completed and
 828        all async read completed */
 829     blk_mig_lock();
 830     assert(block_mig_state.submitted == 0);
 831     blk_mig_unlock();
 832
 833     do {
 834         ret = blk_mig_save_dirty_block(f, 0);
 835         if (ret < 0) {
 836             return ret;
 837         }
 838     } while (ret == 0);
 839
 840     /* report completion */
 841     qemu_put_be64(f, (100 << BDRV_SECTOR_BITS) | BLK_MIG_FLAG_PROGRESS);
 842
 843     trace_migration_block_save_complete();
 844
 845     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 846
 847     /* Make sure that our BlockBackends are gone, so that the block driver
 848      * nodes can be inactivated. */
 849     block_migration_cleanup_bmds();
 850
 851     return 0;
 852 }
 853
 854 static void block_state_pending(void *opaque, uint64_t *must_precopy,
 855                                 uint64_t *can_postcopy)
 856 {
 857     /* Estimate pending number of bytes to send */
 858     uint64_t pending;
 859
 860     qemu_mutex_lock_iothread();
 861     pending = get_remaining_dirty();
 862     qemu_mutex_unlock_iothread();
 863
 864     blk_mig_lock();
 865     pending += block_mig_state.submitted * BLK_MIG_BLOCK_SIZE +
 866                block_mig_state.read_done * BLK_MIG_BLOCK_SIZE;
 867     blk_mig_unlock();
 868
 869     /* Report at least one block pending during bulk phase */
 870     if (!pending && !block_mig_state.bulk_completed) {
 871         pending = BLK_MIG_BLOCK_SIZE;
 872     }
 873
 874     trace_migration_block_state_pending(pending);
 875     /* We don't do postcopy */
 876     *must_precopy += pending;
 877 }
 878
 879 static int block_load(QEMUFile *f, void *opaque, int version_id)
 880 {
 881     static int banner_printed;
 882     int len, flags;
 883     char device_name[256];
 884     int64_t addr;
 885     BlockBackend *blk, *blk_prev = NULL;
 886     Error *local_err = NULL;
 887     uint8_t *buf;
 888     int64_t total_sectors = 0;
 889     int nr_sectors;
 890     int ret;
 891     BlockDriverInfo bdi;
 892     int cluster_size = BLK_MIG_BLOCK_SIZE;
 893
 894     do {
 895         addr = qemu_get_be64(f);
 896
 897         flags = addr & (BDRV_SECTOR_SIZE - 1);
 898         addr >>= BDRV_SECTOR_BITS;
 899
 900         if (flags & BLK_MIG_FLAG_DEVICE_BLOCK) {
 901             /* get device name */
 902             len = qemu_get_byte(f);
 903             qemu_get_buffer(f, (uint8_t *)device_name, len);
 904             device_name[len] = '\0';
 905
 906             blk = blk_by_name(device_name);
 907             if (!blk) {
 908                 fprintf(stderr, "Error unknown block device %s\n",
 909                         device_name);
 910                 return -EINVAL;
 911             }
 912
 913             if (blk != blk_prev) {
 914                 blk_prev = blk;
 915                 total_sectors = blk_nb_sectors(blk);
 916                 if (total_sectors <= 0) {
 917                     error_report("Error getting length of block device %s",
 918                                  device_name);
 919                     return -EINVAL;
 920                 }
 921
 922                 blk_activate(blk, &local_err);
 923                 if (local_err) {
 924                     error_report_err(local_err);
 925                     return -EINVAL;
 926                 }
 927
 928                 ret = bdrv_get_info(blk_bs(blk), &bdi);
 929                 if (ret == 0 && bdi.cluster_size > 0 &&
 930                     bdi.cluster_size <= BLK_MIG_BLOCK_SIZE &&
 931                     BLK_MIG_BLOCK_SIZE % bdi.cluster_size == 0) {
 932                     cluster_size = bdi.cluster_size;
 933                 } else {
 934                     cluster_size = BLK_MIG_BLOCK_SIZE;
 935                 }
 936             }
 937
 938             if (total_sectors - addr < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 939                 nr_sectors = total_sectors - addr;
 940             } else {
 941                 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 942             }
 943
 944             if (flags & BLK_MIG_FLAG_ZERO_BLOCK) {
 945                 ret = blk_pwrite_zeroes(blk, addr * BDRV_SECTOR_SIZE,
 946                                         nr_sectors * BDRV_SECTOR_SIZE,
 947                                         BDRV_REQ_MAY_UNMAP);
 948             } else {
 949                 int i;
 950                 int64_t cur_addr;
 951                 uint8_t *cur_buf;
 952
 953                 buf = g_malloc(BLK_MIG_BLOCK_SIZE);
 954                 qemu_get_buffer(f, buf, BLK_MIG_BLOCK_SIZE);
 955                 for (i = 0; i < BLK_MIG_BLOCK_SIZE / cluster_size; i++) {
 956                     cur_addr = addr * BDRV_SECTOR_SIZE + i * cluster_size;
 957                     cur_buf = buf + i * cluster_size;
 958
 959                     if ((!block_mig_state.zero_blocks ||
 960                         cluster_size < BLK_MIG_BLOCK_SIZE) &&
 961                         buffer_is_zero(cur_buf, cluster_size)) {
 962                         ret = blk_pwrite_zeroes(blk, cur_addr,
 963                                                 cluster_size,
 964                                                 BDRV_REQ_MAY_UNMAP);
 965                     } else {
 966                         ret = blk_pwrite(blk, cur_addr, cluster_size, cur_buf,
 967                                          0);
 968                     }
 969                     if (ret < 0) {
 970                         break;
 971                     }
 972                 }
 973                 g_free(buf);
 974             }
 975
 976             if (ret < 0) {
 977                 return ret;
 978             }
 979         } else if (flags & BLK_MIG_FLAG_PROGRESS) {
 980             if (!banner_printed) {
 981                 printf("Receiving block device images\n");
 982                 banner_printed = 1;
 983             }
 984             printf("Completed %d %%%c", (int)addr,
 985                    (addr == 100) ? '\n' : '\r');
 986             fflush(stdout);
 987         } else if (!(flags & BLK_MIG_FLAG_EOS)) {
 988             fprintf(stderr, "Unknown block migration flags: 0x%x\n", flags);
 989             return -EINVAL;
 990         }
 991         ret = qemu_file_get_error(f);
 992         if (ret != 0) {
 993             return ret;
 994         }
 995     } while (!(flags & BLK_MIG_FLAG_EOS));
 996
 997     return 0;
 998 }
 999
1000 static bool block_is_active(void *opaque)
1001 {
1002     return migrate_block();
1003 }
1004
1005 static SaveVMHandlers savevm_block_handlers = {
1006     .save_setup = block_save_setup,
1007     .save_live_iterate = block_save_iterate,
1008     .save_live_complete_precopy = block_save_complete,
1009     .state_pending_exact = block_state_pending,
1010     .state_pending_estimate = block_state_pending,
1011     .load_state = block_load,
1012     .save_cleanup = block_migration_cleanup,
1013     .is_active = block_is_active,
1014 };
1015
1016 void blk_mig_init(void)
1017 {
1018     QSIMPLEQ_INIT(&block_mig_state.bmds_list);
1019     QSIMPLEQ_INIT(&block_mig_state.blk_list);
1020     qemu_mutex_init(&block_mig_state.lock);
1021
1022     register_savevm_live("block", 0, 1, &savevm_block_handlers,
1023                          &block_mig_state);
1024 }