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         bdrv_release_dirty_bitmap(bmds->dirty_bitmap);
 372     }
 373 }
 374
 375 static int init_blk_migration(QEMUFile *f)
 376 {
 377     BlockDriverState *bs;
 378     BlkMigDevState *bmds;
 379     int64_t sectors;
 380     BdrvNextIterator it;
 381     int i, num_bs = 0;
 382     struct {
 383         BlkMigDevState *bmds;
 384         BlockDriverState *bs;
 385     } *bmds_bs;
 386     Error *local_err = NULL;
 387     int ret;
 388
 389     block_mig_state.submitted = 0;
 390     block_mig_state.read_done = 0;
 391     block_mig_state.transferred = 0;
 392     block_mig_state.total_sector_sum = 0;
 393     block_mig_state.prev_progress = -1;
 394     block_mig_state.bulk_completed = 0;
 395     block_mig_state.zero_blocks = migrate_zero_blocks();
 396
 397     for (bs = bdrv_first(&it); bs; bs = bdrv_next(&it)) {
 398         num_bs++;
 399     }
 400     bmds_bs = g_malloc0(num_bs * sizeof(*bmds_bs));
 401
 402     for (i = 0, bs = bdrv_first(&it); bs; bs = bdrv_next(&it), i++) {
 403         if (bdrv_is_read_only(bs)) {
 404             continue;
 405         }
 406
 407         sectors = bdrv_nb_sectors(bs);
 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         BlkMigDevState *bmds = bmds_bs[i].bmds;
 442         BlockDriverState *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 iothread lock and AioContext 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 iothread lock 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         aio_context_acquire(blk_get_aio_context(bmds->blk));
 606         ret = mig_save_device_dirty(f, bmds, is_async);
 607         aio_context_release(blk_get_aio_context(bmds->blk));
 608         if (ret <= 0) {
 609             break;
 610         }
 611     }
 612
 613     return ret;
 614 }
 615
 616 /* Called with no locks taken.  */
 617
 618 static int flush_blks(QEMUFile *f)
 619 {
 620     BlkMigBlock *blk;
 621     int ret = 0;
 622
 623     trace_migration_block_flush_blks("Enter", block_mig_state.submitted,
 624                                      block_mig_state.read_done,
 625                                      block_mig_state.transferred);
 626
 627     blk_mig_lock();
 628     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 629         if (migration_rate_exceeded(f)) {
 630             break;
 631         }
 632         if (blk->ret < 0) {
 633             ret = blk->ret;
 634             break;
 635         }
 636
 637         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 638         blk_mig_unlock();
 639         blk_send(f, blk);
 640         blk_mig_lock();
 641
 642         g_free(blk->buf);
 643         g_free(blk);
 644
 645         block_mig_state.read_done--;
 646         block_mig_state.transferred++;
 647         assert(block_mig_state.read_done >= 0);
 648     }
 649     blk_mig_unlock();
 650
 651     trace_migration_block_flush_blks("Exit", block_mig_state.submitted,
 652                                      block_mig_state.read_done,
 653                                      block_mig_state.transferred);
 654     return ret;
 655 }
 656
 657 /* Called with iothread lock taken.  */
 658
 659 static int64_t get_remaining_dirty(void)
 660 {
 661     BlkMigDevState *bmds;
 662     int64_t dirty = 0;
 663
 664     QSIMPLEQ_FOREACH(bmds, &block_mig_state.bmds_list, entry) {
 665         aio_context_acquire(blk_get_aio_context(bmds->blk));
 666         dirty += bdrv_get_dirty_count(bmds->dirty_bitmap);
 667         aio_context_release(blk_get_aio_context(bmds->blk));
 668     }
 669
 670     return dirty;
 671 }
 672
 673
 674
 675 /* Called with iothread lock taken.  */
 676 static void block_migration_cleanup_bmds(void)
 677 {
 678     BlkMigDevState *bmds;
 679     AioContext *ctx;
 680
 681     unset_dirty_tracking();
 682
 683     while ((bmds = QSIMPLEQ_FIRST(&block_mig_state.bmds_list)) != NULL) {
 684         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.bmds_list, entry);
 685         bdrv_op_unblock_all(blk_bs(bmds->blk), bmds->blocker);
 686         error_free(bmds->blocker);
 687
 688         /* Save ctx, because bmds->blk can disappear during blk_unref.  */
 689         ctx = blk_get_aio_context(bmds->blk);
 690         aio_context_acquire(ctx);
 691         blk_unref(bmds->blk);
 692         aio_context_release(ctx);
 693
 694         g_free(bmds->blk_name);
 695         g_free(bmds->aio_bitmap);
 696         g_free(bmds);
 697     }
 698 }
 699
 700 /* Called with iothread lock taken.  */
 701 static void block_migration_cleanup(void *opaque)
 702 {
 703     BlkMigBlock *blk;
 704
 705     bdrv_drain_all();
 706
 707     block_migration_cleanup_bmds();
 708
 709     blk_mig_lock();
 710     while ((blk = QSIMPLEQ_FIRST(&block_mig_state.blk_list)) != NULL) {
 711         QSIMPLEQ_REMOVE_HEAD(&block_mig_state.blk_list, entry);
 712         g_free(blk->buf);
 713         g_free(blk);
 714     }
 715     blk_mig_unlock();
 716 }
 717
 718 static int block_save_setup(QEMUFile *f, void *opaque)
 719 {
 720     int ret;
 721
 722     trace_migration_block_save("setup", block_mig_state.submitted,
 723                                block_mig_state.transferred);
 724
 725     qemu_mutex_lock_iothread();
 726     ret = init_blk_migration(f);
 727     if (ret < 0) {
 728         qemu_mutex_unlock_iothread();
 729         return ret;
 730     }
 731
 732     /* start track dirty blocks */
 733     ret = set_dirty_tracking();
 734
 735     qemu_mutex_unlock_iothread();
 736
 737     if (ret) {
 738         return ret;
 739     }
 740
 741     ret = flush_blks(f);
 742     blk_mig_reset_dirty_cursor();
 743     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 744
 745     return ret;
 746 }
 747
 748 static int block_save_iterate(QEMUFile *f, void *opaque)
 749 {
 750     int ret;
 751     uint64_t last_bytes = qemu_file_transferred(f);
 752
 753     trace_migration_block_save("iterate", block_mig_state.submitted,
 754                                block_mig_state.transferred);
 755
 756     ret = flush_blks(f);
 757     if (ret) {
 758         return ret;
 759     }
 760
 761     blk_mig_reset_dirty_cursor();
 762
 763     /* control the rate of transfer */
 764     blk_mig_lock();
 765     while (block_mig_state.read_done * BLK_MIG_BLOCK_SIZE <
 766            migration_rate_get() &&
 767            block_mig_state.submitted < MAX_PARALLEL_IO &&
 768            (block_mig_state.submitted + block_mig_state.read_done) <
 769            MAX_IO_BUFFERS) {
 770         blk_mig_unlock();
 771         if (block_mig_state.bulk_completed == 0) {
 772             /* first finish the bulk phase */
 773             if (blk_mig_save_bulked_block(f) == 0) {
 774                 /* finished saving bulk on all devices */
 775                 block_mig_state.bulk_completed = 1;
 776             }
 777             ret = 0;
 778         } else {
 779             /* Always called with iothread lock taken for
 780              * simplicity, block_save_complete also calls it.
 781              */
 782             qemu_mutex_lock_iothread();
 783             ret = blk_mig_save_dirty_block(f, 1);
 784             qemu_mutex_unlock_iothread();
 785         }
 786         if (ret < 0) {
 787             return ret;
 788         }
 789         blk_mig_lock();
 790         if (ret != 0) {
 791             /* no more dirty blocks */
 792             break;
 793         }
 794     }
 795     blk_mig_unlock();
 796
 797     ret = flush_blks(f);
 798     if (ret) {
 799         return ret;
 800     }
 801
 802     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 803     uint64_t delta_bytes = qemu_file_transferred(f) - last_bytes;
 804     return (delta_bytes > 0);
 805 }
 806
 807 /* Called with iothread lock taken.  */
 808
 809 static int block_save_complete(QEMUFile *f, void *opaque)
 810 {
 811     int ret;
 812
 813     trace_migration_block_save("complete", block_mig_state.submitted,
 814                                block_mig_state.transferred);
 815
 816     ret = flush_blks(f);
 817     if (ret) {
 818         return ret;
 819     }
 820
 821     blk_mig_reset_dirty_cursor();
 822
 823     /* we know for sure that save bulk is completed and
 824        all async read completed */
 825     blk_mig_lock();
 826     assert(block_mig_state.submitted == 0);
 827     blk_mig_unlock();
 828
 829     do {
 830         ret = blk_mig_save_dirty_block(f, 0);
 831         if (ret < 0) {
 832             return ret;
 833         }
 834     } while (ret == 0);
 835
 836     /* report completion */
 837     qemu_put_be64(f, (100 << BDRV_SECTOR_BITS) | BLK_MIG_FLAG_PROGRESS);
 838
 839     trace_migration_block_save_complete();
 840
 841     qemu_put_be64(f, BLK_MIG_FLAG_EOS);
 842
 843     /* Make sure that our BlockBackends are gone, so that the block driver
 844      * nodes can be inactivated. */
 845     block_migration_cleanup_bmds();
 846
 847     return 0;
 848 }
 849
 850 static void block_state_pending(void *opaque, uint64_t *must_precopy,
 851                                 uint64_t *can_postcopy)
 852 {
 853     /* Estimate pending number of bytes to send */
 854     uint64_t pending;
 855
 856     qemu_mutex_lock_iothread();
 857     pending = get_remaining_dirty();
 858     qemu_mutex_unlock_iothread();
 859
 860     blk_mig_lock();
 861     pending += block_mig_state.submitted * BLK_MIG_BLOCK_SIZE +
 862                block_mig_state.read_done * BLK_MIG_BLOCK_SIZE;
 863     blk_mig_unlock();
 864
 865     /* Report at least one block pending during bulk phase */
 866     if (!pending && !block_mig_state.bulk_completed) {
 867         pending = BLK_MIG_BLOCK_SIZE;
 868     }
 869
 870     trace_migration_block_state_pending(pending);
 871     /* We don't do postcopy */
 872     *must_precopy += pending;
 873 }
 874
 875 static int block_load(QEMUFile *f, void *opaque, int version_id)
 876 {
 877     static int banner_printed;
 878     int len, flags;
 879     char device_name[256];
 880     int64_t addr;
 881     BlockBackend *blk, *blk_prev = NULL;
 882     Error *local_err = NULL;
 883     uint8_t *buf;
 884     int64_t total_sectors = 0;
 885     int nr_sectors;
 886     int ret;
 887     BlockDriverInfo bdi;
 888     int cluster_size = BLK_MIG_BLOCK_SIZE;
 889
 890     do {
 891         addr = qemu_get_be64(f);
 892
 893         flags = addr & (BDRV_SECTOR_SIZE - 1);
 894         addr >>= BDRV_SECTOR_BITS;
 895
 896         if (flags & BLK_MIG_FLAG_DEVICE_BLOCK) {
 897             /* get device name */
 898             len = qemu_get_byte(f);
 899             qemu_get_buffer(f, (uint8_t *)device_name, len);
 900             device_name[len] = '\0';
 901
 902             blk = blk_by_name(device_name);
 903             if (!blk) {
 904                 fprintf(stderr, "Error unknown block device %s\n",
 905                         device_name);
 906                 return -EINVAL;
 907             }
 908
 909             if (blk != blk_prev) {
 910                 blk_prev = blk;
 911                 total_sectors = blk_nb_sectors(blk);
 912                 if (total_sectors <= 0) {
 913                     error_report("Error getting length of block device %s",
 914                                  device_name);
 915                     return -EINVAL;
 916                 }
 917
 918                 blk_activate(blk, &local_err);
 919                 if (local_err) {
 920                     error_report_err(local_err);
 921                     return -EINVAL;
 922                 }
 923
 924                 ret = bdrv_get_info(blk_bs(blk), &bdi);
 925                 if (ret == 0 && bdi.cluster_size > 0 &&
 926                     bdi.cluster_size <= BLK_MIG_BLOCK_SIZE &&
 927                     BLK_MIG_BLOCK_SIZE % bdi.cluster_size == 0) {
 928                     cluster_size = bdi.cluster_size;
 929                 } else {
 930                     cluster_size = BLK_MIG_BLOCK_SIZE;
 931                 }
 932             }
 933
 934             if (total_sectors - addr < BDRV_SECTORS_PER_DIRTY_CHUNK) {
 935                 nr_sectors = total_sectors - addr;
 936             } else {
 937                 nr_sectors = BDRV_SECTORS_PER_DIRTY_CHUNK;
 938             }
 939
 940             if (flags & BLK_MIG_FLAG_ZERO_BLOCK) {
 941                 ret = blk_pwrite_zeroes(blk, addr * BDRV_SECTOR_SIZE,
 942                                         nr_sectors * BDRV_SECTOR_SIZE,
 943                                         BDRV_REQ_MAY_UNMAP);
 944             } else {
 945                 int i;
 946                 int64_t cur_addr;
 947                 uint8_t *cur_buf;
 948
 949                 buf = g_malloc(BLK_MIG_BLOCK_SIZE);
 950                 qemu_get_buffer(f, buf, BLK_MIG_BLOCK_SIZE);
 951                 for (i = 0; i < BLK_MIG_BLOCK_SIZE / cluster_size; i++) {
 952                     cur_addr = addr * BDRV_SECTOR_SIZE + i * cluster_size;
 953                     cur_buf = buf + i * cluster_size;
 954
 955                     if ((!block_mig_state.zero_blocks ||
 956                         cluster_size < BLK_MIG_BLOCK_SIZE) &&
 957                         buffer_is_zero(cur_buf, cluster_size)) {
 958                         ret = blk_pwrite_zeroes(blk, cur_addr,
 959                                                 cluster_size,
 960                                                 BDRV_REQ_MAY_UNMAP);
 961                     } else {
 962                         ret = blk_pwrite(blk, cur_addr, cluster_size, cur_buf,
 963                                          0);
 964                     }
 965                     if (ret < 0) {
 966                         break;
 967                     }
 968                 }
 969                 g_free(buf);
 970             }
 971
 972             if (ret < 0) {
 973                 return ret;
 974             }
 975         } else if (flags & BLK_MIG_FLAG_PROGRESS) {
 976             if (!banner_printed) {
 977                 printf("Receiving block device images\n");
 978                 banner_printed = 1;
 979             }
 980             printf("Completed %d %%%c", (int)addr,
 981                    (addr == 100) ? '\n' : '\r');
 982             fflush(stdout);
 983         } else if (!(flags & BLK_MIG_FLAG_EOS)) {
 984             fprintf(stderr, "Unknown block migration flags: 0x%x\n", flags);
 985             return -EINVAL;
 986         }
 987         ret = qemu_file_get_error(f);
 988         if (ret != 0) {
 989             return ret;
 990         }
 991     } while (!(flags & BLK_MIG_FLAG_EOS));
 992
 993     return 0;
 994 }
 995
 996 static bool block_is_active(void *opaque)
 997 {
 998     return migrate_block();
 999 }
1000
1001 static SaveVMHandlers savevm_block_handlers = {
1002     .save_setup = block_save_setup,
1003     .save_live_iterate = block_save_iterate,
1004     .save_live_complete_precopy = block_save_complete,
1005     .state_pending_exact = block_state_pending,
1006     .state_pending_estimate = block_state_pending,
1007     .load_state = block_load,
1008     .save_cleanup = block_migration_cleanup,
1009     .is_active = block_is_active,
1010 };
1011
1012 void blk_mig_init(void)
1013 {
1014     QSIMPLEQ_INIT(&block_mig_state.bmds_list);
1015     QSIMPLEQ_INIT(&block_mig_state.blk_list);
1016     qemu_mutex_init(&block_mig_state.lock);
1017
1018     register_savevm_live("block", 0, 1, &savevm_block_handlers,
1019                          &block_mig_state);
1020 }