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