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